1	// SPDX-License-Identifier: GPL-2.0-only
2	/* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
3	* Copyright (c) 2016 Facebook
4	* Copyright (c) 2018 Covalent IO, Inc. http://covalent.io
5	*/
6	#include <uapi/linux/btf.h>
7	#include <linux/bpf-cgroup.h>
8	#include <linux/kernel.h>
9	#include <linux/types.h>
10	#include <linux/slab.h>
11	#include <linux/bpf.h>
12	#include <linux/btf.h>
13	#include <linux/bpf_verifier.h>
14	#include <linux/filter.h>
15	#include <net/netlink.h>
16	#include <linux/file.h>
17	#include <linux/vmalloc.h>
18	#include <linux/stringify.h>
19	#include <linux/bsearch.h>
20	#include <linux/sort.h>
21	#include <linux/perf_event.h>
22	#include <linux/ctype.h>
23	#include <linux/error-injection.h>
24	#include <linux/bpf_lsm.h>
25	#include <linux/btf_ids.h>
26	#include <linux/poison.h>
27	#include <linux/module.h>
28	#include <linux/cpumask.h>
29	#include <linux/bpf_mem_alloc.h>
30	#include <net/xdp.h>
31	#include <linux/trace_events.h>
32	#include <linux/kallsyms.h>
33
34	#include "disasm.h"
35
36	static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
37	#define BPF_PROG_TYPE(_id, _name, prog_ctx_type, kern_ctx_type) \
38	[_id] = & _name ## _verifier_ops,
39	#define BPF_MAP_TYPE(_id, _ops)
40	#define BPF_LINK_TYPE(_id, _name)
41	#include <linux/bpf_types.h>
42	#undef BPF_PROG_TYPE
43	#undef BPF_MAP_TYPE
44	#undef BPF_LINK_TYPE
45	};
46
47	enum bpf_features {
48	BPF_FEAT_RDONLY_CAST_TO_VOID = 0,
49	BPF_FEAT_STREAMS = 1,
50	__MAX_BPF_FEAT,
51	};
52
53	struct bpf_mem_alloc bpf_global_percpu_ma;
54	static bool bpf_global_percpu_ma_set;
55
56	/* bpf_check() is a static code analyzer that walks eBPF program
57	* instruction by instruction and updates register/stack state.
58	* All paths of conditional branches are analyzed until 'bpf_exit' insn.
59	*
60	* The first pass is depth-first-search to check that the program is a DAG.
61	* It rejects the following programs:
62	* - larger than BPF_MAXINSNS insns
63	* - if loop is present (detected via back-edge)
64	* - unreachable insns exist (shouldn't be a forest. program = one function)
65	* - out of bounds or malformed jumps
66	* The second pass is all possible path descent from the 1st insn.
67	* Since it's analyzing all paths through the program, the length of the
68	* analysis is limited to 64k insn, which may be hit even if total number of
69	* insn is less then 4K, but there are too many branches that change stack/regs.
70	* Number of 'branches to be analyzed' is limited to 1k
71	*
72	* On entry to each instruction, each register has a type, and the instruction
73	* changes the types of the registers depending on instruction semantics.
74	* If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
75	* copied to R1.
76	*
77	* All registers are 64-bit.
78	* R0 - return register
79	* R1-R5 argument passing registers
80	* R6-R9 callee saved registers
81	* R10 - frame pointer read-only
82	*
83	* At the start of BPF program the register R1 contains a pointer to bpf_context
84	* and has type PTR_TO_CTX.
85	*
86	* Verifier tracks arithmetic operations on pointers in case:
87	* BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
88	* BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
89	* 1st insn copies R10 (which has FRAME_PTR) type into R1
90	* and 2nd arithmetic instruction is pattern matched to recognize
91	* that it wants to construct a pointer to some element within stack.
92	* So after 2nd insn, the register R1 has type PTR_TO_STACK
93	* (and -20 constant is saved for further stack bounds checking).
94	* Meaning that this reg is a pointer to stack plus known immediate constant.
95	*
96	* Most of the time the registers have SCALAR_VALUE type, which
97	* means the register has some value, but it's not a valid pointer.
98	* (like pointer plus pointer becomes SCALAR_VALUE type)
99	*
100	* When verifier sees load or store instructions the type of base register
101	* can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK, PTR_TO_SOCKET. These are
102	* four pointer types recognized by check_mem_access() function.
103	*
104	* PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
105	* and the range of [ptr, ptr + map's value_size) is accessible.
106	*
107	* registers used to pass values to function calls are checked against
108	* function argument constraints.
109	*
110	* ARG_PTR_TO_MAP_KEY is one of such argument constraints.
111	* It means that the register type passed to this function must be
112	* PTR_TO_STACK and it will be used inside the function as
113	* 'pointer to map element key'
114	*
115	* For example the argument constraints for bpf_map_lookup_elem():
116	* .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
117	* .arg1_type = ARG_CONST_MAP_PTR,
118	* .arg2_type = ARG_PTR_TO_MAP_KEY,
119	*
120	* ret_type says that this function returns 'pointer to map elem value or null'
121	* function expects 1st argument to be a const pointer to 'struct bpf_map' and
122	* 2nd argument should be a pointer to stack, which will be used inside
123	* the helper function as a pointer to map element key.
124	*
125	* On the kernel side the helper function looks like:
126	* u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
127	* {
128	* struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
129	* void *key = (void *) (unsigned long) r2;
130	* void *value;
131	*
132	* here kernel can access 'key' and 'map' pointers safely, knowing that
133	* [key, key + map->key_size) bytes are valid and were initialized on
134	* the stack of eBPF program.
135	* }
136	*
137	* Corresponding eBPF program may look like:
138	* BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
139	* BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
140	* BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
141	* BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
142	* here verifier looks at prototype of map_lookup_elem() and sees:
143	* .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
144	* Now verifier knows that this map has key of R1->map_ptr->key_size bytes
145	*
146	* Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
147	* Now verifier checks that [R2, R2 + map's key_size) are within stack limits
148	* and were initialized prior to this call.
149	* If it's ok, then verifier allows this BPF_CALL insn and looks at
150	* .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
151	* R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
152	* returns either pointer to map value or NULL.
153	*
154	* When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
155	* insn, the register holding that pointer in the true branch changes state to
156	* PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
157	* branch. See check_cond_jmp_op().
158	*
159	* After the call R0 is set to return type of the function and registers R1-R5
160	* are set to NOT_INIT to indicate that they are no longer readable.
161	*
162	* The following reference types represent a potential reference to a kernel
163	* resource which, after first being allocated, must be checked and freed by
164	* the BPF program:
165	* - PTR_TO_SOCKET_OR_NULL, PTR_TO_SOCKET
166	*
167	* When the verifier sees a helper call return a reference type, it allocates a
168	* pointer id for the reference and stores it in the current function state.
169	* Similar to the way that PTR_TO_MAP_VALUE_OR_NULL is converted into
170	* PTR_TO_MAP_VALUE, PTR_TO_SOCKET_OR_NULL becomes PTR_TO_SOCKET when the type
171	* passes through a NULL-check conditional. For the branch wherein the state is
172	* changed to CONST_IMM, the verifier releases the reference.
173	*
174	* For each helper function that allocates a reference, such as
175	* bpf_sk_lookup_tcp(), there is a corresponding release function, such as
176	* bpf_sk_release(). When a reference type passes into the release function,
177	* the verifier also releases the reference. If any unchecked or unreleased
178	* reference remains at the end of the program, the verifier rejects it.
179	*/
180
181	/* verifier_state + insn_idx are pushed to stack when branch is encountered */
182	struct bpf_verifier_stack_elem {
183	/* verifier state is 'st'
184	* before processing instruction 'insn_idx'
185	* and after processing instruction 'prev_insn_idx'
186	*/
187	struct bpf_verifier_state st;
188	int insn_idx;
189	int prev_insn_idx;
190	struct bpf_verifier_stack_elem *next;
191	/* length of verifier log at the time this state was pushed on stack */
192	u32 log_pos;
193	};
194
195	#define BPF_COMPLEXITY_LIMIT_JMP_SEQ 8192
196	#define BPF_COMPLEXITY_LIMIT_STATES 64
197
198	#define BPF_MAP_KEY_POISON (1ULL << 63)
199	#define BPF_MAP_KEY_SEEN (1ULL << 62)
200
201	#define BPF_GLOBAL_PERCPU_MA_MAX_SIZE 512
202
203	#define BPF_PRIV_STACK_MIN_SIZE 64
204
205	static int acquire_reference(struct bpf_verifier_env *env, int insn_idx);
206	static int release_reference_nomark(struct bpf_verifier_state *state, int ref_obj_id);
207	static int release_reference(struct bpf_verifier_env *env, int ref_obj_id);
208	static void invalidate_non_owning_refs(struct bpf_verifier_env *env);
209	static bool in_rbtree_lock_required_cb(struct bpf_verifier_env *env);
210	static int ref_set_non_owning(struct bpf_verifier_env *env,
211	struct bpf_reg_state *reg);
212	static bool is_trusted_reg(const struct bpf_reg_state *reg);
213
214	static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
215	{
216	return aux->map_ptr_state.poison;
217	}
218
219	static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
220	{
221	return aux->map_ptr_state.unpriv;
222	}
223
224	static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
225	struct bpf_map *map,
226	bool unpriv, bool poison)
227	{
228	unpriv |= bpf_map_ptr_unpriv(aux);
229	aux->map_ptr_state.unpriv = unpriv;
230	aux->map_ptr_state.poison = poison;
231	aux->map_ptr_state.map_ptr = map;
232	}
233
234	static bool bpf_map_key_poisoned(const struct bpf_insn_aux_data *aux)
235	{
236	return aux->map_key_state & BPF_MAP_KEY_POISON;
237	}
238
239	static bool bpf_map_key_unseen(const struct bpf_insn_aux_data *aux)
240	{
241	return !(aux->map_key_state & BPF_MAP_KEY_SEEN);
242	}
243
244	static u64 bpf_map_key_immediate(const struct bpf_insn_aux_data *aux)
245	{
246	return aux->map_key_state & ~(BPF_MAP_KEY_SEEN | BPF_MAP_KEY_POISON);
247	}
248
249	static void bpf_map_key_store(struct bpf_insn_aux_data *aux, u64 state)
250	{
251	bool poisoned = bpf_map_key_poisoned(aux);
252
253	aux->map_key_state = state | BPF_MAP_KEY_SEEN |
254	(poisoned ? BPF_MAP_KEY_POISON : 0ULL);
255	}
256
257	static bool bpf_helper_call(const struct bpf_insn *insn)
258	{
259	return insn->code == (BPF_JMP | BPF_CALL) &&
260	insn->src_reg == 0;
261	}
262
263	static bool bpf_pseudo_call(const struct bpf_insn *insn)
264	{
265	return insn->code == (BPF_JMP | BPF_CALL) &&
266	insn->src_reg == BPF_PSEUDO_CALL;
267	}
268
269	static bool bpf_pseudo_kfunc_call(const struct bpf_insn *insn)
270	{
271	return insn->code == (BPF_JMP | BPF_CALL) &&
272	insn->src_reg == BPF_PSEUDO_KFUNC_CALL;
273	}
274
275	struct bpf_call_arg_meta {
276	struct bpf_map *map_ptr;
277	bool raw_mode;
278	bool pkt_access;
279	u8 release_regno;
280	int regno;
281	int access_size;
282	int mem_size;
283	u64 msize_max_value;
284	int ref_obj_id;
285	int dynptr_id;
286	int map_uid;
287	int func_id;
288	struct btf *btf;
289	u32 btf_id;
290	struct btf *ret_btf;
291	u32 ret_btf_id;
292	u32 subprogno;
293	struct btf_field *kptr_field;
294	s64 const_map_key;
295	};
296
297	struct bpf_kfunc_call_arg_meta {
298	/* In parameters */
299	struct btf *btf;
300	u32 func_id;
301	u32 kfunc_flags;
302	const struct btf_type *func_proto;
303	const char *func_name;
304	/* Out parameters */
305	u32 ref_obj_id;
306	u8 release_regno;
307	bool r0_rdonly;
308	u32 ret_btf_id;
309	u64 r0_size;
310	u32 subprogno;
311	struct {
312	u64 value;
313	bool found;
314	} arg_constant;
315
316	/* arg_{btf,btf_id,owning_ref} are used by kfunc-specific handling,
317	* generally to pass info about user-defined local kptr types to later
318	* verification logic
319	* bpf_obj_drop/bpf_percpu_obj_drop
320	* Record the local kptr type to be drop'd
321	* bpf_refcount_acquire (via KF_ARG_PTR_TO_REFCOUNTED_KPTR arg type)
322	* Record the local kptr type to be refcount_incr'd and use
323	* arg_owning_ref to determine whether refcount_acquire should be
324	* fallible
325	*/
326	struct btf *arg_btf;
327	u32 arg_btf_id;
328	bool arg_owning_ref;
329	bool arg_prog;
330
331	struct {
332	struct btf_field *field;
333	} arg_list_head;
334	struct {
335	struct btf_field *field;
336	} arg_rbtree_root;
337	struct {
338	enum bpf_dynptr_type type;
339	u32 id;
340	u32 ref_obj_id;
341	} initialized_dynptr;
342	struct {
343	u8 spi;
344	u8 frameno;
345	} iter;
346	struct {
347	struct bpf_map *ptr;
348	int uid;
349	} map;
350	u64 mem_size;
351	};
352
353	struct btf *btf_vmlinux;
354
355	static const char *btf_type_name(const struct btf *btf, u32 id)
356	{
357	return btf_name_by_offset(btf, offset: btf_type_by_id(btf, type_id: id)->name_off);
358	}
359
360	static DEFINE_MUTEX(bpf_verifier_lock);
361	static DEFINE_MUTEX(bpf_percpu_ma_lock);
362
363	__printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
364	{
365	struct bpf_verifier_env *env = private_data;
366	va_list args;
367
368	if (!bpf_verifier_log_needed(log: &env->log))
369	return;
370
371	va_start(args, fmt);
372	bpf_verifier_vlog(log: &env->log, fmt, args);
373	va_end(args);
374	}
375
376	static void verbose_invalid_scalar(struct bpf_verifier_env *env,
377	struct bpf_reg_state *reg,
378	struct bpf_retval_range range, const char *ctx,
379	const char *reg_name)
380	{
381	bool unknown = true;
382
383	verbose(private_data: env, fmt: "%s the register %s has", ctx, reg_name);
384	if (reg->smin_value > S64_MIN) {
385	verbose(private_data: env, fmt: " smin=%lld", reg->smin_value);
386	unknown = false;
387	}
388	if (reg->smax_value < S64_MAX) {
389	verbose(private_data: env, fmt: " smax=%lld", reg->smax_value);
390	unknown = false;
391	}
392	if (unknown)
393	verbose(private_data: env, fmt: " unknown scalar value");
394	verbose(private_data: env, fmt: " should have been in [%d, %d]\n", range.minval, range.maxval);
395	}
396
397	static bool reg_not_null(const struct bpf_reg_state *reg)
398	{
399	enum bpf_reg_type type;
400
401	type = reg->type;
402	if (type_may_be_null(type))
403	return false;
404
405	type = base_type(type);
406	return type == PTR_TO_SOCKET ||
407	type == PTR_TO_TCP_SOCK ||
408	type == PTR_TO_MAP_VALUE ||
409	type == PTR_TO_MAP_KEY ||
410	type == PTR_TO_SOCK_COMMON ||
411	(type == PTR_TO_BTF_ID && is_trusted_reg(reg)) ||
412	(type == PTR_TO_MEM && !(reg->type & PTR_UNTRUSTED)) ||
413	type == CONST_PTR_TO_MAP;
414	}
415
416	static struct btf_record *reg_btf_record(const struct bpf_reg_state *reg)
417	{
418	struct btf_record *rec = NULL;
419	struct btf_struct_meta *meta;
420
421	if (reg->type == PTR_TO_MAP_VALUE) {
422	rec = reg->map_ptr->record;
423	} else if (type_is_ptr_alloc_obj(type: reg->type)) {
424	meta = btf_find_struct_meta(btf: reg->btf, btf_id: reg->btf_id);
425	if (meta)
426	rec = meta->record;
427	}
428	return rec;
429	}
430
431	static bool subprog_is_global(const struct bpf_verifier_env *env, int subprog)
432	{
433	struct bpf_func_info_aux *aux = env->prog->aux->func_info_aux;
434
435	return aux && aux[subprog].linkage == BTF_FUNC_GLOBAL;
436	}
437
438	static const char *subprog_name(const struct bpf_verifier_env *env, int subprog)
439	{
440	struct bpf_func_info *info;
441
442	if (!env->prog->aux->func_info)
443	return "";
444
445	info = &env->prog->aux->func_info[subprog];
446	return btf_type_name(btf: env->prog->aux->btf, id: info->type_id);
447	}
448
449	static void mark_subprog_exc_cb(struct bpf_verifier_env *env, int subprog)
450	{
451	struct bpf_subprog_info *info = subprog_info(env, subprog);
452
453	info->is_cb = true;
454	info->is_async_cb = true;
455	info->is_exception_cb = true;
456	}
457
458	static bool subprog_is_exc_cb(struct bpf_verifier_env *env, int subprog)
459	{
460	return subprog_info(env, subprog)->is_exception_cb;
461	}
462
463	static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
464	{
465	return btf_record_has_field(rec: reg_btf_record(reg), type: BPF_SPIN_LOCK | BPF_RES_SPIN_LOCK);
466	}
467
468	static bool type_is_rdonly_mem(u32 type)
469	{
470	return type & MEM_RDONLY;
471	}
472
473	static bool is_acquire_function(enum bpf_func_id func_id,
474	const struct bpf_map *map)
475	{
476	enum bpf_map_type map_type = map ? map->map_type : BPF_MAP_TYPE_UNSPEC;
477
478	if (func_id == BPF_FUNC_sk_lookup_tcp ||
479	func_id == BPF_FUNC_sk_lookup_udp ||
480	func_id == BPF_FUNC_skc_lookup_tcp ||
481	func_id == BPF_FUNC_ringbuf_reserve ||
482	func_id == BPF_FUNC_kptr_xchg)
483	return true;
484
485	if (func_id == BPF_FUNC_map_lookup_elem &&
486	(map_type == BPF_MAP_TYPE_SOCKMAP ||
487	map_type == BPF_MAP_TYPE_SOCKHASH))
488	return true;
489
490	return false;
491	}
492
493	static bool is_ptr_cast_function(enum bpf_func_id func_id)
494	{
495	return func_id == BPF_FUNC_tcp_sock ||
496	func_id == BPF_FUNC_sk_fullsock ||
497	func_id == BPF_FUNC_skc_to_tcp_sock ||
498	func_id == BPF_FUNC_skc_to_tcp6_sock ||
499	func_id == BPF_FUNC_skc_to_udp6_sock ||
500	func_id == BPF_FUNC_skc_to_mptcp_sock ||
501	func_id == BPF_FUNC_skc_to_tcp_timewait_sock ||
502	func_id == BPF_FUNC_skc_to_tcp_request_sock;
503	}
504
505	static bool is_dynptr_ref_function(enum bpf_func_id func_id)
506	{
507	return func_id == BPF_FUNC_dynptr_data;
508	}
509
510	static bool is_sync_callback_calling_kfunc(u32 btf_id);
511	static bool is_async_callback_calling_kfunc(u32 btf_id);
512	static bool is_callback_calling_kfunc(u32 btf_id);
513	static bool is_bpf_throw_kfunc(struct bpf_insn *insn);
514
515	static bool is_bpf_wq_set_callback_impl_kfunc(u32 btf_id);
516	static bool is_task_work_add_kfunc(u32 func_id);
517
518	static bool is_sync_callback_calling_function(enum bpf_func_id func_id)
519	{
520	return func_id == BPF_FUNC_for_each_map_elem ||
521	func_id == BPF_FUNC_find_vma ||
522	func_id == BPF_FUNC_loop ||
523	func_id == BPF_FUNC_user_ringbuf_drain;
524	}
525
526	static bool is_async_callback_calling_function(enum bpf_func_id func_id)
527	{
528	return func_id == BPF_FUNC_timer_set_callback;
529	}
530
531	static bool is_callback_calling_function(enum bpf_func_id func_id)
532	{
533	return is_sync_callback_calling_function(func_id) ||
534	is_async_callback_calling_function(func_id);
535	}
536
537	static bool is_sync_callback_calling_insn(struct bpf_insn *insn)
538	{
539	return (bpf_helper_call(insn) && is_sync_callback_calling_function(func_id: insn->imm)) ||
540	(bpf_pseudo_kfunc_call(insn) && is_sync_callback_calling_kfunc(btf_id: insn->imm));
541	}
542
543	static bool is_async_callback_calling_insn(struct bpf_insn *insn)
544	{
545	return (bpf_helper_call(insn) && is_async_callback_calling_function(func_id: insn->imm)) ||
546	(bpf_pseudo_kfunc_call(insn) && is_async_callback_calling_kfunc(btf_id: insn->imm));
547	}
548
549	static bool is_async_cb_sleepable(struct bpf_verifier_env *env, struct bpf_insn *insn)
550	{
551	/* bpf_timer callbacks are never sleepable. */
552	if (bpf_helper_call(insn) && insn->imm == BPF_FUNC_timer_set_callback)
553	return false;
554
555	/* bpf_wq and bpf_task_work callbacks are always sleepable. */
556	if (bpf_pseudo_kfunc_call(insn) && insn->off == 0 &&
557	(is_bpf_wq_set_callback_impl_kfunc(btf_id: insn->imm) || is_task_work_add_kfunc(func_id: insn->imm)))
558	return true;
559
560	verifier_bug(env, "unhandled async callback in is_async_cb_sleepable");
561	return false;
562	}
563
564	static bool is_may_goto_insn(struct bpf_insn *insn)
565	{
566	return insn->code == (BPF_JMP | BPF_JCOND) && insn->src_reg == BPF_MAY_GOTO;
567	}
568
569	static bool is_may_goto_insn_at(struct bpf_verifier_env *env, int insn_idx)
570	{
571	return is_may_goto_insn(insn: &env->prog->insnsi[insn_idx]);
572	}
573
574	static bool is_storage_get_function(enum bpf_func_id func_id)
575	{
576	return func_id == BPF_FUNC_sk_storage_get ||
577	func_id == BPF_FUNC_inode_storage_get ||
578	func_id == BPF_FUNC_task_storage_get ||
579	func_id == BPF_FUNC_cgrp_storage_get;
580	}
581
582	static bool helper_multiple_ref_obj_use(enum bpf_func_id func_id,
583	const struct bpf_map *map)
584	{
585	int ref_obj_uses = 0;
586
587	if (is_ptr_cast_function(func_id))
588	ref_obj_uses++;
589	if (is_acquire_function(func_id, map))
590	ref_obj_uses++;
591	if (is_dynptr_ref_function(func_id))
592	ref_obj_uses++;
593
594	return ref_obj_uses > 1;
595	}
596
597	static bool is_cmpxchg_insn(const struct bpf_insn *insn)
598	{
599	return BPF_CLASS(insn->code) == BPF_STX &&
600	BPF_MODE(insn->code) == BPF_ATOMIC &&
601	insn->imm == BPF_CMPXCHG;
602	}
603
604	static bool is_atomic_load_insn(const struct bpf_insn *insn)
605	{
606	return BPF_CLASS(insn->code) == BPF_STX &&
607	BPF_MODE(insn->code) == BPF_ATOMIC &&
608	insn->imm == BPF_LOAD_ACQ;
609	}
610
611	static int __get_spi(s32 off)
612	{
613	return (-off - 1) / BPF_REG_SIZE;
614	}
615
616	static struct bpf_func_state *func(struct bpf_verifier_env *env,
617	const struct bpf_reg_state *reg)
618	{
619	struct bpf_verifier_state *cur = env->cur_state;
620
621	return cur->frame[reg->frameno];
622	}
623
624	static bool is_spi_bounds_valid(struct bpf_func_state *state, int spi, int nr_slots)
625	{
626	int allocated_slots = state->allocated_stack / BPF_REG_SIZE;
627
628	/* We need to check that slots between [spi - nr_slots + 1, spi] are
629	* within [0, allocated_stack).
630	*
631	* Please note that the spi grows downwards. For example, a dynptr
632	* takes the size of two stack slots; the first slot will be at
633	* spi and the second slot will be at spi - 1.
634	*/
635	return spi - nr_slots + 1 >= 0 && spi < allocated_slots;
636	}
637
638	static int stack_slot_obj_get_spi(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
639	const char *obj_kind, int nr_slots)
640	{
641	int off, spi;
642
643	if (!tnum_is_const(a: reg->var_off)) {
644	verbose(private_data: env, fmt: "%s has to be at a constant offset\n", obj_kind);
645	return -EINVAL;
646	}
647
648	off = reg->off + reg->var_off.value;
649	if (off % BPF_REG_SIZE) {
650	verbose(private_data: env, fmt: "cannot pass in %s at an offset=%d\n", obj_kind, off);
651	return -EINVAL;
652	}
653
654	spi = __get_spi(off);
655	if (spi + 1 < nr_slots) {
656	verbose(private_data: env, fmt: "cannot pass in %s at an offset=%d\n", obj_kind, off);
657	return -EINVAL;
658	}
659
660	if (!is_spi_bounds_valid(state: func(env, reg), spi, nr_slots))
661	return -ERANGE;
662	return spi;
663	}
664
665	static int dynptr_get_spi(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
666	{
667	return stack_slot_obj_get_spi(env, reg, obj_kind: "dynptr", BPF_DYNPTR_NR_SLOTS);
668	}
669
670	static int iter_get_spi(struct bpf_verifier_env *env, struct bpf_reg_state *reg, int nr_slots)
671	{
672	return stack_slot_obj_get_spi(env, reg, obj_kind: "iter", nr_slots);
673	}
674
675	static int irq_flag_get_spi(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
676	{
677	return stack_slot_obj_get_spi(env, reg, obj_kind: "irq_flag", nr_slots: 1);
678	}
679
680	static enum bpf_dynptr_type arg_to_dynptr_type(enum bpf_arg_type arg_type)
681	{
682	switch (arg_type & DYNPTR_TYPE_FLAG_MASK) {
683	case DYNPTR_TYPE_LOCAL:
684	return BPF_DYNPTR_TYPE_LOCAL;
685	case DYNPTR_TYPE_RINGBUF:
686	return BPF_DYNPTR_TYPE_RINGBUF;
687	case DYNPTR_TYPE_SKB:
688	return BPF_DYNPTR_TYPE_SKB;
689	case DYNPTR_TYPE_XDP:
690	return BPF_DYNPTR_TYPE_XDP;
691	case DYNPTR_TYPE_SKB_META:
692	return BPF_DYNPTR_TYPE_SKB_META;
693	case DYNPTR_TYPE_FILE:
694	return BPF_DYNPTR_TYPE_FILE;
695	default:
696	return BPF_DYNPTR_TYPE_INVALID;
697	}
698	}
699
700	static enum bpf_type_flag get_dynptr_type_flag(enum bpf_dynptr_type type)
701	{
702	switch (type) {
703	case BPF_DYNPTR_TYPE_LOCAL:
704	return DYNPTR_TYPE_LOCAL;
705	case BPF_DYNPTR_TYPE_RINGBUF:
706	return DYNPTR_TYPE_RINGBUF;
707	case BPF_DYNPTR_TYPE_SKB:
708	return DYNPTR_TYPE_SKB;
709	case BPF_DYNPTR_TYPE_XDP:
710	return DYNPTR_TYPE_XDP;
711	case BPF_DYNPTR_TYPE_SKB_META:
712	return DYNPTR_TYPE_SKB_META;
713	case BPF_DYNPTR_TYPE_FILE:
714	return DYNPTR_TYPE_FILE;
715	default:
716	return 0;
717	}
718	}
719
720	static bool dynptr_type_refcounted(enum bpf_dynptr_type type)
721	{
722	return type == BPF_DYNPTR_TYPE_RINGBUF || type == BPF_DYNPTR_TYPE_FILE;
723	}
724
725	static void __mark_dynptr_reg(struct bpf_reg_state *reg,
726	enum bpf_dynptr_type type,
727	bool first_slot, int dynptr_id);
728
729	static void __mark_reg_not_init(const struct bpf_verifier_env *env,
730	struct bpf_reg_state *reg);
731
732	static void mark_dynptr_stack_regs(struct bpf_verifier_env *env,
733	struct bpf_reg_state *sreg1,
734	struct bpf_reg_state *sreg2,
735	enum bpf_dynptr_type type)
736	{
737	int id = ++env->id_gen;
738
739	__mark_dynptr_reg(reg: sreg1, type, first_slot: true, dynptr_id: id);
740	__mark_dynptr_reg(reg: sreg2, type, first_slot: false, dynptr_id: id);
741	}
742
743	static void mark_dynptr_cb_reg(struct bpf_verifier_env *env,
744	struct bpf_reg_state *reg,
745	enum bpf_dynptr_type type)
746	{
747	__mark_dynptr_reg(reg, type, first_slot: true, dynptr_id: ++env->id_gen);
748	}
749
750	static int destroy_if_dynptr_stack_slot(struct bpf_verifier_env *env,
751	struct bpf_func_state *state, int spi);
752
753	static int mark_stack_slots_dynptr(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
754	enum bpf_arg_type arg_type, int insn_idx, int clone_ref_obj_id)
755	{
756	struct bpf_func_state *state = func(env, reg);
757	enum bpf_dynptr_type type;
758	int spi, i, err;
759
760	spi = dynptr_get_spi(env, reg);
761	if (spi < 0)
762	return spi;
763
764	/* We cannot assume both spi and spi - 1 belong to the same dynptr,
765	* hence we need to call destroy_if_dynptr_stack_slot twice for both,
766	* to ensure that for the following example:
767	* [d1][d1][d2][d2]
768	* spi 3 2 1 0
769	* So marking spi = 2 should lead to destruction of both d1 and d2. In
770	* case they do belong to same dynptr, second call won't see slot_type
771	* as STACK_DYNPTR and will simply skip destruction.
772	*/
773	err = destroy_if_dynptr_stack_slot(env, state, spi);
774	if (err)
775	return err;
776	err = destroy_if_dynptr_stack_slot(env, state, spi: spi - 1);
777	if (err)
778	return err;
779
780	for (i = 0; i < BPF_REG_SIZE; i++) {
781	state->stack[spi].slot_type[i] = STACK_DYNPTR;
782	state->stack[spi - 1].slot_type[i] = STACK_DYNPTR;
783	}
784
785	type = arg_to_dynptr_type(arg_type);
786	if (type == BPF_DYNPTR_TYPE_INVALID)
787	return -EINVAL;
788
789	mark_dynptr_stack_regs(env, sreg1: &state->stack[spi].spilled_ptr,
790	sreg2: &state->stack[spi - 1].spilled_ptr, type);
791
792	if (dynptr_type_refcounted(type)) {
793	/* The id is used to track proper releasing */
794	int id;
795
796	if (clone_ref_obj_id)
797	id = clone_ref_obj_id;
798	else
799	id = acquire_reference(env, insn_idx);
800
801	if (id < 0)
802	return id;
803
804	state->stack[spi].spilled_ptr.ref_obj_id = id;
805	state->stack[spi - 1].spilled_ptr.ref_obj_id = id;
806	}
807
808	bpf_mark_stack_write(env, frameno: state->frameno, BIT(spi - 1) | BIT(spi));
809
810	return 0;
811	}
812
813	static void invalidate_dynptr(struct bpf_verifier_env *env, struct bpf_func_state *state, int spi)
814	{
815	int i;
816
817	for (i = 0; i < BPF_REG_SIZE; i++) {
818	state->stack[spi].slot_type[i] = STACK_INVALID;
819	state->stack[spi - 1].slot_type[i] = STACK_INVALID;
820	}
821
822	__mark_reg_not_init(env, reg: &state->stack[spi].spilled_ptr);
823	__mark_reg_not_init(env, reg: &state->stack[spi - 1].spilled_ptr);
824
825	bpf_mark_stack_write(env, frameno: state->frameno, BIT(spi - 1) | BIT(spi));
826	}
827
828	static int unmark_stack_slots_dynptr(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
829	{
830	struct bpf_func_state *state = func(env, reg);
831	int spi, ref_obj_id, i;
832
833	/*
834	* This can only be set for PTR_TO_STACK, as CONST_PTR_TO_DYNPTR cannot
835	* be released by any dynptr helper. Hence, unmark_stack_slots_dynptr
836	* is safe to do directly.
837	*/
838	if (reg->type == CONST_PTR_TO_DYNPTR) {
839	verifier_bug(env, "CONST_PTR_TO_DYNPTR cannot be released");
840	return -EFAULT;
841	}
842	spi = dynptr_get_spi(env, reg);
843	if (spi < 0)
844	return spi;
845
846	if (!dynptr_type_refcounted(type: state->stack[spi].spilled_ptr.dynptr.type)) {
847	invalidate_dynptr(env, state, spi);
848	return 0;
849	}
850
851	ref_obj_id = state->stack[spi].spilled_ptr.ref_obj_id;
852
853	/* If the dynptr has a ref_obj_id, then we need to invalidate
854	* two things:
855	*
856	* 1) Any dynptrs with a matching ref_obj_id (clones)
857	* 2) Any slices derived from this dynptr.
858	*/
859
860	/* Invalidate any slices associated with this dynptr */
861	WARN_ON_ONCE(release_reference(env, ref_obj_id));
862
863	/* Invalidate any dynptr clones */
864	for (i = 1; i < state->allocated_stack / BPF_REG_SIZE; i++) {
865	if (state->stack[i].spilled_ptr.ref_obj_id != ref_obj_id)
866	continue;
867
868	/* it should always be the case that if the ref obj id
869	* matches then the stack slot also belongs to a
870	* dynptr
871	*/
872	if (state->stack[i].slot_type[0] != STACK_DYNPTR) {
873	verifier_bug(env, "misconfigured ref_obj_id");
874	return -EFAULT;
875	}
876	if (state->stack[i].spilled_ptr.dynptr.first_slot)
877	invalidate_dynptr(env, state, spi: i);
878	}
879
880	return 0;
881	}
882
883	static void __mark_reg_unknown(const struct bpf_verifier_env *env,
884	struct bpf_reg_state *reg);
885
886	static void mark_reg_invalid(const struct bpf_verifier_env *env, struct bpf_reg_state *reg)
887	{
888	if (!env->allow_ptr_leaks)
889	__mark_reg_not_init(env, reg);
890	else
891	__mark_reg_unknown(env, reg);
892	}
893
894	static int destroy_if_dynptr_stack_slot(struct bpf_verifier_env *env,
895	struct bpf_func_state *state, int spi)
896	{
897	struct bpf_func_state *fstate;
898	struct bpf_reg_state *dreg;
899	int i, dynptr_id;
900
901	/* We always ensure that STACK_DYNPTR is never set partially,
902	* hence just checking for slot_type[0] is enough. This is
903	* different for STACK_SPILL, where it may be only set for
904	* 1 byte, so code has to use is_spilled_reg.
905	*/
906	if (state->stack[spi].slot_type[0] != STACK_DYNPTR)
907	return 0;
908
909	/* Reposition spi to first slot */
910	if (!state->stack[spi].spilled_ptr.dynptr.first_slot)
911	spi = spi + 1;
912
913	if (dynptr_type_refcounted(type: state->stack[spi].spilled_ptr.dynptr.type)) {
914	verbose(private_data: env, fmt: "cannot overwrite referenced dynptr\n");
915	return -EINVAL;
916	}
917
918	mark_stack_slot_scratched(env, spi);
919	mark_stack_slot_scratched(env, spi: spi - 1);
920
921	/* Writing partially to one dynptr stack slot destroys both. */
922	for (i = 0; i < BPF_REG_SIZE; i++) {
923	state->stack[spi].slot_type[i] = STACK_INVALID;
924	state->stack[spi - 1].slot_type[i] = STACK_INVALID;
925	}
926
927	dynptr_id = state->stack[spi].spilled_ptr.id;
928	/* Invalidate any slices associated with this dynptr */
929	bpf_for_each_reg_in_vstate(env->cur_state, fstate, dreg, ({
930	/* Dynptr slices are only PTR_TO_MEM_OR_NULL and PTR_TO_MEM */
931	if (dreg->type != (PTR_TO_MEM | PTR_MAYBE_NULL) && dreg->type != PTR_TO_MEM)
932	continue;
933	if (dreg->dynptr_id == dynptr_id)
934	mark_reg_invalid(env, dreg);
935	}));
936
937	/* Do not release reference state, we are destroying dynptr on stack,
938	* not using some helper to release it. Just reset register.
939	*/
940	__mark_reg_not_init(env, reg: &state->stack[spi].spilled_ptr);
941	__mark_reg_not_init(env, reg: &state->stack[spi - 1].spilled_ptr);
942
943	bpf_mark_stack_write(env, frameno: state->frameno, BIT(spi - 1) | BIT(spi));
944
945	return 0;
946	}
947
948	static bool is_dynptr_reg_valid_uninit(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
949	{
950	int spi;
951
952	if (reg->type == CONST_PTR_TO_DYNPTR)
953	return false;
954
955	spi = dynptr_get_spi(env, reg);
956
957	/* -ERANGE (i.e. spi not falling into allocated stack slots) isn't an
958	* error because this just means the stack state hasn't been updated yet.
959	* We will do check_mem_access to check and update stack bounds later.
960	*/
961	if (spi < 0 && spi != -ERANGE)
962	return false;
963
964	/* We don't need to check if the stack slots are marked by previous
965	* dynptr initializations because we allow overwriting existing unreferenced
966	* STACK_DYNPTR slots, see mark_stack_slots_dynptr which calls
967	* destroy_if_dynptr_stack_slot to ensure dynptr objects at the slots we are
968	* touching are completely destructed before we reinitialize them for a new
969	* one. For referenced ones, destroy_if_dynptr_stack_slot returns an error early
970	* instead of delaying it until the end where the user will get "Unreleased
971	* reference" error.
972	*/
973	return true;
974	}
975
976	static bool is_dynptr_reg_valid_init(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
977	{
978	struct bpf_func_state *state = func(env, reg);
979	int i, spi;
980
981	/* This already represents first slot of initialized bpf_dynptr.
982	*
983	* CONST_PTR_TO_DYNPTR already has fixed and var_off as 0 due to
984	* check_func_arg_reg_off's logic, so we don't need to check its
985	* offset and alignment.
986	*/
987	if (reg->type == CONST_PTR_TO_DYNPTR)
988	return true;
989
990	spi = dynptr_get_spi(env, reg);
991	if (spi < 0)
992	return false;
993	if (!state->stack[spi].spilled_ptr.dynptr.first_slot)
994	return false;
995
996	for (i = 0; i < BPF_REG_SIZE; i++) {
997	if (state->stack[spi].slot_type[i] != STACK_DYNPTR ||
998	state->stack[spi - 1].slot_type[i] != STACK_DYNPTR)
999	return false;
1000	}
1001
1002	return true;
1003	}
1004
1005	static bool is_dynptr_type_expected(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
1006	enum bpf_arg_type arg_type)
1007	{
1008	struct bpf_func_state *state = func(env, reg);
1009	enum bpf_dynptr_type dynptr_type;
1010	int spi;
1011
1012	/* ARG_PTR_TO_DYNPTR takes any type of dynptr */
1013	if (arg_type == ARG_PTR_TO_DYNPTR)
1014	return true;
1015
1016	dynptr_type = arg_to_dynptr_type(arg_type);
1017	if (reg->type == CONST_PTR_TO_DYNPTR) {
1018	return reg->dynptr.type == dynptr_type;
1019	} else {
1020	spi = dynptr_get_spi(env, reg);
1021	if (spi < 0)
1022	return false;
1023	return state->stack[spi].spilled_ptr.dynptr.type == dynptr_type;
1024	}
1025	}
1026
1027	static void __mark_reg_known_zero(struct bpf_reg_state *reg);
1028
1029	static bool in_rcu_cs(struct bpf_verifier_env *env);
1030
1031	static bool is_kfunc_rcu_protected(struct bpf_kfunc_call_arg_meta *meta);
1032
1033	static int mark_stack_slots_iter(struct bpf_verifier_env *env,
1034	struct bpf_kfunc_call_arg_meta *meta,
1035	struct bpf_reg_state *reg, int insn_idx,
1036	struct btf *btf, u32 btf_id, int nr_slots)
1037	{
1038	struct bpf_func_state *state = func(env, reg);
1039	int spi, i, j, id;
1040
1041	spi = iter_get_spi(env, reg, nr_slots);
1042	if (spi < 0)
1043	return spi;
1044
1045	id = acquire_reference(env, insn_idx);
1046	if (id < 0)
1047	return id;
1048
1049	for (i = 0; i < nr_slots; i++) {
1050	struct bpf_stack_state *slot = &state->stack[spi - i];
1051	struct bpf_reg_state *st = &slot->spilled_ptr;
1052
1053	__mark_reg_known_zero(reg: st);
1054	st->type = PTR_TO_STACK; /* we don't have dedicated reg type */
1055	if (is_kfunc_rcu_protected(meta)) {
1056	if (in_rcu_cs(env))
1057	st->type |= MEM_RCU;
1058	else
1059	st->type |= PTR_UNTRUSTED;
1060	}
1061	st->ref_obj_id = i == 0 ? id : 0;
1062	st->iter.btf = btf;
1063	st->iter.btf_id = btf_id;
1064	st->iter.state = BPF_ITER_STATE_ACTIVE;
1065	st->iter.depth = 0;
1066
1067	for (j = 0; j < BPF_REG_SIZE; j++)
1068	slot->slot_type[j] = STACK_ITER;
1069
1070	bpf_mark_stack_write(env, frameno: state->frameno, BIT(spi - i));
1071	mark_stack_slot_scratched(env, spi: spi - i);
1072	}
1073
1074	return 0;
1075	}
1076
1077	static int unmark_stack_slots_iter(struct bpf_verifier_env *env,
1078	struct bpf_reg_state *reg, int nr_slots)
1079	{
1080	struct bpf_func_state *state = func(env, reg);
1081	int spi, i, j;
1082
1083	spi = iter_get_spi(env, reg, nr_slots);
1084	if (spi < 0)
1085	return spi;
1086
1087	for (i = 0; i < nr_slots; i++) {
1088	struct bpf_stack_state *slot = &state->stack[spi - i];
1089	struct bpf_reg_state *st = &slot->spilled_ptr;
1090
1091	if (i == 0)
1092	WARN_ON_ONCE(release_reference(env, st->ref_obj_id));
1093
1094	__mark_reg_not_init(env, reg: st);
1095
1096	for (j = 0; j < BPF_REG_SIZE; j++)
1097	slot->slot_type[j] = STACK_INVALID;
1098
1099	bpf_mark_stack_write(env, frameno: state->frameno, BIT(spi - i));
1100	mark_stack_slot_scratched(env, spi: spi - i);
1101	}
1102
1103	return 0;
1104	}
1105
1106	static bool is_iter_reg_valid_uninit(struct bpf_verifier_env *env,
1107	struct bpf_reg_state *reg, int nr_slots)
1108	{
1109	struct bpf_func_state *state = func(env, reg);
1110	int spi, i, j;
1111
1112	/* For -ERANGE (i.e. spi not falling into allocated stack slots), we
1113	* will do check_mem_access to check and update stack bounds later, so
1114	* return true for that case.
1115	*/
1116	spi = iter_get_spi(env, reg, nr_slots);
1117	if (spi == -ERANGE)
1118	return true;
1119	if (spi < 0)
1120	return false;
1121
1122	for (i = 0; i < nr_slots; i++) {
1123	struct bpf_stack_state *slot = &state->stack[spi - i];
1124
1125	for (j = 0; j < BPF_REG_SIZE; j++)
1126	if (slot->slot_type[j] == STACK_ITER)
1127	return false;
1128	}
1129
1130	return true;
1131	}
1132
1133	static int is_iter_reg_valid_init(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
1134	struct btf *btf, u32 btf_id, int nr_slots)
1135	{
1136	struct bpf_func_state *state = func(env, reg);
1137	int spi, i, j;
1138
1139	spi = iter_get_spi(env, reg, nr_slots);
1140	if (spi < 0)
1141	return -EINVAL;
1142
1143	for (i = 0; i < nr_slots; i++) {
1144	struct bpf_stack_state *slot = &state->stack[spi - i];
1145	struct bpf_reg_state *st = &slot->spilled_ptr;
1146
1147	if (st->type & PTR_UNTRUSTED)
1148	return -EPROTO;
1149	/* only main (first) slot has ref_obj_id set */
1150	if (i == 0 && !st->ref_obj_id)
1151	return -EINVAL;
1152	if (i != 0 && st->ref_obj_id)
1153	return -EINVAL;
1154	if (st->iter.btf != btf || st->iter.btf_id != btf_id)
1155	return -EINVAL;
1156
1157	for (j = 0; j < BPF_REG_SIZE; j++)
1158	if (slot->slot_type[j] != STACK_ITER)
1159	return -EINVAL;
1160	}
1161
1162	return 0;
1163	}
1164
1165	static int acquire_irq_state(struct bpf_verifier_env *env, int insn_idx);
1166	static int release_irq_state(struct bpf_verifier_state *state, int id);
1167
1168	static int mark_stack_slot_irq_flag(struct bpf_verifier_env *env,
1169	struct bpf_kfunc_call_arg_meta *meta,
1170	struct bpf_reg_state *reg, int insn_idx,
1171	int kfunc_class)
1172	{
1173	struct bpf_func_state *state = func(env, reg);
1174	struct bpf_stack_state *slot;
1175	struct bpf_reg_state *st;
1176	int spi, i, id;
1177
1178	spi = irq_flag_get_spi(env, reg);
1179	if (spi < 0)
1180	return spi;
1181
1182	id = acquire_irq_state(env, insn_idx);
1183	if (id < 0)
1184	return id;
1185
1186	slot = &state->stack[spi];
1187	st = &slot->spilled_ptr;
1188
1189	bpf_mark_stack_write(env, frameno: reg->frameno, BIT(spi));
1190	__mark_reg_known_zero(reg: st);
1191	st->type = PTR_TO_STACK; /* we don't have dedicated reg type */
1192	st->ref_obj_id = id;
1193	st->irq.kfunc_class = kfunc_class;
1194
1195	for (i = 0; i < BPF_REG_SIZE; i++)
1196	slot->slot_type[i] = STACK_IRQ_FLAG;
1197
1198	mark_stack_slot_scratched(env, spi);
1199	return 0;
1200	}
1201
1202	static int unmark_stack_slot_irq_flag(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
1203	int kfunc_class)
1204	{
1205	struct bpf_func_state *state = func(env, reg);
1206	struct bpf_stack_state *slot;
1207	struct bpf_reg_state *st;
1208	int spi, i, err;
1209
1210	spi = irq_flag_get_spi(env, reg);
1211	if (spi < 0)
1212	return spi;
1213
1214	slot = &state->stack[spi];
1215	st = &slot->spilled_ptr;
1216
1217	if (st->irq.kfunc_class != kfunc_class) {
1218	const char *flag_kfunc = st->irq.kfunc_class == IRQ_NATIVE_KFUNC ? "native" : "lock";
1219	const char *used_kfunc = kfunc_class == IRQ_NATIVE_KFUNC ? "native" : "lock";
1220
1221	verbose(private_data: env, fmt: "irq flag acquired by %s kfuncs cannot be restored with %s kfuncs\n",
1222	flag_kfunc, used_kfunc);
1223	return -EINVAL;
1224	}
1225
1226	err = release_irq_state(state: env->cur_state, id: st->ref_obj_id);
1227	WARN_ON_ONCE(err && err != -EACCES);
1228	if (err) {
1229	int insn_idx = 0;
1230
1231	for (int i = 0; i < env->cur_state->acquired_refs; i++) {
1232	if (env->cur_state->refs[i].id == env->cur_state->active_irq_id) {
1233	insn_idx = env->cur_state->refs[i].insn_idx;
1234	break;
1235	}
1236	}
1237
1238	verbose(private_data: env, fmt: "cannot restore irq state out of order, expected id=%d acquired at insn_idx=%d\n",
1239	env->cur_state->active_irq_id, insn_idx);
1240	return err;
1241	}
1242
1243	__mark_reg_not_init(env, reg: st);
1244
1245	bpf_mark_stack_write(env, frameno: reg->frameno, BIT(spi));
1246
1247	for (i = 0; i < BPF_REG_SIZE; i++)
1248	slot->slot_type[i] = STACK_INVALID;
1249
1250	mark_stack_slot_scratched(env, spi);
1251	return 0;
1252	}
1253
1254	static bool is_irq_flag_reg_valid_uninit(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
1255	{
1256	struct bpf_func_state *state = func(env, reg);
1257	struct bpf_stack_state *slot;
1258	int spi, i;
1259
1260	/* For -ERANGE (i.e. spi not falling into allocated stack slots), we
1261	* will do check_mem_access to check and update stack bounds later, so
1262	* return true for that case.
1263	*/
1264	spi = irq_flag_get_spi(env, reg);
1265	if (spi == -ERANGE)
1266	return true;
1267	if (spi < 0)
1268	return false;
1269
1270	slot = &state->stack[spi];
1271
1272	for (i = 0; i < BPF_REG_SIZE; i++)
1273	if (slot->slot_type[i] == STACK_IRQ_FLAG)
1274	return false;
1275	return true;
1276	}
1277
1278	static int is_irq_flag_reg_valid_init(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
1279	{
1280	struct bpf_func_state *state = func(env, reg);
1281	struct bpf_stack_state *slot;
1282	struct bpf_reg_state *st;
1283	int spi, i;
1284
1285	spi = irq_flag_get_spi(env, reg);
1286	if (spi < 0)
1287	return -EINVAL;
1288
1289	slot = &state->stack[spi];
1290	st = &slot->spilled_ptr;
1291
1292	if (!st->ref_obj_id)
1293	return -EINVAL;
1294
1295	for (i = 0; i < BPF_REG_SIZE; i++)
1296	if (slot->slot_type[i] != STACK_IRQ_FLAG)
1297	return -EINVAL;
1298	return 0;
1299	}
1300
1301	/* Check if given stack slot is "special":
1302	* - spilled register state (STACK_SPILL);
1303	* - dynptr state (STACK_DYNPTR);
1304	* - iter state (STACK_ITER).
1305	* - irq flag state (STACK_IRQ_FLAG)
1306	*/
1307	static bool is_stack_slot_special(const struct bpf_stack_state *stack)
1308	{
1309	enum bpf_stack_slot_type type = stack->slot_type[BPF_REG_SIZE - 1];
1310
1311	switch (type) {
1312	case STACK_SPILL:
1313	case STACK_DYNPTR:
1314	case STACK_ITER:
1315	case STACK_IRQ_FLAG:
1316	return true;
1317	case STACK_INVALID:
1318	case STACK_MISC:
1319	case STACK_ZERO:
1320	return false;
1321	default:
1322	WARN_ONCE(1, "unknown stack slot type %d\n", type);
1323	return true;
1324	}
1325	}
1326
1327	/* The reg state of a pointer or a bounded scalar was saved when
1328	* it was spilled to the stack.
1329	*/
1330	static bool is_spilled_reg(const struct bpf_stack_state *stack)
1331	{
1332	return stack->slot_type[BPF_REG_SIZE - 1] == STACK_SPILL;
1333	}
1334
1335	static bool is_spilled_scalar_reg(const struct bpf_stack_state *stack)
1336	{
1337	return stack->slot_type[BPF_REG_SIZE - 1] == STACK_SPILL &&
1338	stack->spilled_ptr.type == SCALAR_VALUE;
1339	}
1340
1341	static bool is_spilled_scalar_reg64(const struct bpf_stack_state *stack)
1342	{
1343	return stack->slot_type[0] == STACK_SPILL &&
1344	stack->spilled_ptr.type == SCALAR_VALUE;
1345	}
1346
1347	/* Mark stack slot as STACK_MISC, unless it is already STACK_INVALID, in which
1348	* case they are equivalent, or it's STACK_ZERO, in which case we preserve
1349	* more precise STACK_ZERO.
1350	* Regardless of allow_ptr_leaks setting (i.e., privileged or unprivileged
1351	* mode), we won't promote STACK_INVALID to STACK_MISC. In privileged case it is
1352	* unnecessary as both are considered equivalent when loading data and pruning,
1353	* in case of unprivileged mode it will be incorrect to allow reads of invalid
1354	* slots.
1355	*/
1356	static void mark_stack_slot_misc(struct bpf_verifier_env *env, u8 *stype)
1357	{
1358	if (*stype == STACK_ZERO)
1359	return;
1360	if (*stype == STACK_INVALID)
1361	return;
1362	*stype = STACK_MISC;
1363	}
1364
1365	static void scrub_spilled_slot(u8 *stype)
1366	{
1367	if (*stype != STACK_INVALID)
1368	*stype = STACK_MISC;
1369	}
1370
1371	/* copy array src of length n * size bytes to dst. dst is reallocated if it's too
1372	* small to hold src. This is different from krealloc since we don't want to preserve
1373	* the contents of dst.
1374	*
1375	* Leaves dst untouched if src is NULL or length is zero. Returns NULL if memory could
1376	* not be allocated.
1377	*/
1378	static void *copy_array(void *dst, const void *src, size_t n, size_t size, gfp_t flags)
1379	{
1380	size_t alloc_bytes;
1381	void *orig = dst;
1382	size_t bytes;
1383
1384	if (ZERO_OR_NULL_PTR(src))
1385	goto out;
1386
1387	if (unlikely(check_mul_overflow(n, size, &bytes)))
1388	return NULL;
1389
1390	alloc_bytes = max(ksize(orig), kmalloc_size_roundup(bytes));
1391	dst = krealloc(orig, alloc_bytes, flags);
1392	if (!dst) {
1393	kfree(objp: orig);
1394	return NULL;
1395	}
1396
1397	memcpy(dst, src, bytes);
1398	out:
1399	return dst ? dst : ZERO_SIZE_PTR;
1400	}
1401
1402	/* resize an array from old_n items to new_n items. the array is reallocated if it's too
1403	* small to hold new_n items. new items are zeroed out if the array grows.
1404	*
1405	* Contrary to krealloc_array, does not free arr if new_n is zero.
1406	*/
1407	static void *realloc_array(void *arr, size_t old_n, size_t new_n, size_t size)
1408	{
1409	size_t alloc_size;
1410	void *new_arr;
1411
1412	if (!new_n || old_n == new_n)
1413	goto out;
1414
1415	alloc_size = kmalloc_size_roundup(size: size_mul(factor1: new_n, factor2: size));
1416	new_arr = krealloc(arr, alloc_size, GFP_KERNEL_ACCOUNT);
1417	if (!new_arr) {
1418	kfree(objp: arr);
1419	return NULL;
1420	}
1421	arr = new_arr;
1422
1423	if (new_n > old_n)
1424	memset(arr + old_n * size, 0, (new_n - old_n) * size);
1425
1426	out:
1427	return arr ? arr : ZERO_SIZE_PTR;
1428	}
1429
1430	static int copy_reference_state(struct bpf_verifier_state *dst, const struct bpf_verifier_state *src)
1431	{
1432	dst->refs = copy_array(dst: dst->refs, src: src->refs, n: src->acquired_refs,
1433	size: sizeof(struct bpf_reference_state), GFP_KERNEL_ACCOUNT);
1434	if (!dst->refs)
1435	return -ENOMEM;
1436
1437	dst->acquired_refs = src->acquired_refs;
1438	dst->active_locks = src->active_locks;
1439	dst->active_preempt_locks = src->active_preempt_locks;
1440	dst->active_rcu_locks = src->active_rcu_locks;
1441	dst->active_irq_id = src->active_irq_id;
1442	dst->active_lock_id = src->active_lock_id;
1443	dst->active_lock_ptr = src->active_lock_ptr;
1444	return 0;
1445	}
1446
1447	static int copy_stack_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
1448	{
1449	size_t n = src->allocated_stack / BPF_REG_SIZE;
1450
1451	dst->stack = copy_array(dst: dst->stack, src: src->stack, n, size: sizeof(struct bpf_stack_state),
1452	GFP_KERNEL_ACCOUNT);
1453	if (!dst->stack)
1454	return -ENOMEM;
1455
1456	dst->allocated_stack = src->allocated_stack;
1457	return 0;
1458	}
1459
1460	static int resize_reference_state(struct bpf_verifier_state *state, size_t n)
1461	{
1462	state->refs = realloc_array(arr: state->refs, old_n: state->acquired_refs, new_n: n,
1463	size: sizeof(struct bpf_reference_state));
1464	if (!state->refs)
1465	return -ENOMEM;
1466
1467	state->acquired_refs = n;
1468	return 0;
1469	}
1470
1471	/* Possibly update state->allocated_stack to be at least size bytes. Also
1472	* possibly update the function's high-water mark in its bpf_subprog_info.
1473	*/
1474	static int grow_stack_state(struct bpf_verifier_env *env, struct bpf_func_state *state, int size)
1475	{
1476	size_t old_n = state->allocated_stack / BPF_REG_SIZE, n;
1477
1478	/* The stack size is always a multiple of BPF_REG_SIZE. */
1479	size = round_up(size, BPF_REG_SIZE);
1480	n = size / BPF_REG_SIZE;
1481
1482	if (old_n >= n)
1483	return 0;
1484
1485	state->stack = realloc_array(arr: state->stack, old_n, new_n: n, size: sizeof(struct bpf_stack_state));
1486	if (!state->stack)
1487	return -ENOMEM;
1488
1489	state->allocated_stack = size;
1490
1491	/* update known max for given subprogram */
1492	if (env->subprog_info[state->subprogno].stack_depth < size)
1493	env->subprog_info[state->subprogno].stack_depth = size;
1494
1495	return 0;
1496	}
1497
1498	/* Acquire a pointer id from the env and update the state->refs to include
1499	* this new pointer reference.
1500	* On success, returns a valid pointer id to associate with the register
1501	* On failure, returns a negative errno.
1502	*/
1503	static struct bpf_reference_state *acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
1504	{
1505	struct bpf_verifier_state *state = env->cur_state;
1506	int new_ofs = state->acquired_refs;
1507	int err;
1508
1509	err = resize_reference_state(state, n: state->acquired_refs + 1);
1510	if (err)
1511	return NULL;
1512	state->refs[new_ofs].insn_idx = insn_idx;
1513
1514	return &state->refs[new_ofs];
1515	}
1516
1517	static int acquire_reference(struct bpf_verifier_env *env, int insn_idx)
1518	{
1519	struct bpf_reference_state *s;
1520
1521	s = acquire_reference_state(env, insn_idx);
1522	if (!s)
1523	return -ENOMEM;
1524	s->type = REF_TYPE_PTR;
1525	s->id = ++env->id_gen;
1526	return s->id;
1527	}
1528
1529	static int acquire_lock_state(struct bpf_verifier_env *env, int insn_idx, enum ref_state_type type,
1530	int id, void *ptr)
1531	{
1532	struct bpf_verifier_state *state = env->cur_state;
1533	struct bpf_reference_state *s;
1534
1535	s = acquire_reference_state(env, insn_idx);
1536	if (!s)
1537	return -ENOMEM;
1538	s->type = type;
1539	s->id = id;
1540	s->ptr = ptr;
1541
1542	state->active_locks++;
1543	state->active_lock_id = id;
1544	state->active_lock_ptr = ptr;
1545	return 0;
1546	}
1547
1548	static int acquire_irq_state(struct bpf_verifier_env *env, int insn_idx)
1549	{
1550	struct bpf_verifier_state *state = env->cur_state;
1551	struct bpf_reference_state *s;
1552
1553	s = acquire_reference_state(env, insn_idx);
1554	if (!s)
1555	return -ENOMEM;
1556	s->type = REF_TYPE_IRQ;
1557	s->id = ++env->id_gen;
1558
1559	state->active_irq_id = s->id;
1560	return s->id;
1561	}
1562
1563	static void release_reference_state(struct bpf_verifier_state *state, int idx)
1564	{
1565	int last_idx;
1566	size_t rem;
1567
1568	/* IRQ state requires the relative ordering of elements remaining the
1569	* same, since it relies on the refs array to behave as a stack, so that
1570	* it can detect out-of-order IRQ restore. Hence use memmove to shift
1571	* the array instead of swapping the final element into the deleted idx.
1572	*/
1573	last_idx = state->acquired_refs - 1;
1574	rem = state->acquired_refs - idx - 1;
1575	if (last_idx && idx != last_idx)
1576	memmove(&state->refs[idx], &state->refs[idx + 1], sizeof(*state->refs) * rem);
1577	memset(&state->refs[last_idx], 0, sizeof(*state->refs));
1578	state->acquired_refs--;
1579	return;
1580	}
1581
1582	static bool find_reference_state(struct bpf_verifier_state *state, int ptr_id)
1583	{
1584	int i;
1585
1586	for (i = 0; i < state->acquired_refs; i++)
1587	if (state->refs[i].id == ptr_id)
1588	return true;
1589
1590	return false;
1591	}
1592
1593	static int release_lock_state(struct bpf_verifier_state *state, int type, int id, void *ptr)
1594	{
1595	void *prev_ptr = NULL;
1596	u32 prev_id = 0;
1597	int i;
1598
1599	for (i = 0; i < state->acquired_refs; i++) {
1600	if (state->refs[i].type == type && state->refs[i].id == id &&
1601	state->refs[i].ptr == ptr) {
1602	release_reference_state(state, idx: i);
1603	state->active_locks--;
1604	/* Reassign active lock (id, ptr). */
1605	state->active_lock_id = prev_id;
1606	state->active_lock_ptr = prev_ptr;
1607	return 0;
1608	}
1609	if (state->refs[i].type & REF_TYPE_LOCK_MASK) {
1610	prev_id = state->refs[i].id;
1611	prev_ptr = state->refs[i].ptr;
1612	}
1613	}
1614	return -EINVAL;
1615	}
1616
1617	static int release_irq_state(struct bpf_verifier_state *state, int id)
1618	{
1619	u32 prev_id = 0;
1620	int i;
1621
1622	if (id != state->active_irq_id)
1623	return -EACCES;
1624
1625	for (i = 0; i < state->acquired_refs; i++) {
1626	if (state->refs[i].type != REF_TYPE_IRQ)
1627	continue;
1628	if (state->refs[i].id == id) {
1629	release_reference_state(state, idx: i);
1630	state->active_irq_id = prev_id;
1631	return 0;
1632	} else {
1633	prev_id = state->refs[i].id;
1634	}
1635	}
1636	return -EINVAL;
1637	}
1638
1639	static struct bpf_reference_state *find_lock_state(struct bpf_verifier_state *state, enum ref_state_type type,
1640	int id, void *ptr)
1641	{
1642	int i;
1643
1644	for (i = 0; i < state->acquired_refs; i++) {
1645	struct bpf_reference_state *s = &state->refs[i];
1646
1647	if (!(s->type & type))
1648	continue;
1649
1650	if (s->id == id && s->ptr == ptr)
1651	return s;
1652	}
1653	return NULL;
1654	}
1655
1656	static void update_peak_states(struct bpf_verifier_env *env)
1657	{
1658	u32 cur_states;
1659
1660	cur_states = env->explored_states_size + env->free_list_size + env->num_backedges;
1661	env->peak_states = max(env->peak_states, cur_states);
1662	}
1663
1664	static void free_func_state(struct bpf_func_state *state)
1665	{
1666	if (!state)
1667	return;
1668	kfree(objp: state->stack);
1669	kfree(objp: state);
1670	}
1671
1672	static void clear_jmp_history(struct bpf_verifier_state *state)
1673	{
1674	kfree(objp: state->jmp_history);
1675	state->jmp_history = NULL;
1676	state->jmp_history_cnt = 0;
1677	}
1678
1679	static void free_verifier_state(struct bpf_verifier_state *state,
1680	bool free_self)
1681	{
1682	int i;
1683
1684	for (i = 0; i <= state->curframe; i++) {
1685	free_func_state(state: state->frame[i]);
1686	state->frame[i] = NULL;
1687	}
1688	kfree(objp: state->refs);
1689	clear_jmp_history(state);
1690	if (free_self)
1691	kfree(objp: state);
1692	}
1693
1694	/* struct bpf_verifier_state->parent refers to states
1695	* that are in either of env->{expored_states,free_list}.
1696	* In both cases the state is contained in struct bpf_verifier_state_list.
1697	*/
1698	static struct bpf_verifier_state_list *state_parent_as_list(struct bpf_verifier_state *st)
1699	{
1700	if (st->parent)
1701	return container_of(st->parent, struct bpf_verifier_state_list, state);
1702	return NULL;
1703	}
1704
1705	static bool incomplete_read_marks(struct bpf_verifier_env *env,
1706	struct bpf_verifier_state *st);
1707
1708	/* A state can be freed if it is no longer referenced:
1709	* - is in the env->free_list;
1710	* - has no children states;
1711	*/
1712	static void maybe_free_verifier_state(struct bpf_verifier_env *env,
1713	struct bpf_verifier_state_list *sl)
1714	{
1715	if (!sl->in_free_list
1716	|| sl->state.branches != 0
1717	|| incomplete_read_marks(env, st: &sl->state))
1718	return;
1719	list_del(entry: &sl->node);
1720	free_verifier_state(state: &sl->state, free_self: false);
1721	kfree(objp: sl);
1722	env->free_list_size--;
1723	}
1724
1725	/* copy verifier state from src to dst growing dst stack space
1726	* when necessary to accommodate larger src stack
1727	*/
1728	static int copy_func_state(struct bpf_func_state *dst,
1729	const struct bpf_func_state *src)
1730	{
1731	memcpy(dst, src, offsetof(struct bpf_func_state, stack));
1732	return copy_stack_state(dst, src);
1733	}
1734
1735	static int copy_verifier_state(struct bpf_verifier_state *dst_state,
1736	const struct bpf_verifier_state *src)
1737	{
1738	struct bpf_func_state *dst;
1739	int i, err;
1740
1741	dst_state->jmp_history = copy_array(dst: dst_state->jmp_history, src: src->jmp_history,
1742	n: src->jmp_history_cnt, size: sizeof(*dst_state->jmp_history),
1743	GFP_KERNEL_ACCOUNT);
1744	if (!dst_state->jmp_history)
1745	return -ENOMEM;
1746	dst_state->jmp_history_cnt = src->jmp_history_cnt;
1747
1748	/* if dst has more stack frames then src frame, free them, this is also
1749	* necessary in case of exceptional exits using bpf_throw.
1750	*/
1751	for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
1752	free_func_state(state: dst_state->frame[i]);
1753	dst_state->frame[i] = NULL;
1754	}
1755	err = copy_reference_state(dst: dst_state, src);
1756	if (err)
1757	return err;
1758	dst_state->speculative = src->speculative;
1759	dst_state->in_sleepable = src->in_sleepable;
1760	dst_state->cleaned = src->cleaned;
1761	dst_state->curframe = src->curframe;
1762	dst_state->branches = src->branches;
1763	dst_state->parent = src->parent;
1764	dst_state->first_insn_idx = src->first_insn_idx;
1765	dst_state->last_insn_idx = src->last_insn_idx;
1766	dst_state->dfs_depth = src->dfs_depth;
1767	dst_state->callback_unroll_depth = src->callback_unroll_depth;
1768	dst_state->may_goto_depth = src->may_goto_depth;
1769	dst_state->equal_state = src->equal_state;
1770	for (i = 0; i <= src->curframe; i++) {
1771	dst = dst_state->frame[i];
1772	if (!dst) {
1773	dst = kzalloc(sizeof(*dst), GFP_KERNEL_ACCOUNT);
1774	if (!dst)
1775	return -ENOMEM;
1776	dst_state->frame[i] = dst;
1777	}
1778	err = copy_func_state(dst, src: src->frame[i]);
1779	if (err)
1780	return err;
1781	}
1782	return 0;
1783	}
1784
1785	static u32 state_htab_size(struct bpf_verifier_env *env)
1786	{
1787	return env->prog->len;
1788	}
1789
1790	static struct list_head *explored_state(struct bpf_verifier_env *env, int idx)
1791	{
1792	struct bpf_verifier_state *cur = env->cur_state;
1793	struct bpf_func_state *state = cur->frame[cur->curframe];
1794
1795	return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
1796	}
1797
1798	static bool same_callsites(struct bpf_verifier_state *a, struct bpf_verifier_state *b)
1799	{
1800	int fr;
1801
1802	if (a->curframe != b->curframe)
1803	return false;
1804
1805	for (fr = a->curframe; fr >= 0; fr--)
1806	if (a->frame[fr]->callsite != b->frame[fr]->callsite)
1807	return false;
1808
1809	return true;
1810	}
1811
1812	/* Return IP for a given frame in a call stack */
1813	static u32 frame_insn_idx(struct bpf_verifier_state *st, u32 frame)
1814	{
1815	return frame == st->curframe
1816	? st->insn_idx
1817	: st->frame[frame + 1]->callsite;
1818	}
1819
1820	/* For state @st look for a topmost frame with frame_insn_idx() in some SCC,
1821	* if such frame exists form a corresponding @callchain as an array of
1822	* call sites leading to this frame and SCC id.
1823	* E.g.:
1824	*
1825	* void foo() { A: loop {... SCC#1 ...}; }
1826	* void bar() { B: loop { C: foo(); ... SCC#2 ... }
1827	* D: loop { E: foo(); ... SCC#3 ... } }
1828	* void main() { F: bar(); }
1829	*
1830	* @callchain at (A) would be either (F,SCC#2) or (F,SCC#3) depending
1831	* on @st frame call sites being (F,C,A) or (F,E,A).
1832	*/
1833	static bool compute_scc_callchain(struct bpf_verifier_env *env,
1834	struct bpf_verifier_state *st,
1835	struct bpf_scc_callchain *callchain)
1836	{
1837	u32 i, scc, insn_idx;
1838
1839	memset(callchain, 0, sizeof(*callchain));
1840	for (i = 0; i <= st->curframe; i++) {
1841	insn_idx = frame_insn_idx(st, frame: i);
1842	scc = env->insn_aux_data[insn_idx].scc;
1843	if (scc) {
1844	callchain->scc = scc;
1845	break;
1846	} else if (i < st->curframe) {
1847	callchain->callsites[i] = insn_idx;
1848	} else {
1849	return false;
1850	}
1851	}
1852	return true;
1853	}
1854
1855	/* Check if bpf_scc_visit instance for @callchain exists. */
1856	static struct bpf_scc_visit *scc_visit_lookup(struct bpf_verifier_env *env,
1857	struct bpf_scc_callchain *callchain)
1858	{
1859	struct bpf_scc_info *info = env->scc_info[callchain->scc];
1860	struct bpf_scc_visit *visits = info->visits;
1861	u32 i;
1862
1863	if (!info)
1864	return NULL;
1865	for (i = 0; i < info->num_visits; i++)
1866	if (memcmp(p: callchain, q: &visits[i].callchain, size: sizeof(*callchain)) == 0)
1867	return &visits[i];
1868	return NULL;
1869	}
1870
1871	/* Allocate a new bpf_scc_visit instance corresponding to @callchain.
1872	* Allocated instances are alive for a duration of the do_check_common()
1873	* call and are freed by free_states().
1874	*/
1875	static struct bpf_scc_visit *scc_visit_alloc(struct bpf_verifier_env *env,
1876	struct bpf_scc_callchain *callchain)
1877	{
1878	struct bpf_scc_visit *visit;
1879	struct bpf_scc_info *info;
1880	u32 scc, num_visits;
1881	u64 new_sz;
1882
1883	scc = callchain->scc;
1884	info = env->scc_info[scc];
1885	num_visits = info ? info->num_visits : 0;
1886	new_sz = sizeof(*info) + sizeof(struct bpf_scc_visit) * (num_visits + 1);
1887	info = kvrealloc(env->scc_info[scc], new_sz, GFP_KERNEL_ACCOUNT);
1888	if (!info)
1889	return NULL;
1890	env->scc_info[scc] = info;
1891	info->num_visits = num_visits + 1;
1892	visit = &info->visits[num_visits];
1893	memset(visit, 0, sizeof(*visit));
1894	memcpy(&visit->callchain, callchain, sizeof(*callchain));
1895	return visit;
1896	}
1897
1898	/* Form a string '(callsite#1,callsite#2,...,scc)' in env->tmp_str_buf */
1899	static char *format_callchain(struct bpf_verifier_env *env, struct bpf_scc_callchain *callchain)
1900	{
1901	char *buf = env->tmp_str_buf;
1902	int i, delta = 0;
1903
1904	delta += snprintf(buf: buf + delta, TMP_STR_BUF_LEN - delta, fmt: "(");
1905	for (i = 0; i < ARRAY_SIZE(callchain->callsites); i++) {
1906	if (!callchain->callsites[i])
1907	break;
1908	delta += snprintf(buf: buf + delta, TMP_STR_BUF_LEN - delta, fmt: "%u,",
1909	callchain->callsites[i]);
1910	}
1911	delta += snprintf(buf: buf + delta, TMP_STR_BUF_LEN - delta, fmt: "%u)", callchain->scc);
1912	return env->tmp_str_buf;
1913	}
1914
1915	/* If callchain for @st exists (@st is in some SCC), ensure that
1916	* bpf_scc_visit instance for this callchain exists.
1917	* If instance does not exist or is empty, assign visit->entry_state to @st.
1918	*/
1919	static int maybe_enter_scc(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
1920	{
1921	struct bpf_scc_callchain *callchain = &env->callchain_buf;
1922	struct bpf_scc_visit *visit;
1923
1924	if (!compute_scc_callchain(env, st, callchain))
1925	return 0;
1926	visit = scc_visit_lookup(env, callchain);
1927	visit = visit ?: scc_visit_alloc(env, callchain);
1928	if (!visit)
1929	return -ENOMEM;
1930	if (!visit->entry_state) {
1931	visit->entry_state = st;
1932	if (env->log.level & BPF_LOG_LEVEL2)
1933	verbose(private_data: env, fmt: "SCC enter %s\n", format_callchain(env, callchain));
1934	}
1935	return 0;
1936	}
1937
1938	static int propagate_backedges(struct bpf_verifier_env *env, struct bpf_scc_visit *visit);
1939
1940	/* If callchain for @st exists (@st is in some SCC), make it empty:
1941	* - set visit->entry_state to NULL;
1942	* - flush accumulated backedges.
1943	*/
1944	static int maybe_exit_scc(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
1945	{
1946	struct bpf_scc_callchain *callchain = &env->callchain_buf;
1947	struct bpf_scc_visit *visit;
1948
1949	if (!compute_scc_callchain(env, st, callchain))
1950	return 0;
1951	visit = scc_visit_lookup(env, callchain);
1952	if (!visit) {
1953	/*
1954	* If path traversal stops inside an SCC, corresponding bpf_scc_visit
1955	* must exist for non-speculative paths. For non-speculative paths
1956	* traversal stops when:
1957	* a. Verification error is found, maybe_exit_scc() is not called.
1958	* b. Top level BPF_EXIT is reached. Top level BPF_EXIT is not a member
1959	* of any SCC.
1960	* c. A checkpoint is reached and matched. Checkpoints are created by
1961	* is_state_visited(), which calls maybe_enter_scc(), which allocates
1962	* bpf_scc_visit instances for checkpoints within SCCs.
1963	* (c) is the only case that can reach this point.
1964	*/
1965	if (!st->speculative) {
1966	verifier_bug(env, "scc exit: no visit info for call chain %s",
1967	format_callchain(env, callchain));
1968	return -EFAULT;
1969	}
1970	return 0;
1971	}
1972	if (visit->entry_state != st)
1973	return 0;
1974	if (env->log.level & BPF_LOG_LEVEL2)
1975	verbose(private_data: env, fmt: "SCC exit %s\n", format_callchain(env, callchain));
1976	visit->entry_state = NULL;
1977	env->num_backedges -= visit->num_backedges;
1978	visit->num_backedges = 0;
1979	update_peak_states(env);
1980	return propagate_backedges(env, visit);
1981	}
1982
1983	/* Lookup an bpf_scc_visit instance corresponding to @st callchain
1984	* and add @backedge to visit->backedges. @st callchain must exist.
1985	*/
1986	static int add_scc_backedge(struct bpf_verifier_env *env,
1987	struct bpf_verifier_state *st,
1988	struct bpf_scc_backedge *backedge)
1989	{
1990	struct bpf_scc_callchain *callchain = &env->callchain_buf;
1991	struct bpf_scc_visit *visit;
1992
1993	if (!compute_scc_callchain(env, st, callchain)) {
1994	verifier_bug(env, "add backedge: no SCC in verification path, insn_idx %d",
1995	st->insn_idx);
1996	return -EFAULT;
1997	}
1998	visit = scc_visit_lookup(env, callchain);
1999	if (!visit) {
2000	verifier_bug(env, "add backedge: no visit info for call chain %s",
2001	format_callchain(env, callchain));
2002	return -EFAULT;
2003	}
2004	if (env->log.level & BPF_LOG_LEVEL2)
2005	verbose(private_data: env, fmt: "SCC backedge %s\n", format_callchain(env, callchain));
2006	backedge->next = visit->backedges;
2007	visit->backedges = backedge;
2008	visit->num_backedges++;
2009	env->num_backedges++;
2010	update_peak_states(env);
2011	return 0;
2012	}
2013
2014	/* bpf_reg_state->live marks for registers in a state @st are incomplete,
2015	* if state @st is in some SCC and not all execution paths starting at this
2016	* SCC are fully explored.
2017	*/
2018	static bool incomplete_read_marks(struct bpf_verifier_env *env,
2019	struct bpf_verifier_state *st)
2020	{
2021	struct bpf_scc_callchain *callchain = &env->callchain_buf;
2022	struct bpf_scc_visit *visit;
2023
2024	if (!compute_scc_callchain(env, st, callchain))
2025	return false;
2026	visit = scc_visit_lookup(env, callchain);
2027	if (!visit)
2028	return false;
2029	return !!visit->backedges;
2030	}
2031
2032	static void free_backedges(struct bpf_scc_visit *visit)
2033	{
2034	struct bpf_scc_backedge *backedge, *next;
2035
2036	for (backedge = visit->backedges; backedge; backedge = next) {
2037	free_verifier_state(state: &backedge->state, free_self: false);
2038	next = backedge->next;
2039	kfree(objp: backedge);
2040	}
2041	visit->backedges = NULL;
2042	}
2043
2044	static int update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
2045	{
2046	struct bpf_verifier_state_list *sl = NULL, *parent_sl;
2047	struct bpf_verifier_state *parent;
2048	int err;
2049
2050	while (st) {
2051	u32 br = --st->branches;
2052
2053	/* verifier_bug_if(br > 1, ...) technically makes sense here,
2054	* but see comment in push_stack(), hence:
2055	*/
2056	verifier_bug_if((int)br < 0, env, "%s:branches_to_explore=%d", __func__, br);
2057	if (br)
2058	break;
2059	err = maybe_exit_scc(env, st);
2060	if (err)
2061	return err;
2062	parent = st->parent;
2063	parent_sl = state_parent_as_list(st);
2064	if (sl)
2065	maybe_free_verifier_state(env, sl);
2066	st = parent;
2067	sl = parent_sl;
2068	}
2069	return 0;
2070	}
2071
2072	static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
2073	int *insn_idx, bool pop_log)
2074	{
2075	struct bpf_verifier_state *cur = env->cur_state;
2076	struct bpf_verifier_stack_elem *elem, *head = env->head;
2077	int err;
2078
2079	if (env->head == NULL)
2080	return -ENOENT;
2081
2082	if (cur) {
2083	err = copy_verifier_state(dst_state: cur, src: &head->st);
2084	if (err)
2085	return err;
2086	}
2087	if (pop_log)
2088	bpf_vlog_reset(log: &env->log, new_pos: head->log_pos);
2089	if (insn_idx)
2090	*insn_idx = head->insn_idx;
2091	if (prev_insn_idx)
2092	*prev_insn_idx = head->prev_insn_idx;
2093	elem = head->next;
2094	free_verifier_state(state: &head->st, free_self: false);
2095	kfree(objp: head);
2096	env->head = elem;
2097	env->stack_size--;
2098	return 0;
2099	}
2100
2101	static bool error_recoverable_with_nospec(int err)
2102	{
2103	/* Should only return true for non-fatal errors that are allowed to
2104	* occur during speculative verification. For these we can insert a
2105	* nospec and the program might still be accepted. Do not include
2106	* something like ENOMEM because it is likely to re-occur for the next
2107	* architectural path once it has been recovered-from in all speculative
2108	* paths.
2109	*/
2110	return err == -EPERM || err == -EACCES || err == -EINVAL;
2111	}
2112
2113	static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
2114	int insn_idx, int prev_insn_idx,
2115	bool speculative)
2116	{
2117	struct bpf_verifier_state *cur = env->cur_state;
2118	struct bpf_verifier_stack_elem *elem;
2119	int err;
2120
2121	elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL_ACCOUNT);
2122	if (!elem)
2123	return ERR_PTR(error: -ENOMEM);
2124
2125	elem->insn_idx = insn_idx;
2126	elem->prev_insn_idx = prev_insn_idx;
2127	elem->next = env->head;
2128	elem->log_pos = env->log.end_pos;
2129	env->head = elem;
2130	env->stack_size++;
2131	err = copy_verifier_state(dst_state: &elem->st, src: cur);
2132	if (err)
2133	return ERR_PTR(error: -ENOMEM);
2134	elem->st.speculative |= speculative;
2135	if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
2136	verbose(private_data: env, fmt: "The sequence of %d jumps is too complex.\n",
2137	env->stack_size);
2138	return ERR_PTR(error: -E2BIG);
2139	}
2140	if (elem->st.parent) {
2141	++elem->st.parent->branches;
2142	/* WARN_ON(branches > 2) technically makes sense here,
2143	* but
2144	* 1. speculative states will bump 'branches' for non-branch
2145	* instructions
2146	* 2. is_state_visited() heuristics may decide not to create
2147	* a new state for a sequence of branches and all such current
2148	* and cloned states will be pointing to a single parent state
2149	* which might have large 'branches' count.
2150	*/
2151	}
2152	return &elem->st;
2153	}
2154
2155	#define CALLER_SAVED_REGS 6
2156	static const int caller_saved[CALLER_SAVED_REGS] = {
2157	BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
2158	};
2159
2160	/* This helper doesn't clear reg->id */
2161	static void ___mark_reg_known(struct bpf_reg_state *reg, u64 imm)
2162	{
2163	reg->var_off = tnum_const(value: imm);
2164	reg->smin_value = (s64)imm;
2165	reg->smax_value = (s64)imm;
2166	reg->umin_value = imm;
2167	reg->umax_value = imm;
2168
2169	reg->s32_min_value = (s32)imm;
2170	reg->s32_max_value = (s32)imm;
2171	reg->u32_min_value = (u32)imm;
2172	reg->u32_max_value = (u32)imm;
2173	}
2174
2175	/* Mark the unknown part of a register (variable offset or scalar value) as
2176	* known to have the value @imm.
2177	*/
2178	static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
2179	{
2180	/* Clear off and union(map_ptr, range) */
2181	memset(((u8 *)reg) + sizeof(reg->type), 0,
2182	offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
2183	reg->id = 0;
2184	reg->ref_obj_id = 0;
2185	___mark_reg_known(reg, imm);
2186	}
2187
2188	static void __mark_reg32_known(struct bpf_reg_state *reg, u64 imm)
2189	{
2190	reg->var_off = tnum_const_subreg(a: reg->var_off, value: imm);
2191	reg->s32_min_value = (s32)imm;
2192	reg->s32_max_value = (s32)imm;
2193	reg->u32_min_value = (u32)imm;
2194	reg->u32_max_value = (u32)imm;
2195	}
2196
2197	/* Mark the 'variable offset' part of a register as zero. This should be
2198	* used only on registers holding a pointer type.
2199	*/
2200	static void __mark_reg_known_zero(struct bpf_reg_state *reg)
2201	{
2202	__mark_reg_known(reg, imm: 0);
2203	}
2204
2205	static void __mark_reg_const_zero(const struct bpf_verifier_env *env, struct bpf_reg_state *reg)
2206	{
2207	__mark_reg_known(reg, imm: 0);
2208	reg->type = SCALAR_VALUE;
2209	/* all scalars are assumed imprecise initially (unless unprivileged,
2210	* in which case everything is forced to be precise)
2211	*/
2212	reg->precise = !env->bpf_capable;
2213	}
2214
2215	static void mark_reg_known_zero(struct bpf_verifier_env *env,
2216	struct bpf_reg_state *regs, u32 regno)
2217	{
2218	if (WARN_ON(regno >= MAX_BPF_REG)) {
2219	verbose(private_data: env, fmt: "mark_reg_known_zero(regs, %u)\n", regno);
2220	/* Something bad happened, let's kill all regs */
2221	for (regno = 0; regno < MAX_BPF_REG; regno++)
2222	__mark_reg_not_init(env, reg: regs + regno);
2223	return;
2224	}
2225	__mark_reg_known_zero(reg: regs + regno);
2226	}
2227
2228	static void __mark_dynptr_reg(struct bpf_reg_state *reg, enum bpf_dynptr_type type,
2229	bool first_slot, int dynptr_id)
2230	{
2231	/* reg->type has no meaning for STACK_DYNPTR, but when we set reg for
2232	* callback arguments, it does need to be CONST_PTR_TO_DYNPTR, so simply
2233	* set it unconditionally as it is ignored for STACK_DYNPTR anyway.
2234	*/
2235	__mark_reg_known_zero(reg);
2236	reg->type = CONST_PTR_TO_DYNPTR;
2237	/* Give each dynptr a unique id to uniquely associate slices to it. */
2238	reg->id = dynptr_id;
2239	reg->dynptr.type = type;
2240	reg->dynptr.first_slot = first_slot;
2241	}
2242
2243	static void mark_ptr_not_null_reg(struct bpf_reg_state *reg)
2244	{
2245	if (base_type(type: reg->type) == PTR_TO_MAP_VALUE) {
2246	const struct bpf_map *map = reg->map_ptr;
2247
2248	if (map->inner_map_meta) {
2249	reg->type = CONST_PTR_TO_MAP;
2250	reg->map_ptr = map->inner_map_meta;
2251	/* transfer reg's id which is unique for every map_lookup_elem
2252	* as UID of the inner map.
2253	*/
2254	if (btf_record_has_field(rec: map->inner_map_meta->record,
2255	type: BPF_TIMER | BPF_WORKQUEUE | BPF_TASK_WORK)) {
2256	reg->map_uid = reg->id;
2257	}
2258	} else if (map->map_type == BPF_MAP_TYPE_XSKMAP) {
2259	reg->type = PTR_TO_XDP_SOCK;
2260	} else if (map->map_type == BPF_MAP_TYPE_SOCKMAP ||
2261	map->map_type == BPF_MAP_TYPE_SOCKHASH) {
2262	reg->type = PTR_TO_SOCKET;
2263	} else {
2264	reg->type = PTR_TO_MAP_VALUE;
2265	}
2266	return;
2267	}
2268
2269	reg->type &= ~PTR_MAYBE_NULL;
2270	}
2271
2272	static void mark_reg_graph_node(struct bpf_reg_state *regs, u32 regno,
2273	struct btf_field_graph_root *ds_head)
2274	{
2275	__mark_reg_known_zero(reg: &regs[regno]);
2276	regs[regno].type = PTR_TO_BTF_ID | MEM_ALLOC;
2277	regs[regno].btf = ds_head->btf;
2278	regs[regno].btf_id = ds_head->value_btf_id;
2279	regs[regno].off = ds_head->node_offset;
2280	}
2281
2282	static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
2283	{
2284	return type_is_pkt_pointer(type: reg->type);
2285	}
2286
2287	static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
2288	{
2289	return reg_is_pkt_pointer(reg) ||
2290	reg->type == PTR_TO_PACKET_END;
2291	}
2292
2293	static bool reg_is_dynptr_slice_pkt(const struct bpf_reg_state *reg)
2294	{
2295	return base_type(type: reg->type) == PTR_TO_MEM &&
2296	(reg->type &
2297	(DYNPTR_TYPE_SKB | DYNPTR_TYPE_XDP | DYNPTR_TYPE_SKB_META));
2298	}
2299
2300	/* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
2301	static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
2302	enum bpf_reg_type which)
2303	{
2304	/* The register can already have a range from prior markings.
2305	* This is fine as long as it hasn't been advanced from its
2306	* origin.
2307	*/
2308	return reg->type == which &&
2309	reg->id == 0 &&
2310	reg->off == 0 &&
2311	tnum_equals_const(a: reg->var_off, b: 0);
2312	}
2313
2314	/* Reset the min/max bounds of a register */
2315	static void __mark_reg_unbounded(struct bpf_reg_state *reg)
2316	{
2317	reg->smin_value = S64_MIN;
2318	reg->smax_value = S64_MAX;
2319	reg->umin_value = 0;
2320	reg->umax_value = U64_MAX;
2321
2322	reg->s32_min_value = S32_MIN;
2323	reg->s32_max_value = S32_MAX;
2324	reg->u32_min_value = 0;
2325	reg->u32_max_value = U32_MAX;
2326	}
2327
2328	static void __mark_reg64_unbounded(struct bpf_reg_state *reg)
2329	{
2330	reg->smin_value = S64_MIN;
2331	reg->smax_value = S64_MAX;
2332	reg->umin_value = 0;
2333	reg->umax_value = U64_MAX;
2334	}
2335
2336	static void __mark_reg32_unbounded(struct bpf_reg_state *reg)
2337	{
2338	reg->s32_min_value = S32_MIN;
2339	reg->s32_max_value = S32_MAX;
2340	reg->u32_min_value = 0;
2341	reg->u32_max_value = U32_MAX;
2342	}
2343
2344	static void __update_reg32_bounds(struct bpf_reg_state *reg)
2345	{
2346	struct tnum var32_off = tnum_subreg(a: reg->var_off);
2347
2348	/* min signed is max(sign bit) | min(other bits) */
2349	reg->s32_min_value = max_t(s32, reg->s32_min_value,
2350	var32_off.value | (var32_off.mask & S32_MIN));
2351	/* max signed is min(sign bit) | max(other bits) */
2352	reg->s32_max_value = min_t(s32, reg->s32_max_value,
2353	var32_off.value | (var32_off.mask & S32_MAX));
2354	reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)var32_off.value);
2355	reg->u32_max_value = min(reg->u32_max_value,
2356	(u32)(var32_off.value | var32_off.mask));
2357	}
2358
2359	static void __update_reg64_bounds(struct bpf_reg_state *reg)
2360	{
2361	/* min signed is max(sign bit) | min(other bits) */
2362	reg->smin_value = max_t(s64, reg->smin_value,
2363	reg->var_off.value | (reg->var_off.mask & S64_MIN));
2364	/* max signed is min(sign bit) | max(other bits) */
2365	reg->smax_value = min_t(s64, reg->smax_value,
2366	reg->var_off.value | (reg->var_off.mask & S64_MAX));
2367	reg->umin_value = max(reg->umin_value, reg->var_off.value);
2368	reg->umax_value = min(reg->umax_value,
2369	reg->var_off.value | reg->var_off.mask);
2370	}
2371
2372	static void __update_reg_bounds(struct bpf_reg_state *reg)
2373	{
2374	__update_reg32_bounds(reg);
2375	__update_reg64_bounds(reg);
2376	}
2377
2378	/* Uses signed min/max values to inform unsigned, and vice-versa */
2379	static void __reg32_deduce_bounds(struct bpf_reg_state *reg)
2380	{
2381	/* If upper 32 bits of u64/s64 range don't change, we can use lower 32
2382	* bits to improve our u32/s32 boundaries.
2383	*
2384	* E.g., the case where we have upper 32 bits as zero ([10, 20] in
2385	* u64) is pretty trivial, it's obvious that in u32 we'll also have
2386	* [10, 20] range. But this property holds for any 64-bit range as
2387	* long as upper 32 bits in that entire range of values stay the same.
2388	*
2389	* E.g., u64 range [0x10000000A, 0x10000000F] ([4294967306, 4294967311]
2390	* in decimal) has the same upper 32 bits throughout all the values in
2391	* that range. As such, lower 32 bits form a valid [0xA, 0xF] ([10, 15])
2392	* range.
2393	*
2394	* Note also, that [0xA, 0xF] is a valid range both in u32 and in s32,
2395	* following the rules outlined below about u64/s64 correspondence
2396	* (which equally applies to u32 vs s32 correspondence). In general it
2397	* depends on actual hexadecimal values of 32-bit range. They can form
2398	* only valid u32, or only valid s32 ranges in some cases.
2399	*
2400	* So we use all these insights to derive bounds for subregisters here.
2401	*/
2402	if ((reg->umin_value >> 32) == (reg->umax_value >> 32)) {
2403	/* u64 to u32 casting preserves validity of low 32 bits as
2404	* a range, if upper 32 bits are the same
2405	*/
2406	reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)reg->umin_value);
2407	reg->u32_max_value = min_t(u32, reg->u32_max_value, (u32)reg->umax_value);
2408
2409	if ((s32)reg->umin_value <= (s32)reg->umax_value) {
2410	reg->s32_min_value = max_t(s32, reg->s32_min_value, (s32)reg->umin_value);
2411	reg->s32_max_value = min_t(s32, reg->s32_max_value, (s32)reg->umax_value);
2412	}
2413	}
2414	if ((reg->smin_value >> 32) == (reg->smax_value >> 32)) {
2415	/* low 32 bits should form a proper u32 range */
2416	if ((u32)reg->smin_value <= (u32)reg->smax_value) {
2417	reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)reg->smin_value);
2418	reg->u32_max_value = min_t(u32, reg->u32_max_value, (u32)reg->smax_value);
2419	}
2420	/* low 32 bits should form a proper s32 range */
2421	if ((s32)reg->smin_value <= (s32)reg->smax_value) {
2422	reg->s32_min_value = max_t(s32, reg->s32_min_value, (s32)reg->smin_value);
2423	reg->s32_max_value = min_t(s32, reg->s32_max_value, (s32)reg->smax_value);
2424	}
2425	}
2426	/* Special case where upper bits form a small sequence of two
2427	* sequential numbers (in 32-bit unsigned space, so 0xffffffff to
2428	* 0x00000000 is also valid), while lower bits form a proper s32 range
2429	* going from negative numbers to positive numbers. E.g., let's say we
2430	* have s64 range [-1, 1] ([0xffffffffffffffff, 0x0000000000000001]).
2431	* Possible s64 values are {-1, 0, 1} ({0xffffffffffffffff,
2432	* 0x0000000000000000, 0x00000000000001}). Ignoring upper 32 bits,
2433	* we still get a valid s32 range [-1, 1] ([0xffffffff, 0x00000001]).
2434	* Note that it doesn't have to be 0xffffffff going to 0x00000000 in
2435	* upper 32 bits. As a random example, s64 range
2436	* [0xfffffff0fffffff0; 0xfffffff100000010], forms a valid s32 range
2437	* [-16, 16] ([0xfffffff0; 0x00000010]) in its 32 bit subregister.
2438	*/
2439	if ((u32)(reg->umin_value >> 32) + 1 == (u32)(reg->umax_value >> 32) &&
2440	(s32)reg->umin_value < 0 && (s32)reg->umax_value >= 0) {
2441	reg->s32_min_value = max_t(s32, reg->s32_min_value, (s32)reg->umin_value);
2442	reg->s32_max_value = min_t(s32, reg->s32_max_value, (s32)reg->umax_value);
2443	}
2444	if ((u32)(reg->smin_value >> 32) + 1 == (u32)(reg->smax_value >> 32) &&
2445	(s32)reg->smin_value < 0 && (s32)reg->smax_value >= 0) {
2446	reg->s32_min_value = max_t(s32, reg->s32_min_value, (s32)reg->smin_value);
2447	reg->s32_max_value = min_t(s32, reg->s32_max_value, (s32)reg->smax_value);
2448	}
2449	/* if u32 range forms a valid s32 range (due to matching sign bit),
2450	* try to learn from that
2451	*/
2452	if ((s32)reg->u32_min_value <= (s32)reg->u32_max_value) {
2453	reg->s32_min_value = max_t(s32, reg->s32_min_value, reg->u32_min_value);
2454	reg->s32_max_value = min_t(s32, reg->s32_max_value, reg->u32_max_value);
2455	}
2456	/* If we cannot cross the sign boundary, then signed and unsigned bounds
2457	* are the same, so combine. This works even in the negative case, e.g.
2458	* -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
2459	*/
2460	if ((u32)reg->s32_min_value <= (u32)reg->s32_max_value) {
2461	reg->u32_min_value = max_t(u32, reg->s32_min_value, reg->u32_min_value);
2462	reg->u32_max_value = min_t(u32, reg->s32_max_value, reg->u32_max_value);
2463	}
2464	}
2465
2466	static void __reg64_deduce_bounds(struct bpf_reg_state *reg)
2467	{
2468	/* If u64 range forms a valid s64 range (due to matching sign bit),
2469	* try to learn from that. Let's do a bit of ASCII art to see when
2470	* this is happening. Let's take u64 range first:
2471	*
2472	* 0 0x7fffffffffffffff 0x8000000000000000 U64_MAX
2473	* |-------------------------------|--------------------------------|
2474	*
2475	* Valid u64 range is formed when umin and umax are anywhere in the
2476	* range [0, U64_MAX], and umin <= umax. u64 case is simple and
2477	* straightforward. Let's see how s64 range maps onto the same range
2478	* of values, annotated below the line for comparison:
2479	*
2480	* 0 0x7fffffffffffffff 0x8000000000000000 U64_MAX
2481	* |-------------------------------|--------------------------------|
2482	* 0 S64_MAX S64_MIN -1
2483	*
2484	* So s64 values basically start in the middle and they are logically
2485	* contiguous to the right of it, wrapping around from -1 to 0, and
2486	* then finishing as S64_MAX (0x7fffffffffffffff) right before
2487	* S64_MIN. We can try drawing the continuity of u64 vs s64 values
2488	* more visually as mapped to sign-agnostic range of hex values.
2489	*
2490	* u64 start u64 end
2491	* _______________________________________________________________
2492	* / \
2493	* 0 0x7fffffffffffffff 0x8000000000000000 U64_MAX
2494	* |-------------------------------|--------------------------------|
2495	* 0 S64_MAX S64_MIN -1
2496	* / \
2497	* >------------------------------ ------------------------------->
2498	* s64 continues... s64 end s64 start s64 "midpoint"
2499	*
2500	* What this means is that, in general, we can't always derive
2501	* something new about u64 from any random s64 range, and vice versa.
2502	*
2503	* But we can do that in two particular cases. One is when entire
2504	* u64/s64 range is *entirely* contained within left half of the above
2505	* diagram or when it is *entirely* contained in the right half. I.e.:
2506	*
2507	* |-------------------------------|--------------------------------|
2508	* ^ ^ ^ ^
2509	* A B C D
2510	*
2511	* [A, B] and [C, D] are contained entirely in their respective halves
2512	* and form valid contiguous ranges as both u64 and s64 values. [A, B]
2513	* will be non-negative both as u64 and s64 (and in fact it will be
2514	* identical ranges no matter the signedness). [C, D] treated as s64
2515	* will be a range of negative values, while in u64 it will be
2516	* non-negative range of values larger than 0x8000000000000000.
2517	*
2518	* Now, any other range here can't be represented in both u64 and s64
2519	* simultaneously. E.g., [A, C], [A, D], [B, C], [B, D] are valid
2520	* contiguous u64 ranges, but they are discontinuous in s64. [B, C]
2521	* in s64 would be properly presented as [S64_MIN, C] and [B, S64_MAX],
2522	* for example. Similarly, valid s64 range [D, A] (going from negative
2523	* to positive values), would be two separate [D, U64_MAX] and [0, A]
2524	* ranges as u64. Currently reg_state can't represent two segments per
2525	* numeric domain, so in such situations we can only derive maximal
2526	* possible range ([0, U64_MAX] for u64, and [S64_MIN, S64_MAX] for s64).
2527	*
2528	* So we use these facts to derive umin/umax from smin/smax and vice
2529	* versa only if they stay within the same "half". This is equivalent
2530	* to checking sign bit: lower half will have sign bit as zero, upper
2531	* half have sign bit 1. Below in code we simplify this by just
2532	* casting umin/umax as smin/smax and checking if they form valid
2533	* range, and vice versa. Those are equivalent checks.
2534	*/
2535	if ((s64)reg->umin_value <= (s64)reg->umax_value) {
2536	reg->smin_value = max_t(s64, reg->smin_value, reg->umin_value);
2537	reg->smax_value = min_t(s64, reg->smax_value, reg->umax_value);
2538	}
2539	/* If we cannot cross the sign boundary, then signed and unsigned bounds
2540	* are the same, so combine. This works even in the negative case, e.g.
2541	* -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
2542	*/
2543	if ((u64)reg->smin_value <= (u64)reg->smax_value) {
2544	reg->umin_value = max_t(u64, reg->smin_value, reg->umin_value);
2545	reg->umax_value = min_t(u64, reg->smax_value, reg->umax_value);
2546	} else {
2547	/* If the s64 range crosses the sign boundary, then it's split
2548	* between the beginning and end of the U64 domain. In that
2549	* case, we can derive new bounds if the u64 range overlaps
2550	* with only one end of the s64 range.
2551	*
2552	* In the following example, the u64 range overlaps only with
2553	* positive portion of the s64 range.
2554	*
2555	* 0 U64_MAX
2556	* | [xxxxxxxxxxxxxx u64 range xxxxxxxxxxxxxx] |
2557	* |----------------------------|----------------------------|
2558	* |xxxxx s64 range xxxxxxxxx] [xxxxxxx|
2559	* 0 S64_MAX S64_MIN -1
2560	*
2561	* We can thus derive the following new s64 and u64 ranges.
2562	*
2563	* 0 U64_MAX
2564	* | [xxxxxx u64 range xxxxx] |
2565	* |----------------------------|----------------------------|
2566	* | [xxxxxx s64 range xxxxx] |
2567	* 0 S64_MAX S64_MIN -1
2568	*
2569	* If they overlap in two places, we can't derive anything
2570	* because reg_state can't represent two ranges per numeric
2571	* domain.
2572	*
2573	* 0 U64_MAX
2574	* | [xxxxxxxxxxxxxxxxx u64 range xxxxxxxxxxxxxxxxx] |
2575	* |----------------------------|----------------------------|
2576	* |xxxxx s64 range xxxxxxxxx] [xxxxxxxxxx|
2577	* 0 S64_MAX S64_MIN -1
2578	*
2579	* The first condition below corresponds to the first diagram
2580	* above.
2581	*/
2582	if (reg->umax_value < (u64)reg->smin_value) {
2583	reg->smin_value = (s64)reg->umin_value;
2584	reg->umax_value = min_t(u64, reg->umax_value, reg->smax_value);
2585	} else if ((u64)reg->smax_value < reg->umin_value) {
2586	/* This second condition considers the case where the u64 range
2587	* overlaps with the negative portion of the s64 range:
2588	*
2589	* 0 U64_MAX
2590	* | [xxxxxxxxxxxxxx u64 range xxxxxxxxxxxxxx] |
2591	* |----------------------------|----------------------------|
2592	* |xxxxxxxxx] [xxxxxxxxxxxx s64 range |
2593	* 0 S64_MAX S64_MIN -1
2594	*/
2595	reg->smax_value = (s64)reg->umax_value;
2596	reg->umin_value = max_t(u64, reg->umin_value, reg->smin_value);
2597	}
2598	}
2599	}
2600
2601	static void __reg_deduce_mixed_bounds(struct bpf_reg_state *reg)
2602	{
2603	/* Try to tighten 64-bit bounds from 32-bit knowledge, using 32-bit
2604	* values on both sides of 64-bit range in hope to have tighter range.
2605	* E.g., if r1 is [0x1'00000000, 0x3'80000000], and we learn from
2606	* 32-bit signed > 0 operation that s32 bounds are now [1; 0x7fffffff].
2607	* With this, we can substitute 1 as low 32-bits of _low_ 64-bit bound
2608	* (0x100000000 -> 0x100000001) and 0x7fffffff as low 32-bits of
2609	* _high_ 64-bit bound (0x380000000 -> 0x37fffffff) and arrive at a
2610	* better overall bounds for r1 as [0x1'000000001; 0x3'7fffffff].
2611	* We just need to make sure that derived bounds we are intersecting
2612	* with are well-formed ranges in respective s64 or u64 domain, just
2613	* like we do with similar kinds of 32-to-64 or 64-to-32 adjustments.
2614	*/
2615	__u64 new_umin, new_umax;
2616	__s64 new_smin, new_smax;
2617
2618	/* u32 -> u64 tightening, it's always well-formed */
2619	new_umin = (reg->umin_value & ~0xffffffffULL) | reg->u32_min_value;
2620	new_umax = (reg->umax_value & ~0xffffffffULL) | reg->u32_max_value;
2621	reg->umin_value = max_t(u64, reg->umin_value, new_umin);
2622	reg->umax_value = min_t(u64, reg->umax_value, new_umax);
2623	/* u32 -> s64 tightening, u32 range embedded into s64 preserves range validity */
2624	new_smin = (reg->smin_value & ~0xffffffffULL) | reg->u32_min_value;
2625	new_smax = (reg->smax_value & ~0xffffffffULL) | reg->u32_max_value;
2626	reg->smin_value = max_t(s64, reg->smin_value, new_smin);
2627	reg->smax_value = min_t(s64, reg->smax_value, new_smax);
2628
2629	/* Here we would like to handle a special case after sign extending load,
2630	* when upper bits for a 64-bit range are all 1s or all 0s.
2631	*
2632	* Upper bits are all 1s when register is in a range:
2633	* [0xffff_ffff_0000_0000, 0xffff_ffff_ffff_ffff]
2634	* Upper bits are all 0s when register is in a range:
2635	* [0x0000_0000_0000_0000, 0x0000_0000_ffff_ffff]
2636	* Together this forms are continuous range:
2637	* [0xffff_ffff_0000_0000, 0x0000_0000_ffff_ffff]
2638	*
2639	* Now, suppose that register range is in fact tighter:
2640	* [0xffff_ffff_8000_0000, 0x0000_0000_ffff_ffff] (R)
2641	* Also suppose that it's 32-bit range is positive,
2642	* meaning that lower 32-bits of the full 64-bit register
2643	* are in the range:
2644	* [0x0000_0000, 0x7fff_ffff] (W)
2645	*
2646	* If this happens, then any value in a range:
2647	* [0xffff_ffff_0000_0000, 0xffff_ffff_7fff_ffff]
2648	* is smaller than a lowest bound of the range (R):
2649	* 0xffff_ffff_8000_0000
2650	* which means that upper bits of the full 64-bit register
2651	* can't be all 1s, when lower bits are in range (W).
2652	*
2653	* Note that:
2654	* - 0xffff_ffff_8000_0000 == (s64)S32_MIN
2655	* - 0x0000_0000_7fff_ffff == (s64)S32_MAX
2656	* These relations are used in the conditions below.
2657	*/
2658	if (reg->s32_min_value >= 0 && reg->smin_value >= S32_MIN && reg->smax_value <= S32_MAX) {
2659	reg->smin_value = reg->s32_min_value;
2660	reg->smax_value = reg->s32_max_value;
2661	reg->umin_value = reg->s32_min_value;
2662	reg->umax_value = reg->s32_max_value;
2663	reg->var_off = tnum_intersect(a: reg->var_off,
2664	b: tnum_range(min: reg->smin_value, max: reg->smax_value));
2665	}
2666	}
2667
2668	static void __reg_deduce_bounds(struct bpf_reg_state *reg)
2669	{
2670	__reg32_deduce_bounds(reg);
2671	__reg64_deduce_bounds(reg);
2672	__reg_deduce_mixed_bounds(reg);
2673	}
2674
2675	/* Attempts to improve var_off based on unsigned min/max information */
2676	static void __reg_bound_offset(struct bpf_reg_state *reg)
2677	{
2678	struct tnum var64_off = tnum_intersect(a: reg->var_off,
2679	b: tnum_range(min: reg->umin_value,
2680	max: reg->umax_value));
2681	struct tnum var32_off = tnum_intersect(a: tnum_subreg(a: var64_off),
2682	b: tnum_range(min: reg->u32_min_value,
2683	max: reg->u32_max_value));
2684
2685	reg->var_off = tnum_or(a: tnum_clear_subreg(a: var64_off), b: var32_off);
2686	}
2687
2688	static void reg_bounds_sync(struct bpf_reg_state *reg)
2689	{
2690	/* We might have learned new bounds from the var_off. */
2691	__update_reg_bounds(reg);
2692	/* We might have learned something about the sign bit. */
2693	__reg_deduce_bounds(reg);
2694	__reg_deduce_bounds(reg);
2695	__reg_deduce_bounds(reg);
2696	/* We might have learned some bits from the bounds. */
2697	__reg_bound_offset(reg);
2698	/* Intersecting with the old var_off might have improved our bounds
2699	* slightly, e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
2700	* then new var_off is (0; 0x7f...fc) which improves our umax.
2701	*/
2702	__update_reg_bounds(reg);
2703	}
2704
2705	static int reg_bounds_sanity_check(struct bpf_verifier_env *env,
2706	struct bpf_reg_state *reg, const char *ctx)
2707	{
2708	const char *msg;
2709
2710	if (reg->umin_value > reg->umax_value ||
2711	reg->smin_value > reg->smax_value ||
2712	reg->u32_min_value > reg->u32_max_value ||
2713	reg->s32_min_value > reg->s32_max_value) {
2714	msg = "range bounds violation";
2715	goto out;
2716	}
2717
2718	if (tnum_is_const(a: reg->var_off)) {
2719	u64 uval = reg->var_off.value;
2720	s64 sval = (s64)uval;
2721
2722	if (reg->umin_value != uval || reg->umax_value != uval ||
2723	reg->smin_value != sval || reg->smax_value != sval) {
2724	msg = "const tnum out of sync with range bounds";
2725	goto out;
2726	}
2727	}
2728
2729	if (tnum_subreg_is_const(a: reg->var_off)) {
2730	u32 uval32 = tnum_subreg(a: reg->var_off).value;
2731	s32 sval32 = (s32)uval32;
2732
2733	if (reg->u32_min_value != uval32 || reg->u32_max_value != uval32 ||
2734	reg->s32_min_value != sval32 || reg->s32_max_value != sval32) {
2735	msg = "const subreg tnum out of sync with range bounds";
2736	goto out;
2737	}
2738	}
2739
2740	return 0;
2741	out:
2742	verifier_bug(env, "REG INVARIANTS VIOLATION (%s): %s u64=[%#llx, %#llx] "
2743	"s64=[%#llx, %#llx] u32=[%#x, %#x] s32=[%#x, %#x] var_off=(%#llx, %#llx)",
2744	ctx, msg, reg->umin_value, reg->umax_value,
2745	reg->smin_value, reg->smax_value,
2746	reg->u32_min_value, reg->u32_max_value,
2747	reg->s32_min_value, reg->s32_max_value,
2748	reg->var_off.value, reg->var_off.mask);
2749	if (env->test_reg_invariants)
2750	return -EFAULT;
2751	__mark_reg_unbounded(reg);
2752	return 0;
2753	}
2754
2755	static bool __reg32_bound_s64(s32 a)
2756	{
2757	return a >= 0 && a <= S32_MAX;
2758	}
2759
2760	static void __reg_assign_32_into_64(struct bpf_reg_state *reg)
2761	{
2762	reg->umin_value = reg->u32_min_value;
2763	reg->umax_value = reg->u32_max_value;
2764
2765	/* Attempt to pull 32-bit signed bounds into 64-bit bounds but must
2766	* be positive otherwise set to worse case bounds and refine later
2767	* from tnum.
2768	*/
2769	if (__reg32_bound_s64(a: reg->s32_min_value) &&
2770	__reg32_bound_s64(a: reg->s32_max_value)) {
2771	reg->smin_value = reg->s32_min_value;
2772	reg->smax_value = reg->s32_max_value;
2773	} else {
2774	reg->smin_value = 0;
2775	reg->smax_value = U32_MAX;
2776	}
2777	}
2778
2779	/* Mark a register as having a completely unknown (scalar) value. */
2780	static void __mark_reg_unknown_imprecise(struct bpf_reg_state *reg)
2781	{
2782	/*
2783	* Clear type, off, and union(map_ptr, range) and
2784	* padding between 'type' and union
2785	*/
2786	memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
2787	reg->type = SCALAR_VALUE;
2788	reg->id = 0;
2789	reg->ref_obj_id = 0;
2790	reg->var_off = tnum_unknown;
2791	reg->frameno = 0;
2792	reg->precise = false;
2793	__mark_reg_unbounded(reg);
2794	}
2795
2796	/* Mark a register as having a completely unknown (scalar) value,
2797	* initialize .precise as true when not bpf capable.
2798	*/
2799	static void __mark_reg_unknown(const struct bpf_verifier_env *env,
2800	struct bpf_reg_state *reg)
2801	{
2802	__mark_reg_unknown_imprecise(reg);
2803	reg->precise = !env->bpf_capable;
2804	}
2805
2806	static void mark_reg_unknown(struct bpf_verifier_env *env,
2807	struct bpf_reg_state *regs, u32 regno)
2808	{
2809	if (WARN_ON(regno >= MAX_BPF_REG)) {
2810	verbose(private_data: env, fmt: "mark_reg_unknown(regs, %u)\n", regno);
2811	/* Something bad happened, let's kill all regs except FP */
2812	for (regno = 0; regno < BPF_REG_FP; regno++)
2813	__mark_reg_not_init(env, reg: regs + regno);
2814	return;
2815	}
2816	__mark_reg_unknown(env, reg: regs + regno);
2817	}
2818
2819	static int __mark_reg_s32_range(struct bpf_verifier_env *env,
2820	struct bpf_reg_state *regs,
2821	u32 regno,
2822	s32 s32_min,
2823	s32 s32_max)
2824	{
2825	struct bpf_reg_state *reg = regs + regno;
2826
2827	reg->s32_min_value = max_t(s32, reg->s32_min_value, s32_min);
2828	reg->s32_max_value = min_t(s32, reg->s32_max_value, s32_max);
2829
2830	reg->smin_value = max_t(s64, reg->smin_value, s32_min);
2831	reg->smax_value = min_t(s64, reg->smax_value, s32_max);
2832
2833	reg_bounds_sync(reg);
2834
2835	return reg_bounds_sanity_check(env, reg, ctx: "s32_range");
2836	}
2837
2838	static void __mark_reg_not_init(const struct bpf_verifier_env *env,
2839	struct bpf_reg_state *reg)
2840	{
2841	__mark_reg_unknown(env, reg);
2842	reg->type = NOT_INIT;
2843	}
2844
2845	static void mark_reg_not_init(struct bpf_verifier_env *env,
2846	struct bpf_reg_state *regs, u32 regno)
2847	{
2848	if (WARN_ON(regno >= MAX_BPF_REG)) {
2849	verbose(private_data: env, fmt: "mark_reg_not_init(regs, %u)\n", regno);
2850	/* Something bad happened, let's kill all regs except FP */
2851	for (regno = 0; regno < BPF_REG_FP; regno++)
2852	__mark_reg_not_init(env, reg: regs + regno);
2853	return;
2854	}
2855	__mark_reg_not_init(env, reg: regs + regno);
2856	}
2857
2858	static int mark_btf_ld_reg(struct bpf_verifier_env *env,
2859	struct bpf_reg_state *regs, u32 regno,
2860	enum bpf_reg_type reg_type,
2861	struct btf *btf, u32 btf_id,
2862	enum bpf_type_flag flag)
2863	{
2864	switch (reg_type) {
2865	case SCALAR_VALUE:
2866	mark_reg_unknown(env, regs, regno);
2867	return 0;
2868	case PTR_TO_BTF_ID:
2869	mark_reg_known_zero(env, regs, regno);
2870	regs[regno].type = PTR_TO_BTF_ID | flag;
2871	regs[regno].btf = btf;
2872	regs[regno].btf_id = btf_id;
2873	if (type_may_be_null(type: flag))
2874	regs[regno].id = ++env->id_gen;
2875	return 0;
2876	case PTR_TO_MEM:
2877	mark_reg_known_zero(env, regs, regno);
2878	regs[regno].type = PTR_TO_MEM | flag;
2879	regs[regno].mem_size = 0;
2880	return 0;
2881	default:
2882	verifier_bug(env, "unexpected reg_type %d in %s\n", reg_type, __func__);
2883	return -EFAULT;
2884	}
2885	}
2886
2887	#define DEF_NOT_SUBREG (0)
2888	static void init_reg_state(struct bpf_verifier_env *env,
2889	struct bpf_func_state *state)
2890	{
2891	struct bpf_reg_state *regs = state->regs;
2892	int i;
2893
2894	for (i = 0; i < MAX_BPF_REG; i++) {
2895	mark_reg_not_init(env, regs, regno: i);
2896	regs[i].subreg_def = DEF_NOT_SUBREG;
2897	}
2898
2899	/* frame pointer */
2900	regs[BPF_REG_FP].type = PTR_TO_STACK;
2901	mark_reg_known_zero(env, regs, BPF_REG_FP);
2902	regs[BPF_REG_FP].frameno = state->frameno;
2903	}
2904
2905	static struct bpf_retval_range retval_range(s32 minval, s32 maxval)
2906	{
2907	return (struct bpf_retval_range){ minval, maxval };
2908	}
2909
2910	#define BPF_MAIN_FUNC (-1)
2911	static void init_func_state(struct bpf_verifier_env *env,
2912	struct bpf_func_state *state,
2913	int callsite, int frameno, int subprogno)
2914	{
2915	state->callsite = callsite;
2916	state->frameno = frameno;
2917	state->subprogno = subprogno;
2918	state->callback_ret_range = retval_range(minval: 0, maxval: 0);
2919	init_reg_state(env, state);
2920	mark_verifier_state_scratched(env);
2921	}
2922
2923	/* Similar to push_stack(), but for async callbacks */
2924	static struct bpf_verifier_state *push_async_cb(struct bpf_verifier_env *env,
2925	int insn_idx, int prev_insn_idx,
2926	int subprog, bool is_sleepable)
2927	{
2928	struct bpf_verifier_stack_elem *elem;
2929	struct bpf_func_state *frame;
2930
2931	elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL_ACCOUNT);
2932	if (!elem)
2933	return ERR_PTR(error: -ENOMEM);
2934
2935	elem->insn_idx = insn_idx;
2936	elem->prev_insn_idx = prev_insn_idx;
2937	elem->next = env->head;
2938	elem->log_pos = env->log.end_pos;
2939	env->head = elem;
2940	env->stack_size++;
2941	if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
2942	verbose(private_data: env,
2943	fmt: "The sequence of %d jumps is too complex for async cb.\n",
2944	env->stack_size);
2945	return ERR_PTR(error: -E2BIG);
2946	}
2947	/* Unlike push_stack() do not copy_verifier_state().
2948	* The caller state doesn't matter.
2949	* This is async callback. It starts in a fresh stack.
2950	* Initialize it similar to do_check_common().
2951	*/
2952	elem->st.branches = 1;
2953	elem->st.in_sleepable = is_sleepable;
2954	frame = kzalloc(sizeof(*frame), GFP_KERNEL_ACCOUNT);
2955	if (!frame)
2956	return ERR_PTR(error: -ENOMEM);
2957	init_func_state(env, state: frame,
2958	BPF_MAIN_FUNC /* callsite */,
2959	frameno: 0 /* frameno within this callchain */,
2960	subprogno: subprog /* subprog number within this prog */);
2961	elem->st.frame[0] = frame;
2962	return &elem->st;
2963	}
2964
2965
2966	enum reg_arg_type {
2967	SRC_OP, /* register is used as source operand */
2968	DST_OP, /* register is used as destination operand */
2969	DST_OP_NO_MARK /* same as above, check only, don't mark */
2970	};
2971
2972	static int cmp_subprogs(const void *a, const void *b)
2973	{
2974	return ((struct bpf_subprog_info *)a)->start -
2975	((struct bpf_subprog_info *)b)->start;
2976	}
2977
2978	/* Find subprogram that contains instruction at 'off' */
2979	struct bpf_subprog_info *bpf_find_containing_subprog(struct bpf_verifier_env *env, int off)
2980	{
2981	struct bpf_subprog_info *vals = env->subprog_info;
2982	int l, r, m;
2983
2984	if (off >= env->prog->len || off < 0 || env->subprog_cnt == 0)
2985	return NULL;
2986
2987	l = 0;
2988	r = env->subprog_cnt - 1;
2989	while (l < r) {
2990	m = l + (r - l + 1) / 2;
2991	if (vals[m].start <= off)
2992	l = m;
2993	else
2994	r = m - 1;
2995	}
2996	return &vals[l];
2997	}
2998
2999	/* Find subprogram that starts exactly at 'off' */
3000	static int find_subprog(struct bpf_verifier_env *env, int off)
3001	{
3002	struct bpf_subprog_info *p;
3003
3004	p = bpf_find_containing_subprog(env, off);
3005	if (!p || p->start != off)
3006	return -ENOENT;
3007	return p - env->subprog_info;
3008	}
3009
3010	static int add_subprog(struct bpf_verifier_env *env, int off)
3011	{
3012	int insn_cnt = env->prog->len;
3013	int ret;
3014
3015	if (off >= insn_cnt || off < 0) {
3016	verbose(private_data: env, fmt: "call to invalid destination\n");
3017	return -EINVAL;
3018	}
3019	ret = find_subprog(env, off);
3020	if (ret >= 0)
3021	return ret;
3022	if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
3023	verbose(private_data: env, fmt: "too many subprograms\n");
3024	return -E2BIG;
3025	}
3026	/* determine subprog starts. The end is one before the next starts */
3027	env->subprog_info[env->subprog_cnt++].start = off;
3028	sort(base: env->subprog_info, num: env->subprog_cnt,
3029	size: sizeof(env->subprog_info[0]), cmp_func: cmp_subprogs, NULL);
3030	return env->subprog_cnt - 1;
3031	}
3032
3033	static int bpf_find_exception_callback_insn_off(struct bpf_verifier_env *env)
3034	{
3035	struct bpf_prog_aux *aux = env->prog->aux;
3036	struct btf *btf = aux->btf;
3037	const struct btf_type *t;
3038	u32 main_btf_id, id;
3039	const char *name;
3040	int ret, i;
3041
3042	/* Non-zero func_info_cnt implies valid btf */
3043	if (!aux->func_info_cnt)
3044	return 0;
3045	main_btf_id = aux->func_info[0].type_id;
3046
3047	t = btf_type_by_id(btf, type_id: main_btf_id);
3048	if (!t) {
3049	verbose(private_data: env, fmt: "invalid btf id for main subprog in func_info\n");
3050	return -EINVAL;
3051	}
3052
3053	name = btf_find_decl_tag_value(btf, pt: t, comp_idx: -1, tag_key: "exception_callback:");
3054	if (IS_ERR(ptr: name)) {
3055	ret = PTR_ERR(ptr: name);
3056	/* If there is no tag present, there is no exception callback */
3057	if (ret == -ENOENT)
3058	ret = 0;
3059	else if (ret == -EEXIST)
3060	verbose(private_data: env, fmt: "multiple exception callback tags for main subprog\n");
3061	return ret;
3062	}
3063
3064	ret = btf_find_by_name_kind(btf, name, kind: BTF_KIND_FUNC);
3065	if (ret < 0) {
3066	verbose(private_data: env, fmt: "exception callback '%s' could not be found in BTF\n", name);
3067	return ret;
3068	}
3069	id = ret;
3070	t = btf_type_by_id(btf, type_id: id);
3071	if (btf_func_linkage(t) != BTF_FUNC_GLOBAL) {
3072	verbose(private_data: env, fmt: "exception callback '%s' must have global linkage\n", name);
3073	return -EINVAL;
3074	}
3075	ret = 0;
3076	for (i = 0; i < aux->func_info_cnt; i++) {
3077	if (aux->func_info[i].type_id != id)
3078	continue;
3079	ret = aux->func_info[i].insn_off;
3080	/* Further func_info and subprog checks will also happen
3081	* later, so assume this is the right insn_off for now.
3082	*/
3083	if (!ret) {
3084	verbose(private_data: env, fmt: "invalid exception callback insn_off in func_info: 0\n");
3085	ret = -EINVAL;
3086	}
3087	}
3088	if (!ret) {
3089	verbose(private_data: env, fmt: "exception callback type id not found in func_info\n");
3090	ret = -EINVAL;
3091	}
3092	return ret;
3093	}
3094
3095	#define MAX_KFUNC_DESCS 256
3096	#define MAX_KFUNC_BTFS 256
3097
3098	struct bpf_kfunc_desc {
3099	struct btf_func_model func_model;
3100	u32 func_id;
3101	s32 imm;
3102	u16 offset;
3103	unsigned long addr;
3104	};
3105
3106	struct bpf_kfunc_btf {
3107	struct btf *btf;
3108	struct module *module;
3109	u16 offset;
3110	};
3111
3112	struct bpf_kfunc_desc_tab {
3113	/* Sorted by func_id (BTF ID) and offset (fd_array offset) during
3114	* verification. JITs do lookups by bpf_insn, where func_id may not be
3115	* available, therefore at the end of verification do_misc_fixups()
3116	* sorts this by imm and offset.
3117	*/
3118	struct bpf_kfunc_desc descs[MAX_KFUNC_DESCS];
3119	u32 nr_descs;
3120	};
3121
3122	struct bpf_kfunc_btf_tab {
3123	struct bpf_kfunc_btf descs[MAX_KFUNC_BTFS];
3124	u32 nr_descs;
3125	};
3126
3127	static int specialize_kfunc(struct bpf_verifier_env *env, struct bpf_kfunc_desc *desc,
3128	int insn_idx);
3129
3130	static int kfunc_desc_cmp_by_id_off(const void *a, const void *b)
3131	{
3132	const struct bpf_kfunc_desc *d0 = a;
3133	const struct bpf_kfunc_desc *d1 = b;
3134
3135	/* func_id is not greater than BTF_MAX_TYPE */
3136	return d0->func_id - d1->func_id ?: d0->offset - d1->offset;
3137	}
3138
3139	static int kfunc_btf_cmp_by_off(const void *a, const void *b)
3140	{
3141	const struct bpf_kfunc_btf *d0 = a;
3142	const struct bpf_kfunc_btf *d1 = b;
3143
3144	return d0->offset - d1->offset;
3145	}
3146
3147	static struct bpf_kfunc_desc *
3148	find_kfunc_desc(const struct bpf_prog *prog, u32 func_id, u16 offset)
3149	{
3150	struct bpf_kfunc_desc desc = {
3151	.func_id = func_id,
3152	.offset = offset,
3153	};
3154	struct bpf_kfunc_desc_tab *tab;
3155
3156	tab = prog->aux->kfunc_tab;
3157	return bsearch(key: &desc, base: tab->descs, num: tab->nr_descs,
3158	size: sizeof(tab->descs[0]), cmp: kfunc_desc_cmp_by_id_off);
3159	}
3160
3161	int bpf_get_kfunc_addr(const struct bpf_prog *prog, u32 func_id,
3162	u16 btf_fd_idx, u8 **func_addr)
3163	{
3164	const struct bpf_kfunc_desc *desc;
3165
3166	desc = find_kfunc_desc(prog, func_id, offset: btf_fd_idx);
3167	if (!desc)
3168	return -EFAULT;
3169
3170	*func_addr = (u8 *)desc->addr;
3171	return 0;
3172	}
3173
3174	static struct btf *__find_kfunc_desc_btf(struct bpf_verifier_env *env,
3175	s16 offset)
3176	{
3177	struct bpf_kfunc_btf kf_btf = { .offset = offset };
3178	struct bpf_kfunc_btf_tab *tab;
3179	struct bpf_kfunc_btf *b;
3180	struct module *mod;
3181	struct btf *btf;
3182	int btf_fd;
3183
3184	tab = env->prog->aux->kfunc_btf_tab;
3185	b = bsearch(key: &kf_btf, base: tab->descs, num: tab->nr_descs,
3186	size: sizeof(tab->descs[0]), cmp: kfunc_btf_cmp_by_off);
3187	if (!b) {
3188	if (tab->nr_descs == MAX_KFUNC_BTFS) {
3189	verbose(private_data: env, fmt: "too many different module BTFs\n");
3190	return ERR_PTR(error: -E2BIG);
3191	}
3192
3193	if (bpfptr_is_null(bpfptr: env->fd_array)) {
3194	verbose(private_data: env, fmt: "kfunc offset > 0 without fd_array is invalid\n");
3195	return ERR_PTR(error: -EPROTO);
3196	}
3197
3198	if (copy_from_bpfptr_offset(dst: &btf_fd, src: env->fd_array,
3199	offset: offset * sizeof(btf_fd),
3200	size: sizeof(btf_fd)))
3201	return ERR_PTR(error: -EFAULT);
3202
3203	btf = btf_get_by_fd(fd: btf_fd);
3204	if (IS_ERR(ptr: btf)) {
3205	verbose(private_data: env, fmt: "invalid module BTF fd specified\n");
3206	return btf;
3207	}
3208
3209	if (!btf_is_module(btf)) {
3210	verbose(private_data: env, fmt: "BTF fd for kfunc is not a module BTF\n");
3211	btf_put(btf);
3212	return ERR_PTR(error: -EINVAL);
3213	}
3214
3215	mod = btf_try_get_module(btf);
3216	if (!mod) {
3217	btf_put(btf);
3218	return ERR_PTR(error: -ENXIO);
3219	}
3220
3221	b = &tab->descs[tab->nr_descs++];
3222	b->btf = btf;
3223	b->module = mod;
3224	b->offset = offset;
3225
3226	/* sort() reorders entries by value, so b may no longer point
3227	* to the right entry after this
3228	*/
3229	sort(base: tab->descs, num: tab->nr_descs, size: sizeof(tab->descs[0]),
3230	cmp_func: kfunc_btf_cmp_by_off, NULL);
3231	} else {
3232	btf = b->btf;
3233	}
3234
3235	return btf;
3236	}
3237
3238	void bpf_free_kfunc_btf_tab(struct bpf_kfunc_btf_tab *tab)
3239	{
3240	if (!tab)
3241	return;
3242
3243	while (tab->nr_descs--) {
3244	module_put(module: tab->descs[tab->nr_descs].module);
3245	btf_put(btf: tab->descs[tab->nr_descs].btf);
3246	}
3247	kfree(objp: tab);
3248	}
3249
3250	static struct btf *find_kfunc_desc_btf(struct bpf_verifier_env *env, s16 offset)
3251	{
3252	if (offset) {
3253	if (offset < 0) {
3254	/* In the future, this can be allowed to increase limit
3255	* of fd index into fd_array, interpreted as u16.
3256	*/
3257	verbose(private_data: env, fmt: "negative offset disallowed for kernel module function call\n");
3258	return ERR_PTR(error: -EINVAL);
3259	}
3260
3261	return __find_kfunc_desc_btf(env, offset);
3262	}
3263	return btf_vmlinux ?: ERR_PTR(error: -ENOENT);
3264	}
3265
3266	static int add_kfunc_call(struct bpf_verifier_env *env, u32 func_id, s16 offset)
3267	{
3268	const struct btf_type *func, *func_proto;
3269	struct bpf_kfunc_btf_tab *btf_tab;
3270	struct btf_func_model func_model;
3271	struct bpf_kfunc_desc_tab *tab;
3272	struct bpf_prog_aux *prog_aux;
3273	struct bpf_kfunc_desc *desc;
3274	const char *func_name;
3275	struct btf *desc_btf;
3276	unsigned long addr;
3277	int err;
3278
3279	prog_aux = env->prog->aux;
3280	tab = prog_aux->kfunc_tab;
3281	btf_tab = prog_aux->kfunc_btf_tab;
3282	if (!tab) {
3283	if (!btf_vmlinux) {
3284	verbose(private_data: env, fmt: "calling kernel function is not supported without CONFIG_DEBUG_INFO_BTF\n");
3285	return -ENOTSUPP;
3286	}
3287
3288	if (!env->prog->jit_requested) {
3289	verbose(private_data: env, fmt: "JIT is required for calling kernel function\n");
3290	return -ENOTSUPP;
3291	}
3292
3293	if (!bpf_jit_supports_kfunc_call()) {
3294	verbose(private_data: env, fmt: "JIT does not support calling kernel function\n");
3295	return -ENOTSUPP;
3296	}
3297
3298	if (!env->prog->gpl_compatible) {
3299	verbose(private_data: env, fmt: "cannot call kernel function from non-GPL compatible program\n");
3300	return -EINVAL;
3301	}
3302
3303	tab = kzalloc(sizeof(*tab), GFP_KERNEL_ACCOUNT);
3304	if (!tab)
3305	return -ENOMEM;
3306	prog_aux->kfunc_tab = tab;
3307	}
3308
3309	/* func_id == 0 is always invalid, but instead of returning an error, be
3310	* conservative and wait until the code elimination pass before returning
3311	* error, so that invalid calls that get pruned out can be in BPF programs
3312	* loaded from userspace. It is also required that offset be untouched
3313	* for such calls.
3314	*/
3315	if (!func_id && !offset)
3316	return 0;
3317
3318	if (!btf_tab && offset) {
3319	btf_tab = kzalloc(sizeof(*btf_tab), GFP_KERNEL_ACCOUNT);
3320	if (!btf_tab)
3321	return -ENOMEM;
3322	prog_aux->kfunc_btf_tab = btf_tab;
3323	}
3324
3325	desc_btf = find_kfunc_desc_btf(env, offset);
3326	if (IS_ERR(ptr: desc_btf)) {
3327	verbose(private_data: env, fmt: "failed to find BTF for kernel function\n");
3328	return PTR_ERR(ptr: desc_btf);
3329	}
3330
3331	if (find_kfunc_desc(prog: env->prog, func_id, offset))
3332	return 0;
3333
3334	if (tab->nr_descs == MAX_KFUNC_DESCS) {
3335	verbose(private_data: env, fmt: "too many different kernel function calls\n");
3336	return -E2BIG;
3337	}
3338
3339	func = btf_type_by_id(btf: desc_btf, type_id: func_id);
3340	if (!func || !btf_type_is_func(t: func)) {
3341	verbose(private_data: env, fmt: "kernel btf_id %u is not a function\n",
3342	func_id);
3343	return -EINVAL;
3344	}
3345	func_proto = btf_type_by_id(btf: desc_btf, type_id: func->type);
3346	if (!func_proto || !btf_type_is_func_proto(t: func_proto)) {
3347	verbose(private_data: env, fmt: "kernel function btf_id %u does not have a valid func_proto\n",
3348	func_id);
3349	return -EINVAL;
3350	}
3351
3352	func_name = btf_name_by_offset(btf: desc_btf, offset: func->name_off);
3353	addr = kallsyms_lookup_name(name: func_name);
3354	if (!addr) {
3355	verbose(private_data: env, fmt: "cannot find address for kernel function %s\n",
3356	func_name);
3357	return -EINVAL;
3358	}
3359
3360	if (bpf_dev_bound_kfunc_id(btf_id: func_id)) {
3361	err = bpf_dev_bound_kfunc_check(log: &env->log, prog_aux);
3362	if (err)
3363	return err;
3364	}
3365
3366	err = btf_distill_func_proto(log: &env->log, btf: desc_btf,
3367	func_proto, func_name,
3368	m: &func_model);
3369	if (err)
3370	return err;
3371
3372	desc = &tab->descs[tab->nr_descs++];
3373	desc->func_id = func_id;
3374	desc->offset = offset;
3375	desc->addr = addr;
3376	desc->func_model = func_model;
3377	sort(base: tab->descs, num: tab->nr_descs, size: sizeof(tab->descs[0]),
3378	cmp_func: kfunc_desc_cmp_by_id_off, NULL);
3379	return 0;
3380	}
3381
3382	static int kfunc_desc_cmp_by_imm_off(const void *a, const void *b)
3383	{
3384	const struct bpf_kfunc_desc *d0 = a;
3385	const struct bpf_kfunc_desc *d1 = b;
3386
3387	if (d0->imm != d1->imm)
3388	return d0->imm < d1->imm ? -1 : 1;
3389	if (d0->offset != d1->offset)
3390	return d0->offset < d1->offset ? -1 : 1;
3391	return 0;
3392	}
3393
3394	static int set_kfunc_desc_imm(struct bpf_verifier_env *env, struct bpf_kfunc_desc *desc)
3395	{
3396	unsigned long call_imm;
3397
3398	if (bpf_jit_supports_far_kfunc_call()) {
3399	call_imm = desc->func_id;
3400	} else {
3401	call_imm = BPF_CALL_IMM(desc->addr);
3402	/* Check whether the relative offset overflows desc->imm */
3403	if ((unsigned long)(s32)call_imm != call_imm) {
3404	verbose(private_data: env, fmt: "address of kernel func_id %u is out of range\n",
3405	desc->func_id);
3406	return -EINVAL;
3407	}
3408	}
3409	desc->imm = call_imm;
3410	return 0;
3411	}
3412
3413	static int sort_kfunc_descs_by_imm_off(struct bpf_verifier_env *env)
3414	{
3415	struct bpf_kfunc_desc_tab *tab;
3416	int i, err;
3417
3418	tab = env->prog->aux->kfunc_tab;
3419	if (!tab)
3420	return 0;
3421
3422	for (i = 0; i < tab->nr_descs; i++) {
3423	err = set_kfunc_desc_imm(env, desc: &tab->descs[i]);
3424	if (err)
3425	return err;
3426	}
3427
3428	sort(base: tab->descs, num: tab->nr_descs, size: sizeof(tab->descs[0]),
3429	cmp_func: kfunc_desc_cmp_by_imm_off, NULL);
3430	return 0;
3431	}
3432
3433	bool bpf_prog_has_kfunc_call(const struct bpf_prog *prog)
3434	{
3435	return !!prog->aux->kfunc_tab;
3436	}
3437
3438	const struct btf_func_model *
3439	bpf_jit_find_kfunc_model(const struct bpf_prog *prog,
3440	const struct bpf_insn *insn)
3441	{
3442	const struct bpf_kfunc_desc desc = {
3443	.imm = insn->imm,
3444	.offset = insn->off,
3445	};
3446	const struct bpf_kfunc_desc *res;
3447	struct bpf_kfunc_desc_tab *tab;
3448
3449	tab = prog->aux->kfunc_tab;
3450	res = bsearch(key: &desc, base: tab->descs, num: tab->nr_descs,
3451	size: sizeof(tab->descs[0]), cmp: kfunc_desc_cmp_by_imm_off);
3452
3453	return res ? &res->func_model : NULL;
3454	}
3455
3456	static int add_kfunc_in_insns(struct bpf_verifier_env *env,
3457	struct bpf_insn *insn, int cnt)
3458	{
3459	int i, ret;
3460
3461	for (i = 0; i < cnt; i++, insn++) {
3462	if (bpf_pseudo_kfunc_call(insn)) {
3463	ret = add_kfunc_call(env, func_id: insn->imm, offset: insn->off);
3464	if (ret < 0)
3465	return ret;
3466	}
3467	}
3468	return 0;
3469	}
3470
3471	static int add_subprog_and_kfunc(struct bpf_verifier_env *env)
3472	{
3473	struct bpf_subprog_info *subprog = env->subprog_info;
3474	int i, ret, insn_cnt = env->prog->len, ex_cb_insn;
3475	struct bpf_insn *insn = env->prog->insnsi;
3476
3477	/* Add entry function. */
3478	ret = add_subprog(env, off: 0);
3479	if (ret)
3480	return ret;
3481
3482	for (i = 0; i < insn_cnt; i++, insn++) {
3483	if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn) &&
3484	!bpf_pseudo_kfunc_call(insn))
3485	continue;
3486
3487	if (!env->bpf_capable) {
3488	verbose(private_data: env, fmt: "loading/calling other bpf or kernel functions are allowed for CAP_BPF and CAP_SYS_ADMIN\n");
3489	return -EPERM;
3490	}
3491
3492	if (bpf_pseudo_func(insn) || bpf_pseudo_call(insn))
3493	ret = add_subprog(env, off: i + insn->imm + 1);
3494	else
3495	ret = add_kfunc_call(env, func_id: insn->imm, offset: insn->off);
3496
3497	if (ret < 0)
3498	return ret;
3499	}
3500
3501	ret = bpf_find_exception_callback_insn_off(env);
3502	if (ret < 0)
3503	return ret;
3504	ex_cb_insn = ret;
3505
3506	/* If ex_cb_insn > 0, this means that the main program has a subprog
3507	* marked using BTF decl tag to serve as the exception callback.
3508	*/
3509	if (ex_cb_insn) {
3510	ret = add_subprog(env, off: ex_cb_insn);
3511	if (ret < 0)
3512	return ret;
3513	for (i = 1; i < env->subprog_cnt; i++) {
3514	if (env->subprog_info[i].start != ex_cb_insn)
3515	continue;
3516	env->exception_callback_subprog = i;
3517	mark_subprog_exc_cb(env, subprog: i);
3518	break;
3519	}
3520	}
3521
3522	/* Add a fake 'exit' subprog which could simplify subprog iteration
3523	* logic. 'subprog_cnt' should not be increased.
3524	*/
3525	subprog[env->subprog_cnt].start = insn_cnt;
3526
3527	if (env->log.level & BPF_LOG_LEVEL2)
3528	for (i = 0; i < env->subprog_cnt; i++)
3529	verbose(private_data: env, fmt: "func#%d @%d\n", i, subprog[i].start);
3530
3531	return 0;
3532	}
3533
3534	static int check_subprogs(struct bpf_verifier_env *env)
3535	{
3536	int i, subprog_start, subprog_end, off, cur_subprog = 0;
3537	struct bpf_subprog_info *subprog = env->subprog_info;
3538	struct bpf_insn *insn = env->prog->insnsi;
3539	int insn_cnt = env->prog->len;
3540
3541	/* now check that all jumps are within the same subprog */
3542	subprog_start = subprog[cur_subprog].start;
3543	subprog_end = subprog[cur_subprog + 1].start;
3544	for (i = 0; i < insn_cnt; i++) {
3545	u8 code = insn[i].code;
3546
3547	if (code == (BPF_JMP | BPF_CALL) &&
3548	insn[i].src_reg == 0 &&
3549	insn[i].imm == BPF_FUNC_tail_call) {
3550	subprog[cur_subprog].has_tail_call = true;
3551	subprog[cur_subprog].tail_call_reachable = true;
3552	}
3553	if (BPF_CLASS(code) == BPF_LD &&
3554	(BPF_MODE(code) == BPF_ABS || BPF_MODE(code) == BPF_IND))
3555	subprog[cur_subprog].has_ld_abs = true;
3556	if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
3557	goto next;
3558	if (BPF_OP(code) == BPF_CALL)
3559	goto next;
3560	if (BPF_OP(code) == BPF_EXIT) {
3561	subprog[cur_subprog].exit_idx = i;
3562	goto next;
3563	}
3564	off = i + bpf_jmp_offset(insn: &insn[i]) + 1;
3565	if (off < subprog_start || off >= subprog_end) {
3566	verbose(private_data: env, fmt: "jump out of range from insn %d to %d\n", i, off);
3567	return -EINVAL;
3568	}
3569	next:
3570	if (i == subprog_end - 1) {
3571	/* to avoid fall-through from one subprog into another
3572	* the last insn of the subprog should be either exit
3573	* or unconditional jump back or bpf_throw call
3574	*/
3575	if (code != (BPF_JMP | BPF_EXIT) &&
3576	code != (BPF_JMP32 | BPF_JA) &&
3577	code != (BPF_JMP | BPF_JA)) {
3578	verbose(private_data: env, fmt: "last insn is not an exit or jmp\n");
3579	return -EINVAL;
3580	}
3581	subprog_start = subprog_end;
3582	cur_subprog++;
3583	if (cur_subprog < env->subprog_cnt)
3584	subprog_end = subprog[cur_subprog + 1].start;
3585	}
3586	}
3587	return 0;
3588	}
3589
3590	static int mark_stack_slot_obj_read(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
3591	int spi, int nr_slots)
3592	{
3593	int err, i;
3594
3595	for (i = 0; i < nr_slots; i++) {
3596	err = bpf_mark_stack_read(env, frameno: reg->frameno, insn_idx: env->insn_idx, BIT(spi - i));
3597	if (err)
3598	return err;
3599	mark_stack_slot_scratched(env, spi: spi - i);
3600	}
3601	return 0;
3602	}
3603
3604	static int mark_dynptr_read(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
3605	{
3606	int spi;
3607
3608	/* For CONST_PTR_TO_DYNPTR, it must have already been done by
3609	* check_reg_arg in check_helper_call and mark_btf_func_reg_size in
3610	* check_kfunc_call.
3611	*/
3612	if (reg->type == CONST_PTR_TO_DYNPTR)
3613	return 0;
3614	spi = dynptr_get_spi(env, reg);
3615	if (spi < 0)
3616	return spi;
3617	/* Caller ensures dynptr is valid and initialized, which means spi is in
3618	* bounds and spi is the first dynptr slot. Simply mark stack slot as
3619	* read.
3620	*/
3621	return mark_stack_slot_obj_read(env, reg, spi, BPF_DYNPTR_NR_SLOTS);
3622	}
3623
3624	static int mark_iter_read(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
3625	int spi, int nr_slots)
3626	{
3627	return mark_stack_slot_obj_read(env, reg, spi, nr_slots);
3628	}
3629
3630	static int mark_irq_flag_read(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
3631	{
3632	int spi;
3633
3634	spi = irq_flag_get_spi(env, reg);
3635	if (spi < 0)
3636	return spi;
3637	return mark_stack_slot_obj_read(env, reg, spi, nr_slots: 1);
3638	}
3639
3640	/* This function is supposed to be used by the following 32-bit optimization
3641	* code only. It returns TRUE if the source or destination register operates
3642	* on 64-bit, otherwise return FALSE.
3643	*/
3644	static bool is_reg64(struct bpf_insn *insn,
3645	u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
3646	{
3647	u8 code, class, op;
3648
3649	code = insn->code;
3650	class = BPF_CLASS(code);
3651	op = BPF_OP(code);
3652	if (class == BPF_JMP) {
3653	/* BPF_EXIT for "main" will reach here. Return TRUE
3654	* conservatively.
3655	*/
3656	if (op == BPF_EXIT)
3657	return true;
3658	if (op == BPF_CALL) {
3659	/* BPF to BPF call will reach here because of marking
3660	* caller saved clobber with DST_OP_NO_MARK for which we
3661	* don't care the register def because they are anyway
3662	* marked as NOT_INIT already.
3663	*/
3664	if (insn->src_reg == BPF_PSEUDO_CALL)
3665	return false;
3666	/* Helper call will reach here because of arg type
3667	* check, conservatively return TRUE.
3668	*/
3669	if (t == SRC_OP)
3670	return true;
3671
3672	return false;
3673	}
3674	}
3675
3676	if (class == BPF_ALU64 && op == BPF_END && (insn->imm == 16 || insn->imm == 32))
3677	return false;
3678
3679	if (class == BPF_ALU64 || class == BPF_JMP ||
3680	(class == BPF_ALU && op == BPF_END && insn->imm == 64))
3681	return true;
3682
3683	if (class == BPF_ALU || class == BPF_JMP32)
3684	return false;
3685
3686	if (class == BPF_LDX) {
3687	if (t != SRC_OP)
3688	return BPF_SIZE(code) == BPF_DW || BPF_MODE(code) == BPF_MEMSX;
3689	/* LDX source must be ptr. */
3690	return true;
3691	}
3692
3693	if (class == BPF_STX) {
3694	/* BPF_STX (including atomic variants) has one or more source
3695	* operands, one of which is a ptr. Check whether the caller is
3696	* asking about it.
3697	*/
3698	if (t == SRC_OP && reg->type != SCALAR_VALUE)
3699	return true;
3700	return BPF_SIZE(code) == BPF_DW;
3701	}
3702
3703	if (class == BPF_LD) {
3704	u8 mode = BPF_MODE(code);
3705
3706	/* LD_IMM64 */
3707	if (mode == BPF_IMM)
3708	return true;
3709
3710	/* Both LD_IND and LD_ABS return 32-bit data. */
3711	if (t != SRC_OP)
3712	return false;
3713
3714	/* Implicit ctx ptr. */
3715	if (regno == BPF_REG_6)
3716	return true;
3717
3718	/* Explicit source could be any width. */
3719	return true;
3720	}
3721
3722	if (class == BPF_ST)
3723	/* The only source register for BPF_ST is a ptr. */
3724	return true;
3725
3726	/* Conservatively return true at default. */
3727	return true;
3728	}
3729
3730	/* Return the regno defined by the insn, or -1. */
3731	static int insn_def_regno(const struct bpf_insn *insn)
3732	{
3733	switch (BPF_CLASS(insn->code)) {
3734	case BPF_JMP:
3735	case BPF_JMP32:
3736	case BPF_ST:
3737	return -1;
3738	case BPF_STX:
3739	if (BPF_MODE(insn->code) == BPF_ATOMIC ||
3740	BPF_MODE(insn->code) == BPF_PROBE_ATOMIC) {
3741	if (insn->imm == BPF_CMPXCHG)
3742	return BPF_REG_0;
3743	else if (insn->imm == BPF_LOAD_ACQ)
3744	return insn->dst_reg;
3745	else if (insn->imm & BPF_FETCH)
3746	return insn->src_reg;
3747	}
3748	return -1;
3749	default:
3750	return insn->dst_reg;
3751	}
3752	}
3753
3754	/* Return TRUE if INSN has defined any 32-bit value explicitly. */
3755	static bool insn_has_def32(struct bpf_insn *insn)
3756	{
3757	int dst_reg = insn_def_regno(insn);
3758
3759	if (dst_reg == -1)
3760	return false;
3761
3762	return !is_reg64(insn, regno: dst_reg, NULL, t: DST_OP);
3763	}
3764
3765	static void mark_insn_zext(struct bpf_verifier_env *env,
3766	struct bpf_reg_state *reg)
3767	{
3768	s32 def_idx = reg->subreg_def;
3769
3770	if (def_idx == DEF_NOT_SUBREG)
3771	return;
3772
3773	env->insn_aux_data[def_idx - 1].zext_dst = true;
3774	/* The dst will be zero extended, so won't be sub-register anymore. */
3775	reg->subreg_def = DEF_NOT_SUBREG;
3776	}
3777
3778	static int __check_reg_arg(struct bpf_verifier_env *env, struct bpf_reg_state *regs, u32 regno,
3779	enum reg_arg_type t)
3780	{
3781	struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
3782	struct bpf_reg_state *reg;
3783	bool rw64;
3784
3785	if (regno >= MAX_BPF_REG) {
3786	verbose(private_data: env, fmt: "R%d is invalid\n", regno);
3787	return -EINVAL;
3788	}
3789
3790	mark_reg_scratched(env, regno);
3791
3792	reg = &regs[regno];
3793	rw64 = is_reg64(insn, regno, reg, t);
3794	if (t == SRC_OP) {
3795	/* check whether register used as source operand can be read */
3796	if (reg->type == NOT_INIT) {
3797	verbose(private_data: env, fmt: "R%d !read_ok\n", regno);
3798	return -EACCES;
3799	}
3800	/* We don't need to worry about FP liveness because it's read-only */
3801	if (regno == BPF_REG_FP)
3802	return 0;
3803
3804	if (rw64)
3805	mark_insn_zext(env, reg);
3806
3807	return 0;
3808	} else {
3809	/* check whether register used as dest operand can be written to */
3810	if (regno == BPF_REG_FP) {
3811	verbose(private_data: env, fmt: "frame pointer is read only\n");
3812	return -EACCES;
3813	}
3814	reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
3815	if (t == DST_OP)
3816	mark_reg_unknown(env, regs, regno);
3817	}
3818	return 0;
3819	}
3820
3821	static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
3822	enum reg_arg_type t)
3823	{
3824	struct bpf_verifier_state *vstate = env->cur_state;
3825	struct bpf_func_state *state = vstate->frame[vstate->curframe];
3826
3827	return __check_reg_arg(env, regs: state->regs, regno, t);
3828	}
3829
3830	static int insn_stack_access_flags(int frameno, int spi)
3831	{
3832	return INSN_F_STACK_ACCESS | (spi << INSN_F_SPI_SHIFT) | frameno;
3833	}
3834
3835	static int insn_stack_access_spi(int insn_flags)
3836	{
3837	return (insn_flags >> INSN_F_SPI_SHIFT) & INSN_F_SPI_MASK;
3838	}
3839
3840	static int insn_stack_access_frameno(int insn_flags)
3841	{
3842	return insn_flags & INSN_F_FRAMENO_MASK;
3843	}
3844
3845	static void mark_jmp_point(struct bpf_verifier_env *env, int idx)
3846	{
3847	env->insn_aux_data[idx].jmp_point = true;
3848	}
3849
3850	static bool is_jmp_point(struct bpf_verifier_env *env, int insn_idx)
3851	{
3852	return env->insn_aux_data[insn_idx].jmp_point;
3853	}
3854
3855	#define LR_FRAMENO_BITS 3
3856	#define LR_SPI_BITS 6
3857	#define LR_ENTRY_BITS (LR_SPI_BITS + LR_FRAMENO_BITS + 1)
3858	#define LR_SIZE_BITS 4
3859	#define LR_FRAMENO_MASK ((1ull << LR_FRAMENO_BITS) - 1)
3860	#define LR_SPI_MASK ((1ull << LR_SPI_BITS) - 1)
3861	#define LR_SIZE_MASK ((1ull << LR_SIZE_BITS) - 1)
3862	#define LR_SPI_OFF LR_FRAMENO_BITS
3863	#define LR_IS_REG_OFF (LR_SPI_BITS + LR_FRAMENO_BITS)
3864	#define LINKED_REGS_MAX 6
3865
3866	struct linked_reg {
3867	u8 frameno;
3868	union {
3869	u8 spi;
3870	u8 regno;
3871	};
3872	bool is_reg;
3873	};
3874
3875	struct linked_regs {
3876	int cnt;
3877	struct linked_reg entries[LINKED_REGS_MAX];
3878	};
3879
3880	static struct linked_reg *linked_regs_push(struct linked_regs *s)
3881	{
3882	if (s->cnt < LINKED_REGS_MAX)
3883	return &s->entries[s->cnt++];
3884
3885	return NULL;
3886	}
3887
3888	/* Use u64 as a vector of 6 10-bit values, use first 4-bits to track
3889	* number of elements currently in stack.
3890	* Pack one history entry for linked registers as 10 bits in the following format:
3891	* - 3-bits frameno
3892	* - 6-bits spi_or_reg
3893	* - 1-bit is_reg
3894	*/
3895	static u64 linked_regs_pack(struct linked_regs *s)
3896	{
3897	u64 val = 0;
3898	int i;
3899
3900	for (i = 0; i < s->cnt; ++i) {
3901	struct linked_reg *e = &s->entries[i];
3902	u64 tmp = 0;
3903
3904	tmp |= e->frameno;
3905	tmp |= e->spi << LR_SPI_OFF;
3906	tmp |= (e->is_reg ? 1 : 0) << LR_IS_REG_OFF;
3907
3908	val <<= LR_ENTRY_BITS;
3909	val |= tmp;
3910	}
3911	val <<= LR_SIZE_BITS;
3912	val |= s->cnt;
3913	return val;
3914	}
3915
3916	static void linked_regs_unpack(u64 val, struct linked_regs *s)
3917	{
3918	int i;
3919
3920	s->cnt = val & LR_SIZE_MASK;
3921	val >>= LR_SIZE_BITS;
3922
3923	for (i = 0; i < s->cnt; ++i) {
3924	struct linked_reg *e = &s->entries[i];
3925
3926	e->frameno = val & LR_FRAMENO_MASK;
3927	e->spi = (val >> LR_SPI_OFF) & LR_SPI_MASK;
3928	e->is_reg = (val >> LR_IS_REG_OFF) & 0x1;
3929	val >>= LR_ENTRY_BITS;
3930	}
3931	}
3932
3933	/* for any branch, call, exit record the history of jmps in the given state */
3934	static int push_jmp_history(struct bpf_verifier_env *env, struct bpf_verifier_state *cur,
3935	int insn_flags, u64 linked_regs)
3936	{
3937	u32 cnt = cur->jmp_history_cnt;
3938	struct bpf_jmp_history_entry *p;
3939	size_t alloc_size;
3940
3941	/* combine instruction flags if we already recorded this instruction */
3942	if (env->cur_hist_ent) {
3943	/* atomic instructions push insn_flags twice, for READ and
3944	* WRITE sides, but they should agree on stack slot
3945	*/
3946	verifier_bug_if((env->cur_hist_ent->flags & insn_flags) &&
3947	(env->cur_hist_ent->flags & insn_flags) != insn_flags,
3948	env, "insn history: insn_idx %d cur flags %x new flags %x",
3949	env->insn_idx, env->cur_hist_ent->flags, insn_flags);
3950	env->cur_hist_ent->flags |= insn_flags;
3951	verifier_bug_if(env->cur_hist_ent->linked_regs != 0, env,
3952	"insn history: insn_idx %d linked_regs: %#llx",
3953	env->insn_idx, env->cur_hist_ent->linked_regs);
3954	env->cur_hist_ent->linked_regs = linked_regs;
3955	return 0;
3956	}
3957
3958	cnt++;
3959	alloc_size = kmalloc_size_roundup(size: size_mul(factor1: cnt, factor2: sizeof(*p)));
3960	p = krealloc(cur->jmp_history, alloc_size, GFP_KERNEL_ACCOUNT);
3961	if (!p)
3962	return -ENOMEM;
3963	cur->jmp_history = p;
3964
3965	p = &cur->jmp_history[cnt - 1];
3966	p->idx = env->insn_idx;
3967	p->prev_idx = env->prev_insn_idx;
3968	p->flags = insn_flags;
3969	p->linked_regs = linked_regs;
3970	cur->jmp_history_cnt = cnt;
3971	env->cur_hist_ent = p;
3972
3973	return 0;
3974	}
3975
3976	static struct bpf_jmp_history_entry *get_jmp_hist_entry(struct bpf_verifier_state *st,
3977	u32 hist_end, int insn_idx)
3978	{
3979	if (hist_end > 0 && st->jmp_history[hist_end - 1].idx == insn_idx)
3980	return &st->jmp_history[hist_end - 1];
3981	return NULL;
3982	}
3983
3984	/* Backtrack one insn at a time. If idx is not at the top of recorded
3985	* history then previous instruction came from straight line execution.
3986	* Return -ENOENT if we exhausted all instructions within given state.
3987	*
3988	* It's legal to have a bit of a looping with the same starting and ending
3989	* insn index within the same state, e.g.: 3->4->5->3, so just because current
3990	* instruction index is the same as state's first_idx doesn't mean we are
3991	* done. If there is still some jump history left, we should keep going. We
3992	* need to take into account that we might have a jump history between given
3993	* state's parent and itself, due to checkpointing. In this case, we'll have
3994	* history entry recording a jump from last instruction of parent state and
3995	* first instruction of given state.
3996	*/
3997	static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
3998	u32 *history)
3999	{
4000	u32 cnt = *history;
4001
4002	if (i == st->first_insn_idx) {
4003	if (cnt == 0)
4004	return -ENOENT;
4005	if (cnt == 1 && st->jmp_history[0].idx == i)
4006	return -ENOENT;
4007	}
4008
4009	if (cnt && st->jmp_history[cnt - 1].idx == i) {
4010	i = st->jmp_history[cnt - 1].prev_idx;
4011	(*history)--;
4012	} else {
4013	i--;
4014	}
4015	return i;
4016	}
4017
4018	static const char *disasm_kfunc_name(void *data, const struct bpf_insn *insn)
4019	{
4020	const struct btf_type *func;
4021	struct btf *desc_btf;
4022
4023	if (insn->src_reg != BPF_PSEUDO_KFUNC_CALL)
4024	return NULL;
4025
4026	desc_btf = find_kfunc_desc_btf(env: data, offset: insn->off);
4027	if (IS_ERR(ptr: desc_btf))
4028	return "<error>";
4029
4030	func = btf_type_by_id(btf: desc_btf, type_id: insn->imm);
4031	return btf_name_by_offset(btf: desc_btf, offset: func->name_off);
4032	}
4033
4034	static void verbose_insn(struct bpf_verifier_env *env, struct bpf_insn *insn)
4035	{
4036	const struct bpf_insn_cbs cbs = {
4037	.cb_call = disasm_kfunc_name,
4038	.cb_print = verbose,
4039	.private_data = env,
4040	};
4041
4042	print_bpf_insn(cbs: &cbs, insn, allow_ptr_leaks: env->allow_ptr_leaks);
4043	}
4044
4045	static inline void bt_init(struct backtrack_state *bt, u32 frame)
4046	{
4047	bt->frame = frame;
4048	}
4049
4050	static inline void bt_reset(struct backtrack_state *bt)
4051	{
4052	struct bpf_verifier_env *env = bt->env;
4053
4054	memset(bt, 0, sizeof(*bt));
4055	bt->env = env;
4056	}
4057
4058	static inline u32 bt_empty(struct backtrack_state *bt)
4059	{
4060	u64 mask = 0;
4061	int i;
4062
4063	for (i = 0; i <= bt->frame; i++)
4064	mask |= bt->reg_masks[i] | bt->stack_masks[i];
4065
4066	return mask == 0;
4067	}
4068
4069	static inline int bt_subprog_enter(struct backtrack_state *bt)
4070	{
4071	if (bt->frame == MAX_CALL_FRAMES - 1) {
4072	verifier_bug(bt->env, "subprog enter from frame %d", bt->frame);
4073	return -EFAULT;
4074	}
4075	bt->frame++;
4076	return 0;
4077	}
4078
4079	static inline int bt_subprog_exit(struct backtrack_state *bt)
4080	{
4081	if (bt->frame == 0) {
4082	verifier_bug(bt->env, "subprog exit from frame 0");
4083	return -EFAULT;
4084	}
4085	bt->frame--;
4086	return 0;
4087	}
4088
4089	static inline void bt_set_frame_reg(struct backtrack_state *bt, u32 frame, u32 reg)
4090	{
4091	bt->reg_masks[frame] |= 1 << reg;
4092	}
4093
4094	static inline void bt_clear_frame_reg(struct backtrack_state *bt, u32 frame, u32 reg)
4095	{
4096	bt->reg_masks[frame] &= ~(1 << reg);
4097	}
4098
4099	static inline void bt_set_reg(struct backtrack_state *bt, u32 reg)
4100	{
4101	bt_set_frame_reg(bt, frame: bt->frame, reg);
4102	}
4103
4104	static inline void bt_clear_reg(struct backtrack_state *bt, u32 reg)
4105	{
4106	bt_clear_frame_reg(bt, frame: bt->frame, reg);
4107	}
4108
4109	static inline void bt_set_frame_slot(struct backtrack_state *bt, u32 frame, u32 slot)
4110	{
4111	bt->stack_masks[frame] |= 1ull << slot;
4112	}
4113
4114	static inline void bt_clear_frame_slot(struct backtrack_state *bt, u32 frame, u32 slot)
4115	{
4116	bt->stack_masks[frame] &= ~(1ull << slot);
4117	}
4118
4119	static inline u32 bt_frame_reg_mask(struct backtrack_state *bt, u32 frame)
4120	{
4121	return bt->reg_masks[frame];
4122	}
4123
4124	static inline u32 bt_reg_mask(struct backtrack_state *bt)
4125	{
4126	return bt->reg_masks[bt->frame];
4127	}
4128
4129	static inline u64 bt_frame_stack_mask(struct backtrack_state *bt, u32 frame)
4130	{
4131	return bt->stack_masks[frame];
4132	}
4133
4134	static inline u64 bt_stack_mask(struct backtrack_state *bt)
4135	{
4136	return bt->stack_masks[bt->frame];
4137	}
4138
4139	static inline bool bt_is_reg_set(struct backtrack_state *bt, u32 reg)
4140	{
4141	return bt->reg_masks[bt->frame] & (1 << reg);
4142	}
4143
4144	static inline bool bt_is_frame_reg_set(struct backtrack_state *bt, u32 frame, u32 reg)
4145	{
4146	return bt->reg_masks[frame] & (1 << reg);
4147	}
4148
4149	static inline bool bt_is_frame_slot_set(struct backtrack_state *bt, u32 frame, u32 slot)
4150	{
4151	return bt->stack_masks[frame] & (1ull << slot);
4152	}
4153
4154	/* format registers bitmask, e.g., "r0,r2,r4" for 0x15 mask */
4155	static void fmt_reg_mask(char *buf, ssize_t buf_sz, u32 reg_mask)
4156	{
4157	DECLARE_BITMAP(mask, 64);
4158	bool first = true;
4159	int i, n;
4160
4161	buf[0] = '\0';
4162
4163	bitmap_from_u64(dst: mask, mask: reg_mask);
4164	for_each_set_bit(i, mask, 32) {
4165	n = snprintf(buf, size: buf_sz, fmt: "%sr%d", first ? "" : ",", i);
4166	first = false;
4167	buf += n;
4168	buf_sz -= n;
4169	if (buf_sz < 0)
4170	break;
4171	}
4172	}
4173	/* format stack slots bitmask, e.g., "-8,-24,-40" for 0x15 mask */
4174	void bpf_fmt_stack_mask(char *buf, ssize_t buf_sz, u64 stack_mask)
4175	{
4176	DECLARE_BITMAP(mask, 64);
4177	bool first = true;
4178	int i, n;
4179
4180	buf[0] = '\0';
4181
4182	bitmap_from_u64(dst: mask, mask: stack_mask);
4183	for_each_set_bit(i, mask, 64) {
4184	n = snprintf(buf, size: buf_sz, fmt: "%s%d", first ? "" : ",", -(i + 1) * 8);
4185	first = false;
4186	buf += n;
4187	buf_sz -= n;
4188	if (buf_sz < 0)
4189	break;
4190	}
4191	}
4192
4193	/* If any register R in hist->linked_regs is marked as precise in bt,
4194	* do bt_set_frame_{reg,slot}(bt, R) for all registers in hist->linked_regs.
4195	*/
4196	static void bt_sync_linked_regs(struct backtrack_state *bt, struct bpf_jmp_history_entry *hist)
4197	{
4198	struct linked_regs linked_regs;
4199	bool some_precise = false;
4200	int i;
4201
4202	if (!hist || hist->linked_regs == 0)
4203	return;
4204
4205	linked_regs_unpack(val: hist->linked_regs, s: &linked_regs);
4206	for (i = 0; i < linked_regs.cnt; ++i) {
4207	struct linked_reg *e = &linked_regs.entries[i];
4208
4209	if ((e->is_reg && bt_is_frame_reg_set(bt, frame: e->frameno, reg: e->regno)) ||
4210	(!e->is_reg && bt_is_frame_slot_set(bt, frame: e->frameno, slot: e->spi))) {
4211	some_precise = true;
4212	break;
4213	}
4214	}
4215
4216	if (!some_precise)
4217	return;
4218
4219	for (i = 0; i < linked_regs.cnt; ++i) {
4220	struct linked_reg *e = &linked_regs.entries[i];
4221
4222	if (e->is_reg)
4223	bt_set_frame_reg(bt, frame: e->frameno, reg: e->regno);
4224	else
4225	bt_set_frame_slot(bt, frame: e->frameno, slot: e->spi);
4226	}
4227	}
4228
4229	/* For given verifier state backtrack_insn() is called from the last insn to
4230	* the first insn. Its purpose is to compute a bitmask of registers and
4231	* stack slots that needs precision in the parent verifier state.
4232	*
4233	* @idx is an index of the instruction we are currently processing;
4234	* @subseq_idx is an index of the subsequent instruction that:
4235	* - *would be* executed next, if jump history is viewed in forward order;
4236	* - *was* processed previously during backtracking.
4237	*/
4238	static int backtrack_insn(struct bpf_verifier_env *env, int idx, int subseq_idx,
4239	struct bpf_jmp_history_entry *hist, struct backtrack_state *bt)
4240	{
4241	struct bpf_insn *insn = env->prog->insnsi + idx;
4242	u8 class = BPF_CLASS(insn->code);
4243	u8 opcode = BPF_OP(insn->code);
4244	u8 mode = BPF_MODE(insn->code);
4245	u32 dreg = insn->dst_reg;
4246	u32 sreg = insn->src_reg;
4247	u32 spi, i, fr;
4248
4249	if (insn->code == 0)
4250	return 0;
4251	if (env->log.level & BPF_LOG_LEVEL2) {
4252	fmt_reg_mask(buf: env->tmp_str_buf, TMP_STR_BUF_LEN, reg_mask: bt_reg_mask(bt));
4253	verbose(private_data: env, fmt: "mark_precise: frame%d: regs=%s ",
4254	bt->frame, env->tmp_str_buf);
4255	bpf_fmt_stack_mask(buf: env->tmp_str_buf, TMP_STR_BUF_LEN, stack_mask: bt_stack_mask(bt));
4256	verbose(private_data: env, fmt: "stack=%s before ", env->tmp_str_buf);
4257	verbose(private_data: env, fmt: "%d: ", idx);
4258	verbose_insn(env, insn);
4259	}
4260
4261	/* If there is a history record that some registers gained range at this insn,
4262	* propagate precision marks to those registers, so that bt_is_reg_set()
4263	* accounts for these registers.
4264	*/
4265	bt_sync_linked_regs(bt, hist);
4266
4267	if (class == BPF_ALU || class == BPF_ALU64) {
4268	if (!bt_is_reg_set(bt, reg: dreg))
4269	return 0;
4270	if (opcode == BPF_END || opcode == BPF_NEG) {
4271	/* sreg is reserved and unused
4272	* dreg still need precision before this insn
4273	*/
4274	return 0;
4275	} else if (opcode == BPF_MOV) {
4276	if (BPF_SRC(insn->code) == BPF_X) {
4277	/* dreg = sreg or dreg = (s8, s16, s32)sreg
4278	* dreg needs precision after this insn
4279	* sreg needs precision before this insn
4280	*/
4281	bt_clear_reg(bt, reg: dreg);
4282	if (sreg != BPF_REG_FP)
4283	bt_set_reg(bt, reg: sreg);
4284	} else {
4285	/* dreg = K
4286	* dreg needs precision after this insn.
4287	* Corresponding register is already marked
4288	* as precise=true in this verifier state.
4289	* No further markings in parent are necessary
4290	*/
4291	bt_clear_reg(bt, reg: dreg);
4292	}
4293	} else {
4294	if (BPF_SRC(insn->code) == BPF_X) {
4295	/* dreg += sreg
4296	* both dreg and sreg need precision
4297	* before this insn
4298	*/
4299	if (sreg != BPF_REG_FP)
4300	bt_set_reg(bt, reg: sreg);
4301	} /* else dreg += K
4302	* dreg still needs precision before this insn
4303	*/
4304	}
4305	} else if (class == BPF_LDX || is_atomic_load_insn(insn)) {
4306	if (!bt_is_reg_set(bt, reg: dreg))
4307	return 0;
4308	bt_clear_reg(bt, reg: dreg);
4309
4310	/* scalars can only be spilled into stack w/o losing precision.
4311	* Load from any other memory can be zero extended.
4312	* The desire to keep that precision is already indicated
4313	* by 'precise' mark in corresponding register of this state.
4314	* No further tracking necessary.
4315	*/
4316	if (!hist || !(hist->flags & INSN_F_STACK_ACCESS))
4317	return 0;
4318	/* dreg = *(u64 *)[fp - off] was a fill from the stack.
4319	* that [fp - off] slot contains scalar that needs to be
4320	* tracked with precision
4321	*/
4322	spi = insn_stack_access_spi(insn_flags: hist->flags);
4323	fr = insn_stack_access_frameno(insn_flags: hist->flags);
4324	bt_set_frame_slot(bt, frame: fr, slot: spi);
4325	} else if (class == BPF_STX || class == BPF_ST) {
4326	if (bt_is_reg_set(bt, reg: dreg))
4327	/* stx & st shouldn't be using _scalar_ dst_reg
4328	* to access memory. It means backtracking
4329	* encountered a case of pointer subtraction.
4330	*/
4331	return -ENOTSUPP;
4332	/* scalars can only be spilled into stack */
4333	if (!hist || !(hist->flags & INSN_F_STACK_ACCESS))
4334	return 0;
4335	spi = insn_stack_access_spi(insn_flags: hist->flags);
4336	fr = insn_stack_access_frameno(insn_flags: hist->flags);
4337	if (!bt_is_frame_slot_set(bt, frame: fr, slot: spi))
4338	return 0;
4339	bt_clear_frame_slot(bt, frame: fr, slot: spi);
4340	if (class == BPF_STX)
4341	bt_set_reg(bt, reg: sreg);
4342	} else if (class == BPF_JMP || class == BPF_JMP32) {
4343	if (bpf_pseudo_call(insn)) {
4344	int subprog_insn_idx, subprog;
4345
4346	subprog_insn_idx = idx + insn->imm + 1;
4347	subprog = find_subprog(env, off: subprog_insn_idx);
4348	if (subprog < 0)
4349	return -EFAULT;
4350
4351	if (subprog_is_global(env, subprog)) {
4352	/* check that jump history doesn't have any
4353	* extra instructions from subprog; the next
4354	* instruction after call to global subprog
4355	* should be literally next instruction in
4356	* caller program
4357	*/
4358	verifier_bug_if(idx + 1 != subseq_idx, env,
4359	"extra insn from subprog");
4360	/* r1-r5 are invalidated after subprog call,
4361	* so for global func call it shouldn't be set
4362	* anymore
4363	*/
4364	if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) {
4365	verifier_bug(env, "global subprog unexpected regs %x",
4366	bt_reg_mask(bt));
4367	return -EFAULT;
4368	}
4369	/* global subprog always sets R0 */
4370	bt_clear_reg(bt, reg: BPF_REG_0);
4371	return 0;
4372	} else {
4373	/* static subprog call instruction, which
4374	* means that we are exiting current subprog,
4375	* so only r1-r5 could be still requested as
4376	* precise, r0 and r6-r10 or any stack slot in
4377	* the current frame should be zero by now
4378	*/
4379	if (bt_reg_mask(bt) & ~BPF_REGMASK_ARGS) {
4380	verifier_bug(env, "static subprog unexpected regs %x",
4381	bt_reg_mask(bt));
4382	return -EFAULT;
4383	}
4384	/* we are now tracking register spills correctly,
4385	* so any instance of leftover slots is a bug
4386	*/
4387	if (bt_stack_mask(bt) != 0) {
4388	verifier_bug(env,
4389	"static subprog leftover stack slots %llx",
4390	bt_stack_mask(bt));
4391	return -EFAULT;
4392	}
4393	/* propagate r1-r5 to the caller */
4394	for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
4395	if (bt_is_reg_set(bt, reg: i)) {
4396	bt_clear_reg(bt, reg: i);
4397	bt_set_frame_reg(bt, frame: bt->frame - 1, reg: i);
4398	}
4399	}
4400	if (bt_subprog_exit(bt))
4401	return -EFAULT;
4402	return 0;
4403	}
4404	} else if (is_sync_callback_calling_insn(insn) && idx != subseq_idx - 1) {
4405	/* exit from callback subprog to callback-calling helper or
4406	* kfunc call. Use idx/subseq_idx check to discern it from
4407	* straight line code backtracking.
4408	* Unlike the subprog call handling above, we shouldn't
4409	* propagate precision of r1-r5 (if any requested), as they are
4410	* not actually arguments passed directly to callback subprogs
4411	*/
4412	if (bt_reg_mask(bt) & ~BPF_REGMASK_ARGS) {
4413	verifier_bug(env, "callback unexpected regs %x",
4414	bt_reg_mask(bt));
4415	return -EFAULT;
4416	}
4417	if (bt_stack_mask(bt) != 0) {
4418	verifier_bug(env, "callback leftover stack slots %llx",
4419	bt_stack_mask(bt));
4420	return -EFAULT;
4421	}
4422	/* clear r1-r5 in callback subprog's mask */
4423	for (i = BPF_REG_1; i <= BPF_REG_5; i++)
4424	bt_clear_reg(bt, reg: i);
4425	if (bt_subprog_exit(bt))
4426	return -EFAULT;
4427	return 0;
4428	} else if (opcode == BPF_CALL) {
4429	/* kfunc with imm==0 is invalid and fixup_kfunc_call will
4430	* catch this error later. Make backtracking conservative
4431	* with ENOTSUPP.
4432	*/
4433	if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL && insn->imm == 0)
4434	return -ENOTSUPP;
4435	/* regular helper call sets R0 */
4436	bt_clear_reg(bt, reg: BPF_REG_0);
4437	if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) {
4438	/* if backtracking was looking for registers R1-R5
4439	* they should have been found already.
4440	*/
4441	verifier_bug(env, "backtracking call unexpected regs %x",
4442	bt_reg_mask(bt));
4443	return -EFAULT;
4444	}
4445	if (insn->src_reg == BPF_REG_0 && insn->imm == BPF_FUNC_tail_call
4446	&& subseq_idx - idx != 1) {
4447	if (bt_subprog_enter(bt))
4448	return -EFAULT;
4449	}
4450	} else if (opcode == BPF_EXIT) {
4451	bool r0_precise;
4452
4453	/* Backtracking to a nested function call, 'idx' is a part of
4454	* the inner frame 'subseq_idx' is a part of the outer frame.
4455	* In case of a regular function call, instructions giving
4456	* precision to registers R1-R5 should have been found already.
4457	* In case of a callback, it is ok to have R1-R5 marked for
4458	* backtracking, as these registers are set by the function
4459	* invoking callback.
4460	*/
4461	if (subseq_idx >= 0 && bpf_calls_callback(env, insn_idx: subseq_idx))
4462	for (i = BPF_REG_1; i <= BPF_REG_5; i++)
4463	bt_clear_reg(bt, reg: i);
4464	if (bt_reg_mask(bt) & BPF_REGMASK_ARGS) {
4465	verifier_bug(env, "backtracking exit unexpected regs %x",
4466	bt_reg_mask(bt));
4467	return -EFAULT;
4468	}
4469
4470	/* BPF_EXIT in subprog or callback always returns
4471	* right after the call instruction, so by checking
4472	* whether the instruction at subseq_idx-1 is subprog
4473	* call or not we can distinguish actual exit from
4474	* *subprog* from exit from *callback*. In the former
4475	* case, we need to propagate r0 precision, if
4476	* necessary. In the former we never do that.
4477	*/
4478	r0_precise = subseq_idx - 1 >= 0 &&
4479	bpf_pseudo_call(insn: &env->prog->insnsi[subseq_idx - 1]) &&
4480	bt_is_reg_set(bt, reg: BPF_REG_0);
4481
4482	bt_clear_reg(bt, reg: BPF_REG_0);
4483	if (bt_subprog_enter(bt))
4484	return -EFAULT;
4485
4486	if (r0_precise)
4487	bt_set_reg(bt, reg: BPF_REG_0);
4488	/* r6-r9 and stack slots will stay set in caller frame
4489	* bitmasks until we return back from callee(s)
4490	*/
4491	return 0;
4492	} else if (BPF_SRC(insn->code) == BPF_X) {
4493	if (!bt_is_reg_set(bt, reg: dreg) && !bt_is_reg_set(bt, reg: sreg))
4494	return 0;
4495	/* dreg <cond> sreg
4496	* Both dreg and sreg need precision before
4497	* this insn. If only sreg was marked precise
4498	* before it would be equally necessary to
4499	* propagate it to dreg.
4500	*/
4501	if (!hist || !(hist->flags & INSN_F_SRC_REG_STACK))
4502	bt_set_reg(bt, reg: sreg);
4503	if (!hist || !(hist->flags & INSN_F_DST_REG_STACK))
4504	bt_set_reg(bt, reg: dreg);
4505	} else if (BPF_SRC(insn->code) == BPF_K) {
4506	/* dreg <cond> K
4507	* Only dreg still needs precision before
4508	* this insn, so for the K-based conditional
4509	* there is nothing new to be marked.
4510	*/
4511	}
4512	} else if (class == BPF_LD) {
4513	if (!bt_is_reg_set(bt, reg: dreg))
4514	return 0;
4515	bt_clear_reg(bt, reg: dreg);
4516	/* It's ld_imm64 or ld_abs or ld_ind.
4517	* For ld_imm64 no further tracking of precision
4518	* into parent is necessary
4519	*/
4520	if (mode == BPF_IND || mode == BPF_ABS)
4521	/* to be analyzed */
4522	return -ENOTSUPP;
4523	}
4524	/* Propagate precision marks to linked registers, to account for
4525	* registers marked as precise in this function.
4526	*/
4527	bt_sync_linked_regs(bt, hist);
4528	return 0;
4529	}
4530
4531	/* the scalar precision tracking algorithm:
4532	* . at the start all registers have precise=false.
4533	* . scalar ranges are tracked as normal through alu and jmp insns.
4534	* . once precise value of the scalar register is used in:
4535	* . ptr + scalar alu
4536	* . if (scalar cond K|scalar)
4537	* . helper_call(.., scalar, ...) where ARG_CONST is expected
4538	* backtrack through the verifier states and mark all registers and
4539	* stack slots with spilled constants that these scalar registers
4540	* should be precise.
4541	* . during state pruning two registers (or spilled stack slots)
4542	* are equivalent if both are not precise.
4543	*
4544	* Note the verifier cannot simply walk register parentage chain,
4545	* since many different registers and stack slots could have been
4546	* used to compute single precise scalar.
4547	*
4548	* The approach of starting with precise=true for all registers and then
4549	* backtrack to mark a register as not precise when the verifier detects
4550	* that program doesn't care about specific value (e.g., when helper
4551	* takes register as ARG_ANYTHING parameter) is not safe.
4552	*
4553	* It's ok to walk single parentage chain of the verifier states.
4554	* It's possible that this backtracking will go all the way till 1st insn.
4555	* All other branches will be explored for needing precision later.
4556	*
4557	* The backtracking needs to deal with cases like:
4558	* R8=map_value(id=0,off=0,ks=4,vs=1952,imm=0) R9_w=map_value(id=0,off=40,ks=4,vs=1952,imm=0)
4559	* r9 -= r8
4560	* r5 = r9
4561	* if r5 > 0x79f goto pc+7
4562	* R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
4563	* r5 += 1
4564	* ...
4565	* call bpf_perf_event_output#25
4566	* where .arg5_type = ARG_CONST_SIZE_OR_ZERO
4567	*
4568	* and this case:
4569	* r6 = 1
4570	* call foo // uses callee's r6 inside to compute r0
4571	* r0 += r6
4572	* if r0 == 0 goto
4573	*
4574	* to track above reg_mask/stack_mask needs to be independent for each frame.
4575	*
4576	* Also if parent's curframe > frame where backtracking started,
4577	* the verifier need to mark registers in both frames, otherwise callees
4578	* may incorrectly prune callers. This is similar to
4579	* commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
4580	*
4581	* For now backtracking falls back into conservative marking.
4582	*/
4583	static void mark_all_scalars_precise(struct bpf_verifier_env *env,
4584	struct bpf_verifier_state *st)
4585	{
4586	struct bpf_func_state *func;
4587	struct bpf_reg_state *reg;
4588	int i, j;
4589
4590	if (env->log.level & BPF_LOG_LEVEL2) {
4591	verbose(private_data: env, fmt: "mark_precise: frame%d: falling back to forcing all scalars precise\n",
4592	st->curframe);
4593	}
4594
4595	/* big hammer: mark all scalars precise in this path.
4596	* pop_stack may still get !precise scalars.
4597	* We also skip current state and go straight to first parent state,
4598	* because precision markings in current non-checkpointed state are
4599	* not needed. See why in the comment in __mark_chain_precision below.
4600	*/
4601	for (st = st->parent; st; st = st->parent) {
4602	for (i = 0; i <= st->curframe; i++) {
4603	func = st->frame[i];
4604	for (j = 0; j < BPF_REG_FP; j++) {
4605	reg = &func->regs[j];
4606	if (reg->type != SCALAR_VALUE || reg->precise)
4607	continue;
4608	reg->precise = true;
4609	if (env->log.level & BPF_LOG_LEVEL2) {
4610	verbose(private_data: env, fmt: "force_precise: frame%d: forcing r%d to be precise\n",
4611	i, j);
4612	}
4613	}
4614	for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
4615	if (!is_spilled_reg(stack: &func->stack[j]))
4616	continue;
4617	reg = &func->stack[j].spilled_ptr;
4618	if (reg->type != SCALAR_VALUE || reg->precise)
4619	continue;
4620	reg->precise = true;
4621	if (env->log.level & BPF_LOG_LEVEL2) {
4622	verbose(private_data: env, fmt: "force_precise: frame%d: forcing fp%d to be precise\n",
4623	i, -(j + 1) * 8);
4624	}
4625	}
4626	}
4627	}
4628	}
4629
4630	static void mark_all_scalars_imprecise(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
4631	{
4632	struct bpf_func_state *func;
4633	struct bpf_reg_state *reg;
4634	int i, j;
4635
4636	for (i = 0; i <= st->curframe; i++) {
4637	func = st->frame[i];
4638	for (j = 0; j < BPF_REG_FP; j++) {
4639	reg = &func->regs[j];
4640	if (reg->type != SCALAR_VALUE)
4641	continue;
4642	reg->precise = false;
4643	}
4644	for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
4645	if (!is_spilled_reg(stack: &func->stack[j]))
4646	continue;
4647	reg = &func->stack[j].spilled_ptr;
4648	if (reg->type != SCALAR_VALUE)
4649	continue;
4650	reg->precise = false;
4651	}
4652	}
4653	}
4654
4655	/*
4656	* __mark_chain_precision() backtracks BPF program instruction sequence and
4657	* chain of verifier states making sure that register *regno* (if regno >= 0)
4658	* and/or stack slot *spi* (if spi >= 0) are marked as precisely tracked
4659	* SCALARS, as well as any other registers and slots that contribute to
4660	* a tracked state of given registers/stack slots, depending on specific BPF
4661	* assembly instructions (see backtrack_insns() for exact instruction handling
4662	* logic). This backtracking relies on recorded jmp_history and is able to
4663	* traverse entire chain of parent states. This process ends only when all the
4664	* necessary registers/slots and their transitive dependencies are marked as
4665	* precise.
4666	*
4667	* One important and subtle aspect is that precise marks *do not matter* in
4668	* the currently verified state (current state). It is important to understand
4669	* why this is the case.
4670	*
4671	* First, note that current state is the state that is not yet "checkpointed",
4672	* i.e., it is not yet put into env->explored_states, and it has no children
4673	* states as well. It's ephemeral, and can end up either a) being discarded if
4674	* compatible explored state is found at some point or BPF_EXIT instruction is
4675	* reached or b) checkpointed and put into env->explored_states, branching out
4676	* into one or more children states.
4677	*
4678	* In the former case, precise markings in current state are completely
4679	* ignored by state comparison code (see regsafe() for details). Only
4680	* checkpointed ("old") state precise markings are important, and if old
4681	* state's register/slot is precise, regsafe() assumes current state's
4682	* register/slot as precise and checks value ranges exactly and precisely. If
4683	* states turn out to be compatible, current state's necessary precise
4684	* markings and any required parent states' precise markings are enforced
4685	* after the fact with propagate_precision() logic, after the fact. But it's
4686	* important to realize that in this case, even after marking current state
4687	* registers/slots as precise, we immediately discard current state. So what
4688	* actually matters is any of the precise markings propagated into current
4689	* state's parent states, which are always checkpointed (due to b) case above).
4690	* As such, for scenario a) it doesn't matter if current state has precise
4691	* markings set or not.
4692	*
4693	* Now, for the scenario b), checkpointing and forking into child(ren)
4694	* state(s). Note that before current state gets to checkpointing step, any
4695	* processed instruction always assumes precise SCALAR register/slot
4696	* knowledge: if precise value or range is useful to prune jump branch, BPF
4697	* verifier takes this opportunity enthusiastically. Similarly, when
4698	* register's value is used to calculate offset or memory address, exact
4699	* knowledge of SCALAR range is assumed, checked, and enforced. So, similar to
4700	* what we mentioned above about state comparison ignoring precise markings
4701	* during state comparison, BPF verifier ignores and also assumes precise
4702	* markings *at will* during instruction verification process. But as verifier
4703	* assumes precision, it also propagates any precision dependencies across
4704	* parent states, which are not yet finalized, so can be further restricted
4705	* based on new knowledge gained from restrictions enforced by their children
4706	* states. This is so that once those parent states are finalized, i.e., when
4707	* they have no more active children state, state comparison logic in
4708	* is_state_visited() would enforce strict and precise SCALAR ranges, if
4709	* required for correctness.
4710	*
4711	* To build a bit more intuition, note also that once a state is checkpointed,
4712	* the path we took to get to that state is not important. This is crucial
4713	* property for state pruning. When state is checkpointed and finalized at
4714	* some instruction index, it can be correctly and safely used to "short
4715	* circuit" any *compatible* state that reaches exactly the same instruction
4716	* index. I.e., if we jumped to that instruction from a completely different
4717	* code path than original finalized state was derived from, it doesn't
4718	* matter, current state can be discarded because from that instruction
4719	* forward having a compatible state will ensure we will safely reach the
4720	* exit. States describe preconditions for further exploration, but completely
4721	* forget the history of how we got here.
4722	*
4723	* This also means that even if we needed precise SCALAR range to get to
4724	* finalized state, but from that point forward *that same* SCALAR register is
4725	* never used in a precise context (i.e., it's precise value is not needed for
4726	* correctness), it's correct and safe to mark such register as "imprecise"
4727	* (i.e., precise marking set to false). This is what we rely on when we do
4728	* not set precise marking in current state. If no child state requires
4729	* precision for any given SCALAR register, it's safe to dictate that it can
4730	* be imprecise. If any child state does require this register to be precise,
4731	* we'll mark it precise later retroactively during precise markings
4732	* propagation from child state to parent states.
4733	*
4734	* Skipping precise marking setting in current state is a mild version of
4735	* relying on the above observation. But we can utilize this property even
4736	* more aggressively by proactively forgetting any precise marking in the
4737	* current state (which we inherited from the parent state), right before we
4738	* checkpoint it and branch off into new child state. This is done by
4739	* mark_all_scalars_imprecise() to hopefully get more permissive and generic
4740	* finalized states which help in short circuiting more future states.
4741	*/
4742	static int __mark_chain_precision(struct bpf_verifier_env *env,
4743	struct bpf_verifier_state *starting_state,
4744	int regno,
4745	bool *changed)
4746	{
4747	struct bpf_verifier_state *st = starting_state;
4748	struct backtrack_state *bt = &env->bt;
4749	int first_idx = st->first_insn_idx;
4750	int last_idx = starting_state->insn_idx;
4751	int subseq_idx = -1;
4752	struct bpf_func_state *func;
4753	bool tmp, skip_first = true;
4754	struct bpf_reg_state *reg;
4755	int i, fr, err;
4756
4757	if (!env->bpf_capable)
4758	return 0;
4759
4760	changed = changed ?: &tmp;
4761	/* set frame number from which we are starting to backtrack */
4762	bt_init(bt, frame: starting_state->curframe);
4763
4764	/* Do sanity checks against current state of register and/or stack
4765	* slot, but don't set precise flag in current state, as precision
4766	* tracking in the current state is unnecessary.
4767	*/
4768	func = st->frame[bt->frame];
4769	if (regno >= 0) {
4770	reg = &func->regs[regno];
4771	if (reg->type != SCALAR_VALUE) {
4772	verifier_bug(env, "backtracking misuse");
4773	return -EFAULT;
4774	}
4775	bt_set_reg(bt, reg: regno);
4776	}
4777
4778	if (bt_empty(bt))
4779	return 0;
4780
4781	for (;;) {
4782	DECLARE_BITMAP(mask, 64);
4783	u32 history = st->jmp_history_cnt;
4784	struct bpf_jmp_history_entry *hist;
4785
4786	if (env->log.level & BPF_LOG_LEVEL2) {
4787	verbose(private_data: env, fmt: "mark_precise: frame%d: last_idx %d first_idx %d subseq_idx %d \n",
4788	bt->frame, last_idx, first_idx, subseq_idx);
4789	}
4790
4791	if (last_idx < 0) {
4792	/* we are at the entry into subprog, which
4793	* is expected for global funcs, but only if
4794	* requested precise registers are R1-R5
4795	* (which are global func's input arguments)
4796	*/
4797	if (st->curframe == 0 &&
4798	st->frame[0]->subprogno > 0 &&
4799	st->frame[0]->callsite == BPF_MAIN_FUNC &&
4800	bt_stack_mask(bt) == 0 &&
4801	(bt_reg_mask(bt) & ~BPF_REGMASK_ARGS) == 0) {
4802	bitmap_from_u64(dst: mask, mask: bt_reg_mask(bt));
4803	for_each_set_bit(i, mask, 32) {
4804	reg = &st->frame[0]->regs[i];
4805	bt_clear_reg(bt, reg: i);
4806	if (reg->type == SCALAR_VALUE) {
4807	reg->precise = true;
4808	*changed = true;
4809	}
4810	}
4811	return 0;
4812	}
4813
4814	verifier_bug(env, "backtracking func entry subprog %d reg_mask %x stack_mask %llx",
4815	st->frame[0]->subprogno, bt_reg_mask(bt), bt_stack_mask(bt));
4816	return -EFAULT;
4817	}
4818
4819	for (i = last_idx;;) {
4820	if (skip_first) {
4821	err = 0;
4822	skip_first = false;
4823	} else {
4824	hist = get_jmp_hist_entry(st, hist_end: history, insn_idx: i);
4825	err = backtrack_insn(env, idx: i, subseq_idx, hist, bt);
4826	}
4827	if (err == -ENOTSUPP) {
4828	mark_all_scalars_precise(env, st: starting_state);
4829	bt_reset(bt);
4830	return 0;
4831	} else if (err) {
4832	return err;
4833	}
4834	if (bt_empty(bt))
4835	/* Found assignment(s) into tracked register in this state.
4836	* Since this state is already marked, just return.
4837	* Nothing to be tracked further in the parent state.
4838	*/
4839	return 0;
4840	subseq_idx = i;
4841	i = get_prev_insn_idx(st, i, history: &history);
4842	if (i == -ENOENT)
4843	break;
4844	if (i >= env->prog->len) {
4845	/* This can happen if backtracking reached insn 0
4846	* and there are still reg_mask or stack_mask
4847	* to backtrack.
4848	* It means the backtracking missed the spot where
4849	* particular register was initialized with a constant.
4850	*/
4851	verifier_bug(env, "backtracking idx %d", i);
4852	return -EFAULT;
4853	}
4854	}
4855	st = st->parent;
4856	if (!st)
4857	break;
4858
4859	for (fr = bt->frame; fr >= 0; fr--) {
4860	func = st->frame[fr];
4861	bitmap_from_u64(dst: mask, mask: bt_frame_reg_mask(bt, frame: fr));
4862	for_each_set_bit(i, mask, 32) {
4863	reg = &func->regs[i];
4864	if (reg->type != SCALAR_VALUE) {
4865	bt_clear_frame_reg(bt, frame: fr, reg: i);
4866	continue;
4867	}
4868	if (reg->precise) {
4869	bt_clear_frame_reg(bt, frame: fr, reg: i);
4870	} else {
4871	reg->precise = true;
4872	*changed = true;
4873	}
4874	}
4875
4876	bitmap_from_u64(dst: mask, mask: bt_frame_stack_mask(bt, frame: fr));
4877	for_each_set_bit(i, mask, 64) {
4878	if (verifier_bug_if(i >= func->allocated_stack / BPF_REG_SIZE,
4879	env, "stack slot %d, total slots %d",
4880	i, func->allocated_stack / BPF_REG_SIZE))
4881	return -EFAULT;
4882
4883	if (!is_spilled_scalar_reg(stack: &func->stack[i])) {
4884	bt_clear_frame_slot(bt, frame: fr, slot: i);
4885	continue;
4886	}
4887	reg = &func->stack[i].spilled_ptr;
4888	if (reg->precise) {
4889	bt_clear_frame_slot(bt, frame: fr, slot: i);
4890	} else {
4891	reg->precise = true;
4892	*changed = true;
4893	}
4894	}
4895	if (env->log.level & BPF_LOG_LEVEL2) {
4896	fmt_reg_mask(buf: env->tmp_str_buf, TMP_STR_BUF_LEN,
4897	reg_mask: bt_frame_reg_mask(bt, frame: fr));
4898	verbose(private_data: env, fmt: "mark_precise: frame%d: parent state regs=%s ",
4899	fr, env->tmp_str_buf);
4900	bpf_fmt_stack_mask(buf: env->tmp_str_buf, TMP_STR_BUF_LEN,
4901	stack_mask: bt_frame_stack_mask(bt, frame: fr));
4902	verbose(private_data: env, fmt: "stack=%s: ", env->tmp_str_buf);
4903	print_verifier_state(env, vstate: st, frameno: fr, print_all: true);
4904	}
4905	}
4906
4907	if (bt_empty(bt))
4908	return 0;
4909
4910	subseq_idx = first_idx;
4911	last_idx = st->last_insn_idx;
4912	first_idx = st->first_insn_idx;
4913	}
4914
4915	/* if we still have requested precise regs or slots, we missed
4916	* something (e.g., stack access through non-r10 register), so
4917	* fallback to marking all precise
4918	*/
4919	if (!bt_empty(bt)) {
4920	mark_all_scalars_precise(env, st: starting_state);
4921	bt_reset(bt);
4922	}
4923
4924	return 0;
4925	}
4926
4927	int mark_chain_precision(struct bpf_verifier_env *env, int regno)
4928	{
4929	return __mark_chain_precision(env, starting_state: env->cur_state, regno, NULL);
4930	}
4931
4932	/* mark_chain_precision_batch() assumes that env->bt is set in the caller to
4933	* desired reg and stack masks across all relevant frames
4934	*/
4935	static int mark_chain_precision_batch(struct bpf_verifier_env *env,
4936	struct bpf_verifier_state *starting_state)
4937	{
4938	return __mark_chain_precision(env, starting_state, regno: -1, NULL);
4939	}
4940
4941	static bool is_spillable_regtype(enum bpf_reg_type type)
4942	{
4943	switch (base_type(type)) {
4944	case PTR_TO_MAP_VALUE:
4945	case PTR_TO_STACK:
4946	case PTR_TO_CTX:
4947	case PTR_TO_PACKET:
4948	case PTR_TO_PACKET_META:
4949	case PTR_TO_PACKET_END:
4950	case PTR_TO_FLOW_KEYS:
4951	case CONST_PTR_TO_MAP:
4952	case PTR_TO_SOCKET:
4953	case PTR_TO_SOCK_COMMON:
4954	case PTR_TO_TCP_SOCK:
4955	case PTR_TO_XDP_SOCK:
4956	case PTR_TO_BTF_ID:
4957	case PTR_TO_BUF:
4958	case PTR_TO_MEM:
4959	case PTR_TO_FUNC:
4960	case PTR_TO_MAP_KEY:
4961	case PTR_TO_ARENA:
4962	return true;
4963	default:
4964	return false;
4965	}
4966	}
4967
4968	/* Does this register contain a constant zero? */
4969	static bool register_is_null(struct bpf_reg_state *reg)
4970	{
4971	return reg->type == SCALAR_VALUE && tnum_equals_const(a: reg->var_off, b: 0);
4972	}
4973
4974	/* check if register is a constant scalar value */
4975	static bool is_reg_const(struct bpf_reg_state *reg, bool subreg32)
4976	{
4977	return reg->type == SCALAR_VALUE &&
4978	tnum_is_const(a: subreg32 ? tnum_subreg(a: reg->var_off) : reg->var_off);
4979	}
4980
4981	/* assuming is_reg_const() is true, return constant value of a register */
4982	static u64 reg_const_value(struct bpf_reg_state *reg, bool subreg32)
4983	{
4984	return subreg32 ? tnum_subreg(a: reg->var_off).value : reg->var_off.value;
4985	}
4986
4987	static bool __is_pointer_value(bool allow_ptr_leaks,
4988	const struct bpf_reg_state *reg)
4989	{
4990	if (allow_ptr_leaks)
4991	return false;
4992
4993	return reg->type != SCALAR_VALUE;
4994	}
4995
4996	static void assign_scalar_id_before_mov(struct bpf_verifier_env *env,
4997	struct bpf_reg_state *src_reg)
4998	{
4999	if (src_reg->type != SCALAR_VALUE)
5000	return;
5001
5002	if (src_reg->id & BPF_ADD_CONST) {
5003	/*
5004	* The verifier is processing rX = rY insn and
5005	* rY->id has special linked register already.
5006	* Cleared it, since multiple rX += const are not supported.
5007	*/
5008	src_reg->id = 0;
5009	src_reg->off = 0;
5010	}
5011
5012	if (!src_reg->id && !tnum_is_const(a: src_reg->var_off))
5013	/* Ensure that src_reg has a valid ID that will be copied to
5014	* dst_reg and then will be used by sync_linked_regs() to
5015	* propagate min/max range.
5016	*/
5017	src_reg->id = ++env->id_gen;
5018	}
5019
5020	/* Copy src state preserving dst->parent and dst->live fields */
5021	static void copy_register_state(struct bpf_reg_state *dst, const struct bpf_reg_state *src)
5022	{
5023	*dst = *src;
5024	}
5025
5026	static void save_register_state(struct bpf_verifier_env *env,
5027	struct bpf_func_state *state,
5028	int spi, struct bpf_reg_state *reg,
5029	int size)
5030	{
5031	int i;
5032
5033	copy_register_state(dst: &state->stack[spi].spilled_ptr, src: reg);
5034
5035	for (i = BPF_REG_SIZE; i > BPF_REG_SIZE - size; i--)
5036	state->stack[spi].slot_type[i - 1] = STACK_SPILL;
5037
5038	/* size < 8 bytes spill */
5039	for (; i; i--)
5040	mark_stack_slot_misc(env, stype: &state->stack[spi].slot_type[i - 1]);
5041	}
5042
5043	static bool is_bpf_st_mem(struct bpf_insn *insn)
5044	{
5045	return BPF_CLASS(insn->code) == BPF_ST && BPF_MODE(insn->code) == BPF_MEM;
5046	}
5047
5048	static int get_reg_width(struct bpf_reg_state *reg)
5049	{
5050	return fls64(x: reg->umax_value);
5051	}
5052
5053	/* See comment for mark_fastcall_pattern_for_call() */
5054	static void check_fastcall_stack_contract(struct bpf_verifier_env *env,
5055	struct bpf_func_state *state, int insn_idx, int off)
5056	{
5057	struct bpf_subprog_info *subprog = &env->subprog_info[state->subprogno];
5058	struct bpf_insn_aux_data *aux = env->insn_aux_data;
5059	int i;
5060
5061	if (subprog->fastcall_stack_off <= off || aux[insn_idx].fastcall_pattern)
5062	return;
5063	/* access to the region [max_stack_depth .. fastcall_stack_off)
5064	* from something that is not a part of the fastcall pattern,
5065	* disable fastcall rewrites for current subprogram by setting
5066	* fastcall_stack_off to a value smaller than any possible offset.
5067	*/
5068	subprog->fastcall_stack_off = S16_MIN;
5069	/* reset fastcall aux flags within subprogram,
5070	* happens at most once per subprogram
5071	*/
5072	for (i = subprog->start; i < (subprog + 1)->start; ++i) {
5073	aux[i].fastcall_spills_num = 0;
5074	aux[i].fastcall_pattern = 0;
5075	}
5076	}
5077
5078	/* check_stack_{read,write}_fixed_off functions track spill/fill of registers,
5079	* stack boundary and alignment are checked in check_mem_access()
5080	*/
5081	static int check_stack_write_fixed_off(struct bpf_verifier_env *env,
5082	/* stack frame we're writing to */
5083	struct bpf_func_state *state,
5084	int off, int size, int value_regno,
5085	int insn_idx)
5086	{
5087	struct bpf_func_state *cur; /* state of the current function */
5088	int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
5089	struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
5090	struct bpf_reg_state *reg = NULL;
5091	int insn_flags = insn_stack_access_flags(frameno: state->frameno, spi);
5092
5093	/* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
5094	* so it's aligned access and [off, off + size) are within stack limits
5095	*/
5096	if (!env->allow_ptr_leaks &&
5097	is_spilled_reg(stack: &state->stack[spi]) &&
5098	!is_spilled_scalar_reg(stack: &state->stack[spi]) &&
5099	size != BPF_REG_SIZE) {
5100	verbose(private_data: env, fmt: "attempt to corrupt spilled pointer on stack\n");
5101	return -EACCES;
5102	}
5103
5104	cur = env->cur_state->frame[env->cur_state->curframe];
5105	if (value_regno >= 0)
5106	reg = &cur->regs[value_regno];
5107	if (!env->bypass_spec_v4) {
5108	bool sanitize = reg && is_spillable_regtype(type: reg->type);
5109
5110	for (i = 0; i < size; i++) {
5111	u8 type = state->stack[spi].slot_type[i];
5112
5113	if (type != STACK_MISC && type != STACK_ZERO) {
5114	sanitize = true;
5115	break;
5116	}
5117	}
5118
5119	if (sanitize)
5120	env->insn_aux_data[insn_idx].nospec_result = true;
5121	}
5122
5123	err = destroy_if_dynptr_stack_slot(env, state, spi);
5124	if (err)
5125	return err;
5126
5127	if (!(off % BPF_REG_SIZE) && size == BPF_REG_SIZE) {
5128	/* only mark the slot as written if all 8 bytes were written
5129	* otherwise read propagation may incorrectly stop too soon
5130	* when stack slots are partially written.
5131	* This heuristic means that read propagation will be
5132	* conservative, since it will add reg_live_read marks
5133	* to stack slots all the way to first state when programs
5134	* writes+reads less than 8 bytes
5135	*/
5136	bpf_mark_stack_write(env, frameno: state->frameno, BIT(spi));
5137	}
5138
5139	check_fastcall_stack_contract(env, state, insn_idx, off);
5140	mark_stack_slot_scratched(env, spi);
5141	if (reg && !(off % BPF_REG_SIZE) && reg->type == SCALAR_VALUE && env->bpf_capable) {
5142	bool reg_value_fits;
5143
5144	reg_value_fits = get_reg_width(reg) <= BITS_PER_BYTE * size;
5145	/* Make sure that reg had an ID to build a relation on spill. */
5146	if (reg_value_fits)
5147	assign_scalar_id_before_mov(env, src_reg: reg);
5148	save_register_state(env, state, spi, reg, size);
5149	/* Break the relation on a narrowing spill. */
5150	if (!reg_value_fits)
5151	state->stack[spi].spilled_ptr.id = 0;
5152	} else if (!reg && !(off % BPF_REG_SIZE) && is_bpf_st_mem(insn) &&
5153	env->bpf_capable) {
5154	struct bpf_reg_state *tmp_reg = &env->fake_reg[0];
5155
5156	memset(tmp_reg, 0, sizeof(*tmp_reg));
5157	__mark_reg_known(reg: tmp_reg, imm: insn->imm);
5158	tmp_reg->type = SCALAR_VALUE;
5159	save_register_state(env, state, spi, reg: tmp_reg, size);
5160	} else if (reg && is_spillable_regtype(type: reg->type)) {
5161	/* register containing pointer is being spilled into stack */
5162	if (size != BPF_REG_SIZE) {
5163	verbose_linfo(env, insn_off: insn_idx, prefix_fmt: "; ");
5164	verbose(private_data: env, fmt: "invalid size of register spill\n");
5165	return -EACCES;
5166	}
5167	if (state != cur && reg->type == PTR_TO_STACK) {
5168	verbose(private_data: env, fmt: "cannot spill pointers to stack into stack frame of the caller\n");
5169	return -EINVAL;
5170	}
5171	save_register_state(env, state, spi, reg, size);
5172	} else {
5173	u8 type = STACK_MISC;
5174
5175	/* regular write of data into stack destroys any spilled ptr */
5176	state->stack[spi].spilled_ptr.type = NOT_INIT;
5177	/* Mark slots as STACK_MISC if they belonged to spilled ptr/dynptr/iter. */
5178	if (is_stack_slot_special(stack: &state->stack[spi]))
5179	for (i = 0; i < BPF_REG_SIZE; i++)
5180	scrub_spilled_slot(stype: &state->stack[spi].slot_type[i]);
5181
5182	/* when we zero initialize stack slots mark them as such */
5183	if ((reg && register_is_null(reg)) ||
5184	(!reg && is_bpf_st_mem(insn) && insn->imm == 0)) {
5185	/* STACK_ZERO case happened because register spill
5186	* wasn't properly aligned at the stack slot boundary,
5187	* so it's not a register spill anymore; force
5188	* originating register to be precise to make
5189	* STACK_ZERO correct for subsequent states
5190	*/
5191	err = mark_chain_precision(env, regno: value_regno);
5192	if (err)
5193	return err;
5194	type = STACK_ZERO;
5195	}
5196
5197	/* Mark slots affected by this stack write. */
5198	for (i = 0; i < size; i++)
5199	state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] = type;
5200	insn_flags = 0; /* not a register spill */
5201	}
5202
5203	if (insn_flags)
5204	return push_jmp_history(env, cur: env->cur_state, insn_flags, linked_regs: 0);
5205	return 0;
5206	}
5207
5208	/* Write the stack: 'stack[ptr_regno + off] = value_regno'. 'ptr_regno' is
5209	* known to contain a variable offset.
5210	* This function checks whether the write is permitted and conservatively
5211	* tracks the effects of the write, considering that each stack slot in the
5212	* dynamic range is potentially written to.
5213	*
5214	* 'off' includes 'regno->off'.
5215	* 'value_regno' can be -1, meaning that an unknown value is being written to
5216	* the stack.
5217	*
5218	* Spilled pointers in range are not marked as written because we don't know
5219	* what's going to be actually written. This means that read propagation for
5220	* future reads cannot be terminated by this write.
5221	*
5222	* For privileged programs, uninitialized stack slots are considered
5223	* initialized by this write (even though we don't know exactly what offsets
5224	* are going to be written to). The idea is that we don't want the verifier to
5225	* reject future reads that access slots written to through variable offsets.
5226	*/
5227	static int check_stack_write_var_off(struct bpf_verifier_env *env,
5228	/* func where register points to */
5229	struct bpf_func_state *state,
5230	int ptr_regno, int off, int size,
5231	int value_regno, int insn_idx)
5232	{
5233	struct bpf_func_state *cur; /* state of the current function */
5234	int min_off, max_off;
5235	int i, err;
5236	struct bpf_reg_state *ptr_reg = NULL, *value_reg = NULL;
5237	struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
5238	bool writing_zero = false;
5239	/* set if the fact that we're writing a zero is used to let any
5240	* stack slots remain STACK_ZERO
5241	*/
5242	bool zero_used = false;
5243
5244	cur = env->cur_state->frame[env->cur_state->curframe];
5245	ptr_reg = &cur->regs[ptr_regno];
5246	min_off = ptr_reg->smin_value + off;
5247	max_off = ptr_reg->smax_value + off + size;
5248	if (value_regno >= 0)
5249	value_reg = &cur->regs[value_regno];
5250	if ((value_reg && register_is_null(reg: value_reg)) ||
5251	(!value_reg && is_bpf_st_mem(insn) && insn->imm == 0))
5252	writing_zero = true;
5253
5254	for (i = min_off; i < max_off; i++) {
5255	int spi;
5256
5257	spi = __get_spi(off: i);
5258	err = destroy_if_dynptr_stack_slot(env, state, spi);
5259	if (err)
5260	return err;
5261	}
5262
5263	check_fastcall_stack_contract(env, state, insn_idx, off: min_off);
5264	/* Variable offset writes destroy any spilled pointers in range. */
5265	for (i = min_off; i < max_off; i++) {
5266	u8 new_type, *stype;
5267	int slot, spi;
5268
5269	slot = -i - 1;
5270	spi = slot / BPF_REG_SIZE;
5271	stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
5272	mark_stack_slot_scratched(env, spi);
5273
5274	if (!env->allow_ptr_leaks && *stype != STACK_MISC && *stype != STACK_ZERO) {
5275	/* Reject the write if range we may write to has not
5276	* been initialized beforehand. If we didn't reject
5277	* here, the ptr status would be erased below (even
5278	* though not all slots are actually overwritten),
5279	* possibly opening the door to leaks.
5280	*
5281	* We do however catch STACK_INVALID case below, and
5282	* only allow reading possibly uninitialized memory
5283	* later for CAP_PERFMON, as the write may not happen to
5284	* that slot.
5285	*/
5286	verbose(private_data: env, fmt: "spilled ptr in range of var-offset stack write; insn %d, ptr off: %d",
5287	insn_idx, i);
5288	return -EINVAL;
5289	}
5290
5291	/* If writing_zero and the spi slot contains a spill of value 0,
5292	* maintain the spill type.
5293	*/
5294	if (writing_zero && *stype == STACK_SPILL &&
5295	is_spilled_scalar_reg(stack: &state->stack[spi])) {
5296	struct bpf_reg_state *spill_reg = &state->stack[spi].spilled_ptr;
5297
5298	if (tnum_is_const(a: spill_reg->var_off) && spill_reg->var_off.value == 0) {
5299	zero_used = true;
5300	continue;
5301	}
5302	}
5303
5304	/* Erase all other spilled pointers. */
5305	state->stack[spi].spilled_ptr.type = NOT_INIT;
5306
5307	/* Update the slot type. */
5308	new_type = STACK_MISC;
5309	if (writing_zero && *stype == STACK_ZERO) {
5310	new_type = STACK_ZERO;
5311	zero_used = true;
5312	}
5313	/* If the slot is STACK_INVALID, we check whether it's OK to
5314	* pretend that it will be initialized by this write. The slot
5315	* might not actually be written to, and so if we mark it as
5316	* initialized future reads might leak uninitialized memory.
5317	* For privileged programs, we will accept such reads to slots
5318	* that may or may not be written because, if we're reject
5319	* them, the error would be too confusing.
5320	*/
5321	if (*stype == STACK_INVALID && !env->allow_uninit_stack) {
5322	verbose(private_data: env, fmt: "uninit stack in range of var-offset write prohibited for !root; insn %d, off: %d",
5323	insn_idx, i);
5324	return -EINVAL;
5325	}
5326	*stype = new_type;
5327	}
5328	if (zero_used) {
5329	/* backtracking doesn't work for STACK_ZERO yet. */
5330	err = mark_chain_precision(env, regno: value_regno);
5331	if (err)
5332	return err;
5333	}
5334	return 0;
5335	}
5336
5337	/* When register 'dst_regno' is assigned some values from stack[min_off,
5338	* max_off), we set the register's type according to the types of the
5339	* respective stack slots. If all the stack values are known to be zeros, then
5340	* so is the destination reg. Otherwise, the register is considered to be
5341	* SCALAR. This function does not deal with register filling; the caller must
5342	* ensure that all spilled registers in the stack range have been marked as
5343	* read.
5344	*/
5345	static void mark_reg_stack_read(struct bpf_verifier_env *env,
5346	/* func where src register points to */
5347	struct bpf_func_state *ptr_state,
5348	int min_off, int max_off, int dst_regno)
5349	{
5350	struct bpf_verifier_state *vstate = env->cur_state;
5351	struct bpf_func_state *state = vstate->frame[vstate->curframe];
5352	int i, slot, spi;
5353	u8 *stype;
5354	int zeros = 0;
5355
5356	for (i = min_off; i < max_off; i++) {
5357	slot = -i - 1;
5358	spi = slot / BPF_REG_SIZE;
5359	mark_stack_slot_scratched(env, spi);
5360	stype = ptr_state->stack[spi].slot_type;
5361	if (stype[slot % BPF_REG_SIZE] != STACK_ZERO)
5362	break;
5363	zeros++;
5364	}
5365	if (zeros == max_off - min_off) {
5366	/* Any access_size read into register is zero extended,
5367	* so the whole register == const_zero.
5368	*/
5369	__mark_reg_const_zero(env, reg: &state->regs[dst_regno]);
5370	} else {
5371	/* have read misc data from the stack */
5372	mark_reg_unknown(env, regs: state->regs, regno: dst_regno);
5373	}
5374	}
5375
5376	/* Read the stack at 'off' and put the results into the register indicated by
5377	* 'dst_regno'. It handles reg filling if the addressed stack slot is a
5378	* spilled reg.
5379	*
5380	* 'dst_regno' can be -1, meaning that the read value is not going to a
5381	* register.
5382	*
5383	* The access is assumed to be within the current stack bounds.
5384	*/
5385	static int check_stack_read_fixed_off(struct bpf_verifier_env *env,
5386	/* func where src register points to */
5387	struct bpf_func_state *reg_state,
5388	int off, int size, int dst_regno)
5389	{
5390	struct bpf_verifier_state *vstate = env->cur_state;
5391	struct bpf_func_state *state = vstate->frame[vstate->curframe];
5392	int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
5393	struct bpf_reg_state *reg;
5394	u8 *stype, type;
5395	int insn_flags = insn_stack_access_flags(frameno: reg_state->frameno, spi);
5396	int err;
5397
5398	stype = reg_state->stack[spi].slot_type;
5399	reg = &reg_state->stack[spi].spilled_ptr;
5400
5401	mark_stack_slot_scratched(env, spi);
5402	check_fastcall_stack_contract(env, state, insn_idx: env->insn_idx, off);
5403	err = bpf_mark_stack_read(env, frameno: reg_state->frameno, insn_idx: env->insn_idx, BIT(spi));
5404	if (err)
5405	return err;
5406
5407	if (is_spilled_reg(stack: &reg_state->stack[spi])) {
5408	u8 spill_size = 1;
5409
5410	for (i = BPF_REG_SIZE - 1; i > 0 && stype[i - 1] == STACK_SPILL; i--)
5411	spill_size++;
5412
5413	if (size != BPF_REG_SIZE || spill_size != BPF_REG_SIZE) {
5414	if (reg->type != SCALAR_VALUE) {
5415	verbose_linfo(env, insn_off: env->insn_idx, prefix_fmt: "; ");
5416	verbose(private_data: env, fmt: "invalid size of register fill\n");
5417	return -EACCES;
5418	}
5419
5420	if (dst_regno < 0)
5421	return 0;
5422
5423	if (size <= spill_size &&
5424	bpf_stack_narrow_access_ok(off, fill_size: size, spill_size)) {
5425	/* The earlier check_reg_arg() has decided the
5426	* subreg_def for this insn. Save it first.
5427	*/
5428	s32 subreg_def = state->regs[dst_regno].subreg_def;
5429
5430	copy_register_state(dst: &state->regs[dst_regno], src: reg);
5431	state->regs[dst_regno].subreg_def = subreg_def;
5432
5433	/* Break the relation on a narrowing fill.
5434	* coerce_reg_to_size will adjust the boundaries.
5435	*/
5436	if (get_reg_width(reg) > size * BITS_PER_BYTE)
5437	state->regs[dst_regno].id = 0;
5438	} else {
5439	int spill_cnt = 0, zero_cnt = 0;
5440
5441	for (i = 0; i < size; i++) {
5442	type = stype[(slot - i) % BPF_REG_SIZE];
5443	if (type == STACK_SPILL) {
5444	spill_cnt++;
5445	continue;
5446	}
5447	if (type == STACK_MISC)
5448	continue;
5449	if (type == STACK_ZERO) {
5450	zero_cnt++;
5451	continue;
5452	}
5453	if (type == STACK_INVALID && env->allow_uninit_stack)
5454	continue;
5455	verbose(private_data: env, fmt: "invalid read from stack off %d+%d size %d\n",
5456	off, i, size);
5457	return -EACCES;
5458	}
5459
5460	if (spill_cnt == size &&
5461	tnum_is_const(a: reg->var_off) && reg->var_off.value == 0) {
5462	__mark_reg_const_zero(env, reg: &state->regs[dst_regno]);
5463	/* this IS register fill, so keep insn_flags */
5464	} else if (zero_cnt == size) {
5465	/* similarly to mark_reg_stack_read(), preserve zeroes */
5466	__mark_reg_const_zero(env, reg: &state->regs[dst_regno]);
5467	insn_flags = 0; /* not restoring original register state */
5468	} else {
5469	mark_reg_unknown(env, regs: state->regs, regno: dst_regno);
5470	insn_flags = 0; /* not restoring original register state */
5471	}
5472	}
5473	} else if (dst_regno >= 0) {
5474	/* restore register state from stack */
5475	copy_register_state(dst: &state->regs[dst_regno], src: reg);
5476	/* mark reg as written since spilled pointer state likely
5477	* has its liveness marks cleared by is_state_visited()
5478	* which resets stack/reg liveness for state transitions
5479	*/
5480	} else if (__is_pointer_value(allow_ptr_leaks: env->allow_ptr_leaks, reg)) {
5481	/* If dst_regno==-1, the caller is asking us whether
5482	* it is acceptable to use this value as a SCALAR_VALUE
5483	* (e.g. for XADD).
5484	* We must not allow unprivileged callers to do that
5485	* with spilled pointers.
5486	*/
5487	verbose(private_data: env, fmt: "leaking pointer from stack off %d\n",
5488	off);
5489	return -EACCES;
5490	}
5491	} else {
5492	for (i = 0; i < size; i++) {
5493	type = stype[(slot - i) % BPF_REG_SIZE];
5494	if (type == STACK_MISC)
5495	continue;
5496	if (type == STACK_ZERO)
5497	continue;
5498	if (type == STACK_INVALID && env->allow_uninit_stack)
5499	continue;
5500	verbose(private_data: env, fmt: "invalid read from stack off %d+%d size %d\n",
5501	off, i, size);
5502	return -EACCES;
5503	}
5504	if (dst_regno >= 0)
5505	mark_reg_stack_read(env, ptr_state: reg_state, min_off: off, max_off: off + size, dst_regno);
5506	insn_flags = 0; /* we are not restoring spilled register */
5507	}
5508	if (insn_flags)
5509	return push_jmp_history(env, cur: env->cur_state, insn_flags, linked_regs: 0);
5510	return 0;
5511	}
5512
5513	enum bpf_access_src {
5514	ACCESS_DIRECT = 1, /* the access is performed by an instruction */
5515	ACCESS_HELPER = 2, /* the access is performed by a helper */
5516	};
5517
5518	static int check_stack_range_initialized(struct bpf_verifier_env *env,
5519	int regno, int off, int access_size,
5520	bool zero_size_allowed,
5521	enum bpf_access_type type,
5522	struct bpf_call_arg_meta *meta);
5523
5524	static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
5525	{
5526	return cur_regs(env) + regno;
5527	}
5528
5529	/* Read the stack at 'ptr_regno + off' and put the result into the register
5530	* 'dst_regno'.
5531	* 'off' includes the pointer register's fixed offset(i.e. 'ptr_regno.off'),
5532	* but not its variable offset.
5533	* 'size' is assumed to be <= reg size and the access is assumed to be aligned.
5534	*
5535	* As opposed to check_stack_read_fixed_off, this function doesn't deal with
5536	* filling registers (i.e. reads of spilled register cannot be detected when
5537	* the offset is not fixed). We conservatively mark 'dst_regno' as containing
5538	* SCALAR_VALUE. That's why we assert that the 'ptr_regno' has a variable
5539	* offset; for a fixed offset check_stack_read_fixed_off should be used
5540	* instead.
5541	*/
5542	static int check_stack_read_var_off(struct bpf_verifier_env *env,
5543	int ptr_regno, int off, int size, int dst_regno)
5544	{
5545	/* The state of the source register. */
5546	struct bpf_reg_state *reg = reg_state(env, regno: ptr_regno);
5547	struct bpf_func_state *ptr_state = func(env, reg);
5548	int err;
5549	int min_off, max_off;
5550
5551	/* Note that we pass a NULL meta, so raw access will not be permitted.
5552	*/
5553	err = check_stack_range_initialized(env, regno: ptr_regno, off, access_size: size,
5554	zero_size_allowed: false, type: BPF_READ, NULL);
5555	if (err)
5556	return err;
5557
5558	min_off = reg->smin_value + off;
5559	max_off = reg->smax_value + off;
5560	mark_reg_stack_read(env, ptr_state, min_off, max_off: max_off + size, dst_regno);
5561	check_fastcall_stack_contract(env, state: ptr_state, insn_idx: env->insn_idx, off: min_off);
5562	return 0;
5563	}
5564
5565	/* check_stack_read dispatches to check_stack_read_fixed_off or
5566	* check_stack_read_var_off.
5567	*
5568	* The caller must ensure that the offset falls within the allocated stack
5569	* bounds.
5570	*
5571	* 'dst_regno' is a register which will receive the value from the stack. It
5572	* can be -1, meaning that the read value is not going to a register.
5573	*/
5574	static int check_stack_read(struct bpf_verifier_env *env,
5575	int ptr_regno, int off, int size,
5576	int dst_regno)
5577	{
5578	struct bpf_reg_state *reg = reg_state(env, regno: ptr_regno);
5579	struct bpf_func_state *state = func(env, reg);
5580	int err;
5581	/* Some accesses are only permitted with a static offset. */
5582	bool var_off = !tnum_is_const(a: reg->var_off);
5583
5584	/* The offset is required to be static when reads don't go to a
5585	* register, in order to not leak pointers (see
5586	* check_stack_read_fixed_off).
5587	*/
5588	if (dst_regno < 0 && var_off) {
5589	char tn_buf[48];
5590
5591	tnum_strn(str: tn_buf, size: sizeof(tn_buf), a: reg->var_off);
5592	verbose(private_data: env, fmt: "variable offset stack pointer cannot be passed into helper function; var_off=%s off=%d size=%d\n",
5593	tn_buf, off, size);
5594	return -EACCES;
5595	}
5596	/* Variable offset is prohibited for unprivileged mode for simplicity
5597	* since it requires corresponding support in Spectre masking for stack
5598	* ALU. See also retrieve_ptr_limit(). The check in
5599	* check_stack_access_for_ptr_arithmetic() called by
5600	* adjust_ptr_min_max_vals() prevents users from creating stack pointers
5601	* with variable offsets, therefore no check is required here. Further,
5602	* just checking it here would be insufficient as speculative stack
5603	* writes could still lead to unsafe speculative behaviour.
5604	*/
5605	if (!var_off) {
5606	off += reg->var_off.value;
5607	err = check_stack_read_fixed_off(env, reg_state: state, off, size,
5608	dst_regno);
5609	} else {
5610	/* Variable offset stack reads need more conservative handling
5611	* than fixed offset ones. Note that dst_regno >= 0 on this
5612	* branch.
5613	*/
5614	err = check_stack_read_var_off(env, ptr_regno, off, size,
5615	dst_regno);
5616	}
5617	return err;
5618	}
5619
5620
5621	/* check_stack_write dispatches to check_stack_write_fixed_off or
5622	* check_stack_write_var_off.
5623	*
5624	* 'ptr_regno' is the register used as a pointer into the stack.
5625	* 'off' includes 'ptr_regno->off', but not its variable offset (if any).
5626	* 'value_regno' is the register whose value we're writing to the stack. It can
5627	* be -1, meaning that we're not writing from a register.
5628	*
5629	* The caller must ensure that the offset falls within the maximum stack size.
5630	*/
5631	static int check_stack_write(struct bpf_verifier_env *env,
5632	int ptr_regno, int off, int size,
5633	int value_regno, int insn_idx)
5634	{
5635	struct bpf_reg_state *reg = reg_state(env, regno: ptr_regno);
5636	struct bpf_func_state *state = func(env, reg);
5637	int err;
5638
5639	if (tnum_is_const(a: reg->var_off)) {
5640	off += reg->var_off.value;
5641	err = check_stack_write_fixed_off(env, state, off, size,
5642	value_regno, insn_idx);
5643	} else {
5644	/* Variable offset stack reads need more conservative handling
5645	* than fixed offset ones.
5646	*/
5647	err = check_stack_write_var_off(env, state,
5648	ptr_regno, off, size,
5649	value_regno, insn_idx);
5650	}
5651	return err;
5652	}
5653
5654	static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
5655	int off, int size, enum bpf_access_type type)
5656	{
5657	struct bpf_reg_state *regs = cur_regs(env);
5658	struct bpf_map *map = regs[regno].map_ptr;
5659	u32 cap = bpf_map_flags_to_cap(map);
5660
5661	if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
5662	verbose(private_data: env, fmt: "write into map forbidden, value_size=%d off=%d size=%d\n",
5663	map->value_size, off, size);
5664	return -EACCES;
5665	}
5666
5667	if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
5668	verbose(private_data: env, fmt: "read from map forbidden, value_size=%d off=%d size=%d\n",
5669	map->value_size, off, size);
5670	return -EACCES;
5671	}
5672
5673	return 0;
5674	}
5675
5676	/* check read/write into memory region (e.g., map value, ringbuf sample, etc) */
5677	static int __check_mem_access(struct bpf_verifier_env *env, int regno,
5678	int off, int size, u32 mem_size,
5679	bool zero_size_allowed)
5680	{
5681	bool size_ok = size > 0 || (size == 0 && zero_size_allowed);
5682	struct bpf_reg_state *reg;
5683
5684	if (off >= 0 && size_ok && (u64)off + size <= mem_size)
5685	return 0;
5686
5687	reg = &cur_regs(env)[regno];
5688	switch (reg->type) {
5689	case PTR_TO_MAP_KEY:
5690	verbose(private_data: env, fmt: "invalid access to map key, key_size=%d off=%d size=%d\n",
5691	mem_size, off, size);
5692	break;
5693	case PTR_TO_MAP_VALUE:
5694	verbose(private_data: env, fmt: "invalid access to map value, value_size=%d off=%d size=%d\n",
5695	mem_size, off, size);
5696	break;
5697	case PTR_TO_PACKET:
5698	case PTR_TO_PACKET_META:
5699	case PTR_TO_PACKET_END:
5700	verbose(private_data: env, fmt: "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
5701	off, size, regno, reg->id, off, mem_size);
5702	break;
5703	case PTR_TO_MEM:
5704	default:
5705	verbose(private_data: env, fmt: "invalid access to memory, mem_size=%u off=%d size=%d\n",
5706	mem_size, off, size);
5707	}
5708
5709	return -EACCES;
5710	}
5711
5712	/* check read/write into a memory region with possible variable offset */
5713	static int check_mem_region_access(struct bpf_verifier_env *env, u32 regno,
5714	int off, int size, u32 mem_size,
5715	bool zero_size_allowed)
5716	{
5717	struct bpf_verifier_state *vstate = env->cur_state;
5718	struct bpf_func_state *state = vstate->frame[vstate->curframe];
5719	struct bpf_reg_state *reg = &state->regs[regno];
5720	int err;
5721
5722	/* We may have adjusted the register pointing to memory region, so we
5723	* need to try adding each of min_value and max_value to off
5724	* to make sure our theoretical access will be safe.
5725	*
5726	* The minimum value is only important with signed
5727	* comparisons where we can't assume the floor of a
5728	* value is 0. If we are using signed variables for our
5729	* index'es we need to make sure that whatever we use
5730	* will have a set floor within our range.
5731	*/
5732	if (reg->smin_value < 0 &&
5733	(reg->smin_value == S64_MIN ||
5734	(off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
5735	reg->smin_value + off < 0)) {
5736	verbose(private_data: env, fmt: "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
5737	regno);
5738	return -EACCES;
5739	}
5740	err = __check_mem_access(env, regno, off: reg->smin_value + off, size,
5741	mem_size, zero_size_allowed);
5742	if (err) {
5743	verbose(private_data: env, fmt: "R%d min value is outside of the allowed memory range\n",
5744	regno);
5745	return err;
5746	}
5747
5748	/* If we haven't set a max value then we need to bail since we can't be
5749	* sure we won't do bad things.
5750	* If reg->umax_value + off could overflow, treat that as unbounded too.
5751	*/
5752	if (reg->umax_value >= BPF_MAX_VAR_OFF) {
5753	verbose(private_data: env, fmt: "R%d unbounded memory access, make sure to bounds check any such access\n",
5754	regno);
5755	return -EACCES;
5756	}
5757	err = __check_mem_access(env, regno, off: reg->umax_value + off, size,
5758	mem_size, zero_size_allowed);
5759	if (err) {
5760	verbose(private_data: env, fmt: "R%d max value is outside of the allowed memory range\n",
5761	regno);
5762	return err;
5763	}
5764
5765	return 0;
5766	}
5767
5768	static int __check_ptr_off_reg(struct bpf_verifier_env *env,
5769	const struct bpf_reg_state *reg, int regno,
5770	bool fixed_off_ok)
5771	{
5772	/* Access to this pointer-typed register or passing it to a helper
5773	* is only allowed in its original, unmodified form.
5774	*/
5775
5776	if (reg->off < 0) {
5777	verbose(private_data: env, fmt: "negative offset %s ptr R%d off=%d disallowed\n",
5778	reg_type_str(env, type: reg->type), regno, reg->off);
5779	return -EACCES;
5780	}
5781
5782	if (!fixed_off_ok && reg->off) {
5783	verbose(private_data: env, fmt: "dereference of modified %s ptr R%d off=%d disallowed\n",
5784	reg_type_str(env, type: reg->type), regno, reg->off);
5785	return -EACCES;
5786	}
5787
5788	if (!tnum_is_const(a: reg->var_off) || reg->var_off.value) {
5789	char tn_buf[48];
5790
5791	tnum_strn(str: tn_buf, size: sizeof(tn_buf), a: reg->var_off);
5792	verbose(private_data: env, fmt: "variable %s access var_off=%s disallowed\n",
5793	reg_type_str(env, type: reg->type), tn_buf);
5794	return -EACCES;
5795	}
5796
5797	return 0;
5798	}
5799
5800	static int check_ptr_off_reg(struct bpf_verifier_env *env,
5801	const struct bpf_reg_state *reg, int regno)
5802	{
5803	return __check_ptr_off_reg(env, reg, regno, fixed_off_ok: false);
5804	}
5805
5806	static int map_kptr_match_type(struct bpf_verifier_env *env,
5807	struct btf_field *kptr_field,
5808	struct bpf_reg_state *reg, u32 regno)
5809	{
5810	const char *targ_name = btf_type_name(btf: kptr_field->kptr.btf, id: kptr_field->kptr.btf_id);
5811	int perm_flags;
5812	const char *reg_name = "";
5813
5814	if (btf_is_kernel(btf: reg->btf)) {
5815	perm_flags = PTR_MAYBE_NULL | PTR_TRUSTED | MEM_RCU;
5816
5817	/* Only unreferenced case accepts untrusted pointers */
5818	if (kptr_field->type == BPF_KPTR_UNREF)
5819	perm_flags |= PTR_UNTRUSTED;
5820	} else {
5821	perm_flags = PTR_MAYBE_NULL | MEM_ALLOC;
5822	if (kptr_field->type == BPF_KPTR_PERCPU)
5823	perm_flags |= MEM_PERCPU;
5824	}
5825
5826	if (base_type(type: reg->type) != PTR_TO_BTF_ID || (type_flag(type: reg->type) & ~perm_flags))
5827	goto bad_type;
5828
5829	/* We need to verify reg->type and reg->btf, before accessing reg->btf */
5830	reg_name = btf_type_name(btf: reg->btf, id: reg->btf_id);
5831
5832	/* For ref_ptr case, release function check should ensure we get one
5833	* referenced PTR_TO_BTF_ID, and that its fixed offset is 0. For the
5834	* normal store of unreferenced kptr, we must ensure var_off is zero.
5835	* Since ref_ptr cannot be accessed directly by BPF insns, checks for
5836	* reg->off and reg->ref_obj_id are not needed here.
5837	*/
5838	if (__check_ptr_off_reg(env, reg, regno, fixed_off_ok: true))
5839	return -EACCES;
5840
5841	/* A full type match is needed, as BTF can be vmlinux, module or prog BTF, and
5842	* we also need to take into account the reg->off.
5843	*
5844	* We want to support cases like:
5845	*
5846	* struct foo {
5847	* struct bar br;
5848	* struct baz bz;
5849	* };
5850	*
5851	* struct foo *v;
5852	* v = func(); // PTR_TO_BTF_ID
5853	* val->foo = v; // reg->off is zero, btf and btf_id match type
5854	* val->bar = &v->br; // reg->off is still zero, but we need to retry with
5855	* // first member type of struct after comparison fails
5856	* val->baz = &v->bz; // reg->off is non-zero, so struct needs to be walked
5857	* // to match type
5858	*
5859	* In the kptr_ref case, check_func_arg_reg_off already ensures reg->off
5860	* is zero. We must also ensure that btf_struct_ids_match does not walk
5861	* the struct to match type against first member of struct, i.e. reject
5862	* second case from above. Hence, when type is BPF_KPTR_REF, we set
5863	* strict mode to true for type match.
5864	*/
5865	if (!btf_struct_ids_match(log: &env->log, btf: reg->btf, id: reg->btf_id, off: reg->off,
5866	need_btf: kptr_field->kptr.btf, need_type_id: kptr_field->kptr.btf_id,
5867	strict: kptr_field->type != BPF_KPTR_UNREF))
5868	goto bad_type;
5869	return 0;
5870	bad_type:
5871	verbose(private_data: env, fmt: "invalid kptr access, R%d type=%s%s ", regno,
5872	reg_type_str(env, type: reg->type), reg_name);
5873	verbose(private_data: env, fmt: "expected=%s%s", reg_type_str(env, type: PTR_TO_BTF_ID), targ_name);
5874	if (kptr_field->type == BPF_KPTR_UNREF)
5875	verbose(private_data: env, fmt: " or %s%s\n", reg_type_str(env, type: PTR_TO_BTF_ID | PTR_UNTRUSTED),
5876	targ_name);
5877	else
5878	verbose(private_data: env, fmt: "\n");
5879	return -EINVAL;
5880	}
5881
5882	static bool in_sleepable(struct bpf_verifier_env *env)
5883	{
5884	return env->cur_state->in_sleepable;
5885	}
5886
5887	/* The non-sleepable programs and sleepable programs with explicit bpf_rcu_read_lock()
5888	* can dereference RCU protected pointers and result is PTR_TRUSTED.
5889	*/
5890	static bool in_rcu_cs(struct bpf_verifier_env *env)
5891	{
5892	return env->cur_state->active_rcu_locks ||
5893	env->cur_state->active_locks ||
5894	!in_sleepable(env);
5895	}
5896
5897	/* Once GCC supports btf_type_tag the following mechanism will be replaced with tag check */
5898	BTF_SET_START(rcu_protected_types)
5899	#ifdef CONFIG_NET
5900	BTF_ID(struct, prog_test_ref_kfunc)
5901	#endif
5902	#ifdef CONFIG_CGROUPS
5903	BTF_ID(struct, cgroup)
5904	#endif
5905	#ifdef CONFIG_BPF_JIT
5906	BTF_ID(struct, bpf_cpumask)
5907	#endif
5908	BTF_ID(struct, task_struct)
5909	#ifdef CONFIG_CRYPTO
5910	BTF_ID(struct, bpf_crypto_ctx)
5911	#endif
5912	BTF_SET_END(rcu_protected_types)
5913
5914	static bool rcu_protected_object(const struct btf *btf, u32 btf_id)
5915	{
5916	if (!btf_is_kernel(btf))
5917	return true;
5918	return btf_id_set_contains(set: &rcu_protected_types, id: btf_id);
5919	}
5920
5921	static struct btf_record *kptr_pointee_btf_record(struct btf_field *kptr_field)
5922	{
5923	struct btf_struct_meta *meta;
5924
5925	if (btf_is_kernel(btf: kptr_field->kptr.btf))
5926	return NULL;
5927
5928	meta = btf_find_struct_meta(btf: kptr_field->kptr.btf,
5929	btf_id: kptr_field->kptr.btf_id);
5930
5931	return meta ? meta->record : NULL;
5932	}
5933
5934	static bool rcu_safe_kptr(const struct btf_field *field)
5935	{
5936	const struct btf_field_kptr *kptr = &field->kptr;
5937
5938	return field->type == BPF_KPTR_PERCPU ||
5939	(field->type == BPF_KPTR_REF && rcu_protected_object(btf: kptr->btf, btf_id: kptr->btf_id));
5940	}
5941
5942	static u32 btf_ld_kptr_type(struct bpf_verifier_env *env, struct btf_field *kptr_field)
5943	{
5944	struct btf_record *rec;
5945	u32 ret;
5946
5947	ret = PTR_MAYBE_NULL;
5948	if (rcu_safe_kptr(field: kptr_field) && in_rcu_cs(env)) {
5949	ret |= MEM_RCU;
5950	if (kptr_field->type == BPF_KPTR_PERCPU)
5951	ret |= MEM_PERCPU;
5952	else if (!btf_is_kernel(btf: kptr_field->kptr.btf))
5953	ret |= MEM_ALLOC;
5954
5955	rec = kptr_pointee_btf_record(kptr_field);
5956	if (rec && btf_record_has_field(rec, type: BPF_GRAPH_NODE))
5957	ret |= NON_OWN_REF;
5958	} else {
5959	ret |= PTR_UNTRUSTED;
5960	}
5961
5962	return ret;
5963	}
5964
5965	static int mark_uptr_ld_reg(struct bpf_verifier_env *env, u32 regno,
5966	struct btf_field *field)
5967	{
5968	struct bpf_reg_state *reg;
5969	const struct btf_type *t;
5970
5971	t = btf_type_by_id(btf: field->kptr.btf, type_id: field->kptr.btf_id);
5972	mark_reg_known_zero(env, regs: cur_regs(env), regno);
5973	reg = reg_state(env, regno);
5974	reg->type = PTR_TO_MEM | PTR_MAYBE_NULL;
5975	reg->mem_size = t->size;
5976	reg->id = ++env->id_gen;
5977
5978	return 0;
5979	}
5980
5981	static int check_map_kptr_access(struct bpf_verifier_env *env, u32 regno,
5982	int value_regno, int insn_idx,
5983	struct btf_field *kptr_field)
5984	{
5985	struct bpf_insn *insn = &env->prog->insnsi[insn_idx];
5986	int class = BPF_CLASS(insn->code);
5987	struct bpf_reg_state *val_reg;
5988	int ret;
5989
5990	/* Things we already checked for in check_map_access and caller:
5991	* - Reject cases where variable offset may touch kptr
5992	* - size of access (must be BPF_DW)
5993	* - tnum_is_const(reg->var_off)
5994	* - kptr_field->offset == off + reg->var_off.value
5995	*/
5996	/* Only BPF_[LDX,STX,ST] | BPF_MEM | BPF_DW is supported */
5997	if (BPF_MODE(insn->code) != BPF_MEM) {
5998	verbose(private_data: env, fmt: "kptr in map can only be accessed using BPF_MEM instruction mode\n");
5999	return -EACCES;
6000	}
6001
6002	/* We only allow loading referenced kptr, since it will be marked as
6003	* untrusted, similar to unreferenced kptr.
6004	*/
6005	if (class != BPF_LDX &&
6006	(kptr_field->type == BPF_KPTR_REF || kptr_field->type == BPF_KPTR_PERCPU)) {
6007	verbose(private_data: env, fmt: "store to referenced kptr disallowed\n");
6008	return -EACCES;
6009	}
6010	if (class != BPF_LDX && kptr_field->type == BPF_UPTR) {
6011	verbose(private_data: env, fmt: "store to uptr disallowed\n");
6012	return -EACCES;
6013	}
6014
6015	if (class == BPF_LDX) {
6016	if (kptr_field->type == BPF_UPTR)
6017	return mark_uptr_ld_reg(env, regno: value_regno, field: kptr_field);
6018
6019	/* We can simply mark the value_regno receiving the pointer
6020	* value from map as PTR_TO_BTF_ID, with the correct type.
6021	*/
6022	ret = mark_btf_ld_reg(env, regs: cur_regs(env), regno: value_regno, reg_type: PTR_TO_BTF_ID,
6023	btf: kptr_field->kptr.btf, btf_id: kptr_field->kptr.btf_id,
6024	flag: btf_ld_kptr_type(env, kptr_field));
6025	if (ret < 0)
6026	return ret;
6027	} else if (class == BPF_STX) {
6028	val_reg = reg_state(env, regno: value_regno);
6029	if (!register_is_null(reg: val_reg) &&
6030	map_kptr_match_type(env, kptr_field, reg: val_reg, regno: value_regno))
6031	return -EACCES;
6032	} else if (class == BPF_ST) {
6033	if (insn->imm) {
6034	verbose(private_data: env, fmt: "BPF_ST imm must be 0 when storing to kptr at off=%u\n",
6035	kptr_field->offset);
6036	return -EACCES;
6037	}
6038	} else {
6039	verbose(private_data: env, fmt: "kptr in map can only be accessed using BPF_LDX/BPF_STX/BPF_ST\n");
6040	return -EACCES;
6041	}
6042	return 0;
6043	}
6044
6045	/*
6046	* Return the size of the memory region accessible from a pointer to map value.
6047	* For INSN_ARRAY maps whole bpf_insn_array->ips array is accessible.
6048	*/
6049	static u32 map_mem_size(const struct bpf_map *map)
6050	{
6051	if (map->map_type == BPF_MAP_TYPE_INSN_ARRAY)
6052	return map->max_entries * sizeof(long);
6053
6054	return map->value_size;
6055	}
6056
6057	/* check read/write into a map element with possible variable offset */
6058	static int check_map_access(struct bpf_verifier_env *env, u32 regno,
6059	int off, int size, bool zero_size_allowed,
6060	enum bpf_access_src src)
6061	{
6062	struct bpf_verifier_state *vstate = env->cur_state;
6063	struct bpf_func_state *state = vstate->frame[vstate->curframe];
6064	struct bpf_reg_state *reg = &state->regs[regno];
6065	struct bpf_map *map = reg->map_ptr;
6066	u32 mem_size = map_mem_size(map);
6067	struct btf_record *rec;
6068	int err, i;
6069
6070	err = check_mem_region_access(env, regno, off, size, mem_size, zero_size_allowed);
6071	if (err)
6072	return err;
6073
6074	if (IS_ERR_OR_NULL(ptr: map->record))
6075	return 0;
6076	rec = map->record;
6077	for (i = 0; i < rec->cnt; i++) {
6078	struct btf_field *field = &rec->fields[i];
6079	u32 p = field->offset;
6080
6081	/* If any part of a field can be touched by load/store, reject
6082	* this program. To check that [x1, x2) overlaps with [y1, y2),
6083	* it is sufficient to check x1 < y2 && y1 < x2.
6084	*/
6085	if (reg->smin_value + off < p + field->size &&
6086	p < reg->umax_value + off + size) {
6087	switch (field->type) {
6088	case BPF_KPTR_UNREF:
6089	case BPF_KPTR_REF:
6090	case BPF_KPTR_PERCPU:
6091	case BPF_UPTR:
6092	if (src != ACCESS_DIRECT) {
6093	verbose(private_data: env, fmt: "%s cannot be accessed indirectly by helper\n",
6094	btf_field_type_name(type: field->type));
6095	return -EACCES;
6096	}
6097	if (!tnum_is_const(a: reg->var_off)) {
6098	verbose(private_data: env, fmt: "%s access cannot have variable offset\n",
6099	btf_field_type_name(type: field->type));
6100	return -EACCES;
6101	}
6102	if (p != off + reg->var_off.value) {
6103	verbose(private_data: env, fmt: "%s access misaligned expected=%u off=%llu\n",
6104	btf_field_type_name(type: field->type),
6105	p, off + reg->var_off.value);
6106	return -EACCES;
6107	}
6108	if (size != bpf_size_to_bytes(BPF_DW)) {
6109	verbose(private_data: env, fmt: "%s access size must be BPF_DW\n",
6110	btf_field_type_name(type: field->type));
6111	return -EACCES;
6112	}
6113	break;
6114	default:
6115	verbose(private_data: env, fmt: "%s cannot be accessed directly by load/store\n",
6116	btf_field_type_name(type: field->type));
6117	return -EACCES;
6118	}
6119	}
6120	}
6121	return 0;
6122	}
6123
6124	#define MAX_PACKET_OFF 0xffff
6125
6126	static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
6127	const struct bpf_call_arg_meta *meta,
6128	enum bpf_access_type t)
6129	{
6130	enum bpf_prog_type prog_type = resolve_prog_type(prog: env->prog);
6131
6132	switch (prog_type) {
6133	/* Program types only with direct read access go here! */
6134	case BPF_PROG_TYPE_LWT_IN:
6135	case BPF_PROG_TYPE_LWT_OUT:
6136	case BPF_PROG_TYPE_LWT_SEG6LOCAL:
6137	case BPF_PROG_TYPE_SK_REUSEPORT:
6138	case BPF_PROG_TYPE_FLOW_DISSECTOR:
6139	case BPF_PROG_TYPE_CGROUP_SKB:
6140	if (t == BPF_WRITE)
6141	return false;
6142	fallthrough;
6143
6144	/* Program types with direct read + write access go here! */
6145	case BPF_PROG_TYPE_SCHED_CLS:
6146	case BPF_PROG_TYPE_SCHED_ACT:
6147	case BPF_PROG_TYPE_XDP:
6148	case BPF_PROG_TYPE_LWT_XMIT:
6149	case BPF_PROG_TYPE_SK_SKB:
6150	case BPF_PROG_TYPE_SK_MSG:
6151	if (meta)
6152	return meta->pkt_access;
6153
6154	env->seen_direct_write = true;
6155	return true;
6156
6157	case BPF_PROG_TYPE_CGROUP_SOCKOPT:
6158	if (t == BPF_WRITE)
6159	env->seen_direct_write = true;
6160
6161	return true;
6162
6163	default:
6164	return false;
6165	}
6166	}
6167
6168	static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
6169	int size, bool zero_size_allowed)
6170	{
6171	struct bpf_reg_state *regs = cur_regs(env);
6172	struct bpf_reg_state *reg = &regs[regno];
6173	int err;
6174
6175	/* We may have added a variable offset to the packet pointer; but any
6176	* reg->range we have comes after that. We are only checking the fixed
6177	* offset.
6178	*/
6179
6180	/* We don't allow negative numbers, because we aren't tracking enough
6181	* detail to prove they're safe.
6182	*/
6183	if (reg->smin_value < 0) {
6184	verbose(private_data: env, fmt: "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
6185	regno);
6186	return -EACCES;
6187	}
6188
6189	err = reg->range < 0 ? -EINVAL :
6190	__check_mem_access(env, regno, off, size, mem_size: reg->range,
6191	zero_size_allowed);
6192	if (err) {
6193	verbose(private_data: env, fmt: "R%d offset is outside of the packet\n", regno);
6194	return err;
6195	}
6196
6197	/* __check_mem_access has made sure "off + size - 1" is within u16.
6198	* reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
6199	* otherwise find_good_pkt_pointers would have refused to set range info
6200	* that __check_mem_access would have rejected this pkt access.
6201	* Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
6202	*/
6203	env->prog->aux->max_pkt_offset =
6204	max_t(u32, env->prog->aux->max_pkt_offset,
6205	off + reg->umax_value + size - 1);
6206
6207	return err;
6208	}
6209
6210	/* check access to 'struct bpf_context' fields. Supports fixed offsets only */
6211	static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
6212	enum bpf_access_type t, struct bpf_insn_access_aux *info)
6213	{
6214	if (env->ops->is_valid_access &&
6215	env->ops->is_valid_access(off, size, t, env->prog, info)) {
6216	/* A non zero info.ctx_field_size indicates that this field is a
6217	* candidate for later verifier transformation to load the whole
6218	* field and then apply a mask when accessed with a narrower
6219	* access than actual ctx access size. A zero info.ctx_field_size
6220	* will only allow for whole field access and rejects any other
6221	* type of narrower access.
6222	*/
6223	if (base_type(type: info->reg_type) == PTR_TO_BTF_ID) {
6224	if (info->ref_obj_id &&
6225	!find_reference_state(state: env->cur_state, ptr_id: info->ref_obj_id)) {
6226	verbose(private_data: env, fmt: "invalid bpf_context access off=%d. Reference may already be released\n",
6227	off);
6228	return -EACCES;
6229	}
6230	} else {
6231	env->insn_aux_data[insn_idx].ctx_field_size = info->ctx_field_size;
6232	}
6233	/* remember the offset of last byte accessed in ctx */
6234	if (env->prog->aux->max_ctx_offset < off + size)
6235	env->prog->aux->max_ctx_offset = off + size;
6236	return 0;
6237	}
6238
6239	verbose(private_data: env, fmt: "invalid bpf_context access off=%d size=%d\n", off, size);
6240	return -EACCES;
6241	}
6242
6243	static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
6244	int size)
6245	{
6246	if (size < 0 || off < 0 ||
6247	(u64)off + size > sizeof(struct bpf_flow_keys)) {
6248	verbose(private_data: env, fmt: "invalid access to flow keys off=%d size=%d\n",
6249	off, size);
6250	return -EACCES;
6251	}
6252	return 0;
6253	}
6254
6255	static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
6256	u32 regno, int off, int size,
6257	enum bpf_access_type t)
6258	{
6259	struct bpf_reg_state *regs = cur_regs(env);
6260	struct bpf_reg_state *reg = &regs[regno];
6261	struct bpf_insn_access_aux info = {};
6262	bool valid;
6263
6264	if (reg->smin_value < 0) {
6265	verbose(private_data: env, fmt: "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
6266	regno);
6267	return -EACCES;
6268	}
6269
6270	switch (reg->type) {
6271	case PTR_TO_SOCK_COMMON:
6272	valid = bpf_sock_common_is_valid_access(off, size, type: t, info: &info);
6273	break;
6274	case PTR_TO_SOCKET:
6275	valid = bpf_sock_is_valid_access(off, size, type: t, info: &info);
6276	break;
6277	case PTR_TO_TCP_SOCK:
6278	valid = bpf_tcp_sock_is_valid_access(off, size, type: t, info: &info);
6279	break;
6280	case PTR_TO_XDP_SOCK:
6281	valid = bpf_xdp_sock_is_valid_access(off, size, type: t, info: &info);
6282	break;
6283	default:
6284	valid = false;
6285	}
6286
6287
6288	if (valid) {
6289	env->insn_aux_data[insn_idx].ctx_field_size =
6290	info.ctx_field_size;
6291	return 0;
6292	}
6293
6294	verbose(private_data: env, fmt: "R%d invalid %s access off=%d size=%d\n",
6295	regno, reg_type_str(env, type: reg->type), off, size);
6296
6297	return -EACCES;
6298	}
6299
6300	static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
6301	{
6302	return __is_pointer_value(allow_ptr_leaks: env->allow_ptr_leaks, reg: reg_state(env, regno));
6303	}
6304
6305	static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
6306	{
6307	const struct bpf_reg_state *reg = reg_state(env, regno);
6308
6309	return reg->type == PTR_TO_CTX;
6310	}
6311
6312	static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
6313	{
6314	const struct bpf_reg_state *reg = reg_state(env, regno);
6315
6316	return type_is_sk_pointer(type: reg->type);
6317	}
6318
6319	static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
6320	{
6321	const struct bpf_reg_state *reg = reg_state(env, regno);
6322
6323	return type_is_pkt_pointer(type: reg->type);
6324	}
6325
6326	static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
6327	{
6328	const struct bpf_reg_state *reg = reg_state(env, regno);
6329
6330	/* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
6331	return reg->type == PTR_TO_FLOW_KEYS;
6332	}
6333
6334	static bool is_arena_reg(struct bpf_verifier_env *env, int regno)
6335	{
6336	const struct bpf_reg_state *reg = reg_state(env, regno);
6337
6338	return reg->type == PTR_TO_ARENA;
6339	}
6340
6341	/* Return false if @regno contains a pointer whose type isn't supported for
6342	* atomic instruction @insn.
6343	*/
6344	static bool atomic_ptr_type_ok(struct bpf_verifier_env *env, int regno,
6345	struct bpf_insn *insn)
6346	{
6347	if (is_ctx_reg(env, regno))
6348	return false;
6349	if (is_pkt_reg(env, regno))
6350	return false;
6351	if (is_flow_key_reg(env, regno))
6352	return false;
6353	if (is_sk_reg(env, regno))
6354	return false;
6355	if (is_arena_reg(env, regno))
6356	return bpf_jit_supports_insn(insn, in_arena: true);
6357
6358	return true;
6359	}
6360
6361	static u32 *reg2btf_ids[__BPF_REG_TYPE_MAX] = {
6362	#ifdef CONFIG_NET
6363	[PTR_TO_SOCKET] = &btf_sock_ids[BTF_SOCK_TYPE_SOCK],
6364	[PTR_TO_SOCK_COMMON] = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
6365	[PTR_TO_TCP_SOCK] = &btf_sock_ids[BTF_SOCK_TYPE_TCP],
6366	#endif
6367	[CONST_PTR_TO_MAP] = btf_bpf_map_id,
6368	};
6369
6370	static bool is_trusted_reg(const struct bpf_reg_state *reg)
6371	{
6372	/* A referenced register is always trusted. */
6373	if (reg->ref_obj_id)
6374	return true;
6375
6376	/* Types listed in the reg2btf_ids are always trusted */
6377	if (reg2btf_ids[base_type(type: reg->type)] &&
6378	!bpf_type_has_unsafe_modifiers(type: reg->type))
6379	return true;
6380
6381	/* If a register is not referenced, it is trusted if it has the
6382	* MEM_ALLOC or PTR_TRUSTED type modifiers, and no others. Some of the
6383	* other type modifiers may be safe, but we elect to take an opt-in
6384	* approach here as some (e.g. PTR_UNTRUSTED and PTR_MAYBE_NULL) are
6385	* not.
6386	*
6387	* Eventually, we should make PTR_TRUSTED the single source of truth
6388	* for whether a register is trusted.
6389	*/
6390	return type_flag(type: reg->type) & BPF_REG_TRUSTED_MODIFIERS &&
6391	!bpf_type_has_unsafe_modifiers(type: reg->type);
6392	}
6393
6394	static bool is_rcu_reg(const struct bpf_reg_state *reg)
6395	{
6396	return reg->type & MEM_RCU;
6397	}
6398
6399	static void clear_trusted_flags(enum bpf_type_flag *flag)
6400	{
6401	*flag &= ~(BPF_REG_TRUSTED_MODIFIERS | MEM_RCU);
6402	}
6403
6404	static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
6405	const struct bpf_reg_state *reg,
6406	int off, int size, bool strict)
6407	{
6408	struct tnum reg_off;
6409	int ip_align;
6410
6411	/* Byte size accesses are always allowed. */
6412	if (!strict || size == 1)
6413	return 0;
6414
6415	/* For platforms that do not have a Kconfig enabling
6416	* CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
6417	* NET_IP_ALIGN is universally set to '2'. And on platforms
6418	* that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
6419	* to this code only in strict mode where we want to emulate
6420	* the NET_IP_ALIGN==2 checking. Therefore use an
6421	* unconditional IP align value of '2'.
6422	*/
6423	ip_align = 2;
6424
6425	reg_off = tnum_add(a: reg->var_off, b: tnum_const(value: ip_align + reg->off + off));
6426	if (!tnum_is_aligned(a: reg_off, size)) {
6427	char tn_buf[48];
6428
6429	tnum_strn(str: tn_buf, size: sizeof(tn_buf), a: reg->var_off);
6430	verbose(private_data: env,
6431	fmt: "misaligned packet access off %d+%s+%d+%d size %d\n",
6432	ip_align, tn_buf, reg->off, off, size);
6433	return -EACCES;
6434	}
6435
6436	return 0;
6437	}
6438
6439	static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
6440	const struct bpf_reg_state *reg,
6441	const char *pointer_desc,
6442	int off, int size, bool strict)
6443	{
6444	struct tnum reg_off;
6445
6446	/* Byte size accesses are always allowed. */
6447	if (!strict || size == 1)
6448	return 0;
6449
6450	reg_off = tnum_add(a: reg->var_off, b: tnum_const(value: reg->off + off));
6451	if (!tnum_is_aligned(a: reg_off, size)) {
6452	char tn_buf[48];
6453
6454	tnum_strn(str: tn_buf, size: sizeof(tn_buf), a: reg->var_off);
6455	verbose(private_data: env, fmt: "misaligned %saccess off %s+%d+%d size %d\n",
6456	pointer_desc, tn_buf, reg->off, off, size);
6457	return -EACCES;
6458	}
6459
6460	return 0;
6461	}
6462
6463	static int check_ptr_alignment(struct bpf_verifier_env *env,
6464	const struct bpf_reg_state *reg, int off,
6465	int size, bool strict_alignment_once)
6466	{
6467	bool strict = env->strict_alignment || strict_alignment_once;
6468	const char *pointer_desc = "";
6469
6470	switch (reg->type) {
6471	case PTR_TO_PACKET:
6472	case PTR_TO_PACKET_META:
6473	/* Special case, because of NET_IP_ALIGN. Given metadata sits
6474	* right in front, treat it the very same way.
6475	*/
6476	return check_pkt_ptr_alignment(env, reg, off, size, strict);
6477	case PTR_TO_FLOW_KEYS:
6478	pointer_desc = "flow keys ";
6479	break;
6480	case PTR_TO_MAP_KEY:
6481	pointer_desc = "key ";
6482	break;
6483	case PTR_TO_MAP_VALUE:
6484	pointer_desc = "value ";
6485	if (reg->map_ptr->map_type == BPF_MAP_TYPE_INSN_ARRAY)
6486	strict = true;
6487	break;
6488	case PTR_TO_CTX:
6489	pointer_desc = "context ";
6490	break;
6491	case PTR_TO_STACK:
6492	pointer_desc = "stack ";
6493	/* The stack spill tracking logic in check_stack_write_fixed_off()
6494	* and check_stack_read_fixed_off() relies on stack accesses being
6495	* aligned.
6496	*/
6497	strict = true;
6498	break;
6499	case PTR_TO_SOCKET:
6500	pointer_desc = "sock ";
6501	break;
6502	case PTR_TO_SOCK_COMMON:
6503	pointer_desc = "sock_common ";
6504	break;
6505	case PTR_TO_TCP_SOCK:
6506	pointer_desc = "tcp_sock ";
6507	break;
6508	case PTR_TO_XDP_SOCK:
6509	pointer_desc = "xdp_sock ";
6510	break;
6511	case PTR_TO_ARENA:
6512	return 0;
6513	default:
6514	break;
6515	}
6516	return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
6517	strict);
6518	}
6519
6520	static enum priv_stack_mode bpf_enable_priv_stack(struct bpf_prog *prog)
6521	{
6522	if (!bpf_jit_supports_private_stack())
6523	return NO_PRIV_STACK;
6524
6525	/* bpf_prog_check_recur() checks all prog types that use bpf trampoline
6526	* while kprobe/tp/perf_event/raw_tp don't use trampoline hence checked
6527	* explicitly.
6528	*/
6529	switch (prog->type) {
6530	case BPF_PROG_TYPE_KPROBE:
6531	case BPF_PROG_TYPE_TRACEPOINT:
6532	case BPF_PROG_TYPE_PERF_EVENT:
6533	case BPF_PROG_TYPE_RAW_TRACEPOINT:
6534	return PRIV_STACK_ADAPTIVE;
6535	case BPF_PROG_TYPE_TRACING:
6536	case BPF_PROG_TYPE_LSM:
6537	case BPF_PROG_TYPE_STRUCT_OPS:
6538	if (prog->aux->priv_stack_requested || bpf_prog_check_recur(prog))
6539	return PRIV_STACK_ADAPTIVE;
6540	fallthrough;
6541	default:
6542	break;
6543	}
6544
6545	return NO_PRIV_STACK;
6546	}
6547
6548	static int round_up_stack_depth(struct bpf_verifier_env *env, int stack_depth)
6549	{
6550	if (env->prog->jit_requested)
6551	return round_up(stack_depth, 16);
6552
6553	/* round up to 32-bytes, since this is granularity
6554	* of interpreter stack size
6555	*/
6556	return round_up(max_t(u32, stack_depth, 1), 32);
6557	}
6558
6559	/* starting from main bpf function walk all instructions of the function
6560	* and recursively walk all callees that given function can call.
6561	* Ignore jump and exit insns.
6562	* Since recursion is prevented by check_cfg() this algorithm
6563	* only needs a local stack of MAX_CALL_FRAMES to remember callsites
6564	*/
6565	static int check_max_stack_depth_subprog(struct bpf_verifier_env *env, int idx,
6566	bool priv_stack_supported)
6567	{
6568	struct bpf_subprog_info *subprog = env->subprog_info;
6569	struct bpf_insn *insn = env->prog->insnsi;
6570	int depth = 0, frame = 0, i, subprog_end, subprog_depth;
6571	bool tail_call_reachable = false;
6572	int ret_insn[MAX_CALL_FRAMES];
6573	int ret_prog[MAX_CALL_FRAMES];
6574	int j;
6575
6576	i = subprog[idx].start;
6577	if (!priv_stack_supported)
6578	subprog[idx].priv_stack_mode = NO_PRIV_STACK;
6579	process_func:
6580	/* protect against potential stack overflow that might happen when
6581	* bpf2bpf calls get combined with tailcalls. Limit the caller's stack
6582	* depth for such case down to 256 so that the worst case scenario
6583	* would result in 8k stack size (32 which is tailcall limit * 256 =
6584	* 8k).
6585	*
6586	* To get the idea what might happen, see an example:
6587	* func1 -> sub rsp, 128
6588	* subfunc1 -> sub rsp, 256
6589	* tailcall1 -> add rsp, 256
6590	* func2 -> sub rsp, 192 (total stack size = 128 + 192 = 320)
6591	* subfunc2 -> sub rsp, 64
6592	* subfunc22 -> sub rsp, 128
6593	* tailcall2 -> add rsp, 128
6594	* func3 -> sub rsp, 32 (total stack size 128 + 192 + 64 + 32 = 416)
6595	*
6596	* tailcall will unwind the current stack frame but it will not get rid
6597	* of caller's stack as shown on the example above.
6598	*/
6599	if (idx && subprog[idx].has_tail_call && depth >= 256) {
6600	verbose(private_data: env,
6601	fmt: "tail_calls are not allowed when call stack of previous frames is %d bytes. Too large\n",
6602	depth);
6603	return -EACCES;
6604	}
6605
6606	subprog_depth = round_up_stack_depth(env, stack_depth: subprog[idx].stack_depth);
6607	if (priv_stack_supported) {
6608	/* Request private stack support only if the subprog stack
6609	* depth is no less than BPF_PRIV_STACK_MIN_SIZE. This is to
6610	* avoid jit penalty if the stack usage is small.
6611	*/
6612	if (subprog[idx].priv_stack_mode == PRIV_STACK_UNKNOWN &&
6613	subprog_depth >= BPF_PRIV_STACK_MIN_SIZE)
6614	subprog[idx].priv_stack_mode = PRIV_STACK_ADAPTIVE;
6615	}
6616
6617	if (subprog[idx].priv_stack_mode == PRIV_STACK_ADAPTIVE) {
6618	if (subprog_depth > MAX_BPF_STACK) {
6619	verbose(private_data: env, fmt: "stack size of subprog %d is %d. Too large\n",
6620	idx, subprog_depth);
6621	return -EACCES;
6622	}
6623	} else {
6624	depth += subprog_depth;
6625	if (depth > MAX_BPF_STACK) {
6626	verbose(private_data: env, fmt: "combined stack size of %d calls is %d. Too large\n",
6627	frame + 1, depth);
6628	return -EACCES;
6629	}
6630	}
6631	continue_func:
6632	subprog_end = subprog[idx + 1].start;
6633	for (; i < subprog_end; i++) {
6634	int next_insn, sidx;
6635
6636	if (bpf_pseudo_kfunc_call(insn: insn + i) && !insn[i].off) {
6637	bool err = false;
6638
6639	if (!is_bpf_throw_kfunc(insn: insn + i))
6640	continue;
6641	if (subprog[idx].is_cb)
6642	err = true;
6643	for (int c = 0; c < frame && !err; c++) {
6644	if (subprog[ret_prog[c]].is_cb) {
6645	err = true;
6646	break;
6647	}
6648	}
6649	if (!err)
6650	continue;
6651	verbose(private_data: env,
6652	fmt: "bpf_throw kfunc (insn %d) cannot be called from callback subprog %d\n",
6653	i, idx);
6654	return -EINVAL;
6655	}
6656
6657	if (!bpf_pseudo_call(insn: insn + i) && !bpf_pseudo_func(insn: insn + i))
6658	continue;
6659	/* remember insn and function to return to */
6660	ret_insn[frame] = i + 1;
6661	ret_prog[frame] = idx;
6662
6663	/* find the callee */
6664	next_insn = i + insn[i].imm + 1;
6665	sidx = find_subprog(env, off: next_insn);
6666	if (verifier_bug_if(sidx < 0, env, "callee not found at insn %d", next_insn))
6667	return -EFAULT;
6668	if (subprog[sidx].is_async_cb) {
6669	if (subprog[sidx].has_tail_call) {
6670	verifier_bug(env, "subprog has tail_call and async cb");
6671	return -EFAULT;
6672	}
6673	/* async callbacks don't increase bpf prog stack size unless called directly */
6674	if (!bpf_pseudo_call(insn: insn + i))
6675	continue;
6676	if (subprog[sidx].is_exception_cb) {
6677	verbose(private_data: env, fmt: "insn %d cannot call exception cb directly", i);
6678	return -EINVAL;
6679	}
6680	}
6681	i = next_insn;
6682	idx = sidx;
6683	if (!priv_stack_supported)
6684	subprog[idx].priv_stack_mode = NO_PRIV_STACK;
6685
6686	if (subprog[idx].has_tail_call)
6687	tail_call_reachable = true;
6688
6689	frame++;
6690	if (frame >= MAX_CALL_FRAMES) {
6691	verbose(private_data: env, fmt: "the call stack of %d frames is too deep !\n",
6692	frame);
6693	return -E2BIG;
6694	}
6695	goto process_func;
6696	}
6697	/* if tail call got detected across bpf2bpf calls then mark each of the
6698	* currently present subprog frames as tail call reachable subprogs;
6699	* this info will be utilized by JIT so that we will be preserving the
6700	* tail call counter throughout bpf2bpf calls combined with tailcalls
6701	*/
6702	if (tail_call_reachable)
6703	for (j = 0; j < frame; j++) {
6704	if (subprog[ret_prog[j]].is_exception_cb) {
6705	verbose(private_data: env, fmt: "cannot tail call within exception cb\n");
6706	return -EINVAL;
6707	}
6708	subprog[ret_prog[j]].tail_call_reachable = true;
6709	}
6710	if (subprog[0].tail_call_reachable)
6711	env->prog->aux->tail_call_reachable = true;
6712
6713	/* end of for() loop means the last insn of the 'subprog'
6714	* was reached. Doesn't matter whether it was JA or EXIT
6715	*/
6716	if (frame == 0)
6717	return 0;
6718	if (subprog[idx].priv_stack_mode != PRIV_STACK_ADAPTIVE)
6719	depth -= round_up_stack_depth(env, stack_depth: subprog[idx].stack_depth);
6720	frame--;
6721	i = ret_insn[frame];
6722	idx = ret_prog[frame];
6723	goto continue_func;
6724	}
6725
6726	static int check_max_stack_depth(struct bpf_verifier_env *env)
6727	{
6728	enum priv_stack_mode priv_stack_mode = PRIV_STACK_UNKNOWN;
6729	struct bpf_subprog_info *si = env->subprog_info;
6730	bool priv_stack_supported;
6731	int ret;
6732
6733	for (int i = 0; i < env->subprog_cnt; i++) {
6734	if (si[i].has_tail_call) {
6735	priv_stack_mode = NO_PRIV_STACK;
6736	break;
6737	}
6738	}
6739
6740	if (priv_stack_mode == PRIV_STACK_UNKNOWN)
6741	priv_stack_mode = bpf_enable_priv_stack(prog: env->prog);
6742
6743	/* All async_cb subprogs use normal kernel stack. If a particular
6744	* subprog appears in both main prog and async_cb subtree, that
6745	* subprog will use normal kernel stack to avoid potential nesting.
6746	* The reverse subprog traversal ensures when main prog subtree is
6747	* checked, the subprogs appearing in async_cb subtrees are already
6748	* marked as using normal kernel stack, so stack size checking can
6749	* be done properly.
6750	*/
6751	for (int i = env->subprog_cnt - 1; i >= 0; i--) {
6752	if (!i || si[i].is_async_cb) {
6753	priv_stack_supported = !i && priv_stack_mode == PRIV_STACK_ADAPTIVE;
6754	ret = check_max_stack_depth_subprog(env, idx: i, priv_stack_supported);
6755	if (ret < 0)
6756	return ret;
6757	}
6758	}
6759
6760	for (int i = 0; i < env->subprog_cnt; i++) {
6761	if (si[i].priv_stack_mode == PRIV_STACK_ADAPTIVE) {
6762	env->prog->aux->jits_use_priv_stack = true;
6763	break;
6764	}
6765	}
6766
6767	return 0;
6768	}
6769
6770	#ifndef CONFIG_BPF_JIT_ALWAYS_ON
6771	static int get_callee_stack_depth(struct bpf_verifier_env *env,
6772	const struct bpf_insn *insn, int idx)
6773	{
6774	int start = idx + insn->imm + 1, subprog;
6775
6776	subprog = find_subprog(env, start);
6777	if (verifier_bug_if(subprog < 0, env, "get stack depth: no program at insn %d", start))
6778	return -EFAULT;
6779	return env->subprog_info[subprog].stack_depth;
6780	}
6781	#endif
6782
6783	static int __check_buffer_access(struct bpf_verifier_env *env,
6784	const char *buf_info,
6785	const struct bpf_reg_state *reg,
6786	int regno, int off, int size)
6787	{
6788	if (off < 0) {
6789	verbose(private_data: env,
6790	fmt: "R%d invalid %s buffer access: off=%d, size=%d\n",
6791	regno, buf_info, off, size);
6792	return -EACCES;
6793	}
6794	if (!tnum_is_const(a: reg->var_off) || reg->var_off.value) {
6795	char tn_buf[48];
6796
6797	tnum_strn(str: tn_buf, size: sizeof(tn_buf), a: reg->var_off);
6798	verbose(private_data: env,
6799	fmt: "R%d invalid variable buffer offset: off=%d, var_off=%s\n",
6800	regno, off, tn_buf);
6801	return -EACCES;
6802	}
6803
6804	return 0;
6805	}
6806
6807	static int check_tp_buffer_access(struct bpf_verifier_env *env,
6808	const struct bpf_reg_state *reg,
6809	int regno, int off, int size)
6810	{
6811	int err;
6812
6813	err = __check_buffer_access(env, buf_info: "tracepoint", reg, regno, off, size);
6814	if (err)
6815	return err;
6816
6817	if (off + size > env->prog->aux->max_tp_access)
6818	env->prog->aux->max_tp_access = off + size;
6819
6820	return 0;
6821	}
6822
6823	static int check_buffer_access(struct bpf_verifier_env *env,
6824	const struct bpf_reg_state *reg,
6825	int regno, int off, int size,
6826	bool zero_size_allowed,
6827	u32 *max_access)
6828	{
6829	const char *buf_info = type_is_rdonly_mem(type: reg->type) ? "rdonly" : "rdwr";
6830	int err;
6831
6832	err = __check_buffer_access(env, buf_info, reg, regno, off, size);
6833	if (err)
6834	return err;
6835
6836	if (off + size > *max_access)
6837	*max_access = off + size;
6838
6839	return 0;
6840	}
6841
6842	/* BPF architecture zero extends alu32 ops into 64-bit registesr */
6843	static void zext_32_to_64(struct bpf_reg_state *reg)
6844	{
6845	reg->var_off = tnum_subreg(a: reg->var_off);
6846	__reg_assign_32_into_64(reg);
6847	}
6848
6849	/* truncate register to smaller size (in bytes)
6850	* must be called with size < BPF_REG_SIZE
6851	*/
6852	static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
6853	{
6854	u64 mask;
6855
6856	/* clear high bits in bit representation */
6857	reg->var_off = tnum_cast(a: reg->var_off, size);
6858
6859	/* fix arithmetic bounds */
6860	mask = ((u64)1 << (size * 8)) - 1;
6861	if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
6862	reg->umin_value &= mask;
6863	reg->umax_value &= mask;
6864	} else {
6865	reg->umin_value = 0;
6866	reg->umax_value = mask;
6867	}
6868	reg->smin_value = reg->umin_value;
6869	reg->smax_value = reg->umax_value;
6870
6871	/* If size is smaller than 32bit register the 32bit register
6872	* values are also truncated so we push 64-bit bounds into
6873	* 32-bit bounds. Above were truncated < 32-bits already.
6874	*/
6875	if (size < 4)
6876	__mark_reg32_unbounded(reg);
6877
6878	reg_bounds_sync(reg);
6879	}
6880
6881	static void set_sext64_default_val(struct bpf_reg_state *reg, int size)
6882	{
6883	if (size == 1) {
6884	reg->smin_value = reg->s32_min_value = S8_MIN;
6885	reg->smax_value = reg->s32_max_value = S8_MAX;
6886	} else if (size == 2) {
6887	reg->smin_value = reg->s32_min_value = S16_MIN;
6888	reg->smax_value = reg->s32_max_value = S16_MAX;
6889	} else {
6890	/* size == 4 */
6891	reg->smin_value = reg->s32_min_value = S32_MIN;
6892	reg->smax_value = reg->s32_max_value = S32_MAX;
6893	}
6894	reg->umin_value = reg->u32_min_value = 0;
6895	reg->umax_value = U64_MAX;
6896	reg->u32_max_value = U32_MAX;
6897	reg->var_off = tnum_unknown;
6898	}
6899
6900	static void coerce_reg_to_size_sx(struct bpf_reg_state *reg, int size)
6901	{
6902	s64 init_s64_max, init_s64_min, s64_max, s64_min, u64_cval;
6903	u64 top_smax_value, top_smin_value;
6904	u64 num_bits = size * 8;
6905
6906	if (tnum_is_const(a: reg->var_off)) {
6907	u64_cval = reg->var_off.value;
6908	if (size == 1)
6909	reg->var_off = tnum_const(value: (s8)u64_cval);
6910	else if (size == 2)
6911	reg->var_off = tnum_const(value: (s16)u64_cval);
6912	else
6913	/* size == 4 */
6914	reg->var_off = tnum_const(value: (s32)u64_cval);
6915
6916	u64_cval = reg->var_off.value;
6917	reg->smax_value = reg->smin_value = u64_cval;
6918	reg->umax_value = reg->umin_value = u64_cval;
6919	reg->s32_max_value = reg->s32_min_value = u64_cval;
6920	reg->u32_max_value = reg->u32_min_value = u64_cval;
6921	return;
6922	}
6923
6924	top_smax_value = ((u64)reg->smax_value >> num_bits) << num_bits;
6925	top_smin_value = ((u64)reg->smin_value >> num_bits) << num_bits;
6926
6927	if (top_smax_value != top_smin_value)
6928	goto out;
6929
6930	/* find the s64_min and s64_min after sign extension */
6931	if (size == 1) {
6932	init_s64_max = (s8)reg->smax_value;
6933	init_s64_min = (s8)reg->smin_value;
6934	} else if (size == 2) {
6935	init_s64_max = (s16)reg->smax_value;
6936	init_s64_min = (s16)reg->smin_value;
6937	} else {
6938	init_s64_max = (s32)reg->smax_value;
6939	init_s64_min = (s32)reg->smin_value;
6940	}
6941
6942	s64_max = max(init_s64_max, init_s64_min);
6943	s64_min = min(init_s64_max, init_s64_min);
6944
6945	/* both of s64_max/s64_min positive or negative */
6946	if ((s64_max >= 0) == (s64_min >= 0)) {
6947	reg->s32_min_value = reg->smin_value = s64_min;
6948	reg->s32_max_value = reg->smax_value = s64_max;
6949	reg->u32_min_value = reg->umin_value = s64_min;
6950	reg->u32_max_value = reg->umax_value = s64_max;
6951	reg->var_off = tnum_range(min: s64_min, max: s64_max);
6952	return;
6953	}
6954
6955	out:
6956	set_sext64_default_val(reg, size);
6957	}
6958
6959	static void set_sext32_default_val(struct bpf_reg_state *reg, int size)
6960	{
6961	if (size == 1) {
6962	reg->s32_min_value = S8_MIN;
6963	reg->s32_max_value = S8_MAX;
6964	} else {
6965	/* size == 2 */
6966	reg->s32_min_value = S16_MIN;
6967	reg->s32_max_value = S16_MAX;
6968	}
6969	reg->u32_min_value = 0;
6970	reg->u32_max_value = U32_MAX;
6971	reg->var_off = tnum_subreg(a: tnum_unknown);
6972	}
6973
6974	static void coerce_subreg_to_size_sx(struct bpf_reg_state *reg, int size)
6975	{
6976	s32 init_s32_max, init_s32_min, s32_max, s32_min, u32_val;
6977	u32 top_smax_value, top_smin_value;
6978	u32 num_bits = size * 8;
6979
6980	if (tnum_is_const(a: reg->var_off)) {
6981	u32_val = reg->var_off.value;
6982	if (size == 1)
6983	reg->var_off = tnum_const(value: (s8)u32_val);
6984	else
6985	reg->var_off = tnum_const(value: (s16)u32_val);
6986
6987	u32_val = reg->var_off.value;
6988	reg->s32_min_value = reg->s32_max_value = u32_val;
6989	reg->u32_min_value = reg->u32_max_value = u32_val;
6990	return;
6991	}
6992
6993	top_smax_value = ((u32)reg->s32_max_value >> num_bits) << num_bits;
6994	top_smin_value = ((u32)reg->s32_min_value >> num_bits) << num_bits;
6995
6996	if (top_smax_value != top_smin_value)
6997	goto out;
6998
6999	/* find the s32_min and s32_min after sign extension */
7000	if (size == 1) {
7001	init_s32_max = (s8)reg->s32_max_value;
7002	init_s32_min = (s8)reg->s32_min_value;
7003	} else {
7004	/* size == 2 */
7005	init_s32_max = (s16)reg->s32_max_value;
7006	init_s32_min = (s16)reg->s32_min_value;
7007	}
7008	s32_max = max(init_s32_max, init_s32_min);
7009	s32_min = min(init_s32_max, init_s32_min);
7010
7011	if ((s32_min >= 0) == (s32_max >= 0)) {
7012	reg->s32_min_value = s32_min;
7013	reg->s32_max_value = s32_max;
7014	reg->u32_min_value = (u32)s32_min;
7015	reg->u32_max_value = (u32)s32_max;
7016	reg->var_off = tnum_subreg(a: tnum_range(min: s32_min, max: s32_max));
7017	return;
7018	}
7019
7020	out:
7021	set_sext32_default_val(reg, size);
7022	}
7023
7024	static bool bpf_map_is_rdonly(const struct bpf_map *map)
7025	{
7026	/* A map is considered read-only if the following condition are true:
7027	*
7028	* 1) BPF program side cannot change any of the map content. The
7029	* BPF_F_RDONLY_PROG flag is throughout the lifetime of a map
7030	* and was set at map creation time.
7031	* 2) The map value(s) have been initialized from user space by a
7032	* loader and then "frozen", such that no new map update/delete
7033	* operations from syscall side are possible for the rest of
7034	* the map's lifetime from that point onwards.
7035	* 3) Any parallel/pending map update/delete operations from syscall
7036	* side have been completed. Only after that point, it's safe to
7037	* assume that map value(s) are immutable.
7038	*/
7039	return (map->map_flags & BPF_F_RDONLY_PROG) &&
7040	READ_ONCE(map->frozen) &&
7041	!bpf_map_write_active(map);
7042	}
7043
7044	static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val,
7045	bool is_ldsx)
7046	{
7047	void *ptr;
7048	u64 addr;
7049	int err;
7050
7051	err = map->ops->map_direct_value_addr(map, &addr, off);
7052	if (err)
7053	return err;
7054	ptr = (void *)(long)addr + off;
7055
7056	switch (size) {
7057	case sizeof(u8):
7058	*val = is_ldsx ? (s64)*(s8 *)ptr : (u64)*(u8 *)ptr;
7059	break;
7060	case sizeof(u16):
7061	*val = is_ldsx ? (s64)*(s16 *)ptr : (u64)*(u16 *)ptr;
7062	break;
7063	case sizeof(u32):
7064	*val = is_ldsx ? (s64)*(s32 *)ptr : (u64)*(u32 *)ptr;
7065	break;
7066	case sizeof(u64):
7067	*val = *(u64 *)ptr;
7068	break;
7069	default:
7070	return -EINVAL;
7071	}
7072	return 0;
7073	}
7074
7075	#define BTF_TYPE_SAFE_RCU(__type) __PASTE(__type, __safe_rcu)
7076	#define BTF_TYPE_SAFE_RCU_OR_NULL(__type) __PASTE(__type, __safe_rcu_or_null)
7077	#define BTF_TYPE_SAFE_TRUSTED(__type) __PASTE(__type, __safe_trusted)
7078	#define BTF_TYPE_SAFE_TRUSTED_OR_NULL(__type) __PASTE(__type, __safe_trusted_or_null)
7079
7080	/*
7081	* Allow list few fields as RCU trusted or full trusted.
7082	* This logic doesn't allow mix tagging and will be removed once GCC supports
7083	* btf_type_tag.
7084	*/
7085
7086	/* RCU trusted: these fields are trusted in RCU CS and never NULL */
7087	BTF_TYPE_SAFE_RCU(struct task_struct) {
7088	const cpumask_t *cpus_ptr;
7089	struct css_set __rcu *cgroups;
7090	struct task_struct __rcu *real_parent;
7091	struct task_struct *group_leader;
7092	};
7093
7094	BTF_TYPE_SAFE_RCU(struct cgroup) {
7095	/* cgrp->kn is always accessible as documented in kernel/cgroup/cgroup.c */
7096	struct kernfs_node *kn;
7097	};
7098
7099	BTF_TYPE_SAFE_RCU(struct css_set) {
7100	struct cgroup *dfl_cgrp;
7101	};
7102
7103	BTF_TYPE_SAFE_RCU(struct cgroup_subsys_state) {
7104	struct cgroup *cgroup;
7105	};
7106
7107	/* RCU trusted: these fields are trusted in RCU CS and can be NULL */
7108	BTF_TYPE_SAFE_RCU_OR_NULL(struct mm_struct) {
7109	struct file __rcu *exe_file;
7110	#ifdef CONFIG_MEMCG
7111	struct task_struct __rcu *owner;
7112	#endif
7113	};
7114
7115	/* skb->sk, req->sk are not RCU protected, but we mark them as such
7116	* because bpf prog accessible sockets are SOCK_RCU_FREE.
7117	*/
7118	BTF_TYPE_SAFE_RCU_OR_NULL(struct sk_buff) {
7119	struct sock *sk;
7120	};
7121
7122	BTF_TYPE_SAFE_RCU_OR_NULL(struct request_sock) {
7123	struct sock *sk;
7124	};
7125
7126	/* full trusted: these fields are trusted even outside of RCU CS and never NULL */
7127	BTF_TYPE_SAFE_TRUSTED(struct bpf_iter_meta) {
7128	struct seq_file *seq;
7129	};
7130
7131	BTF_TYPE_SAFE_TRUSTED(struct bpf_iter__task) {
7132	struct bpf_iter_meta *meta;
7133	struct task_struct *task;
7134	};
7135
7136	BTF_TYPE_SAFE_TRUSTED(struct linux_binprm) {
7137	struct file *file;
7138	};
7139
7140	BTF_TYPE_SAFE_TRUSTED(struct file) {
7141	struct inode *f_inode;
7142	};
7143
7144	BTF_TYPE_SAFE_TRUSTED_OR_NULL(struct dentry) {
7145	struct inode *d_inode;
7146	};
7147
7148	BTF_TYPE_SAFE_TRUSTED_OR_NULL(struct socket) {
7149	struct sock *sk;
7150	};
7151
7152	BTF_TYPE_SAFE_TRUSTED_OR_NULL(struct vm_area_struct) {
7153	struct mm_struct *vm_mm;
7154	struct file *vm_file;
7155	};
7156
7157	static bool type_is_rcu(struct bpf_verifier_env *env,
7158	struct bpf_reg_state *reg,
7159	const char *field_name, u32 btf_id)
7160	{
7161	BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct task_struct));
7162	BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct cgroup));
7163	BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct css_set));
7164	BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU(struct cgroup_subsys_state));
7165
7166	return btf_nested_type_is_trusted(log: &env->log, reg, field_name, btf_id, suffix: "__safe_rcu");
7167	}
7168
7169	static bool type_is_rcu_or_null(struct bpf_verifier_env *env,
7170	struct bpf_reg_state *reg,
7171	const char *field_name, u32 btf_id)
7172	{
7173	BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct mm_struct));
7174	BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct sk_buff));
7175	BTF_TYPE_EMIT(BTF_TYPE_SAFE_RCU_OR_NULL(struct request_sock));
7176
7177	return btf_nested_type_is_trusted(log: &env->log, reg, field_name, btf_id, suffix: "__safe_rcu_or_null");
7178	}
7179
7180	static bool type_is_trusted(struct bpf_verifier_env *env,
7181	struct bpf_reg_state *reg,
7182	const char *field_name, u32 btf_id)
7183	{
7184	BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct bpf_iter_meta));
7185	BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct bpf_iter__task));
7186	BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct linux_binprm));
7187	BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED(struct file));
7188
7189	return btf_nested_type_is_trusted(log: &env->log, reg, field_name, btf_id, suffix: "__safe_trusted");
7190	}
7191
7192	static bool type_is_trusted_or_null(struct bpf_verifier_env *env,
7193	struct bpf_reg_state *reg,
7194	const char *field_name, u32 btf_id)
7195	{
7196	BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED_OR_NULL(struct socket));
7197	BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED_OR_NULL(struct dentry));
7198	BTF_TYPE_EMIT(BTF_TYPE_SAFE_TRUSTED_OR_NULL(struct vm_area_struct));
7199
7200	return btf_nested_type_is_trusted(log: &env->log, reg, field_name, btf_id,
7201	suffix: "__safe_trusted_or_null");
7202	}
7203
7204	static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
7205	struct bpf_reg_state *regs,
7206	int regno, int off, int size,
7207	enum bpf_access_type atype,
7208	int value_regno)
7209	{
7210	struct bpf_reg_state *reg = regs + regno;
7211	const struct btf_type *t = btf_type_by_id(btf: reg->btf, type_id: reg->btf_id);
7212	const char *tname = btf_name_by_offset(btf: reg->btf, offset: t->name_off);
7213	const char *field_name = NULL;
7214	enum bpf_type_flag flag = 0;
7215	u32 btf_id = 0;
7216	int ret;
7217
7218	if (!env->allow_ptr_leaks) {
7219	verbose(private_data: env,
7220	fmt: "'struct %s' access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
7221	tname);
7222	return -EPERM;
7223	}
7224	if (!env->prog->gpl_compatible && btf_is_kernel(btf: reg->btf)) {
7225	verbose(private_data: env,
7226	fmt: "Cannot access kernel 'struct %s' from non-GPL compatible program\n",
7227	tname);
7228	return -EINVAL;
7229	}
7230	if (off < 0) {
7231	verbose(private_data: env,
7232	fmt: "R%d is ptr_%s invalid negative access: off=%d\n",
7233	regno, tname, off);
7234	return -EACCES;
7235	}
7236	if (!tnum_is_const(a: reg->var_off) || reg->var_off.value) {
7237	char tn_buf[48];
7238
7239	tnum_strn(str: tn_buf, size: sizeof(tn_buf), a: reg->var_off);
7240	verbose(private_data: env,
7241	fmt: "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
7242	regno, tname, off, tn_buf);
7243	return -EACCES;
7244	}
7245
7246	if (reg->type & MEM_USER) {
7247	verbose(private_data: env,
7248	fmt: "R%d is ptr_%s access user memory: off=%d\n",
7249	regno, tname, off);
7250	return -EACCES;
7251	}
7252
7253	if (reg->type & MEM_PERCPU) {
7254	verbose(private_data: env,
7255	fmt: "R%d is ptr_%s access percpu memory: off=%d\n",
7256	regno, tname, off);
7257	return -EACCES;
7258	}
7259
7260	if (env->ops->btf_struct_access && !type_is_alloc(type: reg->type) && atype == BPF_WRITE) {
7261	if (!btf_is_kernel(btf: reg->btf)) {
7262	verifier_bug(env, "reg->btf must be kernel btf");
7263	return -EFAULT;
7264	}
7265	ret = env->ops->btf_struct_access(&env->log, reg, off, size);
7266	} else {
7267	/* Writes are permitted with default btf_struct_access for
7268	* program allocated objects (which always have ref_obj_id > 0),
7269	* but not for untrusted PTR_TO_BTF_ID | MEM_ALLOC.
7270	*/
7271	if (atype != BPF_READ && !type_is_ptr_alloc_obj(type: reg->type)) {
7272	verbose(private_data: env, fmt: "only read is supported\n");
7273	return -EACCES;
7274	}
7275
7276	if (type_is_alloc(type: reg->type) && !type_is_non_owning_ref(type: reg->type) &&
7277	!(reg->type & MEM_RCU) && !reg->ref_obj_id) {
7278	verifier_bug(env, "ref_obj_id for allocated object must be non-zero");
7279	return -EFAULT;
7280	}
7281
7282	ret = btf_struct_access(log: &env->log, reg, off, size, atype, next_btf_id: &btf_id, flag: &flag, field_name: &field_name);
7283	}
7284
7285	if (ret < 0)
7286	return ret;
7287
7288	if (ret != PTR_TO_BTF_ID) {
7289	/* just mark; */
7290
7291	} else if (type_flag(type: reg->type) & PTR_UNTRUSTED) {
7292	/* If this is an untrusted pointer, all pointers formed by walking it
7293	* also inherit the untrusted flag.
7294	*/
7295	flag = PTR_UNTRUSTED;
7296
7297	} else if (is_trusted_reg(reg) || is_rcu_reg(reg)) {
7298	/* By default any pointer obtained from walking a trusted pointer is no
7299	* longer trusted, unless the field being accessed has explicitly been
7300	* marked as inheriting its parent's state of trust (either full or RCU).
7301	* For example:
7302	* 'cgroups' pointer is untrusted if task->cgroups dereference
7303	* happened in a sleepable program outside of bpf_rcu_read_lock()
7304	* section. In a non-sleepable program it's trusted while in RCU CS (aka MEM_RCU).
7305	* Note bpf_rcu_read_unlock() converts MEM_RCU pointers to PTR_UNTRUSTED.
7306	*
7307	* A regular RCU-protected pointer with __rcu tag can also be deemed
7308	* trusted if we are in an RCU CS. Such pointer can be NULL.
7309	*/
7310	if (type_is_trusted(env, reg, field_name, btf_id)) {
7311	flag |= PTR_TRUSTED;
7312	} else if (type_is_trusted_or_null(env, reg, field_name, btf_id)) {
7313	flag |= PTR_TRUSTED | PTR_MAYBE_NULL;
7314	} else if (in_rcu_cs(env) && !type_may_be_null(type: reg->type)) {
7315	if (type_is_rcu(env, reg, field_name, btf_id)) {
7316	/* ignore __rcu tag and mark it MEM_RCU */
7317	flag |= MEM_RCU;
7318	} else if (flag & MEM_RCU ||
7319	type_is_rcu_or_null(env, reg, field_name, btf_id)) {
7320	/* __rcu tagged pointers can be NULL */
7321	flag |= MEM_RCU | PTR_MAYBE_NULL;
7322
7323	/* We always trust them */
7324	if (type_is_rcu_or_null(env, reg, field_name, btf_id) &&
7325	flag & PTR_UNTRUSTED)
7326	flag &= ~PTR_UNTRUSTED;
7327	} else if (flag & (MEM_PERCPU | MEM_USER)) {
7328	/* keep as-is */
7329	} else {
7330	/* walking unknown pointers yields old deprecated PTR_TO_BTF_ID */
7331	clear_trusted_flags(flag: &flag);
7332	}
7333	} else {
7334	/*
7335	* If not in RCU CS or MEM_RCU pointer can be NULL then
7336	* aggressively mark as untrusted otherwise such
7337	* pointers will be plain PTR_TO_BTF_ID without flags
7338	* and will be allowed to be passed into helpers for
7339	* compat reasons.
7340	*/
7341	flag = PTR_UNTRUSTED;
7342	}
7343	} else {
7344	/* Old compat. Deprecated */
7345	clear_trusted_flags(flag: &flag);
7346	}
7347
7348	if (atype == BPF_READ && value_regno >= 0) {
7349	ret = mark_btf_ld_reg(env, regs, regno: value_regno, reg_type: ret, btf: reg->btf, btf_id, flag);
7350	if (ret < 0)
7351	return ret;
7352	}
7353
7354	return 0;
7355	}
7356
7357	static int check_ptr_to_map_access(struct bpf_verifier_env *env,
7358	struct bpf_reg_state *regs,
7359	int regno, int off, int size,
7360	enum bpf_access_type atype,
7361	int value_regno)
7362	{
7363	struct bpf_reg_state *reg = regs + regno;
7364	struct bpf_map *map = reg->map_ptr;
7365	struct bpf_reg_state map_reg;
7366	enum bpf_type_flag flag = 0;
7367	const struct btf_type *t;
7368	const char *tname;
7369	u32 btf_id;
7370	int ret;
7371
7372	if (!btf_vmlinux) {
7373	verbose(private_data: env, fmt: "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n");
7374	return -ENOTSUPP;
7375	}
7376
7377	if (!map->ops->map_btf_id || !*map->ops->map_btf_id) {
7378	verbose(private_data: env, fmt: "map_ptr access not supported for map type %d\n",
7379	map->map_type);
7380	return -ENOTSUPP;
7381	}
7382
7383	t = btf_type_by_id(btf: btf_vmlinux, type_id: *map->ops->map_btf_id);
7384	tname = btf_name_by_offset(btf: btf_vmlinux, offset: t->name_off);
7385
7386	if (!env->allow_ptr_leaks) {
7387	verbose(private_data: env,
7388	fmt: "'struct %s' access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
7389	tname);
7390	return -EPERM;
7391	}
7392
7393	if (off < 0) {
7394	verbose(private_data: env, fmt: "R%d is %s invalid negative access: off=%d\n",
7395	regno, tname, off);
7396	return -EACCES;
7397	}
7398
7399	if (atype != BPF_READ) {
7400	verbose(private_data: env, fmt: "only read from %s is supported\n", tname);
7401	return -EACCES;
7402	}
7403
7404	/* Simulate access to a PTR_TO_BTF_ID */
7405	memset(&map_reg, 0, sizeof(map_reg));
7406	ret = mark_btf_ld_reg(env, regs: &map_reg, regno: 0, reg_type: PTR_TO_BTF_ID,
7407	btf: btf_vmlinux, btf_id: *map->ops->map_btf_id, flag: 0);
7408	if (ret < 0)
7409	return ret;
7410	ret = btf_struct_access(log: &env->log, reg: &map_reg, off, size, atype, next_btf_id: &btf_id, flag: &flag, NULL);
7411	if (ret < 0)
7412	return ret;
7413
7414	if (value_regno >= 0) {
7415	ret = mark_btf_ld_reg(env, regs, regno: value_regno, reg_type: ret, btf: btf_vmlinux, btf_id, flag);
7416	if (ret < 0)
7417	return ret;
7418	}
7419
7420	return 0;
7421	}
7422
7423	/* Check that the stack access at the given offset is within bounds. The
7424	* maximum valid offset is -1.
7425	*
7426	* The minimum valid offset is -MAX_BPF_STACK for writes, and
7427	* -state->allocated_stack for reads.
7428	*/
7429	static int check_stack_slot_within_bounds(struct bpf_verifier_env *env,
7430	s64 off,
7431	struct bpf_func_state *state,
7432	enum bpf_access_type t)
7433	{
7434	int min_valid_off;
7435
7436	if (t == BPF_WRITE || env->allow_uninit_stack)
7437	min_valid_off = -MAX_BPF_STACK;
7438	else
7439	min_valid_off = -state->allocated_stack;
7440
7441	if (off < min_valid_off || off > -1)
7442	return -EACCES;
7443	return 0;
7444	}
7445
7446	/* Check that the stack access at 'regno + off' falls within the maximum stack
7447	* bounds.
7448	*
7449	* 'off' includes `regno->offset`, but not its dynamic part (if any).
7450	*/
7451	static int check_stack_access_within_bounds(
7452	struct bpf_verifier_env *env,
7453	int regno, int off, int access_size,
7454	enum bpf_access_type type)
7455	{
7456	struct bpf_reg_state *regs = cur_regs(env);
7457	struct bpf_reg_state *reg = regs + regno;
7458	struct bpf_func_state *state = func(env, reg);
7459	s64 min_off, max_off;
7460	int err;
7461	char *err_extra;
7462
7463	if (type == BPF_READ)
7464	err_extra = " read from";
7465	else
7466	err_extra = " write to";
7467
7468	if (tnum_is_const(a: reg->var_off)) {
7469	min_off = (s64)reg->var_off.value + off;
7470	max_off = min_off + access_size;
7471	} else {
7472	if (reg->smax_value >= BPF_MAX_VAR_OFF ||
7473	reg->smin_value <= -BPF_MAX_VAR_OFF) {
7474	verbose(private_data: env, fmt: "invalid unbounded variable-offset%s stack R%d\n",
7475	err_extra, regno);
7476	return -EACCES;
7477	}
7478	min_off = reg->smin_value + off;
7479	max_off = reg->smax_value + off + access_size;
7480	}
7481
7482	err = check_stack_slot_within_bounds(env, off: min_off, state, t: type);
7483	if (!err && max_off > 0)
7484	err = -EINVAL; /* out of stack access into non-negative offsets */
7485	if (!err && access_size < 0)
7486	/* access_size should not be negative (or overflow an int); others checks
7487	* along the way should have prevented such an access.
7488	*/
7489	err = -EFAULT; /* invalid negative access size; integer overflow? */
7490
7491	if (err) {
7492	if (tnum_is_const(a: reg->var_off)) {
7493	verbose(private_data: env, fmt: "invalid%s stack R%d off=%d size=%d\n",
7494	err_extra, regno, off, access_size);
7495	} else {
7496	char tn_buf[48];
7497
7498	tnum_strn(str: tn_buf, size: sizeof(tn_buf), a: reg->var_off);
7499	verbose(private_data: env, fmt: "invalid variable-offset%s stack R%d var_off=%s off=%d size=%d\n",
7500	err_extra, regno, tn_buf, off, access_size);
7501	}
7502	return err;
7503	}
7504
7505	/* Note that there is no stack access with offset zero, so the needed stack
7506	* size is -min_off, not -min_off+1.
7507	*/
7508	return grow_stack_state(env, state, size: -min_off /* size */);
7509	}
7510
7511	static bool get_func_retval_range(struct bpf_prog *prog,
7512	struct bpf_retval_range *range)
7513	{
7514	if (prog->type == BPF_PROG_TYPE_LSM &&
7515	prog->expected_attach_type == BPF_LSM_MAC &&
7516	!bpf_lsm_get_retval_range(prog, range)) {
7517	return true;
7518	}
7519	return false;
7520	}
7521
7522	/* check whether memory at (regno + off) is accessible for t = (read | write)
7523	* if t==write, value_regno is a register which value is stored into memory
7524	* if t==read, value_regno is a register which will receive the value from memory
7525	* if t==write && value_regno==-1, some unknown value is stored into memory
7526	* if t==read && value_regno==-1, don't care what we read from memory
7527	*/
7528	static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
7529	int off, int bpf_size, enum bpf_access_type t,
7530	int value_regno, bool strict_alignment_once, bool is_ldsx)
7531	{
7532	struct bpf_reg_state *regs = cur_regs(env);
7533	struct bpf_reg_state *reg = regs + regno;
7534	int size, err = 0;
7535
7536	size = bpf_size_to_bytes(bpf_size);
7537	if (size < 0)
7538	return size;
7539
7540	/* alignment checks will add in reg->off themselves */
7541	err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
7542	if (err)
7543	return err;
7544
7545	/* for access checks, reg->off is just part of off */
7546	off += reg->off;
7547
7548	if (reg->type == PTR_TO_MAP_KEY) {
7549	if (t == BPF_WRITE) {
7550	verbose(private_data: env, fmt: "write to change key R%d not allowed\n", regno);
7551	return -EACCES;
7552	}
7553
7554	err = check_mem_region_access(env, regno, off, size,
7555	mem_size: reg->map_ptr->key_size, zero_size_allowed: false);
7556	if (err)
7557	return err;
7558	if (value_regno >= 0)
7559	mark_reg_unknown(env, regs, regno: value_regno);
7560	} else if (reg->type == PTR_TO_MAP_VALUE) {
7561	struct btf_field *kptr_field = NULL;
7562
7563	if (t == BPF_WRITE && value_regno >= 0 &&
7564	is_pointer_value(env, regno: value_regno)) {
7565	verbose(private_data: env, fmt: "R%d leaks addr into map\n", value_regno);
7566	return -EACCES;
7567	}
7568	err = check_map_access_type(env, regno, off, size, type: t);
7569	if (err)
7570	return err;
7571	err = check_map_access(env, regno, off, size, zero_size_allowed: false, src: ACCESS_DIRECT);
7572	if (err)
7573	return err;
7574	if (tnum_is_const(a: reg->var_off))
7575	kptr_field = btf_record_find(rec: reg->map_ptr->record,
7576	offset: off + reg->var_off.value, field_mask: BPF_KPTR | BPF_UPTR);
7577	if (kptr_field) {
7578	err = check_map_kptr_access(env, regno, value_regno, insn_idx, kptr_field);
7579	} else if (t == BPF_READ && value_regno >= 0) {
7580	struct bpf_map *map = reg->map_ptr;
7581
7582	/*
7583	* If map is read-only, track its contents as scalars,
7584	* unless it is an insn array (see the special case below)
7585	*/
7586	if (tnum_is_const(a: reg->var_off) &&
7587	bpf_map_is_rdonly(map) &&
7588	map->ops->map_direct_value_addr &&
7589	map->map_type != BPF_MAP_TYPE_INSN_ARRAY) {
7590	int map_off = off + reg->var_off.value;
7591	u64 val = 0;
7592
7593	err = bpf_map_direct_read(map, off: map_off, size,
7594	val: &val, is_ldsx);
7595	if (err)
7596	return err;
7597
7598	regs[value_regno].type = SCALAR_VALUE;
7599	__mark_reg_known(reg: &regs[value_regno], imm: val);
7600	} else if (map->map_type == BPF_MAP_TYPE_INSN_ARRAY) {
7601	if (bpf_size != BPF_DW) {
7602	verbose(private_data: env, fmt: "Invalid read of %d bytes from insn_array\n",
7603	size);
7604	return -EACCES;
7605	}
7606	copy_register_state(dst: &regs[value_regno], src: reg);
7607	regs[value_regno].type = PTR_TO_INSN;
7608	} else {
7609	mark_reg_unknown(env, regs, regno: value_regno);
7610	}
7611	}
7612	} else if (base_type(type: reg->type) == PTR_TO_MEM) {
7613	bool rdonly_mem = type_is_rdonly_mem(type: reg->type);
7614	bool rdonly_untrusted = rdonly_mem && (reg->type & PTR_UNTRUSTED);
7615
7616	if (type_may_be_null(type: reg->type)) {
7617	verbose(private_data: env, fmt: "R%d invalid mem access '%s'\n", regno,
7618	reg_type_str(env, type: reg->type));
7619	return -EACCES;
7620	}
7621
7622	if (t == BPF_WRITE && rdonly_mem) {
7623	verbose(private_data: env, fmt: "R%d cannot write into %s\n",
7624	regno, reg_type_str(env, type: reg->type));
7625	return -EACCES;
7626	}
7627
7628	if (t == BPF_WRITE && value_regno >= 0 &&
7629	is_pointer_value(env, regno: value_regno)) {
7630	verbose(private_data: env, fmt: "R%d leaks addr into mem\n", value_regno);
7631	return -EACCES;
7632	}
7633
7634	/*
7635	* Accesses to untrusted PTR_TO_MEM are done through probe
7636	* instructions, hence no need to check bounds in that case.
7637	*/
7638	if (!rdonly_untrusted)
7639	err = check_mem_region_access(env, regno, off, size,
7640	mem_size: reg->mem_size, zero_size_allowed: false);
7641	if (!err && value_regno >= 0 && (t == BPF_READ || rdonly_mem))
7642	mark_reg_unknown(env, regs, regno: value_regno);
7643	} else if (reg->type == PTR_TO_CTX) {
7644	struct bpf_retval_range range;
7645	struct bpf_insn_access_aux info = {
7646	.reg_type = SCALAR_VALUE,
7647	.is_ldsx = is_ldsx,
7648	.log = &env->log,
7649	};
7650
7651	if (t == BPF_WRITE && value_regno >= 0 &&
7652	is_pointer_value(env, regno: value_regno)) {
7653	verbose(private_data: env, fmt: "R%d leaks addr into ctx\n", value_regno);
7654	return -EACCES;
7655	}
7656
7657	err = check_ptr_off_reg(env, reg, regno);
7658	if (err < 0)
7659	return err;
7660
7661	err = check_ctx_access(env, insn_idx, off, size, t, info: &info);
7662	if (err)
7663	verbose_linfo(env, insn_off: insn_idx, prefix_fmt: "; ");
7664	if (!err && t == BPF_READ && value_regno >= 0) {
7665	/* ctx access returns either a scalar, or a
7666	* PTR_TO_PACKET[_META,_END]. In the latter
7667	* case, we know the offset is zero.
7668	*/
7669	if (info.reg_type == SCALAR_VALUE) {
7670	if (info.is_retval && get_func_retval_range(prog: env->prog, range: &range)) {
7671	err = __mark_reg_s32_range(env, regs, regno: value_regno,
7672	s32_min: range.minval, s32_max: range.maxval);
7673	if (err)
7674	return err;
7675	} else {
7676	mark_reg_unknown(env, regs, regno: value_regno);
7677	}
7678	} else {
7679	mark_reg_known_zero(env, regs,
7680	regno: value_regno);
7681	if (type_may_be_null(type: info.reg_type))
7682	regs[value_regno].id = ++env->id_gen;
7683	/* A load of ctx field could have different
7684	* actual load size with the one encoded in the
7685	* insn. When the dst is PTR, it is for sure not
7686	* a sub-register.
7687	*/
7688	regs[value_regno].subreg_def = DEF_NOT_SUBREG;
7689	if (base_type(type: info.reg_type) == PTR_TO_BTF_ID) {
7690	regs[value_regno].btf = info.btf;
7691	regs[value_regno].btf_id = info.btf_id;
7692	regs[value_regno].ref_obj_id = info.ref_obj_id;
7693	}
7694	}
7695	regs[value_regno].type = info.reg_type;
7696	}
7697
7698	} else if (reg->type == PTR_TO_STACK) {
7699	/* Basic bounds checks. */
7700	err = check_stack_access_within_bounds(env, regno, off, access_size: size, type: t);
7701	if (err)
7702	return err;
7703
7704	if (t == BPF_READ)
7705	err = check_stack_read(env, ptr_regno: regno, off, size,
7706	dst_regno: value_regno);
7707	else
7708	err = check_stack_write(env, ptr_regno: regno, off, size,
7709	value_regno, insn_idx);
7710	} else if (reg_is_pkt_pointer(reg)) {
7711	if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
7712	verbose(private_data: env, fmt: "cannot write into packet\n");
7713	return -EACCES;
7714	}
7715	if (t == BPF_WRITE && value_regno >= 0 &&
7716	is_pointer_value(env, regno: value_regno)) {
7717	verbose(private_data: env, fmt: "R%d leaks addr into packet\n",
7718	value_regno);
7719	return -EACCES;
7720	}
7721	err = check_packet_access(env, regno, off, size, zero_size_allowed: false);
7722	if (!err && t == BPF_READ && value_regno >= 0)
7723	mark_reg_unknown(env, regs, regno: value_regno);
7724	} else if (reg->type == PTR_TO_FLOW_KEYS) {
7725	if (t == BPF_WRITE && value_regno >= 0 &&
7726	is_pointer_value(env, regno: value_regno)) {
7727	verbose(private_data: env, fmt: "R%d leaks addr into flow keys\n",
7728	value_regno);
7729	return -EACCES;
7730	}
7731
7732	err = check_flow_keys_access(env, off, size);
7733	if (!err && t == BPF_READ && value_regno >= 0)
7734	mark_reg_unknown(env, regs, regno: value_regno);
7735	} else if (type_is_sk_pointer(type: reg->type)) {
7736	if (t == BPF_WRITE) {
7737	verbose(private_data: env, fmt: "R%d cannot write into %s\n",
7738	regno, reg_type_str(env, type: reg->type));
7739	return -EACCES;
7740	}
7741	err = check_sock_access(env, insn_idx, regno, off, size, t);
7742	if (!err && value_regno >= 0)
7743	mark_reg_unknown(env, regs, regno: value_regno);
7744	} else if (reg->type == PTR_TO_TP_BUFFER) {
7745	err = check_tp_buffer_access(env, reg, regno, off, size);
7746	if (!err && t == BPF_READ && value_regno >= 0)
7747	mark_reg_unknown(env, regs, regno: value_regno);
7748	} else if (base_type(type: reg->type) == PTR_TO_BTF_ID &&
7749	!type_may_be_null(type: reg->type)) {
7750	err = check_ptr_to_btf_access(env, regs, regno, off, size, atype: t,
7751	value_regno);
7752	} else if (reg->type == CONST_PTR_TO_MAP) {
7753	err = check_ptr_to_map_access(env, regs, regno, off, size, atype: t,
7754	value_regno);
7755	} else if (base_type(type: reg->type) == PTR_TO_BUF) {
7756	bool rdonly_mem = type_is_rdonly_mem(type: reg->type);
7757	u32 *max_access;
7758
7759	if (rdonly_mem) {
7760	if (t == BPF_WRITE) {
7761	verbose(private_data: env, fmt: "R%d cannot write into %s\n",
7762	regno, reg_type_str(env, type: reg->type));
7763	return -EACCES;
7764	}
7765	max_access = &env->prog->aux->max_rdonly_access;
7766	} else {
7767	max_access = &env->prog->aux->max_rdwr_access;
7768	}
7769
7770	err = check_buffer_access(env, reg, regno, off, size, zero_size_allowed: false,
7771	max_access);
7772
7773	if (!err && value_regno >= 0 && (rdonly_mem || t == BPF_READ))
7774	mark_reg_unknown(env, regs, regno: value_regno);
7775	} else if (reg->type == PTR_TO_ARENA) {
7776	if (t == BPF_READ && value_regno >= 0)
7777	mark_reg_unknown(env, regs, regno: value_regno);
7778	} else {
7779	verbose(private_data: env, fmt: "R%d invalid mem access '%s'\n", regno,
7780	reg_type_str(env, type: reg->type));
7781	return -EACCES;
7782	}
7783
7784	if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
7785	regs[value_regno].type == SCALAR_VALUE) {
7786	if (!is_ldsx)
7787	/* b/h/w load zero-extends, mark upper bits as known 0 */
7788	coerce_reg_to_size(reg: &regs[value_regno], size);
7789	else
7790	coerce_reg_to_size_sx(reg: &regs[value_regno], size);
7791	}
7792	return err;
7793	}
7794
7795	static int save_aux_ptr_type(struct bpf_verifier_env *env, enum bpf_reg_type type,
7796	bool allow_trust_mismatch);
7797
7798	static int check_load_mem(struct bpf_verifier_env *env, struct bpf_insn *insn,
7799	bool strict_alignment_once, bool is_ldsx,
7800	bool allow_trust_mismatch, const char *ctx)
7801	{
7802	struct bpf_reg_state *regs = cur_regs(env);
7803	enum bpf_reg_type src_reg_type;
7804	int err;
7805
7806	/* check src operand */
7807	err = check_reg_arg(env, regno: insn->src_reg, t: SRC_OP);
7808	if (err)
7809	return err;
7810
7811	/* check dst operand */
7812	err = check_reg_arg(env, regno: insn->dst_reg, t: DST_OP_NO_MARK);
7813	if (err)
7814	return err;
7815
7816	src_reg_type = regs[insn->src_reg].type;
7817
7818	/* Check if (src_reg + off) is readable. The state of dst_reg will be
7819	* updated by this call.
7820	*/
7821	err = check_mem_access(env, insn_idx: env->insn_idx, regno: insn->src_reg, off: insn->off,
7822	BPF_SIZE(insn->code), t: BPF_READ, value_regno: insn->dst_reg,
7823	strict_alignment_once, is_ldsx);
7824	err = err ?: save_aux_ptr_type(env, type: src_reg_type,
7825	allow_trust_mismatch);
7826	err = err ?: reg_bounds_sanity_check(env, reg: &regs[insn->dst_reg], ctx);
7827
7828	return err;
7829	}
7830
7831	static int check_store_reg(struct bpf_verifier_env *env, struct bpf_insn *insn,
7832	bool strict_alignment_once)
7833	{
7834	struct bpf_reg_state *regs = cur_regs(env);
7835	enum bpf_reg_type dst_reg_type;
7836	int err;
7837
7838	/* check src1 operand */
7839	err = check_reg_arg(env, regno: insn->src_reg, t: SRC_OP);
7840	if (err)
7841	return err;
7842
7843	/* check src2 operand */
7844	err = check_reg_arg(env, regno: insn->dst_reg, t: SRC_OP);
7845	if (err)
7846	return err;
7847
7848	dst_reg_type = regs[insn->dst_reg].type;
7849
7850	/* Check if (dst_reg + off) is writeable. */
7851	err = check_mem_access(env, insn_idx: env->insn_idx, regno: insn->dst_reg, off: insn->off,
7852	BPF_SIZE(insn->code), t: BPF_WRITE, value_regno: insn->src_reg,
7853	strict_alignment_once, is_ldsx: false);
7854	err = err ?: save_aux_ptr_type(env, type: dst_reg_type, allow_trust_mismatch: false);
7855
7856	return err;
7857	}
7858
7859	static int check_atomic_rmw(struct bpf_verifier_env *env,
7860	struct bpf_insn *insn)
7861	{
7862	int load_reg;
7863	int err;
7864
7865	if (BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) {
7866	verbose(private_data: env, fmt: "invalid atomic operand size\n");
7867	return -EINVAL;
7868	}
7869
7870	/* check src1 operand */
7871	err = check_reg_arg(env, regno: insn->src_reg, t: SRC_OP);
7872	if (err)
7873	return err;
7874
7875	/* check src2 operand */
7876	err = check_reg_arg(env, regno: insn->dst_reg, t: SRC_OP);
7877	if (err)
7878	return err;
7879
7880	if (insn->imm == BPF_CMPXCHG) {
7881	/* Check comparison of R0 with memory location */
7882	const u32 aux_reg = BPF_REG_0;
7883
7884	err = check_reg_arg(env, regno: aux_reg, t: SRC_OP);
7885	if (err)
7886	return err;
7887
7888	if (is_pointer_value(env, regno: aux_reg)) {
7889	verbose(private_data: env, fmt: "R%d leaks addr into mem\n", aux_reg);
7890	return -EACCES;
7891	}
7892	}
7893
7894	if (is_pointer_value(env, regno: insn->src_reg)) {
7895	verbose(private_data: env, fmt: "R%d leaks addr into mem\n", insn->src_reg);
7896	return -EACCES;
7897	}
7898
7899	if (!atomic_ptr_type_ok(env, regno: insn->dst_reg, insn)) {
7900	verbose(private_data: env, fmt: "BPF_ATOMIC stores into R%d %s is not allowed\n",
7901	insn->dst_reg,
7902	reg_type_str(env, type: reg_state(env, regno: insn->dst_reg)->type));
7903	return -EACCES;
7904	}
7905
7906	if (insn->imm & BPF_FETCH) {
7907	if (insn->imm == BPF_CMPXCHG)
7908	load_reg = BPF_REG_0;
7909	else
7910	load_reg = insn->src_reg;
7911
7912	/* check and record load of old value */
7913	err = check_reg_arg(env, regno: load_reg, t: DST_OP);
7914	if (err)
7915	return err;
7916	} else {
7917	/* This instruction accesses a memory location but doesn't
7918	* actually load it into a register.
7919	*/
7920	load_reg = -1;
7921	}
7922
7923	/* Check whether we can read the memory, with second call for fetch
7924	* case to simulate the register fill.
7925	*/
7926	err = check_mem_access(env, insn_idx: env->insn_idx, regno: insn->dst_reg, off: insn->off,
7927	BPF_SIZE(insn->code), t: BPF_READ, value_regno: -1, strict_alignment_once: true, is_ldsx: false);
7928	if (!err && load_reg >= 0)
7929	err = check_mem_access(env, insn_idx: env->insn_idx, regno: insn->dst_reg,
7930	off: insn->off, BPF_SIZE(insn->code),
7931	t: BPF_READ, value_regno: load_reg, strict_alignment_once: true, is_ldsx: false);
7932	if (err)
7933	return err;
7934
7935	if (is_arena_reg(env, regno: insn->dst_reg)) {
7936	err = save_aux_ptr_type(env, type: PTR_TO_ARENA, allow_trust_mismatch: false);
7937	if (err)
7938	return err;
7939	}
7940	/* Check whether we can write into the same memory. */
7941	err = check_mem_access(env, insn_idx: env->insn_idx, regno: insn->dst_reg, off: insn->off,
7942	BPF_SIZE(insn->code), t: BPF_WRITE, value_regno: -1, strict_alignment_once: true, is_ldsx: false);
7943	if (err)
7944	return err;
7945	return 0;
7946	}
7947
7948	static int check_atomic_load(struct bpf_verifier_env *env,
7949	struct bpf_insn *insn)
7950	{
7951	int err;
7952
7953	err = check_load_mem(env, insn, strict_alignment_once: true, is_ldsx: false, allow_trust_mismatch: false, ctx: "atomic_load");
7954	if (err)
7955	return err;
7956
7957	if (!atomic_ptr_type_ok(env, regno: insn->src_reg, insn)) {
7958	verbose(private_data: env, fmt: "BPF_ATOMIC loads from R%d %s is not allowed\n",
7959	insn->src_reg,
7960	reg_type_str(env, type: reg_state(env, regno: insn->src_reg)->type));
7961	return -EACCES;
7962	}
7963
7964	return 0;
7965	}
7966
7967	static int check_atomic_store(struct bpf_verifier_env *env,
7968	struct bpf_insn *insn)
7969	{
7970	int err;
7971
7972	err = check_store_reg(env, insn, strict_alignment_once: true);
7973	if (err)
7974	return err;
7975
7976	if (!atomic_ptr_type_ok(env, regno: insn->dst_reg, insn)) {
7977	verbose(private_data: env, fmt: "BPF_ATOMIC stores into R%d %s is not allowed\n",
7978	insn->dst_reg,
7979	reg_type_str(env, type: reg_state(env, regno: insn->dst_reg)->type));
7980	return -EACCES;
7981	}
7982
7983	return 0;
7984	}
7985
7986	static int check_atomic(struct bpf_verifier_env *env, struct bpf_insn *insn)
7987	{
7988	switch (insn->imm) {
7989	case BPF_ADD:
7990	case BPF_ADD | BPF_FETCH:
7991	case BPF_AND:
7992	case BPF_AND | BPF_FETCH:
7993	case BPF_OR:
7994	case BPF_OR | BPF_FETCH:
7995	case BPF_XOR:
7996	case BPF_XOR | BPF_FETCH:
7997	case BPF_XCHG:
7998	case BPF_CMPXCHG:
7999	return check_atomic_rmw(env, insn);
8000	case BPF_LOAD_ACQ:
8001	if (BPF_SIZE(insn->code) == BPF_DW && BITS_PER_LONG != 64) {
8002	verbose(private_data: env,
8003	fmt: "64-bit load-acquires are only supported on 64-bit arches\n");
8004	return -EOPNOTSUPP;
8005	}
8006	return check_atomic_load(env, insn);
8007	case BPF_STORE_REL:
8008	if (BPF_SIZE(insn->code) == BPF_DW && BITS_PER_LONG != 64) {
8009	verbose(private_data: env,
8010	fmt: "64-bit store-releases are only supported on 64-bit arches\n");
8011	return -EOPNOTSUPP;
8012	}
8013	return check_atomic_store(env, insn);
8014	default:
8015	verbose(private_data: env, fmt: "BPF_ATOMIC uses invalid atomic opcode %02x\n",
8016	insn->imm);
8017	return -EINVAL;
8018	}
8019	}
8020
8021	/* When register 'regno' is used to read the stack (either directly or through
8022	* a helper function) make sure that it's within stack boundary and, depending
8023	* on the access type and privileges, that all elements of the stack are
8024	* initialized.
8025	*
8026	* 'off' includes 'regno->off', but not its dynamic part (if any).
8027	*
8028	* All registers that have been spilled on the stack in the slots within the
8029	* read offsets are marked as read.
8030	*/
8031	static int check_stack_range_initialized(
8032	struct bpf_verifier_env *env, int regno, int off,
8033	int access_size, bool zero_size_allowed,
8034	enum bpf_access_type type, struct bpf_call_arg_meta *meta)
8035	{
8036	struct bpf_reg_state *reg = reg_state(env, regno);
8037	struct bpf_func_state *state = func(env, reg);
8038	int err, min_off, max_off, i, j, slot, spi;
8039	/* Some accesses can write anything into the stack, others are
8040	* read-only.
8041	*/
8042	bool clobber = false;
8043
8044	if (access_size == 0 && !zero_size_allowed) {
8045	verbose(private_data: env, fmt: "invalid zero-sized read\n");
8046	return -EACCES;
8047	}
8048
8049	if (type == BPF_WRITE)
8050	clobber = true;
8051
8052	err = check_stack_access_within_bounds(env, regno, off, access_size, type);
8053	if (err)
8054	return err;
8055
8056
8057	if (tnum_is_const(a: reg->var_off)) {
8058	min_off = max_off = reg->var_off.value + off;
8059	} else {
8060	/* Variable offset is prohibited for unprivileged mode for
8061	* simplicity since it requires corresponding support in
8062	* Spectre masking for stack ALU.
8063	* See also retrieve_ptr_limit().
8064	*/
8065	if (!env->bypass_spec_v1) {
8066	char tn_buf[48];
8067
8068	tnum_strn(str: tn_buf, size: sizeof(tn_buf), a: reg->var_off);
8069	verbose(private_data: env, fmt: "R%d variable offset stack access prohibited for !root, var_off=%s\n",
8070	regno, tn_buf);
8071	return -EACCES;
8072	}
8073	/* Only initialized buffer on stack is allowed to be accessed
8074	* with variable offset. With uninitialized buffer it's hard to
8075	* guarantee that whole memory is marked as initialized on
8076	* helper return since specific bounds are unknown what may
8077	* cause uninitialized stack leaking.
8078	*/
8079	if (meta && meta->raw_mode)
8080	meta = NULL;
8081
8082	min_off = reg->smin_value + off;
8083	max_off = reg->smax_value + off;
8084	}
8085
8086	if (meta && meta->raw_mode) {
8087	/* Ensure we won't be overwriting dynptrs when simulating byte
8088	* by byte access in check_helper_call using meta.access_size.
8089	* This would be a problem if we have a helper in the future
8090	* which takes:
8091	*
8092	* helper(uninit_mem, len, dynptr)
8093	*
8094	* Now, uninint_mem may overlap with dynptr pointer. Hence, it
8095	* may end up writing to dynptr itself when touching memory from
8096	* arg 1. This can be relaxed on a case by case basis for known
8097	* safe cases, but reject due to the possibilitiy of aliasing by
8098	* default.
8099	*/
8100	for (i = min_off; i < max_off + access_size; i++) {
8101	int stack_off = -i - 1;
8102
8103	spi = __get_spi(off: i);
8104	/* raw_mode may write past allocated_stack */
8105	if (state->allocated_stack <= stack_off)
8106	continue;
8107	if (state->stack[spi].slot_type[stack_off % BPF_REG_SIZE] == STACK_DYNPTR) {
8108	verbose(private_data: env, fmt: "potential write to dynptr at off=%d disallowed\n", i);
8109	return -EACCES;
8110	}
8111	}
8112	meta->access_size = access_size;
8113	meta->regno = regno;
8114	return 0;
8115	}
8116
8117	for (i = min_off; i < max_off + access_size; i++) {
8118	u8 *stype;
8119
8120	slot = -i - 1;
8121	spi = slot / BPF_REG_SIZE;
8122	if (state->allocated_stack <= slot) {
8123	verbose(private_data: env, fmt: "allocated_stack too small\n");
8124	return -EFAULT;
8125	}
8126
8127	stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
8128	if (*stype == STACK_MISC)
8129	goto mark;
8130	if ((*stype == STACK_ZERO) ||
8131	(*stype == STACK_INVALID && env->allow_uninit_stack)) {
8132	if (clobber) {
8133	/* helper can write anything into the stack */
8134	*stype = STACK_MISC;
8135	}
8136	goto mark;
8137	}
8138
8139	if (is_spilled_reg(stack: &state->stack[spi]) &&
8140	(state->stack[spi].spilled_ptr.type == SCALAR_VALUE ||
8141	env->allow_ptr_leaks)) {
8142	if (clobber) {
8143	__mark_reg_unknown(env, reg: &state->stack[spi].spilled_ptr);
8144	for (j = 0; j < BPF_REG_SIZE; j++)
8145	scrub_spilled_slot(stype: &state->stack[spi].slot_type[j]);
8146	}
8147	goto mark;
8148	}
8149
8150	if (tnum_is_const(a: reg->var_off)) {
8151	verbose(private_data: env, fmt: "invalid read from stack R%d off %d+%d size %d\n",
8152	regno, min_off, i - min_off, access_size);
8153	} else {
8154	char tn_buf[48];
8155
8156	tnum_strn(str: tn_buf, size: sizeof(tn_buf), a: reg->var_off);
8157	verbose(private_data: env, fmt: "invalid read from stack R%d var_off %s+%d size %d\n",
8158	regno, tn_buf, i - min_off, access_size);
8159	}
8160	return -EACCES;
8161	mark:
8162	/* reading any byte out of 8-byte 'spill_slot' will cause
8163	* the whole slot to be marked as 'read'
8164	*/
8165	err = bpf_mark_stack_read(env, frameno: reg->frameno, insn_idx: env->insn_idx, BIT(spi));
8166	if (err)
8167	return err;
8168	/* We do not call bpf_mark_stack_write(), as we can not
8169	* be sure that whether stack slot is written to or not. Hence,
8170	* we must still conservatively propagate reads upwards even if
8171	* helper may write to the entire memory range.
8172	*/
8173	}
8174	return 0;
8175	}
8176
8177	static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
8178	int access_size, enum bpf_access_type access_type,
8179	bool zero_size_allowed,
8180	struct bpf_call_arg_meta *meta)
8181	{
8182	struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[regno];
8183	u32 *max_access;
8184
8185	switch (base_type(type: reg->type)) {
8186	case PTR_TO_PACKET:
8187	case PTR_TO_PACKET_META:
8188	return check_packet_access(env, regno, off: reg->off, size: access_size,
8189	zero_size_allowed);
8190	case PTR_TO_MAP_KEY:
8191	if (access_type == BPF_WRITE) {
8192	verbose(private_data: env, fmt: "R%d cannot write into %s\n", regno,
8193	reg_type_str(env, type: reg->type));
8194	return -EACCES;
8195	}
8196	return check_mem_region_access(env, regno, off: reg->off, size: access_size,
8197	mem_size: reg->map_ptr->key_size, zero_size_allowed: false);
8198	case PTR_TO_MAP_VALUE:
8199	if (check_map_access_type(env, regno, off: reg->off, size: access_size, type: access_type))
8200	return -EACCES;
8201	return check_map_access(env, regno, off: reg->off, size: access_size,
8202	zero_size_allowed, src: ACCESS_HELPER);
8203	case PTR_TO_MEM:
8204	if (type_is_rdonly_mem(type: reg->type)) {
8205	if (access_type == BPF_WRITE) {
8206	verbose(private_data: env, fmt: "R%d cannot write into %s\n", regno,
8207	reg_type_str(env, type: reg->type));
8208	return -EACCES;
8209	}
8210	}
8211	return check_mem_region_access(env, regno, off: reg->off,
8212	size: access_size, mem_size: reg->mem_size,
8213	zero_size_allowed);
8214	case PTR_TO_BUF:
8215	if (type_is_rdonly_mem(type: reg->type)) {
8216	if (access_type == BPF_WRITE) {
8217	verbose(private_data: env, fmt: "R%d cannot write into %s\n", regno,
8218	reg_type_str(env, type: reg->type));
8219	return -EACCES;
8220	}
8221
8222	max_access = &env->prog->aux->max_rdonly_access;
8223	} else {
8224	max_access = &env->prog->aux->max_rdwr_access;
8225	}
8226	return check_buffer_access(env, reg, regno, off: reg->off,
8227	size: access_size, zero_size_allowed,
8228	max_access);
8229	case PTR_TO_STACK:
8230	return check_stack_range_initialized(
8231	env,
8232	regno, off: reg->off, access_size,
8233	zero_size_allowed, type: access_type, meta);
8234	case PTR_TO_BTF_ID:
8235	return check_ptr_to_btf_access(env, regs, regno, off: reg->off,
8236	size: access_size, atype: BPF_READ, value_regno: -1);
8237	case PTR_TO_CTX:
8238	/* in case the function doesn't know how to access the context,
8239	* (because we are in a program of type SYSCALL for example), we
8240	* can not statically check its size.
8241	* Dynamically check it now.
8242	*/
8243	if (!env->ops->convert_ctx_access) {
8244	int offset = access_size - 1;
8245
8246	/* Allow zero-byte read from PTR_TO_CTX */
8247	if (access_size == 0)
8248	return zero_size_allowed ? 0 : -EACCES;
8249
8250	return check_mem_access(env, insn_idx: env->insn_idx, regno, off: offset, BPF_B,
8251	t: access_type, value_regno: -1, strict_alignment_once: false, is_ldsx: false);
8252	}
8253
8254	fallthrough;
8255	default: /* scalar_value or invalid ptr */
8256	/* Allow zero-byte read from NULL, regardless of pointer type */
8257	if (zero_size_allowed && access_size == 0 &&
8258	register_is_null(reg))
8259	return 0;
8260
8261	verbose(private_data: env, fmt: "R%d type=%s ", regno,
8262	reg_type_str(env, type: reg->type));
8263	verbose(private_data: env, fmt: "expected=%s\n", reg_type_str(env, type: PTR_TO_STACK));
8264	return -EACCES;
8265	}
8266	}
8267
8268	/* verify arguments to helpers or kfuncs consisting of a pointer and an access
8269	* size.
8270	*
8271	* @regno is the register containing the access size. regno-1 is the register
8272	* containing the pointer.
8273	*/
8274	static int check_mem_size_reg(struct bpf_verifier_env *env,
8275	struct bpf_reg_state *reg, u32 regno,
8276	enum bpf_access_type access_type,
8277	bool zero_size_allowed,
8278	struct bpf_call_arg_meta *meta)
8279	{
8280	int err;
8281
8282	/* This is used to refine r0 return value bounds for helpers
8283	* that enforce this value as an upper bound on return values.
8284	* See do_refine_retval_range() for helpers that can refine
8285	* the return value. C type of helper is u32 so we pull register
8286	* bound from umax_value however, if negative verifier errors
8287	* out. Only upper bounds can be learned because retval is an
8288	* int type and negative retvals are allowed.
8289	*/
8290	meta->msize_max_value = reg->umax_value;
8291
8292	/* The register is SCALAR_VALUE; the access check happens using
8293	* its boundaries. For unprivileged variable accesses, disable
8294	* raw mode so that the program is required to initialize all
8295	* the memory that the helper could just partially fill up.
8296	*/
8297	if (!tnum_is_const(a: reg->var_off))
8298	meta = NULL;
8299
8300	if (reg->smin_value < 0) {
8301	verbose(private_data: env, fmt: "R%d min value is negative, either use unsigned or 'var &= const'\n",
8302	regno);
8303	return -EACCES;
8304	}
8305
8306	if (reg->umin_value == 0 && !zero_size_allowed) {
8307	verbose(private_data: env, fmt: "R%d invalid zero-sized read: u64=[%lld,%lld]\n",
8308	regno, reg->umin_value, reg->umax_value);
8309	return -EACCES;
8310	}
8311
8312	if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
8313	verbose(private_data: env, fmt: "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
8314	regno);
8315	return -EACCES;
8316	}
8317	err = check_helper_mem_access(env, regno: regno - 1, access_size: reg->umax_value,
8318	access_type, zero_size_allowed, meta);
8319	if (!err)
8320	err = mark_chain_precision(env, regno);
8321	return err;
8322	}
8323
8324	static int check_mem_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
8325	u32 regno, u32 mem_size)
8326	{
8327	bool may_be_null = type_may_be_null(type: reg->type);
8328	struct bpf_reg_state saved_reg;
8329	int err;
8330
8331	if (register_is_null(reg))
8332	return 0;
8333
8334	/* Assuming that the register contains a value check if the memory
8335	* access is safe. Temporarily save and restore the register's state as
8336	* the conversion shouldn't be visible to a caller.
8337	*/
8338	if (may_be_null) {
8339	saved_reg = *reg;
8340	mark_ptr_not_null_reg(reg);
8341	}
8342
8343	err = check_helper_mem_access(env, regno, access_size: mem_size, access_type: BPF_READ, zero_size_allowed: true, NULL);
8344	err = err ?: check_helper_mem_access(env, regno, access_size: mem_size, access_type: BPF_WRITE, zero_size_allowed: true, NULL);
8345
8346	if (may_be_null)
8347	*reg = saved_reg;
8348
8349	return err;
8350	}
8351
8352	static int check_kfunc_mem_size_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
8353	u32 regno)
8354	{
8355	struct bpf_reg_state *mem_reg = &cur_regs(env)[regno - 1];
8356	bool may_be_null = type_may_be_null(type: mem_reg->type);
8357	struct bpf_reg_state saved_reg;
8358	struct bpf_call_arg_meta meta;
8359	int err;
8360
8361	WARN_ON_ONCE(regno < BPF_REG_2 || regno > BPF_REG_5);
8362
8363	memset(&meta, 0, sizeof(meta));
8364
8365	if (may_be_null) {
8366	saved_reg = *mem_reg;
8367	mark_ptr_not_null_reg(reg: mem_reg);
8368	}
8369
8370	err = check_mem_size_reg(env, reg, regno, access_type: BPF_READ, zero_size_allowed: true, meta: &meta);
8371	err = err ?: check_mem_size_reg(env, reg, regno, access_type: BPF_WRITE, zero_size_allowed: true, meta: &meta);
8372
8373	if (may_be_null)
8374	*mem_reg = saved_reg;
8375
8376	return err;
8377	}
8378
8379	enum {
8380	PROCESS_SPIN_LOCK = (1 << 0),
8381	PROCESS_RES_LOCK = (1 << 1),
8382	PROCESS_LOCK_IRQ = (1 << 2),
8383	};
8384
8385	/* Implementation details:
8386	* bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL.
8387	* bpf_obj_new returns PTR_TO_BTF_ID | MEM_ALLOC | PTR_MAYBE_NULL.
8388	* Two bpf_map_lookups (even with the same key) will have different reg->id.
8389	* Two separate bpf_obj_new will also have different reg->id.
8390	* For traditional PTR_TO_MAP_VALUE or PTR_TO_BTF_ID | MEM_ALLOC, the verifier
8391	* clears reg->id after value_or_null->value transition, since the verifier only
8392	* cares about the range of access to valid map value pointer and doesn't care
8393	* about actual address of the map element.
8394	* For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
8395	* reg->id > 0 after value_or_null->value transition. By doing so
8396	* two bpf_map_lookups will be considered two different pointers that
8397	* point to different bpf_spin_locks. Likewise for pointers to allocated objects
8398	* returned from bpf_obj_new.
8399	* The verifier allows taking only one bpf_spin_lock at a time to avoid
8400	* dead-locks.
8401	* Since only one bpf_spin_lock is allowed the checks are simpler than
8402	* reg_is_refcounted() logic. The verifier needs to remember only
8403	* one spin_lock instead of array of acquired_refs.
8404	* env->cur_state->active_locks remembers which map value element or allocated
8405	* object got locked and clears it after bpf_spin_unlock.
8406	*/
8407	static int process_spin_lock(struct bpf_verifier_env *env, int regno, int flags)
8408	{
8409	bool is_lock = flags & PROCESS_SPIN_LOCK, is_res_lock = flags & PROCESS_RES_LOCK;
8410	const char *lock_str = is_res_lock ? "bpf_res_spin" : "bpf_spin";
8411	struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[regno];
8412	struct bpf_verifier_state *cur = env->cur_state;
8413	bool is_const = tnum_is_const(a: reg->var_off);
8414	bool is_irq = flags & PROCESS_LOCK_IRQ;
8415	u64 val = reg->var_off.value;
8416	struct bpf_map *map = NULL;
8417	struct btf *btf = NULL;
8418	struct btf_record *rec;
8419	u32 spin_lock_off;
8420	int err;
8421
8422	if (!is_const) {
8423	verbose(private_data: env,
8424	fmt: "R%d doesn't have constant offset. %s_lock has to be at the constant offset\n",
8425	regno, lock_str);
8426	return -EINVAL;
8427	}
8428	if (reg->type == PTR_TO_MAP_VALUE) {
8429	map = reg->map_ptr;
8430	if (!map->btf) {
8431	verbose(private_data: env,
8432	fmt: "map '%s' has to have BTF in order to use %s_lock\n",
8433	map->name, lock_str);
8434	return -EINVAL;
8435	}
8436	} else {
8437	btf = reg->btf;
8438	}
8439
8440	rec = reg_btf_record(reg);
8441	if (!btf_record_has_field(rec, type: is_res_lock ? BPF_RES_SPIN_LOCK : BPF_SPIN_LOCK)) {
8442	verbose(private_data: env, fmt: "%s '%s' has no valid %s_lock\n", map ? "map" : "local",
8443	map ? map->name : "kptr", lock_str);
8444	return -EINVAL;
8445	}
8446	spin_lock_off = is_res_lock ? rec->res_spin_lock_off : rec->spin_lock_off;
8447	if (spin_lock_off != val + reg->off) {
8448	verbose(private_data: env, fmt: "off %lld doesn't point to 'struct %s_lock' that is at %d\n",
8449	val + reg->off, lock_str, spin_lock_off);
8450	return -EINVAL;
8451	}
8452	if (is_lock) {
8453	void *ptr;
8454	int type;
8455
8456	if (map)
8457	ptr = map;
8458	else
8459	ptr = btf;
8460
8461	if (!is_res_lock && cur->active_locks) {
8462	if (find_lock_state(state: env->cur_state, type: REF_TYPE_LOCK, id: 0, NULL)) {
8463	verbose(private_data: env,
8464	fmt: "Locking two bpf_spin_locks are not allowed\n");
8465	return -EINVAL;
8466	}
8467	} else if (is_res_lock && cur->active_locks) {
8468	if (find_lock_state(state: env->cur_state, type: REF_TYPE_RES_LOCK | REF_TYPE_RES_LOCK_IRQ, id: reg->id, ptr)) {
8469	verbose(private_data: env, fmt: "Acquiring the same lock again, AA deadlock detected\n");
8470	return -EINVAL;
8471	}
8472	}
8473
8474	if (is_res_lock && is_irq)
8475	type = REF_TYPE_RES_LOCK_IRQ;
8476	else if (is_res_lock)
8477	type = REF_TYPE_RES_LOCK;
8478	else
8479	type = REF_TYPE_LOCK;
8480	err = acquire_lock_state(env, insn_idx: env->insn_idx, type, id: reg->id, ptr);
8481	if (err < 0) {
8482	verbose(private_data: env, fmt: "Failed to acquire lock state\n");
8483	return err;
8484	}
8485	} else {
8486	void *ptr;
8487	int type;
8488
8489	if (map)
8490	ptr = map;
8491	else
8492	ptr = btf;
8493
8494	if (!cur->active_locks) {
8495	verbose(private_data: env, fmt: "%s_unlock without taking a lock\n", lock_str);
8496	return -EINVAL;
8497	}
8498
8499	if (is_res_lock && is_irq)
8500	type = REF_TYPE_RES_LOCK_IRQ;
8501	else if (is_res_lock)
8502	type = REF_TYPE_RES_LOCK;
8503	else
8504	type = REF_TYPE_LOCK;
8505	if (!find_lock_state(state: cur, type, id: reg->id, ptr)) {
8506	verbose(private_data: env, fmt: "%s_unlock of different lock\n", lock_str);
8507	return -EINVAL;
8508	}
8509	if (reg->id != cur->active_lock_id || ptr != cur->active_lock_ptr) {
8510	verbose(private_data: env, fmt: "%s_unlock cannot be out of order\n", lock_str);
8511	return -EINVAL;
8512	}
8513	if (release_lock_state(state: cur, type, id: reg->id, ptr)) {
8514	verbose(private_data: env, fmt: "%s_unlock of different lock\n", lock_str);
8515	return -EINVAL;
8516	}
8517
8518	invalidate_non_owning_refs(env);
8519	}
8520	return 0;
8521	}
8522
8523	/* Check if @regno is a pointer to a specific field in a map value */
8524	static int check_map_field_pointer(struct bpf_verifier_env *env, u32 regno,
8525	enum btf_field_type field_type)
8526	{
8527	struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[regno];
8528	bool is_const = tnum_is_const(a: reg->var_off);
8529	struct bpf_map *map = reg->map_ptr;
8530	u64 val = reg->var_off.value;
8531	const char *struct_name = btf_field_type_name(type: field_type);
8532	int field_off = -1;
8533
8534	if (!is_const) {
8535	verbose(private_data: env,
8536	fmt: "R%d doesn't have constant offset. %s has to be at the constant offset\n",
8537	regno, struct_name);
8538	return -EINVAL;
8539	}
8540	if (!map->btf) {
8541	verbose(private_data: env, fmt: "map '%s' has to have BTF in order to use %s\n", map->name,
8542	struct_name);
8543	return -EINVAL;
8544	}
8545	if (!btf_record_has_field(rec: map->record, type: field_type)) {
8546	verbose(private_data: env, fmt: "map '%s' has no valid %s\n", map->name, struct_name);
8547	return -EINVAL;
8548	}
8549	switch (field_type) {
8550	case BPF_TIMER:
8551	field_off = map->record->timer_off;
8552	break;
8553	case BPF_TASK_WORK:
8554	field_off = map->record->task_work_off;
8555	break;
8556	case BPF_WORKQUEUE:
8557	field_off = map->record->wq_off;
8558	break;
8559	default:
8560	verifier_bug(env, "unsupported BTF field type: %s\n", struct_name);
8561	return -EINVAL;
8562	}
8563	if (field_off != val + reg->off) {
8564	verbose(private_data: env, fmt: "off %lld doesn't point to 'struct %s' that is at %d\n",
8565	val + reg->off, struct_name, field_off);
8566	return -EINVAL;
8567	}
8568	return 0;
8569	}
8570
8571	static int process_timer_func(struct bpf_verifier_env *env, int regno,
8572	struct bpf_call_arg_meta *meta)
8573	{
8574	struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[regno];
8575	struct bpf_map *map = reg->map_ptr;
8576	int err;
8577
8578	err = check_map_field_pointer(env, regno, field_type: BPF_TIMER);
8579	if (err)
8580	return err;
8581
8582	if (meta->map_ptr) {
8583	verifier_bug(env, "Two map pointers in a timer helper");
8584	return -EFAULT;
8585	}
8586	if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
8587	verbose(private_data: env, fmt: "bpf_timer cannot be used for PREEMPT_RT.\n");
8588	return -EOPNOTSUPP;
8589	}
8590	meta->map_uid = reg->map_uid;
8591	meta->map_ptr = map;
8592	return 0;
8593	}
8594
8595	static int process_wq_func(struct bpf_verifier_env *env, int regno,
8596	struct bpf_kfunc_call_arg_meta *meta)
8597	{
8598	struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[regno];
8599	struct bpf_map *map = reg->map_ptr;
8600	int err;
8601
8602	err = check_map_field_pointer(env, regno, field_type: BPF_WORKQUEUE);
8603	if (err)
8604	return err;
8605
8606	if (meta->map.ptr) {
8607	verifier_bug(env, "Two map pointers in a bpf_wq helper");
8608	return -EFAULT;
8609	}
8610
8611	meta->map.uid = reg->map_uid;
8612	meta->map.ptr = map;
8613	return 0;
8614	}
8615
8616	static int process_task_work_func(struct bpf_verifier_env *env, int regno,
8617	struct bpf_kfunc_call_arg_meta *meta)
8618	{
8619	struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[regno];
8620	struct bpf_map *map = reg->map_ptr;
8621	int err;
8622
8623	err = check_map_field_pointer(env, regno, field_type: BPF_TASK_WORK);
8624	if (err)
8625	return err;
8626
8627	if (meta->map.ptr) {
8628	verifier_bug(env, "Two map pointers in a bpf_task_work helper");
8629	return -EFAULT;
8630	}
8631	meta->map.uid = reg->map_uid;
8632	meta->map.ptr = map;
8633	return 0;
8634	}
8635
8636	static int process_kptr_func(struct bpf_verifier_env *env, int regno,
8637	struct bpf_call_arg_meta *meta)
8638	{
8639	struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[regno];
8640	struct btf_field *kptr_field;
8641	struct bpf_map *map_ptr;
8642	struct btf_record *rec;
8643	u32 kptr_off;
8644
8645	if (type_is_ptr_alloc_obj(type: reg->type)) {
8646	rec = reg_btf_record(reg);
8647	} else { /* PTR_TO_MAP_VALUE */
8648	map_ptr = reg->map_ptr;
8649	if (!map_ptr->btf) {
8650	verbose(private_data: env, fmt: "map '%s' has to have BTF in order to use bpf_kptr_xchg\n",
8651	map_ptr->name);
8652	return -EINVAL;
8653	}
8654	rec = map_ptr->record;
8655	meta->map_ptr = map_ptr;
8656	}
8657
8658	if (!tnum_is_const(a: reg->var_off)) {
8659	verbose(private_data: env,
8660	fmt: "R%d doesn't have constant offset. kptr has to be at the constant offset\n",
8661	regno);
8662	return -EINVAL;
8663	}
8664
8665	if (!btf_record_has_field(rec, type: BPF_KPTR)) {
8666	verbose(private_data: env, fmt: "R%d has no valid kptr\n", regno);
8667	return -EINVAL;
8668	}
8669
8670	kptr_off = reg->off + reg->var_off.value;
8671	kptr_field = btf_record_find(rec, offset: kptr_off, field_mask: BPF_KPTR);
8672	if (!kptr_field) {
8673	verbose(private_data: env, fmt: "off=%d doesn't point to kptr\n", kptr_off);
8674	return -EACCES;
8675	}
8676	if (kptr_field->type != BPF_KPTR_REF && kptr_field->type != BPF_KPTR_PERCPU) {
8677	verbose(private_data: env, fmt: "off=%d kptr isn't referenced kptr\n", kptr_off);
8678	return -EACCES;
8679	}
8680	meta->kptr_field = kptr_field;
8681	return 0;
8682	}
8683
8684	/* There are two register types representing a bpf_dynptr, one is PTR_TO_STACK
8685	* which points to a stack slot, and the other is CONST_PTR_TO_DYNPTR.
8686	*
8687	* In both cases we deal with the first 8 bytes, but need to mark the next 8
8688	* bytes as STACK_DYNPTR in case of PTR_TO_STACK. In case of
8689	* CONST_PTR_TO_DYNPTR, we are guaranteed to get the beginning of the object.
8690	*
8691	* Mutability of bpf_dynptr is at two levels, one is at the level of struct
8692	* bpf_dynptr itself, i.e. whether the helper is receiving a pointer to struct
8693	* bpf_dynptr or pointer to const struct bpf_dynptr. In the former case, it can
8694	* mutate the view of the dynptr and also possibly destroy it. In the latter
8695	* case, it cannot mutate the bpf_dynptr itself but it can still mutate the
8696	* memory that dynptr points to.
8697	*
8698	* The verifier will keep track both levels of mutation (bpf_dynptr's in
8699	* reg->type and the memory's in reg->dynptr.type), but there is no support for
8700	* readonly dynptr view yet, hence only the first case is tracked and checked.
8701	*
8702	* This is consistent with how C applies the const modifier to a struct object,
8703	* where the pointer itself inside bpf_dynptr becomes const but not what it
8704	* points to.
8705	*
8706	* Helpers which do not mutate the bpf_dynptr set MEM_RDONLY in their argument
8707	* type, and declare it as 'const struct bpf_dynptr *' in their prototype.
8708	*/
8709	static int process_dynptr_func(struct bpf_verifier_env *env, int regno, int insn_idx,
8710	enum bpf_arg_type arg_type, int clone_ref_obj_id)
8711	{
8712	struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[regno];
8713	int err;
8714
8715	if (reg->type != PTR_TO_STACK && reg->type != CONST_PTR_TO_DYNPTR) {
8716	verbose(private_data: env,
8717	fmt: "arg#%d expected pointer to stack or const struct bpf_dynptr\n",
8718	regno - 1);
8719	return -EINVAL;
8720	}
8721
8722	/* MEM_UNINIT and MEM_RDONLY are exclusive, when applied to an
8723	* ARG_PTR_TO_DYNPTR (or ARG_PTR_TO_DYNPTR | DYNPTR_TYPE_*):
8724	*/
8725	if ((arg_type & (MEM_UNINIT | MEM_RDONLY)) == (MEM_UNINIT | MEM_RDONLY)) {
8726	verifier_bug(env, "misconfigured dynptr helper type flags");
8727	return -EFAULT;
8728	}
8729
8730	/* MEM_UNINIT - Points to memory that is an appropriate candidate for
8731	* constructing a mutable bpf_dynptr object.
8732	*
8733	* Currently, this is only possible with PTR_TO_STACK
8734	* pointing to a region of at least 16 bytes which doesn't
8735	* contain an existing bpf_dynptr.
8736	*
8737	* MEM_RDONLY - Points to a initialized bpf_dynptr that will not be
8738	* mutated or destroyed. However, the memory it points to
8739	* may be mutated.
8740	*
8741	* None - Points to a initialized dynptr that can be mutated and
8742	* destroyed, including mutation of the memory it points
8743	* to.
8744	*/
8745	if (arg_type & MEM_UNINIT) {
8746	int i;
8747
8748	if (!is_dynptr_reg_valid_uninit(env, reg)) {
8749	verbose(private_data: env, fmt: "Dynptr has to be an uninitialized dynptr\n");
8750	return -EINVAL;
8751	}
8752
8753	/* we write BPF_DW bits (8 bytes) at a time */
8754	for (i = 0; i < BPF_DYNPTR_SIZE; i += 8) {
8755	err = check_mem_access(env, insn_idx, regno,
8756	off: i, BPF_DW, t: BPF_WRITE, value_regno: -1, strict_alignment_once: false, is_ldsx: false);
8757	if (err)
8758	return err;
8759	}
8760
8761	err = mark_stack_slots_dynptr(env, reg, arg_type, insn_idx, clone_ref_obj_id);
8762	} else /* MEM_RDONLY and None case from above */ {
8763	/* For the reg->type == PTR_TO_STACK case, bpf_dynptr is never const */
8764	if (reg->type == CONST_PTR_TO_DYNPTR && !(arg_type & MEM_RDONLY)) {
8765	verbose(private_data: env, fmt: "cannot pass pointer to const bpf_dynptr, the helper mutates it\n");
8766	return -EINVAL;
8767	}
8768
8769	if (!is_dynptr_reg_valid_init(env, reg)) {
8770	verbose(private_data: env,
8771	fmt: "Expected an initialized dynptr as arg #%d\n",
8772	regno - 1);
8773	return -EINVAL;
8774	}
8775
8776	/* Fold modifiers (in this case, MEM_RDONLY) when checking expected type */
8777	if (!is_dynptr_type_expected(env, reg, arg_type: arg_type & ~MEM_RDONLY)) {
8778	verbose(private_data: env,
8779	fmt: "Expected a dynptr of type %s as arg #%d\n",
8780	dynptr_type_str(type: arg_to_dynptr_type(arg_type)), regno - 1);
8781	return -EINVAL;
8782	}
8783
8784	err = mark_dynptr_read(env, reg);
8785	}
8786	return err;
8787	}
8788
8789	static u32 iter_ref_obj_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg, int spi)
8790	{
8791	struct bpf_func_state *state = func(env, reg);
8792
8793	return state->stack[spi].spilled_ptr.ref_obj_id;
8794	}
8795
8796	static bool is_iter_kfunc(struct bpf_kfunc_call_arg_meta *meta)
8797	{
8798	return meta->kfunc_flags & (KF_ITER_NEW | KF_ITER_NEXT | KF_ITER_DESTROY);
8799	}
8800
8801	static bool is_iter_new_kfunc(struct bpf_kfunc_call_arg_meta *meta)
8802	{
8803	return meta->kfunc_flags & KF_ITER_NEW;
8804	}
8805
8806	static bool is_iter_next_kfunc(struct bpf_kfunc_call_arg_meta *meta)
8807	{
8808	return meta->kfunc_flags & KF_ITER_NEXT;
8809	}
8810
8811	static bool is_iter_destroy_kfunc(struct bpf_kfunc_call_arg_meta *meta)
8812	{
8813	return meta->kfunc_flags & KF_ITER_DESTROY;
8814	}
8815
8816	static bool is_kfunc_arg_iter(struct bpf_kfunc_call_arg_meta *meta, int arg_idx,
8817	const struct btf_param *arg)
8818	{
8819	/* btf_check_iter_kfuncs() guarantees that first argument of any iter
8820	* kfunc is iter state pointer
8821	*/
8822	if (is_iter_kfunc(meta))
8823	return arg_idx == 0;
8824
8825	/* iter passed as an argument to a generic kfunc */
8826	return btf_param_match_suffix(btf: meta->btf, arg, suffix: "__iter");
8827	}
8828
8829	static int process_iter_arg(struct bpf_verifier_env *env, int regno, int insn_idx,
8830	struct bpf_kfunc_call_arg_meta *meta)
8831	{
8832	struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[regno];
8833	const struct btf_type *t;
8834	int spi, err, i, nr_slots, btf_id;
8835
8836	if (reg->type != PTR_TO_STACK) {
8837	verbose(private_data: env, fmt: "arg#%d expected pointer to an iterator on stack\n", regno - 1);
8838	return -EINVAL;
8839	}
8840
8841	/* For iter_{new,next,destroy} functions, btf_check_iter_kfuncs()
8842	* ensures struct convention, so we wouldn't need to do any BTF
8843	* validation here. But given iter state can be passed as a parameter
8844	* to any kfunc, if arg has "__iter" suffix, we need to be a bit more
8845	* conservative here.
8846	*/
8847	btf_id = btf_check_iter_arg(btf: meta->btf, func: meta->func_proto, arg_idx: regno - 1);
8848	if (btf_id < 0) {
8849	verbose(private_data: env, fmt: "expected valid iter pointer as arg #%d\n", regno - 1);
8850	return -EINVAL;
8851	}
8852	t = btf_type_by_id(btf: meta->btf, type_id: btf_id);
8853	nr_slots = t->size / BPF_REG_SIZE;
8854
8855	if (is_iter_new_kfunc(meta)) {
8856	/* bpf_iter_<type>_new() expects pointer to uninit iter state */
8857	if (!is_iter_reg_valid_uninit(env, reg, nr_slots)) {
8858	verbose(private_data: env, fmt: "expected uninitialized iter_%s as arg #%d\n",
8859	iter_type_str(btf: meta->btf, btf_id), regno - 1);
8860	return -EINVAL;
8861	}
8862
8863	for (i = 0; i < nr_slots * 8; i += BPF_REG_SIZE) {
8864	err = check_mem_access(env, insn_idx, regno,
8865	off: i, BPF_DW, t: BPF_WRITE, value_regno: -1, strict_alignment_once: false, is_ldsx: false);
8866	if (err)
8867	return err;
8868	}
8869
8870	err = mark_stack_slots_iter(env, meta, reg, insn_idx, btf: meta->btf, btf_id, nr_slots);
8871	if (err)
8872	return err;
8873	} else {
8874	/* iter_next() or iter_destroy(), as well as any kfunc
8875	* accepting iter argument, expect initialized iter state
8876	*/
8877	err = is_iter_reg_valid_init(env, reg, btf: meta->btf, btf_id, nr_slots);
8878	switch (err) {
8879	case 0:
8880	break;
8881	case -EINVAL:
8882	verbose(private_data: env, fmt: "expected an initialized iter_%s as arg #%d\n",
8883	iter_type_str(btf: meta->btf, btf_id), regno - 1);
8884	return err;
8885	case -EPROTO:
8886	verbose(private_data: env, fmt: "expected an RCU CS when using %s\n", meta->func_name);
8887	return err;
8888	default:
8889	return err;
8890	}
8891
8892	spi = iter_get_spi(env, reg, nr_slots);
8893	if (spi < 0)
8894	return spi;
8895
8896	err = mark_iter_read(env, reg, spi, nr_slots);
8897	if (err)
8898	return err;
8899
8900	/* remember meta->iter info for process_iter_next_call() */
8901	meta->iter.spi = spi;
8902	meta->iter.frameno = reg->frameno;
8903	meta->ref_obj_id = iter_ref_obj_id(env, reg, spi);
8904
8905	if (is_iter_destroy_kfunc(meta)) {
8906	err = unmark_stack_slots_iter(env, reg, nr_slots);
8907	if (err)
8908	return err;
8909	}
8910	}
8911
8912	return 0;
8913	}
8914
8915	/* Look for a previous loop entry at insn_idx: nearest parent state
8916	* stopped at insn_idx with callsites matching those in cur->frame.
8917	*/
8918	static struct bpf_verifier_state *find_prev_entry(struct bpf_verifier_env *env,
8919	struct bpf_verifier_state *cur,
8920	int insn_idx)
8921	{
8922	struct bpf_verifier_state_list *sl;
8923	struct bpf_verifier_state *st;
8924	struct list_head *pos, *head;
8925
8926	/* Explored states are pushed in stack order, most recent states come first */
8927	head = explored_state(env, idx: insn_idx);
8928	list_for_each(pos, head) {
8929	sl = container_of(pos, struct bpf_verifier_state_list, node);
8930	/* If st->branches != 0 state is a part of current DFS verification path,
8931	* hence cur & st for a loop.
8932	*/
8933	st = &sl->state;
8934	if (st->insn_idx == insn_idx && st->branches && same_callsites(a: st, b: cur) &&
8935	st->dfs_depth < cur->dfs_depth)
8936	return st;
8937	}
8938
8939	return NULL;
8940	}
8941
8942	static void reset_idmap_scratch(struct bpf_verifier_env *env);
8943	static bool regs_exact(const struct bpf_reg_state *rold,
8944	const struct bpf_reg_state *rcur,
8945	struct bpf_idmap *idmap);
8946
8947	static void maybe_widen_reg(struct bpf_verifier_env *env,
8948	struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
8949	struct bpf_idmap *idmap)
8950	{
8951	if (rold->type != SCALAR_VALUE)
8952	return;
8953	if (rold->type != rcur->type)
8954	return;
8955	if (rold->precise || rcur->precise || regs_exact(rold, rcur, idmap))
8956	return;
8957	__mark_reg_unknown(env, reg: rcur);
8958	}
8959
8960	static int widen_imprecise_scalars(struct bpf_verifier_env *env,
8961	struct bpf_verifier_state *old,
8962	struct bpf_verifier_state *cur)
8963	{
8964	struct bpf_func_state *fold, *fcur;
8965	int i, fr, num_slots;
8966
8967	reset_idmap_scratch(env);
8968	for (fr = old->curframe; fr >= 0; fr--) {
8969	fold = old->frame[fr];
8970	fcur = cur->frame[fr];
8971
8972	for (i = 0; i < MAX_BPF_REG; i++)
8973	maybe_widen_reg(env,
8974	rold: &fold->regs[i],
8975	rcur: &fcur->regs[i],
8976	idmap: &env->idmap_scratch);
8977
8978	num_slots = min(fold->allocated_stack / BPF_REG_SIZE,
8979	fcur->allocated_stack / BPF_REG_SIZE);
8980	for (i = 0; i < num_slots; i++) {
8981	if (!is_spilled_reg(stack: &fold->stack[i]) ||
8982	!is_spilled_reg(stack: &fcur->stack[i]))
8983	continue;
8984
8985	maybe_widen_reg(env,
8986	rold: &fold->stack[i].spilled_ptr,
8987	rcur: &fcur->stack[i].spilled_ptr,
8988	idmap: &env->idmap_scratch);
8989	}
8990	}
8991	return 0;
8992	}
8993
8994	static struct bpf_reg_state *get_iter_from_state(struct bpf_verifier_state *cur_st,
8995	struct bpf_kfunc_call_arg_meta *meta)
8996	{
8997	int iter_frameno = meta->iter.frameno;
8998	int iter_spi = meta->iter.spi;
8999
9000	return &cur_st->frame[iter_frameno]->stack[iter_spi].spilled_ptr;
9001	}
9002
9003	/* process_iter_next_call() is called when verifier gets to iterator's next
9004	* "method" (e.g., bpf_iter_num_next() for numbers iterator) call. We'll refer
9005	* to it as just "iter_next()" in comments below.
9006	*
9007	* BPF verifier relies on a crucial contract for any iter_next()
9008	* implementation: it should *eventually* return NULL, and once that happens
9009	* it should keep returning NULL. That is, once iterator exhausts elements to
9010	* iterate, it should never reset or spuriously return new elements.
9011	*
9012	* With the assumption of such contract, process_iter_next_call() simulates
9013	* a fork in the verifier state to validate loop logic correctness and safety
9014	* without having to simulate infinite amount of iterations.
9015	*
9016	* In current state, we first assume that iter_next() returned NULL and
9017	* iterator state is set to DRAINED (BPF_ITER_STATE_DRAINED). In such
9018	* conditions we should not form an infinite loop and should eventually reach
9019	* exit.
9020	*
9021	* Besides that, we also fork current state and enqueue it for later
9022	* verification. In a forked state we keep iterator state as ACTIVE
9023	* (BPF_ITER_STATE_ACTIVE) and assume non-NULL return from iter_next(). We
9024	* also bump iteration depth to prevent erroneous infinite loop detection
9025	* later on (see iter_active_depths_differ() comment for details). In this
9026	* state we assume that we'll eventually loop back to another iter_next()
9027	* calls (it could be in exactly same location or in some other instruction,
9028	* it doesn't matter, we don't make any unnecessary assumptions about this,
9029	* everything revolves around iterator state in a stack slot, not which
9030	* instruction is calling iter_next()). When that happens, we either will come
9031	* to iter_next() with equivalent state and can conclude that next iteration
9032	* will proceed in exactly the same way as we just verified, so it's safe to
9033	* assume that loop converges. If not, we'll go on another iteration
9034	* simulation with a different input state, until all possible starting states
9035	* are validated or we reach maximum number of instructions limit.
9036	*
9037	* This way, we will either exhaustively discover all possible input states
9038	* that iterator loop can start with and eventually will converge, or we'll
9039	* effectively regress into bounded loop simulation logic and either reach
9040	* maximum number of instructions if loop is not provably convergent, or there
9041	* is some statically known limit on number of iterations (e.g., if there is
9042	* an explicit `if n > 100 then break;` statement somewhere in the loop).
9043	*
9044	* Iteration convergence logic in is_state_visited() relies on exact
9045	* states comparison, which ignores read and precision marks.
9046	* This is necessary because read and precision marks are not finalized
9047	* while in the loop. Exact comparison might preclude convergence for
9048	* simple programs like below:
9049	*
9050	* i = 0;
9051	* while(iter_next(&it))
9052	* i++;
9053	*
9054	* At each iteration step i++ would produce a new distinct state and
9055	* eventually instruction processing limit would be reached.
9056	*
9057	* To avoid such behavior speculatively forget (widen) range for
9058	* imprecise scalar registers, if those registers were not precise at the
9059	* end of the previous iteration and do not match exactly.
9060	*
9061	* This is a conservative heuristic that allows to verify wide range of programs,
9062	* however it precludes verification of programs that conjure an
9063	* imprecise value on the first loop iteration and use it as precise on a second.
9064	* For example, the following safe program would fail to verify:
9065	*
9066	* struct bpf_num_iter it;
9067	* int arr[10];
9068	* int i = 0, a = 0;
9069	* bpf_iter_num_new(&it, 0, 10);
9070	* while (bpf_iter_num_next(&it)) {
9071	* if (a == 0) {
9072	* a = 1;
9073	* i = 7; // Because i changed verifier would forget
9074	* // it's range on second loop entry.
9075	* } else {
9076	* arr[i] = 42; // This would fail to verify.
9077	* }
9078	* }
9079	* bpf_iter_num_destroy(&it);
9080	*/
9081	static int process_iter_next_call(struct bpf_verifier_env *env, int insn_idx,
9082	struct bpf_kfunc_call_arg_meta *meta)
9083	{
9084	struct bpf_verifier_state *cur_st = env->cur_state, *queued_st, *prev_st;
9085	struct bpf_func_state *cur_fr = cur_st->frame[cur_st->curframe], *queued_fr;
9086	struct bpf_reg_state *cur_iter, *queued_iter;
9087
9088	BTF_TYPE_EMIT(struct bpf_iter);
9089
9090	cur_iter = get_iter_from_state(cur_st, meta);
9091
9092	if (cur_iter->iter.state != BPF_ITER_STATE_ACTIVE &&
9093	cur_iter->iter.state != BPF_ITER_STATE_DRAINED) {
9094	verifier_bug(env, "unexpected iterator state %d (%s)",
9095	cur_iter->iter.state, iter_state_str(cur_iter->iter.state));
9096	return -EFAULT;
9097	}
9098
9099	if (cur_iter->iter.state == BPF_ITER_STATE_ACTIVE) {
9100	/* Because iter_next() call is a checkpoint is_state_visitied()
9101	* should guarantee parent state with same call sites and insn_idx.
9102	*/
9103	if (!cur_st->parent || cur_st->parent->insn_idx != insn_idx ||
9104	!same_callsites(a: cur_st->parent, b: cur_st)) {
9105	verifier_bug(env, "bad parent state for iter next call");
9106	return -EFAULT;
9107	}
9108	/* Note cur_st->parent in the call below, it is necessary to skip
9109	* checkpoint created for cur_st by is_state_visited()
9110	* right at this instruction.
9111	*/
9112	prev_st = find_prev_entry(env, cur: cur_st->parent, insn_idx);
9113	/* branch out active iter state */
9114	queued_st = push_stack(env, insn_idx: insn_idx + 1, prev_insn_idx: insn_idx, speculative: false);
9115	if (IS_ERR(ptr: queued_st))
9116	return PTR_ERR(ptr: queued_st);
9117
9118	queued_iter = get_iter_from_state(cur_st: queued_st, meta);
9119	queued_iter->iter.state = BPF_ITER_STATE_ACTIVE;
9120	queued_iter->iter.depth++;
9121	if (prev_st)
9122	widen_imprecise_scalars(env, old: prev_st, cur: queued_st);
9123
9124	queued_fr = queued_st->frame[queued_st->curframe];
9125	mark_ptr_not_null_reg(reg: &queued_fr->regs[BPF_REG_0]);
9126	}
9127
9128	/* switch to DRAINED state, but keep the depth unchanged */
9129	/* mark current iter state as drained and assume returned NULL */
9130	cur_iter->iter.state = BPF_ITER_STATE_DRAINED;
9131	__mark_reg_const_zero(env, reg: &cur_fr->regs[BPF_REG_0]);
9132
9133	return 0;
9134	}
9135
9136	static bool arg_type_is_mem_size(enum bpf_arg_type type)
9137	{
9138	return type == ARG_CONST_SIZE ||
9139	type == ARG_CONST_SIZE_OR_ZERO;
9140	}
9141
9142	static bool arg_type_is_raw_mem(enum bpf_arg_type type)
9143	{
9144	return base_type(type) == ARG_PTR_TO_MEM &&
9145	type & MEM_UNINIT;
9146	}
9147
9148	static bool arg_type_is_release(enum bpf_arg_type type)
9149	{
9150	return type & OBJ_RELEASE;
9151	}
9152
9153	static bool arg_type_is_dynptr(enum bpf_arg_type type)
9154	{
9155	return base_type(type) == ARG_PTR_TO_DYNPTR;
9156	}
9157
9158	static int resolve_map_arg_type(struct bpf_verifier_env *env,
9159	const struct bpf_call_arg_meta *meta,
9160	enum bpf_arg_type *arg_type)
9161	{
9162	if (!meta->map_ptr) {
9163	/* kernel subsystem misconfigured verifier */
9164	verifier_bug(env, "invalid map_ptr to access map->type");
9165	return -EFAULT;
9166	}
9167
9168	switch (meta->map_ptr->map_type) {
9169	case BPF_MAP_TYPE_SOCKMAP:
9170	case BPF_MAP_TYPE_SOCKHASH:
9171	if (*arg_type == ARG_PTR_TO_MAP_VALUE) {
9172	*arg_type = ARG_PTR_TO_BTF_ID_SOCK_COMMON;
9173	} else {
9174	verbose(private_data: env, fmt: "invalid arg_type for sockmap/sockhash\n");
9175	return -EINVAL;
9176	}
9177	break;
9178	case BPF_MAP_TYPE_BLOOM_FILTER:
9179	if (meta->func_id == BPF_FUNC_map_peek_elem)
9180	*arg_type = ARG_PTR_TO_MAP_VALUE;
9181	break;
9182	default:
9183	break;
9184	}
9185	return 0;
9186	}
9187
9188	struct bpf_reg_types {
9189	const enum bpf_reg_type types[10];
9190	u32 *btf_id;
9191	};
9192
9193	static const struct bpf_reg_types sock_types = {
9194	.types = {
9195	PTR_TO_SOCK_COMMON,
9196	PTR_TO_SOCKET,
9197	PTR_TO_TCP_SOCK,
9198	PTR_TO_XDP_SOCK,
9199	},
9200	};
9201
9202	#ifdef CONFIG_NET
9203	static const struct bpf_reg_types btf_id_sock_common_types = {
9204	.types = {
9205	PTR_TO_SOCK_COMMON,
9206	PTR_TO_SOCKET,
9207	PTR_TO_TCP_SOCK,
9208	PTR_TO_XDP_SOCK,
9209	PTR_TO_BTF_ID,
9210	PTR_TO_BTF_ID | PTR_TRUSTED,
9211	},
9212	.btf_id = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
9213	};
9214	#endif
9215
9216	static const struct bpf_reg_types mem_types = {
9217	.types = {
9218	PTR_TO_STACK,
9219	PTR_TO_PACKET,
9220	PTR_TO_PACKET_META,
9221	PTR_TO_MAP_KEY,
9222	PTR_TO_MAP_VALUE,
9223	PTR_TO_MEM,
9224	PTR_TO_MEM | MEM_RINGBUF,
9225	PTR_TO_BUF,
9226	PTR_TO_BTF_ID | PTR_TRUSTED,
9227	},
9228	};
9229
9230	static const struct bpf_reg_types spin_lock_types = {
9231	.types = {
9232	PTR_TO_MAP_VALUE,
9233	PTR_TO_BTF_ID | MEM_ALLOC,
9234	}
9235	};
9236
9237	static const struct bpf_reg_types fullsock_types = { .types = { PTR_TO_SOCKET } };
9238	static const struct bpf_reg_types scalar_types = { .types = { SCALAR_VALUE } };
9239	static const struct bpf_reg_types context_types = { .types = { PTR_TO_CTX } };
9240	static const struct bpf_reg_types ringbuf_mem_types = { .types = { PTR_TO_MEM | MEM_RINGBUF } };
9241	static const struct bpf_reg_types const_map_ptr_types = { .types = { CONST_PTR_TO_MAP } };
9242	static const struct bpf_reg_types btf_ptr_types = {
9243	.types = {
9244	PTR_TO_BTF_ID,
9245	PTR_TO_BTF_ID | PTR_TRUSTED,
9246	PTR_TO_BTF_ID | MEM_RCU,
9247	},
9248	};
9249	static const struct bpf_reg_types percpu_btf_ptr_types = {
9250	.types = {
9251	PTR_TO_BTF_ID | MEM_PERCPU,
9252	PTR_TO_BTF_ID | MEM_PERCPU | MEM_RCU,
9253	PTR_TO_BTF_ID | MEM_PERCPU | PTR_TRUSTED,
9254	}
9255	};
9256	static const struct bpf_reg_types func_ptr_types = { .types = { PTR_TO_FUNC } };
9257	static const struct bpf_reg_types stack_ptr_types = { .types = { PTR_TO_STACK } };
9258	static const struct bpf_reg_types const_str_ptr_types = { .types = { PTR_TO_MAP_VALUE } };
9259	static const struct bpf_reg_types timer_types = { .types = { PTR_TO_MAP_VALUE } };
9260	static const struct bpf_reg_types kptr_xchg_dest_types = {
9261	.types = {
9262	PTR_TO_MAP_VALUE,
9263	PTR_TO_BTF_ID | MEM_ALLOC
9264	}
9265	};
9266	static const struct bpf_reg_types dynptr_types = {
9267	.types = {
9268	PTR_TO_STACK,
9269	CONST_PTR_TO_DYNPTR,
9270	}
9271	};
9272
9273	static const struct bpf_reg_types *compatible_reg_types[__BPF_ARG_TYPE_MAX] = {
9274	[ARG_PTR_TO_MAP_KEY] = &mem_types,
9275	[ARG_PTR_TO_MAP_VALUE] = &mem_types,
9276	[ARG_CONST_SIZE] = &scalar_types,
9277	[ARG_CONST_SIZE_OR_ZERO] = &scalar_types,
9278	[ARG_CONST_ALLOC_SIZE_OR_ZERO] = &scalar_types,
9279	[ARG_CONST_MAP_PTR] = &const_map_ptr_types,
9280	[ARG_PTR_TO_CTX] = &context_types,
9281	[ARG_PTR_TO_SOCK_COMMON] = &sock_types,
9282	#ifdef CONFIG_NET
9283	[ARG_PTR_TO_BTF_ID_SOCK_COMMON] = &btf_id_sock_common_types,
9284	#endif
9285	[ARG_PTR_TO_SOCKET] = &fullsock_types,
9286	[ARG_PTR_TO_BTF_ID] = &btf_ptr_types,
9287	[ARG_PTR_TO_SPIN_LOCK] = &spin_lock_types,
9288	[ARG_PTR_TO_MEM] = &mem_types,
9289	[ARG_PTR_TO_RINGBUF_MEM] = &ringbuf_mem_types,
9290	[ARG_PTR_TO_PERCPU_BTF_ID] = &percpu_btf_ptr_types,
9291	[ARG_PTR_TO_FUNC] = &func_ptr_types,
9292	[ARG_PTR_TO_STACK] = &stack_ptr_types,
9293	[ARG_PTR_TO_CONST_STR] = &const_str_ptr_types,
9294	[ARG_PTR_TO_TIMER] = &timer_types,
9295	[ARG_KPTR_XCHG_DEST] = &kptr_xchg_dest_types,
9296	[ARG_PTR_TO_DYNPTR] = &dynptr_types,
9297	};
9298
9299	static int check_reg_type(struct bpf_verifier_env *env, u32 regno,
9300	enum bpf_arg_type arg_type,
9301	const u32 *arg_btf_id,
9302	struct bpf_call_arg_meta *meta)
9303	{
9304	struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[regno];
9305	enum bpf_reg_type expected, type = reg->type;
9306	const struct bpf_reg_types *compatible;
9307	int i, j;
9308
9309	compatible = compatible_reg_types[base_type(type: arg_type)];
9310	if (!compatible) {
9311	verifier_bug(env, "unsupported arg type %d", arg_type);
9312	return -EFAULT;
9313	}
9314
9315	/* ARG_PTR_TO_MEM + RDONLY is compatible with PTR_TO_MEM and PTR_TO_MEM + RDONLY,
9316	* but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM and NOT with PTR_TO_MEM + RDONLY
9317	*
9318	* Same for MAYBE_NULL:
9319	*
9320	* ARG_PTR_TO_MEM + MAYBE_NULL is compatible with PTR_TO_MEM and PTR_TO_MEM + MAYBE_NULL,
9321	* but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM but NOT with PTR_TO_MEM + MAYBE_NULL
9322	*
9323	* ARG_PTR_TO_MEM is compatible with PTR_TO_MEM that is tagged with a dynptr type.
9324	*
9325	* Therefore we fold these flags depending on the arg_type before comparison.
9326	*/
9327	if (arg_type & MEM_RDONLY)
9328	type &= ~MEM_RDONLY;
9329	if (arg_type & PTR_MAYBE_NULL)
9330	type &= ~PTR_MAYBE_NULL;
9331	if (base_type(type: arg_type) == ARG_PTR_TO_MEM)
9332	type &= ~DYNPTR_TYPE_FLAG_MASK;
9333
9334	/* Local kptr types are allowed as the source argument of bpf_kptr_xchg */
9335	if (meta->func_id == BPF_FUNC_kptr_xchg && type_is_alloc(type) && regno == BPF_REG_2) {
9336	type &= ~MEM_ALLOC;
9337	type &= ~MEM_PERCPU;
9338	}
9339
9340	for (i = 0; i < ARRAY_SIZE(compatible->types); i++) {
9341	expected = compatible->types[i];
9342	if (expected == NOT_INIT)
9343	break;
9344
9345	if (type == expected)
9346	goto found;
9347	}
9348
9349	verbose(private_data: env, fmt: "R%d type=%s expected=", regno, reg_type_str(env, type: reg->type));
9350	for (j = 0; j + 1 < i; j++)
9351	verbose(private_data: env, fmt: "%s, ", reg_type_str(env, type: compatible->types[j]));
9352	verbose(private_data: env, fmt: "%s\n", reg_type_str(env, type: compatible->types[j]));
9353	return -EACCES;
9354
9355	found:
9356	if (base_type(type: reg->type) != PTR_TO_BTF_ID)
9357	return 0;
9358
9359	if (compatible == &mem_types) {
9360	if (!(arg_type & MEM_RDONLY)) {
9361	verbose(private_data: env,
9362	fmt: "%s() may write into memory pointed by R%d type=%s\n",
9363	func_id_name(id: meta->func_id),
9364	regno, reg_type_str(env, type: reg->type));
9365	return -EACCES;
9366	}
9367	return 0;
9368	}
9369
9370	switch ((int)reg->type) {
9371	case PTR_TO_BTF_ID:
9372	case PTR_TO_BTF_ID | PTR_TRUSTED:
9373	case PTR_TO_BTF_ID | PTR_TRUSTED | PTR_MAYBE_NULL:
9374	case PTR_TO_BTF_ID | MEM_RCU:
9375	case PTR_TO_BTF_ID | PTR_MAYBE_NULL:
9376	case PTR_TO_BTF_ID | PTR_MAYBE_NULL | MEM_RCU:
9377	{
9378	/* For bpf_sk_release, it needs to match against first member
9379	* 'struct sock_common', hence make an exception for it. This
9380	* allows bpf_sk_release to work for multiple socket types.
9381	*/
9382	bool strict_type_match = arg_type_is_release(type: arg_type) &&
9383	meta->func_id != BPF_FUNC_sk_release;
9384
9385	if (type_may_be_null(type: reg->type) &&
9386	(!type_may_be_null(type: arg_type) || arg_type_is_release(type: arg_type))) {
9387	verbose(private_data: env, fmt: "Possibly NULL pointer passed to helper arg%d\n", regno);
9388	return -EACCES;
9389	}
9390
9391	if (!arg_btf_id) {
9392	if (!compatible->btf_id) {
9393	verifier_bug(env, "missing arg compatible BTF ID");
9394	return -EFAULT;
9395	}
9396	arg_btf_id = compatible->btf_id;
9397	}
9398
9399	if (meta->func_id == BPF_FUNC_kptr_xchg) {
9400	if (map_kptr_match_type(env, kptr_field: meta->kptr_field, reg, regno))
9401	return -EACCES;
9402	} else {
9403	if (arg_btf_id == BPF_PTR_POISON) {
9404	verbose(private_data: env, fmt: "verifier internal error:");
9405	verbose(private_data: env, fmt: "R%d has non-overwritten BPF_PTR_POISON type\n",
9406	regno);
9407	return -EACCES;
9408	}
9409
9410	if (!btf_struct_ids_match(log: &env->log, btf: reg->btf, id: reg->btf_id, off: reg->off,
9411	need_btf: btf_vmlinux, need_type_id: *arg_btf_id,
9412	strict: strict_type_match)) {
9413	verbose(private_data: env, fmt: "R%d is of type %s but %s is expected\n",
9414	regno, btf_type_name(btf: reg->btf, id: reg->btf_id),
9415	btf_type_name(btf: btf_vmlinux, id: *arg_btf_id));
9416	return -EACCES;
9417	}
9418	}
9419	break;
9420	}
9421	case PTR_TO_BTF_ID | MEM_ALLOC:
9422	case PTR_TO_BTF_ID | MEM_PERCPU | MEM_ALLOC:
9423	if (meta->func_id != BPF_FUNC_spin_lock && meta->func_id != BPF_FUNC_spin_unlock &&
9424	meta->func_id != BPF_FUNC_kptr_xchg) {
9425	verifier_bug(env, "unimplemented handling of MEM_ALLOC");
9426	return -EFAULT;
9427	}
9428	/* Check if local kptr in src arg matches kptr in dst arg */
9429	if (meta->func_id == BPF_FUNC_kptr_xchg && regno == BPF_REG_2) {
9430	if (map_kptr_match_type(env, kptr_field: meta->kptr_field, reg, regno))
9431	return -EACCES;
9432	}
9433	break;
9434	case PTR_TO_BTF_ID | MEM_PERCPU:
9435	case PTR_TO_BTF_ID | MEM_PERCPU | MEM_RCU:
9436	case PTR_TO_BTF_ID | MEM_PERCPU | PTR_TRUSTED:
9437	/* Handled by helper specific checks */
9438	break;
9439	default:
9440	verifier_bug(env, "invalid PTR_TO_BTF_ID register for type match");
9441	return -EFAULT;
9442	}
9443	return 0;
9444	}
9445
9446	static struct btf_field *
9447	reg_find_field_offset(const struct bpf_reg_state *reg, s32 off, u32 fields)
9448	{
9449	struct btf_field *field;
9450	struct btf_record *rec;
9451
9452	rec = reg_btf_record(reg);
9453	if (!rec)
9454	return NULL;
9455
9456	field = btf_record_find(rec, offset: off, field_mask: fields);
9457	if (!field)
9458	return NULL;
9459
9460	return field;
9461	}
9462
9463	static int check_func_arg_reg_off(struct bpf_verifier_env *env,
9464	const struct bpf_reg_state *reg, int regno,
9465	enum bpf_arg_type arg_type)
9466	{
9467	u32 type = reg->type;
9468
9469	/* When referenced register is passed to release function, its fixed
9470	* offset must be 0.
9471	*
9472	* We will check arg_type_is_release reg has ref_obj_id when storing
9473	* meta->release_regno.
9474	*/
9475	if (arg_type_is_release(type: arg_type)) {
9476	/* ARG_PTR_TO_DYNPTR with OBJ_RELEASE is a bit special, as it
9477	* may not directly point to the object being released, but to
9478	* dynptr pointing to such object, which might be at some offset
9479	* on the stack. In that case, we simply to fallback to the
9480	* default handling.
9481	*/
9482	if (arg_type_is_dynptr(type: arg_type) && type == PTR_TO_STACK)
9483	return 0;
9484
9485	/* Doing check_ptr_off_reg check for the offset will catch this
9486	* because fixed_off_ok is false, but checking here allows us
9487	* to give the user a better error message.
9488	*/
9489	if (reg->off) {
9490	verbose(private_data: env, fmt: "R%d must have zero offset when passed to release func or trusted arg to kfunc\n",
9491	regno);
9492	return -EINVAL;
9493	}
9494	return __check_ptr_off_reg(env, reg, regno, fixed_off_ok: false);
9495	}
9496
9497	switch (type) {
9498	/* Pointer types where both fixed and variable offset is explicitly allowed: */
9499	case PTR_TO_STACK:
9500	case PTR_TO_PACKET:
9501	case PTR_TO_PACKET_META:
9502	case PTR_TO_MAP_KEY:
9503	case PTR_TO_MAP_VALUE:
9504	case PTR_TO_MEM:
9505	case PTR_TO_MEM | MEM_RDONLY:
9506	case PTR_TO_MEM | MEM_RINGBUF:
9507	case PTR_TO_BUF:
9508	case PTR_TO_BUF | MEM_RDONLY:
9509	case PTR_TO_ARENA:
9510	case SCALAR_VALUE:
9511	return 0;
9512	/* All the rest must be rejected, except PTR_TO_BTF_ID which allows
9513	* fixed offset.
9514	*/
9515	case PTR_TO_BTF_ID:
9516	case PTR_TO_BTF_ID | MEM_ALLOC:
9517	case PTR_TO_BTF_ID | PTR_TRUSTED:
9518	case PTR_TO_BTF_ID | MEM_RCU:
9519	case PTR_TO_BTF_ID | MEM_ALLOC | NON_OWN_REF:
9520	case PTR_TO_BTF_ID | MEM_ALLOC | NON_OWN_REF | MEM_RCU:
9521	/* When referenced PTR_TO_BTF_ID is passed to release function,
9522	* its fixed offset must be 0. In the other cases, fixed offset
9523	* can be non-zero. This was already checked above. So pass
9524	* fixed_off_ok as true to allow fixed offset for all other
9525	* cases. var_off always must be 0 for PTR_TO_BTF_ID, hence we
9526	* still need to do checks instead of returning.
9527	*/
9528	return __check_ptr_off_reg(env, reg, regno, fixed_off_ok: true);
9529	default:
9530	return __check_ptr_off_reg(env, reg, regno, fixed_off_ok: false);
9531	}
9532	}
9533
9534	static struct bpf_reg_state *get_dynptr_arg_reg(struct bpf_verifier_env *env,
9535	const struct bpf_func_proto *fn,
9536	struct bpf_reg_state *regs)
9537	{
9538	struct bpf_reg_state *state = NULL;
9539	int i;
9540
9541	for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++)
9542	if (arg_type_is_dynptr(type: fn->arg_type[i])) {
9543	if (state) {
9544	verbose(private_data: env, fmt: "verifier internal error: multiple dynptr args\n");
9545	return NULL;
9546	}
9547	state = &regs[BPF_REG_1 + i];
9548	}
9549
9550	if (!state)
9551	verbose(private_data: env, fmt: "verifier internal error: no dynptr arg found\n");
9552
9553	return state;
9554	}
9555
9556	static int dynptr_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
9557	{
9558	struct bpf_func_state *state = func(env, reg);
9559	int spi;
9560
9561	if (reg->type == CONST_PTR_TO_DYNPTR)
9562	return reg->id;
9563	spi = dynptr_get_spi(env, reg);
9564	if (spi < 0)
9565	return spi;
9566	return state->stack[spi].spilled_ptr.id;
9567	}
9568
9569	static int dynptr_ref_obj_id(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
9570	{
9571	struct bpf_func_state *state = func(env, reg);
9572	int spi;
9573
9574	if (reg->type == CONST_PTR_TO_DYNPTR)
9575	return reg->ref_obj_id;
9576	spi = dynptr_get_spi(env, reg);
9577	if (spi < 0)
9578	return spi;
9579	return state->stack[spi].spilled_ptr.ref_obj_id;
9580	}
9581
9582	static enum bpf_dynptr_type dynptr_get_type(struct bpf_verifier_env *env,
9583	struct bpf_reg_state *reg)
9584	{
9585	struct bpf_func_state *state = func(env, reg);
9586	int spi;
9587
9588	if (reg->type == CONST_PTR_TO_DYNPTR)
9589	return reg->dynptr.type;
9590
9591	spi = __get_spi(off: reg->off);
9592	if (spi < 0) {
9593	verbose(private_data: env, fmt: "verifier internal error: invalid spi when querying dynptr type\n");
9594	return BPF_DYNPTR_TYPE_INVALID;
9595	}
9596
9597	return state->stack[spi].spilled_ptr.dynptr.type;
9598	}
9599
9600	static int check_reg_const_str(struct bpf_verifier_env *env,
9601	struct bpf_reg_state *reg, u32 regno)
9602	{
9603	struct bpf_map *map = reg->map_ptr;
9604	int err;
9605	int map_off;
9606	u64 map_addr;
9607	char *str_ptr;
9608
9609	if (reg->type != PTR_TO_MAP_VALUE)
9610	return -EINVAL;
9611
9612	if (map->map_type == BPF_MAP_TYPE_INSN_ARRAY) {
9613	verbose(private_data: env, fmt: "R%d points to insn_array map which cannot be used as const string\n", regno);
9614	return -EACCES;
9615	}
9616
9617	if (!bpf_map_is_rdonly(map)) {
9618	verbose(private_data: env, fmt: "R%d does not point to a readonly map'\n", regno);
9619	return -EACCES;
9620	}
9621
9622	if (!tnum_is_const(a: reg->var_off)) {
9623	verbose(private_data: env, fmt: "R%d is not a constant address'\n", regno);
9624	return -EACCES;
9625	}
9626
9627	if (!map->ops->map_direct_value_addr) {
9628	verbose(private_data: env, fmt: "no direct value access support for this map type\n");
9629	return -EACCES;
9630	}
9631
9632	err = check_map_access(env, regno, off: reg->off,
9633	size: map->value_size - reg->off, zero_size_allowed: false,
9634	src: ACCESS_HELPER);
9635	if (err)
9636	return err;
9637
9638	map_off = reg->off + reg->var_off.value;
9639	err = map->ops->map_direct_value_addr(map, &map_addr, map_off);
9640	if (err) {
9641	verbose(private_data: env, fmt: "direct value access on string failed\n");
9642	return err;
9643	}
9644
9645	str_ptr = (char *)(long)(map_addr);
9646	if (!strnchr(str_ptr + map_off, map->value_size - map_off, 0)) {
9647	verbose(private_data: env, fmt: "string is not zero-terminated\n");
9648	return -EINVAL;
9649	}
9650	return 0;
9651	}
9652
9653	/* Returns constant key value in `value` if possible, else negative error */
9654	static int get_constant_map_key(struct bpf_verifier_env *env,
9655	struct bpf_reg_state *key,
9656	u32 key_size,
9657	s64 *value)
9658	{
9659	struct bpf_func_state *state = func(env, reg: key);
9660	struct bpf_reg_state *reg;
9661	int slot, spi, off;
9662	int spill_size = 0;
9663	int zero_size = 0;
9664	int stack_off;
9665	int i, err;
9666	u8 *stype;
9667
9668	if (!env->bpf_capable)
9669	return -EOPNOTSUPP;
9670	if (key->type != PTR_TO_STACK)
9671	return -EOPNOTSUPP;
9672	if (!tnum_is_const(a: key->var_off))
9673	return -EOPNOTSUPP;
9674
9675	stack_off = key->off + key->var_off.value;
9676	slot = -stack_off - 1;
9677	spi = slot / BPF_REG_SIZE;
9678	off = slot % BPF_REG_SIZE;
9679	stype = state->stack[spi].slot_type;
9680
9681	/* First handle precisely tracked STACK_ZERO */
9682	for (i = off; i >= 0 && stype[i] == STACK_ZERO; i--)
9683	zero_size++;
9684	if (zero_size >= key_size) {
9685	*value = 0;
9686	return 0;
9687	}
9688
9689	/* Check that stack contains a scalar spill of expected size */
9690	if (!is_spilled_scalar_reg(stack: &state->stack[spi]))
9691	return -EOPNOTSUPP;
9692	for (i = off; i >= 0 && stype[i] == STACK_SPILL; i--)
9693	spill_size++;
9694	if (spill_size != key_size)
9695	return -EOPNOTSUPP;
9696
9697	reg = &state->stack[spi].spilled_ptr;
9698	if (!tnum_is_const(a: reg->var_off))
9699	/* Stack value not statically known */
9700	return -EOPNOTSUPP;
9701
9702	/* We are relying on a constant value. So mark as precise
9703	* to prevent pruning on it.
9704	*/
9705	bt_set_frame_slot(bt: &env->bt, frame: key->frameno, slot: spi);
9706	err = mark_chain_precision_batch(env, starting_state: env->cur_state);
9707	if (err < 0)
9708	return err;
9709
9710	*value = reg->var_off.value;
9711	return 0;
9712	}
9713
9714	static bool can_elide_value_nullness(enum bpf_map_type type);
9715
9716	static int check_func_arg(struct bpf_verifier_env *env, u32 arg,
9717	struct bpf_call_arg_meta *meta,
9718	const struct bpf_func_proto *fn,
9719	int insn_idx)
9720	{
9721	u32 regno = BPF_REG_1 + arg;
9722	struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[regno];
9723	enum bpf_arg_type arg_type = fn->arg_type[arg];
9724	enum bpf_reg_type type = reg->type;
9725	u32 *arg_btf_id = NULL;
9726	u32 key_size;
9727	int err = 0;
9728
9729	if (arg_type == ARG_DONTCARE)
9730	return 0;
9731
9732	err = check_reg_arg(env, regno, t: SRC_OP);
9733	if (err)
9734	return err;
9735
9736	if (arg_type == ARG_ANYTHING) {
9737	if (is_pointer_value(env, regno)) {
9738	verbose(private_data: env, fmt: "R%d leaks addr into helper function\n",
9739	regno);
9740	return -EACCES;
9741	}
9742	return 0;
9743	}
9744
9745	if (type_is_pkt_pointer(type) &&
9746	!may_access_direct_pkt_data(env, meta, t: BPF_READ)) {
9747	verbose(private_data: env, fmt: "helper access to the packet is not allowed\n");
9748	return -EACCES;
9749	}
9750
9751	if (base_type(type: arg_type) == ARG_PTR_TO_MAP_VALUE) {
9752	err = resolve_map_arg_type(env, meta, arg_type: &arg_type);
9753	if (err)
9754	return err;
9755	}
9756
9757	if (register_is_null(reg) && type_may_be_null(type: arg_type))
9758	/* A NULL register has a SCALAR_VALUE type, so skip
9759	* type checking.
9760	*/
9761	goto skip_type_check;
9762
9763	/* arg_btf_id and arg_size are in a union. */
9764	if (base_type(type: arg_type) == ARG_PTR_TO_BTF_ID ||
9765	base_type(type: arg_type) == ARG_PTR_TO_SPIN_LOCK)
9766	arg_btf_id = fn->arg_btf_id[arg];
9767
9768	err = check_reg_type(env, regno, arg_type, arg_btf_id, meta);
9769	if (err)
9770	return err;
9771
9772	err = check_func_arg_reg_off(env, reg, regno, arg_type);
9773	if (err)
9774	return err;
9775
9776	skip_type_check:
9777	if (arg_type_is_release(type: arg_type)) {
9778	if (arg_type_is_dynptr(type: arg_type)) {
9779	struct bpf_func_state *state = func(env, reg);
9780	int spi;
9781
9782	/* Only dynptr created on stack can be released, thus
9783	* the get_spi and stack state checks for spilled_ptr
9784	* should only be done before process_dynptr_func for
9785	* PTR_TO_STACK.
9786	*/
9787	if (reg->type == PTR_TO_STACK) {
9788	spi = dynptr_get_spi(env, reg);
9789	if (spi < 0 || !state->stack[spi].spilled_ptr.ref_obj_id) {
9790	verbose(private_data: env, fmt: "arg %d is an unacquired reference\n", regno);
9791	return -EINVAL;
9792	}
9793	} else {
9794	verbose(private_data: env, fmt: "cannot release unowned const bpf_dynptr\n");
9795	return -EINVAL;
9796	}
9797	} else if (!reg->ref_obj_id && !register_is_null(reg)) {
9798	verbose(private_data: env, fmt: "R%d must be referenced when passed to release function\n",
9799	regno);
9800	return -EINVAL;
9801	}
9802	if (meta->release_regno) {
9803	verifier_bug(env, "more than one release argument");
9804	return -EFAULT;
9805	}
9806	meta->release_regno = regno;
9807	}
9808
9809	if (reg->ref_obj_id && base_type(type: arg_type) != ARG_KPTR_XCHG_DEST) {
9810	if (meta->ref_obj_id) {
9811	verbose(private_data: env, fmt: "more than one arg with ref_obj_id R%d %u %u",
9812	regno, reg->ref_obj_id,
9813	meta->ref_obj_id);
9814	return -EACCES;
9815	}
9816	meta->ref_obj_id = reg->ref_obj_id;
9817	}
9818
9819	switch (base_type(type: arg_type)) {
9820	case ARG_CONST_MAP_PTR:
9821	/* bpf_map_xxx(map_ptr) call: remember that map_ptr */
9822	if (meta->map_ptr) {
9823	/* Use map_uid (which is unique id of inner map) to reject:
9824	* inner_map1 = bpf_map_lookup_elem(outer_map, key1)
9825	* inner_map2 = bpf_map_lookup_elem(outer_map, key2)
9826	* if (inner_map1 && inner_map2) {
9827	* timer = bpf_map_lookup_elem(inner_map1);
9828	* if (timer)
9829	* // mismatch would have been allowed
9830	* bpf_timer_init(timer, inner_map2);
9831	* }
9832	*
9833	* Comparing map_ptr is enough to distinguish normal and outer maps.
9834	*/
9835	if (meta->map_ptr != reg->map_ptr ||
9836	meta->map_uid != reg->map_uid) {
9837	verbose(private_data: env,
9838	fmt: "timer pointer in R1 map_uid=%d doesn't match map pointer in R2 map_uid=%d\n",
9839	meta->map_uid, reg->map_uid);
9840	return -EINVAL;
9841	}
9842	}
9843	meta->map_ptr = reg->map_ptr;
9844	meta->map_uid = reg->map_uid;
9845	break;
9846	case ARG_PTR_TO_MAP_KEY:
9847	/* bpf_map_xxx(..., map_ptr, ..., key) call:
9848	* check that [key, key + map->key_size) are within
9849	* stack limits and initialized
9850	*/
9851	if (!meta->map_ptr) {
9852	/* in function declaration map_ptr must come before
9853	* map_key, so that it's verified and known before
9854	* we have to check map_key here. Otherwise it means
9855	* that kernel subsystem misconfigured verifier
9856	*/
9857	verifier_bug(env, "invalid map_ptr to access map->key");
9858	return -EFAULT;
9859	}
9860	key_size = meta->map_ptr->key_size;
9861	err = check_helper_mem_access(env, regno, access_size: key_size, access_type: BPF_READ, zero_size_allowed: false, NULL);
9862	if (err)
9863	return err;
9864	if (can_elide_value_nullness(type: meta->map_ptr->map_type)) {
9865	err = get_constant_map_key(env, key: reg, key_size, value: &meta->const_map_key);
9866	if (err < 0) {
9867	meta->const_map_key = -1;
9868	if (err == -EOPNOTSUPP)
9869	err = 0;
9870	else
9871	return err;
9872	}
9873	}
9874	break;
9875	case ARG_PTR_TO_MAP_VALUE:
9876	if (type_may_be_null(type: arg_type) && register_is_null(reg))
9877	return 0;
9878
9879	/* bpf_map_xxx(..., map_ptr, ..., value) call:
9880	* check [value, value + map->value_size) validity
9881	*/
9882	if (!meta->map_ptr) {
9883	/* kernel subsystem misconfigured verifier */
9884	verifier_bug(env, "invalid map_ptr to access map->value");
9885	return -EFAULT;
9886	}
9887	meta->raw_mode = arg_type & MEM_UNINIT;
9888	err = check_helper_mem_access(env, regno, access_size: meta->map_ptr->value_size,
9889	access_type: arg_type & MEM_WRITE ? BPF_WRITE : BPF_READ,
9890	zero_size_allowed: false, meta);
9891	break;
9892	case ARG_PTR_TO_PERCPU_BTF_ID:
9893	if (!reg->btf_id) {
9894	verbose(private_data: env, fmt: "Helper has invalid btf_id in R%d\n", regno);
9895	return -EACCES;
9896	}
9897	meta->ret_btf = reg->btf;
9898	meta->ret_btf_id = reg->btf_id;
9899	break;
9900	case ARG_PTR_TO_SPIN_LOCK:
9901	if (in_rbtree_lock_required_cb(env)) {
9902	verbose(private_data: env, fmt: "can't spin_{lock,unlock} in rbtree cb\n");
9903	return -EACCES;
9904	}
9905	if (meta->func_id == BPF_FUNC_spin_lock) {
9906	err = process_spin_lock(env, regno, flags: PROCESS_SPIN_LOCK);
9907	if (err)
9908	return err;
9909	} else if (meta->func_id == BPF_FUNC_spin_unlock) {
9910	err = process_spin_lock(env, regno, flags: 0);
9911	if (err)
9912	return err;
9913	} else {
9914	verifier_bug(env, "spin lock arg on unexpected helper");
9915	return -EFAULT;
9916	}
9917	break;
9918	case ARG_PTR_TO_TIMER:
9919	err = process_timer_func(env, regno, meta);
9920	if (err)
9921	return err;
9922	break;
9923	case ARG_PTR_TO_FUNC:
9924	meta->subprogno = reg->subprogno;
9925	break;
9926	case ARG_PTR_TO_MEM:
9927	/* The access to this pointer is only checked when we hit the
9928	* next is_mem_size argument below.
9929	*/
9930	meta->raw_mode = arg_type & MEM_UNINIT;
9931	if (arg_type & MEM_FIXED_SIZE) {
9932	err = check_helper_mem_access(env, regno, access_size: fn->arg_size[arg],
9933	access_type: arg_type & MEM_WRITE ? BPF_WRITE : BPF_READ,
9934	zero_size_allowed: false, meta);
9935	if (err)
9936	return err;
9937	if (arg_type & MEM_ALIGNED)
9938	err = check_ptr_alignment(env, reg, off: 0, size: fn->arg_size[arg], strict_alignment_once: true);
9939	}
9940	break;
9941	case ARG_CONST_SIZE:
9942	err = check_mem_size_reg(env, reg, regno,
9943	access_type: fn->arg_type[arg - 1] & MEM_WRITE ?
9944	BPF_WRITE : BPF_READ,
9945	zero_size_allowed: false, meta);
9946	break;
9947	case ARG_CONST_SIZE_OR_ZERO:
9948	err = check_mem_size_reg(env, reg, regno,
9949	access_type: fn->arg_type[arg - 1] & MEM_WRITE ?
9950	BPF_WRITE : BPF_READ,
9951	zero_size_allowed: true, meta);
9952	break;
9953	case ARG_PTR_TO_DYNPTR:
9954	err = process_dynptr_func(env, regno, insn_idx, arg_type, clone_ref_obj_id: 0);
9955	if (err)
9956	return err;
9957	break;
9958	case ARG_CONST_ALLOC_SIZE_OR_ZERO:
9959	if (!tnum_is_const(a: reg->var_off)) {
9960	verbose(private_data: env, fmt: "R%d is not a known constant'\n",
9961	regno);
9962	return -EACCES;
9963	}
9964	meta->mem_size = reg->var_off.value;
9965	err = mark_chain_precision(env, regno);
9966	if (err)
9967	return err;
9968	break;
9969	case ARG_PTR_TO_CONST_STR:
9970	{
9971	err = check_reg_const_str(env, reg, regno);
9972	if (err)
9973	return err;
9974	break;
9975	}
9976	case ARG_KPTR_XCHG_DEST:
9977	err = process_kptr_func(env, regno, meta);
9978	if (err)
9979	return err;
9980	break;
9981	}
9982
9983	return err;
9984	}
9985
9986	static bool may_update_sockmap(struct bpf_verifier_env *env, int func_id)
9987	{
9988	enum bpf_attach_type eatype = env->prog->expected_attach_type;
9989	enum bpf_prog_type type = resolve_prog_type(prog: env->prog);
9990
9991	if (func_id != BPF_FUNC_map_update_elem &&
9992	func_id != BPF_FUNC_map_delete_elem)
9993	return false;
9994
9995	/* It's not possible to get access to a locked struct sock in these
9996	* contexts, so updating is safe.
9997	*/
9998	switch (type) {
9999	case BPF_PROG_TYPE_TRACING:
10000	if (eatype == BPF_TRACE_ITER)
10001	return true;
10002	break;
10003	case BPF_PROG_TYPE_SOCK_OPS:
10004	/* map_update allowed only via dedicated helpers with event type checks */
10005	if (func_id == BPF_FUNC_map_delete_elem)
10006	return true;
10007	break;
10008	case BPF_PROG_TYPE_SOCKET_FILTER:
10009	case BPF_PROG_TYPE_SCHED_CLS:
10010	case BPF_PROG_TYPE_SCHED_ACT:
10011	case BPF_PROG_TYPE_XDP:
10012	case BPF_PROG_TYPE_SK_REUSEPORT:
10013	case BPF_PROG_TYPE_FLOW_DISSECTOR:
10014	case BPF_PROG_TYPE_SK_LOOKUP:
10015	return true;
10016	default:
10017	break;
10018	}
10019
10020	verbose(private_data: env, fmt: "cannot update sockmap in this context\n");
10021	return false;
10022	}
10023
10024	static bool allow_tail_call_in_subprogs(struct bpf_verifier_env *env)
10025	{
10026	return env->prog->jit_requested &&
10027	bpf_jit_supports_subprog_tailcalls();
10028	}
10029
10030	static int check_map_func_compatibility(struct bpf_verifier_env *env,
10031	struct bpf_map *map, int func_id)
10032	{
10033	if (!map)
10034	return 0;
10035
10036	/* We need a two way check, first is from map perspective ... */
10037	switch (map->map_type) {
10038	case BPF_MAP_TYPE_PROG_ARRAY:
10039	if (func_id != BPF_FUNC_tail_call)
10040	goto error;
10041	break;
10042	case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
10043	if (func_id != BPF_FUNC_perf_event_read &&
10044	func_id != BPF_FUNC_perf_event_output &&
10045	func_id != BPF_FUNC_skb_output &&
10046	func_id != BPF_FUNC_perf_event_read_value &&
10047	func_id != BPF_FUNC_xdp_output)
10048	goto error;
10049	break;
10050	case BPF_MAP_TYPE_RINGBUF:
10051	if (func_id != BPF_FUNC_ringbuf_output &&
10052	func_id != BPF_FUNC_ringbuf_reserve &&
10053	func_id != BPF_FUNC_ringbuf_query &&
10054	func_id != BPF_FUNC_ringbuf_reserve_dynptr &&
10055	func_id != BPF_FUNC_ringbuf_submit_dynptr &&
10056	func_id != BPF_FUNC_ringbuf_discard_dynptr)
10057	goto error;
10058	break;
10059	case BPF_MAP_TYPE_USER_RINGBUF:
10060	if (func_id != BPF_FUNC_user_ringbuf_drain)
10061	goto error;
10062	break;
10063	case BPF_MAP_TYPE_STACK_TRACE:
10064	if (func_id != BPF_FUNC_get_stackid)
10065	goto error;
10066	break;
10067	case BPF_MAP_TYPE_CGROUP_ARRAY:
10068	if (func_id != BPF_FUNC_skb_under_cgroup &&
10069	func_id != BPF_FUNC_current_task_under_cgroup)
10070	goto error;
10071	break;
10072	case BPF_MAP_TYPE_CGROUP_STORAGE:
10073	case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
10074	if (func_id != BPF_FUNC_get_local_storage)
10075	goto error;
10076	break;
10077	case BPF_MAP_TYPE_DEVMAP:
10078	case BPF_MAP_TYPE_DEVMAP_HASH:
10079	if (func_id != BPF_FUNC_redirect_map &&
10080	func_id != BPF_FUNC_map_lookup_elem)
10081	goto error;
10082	break;
10083	/* Restrict bpf side of cpumap and xskmap, open when use-cases
10084	* appear.
10085	*/
10086	case BPF_MAP_TYPE_CPUMAP:
10087	if (func_id != BPF_FUNC_redirect_map)
10088	goto error;
10089	break;
10090	case BPF_MAP_TYPE_XSKMAP:
10091	if (func_id != BPF_FUNC_redirect_map &&
10092	func_id != BPF_FUNC_map_lookup_elem)
10093	goto error;
10094	break;
10095	case BPF_MAP_TYPE_ARRAY_OF_MAPS:
10096	case BPF_MAP_TYPE_HASH_OF_MAPS:
10097	if (func_id != BPF_FUNC_map_lookup_elem)
10098	goto error;
10099	break;
10100	case BPF_MAP_TYPE_SOCKMAP:
10101	if (func_id != BPF_FUNC_sk_redirect_map &&
10102	func_id != BPF_FUNC_sock_map_update &&
10103	func_id != BPF_FUNC_msg_redirect_map &&
10104	func_id != BPF_FUNC_sk_select_reuseport &&
10105	func_id != BPF_FUNC_map_lookup_elem &&
10106	!may_update_sockmap(env, func_id))
10107	goto error;
10108	break;
10109	case BPF_MAP_TYPE_SOCKHASH:
10110	if (func_id != BPF_FUNC_sk_redirect_hash &&
10111	func_id != BPF_FUNC_sock_hash_update &&
10112	func_id != BPF_FUNC_msg_redirect_hash &&
10113	func_id != BPF_FUNC_sk_select_reuseport &&
10114	func_id != BPF_FUNC_map_lookup_elem &&
10115	!may_update_sockmap(env, func_id))
10116	goto error;
10117	break;
10118	case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
10119	if (func_id != BPF_FUNC_sk_select_reuseport)
10120	goto error;
10121	break;
10122	case BPF_MAP_TYPE_QUEUE:
10123	case BPF_MAP_TYPE_STACK:
10124	if (func_id != BPF_FUNC_map_peek_elem &&
10125	func_id != BPF_FUNC_map_pop_elem &&
10126	func_id != BPF_FUNC_map_push_elem)
10127	goto error;
10128	break;
10129	case BPF_MAP_TYPE_SK_STORAGE:
10130	if (func_id != BPF_FUNC_sk_storage_get &&
10131	func_id != BPF_FUNC_sk_storage_delete &&
10132	func_id != BPF_FUNC_kptr_xchg)
10133	goto error;
10134	break;
10135	case BPF_MAP_TYPE_INODE_STORAGE:
10136	if (func_id != BPF_FUNC_inode_storage_get &&
10137	func_id != BPF_FUNC_inode_storage_delete &&
10138	func_id != BPF_FUNC_kptr_xchg)
10139	goto error;
10140	break;
10141	case BPF_MAP_TYPE_TASK_STORAGE:
10142	if (func_id != BPF_FUNC_task_storage_get &&
10143	func_id != BPF_FUNC_task_storage_delete &&
10144	func_id != BPF_FUNC_kptr_xchg)
10145	goto error;
10146	break;
10147	case BPF_MAP_TYPE_CGRP_STORAGE:
10148	if (func_id != BPF_FUNC_cgrp_storage_get &&
10149	func_id != BPF_FUNC_cgrp_storage_delete &&
10150	func_id != BPF_FUNC_kptr_xchg)
10151	goto error;
10152	break;
10153	case BPF_MAP_TYPE_BLOOM_FILTER:
10154	if (func_id != BPF_FUNC_map_peek_elem &&
10155	func_id != BPF_FUNC_map_push_elem)
10156	goto error;
10157	break;
10158	case BPF_MAP_TYPE_INSN_ARRAY:
10159	goto error;
10160	default:
10161	break;
10162	}
10163
10164	/* ... and second from the function itself. */
10165	switch (func_id) {
10166	case BPF_FUNC_tail_call:
10167	if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
10168	goto error;
10169	if (env->subprog_cnt > 1 && !allow_tail_call_in_subprogs(env)) {
10170	verbose(private_data: env, fmt: "mixing of tail_calls and bpf-to-bpf calls is not supported\n");
10171	return -EINVAL;
10172	}
10173	break;
10174	case BPF_FUNC_perf_event_read:
10175	case BPF_FUNC_perf_event_output:
10176	case BPF_FUNC_perf_event_read_value:
10177	case BPF_FUNC_skb_output:
10178	case BPF_FUNC_xdp_output:
10179	if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
10180	goto error;
10181	break;
10182	case BPF_FUNC_ringbuf_output:
10183	case BPF_FUNC_ringbuf_reserve:
10184	case BPF_FUNC_ringbuf_query:
10185	case BPF_FUNC_ringbuf_reserve_dynptr:
10186	case BPF_FUNC_ringbuf_submit_dynptr:
10187	case BPF_FUNC_ringbuf_discard_dynptr:
10188	if (map->map_type != BPF_MAP_TYPE_RINGBUF)
10189	goto error;
10190	break;
10191	case BPF_FUNC_user_ringbuf_drain:
10192	if (map->map_type != BPF_MAP_TYPE_USER_RINGBUF)
10193	goto error;
10194	break;
10195	case BPF_FUNC_get_stackid:
10196	if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
10197	goto error;
10198	break;
10199	case BPF_FUNC_current_task_under_cgroup:
10200	case BPF_FUNC_skb_under_cgroup:
10201	if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
10202	goto error;
10203	break;
10204	case BPF_FUNC_redirect_map:
10205	if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
10206	map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
10207	map->map_type != BPF_MAP_TYPE_CPUMAP &&
10208	map->map_type != BPF_MAP_TYPE_XSKMAP)
10209	goto error;
10210	break;
10211	case BPF_FUNC_sk_redirect_map:
10212	case BPF_FUNC_msg_redirect_map:
10213	case BPF_FUNC_sock_map_update:
10214	if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
10215	goto error;
10216	break;
10217	case BPF_FUNC_sk_redirect_hash:
10218	case BPF_FUNC_msg_redirect_hash:
10219	case BPF_FUNC_sock_hash_update:
10220	if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
10221	goto error;
10222	break;
10223	case BPF_FUNC_get_local_storage:
10224	if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
10225	map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
10226	goto error;
10227	break;
10228	case BPF_FUNC_sk_select_reuseport:
10229	if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
10230	map->map_type != BPF_MAP_TYPE_SOCKMAP &&
10231	map->map_type != BPF_MAP_TYPE_SOCKHASH)
10232	goto error;
10233	break;
10234	case BPF_FUNC_map_pop_elem:
10235	if (map->map_type != BPF_MAP_TYPE_QUEUE &&
10236	map->map_type != BPF_MAP_TYPE_STACK)
10237	goto error;
10238	break;
10239	case BPF_FUNC_map_peek_elem:
10240	case BPF_FUNC_map_push_elem:
10241	if (map->map_type != BPF_MAP_TYPE_QUEUE &&
10242	map->map_type != BPF_MAP_TYPE_STACK &&
10243	map->map_type != BPF_MAP_TYPE_BLOOM_FILTER)
10244	goto error;
10245	break;
10246	case BPF_FUNC_map_lookup_percpu_elem:
10247	if (map->map_type != BPF_MAP_TYPE_PERCPU_ARRAY &&
10248	map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
10249	map->map_type != BPF_MAP_TYPE_LRU_PERCPU_HASH)
10250	goto error;
10251	break;
10252	case BPF_FUNC_sk_storage_get:
10253	case BPF_FUNC_sk_storage_delete:
10254	if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
10255	goto error;
10256	break;
10257	case BPF_FUNC_inode_storage_get:
10258	case BPF_FUNC_inode_storage_delete:
10259	if (map->map_type != BPF_MAP_TYPE_INODE_STORAGE)
10260	goto error;
10261	break;
10262	case BPF_FUNC_task_storage_get:
10263	case BPF_FUNC_task_storage_delete:
10264	if (map->map_type != BPF_MAP_TYPE_TASK_STORAGE)
10265	goto error;
10266	break;
10267	case BPF_FUNC_cgrp_storage_get:
10268	case BPF_FUNC_cgrp_storage_delete:
10269	if (map->map_type != BPF_MAP_TYPE_CGRP_STORAGE)
10270	goto error;
10271	break;
10272	default:
10273	break;
10274	}
10275
10276	return 0;
10277	error:
10278	verbose(private_data: env, fmt: "cannot pass map_type %d into func %s#%d\n",
10279	map->map_type, func_id_name(id: func_id), func_id);
10280	return -EINVAL;
10281	}
10282
10283	static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
10284	{
10285	int count = 0;
10286
10287	if (arg_type_is_raw_mem(type: fn->arg1_type))
10288	count++;
10289	if (arg_type_is_raw_mem(type: fn->arg2_type))
10290	count++;
10291	if (arg_type_is_raw_mem(type: fn->arg3_type))
10292	count++;
10293	if (arg_type_is_raw_mem(type: fn->arg4_type))
10294	count++;
10295	if (arg_type_is_raw_mem(type: fn->arg5_type))
10296	count++;
10297
10298	/* We only support one arg being in raw mode at the moment,
10299	* which is sufficient for the helper functions we have
10300	* right now.
10301	*/
10302	return count <= 1;
10303	}
10304
10305	static bool check_args_pair_invalid(const struct bpf_func_proto *fn, int arg)
10306	{
10307	bool is_fixed = fn->arg_type[arg] & MEM_FIXED_SIZE;
10308	bool has_size = fn->arg_size[arg] != 0;
10309	bool is_next_size = false;
10310
10311	if (arg + 1 < ARRAY_SIZE(fn->arg_type))
10312	is_next_size = arg_type_is_mem_size(type: fn->arg_type[arg + 1]);
10313
10314	if (base_type(type: fn->arg_type[arg]) != ARG_PTR_TO_MEM)
10315	return is_next_size;
10316
10317	return has_size == is_next_size || is_next_size == is_fixed;
10318	}
10319
10320	static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
10321	{
10322	/* bpf_xxx(..., buf, len) call will access 'len'
10323	* bytes from memory 'buf'. Both arg types need
10324	* to be paired, so make sure there's no buggy
10325	* helper function specification.
10326	*/
10327	if (arg_type_is_mem_size(type: fn->arg1_type) ||
10328	check_args_pair_invalid(fn, arg: 0) ||
10329	check_args_pair_invalid(fn, arg: 1) ||
10330	check_args_pair_invalid(fn, arg: 2) ||
10331	check_args_pair_invalid(fn, arg: 3) ||
10332	check_args_pair_invalid(fn, arg: 4))
10333	return false;
10334
10335	return true;
10336	}
10337
10338	static bool check_btf_id_ok(const struct bpf_func_proto *fn)
10339	{
10340	int i;
10341
10342	for (i = 0; i < ARRAY_SIZE(fn->arg_type); i++) {
10343	if (base_type(type: fn->arg_type[i]) == ARG_PTR_TO_BTF_ID)
10344	return !!fn->arg_btf_id[i];
10345	if (base_type(type: fn->arg_type[i]) == ARG_PTR_TO_SPIN_LOCK)
10346	return fn->arg_btf_id[i] == BPF_PTR_POISON;
10347	if (base_type(type: fn->arg_type[i]) != ARG_PTR_TO_BTF_ID && fn->arg_btf_id[i] &&
10348	/* arg_btf_id and arg_size are in a union. */
10349	(base_type(type: fn->arg_type[i]) != ARG_PTR_TO_MEM ||
10350	!(fn->arg_type[i] & MEM_FIXED_SIZE)))
10351	return false;
10352	}
10353
10354	return true;
10355	}
10356
10357	static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
10358	{
10359	return check_raw_mode_ok(fn) &&
10360	check_arg_pair_ok(fn) &&
10361	check_btf_id_ok(fn) ? 0 : -EINVAL;
10362	}
10363
10364	/* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
10365	* are now invalid, so turn them into unknown SCALAR_VALUE.
10366	*
10367	* This also applies to dynptr slices belonging to skb and xdp dynptrs,
10368	* since these slices point to packet data.
10369	*/
10370	static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
10371	{
10372	struct bpf_func_state *state;
10373	struct bpf_reg_state *reg;
10374
10375	bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
10376	if (reg_is_pkt_pointer_any(reg) || reg_is_dynptr_slice_pkt(reg))
10377	mark_reg_invalid(env, reg);
10378	}));
10379	}
10380
10381	enum {
10382	AT_PKT_END = -1,
10383	BEYOND_PKT_END = -2,
10384	};
10385
10386	static void mark_pkt_end(struct bpf_verifier_state *vstate, int regn, bool range_open)
10387	{
10388	struct bpf_func_state *state = vstate->frame[vstate->curframe];
10389	struct bpf_reg_state *reg = &state->regs[regn];
10390
10391	if (reg->type != PTR_TO_PACKET)
10392	/* PTR_TO_PACKET_META is not supported yet */
10393	return;
10394
10395	/* The 'reg' is pkt > pkt_end or pkt >= pkt_end.
10396	* How far beyond pkt_end it goes is unknown.
10397	* if (!range_open) it's the case of pkt >= pkt_end
10398	* if (range_open) it's the case of pkt > pkt_end
10399	* hence this pointer is at least 1 byte bigger than pkt_end
10400	*/
10401	if (range_open)
10402	reg->range = BEYOND_PKT_END;
10403	else
10404	reg->range = AT_PKT_END;
10405	}
10406
10407	static int release_reference_nomark(struct bpf_verifier_state *state, int ref_obj_id)
10408	{
10409	int i;
10410
10411	for (i = 0; i < state->acquired_refs; i++) {
10412	if (state->refs[i].type != REF_TYPE_PTR)
10413	continue;
10414	if (state->refs[i].id == ref_obj_id) {
10415	release_reference_state(state, idx: i);
10416	return 0;
10417	}
10418	}
10419	return -EINVAL;
10420	}
10421
10422	/* The pointer with the specified id has released its reference to kernel
10423	* resources. Identify all copies of the same pointer and clear the reference.
10424	*
10425	* This is the release function corresponding to acquire_reference(). Idempotent.
10426	*/
10427	static int release_reference(struct bpf_verifier_env *env, int ref_obj_id)
10428	{
10429	struct bpf_verifier_state *vstate = env->cur_state;
10430	struct bpf_func_state *state;
10431	struct bpf_reg_state *reg;
10432	int err;
10433
10434	err = release_reference_nomark(state: vstate, ref_obj_id);
10435	if (err)
10436	return err;
10437
10438	bpf_for_each_reg_in_vstate(vstate, state, reg, ({
10439	if (reg->ref_obj_id == ref_obj_id)
10440	mark_reg_invalid(env, reg);
10441	}));
10442
10443	return 0;
10444	}
10445
10446	static void invalidate_non_owning_refs(struct bpf_verifier_env *env)
10447	{
10448	struct bpf_func_state *unused;
10449	struct bpf_reg_state *reg;
10450
10451	bpf_for_each_reg_in_vstate(env->cur_state, unused, reg, ({
10452	if (type_is_non_owning_ref(reg->type))
10453	mark_reg_invalid(env, reg);
10454	}));
10455	}
10456
10457	static void clear_caller_saved_regs(struct bpf_verifier_env *env,
10458	struct bpf_reg_state *regs)
10459	{
10460	int i;
10461
10462	/* after the call registers r0 - r5 were scratched */
10463	for (i = 0; i < CALLER_SAVED_REGS; i++) {
10464	mark_reg_not_init(env, regs, regno: caller_saved[i]);
10465	__check_reg_arg(env, regs, regno: caller_saved[i], t: DST_OP_NO_MARK);
10466	}
10467	}
10468
10469	typedef int (*set_callee_state_fn)(struct bpf_verifier_env *env,
10470	struct bpf_func_state *caller,
10471	struct bpf_func_state *callee,
10472	int insn_idx);
10473
10474	static int set_callee_state(struct bpf_verifier_env *env,
10475	struct bpf_func_state *caller,
10476	struct bpf_func_state *callee, int insn_idx);
10477
10478	static int setup_func_entry(struct bpf_verifier_env *env, int subprog, int callsite,
10479	set_callee_state_fn set_callee_state_cb,
10480	struct bpf_verifier_state *state)
10481	{
10482	struct bpf_func_state *caller, *callee;
10483	int err;
10484
10485	if (state->curframe + 1 >= MAX_CALL_FRAMES) {
10486	verbose(private_data: env, fmt: "the call stack of %d frames is too deep\n",
10487	state->curframe + 2);
10488	return -E2BIG;
10489	}
10490
10491	if (state->frame[state->curframe + 1]) {
10492	verifier_bug(env, "Frame %d already allocated", state->curframe + 1);
10493	return -EFAULT;
10494	}
10495
10496	caller = state->frame[state->curframe];
10497	callee = kzalloc(sizeof(*callee), GFP_KERNEL_ACCOUNT);
10498	if (!callee)
10499	return -ENOMEM;
10500	state->frame[state->curframe + 1] = callee;
10501
10502	/* callee cannot access r0, r6 - r9 for reading and has to write
10503	* into its own stack before reading from it.
10504	* callee can read/write into caller's stack
10505	*/
10506	init_func_state(env, state: callee,
10507	/* remember the callsite, it will be used by bpf_exit */
10508	callsite,
10509	frameno: state->curframe + 1 /* frameno within this callchain */,
10510	subprogno: subprog /* subprog number within this prog */);
10511	err = set_callee_state_cb(env, caller, callee, callsite);
10512	if (err)
10513	goto err_out;
10514
10515	/* only increment it after check_reg_arg() finished */
10516	state->curframe++;
10517
10518	return 0;
10519
10520	err_out:
10521	free_func_state(state: callee);
10522	state->frame[state->curframe + 1] = NULL;
10523	return err;
10524	}
10525
10526	static int btf_check_func_arg_match(struct bpf_verifier_env *env, int subprog,
10527	const struct btf *btf,
10528	struct bpf_reg_state *regs)
10529	{
10530	struct bpf_subprog_info *sub = subprog_info(env, subprog);
10531	struct bpf_verifier_log *log = &env->log;
10532	u32 i;
10533	int ret;
10534
10535	ret = btf_prepare_func_args(env, subprog);
10536	if (ret)
10537	return ret;
10538
10539	/* check that BTF function arguments match actual types that the
10540	* verifier sees.
10541	*/
10542	for (i = 0; i < sub->arg_cnt; i++) {
10543	u32 regno = i + 1;
10544	struct bpf_reg_state *reg = &regs[regno];
10545	struct bpf_subprog_arg_info *arg = &sub->args[i];
10546
10547	if (arg->arg_type == ARG_ANYTHING) {
10548	if (reg->type != SCALAR_VALUE) {
10549	bpf_log(log, fmt: "R%d is not a scalar\n", regno);
10550	return -EINVAL;
10551	}
10552	} else if (arg->arg_type & PTR_UNTRUSTED) {
10553	/*
10554	* Anything is allowed for untrusted arguments, as these are
10555	* read-only and probe read instructions would protect against
10556	* invalid memory access.
10557	*/
10558	} else if (arg->arg_type == ARG_PTR_TO_CTX) {
10559	ret = check_func_arg_reg_off(env, reg, regno, arg_type: ARG_DONTCARE);
10560	if (ret < 0)
10561	return ret;
10562	/* If function expects ctx type in BTF check that caller
10563	* is passing PTR_TO_CTX.
10564	*/
10565	if (reg->type != PTR_TO_CTX) {
10566	bpf_log(log, fmt: "arg#%d expects pointer to ctx\n", i);
10567	return -EINVAL;
10568	}
10569	} else if (base_type(type: arg->arg_type) == ARG_PTR_TO_MEM) {
10570	ret = check_func_arg_reg_off(env, reg, regno, arg_type: ARG_DONTCARE);
10571	if (ret < 0)
10572	return ret;
10573	if (check_mem_reg(env, reg, regno, mem_size: arg->mem_size))
10574	return -EINVAL;
10575	if (!(arg->arg_type & PTR_MAYBE_NULL) && (reg->type & PTR_MAYBE_NULL)) {
10576	bpf_log(log, fmt: "arg#%d is expected to be non-NULL\n", i);
10577	return -EINVAL;
10578	}
10579	} else if (base_type(type: arg->arg_type) == ARG_PTR_TO_ARENA) {
10580	/*
10581	* Can pass any value and the kernel won't crash, but
10582	* only PTR_TO_ARENA or SCALAR make sense. Everything
10583	* else is a bug in the bpf program. Point it out to
10584	* the user at the verification time instead of
10585	* run-time debug nightmare.
10586	*/
10587	if (reg->type != PTR_TO_ARENA && reg->type != SCALAR_VALUE) {
10588	bpf_log(log, fmt: "R%d is not a pointer to arena or scalar.\n", regno);
10589	return -EINVAL;
10590	}
10591	} else if (arg->arg_type == (ARG_PTR_TO_DYNPTR | MEM_RDONLY)) {
10592	ret = check_func_arg_reg_off(env, reg, regno, arg_type: ARG_PTR_TO_DYNPTR);
10593	if (ret)
10594	return ret;
10595
10596	ret = process_dynptr_func(env, regno, insn_idx: -1, arg_type: arg->arg_type, clone_ref_obj_id: 0);
10597	if (ret)
10598	return ret;
10599	} else if (base_type(type: arg->arg_type) == ARG_PTR_TO_BTF_ID) {
10600	struct bpf_call_arg_meta meta;
10601	int err;
10602
10603	if (register_is_null(reg) && type_may_be_null(type: arg->arg_type))
10604	continue;
10605
10606	memset(&meta, 0, sizeof(meta)); /* leave func_id as zero */
10607	err = check_reg_type(env, regno, arg_type: arg->arg_type, arg_btf_id: &arg->btf_id, meta: &meta);
10608	err = err ?: check_func_arg_reg_off(env, reg, regno, arg_type: arg->arg_type);
10609	if (err)
10610	return err;
10611	} else {
10612	verifier_bug(env, "unrecognized arg#%d type %d", i, arg->arg_type);
10613	return -EFAULT;
10614	}
10615	}
10616
10617	return 0;
10618	}
10619
10620	/* Compare BTF of a function call with given bpf_reg_state.
10621	* Returns:
10622	* EFAULT - there is a verifier bug. Abort verification.
10623	* EINVAL - there is a type mismatch or BTF is not available.
10624	* 0 - BTF matches with what bpf_reg_state expects.
10625	* Only PTR_TO_CTX and SCALAR_VALUE states are recognized.
10626	*/
10627	static int btf_check_subprog_call(struct bpf_verifier_env *env, int subprog,
10628	struct bpf_reg_state *regs)
10629	{
10630	struct bpf_prog *prog = env->prog;
10631	struct btf *btf = prog->aux->btf;
10632	u32 btf_id;
10633	int err;
10634
10635	if (!prog->aux->func_info)
10636	return -EINVAL;
10637
10638	btf_id = prog->aux->func_info[subprog].type_id;
10639	if (!btf_id)
10640	return -EFAULT;
10641
10642	if (prog->aux->func_info_aux[subprog].unreliable)
10643	return -EINVAL;
10644
10645	err = btf_check_func_arg_match(env, subprog, btf, regs);
10646	/* Compiler optimizations can remove arguments from static functions
10647	* or mismatched type can be passed into a global function.
10648	* In such cases mark the function as unreliable from BTF point of view.
10649	*/
10650	if (err)
10651	prog->aux->func_info_aux[subprog].unreliable = true;
10652	return err;
10653	}
10654
10655	static int push_callback_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
10656	int insn_idx, int subprog,
10657	set_callee_state_fn set_callee_state_cb)
10658	{
10659	struct bpf_verifier_state *state = env->cur_state, *callback_state;
10660	struct bpf_func_state *caller, *callee;
10661	int err;
10662
10663	caller = state->frame[state->curframe];
10664	err = btf_check_subprog_call(env, subprog, regs: caller->regs);
10665	if (err == -EFAULT)
10666	return err;
10667
10668	/* set_callee_state is used for direct subprog calls, but we are
10669	* interested in validating only BPF helpers that can call subprogs as
10670	* callbacks
10671	*/
10672	env->subprog_info[subprog].is_cb = true;
10673	if (bpf_pseudo_kfunc_call(insn) &&
10674	!is_callback_calling_kfunc(btf_id: insn->imm)) {
10675	verifier_bug(env, "kfunc %s#%d not marked as callback-calling",
10676	func_id_name(insn->imm), insn->imm);
10677	return -EFAULT;
10678	} else if (!bpf_pseudo_kfunc_call(insn) &&
10679	!is_callback_calling_function(func_id: insn->imm)) { /* helper */
10680	verifier_bug(env, "helper %s#%d not marked as callback-calling",
10681	func_id_name(insn->imm), insn->imm);
10682	return -EFAULT;
10683	}
10684
10685	if (is_async_callback_calling_insn(insn)) {
10686	struct bpf_verifier_state *async_cb;
10687
10688	/* there is no real recursion here. timer and workqueue callbacks are async */
10689	env->subprog_info[subprog].is_async_cb = true;
10690	async_cb = push_async_cb(env, insn_idx: env->subprog_info[subprog].start,
10691	prev_insn_idx: insn_idx, subprog,
10692	is_sleepable: is_async_cb_sleepable(env, insn));
10693	if (IS_ERR(ptr: async_cb))
10694	return PTR_ERR(ptr: async_cb);
10695	callee = async_cb->frame[0];
10696	callee->async_entry_cnt = caller->async_entry_cnt + 1;
10697
10698	/* Convert bpf_timer_set_callback() args into timer callback args */
10699	err = set_callee_state_cb(env, caller, callee, insn_idx);
10700	if (err)
10701	return err;
10702
10703	return 0;
10704	}
10705
10706	/* for callback functions enqueue entry to callback and
10707	* proceed with next instruction within current frame.
10708	*/
10709	callback_state = push_stack(env, insn_idx: env->subprog_info[subprog].start, prev_insn_idx: insn_idx, speculative: false);
10710	if (IS_ERR(ptr: callback_state))
10711	return PTR_ERR(ptr: callback_state);
10712
10713	err = setup_func_entry(env, subprog, callsite: insn_idx, set_callee_state_cb,
10714	state: callback_state);
10715	if (err)
10716	return err;
10717
10718	callback_state->callback_unroll_depth++;
10719	callback_state->frame[callback_state->curframe - 1]->callback_depth++;
10720	caller->callback_depth = 0;
10721	return 0;
10722	}
10723
10724	static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
10725	int *insn_idx)
10726	{
10727	struct bpf_verifier_state *state = env->cur_state;
10728	struct bpf_func_state *caller;
10729	int err, subprog, target_insn;
10730
10731	target_insn = *insn_idx + insn->imm + 1;
10732	subprog = find_subprog(env, off: target_insn);
10733	if (verifier_bug_if(subprog < 0, env, "target of func call at insn %d is not a program",
10734	target_insn))
10735	return -EFAULT;
10736
10737	caller = state->frame[state->curframe];
10738	err = btf_check_subprog_call(env, subprog, regs: caller->regs);
10739	if (err == -EFAULT)
10740	return err;
10741	if (subprog_is_global(env, subprog)) {
10742	const char *sub_name = subprog_name(env, subprog);
10743
10744	if (env->cur_state->active_locks) {
10745	verbose(private_data: env, fmt: "global function calls are not allowed while holding a lock,\n"
10746	"use static function instead\n");
10747	return -EINVAL;
10748	}
10749
10750	if (env->subprog_info[subprog].might_sleep &&
10751	(env->cur_state->active_rcu_locks || env->cur_state->active_preempt_locks ||
10752	env->cur_state->active_irq_id || !in_sleepable(env))) {
10753	verbose(private_data: env, fmt: "global functions that may sleep are not allowed in non-sleepable context,\n"
10754	"i.e., in a RCU/IRQ/preempt-disabled section, or in\n"
10755	"a non-sleepable BPF program context\n");
10756	return -EINVAL;
10757	}
10758
10759	if (err) {
10760	verbose(private_data: env, fmt: "Caller passes invalid args into func#%d ('%s')\n",
10761	subprog, sub_name);
10762	return err;
10763	}
10764
10765	if (env->log.level & BPF_LOG_LEVEL)
10766	verbose(private_data: env, fmt: "Func#%d ('%s') is global and assumed valid.\n",
10767	subprog, sub_name);
10768	if (env->subprog_info[subprog].changes_pkt_data)
10769	clear_all_pkt_pointers(env);
10770	/* mark global subprog for verifying after main prog */
10771	subprog_aux(env, subprog)->called = true;
10772	clear_caller_saved_regs(env, regs: caller->regs);
10773
10774	/* All global functions return a 64-bit SCALAR_VALUE */
10775	mark_reg_unknown(env, regs: caller->regs, regno: BPF_REG_0);
10776	caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
10777
10778	/* continue with next insn after call */
10779	return 0;
10780	}
10781
10782	/* for regular function entry setup new frame and continue
10783	* from that frame.
10784	*/
10785	err = setup_func_entry(env, subprog, callsite: *insn_idx, set_callee_state_cb: set_callee_state, state);
10786	if (err)
10787	return err;
10788
10789	clear_caller_saved_regs(env, regs: caller->regs);
10790
10791	/* and go analyze first insn of the callee */
10792	*insn_idx = env->subprog_info[subprog].start - 1;
10793
10794	bpf_reset_live_stack_callchain(env);
10795
10796	if (env->log.level & BPF_LOG_LEVEL) {
10797	verbose(private_data: env, fmt: "caller:\n");
10798	print_verifier_state(env, vstate: state, frameno: caller->frameno, print_all: true);
10799	verbose(private_data: env, fmt: "callee:\n");
10800	print_verifier_state(env, vstate: state, frameno: state->curframe, print_all: true);
10801	}
10802
10803	return 0;
10804	}
10805
10806	int map_set_for_each_callback_args(struct bpf_verifier_env *env,
10807	struct bpf_func_state *caller,
10808	struct bpf_func_state *callee)
10809	{
10810	/* bpf_for_each_map_elem(struct bpf_map *map, void *callback_fn,
10811	* void *callback_ctx, u64 flags);
10812	* callback_fn(struct bpf_map *map, void *key, void *value,
10813	* void *callback_ctx);
10814	*/
10815	callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
10816
10817	callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
10818	__mark_reg_known_zero(reg: &callee->regs[BPF_REG_2]);
10819	callee->regs[BPF_REG_2].map_ptr = caller->regs[BPF_REG_1].map_ptr;
10820
10821	callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
10822	__mark_reg_known_zero(reg: &callee->regs[BPF_REG_3]);
10823	callee->regs[BPF_REG_3].map_ptr = caller->regs[BPF_REG_1].map_ptr;
10824
10825	/* pointer to stack or null */
10826	callee->regs[BPF_REG_4] = caller->regs[BPF_REG_3];
10827
10828	/* unused */
10829	__mark_reg_not_init(env, reg: &callee->regs[BPF_REG_5]);
10830	return 0;
10831	}
10832
10833	static int set_callee_state(struct bpf_verifier_env *env,
10834	struct bpf_func_state *caller,
10835	struct bpf_func_state *callee, int insn_idx)
10836	{
10837	int i;
10838
10839	/* copy r1 - r5 args that callee can access. The copy includes parent
10840	* pointers, which connects us up to the liveness chain
10841	*/
10842	for (i = BPF_REG_1; i <= BPF_REG_5; i++)
10843	callee->regs[i] = caller->regs[i];
10844	return 0;
10845	}
10846
10847	static int set_map_elem_callback_state(struct bpf_verifier_env *env,
10848	struct bpf_func_state *caller,
10849	struct bpf_func_state *callee,
10850	int insn_idx)
10851	{
10852	struct bpf_insn_aux_data *insn_aux = &env->insn_aux_data[insn_idx];
10853	struct bpf_map *map;
10854	int err;
10855
10856	/* valid map_ptr and poison value does not matter */
10857	map = insn_aux->map_ptr_state.map_ptr;
10858	if (!map->ops->map_set_for_each_callback_args ||
10859	!map->ops->map_for_each_callback) {
10860	verbose(private_data: env, fmt: "callback function not allowed for map\n");
10861	return -ENOTSUPP;
10862	}
10863
10864	err = map->ops->map_set_for_each_callback_args(env, caller, callee);
10865	if (err)
10866	return err;
10867
10868	callee->in_callback_fn = true;
10869	callee->callback_ret_range = retval_range(minval: 0, maxval: 1);
10870	return 0;
10871	}
10872
10873	static int set_loop_callback_state(struct bpf_verifier_env *env,
10874	struct bpf_func_state *caller,
10875	struct bpf_func_state *callee,
10876	int insn_idx)
10877	{
10878	/* bpf_loop(u32 nr_loops, void *callback_fn, void *callback_ctx,
10879	* u64 flags);
10880	* callback_fn(u64 index, void *callback_ctx);
10881	*/
10882	callee->regs[BPF_REG_1].type = SCALAR_VALUE;
10883	callee->regs[BPF_REG_2] = caller->regs[BPF_REG_3];
10884
10885	/* unused */
10886	__mark_reg_not_init(env, reg: &callee->regs[BPF_REG_3]);
10887	__mark_reg_not_init(env, reg: &callee->regs[BPF_REG_4]);
10888	__mark_reg_not_init(env, reg: &callee->regs[BPF_REG_5]);
10889
10890	callee->in_callback_fn = true;
10891	callee->callback_ret_range = retval_range(minval: 0, maxval: 1);
10892	return 0;
10893	}
10894
10895	static int set_timer_callback_state(struct bpf_verifier_env *env,
10896	struct bpf_func_state *caller,
10897	struct bpf_func_state *callee,
10898	int insn_idx)
10899	{
10900	struct bpf_map *map_ptr = caller->regs[BPF_REG_1].map_ptr;
10901
10902	/* bpf_timer_set_callback(struct bpf_timer *timer, void *callback_fn);
10903	* callback_fn(struct bpf_map *map, void *key, void *value);
10904	*/
10905	callee->regs[BPF_REG_1].type = CONST_PTR_TO_MAP;
10906	__mark_reg_known_zero(reg: &callee->regs[BPF_REG_1]);
10907	callee->regs[BPF_REG_1].map_ptr = map_ptr;
10908
10909	callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
10910	__mark_reg_known_zero(reg: &callee->regs[BPF_REG_2]);
10911	callee->regs[BPF_REG_2].map_ptr = map_ptr;
10912
10913	callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
10914	__mark_reg_known_zero(reg: &callee->regs[BPF_REG_3]);
10915	callee->regs[BPF_REG_3].map_ptr = map_ptr;
10916
10917	/* unused */
10918	__mark_reg_not_init(env, reg: &callee->regs[BPF_REG_4]);
10919	__mark_reg_not_init(env, reg: &callee->regs[BPF_REG_5]);
10920	callee->in_async_callback_fn = true;
10921	callee->callback_ret_range = retval_range(minval: 0, maxval: 0);
10922	return 0;
10923	}
10924
10925	static int set_find_vma_callback_state(struct bpf_verifier_env *env,
10926	struct bpf_func_state *caller,
10927	struct bpf_func_state *callee,
10928	int insn_idx)
10929	{
10930	/* bpf_find_vma(struct task_struct *task, u64 addr,
10931	* void *callback_fn, void *callback_ctx, u64 flags)
10932	* (callback_fn)(struct task_struct *task,
10933	* struct vm_area_struct *vma, void *callback_ctx);
10934	*/
10935	callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
10936
10937	callee->regs[BPF_REG_2].type = PTR_TO_BTF_ID;
10938	__mark_reg_known_zero(reg: &callee->regs[BPF_REG_2]);
10939	callee->regs[BPF_REG_2].btf = btf_vmlinux;
10940	callee->regs[BPF_REG_2].btf_id = btf_tracing_ids[BTF_TRACING_TYPE_VMA];
10941
10942	/* pointer to stack or null */
10943	callee->regs[BPF_REG_3] = caller->regs[BPF_REG_4];
10944
10945	/* unused */
10946	__mark_reg_not_init(env, reg: &callee->regs[BPF_REG_4]);
10947	__mark_reg_not_init(env, reg: &callee->regs[BPF_REG_5]);
10948	callee->in_callback_fn = true;
10949	callee->callback_ret_range = retval_range(minval: 0, maxval: 1);
10950	return 0;
10951	}
10952
10953	static int set_user_ringbuf_callback_state(struct bpf_verifier_env *env,
10954	struct bpf_func_state *caller,
10955	struct bpf_func_state *callee,
10956	int insn_idx)
10957	{
10958	/* bpf_user_ringbuf_drain(struct bpf_map *map, void *callback_fn, void
10959	* callback_ctx, u64 flags);
10960	* callback_fn(const struct bpf_dynptr_t* dynptr, void *callback_ctx);
10961	*/
10962	__mark_reg_not_init(env, reg: &callee->regs[BPF_REG_0]);
10963	mark_dynptr_cb_reg(env, reg: &callee->regs[BPF_REG_1], type: BPF_DYNPTR_TYPE_LOCAL);
10964	callee->regs[BPF_REG_2] = caller->regs[BPF_REG_3];
10965
10966	/* unused */
10967	__mark_reg_not_init(env, reg: &callee->regs[BPF_REG_3]);
10968	__mark_reg_not_init(env, reg: &callee->regs[BPF_REG_4]);
10969	__mark_reg_not_init(env, reg: &callee->regs[BPF_REG_5]);
10970
10971	callee->in_callback_fn = true;
10972	callee->callback_ret_range = retval_range(minval: 0, maxval: 1);
10973	return 0;
10974	}
10975
10976	static int set_rbtree_add_callback_state(struct bpf_verifier_env *env,
10977	struct bpf_func_state *caller,
10978	struct bpf_func_state *callee,
10979	int insn_idx)
10980	{
10981	/* void bpf_rbtree_add_impl(struct bpf_rb_root *root, struct bpf_rb_node *node,
10982	* bool (less)(struct bpf_rb_node *a, const struct bpf_rb_node *b));
10983	*
10984	* 'struct bpf_rb_node *node' arg to bpf_rbtree_add_impl is the same PTR_TO_BTF_ID w/ offset
10985	* that 'less' callback args will be receiving. However, 'node' arg was release_reference'd
10986	* by this point, so look at 'root'
10987	*/
10988	struct btf_field *field;
10989
10990	field = reg_find_field_offset(reg: &caller->regs[BPF_REG_1], off: caller->regs[BPF_REG_1].off,
10991	fields: BPF_RB_ROOT);
10992	if (!field || !field->graph_root.value_btf_id)
10993	return -EFAULT;
10994
10995	mark_reg_graph_node(regs: callee->regs, regno: BPF_REG_1, ds_head: &field->graph_root);
10996	ref_set_non_owning(env, reg: &callee->regs[BPF_REG_1]);
10997	mark_reg_graph_node(regs: callee->regs, regno: BPF_REG_2, ds_head: &field->graph_root);
10998	ref_set_non_owning(env, reg: &callee->regs[BPF_REG_2]);
10999
11000	__mark_reg_not_init(env, reg: &callee->regs[BPF_REG_3]);
11001	__mark_reg_not_init(env, reg: &callee->regs[BPF_REG_4]);
11002	__mark_reg_not_init(env, reg: &callee->regs[BPF_REG_5]);
11003	callee->in_callback_fn = true;
11004	callee->callback_ret_range = retval_range(minval: 0, maxval: 1);
11005	return 0;
11006	}
11007
11008	static int set_task_work_schedule_callback_state(struct bpf_verifier_env *env,
11009	struct bpf_func_state *caller,
11010	struct bpf_func_state *callee,
11011	int insn_idx)
11012	{
11013	struct bpf_map *map_ptr = caller->regs[BPF_REG_3].map_ptr;
11014
11015	/*
11016	* callback_fn(struct bpf_map *map, void *key, void *value);
11017	*/
11018	callee->regs[BPF_REG_1].type = CONST_PTR_TO_MAP;
11019	__mark_reg_known_zero(reg: &callee->regs[BPF_REG_1]);
11020	callee->regs[BPF_REG_1].map_ptr = map_ptr;
11021
11022	callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
11023	__mark_reg_known_zero(reg: &callee->regs[BPF_REG_2]);
11024	callee->regs[BPF_REG_2].map_ptr = map_ptr;
11025
11026	callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
11027	__mark_reg_known_zero(reg: &callee->regs[BPF_REG_3]);
11028	callee->regs[BPF_REG_3].map_ptr = map_ptr;
11029
11030	/* unused */
11031	__mark_reg_not_init(env, reg: &callee->regs[BPF_REG_4]);
11032	__mark_reg_not_init(env, reg: &callee->regs[BPF_REG_5]);
11033	callee->in_async_callback_fn = true;
11034	callee->callback_ret_range = retval_range(S32_MIN, S32_MAX);
11035	return 0;
11036	}
11037
11038	static bool is_rbtree_lock_required_kfunc(u32 btf_id);
11039
11040	/* Are we currently verifying the callback for a rbtree helper that must
11041	* be called with lock held? If so, no need to complain about unreleased
11042	* lock
11043	*/
11044	static bool in_rbtree_lock_required_cb(struct bpf_verifier_env *env)
11045	{
11046	struct bpf_verifier_state *state = env->cur_state;
11047	struct bpf_insn *insn = env->prog->insnsi;
11048	struct bpf_func_state *callee;
11049	int kfunc_btf_id;
11050
11051	if (!state->curframe)
11052	return false;
11053
11054	callee = state->frame[state->curframe];
11055
11056	if (!callee->in_callback_fn)
11057	return false;
11058
11059	kfunc_btf_id = insn[callee->callsite].imm;
11060	return is_rbtree_lock_required_kfunc(btf_id: kfunc_btf_id);
11061	}
11062
11063	static bool retval_range_within(struct bpf_retval_range range, const struct bpf_reg_state *reg,
11064	bool return_32bit)
11065	{
11066	if (return_32bit)
11067	return range.minval <= reg->s32_min_value && reg->s32_max_value <= range.maxval;
11068	else
11069	return range.minval <= reg->smin_value && reg->smax_value <= range.maxval;
11070	}
11071
11072	static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
11073	{
11074	struct bpf_verifier_state *state = env->cur_state, *prev_st;
11075	struct bpf_func_state *caller, *callee;
11076	struct bpf_reg_state *r0;
11077	bool in_callback_fn;
11078	int err;
11079
11080	err = bpf_update_live_stack(env);
11081	if (err)
11082	return err;
11083
11084	callee = state->frame[state->curframe];
11085	r0 = &callee->regs[BPF_REG_0];
11086	if (r0->type == PTR_TO_STACK) {
11087	/* technically it's ok to return caller's stack pointer
11088	* (or caller's caller's pointer) back to the caller,
11089	* since these pointers are valid. Only current stack
11090	* pointer will be invalid as soon as function exits,
11091	* but let's be conservative
11092	*/
11093	verbose(private_data: env, fmt: "cannot return stack pointer to the caller\n");
11094	return -EINVAL;
11095	}
11096
11097	caller = state->frame[state->curframe - 1];
11098	if (callee->in_callback_fn) {
11099	if (r0->type != SCALAR_VALUE) {
11100	verbose(private_data: env, fmt: "R0 not a scalar value\n");
11101	return -EACCES;
11102	}
11103
11104	/* we are going to rely on register's precise value */
11105	err = mark_chain_precision(env, regno: BPF_REG_0);
11106	if (err)
11107	return err;
11108
11109	/* enforce R0 return value range, and bpf_callback_t returns 64bit */
11110	if (!retval_range_within(range: callee->callback_ret_range, reg: r0, return_32bit: false)) {
11111	verbose_invalid_scalar(env, reg: r0, range: callee->callback_ret_range,
11112	ctx: "At callback return", reg_name: "R0");
11113	return -EINVAL;
11114	}
11115	if (!bpf_calls_callback(env, insn_idx: callee->callsite)) {
11116	verifier_bug(env, "in callback at %d, callsite %d !calls_callback",
11117	*insn_idx, callee->callsite);
11118	return -EFAULT;
11119	}
11120	} else {
11121	/* return to the caller whatever r0 had in the callee */
11122	caller->regs[BPF_REG_0] = *r0;
11123	}
11124
11125	/* for callbacks like bpf_loop or bpf_for_each_map_elem go back to callsite,
11126	* there function call logic would reschedule callback visit. If iteration
11127	* converges is_state_visited() would prune that visit eventually.
11128	*/
11129	in_callback_fn = callee->in_callback_fn;
11130	if (in_callback_fn)
11131	*insn_idx = callee->callsite;
11132	else
11133	*insn_idx = callee->callsite + 1;
11134
11135	if (env->log.level & BPF_LOG_LEVEL) {
11136	verbose(private_data: env, fmt: "returning from callee:\n");
11137	print_verifier_state(env, vstate: state, frameno: callee->frameno, print_all: true);
11138	verbose(private_data: env, fmt: "to caller at %d:\n", *insn_idx);
11139	print_verifier_state(env, vstate: state, frameno: caller->frameno, print_all: true);
11140	}
11141	/* clear everything in the callee. In case of exceptional exits using
11142	* bpf_throw, this will be done by copy_verifier_state for extra frames. */
11143	free_func_state(state: callee);
11144	state->frame[state->curframe--] = NULL;
11145
11146	/* for callbacks widen imprecise scalars to make programs like below verify:
11147	*
11148	* struct ctx { int i; }
11149	* void cb(int idx, struct ctx *ctx) { ctx->i++; ... }
11150	* ...
11151	* struct ctx = { .i = 0; }
11152	* bpf_loop(100, cb, &ctx, 0);
11153	*
11154	* This is similar to what is done in process_iter_next_call() for open
11155	* coded iterators.
11156	*/
11157	prev_st = in_callback_fn ? find_prev_entry(env, cur: state, insn_idx: *insn_idx) : NULL;
11158	if (prev_st) {
11159	err = widen_imprecise_scalars(env, old: prev_st, cur: state);
11160	if (err)
11161	return err;
11162	}
11163	return 0;
11164	}
11165
11166	static int do_refine_retval_range(struct bpf_verifier_env *env,
11167	struct bpf_reg_state *regs, int ret_type,
11168	int func_id,
11169	struct bpf_call_arg_meta *meta)
11170	{
11171	struct bpf_reg_state *ret_reg = &regs[BPF_REG_0];
11172
11173	if (ret_type != RET_INTEGER)
11174	return 0;
11175
11176	switch (func_id) {
11177	case BPF_FUNC_get_stack:
11178	case BPF_FUNC_get_task_stack:
11179	case BPF_FUNC_probe_read_str:
11180	case BPF_FUNC_probe_read_kernel_str:
11181	case BPF_FUNC_probe_read_user_str:
11182	ret_reg->smax_value = meta->msize_max_value;
11183	ret_reg->s32_max_value = meta->msize_max_value;
11184	ret_reg->smin_value = -MAX_ERRNO;
11185	ret_reg->s32_min_value = -MAX_ERRNO;
11186	reg_bounds_sync(reg: ret_reg);
11187	break;
11188	case BPF_FUNC_get_smp_processor_id:
11189	ret_reg->umax_value = nr_cpu_ids - 1;
11190	ret_reg->u32_max_value = nr_cpu_ids - 1;
11191	ret_reg->smax_value = nr_cpu_ids - 1;
11192	ret_reg->s32_max_value = nr_cpu_ids - 1;
11193	ret_reg->umin_value = 0;
11194	ret_reg->u32_min_value = 0;
11195	ret_reg->smin_value = 0;
11196	ret_reg->s32_min_value = 0;
11197	reg_bounds_sync(reg: ret_reg);
11198	break;
11199	}
11200
11201	return reg_bounds_sanity_check(env, reg: ret_reg, ctx: "retval");
11202	}
11203
11204	static int
11205	record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
11206	int func_id, int insn_idx)
11207	{
11208	struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
11209	struct bpf_map *map = meta->map_ptr;
11210
11211	if (func_id != BPF_FUNC_tail_call &&
11212	func_id != BPF_FUNC_map_lookup_elem &&
11213	func_id != BPF_FUNC_map_update_elem &&
11214	func_id != BPF_FUNC_map_delete_elem &&
11215	func_id != BPF_FUNC_map_push_elem &&
11216	func_id != BPF_FUNC_map_pop_elem &&
11217	func_id != BPF_FUNC_map_peek_elem &&
11218	func_id != BPF_FUNC_for_each_map_elem &&
11219	func_id != BPF_FUNC_redirect_map &&
11220	func_id != BPF_FUNC_map_lookup_percpu_elem)
11221	return 0;
11222
11223	if (map == NULL) {
11224	verifier_bug(env, "expected map for helper call");
11225	return -EFAULT;
11226	}
11227
11228	/* In case of read-only, some additional restrictions
11229	* need to be applied in order to prevent altering the
11230	* state of the map from program side.
11231	*/
11232	if ((map->map_flags & BPF_F_RDONLY_PROG) &&
11233	(func_id == BPF_FUNC_map_delete_elem ||
11234	func_id == BPF_FUNC_map_update_elem ||
11235	func_id == BPF_FUNC_map_push_elem ||
11236	func_id == BPF_FUNC_map_pop_elem)) {
11237	verbose(private_data: env, fmt: "write into map forbidden\n");
11238	return -EACCES;
11239	}
11240
11241	if (!aux->map_ptr_state.map_ptr)
11242	bpf_map_ptr_store(aux, map: meta->map_ptr,
11243	unpriv: !meta->map_ptr->bypass_spec_v1, poison: false);
11244	else if (aux->map_ptr_state.map_ptr != meta->map_ptr)
11245	bpf_map_ptr_store(aux, map: meta->map_ptr,
11246	unpriv: !meta->map_ptr->bypass_spec_v1, poison: true);
11247	return 0;
11248	}
11249
11250	static int
11251	record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
11252	int func_id, int insn_idx)
11253	{
11254	struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
11255	struct bpf_reg_state *regs = cur_regs(env), *reg;
11256	struct bpf_map *map = meta->map_ptr;
11257	u64 val, max;
11258	int err;
11259
11260	if (func_id != BPF_FUNC_tail_call)
11261	return 0;
11262	if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
11263	verbose(private_data: env, fmt: "expected prog array map for tail call");
11264	return -EINVAL;
11265	}
11266
11267	reg = &regs[BPF_REG_3];
11268	val = reg->var_off.value;
11269	max = map->max_entries;
11270
11271	if (!(is_reg_const(reg, subreg32: false) && val < max)) {
11272	bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
11273	return 0;
11274	}
11275
11276	err = mark_chain_precision(env, regno: BPF_REG_3);
11277	if (err)
11278	return err;
11279	if (bpf_map_key_unseen(aux))
11280	bpf_map_key_store(aux, state: val);
11281	else if (!bpf_map_key_poisoned(aux) &&
11282	bpf_map_key_immediate(aux) != val)
11283	bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
11284	return 0;
11285	}
11286
11287	static int check_reference_leak(struct bpf_verifier_env *env, bool exception_exit)
11288	{
11289	struct bpf_verifier_state *state = env->cur_state;
11290	enum bpf_prog_type type = resolve_prog_type(prog: env->prog);
11291	struct bpf_reg_state *reg = reg_state(env, regno: BPF_REG_0);
11292	bool refs_lingering = false;
11293	int i;
11294
11295	if (!exception_exit && cur_func(env)->frameno)
11296	return 0;
11297
11298	for (i = 0; i < state->acquired_refs; i++) {
11299	if (state->refs[i].type != REF_TYPE_PTR)
11300	continue;
11301	/* Allow struct_ops programs to return a referenced kptr back to
11302	* kernel. Type checks are performed later in check_return_code.
11303	*/
11304	if (type == BPF_PROG_TYPE_STRUCT_OPS && !exception_exit &&
11305	reg->ref_obj_id == state->refs[i].id)
11306	continue;
11307	verbose(private_data: env, fmt: "Unreleased reference id=%d alloc_insn=%d\n",
11308	state->refs[i].id, state->refs[i].insn_idx);
11309	refs_lingering = true;
11310	}
11311	return refs_lingering ? -EINVAL : 0;
11312	}
11313
11314	static int check_resource_leak(struct bpf_verifier_env *env, bool exception_exit, bool check_lock, const char *prefix)
11315	{
11316	int err;
11317
11318	if (check_lock && env->cur_state->active_locks) {
11319	verbose(private_data: env, fmt: "%s cannot be used inside bpf_spin_lock-ed region\n", prefix);
11320	return -EINVAL;
11321	}
11322
11323	err = check_reference_leak(env, exception_exit);
11324	if (err) {
11325	verbose(private_data: env, fmt: "%s would lead to reference leak\n", prefix);
11326	return err;
11327	}
11328
11329	if (check_lock && env->cur_state->active_irq_id) {
11330	verbose(private_data: env, fmt: "%s cannot be used inside bpf_local_irq_save-ed region\n", prefix);
11331	return -EINVAL;
11332	}
11333
11334	if (check_lock && env->cur_state->active_rcu_locks) {
11335	verbose(private_data: env, fmt: "%s cannot be used inside bpf_rcu_read_lock-ed region\n", prefix);
11336	return -EINVAL;
11337	}
11338
11339	if (check_lock && env->cur_state->active_preempt_locks) {
11340	verbose(private_data: env, fmt: "%s cannot be used inside bpf_preempt_disable-ed region\n", prefix);
11341	return -EINVAL;
11342	}
11343
11344	return 0;
11345	}
11346
11347	static int check_bpf_snprintf_call(struct bpf_verifier_env *env,
11348	struct bpf_reg_state *regs)
11349	{
11350	struct bpf_reg_state *fmt_reg = &regs[BPF_REG_3];
11351	struct bpf_reg_state *data_len_reg = &regs[BPF_REG_5];
11352	struct bpf_map *fmt_map = fmt_reg->map_ptr;
11353	struct bpf_bprintf_data data = {};
11354	int err, fmt_map_off, num_args;
11355	u64 fmt_addr;
11356	char *fmt;
11357
11358	/* data must be an array of u64 */
11359	if (data_len_reg->var_off.value % 8)
11360	return -EINVAL;
11361	num_args = data_len_reg->var_off.value / 8;
11362
11363	/* fmt being ARG_PTR_TO_CONST_STR guarantees that var_off is const
11364	* and map_direct_value_addr is set.
11365	*/
11366	fmt_map_off = fmt_reg->off + fmt_reg->var_off.value;
11367	err = fmt_map->ops->map_direct_value_addr(fmt_map, &fmt_addr,
11368	fmt_map_off);
11369	if (err) {
11370	verbose(private_data: env, fmt: "failed to retrieve map value address\n");
11371	return -EFAULT;
11372	}
11373	fmt = (char *)(long)fmt_addr + fmt_map_off;
11374
11375	/* We are also guaranteed that fmt+fmt_map_off is NULL terminated, we
11376	* can focus on validating the format specifiers.
11377	*/
11378	err = bpf_bprintf_prepare(fmt, UINT_MAX, NULL, num_args, data: &data);
11379	if (err < 0)
11380	verbose(private_data: env, fmt: "Invalid format string\n");
11381
11382	return err;
11383	}
11384
11385	static int check_get_func_ip(struct bpf_verifier_env *env)
11386	{
11387	enum bpf_prog_type type = resolve_prog_type(prog: env->prog);
11388	int func_id = BPF_FUNC_get_func_ip;
11389
11390	if (type == BPF_PROG_TYPE_TRACING) {
11391	if (!bpf_prog_has_trampoline(prog: env->prog)) {
11392	verbose(private_data: env, fmt: "func %s#%d supported only for fentry/fexit/fmod_ret programs\n",
11393	func_id_name(id: func_id), func_id);
11394	return -ENOTSUPP;
11395	}
11396	return 0;
11397	} else if (type == BPF_PROG_TYPE_KPROBE) {
11398	return 0;
11399	}
11400
11401	verbose(private_data: env, fmt: "func %s#%d not supported for program type %d\n",
11402	func_id_name(id: func_id), func_id, type);
11403	return -ENOTSUPP;
11404	}
11405
11406	static struct bpf_insn_aux_data *cur_aux(const struct bpf_verifier_env *env)
11407	{
11408	return &env->insn_aux_data[env->insn_idx];
11409	}
11410
11411	static bool loop_flag_is_zero(struct bpf_verifier_env *env)
11412	{
11413	struct bpf_reg_state *regs = cur_regs(env);
11414	struct bpf_reg_state *reg = &regs[BPF_REG_4];
11415	bool reg_is_null = register_is_null(reg);
11416
11417	if (reg_is_null)
11418	mark_chain_precision(env, regno: BPF_REG_4);
11419
11420	return reg_is_null;
11421	}
11422
11423	static void update_loop_inline_state(struct bpf_verifier_env *env, u32 subprogno)
11424	{
11425	struct bpf_loop_inline_state *state = &cur_aux(env)->loop_inline_state;
11426
11427	if (!state->initialized) {
11428	state->initialized = 1;
11429	state->fit_for_inline = loop_flag_is_zero(env);
11430	state->callback_subprogno = subprogno;
11431	return;
11432	}
11433
11434	if (!state->fit_for_inline)
11435	return;
11436
11437	state->fit_for_inline = (loop_flag_is_zero(env) &&
11438	state->callback_subprogno == subprogno);
11439	}
11440
11441	/* Returns whether or not the given map type can potentially elide
11442	* lookup return value nullness check. This is possible if the key
11443	* is statically known.
11444	*/
11445	static bool can_elide_value_nullness(enum bpf_map_type type)
11446	{
11447	switch (type) {
11448	case BPF_MAP_TYPE_ARRAY:
11449	case BPF_MAP_TYPE_PERCPU_ARRAY:
11450	return true;
11451	default:
11452	return false;
11453	}
11454	}
11455
11456	static int get_helper_proto(struct bpf_verifier_env *env, int func_id,
11457	const struct bpf_func_proto **ptr)
11458	{
11459	if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID)
11460	return -ERANGE;
11461
11462	if (!env->ops->get_func_proto)
11463	return -EINVAL;
11464
11465	*ptr = env->ops->get_func_proto(func_id, env->prog);
11466	return *ptr && (*ptr)->func ? 0 : -EINVAL;
11467	}
11468
11469	/* Check if we're in a sleepable context. */
11470	static inline bool in_sleepable_context(struct bpf_verifier_env *env)
11471	{
11472	return !env->cur_state->active_rcu_locks &&
11473	!env->cur_state->active_preempt_locks &&
11474	!env->cur_state->active_irq_id &&
11475	in_sleepable(env);
11476	}
11477
11478	static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
11479	int *insn_idx_p)
11480	{
11481	enum bpf_prog_type prog_type = resolve_prog_type(prog: env->prog);
11482	bool returns_cpu_specific_alloc_ptr = false;
11483	const struct bpf_func_proto *fn = NULL;
11484	enum bpf_return_type ret_type;
11485	enum bpf_type_flag ret_flag;
11486	struct bpf_reg_state *regs;
11487	struct bpf_call_arg_meta meta;
11488	int insn_idx = *insn_idx_p;
11489	bool changes_data;
11490	int i, err, func_id;
11491
11492	/* find function prototype */
11493	func_id = insn->imm;
11494	err = get_helper_proto(env, func_id: insn->imm, ptr: &fn);
11495	if (err == -ERANGE) {
11496	verbose(private_data: env, fmt: "invalid func %s#%d\n", func_id_name(id: func_id), func_id);
11497	return -EINVAL;
11498	}
11499
11500	if (err) {
11501	verbose(private_data: env, fmt: "program of this type cannot use helper %s#%d\n",
11502	func_id_name(id: func_id), func_id);
11503	return err;
11504	}
11505
11506	/* eBPF programs must be GPL compatible to use GPL-ed functions */
11507	if (!env->prog->gpl_compatible && fn->gpl_only) {
11508	verbose(private_data: env, fmt: "cannot call GPL-restricted function from non-GPL compatible program\n");
11509	return -EINVAL;
11510	}
11511
11512	if (fn->allowed && !fn->allowed(env->prog)) {
11513	verbose(private_data: env, fmt: "helper call is not allowed in probe\n");
11514	return -EINVAL;
11515	}
11516
11517	if (!in_sleepable(env) && fn->might_sleep) {
11518	verbose(private_data: env, fmt: "helper call might sleep in a non-sleepable prog\n");
11519	return -EINVAL;
11520	}
11521
11522	/* With LD_ABS/IND some JITs save/restore skb from r1. */
11523	changes_data = bpf_helper_changes_pkt_data(func_id);
11524	if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
11525	verifier_bug(env, "func %s#%d: r1 != ctx", func_id_name(func_id), func_id);
11526	return -EFAULT;
11527	}
11528
11529	memset(&meta, 0, sizeof(meta));
11530	meta.pkt_access = fn->pkt_access;
11531
11532	err = check_func_proto(fn, func_id);
11533	if (err) {
11534	verifier_bug(env, "incorrect func proto %s#%d", func_id_name(func_id), func_id);
11535	return err;
11536	}
11537
11538	if (env->cur_state->active_rcu_locks) {
11539	if (fn->might_sleep) {
11540	verbose(private_data: env, fmt: "sleepable helper %s#%d in rcu_read_lock region\n",
11541	func_id_name(id: func_id), func_id);
11542	return -EINVAL;
11543	}
11544	}
11545
11546	if (env->cur_state->active_preempt_locks) {
11547	if (fn->might_sleep) {
11548	verbose(private_data: env, fmt: "sleepable helper %s#%d in non-preemptible region\n",
11549	func_id_name(id: func_id), func_id);
11550	return -EINVAL;
11551	}
11552	}
11553
11554	if (env->cur_state->active_irq_id) {
11555	if (fn->might_sleep) {
11556	verbose(private_data: env, fmt: "sleepable helper %s#%d in IRQ-disabled region\n",
11557	func_id_name(id: func_id), func_id);
11558	return -EINVAL;
11559	}
11560	}
11561
11562	/* Track non-sleepable context for helpers. */
11563	if (!in_sleepable_context(env))
11564	env->insn_aux_data[insn_idx].non_sleepable = true;
11565
11566	meta.func_id = func_id;
11567	/* check args */
11568	for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++) {
11569	err = check_func_arg(env, arg: i, meta: &meta, fn, insn_idx);
11570	if (err)
11571	return err;
11572	}
11573
11574	err = record_func_map(env, meta: &meta, func_id, insn_idx);
11575	if (err)
11576	return err;
11577
11578	err = record_func_key(env, meta: &meta, func_id, insn_idx);
11579	if (err)
11580	return err;
11581
11582	/* Mark slots with STACK_MISC in case of raw mode, stack offset
11583	* is inferred from register state.
11584	*/
11585	for (i = 0; i < meta.access_size; i++) {
11586	err = check_mem_access(env, insn_idx, regno: meta.regno, off: i, BPF_B,
11587	t: BPF_WRITE, value_regno: -1, strict_alignment_once: false, is_ldsx: false);
11588	if (err)
11589	return err;
11590	}
11591
11592	regs = cur_regs(env);
11593
11594	if (meta.release_regno) {
11595	err = -EINVAL;
11596	if (arg_type_is_dynptr(type: fn->arg_type[meta.release_regno - BPF_REG_1])) {
11597	err = unmark_stack_slots_dynptr(env, reg: &regs[meta.release_regno]);
11598	} else if (func_id == BPF_FUNC_kptr_xchg && meta.ref_obj_id) {
11599	u32 ref_obj_id = meta.ref_obj_id;
11600	bool in_rcu = in_rcu_cs(env);
11601	struct bpf_func_state *state;
11602	struct bpf_reg_state *reg;
11603
11604	err = release_reference_nomark(state: env->cur_state, ref_obj_id);
11605	if (!err) {
11606	bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
11607	if (reg->ref_obj_id == ref_obj_id) {
11608	if (in_rcu && (reg->type & MEM_ALLOC) && (reg->type & MEM_PERCPU)) {
11609	reg->ref_obj_id = 0;
11610	reg->type &= ~MEM_ALLOC;
11611	reg->type |= MEM_RCU;
11612	} else {
11613	mark_reg_invalid(env, reg);
11614	}
11615	}
11616	}));
11617	}
11618	} else if (meta.ref_obj_id) {
11619	err = release_reference(env, ref_obj_id: meta.ref_obj_id);
11620	} else if (register_is_null(reg: &regs[meta.release_regno])) {
11621	/* meta.ref_obj_id can only be 0 if register that is meant to be
11622	* released is NULL, which must be > R0.
11623	*/
11624	err = 0;
11625	}
11626	if (err) {
11627	verbose(private_data: env, fmt: "func %s#%d reference has not been acquired before\n",
11628	func_id_name(id: func_id), func_id);
11629	return err;
11630	}
11631	}
11632
11633	switch (func_id) {
11634	case BPF_FUNC_tail_call:
11635	err = check_resource_leak(env, exception_exit: false, check_lock: true, prefix: "tail_call");
11636	if (err)
11637	return err;
11638	break;
11639	case BPF_FUNC_get_local_storage:
11640	/* check that flags argument in get_local_storage(map, flags) is 0,
11641	* this is required because get_local_storage() can't return an error.
11642	*/
11643	if (!register_is_null(reg: &regs[BPF_REG_2])) {
11644	verbose(private_data: env, fmt: "get_local_storage() doesn't support non-zero flags\n");
11645	return -EINVAL;
11646	}
11647	break;
11648	case BPF_FUNC_for_each_map_elem:
11649	err = push_callback_call(env, insn, insn_idx, subprog: meta.subprogno,
11650	set_callee_state_cb: set_map_elem_callback_state);
11651	break;
11652	case BPF_FUNC_timer_set_callback:
11653	err = push_callback_call(env, insn, insn_idx, subprog: meta.subprogno,
11654	set_callee_state_cb: set_timer_callback_state);
11655	break;
11656	case BPF_FUNC_find_vma:
11657	err = push_callback_call(env, insn, insn_idx, subprog: meta.subprogno,
11658	set_callee_state_cb: set_find_vma_callback_state);
11659	break;
11660	case BPF_FUNC_snprintf:
11661	err = check_bpf_snprintf_call(env, regs);
11662	break;
11663	case BPF_FUNC_loop:
11664	update_loop_inline_state(env, subprogno: meta.subprogno);
11665	/* Verifier relies on R1 value to determine if bpf_loop() iteration
11666	* is finished, thus mark it precise.
11667	*/
11668	err = mark_chain_precision(env, regno: BPF_REG_1);
11669	if (err)
11670	return err;
11671	if (cur_func(env)->callback_depth < regs[BPF_REG_1].umax_value) {
11672	err = push_callback_call(env, insn, insn_idx, subprog: meta.subprogno,
11673	set_callee_state_cb: set_loop_callback_state);
11674	} else {
11675	cur_func(env)->callback_depth = 0;
11676	if (env->log.level & BPF_LOG_LEVEL2)
11677	verbose(private_data: env, fmt: "frame%d bpf_loop iteration limit reached\n",
11678	env->cur_state->curframe);
11679	}
11680	break;
11681	case BPF_FUNC_dynptr_from_mem:
11682	if (regs[BPF_REG_1].type != PTR_TO_MAP_VALUE) {
11683	verbose(private_data: env, fmt: "Unsupported reg type %s for bpf_dynptr_from_mem data\n",
11684	reg_type_str(env, type: regs[BPF_REG_1].type));
11685	return -EACCES;
11686	}
11687	break;
11688	case BPF_FUNC_set_retval:
11689	if (prog_type == BPF_PROG_TYPE_LSM &&
11690	env->prog->expected_attach_type == BPF_LSM_CGROUP) {
11691	if (!env->prog->aux->attach_func_proto->type) {
11692	/* Make sure programs that attach to void
11693	* hooks don't try to modify return value.
11694	*/
11695	verbose(private_data: env, fmt: "BPF_LSM_CGROUP that attach to void LSM hooks can't modify return value!\n");
11696	return -EINVAL;
11697	}
11698	}
11699	break;
11700	case BPF_FUNC_dynptr_data:
11701	{
11702	struct bpf_reg_state *reg;
11703	int id, ref_obj_id;
11704
11705	reg = get_dynptr_arg_reg(env, fn, regs);
11706	if (!reg)
11707	return -EFAULT;
11708
11709
11710	if (meta.dynptr_id) {
11711	verifier_bug(env, "meta.dynptr_id already set");
11712	return -EFAULT;
11713	}
11714	if (meta.ref_obj_id) {
11715	verifier_bug(env, "meta.ref_obj_id already set");
11716	return -EFAULT;
11717	}
11718
11719	id = dynptr_id(env, reg);
11720	if (id < 0) {
11721	verifier_bug(env, "failed to obtain dynptr id");
11722	return id;
11723	}
11724
11725	ref_obj_id = dynptr_ref_obj_id(env, reg);
11726	if (ref_obj_id < 0) {
11727	verifier_bug(env, "failed to obtain dynptr ref_obj_id");
11728	return ref_obj_id;
11729	}
11730
11731	meta.dynptr_id = id;
11732	meta.ref_obj_id = ref_obj_id;
11733
11734	break;
11735	}
11736	case BPF_FUNC_dynptr_write:
11737	{
11738	enum bpf_dynptr_type dynptr_type;
11739	struct bpf_reg_state *reg;
11740
11741	reg = get_dynptr_arg_reg(env, fn, regs);
11742	if (!reg)
11743	return -EFAULT;
11744
11745	dynptr_type = dynptr_get_type(env, reg);
11746	if (dynptr_type == BPF_DYNPTR_TYPE_INVALID)
11747	return -EFAULT;
11748
11749	if (dynptr_type == BPF_DYNPTR_TYPE_SKB ||
11750	dynptr_type == BPF_DYNPTR_TYPE_SKB_META)
11751	/* this will trigger clear_all_pkt_pointers(), which will
11752	* invalidate all dynptr slices associated with the skb
11753	*/
11754	changes_data = true;
11755
11756	break;
11757	}
11758	case BPF_FUNC_per_cpu_ptr:
11759	case BPF_FUNC_this_cpu_ptr:
11760	{
11761	struct bpf_reg_state *reg = &regs[BPF_REG_1];
11762	const struct btf_type *type;
11763
11764	if (reg->type & MEM_RCU) {
11765	type = btf_type_by_id(btf: reg->btf, type_id: reg->btf_id);
11766	if (!type || !btf_type_is_struct(t: type)) {
11767	verbose(private_data: env, fmt: "Helper has invalid btf/btf_id in R1\n");
11768	return -EFAULT;
11769	}
11770	returns_cpu_specific_alloc_ptr = true;
11771	env->insn_aux_data[insn_idx].call_with_percpu_alloc_ptr = true;
11772	}
11773	break;
11774	}
11775	case BPF_FUNC_user_ringbuf_drain:
11776	err = push_callback_call(env, insn, insn_idx, subprog: meta.subprogno,
11777	set_callee_state_cb: set_user_ringbuf_callback_state);
11778	break;
11779	}
11780
11781	if (err)
11782	return err;
11783
11784	/* reset caller saved regs */
11785	for (i = 0; i < CALLER_SAVED_REGS; i++) {
11786	mark_reg_not_init(env, regs, regno: caller_saved[i]);
11787	check_reg_arg(env, regno: caller_saved[i], t: DST_OP_NO_MARK);
11788	}
11789
11790	/* helper call returns 64-bit value. */
11791	regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
11792
11793	/* update return register (already marked as written above) */
11794	ret_type = fn->ret_type;
11795	ret_flag = type_flag(type: ret_type);
11796
11797	switch (base_type(type: ret_type)) {
11798	case RET_INTEGER:
11799	/* sets type to SCALAR_VALUE */
11800	mark_reg_unknown(env, regs, regno: BPF_REG_0);
11801	break;
11802	case RET_VOID:
11803	regs[BPF_REG_0].type = NOT_INIT;
11804	break;
11805	case RET_PTR_TO_MAP_VALUE:
11806	/* There is no offset yet applied, variable or fixed */
11807	mark_reg_known_zero(env, regs, regno: BPF_REG_0);
11808	/* remember map_ptr, so that check_map_access()
11809	* can check 'value_size' boundary of memory access
11810	* to map element returned from bpf_map_lookup_elem()
11811	*/
11812	if (meta.map_ptr == NULL) {
11813	verifier_bug(env, "unexpected null map_ptr");
11814	return -EFAULT;
11815	}
11816
11817	if (func_id == BPF_FUNC_map_lookup_elem &&
11818	can_elide_value_nullness(type: meta.map_ptr->map_type) &&
11819	meta.const_map_key >= 0 &&
11820	meta.const_map_key < meta.map_ptr->max_entries)
11821	ret_flag &= ~PTR_MAYBE_NULL;
11822
11823	regs[BPF_REG_0].map_ptr = meta.map_ptr;
11824	regs[BPF_REG_0].map_uid = meta.map_uid;
11825	regs[BPF_REG_0].type = PTR_TO_MAP_VALUE | ret_flag;
11826	if (!type_may_be_null(type: ret_flag) &&
11827	btf_record_has_field(rec: meta.map_ptr->record, type: BPF_SPIN_LOCK | BPF_RES_SPIN_LOCK)) {
11828	regs[BPF_REG_0].id = ++env->id_gen;
11829	}
11830	break;
11831	case RET_PTR_TO_SOCKET:
11832	mark_reg_known_zero(env, regs, regno: BPF_REG_0);
11833	regs[BPF_REG_0].type = PTR_TO_SOCKET | ret_flag;
11834	break;
11835	case RET_PTR_TO_SOCK_COMMON:
11836	mark_reg_known_zero(env, regs, regno: BPF_REG_0);
11837	regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON | ret_flag;
11838	break;
11839	case RET_PTR_TO_TCP_SOCK:
11840	mark_reg_known_zero(env, regs, regno: BPF_REG_0);
11841	regs[BPF_REG_0].type = PTR_TO_TCP_SOCK | ret_flag;
11842	break;
11843	case RET_PTR_TO_MEM:
11844	mark_reg_known_zero(env, regs, regno: BPF_REG_0);
11845	regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
11846	regs[BPF_REG_0].mem_size = meta.mem_size;
11847	break;
11848	case RET_PTR_TO_MEM_OR_BTF_ID:
11849	{
11850	const struct btf_type *t;
11851
11852	mark_reg_known_zero(env, regs, regno: BPF_REG_0);
11853	t = btf_type_skip_modifiers(btf: meta.ret_btf, id: meta.ret_btf_id, NULL);
11854	if (!btf_type_is_struct(t)) {
11855	u32 tsize;
11856	const struct btf_type *ret;
11857	const char *tname;
11858
11859	/* resolve the type size of ksym. */
11860	ret = btf_resolve_size(btf: meta.ret_btf, type: t, type_size: &tsize);
11861	if (IS_ERR(ptr: ret)) {
11862	tname = btf_name_by_offset(btf: meta.ret_btf, offset: t->name_off);
11863	verbose(private_data: env, fmt: "unable to resolve the size of type '%s': %ld\n",
11864	tname, PTR_ERR(ptr: ret));
11865	return -EINVAL;
11866	}
11867	regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
11868	regs[BPF_REG_0].mem_size = tsize;
11869	} else {
11870	if (returns_cpu_specific_alloc_ptr) {
11871	regs[BPF_REG_0].type = PTR_TO_BTF_ID | MEM_ALLOC | MEM_RCU;
11872	} else {
11873	/* MEM_RDONLY may be carried from ret_flag, but it
11874	* doesn't apply on PTR_TO_BTF_ID. Fold it, otherwise
11875	* it will confuse the check of PTR_TO_BTF_ID in
11876	* check_mem_access().
11877	*/
11878	ret_flag &= ~MEM_RDONLY;
11879	regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
11880	}
11881
11882	regs[BPF_REG_0].btf = meta.ret_btf;
11883	regs[BPF_REG_0].btf_id = meta.ret_btf_id;
11884	}
11885	break;
11886	}
11887	case RET_PTR_TO_BTF_ID:
11888	{
11889	struct btf *ret_btf;
11890	int ret_btf_id;
11891
11892	mark_reg_known_zero(env, regs, regno: BPF_REG_0);
11893	regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
11894	if (func_id == BPF_FUNC_kptr_xchg) {
11895	ret_btf = meta.kptr_field->kptr.btf;
11896	ret_btf_id = meta.kptr_field->kptr.btf_id;
11897	if (!btf_is_kernel(btf: ret_btf)) {
11898	regs[BPF_REG_0].type |= MEM_ALLOC;
11899	if (meta.kptr_field->type == BPF_KPTR_PERCPU)
11900	regs[BPF_REG_0].type |= MEM_PERCPU;
11901	}
11902	} else {
11903	if (fn->ret_btf_id == BPF_PTR_POISON) {
11904	verifier_bug(env, "func %s has non-overwritten BPF_PTR_POISON return type",
11905	func_id_name(func_id));
11906	return -EFAULT;
11907	}
11908	ret_btf = btf_vmlinux;
11909	ret_btf_id = *fn->ret_btf_id;
11910	}
11911	if (ret_btf_id == 0) {
11912	verbose(private_data: env, fmt: "invalid return type %u of func %s#%d\n",
11913	base_type(type: ret_type), func_id_name(id: func_id),
11914	func_id);
11915	return -EINVAL;
11916	}
11917	regs[BPF_REG_0].btf = ret_btf;
11918	regs[BPF_REG_0].btf_id = ret_btf_id;
11919	break;
11920	}
11921	default:
11922	verbose(private_data: env, fmt: "unknown return type %u of func %s#%d\n",
11923	base_type(type: ret_type), func_id_name(id: func_id), func_id);
11924	return -EINVAL;
11925	}
11926
11927	if (type_may_be_null(type: regs[BPF_REG_0].type))
11928	regs[BPF_REG_0].id = ++env->id_gen;
11929
11930	if (helper_multiple_ref_obj_use(func_id, map: meta.map_ptr)) {
11931	verifier_bug(env, "func %s#%d sets ref_obj_id more than once",
11932	func_id_name(func_id), func_id);
11933	return -EFAULT;
11934	}
11935
11936	if (is_dynptr_ref_function(func_id))
11937	regs[BPF_REG_0].dynptr_id = meta.dynptr_id;
11938
11939	if (is_ptr_cast_function(func_id) || is_dynptr_ref_function(func_id)) {
11940	/* For release_reference() */
11941	regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
11942	} else if (is_acquire_function(func_id, map: meta.map_ptr)) {
11943	int id = acquire_reference(env, insn_idx);
11944
11945	if (id < 0)
11946	return id;
11947	/* For mark_ptr_or_null_reg() */
11948	regs[BPF_REG_0].id = id;
11949	/* For release_reference() */
11950	regs[BPF_REG_0].ref_obj_id = id;
11951	}
11952
11953	err = do_refine_retval_range(env, regs, ret_type: fn->ret_type, func_id, meta: &meta);
11954	if (err)
11955	return err;
11956
11957	err = check_map_func_compatibility(env, map: meta.map_ptr, func_id);
11958	if (err)
11959	return err;
11960
11961	if ((func_id == BPF_FUNC_get_stack ||
11962	func_id == BPF_FUNC_get_task_stack) &&
11963	!env->prog->has_callchain_buf) {
11964	const char *err_str;
11965
11966	#ifdef CONFIG_PERF_EVENTS
11967	err = get_callchain_buffers(max_stack: sysctl_perf_event_max_stack);
11968	err_str = "cannot get callchain buffer for func %s#%d\n";
11969	#else
11970	err = -ENOTSUPP;
11971	err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
11972	#endif
11973	if (err) {
11974	verbose(private_data: env, fmt: err_str, func_id_name(id: func_id), func_id);
11975	return err;
11976	}
11977
11978	env->prog->has_callchain_buf = true;
11979	}
11980
11981	if (func_id == BPF_FUNC_get_stackid || func_id == BPF_FUNC_get_stack)
11982	env->prog->call_get_stack = true;
11983
11984	if (func_id == BPF_FUNC_get_func_ip) {
11985	if (check_get_func_ip(env))
11986	return -ENOTSUPP;
11987	env->prog->call_get_func_ip = true;
11988	}
11989
11990	if (func_id == BPF_FUNC_tail_call) {
11991	if (env->cur_state->curframe) {
11992	struct bpf_verifier_state *branch;
11993
11994	mark_reg_scratched(env, regno: BPF_REG_0);
11995	branch = push_stack(env, insn_idx: env->insn_idx + 1, prev_insn_idx: env->insn_idx, speculative: false);
11996	if (IS_ERR(ptr: branch))
11997	return PTR_ERR(ptr: branch);
11998	clear_all_pkt_pointers(env);
11999	mark_reg_unknown(env, regs, regno: BPF_REG_0);
12000	err = prepare_func_exit(env, insn_idx: &env->insn_idx);
12001	if (err)
12002	return err;
12003	env->insn_idx--;
12004	} else {
12005	changes_data = false;
12006	}
12007	}
12008
12009	if (changes_data)
12010	clear_all_pkt_pointers(env);
12011	return 0;
12012	}
12013
12014	/* mark_btf_func_reg_size() is used when the reg size is determined by
12015	* the BTF func_proto's return value size and argument.
12016	*/
12017	static void __mark_btf_func_reg_size(struct bpf_verifier_env *env, struct bpf_reg_state *regs,
12018	u32 regno, size_t reg_size)
12019	{
12020	struct bpf_reg_state *reg = &regs[regno];
12021
12022	if (regno == BPF_REG_0) {
12023	/* Function return value */
12024	reg->subreg_def = reg_size == sizeof(u64) ?
12025	DEF_NOT_SUBREG : env->insn_idx + 1;
12026	} else if (reg_size == sizeof(u64)) {
12027	/* Function argument */
12028	mark_insn_zext(env, reg);
12029	}
12030	}
12031
12032	static void mark_btf_func_reg_size(struct bpf_verifier_env *env, u32 regno,
12033	size_t reg_size)
12034	{
12035	return __mark_btf_func_reg_size(env, regs: cur_regs(env), regno, reg_size);
12036	}
12037
12038	static bool is_kfunc_acquire(struct bpf_kfunc_call_arg_meta *meta)
12039	{
12040	return meta->kfunc_flags & KF_ACQUIRE;
12041	}
12042
12043	static bool is_kfunc_release(struct bpf_kfunc_call_arg_meta *meta)
12044	{
12045	return meta->kfunc_flags & KF_RELEASE;
12046	}
12047
12048	static bool is_kfunc_trusted_args(struct bpf_kfunc_call_arg_meta *meta)
12049	{
12050	return (meta->kfunc_flags & KF_TRUSTED_ARGS) || is_kfunc_release(meta);
12051	}
12052
12053	static bool is_kfunc_sleepable(struct bpf_kfunc_call_arg_meta *meta)
12054	{
12055	return meta->kfunc_flags & KF_SLEEPABLE;
12056	}
12057
12058	static bool is_kfunc_destructive(struct bpf_kfunc_call_arg_meta *meta)
12059	{
12060	return meta->kfunc_flags & KF_DESTRUCTIVE;
12061	}
12062
12063	static bool is_kfunc_rcu(struct bpf_kfunc_call_arg_meta *meta)
12064	{
12065	return meta->kfunc_flags & KF_RCU;
12066	}
12067
12068	static bool is_kfunc_rcu_protected(struct bpf_kfunc_call_arg_meta *meta)
12069	{
12070	return meta->kfunc_flags & KF_RCU_PROTECTED;
12071	}
12072
12073	static bool is_kfunc_arg_mem_size(const struct btf *btf,
12074	const struct btf_param *arg,
12075	const struct bpf_reg_state *reg)
12076	{
12077	const struct btf_type *t;
12078
12079	t = btf_type_skip_modifiers(btf, id: arg->type, NULL);
12080	if (!btf_type_is_scalar(t) || reg->type != SCALAR_VALUE)
12081	return false;
12082
12083	return btf_param_match_suffix(btf, arg, suffix: "__sz");
12084	}
12085
12086	static bool is_kfunc_arg_const_mem_size(const struct btf *btf,
12087	const struct btf_param *arg,
12088	const struct bpf_reg_state *reg)
12089	{
12090	const struct btf_type *t;
12091
12092	t = btf_type_skip_modifiers(btf, id: arg->type, NULL);
12093	if (!btf_type_is_scalar(t) || reg->type != SCALAR_VALUE)
12094	return false;
12095
12096	return btf_param_match_suffix(btf, arg, suffix: "__szk");
12097	}
12098
12099	static bool is_kfunc_arg_optional(const struct btf *btf, const struct btf_param *arg)
12100	{
12101	return btf_param_match_suffix(btf, arg, suffix: "__opt");
12102	}
12103
12104	static bool is_kfunc_arg_constant(const struct btf *btf, const struct btf_param *arg)
12105	{
12106	return btf_param_match_suffix(btf, arg, suffix: "__k");
12107	}
12108
12109	static bool is_kfunc_arg_ignore(const struct btf *btf, const struct btf_param *arg)
12110	{
12111	return btf_param_match_suffix(btf, arg, suffix: "__ign");
12112	}
12113
12114	static bool is_kfunc_arg_map(const struct btf *btf, const struct btf_param *arg)
12115	{
12116	return btf_param_match_suffix(btf, arg, suffix: "__map");
12117	}
12118
12119	static bool is_kfunc_arg_alloc_obj(const struct btf *btf, const struct btf_param *arg)
12120	{
12121	return btf_param_match_suffix(btf, arg, suffix: "__alloc");
12122	}
12123
12124	static bool is_kfunc_arg_uninit(const struct btf *btf, const struct btf_param *arg)
12125	{
12126	return btf_param_match_suffix(btf, arg, suffix: "__uninit");
12127	}
12128
12129	static bool is_kfunc_arg_refcounted_kptr(const struct btf *btf, const struct btf_param *arg)
12130	{
12131	return btf_param_match_suffix(btf, arg, suffix: "__refcounted_kptr");
12132	}
12133
12134	static bool is_kfunc_arg_nullable(const struct btf *btf, const struct btf_param *arg)
12135	{
12136	return btf_param_match_suffix(btf, arg, suffix: "__nullable");
12137	}
12138
12139	static bool is_kfunc_arg_const_str(const struct btf *btf, const struct btf_param *arg)
12140	{
12141	return btf_param_match_suffix(btf, arg, suffix: "__str");
12142	}
12143
12144	static bool is_kfunc_arg_irq_flag(const struct btf *btf, const struct btf_param *arg)
12145	{
12146	return btf_param_match_suffix(btf, arg, suffix: "__irq_flag");
12147	}
12148
12149	static bool is_kfunc_arg_prog(const struct btf *btf, const struct btf_param *arg)
12150	{
12151	return btf_param_match_suffix(btf, arg, suffix: "__prog");
12152	}
12153
12154	static bool is_kfunc_arg_scalar_with_name(const struct btf *btf,
12155	const struct btf_param *arg,
12156	const char *name)
12157	{
12158	int len, target_len = strlen(name);
12159	const char *param_name;
12160
12161	param_name = btf_name_by_offset(btf, offset: arg->name_off);
12162	if (str_is_empty(s: param_name))
12163	return false;
12164	len = strlen(param_name);
12165	if (len != target_len)
12166	return false;
12167	if (strcmp(param_name, name))
12168	return false;
12169
12170	return true;
12171	}
12172
12173	enum {
12174	KF_ARG_DYNPTR_ID,
12175	KF_ARG_LIST_HEAD_ID,
12176	KF_ARG_LIST_NODE_ID,
12177	KF_ARG_RB_ROOT_ID,
12178	KF_ARG_RB_NODE_ID,
12179	KF_ARG_WORKQUEUE_ID,
12180	KF_ARG_RES_SPIN_LOCK_ID,
12181	KF_ARG_TASK_WORK_ID,
12182	};
12183
12184	BTF_ID_LIST(kf_arg_btf_ids)
12185	BTF_ID(struct, bpf_dynptr)
12186	BTF_ID(struct, bpf_list_head)
12187	BTF_ID(struct, bpf_list_node)
12188	BTF_ID(struct, bpf_rb_root)
12189	BTF_ID(struct, bpf_rb_node)
12190	BTF_ID(struct, bpf_wq)
12191	BTF_ID(struct, bpf_res_spin_lock)
12192	BTF_ID(struct, bpf_task_work)
12193
12194	static bool __is_kfunc_ptr_arg_type(const struct btf *btf,
12195	const struct btf_param *arg, int type)
12196	{
12197	const struct btf_type *t;
12198	u32 res_id;
12199
12200	t = btf_type_skip_modifiers(btf, id: arg->type, NULL);
12201	if (!t)
12202	return false;
12203	if (!btf_type_is_ptr(t))
12204	return false;
12205	t = btf_type_skip_modifiers(btf, id: t->type, res_id: &res_id);
12206	if (!t)
12207	return false;
12208	return btf_types_are_same(btf1: btf, id1: res_id, btf2: btf_vmlinux, id2: kf_arg_btf_ids[type]);
12209	}
12210
12211	static bool is_kfunc_arg_dynptr(const struct btf *btf, const struct btf_param *arg)
12212	{
12213	return __is_kfunc_ptr_arg_type(btf, arg, type: KF_ARG_DYNPTR_ID);
12214	}
12215
12216	static bool is_kfunc_arg_list_head(const struct btf *btf, const struct btf_param *arg)
12217	{
12218	return __is_kfunc_ptr_arg_type(btf, arg, type: KF_ARG_LIST_HEAD_ID);
12219	}
12220
12221	static bool is_kfunc_arg_list_node(const struct btf *btf, const struct btf_param *arg)
12222	{
12223	return __is_kfunc_ptr_arg_type(btf, arg, type: KF_ARG_LIST_NODE_ID);
12224	}
12225
12226	static bool is_kfunc_arg_rbtree_root(const struct btf *btf, const struct btf_param *arg)
12227	{
12228	return __is_kfunc_ptr_arg_type(btf, arg, type: KF_ARG_RB_ROOT_ID);
12229	}
12230
12231	static bool is_kfunc_arg_rbtree_node(const struct btf *btf, const struct btf_param *arg)
12232	{
12233	return __is_kfunc_ptr_arg_type(btf, arg, type: KF_ARG_RB_NODE_ID);
12234	}
12235
12236	static bool is_kfunc_arg_wq(const struct btf *btf, const struct btf_param *arg)
12237	{
12238	return __is_kfunc_ptr_arg_type(btf, arg, type: KF_ARG_WORKQUEUE_ID);
12239	}
12240
12241	static bool is_kfunc_arg_task_work(const struct btf *btf, const struct btf_param *arg)
12242	{
12243	return __is_kfunc_ptr_arg_type(btf, arg, type: KF_ARG_TASK_WORK_ID);
12244	}
12245
12246	static bool is_kfunc_arg_res_spin_lock(const struct btf *btf, const struct btf_param *arg)
12247	{
12248	return __is_kfunc_ptr_arg_type(btf, arg, type: KF_ARG_RES_SPIN_LOCK_ID);
12249	}
12250
12251	static bool is_rbtree_node_type(const struct btf_type *t)
12252	{
12253	return t == btf_type_by_id(btf: btf_vmlinux, type_id: kf_arg_btf_ids[KF_ARG_RB_NODE_ID]);
12254	}
12255
12256	static bool is_list_node_type(const struct btf_type *t)
12257	{
12258	return t == btf_type_by_id(btf: btf_vmlinux, type_id: kf_arg_btf_ids[KF_ARG_LIST_NODE_ID]);
12259	}
12260
12261	static bool is_kfunc_arg_callback(struct bpf_verifier_env *env, const struct btf *btf,
12262	const struct btf_param *arg)
12263	{
12264	const struct btf_type *t;
12265
12266	t = btf_type_resolve_func_ptr(btf, id: arg->type, NULL);
12267	if (!t)
12268	return false;
12269
12270	return true;
12271	}
12272
12273	/* Returns true if struct is composed of scalars, 4 levels of nesting allowed */
12274	static bool __btf_type_is_scalar_struct(struct bpf_verifier_env *env,
12275	const struct btf *btf,
12276	const struct btf_type *t, int rec)
12277	{
12278	const struct btf_type *member_type;
12279	const struct btf_member *member;
12280	u32 i;
12281
12282	if (!btf_type_is_struct(t))
12283	return false;
12284
12285	for_each_member(i, t, member) {
12286	const struct btf_array *array;
12287
12288	member_type = btf_type_skip_modifiers(btf, id: member->type, NULL);
12289	if (btf_type_is_struct(t: member_type)) {
12290	if (rec >= 3) {
12291	verbose(private_data: env, fmt: "max struct nesting depth exceeded\n");
12292	return false;
12293	}
12294	if (!__btf_type_is_scalar_struct(env, btf, t: member_type, rec: rec + 1))
12295	return false;
12296	continue;
12297	}
12298	if (btf_type_is_array(t: member_type)) {
12299	array = btf_array(t: member_type);
12300	if (!array->nelems)
12301	return false;
12302	member_type = btf_type_skip_modifiers(btf, id: array->type, NULL);
12303	if (!btf_type_is_scalar(t: member_type))
12304	return false;
12305	continue;
12306	}
12307	if (!btf_type_is_scalar(t: member_type))
12308	return false;
12309	}
12310	return true;
12311	}
12312
12313	enum kfunc_ptr_arg_type {
12314	KF_ARG_PTR_TO_CTX,
12315	KF_ARG_PTR_TO_ALLOC_BTF_ID, /* Allocated object */
12316	KF_ARG_PTR_TO_REFCOUNTED_KPTR, /* Refcounted local kptr */
12317	KF_ARG_PTR_TO_DYNPTR,
12318	KF_ARG_PTR_TO_ITER,
12319	KF_ARG_PTR_TO_LIST_HEAD,
12320	KF_ARG_PTR_TO_LIST_NODE,
12321	KF_ARG_PTR_TO_BTF_ID, /* Also covers reg2btf_ids conversions */
12322	KF_ARG_PTR_TO_MEM,
12323	KF_ARG_PTR_TO_MEM_SIZE, /* Size derived from next argument, skip it */
12324	KF_ARG_PTR_TO_CALLBACK,
12325	KF_ARG_PTR_TO_RB_ROOT,
12326	KF_ARG_PTR_TO_RB_NODE,
12327	KF_ARG_PTR_TO_NULL,
12328	KF_ARG_PTR_TO_CONST_STR,
12329	KF_ARG_PTR_TO_MAP,
12330	KF_ARG_PTR_TO_WORKQUEUE,
12331	KF_ARG_PTR_TO_IRQ_FLAG,
12332	KF_ARG_PTR_TO_RES_SPIN_LOCK,
12333	KF_ARG_PTR_TO_TASK_WORK,
12334	};
12335
12336	enum special_kfunc_type {
12337	KF_bpf_obj_new_impl,
12338	KF_bpf_obj_drop_impl,
12339	KF_bpf_refcount_acquire_impl,
12340	KF_bpf_list_push_front_impl,
12341	KF_bpf_list_push_back_impl,
12342	KF_bpf_list_pop_front,
12343	KF_bpf_list_pop_back,
12344	KF_bpf_list_front,
12345	KF_bpf_list_back,
12346	KF_bpf_cast_to_kern_ctx,
12347	KF_bpf_rdonly_cast,
12348	KF_bpf_rcu_read_lock,
12349	KF_bpf_rcu_read_unlock,
12350	KF_bpf_rbtree_remove,
12351	KF_bpf_rbtree_add_impl,
12352	KF_bpf_rbtree_first,
12353	KF_bpf_rbtree_root,
12354	KF_bpf_rbtree_left,
12355	KF_bpf_rbtree_right,
12356	KF_bpf_dynptr_from_skb,
12357	KF_bpf_dynptr_from_xdp,
12358	KF_bpf_dynptr_from_skb_meta,
12359	KF_bpf_xdp_pull_data,
12360	KF_bpf_dynptr_slice,
12361	KF_bpf_dynptr_slice_rdwr,
12362	KF_bpf_dynptr_clone,
12363	KF_bpf_percpu_obj_new_impl,
12364	KF_bpf_percpu_obj_drop_impl,
12365	KF_bpf_throw,
12366	KF_bpf_wq_set_callback_impl,
12367	KF_bpf_preempt_disable,
12368	KF_bpf_preempt_enable,
12369	KF_bpf_iter_css_task_new,
12370	KF_bpf_session_cookie,
12371	KF_bpf_get_kmem_cache,
12372	KF_bpf_local_irq_save,
12373	KF_bpf_local_irq_restore,
12374	KF_bpf_iter_num_new,
12375	KF_bpf_iter_num_next,
12376	KF_bpf_iter_num_destroy,
12377	KF_bpf_set_dentry_xattr,
12378	KF_bpf_remove_dentry_xattr,
12379	KF_bpf_res_spin_lock,
12380	KF_bpf_res_spin_unlock,
12381	KF_bpf_res_spin_lock_irqsave,
12382	KF_bpf_res_spin_unlock_irqrestore,
12383	KF_bpf_dynptr_from_file,
12384	KF_bpf_dynptr_file_discard,
12385	KF___bpf_trap,
12386	KF_bpf_task_work_schedule_signal_impl,
12387	KF_bpf_task_work_schedule_resume_impl,
12388	};
12389
12390	BTF_ID_LIST(special_kfunc_list)
12391	BTF_ID(func, bpf_obj_new_impl)
12392	BTF_ID(func, bpf_obj_drop_impl)
12393	BTF_ID(func, bpf_refcount_acquire_impl)
12394	BTF_ID(func, bpf_list_push_front_impl)
12395	BTF_ID(func, bpf_list_push_back_impl)
12396	BTF_ID(func, bpf_list_pop_front)
12397	BTF_ID(func, bpf_list_pop_back)
12398	BTF_ID(func, bpf_list_front)
12399	BTF_ID(func, bpf_list_back)
12400	BTF_ID(func, bpf_cast_to_kern_ctx)
12401	BTF_ID(func, bpf_rdonly_cast)
12402	BTF_ID(func, bpf_rcu_read_lock)
12403	BTF_ID(func, bpf_rcu_read_unlock)
12404	BTF_ID(func, bpf_rbtree_remove)
12405	BTF_ID(func, bpf_rbtree_add_impl)
12406	BTF_ID(func, bpf_rbtree_first)
12407	BTF_ID(func, bpf_rbtree_root)
12408	BTF_ID(func, bpf_rbtree_left)
12409	BTF_ID(func, bpf_rbtree_right)
12410	#ifdef CONFIG_NET
12411	BTF_ID(func, bpf_dynptr_from_skb)
12412	BTF_ID(func, bpf_dynptr_from_xdp)
12413	BTF_ID(func, bpf_dynptr_from_skb_meta)
12414	BTF_ID(func, bpf_xdp_pull_data)
12415	#else
12416	BTF_ID_UNUSED
12417	BTF_ID_UNUSED
12418	BTF_ID_UNUSED
12419	BTF_ID_UNUSED
12420	#endif
12421	BTF_ID(func, bpf_dynptr_slice)
12422	BTF_ID(func, bpf_dynptr_slice_rdwr)
12423	BTF_ID(func, bpf_dynptr_clone)
12424	BTF_ID(func, bpf_percpu_obj_new_impl)
12425	BTF_ID(func, bpf_percpu_obj_drop_impl)
12426	BTF_ID(func, bpf_throw)
12427	BTF_ID(func, bpf_wq_set_callback_impl)
12428	BTF_ID(func, bpf_preempt_disable)
12429	BTF_ID(func, bpf_preempt_enable)
12430	#ifdef CONFIG_CGROUPS
12431	BTF_ID(func, bpf_iter_css_task_new)
12432	#else
12433	BTF_ID_UNUSED
12434	#endif
12435	#ifdef CONFIG_BPF_EVENTS
12436	BTF_ID(func, bpf_session_cookie)
12437	#else
12438	BTF_ID_UNUSED
12439	#endif
12440	BTF_ID(func, bpf_get_kmem_cache)
12441	BTF_ID(func, bpf_local_irq_save)
12442	BTF_ID(func, bpf_local_irq_restore)
12443	BTF_ID(func, bpf_iter_num_new)
12444	BTF_ID(func, bpf_iter_num_next)
12445	BTF_ID(func, bpf_iter_num_destroy)
12446	#ifdef CONFIG_BPF_LSM
12447	BTF_ID(func, bpf_set_dentry_xattr)
12448	BTF_ID(func, bpf_remove_dentry_xattr)
12449	#else
12450	BTF_ID_UNUSED
12451	BTF_ID_UNUSED
12452	#endif
12453	BTF_ID(func, bpf_res_spin_lock)
12454	BTF_ID(func, bpf_res_spin_unlock)
12455	BTF_ID(func, bpf_res_spin_lock_irqsave)
12456	BTF_ID(func, bpf_res_spin_unlock_irqrestore)
12457	BTF_ID(func, bpf_dynptr_from_file)
12458	BTF_ID(func, bpf_dynptr_file_discard)
12459	BTF_ID(func, __bpf_trap)
12460	BTF_ID(func, bpf_task_work_schedule_signal_impl)
12461	BTF_ID(func, bpf_task_work_schedule_resume_impl)
12462
12463	static bool is_task_work_add_kfunc(u32 func_id)
12464	{
12465	return func_id == special_kfunc_list[KF_bpf_task_work_schedule_signal_impl] ||
12466	func_id == special_kfunc_list[KF_bpf_task_work_schedule_resume_impl];
12467	}
12468
12469	static bool is_kfunc_ret_null(struct bpf_kfunc_call_arg_meta *meta)
12470	{
12471	if (meta->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl] &&
12472	meta->arg_owning_ref) {
12473	return false;
12474	}
12475
12476	return meta->kfunc_flags & KF_RET_NULL;
12477	}
12478
12479	static bool is_kfunc_bpf_rcu_read_lock(struct bpf_kfunc_call_arg_meta *meta)
12480	{
12481	return meta->func_id == special_kfunc_list[KF_bpf_rcu_read_lock];
12482	}
12483
12484	static bool is_kfunc_bpf_rcu_read_unlock(struct bpf_kfunc_call_arg_meta *meta)
12485	{
12486	return meta->func_id == special_kfunc_list[KF_bpf_rcu_read_unlock];
12487	}
12488
12489	static bool is_kfunc_bpf_preempt_disable(struct bpf_kfunc_call_arg_meta *meta)
12490	{
12491	return meta->func_id == special_kfunc_list[KF_bpf_preempt_disable];
12492	}
12493
12494	static bool is_kfunc_bpf_preempt_enable(struct bpf_kfunc_call_arg_meta *meta)
12495	{
12496	return meta->func_id == special_kfunc_list[KF_bpf_preempt_enable];
12497	}
12498
12499	static bool is_kfunc_pkt_changing(struct bpf_kfunc_call_arg_meta *meta)
12500	{
12501	return meta->func_id == special_kfunc_list[KF_bpf_xdp_pull_data];
12502	}
12503
12504	static enum kfunc_ptr_arg_type
12505	get_kfunc_ptr_arg_type(struct bpf_verifier_env *env,
12506	struct bpf_kfunc_call_arg_meta *meta,
12507	const struct btf_type *t, const struct btf_type *ref_t,
12508	const char *ref_tname, const struct btf_param *args,
12509	int argno, int nargs)
12510	{
12511	u32 regno = argno + 1;
12512	struct bpf_reg_state *regs = cur_regs(env);
12513	struct bpf_reg_state *reg = &regs[regno];
12514	bool arg_mem_size = false;
12515
12516	if (meta->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx])
12517	return KF_ARG_PTR_TO_CTX;
12518
12519	/* In this function, we verify the kfunc's BTF as per the argument type,
12520	* leaving the rest of the verification with respect to the register
12521	* type to our caller. When a set of conditions hold in the BTF type of
12522	* arguments, we resolve it to a known kfunc_ptr_arg_type.
12523	*/
12524	if (btf_is_prog_ctx_type(log: &env->log, btf: meta->btf, t, prog_type: resolve_prog_type(prog: env->prog), arg: argno))
12525	return KF_ARG_PTR_TO_CTX;
12526
12527	if (is_kfunc_arg_nullable(btf: meta->btf, arg: &args[argno]) && register_is_null(reg))
12528	return KF_ARG_PTR_TO_NULL;
12529
12530	if (is_kfunc_arg_alloc_obj(btf: meta->btf, arg: &args[argno]))
12531	return KF_ARG_PTR_TO_ALLOC_BTF_ID;
12532
12533	if (is_kfunc_arg_refcounted_kptr(btf: meta->btf, arg: &args[argno]))
12534	return KF_ARG_PTR_TO_REFCOUNTED_KPTR;
12535
12536	if (is_kfunc_arg_dynptr(btf: meta->btf, arg: &args[argno]))
12537	return KF_ARG_PTR_TO_DYNPTR;
12538
12539	if (is_kfunc_arg_iter(meta, arg_idx: argno, arg: &args[argno]))
12540	return KF_ARG_PTR_TO_ITER;
12541
12542	if (is_kfunc_arg_list_head(btf: meta->btf, arg: &args[argno]))
12543	return KF_ARG_PTR_TO_LIST_HEAD;
12544
12545	if (is_kfunc_arg_list_node(btf: meta->btf, arg: &args[argno]))
12546	return KF_ARG_PTR_TO_LIST_NODE;
12547
12548	if (is_kfunc_arg_rbtree_root(btf: meta->btf, arg: &args[argno]))
12549	return KF_ARG_PTR_TO_RB_ROOT;
12550
12551	if (is_kfunc_arg_rbtree_node(btf: meta->btf, arg: &args[argno]))
12552	return KF_ARG_PTR_TO_RB_NODE;
12553
12554	if (is_kfunc_arg_const_str(btf: meta->btf, arg: &args[argno]))
12555	return KF_ARG_PTR_TO_CONST_STR;
12556
12557	if (is_kfunc_arg_map(btf: meta->btf, arg: &args[argno]))
12558	return KF_ARG_PTR_TO_MAP;
12559
12560	if (is_kfunc_arg_wq(btf: meta->btf, arg: &args[argno]))
12561	return KF_ARG_PTR_TO_WORKQUEUE;
12562
12563	if (is_kfunc_arg_task_work(btf: meta->btf, arg: &args[argno]))
12564	return KF_ARG_PTR_TO_TASK_WORK;
12565
12566	if (is_kfunc_arg_irq_flag(btf: meta->btf, arg: &args[argno]))
12567	return KF_ARG_PTR_TO_IRQ_FLAG;
12568
12569	if (is_kfunc_arg_res_spin_lock(btf: meta->btf, arg: &args[argno]))
12570	return KF_ARG_PTR_TO_RES_SPIN_LOCK;
12571
12572	if ((base_type(type: reg->type) == PTR_TO_BTF_ID || reg2btf_ids[base_type(type: reg->type)])) {
12573	if (!btf_type_is_struct(t: ref_t)) {
12574	verbose(private_data: env, fmt: "kernel function %s args#%d pointer type %s %s is not supported\n",
12575	meta->func_name, argno, btf_type_str(t: ref_t), ref_tname);
12576	return -EINVAL;
12577	}
12578	return KF_ARG_PTR_TO_BTF_ID;
12579	}
12580
12581	if (is_kfunc_arg_callback(env, btf: meta->btf, arg: &args[argno]))
12582	return KF_ARG_PTR_TO_CALLBACK;
12583
12584	if (argno + 1 < nargs &&
12585	(is_kfunc_arg_mem_size(btf: meta->btf, arg: &args[argno + 1], reg: &regs[regno + 1]) ||
12586	is_kfunc_arg_const_mem_size(btf: meta->btf, arg: &args[argno + 1], reg: &regs[regno + 1])))
12587	arg_mem_size = true;
12588
12589	/* This is the catch all argument type of register types supported by
12590	* check_helper_mem_access. However, we only allow when argument type is
12591	* pointer to scalar, or struct composed (recursively) of scalars. When
12592	* arg_mem_size is true, the pointer can be void *.
12593	*/
12594	if (!btf_type_is_scalar(t: ref_t) && !__btf_type_is_scalar_struct(env, btf: meta->btf, t: ref_t, rec: 0) &&
12595	(arg_mem_size ? !btf_type_is_void(t: ref_t) : 1)) {
12596	verbose(private_data: env, fmt: "arg#%d pointer type %s %s must point to %sscalar, or struct with scalar\n",
12597	argno, btf_type_str(t: ref_t), ref_tname, arg_mem_size ? "void, " : "");
12598	return -EINVAL;
12599	}
12600	return arg_mem_size ? KF_ARG_PTR_TO_MEM_SIZE : KF_ARG_PTR_TO_MEM;
12601	}
12602
12603	static int process_kf_arg_ptr_to_btf_id(struct bpf_verifier_env *env,
12604	struct bpf_reg_state *reg,
12605	const struct btf_type *ref_t,
12606	const char *ref_tname, u32 ref_id,
12607	struct bpf_kfunc_call_arg_meta *meta,
12608	int argno)
12609	{
12610	const struct btf_type *reg_ref_t;
12611	bool strict_type_match = false;
12612	const struct btf *reg_btf;
12613	const char *reg_ref_tname;
12614	bool taking_projection;
12615	bool struct_same;
12616	u32 reg_ref_id;
12617
12618	if (base_type(type: reg->type) == PTR_TO_BTF_ID) {
12619	reg_btf = reg->btf;
12620	reg_ref_id = reg->btf_id;
12621	} else {
12622	reg_btf = btf_vmlinux;
12623	reg_ref_id = *reg2btf_ids[base_type(type: reg->type)];
12624	}
12625
12626	/* Enforce strict type matching for calls to kfuncs that are acquiring
12627	* or releasing a reference, or are no-cast aliases. We do _not_
12628	* enforce strict matching for plain KF_TRUSTED_ARGS kfuncs by default,
12629	* as we want to enable BPF programs to pass types that are bitwise
12630	* equivalent without forcing them to explicitly cast with something
12631	* like bpf_cast_to_kern_ctx().
12632	*
12633	* For example, say we had a type like the following:
12634	*
12635	* struct bpf_cpumask {
12636	* cpumask_t cpumask;
12637	* refcount_t usage;
12638	* };
12639	*
12640	* Note that as specified in <linux/cpumask.h>, cpumask_t is typedef'ed
12641	* to a struct cpumask, so it would be safe to pass a struct
12642	* bpf_cpumask * to a kfunc expecting a struct cpumask *.
12643	*
12644	* The philosophy here is similar to how we allow scalars of different
12645	* types to be passed to kfuncs as long as the size is the same. The
12646	* only difference here is that we're simply allowing
12647	* btf_struct_ids_match() to walk the struct at the 0th offset, and
12648	* resolve types.
12649	*/
12650	if ((is_kfunc_release(meta) && reg->ref_obj_id) ||
12651	btf_type_ids_nocast_alias(log: &env->log, reg_btf, reg_id: reg_ref_id, arg_btf: meta->btf, arg_id: ref_id))
12652	strict_type_match = true;
12653
12654	WARN_ON_ONCE(is_kfunc_release(meta) &&
12655	(reg->off || !tnum_is_const(reg->var_off) ||
12656	reg->var_off.value));
12657
12658	reg_ref_t = btf_type_skip_modifiers(btf: reg_btf, id: reg_ref_id, res_id: &reg_ref_id);
12659	reg_ref_tname = btf_name_by_offset(btf: reg_btf, offset: reg_ref_t->name_off);
12660	struct_same = btf_struct_ids_match(log: &env->log, btf: reg_btf, id: reg_ref_id, off: reg->off, need_btf: meta->btf, need_type_id: ref_id, strict: strict_type_match);
12661	/* If kfunc is accepting a projection type (ie. __sk_buff), it cannot
12662	* actually use it -- it must cast to the underlying type. So we allow
12663	* caller to pass in the underlying type.
12664	*/
12665	taking_projection = btf_is_projection_of(pname: ref_tname, tname: reg_ref_tname);
12666	if (!taking_projection && !struct_same) {
12667	verbose(private_data: env, fmt: "kernel function %s args#%d expected pointer to %s %s but R%d has a pointer to %s %s\n",
12668	meta->func_name, argno, btf_type_str(t: ref_t), ref_tname, argno + 1,
12669	btf_type_str(t: reg_ref_t), reg_ref_tname);
12670	return -EINVAL;
12671	}
12672	return 0;
12673	}
12674
12675	static int process_irq_flag(struct bpf_verifier_env *env, int regno,
12676	struct bpf_kfunc_call_arg_meta *meta)
12677	{
12678	struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[regno];
12679	int err, kfunc_class = IRQ_NATIVE_KFUNC;
12680	bool irq_save;
12681
12682	if (meta->func_id == special_kfunc_list[KF_bpf_local_irq_save] ||
12683	meta->func_id == special_kfunc_list[KF_bpf_res_spin_lock_irqsave]) {
12684	irq_save = true;
12685	if (meta->func_id == special_kfunc_list[KF_bpf_res_spin_lock_irqsave])
12686	kfunc_class = IRQ_LOCK_KFUNC;
12687	} else if (meta->func_id == special_kfunc_list[KF_bpf_local_irq_restore] ||
12688	meta->func_id == special_kfunc_list[KF_bpf_res_spin_unlock_irqrestore]) {
12689	irq_save = false;
12690	if (meta->func_id == special_kfunc_list[KF_bpf_res_spin_unlock_irqrestore])
12691	kfunc_class = IRQ_LOCK_KFUNC;
12692	} else {
12693	verifier_bug(env, "unknown irq flags kfunc");
12694	return -EFAULT;
12695	}
12696
12697	if (irq_save) {
12698	if (!is_irq_flag_reg_valid_uninit(env, reg)) {
12699	verbose(private_data: env, fmt: "expected uninitialized irq flag as arg#%d\n", regno - 1);
12700	return -EINVAL;
12701	}
12702
12703	err = check_mem_access(env, insn_idx: env->insn_idx, regno, off: 0, BPF_DW, t: BPF_WRITE, value_regno: -1, strict_alignment_once: false, is_ldsx: false);
12704	if (err)
12705	return err;
12706
12707	err = mark_stack_slot_irq_flag(env, meta, reg, insn_idx: env->insn_idx, kfunc_class);
12708	if (err)
12709	return err;
12710	} else {
12711	err = is_irq_flag_reg_valid_init(env, reg);
12712	if (err) {
12713	verbose(private_data: env, fmt: "expected an initialized irq flag as arg#%d\n", regno - 1);
12714	return err;
12715	}
12716
12717	err = mark_irq_flag_read(env, reg);
12718	if (err)
12719	return err;
12720
12721	err = unmark_stack_slot_irq_flag(env, reg, kfunc_class);
12722	if (err)
12723	return err;
12724	}
12725	return 0;
12726	}
12727
12728
12729	static int ref_set_non_owning(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
12730	{
12731	struct btf_record *rec = reg_btf_record(reg);
12732
12733	if (!env->cur_state->active_locks) {
12734	verifier_bug(env, "%s w/o active lock", __func__);
12735	return -EFAULT;
12736	}
12737
12738	if (type_flag(type: reg->type) & NON_OWN_REF) {
12739	verifier_bug(env, "NON_OWN_REF already set");
12740	return -EFAULT;
12741	}
12742
12743	reg->type |= NON_OWN_REF;
12744	if (rec->refcount_off >= 0)
12745	reg->type |= MEM_RCU;
12746
12747	return 0;
12748	}
12749
12750	static int ref_convert_owning_non_owning(struct bpf_verifier_env *env, u32 ref_obj_id)
12751	{
12752	struct bpf_verifier_state *state = env->cur_state;
12753	struct bpf_func_state *unused;
12754	struct bpf_reg_state *reg;
12755	int i;
12756
12757	if (!ref_obj_id) {
12758	verifier_bug(env, "ref_obj_id is zero for owning -> non-owning conversion");
12759	return -EFAULT;
12760	}
12761
12762	for (i = 0; i < state->acquired_refs; i++) {
12763	if (state->refs[i].id != ref_obj_id)
12764	continue;
12765
12766	/* Clear ref_obj_id here so release_reference doesn't clobber
12767	* the whole reg
12768	*/
12769	bpf_for_each_reg_in_vstate(env->cur_state, unused, reg, ({
12770	if (reg->ref_obj_id == ref_obj_id) {
12771	reg->ref_obj_id = 0;
12772	ref_set_non_owning(env, reg);
12773	}
12774	}));
12775	return 0;
12776	}
12777
12778	verifier_bug(env, "ref state missing for ref_obj_id");
12779	return -EFAULT;
12780	}
12781
12782	/* Implementation details:
12783	*
12784	* Each register points to some region of memory, which we define as an
12785	* allocation. Each allocation may embed a bpf_spin_lock which protects any
12786	* special BPF objects (bpf_list_head, bpf_rb_root, etc.) part of the same
12787	* allocation. The lock and the data it protects are colocated in the same
12788	* memory region.
12789	*
12790	* Hence, everytime a register holds a pointer value pointing to such
12791	* allocation, the verifier preserves a unique reg->id for it.
12792	*
12793	* The verifier remembers the lock 'ptr' and the lock 'id' whenever
12794	* bpf_spin_lock is called.
12795	*
12796	* To enable this, lock state in the verifier captures two values:
12797	* active_lock.ptr = Register's type specific pointer
12798	* active_lock.id = A unique ID for each register pointer value
12799	*
12800	* Currently, PTR_TO_MAP_VALUE and PTR_TO_BTF_ID | MEM_ALLOC are the two
12801	* supported register types.
12802	*
12803	* The active_lock.ptr in case of map values is the reg->map_ptr, and in case of
12804	* allocated objects is the reg->btf pointer.
12805	*
12806	* The active_lock.id is non-unique for maps supporting direct_value_addr, as we
12807	* can establish the provenance of the map value statically for each distinct
12808	* lookup into such maps. They always contain a single map value hence unique
12809	* IDs for each pseudo load pessimizes the algorithm and rejects valid programs.
12810	*
12811	* So, in case of global variables, they use array maps with max_entries = 1,
12812	* hence their active_lock.ptr becomes map_ptr and id = 0 (since they all point
12813	* into the same map value as max_entries is 1, as described above).
12814	*
12815	* In case of inner map lookups, the inner map pointer has same map_ptr as the
12816	* outer map pointer (in verifier context), but each lookup into an inner map
12817	* assigns a fresh reg->id to the lookup, so while lookups into distinct inner
12818	* maps from the same outer map share the same map_ptr as active_lock.ptr, they
12819	* will get different reg->id assigned to each lookup, hence different
12820	* active_lock.id.
12821	*
12822	* In case of allocated objects, active_lock.ptr is the reg->btf, and the
12823	* reg->id is a unique ID preserved after the NULL pointer check on the pointer
12824	* returned from bpf_obj_new. Each allocation receives a new reg->id.
12825	*/
12826	static int check_reg_allocation_locked(struct bpf_verifier_env *env, struct bpf_reg_state *reg)
12827	{
12828	struct bpf_reference_state *s;
12829	void *ptr;
12830	u32 id;
12831
12832	switch ((int)reg->type) {
12833	case PTR_TO_MAP_VALUE:
12834	ptr = reg->map_ptr;
12835	break;
12836	case PTR_TO_BTF_ID | MEM_ALLOC:
12837	ptr = reg->btf;
12838	break;
12839	default:
12840	verifier_bug(env, "unknown reg type for lock check");
12841	return -EFAULT;
12842	}
12843	id = reg->id;
12844
12845	if (!env->cur_state->active_locks)
12846	return -EINVAL;
12847	s = find_lock_state(state: env->cur_state, type: REF_TYPE_LOCK_MASK, id, ptr);
12848	if (!s) {
12849	verbose(private_data: env, fmt: "held lock and object are not in the same allocation\n");
12850	return -EINVAL;
12851	}
12852	return 0;
12853	}
12854
12855	static bool is_bpf_list_api_kfunc(u32 btf_id)
12856	{
12857	return btf_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
12858	btf_id == special_kfunc_list[KF_bpf_list_push_back_impl] ||
12859	btf_id == special_kfunc_list[KF_bpf_list_pop_front] ||
12860	btf_id == special_kfunc_list[KF_bpf_list_pop_back] ||
12861	btf_id == special_kfunc_list[KF_bpf_list_front] ||
12862	btf_id == special_kfunc_list[KF_bpf_list_back];
12863	}
12864
12865	static bool is_bpf_rbtree_api_kfunc(u32 btf_id)
12866	{
12867	return btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl] ||
12868	btf_id == special_kfunc_list[KF_bpf_rbtree_remove] ||
12869	btf_id == special_kfunc_list[KF_bpf_rbtree_first] ||
12870	btf_id == special_kfunc_list[KF_bpf_rbtree_root] ||
12871	btf_id == special_kfunc_list[KF_bpf_rbtree_left] ||
12872	btf_id == special_kfunc_list[KF_bpf_rbtree_right];
12873	}
12874
12875	static bool is_bpf_iter_num_api_kfunc(u32 btf_id)
12876	{
12877	return btf_id == special_kfunc_list[KF_bpf_iter_num_new] ||
12878	btf_id == special_kfunc_list[KF_bpf_iter_num_next] ||
12879	btf_id == special_kfunc_list[KF_bpf_iter_num_destroy];
12880	}
12881
12882	static bool is_bpf_graph_api_kfunc(u32 btf_id)
12883	{
12884	return is_bpf_list_api_kfunc(btf_id) || is_bpf_rbtree_api_kfunc(btf_id) ||
12885	btf_id == special_kfunc_list[KF_bpf_refcount_acquire_impl];
12886	}
12887
12888	static bool is_bpf_res_spin_lock_kfunc(u32 btf_id)
12889	{
12890	return btf_id == special_kfunc_list[KF_bpf_res_spin_lock] ||
12891	btf_id == special_kfunc_list[KF_bpf_res_spin_unlock] ||
12892	btf_id == special_kfunc_list[KF_bpf_res_spin_lock_irqsave] ||
12893	btf_id == special_kfunc_list[KF_bpf_res_spin_unlock_irqrestore];
12894	}
12895
12896	static bool kfunc_spin_allowed(u32 btf_id)
12897	{
12898	return is_bpf_graph_api_kfunc(btf_id) || is_bpf_iter_num_api_kfunc(btf_id) ||
12899	is_bpf_res_spin_lock_kfunc(btf_id);
12900	}
12901
12902	static bool is_sync_callback_calling_kfunc(u32 btf_id)
12903	{
12904	return btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl];
12905	}
12906
12907	static bool is_async_callback_calling_kfunc(u32 btf_id)
12908	{
12909	return btf_id == special_kfunc_list[KF_bpf_wq_set_callback_impl] ||
12910	is_task_work_add_kfunc(func_id: btf_id);
12911	}
12912
12913	static bool is_bpf_throw_kfunc(struct bpf_insn *insn)
12914	{
12915	return bpf_pseudo_kfunc_call(insn) && insn->off == 0 &&
12916	insn->imm == special_kfunc_list[KF_bpf_throw];
12917	}
12918
12919	static bool is_bpf_wq_set_callback_impl_kfunc(u32 btf_id)
12920	{
12921	return btf_id == special_kfunc_list[KF_bpf_wq_set_callback_impl];
12922	}
12923
12924	static bool is_callback_calling_kfunc(u32 btf_id)
12925	{
12926	return is_sync_callback_calling_kfunc(btf_id) ||
12927	is_async_callback_calling_kfunc(btf_id);
12928	}
12929
12930	static bool is_rbtree_lock_required_kfunc(u32 btf_id)
12931	{
12932	return is_bpf_rbtree_api_kfunc(btf_id);
12933	}
12934
12935	static bool check_kfunc_is_graph_root_api(struct bpf_verifier_env *env,
12936	enum btf_field_type head_field_type,
12937	u32 kfunc_btf_id)
12938	{
12939	bool ret;
12940
12941	switch (head_field_type) {
12942	case BPF_LIST_HEAD:
12943	ret = is_bpf_list_api_kfunc(btf_id: kfunc_btf_id);
12944	break;
12945	case BPF_RB_ROOT:
12946	ret = is_bpf_rbtree_api_kfunc(btf_id: kfunc_btf_id);
12947	break;
12948	default:
12949	verbose(private_data: env, fmt: "verifier internal error: unexpected graph root argument type %s\n",
12950	btf_field_type_name(type: head_field_type));
12951	return false;
12952	}
12953
12954	if (!ret)
12955	verbose(private_data: env, fmt: "verifier internal error: %s head arg for unknown kfunc\n",
12956	btf_field_type_name(type: head_field_type));
12957	return ret;
12958	}
12959
12960	static bool check_kfunc_is_graph_node_api(struct bpf_verifier_env *env,
12961	enum btf_field_type node_field_type,
12962	u32 kfunc_btf_id)
12963	{
12964	bool ret;
12965
12966	switch (node_field_type) {
12967	case BPF_LIST_NODE:
12968	ret = (kfunc_btf_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
12969	kfunc_btf_id == special_kfunc_list[KF_bpf_list_push_back_impl]);
12970	break;
12971	case BPF_RB_NODE:
12972	ret = (kfunc_btf_id == special_kfunc_list[KF_bpf_rbtree_remove] ||
12973	kfunc_btf_id == special_kfunc_list[KF_bpf_rbtree_add_impl] ||
12974	kfunc_btf_id == special_kfunc_list[KF_bpf_rbtree_left] ||
12975	kfunc_btf_id == special_kfunc_list[KF_bpf_rbtree_right]);
12976	break;
12977	default:
12978	verbose(private_data: env, fmt: "verifier internal error: unexpected graph node argument type %s\n",
12979	btf_field_type_name(type: node_field_type));
12980	return false;
12981	}
12982
12983	if (!ret)
12984	verbose(private_data: env, fmt: "verifier internal error: %s node arg for unknown kfunc\n",
12985	btf_field_type_name(type: node_field_type));
12986	return ret;
12987	}
12988
12989	static int
12990	__process_kf_arg_ptr_to_graph_root(struct bpf_verifier_env *env,
12991	struct bpf_reg_state *reg, u32 regno,
12992	struct bpf_kfunc_call_arg_meta *meta,
12993	enum btf_field_type head_field_type,
12994	struct btf_field **head_field)
12995	{
12996	const char *head_type_name;
12997	struct btf_field *field;
12998	struct btf_record *rec;
12999	u32 head_off;
13000
13001	if (meta->btf != btf_vmlinux) {
13002	verifier_bug(env, "unexpected btf mismatch in kfunc call");
13003	return -EFAULT;
13004	}
13005
13006	if (!check_kfunc_is_graph_root_api(env, head_field_type, kfunc_btf_id: meta->func_id))
13007	return -EFAULT;
13008
13009	head_type_name = btf_field_type_name(type: head_field_type);
13010	if (!tnum_is_const(a: reg->var_off)) {
13011	verbose(private_data: env,
13012	fmt: "R%d doesn't have constant offset. %s has to be at the constant offset\n",
13013	regno, head_type_name);
13014	return -EINVAL;
13015	}
13016
13017	rec = reg_btf_record(reg);
13018	head_off = reg->off + reg->var_off.value;
13019	field = btf_record_find(rec, offset: head_off, field_mask: head_field_type);
13020	if (!field) {
13021	verbose(private_data: env, fmt: "%s not found at offset=%u\n", head_type_name, head_off);
13022	return -EINVAL;
13023	}
13024
13025	/* All functions require bpf_list_head to be protected using a bpf_spin_lock */
13026	if (check_reg_allocation_locked(env, reg)) {
13027	verbose(private_data: env, fmt: "bpf_spin_lock at off=%d must be held for %s\n",
13028	rec->spin_lock_off, head_type_name);
13029	return -EINVAL;
13030	}
13031
13032	if (*head_field) {
13033	verifier_bug(env, "repeating %s arg", head_type_name);
13034	return -EFAULT;
13035	}
13036	*head_field = field;
13037	return 0;
13038	}
13039
13040	static int process_kf_arg_ptr_to_list_head(struct bpf_verifier_env *env,
13041	struct bpf_reg_state *reg, u32 regno,
13042	struct bpf_kfunc_call_arg_meta *meta)
13043	{
13044	return __process_kf_arg_ptr_to_graph_root(env, reg, regno, meta, head_field_type: BPF_LIST_HEAD,
13045	head_field: &meta->arg_list_head.field);
13046	}
13047
13048	static int process_kf_arg_ptr_to_rbtree_root(struct bpf_verifier_env *env,
13049	struct bpf_reg_state *reg, u32 regno,
13050	struct bpf_kfunc_call_arg_meta *meta)
13051	{
13052	return __process_kf_arg_ptr_to_graph_root(env, reg, regno, meta, head_field_type: BPF_RB_ROOT,
13053	head_field: &meta->arg_rbtree_root.field);
13054	}
13055
13056	static int
13057	__process_kf_arg_ptr_to_graph_node(struct bpf_verifier_env *env,
13058	struct bpf_reg_state *reg, u32 regno,
13059	struct bpf_kfunc_call_arg_meta *meta,
13060	enum btf_field_type head_field_type,
13061	enum btf_field_type node_field_type,
13062	struct btf_field **node_field)
13063	{
13064	const char *node_type_name;
13065	const struct btf_type *et, *t;
13066	struct btf_field *field;
13067	u32 node_off;
13068
13069	if (meta->btf != btf_vmlinux) {
13070	verifier_bug(env, "unexpected btf mismatch in kfunc call");
13071	return -EFAULT;
13072	}
13073
13074	if (!check_kfunc_is_graph_node_api(env, node_field_type, kfunc_btf_id: meta->func_id))
13075	return -EFAULT;
13076
13077	node_type_name = btf_field_type_name(type: node_field_type);
13078	if (!tnum_is_const(a: reg->var_off)) {
13079	verbose(private_data: env,
13080	fmt: "R%d doesn't have constant offset. %s has to be at the constant offset\n",
13081	regno, node_type_name);
13082	return -EINVAL;
13083	}
13084
13085	node_off = reg->off + reg->var_off.value;
13086	field = reg_find_field_offset(reg, off: node_off, fields: node_field_type);
13087	if (!field) {
13088	verbose(private_data: env, fmt: "%s not found at offset=%u\n", node_type_name, node_off);
13089	return -EINVAL;
13090	}
13091
13092	field = *node_field;
13093
13094	et = btf_type_by_id(btf: field->graph_root.btf, type_id: field->graph_root.value_btf_id);
13095	t = btf_type_by_id(btf: reg->btf, type_id: reg->btf_id);
13096	if (!btf_struct_ids_match(log: &env->log, btf: reg->btf, id: reg->btf_id, off: 0, need_btf: field->graph_root.btf,
13097	need_type_id: field->graph_root.value_btf_id, strict: true)) {
13098	verbose(private_data: env, fmt: "operation on %s expects arg#1 %s at offset=%d "
13099	"in struct %s, but arg is at offset=%d in struct %s\n",
13100	btf_field_type_name(type: head_field_type),
13101	btf_field_type_name(type: node_field_type),
13102	field->graph_root.node_offset,
13103	btf_name_by_offset(btf: field->graph_root.btf, offset: et->name_off),
13104	node_off, btf_name_by_offset(btf: reg->btf, offset: t->name_off));
13105	return -EINVAL;
13106	}
13107	meta->arg_btf = reg->btf;
13108	meta->arg_btf_id = reg->btf_id;
13109
13110	if (node_off != field->graph_root.node_offset) {
13111	verbose(private_data: env, fmt: "arg#1 offset=%d, but expected %s at offset=%d in struct %s\n",
13112	node_off, btf_field_type_name(type: node_field_type),
13113	field->graph_root.node_offset,
13114	btf_name_by_offset(btf: field->graph_root.btf, offset: et->name_off));
13115	return -EINVAL;
13116	}
13117
13118	return 0;
13119	}
13120
13121	static int process_kf_arg_ptr_to_list_node(struct bpf_verifier_env *env,
13122	struct bpf_reg_state *reg, u32 regno,
13123	struct bpf_kfunc_call_arg_meta *meta)
13124	{
13125	return __process_kf_arg_ptr_to_graph_node(env, reg, regno, meta,
13126	head_field_type: BPF_LIST_HEAD, node_field_type: BPF_LIST_NODE,
13127	node_field: &meta->arg_list_head.field);
13128	}
13129
13130	static int process_kf_arg_ptr_to_rbtree_node(struct bpf_verifier_env *env,
13131	struct bpf_reg_state *reg, u32 regno,
13132	struct bpf_kfunc_call_arg_meta *meta)
13133	{
13134	return __process_kf_arg_ptr_to_graph_node(env, reg, regno, meta,
13135	head_field_type: BPF_RB_ROOT, node_field_type: BPF_RB_NODE,
13136	node_field: &meta->arg_rbtree_root.field);
13137	}
13138
13139	/*
13140	* css_task iter allowlist is needed to avoid dead locking on css_set_lock.
13141	* LSM hooks and iters (both sleepable and non-sleepable) are safe.
13142	* Any sleepable progs are also safe since bpf_check_attach_target() enforce
13143	* them can only be attached to some specific hook points.
13144	*/
13145	static bool check_css_task_iter_allowlist(struct bpf_verifier_env *env)
13146	{
13147	enum bpf_prog_type prog_type = resolve_prog_type(prog: env->prog);
13148
13149	switch (prog_type) {
13150	case BPF_PROG_TYPE_LSM:
13151	return true;
13152	case BPF_PROG_TYPE_TRACING:
13153	if (env->prog->expected_attach_type == BPF_TRACE_ITER)
13154	return true;
13155	fallthrough;
13156	default:
13157	return in_sleepable(env);
13158	}
13159	}
13160
13161	static int check_kfunc_args(struct bpf_verifier_env *env, struct bpf_kfunc_call_arg_meta *meta,
13162	int insn_idx)
13163	{
13164	const char *func_name = meta->func_name, *ref_tname;
13165	const struct btf *btf = meta->btf;
13166	const struct btf_param *args;
13167	struct btf_record *rec;
13168	u32 i, nargs;
13169	int ret;
13170
13171	args = (const struct btf_param *)(meta->func_proto + 1);
13172	nargs = btf_type_vlen(t: meta->func_proto);
13173	if (nargs > MAX_BPF_FUNC_REG_ARGS) {
13174	verbose(private_data: env, fmt: "Function %s has %d > %d args\n", func_name, nargs,
13175	MAX_BPF_FUNC_REG_ARGS);
13176	return -EINVAL;
13177	}
13178
13179	/* Check that BTF function arguments match actual types that the
13180	* verifier sees.
13181	*/
13182	for (i = 0; i < nargs; i++) {
13183	struct bpf_reg_state *regs = cur_regs(env), *reg = &regs[i + 1];
13184	const struct btf_type *t, *ref_t, *resolve_ret;
13185	enum bpf_arg_type arg_type = ARG_DONTCARE;
13186	u32 regno = i + 1, ref_id, type_size;
13187	bool is_ret_buf_sz = false;
13188	int kf_arg_type;
13189
13190	t = btf_type_skip_modifiers(btf, id: args[i].type, NULL);
13191
13192	if (is_kfunc_arg_ignore(btf, arg: &args[i]))
13193	continue;
13194
13195	if (is_kfunc_arg_prog(btf, arg: &args[i])) {
13196	/* Used to reject repeated use of __prog. */
13197	if (meta->arg_prog) {
13198	verifier_bug(env, "Only 1 prog->aux argument supported per-kfunc");
13199	return -EFAULT;
13200	}
13201	meta->arg_prog = true;
13202	cur_aux(env)->arg_prog = regno;
13203	continue;
13204	}
13205
13206	if (btf_type_is_scalar(t)) {
13207	if (reg->type != SCALAR_VALUE) {
13208	verbose(private_data: env, fmt: "R%d is not a scalar\n", regno);
13209	return -EINVAL;
13210	}
13211
13212	if (is_kfunc_arg_constant(btf: meta->btf, arg: &args[i])) {
13213	if (meta->arg_constant.found) {
13214	verifier_bug(env, "only one constant argument permitted");
13215	return -EFAULT;
13216	}
13217	if (!tnum_is_const(a: reg->var_off)) {
13218	verbose(private_data: env, fmt: "R%d must be a known constant\n", regno);
13219	return -EINVAL;
13220	}
13221	ret = mark_chain_precision(env, regno);
13222	if (ret < 0)
13223	return ret;
13224	meta->arg_constant.found = true;
13225	meta->arg_constant.value = reg->var_off.value;
13226	} else if (is_kfunc_arg_scalar_with_name(btf, arg: &args[i], name: "rdonly_buf_size")) {
13227	meta->r0_rdonly = true;
13228	is_ret_buf_sz = true;
13229	} else if (is_kfunc_arg_scalar_with_name(btf, arg: &args[i], name: "rdwr_buf_size")) {
13230	is_ret_buf_sz = true;
13231	}
13232
13233	if (is_ret_buf_sz) {
13234	if (meta->r0_size) {
13235	verbose(private_data: env, fmt: "2 or more rdonly/rdwr_buf_size parameters for kfunc");
13236	return -EINVAL;
13237	}
13238
13239	if (!tnum_is_const(a: reg->var_off)) {
13240	verbose(private_data: env, fmt: "R%d is not a const\n", regno);
13241	return -EINVAL;
13242	}
13243
13244	meta->r0_size = reg->var_off.value;
13245	ret = mark_chain_precision(env, regno);
13246	if (ret)
13247	return ret;
13248	}
13249	continue;
13250	}
13251
13252	if (!btf_type_is_ptr(t)) {
13253	verbose(private_data: env, fmt: "Unrecognized arg#%d type %s\n", i, btf_type_str(t));
13254	return -EINVAL;
13255	}
13256
13257	if ((is_kfunc_trusted_args(meta) || is_kfunc_rcu(meta)) &&
13258	(register_is_null(reg) || type_may_be_null(type: reg->type)) &&
13259	!is_kfunc_arg_nullable(btf: meta->btf, arg: &args[i])) {
13260	verbose(private_data: env, fmt: "Possibly NULL pointer passed to trusted arg%d\n", i);
13261	return -EACCES;
13262	}
13263
13264	if (reg->ref_obj_id) {
13265	if (is_kfunc_release(meta) && meta->ref_obj_id) {
13266	verifier_bug(env, "more than one arg with ref_obj_id R%d %u %u",
13267	regno, reg->ref_obj_id,
13268	meta->ref_obj_id);
13269	return -EFAULT;
13270	}
13271	meta->ref_obj_id = reg->ref_obj_id;
13272	if (is_kfunc_release(meta))
13273	meta->release_regno = regno;
13274	}
13275
13276	ref_t = btf_type_skip_modifiers(btf, id: t->type, res_id: &ref_id);
13277	ref_tname = btf_name_by_offset(btf, offset: ref_t->name_off);
13278
13279	kf_arg_type = get_kfunc_ptr_arg_type(env, meta, t, ref_t, ref_tname, args, argno: i, nargs);
13280	if (kf_arg_type < 0)
13281	return kf_arg_type;
13282
13283	switch (kf_arg_type) {
13284	case KF_ARG_PTR_TO_NULL:
13285	continue;
13286	case KF_ARG_PTR_TO_MAP:
13287	if (!reg->map_ptr) {
13288	verbose(private_data: env, fmt: "pointer in R%d isn't map pointer\n", regno);
13289	return -EINVAL;
13290	}
13291	if (meta->map.ptr && (reg->map_ptr->record->wq_off >= 0 ||
13292	reg->map_ptr->record->task_work_off >= 0)) {
13293	/* Use map_uid (which is unique id of inner map) to reject:
13294	* inner_map1 = bpf_map_lookup_elem(outer_map, key1)
13295	* inner_map2 = bpf_map_lookup_elem(outer_map, key2)
13296	* if (inner_map1 && inner_map2) {
13297	* wq = bpf_map_lookup_elem(inner_map1);
13298	* if (wq)
13299	* // mismatch would have been allowed
13300	* bpf_wq_init(wq, inner_map2);
13301	* }
13302	*
13303	* Comparing map_ptr is enough to distinguish normal and outer maps.
13304	*/
13305	if (meta->map.ptr != reg->map_ptr ||
13306	meta->map.uid != reg->map_uid) {
13307	if (reg->map_ptr->record->task_work_off >= 0) {
13308	verbose(private_data: env,
13309	fmt: "bpf_task_work pointer in R2 map_uid=%d doesn't match map pointer in R3 map_uid=%d\n",
13310	meta->map.uid, reg->map_uid);
13311	return -EINVAL;
13312	}
13313	verbose(private_data: env,
13314	fmt: "workqueue pointer in R1 map_uid=%d doesn't match map pointer in R2 map_uid=%d\n",
13315	meta->map.uid, reg->map_uid);
13316	return -EINVAL;
13317	}
13318	}
13319	meta->map.ptr = reg->map_ptr;
13320	meta->map.uid = reg->map_uid;
13321	fallthrough;
13322	case KF_ARG_PTR_TO_ALLOC_BTF_ID:
13323	case KF_ARG_PTR_TO_BTF_ID:
13324	if (!is_kfunc_trusted_args(meta) && !is_kfunc_rcu(meta))
13325	break;
13326
13327	if (!is_trusted_reg(reg)) {
13328	if (!is_kfunc_rcu(meta)) {
13329	verbose(private_data: env, fmt: "R%d must be referenced or trusted\n", regno);
13330	return -EINVAL;
13331	}
13332	if (!is_rcu_reg(reg)) {
13333	verbose(private_data: env, fmt: "R%d must be a rcu pointer\n", regno);
13334	return -EINVAL;
13335	}
13336	}
13337	fallthrough;
13338	case KF_ARG_PTR_TO_CTX:
13339	case KF_ARG_PTR_TO_DYNPTR:
13340	case KF_ARG_PTR_TO_ITER:
13341	case KF_ARG_PTR_TO_LIST_HEAD:
13342	case KF_ARG_PTR_TO_LIST_NODE:
13343	case KF_ARG_PTR_TO_RB_ROOT:
13344	case KF_ARG_PTR_TO_RB_NODE:
13345	case KF_ARG_PTR_TO_MEM:
13346	case KF_ARG_PTR_TO_MEM_SIZE:
13347	case KF_ARG_PTR_TO_CALLBACK:
13348	case KF_ARG_PTR_TO_REFCOUNTED_KPTR:
13349	case KF_ARG_PTR_TO_CONST_STR:
13350	case KF_ARG_PTR_TO_WORKQUEUE:
13351	case KF_ARG_PTR_TO_TASK_WORK:
13352	case KF_ARG_PTR_TO_IRQ_FLAG:
13353	case KF_ARG_PTR_TO_RES_SPIN_LOCK:
13354	break;
13355	default:
13356	verifier_bug(env, "unknown kfunc arg type %d", kf_arg_type);
13357	return -EFAULT;
13358	}
13359
13360	if (is_kfunc_release(meta) && reg->ref_obj_id)
13361	arg_type |= OBJ_RELEASE;
13362	ret = check_func_arg_reg_off(env, reg, regno, arg_type);
13363	if (ret < 0)
13364	return ret;
13365
13366	switch (kf_arg_type) {
13367	case KF_ARG_PTR_TO_CTX:
13368	if (reg->type != PTR_TO_CTX) {
13369	verbose(private_data: env, fmt: "arg#%d expected pointer to ctx, but got %s\n",
13370	i, reg_type_str(env, type: reg->type));
13371	return -EINVAL;
13372	}
13373
13374	if (meta->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx]) {
13375	ret = get_kern_ctx_btf_id(log: &env->log, prog_type: resolve_prog_type(prog: env->prog));
13376	if (ret < 0)
13377	return -EINVAL;
13378	meta->ret_btf_id = ret;
13379	}
13380	break;
13381	case KF_ARG_PTR_TO_ALLOC_BTF_ID:
13382	if (reg->type == (PTR_TO_BTF_ID | MEM_ALLOC)) {
13383	if (meta->func_id != special_kfunc_list[KF_bpf_obj_drop_impl]) {
13384	verbose(private_data: env, fmt: "arg#%d expected for bpf_obj_drop_impl()\n", i);
13385	return -EINVAL;
13386	}
13387	} else if (reg->type == (PTR_TO_BTF_ID | MEM_ALLOC | MEM_PERCPU)) {
13388	if (meta->func_id != special_kfunc_list[KF_bpf_percpu_obj_drop_impl]) {
13389	verbose(private_data: env, fmt: "arg#%d expected for bpf_percpu_obj_drop_impl()\n", i);
13390	return -EINVAL;
13391	}
13392	} else {
13393	verbose(private_data: env, fmt: "arg#%d expected pointer to allocated object\n", i);
13394	return -EINVAL;
13395	}
13396	if (!reg->ref_obj_id) {
13397	verbose(private_data: env, fmt: "allocated object must be referenced\n");
13398	return -EINVAL;
13399	}
13400	if (meta->btf == btf_vmlinux) {
13401	meta->arg_btf = reg->btf;
13402	meta->arg_btf_id = reg->btf_id;
13403	}
13404	break;
13405	case KF_ARG_PTR_TO_DYNPTR:
13406	{
13407	enum bpf_arg_type dynptr_arg_type = ARG_PTR_TO_DYNPTR;
13408	int clone_ref_obj_id = 0;
13409
13410	if (reg->type == CONST_PTR_TO_DYNPTR)
13411	dynptr_arg_type |= MEM_RDONLY;
13412
13413	if (is_kfunc_arg_uninit(btf, arg: &args[i]))
13414	dynptr_arg_type |= MEM_UNINIT;
13415
13416	if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_from_skb]) {
13417	dynptr_arg_type |= DYNPTR_TYPE_SKB;
13418	} else if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_from_xdp]) {
13419	dynptr_arg_type |= DYNPTR_TYPE_XDP;
13420	} else if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_from_skb_meta]) {
13421	dynptr_arg_type |= DYNPTR_TYPE_SKB_META;
13422	} else if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_from_file]) {
13423	dynptr_arg_type |= DYNPTR_TYPE_FILE;
13424	} else if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_file_discard]) {
13425	dynptr_arg_type |= DYNPTR_TYPE_FILE;
13426	meta->release_regno = regno;
13427	} else if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_clone] &&
13428	(dynptr_arg_type & MEM_UNINIT)) {
13429	enum bpf_dynptr_type parent_type = meta->initialized_dynptr.type;
13430
13431	if (parent_type == BPF_DYNPTR_TYPE_INVALID) {
13432	verifier_bug(env, "no dynptr type for parent of clone");
13433	return -EFAULT;
13434	}
13435
13436	dynptr_arg_type |= (unsigned int)get_dynptr_type_flag(type: parent_type);
13437	clone_ref_obj_id = meta->initialized_dynptr.ref_obj_id;
13438	if (dynptr_type_refcounted(type: parent_type) && !clone_ref_obj_id) {
13439	verifier_bug(env, "missing ref obj id for parent of clone");
13440	return -EFAULT;
13441	}
13442	}
13443
13444	ret = process_dynptr_func(env, regno, insn_idx, arg_type: dynptr_arg_type, clone_ref_obj_id);
13445	if (ret < 0)
13446	return ret;
13447
13448	if (!(dynptr_arg_type & MEM_UNINIT)) {
13449	int id = dynptr_id(env, reg);
13450
13451	if (id < 0) {
13452	verifier_bug(env, "failed to obtain dynptr id");
13453	return id;
13454	}
13455	meta->initialized_dynptr.id = id;
13456	meta->initialized_dynptr.type = dynptr_get_type(env, reg);
13457	meta->initialized_dynptr.ref_obj_id = dynptr_ref_obj_id(env, reg);
13458	}
13459
13460	break;
13461	}
13462	case KF_ARG_PTR_TO_ITER:
13463	if (meta->func_id == special_kfunc_list[KF_bpf_iter_css_task_new]) {
13464	if (!check_css_task_iter_allowlist(env)) {
13465	verbose(private_data: env, fmt: "css_task_iter is only allowed in bpf_lsm, bpf_iter and sleepable progs\n");
13466	return -EINVAL;
13467	}
13468	}
13469	ret = process_iter_arg(env, regno, insn_idx, meta);
13470	if (ret < 0)
13471	return ret;
13472	break;
13473	case KF_ARG_PTR_TO_LIST_HEAD:
13474	if (reg->type != PTR_TO_MAP_VALUE &&
13475	reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
13476	verbose(private_data: env, fmt: "arg#%d expected pointer to map value or allocated object\n", i);
13477	return -EINVAL;
13478	}
13479	if (reg->type == (PTR_TO_BTF_ID | MEM_ALLOC) && !reg->ref_obj_id) {
13480	verbose(private_data: env, fmt: "allocated object must be referenced\n");
13481	return -EINVAL;
13482	}
13483	ret = process_kf_arg_ptr_to_list_head(env, reg, regno, meta);
13484	if (ret < 0)
13485	return ret;
13486	break;
13487	case KF_ARG_PTR_TO_RB_ROOT:
13488	if (reg->type != PTR_TO_MAP_VALUE &&
13489	reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
13490	verbose(private_data: env, fmt: "arg#%d expected pointer to map value or allocated object\n", i);
13491	return -EINVAL;
13492	}
13493	if (reg->type == (PTR_TO_BTF_ID | MEM_ALLOC) && !reg->ref_obj_id) {
13494	verbose(private_data: env, fmt: "allocated object must be referenced\n");
13495	return -EINVAL;
13496	}
13497	ret = process_kf_arg_ptr_to_rbtree_root(env, reg, regno, meta);
13498	if (ret < 0)
13499	return ret;
13500	break;
13501	case KF_ARG_PTR_TO_LIST_NODE:
13502	if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
13503	verbose(private_data: env, fmt: "arg#%d expected pointer to allocated object\n", i);
13504	return -EINVAL;
13505	}
13506	if (!reg->ref_obj_id) {
13507	verbose(private_data: env, fmt: "allocated object must be referenced\n");
13508	return -EINVAL;
13509	}
13510	ret = process_kf_arg_ptr_to_list_node(env, reg, regno, meta);
13511	if (ret < 0)
13512	return ret;
13513	break;
13514	case KF_ARG_PTR_TO_RB_NODE:
13515	if (meta->func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
13516	if (reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
13517	verbose(private_data: env, fmt: "arg#%d expected pointer to allocated object\n", i);
13518	return -EINVAL;
13519	}
13520	if (!reg->ref_obj_id) {
13521	verbose(private_data: env, fmt: "allocated object must be referenced\n");
13522	return -EINVAL;
13523	}
13524	} else {
13525	if (!type_is_non_owning_ref(type: reg->type) && !reg->ref_obj_id) {
13526	verbose(private_data: env, fmt: "%s can only take non-owning or refcounted bpf_rb_node pointer\n", func_name);
13527	return -EINVAL;
13528	}
13529	if (in_rbtree_lock_required_cb(env)) {
13530	verbose(private_data: env, fmt: "%s not allowed in rbtree cb\n", func_name);
13531	return -EINVAL;
13532	}
13533	}
13534
13535	ret = process_kf_arg_ptr_to_rbtree_node(env, reg, regno, meta);
13536	if (ret < 0)
13537	return ret;
13538	break;
13539	case KF_ARG_PTR_TO_MAP:
13540	/* If argument has '__map' suffix expect 'struct bpf_map *' */
13541	ref_id = *reg2btf_ids[CONST_PTR_TO_MAP];
13542	ref_t = btf_type_by_id(btf: btf_vmlinux, type_id: ref_id);
13543	ref_tname = btf_name_by_offset(btf, offset: ref_t->name_off);
13544	fallthrough;
13545	case KF_ARG_PTR_TO_BTF_ID:
13546	/* Only base_type is checked, further checks are done here */
13547	if ((base_type(type: reg->type) != PTR_TO_BTF_ID ||
13548	(bpf_type_has_unsafe_modifiers(type: reg->type) && !is_rcu_reg(reg))) &&
13549	!reg2btf_ids[base_type(type: reg->type)]) {
13550	verbose(private_data: env, fmt: "arg#%d is %s ", i, reg_type_str(env, type: reg->type));
13551	verbose(private_data: env, fmt: "expected %s or socket\n",
13552	reg_type_str(env, type: base_type(type: reg->type) |
13553	(type_flag(type: reg->type) & BPF_REG_TRUSTED_MODIFIERS)));
13554	return -EINVAL;
13555	}
13556	ret = process_kf_arg_ptr_to_btf_id(env, reg, ref_t, ref_tname, ref_id, meta, argno: i);
13557	if (ret < 0)
13558	return ret;
13559	break;
13560	case KF_ARG_PTR_TO_MEM:
13561	resolve_ret = btf_resolve_size(btf, type: ref_t, type_size: &type_size);
13562	if (IS_ERR(ptr: resolve_ret)) {
13563	verbose(private_data: env, fmt: "arg#%d reference type('%s %s') size cannot be determined: %ld\n",
13564	i, btf_type_str(t: ref_t), ref_tname, PTR_ERR(ptr: resolve_ret));
13565	return -EINVAL;
13566	}
13567	ret = check_mem_reg(env, reg, regno, mem_size: type_size);
13568	if (ret < 0)
13569	return ret;
13570	break;
13571	case KF_ARG_PTR_TO_MEM_SIZE:
13572	{
13573	struct bpf_reg_state *buff_reg = &regs[regno];
13574	const struct btf_param *buff_arg = &args[i];
13575	struct bpf_reg_state *size_reg = &regs[regno + 1];
13576	const struct btf_param *size_arg = &args[i + 1];
13577
13578	if (!register_is_null(reg: buff_reg) || !is_kfunc_arg_optional(btf: meta->btf, arg: buff_arg)) {
13579	ret = check_kfunc_mem_size_reg(env, reg: size_reg, regno: regno + 1);
13580	if (ret < 0) {
13581	verbose(private_data: env, fmt: "arg#%d arg#%d memory, len pair leads to invalid memory access\n", i, i + 1);
13582	return ret;
13583	}
13584	}
13585
13586	if (is_kfunc_arg_const_mem_size(btf: meta->btf, arg: size_arg, reg: size_reg)) {
13587	if (meta->arg_constant.found) {
13588	verifier_bug(env, "only one constant argument permitted");
13589	return -EFAULT;
13590	}
13591	if (!tnum_is_const(a: size_reg->var_off)) {
13592	verbose(private_data: env, fmt: "R%d must be a known constant\n", regno + 1);
13593	return -EINVAL;
13594	}
13595	meta->arg_constant.found = true;
13596	meta->arg_constant.value = size_reg->var_off.value;
13597	}
13598
13599	/* Skip next '__sz' or '__szk' argument */
13600	i++;
13601	break;
13602	}
13603	case KF_ARG_PTR_TO_CALLBACK:
13604	if (reg->type != PTR_TO_FUNC) {
13605	verbose(private_data: env, fmt: "arg%d expected pointer to func\n", i);
13606	return -EINVAL;
13607	}
13608	meta->subprogno = reg->subprogno;
13609	break;
13610	case KF_ARG_PTR_TO_REFCOUNTED_KPTR:
13611	if (!type_is_ptr_alloc_obj(type: reg->type)) {
13612	verbose(private_data: env, fmt: "arg#%d is neither owning or non-owning ref\n", i);
13613	return -EINVAL;
13614	}
13615	if (!type_is_non_owning_ref(type: reg->type))
13616	meta->arg_owning_ref = true;
13617
13618	rec = reg_btf_record(reg);
13619	if (!rec) {
13620	verifier_bug(env, "Couldn't find btf_record");
13621	return -EFAULT;
13622	}
13623
13624	if (rec->refcount_off < 0) {
13625	verbose(private_data: env, fmt: "arg#%d doesn't point to a type with bpf_refcount field\n", i);
13626	return -EINVAL;
13627	}
13628
13629	meta->arg_btf = reg->btf;
13630	meta->arg_btf_id = reg->btf_id;
13631	break;
13632	case KF_ARG_PTR_TO_CONST_STR:
13633	if (reg->type != PTR_TO_MAP_VALUE) {
13634	verbose(private_data: env, fmt: "arg#%d doesn't point to a const string\n", i);
13635	return -EINVAL;
13636	}
13637	ret = check_reg_const_str(env, reg, regno);
13638	if (ret)
13639	return ret;
13640	break;
13641	case KF_ARG_PTR_TO_WORKQUEUE:
13642	if (reg->type != PTR_TO_MAP_VALUE) {
13643	verbose(private_data: env, fmt: "arg#%d doesn't point to a map value\n", i);
13644	return -EINVAL;
13645	}
13646	ret = process_wq_func(env, regno, meta);
13647	if (ret < 0)
13648	return ret;
13649	break;
13650	case KF_ARG_PTR_TO_TASK_WORK:
13651	if (reg->type != PTR_TO_MAP_VALUE) {
13652	verbose(private_data: env, fmt: "arg#%d doesn't point to a map value\n", i);
13653	return -EINVAL;
13654	}
13655	ret = process_task_work_func(env, regno, meta);
13656	if (ret < 0)
13657	return ret;
13658	break;
13659	case KF_ARG_PTR_TO_IRQ_FLAG:
13660	if (reg->type != PTR_TO_STACK) {
13661	verbose(private_data: env, fmt: "arg#%d doesn't point to an irq flag on stack\n", i);
13662	return -EINVAL;
13663	}
13664	ret = process_irq_flag(env, regno, meta);
13665	if (ret < 0)
13666	return ret;
13667	break;
13668	case KF_ARG_PTR_TO_RES_SPIN_LOCK:
13669	{
13670	int flags = PROCESS_RES_LOCK;
13671
13672	if (reg->type != PTR_TO_MAP_VALUE && reg->type != (PTR_TO_BTF_ID | MEM_ALLOC)) {
13673	verbose(private_data: env, fmt: "arg#%d doesn't point to map value or allocated object\n", i);
13674	return -EINVAL;
13675	}
13676
13677	if (!is_bpf_res_spin_lock_kfunc(btf_id: meta->func_id))
13678	return -EFAULT;
13679	if (meta->func_id == special_kfunc_list[KF_bpf_res_spin_lock] ||
13680	meta->func_id == special_kfunc_list[KF_bpf_res_spin_lock_irqsave])
13681	flags |= PROCESS_SPIN_LOCK;
13682	if (meta->func_id == special_kfunc_list[KF_bpf_res_spin_lock_irqsave] ||
13683	meta->func_id == special_kfunc_list[KF_bpf_res_spin_unlock_irqrestore])
13684	flags |= PROCESS_LOCK_IRQ;
13685	ret = process_spin_lock(env, regno, flags);
13686	if (ret < 0)
13687	return ret;
13688	break;
13689	}
13690	}
13691	}
13692
13693	if (is_kfunc_release(meta) && !meta->release_regno) {
13694	verbose(private_data: env, fmt: "release kernel function %s expects refcounted PTR_TO_BTF_ID\n",
13695	func_name);
13696	return -EINVAL;
13697	}
13698
13699	return 0;
13700	}
13701
13702	static int fetch_kfunc_meta(struct bpf_verifier_env *env,
13703	struct bpf_insn *insn,
13704	struct bpf_kfunc_call_arg_meta *meta,
13705	const char **kfunc_name)
13706	{
13707	const struct btf_type *func, *func_proto;
13708	u32 func_id, *kfunc_flags;
13709	const char *func_name;
13710	struct btf *desc_btf;
13711
13712	if (kfunc_name)
13713	*kfunc_name = NULL;
13714
13715	if (!insn->imm)
13716	return -EINVAL;
13717
13718	desc_btf = find_kfunc_desc_btf(env, offset: insn->off);
13719	if (IS_ERR(ptr: desc_btf))
13720	return PTR_ERR(ptr: desc_btf);
13721
13722	func_id = insn->imm;
13723	func = btf_type_by_id(btf: desc_btf, type_id: func_id);
13724	func_name = btf_name_by_offset(btf: desc_btf, offset: func->name_off);
13725	if (kfunc_name)
13726	*kfunc_name = func_name;
13727	func_proto = btf_type_by_id(btf: desc_btf, type_id: func->type);
13728
13729	kfunc_flags = btf_kfunc_id_set_contains(btf: desc_btf, kfunc_btf_id: func_id, prog: env->prog);
13730	if (!kfunc_flags) {
13731	return -EACCES;
13732	}
13733
13734	memset(meta, 0, sizeof(*meta));
13735	meta->btf = desc_btf;
13736	meta->func_id = func_id;
13737	meta->kfunc_flags = *kfunc_flags;
13738	meta->func_proto = func_proto;
13739	meta->func_name = func_name;
13740
13741	return 0;
13742	}
13743
13744	/* check special kfuncs and return:
13745	* 1 - not fall-through to 'else' branch, continue verification
13746	* 0 - fall-through to 'else' branch
13747	* < 0 - not fall-through to 'else' branch, return error
13748	*/
13749	static int check_special_kfunc(struct bpf_verifier_env *env, struct bpf_kfunc_call_arg_meta *meta,
13750	struct bpf_reg_state *regs, struct bpf_insn_aux_data *insn_aux,
13751	const struct btf_type *ptr_type, struct btf *desc_btf)
13752	{
13753	const struct btf_type *ret_t;
13754	int err = 0;
13755
13756	if (meta->btf != btf_vmlinux)
13757	return 0;
13758
13759	if (meta->func_id == special_kfunc_list[KF_bpf_obj_new_impl] ||
13760	meta->func_id == special_kfunc_list[KF_bpf_percpu_obj_new_impl]) {
13761	struct btf_struct_meta *struct_meta;
13762	struct btf *ret_btf;
13763	u32 ret_btf_id;
13764
13765	if (meta->func_id == special_kfunc_list[KF_bpf_obj_new_impl] && !bpf_global_ma_set)
13766	return -ENOMEM;
13767
13768	if (((u64)(u32)meta->arg_constant.value) != meta->arg_constant.value) {
13769	verbose(private_data: env, fmt: "local type ID argument must be in range [0, U32_MAX]\n");
13770	return -EINVAL;
13771	}
13772
13773	ret_btf = env->prog->aux->btf;
13774	ret_btf_id = meta->arg_constant.value;
13775
13776	/* This may be NULL due to user not supplying a BTF */
13777	if (!ret_btf) {
13778	verbose(private_data: env, fmt: "bpf_obj_new/bpf_percpu_obj_new requires prog BTF\n");
13779	return -EINVAL;
13780	}
13781
13782	ret_t = btf_type_by_id(btf: ret_btf, type_id: ret_btf_id);
13783	if (!ret_t || !__btf_type_is_struct(t: ret_t)) {
13784	verbose(private_data: env, fmt: "bpf_obj_new/bpf_percpu_obj_new type ID argument must be of a struct\n");
13785	return -EINVAL;
13786	}
13787
13788	if (meta->func_id == special_kfunc_list[KF_bpf_percpu_obj_new_impl]) {
13789	if (ret_t->size > BPF_GLOBAL_PERCPU_MA_MAX_SIZE) {
13790	verbose(private_data: env, fmt: "bpf_percpu_obj_new type size (%d) is greater than %d\n",
13791	ret_t->size, BPF_GLOBAL_PERCPU_MA_MAX_SIZE);
13792	return -EINVAL;
13793	}
13794
13795	if (!bpf_global_percpu_ma_set) {
13796	mutex_lock(&bpf_percpu_ma_lock);
13797	if (!bpf_global_percpu_ma_set) {
13798	/* Charge memory allocated with bpf_global_percpu_ma to
13799	* root memcg. The obj_cgroup for root memcg is NULL.
13800	*/
13801	err = bpf_mem_alloc_percpu_init(ma: &bpf_global_percpu_ma, NULL);
13802	if (!err)
13803	bpf_global_percpu_ma_set = true;
13804	}
13805	mutex_unlock(lock: &bpf_percpu_ma_lock);
13806	if (err)
13807	return err;
13808	}
13809
13810	mutex_lock(&bpf_percpu_ma_lock);
13811	err = bpf_mem_alloc_percpu_unit_init(ma: &bpf_global_percpu_ma, size: ret_t->size);
13812	mutex_unlock(lock: &bpf_percpu_ma_lock);
13813	if (err)
13814	return err;
13815	}
13816
13817	struct_meta = btf_find_struct_meta(btf: ret_btf, btf_id: ret_btf_id);
13818	if (meta->func_id == special_kfunc_list[KF_bpf_percpu_obj_new_impl]) {
13819	if (!__btf_type_is_scalar_struct(env, btf: ret_btf, t: ret_t, rec: 0)) {
13820	verbose(private_data: env, fmt: "bpf_percpu_obj_new type ID argument must be of a struct of scalars\n");
13821	return -EINVAL;
13822	}
13823
13824	if (struct_meta) {
13825	verbose(private_data: env, fmt: "bpf_percpu_obj_new type ID argument must not contain special fields\n");
13826	return -EINVAL;
13827	}
13828	}
13829
13830	mark_reg_known_zero(env, regs, regno: BPF_REG_0);
13831	regs[BPF_REG_0].type = PTR_TO_BTF_ID | MEM_ALLOC;
13832	regs[BPF_REG_0].btf = ret_btf;
13833	regs[BPF_REG_0].btf_id = ret_btf_id;
13834	if (meta->func_id == special_kfunc_list[KF_bpf_percpu_obj_new_impl])
13835	regs[BPF_REG_0].type |= MEM_PERCPU;
13836
13837	insn_aux->obj_new_size = ret_t->size;
13838	insn_aux->kptr_struct_meta = struct_meta;
13839	} else if (meta->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl]) {
13840	mark_reg_known_zero(env, regs, regno: BPF_REG_0);
13841	regs[BPF_REG_0].type = PTR_TO_BTF_ID | MEM_ALLOC;
13842	regs[BPF_REG_0].btf = meta->arg_btf;
13843	regs[BPF_REG_0].btf_id = meta->arg_btf_id;
13844
13845	insn_aux->kptr_struct_meta =
13846	btf_find_struct_meta(btf: meta->arg_btf,
13847	btf_id: meta->arg_btf_id);
13848	} else if (is_list_node_type(t: ptr_type)) {
13849	struct btf_field *field = meta->arg_list_head.field;
13850
13851	mark_reg_graph_node(regs, regno: BPF_REG_0, ds_head: &field->graph_root);
13852	} else if (is_rbtree_node_type(t: ptr_type)) {
13853	struct btf_field *field = meta->arg_rbtree_root.field;
13854
13855	mark_reg_graph_node(regs, regno: BPF_REG_0, ds_head: &field->graph_root);
13856	} else if (meta->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx]) {
13857	mark_reg_known_zero(env, regs, regno: BPF_REG_0);
13858	regs[BPF_REG_0].type = PTR_TO_BTF_ID | PTR_TRUSTED;
13859	regs[BPF_REG_0].btf = desc_btf;
13860	regs[BPF_REG_0].btf_id = meta->ret_btf_id;
13861	} else if (meta->func_id == special_kfunc_list[KF_bpf_rdonly_cast]) {
13862	ret_t = btf_type_by_id(btf: desc_btf, type_id: meta->arg_constant.value);
13863	if (!ret_t) {
13864	verbose(private_data: env, fmt: "Unknown type ID %lld passed to kfunc bpf_rdonly_cast\n",
13865	meta->arg_constant.value);
13866	return -EINVAL;
13867	} else if (btf_type_is_struct(t: ret_t)) {
13868	mark_reg_known_zero(env, regs, regno: BPF_REG_0);
13869	regs[BPF_REG_0].type = PTR_TO_BTF_ID | PTR_UNTRUSTED;
13870	regs[BPF_REG_0].btf = desc_btf;
13871	regs[BPF_REG_0].btf_id = meta->arg_constant.value;
13872	} else if (btf_type_is_void(t: ret_t)) {
13873	mark_reg_known_zero(env, regs, regno: BPF_REG_0);
13874	regs[BPF_REG_0].type = PTR_TO_MEM | MEM_RDONLY | PTR_UNTRUSTED;
13875	regs[BPF_REG_0].mem_size = 0;
13876	} else {
13877	verbose(private_data: env,
13878	fmt: "kfunc bpf_rdonly_cast type ID argument must be of a struct or void\n");
13879	return -EINVAL;
13880	}
13881	} else if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_slice] ||
13882	meta->func_id == special_kfunc_list[KF_bpf_dynptr_slice_rdwr]) {
13883	enum bpf_type_flag type_flag = get_dynptr_type_flag(type: meta->initialized_dynptr.type);
13884
13885	mark_reg_known_zero(env, regs, regno: BPF_REG_0);
13886
13887	if (!meta->arg_constant.found) {
13888	verifier_bug(env, "bpf_dynptr_slice(_rdwr) no constant size");
13889	return -EFAULT;
13890	}
13891
13892	regs[BPF_REG_0].mem_size = meta->arg_constant.value;
13893
13894	/* PTR_MAYBE_NULL will be added when is_kfunc_ret_null is checked */
13895	regs[BPF_REG_0].type = PTR_TO_MEM | type_flag;
13896
13897	if (meta->func_id == special_kfunc_list[KF_bpf_dynptr_slice]) {
13898	regs[BPF_REG_0].type |= MEM_RDONLY;
13899	} else {
13900	/* this will set env->seen_direct_write to true */
13901	if (!may_access_direct_pkt_data(env, NULL, t: BPF_WRITE)) {
13902	verbose(private_data: env, fmt: "the prog does not allow writes to packet data\n");
13903	return -EINVAL;
13904	}
13905	}
13906
13907	if (!meta->initialized_dynptr.id) {
13908	verifier_bug(env, "no dynptr id");
13909	return -EFAULT;
13910	}
13911	regs[BPF_REG_0].dynptr_id = meta->initialized_dynptr.id;
13912
13913	/* we don't need to set BPF_REG_0's ref obj id
13914	* because packet slices are not refcounted (see
13915	* dynptr_type_refcounted)
13916	*/
13917	} else {
13918	return 0;
13919	}
13920
13921	return 1;
13922	}
13923
13924	static int check_return_code(struct bpf_verifier_env *env, int regno, const char *reg_name);
13925
13926	static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
13927	int *insn_idx_p)
13928	{
13929	bool sleepable, rcu_lock, rcu_unlock, preempt_disable, preempt_enable;
13930	u32 i, nargs, ptr_type_id, release_ref_obj_id;
13931	struct bpf_reg_state *regs = cur_regs(env);
13932	const char *func_name, *ptr_type_name;
13933	const struct btf_type *t, *ptr_type;
13934	struct bpf_kfunc_call_arg_meta meta;
13935	struct bpf_insn_aux_data *insn_aux;
13936	int err, insn_idx = *insn_idx_p;
13937	const struct btf_param *args;
13938	struct btf *desc_btf;
13939
13940	/* skip for now, but return error when we find this in fixup_kfunc_call */
13941	if (!insn->imm)
13942	return 0;
13943
13944	err = fetch_kfunc_meta(env, insn, meta: &meta, kfunc_name: &func_name);
13945	if (err == -EACCES && func_name)
13946	verbose(private_data: env, fmt: "calling kernel function %s is not allowed\n", func_name);
13947	if (err)
13948	return err;
13949	desc_btf = meta.btf;
13950	insn_aux = &env->insn_aux_data[insn_idx];
13951
13952	insn_aux->is_iter_next = is_iter_next_kfunc(meta: &meta);
13953
13954	if (!insn->off &&
13955	(insn->imm == special_kfunc_list[KF_bpf_res_spin_lock] ||
13956	insn->imm == special_kfunc_list[KF_bpf_res_spin_lock_irqsave])) {
13957	struct bpf_verifier_state *branch;
13958	struct bpf_reg_state *regs;
13959
13960	branch = push_stack(env, insn_idx: env->insn_idx + 1, prev_insn_idx: env->insn_idx, speculative: false);
13961	if (IS_ERR(ptr: branch)) {
13962	verbose(private_data: env, fmt: "failed to push state for failed lock acquisition\n");
13963	return PTR_ERR(ptr: branch);
13964	}
13965
13966	regs = branch->frame[branch->curframe]->regs;
13967
13968	/* Clear r0-r5 registers in forked state */
13969	for (i = 0; i < CALLER_SAVED_REGS; i++)
13970	mark_reg_not_init(env, regs, regno: caller_saved[i]);
13971
13972	mark_reg_unknown(env, regs, regno: BPF_REG_0);
13973	err = __mark_reg_s32_range(env, regs, regno: BPF_REG_0, s32_min: -MAX_ERRNO, s32_max: -1);
13974	if (err) {
13975	verbose(private_data: env, fmt: "failed to mark s32 range for retval in forked state for lock\n");
13976	return err;
13977	}
13978	__mark_btf_func_reg_size(env, regs, regno: BPF_REG_0, reg_size: sizeof(u32));
13979	} else if (!insn->off && insn->imm == special_kfunc_list[KF___bpf_trap]) {
13980	verbose(private_data: env, fmt: "unexpected __bpf_trap() due to uninitialized variable?\n");
13981	return -EFAULT;
13982	}
13983
13984	if (is_kfunc_destructive(meta: &meta) && !capable(CAP_SYS_BOOT)) {
13985	verbose(private_data: env, fmt: "destructive kfunc calls require CAP_SYS_BOOT capability\n");
13986	return -EACCES;
13987	}
13988
13989	sleepable = is_kfunc_sleepable(meta: &meta);
13990	if (sleepable && !in_sleepable(env)) {
13991	verbose(private_data: env, fmt: "program must be sleepable to call sleepable kfunc %s\n", func_name);
13992	return -EACCES;
13993	}
13994
13995	/* Track non-sleepable context for kfuncs, same as for helpers. */
13996	if (!in_sleepable_context(env))
13997	insn_aux->non_sleepable = true;
13998
13999	/* Check the arguments */
14000	err = check_kfunc_args(env, meta: &meta, insn_idx);
14001	if (err < 0)
14002	return err;
14003
14004	if (meta.func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
14005	err = push_callback_call(env, insn, insn_idx, subprog: meta.subprogno,
14006	set_callee_state_cb: set_rbtree_add_callback_state);
14007	if (err) {
14008	verbose(private_data: env, fmt: "kfunc %s#%d failed callback verification\n",
14009	func_name, meta.func_id);
14010	return err;
14011	}
14012	}
14013
14014	if (meta.func_id == special_kfunc_list[KF_bpf_session_cookie]) {
14015	meta.r0_size = sizeof(u64);
14016	meta.r0_rdonly = false;
14017	}
14018
14019	if (is_bpf_wq_set_callback_impl_kfunc(btf_id: meta.func_id)) {
14020	err = push_callback_call(env, insn, insn_idx, subprog: meta.subprogno,
14021	set_callee_state_cb: set_timer_callback_state);
14022	if (err) {
14023	verbose(private_data: env, fmt: "kfunc %s#%d failed callback verification\n",
14024	func_name, meta.func_id);
14025	return err;
14026	}
14027	}
14028
14029	if (is_task_work_add_kfunc(func_id: meta.func_id)) {
14030	err = push_callback_call(env, insn, insn_idx, subprog: meta.subprogno,
14031	set_callee_state_cb: set_task_work_schedule_callback_state);
14032	if (err) {
14033	verbose(private_data: env, fmt: "kfunc %s#%d failed callback verification\n",
14034	func_name, meta.func_id);
14035	return err;
14036	}
14037	}
14038
14039	rcu_lock = is_kfunc_bpf_rcu_read_lock(meta: &meta);
14040	rcu_unlock = is_kfunc_bpf_rcu_read_unlock(meta: &meta);
14041
14042	preempt_disable = is_kfunc_bpf_preempt_disable(meta: &meta);
14043	preempt_enable = is_kfunc_bpf_preempt_enable(meta: &meta);
14044
14045	if (rcu_lock) {
14046	env->cur_state->active_rcu_locks++;
14047	} else if (rcu_unlock) {
14048	struct bpf_func_state *state;
14049	struct bpf_reg_state *reg;
14050	u32 clear_mask = (1 << STACK_SPILL) | (1 << STACK_ITER);
14051
14052	if (env->cur_state->active_rcu_locks == 0) {
14053	verbose(private_data: env, fmt: "unmatched rcu read unlock (kernel function %s)\n", func_name);
14054	return -EINVAL;
14055	}
14056	if (--env->cur_state->active_rcu_locks == 0) {
14057	bpf_for_each_reg_in_vstate_mask(env->cur_state, state, reg, clear_mask, ({
14058	if (reg->type & MEM_RCU) {
14059	reg->type &= ~(MEM_RCU | PTR_MAYBE_NULL);
14060	reg->type |= PTR_UNTRUSTED;
14061	}
14062	}));
14063	}
14064	} else if (sleepable && env->cur_state->active_rcu_locks) {
14065	verbose(private_data: env, fmt: "kernel func %s is sleepable within rcu_read_lock region\n", func_name);
14066	return -EACCES;
14067	}
14068
14069	if (in_rbtree_lock_required_cb(env) && (rcu_lock || rcu_unlock)) {
14070	verbose(private_data: env, fmt: "Calling bpf_rcu_read_{lock,unlock} in unnecessary rbtree callback\n");
14071	return -EACCES;
14072	}
14073
14074	if (env->cur_state->active_preempt_locks) {
14075	if (preempt_disable) {
14076	env->cur_state->active_preempt_locks++;
14077	} else if (preempt_enable) {
14078	env->cur_state->active_preempt_locks--;
14079	} else if (sleepable) {
14080	verbose(private_data: env, fmt: "kernel func %s is sleepable within non-preemptible region\n", func_name);
14081	return -EACCES;
14082	}
14083	} else if (preempt_disable) {
14084	env->cur_state->active_preempt_locks++;
14085	} else if (preempt_enable) {
14086	verbose(private_data: env, fmt: "unmatched attempt to enable preemption (kernel function %s)\n", func_name);
14087	return -EINVAL;
14088	}
14089
14090	if (env->cur_state->active_irq_id && sleepable) {
14091	verbose(private_data: env, fmt: "kernel func %s is sleepable within IRQ-disabled region\n", func_name);
14092	return -EACCES;
14093	}
14094
14095	if (is_kfunc_rcu_protected(meta: &meta) && !in_rcu_cs(env)) {
14096	verbose(private_data: env, fmt: "kernel func %s requires RCU critical section protection\n", func_name);
14097	return -EACCES;
14098	}
14099
14100	/* In case of release function, we get register number of refcounted
14101	* PTR_TO_BTF_ID in bpf_kfunc_arg_meta, do the release now.
14102	*/
14103	if (meta.release_regno) {
14104	struct bpf_reg_state *reg = &regs[meta.release_regno];
14105
14106	if (meta.initialized_dynptr.ref_obj_id) {
14107	err = unmark_stack_slots_dynptr(env, reg);
14108	} else {
14109	err = release_reference(env, ref_obj_id: reg->ref_obj_id);
14110	if (err)
14111	verbose(private_data: env, fmt: "kfunc %s#%d reference has not been acquired before\n",
14112	func_name, meta.func_id);
14113	}
14114	if (err)
14115	return err;
14116	}
14117
14118	if (meta.func_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
14119	meta.func_id == special_kfunc_list[KF_bpf_list_push_back_impl] ||
14120	meta.func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
14121	release_ref_obj_id = regs[BPF_REG_2].ref_obj_id;
14122	insn_aux->insert_off = regs[BPF_REG_2].off;
14123	insn_aux->kptr_struct_meta = btf_find_struct_meta(btf: meta.arg_btf, btf_id: meta.arg_btf_id);
14124	err = ref_convert_owning_non_owning(env, ref_obj_id: release_ref_obj_id);
14125	if (err) {
14126	verbose(private_data: env, fmt: "kfunc %s#%d conversion of owning ref to non-owning failed\n",
14127	func_name, meta.func_id);
14128	return err;
14129	}
14130
14131	err = release_reference(env, ref_obj_id: release_ref_obj_id);
14132	if (err) {
14133	verbose(private_data: env, fmt: "kfunc %s#%d reference has not been acquired before\n",
14134	func_name, meta.func_id);
14135	return err;
14136	}
14137	}
14138
14139	if (meta.func_id == special_kfunc_list[KF_bpf_throw]) {
14140	if (!bpf_jit_supports_exceptions()) {
14141	verbose(private_data: env, fmt: "JIT does not support calling kfunc %s#%d\n",
14142	func_name, meta.func_id);
14143	return -ENOTSUPP;
14144	}
14145	env->seen_exception = true;
14146
14147	/* In the case of the default callback, the cookie value passed
14148	* to bpf_throw becomes the return value of the program.
14149	*/
14150	if (!env->exception_callback_subprog) {
14151	err = check_return_code(env, regno: BPF_REG_1, reg_name: "R1");
14152	if (err < 0)
14153	return err;
14154	}
14155	}
14156
14157	for (i = 0; i < CALLER_SAVED_REGS; i++)
14158	mark_reg_not_init(env, regs, regno: caller_saved[i]);
14159
14160	/* Check return type */
14161	t = btf_type_skip_modifiers(btf: desc_btf, id: meta.func_proto->type, NULL);
14162
14163	if (is_kfunc_acquire(meta: &meta) && !btf_type_is_struct_ptr(btf: meta.btf, t)) {
14164	/* Only exception is bpf_obj_new_impl */
14165	if (meta.btf != btf_vmlinux ||
14166	(meta.func_id != special_kfunc_list[KF_bpf_obj_new_impl] &&
14167	meta.func_id != special_kfunc_list[KF_bpf_percpu_obj_new_impl] &&
14168	meta.func_id != special_kfunc_list[KF_bpf_refcount_acquire_impl])) {
14169	verbose(private_data: env, fmt: "acquire kernel function does not return PTR_TO_BTF_ID\n");
14170	return -EINVAL;
14171	}
14172	}
14173
14174	if (btf_type_is_scalar(t)) {
14175	mark_reg_unknown(env, regs, regno: BPF_REG_0);
14176	if (meta.btf == btf_vmlinux && (meta.func_id == special_kfunc_list[KF_bpf_res_spin_lock] ||
14177	meta.func_id == special_kfunc_list[KF_bpf_res_spin_lock_irqsave]))
14178	__mark_reg_const_zero(env, reg: &regs[BPF_REG_0]);
14179	mark_btf_func_reg_size(env, regno: BPF_REG_0, reg_size: t->size);
14180	} else if (btf_type_is_ptr(t)) {
14181	ptr_type = btf_type_skip_modifiers(btf: desc_btf, id: t->type, res_id: &ptr_type_id);
14182	err = check_special_kfunc(env, meta: &meta, regs, insn_aux, ptr_type, desc_btf);
14183	if (err) {
14184	if (err < 0)
14185	return err;
14186	} else if (btf_type_is_void(t: ptr_type)) {
14187	/* kfunc returning 'void *' is equivalent to returning scalar */
14188	mark_reg_unknown(env, regs, regno: BPF_REG_0);
14189	} else if (!__btf_type_is_struct(t: ptr_type)) {
14190	if (!meta.r0_size) {
14191	__u32 sz;
14192
14193	if (!IS_ERR(ptr: btf_resolve_size(btf: desc_btf, type: ptr_type, type_size: &sz))) {
14194	meta.r0_size = sz;
14195	meta.r0_rdonly = true;
14196	}
14197	}
14198	if (!meta.r0_size) {
14199	ptr_type_name = btf_name_by_offset(btf: desc_btf,
14200	offset: ptr_type->name_off);
14201	verbose(private_data: env,
14202	fmt: "kernel function %s returns pointer type %s %s is not supported\n",
14203	func_name,
14204	btf_type_str(t: ptr_type),
14205	ptr_type_name);
14206	return -EINVAL;
14207	}
14208
14209	mark_reg_known_zero(env, regs, regno: BPF_REG_0);
14210	regs[BPF_REG_0].type = PTR_TO_MEM;
14211	regs[BPF_REG_0].mem_size = meta.r0_size;
14212
14213	if (meta.r0_rdonly)
14214	regs[BPF_REG_0].type |= MEM_RDONLY;
14215
14216	/* Ensures we don't access the memory after a release_reference() */
14217	if (meta.ref_obj_id)
14218	regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
14219
14220	if (is_kfunc_rcu_protected(meta: &meta))
14221	regs[BPF_REG_0].type |= MEM_RCU;
14222	} else {
14223	mark_reg_known_zero(env, regs, regno: BPF_REG_0);
14224	regs[BPF_REG_0].btf = desc_btf;
14225	regs[BPF_REG_0].type = PTR_TO_BTF_ID;
14226	regs[BPF_REG_0].btf_id = ptr_type_id;
14227
14228	if (meta.func_id == special_kfunc_list[KF_bpf_get_kmem_cache])
14229	regs[BPF_REG_0].type |= PTR_UNTRUSTED;
14230	else if (is_kfunc_rcu_protected(meta: &meta))
14231	regs[BPF_REG_0].type |= MEM_RCU;
14232
14233	if (is_iter_next_kfunc(meta: &meta)) {
14234	struct bpf_reg_state *cur_iter;
14235
14236	cur_iter = get_iter_from_state(cur_st: env->cur_state, meta: &meta);
14237
14238	if (cur_iter->type & MEM_RCU) /* KF_RCU_PROTECTED */
14239	regs[BPF_REG_0].type |= MEM_RCU;
14240	else
14241	regs[BPF_REG_0].type |= PTR_TRUSTED;
14242	}
14243	}
14244
14245	if (is_kfunc_ret_null(meta: &meta)) {
14246	regs[BPF_REG_0].type |= PTR_MAYBE_NULL;
14247	/* For mark_ptr_or_null_reg, see 93c230e3f5bd6 */
14248	regs[BPF_REG_0].id = ++env->id_gen;
14249	}
14250	mark_btf_func_reg_size(env, regno: BPF_REG_0, reg_size: sizeof(void *));
14251	if (is_kfunc_acquire(meta: &meta)) {
14252	int id = acquire_reference(env, insn_idx);
14253
14254	if (id < 0)
14255	return id;
14256	if (is_kfunc_ret_null(meta: &meta))
14257	regs[BPF_REG_0].id = id;
14258	regs[BPF_REG_0].ref_obj_id = id;
14259	} else if (is_rbtree_node_type(t: ptr_type) || is_list_node_type(t: ptr_type)) {
14260	ref_set_non_owning(env, reg: &regs[BPF_REG_0]);
14261	}
14262
14263	if (reg_may_point_to_spin_lock(reg: &regs[BPF_REG_0]) && !regs[BPF_REG_0].id)
14264	regs[BPF_REG_0].id = ++env->id_gen;
14265	} else if (btf_type_is_void(t)) {
14266	if (meta.btf == btf_vmlinux) {
14267	if (meta.func_id == special_kfunc_list[KF_bpf_obj_drop_impl] ||
14268	meta.func_id == special_kfunc_list[KF_bpf_percpu_obj_drop_impl]) {
14269	insn_aux->kptr_struct_meta =
14270	btf_find_struct_meta(btf: meta.arg_btf,
14271	btf_id: meta.arg_btf_id);
14272	}
14273	}
14274	}
14275
14276	if (is_kfunc_pkt_changing(meta: &meta))
14277	clear_all_pkt_pointers(env);
14278
14279	nargs = btf_type_vlen(t: meta.func_proto);
14280	args = (const struct btf_param *)(meta.func_proto + 1);
14281	for (i = 0; i < nargs; i++) {
14282	u32 regno = i + 1;
14283
14284	t = btf_type_skip_modifiers(btf: desc_btf, id: args[i].type, NULL);
14285	if (btf_type_is_ptr(t))
14286	mark_btf_func_reg_size(env, regno, reg_size: sizeof(void *));
14287	else
14288	/* scalar. ensured by btf_check_kfunc_arg_match() */
14289	mark_btf_func_reg_size(env, regno, reg_size: t->size);
14290	}
14291
14292	if (is_iter_next_kfunc(meta: &meta)) {
14293	err = process_iter_next_call(env, insn_idx, meta: &meta);
14294	if (err)
14295	return err;
14296	}
14297
14298	return 0;
14299	}
14300
14301	static bool check_reg_sane_offset(struct bpf_verifier_env *env,
14302	const struct bpf_reg_state *reg,
14303	enum bpf_reg_type type)
14304	{
14305	bool known = tnum_is_const(a: reg->var_off);
14306	s64 val = reg->var_off.value;
14307	s64 smin = reg->smin_value;
14308
14309	if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
14310	verbose(private_data: env, fmt: "math between %s pointer and %lld is not allowed\n",
14311	reg_type_str(env, type), val);
14312	return false;
14313	}
14314
14315	if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
14316	verbose(private_data: env, fmt: "%s pointer offset %d is not allowed\n",
14317	reg_type_str(env, type), reg->off);
14318	return false;
14319	}
14320
14321	if (smin == S64_MIN) {
14322	verbose(private_data: env, fmt: "math between %s pointer and register with unbounded min value is not allowed\n",
14323	reg_type_str(env, type));
14324	return false;
14325	}
14326
14327	if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
14328	verbose(private_data: env, fmt: "value %lld makes %s pointer be out of bounds\n",
14329	smin, reg_type_str(env, type));
14330	return false;
14331	}
14332
14333	return true;
14334	}
14335
14336	enum {
14337	REASON_BOUNDS = -1,
14338	REASON_TYPE = -2,
14339	REASON_PATHS = -3,
14340	REASON_LIMIT = -4,
14341	REASON_STACK = -5,
14342	};
14343
14344	static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
14345	u32 *alu_limit, bool mask_to_left)
14346	{
14347	u32 max = 0, ptr_limit = 0;
14348
14349	switch (ptr_reg->type) {
14350	case PTR_TO_STACK:
14351	/* Offset 0 is out-of-bounds, but acceptable start for the
14352	* left direction, see BPF_REG_FP. Also, unknown scalar
14353	* offset where we would need to deal with min/max bounds is
14354	* currently prohibited for unprivileged.
14355	*/
14356	max = MAX_BPF_STACK + mask_to_left;
14357	ptr_limit = -(ptr_reg->var_off.value + ptr_reg->off);
14358	break;
14359	case PTR_TO_MAP_VALUE:
14360	max = ptr_reg->map_ptr->value_size;
14361	ptr_limit = (mask_to_left ?
14362	ptr_reg->smin_value :
14363	ptr_reg->umax_value) + ptr_reg->off;
14364	break;
14365	default:
14366	return REASON_TYPE;
14367	}
14368
14369	if (ptr_limit >= max)
14370	return REASON_LIMIT;
14371	*alu_limit = ptr_limit;
14372	return 0;
14373	}
14374
14375	static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
14376	const struct bpf_insn *insn)
14377	{
14378	return env->bypass_spec_v1 ||
14379	BPF_SRC(insn->code) == BPF_K ||
14380	cur_aux(env)->nospec;
14381	}
14382
14383	static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
14384	u32 alu_state, u32 alu_limit)
14385	{
14386	/* If we arrived here from different branches with different
14387	* state or limits to sanitize, then this won't work.
14388	*/
14389	if (aux->alu_state &&
14390	(aux->alu_state != alu_state ||
14391	aux->alu_limit != alu_limit))
14392	return REASON_PATHS;
14393
14394	/* Corresponding fixup done in do_misc_fixups(). */
14395	aux->alu_state = alu_state;
14396	aux->alu_limit = alu_limit;
14397	return 0;
14398	}
14399
14400	static int sanitize_val_alu(struct bpf_verifier_env *env,
14401	struct bpf_insn *insn)
14402	{
14403	struct bpf_insn_aux_data *aux = cur_aux(env);
14404
14405	if (can_skip_alu_sanitation(env, insn))
14406	return 0;
14407
14408	return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, alu_limit: 0);
14409	}
14410
14411	static bool sanitize_needed(u8 opcode)
14412	{
14413	return opcode == BPF_ADD || opcode == BPF_SUB;
14414	}
14415
14416	struct bpf_sanitize_info {
14417	struct bpf_insn_aux_data aux;
14418	bool mask_to_left;
14419	};
14420
14421	static int sanitize_speculative_path(struct bpf_verifier_env *env,
14422	const struct bpf_insn *insn,
14423	u32 next_idx, u32 curr_idx)
14424	{
14425	struct bpf_verifier_state *branch;
14426	struct bpf_reg_state *regs;
14427
14428	branch = push_stack(env, insn_idx: next_idx, prev_insn_idx: curr_idx, speculative: true);
14429	if (!IS_ERR(ptr: branch) && insn) {
14430	regs = branch->frame[branch->curframe]->regs;
14431	if (BPF_SRC(insn->code) == BPF_K) {
14432	mark_reg_unknown(env, regs, regno: insn->dst_reg);
14433	} else if (BPF_SRC(insn->code) == BPF_X) {
14434	mark_reg_unknown(env, regs, regno: insn->dst_reg);
14435	mark_reg_unknown(env, regs, regno: insn->src_reg);
14436	}
14437	}
14438	return PTR_ERR_OR_ZERO(ptr: branch);
14439	}
14440
14441	static int sanitize_ptr_alu(struct bpf_verifier_env *env,
14442	struct bpf_insn *insn,
14443	const struct bpf_reg_state *ptr_reg,
14444	const struct bpf_reg_state *off_reg,
14445	struct bpf_reg_state *dst_reg,
14446	struct bpf_sanitize_info *info,
14447	const bool commit_window)
14448	{
14449	struct bpf_insn_aux_data *aux = commit_window ? cur_aux(env) : &info->aux;
14450	struct bpf_verifier_state *vstate = env->cur_state;
14451	bool off_is_imm = tnum_is_const(a: off_reg->var_off);
14452	bool off_is_neg = off_reg->smin_value < 0;
14453	bool ptr_is_dst_reg = ptr_reg == dst_reg;
14454	u8 opcode = BPF_OP(insn->code);
14455	u32 alu_state, alu_limit;
14456	struct bpf_reg_state tmp;
14457	int err;
14458
14459	if (can_skip_alu_sanitation(env, insn))
14460	return 0;
14461
14462	/* We already marked aux for masking from non-speculative
14463	* paths, thus we got here in the first place. We only care
14464	* to explore bad access from here.
14465	*/
14466	if (vstate->speculative)
14467	goto do_sim;
14468
14469	if (!commit_window) {
14470	if (!tnum_is_const(a: off_reg->var_off) &&
14471	(off_reg->smin_value < 0) != (off_reg->smax_value < 0))
14472	return REASON_BOUNDS;
14473
14474	info->mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
14475	(opcode == BPF_SUB && !off_is_neg);
14476	}
14477
14478	err = retrieve_ptr_limit(ptr_reg, alu_limit: &alu_limit, mask_to_left: info->mask_to_left);
14479	if (err < 0)
14480	return err;
14481
14482	if (commit_window) {
14483	/* In commit phase we narrow the masking window based on
14484	* the observed pointer move after the simulated operation.
14485	*/
14486	alu_state = info->aux.alu_state;
14487	alu_limit = abs(info->aux.alu_limit - alu_limit);
14488	} else {
14489	alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
14490	alu_state |= off_is_imm ? BPF_ALU_IMMEDIATE : 0;
14491	alu_state |= ptr_is_dst_reg ?
14492	BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
14493
14494	/* Limit pruning on unknown scalars to enable deep search for
14495	* potential masking differences from other program paths.
14496	*/
14497	if (!off_is_imm)
14498	env->explore_alu_limits = true;
14499	}
14500
14501	err = update_alu_sanitation_state(aux, alu_state, alu_limit);
14502	if (err < 0)
14503	return err;
14504	do_sim:
14505	/* If we're in commit phase, we're done here given we already
14506	* pushed the truncated dst_reg into the speculative verification
14507	* stack.
14508	*
14509	* Also, when register is a known constant, we rewrite register-based
14510	* operation to immediate-based, and thus do not need masking (and as
14511	* a consequence, do not need to simulate the zero-truncation either).
14512	*/
14513	if (commit_window || off_is_imm)
14514	return 0;
14515
14516	/* Simulate and find potential out-of-bounds access under
14517	* speculative execution from truncation as a result of
14518	* masking when off was not within expected range. If off
14519	* sits in dst, then we temporarily need to move ptr there
14520	* to simulate dst (== 0) +/-= ptr. Needed, for example,
14521	* for cases where we use K-based arithmetic in one direction
14522	* and truncated reg-based in the other in order to explore
14523	* bad access.
14524	*/
14525	if (!ptr_is_dst_reg) {
14526	tmp = *dst_reg;
14527	copy_register_state(dst: dst_reg, src: ptr_reg);
14528	}
14529	err = sanitize_speculative_path(env, NULL, next_idx: env->insn_idx + 1, curr_idx: env->insn_idx);
14530	if (err < 0)
14531	return REASON_STACK;
14532	if (!ptr_is_dst_reg)
14533	*dst_reg = tmp;
14534	return 0;
14535	}
14536
14537	static void sanitize_mark_insn_seen(struct bpf_verifier_env *env)
14538	{
14539	struct bpf_verifier_state *vstate = env->cur_state;
14540
14541	/* If we simulate paths under speculation, we don't update the
14542	* insn as 'seen' such that when we verify unreachable paths in
14543	* the non-speculative domain, sanitize_dead_code() can still
14544	* rewrite/sanitize them.
14545	*/
14546	if (!vstate->speculative)
14547	env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
14548	}
14549
14550	static int sanitize_err(struct bpf_verifier_env *env,
14551	const struct bpf_insn *insn, int reason,
14552	const struct bpf_reg_state *off_reg,
14553	const struct bpf_reg_state *dst_reg)
14554	{
14555	static const char *err = "pointer arithmetic with it prohibited for !root";
14556	const char *op = BPF_OP(insn->code) == BPF_ADD ? "add" : "sub";
14557	u32 dst = insn->dst_reg, src = insn->src_reg;
14558
14559	switch (reason) {
14560	case REASON_BOUNDS:
14561	verbose(private_data: env, fmt: "R%d has unknown scalar with mixed signed bounds, %s\n",
14562	off_reg == dst_reg ? dst : src, err);
14563	break;
14564	case REASON_TYPE:
14565	verbose(private_data: env, fmt: "R%d has pointer with unsupported alu operation, %s\n",
14566	off_reg == dst_reg ? src : dst, err);
14567	break;
14568	case REASON_PATHS:
14569	verbose(private_data: env, fmt: "R%d tried to %s from different maps, paths or scalars, %s\n",
14570	dst, op, err);
14571	break;
14572	case REASON_LIMIT:
14573	verbose(private_data: env, fmt: "R%d tried to %s beyond pointer bounds, %s\n",
14574	dst, op, err);
14575	break;
14576	case REASON_STACK:
14577	verbose(private_data: env, fmt: "R%d could not be pushed for speculative verification, %s\n",
14578	dst, err);
14579	return -ENOMEM;
14580	default:
14581	verifier_bug(env, "unknown reason (%d)", reason);
14582	break;
14583	}
14584
14585	return -EACCES;
14586	}
14587
14588	/* check that stack access falls within stack limits and that 'reg' doesn't
14589	* have a variable offset.
14590	*
14591	* Variable offset is prohibited for unprivileged mode for simplicity since it
14592	* requires corresponding support in Spectre masking for stack ALU. See also
14593	* retrieve_ptr_limit().
14594	*
14595	*
14596	* 'off' includes 'reg->off'.
14597	*/
14598	static int check_stack_access_for_ptr_arithmetic(
14599	struct bpf_verifier_env *env,
14600	int regno,
14601	const struct bpf_reg_state *reg,
14602	int off)
14603	{
14604	if (!tnum_is_const(a: reg->var_off)) {
14605	char tn_buf[48];
14606
14607	tnum_strn(str: tn_buf, size: sizeof(tn_buf), a: reg->var_off);
14608	verbose(private_data: env, fmt: "R%d variable stack access prohibited for !root, var_off=%s off=%d\n",
14609	regno, tn_buf, off);
14610	return -EACCES;
14611	}
14612
14613	if (off >= 0 || off < -MAX_BPF_STACK) {
14614	verbose(private_data: env, fmt: "R%d stack pointer arithmetic goes out of range, "
14615	"prohibited for !root; off=%d\n", regno, off);
14616	return -EACCES;
14617	}
14618
14619	return 0;
14620	}
14621
14622	static int sanitize_check_bounds(struct bpf_verifier_env *env,
14623	const struct bpf_insn *insn,
14624	const struct bpf_reg_state *dst_reg)
14625	{
14626	u32 dst = insn->dst_reg;
14627
14628	/* For unprivileged we require that resulting offset must be in bounds
14629	* in order to be able to sanitize access later on.
14630	*/
14631	if (env->bypass_spec_v1)
14632	return 0;
14633
14634	switch (dst_reg->type) {
14635	case PTR_TO_STACK:
14636	if (check_stack_access_for_ptr_arithmetic(env, regno: dst, reg: dst_reg,
14637	off: dst_reg->off + dst_reg->var_off.value))
14638	return -EACCES;
14639	break;
14640	case PTR_TO_MAP_VALUE:
14641	if (check_map_access(env, regno: dst, off: dst_reg->off, size: 1, zero_size_allowed: false, src: ACCESS_HELPER)) {
14642	verbose(private_data: env, fmt: "R%d pointer arithmetic of map value goes out of range, "
14643	"prohibited for !root\n", dst);
14644	return -EACCES;
14645	}
14646	break;
14647	default:
14648	return -EOPNOTSUPP;
14649	}
14650
14651	return 0;
14652	}
14653
14654	/* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
14655	* Caller should also handle BPF_MOV case separately.
14656	* If we return -EACCES, caller may want to try again treating pointer as a
14657	* scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
14658	*/
14659	static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
14660	struct bpf_insn *insn,
14661	const struct bpf_reg_state *ptr_reg,
14662	const struct bpf_reg_state *off_reg)
14663	{
14664	struct bpf_verifier_state *vstate = env->cur_state;
14665	struct bpf_func_state *state = vstate->frame[vstate->curframe];
14666	struct bpf_reg_state *regs = state->regs, *dst_reg;
14667	bool known = tnum_is_const(a: off_reg->var_off);
14668	s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
14669	smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
14670	u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
14671	umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
14672	struct bpf_sanitize_info info = {};
14673	u8 opcode = BPF_OP(insn->code);
14674	u32 dst = insn->dst_reg;
14675	int ret, bounds_ret;
14676
14677	dst_reg = &regs[dst];
14678
14679	if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
14680	smin_val > smax_val || umin_val > umax_val) {
14681	/* Taint dst register if offset had invalid bounds derived from
14682	* e.g. dead branches.
14683	*/
14684	__mark_reg_unknown(env, reg: dst_reg);
14685	return 0;
14686	}
14687
14688	if (BPF_CLASS(insn->code) != BPF_ALU64) {
14689	/* 32-bit ALU ops on pointers produce (meaningless) scalars */
14690	if (opcode == BPF_SUB && env->allow_ptr_leaks) {
14691	__mark_reg_unknown(env, reg: dst_reg);
14692	return 0;
14693	}
14694
14695	verbose(private_data: env,
14696	fmt: "R%d 32-bit pointer arithmetic prohibited\n",
14697	dst);
14698	return -EACCES;
14699	}
14700
14701	if (ptr_reg->type & PTR_MAYBE_NULL) {
14702	verbose(private_data: env, fmt: "R%d pointer arithmetic on %s prohibited, null-check it first\n",
14703	dst, reg_type_str(env, type: ptr_reg->type));
14704	return -EACCES;
14705	}
14706
14707	/*
14708	* Accesses to untrusted PTR_TO_MEM are done through probe
14709	* instructions, hence no need to track offsets.
14710	*/
14711	if (base_type(type: ptr_reg->type) == PTR_TO_MEM && (ptr_reg->type & PTR_UNTRUSTED))
14712	return 0;
14713
14714	switch (base_type(type: ptr_reg->type)) {
14715	case PTR_TO_CTX:
14716	case PTR_TO_MAP_VALUE:
14717	case PTR_TO_MAP_KEY:
14718	case PTR_TO_STACK:
14719	case PTR_TO_PACKET_META:
14720	case PTR_TO_PACKET:
14721	case PTR_TO_TP_BUFFER:
14722	case PTR_TO_BTF_ID:
14723	case PTR_TO_MEM:
14724	case PTR_TO_BUF:
14725	case PTR_TO_FUNC:
14726	case CONST_PTR_TO_DYNPTR:
14727	break;
14728	case PTR_TO_FLOW_KEYS:
14729	if (known)
14730	break;
14731	fallthrough;
14732	case CONST_PTR_TO_MAP:
14733	/* smin_val represents the known value */
14734	if (known && smin_val == 0 && opcode == BPF_ADD)
14735	break;
14736	fallthrough;
14737	default:
14738	verbose(private_data: env, fmt: "R%d pointer arithmetic on %s prohibited\n",
14739	dst, reg_type_str(env, type: ptr_reg->type));
14740	return -EACCES;
14741	}
14742
14743	/* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
14744	* The id may be overwritten later if we create a new variable offset.
14745	*/
14746	dst_reg->type = ptr_reg->type;
14747	dst_reg->id = ptr_reg->id;
14748
14749	if (!check_reg_sane_offset(env, reg: off_reg, type: ptr_reg->type) ||
14750	!check_reg_sane_offset(env, reg: ptr_reg, type: ptr_reg->type))
14751	return -EINVAL;
14752
14753	/* pointer types do not carry 32-bit bounds at the moment. */
14754	__mark_reg32_unbounded(reg: dst_reg);
14755
14756	if (sanitize_needed(opcode)) {
14757	ret = sanitize_ptr_alu(env, insn, ptr_reg, off_reg, dst_reg,
14758	info: &info, commit_window: false);
14759	if (ret < 0)
14760	return sanitize_err(env, insn, reason: ret, off_reg, dst_reg);
14761	}
14762
14763	switch (opcode) {
14764	case BPF_ADD:
14765	/* We can take a fixed offset as long as it doesn't overflow
14766	* the s32 'off' field
14767	*/
14768	if (known && (ptr_reg->off + smin_val ==
14769	(s64)(s32)(ptr_reg->off + smin_val))) {
14770	/* pointer += K. Accumulate it into fixed offset */
14771	dst_reg->smin_value = smin_ptr;
14772	dst_reg->smax_value = smax_ptr;
14773	dst_reg->umin_value = umin_ptr;
14774	dst_reg->umax_value = umax_ptr;
14775	dst_reg->var_off = ptr_reg->var_off;
14776	dst_reg->off = ptr_reg->off + smin_val;
14777	dst_reg->raw = ptr_reg->raw;
14778	break;
14779	}
14780	/* A new variable offset is created. Note that off_reg->off
14781	* == 0, since it's a scalar.
14782	* dst_reg gets the pointer type and since some positive
14783	* integer value was added to the pointer, give it a new 'id'
14784	* if it's a PTR_TO_PACKET.
14785	* this creates a new 'base' pointer, off_reg (variable) gets
14786	* added into the variable offset, and we copy the fixed offset
14787	* from ptr_reg.
14788	*/
14789	if (check_add_overflow(smin_ptr, smin_val, &dst_reg->smin_value) ||
14790	check_add_overflow(smax_ptr, smax_val, &dst_reg->smax_value)) {
14791	dst_reg->smin_value = S64_MIN;
14792	dst_reg->smax_value = S64_MAX;
14793	}
14794	if (check_add_overflow(umin_ptr, umin_val, &dst_reg->umin_value) ||
14795	check_add_overflow(umax_ptr, umax_val, &dst_reg->umax_value)) {
14796	dst_reg->umin_value = 0;
14797	dst_reg->umax_value = U64_MAX;
14798	}
14799	dst_reg->var_off = tnum_add(a: ptr_reg->var_off, b: off_reg->var_off);
14800	dst_reg->off = ptr_reg->off;
14801	dst_reg->raw = ptr_reg->raw;
14802	if (reg_is_pkt_pointer(reg: ptr_reg)) {
14803	dst_reg->id = ++env->id_gen;
14804	/* something was added to pkt_ptr, set range to zero */
14805	memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
14806	}
14807	break;
14808	case BPF_SUB:
14809	if (dst_reg == off_reg) {
14810	/* scalar -= pointer. Creates an unknown scalar */
14811	verbose(private_data: env, fmt: "R%d tried to subtract pointer from scalar\n",
14812	dst);
14813	return -EACCES;
14814	}
14815	/* We don't allow subtraction from FP, because (according to
14816	* test_verifier.c test "invalid fp arithmetic", JITs might not
14817	* be able to deal with it.
14818	*/
14819	if (ptr_reg->type == PTR_TO_STACK) {
14820	verbose(private_data: env, fmt: "R%d subtraction from stack pointer prohibited\n",
14821	dst);
14822	return -EACCES;
14823	}
14824	if (known && (ptr_reg->off - smin_val ==
14825	(s64)(s32)(ptr_reg->off - smin_val))) {
14826	/* pointer -= K. Subtract it from fixed offset */
14827	dst_reg->smin_value = smin_ptr;
14828	dst_reg->smax_value = smax_ptr;
14829	dst_reg->umin_value = umin_ptr;
14830	dst_reg->umax_value = umax_ptr;
14831	dst_reg->var_off = ptr_reg->var_off;
14832	dst_reg->id = ptr_reg->id;
14833	dst_reg->off = ptr_reg->off - smin_val;
14834	dst_reg->raw = ptr_reg->raw;
14835	break;
14836	}
14837	/* A new variable offset is created. If the subtrahend is known
14838	* nonnegative, then any reg->range we had before is still good.
14839	*/
14840	if (check_sub_overflow(smin_ptr, smax_val, &dst_reg->smin_value) ||
14841	check_sub_overflow(smax_ptr, smin_val, &dst_reg->smax_value)) {
14842	/* Overflow possible, we know nothing */
14843	dst_reg->smin_value = S64_MIN;
14844	dst_reg->smax_value = S64_MAX;
14845	}
14846	if (umin_ptr < umax_val) {
14847	/* Overflow possible, we know nothing */
14848	dst_reg->umin_value = 0;
14849	dst_reg->umax_value = U64_MAX;
14850	} else {
14851	/* Cannot overflow (as long as bounds are consistent) */
14852	dst_reg->umin_value = umin_ptr - umax_val;
14853	dst_reg->umax_value = umax_ptr - umin_val;
14854	}
14855	dst_reg->var_off = tnum_sub(a: ptr_reg->var_off, b: off_reg->var_off);
14856	dst_reg->off = ptr_reg->off;
14857	dst_reg->raw = ptr_reg->raw;
14858	if (reg_is_pkt_pointer(reg: ptr_reg)) {
14859	dst_reg->id = ++env->id_gen;
14860	/* something was added to pkt_ptr, set range to zero */
14861	if (smin_val < 0)
14862	memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
14863	}
14864	break;
14865	case BPF_AND:
14866	case BPF_OR:
14867	case BPF_XOR:
14868	/* bitwise ops on pointers are troublesome, prohibit. */
14869	verbose(private_data: env, fmt: "R%d bitwise operator %s on pointer prohibited\n",
14870	dst, bpf_alu_string[opcode >> 4]);
14871	return -EACCES;
14872	default:
14873	/* other operators (e.g. MUL,LSH) produce non-pointer results */
14874	verbose(private_data: env, fmt: "R%d pointer arithmetic with %s operator prohibited\n",
14875	dst, bpf_alu_string[opcode >> 4]);
14876	return -EACCES;
14877	}
14878
14879	if (!check_reg_sane_offset(env, reg: dst_reg, type: ptr_reg->type))
14880	return -EINVAL;
14881	reg_bounds_sync(reg: dst_reg);
14882	bounds_ret = sanitize_check_bounds(env, insn, dst_reg);
14883	if (bounds_ret == -EACCES)
14884	return bounds_ret;
14885	if (sanitize_needed(opcode)) {
14886	ret = sanitize_ptr_alu(env, insn, ptr_reg: dst_reg, off_reg, dst_reg,
14887	info: &info, commit_window: true);
14888	if (verifier_bug_if(!can_skip_alu_sanitation(env, insn)
14889	&& !env->cur_state->speculative
14890	&& bounds_ret
14891	&& !ret,
14892	env, "Pointer type unsupported by sanitize_check_bounds() not rejected by retrieve_ptr_limit() as required")) {
14893	return -EFAULT;
14894	}
14895	if (ret < 0)
14896	return sanitize_err(env, insn, reason: ret, off_reg, dst_reg);
14897	}
14898
14899	return 0;
14900	}
14901
14902	static void scalar32_min_max_add(struct bpf_reg_state *dst_reg,
14903	struct bpf_reg_state *src_reg)
14904	{
14905	s32 *dst_smin = &dst_reg->s32_min_value;
14906	s32 *dst_smax = &dst_reg->s32_max_value;
14907	u32 *dst_umin = &dst_reg->u32_min_value;
14908	u32 *dst_umax = &dst_reg->u32_max_value;
14909	u32 umin_val = src_reg->u32_min_value;
14910	u32 umax_val = src_reg->u32_max_value;
14911	bool min_overflow, max_overflow;
14912
14913	if (check_add_overflow(*dst_smin, src_reg->s32_min_value, dst_smin) ||
14914	check_add_overflow(*dst_smax, src_reg->s32_max_value, dst_smax)) {
14915	*dst_smin = S32_MIN;
14916	*dst_smax = S32_MAX;
14917	}
14918
14919	/* If either all additions overflow or no additions overflow, then
14920	* it is okay to set: dst_umin = dst_umin + src_umin, dst_umax =
14921	* dst_umax + src_umax. Otherwise (some additions overflow), set
14922	* the output bounds to unbounded.
14923	*/
14924	min_overflow = check_add_overflow(*dst_umin, umin_val, dst_umin);
14925	max_overflow = check_add_overflow(*dst_umax, umax_val, dst_umax);
14926
14927	if (!min_overflow && max_overflow) {
14928	*dst_umin = 0;
14929	*dst_umax = U32_MAX;
14930	}
14931	}
14932
14933	static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
14934	struct bpf_reg_state *src_reg)
14935	{
14936	s64 *dst_smin = &dst_reg->smin_value;
14937	s64 *dst_smax = &dst_reg->smax_value;
14938	u64 *dst_umin = &dst_reg->umin_value;
14939	u64 *dst_umax = &dst_reg->umax_value;
14940	u64 umin_val = src_reg->umin_value;
14941	u64 umax_val = src_reg->umax_value;
14942	bool min_overflow, max_overflow;
14943
14944	if (check_add_overflow(*dst_smin, src_reg->smin_value, dst_smin) ||
14945	check_add_overflow(*dst_smax, src_reg->smax_value, dst_smax)) {
14946	*dst_smin = S64_MIN;
14947	*dst_smax = S64_MAX;
14948	}
14949
14950	/* If either all additions overflow or no additions overflow, then
14951	* it is okay to set: dst_umin = dst_umin + src_umin, dst_umax =
14952	* dst_umax + src_umax. Otherwise (some additions overflow), set
14953	* the output bounds to unbounded.
14954	*/
14955	min_overflow = check_add_overflow(*dst_umin, umin_val, dst_umin);
14956	max_overflow = check_add_overflow(*dst_umax, umax_val, dst_umax);
14957
14958	if (!min_overflow && max_overflow) {
14959	*dst_umin = 0;
14960	*dst_umax = U64_MAX;
14961	}
14962	}
14963
14964	static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg,
14965	struct bpf_reg_state *src_reg)
14966	{
14967	s32 *dst_smin = &dst_reg->s32_min_value;
14968	s32 *dst_smax = &dst_reg->s32_max_value;
14969	u32 *dst_umin = &dst_reg->u32_min_value;
14970	u32 *dst_umax = &dst_reg->u32_max_value;
14971	u32 umin_val = src_reg->u32_min_value;
14972	u32 umax_val = src_reg->u32_max_value;
14973	bool min_underflow, max_underflow;
14974
14975	if (check_sub_overflow(*dst_smin, src_reg->s32_max_value, dst_smin) ||
14976	check_sub_overflow(*dst_smax, src_reg->s32_min_value, dst_smax)) {
14977	/* Overflow possible, we know nothing */
14978	*dst_smin = S32_MIN;
14979	*dst_smax = S32_MAX;
14980	}
14981
14982	/* If either all subtractions underflow or no subtractions
14983	* underflow, it is okay to set: dst_umin = dst_umin - src_umax,
14984	* dst_umax = dst_umax - src_umin. Otherwise (some subtractions
14985	* underflow), set the output bounds to unbounded.
14986	*/
14987	min_underflow = check_sub_overflow(*dst_umin, umax_val, dst_umin);
14988	max_underflow = check_sub_overflow(*dst_umax, umin_val, dst_umax);
14989
14990	if (min_underflow && !max_underflow) {
14991	*dst_umin = 0;
14992	*dst_umax = U32_MAX;
14993	}
14994	}
14995
14996	static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
14997	struct bpf_reg_state *src_reg)
14998	{
14999	s64 *dst_smin = &dst_reg->smin_value;
15000	s64 *dst_smax = &dst_reg->smax_value;
15001	u64 *dst_umin = &dst_reg->umin_value;
15002	u64 *dst_umax = &dst_reg->umax_value;
15003	u64 umin_val = src_reg->umin_value;
15004	u64 umax_val = src_reg->umax_value;
15005	bool min_underflow, max_underflow;
15006
15007	if (check_sub_overflow(*dst_smin, src_reg->smax_value, dst_smin) ||
15008	check_sub_overflow(*dst_smax, src_reg->smin_value, dst_smax)) {
15009	/* Overflow possible, we know nothing */
15010	*dst_smin = S64_MIN;
15011	*dst_smax = S64_MAX;
15012	}
15013
15014	/* If either all subtractions underflow or no subtractions
15015	* underflow, it is okay to set: dst_umin = dst_umin - src_umax,
15016	* dst_umax = dst_umax - src_umin. Otherwise (some subtractions
15017	* underflow), set the output bounds to unbounded.
15018	*/
15019	min_underflow = check_sub_overflow(*dst_umin, umax_val, dst_umin);
15020	max_underflow = check_sub_overflow(*dst_umax, umin_val, dst_umax);
15021
15022	if (min_underflow && !max_underflow) {
15023	*dst_umin = 0;
15024	*dst_umax = U64_MAX;
15025	}
15026	}
15027
15028	static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg,
15029	struct bpf_reg_state *src_reg)
15030	{
15031	s32 *dst_smin = &dst_reg->s32_min_value;
15032	s32 *dst_smax = &dst_reg->s32_max_value;
15033	u32 *dst_umin = &dst_reg->u32_min_value;
15034	u32 *dst_umax = &dst_reg->u32_max_value;
15035	s32 tmp_prod[4];
15036
15037	if (check_mul_overflow(*dst_umax, src_reg->u32_max_value, dst_umax) ||
15038	check_mul_overflow(*dst_umin, src_reg->u32_min_value, dst_umin)) {
15039	/* Overflow possible, we know nothing */
15040	*dst_umin = 0;
15041	*dst_umax = U32_MAX;
15042	}
15043	if (check_mul_overflow(*dst_smin, src_reg->s32_min_value, &tmp_prod[0]) ||
15044	check_mul_overflow(*dst_smin, src_reg->s32_max_value, &tmp_prod[1]) ||
15045	check_mul_overflow(*dst_smax, src_reg->s32_min_value, &tmp_prod[2]) ||
15046	check_mul_overflow(*dst_smax, src_reg->s32_max_value, &tmp_prod[3])) {
15047	/* Overflow possible, we know nothing */
15048	*dst_smin = S32_MIN;
15049	*dst_smax = S32_MAX;
15050	} else {
15051	*dst_smin = min_array(tmp_prod, 4);
15052	*dst_smax = max_array(tmp_prod, 4);
15053	}
15054	}
15055
15056	static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
15057	struct bpf_reg_state *src_reg)
15058	{
15059	s64 *dst_smin = &dst_reg->smin_value;
15060	s64 *dst_smax = &dst_reg->smax_value;
15061	u64 *dst_umin = &dst_reg->umin_value;
15062	u64 *dst_umax = &dst_reg->umax_value;
15063	s64 tmp_prod[4];
15064
15065	if (check_mul_overflow(*dst_umax, src_reg->umax_value, dst_umax) ||
15066	check_mul_overflow(*dst_umin, src_reg->umin_value, dst_umin)) {
15067	/* Overflow possible, we know nothing */
15068	*dst_umin = 0;
15069	*dst_umax = U64_MAX;
15070	}
15071	if (check_mul_overflow(*dst_smin, src_reg->smin_value, &tmp_prod[0]) ||
15072	check_mul_overflow(*dst_smin, src_reg->smax_value, &tmp_prod[1]) ||
15073	check_mul_overflow(*dst_smax, src_reg->smin_value, &tmp_prod[2]) ||
15074	check_mul_overflow(*dst_smax, src_reg->smax_value, &tmp_prod[3])) {
15075	/* Overflow possible, we know nothing */
15076	*dst_smin = S64_MIN;
15077	*dst_smax = S64_MAX;
15078	} else {
15079	*dst_smin = min_array(tmp_prod, 4);
15080	*dst_smax = max_array(tmp_prod, 4);
15081	}
15082	}
15083
15084	static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
15085	struct bpf_reg_state *src_reg)
15086	{
15087	bool src_known = tnum_subreg_is_const(a: src_reg->var_off);
15088	bool dst_known = tnum_subreg_is_const(a: dst_reg->var_off);
15089	struct tnum var32_off = tnum_subreg(a: dst_reg->var_off);
15090	u32 umax_val = src_reg->u32_max_value;
15091
15092	if (src_known && dst_known) {
15093	__mark_reg32_known(reg: dst_reg, imm: var32_off.value);
15094	return;
15095	}
15096
15097	/* We get our minimum from the var_off, since that's inherently
15098	* bitwise. Our maximum is the minimum of the operands' maxima.
15099	*/
15100	dst_reg->u32_min_value = var32_off.value;
15101	dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val);
15102
15103	/* Safe to set s32 bounds by casting u32 result into s32 when u32
15104	* doesn't cross sign boundary. Otherwise set s32 bounds to unbounded.
15105	*/
15106	if ((s32)dst_reg->u32_min_value <= (s32)dst_reg->u32_max_value) {
15107	dst_reg->s32_min_value = dst_reg->u32_min_value;
15108	dst_reg->s32_max_value = dst_reg->u32_max_value;
15109	} else {
15110	dst_reg->s32_min_value = S32_MIN;
15111	dst_reg->s32_max_value = S32_MAX;
15112	}
15113	}
15114
15115	static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
15116	struct bpf_reg_state *src_reg)
15117	{
15118	bool src_known = tnum_is_const(a: src_reg->var_off);
15119	bool dst_known = tnum_is_const(a: dst_reg->var_off);
15120	u64 umax_val = src_reg->umax_value;
15121
15122	if (src_known && dst_known) {
15123	__mark_reg_known(reg: dst_reg, imm: dst_reg->var_off.value);
15124	return;
15125	}
15126
15127	/* We get our minimum from the var_off, since that's inherently
15128	* bitwise. Our maximum is the minimum of the operands' maxima.
15129	*/
15130	dst_reg->umin_value = dst_reg->var_off.value;
15131	dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
15132
15133	/* Safe to set s64 bounds by casting u64 result into s64 when u64
15134	* doesn't cross sign boundary. Otherwise set s64 bounds to unbounded.
15135	*/
15136	if ((s64)dst_reg->umin_value <= (s64)dst_reg->umax_value) {
15137	dst_reg->smin_value = dst_reg->umin_value;
15138	dst_reg->smax_value = dst_reg->umax_value;
15139	} else {
15140	dst_reg->smin_value = S64_MIN;
15141	dst_reg->smax_value = S64_MAX;
15142	}
15143	/* We may learn something more from the var_off */
15144	__update_reg_bounds(reg: dst_reg);
15145	}
15146
15147	static void scalar32_min_max_or(struct bpf_reg_state *dst_reg,
15148	struct bpf_reg_state *src_reg)
15149	{
15150	bool src_known = tnum_subreg_is_const(a: src_reg->var_off);
15151	bool dst_known = tnum_subreg_is_const(a: dst_reg->var_off);
15152	struct tnum var32_off = tnum_subreg(a: dst_reg->var_off);
15153	u32 umin_val = src_reg->u32_min_value;
15154
15155	if (src_known && dst_known) {
15156	__mark_reg32_known(reg: dst_reg, imm: var32_off.value);
15157	return;
15158	}
15159
15160	/* We get our maximum from the var_off, and our minimum is the
15161	* maximum of the operands' minima
15162	*/
15163	dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val);
15164	dst_reg->u32_max_value = var32_off.value | var32_off.mask;
15165
15166	/* Safe to set s32 bounds by casting u32 result into s32 when u32
15167	* doesn't cross sign boundary. Otherwise set s32 bounds to unbounded.
15168	*/
15169	if ((s32)dst_reg->u32_min_value <= (s32)dst_reg->u32_max_value) {
15170	dst_reg->s32_min_value = dst_reg->u32_min_value;
15171	dst_reg->s32_max_value = dst_reg->u32_max_value;
15172	} else {
15173	dst_reg->s32_min_value = S32_MIN;
15174	dst_reg->s32_max_value = S32_MAX;
15175	}
15176	}
15177
15178	static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
15179	struct bpf_reg_state *src_reg)
15180	{
15181	bool src_known = tnum_is_const(a: src_reg->var_off);
15182	bool dst_known = tnum_is_const(a: dst_reg->var_off);
15183	u64 umin_val = src_reg->umin_value;
15184
15185	if (src_known && dst_known) {
15186	__mark_reg_known(reg: dst_reg, imm: dst_reg->var_off.value);
15187	return;
15188	}
15189
15190	/* We get our maximum from the var_off, and our minimum is the
15191	* maximum of the operands' minima
15192	*/
15193	dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
15194	dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
15195
15196	/* Safe to set s64 bounds by casting u64 result into s64 when u64
15197	* doesn't cross sign boundary. Otherwise set s64 bounds to unbounded.
15198	*/
15199	if ((s64)dst_reg->umin_value <= (s64)dst_reg->umax_value) {
15200	dst_reg->smin_value = dst_reg->umin_value;
15201	dst_reg->smax_value = dst_reg->umax_value;
15202	} else {
15203	dst_reg->smin_value = S64_MIN;
15204	dst_reg->smax_value = S64_MAX;
15205	}
15206	/* We may learn something more from the var_off */
15207	__update_reg_bounds(reg: dst_reg);
15208	}
15209
15210	static void scalar32_min_max_xor(struct bpf_reg_state *dst_reg,
15211	struct bpf_reg_state *src_reg)
15212	{
15213	bool src_known = tnum_subreg_is_const(a: src_reg->var_off);
15214	bool dst_known = tnum_subreg_is_const(a: dst_reg->var_off);
15215	struct tnum var32_off = tnum_subreg(a: dst_reg->var_off);
15216
15217	if (src_known && dst_known) {
15218	__mark_reg32_known(reg: dst_reg, imm: var32_off.value);
15219	return;
15220	}
15221
15222	/* We get both minimum and maximum from the var32_off. */
15223	dst_reg->u32_min_value = var32_off.value;
15224	dst_reg->u32_max_value = var32_off.value | var32_off.mask;
15225
15226	/* Safe to set s32 bounds by casting u32 result into s32 when u32
15227	* doesn't cross sign boundary. Otherwise set s32 bounds to unbounded.
15228	*/
15229	if ((s32)dst_reg->u32_min_value <= (s32)dst_reg->u32_max_value) {
15230	dst_reg->s32_min_value = dst_reg->u32_min_value;
15231	dst_reg->s32_max_value = dst_reg->u32_max_value;
15232	} else {
15233	dst_reg->s32_min_value = S32_MIN;
15234	dst_reg->s32_max_value = S32_MAX;
15235	}
15236	}
15237
15238	static void scalar_min_max_xor(struct bpf_reg_state *dst_reg,
15239	struct bpf_reg_state *src_reg)
15240	{
15241	bool src_known = tnum_is_const(a: src_reg->var_off);
15242	bool dst_known = tnum_is_const(a: dst_reg->var_off);
15243
15244	if (src_known && dst_known) {
15245	/* dst_reg->var_off.value has been updated earlier */
15246	__mark_reg_known(reg: dst_reg, imm: dst_reg->var_off.value);
15247	return;
15248	}
15249
15250	/* We get both minimum and maximum from the var_off. */
15251	dst_reg->umin_value = dst_reg->var_off.value;
15252	dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
15253
15254	/* Safe to set s64 bounds by casting u64 result into s64 when u64
15255	* doesn't cross sign boundary. Otherwise set s64 bounds to unbounded.
15256	*/
15257	if ((s64)dst_reg->umin_value <= (s64)dst_reg->umax_value) {
15258	dst_reg->smin_value = dst_reg->umin_value;
15259	dst_reg->smax_value = dst_reg->umax_value;
15260	} else {
15261	dst_reg->smin_value = S64_MIN;
15262	dst_reg->smax_value = S64_MAX;
15263	}
15264
15265	__update_reg_bounds(reg: dst_reg);
15266	}
15267
15268	static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
15269	u64 umin_val, u64 umax_val)
15270	{
15271	/* We lose all sign bit information (except what we can pick
15272	* up from var_off)
15273	*/
15274	dst_reg->s32_min_value = S32_MIN;
15275	dst_reg->s32_max_value = S32_MAX;
15276	/* If we might shift our top bit out, then we know nothing */
15277	if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) {
15278	dst_reg->u32_min_value = 0;
15279	dst_reg->u32_max_value = U32_MAX;
15280	} else {
15281	dst_reg->u32_min_value <<= umin_val;
15282	dst_reg->u32_max_value <<= umax_val;
15283	}
15284	}
15285
15286	static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
15287	struct bpf_reg_state *src_reg)
15288	{
15289	u32 umax_val = src_reg->u32_max_value;
15290	u32 umin_val = src_reg->u32_min_value;
15291	/* u32 alu operation will zext upper bits */
15292	struct tnum subreg = tnum_subreg(a: dst_reg->var_off);
15293
15294	__scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
15295	dst_reg->var_off = tnum_subreg(a: tnum_lshift(a: subreg, shift: umin_val));
15296	/* Not required but being careful mark reg64 bounds as unknown so
15297	* that we are forced to pick them up from tnum and zext later and
15298	* if some path skips this step we are still safe.
15299	*/
15300	__mark_reg64_unbounded(reg: dst_reg);
15301	__update_reg32_bounds(reg: dst_reg);
15302	}
15303
15304	static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg,
15305	u64 umin_val, u64 umax_val)
15306	{
15307	/* Special case <<32 because it is a common compiler pattern to sign
15308	* extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
15309	* positive we know this shift will also be positive so we can track
15310	* bounds correctly. Otherwise we lose all sign bit information except
15311	* what we can pick up from var_off. Perhaps we can generalize this
15312	* later to shifts of any length.
15313	*/
15314	if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0)
15315	dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32;
15316	else
15317	dst_reg->smax_value = S64_MAX;
15318
15319	if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0)
15320	dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32;
15321	else
15322	dst_reg->smin_value = S64_MIN;
15323
15324	/* If we might shift our top bit out, then we know nothing */
15325	if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
15326	dst_reg->umin_value = 0;
15327	dst_reg->umax_value = U64_MAX;
15328	} else {
15329	dst_reg->umin_value <<= umin_val;
15330	dst_reg->umax_value <<= umax_val;
15331	}
15332	}
15333
15334	static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
15335	struct bpf_reg_state *src_reg)
15336	{
15337	u64 umax_val = src_reg->umax_value;
15338	u64 umin_val = src_reg->umin_value;
15339
15340	/* scalar64 calc uses 32bit unshifted bounds so must be called first */
15341	__scalar64_min_max_lsh(dst_reg, umin_val, umax_val);
15342	__scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
15343
15344	dst_reg->var_off = tnum_lshift(a: dst_reg->var_off, shift: umin_val);
15345	/* We may learn something more from the var_off */
15346	__update_reg_bounds(reg: dst_reg);
15347	}
15348
15349	static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg,
15350	struct bpf_reg_state *src_reg)
15351	{
15352	struct tnum subreg = tnum_subreg(a: dst_reg->var_off);
15353	u32 umax_val = src_reg->u32_max_value;
15354	u32 umin_val = src_reg->u32_min_value;
15355
15356	/* BPF_RSH is an unsigned shift. If the value in dst_reg might
15357	* be negative, then either:
15358	* 1) src_reg might be zero, so the sign bit of the result is
15359	* unknown, so we lose our signed bounds
15360	* 2) it's known negative, thus the unsigned bounds capture the
15361	* signed bounds
15362	* 3) the signed bounds cross zero, so they tell us nothing
15363	* about the result
15364	* If the value in dst_reg is known nonnegative, then again the
15365	* unsigned bounds capture the signed bounds.
15366	* Thus, in all cases it suffices to blow away our signed bounds
15367	* and rely on inferring new ones from the unsigned bounds and
15368	* var_off of the result.
15369	*/
15370	dst_reg->s32_min_value = S32_MIN;
15371	dst_reg->s32_max_value = S32_MAX;
15372
15373	dst_reg->var_off = tnum_rshift(a: subreg, shift: umin_val);
15374	dst_reg->u32_min_value >>= umax_val;
15375	dst_reg->u32_max_value >>= umin_val;
15376
15377	__mark_reg64_unbounded(reg: dst_reg);
15378	__update_reg32_bounds(reg: dst_reg);
15379	}
15380
15381	static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
15382	struct bpf_reg_state *src_reg)
15383	{
15384	u64 umax_val = src_reg->umax_value;
15385	u64 umin_val = src_reg->umin_value;
15386
15387	/* BPF_RSH is an unsigned shift. If the value in dst_reg might
15388	* be negative, then either:
15389	* 1) src_reg might be zero, so the sign bit of the result is
15390	* unknown, so we lose our signed bounds
15391	* 2) it's known negative, thus the unsigned bounds capture the
15392	* signed bounds
15393	* 3) the signed bounds cross zero, so they tell us nothing
15394	* about the result
15395	* If the value in dst_reg is known nonnegative, then again the
15396	* unsigned bounds capture the signed bounds.
15397	* Thus, in all cases it suffices to blow away our signed bounds
15398	* and rely on inferring new ones from the unsigned bounds and
15399	* var_off of the result.
15400	*/
15401	dst_reg->smin_value = S64_MIN;
15402	dst_reg->smax_value = S64_MAX;
15403	dst_reg->var_off = tnum_rshift(a: dst_reg->var_off, shift: umin_val);
15404	dst_reg->umin_value >>= umax_val;
15405	dst_reg->umax_value >>= umin_val;
15406
15407	/* Its not easy to operate on alu32 bounds here because it depends
15408	* on bits being shifted in. Take easy way out and mark unbounded
15409	* so we can recalculate later from tnum.
15410	*/
15411	__mark_reg32_unbounded(reg: dst_reg);
15412	__update_reg_bounds(reg: dst_reg);
15413	}
15414
15415	static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg,
15416	struct bpf_reg_state *src_reg)
15417	{
15418	u64 umin_val = src_reg->u32_min_value;
15419
15420	/* Upon reaching here, src_known is true and
15421	* umax_val is equal to umin_val.
15422	*/
15423	dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val);
15424	dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val);
15425
15426	dst_reg->var_off = tnum_arshift(a: tnum_subreg(a: dst_reg->var_off), min_shift: umin_val, insn_bitness: 32);
15427
15428	/* blow away the dst_reg umin_value/umax_value and rely on
15429	* dst_reg var_off to refine the result.
15430	*/
15431	dst_reg->u32_min_value = 0;
15432	dst_reg->u32_max_value = U32_MAX;
15433
15434	__mark_reg64_unbounded(reg: dst_reg);
15435	__update_reg32_bounds(reg: dst_reg);
15436	}
15437
15438	static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
15439	struct bpf_reg_state *src_reg)
15440	{
15441	u64 umin_val = src_reg->umin_value;
15442
15443	/* Upon reaching here, src_known is true and umax_val is equal
15444	* to umin_val.
15445	*/
15446	dst_reg->smin_value >>= umin_val;
15447	dst_reg->smax_value >>= umin_val;
15448
15449	dst_reg->var_off = tnum_arshift(a: dst_reg->var_off, min_shift: umin_val, insn_bitness: 64);
15450
15451	/* blow away the dst_reg umin_value/umax_value and rely on
15452	* dst_reg var_off to refine the result.
15453	*/
15454	dst_reg->umin_value = 0;
15455	dst_reg->umax_value = U64_MAX;
15456
15457	/* Its not easy to operate on alu32 bounds here because it depends
15458	* on bits being shifted in from upper 32-bits. Take easy way out
15459	* and mark unbounded so we can recalculate later from tnum.
15460	*/
15461	__mark_reg32_unbounded(reg: dst_reg);
15462	__update_reg_bounds(reg: dst_reg);
15463	}
15464
15465	static bool is_safe_to_compute_dst_reg_range(struct bpf_insn *insn,
15466	const struct bpf_reg_state *src_reg)
15467	{
15468	bool src_is_const = false;
15469	u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
15470
15471	if (insn_bitness == 32) {
15472	if (tnum_subreg_is_const(a: src_reg->var_off)
15473	&& src_reg->s32_min_value == src_reg->s32_max_value
15474	&& src_reg->u32_min_value == src_reg->u32_max_value)
15475	src_is_const = true;
15476	} else {
15477	if (tnum_is_const(a: src_reg->var_off)
15478	&& src_reg->smin_value == src_reg->smax_value
15479	&& src_reg->umin_value == src_reg->umax_value)
15480	src_is_const = true;
15481	}
15482
15483	switch (BPF_OP(insn->code)) {
15484	case BPF_ADD:
15485	case BPF_SUB:
15486	case BPF_NEG:
15487	case BPF_AND:
15488	case BPF_XOR:
15489	case BPF_OR:
15490	case BPF_MUL:
15491	return true;
15492
15493	/* Shift operators range is only computable if shift dimension operand
15494	* is a constant. Shifts greater than 31 or 63 are undefined. This
15495	* includes shifts by a negative number.
15496	*/
15497	case BPF_LSH:
15498	case BPF_RSH:
15499	case BPF_ARSH:
15500	return (src_is_const && src_reg->umax_value < insn_bitness);
15501	default:
15502	return false;
15503	}
15504	}
15505
15506	/* WARNING: This function does calculations on 64-bit values, but the actual
15507	* execution may occur on 32-bit values. Therefore, things like bitshifts
15508	* need extra checks in the 32-bit case.
15509	*/
15510	static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
15511	struct bpf_insn *insn,
15512	struct bpf_reg_state *dst_reg,
15513	struct bpf_reg_state src_reg)
15514	{
15515	u8 opcode = BPF_OP(insn->code);
15516	bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
15517	int ret;
15518
15519	if (!is_safe_to_compute_dst_reg_range(insn, src_reg: &src_reg)) {
15520	__mark_reg_unknown(env, reg: dst_reg);
15521	return 0;
15522	}
15523
15524	if (sanitize_needed(opcode)) {
15525	ret = sanitize_val_alu(env, insn);
15526	if (ret < 0)
15527	return sanitize_err(env, insn, reason: ret, NULL, NULL);
15528	}
15529
15530	/* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
15531	* There are two classes of instructions: The first class we track both
15532	* alu32 and alu64 sign/unsigned bounds independently this provides the
15533	* greatest amount of precision when alu operations are mixed with jmp32
15534	* operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
15535	* and BPF_OR. This is possible because these ops have fairly easy to
15536	* understand and calculate behavior in both 32-bit and 64-bit alu ops.
15537	* See alu32 verifier tests for examples. The second class of
15538	* operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
15539	* with regards to tracking sign/unsigned bounds because the bits may
15540	* cross subreg boundaries in the alu64 case. When this happens we mark
15541	* the reg unbounded in the subreg bound space and use the resulting
15542	* tnum to calculate an approximation of the sign/unsigned bounds.
15543	*/
15544	switch (opcode) {
15545	case BPF_ADD:
15546	scalar32_min_max_add(dst_reg, src_reg: &src_reg);
15547	scalar_min_max_add(dst_reg, src_reg: &src_reg);
15548	dst_reg->var_off = tnum_add(a: dst_reg->var_off, b: src_reg.var_off);
15549	break;
15550	case BPF_SUB:
15551	scalar32_min_max_sub(dst_reg, src_reg: &src_reg);
15552	scalar_min_max_sub(dst_reg, src_reg: &src_reg);
15553	dst_reg->var_off = tnum_sub(a: dst_reg->var_off, b: src_reg.var_off);
15554	break;
15555	case BPF_NEG:
15556	env->fake_reg[0] = *dst_reg;
15557	__mark_reg_known(reg: dst_reg, imm: 0);
15558	scalar32_min_max_sub(dst_reg, src_reg: &env->fake_reg[0]);
15559	scalar_min_max_sub(dst_reg, src_reg: &env->fake_reg[0]);
15560	dst_reg->var_off = tnum_neg(a: env->fake_reg[0].var_off);
15561	break;
15562	case BPF_MUL:
15563	dst_reg->var_off = tnum_mul(a: dst_reg->var_off, b: src_reg.var_off);
15564	scalar32_min_max_mul(dst_reg, src_reg: &src_reg);
15565	scalar_min_max_mul(dst_reg, src_reg: &src_reg);
15566	break;
15567	case BPF_AND:
15568	dst_reg->var_off = tnum_and(a: dst_reg->var_off, b: src_reg.var_off);
15569	scalar32_min_max_and(dst_reg, src_reg: &src_reg);
15570	scalar_min_max_and(dst_reg, src_reg: &src_reg);
15571	break;
15572	case BPF_OR:
15573	dst_reg->var_off = tnum_or(a: dst_reg->var_off, b: src_reg.var_off);
15574	scalar32_min_max_or(dst_reg, src_reg: &src_reg);
15575	scalar_min_max_or(dst_reg, src_reg: &src_reg);
15576	break;
15577	case BPF_XOR:
15578	dst_reg->var_off = tnum_xor(a: dst_reg->var_off, b: src_reg.var_off);
15579	scalar32_min_max_xor(dst_reg, src_reg: &src_reg);
15580	scalar_min_max_xor(dst_reg, src_reg: &src_reg);
15581	break;
15582	case BPF_LSH:
15583	if (alu32)
15584	scalar32_min_max_lsh(dst_reg, src_reg: &src_reg);
15585	else
15586	scalar_min_max_lsh(dst_reg, src_reg: &src_reg);
15587	break;
15588	case BPF_RSH:
15589	if (alu32)
15590	scalar32_min_max_rsh(dst_reg, src_reg: &src_reg);
15591	else
15592	scalar_min_max_rsh(dst_reg, src_reg: &src_reg);
15593	break;
15594	case BPF_ARSH:
15595	if (alu32)
15596	scalar32_min_max_arsh(dst_reg, src_reg: &src_reg);
15597	else
15598	scalar_min_max_arsh(dst_reg, src_reg: &src_reg);
15599	break;
15600	default:
15601	break;
15602	}
15603
15604	/* ALU32 ops are zero extended into 64bit register */
15605	if (alu32)
15606	zext_32_to_64(reg: dst_reg);
15607	reg_bounds_sync(reg: dst_reg);
15608	return 0;
15609	}
15610
15611	/* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
15612	* and var_off.
15613	*/
15614	static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
15615	struct bpf_insn *insn)
15616	{
15617	struct bpf_verifier_state *vstate = env->cur_state;
15618	struct bpf_func_state *state = vstate->frame[vstate->curframe];
15619	struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
15620	struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
15621	bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
15622	u8 opcode = BPF_OP(insn->code);
15623	int err;
15624
15625	dst_reg = &regs[insn->dst_reg];
15626	src_reg = NULL;
15627
15628	if (dst_reg->type == PTR_TO_ARENA) {
15629	struct bpf_insn_aux_data *aux = cur_aux(env);
15630
15631	if (BPF_CLASS(insn->code) == BPF_ALU64)
15632	/*
15633	* 32-bit operations zero upper bits automatically.
15634	* 64-bit operations need to be converted to 32.
15635	*/
15636	aux->needs_zext = true;
15637
15638	/* Any arithmetic operations are allowed on arena pointers */
15639	return 0;
15640	}
15641
15642	if (dst_reg->type != SCALAR_VALUE)
15643	ptr_reg = dst_reg;
15644
15645	if (BPF_SRC(insn->code) == BPF_X) {
15646	src_reg = &regs[insn->src_reg];
15647	if (src_reg->type != SCALAR_VALUE) {
15648	if (dst_reg->type != SCALAR_VALUE) {
15649	/* Combining two pointers by any ALU op yields
15650	* an arbitrary scalar. Disallow all math except
15651	* pointer subtraction
15652	*/
15653	if (opcode == BPF_SUB && env->allow_ptr_leaks) {
15654	mark_reg_unknown(env, regs, regno: insn->dst_reg);
15655	return 0;
15656	}
15657	verbose(private_data: env, fmt: "R%d pointer %s pointer prohibited\n",
15658	insn->dst_reg,
15659	bpf_alu_string[opcode >> 4]);
15660	return -EACCES;
15661	} else {
15662	/* scalar += pointer
15663	* This is legal, but we have to reverse our
15664	* src/dest handling in computing the range
15665	*/
15666	err = mark_chain_precision(env, regno: insn->dst_reg);
15667	if (err)
15668	return err;
15669	return adjust_ptr_min_max_vals(env, insn,
15670	ptr_reg: src_reg, off_reg: dst_reg);
15671	}
15672	} else if (ptr_reg) {
15673	/* pointer += scalar */
15674	err = mark_chain_precision(env, regno: insn->src_reg);
15675	if (err)
15676	return err;
15677	return adjust_ptr_min_max_vals(env, insn,
15678	ptr_reg: dst_reg, off_reg: src_reg);
15679	} else if (dst_reg->precise) {
15680	/* if dst_reg is precise, src_reg should be precise as well */
15681	err = mark_chain_precision(env, regno: insn->src_reg);
15682	if (err)
15683	return err;
15684	}
15685	} else {
15686	/* Pretend the src is a reg with a known value, since we only
15687	* need to be able to read from this state.
15688	*/
15689	off_reg.type = SCALAR_VALUE;
15690	__mark_reg_known(reg: &off_reg, imm: insn->imm);
15691	src_reg = &off_reg;
15692	if (ptr_reg) /* pointer += K */
15693	return adjust_ptr_min_max_vals(env, insn,
15694	ptr_reg, off_reg: src_reg);
15695	}
15696
15697	/* Got here implies adding two SCALAR_VALUEs */
15698	if (WARN_ON_ONCE(ptr_reg)) {
15699	print_verifier_state(env, vstate, frameno: vstate->curframe, print_all: true);
15700	verbose(private_data: env, fmt: "verifier internal error: unexpected ptr_reg\n");
15701	return -EFAULT;
15702	}
15703	if (WARN_ON(!src_reg)) {
15704	print_verifier_state(env, vstate, frameno: vstate->curframe, print_all: true);
15705	verbose(private_data: env, fmt: "verifier internal error: no src_reg\n");
15706	return -EFAULT;
15707	}
15708	err = adjust_scalar_min_max_vals(env, insn, dst_reg, src_reg: *src_reg);
15709	if (err)
15710	return err;
15711	/*
15712	* Compilers can generate the code
15713	* r1 = r2
15714	* r1 += 0x1
15715	* if r2 < 1000 goto ...
15716	* use r1 in memory access
15717	* So for 64-bit alu remember constant delta between r2 and r1 and
15718	* update r1 after 'if' condition.
15719	*/
15720	if (env->bpf_capable &&
15721	BPF_OP(insn->code) == BPF_ADD && !alu32 &&
15722	dst_reg->id && is_reg_const(reg: src_reg, subreg32: false)) {
15723	u64 val = reg_const_value(reg: src_reg, subreg32: false);
15724
15725	if ((dst_reg->id & BPF_ADD_CONST) ||
15726	/* prevent overflow in sync_linked_regs() later */
15727	val > (u32)S32_MAX) {
15728	/*
15729	* If the register already went through rX += val
15730	* we cannot accumulate another val into rx->off.
15731	*/
15732	dst_reg->off = 0;
15733	dst_reg->id = 0;
15734	} else {
15735	dst_reg->id |= BPF_ADD_CONST;
15736	dst_reg->off = val;
15737	}
15738	} else {
15739	/*
15740	* Make sure ID is cleared otherwise dst_reg min/max could be
15741	* incorrectly propagated into other registers by sync_linked_regs()
15742	*/
15743	dst_reg->id = 0;
15744	}
15745	return 0;
15746	}
15747
15748	/* check validity of 32-bit and 64-bit arithmetic operations */
15749	static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
15750	{
15751	struct bpf_reg_state *regs = cur_regs(env);
15752	u8 opcode = BPF_OP(insn->code);
15753	int err;
15754
15755	if (opcode == BPF_END || opcode == BPF_NEG) {
15756	if (opcode == BPF_NEG) {
15757	if (BPF_SRC(insn->code) != BPF_K ||
15758	insn->src_reg != BPF_REG_0 ||
15759	insn->off != 0 || insn->imm != 0) {
15760	verbose(private_data: env, fmt: "BPF_NEG uses reserved fields\n");
15761	return -EINVAL;
15762	}
15763	} else {
15764	if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
15765	(insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
15766	(BPF_CLASS(insn->code) == BPF_ALU64 &&
15767	BPF_SRC(insn->code) != BPF_TO_LE)) {
15768	verbose(private_data: env, fmt: "BPF_END uses reserved fields\n");
15769	return -EINVAL;
15770	}
15771	}
15772
15773	/* check src operand */
15774	err = check_reg_arg(env, regno: insn->dst_reg, t: SRC_OP);
15775	if (err)
15776	return err;
15777
15778	if (is_pointer_value(env, regno: insn->dst_reg)) {
15779	verbose(private_data: env, fmt: "R%d pointer arithmetic prohibited\n",
15780	insn->dst_reg);
15781	return -EACCES;
15782	}
15783
15784	/* check dest operand */
15785	if (opcode == BPF_NEG &&
15786	regs[insn->dst_reg].type == SCALAR_VALUE) {
15787	err = check_reg_arg(env, regno: insn->dst_reg, t: DST_OP_NO_MARK);
15788	err = err ?: adjust_scalar_min_max_vals(env, insn,
15789	dst_reg: &regs[insn->dst_reg],
15790	src_reg: regs[insn->dst_reg]);
15791	} else {
15792	err = check_reg_arg(env, regno: insn->dst_reg, t: DST_OP);
15793	}
15794	if (err)
15795	return err;
15796
15797	} else if (opcode == BPF_MOV) {
15798
15799	if (BPF_SRC(insn->code) == BPF_X) {
15800	if (BPF_CLASS(insn->code) == BPF_ALU) {
15801	if ((insn->off != 0 && insn->off != 8 && insn->off != 16) ||
15802	insn->imm) {
15803	verbose(private_data: env, fmt: "BPF_MOV uses reserved fields\n");
15804	return -EINVAL;
15805	}
15806	} else if (insn->off == BPF_ADDR_SPACE_CAST) {
15807	if (insn->imm != 1 && insn->imm != 1u << 16) {
15808	verbose(private_data: env, fmt: "addr_space_cast insn can only convert between address space 1 and 0\n");
15809	return -EINVAL;
15810	}
15811	if (!env->prog->aux->arena) {
15812	verbose(private_data: env, fmt: "addr_space_cast insn can only be used in a program that has an associated arena\n");
15813	return -EINVAL;
15814	}
15815	} else {
15816	if ((insn->off != 0 && insn->off != 8 && insn->off != 16 &&
15817	insn->off != 32) || insn->imm) {
15818	verbose(private_data: env, fmt: "BPF_MOV uses reserved fields\n");
15819	return -EINVAL;
15820	}
15821	}
15822
15823	/* check src operand */
15824	err = check_reg_arg(env, regno: insn->src_reg, t: SRC_OP);
15825	if (err)
15826	return err;
15827	} else {
15828	if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
15829	verbose(private_data: env, fmt: "BPF_MOV uses reserved fields\n");
15830	return -EINVAL;
15831	}
15832	}
15833
15834	/* check dest operand, mark as required later */
15835	err = check_reg_arg(env, regno: insn->dst_reg, t: DST_OP_NO_MARK);
15836	if (err)
15837	return err;
15838
15839	if (BPF_SRC(insn->code) == BPF_X) {
15840	struct bpf_reg_state *src_reg = regs + insn->src_reg;
15841	struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
15842
15843	if (BPF_CLASS(insn->code) == BPF_ALU64) {
15844	if (insn->imm) {
15845	/* off == BPF_ADDR_SPACE_CAST */
15846	mark_reg_unknown(env, regs, regno: insn->dst_reg);
15847	if (insn->imm == 1) { /* cast from as(1) to as(0) */
15848	dst_reg->type = PTR_TO_ARENA;
15849	/* PTR_TO_ARENA is 32-bit */
15850	dst_reg->subreg_def = env->insn_idx + 1;
15851	}
15852	} else if (insn->off == 0) {
15853	/* case: R1 = R2
15854	* copy register state to dest reg
15855	*/
15856	assign_scalar_id_before_mov(env, src_reg);
15857	copy_register_state(dst: dst_reg, src: src_reg);
15858	dst_reg->subreg_def = DEF_NOT_SUBREG;
15859	} else {
15860	/* case: R1 = (s8, s16 s32)R2 */
15861	if (is_pointer_value(env, regno: insn->src_reg)) {
15862	verbose(private_data: env,
15863	fmt: "R%d sign-extension part of pointer\n",
15864	insn->src_reg);
15865	return -EACCES;
15866	} else if (src_reg->type == SCALAR_VALUE) {
15867	bool no_sext;
15868
15869	no_sext = src_reg->umax_value < (1ULL << (insn->off - 1));
15870	if (no_sext)
15871	assign_scalar_id_before_mov(env, src_reg);
15872	copy_register_state(dst: dst_reg, src: src_reg);
15873	if (!no_sext)
15874	dst_reg->id = 0;
15875	coerce_reg_to_size_sx(reg: dst_reg, size: insn->off >> 3);
15876	dst_reg->subreg_def = DEF_NOT_SUBREG;
15877	} else {
15878	mark_reg_unknown(env, regs, regno: insn->dst_reg);
15879	}
15880	}
15881	} else {
15882	/* R1 = (u32) R2 */
15883	if (is_pointer_value(env, regno: insn->src_reg)) {
15884	verbose(private_data: env,
15885	fmt: "R%d partial copy of pointer\n",
15886	insn->src_reg);
15887	return -EACCES;
15888	} else if (src_reg->type == SCALAR_VALUE) {
15889	if (insn->off == 0) {
15890	bool is_src_reg_u32 = get_reg_width(reg: src_reg) <= 32;
15891
15892	if (is_src_reg_u32)
15893	assign_scalar_id_before_mov(env, src_reg);
15894	copy_register_state(dst: dst_reg, src: src_reg);
15895	/* Make sure ID is cleared if src_reg is not in u32
15896	* range otherwise dst_reg min/max could be incorrectly
15897	* propagated into src_reg by sync_linked_regs()
15898	*/
15899	if (!is_src_reg_u32)
15900	dst_reg->id = 0;
15901	dst_reg->subreg_def = env->insn_idx + 1;
15902	} else {
15903	/* case: W1 = (s8, s16)W2 */
15904	bool no_sext = src_reg->umax_value < (1ULL << (insn->off - 1));
15905
15906	if (no_sext)
15907	assign_scalar_id_before_mov(env, src_reg);
15908	copy_register_state(dst: dst_reg, src: src_reg);
15909	if (!no_sext)
15910	dst_reg->id = 0;
15911	dst_reg->subreg_def = env->insn_idx + 1;
15912	coerce_subreg_to_size_sx(reg: dst_reg, size: insn->off >> 3);
15913	}
15914	} else {
15915	mark_reg_unknown(env, regs,
15916	regno: insn->dst_reg);
15917	}
15918	zext_32_to_64(reg: dst_reg);
15919	reg_bounds_sync(reg: dst_reg);
15920	}
15921	} else {
15922	/* case: R = imm
15923	* remember the value we stored into this reg
15924	*/
15925	/* clear any state __mark_reg_known doesn't set */
15926	mark_reg_unknown(env, regs, regno: insn->dst_reg);
15927	regs[insn->dst_reg].type = SCALAR_VALUE;
15928	if (BPF_CLASS(insn->code) == BPF_ALU64) {
15929	__mark_reg_known(reg: regs + insn->dst_reg,
15930	imm: insn->imm);
15931	} else {
15932	__mark_reg_known(reg: regs + insn->dst_reg,
15933	imm: (u32)insn->imm);
15934	}
15935	}
15936
15937	} else if (opcode > BPF_END) {
15938	verbose(private_data: env, fmt: "invalid BPF_ALU opcode %x\n", opcode);
15939	return -EINVAL;
15940
15941	} else { /* all other ALU ops: and, sub, xor, add, ... */
15942
15943	if (BPF_SRC(insn->code) == BPF_X) {
15944	if (insn->imm != 0 || (insn->off != 0 && insn->off != 1) ||
15945	(insn->off == 1 && opcode != BPF_MOD && opcode != BPF_DIV)) {
15946	verbose(private_data: env, fmt: "BPF_ALU uses reserved fields\n");
15947	return -EINVAL;
15948	}
15949	/* check src1 operand */
15950	err = check_reg_arg(env, regno: insn->src_reg, t: SRC_OP);
15951	if (err)
15952	return err;
15953	} else {
15954	if (insn->src_reg != BPF_REG_0 || (insn->off != 0 && insn->off != 1) ||
15955	(insn->off == 1 && opcode != BPF_MOD && opcode != BPF_DIV)) {
15956	verbose(private_data: env, fmt: "BPF_ALU uses reserved fields\n");
15957	return -EINVAL;
15958	}
15959	}
15960
15961	/* check src2 operand */
15962	err = check_reg_arg(env, regno: insn->dst_reg, t: SRC_OP);
15963	if (err)
15964	return err;
15965
15966	if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
15967	BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
15968	verbose(private_data: env, fmt: "div by zero\n");
15969	return -EINVAL;
15970	}
15971
15972	if ((opcode == BPF_LSH || opcode == BPF_RSH ||
15973	opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
15974	int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
15975
15976	if (insn->imm < 0 || insn->imm >= size) {
15977	verbose(private_data: env, fmt: "invalid shift %d\n", insn->imm);
15978	return -EINVAL;
15979	}
15980	}
15981
15982	/* check dest operand */
15983	err = check_reg_arg(env, regno: insn->dst_reg, t: DST_OP_NO_MARK);
15984	err = err ?: adjust_reg_min_max_vals(env, insn);
15985	if (err)
15986	return err;
15987	}
15988
15989	return reg_bounds_sanity_check(env, reg: &regs[insn->dst_reg], ctx: "alu");
15990	}
15991
15992	static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
15993	struct bpf_reg_state *dst_reg,
15994	enum bpf_reg_type type,
15995	bool range_right_open)
15996	{
15997	struct bpf_func_state *state;
15998	struct bpf_reg_state *reg;
15999	int new_range;
16000
16001	if (dst_reg->off < 0 ||
16002	(dst_reg->off == 0 && range_right_open))
16003	/* This doesn't give us any range */
16004	return;
16005
16006	if (dst_reg->umax_value > MAX_PACKET_OFF ||
16007	dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
16008	/* Risk of overflow. For instance, ptr + (1<<63) may be less
16009	* than pkt_end, but that's because it's also less than pkt.
16010	*/
16011	return;
16012
16013	new_range = dst_reg->off;
16014	if (range_right_open)
16015	new_range++;
16016
16017	/* Examples for register markings:
16018	*
16019	* pkt_data in dst register:
16020	*
16021	* r2 = r3;
16022	* r2 += 8;
16023	* if (r2 > pkt_end) goto <handle exception>
16024	* <access okay>
16025	*
16026	* r2 = r3;
16027	* r2 += 8;
16028	* if (r2 < pkt_end) goto <access okay>
16029	* <handle exception>
16030	*
16031	* Where:
16032	* r2 == dst_reg, pkt_end == src_reg
16033	* r2=pkt(id=n,off=8,r=0)
16034	* r3=pkt(id=n,off=0,r=0)
16035	*
16036	* pkt_data in src register:
16037	*
16038	* r2 = r3;
16039	* r2 += 8;
16040	* if (pkt_end >= r2) goto <access okay>
16041	* <handle exception>
16042	*
16043	* r2 = r3;
16044	* r2 += 8;
16045	* if (pkt_end <= r2) goto <handle exception>
16046	* <access okay>
16047	*
16048	* Where:
16049	* pkt_end == dst_reg, r2 == src_reg
16050	* r2=pkt(id=n,off=8,r=0)
16051	* r3=pkt(id=n,off=0,r=0)
16052	*
16053	* Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
16054	* or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
16055	* and [r3, r3 + 8-1) respectively is safe to access depending on
16056	* the check.
16057	*/
16058
16059	/* If our ids match, then we must have the same max_value. And we
16060	* don't care about the other reg's fixed offset, since if it's too big
16061	* the range won't allow anything.
16062	* dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
16063	*/
16064	bpf_for_each_reg_in_vstate(vstate, state, reg, ({
16065	if (reg->type == type && reg->id == dst_reg->id)
16066	/* keep the maximum range already checked */
16067	reg->range = max(reg->range, new_range);
16068	}));
16069	}
16070
16071	/*
16072	* <reg1> <op> <reg2>, currently assuming reg2 is a constant
16073	*/
16074	static int is_scalar_branch_taken(struct bpf_reg_state *reg1, struct bpf_reg_state *reg2,
16075	u8 opcode, bool is_jmp32)
16076	{
16077	struct tnum t1 = is_jmp32 ? tnum_subreg(a: reg1->var_off) : reg1->var_off;
16078	struct tnum t2 = is_jmp32 ? tnum_subreg(a: reg2->var_off) : reg2->var_off;
16079	u64 umin1 = is_jmp32 ? (u64)reg1->u32_min_value : reg1->umin_value;
16080	u64 umax1 = is_jmp32 ? (u64)reg1->u32_max_value : reg1->umax_value;
16081	s64 smin1 = is_jmp32 ? (s64)reg1->s32_min_value : reg1->smin_value;
16082	s64 smax1 = is_jmp32 ? (s64)reg1->s32_max_value : reg1->smax_value;
16083	u64 umin2 = is_jmp32 ? (u64)reg2->u32_min_value : reg2->umin_value;
16084	u64 umax2 = is_jmp32 ? (u64)reg2->u32_max_value : reg2->umax_value;
16085	s64 smin2 = is_jmp32 ? (s64)reg2->s32_min_value : reg2->smin_value;
16086	s64 smax2 = is_jmp32 ? (s64)reg2->s32_max_value : reg2->smax_value;
16087
16088	if (reg1 == reg2) {
16089	switch (opcode) {
16090	case BPF_JGE:
16091	case BPF_JLE:
16092	case BPF_JSGE:
16093	case BPF_JSLE:
16094	case BPF_JEQ:
16095	return 1;
16096	case BPF_JGT:
16097	case BPF_JLT:
16098	case BPF_JSGT:
16099	case BPF_JSLT:
16100	case BPF_JNE:
16101	return 0;
16102	case BPF_JSET:
16103	if (tnum_is_const(a: t1))
16104	return t1.value != 0;
16105	else
16106	return (smin1 <= 0 && smax1 >= 0) ? -1 : 1;
16107	default:
16108	return -1;
16109	}
16110	}
16111
16112	switch (opcode) {
16113	case BPF_JEQ:
16114	/* constants, umin/umax and smin/smax checks would be
16115	* redundant in this case because they all should match
16116	*/
16117	if (tnum_is_const(a: t1) && tnum_is_const(a: t2))
16118	return t1.value == t2.value;
16119	if (!tnum_overlap(a: t1, b: t2))
16120	return 0;
16121	/* non-overlapping ranges */
16122	if (umin1 > umax2 || umax1 < umin2)
16123	return 0;
16124	if (smin1 > smax2 || smax1 < smin2)
16125	return 0;
16126	if (!is_jmp32) {
16127	/* if 64-bit ranges are inconclusive, see if we can
16128	* utilize 32-bit subrange knowledge to eliminate
16129	* branches that can't be taken a priori
16130	*/
16131	if (reg1->u32_min_value > reg2->u32_max_value ||
16132	reg1->u32_max_value < reg2->u32_min_value)
16133	return 0;
16134	if (reg1->s32_min_value > reg2->s32_max_value ||
16135	reg1->s32_max_value < reg2->s32_min_value)
16136	return 0;
16137	}
16138	break;
16139	case BPF_JNE:
16140	/* constants, umin/umax and smin/smax checks would be
16141	* redundant in this case because they all should match
16142	*/
16143	if (tnum_is_const(a: t1) && tnum_is_const(a: t2))
16144	return t1.value != t2.value;
16145	if (!tnum_overlap(a: t1, b: t2))
16146	return 1;
16147	/* non-overlapping ranges */
16148	if (umin1 > umax2 || umax1 < umin2)
16149	return 1;
16150	if (smin1 > smax2 || smax1 < smin2)
16151	return 1;
16152	if (!is_jmp32) {
16153	/* if 64-bit ranges are inconclusive, see if we can
16154	* utilize 32-bit subrange knowledge to eliminate
16155	* branches that can't be taken a priori
16156	*/
16157	if (reg1->u32_min_value > reg2->u32_max_value ||
16158	reg1->u32_max_value < reg2->u32_min_value)
16159	return 1;
16160	if (reg1->s32_min_value > reg2->s32_max_value ||
16161	reg1->s32_max_value < reg2->s32_min_value)
16162	return 1;
16163	}
16164	break;
16165	case BPF_JSET:
16166	if (!is_reg_const(reg: reg2, subreg32: is_jmp32)) {
16167	swap(reg1, reg2);
16168	swap(t1, t2);
16169	}
16170	if (!is_reg_const(reg: reg2, subreg32: is_jmp32))
16171	return -1;
16172	if ((~t1.mask & t1.value) & t2.value)
16173	return 1;
16174	if (!((t1.mask | t1.value) & t2.value))
16175	return 0;
16176	break;
16177	case BPF_JGT:
16178	if (umin1 > umax2)
16179	return 1;
16180	else if (umax1 <= umin2)
16181	return 0;
16182	break;
16183	case BPF_JSGT:
16184	if (smin1 > smax2)
16185	return 1;
16186	else if (smax1 <= smin2)
16187	return 0;
16188	break;
16189	case BPF_JLT:
16190	if (umax1 < umin2)
16191	return 1;
16192	else if (umin1 >= umax2)
16193	return 0;
16194	break;
16195	case BPF_JSLT:
16196	if (smax1 < smin2)
16197	return 1;
16198	else if (smin1 >= smax2)
16199	return 0;
16200	break;
16201	case BPF_JGE:
16202	if (umin1 >= umax2)
16203	return 1;
16204	else if (umax1 < umin2)
16205	return 0;
16206	break;
16207	case BPF_JSGE:
16208	if (smin1 >= smax2)
16209	return 1;
16210	else if (smax1 < smin2)
16211	return 0;
16212	break;
16213	case BPF_JLE:
16214	if (umax1 <= umin2)
16215	return 1;
16216	else if (umin1 > umax2)
16217	return 0;
16218	break;
16219	case BPF_JSLE:
16220	if (smax1 <= smin2)
16221	return 1;
16222	else if (smin1 > smax2)
16223	return 0;
16224	break;
16225	}
16226
16227	return -1;
16228	}
16229
16230	static int flip_opcode(u32 opcode)
16231	{
16232	/* How can we transform "a <op> b" into "b <op> a"? */
16233	static const u8 opcode_flip[16] = {
16234	/* these stay the same */
16235	[BPF_JEQ >> 4] = BPF_JEQ,
16236	[BPF_JNE >> 4] = BPF_JNE,
16237	[BPF_JSET >> 4] = BPF_JSET,
16238	/* these swap "lesser" and "greater" (L and G in the opcodes) */
16239	[BPF_JGE >> 4] = BPF_JLE,
16240	[BPF_JGT >> 4] = BPF_JLT,
16241	[BPF_JLE >> 4] = BPF_JGE,
16242	[BPF_JLT >> 4] = BPF_JGT,
16243	[BPF_JSGE >> 4] = BPF_JSLE,
16244	[BPF_JSGT >> 4] = BPF_JSLT,
16245	[BPF_JSLE >> 4] = BPF_JSGE,
16246	[BPF_JSLT >> 4] = BPF_JSGT
16247	};
16248	return opcode_flip[opcode >> 4];
16249	}
16250
16251	static int is_pkt_ptr_branch_taken(struct bpf_reg_state *dst_reg,
16252	struct bpf_reg_state *src_reg,
16253	u8 opcode)
16254	{
16255	struct bpf_reg_state *pkt;
16256
16257	if (src_reg->type == PTR_TO_PACKET_END) {
16258	pkt = dst_reg;
16259	} else if (dst_reg->type == PTR_TO_PACKET_END) {
16260	pkt = src_reg;
16261	opcode = flip_opcode(opcode);
16262	} else {
16263	return -1;
16264	}
16265
16266	if (pkt->range >= 0)
16267	return -1;
16268
16269	switch (opcode) {
16270	case BPF_JLE:
16271	/* pkt <= pkt_end */
16272	fallthrough;
16273	case BPF_JGT:
16274	/* pkt > pkt_end */
16275	if (pkt->range == BEYOND_PKT_END)
16276	/* pkt has at last one extra byte beyond pkt_end */
16277	return opcode == BPF_JGT;
16278	break;
16279	case BPF_JLT:
16280	/* pkt < pkt_end */
16281	fallthrough;
16282	case BPF_JGE:
16283	/* pkt >= pkt_end */
16284	if (pkt->range == BEYOND_PKT_END || pkt->range == AT_PKT_END)
16285	return opcode == BPF_JGE;
16286	break;
16287	}
16288	return -1;
16289	}
16290
16291	/* compute branch direction of the expression "if (<reg1> opcode <reg2>) goto target;"
16292	* and return:
16293	* 1 - branch will be taken and "goto target" will be executed
16294	* 0 - branch will not be taken and fall-through to next insn
16295	* -1 - unknown. Example: "if (reg1 < 5)" is unknown when register value
16296	* range [0,10]
16297	*/
16298	static int is_branch_taken(struct bpf_reg_state *reg1, struct bpf_reg_state *reg2,
16299	u8 opcode, bool is_jmp32)
16300	{
16301	if (reg_is_pkt_pointer_any(reg: reg1) && reg_is_pkt_pointer_any(reg: reg2) && !is_jmp32)
16302	return is_pkt_ptr_branch_taken(dst_reg: reg1, src_reg: reg2, opcode);
16303
16304	if (__is_pointer_value(allow_ptr_leaks: false, reg: reg1) || __is_pointer_value(allow_ptr_leaks: false, reg: reg2)) {
16305	u64 val;
16306
16307	/* arrange that reg2 is a scalar, and reg1 is a pointer */
16308	if (!is_reg_const(reg: reg2, subreg32: is_jmp32)) {
16309	opcode = flip_opcode(opcode);
16310	swap(reg1, reg2);
16311	}
16312	/* and ensure that reg2 is a constant */
16313	if (!is_reg_const(reg: reg2, subreg32: is_jmp32))
16314	return -1;
16315
16316	if (!reg_not_null(reg: reg1))
16317	return -1;
16318
16319	/* If pointer is valid tests against zero will fail so we can
16320	* use this to direct branch taken.
16321	*/
16322	val = reg_const_value(reg: reg2, subreg32: is_jmp32);
16323	if (val != 0)
16324	return -1;
16325
16326	switch (opcode) {
16327	case BPF_JEQ:
16328	return 0;
16329	case BPF_JNE:
16330	return 1;
16331	default:
16332	return -1;
16333	}
16334	}
16335
16336	/* now deal with two scalars, but not necessarily constants */
16337	return is_scalar_branch_taken(reg1, reg2, opcode, is_jmp32);
16338	}
16339
16340	/* Opcode that corresponds to a *false* branch condition.
16341	* E.g., if r1 < r2, then reverse (false) condition is r1 >= r2
16342	*/
16343	static u8 rev_opcode(u8 opcode)
16344	{
16345	switch (opcode) {
16346	case BPF_JEQ: return BPF_JNE;
16347	case BPF_JNE: return BPF_JEQ;
16348	/* JSET doesn't have it's reverse opcode in BPF, so add
16349	* BPF_X flag to denote the reverse of that operation
16350	*/
16351	case BPF_JSET: return BPF_JSET | BPF_X;
16352	case BPF_JSET | BPF_X: return BPF_JSET;
16353	case BPF_JGE: return BPF_JLT;
16354	case BPF_JGT: return BPF_JLE;
16355	case BPF_JLE: return BPF_JGT;
16356	case BPF_JLT: return BPF_JGE;
16357	case BPF_JSGE: return BPF_JSLT;
16358	case BPF_JSGT: return BPF_JSLE;
16359	case BPF_JSLE: return BPF_JSGT;
16360	case BPF_JSLT: return BPF_JSGE;
16361	default: return 0;
16362	}
16363	}
16364
16365	/* Refine range knowledge for <reg1> <op> <reg>2 conditional operation. */
16366	static void regs_refine_cond_op(struct bpf_reg_state *reg1, struct bpf_reg_state *reg2,
16367	u8 opcode, bool is_jmp32)
16368	{
16369	struct tnum t;
16370	u64 val;
16371
16372	/* In case of GE/GT/SGE/JST, reuse LE/LT/SLE/SLT logic from below */
16373	switch (opcode) {
16374	case BPF_JGE:
16375	case BPF_JGT:
16376	case BPF_JSGE:
16377	case BPF_JSGT:
16378	opcode = flip_opcode(opcode);
16379	swap(reg1, reg2);
16380	break;
16381	default:
16382	break;
16383	}
16384
16385	switch (opcode) {
16386	case BPF_JEQ:
16387	if (is_jmp32) {
16388	reg1->u32_min_value = max(reg1->u32_min_value, reg2->u32_min_value);
16389	reg1->u32_max_value = min(reg1->u32_max_value, reg2->u32_max_value);
16390	reg1->s32_min_value = max(reg1->s32_min_value, reg2->s32_min_value);
16391	reg1->s32_max_value = min(reg1->s32_max_value, reg2->s32_max_value);
16392	reg2->u32_min_value = reg1->u32_min_value;
16393	reg2->u32_max_value = reg1->u32_max_value;
16394	reg2->s32_min_value = reg1->s32_min_value;
16395	reg2->s32_max_value = reg1->s32_max_value;
16396
16397	t = tnum_intersect(a: tnum_subreg(a: reg1->var_off), b: tnum_subreg(a: reg2->var_off));
16398	reg1->var_off = tnum_with_subreg(reg: reg1->var_off, subreg: t);
16399	reg2->var_off = tnum_with_subreg(reg: reg2->var_off, subreg: t);
16400	} else {
16401	reg1->umin_value = max(reg1->umin_value, reg2->umin_value);
16402	reg1->umax_value = min(reg1->umax_value, reg2->umax_value);
16403	reg1->smin_value = max(reg1->smin_value, reg2->smin_value);
16404	reg1->smax_value = min(reg1->smax_value, reg2->smax_value);
16405	reg2->umin_value = reg1->umin_value;
16406	reg2->umax_value = reg1->umax_value;
16407	reg2->smin_value = reg1->smin_value;
16408	reg2->smax_value = reg1->smax_value;
16409
16410	reg1->var_off = tnum_intersect(a: reg1->var_off, b: reg2->var_off);
16411	reg2->var_off = reg1->var_off;
16412	}
16413	break;
16414	case BPF_JNE:
16415	if (!is_reg_const(reg: reg2, subreg32: is_jmp32))
16416	swap(reg1, reg2);
16417	if (!is_reg_const(reg: reg2, subreg32: is_jmp32))
16418	break;
16419
16420	/* try to recompute the bound of reg1 if reg2 is a const and
16421	* is exactly the edge of reg1.
16422	*/
16423	val = reg_const_value(reg: reg2, subreg32: is_jmp32);
16424	if (is_jmp32) {
16425	/* u32_min_value is not equal to 0xffffffff at this point,
16426	* because otherwise u32_max_value is 0xffffffff as well,
16427	* in such a case both reg1 and reg2 would be constants,
16428	* jump would be predicted and reg_set_min_max() won't
16429	* be called.
16430	*
16431	* Same reasoning works for all {u,s}{min,max}{32,64} cases
16432	* below.
16433	*/
16434	if (reg1->u32_min_value == (u32)val)
16435	reg1->u32_min_value++;
16436	if (reg1->u32_max_value == (u32)val)
16437	reg1->u32_max_value--;
16438	if (reg1->s32_min_value == (s32)val)
16439	reg1->s32_min_value++;
16440	if (reg1->s32_max_value == (s32)val)
16441	reg1->s32_max_value--;
16442	} else {
16443	if (reg1->umin_value == (u64)val)
16444	reg1->umin_value++;
16445	if (reg1->umax_value == (u64)val)
16446	reg1->umax_value--;
16447	if (reg1->smin_value == (s64)val)
16448	reg1->smin_value++;
16449	if (reg1->smax_value == (s64)val)
16450	reg1->smax_value--;
16451	}
16452	break;
16453	case BPF_JSET:
16454	if (!is_reg_const(reg: reg2, subreg32: is_jmp32))
16455	swap(reg1, reg2);
16456	if (!is_reg_const(reg: reg2, subreg32: is_jmp32))
16457	break;
16458	val = reg_const_value(reg: reg2, subreg32: is_jmp32);
16459	/* BPF_JSET (i.e., TRUE branch, *not* BPF_JSET | BPF_X)
16460	* requires single bit to learn something useful. E.g., if we
16461	* know that `r1 & 0x3` is true, then which bits (0, 1, or both)
16462	* are actually set? We can learn something definite only if
16463	* it's a single-bit value to begin with.
16464	*
16465	* BPF_JSET | BPF_X (i.e., negation of BPF_JSET) doesn't have
16466	* this restriction. I.e., !(r1 & 0x3) means neither bit 0 nor
16467	* bit 1 is set, which we can readily use in adjustments.
16468	*/
16469	if (!is_power_of_2(n: val))
16470	break;
16471	if (is_jmp32) {
16472	t = tnum_or(a: tnum_subreg(a: reg1->var_off), b: tnum_const(value: val));
16473	reg1->var_off = tnum_with_subreg(reg: reg1->var_off, subreg: t);
16474	} else {
16475	reg1->var_off = tnum_or(a: reg1->var_off, b: tnum_const(value: val));
16476	}
16477	break;
16478	case BPF_JSET | BPF_X: /* reverse of BPF_JSET, see rev_opcode() */
16479	if (!is_reg_const(reg: reg2, subreg32: is_jmp32))
16480	swap(reg1, reg2);
16481	if (!is_reg_const(reg: reg2, subreg32: is_jmp32))
16482	break;
16483	val = reg_const_value(reg: reg2, subreg32: is_jmp32);
16484	/* Forget the ranges before narrowing tnums, to avoid invariant
16485	* violations if we're on a dead branch.
16486	*/
16487	__mark_reg_unbounded(reg: reg1);
16488	if (is_jmp32) {
16489	t = tnum_and(a: tnum_subreg(a: reg1->var_off), b: tnum_const(value: ~val));
16490	reg1->var_off = tnum_with_subreg(reg: reg1->var_off, subreg: t);
16491	} else {
16492	reg1->var_off = tnum_and(a: reg1->var_off, b: tnum_const(value: ~val));
16493	}
16494	break;
16495	case BPF_JLE:
16496	if (is_jmp32) {
16497	reg1->u32_max_value = min(reg1->u32_max_value, reg2->u32_max_value);
16498	reg2->u32_min_value = max(reg1->u32_min_value, reg2->u32_min_value);
16499	} else {
16500	reg1->umax_value = min(reg1->umax_value, reg2->umax_value);
16501	reg2->umin_value = max(reg1->umin_value, reg2->umin_value);
16502	}
16503	break;
16504	case BPF_JLT:
16505	if (is_jmp32) {
16506	reg1->u32_max_value = min(reg1->u32_max_value, reg2->u32_max_value - 1);
16507	reg2->u32_min_value = max(reg1->u32_min_value + 1, reg2->u32_min_value);
16508	} else {
16509	reg1->umax_value = min(reg1->umax_value, reg2->umax_value - 1);
16510	reg2->umin_value = max(reg1->umin_value + 1, reg2->umin_value);
16511	}
16512	break;
16513	case BPF_JSLE:
16514	if (is_jmp32) {
16515	reg1->s32_max_value = min(reg1->s32_max_value, reg2->s32_max_value);
16516	reg2->s32_min_value = max(reg1->s32_min_value, reg2->s32_min_value);
16517	} else {
16518	reg1->smax_value = min(reg1->smax_value, reg2->smax_value);
16519	reg2->smin_value = max(reg1->smin_value, reg2->smin_value);
16520	}
16521	break;
16522	case BPF_JSLT:
16523	if (is_jmp32) {
16524	reg1->s32_max_value = min(reg1->s32_max_value, reg2->s32_max_value - 1);
16525	reg2->s32_min_value = max(reg1->s32_min_value + 1, reg2->s32_min_value);
16526	} else {
16527	reg1->smax_value = min(reg1->smax_value, reg2->smax_value - 1);
16528	reg2->smin_value = max(reg1->smin_value + 1, reg2->smin_value);
16529	}
16530	break;
16531	default:
16532	return;
16533	}
16534	}
16535
16536	/* Adjusts the register min/max values in the case that the dst_reg and
16537	* src_reg are both SCALAR_VALUE registers (or we are simply doing a BPF_K
16538	* check, in which case we have a fake SCALAR_VALUE representing insn->imm).
16539	* Technically we can do similar adjustments for pointers to the same object,
16540	* but we don't support that right now.
16541	*/
16542	static int reg_set_min_max(struct bpf_verifier_env *env,
16543	struct bpf_reg_state *true_reg1,
16544	struct bpf_reg_state *true_reg2,
16545	struct bpf_reg_state *false_reg1,
16546	struct bpf_reg_state *false_reg2,
16547	u8 opcode, bool is_jmp32)
16548	{
16549	int err;
16550
16551	/* If either register is a pointer, we can't learn anything about its
16552	* variable offset from the compare (unless they were a pointer into
16553	* the same object, but we don't bother with that).
16554	*/
16555	if (false_reg1->type != SCALAR_VALUE || false_reg2->type != SCALAR_VALUE)
16556	return 0;
16557
16558	/* We compute branch direction for same SCALAR_VALUE registers in
16559	* is_scalar_branch_taken(). For unknown branch directions (e.g., BPF_JSET)
16560	* on the same registers, we don't need to adjust the min/max values.
16561	*/
16562	if (false_reg1 == false_reg2)
16563	return 0;
16564
16565	/* fallthrough (FALSE) branch */
16566	regs_refine_cond_op(reg1: false_reg1, reg2: false_reg2, opcode: rev_opcode(opcode), is_jmp32);
16567	reg_bounds_sync(reg: false_reg1);
16568	reg_bounds_sync(reg: false_reg2);
16569
16570	/* jump (TRUE) branch */
16571	regs_refine_cond_op(reg1: true_reg1, reg2: true_reg2, opcode, is_jmp32);
16572	reg_bounds_sync(reg: true_reg1);
16573	reg_bounds_sync(reg: true_reg2);
16574
16575	err = reg_bounds_sanity_check(env, reg: true_reg1, ctx: "true_reg1");
16576	err = err ?: reg_bounds_sanity_check(env, reg: true_reg2, ctx: "true_reg2");
16577	err = err ?: reg_bounds_sanity_check(env, reg: false_reg1, ctx: "false_reg1");
16578	err = err ?: reg_bounds_sanity_check(env, reg: false_reg2, ctx: "false_reg2");
16579	return err;
16580	}
16581
16582	static void mark_ptr_or_null_reg(struct bpf_func_state *state,
16583	struct bpf_reg_state *reg, u32 id,
16584	bool is_null)
16585	{
16586	if (type_may_be_null(type: reg->type) && reg->id == id &&
16587	(is_rcu_reg(reg) || !WARN_ON_ONCE(!reg->id))) {
16588	/* Old offset (both fixed and variable parts) should have been
16589	* known-zero, because we don't allow pointer arithmetic on
16590	* pointers that might be NULL. If we see this happening, don't
16591	* convert the register.
16592	*
16593	* But in some cases, some helpers that return local kptrs
16594	* advance offset for the returned pointer. In those cases, it
16595	* is fine to expect to see reg->off.
16596	*/
16597	if (WARN_ON_ONCE(reg->smin_value || reg->smax_value || !tnum_equals_const(reg->var_off, 0)))
16598	return;
16599	if (!(type_is_ptr_alloc_obj(type: reg->type) || type_is_non_owning_ref(type: reg->type)) &&
16600	WARN_ON_ONCE(reg->off))
16601	return;
16602
16603	if (is_null) {
16604	reg->type = SCALAR_VALUE;
16605	/* We don't need id and ref_obj_id from this point
16606	* onwards anymore, thus we should better reset it,
16607	* so that state pruning has chances to take effect.
16608	*/
16609	reg->id = 0;
16610	reg->ref_obj_id = 0;
16611
16612	return;
16613	}
16614
16615	mark_ptr_not_null_reg(reg);
16616
16617	if (!reg_may_point_to_spin_lock(reg)) {
16618	/* For not-NULL ptr, reg->ref_obj_id will be reset
16619	* in release_reference().
16620	*
16621	* reg->id is still used by spin_lock ptr. Other
16622	* than spin_lock ptr type, reg->id can be reset.
16623	*/
16624	reg->id = 0;
16625	}
16626	}
16627	}
16628
16629	/* The logic is similar to find_good_pkt_pointers(), both could eventually
16630	* be folded together at some point.
16631	*/
16632	static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
16633	bool is_null)
16634	{
16635	struct bpf_func_state *state = vstate->frame[vstate->curframe];
16636	struct bpf_reg_state *regs = state->regs, *reg;
16637	u32 ref_obj_id = regs[regno].ref_obj_id;
16638	u32 id = regs[regno].id;
16639
16640	if (ref_obj_id && ref_obj_id == id && is_null)
16641	/* regs[regno] is in the " == NULL" branch.
16642	* No one could have freed the reference state before
16643	* doing the NULL check.
16644	*/
16645	WARN_ON_ONCE(release_reference_nomark(vstate, id));
16646
16647	bpf_for_each_reg_in_vstate(vstate, state, reg, ({
16648	mark_ptr_or_null_reg(state, reg, id, is_null);
16649	}));
16650	}
16651
16652	static bool try_match_pkt_pointers(const struct bpf_insn *insn,
16653	struct bpf_reg_state *dst_reg,
16654	struct bpf_reg_state *src_reg,
16655	struct bpf_verifier_state *this_branch,
16656	struct bpf_verifier_state *other_branch)
16657	{
16658	if (BPF_SRC(insn->code) != BPF_X)
16659	return false;
16660
16661	/* Pointers are always 64-bit. */
16662	if (BPF_CLASS(insn->code) == BPF_JMP32)
16663	return false;
16664
16665	switch (BPF_OP(insn->code)) {
16666	case BPF_JGT:
16667	if ((dst_reg->type == PTR_TO_PACKET &&
16668	src_reg->type == PTR_TO_PACKET_END) ||
16669	(dst_reg->type == PTR_TO_PACKET_META &&
16670	reg_is_init_pkt_pointer(reg: src_reg, which: PTR_TO_PACKET))) {
16671	/* pkt_data' > pkt_end, pkt_meta' > pkt_data */
16672	find_good_pkt_pointers(vstate: this_branch, dst_reg,
16673	type: dst_reg->type, range_right_open: false);
16674	mark_pkt_end(vstate: other_branch, regn: insn->dst_reg, range_open: true);
16675	} else if ((dst_reg->type == PTR_TO_PACKET_END &&
16676	src_reg->type == PTR_TO_PACKET) ||
16677	(reg_is_init_pkt_pointer(reg: dst_reg, which: PTR_TO_PACKET) &&
16678	src_reg->type == PTR_TO_PACKET_META)) {
16679	/* pkt_end > pkt_data', pkt_data > pkt_meta' */
16680	find_good_pkt_pointers(vstate: other_branch, dst_reg: src_reg,
16681	type: src_reg->type, range_right_open: true);
16682	mark_pkt_end(vstate: this_branch, regn: insn->src_reg, range_open: false);
16683	} else {
16684	return false;
16685	}
16686	break;
16687	case BPF_JLT:
16688	if ((dst_reg->type == PTR_TO_PACKET &&
16689	src_reg->type == PTR_TO_PACKET_END) ||
16690	(dst_reg->type == PTR_TO_PACKET_META &&
16691	reg_is_init_pkt_pointer(reg: src_reg, which: PTR_TO_PACKET))) {
16692	/* pkt_data' < pkt_end, pkt_meta' < pkt_data */
16693	find_good_pkt_pointers(vstate: other_branch, dst_reg,
16694	type: dst_reg->type, range_right_open: true);
16695	mark_pkt_end(vstate: this_branch, regn: insn->dst_reg, range_open: false);
16696	} else if ((dst_reg->type == PTR_TO_PACKET_END &&
16697	src_reg->type == PTR_TO_PACKET) ||
16698	(reg_is_init_pkt_pointer(reg: dst_reg, which: PTR_TO_PACKET) &&
16699	src_reg->type == PTR_TO_PACKET_META)) {
16700	/* pkt_end < pkt_data', pkt_data > pkt_meta' */
16701	find_good_pkt_pointers(vstate: this_branch, dst_reg: src_reg,
16702	type: src_reg->type, range_right_open: false);
16703	mark_pkt_end(vstate: other_branch, regn: insn->src_reg, range_open: true);
16704	} else {
16705	return false;
16706	}
16707	break;
16708	case BPF_JGE:
16709	if ((dst_reg->type == PTR_TO_PACKET &&
16710	src_reg->type == PTR_TO_PACKET_END) ||
16711	(dst_reg->type == PTR_TO_PACKET_META &&
16712	reg_is_init_pkt_pointer(reg: src_reg, which: PTR_TO_PACKET))) {
16713	/* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
16714	find_good_pkt_pointers(vstate: this_branch, dst_reg,
16715	type: dst_reg->type, range_right_open: true);
16716	mark_pkt_end(vstate: other_branch, regn: insn->dst_reg, range_open: false);
16717	} else if ((dst_reg->type == PTR_TO_PACKET_END &&
16718	src_reg->type == PTR_TO_PACKET) ||
16719	(reg_is_init_pkt_pointer(reg: dst_reg, which: PTR_TO_PACKET) &&
16720	src_reg->type == PTR_TO_PACKET_META)) {
16721	/* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
16722	find_good_pkt_pointers(vstate: other_branch, dst_reg: src_reg,
16723	type: src_reg->type, range_right_open: false);
16724	mark_pkt_end(vstate: this_branch, regn: insn->src_reg, range_open: true);
16725	} else {
16726	return false;
16727	}
16728	break;
16729	case BPF_JLE:
16730	if ((dst_reg->type == PTR_TO_PACKET &&
16731	src_reg->type == PTR_TO_PACKET_END) ||
16732	(dst_reg->type == PTR_TO_PACKET_META &&
16733	reg_is_init_pkt_pointer(reg: src_reg, which: PTR_TO_PACKET))) {
16734	/* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
16735	find_good_pkt_pointers(vstate: other_branch, dst_reg,
16736	type: dst_reg->type, range_right_open: false);
16737	mark_pkt_end(vstate: this_branch, regn: insn->dst_reg, range_open: true);
16738	} else if ((dst_reg->type == PTR_TO_PACKET_END &&
16739	src_reg->type == PTR_TO_PACKET) ||
16740	(reg_is_init_pkt_pointer(reg: dst_reg, which: PTR_TO_PACKET) &&
16741	src_reg->type == PTR_TO_PACKET_META)) {
16742	/* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
16743	find_good_pkt_pointers(vstate: this_branch, dst_reg: src_reg,
16744	type: src_reg->type, range_right_open: true);
16745	mark_pkt_end(vstate: other_branch, regn: insn->src_reg, range_open: false);
16746	} else {
16747	return false;
16748	}
16749	break;
16750	default:
16751	return false;
16752	}
16753
16754	return true;
16755	}
16756
16757	static void __collect_linked_regs(struct linked_regs *reg_set, struct bpf_reg_state *reg,
16758	u32 id, u32 frameno, u32 spi_or_reg, bool is_reg)
16759	{
16760	struct linked_reg *e;
16761
16762	if (reg->type != SCALAR_VALUE || (reg->id & ~BPF_ADD_CONST) != id)
16763	return;
16764
16765	e = linked_regs_push(s: reg_set);
16766	if (e) {
16767	e->frameno = frameno;
16768	e->is_reg = is_reg;
16769	e->regno = spi_or_reg;
16770	} else {
16771	reg->id = 0;
16772	}
16773	}
16774
16775	/* For all R being scalar registers or spilled scalar registers
16776	* in verifier state, save R in linked_regs if R->id == id.
16777	* If there are too many Rs sharing same id, reset id for leftover Rs.
16778	*/
16779	static void collect_linked_regs(struct bpf_verifier_state *vstate, u32 id,
16780	struct linked_regs *linked_regs)
16781	{
16782	struct bpf_func_state *func;
16783	struct bpf_reg_state *reg;
16784	int i, j;
16785
16786	id = id & ~BPF_ADD_CONST;
16787	for (i = vstate->curframe; i >= 0; i--) {
16788	func = vstate->frame[i];
16789	for (j = 0; j < BPF_REG_FP; j++) {
16790	reg = &func->regs[j];
16791	__collect_linked_regs(reg_set: linked_regs, reg, id, frameno: i, spi_or_reg: j, is_reg: true);
16792	}
16793	for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
16794	if (!is_spilled_reg(stack: &func->stack[j]))
16795	continue;
16796	reg = &func->stack[j].spilled_ptr;
16797	__collect_linked_regs(reg_set: linked_regs, reg, id, frameno: i, spi_or_reg: j, is_reg: false);
16798	}
16799	}
16800	}
16801
16802	/* For all R in linked_regs, copy known_reg range into R
16803	* if R->id == known_reg->id.
16804	*/
16805	static void sync_linked_regs(struct bpf_verifier_state *vstate, struct bpf_reg_state *known_reg,
16806	struct linked_regs *linked_regs)
16807	{
16808	struct bpf_reg_state fake_reg;
16809	struct bpf_reg_state *reg;
16810	struct linked_reg *e;
16811	int i;
16812
16813	for (i = 0; i < linked_regs->cnt; ++i) {
16814	e = &linked_regs->entries[i];
16815	reg = e->is_reg ? &vstate->frame[e->frameno]->regs[e->regno]
16816	: &vstate->frame[e->frameno]->stack[e->spi].spilled_ptr;
16817	if (reg->type != SCALAR_VALUE || reg == known_reg)
16818	continue;
16819	if ((reg->id & ~BPF_ADD_CONST) != (known_reg->id & ~BPF_ADD_CONST))
16820	continue;
16821	if ((!(reg->id & BPF_ADD_CONST) && !(known_reg->id & BPF_ADD_CONST)) ||
16822	reg->off == known_reg->off) {
16823	s32 saved_subreg_def = reg->subreg_def;
16824
16825	copy_register_state(dst: reg, src: known_reg);
16826	reg->subreg_def = saved_subreg_def;
16827	} else {
16828	s32 saved_subreg_def = reg->subreg_def;
16829	s32 saved_off = reg->off;
16830
16831	fake_reg.type = SCALAR_VALUE;
16832	__mark_reg_known(reg: &fake_reg, imm: (s32)reg->off - (s32)known_reg->off);
16833
16834	/* reg = known_reg; reg += delta */
16835	copy_register_state(dst: reg, src: known_reg);
16836	/*
16837	* Must preserve off, id and add_const flag,
16838	* otherwise another sync_linked_regs() will be incorrect.
16839	*/
16840	reg->off = saved_off;
16841	reg->subreg_def = saved_subreg_def;
16842
16843	scalar32_min_max_add(dst_reg: reg, src_reg: &fake_reg);
16844	scalar_min_max_add(dst_reg: reg, src_reg: &fake_reg);
16845	reg->var_off = tnum_add(a: reg->var_off, b: fake_reg.var_off);
16846	}
16847	}
16848	}
16849
16850	static int check_cond_jmp_op(struct bpf_verifier_env *env,
16851	struct bpf_insn *insn, int *insn_idx)
16852	{
16853	struct bpf_verifier_state *this_branch = env->cur_state;
16854	struct bpf_verifier_state *other_branch;
16855	struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
16856	struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
16857	struct bpf_reg_state *eq_branch_regs;
16858	struct linked_regs linked_regs = {};
16859	u8 opcode = BPF_OP(insn->code);
16860	int insn_flags = 0;
16861	bool is_jmp32;
16862	int pred = -1;
16863	int err;
16864
16865	/* Only conditional jumps are expected to reach here. */
16866	if (opcode == BPF_JA || opcode > BPF_JCOND) {
16867	verbose(private_data: env, fmt: "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
16868	return -EINVAL;
16869	}
16870
16871	if (opcode == BPF_JCOND) {
16872	struct bpf_verifier_state *cur_st = env->cur_state, *queued_st, *prev_st;
16873	int idx = *insn_idx;
16874
16875	if (insn->code != (BPF_JMP | BPF_JCOND) ||
16876	insn->src_reg != BPF_MAY_GOTO ||
16877	insn->dst_reg || insn->imm) {
16878	verbose(private_data: env, fmt: "invalid may_goto imm %d\n", insn->imm);
16879	return -EINVAL;
16880	}
16881	prev_st = find_prev_entry(env, cur: cur_st->parent, insn_idx: idx);
16882
16883	/* branch out 'fallthrough' insn as a new state to explore */
16884	queued_st = push_stack(env, insn_idx: idx + 1, prev_insn_idx: idx, speculative: false);
16885	if (IS_ERR(ptr: queued_st))
16886	return PTR_ERR(ptr: queued_st);
16887
16888	queued_st->may_goto_depth++;
16889	if (prev_st)
16890	widen_imprecise_scalars(env, old: prev_st, cur: queued_st);
16891	*insn_idx += insn->off;
16892	return 0;
16893	}
16894
16895	/* check src2 operand */
16896	err = check_reg_arg(env, regno: insn->dst_reg, t: SRC_OP);
16897	if (err)
16898	return err;
16899
16900	dst_reg = &regs[insn->dst_reg];
16901	if (BPF_SRC(insn->code) == BPF_X) {
16902	if (insn->imm != 0) {
16903	verbose(private_data: env, fmt: "BPF_JMP/JMP32 uses reserved fields\n");
16904	return -EINVAL;
16905	}
16906
16907	/* check src1 operand */
16908	err = check_reg_arg(env, regno: insn->src_reg, t: SRC_OP);
16909	if (err)
16910	return err;
16911
16912	src_reg = &regs[insn->src_reg];
16913	if (!(reg_is_pkt_pointer_any(reg: dst_reg) && reg_is_pkt_pointer_any(reg: src_reg)) &&
16914	is_pointer_value(env, regno: insn->src_reg)) {
16915	verbose(private_data: env, fmt: "R%d pointer comparison prohibited\n",
16916	insn->src_reg);
16917	return -EACCES;
16918	}
16919
16920	if (src_reg->type == PTR_TO_STACK)
16921	insn_flags |= INSN_F_SRC_REG_STACK;
16922	if (dst_reg->type == PTR_TO_STACK)
16923	insn_flags |= INSN_F_DST_REG_STACK;
16924	} else {
16925	if (insn->src_reg != BPF_REG_0) {
16926	verbose(private_data: env, fmt: "BPF_JMP/JMP32 uses reserved fields\n");
16927	return -EINVAL;
16928	}
16929	src_reg = &env->fake_reg[0];
16930	memset(src_reg, 0, sizeof(*src_reg));
16931	src_reg->type = SCALAR_VALUE;
16932	__mark_reg_known(reg: src_reg, imm: insn->imm);
16933
16934	if (dst_reg->type == PTR_TO_STACK)
16935	insn_flags |= INSN_F_DST_REG_STACK;
16936	}
16937
16938	if (insn_flags) {
16939	err = push_jmp_history(env, cur: this_branch, insn_flags, linked_regs: 0);
16940	if (err)
16941	return err;
16942	}
16943
16944	is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
16945	pred = is_branch_taken(reg1: dst_reg, reg2: src_reg, opcode, is_jmp32);
16946	if (pred >= 0) {
16947	/* If we get here with a dst_reg pointer type it is because
16948	* above is_branch_taken() special cased the 0 comparison.
16949	*/
16950	if (!__is_pointer_value(allow_ptr_leaks: false, reg: dst_reg))
16951	err = mark_chain_precision(env, regno: insn->dst_reg);
16952	if (BPF_SRC(insn->code) == BPF_X && !err &&
16953	!__is_pointer_value(allow_ptr_leaks: false, reg: src_reg))
16954	err = mark_chain_precision(env, regno: insn->src_reg);
16955	if (err)
16956	return err;
16957	}
16958
16959	if (pred == 1) {
16960	/* Only follow the goto, ignore fall-through. If needed, push
16961	* the fall-through branch for simulation under speculative
16962	* execution.
16963	*/
16964	if (!env->bypass_spec_v1) {
16965	err = sanitize_speculative_path(env, insn, next_idx: *insn_idx + 1, curr_idx: *insn_idx);
16966	if (err < 0)
16967	return err;
16968	}
16969	if (env->log.level & BPF_LOG_LEVEL)
16970	print_insn_state(env, vstate: this_branch, frameno: this_branch->curframe);
16971	*insn_idx += insn->off;
16972	return 0;
16973	} else if (pred == 0) {
16974	/* Only follow the fall-through branch, since that's where the
16975	* program will go. If needed, push the goto branch for
16976	* simulation under speculative execution.
16977	*/
16978	if (!env->bypass_spec_v1) {
16979	err = sanitize_speculative_path(env, insn, next_idx: *insn_idx + insn->off + 1,
16980	curr_idx: *insn_idx);
16981	if (err < 0)
16982	return err;
16983	}
16984	if (env->log.level & BPF_LOG_LEVEL)
16985	print_insn_state(env, vstate: this_branch, frameno: this_branch->curframe);
16986	return 0;
16987	}
16988
16989	/* Push scalar registers sharing same ID to jump history,
16990	* do this before creating 'other_branch', so that both
16991	* 'this_branch' and 'other_branch' share this history
16992	* if parent state is created.
16993	*/
16994	if (BPF_SRC(insn->code) == BPF_X && src_reg->type == SCALAR_VALUE && src_reg->id)
16995	collect_linked_regs(vstate: this_branch, id: src_reg->id, linked_regs: &linked_regs);
16996	if (dst_reg->type == SCALAR_VALUE && dst_reg->id)
16997	collect_linked_regs(vstate: this_branch, id: dst_reg->id, linked_regs: &linked_regs);
16998	if (linked_regs.cnt > 1) {
16999	err = push_jmp_history(env, cur: this_branch, insn_flags: 0, linked_regs: linked_regs_pack(s: &linked_regs));
17000	if (err)
17001	return err;
17002	}
17003
17004	other_branch = push_stack(env, insn_idx: *insn_idx + insn->off + 1, prev_insn_idx: *insn_idx, speculative: false);
17005	if (IS_ERR(ptr: other_branch))
17006	return PTR_ERR(ptr: other_branch);
17007	other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
17008
17009	if (BPF_SRC(insn->code) == BPF_X) {
17010	err = reg_set_min_max(env,
17011	true_reg1: &other_branch_regs[insn->dst_reg],
17012	true_reg2: &other_branch_regs[insn->src_reg],
17013	false_reg1: dst_reg, false_reg2: src_reg, opcode, is_jmp32);
17014	} else /* BPF_SRC(insn->code) == BPF_K */ {
17015	/* reg_set_min_max() can mangle the fake_reg. Make a copy
17016	* so that these are two different memory locations. The
17017	* src_reg is not used beyond here in context of K.
17018	*/
17019	memcpy(&env->fake_reg[1], &env->fake_reg[0],
17020	sizeof(env->fake_reg[0]));
17021	err = reg_set_min_max(env,
17022	true_reg1: &other_branch_regs[insn->dst_reg],
17023	true_reg2: &env->fake_reg[0],
17024	false_reg1: dst_reg, false_reg2: &env->fake_reg[1],
17025	opcode, is_jmp32);
17026	}
17027	if (err)
17028	return err;
17029
17030	if (BPF_SRC(insn->code) == BPF_X &&
17031	src_reg->type == SCALAR_VALUE && src_reg->id &&
17032	!WARN_ON_ONCE(src_reg->id != other_branch_regs[insn->src_reg].id)) {
17033	sync_linked_regs(vstate: this_branch, known_reg: src_reg, linked_regs: &linked_regs);
17034	sync_linked_regs(vstate: other_branch, known_reg: &other_branch_regs[insn->src_reg], linked_regs: &linked_regs);
17035	}
17036	if (dst_reg->type == SCALAR_VALUE && dst_reg->id &&
17037	!WARN_ON_ONCE(dst_reg->id != other_branch_regs[insn->dst_reg].id)) {
17038	sync_linked_regs(vstate: this_branch, known_reg: dst_reg, linked_regs: &linked_regs);
17039	sync_linked_regs(vstate: other_branch, known_reg: &other_branch_regs[insn->dst_reg], linked_regs: &linked_regs);
17040	}
17041
17042	/* if one pointer register is compared to another pointer
17043	* register check if PTR_MAYBE_NULL could be lifted.
17044	* E.g. register A - maybe null
17045	* register B - not null
17046	* for JNE A, B, ... - A is not null in the false branch;
17047	* for JEQ A, B, ... - A is not null in the true branch.
17048	*
17049	* Since PTR_TO_BTF_ID points to a kernel struct that does
17050	* not need to be null checked by the BPF program, i.e.,
17051	* could be null even without PTR_MAYBE_NULL marking, so
17052	* only propagate nullness when neither reg is that type.
17053	*/
17054	if (!is_jmp32 && BPF_SRC(insn->code) == BPF_X &&
17055	__is_pointer_value(allow_ptr_leaks: false, reg: src_reg) && __is_pointer_value(allow_ptr_leaks: false, reg: dst_reg) &&
17056	type_may_be_null(type: src_reg->type) != type_may_be_null(type: dst_reg->type) &&
17057	base_type(type: src_reg->type) != PTR_TO_BTF_ID &&
17058	base_type(type: dst_reg->type) != PTR_TO_BTF_ID) {
17059	eq_branch_regs = NULL;
17060	switch (opcode) {
17061	case BPF_JEQ:
17062	eq_branch_regs = other_branch_regs;
17063	break;
17064	case BPF_JNE:
17065	eq_branch_regs = regs;
17066	break;
17067	default:
17068	/* do nothing */
17069	break;
17070	}
17071	if (eq_branch_regs) {
17072	if (type_may_be_null(type: src_reg->type))
17073	mark_ptr_not_null_reg(reg: &eq_branch_regs[insn->src_reg]);
17074	else
17075	mark_ptr_not_null_reg(reg: &eq_branch_regs[insn->dst_reg]);
17076	}
17077	}
17078
17079	/* detect if R == 0 where R is returned from bpf_map_lookup_elem().
17080	* NOTE: these optimizations below are related with pointer comparison
17081	* which will never be JMP32.
17082	*/
17083	if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
17084	insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
17085	type_may_be_null(type: dst_reg->type)) {
17086	/* Mark all identical registers in each branch as either
17087	* safe or unknown depending R == 0 or R != 0 conditional.
17088	*/
17089	mark_ptr_or_null_regs(vstate: this_branch, regno: insn->dst_reg,
17090	is_null: opcode == BPF_JNE);
17091	mark_ptr_or_null_regs(vstate: other_branch, regno: insn->dst_reg,
17092	is_null: opcode == BPF_JEQ);
17093	} else if (!try_match_pkt_pointers(insn, dst_reg, src_reg: &regs[insn->src_reg],
17094	this_branch, other_branch) &&
17095	is_pointer_value(env, regno: insn->dst_reg)) {
17096	verbose(private_data: env, fmt: "R%d pointer comparison prohibited\n",
17097	insn->dst_reg);
17098	return -EACCES;
17099	}
17100	if (env->log.level & BPF_LOG_LEVEL)
17101	print_insn_state(env, vstate: this_branch, frameno: this_branch->curframe);
17102	return 0;
17103	}
17104
17105	/* verify BPF_LD_IMM64 instruction */
17106	static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
17107	{
17108	struct bpf_insn_aux_data *aux = cur_aux(env);
17109	struct bpf_reg_state *regs = cur_regs(env);
17110	struct bpf_reg_state *dst_reg;
17111	struct bpf_map *map;
17112	int err;
17113
17114	if (BPF_SIZE(insn->code) != BPF_DW) {
17115	verbose(private_data: env, fmt: "invalid BPF_LD_IMM insn\n");
17116	return -EINVAL;
17117	}
17118	if (insn->off != 0) {
17119	verbose(private_data: env, fmt: "BPF_LD_IMM64 uses reserved fields\n");
17120	return -EINVAL;
17121	}
17122
17123	err = check_reg_arg(env, regno: insn->dst_reg, t: DST_OP);
17124	if (err)
17125	return err;
17126
17127	dst_reg = &regs[insn->dst_reg];
17128	if (insn->src_reg == 0) {
17129	u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
17130
17131	dst_reg->type = SCALAR_VALUE;
17132	__mark_reg_known(reg: &regs[insn->dst_reg], imm);
17133	return 0;
17134	}
17135
17136	/* All special src_reg cases are listed below. From this point onwards
17137	* we either succeed and assign a corresponding dst_reg->type after
17138	* zeroing the offset, or fail and reject the program.
17139	*/
17140	mark_reg_known_zero(env, regs, regno: insn->dst_reg);
17141
17142	if (insn->src_reg == BPF_PSEUDO_BTF_ID) {
17143	dst_reg->type = aux->btf_var.reg_type;
17144	switch (base_type(type: dst_reg->type)) {
17145	case PTR_TO_MEM:
17146	dst_reg->mem_size = aux->btf_var.mem_size;
17147	break;
17148	case PTR_TO_BTF_ID:
17149	dst_reg->btf = aux->btf_var.btf;
17150	dst_reg->btf_id = aux->btf_var.btf_id;
17151	break;
17152	default:
17153	verifier_bug(env, "pseudo btf id: unexpected dst reg type");
17154	return -EFAULT;
17155	}
17156	return 0;
17157	}
17158
17159	if (insn->src_reg == BPF_PSEUDO_FUNC) {
17160	struct bpf_prog_aux *aux = env->prog->aux;
17161	u32 subprogno = find_subprog(env,
17162	off: env->insn_idx + insn->imm + 1);
17163
17164	if (!aux->func_info) {
17165	verbose(private_data: env, fmt: "missing btf func_info\n");
17166	return -EINVAL;
17167	}
17168	if (aux->func_info_aux[subprogno].linkage != BTF_FUNC_STATIC) {
17169	verbose(private_data: env, fmt: "callback function not static\n");
17170	return -EINVAL;
17171	}
17172
17173	dst_reg->type = PTR_TO_FUNC;
17174	dst_reg->subprogno = subprogno;
17175	return 0;
17176	}
17177
17178	map = env->used_maps[aux->map_index];
17179	dst_reg->map_ptr = map;
17180
17181	if (insn->src_reg == BPF_PSEUDO_MAP_VALUE ||
17182	insn->src_reg == BPF_PSEUDO_MAP_IDX_VALUE) {
17183	if (map->map_type == BPF_MAP_TYPE_ARENA) {
17184	__mark_reg_unknown(env, reg: dst_reg);
17185	return 0;
17186	}
17187	dst_reg->type = PTR_TO_MAP_VALUE;
17188	dst_reg->off = aux->map_off;
17189	WARN_ON_ONCE(map->map_type != BPF_MAP_TYPE_INSN_ARRAY &&
17190	map->max_entries != 1);
17191	/* We want reg->id to be same (0) as map_value is not distinct */
17192	} else if (insn->src_reg == BPF_PSEUDO_MAP_FD ||
17193	insn->src_reg == BPF_PSEUDO_MAP_IDX) {
17194	dst_reg->type = CONST_PTR_TO_MAP;
17195	} else {
17196	verifier_bug(env, "unexpected src reg value for ldimm64");
17197	return -EFAULT;
17198	}
17199
17200	return 0;
17201	}
17202
17203	static bool may_access_skb(enum bpf_prog_type type)
17204	{
17205	switch (type) {
17206	case BPF_PROG_TYPE_SOCKET_FILTER:
17207	case BPF_PROG_TYPE_SCHED_CLS:
17208	case BPF_PROG_TYPE_SCHED_ACT:
17209	return true;
17210	default:
17211	return false;
17212	}
17213	}
17214
17215	/* verify safety of LD_ABS|LD_IND instructions:
17216	* - they can only appear in the programs where ctx == skb
17217	* - since they are wrappers of function calls, they scratch R1-R5 registers,
17218	* preserve R6-R9, and store return value into R0
17219	*
17220	* Implicit input:
17221	* ctx == skb == R6 == CTX
17222	*
17223	* Explicit input:
17224	* SRC == any register
17225	* IMM == 32-bit immediate
17226	*
17227	* Output:
17228	* R0 - 8/16/32-bit skb data converted to cpu endianness
17229	*/
17230	static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
17231	{
17232	struct bpf_reg_state *regs = cur_regs(env);
17233	static const int ctx_reg = BPF_REG_6;
17234	u8 mode = BPF_MODE(insn->code);
17235	int i, err;
17236
17237	if (!may_access_skb(type: resolve_prog_type(prog: env->prog))) {
17238	verbose(private_data: env, fmt: "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
17239	return -EINVAL;
17240	}
17241
17242	if (!env->ops->gen_ld_abs) {
17243	verifier_bug(env, "gen_ld_abs is null");
17244	return -EFAULT;
17245	}
17246
17247	if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
17248	BPF_SIZE(insn->code) == BPF_DW ||
17249	(mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
17250	verbose(private_data: env, fmt: "BPF_LD_[ABS|IND] uses reserved fields\n");
17251	return -EINVAL;
17252	}
17253
17254	/* check whether implicit source operand (register R6) is readable */
17255	err = check_reg_arg(env, regno: ctx_reg, t: SRC_OP);
17256	if (err)
17257	return err;
17258
17259	/* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
17260	* gen_ld_abs() may terminate the program at runtime, leading to
17261	* reference leak.
17262	*/
17263	err = check_resource_leak(env, exception_exit: false, check_lock: true, prefix: "BPF_LD_[ABS|IND]");
17264	if (err)
17265	return err;
17266
17267	if (regs[ctx_reg].type != PTR_TO_CTX) {
17268	verbose(private_data: env,
17269	fmt: "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
17270	return -EINVAL;
17271	}
17272
17273	if (mode == BPF_IND) {
17274	/* check explicit source operand */
17275	err = check_reg_arg(env, regno: insn->src_reg, t: SRC_OP);
17276	if (err)
17277	return err;
17278	}
17279
17280	err = check_ptr_off_reg(env, reg: &regs[ctx_reg], regno: ctx_reg);
17281	if (err < 0)
17282	return err;
17283
17284	/* reset caller saved regs to unreadable */
17285	for (i = 0; i < CALLER_SAVED_REGS; i++) {
17286	mark_reg_not_init(env, regs, regno: caller_saved[i]);
17287	check_reg_arg(env, regno: caller_saved[i], t: DST_OP_NO_MARK);
17288	}
17289
17290	/* mark destination R0 register as readable, since it contains
17291	* the value fetched from the packet.
17292	* Already marked as written above.
17293	*/
17294	mark_reg_unknown(env, regs, regno: BPF_REG_0);
17295	/* ld_abs load up to 32-bit skb data. */
17296	regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
17297	return 0;
17298	}
17299
17300	static int check_return_code(struct bpf_verifier_env *env, int regno, const char *reg_name)
17301	{
17302	const char *exit_ctx = "At program exit";
17303	struct tnum enforce_attach_type_range = tnum_unknown;
17304	const struct bpf_prog *prog = env->prog;
17305	struct bpf_reg_state *reg = reg_state(env, regno);
17306	struct bpf_retval_range range = retval_range(minval: 0, maxval: 1);
17307	enum bpf_prog_type prog_type = resolve_prog_type(prog: env->prog);
17308	int err;
17309	struct bpf_func_state *frame = env->cur_state->frame[0];
17310	const bool is_subprog = frame->subprogno;
17311	bool return_32bit = false;
17312	const struct btf_type *reg_type, *ret_type = NULL;
17313
17314	/* LSM and struct_ops func-ptr's return type could be "void" */
17315	if (!is_subprog || frame->in_exception_callback_fn) {
17316	switch (prog_type) {
17317	case BPF_PROG_TYPE_LSM:
17318	if (prog->expected_attach_type == BPF_LSM_CGROUP)
17319	/* See below, can be 0 or 0-1 depending on hook. */
17320	break;
17321	if (!prog->aux->attach_func_proto->type)
17322	return 0;
17323	break;
17324	case BPF_PROG_TYPE_STRUCT_OPS:
17325	if (!prog->aux->attach_func_proto->type)
17326	return 0;
17327
17328	if (frame->in_exception_callback_fn)
17329	break;
17330
17331	/* Allow a struct_ops program to return a referenced kptr if it
17332	* matches the operator's return type and is in its unmodified
17333	* form. A scalar zero (i.e., a null pointer) is also allowed.
17334	*/
17335	reg_type = reg->btf ? btf_type_by_id(btf: reg->btf, type_id: reg->btf_id) : NULL;
17336	ret_type = btf_type_resolve_ptr(btf: prog->aux->attach_btf,
17337	id: prog->aux->attach_func_proto->type,
17338	NULL);
17339	if (ret_type && ret_type == reg_type && reg->ref_obj_id)
17340	return __check_ptr_off_reg(env, reg, regno, fixed_off_ok: false);
17341	break;
17342	default:
17343	break;
17344	}
17345	}
17346
17347	/* eBPF calling convention is such that R0 is used
17348	* to return the value from eBPF program.
17349	* Make sure that it's readable at this time
17350	* of bpf_exit, which means that program wrote
17351	* something into it earlier
17352	*/
17353	err = check_reg_arg(env, regno, t: SRC_OP);
17354	if (err)
17355	return err;
17356
17357	if (is_pointer_value(env, regno)) {
17358	verbose(private_data: env, fmt: "R%d leaks addr as return value\n", regno);
17359	return -EACCES;
17360	}
17361
17362	if (frame->in_async_callback_fn) {
17363	exit_ctx = "At async callback return";
17364	range = frame->callback_ret_range;
17365	goto enforce_retval;
17366	}
17367
17368	if (is_subprog && !frame->in_exception_callback_fn) {
17369	if (reg->type != SCALAR_VALUE) {
17370	verbose(private_data: env, fmt: "At subprogram exit the register R%d is not a scalar value (%s)\n",
17371	regno, reg_type_str(env, type: reg->type));
17372	return -EINVAL;
17373	}
17374	return 0;
17375	}
17376
17377	switch (prog_type) {
17378	case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
17379	if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
17380	env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG ||
17381	env->prog->expected_attach_type == BPF_CGROUP_UNIX_RECVMSG ||
17382	env->prog->expected_attach_type == BPF_CGROUP_INET4_GETPEERNAME ||
17383	env->prog->expected_attach_type == BPF_CGROUP_INET6_GETPEERNAME ||
17384	env->prog->expected_attach_type == BPF_CGROUP_UNIX_GETPEERNAME ||
17385	env->prog->expected_attach_type == BPF_CGROUP_INET4_GETSOCKNAME ||
17386	env->prog->expected_attach_type == BPF_CGROUP_INET6_GETSOCKNAME ||
17387	env->prog->expected_attach_type == BPF_CGROUP_UNIX_GETSOCKNAME)
17388	range = retval_range(minval: 1, maxval: 1);
17389	if (env->prog->expected_attach_type == BPF_CGROUP_INET4_BIND ||
17390	env->prog->expected_attach_type == BPF_CGROUP_INET6_BIND)
17391	range = retval_range(minval: 0, maxval: 3);
17392	break;
17393	case BPF_PROG_TYPE_CGROUP_SKB:
17394	if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
17395	range = retval_range(minval: 0, maxval: 3);
17396	enforce_attach_type_range = tnum_range(min: 2, max: 3);
17397	}
17398	break;
17399	case BPF_PROG_TYPE_CGROUP_SOCK:
17400	case BPF_PROG_TYPE_SOCK_OPS:
17401	case BPF_PROG_TYPE_CGROUP_DEVICE:
17402	case BPF_PROG_TYPE_CGROUP_SYSCTL:
17403	case BPF_PROG_TYPE_CGROUP_SOCKOPT:
17404	break;
17405	case BPF_PROG_TYPE_RAW_TRACEPOINT:
17406	if (!env->prog->aux->attach_btf_id)
17407	return 0;
17408	range = retval_range(minval: 0, maxval: 0);
17409	break;
17410	case BPF_PROG_TYPE_TRACING:
17411	switch (env->prog->expected_attach_type) {
17412	case BPF_TRACE_FENTRY:
17413	case BPF_TRACE_FEXIT:
17414	range = retval_range(minval: 0, maxval: 0);
17415	break;
17416	case BPF_TRACE_RAW_TP:
17417	case BPF_MODIFY_RETURN:
17418	return 0;
17419	case BPF_TRACE_ITER:
17420	break;
17421	default:
17422	return -ENOTSUPP;
17423	}
17424	break;
17425	case BPF_PROG_TYPE_KPROBE:
17426	switch (env->prog->expected_attach_type) {
17427	case BPF_TRACE_KPROBE_SESSION:
17428	case BPF_TRACE_UPROBE_SESSION:
17429	range = retval_range(minval: 0, maxval: 1);
17430	break;
17431	default:
17432	return 0;
17433	}
17434	break;
17435	case BPF_PROG_TYPE_SK_LOOKUP:
17436	range = retval_range(minval: SK_DROP, maxval: SK_PASS);
17437	break;
17438
17439	case BPF_PROG_TYPE_LSM:
17440	if (env->prog->expected_attach_type != BPF_LSM_CGROUP) {
17441	/* no range found, any return value is allowed */
17442	if (!get_func_retval_range(prog: env->prog, range: &range))
17443	return 0;
17444	/* no restricted range, any return value is allowed */
17445	if (range.minval == S32_MIN && range.maxval == S32_MAX)
17446	return 0;
17447	return_32bit = true;
17448	} else if (!env->prog->aux->attach_func_proto->type) {
17449	/* Make sure programs that attach to void
17450	* hooks don't try to modify return value.
17451	*/
17452	range = retval_range(minval: 1, maxval: 1);
17453	}
17454	break;
17455
17456	case BPF_PROG_TYPE_NETFILTER:
17457	range = retval_range(NF_DROP, NF_ACCEPT);
17458	break;
17459	case BPF_PROG_TYPE_STRUCT_OPS:
17460	if (!ret_type)
17461	return 0;
17462	range = retval_range(minval: 0, maxval: 0);
17463	break;
17464	case BPF_PROG_TYPE_EXT:
17465	/* freplace program can return anything as its return value
17466	* depends on the to-be-replaced kernel func or bpf program.
17467	*/
17468	default:
17469	return 0;
17470	}
17471
17472	enforce_retval:
17473	if (reg->type != SCALAR_VALUE) {
17474	verbose(private_data: env, fmt: "%s the register R%d is not a known value (%s)\n",
17475	exit_ctx, regno, reg_type_str(env, type: reg->type));
17476	return -EINVAL;
17477	}
17478
17479	err = mark_chain_precision(env, regno);
17480	if (err)
17481	return err;
17482
17483	if (!retval_range_within(range, reg, return_32bit)) {
17484	verbose_invalid_scalar(env, reg, range, ctx: exit_ctx, reg_name);
17485	if (!is_subprog &&
17486	prog->expected_attach_type == BPF_LSM_CGROUP &&
17487	prog_type == BPF_PROG_TYPE_LSM &&
17488	!prog->aux->attach_func_proto->type)
17489	verbose(private_data: env, fmt: "Note, BPF_LSM_CGROUP that attach to void LSM hooks can't modify return value!\n");
17490	return -EINVAL;
17491	}
17492
17493	if (!tnum_is_unknown(a: enforce_attach_type_range) &&
17494	tnum_in(a: enforce_attach_type_range, b: reg->var_off))
17495	env->prog->enforce_expected_attach_type = 1;
17496	return 0;
17497	}
17498
17499	static void mark_subprog_changes_pkt_data(struct bpf_verifier_env *env, int off)
17500	{
17501	struct bpf_subprog_info *subprog;
17502
17503	subprog = bpf_find_containing_subprog(env, off);
17504	subprog->changes_pkt_data = true;
17505	}
17506
17507	static void mark_subprog_might_sleep(struct bpf_verifier_env *env, int off)
17508	{
17509	struct bpf_subprog_info *subprog;
17510
17511	subprog = bpf_find_containing_subprog(env, off);
17512	subprog->might_sleep = true;
17513	}
17514
17515	/* 't' is an index of a call-site.
17516	* 'w' is a callee entry point.
17517	* Eventually this function would be called when env->cfg.insn_state[w] == EXPLORED.
17518	* Rely on DFS traversal order and absence of recursive calls to guarantee that
17519	* callee's change_pkt_data marks would be correct at that moment.
17520	*/
17521	static void merge_callee_effects(struct bpf_verifier_env *env, int t, int w)
17522	{
17523	struct bpf_subprog_info *caller, *callee;
17524
17525	caller = bpf_find_containing_subprog(env, off: t);
17526	callee = bpf_find_containing_subprog(env, off: w);
17527	caller->changes_pkt_data |= callee->changes_pkt_data;
17528	caller->might_sleep |= callee->might_sleep;
17529	}
17530
17531	/* non-recursive DFS pseudo code
17532	* 1 procedure DFS-iterative(G,v):
17533	* 2 label v as discovered
17534	* 3 let S be a stack
17535	* 4 S.push(v)
17536	* 5 while S is not empty
17537	* 6 t <- S.peek()
17538	* 7 if t is what we're looking for:
17539	* 8 return t
17540	* 9 for all edges e in G.adjacentEdges(t) do
17541	* 10 if edge e is already labelled
17542	* 11 continue with the next edge
17543	* 12 w <- G.adjacentVertex(t,e)
17544	* 13 if vertex w is not discovered and not explored
17545	* 14 label e as tree-edge
17546	* 15 label w as discovered
17547	* 16 S.push(w)
17548	* 17 continue at 5
17549	* 18 else if vertex w is discovered
17550	* 19 label e as back-edge
17551	* 20 else
17552	* 21 // vertex w is explored
17553	* 22 label e as forward- or cross-edge
17554	* 23 label t as explored
17555	* 24 S.pop()
17556	*
17557	* convention:
17558	* 0x10 - discovered
17559	* 0x11 - discovered and fall-through edge labelled
17560	* 0x12 - discovered and fall-through and branch edges labelled
17561	* 0x20 - explored
17562	*/
17563
17564	enum {
17565	DISCOVERED = 0x10,
17566	EXPLORED = 0x20,
17567	FALLTHROUGH = 1,
17568	BRANCH = 2,
17569	};
17570
17571	static void mark_prune_point(struct bpf_verifier_env *env, int idx)
17572	{
17573	env->insn_aux_data[idx].prune_point = true;
17574	}
17575
17576	static bool is_prune_point(struct bpf_verifier_env *env, int insn_idx)
17577	{
17578	return env->insn_aux_data[insn_idx].prune_point;
17579	}
17580
17581	static void mark_force_checkpoint(struct bpf_verifier_env *env, int idx)
17582	{
17583	env->insn_aux_data[idx].force_checkpoint = true;
17584	}
17585
17586	static bool is_force_checkpoint(struct bpf_verifier_env *env, int insn_idx)
17587	{
17588	return env->insn_aux_data[insn_idx].force_checkpoint;
17589	}
17590
17591	static void mark_calls_callback(struct bpf_verifier_env *env, int idx)
17592	{
17593	env->insn_aux_data[idx].calls_callback = true;
17594	}
17595
17596	bool bpf_calls_callback(struct bpf_verifier_env *env, int insn_idx)
17597	{
17598	return env->insn_aux_data[insn_idx].calls_callback;
17599	}
17600
17601	enum {
17602	DONE_EXPLORING = 0,
17603	KEEP_EXPLORING = 1,
17604	};
17605
17606	/* t, w, e - match pseudo-code above:
17607	* t - index of current instruction
17608	* w - next instruction
17609	* e - edge
17610	*/
17611	static int push_insn(int t, int w, int e, struct bpf_verifier_env *env)
17612	{
17613	int *insn_stack = env->cfg.insn_stack;
17614	int *insn_state = env->cfg.insn_state;
17615
17616	if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
17617	return DONE_EXPLORING;
17618
17619	if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
17620	return DONE_EXPLORING;
17621
17622	if (w < 0 || w >= env->prog->len) {
17623	verbose_linfo(env, insn_off: t, prefix_fmt: "%d: ", t);
17624	verbose(private_data: env, fmt: "jump out of range from insn %d to %d\n", t, w);
17625	return -EINVAL;
17626	}
17627
17628	if (e == BRANCH) {
17629	/* mark branch target for state pruning */
17630	mark_prune_point(env, idx: w);
17631	mark_jmp_point(env, idx: w);
17632	}
17633
17634	if (insn_state[w] == 0) {
17635	/* tree-edge */
17636	insn_state[t] = DISCOVERED | e;
17637	insn_state[w] = DISCOVERED;
17638	if (env->cfg.cur_stack >= env->prog->len)
17639	return -E2BIG;
17640	insn_stack[env->cfg.cur_stack++] = w;
17641	return KEEP_EXPLORING;
17642	} else if ((insn_state[w] & 0xF0) == DISCOVERED) {
17643	if (env->bpf_capable)
17644	return DONE_EXPLORING;
17645	verbose_linfo(env, insn_off: t, prefix_fmt: "%d: ", t);
17646	verbose_linfo(env, insn_off: w, prefix_fmt: "%d: ", w);
17647	verbose(private_data: env, fmt: "back-edge from insn %d to %d\n", t, w);
17648	return -EINVAL;
17649	} else if (insn_state[w] == EXPLORED) {
17650	/* forward- or cross-edge */
17651	insn_state[t] = DISCOVERED | e;
17652	} else {
17653	verifier_bug(env, "insn state internal bug");
17654	return -EFAULT;
17655	}
17656	return DONE_EXPLORING;
17657	}
17658
17659	static int visit_func_call_insn(int t, struct bpf_insn *insns,
17660	struct bpf_verifier_env *env,
17661	bool visit_callee)
17662	{
17663	int ret, insn_sz;
17664	int w;
17665
17666	insn_sz = bpf_is_ldimm64(insn: &insns[t]) ? 2 : 1;
17667	ret = push_insn(t, w: t + insn_sz, e: FALLTHROUGH, env);
17668	if (ret)
17669	return ret;
17670
17671	mark_prune_point(env, idx: t + insn_sz);
17672	/* when we exit from subprog, we need to record non-linear history */
17673	mark_jmp_point(env, idx: t + insn_sz);
17674
17675	if (visit_callee) {
17676	w = t + insns[t].imm + 1;
17677	mark_prune_point(env, idx: t);
17678	merge_callee_effects(env, t, w);
17679	ret = push_insn(t, w, e: BRANCH, env);
17680	}
17681	return ret;
17682	}
17683
17684	/* Bitmask with 1s for all caller saved registers */
17685	#define ALL_CALLER_SAVED_REGS ((1u << CALLER_SAVED_REGS) - 1)
17686
17687	/* True if do_misc_fixups() replaces calls to helper number 'imm',
17688	* replacement patch is presumed to follow bpf_fastcall contract
17689	* (see mark_fastcall_pattern_for_call() below).
17690	*/
17691	static bool verifier_inlines_helper_call(struct bpf_verifier_env *env, s32 imm)
17692	{
17693	switch (imm) {
17694	#ifdef CONFIG_X86_64
17695	case BPF_FUNC_get_smp_processor_id:
17696	return env->prog->jit_requested && bpf_jit_supports_percpu_insn();
17697	#endif
17698	default:
17699	return false;
17700	}
17701	}
17702
17703	struct call_summary {
17704	u8 num_params;
17705	bool is_void;
17706	bool fastcall;
17707	};
17708
17709	/* If @call is a kfunc or helper call, fills @cs and returns true,
17710	* otherwise returns false.
17711	*/
17712	static bool get_call_summary(struct bpf_verifier_env *env, struct bpf_insn *call,
17713	struct call_summary *cs)
17714	{
17715	struct bpf_kfunc_call_arg_meta meta;
17716	const struct bpf_func_proto *fn;
17717	int i;
17718
17719	if (bpf_helper_call(insn: call)) {
17720
17721	if (get_helper_proto(env, func_id: call->imm, ptr: &fn) < 0)
17722	/* error would be reported later */
17723	return false;
17724	cs->fastcall = fn->allow_fastcall &&
17725	(verifier_inlines_helper_call(env, imm: call->imm) ||
17726	bpf_jit_inlines_helper_call(imm: call->imm));
17727	cs->is_void = fn->ret_type == RET_VOID;
17728	cs->num_params = 0;
17729	for (i = 0; i < ARRAY_SIZE(fn->arg_type); ++i) {
17730	if (fn->arg_type[i] == ARG_DONTCARE)
17731	break;
17732	cs->num_params++;
17733	}
17734	return true;
17735	}
17736
17737	if (bpf_pseudo_kfunc_call(insn: call)) {
17738	int err;
17739
17740	err = fetch_kfunc_meta(env, insn: call, meta: &meta, NULL);
17741	if (err < 0)
17742	/* error would be reported later */
17743	return false;
17744	cs->num_params = btf_type_vlen(t: meta.func_proto);
17745	cs->fastcall = meta.kfunc_flags & KF_FASTCALL;
17746	cs->is_void = btf_type_is_void(t: btf_type_by_id(btf: meta.btf, type_id: meta.func_proto->type));
17747	return true;
17748	}
17749
17750	return false;
17751	}
17752
17753	/* LLVM define a bpf_fastcall function attribute.
17754	* This attribute means that function scratches only some of
17755	* the caller saved registers defined by ABI.
17756	* For BPF the set of such registers could be defined as follows:
17757	* - R0 is scratched only if function is non-void;
17758	* - R1-R5 are scratched only if corresponding parameter type is defined
17759	* in the function prototype.
17760	*
17761	* The contract between kernel and clang allows to simultaneously use
17762	* such functions and maintain backwards compatibility with old
17763	* kernels that don't understand bpf_fastcall calls:
17764	*
17765	* - for bpf_fastcall calls clang allocates registers as-if relevant r0-r5
17766	* registers are not scratched by the call;
17767	*
17768	* - as a post-processing step, clang visits each bpf_fastcall call and adds
17769	* spill/fill for every live r0-r5;
17770	*
17771	* - stack offsets used for the spill/fill are allocated as lowest
17772	* stack offsets in whole function and are not used for any other
17773	* purposes;
17774	*
17775	* - when kernel loads a program, it looks for such patterns
17776	* (bpf_fastcall function surrounded by spills/fills) and checks if
17777	* spill/fill stack offsets are used exclusively in fastcall patterns;
17778	*
17779	* - if so, and if verifier or current JIT inlines the call to the
17780	* bpf_fastcall function (e.g. a helper call), kernel removes unnecessary
17781	* spill/fill pairs;
17782	*
17783	* - when old kernel loads a program, presence of spill/fill pairs
17784	* keeps BPF program valid, albeit slightly less efficient.
17785	*
17786	* For example:
17787	*
17788	* r1 = 1;
17789	* r2 = 2;
17790	* *(u64 *)(r10 - 8) = r1; r1 = 1;
17791	* *(u64 *)(r10 - 16) = r2; r2 = 2;
17792	* call %[to_be_inlined] --> call %[to_be_inlined]
17793	* r2 = *(u64 *)(r10 - 16); r0 = r1;
17794	* r1 = *(u64 *)(r10 - 8); r0 += r2;
17795	* r0 = r1; exit;
17796	* r0 += r2;
17797	* exit;
17798	*
17799	* The purpose of mark_fastcall_pattern_for_call is to:
17800	* - look for such patterns;
17801	* - mark spill and fill instructions in env->insn_aux_data[*].fastcall_pattern;
17802	* - mark set env->insn_aux_data[*].fastcall_spills_num for call instruction;
17803	* - update env->subprog_info[*]->fastcall_stack_off to find an offset
17804	* at which bpf_fastcall spill/fill stack slots start;
17805	* - update env->subprog_info[*]->keep_fastcall_stack.
17806	*
17807	* The .fastcall_pattern and .fastcall_stack_off are used by
17808	* check_fastcall_stack_contract() to check if every stack access to
17809	* fastcall spill/fill stack slot originates from spill/fill
17810	* instructions, members of fastcall patterns.
17811	*
17812	* If such condition holds true for a subprogram, fastcall patterns could
17813	* be rewritten by remove_fastcall_spills_fills().
17814	* Otherwise bpf_fastcall patterns are not changed in the subprogram
17815	* (code, presumably, generated by an older clang version).
17816	*
17817	* For example, it is *not* safe to remove spill/fill below:
17818	*
17819	* r1 = 1;
17820	* *(u64 *)(r10 - 8) = r1; r1 = 1;
17821	* call %[to_be_inlined] --> call %[to_be_inlined]
17822	* r1 = *(u64 *)(r10 - 8); r0 = *(u64 *)(r10 - 8); <---- wrong !!!
17823	* r0 = *(u64 *)(r10 - 8); r0 += r1;
17824	* r0 += r1; exit;
17825	* exit;
17826	*/
17827	static void mark_fastcall_pattern_for_call(struct bpf_verifier_env *env,
17828	struct bpf_subprog_info *subprog,
17829	int insn_idx, s16 lowest_off)
17830	{
17831	struct bpf_insn *insns = env->prog->insnsi, *stx, *ldx;
17832	struct bpf_insn *call = &env->prog->insnsi[insn_idx];
17833	u32 clobbered_regs_mask;
17834	struct call_summary cs;
17835	u32 expected_regs_mask;
17836	s16 off;
17837	int i;
17838
17839	if (!get_call_summary(env, call, cs: &cs))
17840	return;
17841
17842	/* A bitmask specifying which caller saved registers are clobbered
17843	* by a call to a helper/kfunc *as if* this helper/kfunc follows
17844	* bpf_fastcall contract:
17845	* - includes R0 if function is non-void;
17846	* - includes R1-R5 if corresponding parameter has is described
17847	* in the function prototype.
17848	*/
17849	clobbered_regs_mask = GENMASK(cs.num_params, cs.is_void ? 1 : 0);
17850	/* e.g. if helper call clobbers r{0,1}, expect r{2,3,4,5} in the pattern */
17851	expected_regs_mask = ~clobbered_regs_mask & ALL_CALLER_SAVED_REGS;
17852
17853	/* match pairs of form:
17854	*
17855	* *(u64 *)(r10 - Y) = rX (where Y % 8 == 0)
17856	* ...
17857	* call %[to_be_inlined]
17858	* ...
17859	* rX = *(u64 *)(r10 - Y)
17860	*/
17861	for (i = 1, off = lowest_off; i <= ARRAY_SIZE(caller_saved); ++i, off += BPF_REG_SIZE) {
17862	if (insn_idx - i < 0 || insn_idx + i >= env->prog->len)
17863	break;
17864	stx = &insns[insn_idx - i];
17865	ldx = &insns[insn_idx + i];
17866	/* must be a stack spill/fill pair */
17867	if (stx->code != (BPF_STX | BPF_MEM | BPF_DW) ||
17868	ldx->code != (BPF_LDX | BPF_MEM | BPF_DW) ||
17869	stx->dst_reg != BPF_REG_10 ||
17870	ldx->src_reg != BPF_REG_10)
17871	break;
17872	/* must be a spill/fill for the same reg */
17873	if (stx->src_reg != ldx->dst_reg)
17874	break;
17875	/* must be one of the previously unseen registers */
17876	if ((BIT(stx->src_reg) & expected_regs_mask) == 0)
17877	break;
17878	/* must be a spill/fill for the same expected offset,
17879	* no need to check offset alignment, BPF_DW stack access
17880	* is always 8-byte aligned.
17881	*/
17882	if (stx->off != off || ldx->off != off)
17883	break;
17884	expected_regs_mask &= ~BIT(stx->src_reg);
17885	env->insn_aux_data[insn_idx - i].fastcall_pattern = 1;
17886	env->insn_aux_data[insn_idx + i].fastcall_pattern = 1;
17887	}
17888	if (i == 1)
17889	return;
17890
17891	/* Conditionally set 'fastcall_spills_num' to allow forward
17892	* compatibility when more helper functions are marked as
17893	* bpf_fastcall at compile time than current kernel supports, e.g:
17894	*
17895	* 1: *(u64 *)(r10 - 8) = r1
17896	* 2: call A ;; assume A is bpf_fastcall for current kernel
17897	* 3: r1 = *(u64 *)(r10 - 8)
17898	* 4: *(u64 *)(r10 - 8) = r1
17899	* 5: call B ;; assume B is not bpf_fastcall for current kernel
17900	* 6: r1 = *(u64 *)(r10 - 8)
17901	*
17902	* There is no need to block bpf_fastcall rewrite for such program.
17903	* Set 'fastcall_pattern' for both calls to keep check_fastcall_stack_contract() happy,
17904	* don't set 'fastcall_spills_num' for call B so that remove_fastcall_spills_fills()
17905	* does not remove spill/fill pair {4,6}.
17906	*/
17907	if (cs.fastcall)
17908	env->insn_aux_data[insn_idx].fastcall_spills_num = i - 1;
17909	else
17910	subprog->keep_fastcall_stack = 1;
17911	subprog->fastcall_stack_off = min(subprog->fastcall_stack_off, off);
17912	}
17913
17914	static int mark_fastcall_patterns(struct bpf_verifier_env *env)
17915	{
17916	struct bpf_subprog_info *subprog = env->subprog_info;
17917	struct bpf_insn *insn;
17918	s16 lowest_off;
17919	int s, i;
17920
17921	for (s = 0; s < env->subprog_cnt; ++s, ++subprog) {
17922	/* find lowest stack spill offset used in this subprog */
17923	lowest_off = 0;
17924	for (i = subprog->start; i < (subprog + 1)->start; ++i) {
17925	insn = env->prog->insnsi + i;
17926	if (insn->code != (BPF_STX | BPF_MEM | BPF_DW) ||
17927	insn->dst_reg != BPF_REG_10)
17928	continue;
17929	lowest_off = min(lowest_off, insn->off);
17930	}
17931	/* use this offset to find fastcall patterns */
17932	for (i = subprog->start; i < (subprog + 1)->start; ++i) {
17933	insn = env->prog->insnsi + i;
17934	if (insn->code != (BPF_JMP | BPF_CALL))
17935	continue;
17936	mark_fastcall_pattern_for_call(env, subprog, insn_idx: i, lowest_off);
17937	}
17938	}
17939	return 0;
17940	}
17941
17942	static struct bpf_iarray *iarray_realloc(struct bpf_iarray *old, size_t n_elem)
17943	{
17944	size_t new_size = sizeof(struct bpf_iarray) + n_elem * sizeof(old->items[0]);
17945	struct bpf_iarray *new;
17946
17947	new = kvrealloc(old, new_size, GFP_KERNEL_ACCOUNT);
17948	if (!new) {
17949	/* this is what callers always want, so simplify the call site */
17950	kvfree(addr: old);
17951	return NULL;
17952	}
17953
17954	new->cnt = n_elem;
17955	return new;
17956	}
17957
17958	static int copy_insn_array(struct bpf_map *map, u32 start, u32 end, u32 *items)
17959	{
17960	struct bpf_insn_array_value *value;
17961	u32 i;
17962
17963	for (i = start; i <= end; i++) {
17964	value = map->ops->map_lookup_elem(map, &i);
17965	/*
17966	* map_lookup_elem of an array map will never return an error,
17967	* but not checking it makes some static analysers to worry
17968	*/
17969	if (IS_ERR(ptr: value))
17970	return PTR_ERR(ptr: value);
17971	else if (!value)
17972	return -EINVAL;
17973	items[i - start] = value->xlated_off;
17974	}
17975	return 0;
17976	}
17977
17978	static int cmp_ptr_to_u32(const void *a, const void *b)
17979	{
17980	return *(u32 *)a - *(u32 *)b;
17981	}
17982
17983	static int sort_insn_array_uniq(u32 *items, int cnt)
17984	{
17985	int unique = 1;
17986	int i;
17987
17988	sort(base: items, num: cnt, size: sizeof(items[0]), cmp_func: cmp_ptr_to_u32, NULL);
17989
17990	for (i = 1; i < cnt; i++)
17991	if (items[i] != items[unique - 1])
17992	items[unique++] = items[i];
17993
17994	return unique;
17995	}
17996
17997	/*
17998	* sort_unique({map[start], ..., map[end]}) into off
17999	*/
18000	static int copy_insn_array_uniq(struct bpf_map *map, u32 start, u32 end, u32 *off)
18001	{
18002	u32 n = end - start + 1;
18003	int err;
18004
18005	err = copy_insn_array(map, start, end, items: off);
18006	if (err)
18007	return err;
18008
18009	return sort_insn_array_uniq(items: off, cnt: n);
18010	}
18011
18012	/*
18013	* Copy all unique offsets from the map
18014	*/
18015	static struct bpf_iarray *jt_from_map(struct bpf_map *map)
18016	{
18017	struct bpf_iarray *jt;
18018	int err;
18019	int n;
18020
18021	jt = iarray_realloc(NULL, n_elem: map->max_entries);
18022	if (!jt)
18023	return ERR_PTR(error: -ENOMEM);
18024
18025	n = copy_insn_array_uniq(map, start: 0, end: map->max_entries - 1, off: jt->items);
18026	if (n < 0) {
18027	err = n;
18028	goto err_free;
18029	}
18030	if (n == 0) {
18031	err = -EINVAL;
18032	goto err_free;
18033	}
18034	jt->cnt = n;
18035	return jt;
18036
18037	err_free:
18038	kvfree(addr: jt);
18039	return ERR_PTR(error: err);
18040	}
18041
18042	/*
18043	* Find and collect all maps which fit in the subprog. Return the result as one
18044	* combined jump table in jt->items (allocated with kvcalloc)
18045	*/
18046	static struct bpf_iarray *jt_from_subprog(struct bpf_verifier_env *env,
18047	int subprog_start, int subprog_end)
18048	{
18049	struct bpf_iarray *jt = NULL;
18050	struct bpf_map *map;
18051	struct bpf_iarray *jt_cur;
18052	int i;
18053
18054	for (i = 0; i < env->insn_array_map_cnt; i++) {
18055	/*
18056	* TODO (when needed): collect only jump tables, not static keys
18057	* or maps for indirect calls
18058	*/
18059	map = env->insn_array_maps[i];
18060
18061	jt_cur = jt_from_map(map);
18062	if (IS_ERR(ptr: jt_cur)) {
18063	kvfree(addr: jt);
18064	return jt_cur;
18065	}
18066
18067	/*
18068	* This is enough to check one element. The full table is
18069	* checked to fit inside the subprog later in create_jt()
18070	*/
18071	if (jt_cur->items[0] >= subprog_start && jt_cur->items[0] < subprog_end) {
18072	u32 old_cnt = jt ? jt->cnt : 0;
18073	jt = iarray_realloc(old: jt, n_elem: old_cnt + jt_cur->cnt);
18074	if (!jt) {
18075	kvfree(addr: jt_cur);
18076	return ERR_PTR(error: -ENOMEM);
18077	}
18078	memcpy(jt->items + old_cnt, jt_cur->items, jt_cur->cnt << 2);
18079	}
18080
18081	kvfree(addr: jt_cur);
18082	}
18083
18084	if (!jt) {
18085	verbose(private_data: env, fmt: "no jump tables found for subprog starting at %u\n", subprog_start);
18086	return ERR_PTR(error: -EINVAL);
18087	}
18088
18089	jt->cnt = sort_insn_array_uniq(items: jt->items, cnt: jt->cnt);
18090	return jt;
18091	}
18092
18093	static struct bpf_iarray *
18094	create_jt(int t, struct bpf_verifier_env *env)
18095	{
18096	static struct bpf_subprog_info *subprog;
18097	int subprog_start, subprog_end;
18098	struct bpf_iarray *jt;
18099	int i;
18100
18101	subprog = bpf_find_containing_subprog(env, off: t);
18102	subprog_start = subprog->start;
18103	subprog_end = (subprog + 1)->start;
18104	jt = jt_from_subprog(env, subprog_start, subprog_end);
18105	if (IS_ERR(ptr: jt))
18106	return jt;
18107
18108	/* Check that the every element of the jump table fits within the given subprogram */
18109	for (i = 0; i < jt->cnt; i++) {
18110	if (jt->items[i] < subprog_start || jt->items[i] >= subprog_end) {
18111	verbose(private_data: env, fmt: "jump table for insn %d points outside of the subprog [%u,%u]\n",
18112	t, subprog_start, subprog_end);
18113	kvfree(addr: jt);
18114	return ERR_PTR(error: -EINVAL);
18115	}
18116	}
18117
18118	return jt;
18119	}
18120
18121	/* "conditional jump with N edges" */
18122	static int visit_gotox_insn(int t, struct bpf_verifier_env *env)
18123	{
18124	int *insn_stack = env->cfg.insn_stack;
18125	int *insn_state = env->cfg.insn_state;
18126	bool keep_exploring = false;
18127	struct bpf_iarray *jt;
18128	int i, w;
18129
18130	jt = env->insn_aux_data[t].jt;
18131	if (!jt) {
18132	jt = create_jt(t, env);
18133	if (IS_ERR(ptr: jt))
18134	return PTR_ERR(ptr: jt);
18135
18136	env->insn_aux_data[t].jt = jt;
18137	}
18138
18139	mark_prune_point(env, idx: t);
18140	for (i = 0; i < jt->cnt; i++) {
18141	w = jt->items[i];
18142	if (w < 0 || w >= env->prog->len) {
18143	verbose(private_data: env, fmt: "indirect jump out of range from insn %d to %d\n", t, w);
18144	return -EINVAL;
18145	}
18146
18147	mark_jmp_point(env, idx: w);
18148
18149	/* EXPLORED || DISCOVERED */
18150	if (insn_state[w])
18151	continue;
18152
18153	if (env->cfg.cur_stack >= env->prog->len)
18154	return -E2BIG;
18155
18156	insn_stack[env->cfg.cur_stack++] = w;
18157	insn_state[w] |= DISCOVERED;
18158	keep_exploring = true;
18159	}
18160
18161	return keep_exploring ? KEEP_EXPLORING : DONE_EXPLORING;
18162	}
18163
18164	static int visit_tailcall_insn(struct bpf_verifier_env *env, int t)
18165	{
18166	static struct bpf_subprog_info *subprog;
18167	struct bpf_iarray *jt;
18168
18169	if (env->insn_aux_data[t].jt)
18170	return 0;
18171
18172	jt = iarray_realloc(NULL, n_elem: 2);
18173	if (!jt)
18174	return -ENOMEM;
18175
18176	subprog = bpf_find_containing_subprog(env, off: t);
18177	jt->items[0] = t + 1;
18178	jt->items[1] = subprog->exit_idx;
18179	env->insn_aux_data[t].jt = jt;
18180	return 0;
18181	}
18182
18183	/* Visits the instruction at index t and returns one of the following:
18184	* < 0 - an error occurred
18185	* DONE_EXPLORING - the instruction was fully explored
18186	* KEEP_EXPLORING - there is still work to be done before it is fully explored
18187	*/
18188	static int visit_insn(int t, struct bpf_verifier_env *env)
18189	{
18190	struct bpf_insn *insns = env->prog->insnsi, *insn = &insns[t];
18191	int ret, off, insn_sz;
18192
18193	if (bpf_pseudo_func(insn))
18194	return visit_func_call_insn(t, insns, env, visit_callee: true);
18195
18196	/* All non-branch instructions have a single fall-through edge. */
18197	if (BPF_CLASS(insn->code) != BPF_JMP &&
18198	BPF_CLASS(insn->code) != BPF_JMP32) {
18199	insn_sz = bpf_is_ldimm64(insn) ? 2 : 1;
18200	return push_insn(t, w: t + insn_sz, e: FALLTHROUGH, env);
18201	}
18202
18203	switch (BPF_OP(insn->code)) {
18204	case BPF_EXIT:
18205	return DONE_EXPLORING;
18206
18207	case BPF_CALL:
18208	if (is_async_callback_calling_insn(insn))
18209	/* Mark this call insn as a prune point to trigger
18210	* is_state_visited() check before call itself is
18211	* processed by __check_func_call(). Otherwise new
18212	* async state will be pushed for further exploration.
18213	*/
18214	mark_prune_point(env, idx: t);
18215	/* For functions that invoke callbacks it is not known how many times
18216	* callback would be called. Verifier models callback calling functions
18217	* by repeatedly visiting callback bodies and returning to origin call
18218	* instruction.
18219	* In order to stop such iteration verifier needs to identify when a
18220	* state identical some state from a previous iteration is reached.
18221	* Check below forces creation of checkpoint before callback calling
18222	* instruction to allow search for such identical states.
18223	*/
18224	if (is_sync_callback_calling_insn(insn)) {
18225	mark_calls_callback(env, idx: t);
18226	mark_force_checkpoint(env, idx: t);
18227	mark_prune_point(env, idx: t);
18228	mark_jmp_point(env, idx: t);
18229	}
18230	if (bpf_helper_call(insn)) {
18231	const struct bpf_func_proto *fp;
18232
18233	ret = get_helper_proto(env, func_id: insn->imm, ptr: &fp);
18234	/* If called in a non-sleepable context program will be
18235	* rejected anyway, so we should end up with precise
18236	* sleepable marks on subprogs, except for dead code
18237	* elimination.
18238	*/
18239	if (ret == 0 && fp->might_sleep)
18240	mark_subprog_might_sleep(env, off: t);
18241	if (bpf_helper_changes_pkt_data(func_id: insn->imm))
18242	mark_subprog_changes_pkt_data(env, off: t);
18243	if (insn->imm == BPF_FUNC_tail_call)
18244	visit_tailcall_insn(env, t);
18245	} else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
18246	struct bpf_kfunc_call_arg_meta meta;
18247
18248	ret = fetch_kfunc_meta(env, insn, meta: &meta, NULL);
18249	if (ret == 0 && is_iter_next_kfunc(meta: &meta)) {
18250	mark_prune_point(env, idx: t);
18251	/* Checking and saving state checkpoints at iter_next() call
18252	* is crucial for fast convergence of open-coded iterator loop
18253	* logic, so we need to force it. If we don't do that,
18254	* is_state_visited() might skip saving a checkpoint, causing
18255	* unnecessarily long sequence of not checkpointed
18256	* instructions and jumps, leading to exhaustion of jump
18257	* history buffer, and potentially other undesired outcomes.
18258	* It is expected that with correct open-coded iterators
18259	* convergence will happen quickly, so we don't run a risk of
18260	* exhausting memory.
18261	*/
18262	mark_force_checkpoint(env, idx: t);
18263	}
18264	/* Same as helpers, if called in a non-sleepable context
18265	* program will be rejected anyway, so we should end up
18266	* with precise sleepable marks on subprogs, except for
18267	* dead code elimination.
18268	*/
18269	if (ret == 0 && is_kfunc_sleepable(meta: &meta))
18270	mark_subprog_might_sleep(env, off: t);
18271	if (ret == 0 && is_kfunc_pkt_changing(meta: &meta))
18272	mark_subprog_changes_pkt_data(env, off: t);
18273	}
18274	return visit_func_call_insn(t, insns, env, visit_callee: insn->src_reg == BPF_PSEUDO_CALL);
18275
18276	case BPF_JA:
18277	if (BPF_SRC(insn->code) == BPF_X)
18278	return visit_gotox_insn(t, env);
18279
18280	if (BPF_CLASS(insn->code) == BPF_JMP)
18281	off = insn->off;
18282	else
18283	off = insn->imm;
18284
18285	/* unconditional jump with single edge */
18286	ret = push_insn(t, w: t + off + 1, e: FALLTHROUGH, env);
18287	if (ret)
18288	return ret;
18289
18290	mark_prune_point(env, idx: t + off + 1);
18291	mark_jmp_point(env, idx: t + off + 1);
18292
18293	return ret;
18294
18295	default:
18296	/* conditional jump with two edges */
18297	mark_prune_point(env, idx: t);
18298	if (is_may_goto_insn(insn))
18299	mark_force_checkpoint(env, idx: t);
18300
18301	ret = push_insn(t, w: t + 1, e: FALLTHROUGH, env);
18302	if (ret)
18303	return ret;
18304
18305	return push_insn(t, w: t + insn->off + 1, e: BRANCH, env);
18306	}
18307	}
18308
18309	/* non-recursive depth-first-search to detect loops in BPF program
18310	* loop == back-edge in directed graph
18311	*/
18312	static int check_cfg(struct bpf_verifier_env *env)
18313	{
18314	int insn_cnt = env->prog->len;
18315	int *insn_stack, *insn_state;
18316	int ex_insn_beg, i, ret = 0;
18317
18318	insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL_ACCOUNT);
18319	if (!insn_state)
18320	return -ENOMEM;
18321
18322	insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL_ACCOUNT);
18323	if (!insn_stack) {
18324	kvfree(addr: insn_state);
18325	return -ENOMEM;
18326	}
18327
18328	ex_insn_beg = env->exception_callback_subprog
18329	? env->subprog_info[env->exception_callback_subprog].start
18330	: 0;
18331
18332	insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
18333	insn_stack[0] = 0; /* 0 is the first instruction */
18334	env->cfg.cur_stack = 1;
18335
18336	walk_cfg:
18337	while (env->cfg.cur_stack > 0) {
18338	int t = insn_stack[env->cfg.cur_stack - 1];
18339
18340	ret = visit_insn(t, env);
18341	switch (ret) {
18342	case DONE_EXPLORING:
18343	insn_state[t] = EXPLORED;
18344	env->cfg.cur_stack--;
18345	break;
18346	case KEEP_EXPLORING:
18347	break;
18348	default:
18349	if (ret > 0) {
18350	verifier_bug(env, "visit_insn internal bug");
18351	ret = -EFAULT;
18352	}
18353	goto err_free;
18354	}
18355	}
18356
18357	if (env->cfg.cur_stack < 0) {
18358	verifier_bug(env, "pop stack internal bug");
18359	ret = -EFAULT;
18360	goto err_free;
18361	}
18362
18363	if (ex_insn_beg && insn_state[ex_insn_beg] != EXPLORED) {
18364	insn_state[ex_insn_beg] = DISCOVERED;
18365	insn_stack[0] = ex_insn_beg;
18366	env->cfg.cur_stack = 1;
18367	goto walk_cfg;
18368	}
18369
18370	for (i = 0; i < insn_cnt; i++) {
18371	struct bpf_insn *insn = &env->prog->insnsi[i];
18372
18373	if (insn_state[i] != EXPLORED) {
18374	verbose(private_data: env, fmt: "unreachable insn %d\n", i);
18375	ret = -EINVAL;
18376	goto err_free;
18377	}
18378	if (bpf_is_ldimm64(insn)) {
18379	if (insn_state[i + 1] != 0) {
18380	verbose(private_data: env, fmt: "jump into the middle of ldimm64 insn %d\n", i);
18381	ret = -EINVAL;
18382	goto err_free;
18383	}
18384	i++; /* skip second half of ldimm64 */
18385	}
18386	}
18387	ret = 0; /* cfg looks good */
18388	env->prog->aux->changes_pkt_data = env->subprog_info[0].changes_pkt_data;
18389	env->prog->aux->might_sleep = env->subprog_info[0].might_sleep;
18390
18391	err_free:
18392	kvfree(addr: insn_state);
18393	kvfree(addr: insn_stack);
18394	env->cfg.insn_state = env->cfg.insn_stack = NULL;
18395	return ret;
18396	}
18397
18398	/*
18399	* For each subprogram 'i' fill array env->cfg.insn_subprogram sub-range
18400	* [env->subprog_info[i].postorder_start, env->subprog_info[i+1].postorder_start)
18401	* with indices of 'i' instructions in postorder.
18402	*/
18403	static int compute_postorder(struct bpf_verifier_env *env)
18404	{
18405	u32 cur_postorder, i, top, stack_sz, s;
18406	int *stack = NULL, *postorder = NULL, *state = NULL;
18407	struct bpf_iarray *succ;
18408
18409	postorder = kvcalloc(env->prog->len, sizeof(int), GFP_KERNEL_ACCOUNT);
18410	state = kvcalloc(env->prog->len, sizeof(int), GFP_KERNEL_ACCOUNT);
18411	stack = kvcalloc(env->prog->len, sizeof(int), GFP_KERNEL_ACCOUNT);
18412	if (!postorder || !state || !stack) {
18413	kvfree(addr: postorder);
18414	kvfree(addr: state);
18415	kvfree(addr: stack);
18416	return -ENOMEM;
18417	}
18418	cur_postorder = 0;
18419	for (i = 0; i < env->subprog_cnt; i++) {
18420	env->subprog_info[i].postorder_start = cur_postorder;
18421	stack[0] = env->subprog_info[i].start;
18422	stack_sz = 1;
18423	do {
18424	top = stack[stack_sz - 1];
18425	state[top] |= DISCOVERED;
18426	if (state[top] & EXPLORED) {
18427	postorder[cur_postorder++] = top;
18428	stack_sz--;
18429	continue;
18430	}
18431	succ = bpf_insn_successors(env, idx: top);
18432	for (s = 0; s < succ->cnt; ++s) {
18433	if (!state[succ->items[s]]) {
18434	stack[stack_sz++] = succ->items[s];
18435	state[succ->items[s]] |= DISCOVERED;
18436	}
18437	}
18438	state[top] |= EXPLORED;
18439	} while (stack_sz);
18440	}
18441	env->subprog_info[i].postorder_start = cur_postorder;
18442	env->cfg.insn_postorder = postorder;
18443	env->cfg.cur_postorder = cur_postorder;
18444	kvfree(addr: stack);
18445	kvfree(addr: state);
18446	return 0;
18447	}
18448
18449	static int check_abnormal_return(struct bpf_verifier_env *env)
18450	{
18451	int i;
18452
18453	for (i = 1; i < env->subprog_cnt; i++) {
18454	if (env->subprog_info[i].has_ld_abs) {
18455	verbose(private_data: env, fmt: "LD_ABS is not allowed in subprogs without BTF\n");
18456	return -EINVAL;
18457	}
18458	if (env->subprog_info[i].has_tail_call) {
18459	verbose(private_data: env, fmt: "tail_call is not allowed in subprogs without BTF\n");
18460	return -EINVAL;
18461	}
18462	}
18463	return 0;
18464	}
18465
18466	/* The minimum supported BTF func info size */
18467	#define MIN_BPF_FUNCINFO_SIZE 8
18468	#define MAX_FUNCINFO_REC_SIZE 252
18469
18470	static int check_btf_func_early(struct bpf_verifier_env *env,
18471	const union bpf_attr *attr,
18472	bpfptr_t uattr)
18473	{
18474	u32 krec_size = sizeof(struct bpf_func_info);
18475	const struct btf_type *type, *func_proto;
18476	u32 i, nfuncs, urec_size, min_size;
18477	struct bpf_func_info *krecord;
18478	struct bpf_prog *prog;
18479	const struct btf *btf;
18480	u32 prev_offset = 0;
18481	bpfptr_t urecord;
18482	int ret = -ENOMEM;
18483
18484	nfuncs = attr->func_info_cnt;
18485	if (!nfuncs) {
18486	if (check_abnormal_return(env))
18487	return -EINVAL;
18488	return 0;
18489	}
18490
18491	urec_size = attr->func_info_rec_size;
18492	if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
18493	urec_size > MAX_FUNCINFO_REC_SIZE ||
18494	urec_size % sizeof(u32)) {
18495	verbose(private_data: env, fmt: "invalid func info rec size %u\n", urec_size);
18496	return -EINVAL;
18497	}
18498
18499	prog = env->prog;
18500	btf = prog->aux->btf;
18501
18502	urecord = make_bpfptr(addr: attr->func_info, is_kernel: uattr.is_kernel);
18503	min_size = min_t(u32, krec_size, urec_size);
18504
18505	krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
18506	if (!krecord)
18507	return -ENOMEM;
18508
18509	for (i = 0; i < nfuncs; i++) {
18510	ret = bpf_check_uarg_tail_zero(uaddr: urecord, expected_size: krec_size, actual_size: urec_size);
18511	if (ret) {
18512	if (ret == -E2BIG) {
18513	verbose(private_data: env, fmt: "nonzero tailing record in func info");
18514	/* set the size kernel expects so loader can zero
18515	* out the rest of the record.
18516	*/
18517	if (copy_to_bpfptr_offset(dst: uattr,
18518	offsetof(union bpf_attr, func_info_rec_size),
18519	src: &min_size, size: sizeof(min_size)))
18520	ret = -EFAULT;
18521	}
18522	goto err_free;
18523	}
18524
18525	if (copy_from_bpfptr(dst: &krecord[i], src: urecord, size: min_size)) {
18526	ret = -EFAULT;
18527	goto err_free;
18528	}
18529
18530	/* check insn_off */
18531	ret = -EINVAL;
18532	if (i == 0) {
18533	if (krecord[i].insn_off) {
18534	verbose(private_data: env,
18535	fmt: "nonzero insn_off %u for the first func info record",
18536	krecord[i].insn_off);
18537	goto err_free;
18538	}
18539	} else if (krecord[i].insn_off <= prev_offset) {
18540	verbose(private_data: env,
18541	fmt: "same or smaller insn offset (%u) than previous func info record (%u)",
18542	krecord[i].insn_off, prev_offset);
18543	goto err_free;
18544	}
18545
18546	/* check type_id */
18547	type = btf_type_by_id(btf, type_id: krecord[i].type_id);
18548	if (!type || !btf_type_is_func(t: type)) {
18549	verbose(private_data: env, fmt: "invalid type id %d in func info",
18550	krecord[i].type_id);
18551	goto err_free;
18552	}
18553
18554	func_proto = btf_type_by_id(btf, type_id: type->type);
18555	if (unlikely(!func_proto || !btf_type_is_func_proto(func_proto)))
18556	/* btf_func_check() already verified it during BTF load */
18557	goto err_free;
18558
18559	prev_offset = krecord[i].insn_off;
18560	bpfptr_add(bpfptr: &urecord, val: urec_size);
18561	}
18562
18563	prog->aux->func_info = krecord;
18564	prog->aux->func_info_cnt = nfuncs;
18565	return 0;
18566
18567	err_free:
18568	kvfree(addr: krecord);
18569	return ret;
18570	}
18571
18572	static int check_btf_func(struct bpf_verifier_env *env,
18573	const union bpf_attr *attr,
18574	bpfptr_t uattr)
18575	{
18576	const struct btf_type *type, *func_proto, *ret_type;
18577	u32 i, nfuncs, urec_size;
18578	struct bpf_func_info *krecord;
18579	struct bpf_func_info_aux *info_aux = NULL;
18580	struct bpf_prog *prog;
18581	const struct btf *btf;
18582	bpfptr_t urecord;
18583	bool scalar_return;
18584	int ret = -ENOMEM;
18585
18586	nfuncs = attr->func_info_cnt;
18587	if (!nfuncs) {
18588	if (check_abnormal_return(env))
18589	return -EINVAL;
18590	return 0;
18591	}
18592	if (nfuncs != env->subprog_cnt) {
18593	verbose(private_data: env, fmt: "number of funcs in func_info doesn't match number of subprogs\n");
18594	return -EINVAL;
18595	}
18596
18597	urec_size = attr->func_info_rec_size;
18598
18599	prog = env->prog;
18600	btf = prog->aux->btf;
18601
18602	urecord = make_bpfptr(addr: attr->func_info, is_kernel: uattr.is_kernel);
18603
18604	krecord = prog->aux->func_info;
18605	info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
18606	if (!info_aux)
18607	return -ENOMEM;
18608
18609	for (i = 0; i < nfuncs; i++) {
18610	/* check insn_off */
18611	ret = -EINVAL;
18612
18613	if (env->subprog_info[i].start != krecord[i].insn_off) {
18614	verbose(private_data: env, fmt: "func_info BTF section doesn't match subprog layout in BPF program\n");
18615	goto err_free;
18616	}
18617
18618	/* Already checked type_id */
18619	type = btf_type_by_id(btf, type_id: krecord[i].type_id);
18620	info_aux[i].linkage = BTF_INFO_VLEN(type->info);
18621	/* Already checked func_proto */
18622	func_proto = btf_type_by_id(btf, type_id: type->type);
18623
18624	ret_type = btf_type_skip_modifiers(btf, id: func_proto->type, NULL);
18625	scalar_return =
18626	btf_type_is_small_int(t: ret_type) || btf_is_any_enum(t: ret_type);
18627	if (i && !scalar_return && env->subprog_info[i].has_ld_abs) {
18628	verbose(private_data: env, fmt: "LD_ABS is only allowed in functions that return 'int'.\n");
18629	goto err_free;
18630	}
18631	if (i && !scalar_return && env->subprog_info[i].has_tail_call) {
18632	verbose(private_data: env, fmt: "tail_call is only allowed in functions that return 'int'.\n");
18633	goto err_free;
18634	}
18635
18636	bpfptr_add(bpfptr: &urecord, val: urec_size);
18637	}
18638
18639	prog->aux->func_info_aux = info_aux;
18640	return 0;
18641
18642	err_free:
18643	kfree(objp: info_aux);
18644	return ret;
18645	}
18646
18647	static void adjust_btf_func(struct bpf_verifier_env *env)
18648	{
18649	struct bpf_prog_aux *aux = env->prog->aux;
18650	int i;
18651
18652	if (!aux->func_info)
18653	return;
18654
18655	/* func_info is not available for hidden subprogs */
18656	for (i = 0; i < env->subprog_cnt - env->hidden_subprog_cnt; i++)
18657	aux->func_info[i].insn_off = env->subprog_info[i].start;
18658	}
18659
18660	#define MIN_BPF_LINEINFO_SIZE offsetofend(struct bpf_line_info, line_col)
18661	#define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
18662
18663	static int check_btf_line(struct bpf_verifier_env *env,
18664	const union bpf_attr *attr,
18665	bpfptr_t uattr)
18666	{
18667	u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
18668	struct bpf_subprog_info *sub;
18669	struct bpf_line_info *linfo;
18670	struct bpf_prog *prog;
18671	const struct btf *btf;
18672	bpfptr_t ulinfo;
18673	int err;
18674
18675	nr_linfo = attr->line_info_cnt;
18676	if (!nr_linfo)
18677	return 0;
18678	if (nr_linfo > INT_MAX / sizeof(struct bpf_line_info))
18679	return -EINVAL;
18680
18681	rec_size = attr->line_info_rec_size;
18682	if (rec_size < MIN_BPF_LINEINFO_SIZE ||
18683	rec_size > MAX_LINEINFO_REC_SIZE ||
18684	rec_size & (sizeof(u32) - 1))
18685	return -EINVAL;
18686
18687	/* Need to zero it in case the userspace may
18688	* pass in a smaller bpf_line_info object.
18689	*/
18690	linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
18691	GFP_KERNEL_ACCOUNT | __GFP_NOWARN);
18692	if (!linfo)
18693	return -ENOMEM;
18694
18695	prog = env->prog;
18696	btf = prog->aux->btf;
18697
18698	s = 0;
18699	sub = env->subprog_info;
18700	ulinfo = make_bpfptr(addr: attr->line_info, is_kernel: uattr.is_kernel);
18701	expected_size = sizeof(struct bpf_line_info);
18702	ncopy = min_t(u32, expected_size, rec_size);
18703	for (i = 0; i < nr_linfo; i++) {
18704	err = bpf_check_uarg_tail_zero(uaddr: ulinfo, expected_size, actual_size: rec_size);
18705	if (err) {
18706	if (err == -E2BIG) {
18707	verbose(private_data: env, fmt: "nonzero tailing record in line_info");
18708	if (copy_to_bpfptr_offset(dst: uattr,
18709	offsetof(union bpf_attr, line_info_rec_size),
18710	src: &expected_size, size: sizeof(expected_size)))
18711	err = -EFAULT;
18712	}
18713	goto err_free;
18714	}
18715
18716	if (copy_from_bpfptr(dst: &linfo[i], src: ulinfo, size: ncopy)) {
18717	err = -EFAULT;
18718	goto err_free;
18719	}
18720
18721	/*
18722	* Check insn_off to ensure
18723	* 1) strictly increasing AND
18724	* 2) bounded by prog->len
18725	*
18726	* The linfo[0].insn_off == 0 check logically falls into
18727	* the later "missing bpf_line_info for func..." case
18728	* because the first linfo[0].insn_off must be the
18729	* first sub also and the first sub must have
18730	* subprog_info[0].start == 0.
18731	*/
18732	if ((i && linfo[i].insn_off <= prev_offset) ||
18733	linfo[i].insn_off >= prog->len) {
18734	verbose(private_data: env, fmt: "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
18735	i, linfo[i].insn_off, prev_offset,
18736	prog->len);
18737	err = -EINVAL;
18738	goto err_free;
18739	}
18740
18741	if (!prog->insnsi[linfo[i].insn_off].code) {
18742	verbose(private_data: env,
18743	fmt: "Invalid insn code at line_info[%u].insn_off\n",
18744	i);
18745	err = -EINVAL;
18746	goto err_free;
18747	}
18748
18749	if (!btf_name_by_offset(btf, offset: linfo[i].line_off) ||
18750	!btf_name_by_offset(btf, offset: linfo[i].file_name_off)) {
18751	verbose(private_data: env, fmt: "Invalid line_info[%u].line_off or .file_name_off\n", i);
18752	err = -EINVAL;
18753	goto err_free;
18754	}
18755
18756	if (s != env->subprog_cnt) {
18757	if (linfo[i].insn_off == sub[s].start) {
18758	sub[s].linfo_idx = i;
18759	s++;
18760	} else if (sub[s].start < linfo[i].insn_off) {
18761	verbose(private_data: env, fmt: "missing bpf_line_info for func#%u\n", s);
18762	err = -EINVAL;
18763	goto err_free;
18764	}
18765	}
18766
18767	prev_offset = linfo[i].insn_off;
18768	bpfptr_add(bpfptr: &ulinfo, val: rec_size);
18769	}
18770
18771	if (s != env->subprog_cnt) {
18772	verbose(private_data: env, fmt: "missing bpf_line_info for %u funcs starting from func#%u\n",
18773	env->subprog_cnt - s, s);
18774	err = -EINVAL;
18775	goto err_free;
18776	}
18777
18778	prog->aux->linfo = linfo;
18779	prog->aux->nr_linfo = nr_linfo;
18780
18781	return 0;
18782
18783	err_free:
18784	kvfree(addr: linfo);
18785	return err;
18786	}
18787
18788	#define MIN_CORE_RELO_SIZE sizeof(struct bpf_core_relo)
18789	#define MAX_CORE_RELO_SIZE MAX_FUNCINFO_REC_SIZE
18790
18791	static int check_core_relo(struct bpf_verifier_env *env,
18792	const union bpf_attr *attr,
18793	bpfptr_t uattr)
18794	{
18795	u32 i, nr_core_relo, ncopy, expected_size, rec_size;
18796	struct bpf_core_relo core_relo = {};
18797	struct bpf_prog *prog = env->prog;
18798	const struct btf *btf = prog->aux->btf;
18799	struct bpf_core_ctx ctx = {
18800	.log = &env->log,
18801	.btf = btf,
18802	};
18803	bpfptr_t u_core_relo;
18804	int err;
18805
18806	nr_core_relo = attr->core_relo_cnt;
18807	if (!nr_core_relo)
18808	return 0;
18809	if (nr_core_relo > INT_MAX / sizeof(struct bpf_core_relo))
18810	return -EINVAL;
18811
18812	rec_size = attr->core_relo_rec_size;
18813	if (rec_size < MIN_CORE_RELO_SIZE ||
18814	rec_size > MAX_CORE_RELO_SIZE ||
18815	rec_size % sizeof(u32))
18816	return -EINVAL;
18817
18818	u_core_relo = make_bpfptr(addr: attr->core_relos, is_kernel: uattr.is_kernel);
18819	expected_size = sizeof(struct bpf_core_relo);
18820	ncopy = min_t(u32, expected_size, rec_size);
18821
18822	/* Unlike func_info and line_info, copy and apply each CO-RE
18823	* relocation record one at a time.
18824	*/
18825	for (i = 0; i < nr_core_relo; i++) {
18826	/* future proofing when sizeof(bpf_core_relo) changes */
18827	err = bpf_check_uarg_tail_zero(uaddr: u_core_relo, expected_size, actual_size: rec_size);
18828	if (err) {
18829	if (err == -E2BIG) {
18830	verbose(private_data: env, fmt: "nonzero tailing record in core_relo");
18831	if (copy_to_bpfptr_offset(dst: uattr,
18832	offsetof(union bpf_attr, core_relo_rec_size),
18833	src: &expected_size, size: sizeof(expected_size)))
18834	err = -EFAULT;
18835	}
18836	break;
18837	}
18838
18839	if (copy_from_bpfptr(dst: &core_relo, src: u_core_relo, size: ncopy)) {
18840	err = -EFAULT;
18841	break;
18842	}
18843
18844	if (core_relo.insn_off % 8 || core_relo.insn_off / 8 >= prog->len) {
18845	verbose(private_data: env, fmt: "Invalid core_relo[%u].insn_off:%u prog->len:%u\n",
18846	i, core_relo.insn_off, prog->len);
18847	err = -EINVAL;
18848	break;
18849	}
18850
18851	err = bpf_core_apply(ctx: &ctx, relo: &core_relo, relo_idx: i,
18852	insn: &prog->insnsi[core_relo.insn_off / 8]);
18853	if (err)
18854	break;
18855	bpfptr_add(bpfptr: &u_core_relo, val: rec_size);
18856	}
18857	return err;
18858	}
18859
18860	static int check_btf_info_early(struct bpf_verifier_env *env,
18861	const union bpf_attr *attr,
18862	bpfptr_t uattr)
18863	{
18864	struct btf *btf;
18865	int err;
18866
18867	if (!attr->func_info_cnt && !attr->line_info_cnt) {
18868	if (check_abnormal_return(env))
18869	return -EINVAL;
18870	return 0;
18871	}
18872
18873	btf = btf_get_by_fd(fd: attr->prog_btf_fd);
18874	if (IS_ERR(ptr: btf))
18875	return PTR_ERR(ptr: btf);
18876	if (btf_is_kernel(btf)) {
18877	btf_put(btf);
18878	return -EACCES;
18879	}
18880	env->prog->aux->btf = btf;
18881
18882	err = check_btf_func_early(env, attr, uattr);
18883	if (err)
18884	return err;
18885	return 0;
18886	}
18887
18888	static int check_btf_info(struct bpf_verifier_env *env,
18889	const union bpf_attr *attr,
18890	bpfptr_t uattr)
18891	{
18892	int err;
18893
18894	if (!attr->func_info_cnt && !attr->line_info_cnt) {
18895	if (check_abnormal_return(env))
18896	return -EINVAL;
18897	return 0;
18898	}
18899
18900	err = check_btf_func(env, attr, uattr);
18901	if (err)
18902	return err;
18903
18904	err = check_btf_line(env, attr, uattr);
18905	if (err)
18906	return err;
18907
18908	err = check_core_relo(env, attr, uattr);
18909	if (err)
18910	return err;
18911
18912	return 0;
18913	}
18914
18915	/* check %cur's range satisfies %old's */
18916	static bool range_within(const struct bpf_reg_state *old,
18917	const struct bpf_reg_state *cur)
18918	{
18919	return old->umin_value <= cur->umin_value &&
18920	old->umax_value >= cur->umax_value &&
18921	old->smin_value <= cur->smin_value &&
18922	old->smax_value >= cur->smax_value &&
18923	old->u32_min_value <= cur->u32_min_value &&
18924	old->u32_max_value >= cur->u32_max_value &&
18925	old->s32_min_value <= cur->s32_min_value &&
18926	old->s32_max_value >= cur->s32_max_value;
18927	}
18928
18929	/* If in the old state two registers had the same id, then they need to have
18930	* the same id in the new state as well. But that id could be different from
18931	* the old state, so we need to track the mapping from old to new ids.
18932	* Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
18933	* regs with old id 5 must also have new id 9 for the new state to be safe. But
18934	* regs with a different old id could still have new id 9, we don't care about
18935	* that.
18936	* So we look through our idmap to see if this old id has been seen before. If
18937	* so, we require the new id to match; otherwise, we add the id pair to the map.
18938	*/
18939	static bool check_ids(u32 old_id, u32 cur_id, struct bpf_idmap *idmap)
18940	{
18941	struct bpf_id_pair *map = idmap->map;
18942	unsigned int i;
18943
18944	/* either both IDs should be set or both should be zero */
18945	if (!!old_id != !!cur_id)
18946	return false;
18947
18948	if (old_id == 0) /* cur_id == 0 as well */
18949	return true;
18950
18951	for (i = 0; i < BPF_ID_MAP_SIZE; i++) {
18952	if (!map[i].old) {
18953	/* Reached an empty slot; haven't seen this id before */
18954	map[i].old = old_id;
18955	map[i].cur = cur_id;
18956	return true;
18957	}
18958	if (map[i].old == old_id)
18959	return map[i].cur == cur_id;
18960	if (map[i].cur == cur_id)
18961	return false;
18962	}
18963	/* We ran out of idmap slots, which should be impossible */
18964	WARN_ON_ONCE(1);
18965	return false;
18966	}
18967
18968	/* Similar to check_ids(), but allocate a unique temporary ID
18969	* for 'old_id' or 'cur_id' of zero.
18970	* This makes pairs like '0 vs unique ID', 'unique ID vs 0' valid.
18971	*/
18972	static bool check_scalar_ids(u32 old_id, u32 cur_id, struct bpf_idmap *idmap)
18973	{
18974	old_id = old_id ? old_id : ++idmap->tmp_id_gen;
18975	cur_id = cur_id ? cur_id : ++idmap->tmp_id_gen;
18976
18977	return check_ids(old_id, cur_id, idmap);
18978	}
18979
18980	static void clean_func_state(struct bpf_verifier_env *env,
18981	struct bpf_func_state *st,
18982	u32 ip)
18983	{
18984	u16 live_regs = env->insn_aux_data[ip].live_regs_before;
18985	int i, j;
18986
18987	for (i = 0; i < BPF_REG_FP; i++) {
18988	/* liveness must not touch this register anymore */
18989	if (!(live_regs & BIT(i)))
18990	/* since the register is unused, clear its state
18991	* to make further comparison simpler
18992	*/
18993	__mark_reg_not_init(env, reg: &st->regs[i]);
18994	}
18995
18996	for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
18997	if (!bpf_stack_slot_alive(env, frameno: st->frameno, spi: i)) {
18998	__mark_reg_not_init(env, reg: &st->stack[i].spilled_ptr);
18999	for (j = 0; j < BPF_REG_SIZE; j++)
19000	st->stack[i].slot_type[j] = STACK_INVALID;
19001	}
19002	}
19003	}
19004
19005	static void clean_verifier_state(struct bpf_verifier_env *env,
19006	struct bpf_verifier_state *st)
19007	{
19008	int i, ip;
19009
19010	bpf_live_stack_query_init(env, st);
19011	st->cleaned = true;
19012	for (i = 0; i <= st->curframe; i++) {
19013	ip = frame_insn_idx(st, frame: i);
19014	clean_func_state(env, st: st->frame[i], ip);
19015	}
19016	}
19017
19018	/* the parentage chains form a tree.
19019	* the verifier states are added to state lists at given insn and
19020	* pushed into state stack for future exploration.
19021	* when the verifier reaches bpf_exit insn some of the verifier states
19022	* stored in the state lists have their final liveness state already,
19023	* but a lot of states will get revised from liveness point of view when
19024	* the verifier explores other branches.
19025	* Example:
19026	* 1: *(u64)(r10 - 8) = 1
19027	* 2: if r1 == 100 goto pc+1
19028	* 3: *(u64)(r10 - 8) = 2
19029	* 4: r0 = *(u64)(r10 - 8)
19030	* 5: exit
19031	* when the verifier reaches exit insn the stack slot -8 in the state list of
19032	* insn 2 is not yet marked alive. Then the verifier pops the other_branch
19033	* of insn 2 and goes exploring further. After the insn 4 read, liveness
19034	* analysis would propagate read mark for -8 at insn 2.
19035	*
19036	* Since the verifier pushes the branch states as it sees them while exploring
19037	* the program the condition of walking the branch instruction for the second
19038	* time means that all states below this branch were already explored and
19039	* their final liveness marks are already propagated.
19040	* Hence when the verifier completes the search of state list in is_state_visited()
19041	* we can call this clean_live_states() function to clear dead the registers and stack
19042	* slots to simplify state merging.
19043	*
19044	* Important note here that walking the same branch instruction in the callee
19045	* doesn't meant that the states are DONE. The verifier has to compare
19046	* the callsites
19047	*/
19048	static void clean_live_states(struct bpf_verifier_env *env, int insn,
19049	struct bpf_verifier_state *cur)
19050	{
19051	struct bpf_verifier_state_list *sl;
19052	struct list_head *pos, *head;
19053
19054	head = explored_state(env, idx: insn);
19055	list_for_each(pos, head) {
19056	sl = container_of(pos, struct bpf_verifier_state_list, node);
19057	if (sl->state.branches)
19058	continue;
19059	if (sl->state.insn_idx != insn ||
19060	!same_callsites(a: &sl->state, b: cur))
19061	continue;
19062	if (sl->state.cleaned)
19063	/* all regs in this state in all frames were already marked */
19064	continue;
19065	if (incomplete_read_marks(env, st: &sl->state))
19066	continue;
19067	clean_verifier_state(env, st: &sl->state);
19068	}
19069	}
19070
19071	static bool regs_exact(const struct bpf_reg_state *rold,
19072	const struct bpf_reg_state *rcur,
19073	struct bpf_idmap *idmap)
19074	{
19075	return memcmp(p: rold, q: rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
19076	check_ids(old_id: rold->id, cur_id: rcur->id, idmap) &&
19077	check_ids(old_id: rold->ref_obj_id, cur_id: rcur->ref_obj_id, idmap);
19078	}
19079
19080	enum exact_level {
19081	NOT_EXACT,
19082	EXACT,
19083	RANGE_WITHIN
19084	};
19085
19086	/* Returns true if (rold safe implies rcur safe) */
19087	static bool regsafe(struct bpf_verifier_env *env, struct bpf_reg_state *rold,
19088	struct bpf_reg_state *rcur, struct bpf_idmap *idmap,
19089	enum exact_level exact)
19090	{
19091	if (exact == EXACT)
19092	return regs_exact(rold, rcur, idmap);
19093
19094	if (rold->type == NOT_INIT) {
19095	if (exact == NOT_EXACT || rcur->type == NOT_INIT)
19096	/* explored state can't have used this */
19097	return true;
19098	}
19099
19100	/* Enforce that register types have to match exactly, including their
19101	* modifiers (like PTR_MAYBE_NULL, MEM_RDONLY, etc), as a general
19102	* rule.
19103	*
19104	* One can make a point that using a pointer register as unbounded
19105	* SCALAR would be technically acceptable, but this could lead to
19106	* pointer leaks because scalars are allowed to leak while pointers
19107	* are not. We could make this safe in special cases if root is
19108	* calling us, but it's probably not worth the hassle.
19109	*
19110	* Also, register types that are *not* MAYBE_NULL could technically be
19111	* safe to use as their MAYBE_NULL variants (e.g., PTR_TO_MAP_VALUE
19112	* is safe to be used as PTR_TO_MAP_VALUE_OR_NULL, provided both point
19113	* to the same map).
19114	* However, if the old MAYBE_NULL register then got NULL checked,
19115	* doing so could have affected others with the same id, and we can't
19116	* check for that because we lost the id when we converted to
19117	* a non-MAYBE_NULL variant.
19118	* So, as a general rule we don't allow mixing MAYBE_NULL and
19119	* non-MAYBE_NULL registers as well.
19120	*/
19121	if (rold->type != rcur->type)
19122	return false;
19123
19124	switch (base_type(type: rold->type)) {
19125	case SCALAR_VALUE:
19126	if (env->explore_alu_limits) {
19127	/* explore_alu_limits disables tnum_in() and range_within()
19128	* logic and requires everything to be strict
19129	*/
19130	return memcmp(p: rold, q: rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
19131	check_scalar_ids(old_id: rold->id, cur_id: rcur->id, idmap);
19132	}
19133	if (!rold->precise && exact == NOT_EXACT)
19134	return true;
19135	if ((rold->id & BPF_ADD_CONST) != (rcur->id & BPF_ADD_CONST))
19136	return false;
19137	if ((rold->id & BPF_ADD_CONST) && (rold->off != rcur->off))
19138	return false;
19139	/* Why check_ids() for scalar registers?
19140	*
19141	* Consider the following BPF code:
19142	* 1: r6 = ... unbound scalar, ID=a ...
19143	* 2: r7 = ... unbound scalar, ID=b ...
19144	* 3: if (r6 > r7) goto +1
19145	* 4: r6 = r7
19146	* 5: if (r6 > X) goto ...
19147	* 6: ... memory operation using r7 ...
19148	*
19149	* First verification path is [1-6]:
19150	* - at (4) same bpf_reg_state::id (b) would be assigned to r6 and r7;
19151	* - at (5) r6 would be marked <= X, sync_linked_regs() would also mark
19152	* r7 <= X, because r6 and r7 share same id.
19153	* Next verification path is [1-4, 6].
19154	*
19155	* Instruction (6) would be reached in two states:
19156	* I. r6{.id=b}, r7{.id=b} via path 1-6;
19157	* II. r6{.id=a}, r7{.id=b} via path 1-4, 6.
19158	*
19159	* Use check_ids() to distinguish these states.
19160	* ---
19161	* Also verify that new value satisfies old value range knowledge.
19162	*/
19163	return range_within(old: rold, cur: rcur) &&
19164	tnum_in(a: rold->var_off, b: rcur->var_off) &&
19165	check_scalar_ids(old_id: rold->id, cur_id: rcur->id, idmap);
19166	case PTR_TO_MAP_KEY:
19167	case PTR_TO_MAP_VALUE:
19168	case PTR_TO_MEM:
19169	case PTR_TO_BUF:
19170	case PTR_TO_TP_BUFFER:
19171	/* If the new min/max/var_off satisfy the old ones and
19172	* everything else matches, we are OK.
19173	*/
19174	return memcmp(p: rold, q: rcur, offsetof(struct bpf_reg_state, var_off)) == 0 &&
19175	range_within(old: rold, cur: rcur) &&
19176	tnum_in(a: rold->var_off, b: rcur->var_off) &&
19177	check_ids(old_id: rold->id, cur_id: rcur->id, idmap) &&
19178	check_ids(old_id: rold->ref_obj_id, cur_id: rcur->ref_obj_id, idmap);
19179	case PTR_TO_PACKET_META:
19180	case PTR_TO_PACKET:
19181	/* We must have at least as much range as the old ptr
19182	* did, so that any accesses which were safe before are
19183	* still safe. This is true even if old range < old off,
19184	* since someone could have accessed through (ptr - k), or
19185	* even done ptr -= k in a register, to get a safe access.
19186	*/
19187	if (rold->range > rcur->range)
19188	return false;
19189	/* If the offsets don't match, we can't trust our alignment;
19190	* nor can we be sure that we won't fall out of range.
19191	*/
19192	if (rold->off != rcur->off)
19193	return false;
19194	/* id relations must be preserved */
19195	if (!check_ids(old_id: rold->id, cur_id: rcur->id, idmap))
19196	return false;
19197	/* new val must satisfy old val knowledge */
19198	return range_within(old: rold, cur: rcur) &&
19199	tnum_in(a: rold->var_off, b: rcur->var_off);
19200	case PTR_TO_STACK:
19201	/* two stack pointers are equal only if they're pointing to
19202	* the same stack frame, since fp-8 in foo != fp-8 in bar
19203	*/
19204	return regs_exact(rold, rcur, idmap) && rold->frameno == rcur->frameno;
19205	case PTR_TO_ARENA:
19206	return true;
19207	case PTR_TO_INSN:
19208	return memcmp(p: rold, q: rcur, offsetof(struct bpf_reg_state, var_off)) == 0 &&
19209	rold->off == rcur->off && range_within(old: rold, cur: rcur) &&
19210	tnum_in(a: rold->var_off, b: rcur->var_off);
19211	default:
19212	return regs_exact(rold, rcur, idmap);
19213	}
19214	}
19215
19216	static struct bpf_reg_state unbound_reg;
19217
19218	static __init int unbound_reg_init(void)
19219	{
19220	__mark_reg_unknown_imprecise(reg: &unbound_reg);
19221	return 0;
19222	}
19223	late_initcall(unbound_reg_init);
19224
19225	static bool is_stack_all_misc(struct bpf_verifier_env *env,
19226	struct bpf_stack_state *stack)
19227	{
19228	u32 i;
19229
19230	for (i = 0; i < ARRAY_SIZE(stack->slot_type); ++i) {
19231	if ((stack->slot_type[i] == STACK_MISC) ||
19232	(stack->slot_type[i] == STACK_INVALID && env->allow_uninit_stack))
19233	continue;
19234	return false;
19235	}
19236
19237	return true;
19238	}
19239
19240	static struct bpf_reg_state *scalar_reg_for_stack(struct bpf_verifier_env *env,
19241	struct bpf_stack_state *stack)
19242	{
19243	if (is_spilled_scalar_reg64(stack))
19244	return &stack->spilled_ptr;
19245
19246	if (is_stack_all_misc(env, stack))
19247	return &unbound_reg;
19248
19249	return NULL;
19250	}
19251
19252	static bool stacksafe(struct bpf_verifier_env *env, struct bpf_func_state *old,
19253	struct bpf_func_state *cur, struct bpf_idmap *idmap,
19254	enum exact_level exact)
19255	{
19256	int i, spi;
19257
19258	/* walk slots of the explored stack and ignore any additional
19259	* slots in the current stack, since explored(safe) state
19260	* didn't use them
19261	*/
19262	for (i = 0; i < old->allocated_stack; i++) {
19263	struct bpf_reg_state *old_reg, *cur_reg;
19264
19265	spi = i / BPF_REG_SIZE;
19266
19267	if (exact != NOT_EXACT &&
19268	(i >= cur->allocated_stack ||
19269	old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
19270	cur->stack[spi].slot_type[i % BPF_REG_SIZE]))
19271	return false;
19272
19273	if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
19274	continue;
19275
19276	if (env->allow_uninit_stack &&
19277	old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC)
19278	continue;
19279
19280	/* explored stack has more populated slots than current stack
19281	* and these slots were used
19282	*/
19283	if (i >= cur->allocated_stack)
19284	return false;
19285
19286	/* 64-bit scalar spill vs all slots MISC and vice versa.
19287	* Load from all slots MISC produces unbound scalar.
19288	* Construct a fake register for such stack and call
19289	* regsafe() to ensure scalar ids are compared.
19290	*/
19291	old_reg = scalar_reg_for_stack(env, stack: &old->stack[spi]);
19292	cur_reg = scalar_reg_for_stack(env, stack: &cur->stack[spi]);
19293	if (old_reg && cur_reg) {
19294	if (!regsafe(env, rold: old_reg, rcur: cur_reg, idmap, exact))
19295	return false;
19296	i += BPF_REG_SIZE - 1;
19297	continue;
19298	}
19299
19300	/* if old state was safe with misc data in the stack
19301	* it will be safe with zero-initialized stack.
19302	* The opposite is not true
19303	*/
19304	if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
19305	cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
19306	continue;
19307	if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
19308	cur->stack[spi].slot_type[i % BPF_REG_SIZE])
19309	/* Ex: old explored (safe) state has STACK_SPILL in
19310	* this stack slot, but current has STACK_MISC ->
19311	* this verifier states are not equivalent,
19312	* return false to continue verification of this path
19313	*/
19314	return false;
19315	if (i % BPF_REG_SIZE != BPF_REG_SIZE - 1)
19316	continue;
19317	/* Both old and cur are having same slot_type */
19318	switch (old->stack[spi].slot_type[BPF_REG_SIZE - 1]) {
19319	case STACK_SPILL:
19320	/* when explored and current stack slot are both storing
19321	* spilled registers, check that stored pointers types
19322	* are the same as well.
19323	* Ex: explored safe path could have stored
19324	* (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
19325	* but current path has stored:
19326	* (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
19327	* such verifier states are not equivalent.
19328	* return false to continue verification of this path
19329	*/
19330	if (!regsafe(env, rold: &old->stack[spi].spilled_ptr,
19331	rcur: &cur->stack[spi].spilled_ptr, idmap, exact))
19332	return false;
19333	break;
19334	case STACK_DYNPTR:
19335	old_reg = &old->stack[spi].spilled_ptr;
19336	cur_reg = &cur->stack[spi].spilled_ptr;
19337	if (old_reg->dynptr.type != cur_reg->dynptr.type ||
19338	old_reg->dynptr.first_slot != cur_reg->dynptr.first_slot ||
19339	!check_ids(old_id: old_reg->ref_obj_id, cur_id: cur_reg->ref_obj_id, idmap))
19340	return false;
19341	break;
19342	case STACK_ITER:
19343	old_reg = &old->stack[spi].spilled_ptr;
19344	cur_reg = &cur->stack[spi].spilled_ptr;
19345	/* iter.depth is not compared between states as it
19346	* doesn't matter for correctness and would otherwise
19347	* prevent convergence; we maintain it only to prevent
19348	* infinite loop check triggering, see
19349	* iter_active_depths_differ()
19350	*/
19351	if (old_reg->iter.btf != cur_reg->iter.btf ||
19352	old_reg->iter.btf_id != cur_reg->iter.btf_id ||
19353	old_reg->iter.state != cur_reg->iter.state ||
19354	/* ignore {old_reg,cur_reg}->iter.depth, see above */
19355	!check_ids(old_id: old_reg->ref_obj_id, cur_id: cur_reg->ref_obj_id, idmap))
19356	return false;
19357	break;
19358	case STACK_IRQ_FLAG:
19359	old_reg = &old->stack[spi].spilled_ptr;
19360	cur_reg = &cur->stack[spi].spilled_ptr;
19361	if (!check_ids(old_id: old_reg->ref_obj_id, cur_id: cur_reg->ref_obj_id, idmap) ||
19362	old_reg->irq.kfunc_class != cur_reg->irq.kfunc_class)
19363	return false;
19364	break;
19365	case STACK_MISC:
19366	case STACK_ZERO:
19367	case STACK_INVALID:
19368	continue;
19369	/* Ensure that new unhandled slot types return false by default */
19370	default:
19371	return false;
19372	}
19373	}
19374	return true;
19375	}
19376
19377	static bool refsafe(struct bpf_verifier_state *old, struct bpf_verifier_state *cur,
19378	struct bpf_idmap *idmap)
19379	{
19380	int i;
19381
19382	if (old->acquired_refs != cur->acquired_refs)
19383	return false;
19384
19385	if (old->active_locks != cur->active_locks)
19386	return false;
19387
19388	if (old->active_preempt_locks != cur->active_preempt_locks)
19389	return false;
19390
19391	if (old->active_rcu_locks != cur->active_rcu_locks)
19392	return false;
19393
19394	if (!check_ids(old_id: old->active_irq_id, cur_id: cur->active_irq_id, idmap))
19395	return false;
19396
19397	if (!check_ids(old_id: old->active_lock_id, cur_id: cur->active_lock_id, idmap) ||
19398	old->active_lock_ptr != cur->active_lock_ptr)
19399	return false;
19400
19401	for (i = 0; i < old->acquired_refs; i++) {
19402	if (!check_ids(old_id: old->refs[i].id, cur_id: cur->refs[i].id, idmap) ||
19403	old->refs[i].type != cur->refs[i].type)
19404	return false;
19405	switch (old->refs[i].type) {
19406	case REF_TYPE_PTR:
19407	case REF_TYPE_IRQ:
19408	break;
19409	case REF_TYPE_LOCK:
19410	case REF_TYPE_RES_LOCK:
19411	case REF_TYPE_RES_LOCK_IRQ:
19412	if (old->refs[i].ptr != cur->refs[i].ptr)
19413	return false;
19414	break;
19415	default:
19416	WARN_ONCE(1, "Unhandled enum type for reference state: %d\n", old->refs[i].type);
19417	return false;
19418	}
19419	}
19420
19421	return true;
19422	}
19423
19424	/* compare two verifier states
19425	*
19426	* all states stored in state_list are known to be valid, since
19427	* verifier reached 'bpf_exit' instruction through them
19428	*
19429	* this function is called when verifier exploring different branches of
19430	* execution popped from the state stack. If it sees an old state that has
19431	* more strict register state and more strict stack state then this execution
19432	* branch doesn't need to be explored further, since verifier already
19433	* concluded that more strict state leads to valid finish.
19434	*
19435	* Therefore two states are equivalent if register state is more conservative
19436	* and explored stack state is more conservative than the current one.
19437	* Example:
19438	* explored current
19439	* (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
19440	* (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
19441	*
19442	* In other words if current stack state (one being explored) has more
19443	* valid slots than old one that already passed validation, it means
19444	* the verifier can stop exploring and conclude that current state is valid too
19445	*
19446	* Similarly with registers. If explored state has register type as invalid
19447	* whereas register type in current state is meaningful, it means that
19448	* the current state will reach 'bpf_exit' instruction safely
19449	*/
19450	static bool func_states_equal(struct bpf_verifier_env *env, struct bpf_func_state *old,
19451	struct bpf_func_state *cur, u32 insn_idx, enum exact_level exact)
19452	{
19453	u16 live_regs = env->insn_aux_data[insn_idx].live_regs_before;
19454	u16 i;
19455
19456	if (old->callback_depth > cur->callback_depth)
19457	return false;
19458
19459	for (i = 0; i < MAX_BPF_REG; i++)
19460	if (((1 << i) & live_regs) &&
19461	!regsafe(env, rold: &old->regs[i], rcur: &cur->regs[i],
19462	idmap: &env->idmap_scratch, exact))
19463	return false;
19464
19465	if (!stacksafe(env, old, cur, idmap: &env->idmap_scratch, exact))
19466	return false;
19467
19468	return true;
19469	}
19470
19471	static void reset_idmap_scratch(struct bpf_verifier_env *env)
19472	{
19473	env->idmap_scratch.tmp_id_gen = env->id_gen;
19474	memset(&env->idmap_scratch.map, 0, sizeof(env->idmap_scratch.map));
19475	}
19476
19477	static bool states_equal(struct bpf_verifier_env *env,
19478	struct bpf_verifier_state *old,
19479	struct bpf_verifier_state *cur,
19480	enum exact_level exact)
19481	{
19482	u32 insn_idx;
19483	int i;
19484
19485	if (old->curframe != cur->curframe)
19486	return false;
19487
19488	reset_idmap_scratch(env);
19489
19490	/* Verification state from speculative execution simulation
19491	* must never prune a non-speculative execution one.
19492	*/
19493	if (old->speculative && !cur->speculative)
19494	return false;
19495
19496	if (old->in_sleepable != cur->in_sleepable)
19497	return false;
19498
19499	if (!refsafe(old, cur, idmap: &env->idmap_scratch))
19500	return false;
19501
19502	/* for states to be equal callsites have to be the same
19503	* and all frame states need to be equivalent
19504	*/
19505	for (i = 0; i <= old->curframe; i++) {
19506	insn_idx = frame_insn_idx(st: old, frame: i);
19507	if (old->frame[i]->callsite != cur->frame[i]->callsite)
19508	return false;
19509	if (!func_states_equal(env, old: old->frame[i], cur: cur->frame[i], insn_idx, exact))
19510	return false;
19511	}
19512	return true;
19513	}
19514
19515	/* find precise scalars in the previous equivalent state and
19516	* propagate them into the current state
19517	*/
19518	static int propagate_precision(struct bpf_verifier_env *env,
19519	const struct bpf_verifier_state *old,
19520	struct bpf_verifier_state *cur,
19521	bool *changed)
19522	{
19523	struct bpf_reg_state *state_reg;
19524	struct bpf_func_state *state;
19525	int i, err = 0, fr;
19526	bool first;
19527
19528	for (fr = old->curframe; fr >= 0; fr--) {
19529	state = old->frame[fr];
19530	state_reg = state->regs;
19531	first = true;
19532	for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
19533	if (state_reg->type != SCALAR_VALUE ||
19534	!state_reg->precise)
19535	continue;
19536	if (env->log.level & BPF_LOG_LEVEL2) {
19537	if (first)
19538	verbose(private_data: env, fmt: "frame %d: propagating r%d", fr, i);
19539	else
19540	verbose(private_data: env, fmt: ",r%d", i);
19541	}
19542	bt_set_frame_reg(bt: &env->bt, frame: fr, reg: i);
19543	first = false;
19544	}
19545
19546	for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
19547	if (!is_spilled_reg(stack: &state->stack[i]))
19548	continue;
19549	state_reg = &state->stack[i].spilled_ptr;
19550	if (state_reg->type != SCALAR_VALUE ||
19551	!state_reg->precise)
19552	continue;
19553	if (env->log.level & BPF_LOG_LEVEL2) {
19554	if (first)
19555	verbose(private_data: env, fmt: "frame %d: propagating fp%d",
19556	fr, (-i - 1) * BPF_REG_SIZE);
19557	else
19558	verbose(private_data: env, fmt: ",fp%d", (-i - 1) * BPF_REG_SIZE);
19559	}
19560	bt_set_frame_slot(bt: &env->bt, frame: fr, slot: i);
19561	first = false;
19562	}
19563	if (!first && (env->log.level & BPF_LOG_LEVEL2))
19564	verbose(private_data: env, fmt: "\n");
19565	}
19566
19567	err = __mark_chain_precision(env, starting_state: cur, regno: -1, changed);
19568	if (err < 0)
19569	return err;
19570
19571	return 0;
19572	}
19573
19574	#define MAX_BACKEDGE_ITERS 64
19575
19576	/* Propagate read and precision marks from visit->backedges[*].state->equal_state
19577	* to corresponding parent states of visit->backedges[*].state until fixed point is reached,
19578	* then free visit->backedges.
19579	* After execution of this function incomplete_read_marks() will return false
19580	* for all states corresponding to @visit->callchain.
19581	*/
19582	static int propagate_backedges(struct bpf_verifier_env *env, struct bpf_scc_visit *visit)
19583	{
19584	struct bpf_scc_backedge *backedge;
19585	struct bpf_verifier_state *st;
19586	bool changed;
19587	int i, err;
19588
19589	i = 0;
19590	do {
19591	if (i++ > MAX_BACKEDGE_ITERS) {
19592	if (env->log.level & BPF_LOG_LEVEL2)
19593	verbose(private_data: env, fmt: "%s: too many iterations\n", __func__);
19594	for (backedge = visit->backedges; backedge; backedge = backedge->next)
19595	mark_all_scalars_precise(env, st: &backedge->state);
19596	break;
19597	}
19598	changed = false;
19599	for (backedge = visit->backedges; backedge; backedge = backedge->next) {
19600	st = &backedge->state;
19601	err = propagate_precision(env, old: st->equal_state, cur: st, changed: &changed);
19602	if (err)
19603	return err;
19604	}
19605	} while (changed);
19606
19607	free_backedges(visit);
19608	return 0;
19609	}
19610
19611	static bool states_maybe_looping(struct bpf_verifier_state *old,
19612	struct bpf_verifier_state *cur)
19613	{
19614	struct bpf_func_state *fold, *fcur;
19615	int i, fr = cur->curframe;
19616
19617	if (old->curframe != fr)
19618	return false;
19619
19620	fold = old->frame[fr];
19621	fcur = cur->frame[fr];
19622	for (i = 0; i < MAX_BPF_REG; i++)
19623	if (memcmp(p: &fold->regs[i], q: &fcur->regs[i],
19624	offsetof(struct bpf_reg_state, frameno)))
19625	return false;
19626	return true;
19627	}
19628
19629	static bool is_iter_next_insn(struct bpf_verifier_env *env, int insn_idx)
19630	{
19631	return env->insn_aux_data[insn_idx].is_iter_next;
19632	}
19633
19634	/* is_state_visited() handles iter_next() (see process_iter_next_call() for
19635	* terminology) calls specially: as opposed to bounded BPF loops, it *expects*
19636	* states to match, which otherwise would look like an infinite loop. So while
19637	* iter_next() calls are taken care of, we still need to be careful and
19638	* prevent erroneous and too eager declaration of "infinite loop", when
19639	* iterators are involved.
19640	*
19641	* Here's a situation in pseudo-BPF assembly form:
19642	*
19643	* 0: again: ; set up iter_next() call args
19644	* 1: r1 = &it ; <CHECKPOINT HERE>
19645	* 2: call bpf_iter_num_next ; this is iter_next() call
19646	* 3: if r0 == 0 goto done
19647	* 4: ... something useful here ...
19648	* 5: goto again ; another iteration
19649	* 6: done:
19650	* 7: r1 = &it
19651	* 8: call bpf_iter_num_destroy ; clean up iter state
19652	* 9: exit
19653	*
19654	* This is a typical loop. Let's assume that we have a prune point at 1:,
19655	* before we get to `call bpf_iter_num_next` (e.g., because of that `goto
19656	* again`, assuming other heuristics don't get in a way).
19657	*
19658	* When we first time come to 1:, let's say we have some state X. We proceed
19659	* to 2:, fork states, enqueue ACTIVE, validate NULL case successfully, exit.
19660	* Now we come back to validate that forked ACTIVE state. We proceed through
19661	* 3-5, come to goto, jump to 1:. Let's assume our state didn't change, so we
19662	* are converging. But the problem is that we don't know that yet, as this
19663	* convergence has to happen at iter_next() call site only. So if nothing is
19664	* done, at 1: verifier will use bounded loop logic and declare infinite
19665	* looping (and would be *technically* correct, if not for iterator's
19666	* "eventual sticky NULL" contract, see process_iter_next_call()). But we
19667	* don't want that. So what we do in process_iter_next_call() when we go on
19668	* another ACTIVE iteration, we bump slot->iter.depth, to mark that it's
19669	* a different iteration. So when we suspect an infinite loop, we additionally
19670	* check if any of the *ACTIVE* iterator states depths differ. If yes, we
19671	* pretend we are not looping and wait for next iter_next() call.
19672	*
19673	* This only applies to ACTIVE state. In DRAINED state we don't expect to
19674	* loop, because that would actually mean infinite loop, as DRAINED state is
19675	* "sticky", and so we'll keep returning into the same instruction with the
19676	* same state (at least in one of possible code paths).
19677	*
19678	* This approach allows to keep infinite loop heuristic even in the face of
19679	* active iterator. E.g., C snippet below is and will be detected as
19680	* infinitely looping:
19681	*
19682	* struct bpf_iter_num it;
19683	* int *p, x;
19684	*
19685	* bpf_iter_num_new(&it, 0, 10);
19686	* while ((p = bpf_iter_num_next(&t))) {
19687	* x = p;
19688	* while (x--) {} // <<-- infinite loop here
19689	* }
19690	*
19691	*/
19692	static bool iter_active_depths_differ(struct bpf_verifier_state *old, struct bpf_verifier_state *cur)
19693	{
19694	struct bpf_reg_state *slot, *cur_slot;
19695	struct bpf_func_state *state;
19696	int i, fr;
19697
19698	for (fr = old->curframe; fr >= 0; fr--) {
19699	state = old->frame[fr];
19700	for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
19701	if (state->stack[i].slot_type[0] != STACK_ITER)
19702	continue;
19703
19704	slot = &state->stack[i].spilled_ptr;
19705	if (slot->iter.state != BPF_ITER_STATE_ACTIVE)
19706	continue;
19707
19708	cur_slot = &cur->frame[fr]->stack[i].spilled_ptr;
19709	if (cur_slot->iter.depth != slot->iter.depth)
19710	return true;
19711	}
19712	}
19713	return false;
19714	}
19715
19716	static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
19717	{
19718	struct bpf_verifier_state_list *new_sl;
19719	struct bpf_verifier_state_list *sl;
19720	struct bpf_verifier_state *cur = env->cur_state, *new;
19721	bool force_new_state, add_new_state, loop;
19722	int n, err, states_cnt = 0;
19723	struct list_head *pos, *tmp, *head;
19724
19725	force_new_state = env->test_state_freq || is_force_checkpoint(env, insn_idx) ||
19726	/* Avoid accumulating infinitely long jmp history */
19727	cur->jmp_history_cnt > 40;
19728
19729	/* bpf progs typically have pruning point every 4 instructions
19730	* http://vger.kernel.org/bpfconf2019.html#session-1
19731	* Do not add new state for future pruning if the verifier hasn't seen
19732	* at least 2 jumps and at least 8 instructions.
19733	* This heuristics helps decrease 'total_states' and 'peak_states' metric.
19734	* In tests that amounts to up to 50% reduction into total verifier
19735	* memory consumption and 20% verifier time speedup.
19736	*/
19737	add_new_state = force_new_state;
19738	if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
19739	env->insn_processed - env->prev_insn_processed >= 8)
19740	add_new_state = true;
19741
19742	clean_live_states(env, insn: insn_idx, cur);
19743
19744	loop = false;
19745	head = explored_state(env, idx: insn_idx);
19746	list_for_each_safe(pos, tmp, head) {
19747	sl = container_of(pos, struct bpf_verifier_state_list, node);
19748	states_cnt++;
19749	if (sl->state.insn_idx != insn_idx)
19750	continue;
19751
19752	if (sl->state.branches) {
19753	struct bpf_func_state *frame = sl->state.frame[sl->state.curframe];
19754
19755	if (frame->in_async_callback_fn &&
19756	frame->async_entry_cnt != cur->frame[cur->curframe]->async_entry_cnt) {
19757	/* Different async_entry_cnt means that the verifier is
19758	* processing another entry into async callback.
19759	* Seeing the same state is not an indication of infinite
19760	* loop or infinite recursion.
19761	* But finding the same state doesn't mean that it's safe
19762	* to stop processing the current state. The previous state
19763	* hasn't yet reached bpf_exit, since state.branches > 0.
19764	* Checking in_async_callback_fn alone is not enough either.
19765	* Since the verifier still needs to catch infinite loops
19766	* inside async callbacks.
19767	*/
19768	goto skip_inf_loop_check;
19769	}
19770	/* BPF open-coded iterators loop detection is special.
19771	* states_maybe_looping() logic is too simplistic in detecting
19772	* states that *might* be equivalent, because it doesn't know
19773	* about ID remapping, so don't even perform it.
19774	* See process_iter_next_call() and iter_active_depths_differ()
19775	* for overview of the logic. When current and one of parent
19776	* states are detected as equivalent, it's a good thing: we prove
19777	* convergence and can stop simulating further iterations.
19778	* It's safe to assume that iterator loop will finish, taking into
19779	* account iter_next() contract of eventually returning
19780	* sticky NULL result.
19781	*
19782	* Note, that states have to be compared exactly in this case because
19783	* read and precision marks might not be finalized inside the loop.
19784	* E.g. as in the program below:
19785	*
19786	* 1. r7 = -16
19787	* 2. r6 = bpf_get_prandom_u32()
19788	* 3. while (bpf_iter_num_next(&fp[-8])) {
19789	* 4. if (r6 != 42) {
19790	* 5. r7 = -32
19791	* 6. r6 = bpf_get_prandom_u32()
19792	* 7. continue
19793	* 8. }
19794	* 9. r0 = r10
19795	* 10. r0 += r7
19796	* 11. r8 = *(u64 *)(r0 + 0)
19797	* 12. r6 = bpf_get_prandom_u32()
19798	* 13. }
19799	*
19800	* Here verifier would first visit path 1-3, create a checkpoint at 3
19801	* with r7=-16, continue to 4-7,3. Existing checkpoint at 3 does
19802	* not have read or precision mark for r7 yet, thus inexact states
19803	* comparison would discard current state with r7=-32
19804	* => unsafe memory access at 11 would not be caught.
19805	*/
19806	if (is_iter_next_insn(env, insn_idx)) {
19807	if (states_equal(env, old: &sl->state, cur, exact: RANGE_WITHIN)) {
19808	struct bpf_func_state *cur_frame;
19809	struct bpf_reg_state *iter_state, *iter_reg;
19810	int spi;
19811
19812	cur_frame = cur->frame[cur->curframe];
19813	/* btf_check_iter_kfuncs() enforces that
19814	* iter state pointer is always the first arg
19815	*/
19816	iter_reg = &cur_frame->regs[BPF_REG_1];
19817	/* current state is valid due to states_equal(),
19818	* so we can assume valid iter and reg state,
19819	* no need for extra (re-)validations
19820	*/
19821	spi = __get_spi(off: iter_reg->off + iter_reg->var_off.value);
19822	iter_state = &func(env, reg: iter_reg)->stack[spi].spilled_ptr;
19823	if (iter_state->iter.state == BPF_ITER_STATE_ACTIVE) {
19824	loop = true;
19825	goto hit;
19826	}
19827	}
19828	goto skip_inf_loop_check;
19829	}
19830	if (is_may_goto_insn_at(env, insn_idx)) {
19831	if (sl->state.may_goto_depth != cur->may_goto_depth &&
19832	states_equal(env, old: &sl->state, cur, exact: RANGE_WITHIN)) {
19833	loop = true;
19834	goto hit;
19835	}
19836	}
19837	if (bpf_calls_callback(env, insn_idx)) {
19838	if (states_equal(env, old: &sl->state, cur, exact: RANGE_WITHIN))
19839	goto hit;
19840	goto skip_inf_loop_check;
19841	}
19842	/* attempt to detect infinite loop to avoid unnecessary doomed work */
19843	if (states_maybe_looping(old: &sl->state, cur) &&
19844	states_equal(env, old: &sl->state, cur, exact: EXACT) &&
19845	!iter_active_depths_differ(old: &sl->state, cur) &&
19846	sl->state.may_goto_depth == cur->may_goto_depth &&
19847	sl->state.callback_unroll_depth == cur->callback_unroll_depth) {
19848	verbose_linfo(env, insn_off: insn_idx, prefix_fmt: "; ");
19849	verbose(private_data: env, fmt: "infinite loop detected at insn %d\n", insn_idx);
19850	verbose(private_data: env, fmt: "cur state:");
19851	print_verifier_state(env, vstate: cur, frameno: cur->curframe, print_all: true);
19852	verbose(private_data: env, fmt: "old state:");
19853	print_verifier_state(env, vstate: &sl->state, frameno: cur->curframe, print_all: true);
19854	return -EINVAL;
19855	}
19856	/* if the verifier is processing a loop, avoid adding new state
19857	* too often, since different loop iterations have distinct
19858	* states and may not help future pruning.
19859	* This threshold shouldn't be too low to make sure that
19860	* a loop with large bound will be rejected quickly.
19861	* The most abusive loop will be:
19862	* r1 += 1
19863	* if r1 < 1000000 goto pc-2
19864	* 1M insn_procssed limit / 100 == 10k peak states.
19865	* This threshold shouldn't be too high either, since states
19866	* at the end of the loop are likely to be useful in pruning.
19867	*/
19868	skip_inf_loop_check:
19869	if (!force_new_state &&
19870	env->jmps_processed - env->prev_jmps_processed < 20 &&
19871	env->insn_processed - env->prev_insn_processed < 100)
19872	add_new_state = false;
19873	goto miss;
19874	}
19875	/* See comments for mark_all_regs_read_and_precise() */
19876	loop = incomplete_read_marks(env, st: &sl->state);
19877	if (states_equal(env, old: &sl->state, cur, exact: loop ? RANGE_WITHIN : NOT_EXACT)) {
19878	hit:
19879	sl->hit_cnt++;
19880
19881	/* if previous state reached the exit with precision and
19882	* current state is equivalent to it (except precision marks)
19883	* the precision needs to be propagated back in
19884	* the current state.
19885	*/
19886	err = 0;
19887	if (is_jmp_point(env, insn_idx: env->insn_idx))
19888	err = push_jmp_history(env, cur, insn_flags: 0, linked_regs: 0);
19889	err = err ? : propagate_precision(env, old: &sl->state, cur, NULL);
19890	if (err)
19891	return err;
19892	/* When processing iterator based loops above propagate_liveness and
19893	* propagate_precision calls are not sufficient to transfer all relevant
19894	* read and precision marks. E.g. consider the following case:
19895	*
19896	* .-> A --. Assume the states are visited in the order A, B, C.
19897	* | | | Assume that state B reaches a state equivalent to state A.
19898	* | v v At this point, state C is not processed yet, so state A
19899	* '-- B C has not received any read or precision marks from C.
19900	* Thus, marks propagated from A to B are incomplete.
19901	*
19902	* The verifier mitigates this by performing the following steps:
19903	*
19904	* - Prior to the main verification pass, strongly connected components
19905	* (SCCs) are computed over the program's control flow graph,
19906	* intraprocedurally.
19907	*
19908	* - During the main verification pass, `maybe_enter_scc()` checks
19909	* whether the current verifier state is entering an SCC. If so, an
19910	* instance of a `bpf_scc_visit` object is created, and the state
19911	* entering the SCC is recorded as the entry state.
19912	*
19913	* - This instance is associated not with the SCC itself, but with a
19914	* `bpf_scc_callchain`: a tuple consisting of the call sites leading to
19915	* the SCC and the SCC id. See `compute_scc_callchain()`.
19916	*
19917	* - When a verification path encounters a `states_equal(...,
19918	* RANGE_WITHIN)` condition, there exists a call chain describing the
19919	* current state and a corresponding `bpf_scc_visit` instance. A copy
19920	* of the current state is created and added to
19921	* `bpf_scc_visit->backedges`.
19922	*
19923	* - When a verification path terminates, `maybe_exit_scc()` is called
19924	* from `update_branch_counts()`. For states with `branches == 0`, it
19925	* checks whether the state is the entry state of any `bpf_scc_visit`
19926	* instance. If it is, this indicates that all paths originating from
19927	* this SCC visit have been explored. `propagate_backedges()` is then
19928	* called, which propagates read and precision marks through the
19929	* backedges until a fixed point is reached.
19930	* (In the earlier example, this would propagate marks from A to B,
19931	* from C to A, and then again from A to B.)
19932	*
19933	* A note on callchains
19934	* --------------------
19935	*
19936	* Consider the following example:
19937	*
19938	* void foo() { loop { ... SCC#1 ... } }
19939	* void main() {
19940	* A: foo();
19941	* B: ...
19942	* C: foo();
19943	* }
19944	*
19945	* Here, there are two distinct callchains leading to SCC#1:
19946	* - (A, SCC#1)
19947	* - (C, SCC#1)
19948	*
19949	* Each callchain identifies a separate `bpf_scc_visit` instance that
19950	* accumulates backedge states. The `propagate_{liveness,precision}()`
19951	* functions traverse the parent state of each backedge state, which
19952	* means these parent states must remain valid (i.e., not freed) while
19953	* the corresponding `bpf_scc_visit` instance exists.
19954	*
19955	* Associating `bpf_scc_visit` instances directly with SCCs instead of
19956	* callchains would break this invariant:
19957	* - States explored during `C: foo()` would contribute backedges to
19958	* SCC#1, but SCC#1 would only be exited once the exploration of
19959	* `A: foo()` completes.
19960	* - By that time, the states explored between `A: foo()` and `C: foo()`
19961	* (i.e., `B: ...`) may have already been freed, causing the parent
19962	* links for states from `C: foo()` to become invalid.
19963	*/
19964	if (loop) {
19965	struct bpf_scc_backedge *backedge;
19966
19967	backedge = kzalloc(sizeof(*backedge), GFP_KERNEL_ACCOUNT);
19968	if (!backedge)
19969	return -ENOMEM;
19970	err = copy_verifier_state(dst_state: &backedge->state, src: cur);
19971	backedge->state.equal_state = &sl->state;
19972	backedge->state.insn_idx = insn_idx;
19973	err = err ?: add_scc_backedge(env, st: &sl->state, backedge);
19974	if (err) {
19975	free_verifier_state(state: &backedge->state, free_self: false);
19976	kfree(objp: backedge);
19977	return err;
19978	}
19979	}
19980	return 1;
19981	}
19982	miss:
19983	/* when new state is not going to be added do not increase miss count.
19984	* Otherwise several loop iterations will remove the state
19985	* recorded earlier. The goal of these heuristics is to have
19986	* states from some iterations of the loop (some in the beginning
19987	* and some at the end) to help pruning.
19988	*/
19989	if (add_new_state)
19990	sl->miss_cnt++;
19991	/* heuristic to determine whether this state is beneficial
19992	* to keep checking from state equivalence point of view.
19993	* Higher numbers increase max_states_per_insn and verification time,
19994	* but do not meaningfully decrease insn_processed.
19995	* 'n' controls how many times state could miss before eviction.
19996	* Use bigger 'n' for checkpoints because evicting checkpoint states
19997	* too early would hinder iterator convergence.
19998	*/
19999	n = is_force_checkpoint(env, insn_idx) && sl->state.branches > 0 ? 64 : 3;
20000	if (sl->miss_cnt > sl->hit_cnt * n + n) {
20001	/* the state is unlikely to be useful. Remove it to
20002	* speed up verification
20003	*/
20004	sl->in_free_list = true;
20005	list_del(entry: &sl->node);
20006	list_add(new: &sl->node, head: &env->free_list);
20007	env->free_list_size++;
20008	env->explored_states_size--;
20009	maybe_free_verifier_state(env, sl);
20010	}
20011	}
20012
20013	if (env->max_states_per_insn < states_cnt)
20014	env->max_states_per_insn = states_cnt;
20015
20016	if (!env->bpf_capable && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
20017	return 0;
20018
20019	if (!add_new_state)
20020	return 0;
20021
20022	/* There were no equivalent states, remember the current one.
20023	* Technically the current state is not proven to be safe yet,
20024	* but it will either reach outer most bpf_exit (which means it's safe)
20025	* or it will be rejected. When there are no loops the verifier won't be
20026	* seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
20027	* again on the way to bpf_exit.
20028	* When looping the sl->state.branches will be > 0 and this state
20029	* will not be considered for equivalence until branches == 0.
20030	*/
20031	new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL_ACCOUNT);
20032	if (!new_sl)
20033	return -ENOMEM;
20034	env->total_states++;
20035	env->explored_states_size++;
20036	update_peak_states(env);
20037	env->prev_jmps_processed = env->jmps_processed;
20038	env->prev_insn_processed = env->insn_processed;
20039
20040	/* forget precise markings we inherited, see __mark_chain_precision */
20041	if (env->bpf_capable)
20042	mark_all_scalars_imprecise(env, st: cur);
20043
20044	/* add new state to the head of linked list */
20045	new = &new_sl->state;
20046	err = copy_verifier_state(dst_state: new, src: cur);
20047	if (err) {
20048	free_verifier_state(state: new, free_self: false);
20049	kfree(objp: new_sl);
20050	return err;
20051	}
20052	new->insn_idx = insn_idx;
20053	verifier_bug_if(new->branches != 1, env,
20054	"%s:branches_to_explore=%d insn %d",
20055	__func__, new->branches, insn_idx);
20056	err = maybe_enter_scc(env, st: new);
20057	if (err) {
20058	free_verifier_state(state: new, free_self: false);
20059	kfree(objp: new_sl);
20060	return err;
20061	}
20062
20063	cur->parent = new;
20064	cur->first_insn_idx = insn_idx;
20065	cur->dfs_depth = new->dfs_depth + 1;
20066	clear_jmp_history(state: cur);
20067	list_add(new: &new_sl->node, head);
20068	return 0;
20069	}
20070
20071	/* Return true if it's OK to have the same insn return a different type. */
20072	static bool reg_type_mismatch_ok(enum bpf_reg_type type)
20073	{
20074	switch (base_type(type)) {
20075	case PTR_TO_CTX:
20076	case PTR_TO_SOCKET:
20077	case PTR_TO_SOCK_COMMON:
20078	case PTR_TO_TCP_SOCK:
20079	case PTR_TO_XDP_SOCK:
20080	case PTR_TO_BTF_ID:
20081	case PTR_TO_ARENA:
20082	return false;
20083	default:
20084	return true;
20085	}
20086	}
20087
20088	/* If an instruction was previously used with particular pointer types, then we
20089	* need to be careful to avoid cases such as the below, where it may be ok
20090	* for one branch accessing the pointer, but not ok for the other branch:
20091	*
20092	* R1 = sock_ptr
20093	* goto X;
20094	* ...
20095	* R1 = some_other_valid_ptr;
20096	* goto X;
20097	* ...
20098	* R2 = *(u32 *)(R1 + 0);
20099	*/
20100	static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
20101	{
20102	return src != prev && (!reg_type_mismatch_ok(type: src) ||
20103	!reg_type_mismatch_ok(type: prev));
20104	}
20105
20106	static bool is_ptr_to_mem_or_btf_id(enum bpf_reg_type type)
20107	{
20108	switch (base_type(type)) {
20109	case PTR_TO_MEM:
20110	case PTR_TO_BTF_ID:
20111	return true;
20112	default:
20113	return false;
20114	}
20115	}
20116
20117	static bool is_ptr_to_mem(enum bpf_reg_type type)
20118	{
20119	return base_type(type) == PTR_TO_MEM;
20120	}
20121
20122	static int save_aux_ptr_type(struct bpf_verifier_env *env, enum bpf_reg_type type,
20123	bool allow_trust_mismatch)
20124	{
20125	enum bpf_reg_type *prev_type = &env->insn_aux_data[env->insn_idx].ptr_type;
20126	enum bpf_reg_type merged_type;
20127
20128	if (*prev_type == NOT_INIT) {
20129	/* Saw a valid insn
20130	* dst_reg = *(u32 *)(src_reg + off)
20131	* save type to validate intersecting paths
20132	*/
20133	*prev_type = type;
20134	} else if (reg_type_mismatch(src: type, prev: *prev_type)) {
20135	/* Abuser program is trying to use the same insn
20136	* dst_reg = *(u32*) (src_reg + off)
20137	* with different pointer types:
20138	* src_reg == ctx in one branch and
20139	* src_reg == stack|map in some other branch.
20140	* Reject it.
20141	*/
20142	if (allow_trust_mismatch &&
20143	is_ptr_to_mem_or_btf_id(type) &&
20144	is_ptr_to_mem_or_btf_id(type: *prev_type)) {
20145	/*
20146	* Have to support a use case when one path through
20147	* the program yields TRUSTED pointer while another
20148	* is UNTRUSTED. Fallback to UNTRUSTED to generate
20149	* BPF_PROBE_MEM/BPF_PROBE_MEMSX.
20150	* Same behavior of MEM_RDONLY flag.
20151	*/
20152	if (is_ptr_to_mem(type) || is_ptr_to_mem(type: *prev_type))
20153	merged_type = PTR_TO_MEM;
20154	else
20155	merged_type = PTR_TO_BTF_ID;
20156	if ((type & PTR_UNTRUSTED) || (*prev_type & PTR_UNTRUSTED))
20157	merged_type |= PTR_UNTRUSTED;
20158	if ((type & MEM_RDONLY) || (*prev_type & MEM_RDONLY))
20159	merged_type |= MEM_RDONLY;
20160	*prev_type = merged_type;
20161	} else {
20162	verbose(private_data: env, fmt: "same insn cannot be used with different pointers\n");
20163	return -EINVAL;
20164	}
20165	}
20166
20167	return 0;
20168	}
20169
20170	enum {
20171	PROCESS_BPF_EXIT = 1
20172	};
20173
20174	static int process_bpf_exit_full(struct bpf_verifier_env *env,
20175	bool *do_print_state,
20176	bool exception_exit)
20177	{
20178	/* We must do check_reference_leak here before
20179	* prepare_func_exit to handle the case when
20180	* state->curframe > 0, it may be a callback function,
20181	* for which reference_state must match caller reference
20182	* state when it exits.
20183	*/
20184	int err = check_resource_leak(env, exception_exit,
20185	check_lock: !env->cur_state->curframe,
20186	prefix: "BPF_EXIT instruction in main prog");
20187	if (err)
20188	return err;
20189
20190	/* The side effect of the prepare_func_exit which is
20191	* being skipped is that it frees bpf_func_state.
20192	* Typically, process_bpf_exit will only be hit with
20193	* outermost exit. copy_verifier_state in pop_stack will
20194	* handle freeing of any extra bpf_func_state left over
20195	* from not processing all nested function exits. We
20196	* also skip return code checks as they are not needed
20197	* for exceptional exits.
20198	*/
20199	if (exception_exit)
20200	return PROCESS_BPF_EXIT;
20201
20202	if (env->cur_state->curframe) {
20203	/* exit from nested function */
20204	err = prepare_func_exit(env, insn_idx: &env->insn_idx);
20205	if (err)
20206	return err;
20207	*do_print_state = true;
20208	return 0;
20209	}
20210
20211	err = check_return_code(env, regno: BPF_REG_0, reg_name: "R0");
20212	if (err)
20213	return err;
20214	return PROCESS_BPF_EXIT;
20215	}
20216
20217	static int indirect_jump_min_max_index(struct bpf_verifier_env *env,
20218	int regno,
20219	struct bpf_map *map,
20220	u32 *pmin_index, u32 *pmax_index)
20221	{
20222	struct bpf_reg_state *reg = reg_state(env, regno);
20223	u64 min_index, max_index;
20224	const u32 size = 8;
20225
20226	if (check_add_overflow(reg->umin_value, reg->off, &min_index) ||
20227	(min_index > (u64) U32_MAX * size)) {
20228	verbose(private_data: env, fmt: "the sum of R%u umin_value %llu and off %u is too big\n",
20229	regno, reg->umin_value, reg->off);
20230	return -ERANGE;
20231	}
20232	if (check_add_overflow(reg->umax_value, reg->off, &max_index) ||
20233	(max_index > (u64) U32_MAX * size)) {
20234	verbose(private_data: env, fmt: "the sum of R%u umax_value %llu and off %u is too big\n",
20235	regno, reg->umax_value, reg->off);
20236	return -ERANGE;
20237	}
20238
20239	min_index /= size;
20240	max_index /= size;
20241
20242	if (max_index >= map->max_entries) {
20243	verbose(private_data: env, fmt: "R%u points to outside of jump table: [%llu,%llu] max_entries %u\n",
20244	regno, min_index, max_index, map->max_entries);
20245	return -EINVAL;
20246	}
20247
20248	*pmin_index = min_index;
20249	*pmax_index = max_index;
20250	return 0;
20251	}
20252
20253	/* gotox *dst_reg */
20254	static int check_indirect_jump(struct bpf_verifier_env *env, struct bpf_insn *insn)
20255	{
20256	struct bpf_verifier_state *other_branch;
20257	struct bpf_reg_state *dst_reg;
20258	struct bpf_map *map;
20259	u32 min_index, max_index;
20260	int err = 0;
20261	int n;
20262	int i;
20263
20264	dst_reg = reg_state(env, regno: insn->dst_reg);
20265	if (dst_reg->type != PTR_TO_INSN) {
20266	verbose(private_data: env, fmt: "R%d has type %s, expected PTR_TO_INSN\n",
20267	insn->dst_reg, reg_type_str(env, type: dst_reg->type));
20268	return -EINVAL;
20269	}
20270
20271	map = dst_reg->map_ptr;
20272	if (verifier_bug_if(!map, env, "R%d has an empty map pointer", insn->dst_reg))
20273	return -EFAULT;
20274
20275	if (verifier_bug_if(map->map_type != BPF_MAP_TYPE_INSN_ARRAY, env,
20276	"R%d has incorrect map type %d", insn->dst_reg, map->map_type))
20277	return -EFAULT;
20278
20279	err = indirect_jump_min_max_index(env, regno: insn->dst_reg, map, pmin_index: &min_index, pmax_index: &max_index);
20280	if (err)
20281	return err;
20282
20283	/* Ensure that the buffer is large enough */
20284	if (!env->gotox_tmp_buf || env->gotox_tmp_buf->cnt < max_index - min_index + 1) {
20285	env->gotox_tmp_buf = iarray_realloc(old: env->gotox_tmp_buf,
20286	n_elem: max_index - min_index + 1);
20287	if (!env->gotox_tmp_buf)
20288	return -ENOMEM;
20289	}
20290
20291	n = copy_insn_array_uniq(map, start: min_index, end: max_index, off: env->gotox_tmp_buf->items);
20292	if (n < 0)
20293	return n;
20294	if (n == 0) {
20295	verbose(private_data: env, fmt: "register R%d doesn't point to any offset in map id=%d\n",
20296	insn->dst_reg, map->id);
20297	return -EINVAL;
20298	}
20299
20300	for (i = 0; i < n - 1; i++) {
20301	other_branch = push_stack(env, insn_idx: env->gotox_tmp_buf->items[i],
20302	prev_insn_idx: env->insn_idx, speculative: env->cur_state->speculative);
20303	if (IS_ERR(ptr: other_branch))
20304	return PTR_ERR(ptr: other_branch);
20305	}
20306	env->insn_idx = env->gotox_tmp_buf->items[n-1];
20307	return 0;
20308	}
20309
20310	static int do_check_insn(struct bpf_verifier_env *env, bool *do_print_state)
20311	{
20312	int err;
20313	struct bpf_insn *insn = &env->prog->insnsi[env->insn_idx];
20314	u8 class = BPF_CLASS(insn->code);
20315
20316	if (class == BPF_ALU || class == BPF_ALU64) {
20317	err = check_alu_op(env, insn);
20318	if (err)
20319	return err;
20320
20321	} else if (class == BPF_LDX) {
20322	bool is_ldsx = BPF_MODE(insn->code) == BPF_MEMSX;
20323
20324	/* Check for reserved fields is already done in
20325	* resolve_pseudo_ldimm64().
20326	*/
20327	err = check_load_mem(env, insn, strict_alignment_once: false, is_ldsx, allow_trust_mismatch: true, ctx: "ldx");
20328	if (err)
20329	return err;
20330	} else if (class == BPF_STX) {
20331	if (BPF_MODE(insn->code) == BPF_ATOMIC) {
20332	err = check_atomic(env, insn);
20333	if (err)
20334	return err;
20335	env->insn_idx++;
20336	return 0;
20337	}
20338
20339	if (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0) {
20340	verbose(private_data: env, fmt: "BPF_STX uses reserved fields\n");
20341	return -EINVAL;
20342	}
20343
20344	err = check_store_reg(env, insn, strict_alignment_once: false);
20345	if (err)
20346	return err;
20347	} else if (class == BPF_ST) {
20348	enum bpf_reg_type dst_reg_type;
20349
20350	if (BPF_MODE(insn->code) != BPF_MEM ||
20351	insn->src_reg != BPF_REG_0) {
20352	verbose(private_data: env, fmt: "BPF_ST uses reserved fields\n");
20353	return -EINVAL;
20354	}
20355	/* check src operand */
20356	err = check_reg_arg(env, regno: insn->dst_reg, t: SRC_OP);
20357	if (err)
20358	return err;
20359
20360	dst_reg_type = cur_regs(env)[insn->dst_reg].type;
20361
20362	/* check that memory (dst_reg + off) is writeable */
20363	err = check_mem_access(env, insn_idx: env->insn_idx, regno: insn->dst_reg,
20364	off: insn->off, BPF_SIZE(insn->code),
20365	t: BPF_WRITE, value_regno: -1, strict_alignment_once: false, is_ldsx: false);
20366	if (err)
20367	return err;
20368
20369	err = save_aux_ptr_type(env, type: dst_reg_type, allow_trust_mismatch: false);
20370	if (err)
20371	return err;
20372	} else if (class == BPF_JMP || class == BPF_JMP32) {
20373	u8 opcode = BPF_OP(insn->code);
20374
20375	env->jmps_processed++;
20376	if (opcode == BPF_CALL) {
20377	if (BPF_SRC(insn->code) != BPF_K ||
20378	(insn->src_reg != BPF_PSEUDO_KFUNC_CALL &&
20379	insn->off != 0) ||
20380	(insn->src_reg != BPF_REG_0 &&
20381	insn->src_reg != BPF_PSEUDO_CALL &&
20382	insn->src_reg != BPF_PSEUDO_KFUNC_CALL) ||
20383	insn->dst_reg != BPF_REG_0 || class == BPF_JMP32) {
20384	verbose(private_data: env, fmt: "BPF_CALL uses reserved fields\n");
20385	return -EINVAL;
20386	}
20387
20388	if (env->cur_state->active_locks) {
20389	if ((insn->src_reg == BPF_REG_0 &&
20390	insn->imm != BPF_FUNC_spin_unlock) ||
20391	(insn->src_reg == BPF_PSEUDO_KFUNC_CALL &&
20392	(insn->off != 0 || !kfunc_spin_allowed(btf_id: insn->imm)))) {
20393	verbose(private_data: env,
20394	fmt: "function calls are not allowed while holding a lock\n");
20395	return -EINVAL;
20396	}
20397	}
20398	if (insn->src_reg == BPF_PSEUDO_CALL) {
20399	err = check_func_call(env, insn, insn_idx: &env->insn_idx);
20400	} else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
20401	err = check_kfunc_call(env, insn, insn_idx_p: &env->insn_idx);
20402	if (!err && is_bpf_throw_kfunc(insn))
20403	return process_bpf_exit_full(env, do_print_state, exception_exit: true);
20404	} else {
20405	err = check_helper_call(env, insn, insn_idx_p: &env->insn_idx);
20406	}
20407	if (err)
20408	return err;
20409
20410	mark_reg_scratched(env, regno: BPF_REG_0);
20411	} else if (opcode == BPF_JA) {
20412	if (BPF_SRC(insn->code) == BPF_X) {
20413	if (insn->src_reg != BPF_REG_0 ||
20414	insn->imm != 0 || insn->off != 0) {
20415	verbose(private_data: env, fmt: "BPF_JA|BPF_X uses reserved fields\n");
20416	return -EINVAL;
20417	}
20418	return check_indirect_jump(env, insn);
20419	}
20420
20421	if (BPF_SRC(insn->code) != BPF_K ||
20422	insn->src_reg != BPF_REG_0 ||
20423	insn->dst_reg != BPF_REG_0 ||
20424	(class == BPF_JMP && insn->imm != 0) ||
20425	(class == BPF_JMP32 && insn->off != 0)) {
20426	verbose(private_data: env, fmt: "BPF_JA uses reserved fields\n");
20427	return -EINVAL;
20428	}
20429
20430	if (class == BPF_JMP)
20431	env->insn_idx += insn->off + 1;
20432	else
20433	env->insn_idx += insn->imm + 1;
20434	return 0;
20435	} else if (opcode == BPF_EXIT) {
20436	if (BPF_SRC(insn->code) != BPF_K ||
20437	insn->imm != 0 ||
20438	insn->src_reg != BPF_REG_0 ||
20439	insn->dst_reg != BPF_REG_0 ||
20440	class == BPF_JMP32) {
20441	verbose(private_data: env, fmt: "BPF_EXIT uses reserved fields\n");
20442	return -EINVAL;
20443	}
20444	return process_bpf_exit_full(env, do_print_state, exception_exit: false);
20445	} else {
20446	err = check_cond_jmp_op(env, insn, insn_idx: &env->insn_idx);
20447	if (err)
20448	return err;
20449	}
20450	} else if (class == BPF_LD) {
20451	u8 mode = BPF_MODE(insn->code);
20452
20453	if (mode == BPF_ABS || mode == BPF_IND) {
20454	err = check_ld_abs(env, insn);
20455	if (err)
20456	return err;
20457
20458	} else if (mode == BPF_IMM) {
20459	err = check_ld_imm(env, insn);
20460	if (err)
20461	return err;
20462
20463	env->insn_idx++;
20464	sanitize_mark_insn_seen(env);
20465	} else {
20466	verbose(private_data: env, fmt: "invalid BPF_LD mode\n");
20467	return -EINVAL;
20468	}
20469	} else {
20470	verbose(private_data: env, fmt: "unknown insn class %d\n", class);
20471	return -EINVAL;
20472	}
20473
20474	env->insn_idx++;
20475	return 0;
20476	}
20477
20478	static int do_check(struct bpf_verifier_env *env)
20479	{
20480	bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
20481	struct bpf_verifier_state *state = env->cur_state;
20482	struct bpf_insn *insns = env->prog->insnsi;
20483	int insn_cnt = env->prog->len;
20484	bool do_print_state = false;
20485	int prev_insn_idx = -1;
20486
20487	for (;;) {
20488	struct bpf_insn *insn;
20489	struct bpf_insn_aux_data *insn_aux;
20490	int err, marks_err;
20491
20492	/* reset current history entry on each new instruction */
20493	env->cur_hist_ent = NULL;
20494
20495	env->prev_insn_idx = prev_insn_idx;
20496	if (env->insn_idx >= insn_cnt) {
20497	verbose(private_data: env, fmt: "invalid insn idx %d insn_cnt %d\n",
20498	env->insn_idx, insn_cnt);
20499	return -EFAULT;
20500	}
20501
20502	insn = &insns[env->insn_idx];
20503	insn_aux = &env->insn_aux_data[env->insn_idx];
20504
20505	if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
20506	verbose(private_data: env,
20507	fmt: "BPF program is too large. Processed %d insn\n",
20508	env->insn_processed);
20509	return -E2BIG;
20510	}
20511
20512	state->last_insn_idx = env->prev_insn_idx;
20513	state->insn_idx = env->insn_idx;
20514
20515	if (is_prune_point(env, insn_idx: env->insn_idx)) {
20516	err = is_state_visited(env, insn_idx: env->insn_idx);
20517	if (err < 0)
20518	return err;
20519	if (err == 1) {
20520	/* found equivalent state, can prune the search */
20521	if (env->log.level & BPF_LOG_LEVEL) {
20522	if (do_print_state)
20523	verbose(private_data: env, fmt: "\nfrom %d to %d%s: safe\n",
20524	env->prev_insn_idx, env->insn_idx,
20525	env->cur_state->speculative ?
20526	" (speculative execution)" : "");
20527	else
20528	verbose(private_data: env, fmt: "%d: safe\n", env->insn_idx);
20529	}
20530	goto process_bpf_exit;
20531	}
20532	}
20533
20534	if (is_jmp_point(env, insn_idx: env->insn_idx)) {
20535	err = push_jmp_history(env, cur: state, insn_flags: 0, linked_regs: 0);
20536	if (err)
20537	return err;
20538	}
20539
20540	if (signal_pending(current))
20541	return -EAGAIN;
20542
20543	if (need_resched())
20544	cond_resched();
20545
20546	if (env->log.level & BPF_LOG_LEVEL2 && do_print_state) {
20547	verbose(private_data: env, fmt: "\nfrom %d to %d%s:",
20548	env->prev_insn_idx, env->insn_idx,
20549	env->cur_state->speculative ?
20550	" (speculative execution)" : "");
20551	print_verifier_state(env, vstate: state, frameno: state->curframe, print_all: true);
20552	do_print_state = false;
20553	}
20554
20555	if (env->log.level & BPF_LOG_LEVEL) {
20556	if (verifier_state_scratched(env))
20557	print_insn_state(env, vstate: state, frameno: state->curframe);
20558
20559	verbose_linfo(env, insn_off: env->insn_idx, prefix_fmt: "; ");
20560	env->prev_log_pos = env->log.end_pos;
20561	verbose(private_data: env, fmt: "%d: ", env->insn_idx);
20562	verbose_insn(env, insn);
20563	env->prev_insn_print_pos = env->log.end_pos - env->prev_log_pos;
20564	env->prev_log_pos = env->log.end_pos;
20565	}
20566
20567	if (bpf_prog_is_offloaded(aux: env->prog->aux)) {
20568	err = bpf_prog_offload_verify_insn(env, insn_idx: env->insn_idx,
20569	prev_insn_idx: env->prev_insn_idx);
20570	if (err)
20571	return err;
20572	}
20573
20574	sanitize_mark_insn_seen(env);
20575	prev_insn_idx = env->insn_idx;
20576
20577	/* Reduce verification complexity by stopping speculative path
20578	* verification when a nospec is encountered.
20579	*/
20580	if (state->speculative && insn_aux->nospec)
20581	goto process_bpf_exit;
20582
20583	err = bpf_reset_stack_write_marks(env, insn_idx: env->insn_idx);
20584	if (err)
20585	return err;
20586	err = do_check_insn(env, do_print_state: &do_print_state);
20587	if (err >= 0 || error_recoverable_with_nospec(err)) {
20588	marks_err = bpf_commit_stack_write_marks(env);
20589	if (marks_err)
20590	return marks_err;
20591	}
20592	if (error_recoverable_with_nospec(err) && state->speculative) {
20593	/* Prevent this speculative path from ever reaching the
20594	* insn that would have been unsafe to execute.
20595	*/
20596	insn_aux->nospec = true;
20597	/* If it was an ADD/SUB insn, potentially remove any
20598	* markings for alu sanitization.
20599	*/
20600	insn_aux->alu_state = 0;
20601	goto process_bpf_exit;
20602	} else if (err < 0) {
20603	return err;
20604	} else if (err == PROCESS_BPF_EXIT) {
20605	goto process_bpf_exit;
20606	}
20607	WARN_ON_ONCE(err);
20608
20609	if (state->speculative && insn_aux->nospec_result) {
20610	/* If we are on a path that performed a jump-op, this
20611	* may skip a nospec patched-in after the jump. This can
20612	* currently never happen because nospec_result is only
20613	* used for the write-ops
20614	* `*(size*)(dst_reg+off)=src_reg|imm32` which must
20615	* never skip the following insn. Still, add a warning
20616	* to document this in case nospec_result is used
20617	* elsewhere in the future.
20618	*
20619	* All non-branch instructions have a single
20620	* fall-through edge. For these, nospec_result should
20621	* already work.
20622	*/
20623	if (verifier_bug_if(BPF_CLASS(insn->code) == BPF_JMP ||
20624	BPF_CLASS(insn->code) == BPF_JMP32, env,
20625	"speculation barrier after jump instruction may not have the desired effect"))
20626	return -EFAULT;
20627	process_bpf_exit:
20628	mark_verifier_state_scratched(env);
20629	err = update_branch_counts(env, st: env->cur_state);
20630	if (err)
20631	return err;
20632	err = bpf_update_live_stack(env);
20633	if (err)
20634	return err;
20635	err = pop_stack(env, prev_insn_idx: &prev_insn_idx, insn_idx: &env->insn_idx,
20636	pop_log);
20637	if (err < 0) {
20638	if (err != -ENOENT)
20639	return err;
20640	break;
20641	} else {
20642	do_print_state = true;
20643	continue;
20644	}
20645	}
20646	}
20647
20648	return 0;
20649	}
20650
20651	static int find_btf_percpu_datasec(struct btf *btf)
20652	{
20653	const struct btf_type *t;
20654	const char *tname;
20655	int i, n;
20656
20657	/*
20658	* Both vmlinux and module each have their own ".data..percpu"
20659	* DATASECs in BTF. So for module's case, we need to skip vmlinux BTF
20660	* types to look at only module's own BTF types.
20661	*/
20662	n = btf_nr_types(btf);
20663	if (btf_is_module(btf))
20664	i = btf_nr_types(btf: btf_vmlinux);
20665	else
20666	i = 1;
20667
20668	for(; i < n; i++) {
20669	t = btf_type_by_id(btf, type_id: i);
20670	if (BTF_INFO_KIND(t->info) != BTF_KIND_DATASEC)
20671	continue;
20672
20673	tname = btf_name_by_offset(btf, offset: t->name_off);
20674	if (!strcmp(tname, ".data..percpu"))
20675	return i;
20676	}
20677
20678	return -ENOENT;
20679	}
20680
20681	/*
20682	* Add btf to the used_btfs array and return the index. (If the btf was
20683	* already added, then just return the index.) Upon successful insertion
20684	* increase btf refcnt, and, if present, also refcount the corresponding
20685	* kernel module.
20686	*/
20687	static int __add_used_btf(struct bpf_verifier_env *env, struct btf *btf)
20688	{
20689	struct btf_mod_pair *btf_mod;
20690	int i;
20691
20692	/* check whether we recorded this BTF (and maybe module) already */
20693	for (i = 0; i < env->used_btf_cnt; i++)
20694	if (env->used_btfs[i].btf == btf)
20695	return i;
20696
20697	if (env->used_btf_cnt >= MAX_USED_BTFS) {
20698	verbose(private_data: env, fmt: "The total number of btfs per program has reached the limit of %u\n",
20699	MAX_USED_BTFS);
20700	return -E2BIG;
20701	}
20702
20703	btf_get(btf);
20704
20705	btf_mod = &env->used_btfs[env->used_btf_cnt];
20706	btf_mod->btf = btf;
20707	btf_mod->module = NULL;
20708
20709	/* if we reference variables from kernel module, bump its refcount */
20710	if (btf_is_module(btf)) {
20711	btf_mod->module = btf_try_get_module(btf);
20712	if (!btf_mod->module) {
20713	btf_put(btf);
20714	return -ENXIO;
20715	}
20716	}
20717
20718	return env->used_btf_cnt++;
20719	}
20720
20721	/* replace pseudo btf_id with kernel symbol address */
20722	static int __check_pseudo_btf_id(struct bpf_verifier_env *env,
20723	struct bpf_insn *insn,
20724	struct bpf_insn_aux_data *aux,
20725	struct btf *btf)
20726	{
20727	const struct btf_var_secinfo *vsi;
20728	const struct btf_type *datasec;
20729	const struct btf_type *t;
20730	const char *sym_name;
20731	bool percpu = false;
20732	u32 type, id = insn->imm;
20733	s32 datasec_id;
20734	u64 addr;
20735	int i;
20736
20737	t = btf_type_by_id(btf, type_id: id);
20738	if (!t) {
20739	verbose(private_data: env, fmt: "ldimm64 insn specifies invalid btf_id %d.\n", id);
20740	return -ENOENT;
20741	}
20742
20743	if (!btf_type_is_var(t) && !btf_type_is_func(t)) {
20744	verbose(private_data: env, fmt: "pseudo btf_id %d in ldimm64 isn't KIND_VAR or KIND_FUNC\n", id);
20745	return -EINVAL;
20746	}
20747
20748	sym_name = btf_name_by_offset(btf, offset: t->name_off);
20749	addr = kallsyms_lookup_name(name: sym_name);
20750	if (!addr) {
20751	verbose(private_data: env, fmt: "ldimm64 failed to find the address for kernel symbol '%s'.\n",
20752	sym_name);
20753	return -ENOENT;
20754	}
20755	insn[0].imm = (u32)addr;
20756	insn[1].imm = addr >> 32;
20757
20758	if (btf_type_is_func(t)) {
20759	aux->btf_var.reg_type = PTR_TO_MEM | MEM_RDONLY;
20760	aux->btf_var.mem_size = 0;
20761	return 0;
20762	}
20763
20764	datasec_id = find_btf_percpu_datasec(btf);
20765	if (datasec_id > 0) {
20766	datasec = btf_type_by_id(btf, type_id: datasec_id);
20767	for_each_vsi(i, datasec, vsi) {
20768	if (vsi->type == id) {
20769	percpu = true;
20770	break;
20771	}
20772	}
20773	}
20774
20775	type = t->type;
20776	t = btf_type_skip_modifiers(btf, id: type, NULL);
20777	if (percpu) {
20778	aux->btf_var.reg_type = PTR_TO_BTF_ID | MEM_PERCPU;
20779	aux->btf_var.btf = btf;
20780	aux->btf_var.btf_id = type;
20781	} else if (!btf_type_is_struct(t)) {
20782	const struct btf_type *ret;
20783	const char *tname;
20784	u32 tsize;
20785
20786	/* resolve the type size of ksym. */
20787	ret = btf_resolve_size(btf, type: t, type_size: &tsize);
20788	if (IS_ERR(ptr: ret)) {
20789	tname = btf_name_by_offset(btf, offset: t->name_off);
20790	verbose(private_data: env, fmt: "ldimm64 unable to resolve the size of type '%s': %ld\n",
20791	tname, PTR_ERR(ptr: ret));
20792	return -EINVAL;
20793	}
20794	aux->btf_var.reg_type = PTR_TO_MEM | MEM_RDONLY;
20795	aux->btf_var.mem_size = tsize;
20796	} else {
20797	aux->btf_var.reg_type = PTR_TO_BTF_ID;
20798	aux->btf_var.btf = btf;
20799	aux->btf_var.btf_id = type;
20800	}
20801
20802	return 0;
20803	}
20804
20805	static int check_pseudo_btf_id(struct bpf_verifier_env *env,
20806	struct bpf_insn *insn,
20807	struct bpf_insn_aux_data *aux)
20808	{
20809	struct btf *btf;
20810	int btf_fd;
20811	int err;
20812
20813	btf_fd = insn[1].imm;
20814	if (btf_fd) {
20815	CLASS(fd, f)(fd: btf_fd);
20816
20817	btf = __btf_get_by_fd(f);
20818	if (IS_ERR(ptr: btf)) {
20819	verbose(private_data: env, fmt: "invalid module BTF object FD specified.\n");
20820	return -EINVAL;
20821	}
20822	} else {
20823	if (!btf_vmlinux) {
20824	verbose(private_data: env, fmt: "kernel is missing BTF, make sure CONFIG_DEBUG_INFO_BTF=y is specified in Kconfig.\n");
20825	return -EINVAL;
20826	}
20827	btf = btf_vmlinux;
20828	}
20829
20830	err = __check_pseudo_btf_id(env, insn, aux, btf);
20831	if (err)
20832	return err;
20833
20834	err = __add_used_btf(env, btf);
20835	if (err < 0)
20836	return err;
20837	return 0;
20838	}
20839
20840	static bool is_tracing_prog_type(enum bpf_prog_type type)
20841	{
20842	switch (type) {
20843	case BPF_PROG_TYPE_KPROBE:
20844	case BPF_PROG_TYPE_TRACEPOINT:
20845	case BPF_PROG_TYPE_PERF_EVENT:
20846	case BPF_PROG_TYPE_RAW_TRACEPOINT:
20847	case BPF_PROG_TYPE_RAW_TRACEPOINT_WRITABLE:
20848	return true;
20849	default:
20850	return false;
20851	}
20852	}
20853
20854	static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
20855	{
20856	return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
20857	map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
20858	}
20859
20860	static int check_map_prog_compatibility(struct bpf_verifier_env *env,
20861	struct bpf_map *map,
20862	struct bpf_prog *prog)
20863
20864	{
20865	enum bpf_prog_type prog_type = resolve_prog_type(prog);
20866
20867	if (map->excl_prog_sha &&
20868	memcmp(p: map->excl_prog_sha, q: prog->digest, SHA256_DIGEST_SIZE)) {
20869	verbose(private_data: env, fmt: "program's hash doesn't match map's excl_prog_hash\n");
20870	return -EACCES;
20871	}
20872
20873	if (btf_record_has_field(rec: map->record, type: BPF_LIST_HEAD) ||
20874	btf_record_has_field(rec: map->record, type: BPF_RB_ROOT)) {
20875	if (is_tracing_prog_type(type: prog_type)) {
20876	verbose(private_data: env, fmt: "tracing progs cannot use bpf_{list_head,rb_root} yet\n");
20877	return -EINVAL;
20878	}
20879	}
20880
20881	if (btf_record_has_field(rec: map->record, type: BPF_SPIN_LOCK | BPF_RES_SPIN_LOCK)) {
20882	if (prog_type == BPF_PROG_TYPE_SOCKET_FILTER) {
20883	verbose(private_data: env, fmt: "socket filter progs cannot use bpf_spin_lock yet\n");
20884	return -EINVAL;
20885	}
20886
20887	if (is_tracing_prog_type(type: prog_type)) {
20888	verbose(private_data: env, fmt: "tracing progs cannot use bpf_spin_lock yet\n");
20889	return -EINVAL;
20890	}
20891	}
20892
20893	if (btf_record_has_field(rec: map->record, type: BPF_TIMER)) {
20894	if (is_tracing_prog_type(type: prog_type)) {
20895	verbose(private_data: env, fmt: "tracing progs cannot use bpf_timer yet\n");
20896	return -EINVAL;
20897	}
20898	}
20899
20900	if (btf_record_has_field(rec: map->record, type: BPF_WORKQUEUE)) {
20901	if (is_tracing_prog_type(type: prog_type)) {
20902	verbose(private_data: env, fmt: "tracing progs cannot use bpf_wq yet\n");
20903	return -EINVAL;
20904	}
20905	}
20906
20907	if ((bpf_prog_is_offloaded(aux: prog->aux) || bpf_map_is_offloaded(map)) &&
20908	!bpf_offload_prog_map_match(prog, map)) {
20909	verbose(private_data: env, fmt: "offload device mismatch between prog and map\n");
20910	return -EINVAL;
20911	}
20912
20913	if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
20914	verbose(private_data: env, fmt: "bpf_struct_ops map cannot be used in prog\n");
20915	return -EINVAL;
20916	}
20917
20918	if (prog->sleepable)
20919	switch (map->map_type) {
20920	case BPF_MAP_TYPE_HASH:
20921	case BPF_MAP_TYPE_LRU_HASH:
20922	case BPF_MAP_TYPE_ARRAY:
20923	case BPF_MAP_TYPE_PERCPU_HASH:
20924	case BPF_MAP_TYPE_PERCPU_ARRAY:
20925	case BPF_MAP_TYPE_LRU_PERCPU_HASH:
20926	case BPF_MAP_TYPE_ARRAY_OF_MAPS:
20927	case BPF_MAP_TYPE_HASH_OF_MAPS:
20928	case BPF_MAP_TYPE_RINGBUF:
20929	case BPF_MAP_TYPE_USER_RINGBUF:
20930	case BPF_MAP_TYPE_INODE_STORAGE:
20931	case BPF_MAP_TYPE_SK_STORAGE:
20932	case BPF_MAP_TYPE_TASK_STORAGE:
20933	case BPF_MAP_TYPE_CGRP_STORAGE:
20934	case BPF_MAP_TYPE_QUEUE:
20935	case BPF_MAP_TYPE_STACK:
20936	case BPF_MAP_TYPE_ARENA:
20937	case BPF_MAP_TYPE_INSN_ARRAY:
20938	break;
20939	default:
20940	verbose(private_data: env,
20941	fmt: "Sleepable programs can only use array, hash, ringbuf and local storage maps\n");
20942	return -EINVAL;
20943	}
20944
20945	if (bpf_map_is_cgroup_storage(map) &&
20946	bpf_cgroup_storage_assign(aux: env->prog->aux, map)) {
20947	verbose(private_data: env, fmt: "only one cgroup storage of each type is allowed\n");
20948	return -EBUSY;
20949	}
20950
20951	if (map->map_type == BPF_MAP_TYPE_ARENA) {
20952	if (env->prog->aux->arena) {
20953	verbose(private_data: env, fmt: "Only one arena per program\n");
20954	return -EBUSY;
20955	}
20956	if (!env->allow_ptr_leaks || !env->bpf_capable) {
20957	verbose(private_data: env, fmt: "CAP_BPF and CAP_PERFMON are required to use arena\n");
20958	return -EPERM;
20959	}
20960	if (!env->prog->jit_requested) {
20961	verbose(private_data: env, fmt: "JIT is required to use arena\n");
20962	return -EOPNOTSUPP;
20963	}
20964	if (!bpf_jit_supports_arena()) {
20965	verbose(private_data: env, fmt: "JIT doesn't support arena\n");
20966	return -EOPNOTSUPP;
20967	}
20968	env->prog->aux->arena = (void *)map;
20969	if (!bpf_arena_get_user_vm_start(arena: env->prog->aux->arena)) {
20970	verbose(private_data: env, fmt: "arena's user address must be set via map_extra or mmap()\n");
20971	return -EINVAL;
20972	}
20973	}
20974
20975	return 0;
20976	}
20977
20978	static int __add_used_map(struct bpf_verifier_env *env, struct bpf_map *map)
20979	{
20980	int i, err;
20981
20982	/* check whether we recorded this map already */
20983	for (i = 0; i < env->used_map_cnt; i++)
20984	if (env->used_maps[i] == map)
20985	return i;
20986
20987	if (env->used_map_cnt >= MAX_USED_MAPS) {
20988	verbose(private_data: env, fmt: "The total number of maps per program has reached the limit of %u\n",
20989	MAX_USED_MAPS);
20990	return -E2BIG;
20991	}
20992
20993	err = check_map_prog_compatibility(env, map, prog: env->prog);
20994	if (err)
20995	return err;
20996
20997	if (env->prog->sleepable)
20998	atomic64_inc(v: &map->sleepable_refcnt);
20999
21000	/* hold the map. If the program is rejected by verifier,
21001	* the map will be released by release_maps() or it
21002	* will be used by the valid program until it's unloaded
21003	* and all maps are released in bpf_free_used_maps()
21004	*/
21005	bpf_map_inc(map);
21006
21007	env->used_maps[env->used_map_cnt++] = map;
21008
21009	if (map->map_type == BPF_MAP_TYPE_INSN_ARRAY) {
21010	err = bpf_insn_array_init(map, prog: env->prog);
21011	if (err) {
21012	verbose(private_data: env, fmt: "Failed to properly initialize insn array\n");
21013	return err;
21014	}
21015	env->insn_array_maps[env->insn_array_map_cnt++] = map;
21016	}
21017
21018	return env->used_map_cnt - 1;
21019	}
21020
21021	/* Add map behind fd to used maps list, if it's not already there, and return
21022	* its index.
21023	* Returns <0 on error, or >= 0 index, on success.
21024	*/
21025	static int add_used_map(struct bpf_verifier_env *env, int fd)
21026	{
21027	struct bpf_map *map;
21028	CLASS(fd, f)(fd);
21029
21030	map = __bpf_map_get(f);
21031	if (IS_ERR(ptr: map)) {
21032	verbose(private_data: env, fmt: "fd %d is not pointing to valid bpf_map\n", fd);
21033	return PTR_ERR(ptr: map);
21034	}
21035
21036	return __add_used_map(env, map);
21037	}
21038
21039	/* find and rewrite pseudo imm in ld_imm64 instructions:
21040	*
21041	* 1. if it accesses map FD, replace it with actual map pointer.
21042	* 2. if it accesses btf_id of a VAR, replace it with pointer to the var.
21043	*
21044	* NOTE: btf_vmlinux is required for converting pseudo btf_id.
21045	*/
21046	static int resolve_pseudo_ldimm64(struct bpf_verifier_env *env)
21047	{
21048	struct bpf_insn *insn = env->prog->insnsi;
21049	int insn_cnt = env->prog->len;
21050	int i, err;
21051
21052	err = bpf_prog_calc_tag(fp: env->prog);
21053	if (err)
21054	return err;
21055
21056	for (i = 0; i < insn_cnt; i++, insn++) {
21057	if (BPF_CLASS(insn->code) == BPF_LDX &&
21058	((BPF_MODE(insn->code) != BPF_MEM && BPF_MODE(insn->code) != BPF_MEMSX) ||
21059	insn->imm != 0)) {
21060	verbose(private_data: env, fmt: "BPF_LDX uses reserved fields\n");
21061	return -EINVAL;
21062	}
21063
21064	if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
21065	struct bpf_insn_aux_data *aux;
21066	struct bpf_map *map;
21067	int map_idx;
21068	u64 addr;
21069	u32 fd;
21070
21071	if (i == insn_cnt - 1 || insn[1].code != 0 ||
21072	insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
21073	insn[1].off != 0) {
21074	verbose(private_data: env, fmt: "invalid bpf_ld_imm64 insn\n");
21075	return -EINVAL;
21076	}
21077
21078	if (insn[0].src_reg == 0)
21079	/* valid generic load 64-bit imm */
21080	goto next_insn;
21081
21082	if (insn[0].src_reg == BPF_PSEUDO_BTF_ID) {
21083	aux = &env->insn_aux_data[i];
21084	err = check_pseudo_btf_id(env, insn, aux);
21085	if (err)
21086	return err;
21087	goto next_insn;
21088	}
21089
21090	if (insn[0].src_reg == BPF_PSEUDO_FUNC) {
21091	aux = &env->insn_aux_data[i];
21092	aux->ptr_type = PTR_TO_FUNC;
21093	goto next_insn;
21094	}
21095
21096	/* In final convert_pseudo_ld_imm64() step, this is
21097	* converted into regular 64-bit imm load insn.
21098	*/
21099	switch (insn[0].src_reg) {
21100	case BPF_PSEUDO_MAP_VALUE:
21101	case BPF_PSEUDO_MAP_IDX_VALUE:
21102	break;
21103	case BPF_PSEUDO_MAP_FD:
21104	case BPF_PSEUDO_MAP_IDX:
21105	if (insn[1].imm == 0)
21106	break;
21107	fallthrough;
21108	default:
21109	verbose(private_data: env, fmt: "unrecognized bpf_ld_imm64 insn\n");
21110	return -EINVAL;
21111	}
21112
21113	switch (insn[0].src_reg) {
21114	case BPF_PSEUDO_MAP_IDX_VALUE:
21115	case BPF_PSEUDO_MAP_IDX:
21116	if (bpfptr_is_null(bpfptr: env->fd_array)) {
21117	verbose(private_data: env, fmt: "fd_idx without fd_array is invalid\n");
21118	return -EPROTO;
21119	}
21120	if (copy_from_bpfptr_offset(dst: &fd, src: env->fd_array,
21121	offset: insn[0].imm * sizeof(fd),
21122	size: sizeof(fd)))
21123	return -EFAULT;
21124	break;
21125	default:
21126	fd = insn[0].imm;
21127	break;
21128	}
21129
21130	map_idx = add_used_map(env, fd);
21131	if (map_idx < 0)
21132	return map_idx;
21133	map = env->used_maps[map_idx];
21134
21135	aux = &env->insn_aux_data[i];
21136	aux->map_index = map_idx;
21137
21138	if (insn[0].src_reg == BPF_PSEUDO_MAP_FD ||
21139	insn[0].src_reg == BPF_PSEUDO_MAP_IDX) {
21140	addr = (unsigned long)map;
21141	} else {
21142	u32 off = insn[1].imm;
21143
21144	if (off >= BPF_MAX_VAR_OFF) {
21145	verbose(private_data: env, fmt: "direct value offset of %u is not allowed\n", off);
21146	return -EINVAL;
21147	}
21148
21149	if (!map->ops->map_direct_value_addr) {
21150	verbose(private_data: env, fmt: "no direct value access support for this map type\n");
21151	return -EINVAL;
21152	}
21153
21154	err = map->ops->map_direct_value_addr(map, &addr, off);
21155	if (err) {
21156	verbose(private_data: env, fmt: "invalid access to map value pointer, value_size=%u off=%u\n",
21157	map->value_size, off);
21158	return err;
21159	}
21160
21161	aux->map_off = off;
21162	addr += off;
21163	}
21164
21165	insn[0].imm = (u32)addr;
21166	insn[1].imm = addr >> 32;
21167
21168	next_insn:
21169	insn++;
21170	i++;
21171	continue;
21172	}
21173
21174	/* Basic sanity check before we invest more work here. */
21175	if (!bpf_opcode_in_insntable(code: insn->code)) {
21176	verbose(private_data: env, fmt: "unknown opcode %02x\n", insn->code);
21177	return -EINVAL;
21178	}
21179	}
21180
21181	/* now all pseudo BPF_LD_IMM64 instructions load valid
21182	* 'struct bpf_map *' into a register instead of user map_fd.
21183	* These pointers will be used later by verifier to validate map access.
21184	*/
21185	return 0;
21186	}
21187
21188	/* drop refcnt of maps used by the rejected program */
21189	static void release_maps(struct bpf_verifier_env *env)
21190	{
21191	__bpf_free_used_maps(aux: env->prog->aux, used_maps: env->used_maps,
21192	len: env->used_map_cnt);
21193	}
21194
21195	/* drop refcnt of maps used by the rejected program */
21196	static void release_btfs(struct bpf_verifier_env *env)
21197	{
21198	__bpf_free_used_btfs(used_btfs: env->used_btfs, len: env->used_btf_cnt);
21199	}
21200
21201	/* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
21202	static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
21203	{
21204	struct bpf_insn *insn = env->prog->insnsi;
21205	int insn_cnt = env->prog->len;
21206	int i;
21207
21208	for (i = 0; i < insn_cnt; i++, insn++) {
21209	if (insn->code != (BPF_LD | BPF_IMM | BPF_DW))
21210	continue;
21211	if (insn->src_reg == BPF_PSEUDO_FUNC)
21212	continue;
21213	insn->src_reg = 0;
21214	}
21215	}
21216
21217	/* single env->prog->insni[off] instruction was replaced with the range
21218	* insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
21219	* [0, off) and [off, end) to new locations, so the patched range stays zero
21220	*/
21221	static void adjust_insn_aux_data(struct bpf_verifier_env *env,
21222	struct bpf_prog *new_prog, u32 off, u32 cnt)
21223	{
21224	struct bpf_insn_aux_data *data = env->insn_aux_data;
21225	struct bpf_insn *insn = new_prog->insnsi;
21226	u32 old_seen = data[off].seen;
21227	u32 prog_len;
21228	int i;
21229
21230	/* aux info at OFF always needs adjustment, no matter fast path
21231	* (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
21232	* original insn at old prog.
21233	*/
21234	data[off].zext_dst = insn_has_def32(insn: insn + off + cnt - 1);
21235
21236	if (cnt == 1)
21237	return;
21238	prog_len = new_prog->len;
21239
21240	memmove(data + off + cnt - 1, data + off,
21241	sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
21242	memset(data + off, 0, sizeof(struct bpf_insn_aux_data) * (cnt - 1));
21243	for (i = off; i < off + cnt - 1; i++) {
21244	/* Expand insni[off]'s seen count to the patched range. */
21245	data[i].seen = old_seen;
21246	data[i].zext_dst = insn_has_def32(insn: insn + i);
21247	}
21248	}
21249
21250	static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
21251	{
21252	int i;
21253
21254	if (len == 1)
21255	return;
21256	/* NOTE: fake 'exit' subprog should be updated as well. */
21257	for (i = 0; i <= env->subprog_cnt; i++) {
21258	if (env->subprog_info[i].start <= off)
21259	continue;
21260	env->subprog_info[i].start += len - 1;
21261	}
21262	}
21263
21264	static void release_insn_arrays(struct bpf_verifier_env *env)
21265	{
21266	int i;
21267
21268	for (i = 0; i < env->insn_array_map_cnt; i++)
21269	bpf_insn_array_release(map: env->insn_array_maps[i]);
21270	}
21271
21272	static void adjust_insn_arrays(struct bpf_verifier_env *env, u32 off, u32 len)
21273	{
21274	int i;
21275
21276	if (len == 1)
21277	return;
21278
21279	for (i = 0; i < env->insn_array_map_cnt; i++)
21280	bpf_insn_array_adjust(map: env->insn_array_maps[i], off, len);
21281	}
21282
21283	static void adjust_insn_arrays_after_remove(struct bpf_verifier_env *env, u32 off, u32 len)
21284	{
21285	int i;
21286
21287	for (i = 0; i < env->insn_array_map_cnt; i++)
21288	bpf_insn_array_adjust_after_remove(map: env->insn_array_maps[i], off, len);
21289	}
21290
21291	static void adjust_poke_descs(struct bpf_prog *prog, u32 off, u32 len)
21292	{
21293	struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
21294	int i, sz = prog->aux->size_poke_tab;
21295	struct bpf_jit_poke_descriptor *desc;
21296
21297	for (i = 0; i < sz; i++) {
21298	desc = &tab[i];
21299	if (desc->insn_idx <= off)
21300	continue;
21301	desc->insn_idx += len - 1;
21302	}
21303	}
21304
21305	static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
21306	const struct bpf_insn *patch, u32 len)
21307	{
21308	struct bpf_prog *new_prog;
21309	struct bpf_insn_aux_data *new_data = NULL;
21310
21311	if (len > 1) {
21312	new_data = vrealloc(env->insn_aux_data,
21313	array_size(env->prog->len + len - 1,
21314	sizeof(struct bpf_insn_aux_data)),
21315	GFP_KERNEL_ACCOUNT | __GFP_ZERO);
21316	if (!new_data)
21317	return NULL;
21318
21319	env->insn_aux_data = new_data;
21320	}
21321
21322	new_prog = bpf_patch_insn_single(prog: env->prog, off, patch, len);
21323	if (IS_ERR(ptr: new_prog)) {
21324	if (PTR_ERR(ptr: new_prog) == -ERANGE)
21325	verbose(private_data: env,
21326	fmt: "insn %d cannot be patched due to 16-bit range\n",
21327	env->insn_aux_data[off].orig_idx);
21328	return NULL;
21329	}
21330	adjust_insn_aux_data(env, new_prog, off, cnt: len);
21331	adjust_subprog_starts(env, off, len);
21332	adjust_insn_arrays(env, off, len);
21333	adjust_poke_descs(prog: new_prog, off, len);
21334	return new_prog;
21335	}
21336
21337	/*
21338	* For all jmp insns in a given 'prog' that point to 'tgt_idx' insn adjust the
21339	* jump offset by 'delta'.
21340	*/
21341	static int adjust_jmp_off(struct bpf_prog *prog, u32 tgt_idx, u32 delta)
21342	{
21343	struct bpf_insn *insn = prog->insnsi;
21344	u32 insn_cnt = prog->len, i;
21345	s32 imm;
21346	s16 off;
21347
21348	for (i = 0; i < insn_cnt; i++, insn++) {
21349	u8 code = insn->code;
21350
21351	if (tgt_idx <= i && i < tgt_idx + delta)
21352	continue;
21353
21354	if ((BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32) ||
21355	BPF_OP(code) == BPF_CALL || BPF_OP(code) == BPF_EXIT)
21356	continue;
21357
21358	if (insn->code == (BPF_JMP32 | BPF_JA)) {
21359	if (i + 1 + insn->imm != tgt_idx)
21360	continue;
21361	if (check_add_overflow(insn->imm, delta, &imm))
21362	return -ERANGE;
21363	insn->imm = imm;
21364	} else {
21365	if (i + 1 + insn->off != tgt_idx)
21366	continue;
21367	if (check_add_overflow(insn->off, delta, &off))
21368	return -ERANGE;
21369	insn->off = off;
21370	}
21371	}
21372	return 0;
21373	}
21374
21375	static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
21376	u32 off, u32 cnt)
21377	{
21378	int i, j;
21379
21380	/* find first prog starting at or after off (first to remove) */
21381	for (i = 0; i < env->subprog_cnt; i++)
21382	if (env->subprog_info[i].start >= off)
21383	break;
21384	/* find first prog starting at or after off + cnt (first to stay) */
21385	for (j = i; j < env->subprog_cnt; j++)
21386	if (env->subprog_info[j].start >= off + cnt)
21387	break;
21388	/* if j doesn't start exactly at off + cnt, we are just removing
21389	* the front of previous prog
21390	*/
21391	if (env->subprog_info[j].start != off + cnt)
21392	j--;
21393
21394	if (j > i) {
21395	struct bpf_prog_aux *aux = env->prog->aux;
21396	int move;
21397
21398	/* move fake 'exit' subprog as well */
21399	move = env->subprog_cnt + 1 - j;
21400
21401	memmove(env->subprog_info + i,
21402	env->subprog_info + j,
21403	sizeof(*env->subprog_info) * move);
21404	env->subprog_cnt -= j - i;
21405
21406	/* remove func_info */
21407	if (aux->func_info) {
21408	move = aux->func_info_cnt - j;
21409
21410	memmove(aux->func_info + i,
21411	aux->func_info + j,
21412	sizeof(*aux->func_info) * move);
21413	aux->func_info_cnt -= j - i;
21414	/* func_info->insn_off is set after all code rewrites,
21415	* in adjust_btf_func() - no need to adjust
21416	*/
21417	}
21418	} else {
21419	/* convert i from "first prog to remove" to "first to adjust" */
21420	if (env->subprog_info[i].start == off)
21421	i++;
21422	}
21423
21424	/* update fake 'exit' subprog as well */
21425	for (; i <= env->subprog_cnt; i++)
21426	env->subprog_info[i].start -= cnt;
21427
21428	return 0;
21429	}
21430
21431	static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
21432	u32 cnt)
21433	{
21434	struct bpf_prog *prog = env->prog;
21435	u32 i, l_off, l_cnt, nr_linfo;
21436	struct bpf_line_info *linfo;
21437
21438	nr_linfo = prog->aux->nr_linfo;
21439	if (!nr_linfo)
21440	return 0;
21441
21442	linfo = prog->aux->linfo;
21443
21444	/* find first line info to remove, count lines to be removed */
21445	for (i = 0; i < nr_linfo; i++)
21446	if (linfo[i].insn_off >= off)
21447	break;
21448
21449	l_off = i;
21450	l_cnt = 0;
21451	for (; i < nr_linfo; i++)
21452	if (linfo[i].insn_off < off + cnt)
21453	l_cnt++;
21454	else
21455	break;
21456
21457	/* First live insn doesn't match first live linfo, it needs to "inherit"
21458	* last removed linfo. prog is already modified, so prog->len == off
21459	* means no live instructions after (tail of the program was removed).
21460	*/
21461	if (prog->len != off && l_cnt &&
21462	(i == nr_linfo || linfo[i].insn_off != off + cnt)) {
21463	l_cnt--;
21464	linfo[--i].insn_off = off + cnt;
21465	}
21466
21467	/* remove the line info which refer to the removed instructions */
21468	if (l_cnt) {
21469	memmove(linfo + l_off, linfo + i,
21470	sizeof(*linfo) * (nr_linfo - i));
21471
21472	prog->aux->nr_linfo -= l_cnt;
21473	nr_linfo = prog->aux->nr_linfo;
21474	}
21475
21476	/* pull all linfo[i].insn_off >= off + cnt in by cnt */
21477	for (i = l_off; i < nr_linfo; i++)
21478	linfo[i].insn_off -= cnt;
21479
21480	/* fix up all subprogs (incl. 'exit') which start >= off */
21481	for (i = 0; i <= env->subprog_cnt; i++)
21482	if (env->subprog_info[i].linfo_idx > l_off) {
21483	/* program may have started in the removed region but
21484	* may not be fully removed
21485	*/
21486	if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
21487	env->subprog_info[i].linfo_idx -= l_cnt;
21488	else
21489	env->subprog_info[i].linfo_idx = l_off;
21490	}
21491
21492	return 0;
21493	}
21494
21495	/*
21496	* Clean up dynamically allocated fields of aux data for instructions [start, ...]
21497	*/
21498	static void clear_insn_aux_data(struct bpf_verifier_env *env, int start, int len)
21499	{
21500	struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
21501	struct bpf_insn *insns = env->prog->insnsi;
21502	int end = start + len;
21503	int i;
21504
21505	for (i = start; i < end; i++) {
21506	if (aux_data[i].jt) {
21507	kvfree(addr: aux_data[i].jt);
21508	aux_data[i].jt = NULL;
21509	}
21510
21511	if (bpf_is_ldimm64(insn: &insns[i]))
21512	i++;
21513	}
21514	}
21515
21516	static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
21517	{
21518	struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
21519	unsigned int orig_prog_len = env->prog->len;
21520	int err;
21521
21522	if (bpf_prog_is_offloaded(aux: env->prog->aux))
21523	bpf_prog_offload_remove_insns(env, off, cnt);
21524
21525	/* Should be called before bpf_remove_insns, as it uses prog->insnsi */
21526	clear_insn_aux_data(env, start: off, len: cnt);
21527
21528	err = bpf_remove_insns(prog: env->prog, off, cnt);
21529	if (err)
21530	return err;
21531
21532	err = adjust_subprog_starts_after_remove(env, off, cnt);
21533	if (err)
21534	return err;
21535
21536	err = bpf_adj_linfo_after_remove(env, off, cnt);
21537	if (err)
21538	return err;
21539
21540	adjust_insn_arrays_after_remove(env, off, len: cnt);
21541
21542	memmove(aux_data + off, aux_data + off + cnt,
21543	sizeof(*aux_data) * (orig_prog_len - off - cnt));
21544
21545	return 0;
21546	}
21547
21548	/* The verifier does more data flow analysis than llvm and will not
21549	* explore branches that are dead at run time. Malicious programs can
21550	* have dead code too. Therefore replace all dead at-run-time code
21551	* with 'ja -1'.
21552	*
21553	* Just nops are not optimal, e.g. if they would sit at the end of the
21554	* program and through another bug we would manage to jump there, then
21555	* we'd execute beyond program memory otherwise. Returning exception
21556	* code also wouldn't work since we can have subprogs where the dead
21557	* code could be located.
21558	*/
21559	static void sanitize_dead_code(struct bpf_verifier_env *env)
21560	{
21561	struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
21562	struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
21563	struct bpf_insn *insn = env->prog->insnsi;
21564	const int insn_cnt = env->prog->len;
21565	int i;
21566
21567	for (i = 0; i < insn_cnt; i++) {
21568	if (aux_data[i].seen)
21569	continue;
21570	memcpy(insn + i, &trap, sizeof(trap));
21571	aux_data[i].zext_dst = false;
21572	}
21573	}
21574
21575	static bool insn_is_cond_jump(u8 code)
21576	{
21577	u8 op;
21578
21579	op = BPF_OP(code);
21580	if (BPF_CLASS(code) == BPF_JMP32)
21581	return op != BPF_JA;
21582
21583	if (BPF_CLASS(code) != BPF_JMP)
21584	return false;
21585
21586	return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
21587	}
21588
21589	static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
21590	{
21591	struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
21592	struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
21593	struct bpf_insn *insn = env->prog->insnsi;
21594	const int insn_cnt = env->prog->len;
21595	int i;
21596
21597	for (i = 0; i < insn_cnt; i++, insn++) {
21598	if (!insn_is_cond_jump(code: insn->code))
21599	continue;
21600
21601	if (!aux_data[i + 1].seen)
21602	ja.off = insn->off;
21603	else if (!aux_data[i + 1 + insn->off].seen)
21604	ja.off = 0;
21605	else
21606	continue;
21607
21608	if (bpf_prog_is_offloaded(aux: env->prog->aux))
21609	bpf_prog_offload_replace_insn(env, off: i, insn: &ja);
21610
21611	memcpy(insn, &ja, sizeof(ja));
21612	}
21613	}
21614
21615	static int opt_remove_dead_code(struct bpf_verifier_env *env)
21616	{
21617	struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
21618	int insn_cnt = env->prog->len;
21619	int i, err;
21620
21621	for (i = 0; i < insn_cnt; i++) {
21622	int j;
21623
21624	j = 0;
21625	while (i + j < insn_cnt && !aux_data[i + j].seen)
21626	j++;
21627	if (!j)
21628	continue;
21629
21630	err = verifier_remove_insns(env, off: i, cnt: j);
21631	if (err)
21632	return err;
21633	insn_cnt = env->prog->len;
21634	}
21635
21636	return 0;
21637	}
21638
21639	static const struct bpf_insn NOP = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
21640	static const struct bpf_insn MAY_GOTO_0 = BPF_RAW_INSN(BPF_JMP | BPF_JCOND, 0, 0, 0, 0);
21641
21642	static int opt_remove_nops(struct bpf_verifier_env *env)
21643	{
21644	struct bpf_insn *insn = env->prog->insnsi;
21645	int insn_cnt = env->prog->len;
21646	bool is_may_goto_0, is_ja;
21647	int i, err;
21648
21649	for (i = 0; i < insn_cnt; i++) {
21650	is_may_goto_0 = !memcmp(p: &insn[i], q: &MAY_GOTO_0, size: sizeof(MAY_GOTO_0));
21651	is_ja = !memcmp(p: &insn[i], q: &NOP, size: sizeof(NOP));
21652
21653	if (!is_may_goto_0 && !is_ja)
21654	continue;
21655
21656	err = verifier_remove_insns(env, off: i, cnt: 1);
21657	if (err)
21658	return err;
21659	insn_cnt--;
21660	/* Go back one insn to catch may_goto +1; may_goto +0 sequence */
21661	i -= (is_may_goto_0 && i > 0) ? 2 : 1;
21662	}
21663
21664	return 0;
21665	}
21666
21667	static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
21668	const union bpf_attr *attr)
21669	{
21670	struct bpf_insn *patch;
21671	/* use env->insn_buf as two independent buffers */
21672	struct bpf_insn *zext_patch = env->insn_buf;
21673	struct bpf_insn *rnd_hi32_patch = &env->insn_buf[2];
21674	struct bpf_insn_aux_data *aux = env->insn_aux_data;
21675	int i, patch_len, delta = 0, len = env->prog->len;
21676	struct bpf_insn *insns = env->prog->insnsi;
21677	struct bpf_prog *new_prog;
21678	bool rnd_hi32;
21679
21680	rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
21681	zext_patch[1] = BPF_ZEXT_REG(0);
21682	rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
21683	rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
21684	rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
21685	for (i = 0; i < len; i++) {
21686	int adj_idx = i + delta;
21687	struct bpf_insn insn;
21688	int load_reg;
21689
21690	insn = insns[adj_idx];
21691	load_reg = insn_def_regno(insn: &insn);
21692	if (!aux[adj_idx].zext_dst) {
21693	u8 code, class;
21694	u32 imm_rnd;
21695
21696	if (!rnd_hi32)
21697	continue;
21698
21699	code = insn.code;
21700	class = BPF_CLASS(code);
21701	if (load_reg == -1)
21702	continue;
21703
21704	/* NOTE: arg "reg" (the fourth one) is only used for
21705	* BPF_STX + SRC_OP, so it is safe to pass NULL
21706	* here.
21707	*/
21708	if (is_reg64(insn: &insn, regno: load_reg, NULL, t: DST_OP)) {
21709	if (class == BPF_LD &&
21710	BPF_MODE(code) == BPF_IMM)
21711	i++;
21712	continue;
21713	}
21714
21715	/* ctx load could be transformed into wider load. */
21716	if (class == BPF_LDX &&
21717	aux[adj_idx].ptr_type == PTR_TO_CTX)
21718	continue;
21719
21720	imm_rnd = get_random_u32();
21721	rnd_hi32_patch[0] = insn;
21722	rnd_hi32_patch[1].imm = imm_rnd;
21723	rnd_hi32_patch[3].dst_reg = load_reg;
21724	patch = rnd_hi32_patch;
21725	patch_len = 4;
21726	goto apply_patch_buffer;
21727	}
21728
21729	/* Add in an zero-extend instruction if a) the JIT has requested
21730	* it or b) it's a CMPXCHG.
21731	*
21732	* The latter is because: BPF_CMPXCHG always loads a value into
21733	* R0, therefore always zero-extends. However some archs'
21734	* equivalent instruction only does this load when the
21735	* comparison is successful. This detail of CMPXCHG is
21736	* orthogonal to the general zero-extension behaviour of the
21737	* CPU, so it's treated independently of bpf_jit_needs_zext.
21738	*/
21739	if (!bpf_jit_needs_zext() && !is_cmpxchg_insn(insn: &insn))
21740	continue;
21741
21742	/* Zero-extension is done by the caller. */
21743	if (bpf_pseudo_kfunc_call(insn: &insn))
21744	continue;
21745
21746	if (verifier_bug_if(load_reg == -1, env,
21747	"zext_dst is set, but no reg is defined"))
21748	return -EFAULT;
21749
21750	zext_patch[0] = insn;
21751	zext_patch[1].dst_reg = load_reg;
21752	zext_patch[1].src_reg = load_reg;
21753	patch = zext_patch;
21754	patch_len = 2;
21755	apply_patch_buffer:
21756	new_prog = bpf_patch_insn_data(env, off: adj_idx, patch, len: patch_len);
21757	if (!new_prog)
21758	return -ENOMEM;
21759	env->prog = new_prog;
21760	insns = new_prog->insnsi;
21761	aux = env->insn_aux_data;
21762	delta += patch_len - 1;
21763	}
21764
21765	return 0;
21766	}
21767
21768	/* convert load instructions that access fields of a context type into a
21769	* sequence of instructions that access fields of the underlying structure:
21770	* struct __sk_buff -> struct sk_buff
21771	* struct bpf_sock_ops -> struct sock
21772	*/
21773	static int convert_ctx_accesses(struct bpf_verifier_env *env)
21774	{
21775	struct bpf_subprog_info *subprogs = env->subprog_info;
21776	const struct bpf_verifier_ops *ops = env->ops;
21777	int i, cnt, size, ctx_field_size, ret, delta = 0, epilogue_cnt = 0;
21778	const int insn_cnt = env->prog->len;
21779	struct bpf_insn *epilogue_buf = env->epilogue_buf;
21780	struct bpf_insn *insn_buf = env->insn_buf;
21781	struct bpf_insn *insn;
21782	u32 target_size, size_default, off;
21783	struct bpf_prog *new_prog;
21784	enum bpf_access_type type;
21785	bool is_narrower_load;
21786	int epilogue_idx = 0;
21787
21788	if (ops->gen_epilogue) {
21789	epilogue_cnt = ops->gen_epilogue(epilogue_buf, env->prog,
21790	-(subprogs[0].stack_depth + 8));
21791	if (epilogue_cnt >= INSN_BUF_SIZE) {
21792	verifier_bug(env, "epilogue is too long");
21793	return -EFAULT;
21794	} else if (epilogue_cnt) {
21795	/* Save the ARG_PTR_TO_CTX for the epilogue to use */
21796	cnt = 0;
21797	subprogs[0].stack_depth += 8;
21798	insn_buf[cnt++] = BPF_STX_MEM(BPF_DW, BPF_REG_FP, BPF_REG_1,
21799	-subprogs[0].stack_depth);
21800	insn_buf[cnt++] = env->prog->insnsi[0];
21801	new_prog = bpf_patch_insn_data(env, off: 0, patch: insn_buf, len: cnt);
21802	if (!new_prog)
21803	return -ENOMEM;
21804	env->prog = new_prog;
21805	delta += cnt - 1;
21806
21807	ret = add_kfunc_in_insns(env, insn: epilogue_buf, cnt: epilogue_cnt - 1);
21808	if (ret < 0)
21809	return ret;
21810	}
21811	}
21812
21813	if (ops->gen_prologue || env->seen_direct_write) {
21814	if (!ops->gen_prologue) {
21815	verifier_bug(env, "gen_prologue is null");
21816	return -EFAULT;
21817	}
21818	cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
21819	env->prog);
21820	if (cnt >= INSN_BUF_SIZE) {
21821	verifier_bug(env, "prologue is too long");
21822	return -EFAULT;
21823	} else if (cnt) {
21824	new_prog = bpf_patch_insn_data(env, off: 0, patch: insn_buf, len: cnt);
21825	if (!new_prog)
21826	return -ENOMEM;
21827
21828	env->prog = new_prog;
21829	delta += cnt - 1;
21830
21831	ret = add_kfunc_in_insns(env, insn: insn_buf, cnt: cnt - 1);
21832	if (ret < 0)
21833	return ret;
21834	}
21835	}
21836
21837	if (delta)
21838	WARN_ON(adjust_jmp_off(env->prog, 0, delta));
21839
21840	if (bpf_prog_is_offloaded(aux: env->prog->aux))
21841	return 0;
21842
21843	insn = env->prog->insnsi + delta;
21844
21845	for (i = 0; i < insn_cnt; i++, insn++) {
21846	bpf_convert_ctx_access_t convert_ctx_access;
21847	u8 mode;
21848
21849	if (env->insn_aux_data[i + delta].nospec) {
21850	WARN_ON_ONCE(env->insn_aux_data[i + delta].alu_state);
21851	struct bpf_insn *patch = insn_buf;
21852
21853	*patch++ = BPF_ST_NOSPEC();
21854	*patch++ = *insn;
21855	cnt = patch - insn_buf;
21856	new_prog = bpf_patch_insn_data(env, off: i + delta, patch: insn_buf, len: cnt);
21857	if (!new_prog)
21858	return -ENOMEM;
21859
21860	delta += cnt - 1;
21861	env->prog = new_prog;
21862	insn = new_prog->insnsi + i + delta;
21863	/* This can not be easily merged with the
21864	* nospec_result-case, because an insn may require a
21865	* nospec before and after itself. Therefore also do not
21866	* 'continue' here but potentially apply further
21867	* patching to insn. *insn should equal patch[1] now.
21868	*/
21869	}
21870
21871	if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
21872	insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
21873	insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
21874	insn->code == (BPF_LDX | BPF_MEM | BPF_DW) ||
21875	insn->code == (BPF_LDX | BPF_MEMSX | BPF_B) ||
21876	insn->code == (BPF_LDX | BPF_MEMSX | BPF_H) ||
21877	insn->code == (BPF_LDX | BPF_MEMSX | BPF_W)) {
21878	type = BPF_READ;
21879	} else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
21880	insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
21881	insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
21882	insn->code == (BPF_STX | BPF_MEM | BPF_DW) ||
21883	insn->code == (BPF_ST | BPF_MEM | BPF_B) ||
21884	insn->code == (BPF_ST | BPF_MEM | BPF_H) ||
21885	insn->code == (BPF_ST | BPF_MEM | BPF_W) ||
21886	insn->code == (BPF_ST | BPF_MEM | BPF_DW)) {
21887	type = BPF_WRITE;
21888	} else if ((insn->code == (BPF_STX | BPF_ATOMIC | BPF_B) ||
21889	insn->code == (BPF_STX | BPF_ATOMIC | BPF_H) ||
21890	insn->code == (BPF_STX | BPF_ATOMIC | BPF_W) ||
21891	insn->code == (BPF_STX | BPF_ATOMIC | BPF_DW)) &&
21892	env->insn_aux_data[i + delta].ptr_type == PTR_TO_ARENA) {
21893	insn->code = BPF_STX | BPF_PROBE_ATOMIC | BPF_SIZE(insn->code);
21894	env->prog->aux->num_exentries++;
21895	continue;
21896	} else if (insn->code == (BPF_JMP | BPF_EXIT) &&
21897	epilogue_cnt &&
21898	i + delta < subprogs[1].start) {
21899	/* Generate epilogue for the main prog */
21900	if (epilogue_idx) {
21901	/* jump back to the earlier generated epilogue */
21902	insn_buf[0] = BPF_JMP32_A(epilogue_idx - i - delta - 1);
21903	cnt = 1;
21904	} else {
21905	memcpy(insn_buf, epilogue_buf,
21906	epilogue_cnt * sizeof(*epilogue_buf));
21907	cnt = epilogue_cnt;
21908	/* epilogue_idx cannot be 0. It must have at
21909	* least one ctx ptr saving insn before the
21910	* epilogue.
21911	*/
21912	epilogue_idx = i + delta;
21913	}
21914	goto patch_insn_buf;
21915	} else {
21916	continue;
21917	}
21918
21919	if (type == BPF_WRITE &&
21920	env->insn_aux_data[i + delta].nospec_result) {
21921	/* nospec_result is only used to mitigate Spectre v4 and
21922	* to limit verification-time for Spectre v1.
21923	*/
21924	struct bpf_insn *patch = insn_buf;
21925
21926	*patch++ = *insn;
21927	*patch++ = BPF_ST_NOSPEC();
21928	cnt = patch - insn_buf;
21929	new_prog = bpf_patch_insn_data(env, off: i + delta, patch: insn_buf, len: cnt);
21930	if (!new_prog)
21931	return -ENOMEM;
21932
21933	delta += cnt - 1;
21934	env->prog = new_prog;
21935	insn = new_prog->insnsi + i + delta;
21936	continue;
21937	}
21938
21939	switch ((int)env->insn_aux_data[i + delta].ptr_type) {
21940	case PTR_TO_CTX:
21941	if (!ops->convert_ctx_access)
21942	continue;
21943	convert_ctx_access = ops->convert_ctx_access;
21944	break;
21945	case PTR_TO_SOCKET:
21946	case PTR_TO_SOCK_COMMON:
21947	convert_ctx_access = bpf_sock_convert_ctx_access;
21948	break;
21949	case PTR_TO_TCP_SOCK:
21950	convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
21951	break;
21952	case PTR_TO_XDP_SOCK:
21953	convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
21954	break;
21955	case PTR_TO_BTF_ID:
21956	case PTR_TO_BTF_ID | PTR_UNTRUSTED:
21957	/* PTR_TO_BTF_ID | MEM_ALLOC always has a valid lifetime, unlike
21958	* PTR_TO_BTF_ID, and an active ref_obj_id, but the same cannot
21959	* be said once it is marked PTR_UNTRUSTED, hence we must handle
21960	* any faults for loads into such types. BPF_WRITE is disallowed
21961	* for this case.
21962	*/
21963	case PTR_TO_BTF_ID | MEM_ALLOC | PTR_UNTRUSTED:
21964	case PTR_TO_MEM | MEM_RDONLY | PTR_UNTRUSTED:
21965	if (type == BPF_READ) {
21966	if (BPF_MODE(insn->code) == BPF_MEM)
21967	insn->code = BPF_LDX | BPF_PROBE_MEM |
21968	BPF_SIZE((insn)->code);
21969	else
21970	insn->code = BPF_LDX | BPF_PROBE_MEMSX |
21971	BPF_SIZE((insn)->code);
21972	env->prog->aux->num_exentries++;
21973	}
21974	continue;
21975	case PTR_TO_ARENA:
21976	if (BPF_MODE(insn->code) == BPF_MEMSX) {
21977	if (!bpf_jit_supports_insn(insn, in_arena: true)) {
21978	verbose(private_data: env, fmt: "sign extending loads from arena are not supported yet\n");
21979	return -EOPNOTSUPP;
21980	}
21981	insn->code = BPF_CLASS(insn->code) | BPF_PROBE_MEM32SX | BPF_SIZE(insn->code);
21982	} else {
21983	insn->code = BPF_CLASS(insn->code) | BPF_PROBE_MEM32 | BPF_SIZE(insn->code);
21984	}
21985	env->prog->aux->num_exentries++;
21986	continue;
21987	default:
21988	continue;
21989	}
21990
21991	ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
21992	size = BPF_LDST_BYTES(insn);
21993	mode = BPF_MODE(insn->code);
21994
21995	/* If the read access is a narrower load of the field,
21996	* convert to a 4/8-byte load, to minimum program type specific
21997	* convert_ctx_access changes. If conversion is successful,
21998	* we will apply proper mask to the result.
21999	*/
22000	is_narrower_load = size < ctx_field_size;
22001	size_default = bpf_ctx_off_adjust_machine(size: ctx_field_size);
22002	off = insn->off;
22003	if (is_narrower_load) {
22004	u8 size_code;
22005
22006	if (type == BPF_WRITE) {
22007	verifier_bug(env, "narrow ctx access misconfigured");
22008	return -EFAULT;
22009	}
22010
22011	size_code = BPF_H;
22012	if (ctx_field_size == 4)
22013	size_code = BPF_W;
22014	else if (ctx_field_size == 8)
22015	size_code = BPF_DW;
22016
22017	insn->off = off & ~(size_default - 1);
22018	insn->code = BPF_LDX | BPF_MEM | size_code;
22019	}
22020
22021	target_size = 0;
22022	cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
22023	&target_size);
22024	if (cnt == 0 || cnt >= INSN_BUF_SIZE ||
22025	(ctx_field_size && !target_size)) {
22026	verifier_bug(env, "error during ctx access conversion (%d)", cnt);
22027	return -EFAULT;
22028	}
22029
22030	if (is_narrower_load && size < target_size) {
22031	u8 shift = bpf_ctx_narrow_access_offset(
22032	off, size, size_default) * 8;
22033	if (shift && cnt + 1 >= INSN_BUF_SIZE) {
22034	verifier_bug(env, "narrow ctx load misconfigured");
22035	return -EFAULT;
22036	}
22037	if (ctx_field_size <= 4) {
22038	if (shift)
22039	insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
22040	insn->dst_reg,
22041	shift);
22042	insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
22043	(1 << size * 8) - 1);
22044	} else {
22045	if (shift)
22046	insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
22047	insn->dst_reg,
22048	shift);
22049	insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
22050	(1ULL << size * 8) - 1);
22051	}
22052	}
22053	if (mode == BPF_MEMSX)
22054	insn_buf[cnt++] = BPF_RAW_INSN(BPF_ALU64 | BPF_MOV | BPF_X,
22055	insn->dst_reg, insn->dst_reg,
22056	size * 8, 0);
22057
22058	patch_insn_buf:
22059	new_prog = bpf_patch_insn_data(env, off: i + delta, patch: insn_buf, len: cnt);
22060	if (!new_prog)
22061	return -ENOMEM;
22062
22063	delta += cnt - 1;
22064
22065	/* keep walking new program and skip insns we just inserted */
22066	env->prog = new_prog;
22067	insn = new_prog->insnsi + i + delta;
22068	}
22069
22070	return 0;
22071	}
22072
22073	static int jit_subprogs(struct bpf_verifier_env *env)
22074	{
22075	struct bpf_prog *prog = env->prog, **func, *tmp;
22076	int i, j, subprog_start, subprog_end = 0, len, subprog;
22077	struct bpf_map *map_ptr;
22078	struct bpf_insn *insn;
22079	void *old_bpf_func;
22080	int err, num_exentries;
22081	int old_len, subprog_start_adjustment = 0;
22082
22083	if (env->subprog_cnt <= 1)
22084	return 0;
22085
22086	for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
22087	if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn))
22088	continue;
22089
22090	/* Upon error here we cannot fall back to interpreter but
22091	* need a hard reject of the program. Thus -EFAULT is
22092	* propagated in any case.
22093	*/
22094	subprog = find_subprog(env, off: i + insn->imm + 1);
22095	if (verifier_bug_if(subprog < 0, env, "No program to jit at insn %d",
22096	i + insn->imm + 1))
22097	return -EFAULT;
22098	/* temporarily remember subprog id inside insn instead of
22099	* aux_data, since next loop will split up all insns into funcs
22100	*/
22101	insn->off = subprog;
22102	/* remember original imm in case JIT fails and fallback
22103	* to interpreter will be needed
22104	*/
22105	env->insn_aux_data[i].call_imm = insn->imm;
22106	/* point imm to __bpf_call_base+1 from JITs point of view */
22107	insn->imm = 1;
22108	if (bpf_pseudo_func(insn)) {
22109	#if defined(MODULES_VADDR)
22110	u64 addr = MODULES_VADDR;
22111	#else
22112	u64 addr = VMALLOC_START;
22113	#endif
22114	/* jit (e.g. x86_64) may emit fewer instructions
22115	* if it learns a u32 imm is the same as a u64 imm.
22116	* Set close enough to possible prog address.
22117	*/
22118	insn[0].imm = (u32)addr;
22119	insn[1].imm = addr >> 32;
22120	}
22121	}
22122
22123	err = bpf_prog_alloc_jited_linfo(prog);
22124	if (err)
22125	goto out_undo_insn;
22126
22127	err = -ENOMEM;
22128	func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
22129	if (!func)
22130	goto out_undo_insn;
22131
22132	for (i = 0; i < env->subprog_cnt; i++) {
22133	subprog_start = subprog_end;
22134	subprog_end = env->subprog_info[i + 1].start;
22135
22136	len = subprog_end - subprog_start;
22137	/* bpf_prog_run() doesn't call subprogs directly,
22138	* hence main prog stats include the runtime of subprogs.
22139	* subprogs don't have IDs and not reachable via prog_get_next_id
22140	* func[i]->stats will never be accessed and stays NULL
22141	*/
22142	func[i] = bpf_prog_alloc_no_stats(size: bpf_prog_size(proglen: len), GFP_USER);
22143	if (!func[i])
22144	goto out_free;
22145	memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
22146	len * sizeof(struct bpf_insn));
22147	func[i]->type = prog->type;
22148	func[i]->len = len;
22149	if (bpf_prog_calc_tag(fp: func[i]))
22150	goto out_free;
22151	func[i]->is_func = 1;
22152	func[i]->sleepable = prog->sleepable;
22153	func[i]->aux->func_idx = i;
22154	/* Below members will be freed only at prog->aux */
22155	func[i]->aux->btf = prog->aux->btf;
22156	func[i]->aux->subprog_start = subprog_start + subprog_start_adjustment;
22157	func[i]->aux->func_info = prog->aux->func_info;
22158	func[i]->aux->func_info_cnt = prog->aux->func_info_cnt;
22159	func[i]->aux->poke_tab = prog->aux->poke_tab;
22160	func[i]->aux->size_poke_tab = prog->aux->size_poke_tab;
22161	func[i]->aux->main_prog_aux = prog->aux;
22162
22163	for (j = 0; j < prog->aux->size_poke_tab; j++) {
22164	struct bpf_jit_poke_descriptor *poke;
22165
22166	poke = &prog->aux->poke_tab[j];
22167	if (poke->insn_idx < subprog_end &&
22168	poke->insn_idx >= subprog_start)
22169	poke->aux = func[i]->aux;
22170	}
22171
22172	func[i]->aux->name[0] = 'F';
22173	func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
22174	if (env->subprog_info[i].priv_stack_mode == PRIV_STACK_ADAPTIVE)
22175	func[i]->aux->jits_use_priv_stack = true;
22176
22177	func[i]->jit_requested = 1;
22178	func[i]->blinding_requested = prog->blinding_requested;
22179	func[i]->aux->kfunc_tab = prog->aux->kfunc_tab;
22180	func[i]->aux->kfunc_btf_tab = prog->aux->kfunc_btf_tab;
22181	func[i]->aux->linfo = prog->aux->linfo;
22182	func[i]->aux->nr_linfo = prog->aux->nr_linfo;
22183	func[i]->aux->jited_linfo = prog->aux->jited_linfo;
22184	func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
22185	func[i]->aux->arena = prog->aux->arena;
22186	func[i]->aux->used_maps = env->used_maps;
22187	func[i]->aux->used_map_cnt = env->used_map_cnt;
22188	num_exentries = 0;
22189	insn = func[i]->insnsi;
22190	for (j = 0; j < func[i]->len; j++, insn++) {
22191	if (BPF_CLASS(insn->code) == BPF_LDX &&
22192	(BPF_MODE(insn->code) == BPF_PROBE_MEM ||
22193	BPF_MODE(insn->code) == BPF_PROBE_MEM32 ||
22194	BPF_MODE(insn->code) == BPF_PROBE_MEM32SX ||
22195	BPF_MODE(insn->code) == BPF_PROBE_MEMSX))
22196	num_exentries++;
22197	if ((BPF_CLASS(insn->code) == BPF_STX ||
22198	BPF_CLASS(insn->code) == BPF_ST) &&
22199	BPF_MODE(insn->code) == BPF_PROBE_MEM32)
22200	num_exentries++;
22201	if (BPF_CLASS(insn->code) == BPF_STX &&
22202	BPF_MODE(insn->code) == BPF_PROBE_ATOMIC)
22203	num_exentries++;
22204	}
22205	func[i]->aux->num_exentries = num_exentries;
22206	func[i]->aux->tail_call_reachable = env->subprog_info[i].tail_call_reachable;
22207	func[i]->aux->exception_cb = env->subprog_info[i].is_exception_cb;
22208	func[i]->aux->changes_pkt_data = env->subprog_info[i].changes_pkt_data;
22209	func[i]->aux->might_sleep = env->subprog_info[i].might_sleep;
22210	if (!i)
22211	func[i]->aux->exception_boundary = env->seen_exception;
22212
22213	/*
22214	* To properly pass the absolute subprog start to jit
22215	* all instruction adjustments should be accumulated
22216	*/
22217	old_len = func[i]->len;
22218	func[i] = bpf_int_jit_compile(prog: func[i]);
22219	subprog_start_adjustment += func[i]->len - old_len;
22220
22221	if (!func[i]->jited) {
22222	err = -ENOTSUPP;
22223	goto out_free;
22224	}
22225	cond_resched();
22226	}
22227
22228	/* at this point all bpf functions were successfully JITed
22229	* now populate all bpf_calls with correct addresses and
22230	* run last pass of JIT
22231	*/
22232	for (i = 0; i < env->subprog_cnt; i++) {
22233	insn = func[i]->insnsi;
22234	for (j = 0; j < func[i]->len; j++, insn++) {
22235	if (bpf_pseudo_func(insn)) {
22236	subprog = insn->off;
22237	insn[0].imm = (u32)(long)func[subprog]->bpf_func;
22238	insn[1].imm = ((u64)(long)func[subprog]->bpf_func) >> 32;
22239	continue;
22240	}
22241	if (!bpf_pseudo_call(insn))
22242	continue;
22243	subprog = insn->off;
22244	insn->imm = BPF_CALL_IMM(func[subprog]->bpf_func);
22245	}
22246
22247	/* we use the aux data to keep a list of the start addresses
22248	* of the JITed images for each function in the program
22249	*
22250	* for some architectures, such as powerpc64, the imm field
22251	* might not be large enough to hold the offset of the start
22252	* address of the callee's JITed image from __bpf_call_base
22253	*
22254	* in such cases, we can lookup the start address of a callee
22255	* by using its subprog id, available from the off field of
22256	* the call instruction, as an index for this list
22257	*/
22258	func[i]->aux->func = func;
22259	func[i]->aux->func_cnt = env->subprog_cnt - env->hidden_subprog_cnt;
22260	func[i]->aux->real_func_cnt = env->subprog_cnt;
22261	}
22262	for (i = 0; i < env->subprog_cnt; i++) {
22263	old_bpf_func = func[i]->bpf_func;
22264	tmp = bpf_int_jit_compile(prog: func[i]);
22265	if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
22266	verbose(private_data: env, fmt: "JIT doesn't support bpf-to-bpf calls\n");
22267	err = -ENOTSUPP;
22268	goto out_free;
22269	}
22270	cond_resched();
22271	}
22272
22273	/*
22274	* Cleanup func[i]->aux fields which aren't required
22275	* or can become invalid in future
22276	*/
22277	for (i = 0; i < env->subprog_cnt; i++) {
22278	func[i]->aux->used_maps = NULL;
22279	func[i]->aux->used_map_cnt = 0;
22280	}
22281
22282	/* finally lock prog and jit images for all functions and
22283	* populate kallsysm. Begin at the first subprogram, since
22284	* bpf_prog_load will add the kallsyms for the main program.
22285	*/
22286	for (i = 1; i < env->subprog_cnt; i++) {
22287	err = bpf_prog_lock_ro(fp: func[i]);
22288	if (err)
22289	goto out_free;
22290	}
22291
22292	for (i = 1; i < env->subprog_cnt; i++)
22293	bpf_prog_kallsyms_add(fp: func[i]);
22294
22295	/* Last step: make now unused interpreter insns from main
22296	* prog consistent for later dump requests, so they can
22297	* later look the same as if they were interpreted only.
22298	*/
22299	for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
22300	if (bpf_pseudo_func(insn)) {
22301	insn[0].imm = env->insn_aux_data[i].call_imm;
22302	insn[1].imm = insn->off;
22303	insn->off = 0;
22304	continue;
22305	}
22306	if (!bpf_pseudo_call(insn))
22307	continue;
22308	insn->off = env->insn_aux_data[i].call_imm;
22309	subprog = find_subprog(env, off: i + insn->off + 1);
22310	insn->imm = subprog;
22311	}
22312
22313	prog->jited = 1;
22314	prog->bpf_func = func[0]->bpf_func;
22315	prog->jited_len = func[0]->jited_len;
22316	prog->aux->extable = func[0]->aux->extable;
22317	prog->aux->num_exentries = func[0]->aux->num_exentries;
22318	prog->aux->func = func;
22319	prog->aux->func_cnt = env->subprog_cnt - env->hidden_subprog_cnt;
22320	prog->aux->real_func_cnt = env->subprog_cnt;
22321	prog->aux->bpf_exception_cb = (void *)func[env->exception_callback_subprog]->bpf_func;
22322	prog->aux->exception_boundary = func[0]->aux->exception_boundary;
22323	bpf_prog_jit_attempt_done(prog);
22324	return 0;
22325	out_free:
22326	/* We failed JIT'ing, so at this point we need to unregister poke
22327	* descriptors from subprogs, so that kernel is not attempting to
22328	* patch it anymore as we're freeing the subprog JIT memory.
22329	*/
22330	for (i = 0; i < prog->aux->size_poke_tab; i++) {
22331	map_ptr = prog->aux->poke_tab[i].tail_call.map;
22332	map_ptr->ops->map_poke_untrack(map_ptr, prog->aux);
22333	}
22334	/* At this point we're guaranteed that poke descriptors are not
22335	* live anymore. We can just unlink its descriptor table as it's
22336	* released with the main prog.
22337	*/
22338	for (i = 0; i < env->subprog_cnt; i++) {
22339	if (!func[i])
22340	continue;
22341	func[i]->aux->poke_tab = NULL;
22342	bpf_jit_free(fp: func[i]);
22343	}
22344	kfree(objp: func);
22345	out_undo_insn:
22346	/* cleanup main prog to be interpreted */
22347	prog->jit_requested = 0;
22348	prog->blinding_requested = 0;
22349	for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
22350	if (!bpf_pseudo_call(insn))
22351	continue;
22352	insn->off = 0;
22353	insn->imm = env->insn_aux_data[i].call_imm;
22354	}
22355	bpf_prog_jit_attempt_done(prog);
22356	return err;
22357	}
22358
22359	static int fixup_call_args(struct bpf_verifier_env *env)
22360	{
22361	#ifndef CONFIG_BPF_JIT_ALWAYS_ON
22362	struct bpf_prog *prog = env->prog;
22363	struct bpf_insn *insn = prog->insnsi;
22364	bool has_kfunc_call = bpf_prog_has_kfunc_call(prog);
22365	int i, depth;
22366	#endif
22367	int err = 0;
22368
22369	if (env->prog->jit_requested &&
22370	!bpf_prog_is_offloaded(aux: env->prog->aux)) {
22371	err = jit_subprogs(env);
22372	if (err == 0)
22373	return 0;
22374	if (err == -EFAULT)
22375	return err;
22376	}
22377	#ifndef CONFIG_BPF_JIT_ALWAYS_ON
22378	if (has_kfunc_call) {
22379	verbose(env, "calling kernel functions are not allowed in non-JITed programs\n");
22380	return -EINVAL;
22381	}
22382	if (env->subprog_cnt > 1 && env->prog->aux->tail_call_reachable) {
22383	/* When JIT fails the progs with bpf2bpf calls and tail_calls
22384	* have to be rejected, since interpreter doesn't support them yet.
22385	*/
22386	verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
22387	return -EINVAL;
22388	}
22389	for (i = 0; i < prog->len; i++, insn++) {
22390	if (bpf_pseudo_func(insn)) {
22391	/* When JIT fails the progs with callback calls
22392	* have to be rejected, since interpreter doesn't support them yet.
22393	*/
22394	verbose(env, "callbacks are not allowed in non-JITed programs\n");
22395	return -EINVAL;
22396	}
22397
22398	if (!bpf_pseudo_call(insn))
22399	continue;
22400	depth = get_callee_stack_depth(env, insn, i);
22401	if (depth < 0)
22402	return depth;
22403	bpf_patch_call_args(insn, depth);
22404	}
22405	err = 0;
22406	#endif
22407	return err;
22408	}
22409
22410	/* replace a generic kfunc with a specialized version if necessary */
22411	static int specialize_kfunc(struct bpf_verifier_env *env, struct bpf_kfunc_desc *desc, int insn_idx)
22412	{
22413	struct bpf_prog *prog = env->prog;
22414	bool seen_direct_write;
22415	void *xdp_kfunc;
22416	bool is_rdonly;
22417	u32 func_id = desc->func_id;
22418	u16 offset = desc->offset;
22419	unsigned long addr = desc->addr;
22420
22421	if (offset) /* return if module BTF is used */
22422	return 0;
22423
22424	if (bpf_dev_bound_kfunc_id(btf_id: func_id)) {
22425	xdp_kfunc = bpf_dev_bound_resolve_kfunc(prog, func_id);
22426	if (xdp_kfunc)
22427	addr = (unsigned long)xdp_kfunc;
22428	/* fallback to default kfunc when not supported by netdev */
22429	} else if (func_id == special_kfunc_list[KF_bpf_dynptr_from_skb]) {
22430	seen_direct_write = env->seen_direct_write;
22431	is_rdonly = !may_access_direct_pkt_data(env, NULL, t: BPF_WRITE);
22432
22433	if (is_rdonly)
22434	addr = (unsigned long)bpf_dynptr_from_skb_rdonly;
22435
22436	/* restore env->seen_direct_write to its original value, since
22437	* may_access_direct_pkt_data mutates it
22438	*/
22439	env->seen_direct_write = seen_direct_write;
22440	} else if (func_id == special_kfunc_list[KF_bpf_set_dentry_xattr]) {
22441	if (bpf_lsm_has_d_inode_locked(prog))
22442	addr = (unsigned long)bpf_set_dentry_xattr_locked;
22443	} else if (func_id == special_kfunc_list[KF_bpf_remove_dentry_xattr]) {
22444	if (bpf_lsm_has_d_inode_locked(prog))
22445	addr = (unsigned long)bpf_remove_dentry_xattr_locked;
22446	} else if (func_id == special_kfunc_list[KF_bpf_dynptr_from_file]) {
22447	if (!env->insn_aux_data[insn_idx].non_sleepable)
22448	addr = (unsigned long)bpf_dynptr_from_file_sleepable;
22449	}
22450	desc->addr = addr;
22451	return 0;
22452	}
22453
22454	static void __fixup_collection_insert_kfunc(struct bpf_insn_aux_data *insn_aux,
22455	u16 struct_meta_reg,
22456	u16 node_offset_reg,
22457	struct bpf_insn *insn,
22458	struct bpf_insn *insn_buf,
22459	int *cnt)
22460	{
22461	struct btf_struct_meta *kptr_struct_meta = insn_aux->kptr_struct_meta;
22462	struct bpf_insn addr[2] = { BPF_LD_IMM64(struct_meta_reg, (long)kptr_struct_meta) };
22463
22464	insn_buf[0] = addr[0];
22465	insn_buf[1] = addr[1];
22466	insn_buf[2] = BPF_MOV64_IMM(node_offset_reg, insn_aux->insert_off);
22467	insn_buf[3] = *insn;
22468	*cnt = 4;
22469	}
22470
22471	static int fixup_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
22472	struct bpf_insn *insn_buf, int insn_idx, int *cnt)
22473	{
22474	struct bpf_kfunc_desc *desc;
22475	int err;
22476
22477	if (!insn->imm) {
22478	verbose(private_data: env, fmt: "invalid kernel function call not eliminated in verifier pass\n");
22479	return -EINVAL;
22480	}
22481
22482	*cnt = 0;
22483
22484	/* insn->imm has the btf func_id. Replace it with an offset relative to
22485	* __bpf_call_base, unless the JIT needs to call functions that are
22486	* further than 32 bits away (bpf_jit_supports_far_kfunc_call()).
22487	*/
22488	desc = find_kfunc_desc(prog: env->prog, func_id: insn->imm, offset: insn->off);
22489	if (!desc) {
22490	verifier_bug(env, "kernel function descriptor not found for func_id %u",
22491	insn->imm);
22492	return -EFAULT;
22493	}
22494
22495	err = specialize_kfunc(env, desc, insn_idx);
22496	if (err)
22497	return err;
22498
22499	if (!bpf_jit_supports_far_kfunc_call())
22500	insn->imm = BPF_CALL_IMM(desc->addr);
22501	if (insn->off)
22502	return 0;
22503	if (desc->func_id == special_kfunc_list[KF_bpf_obj_new_impl] ||
22504	desc->func_id == special_kfunc_list[KF_bpf_percpu_obj_new_impl]) {
22505	struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
22506	struct bpf_insn addr[2] = { BPF_LD_IMM64(BPF_REG_2, (long)kptr_struct_meta) };
22507	u64 obj_new_size = env->insn_aux_data[insn_idx].obj_new_size;
22508
22509	if (desc->func_id == special_kfunc_list[KF_bpf_percpu_obj_new_impl] && kptr_struct_meta) {
22510	verifier_bug(env, "NULL kptr_struct_meta expected at insn_idx %d",
22511	insn_idx);
22512	return -EFAULT;
22513	}
22514
22515	insn_buf[0] = BPF_MOV64_IMM(BPF_REG_1, obj_new_size);
22516	insn_buf[1] = addr[0];
22517	insn_buf[2] = addr[1];
22518	insn_buf[3] = *insn;
22519	*cnt = 4;
22520	} else if (desc->func_id == special_kfunc_list[KF_bpf_obj_drop_impl] ||
22521	desc->func_id == special_kfunc_list[KF_bpf_percpu_obj_drop_impl] ||
22522	desc->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl]) {
22523	struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
22524	struct bpf_insn addr[2] = { BPF_LD_IMM64(BPF_REG_2, (long)kptr_struct_meta) };
22525
22526	if (desc->func_id == special_kfunc_list[KF_bpf_percpu_obj_drop_impl] && kptr_struct_meta) {
22527	verifier_bug(env, "NULL kptr_struct_meta expected at insn_idx %d",
22528	insn_idx);
22529	return -EFAULT;
22530	}
22531
22532	if (desc->func_id == special_kfunc_list[KF_bpf_refcount_acquire_impl] &&
22533	!kptr_struct_meta) {
22534	verifier_bug(env, "kptr_struct_meta expected at insn_idx %d",
22535	insn_idx);
22536	return -EFAULT;
22537	}
22538
22539	insn_buf[0] = addr[0];
22540	insn_buf[1] = addr[1];
22541	insn_buf[2] = *insn;
22542	*cnt = 3;
22543	} else if (desc->func_id == special_kfunc_list[KF_bpf_list_push_back_impl] ||
22544	desc->func_id == special_kfunc_list[KF_bpf_list_push_front_impl] ||
22545	desc->func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
22546	struct btf_struct_meta *kptr_struct_meta = env->insn_aux_data[insn_idx].kptr_struct_meta;
22547	int struct_meta_reg = BPF_REG_3;
22548	int node_offset_reg = BPF_REG_4;
22549
22550	/* rbtree_add has extra 'less' arg, so args-to-fixup are in diff regs */
22551	if (desc->func_id == special_kfunc_list[KF_bpf_rbtree_add_impl]) {
22552	struct_meta_reg = BPF_REG_4;
22553	node_offset_reg = BPF_REG_5;
22554	}
22555
22556	if (!kptr_struct_meta) {
22557	verifier_bug(env, "kptr_struct_meta expected at insn_idx %d",
22558	insn_idx);
22559	return -EFAULT;
22560	}
22561
22562	__fixup_collection_insert_kfunc(insn_aux: &env->insn_aux_data[insn_idx], struct_meta_reg,
22563	node_offset_reg, insn, insn_buf, cnt);
22564	} else if (desc->func_id == special_kfunc_list[KF_bpf_cast_to_kern_ctx] ||
22565	desc->func_id == special_kfunc_list[KF_bpf_rdonly_cast]) {
22566	insn_buf[0] = BPF_MOV64_REG(BPF_REG_0, BPF_REG_1);
22567	*cnt = 1;
22568	}
22569
22570	if (env->insn_aux_data[insn_idx].arg_prog) {
22571	u32 regno = env->insn_aux_data[insn_idx].arg_prog;
22572	struct bpf_insn ld_addrs[2] = { BPF_LD_IMM64(regno, (long)env->prog->aux) };
22573	int idx = *cnt;
22574
22575	insn_buf[idx++] = ld_addrs[0];
22576	insn_buf[idx++] = ld_addrs[1];
22577	insn_buf[idx++] = *insn;
22578	*cnt = idx;
22579	}
22580	return 0;
22581	}
22582
22583	/* The function requires that first instruction in 'patch' is insnsi[prog->len - 1] */
22584	static int add_hidden_subprog(struct bpf_verifier_env *env, struct bpf_insn *patch, int len)
22585	{
22586	struct bpf_subprog_info *info = env->subprog_info;
22587	int cnt = env->subprog_cnt;
22588	struct bpf_prog *prog;
22589
22590	/* We only reserve one slot for hidden subprogs in subprog_info. */
22591	if (env->hidden_subprog_cnt) {
22592	verifier_bug(env, "only one hidden subprog supported");
22593	return -EFAULT;
22594	}
22595	/* We're not patching any existing instruction, just appending the new
22596	* ones for the hidden subprog. Hence all of the adjustment operations
22597	* in bpf_patch_insn_data are no-ops.
22598	*/
22599	prog = bpf_patch_insn_data(env, off: env->prog->len - 1, patch, len);
22600	if (!prog)
22601	return -ENOMEM;
22602	env->prog = prog;
22603	info[cnt + 1].start = info[cnt].start;
22604	info[cnt].start = prog->len - len + 1;
22605	env->subprog_cnt++;
22606	env->hidden_subprog_cnt++;
22607	return 0;
22608	}
22609
22610	/* Do various post-verification rewrites in a single program pass.
22611	* These rewrites simplify JIT and interpreter implementations.
22612	*/
22613	static int do_misc_fixups(struct bpf_verifier_env *env)
22614	{
22615	struct bpf_prog *prog = env->prog;
22616	enum bpf_attach_type eatype = prog->expected_attach_type;
22617	enum bpf_prog_type prog_type = resolve_prog_type(prog);
22618	struct bpf_insn *insn = prog->insnsi;
22619	const struct bpf_func_proto *fn;
22620	const int insn_cnt = prog->len;
22621	const struct bpf_map_ops *ops;
22622	struct bpf_insn_aux_data *aux;
22623	struct bpf_insn *insn_buf = env->insn_buf;
22624	struct bpf_prog *new_prog;
22625	struct bpf_map *map_ptr;
22626	int i, ret, cnt, delta = 0, cur_subprog = 0;
22627	struct bpf_subprog_info *subprogs = env->subprog_info;
22628	u16 stack_depth = subprogs[cur_subprog].stack_depth;
22629	u16 stack_depth_extra = 0;
22630
22631	if (env->seen_exception && !env->exception_callback_subprog) {
22632	struct bpf_insn *patch = insn_buf;
22633
22634	*patch++ = env->prog->insnsi[insn_cnt - 1];
22635	*patch++ = BPF_MOV64_REG(BPF_REG_0, BPF_REG_1);
22636	*patch++ = BPF_EXIT_INSN();
22637	ret = add_hidden_subprog(env, patch: insn_buf, len: patch - insn_buf);
22638	if (ret < 0)
22639	return ret;
22640	prog = env->prog;
22641	insn = prog->insnsi;
22642
22643	env->exception_callback_subprog = env->subprog_cnt - 1;
22644	/* Don't update insn_cnt, as add_hidden_subprog always appends insns */
22645	mark_subprog_exc_cb(env, subprog: env->exception_callback_subprog);
22646	}
22647
22648	for (i = 0; i < insn_cnt;) {
22649	if (insn->code == (BPF_ALU64 | BPF_MOV | BPF_X) && insn->imm) {
22650	if ((insn->off == BPF_ADDR_SPACE_CAST && insn->imm == 1) ||
22651	(((struct bpf_map *)env->prog->aux->arena)->map_flags & BPF_F_NO_USER_CONV)) {
22652	/* convert to 32-bit mov that clears upper 32-bit */
22653	insn->code = BPF_ALU | BPF_MOV | BPF_X;
22654	/* clear off and imm, so it's a normal 'wX = wY' from JIT pov */
22655	insn->off = 0;
22656	insn->imm = 0;
22657	} /* cast from as(0) to as(1) should be handled by JIT */
22658	goto next_insn;
22659	}
22660
22661	if (env->insn_aux_data[i + delta].needs_zext)
22662	/* Convert BPF_CLASS(insn->code) == BPF_ALU64 to 32-bit ALU */
22663	insn->code = BPF_ALU | BPF_OP(insn->code) | BPF_SRC(insn->code);
22664
22665	/* Make sdiv/smod divide-by-minus-one exceptions impossible. */
22666	if ((insn->code == (BPF_ALU64 | BPF_MOD | BPF_K) ||
22667	insn->code == (BPF_ALU64 | BPF_DIV | BPF_K) ||
22668	insn->code == (BPF_ALU | BPF_MOD | BPF_K) ||
22669	insn->code == (BPF_ALU | BPF_DIV | BPF_K)) &&
22670	insn->off == 1 && insn->imm == -1) {
22671	bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
22672	bool isdiv = BPF_OP(insn->code) == BPF_DIV;
22673	struct bpf_insn *patch = insn_buf;
22674
22675	if (isdiv)
22676	*patch++ = BPF_RAW_INSN((is64 ? BPF_ALU64 : BPF_ALU) |
22677	BPF_NEG | BPF_K, insn->dst_reg,
22678	0, 0, 0);
22679	else
22680	*patch++ = BPF_MOV32_IMM(insn->dst_reg, 0);
22681
22682	cnt = patch - insn_buf;
22683
22684	new_prog = bpf_patch_insn_data(env, off: i + delta, patch: insn_buf, len: cnt);
22685	if (!new_prog)
22686	return -ENOMEM;
22687
22688	delta += cnt - 1;
22689	env->prog = prog = new_prog;
22690	insn = new_prog->insnsi + i + delta;
22691	goto next_insn;
22692	}
22693
22694	/* Make divide-by-zero and divide-by-minus-one exceptions impossible. */
22695	if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
22696	insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
22697	insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
22698	insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
22699	bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
22700	bool isdiv = BPF_OP(insn->code) == BPF_DIV;
22701	bool is_sdiv = isdiv && insn->off == 1;
22702	bool is_smod = !isdiv && insn->off == 1;
22703	struct bpf_insn *patch = insn_buf;
22704
22705	if (is_sdiv) {
22706	/* [R,W]x sdiv 0 -> 0
22707	* LLONG_MIN sdiv -1 -> LLONG_MIN
22708	* INT_MIN sdiv -1 -> INT_MIN
22709	*/
22710	*patch++ = BPF_MOV64_REG(BPF_REG_AX, insn->src_reg);
22711	*patch++ = BPF_RAW_INSN((is64 ? BPF_ALU64 : BPF_ALU) |
22712	BPF_ADD | BPF_K, BPF_REG_AX,
22713	0, 0, 1);
22714	*patch++ = BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
22715	BPF_JGT | BPF_K, BPF_REG_AX,
22716	0, 4, 1);
22717	*patch++ = BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
22718	BPF_JEQ | BPF_K, BPF_REG_AX,
22719	0, 1, 0);
22720	*patch++ = BPF_RAW_INSN((is64 ? BPF_ALU64 : BPF_ALU) |
22721	BPF_MOV | BPF_K, insn->dst_reg,
22722	0, 0, 0);
22723	/* BPF_NEG(LLONG_MIN) == -LLONG_MIN == LLONG_MIN */
22724	*patch++ = BPF_RAW_INSN((is64 ? BPF_ALU64 : BPF_ALU) |
22725	BPF_NEG | BPF_K, insn->dst_reg,
22726	0, 0, 0);
22727	*patch++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1);
22728	*patch++ = *insn;
22729	cnt = patch - insn_buf;
22730	} else if (is_smod) {
22731	/* [R,W]x mod 0 -> [R,W]x */
22732	/* [R,W]x mod -1 -> 0 */
22733	*patch++ = BPF_MOV64_REG(BPF_REG_AX, insn->src_reg);
22734	*patch++ = BPF_RAW_INSN((is64 ? BPF_ALU64 : BPF_ALU) |
22735	BPF_ADD | BPF_K, BPF_REG_AX,
22736	0, 0, 1);
22737	*patch++ = BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
22738	BPF_JGT | BPF_K, BPF_REG_AX,
22739	0, 3, 1);
22740	*patch++ = BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
22741	BPF_JEQ | BPF_K, BPF_REG_AX,
22742	0, 3 + (is64 ? 0 : 1), 1);
22743	*patch++ = BPF_MOV32_IMM(insn->dst_reg, 0);
22744	*patch++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1);
22745	*patch++ = *insn;
22746
22747	if (!is64) {
22748	*patch++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1);
22749	*patch++ = BPF_MOV32_REG(insn->dst_reg, insn->dst_reg);
22750	}
22751	cnt = patch - insn_buf;
22752	} else if (isdiv) {
22753	/* [R,W]x div 0 -> 0 */
22754	*patch++ = BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
22755	BPF_JNE | BPF_K, insn->src_reg,
22756	0, 2, 0);
22757	*patch++ = BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg);
22758	*patch++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1);
22759	*patch++ = *insn;
22760	cnt = patch - insn_buf;
22761	} else {
22762	/* [R,W]x mod 0 -> [R,W]x */
22763	*patch++ = BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
22764	BPF_JEQ | BPF_K, insn->src_reg,
22765	0, 1 + (is64 ? 0 : 1), 0);
22766	*patch++ = *insn;
22767
22768	if (!is64) {
22769	*patch++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1);
22770	*patch++ = BPF_MOV32_REG(insn->dst_reg, insn->dst_reg);
22771	}
22772	cnt = patch - insn_buf;
22773	}
22774
22775	new_prog = bpf_patch_insn_data(env, off: i + delta, patch: insn_buf, len: cnt);
22776	if (!new_prog)
22777	return -ENOMEM;
22778
22779	delta += cnt - 1;
22780	env->prog = prog = new_prog;
22781	insn = new_prog->insnsi + i + delta;
22782	goto next_insn;
22783	}
22784
22785	/* Make it impossible to de-reference a userspace address */
22786	if (BPF_CLASS(insn->code) == BPF_LDX &&
22787	(BPF_MODE(insn->code) == BPF_PROBE_MEM ||
22788	BPF_MODE(insn->code) == BPF_PROBE_MEMSX)) {
22789	struct bpf_insn *patch = insn_buf;
22790	u64 uaddress_limit = bpf_arch_uaddress_limit();
22791
22792	if (!uaddress_limit)
22793	goto next_insn;
22794
22795	*patch++ = BPF_MOV64_REG(BPF_REG_AX, insn->src_reg);
22796	if (insn->off)
22797	*patch++ = BPF_ALU64_IMM(BPF_ADD, BPF_REG_AX, insn->off);
22798	*patch++ = BPF_ALU64_IMM(BPF_RSH, BPF_REG_AX, 32);
22799	*patch++ = BPF_JMP_IMM(BPF_JLE, BPF_REG_AX, uaddress_limit >> 32, 2);
22800	*patch++ = *insn;
22801	*patch++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1);
22802	*patch++ = BPF_MOV64_IMM(insn->dst_reg, 0);
22803
22804	cnt = patch - insn_buf;
22805	new_prog = bpf_patch_insn_data(env, off: i + delta, patch: insn_buf, len: cnt);
22806	if (!new_prog)
22807	return -ENOMEM;
22808
22809	delta += cnt - 1;
22810	env->prog = prog = new_prog;
22811	insn = new_prog->insnsi + i + delta;
22812	goto next_insn;
22813	}
22814
22815	/* Implement LD_ABS and LD_IND with a rewrite, if supported by the program type. */
22816	if (BPF_CLASS(insn->code) == BPF_LD &&
22817	(BPF_MODE(insn->code) == BPF_ABS ||
22818	BPF_MODE(insn->code) == BPF_IND)) {
22819	cnt = env->ops->gen_ld_abs(insn, insn_buf);
22820	if (cnt == 0 || cnt >= INSN_BUF_SIZE) {
22821	verifier_bug(env, "%d insns generated for ld_abs", cnt);
22822	return -EFAULT;
22823	}
22824
22825	new_prog = bpf_patch_insn_data(env, off: i + delta, patch: insn_buf, len: cnt);
22826	if (!new_prog)
22827	return -ENOMEM;
22828
22829	delta += cnt - 1;
22830	env->prog = prog = new_prog;
22831	insn = new_prog->insnsi + i + delta;
22832	goto next_insn;
22833	}
22834
22835	/* Rewrite pointer arithmetic to mitigate speculation attacks. */
22836	if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
22837	insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
22838	const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
22839	const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
22840	struct bpf_insn *patch = insn_buf;
22841	bool issrc, isneg, isimm;
22842	u32 off_reg;
22843
22844	aux = &env->insn_aux_data[i + delta];
22845	if (!aux->alu_state ||
22846	aux->alu_state == BPF_ALU_NON_POINTER)
22847	goto next_insn;
22848
22849	isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
22850	issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
22851	BPF_ALU_SANITIZE_SRC;
22852	isimm = aux->alu_state & BPF_ALU_IMMEDIATE;
22853
22854	off_reg = issrc ? insn->src_reg : insn->dst_reg;
22855	if (isimm) {
22856	*patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
22857	} else {
22858	if (isneg)
22859	*patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
22860	*patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
22861	*patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
22862	*patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
22863	*patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
22864	*patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
22865	*patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX, off_reg);
22866	}
22867	if (!issrc)
22868	*patch++ = BPF_MOV64_REG(insn->dst_reg, insn->src_reg);
22869	insn->src_reg = BPF_REG_AX;
22870	if (isneg)
22871	insn->code = insn->code == code_add ?
22872	code_sub : code_add;
22873	*patch++ = *insn;
22874	if (issrc && isneg && !isimm)
22875	*patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
22876	cnt = patch - insn_buf;
22877
22878	new_prog = bpf_patch_insn_data(env, off: i + delta, patch: insn_buf, len: cnt);
22879	if (!new_prog)
22880	return -ENOMEM;
22881
22882	delta += cnt - 1;
22883	env->prog = prog = new_prog;
22884	insn = new_prog->insnsi + i + delta;
22885	goto next_insn;
22886	}
22887
22888	if (is_may_goto_insn(insn) && bpf_jit_supports_timed_may_goto()) {
22889	int stack_off_cnt = -stack_depth - 16;
22890
22891	/*
22892	* Two 8 byte slots, depth-16 stores the count, and
22893	* depth-8 stores the start timestamp of the loop.
22894	*
22895	* The starting value of count is BPF_MAX_TIMED_LOOPS
22896	* (0xffff). Every iteration loads it and subs it by 1,
22897	* until the value becomes 0 in AX (thus, 1 in stack),
22898	* after which we call arch_bpf_timed_may_goto, which
22899	* either sets AX to 0xffff to keep looping, or to 0
22900	* upon timeout. AX is then stored into the stack. In
22901	* the next iteration, we either see 0 and break out, or
22902	* continue iterating until the next time value is 0
22903	* after subtraction, rinse and repeat.
22904	*/
22905	stack_depth_extra = 16;
22906	insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_AX, BPF_REG_10, stack_off_cnt);
22907	if (insn->off >= 0)
22908	insn_buf[1] = BPF_JMP_IMM(BPF_JEQ, BPF_REG_AX, 0, insn->off + 5);
22909	else
22910	insn_buf[1] = BPF_JMP_IMM(BPF_JEQ, BPF_REG_AX, 0, insn->off - 1);
22911	insn_buf[2] = BPF_ALU64_IMM(BPF_SUB, BPF_REG_AX, 1);
22912	insn_buf[3] = BPF_JMP_IMM(BPF_JNE, BPF_REG_AX, 0, 2);
22913	/*
22914	* AX is used as an argument to pass in stack_off_cnt
22915	* (to add to r10/fp), and also as the return value of
22916	* the call to arch_bpf_timed_may_goto.
22917	*/
22918	insn_buf[4] = BPF_MOV64_IMM(BPF_REG_AX, stack_off_cnt);
22919	insn_buf[5] = BPF_EMIT_CALL(arch_bpf_timed_may_goto);
22920	insn_buf[6] = BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_AX, stack_off_cnt);
22921	cnt = 7;
22922
22923	new_prog = bpf_patch_insn_data(env, off: i + delta, patch: insn_buf, len: cnt);
22924	if (!new_prog)
22925	return -ENOMEM;
22926
22927	delta += cnt - 1;
22928	env->prog = prog = new_prog;
22929	insn = new_prog->insnsi + i + delta;
22930	goto next_insn;
22931	} else if (is_may_goto_insn(insn)) {
22932	int stack_off = -stack_depth - 8;
22933
22934	stack_depth_extra = 8;
22935	insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_AX, BPF_REG_10, stack_off);
22936	if (insn->off >= 0)
22937	insn_buf[1] = BPF_JMP_IMM(BPF_JEQ, BPF_REG_AX, 0, insn->off + 2);
22938	else
22939	insn_buf[1] = BPF_JMP_IMM(BPF_JEQ, BPF_REG_AX, 0, insn->off - 1);
22940	insn_buf[2] = BPF_ALU64_IMM(BPF_SUB, BPF_REG_AX, 1);
22941	insn_buf[3] = BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_AX, stack_off);
22942	cnt = 4;
22943
22944	new_prog = bpf_patch_insn_data(env, off: i + delta, patch: insn_buf, len: cnt);
22945	if (!new_prog)
22946	return -ENOMEM;
22947
22948	delta += cnt - 1;
22949	env->prog = prog = new_prog;
22950	insn = new_prog->insnsi + i + delta;
22951	goto next_insn;
22952	}
22953
22954	if (insn->code != (BPF_JMP | BPF_CALL))
22955	goto next_insn;
22956	if (insn->src_reg == BPF_PSEUDO_CALL)
22957	goto next_insn;
22958	if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
22959	ret = fixup_kfunc_call(env, insn, insn_buf, insn_idx: i + delta, cnt: &cnt);
22960	if (ret)
22961	return ret;
22962	if (cnt == 0)
22963	goto next_insn;
22964
22965	new_prog = bpf_patch_insn_data(env, off: i + delta, patch: insn_buf, len: cnt);
22966	if (!new_prog)
22967	return -ENOMEM;
22968
22969	delta += cnt - 1;
22970	env->prog = prog = new_prog;
22971	insn = new_prog->insnsi + i + delta;
22972	goto next_insn;
22973	}
22974
22975	/* Skip inlining the helper call if the JIT does it. */
22976	if (bpf_jit_inlines_helper_call(imm: insn->imm))
22977	goto next_insn;
22978
22979	if (insn->imm == BPF_FUNC_get_route_realm)
22980	prog->dst_needed = 1;
22981	if (insn->imm == BPF_FUNC_get_prandom_u32)
22982	bpf_user_rnd_init_once();
22983	if (insn->imm == BPF_FUNC_override_return)
22984	prog->kprobe_override = 1;
22985	if (insn->imm == BPF_FUNC_tail_call) {
22986	/* If we tail call into other programs, we
22987	* cannot make any assumptions since they can
22988	* be replaced dynamically during runtime in
22989	* the program array.
22990	*/
22991	prog->cb_access = 1;
22992	if (!allow_tail_call_in_subprogs(env))
22993	prog->aux->stack_depth = MAX_BPF_STACK;
22994	prog->aux->max_pkt_offset = MAX_PACKET_OFF;
22995
22996	/* mark bpf_tail_call as different opcode to avoid
22997	* conditional branch in the interpreter for every normal
22998	* call and to prevent accidental JITing by JIT compiler
22999	* that doesn't support bpf_tail_call yet
23000	*/
23001	insn->imm = 0;
23002	insn->code = BPF_JMP | BPF_TAIL_CALL;
23003
23004	aux = &env->insn_aux_data[i + delta];
23005	if (env->bpf_capable && !prog->blinding_requested &&
23006	prog->jit_requested &&
23007	!bpf_map_key_poisoned(aux) &&
23008	!bpf_map_ptr_poisoned(aux) &&
23009	!bpf_map_ptr_unpriv(aux)) {
23010	struct bpf_jit_poke_descriptor desc = {
23011	.reason = BPF_POKE_REASON_TAIL_CALL,
23012	.tail_call.map = aux->map_ptr_state.map_ptr,
23013	.tail_call.key = bpf_map_key_immediate(aux),
23014	.insn_idx = i + delta,
23015	};
23016
23017	ret = bpf_jit_add_poke_descriptor(prog, poke: &desc);
23018	if (ret < 0) {
23019	verbose(private_data: env, fmt: "adding tail call poke descriptor failed\n");
23020	return ret;
23021	}
23022
23023	insn->imm = ret + 1;
23024	goto next_insn;
23025	}
23026
23027	if (!bpf_map_ptr_unpriv(aux))
23028	goto next_insn;
23029
23030	/* instead of changing every JIT dealing with tail_call
23031	* emit two extra insns:
23032	* if (index >= max_entries) goto out;
23033	* index &= array->index_mask;
23034	* to avoid out-of-bounds cpu speculation
23035	*/
23036	if (bpf_map_ptr_poisoned(aux)) {
23037	verbose(private_data: env, fmt: "tail_call abusing map_ptr\n");
23038	return -EINVAL;
23039	}
23040
23041	map_ptr = aux->map_ptr_state.map_ptr;
23042	insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
23043	map_ptr->max_entries, 2);
23044	insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
23045	container_of(map_ptr,
23046	struct bpf_array,
23047	map)->index_mask);
23048	insn_buf[2] = *insn;
23049	cnt = 3;
23050	new_prog = bpf_patch_insn_data(env, off: i + delta, patch: insn_buf, len: cnt);
23051	if (!new_prog)
23052	return -ENOMEM;
23053
23054	delta += cnt - 1;
23055	env->prog = prog = new_prog;
23056	insn = new_prog->insnsi + i + delta;
23057	goto next_insn;
23058	}
23059
23060	if (insn->imm == BPF_FUNC_timer_set_callback) {
23061	/* The verifier will process callback_fn as many times as necessary
23062	* with different maps and the register states prepared by
23063	* set_timer_callback_state will be accurate.
23064	*
23065	* The following use case is valid:
23066	* map1 is shared by prog1, prog2, prog3.
23067	* prog1 calls bpf_timer_init for some map1 elements
23068	* prog2 calls bpf_timer_set_callback for some map1 elements.
23069	* Those that were not bpf_timer_init-ed will return -EINVAL.
23070	* prog3 calls bpf_timer_start for some map1 elements.
23071	* Those that were not both bpf_timer_init-ed and
23072	* bpf_timer_set_callback-ed will return -EINVAL.
23073	*/
23074	struct bpf_insn ld_addrs[2] = {
23075	BPF_LD_IMM64(BPF_REG_3, (long)prog->aux),
23076	};
23077
23078	insn_buf[0] = ld_addrs[0];
23079	insn_buf[1] = ld_addrs[1];
23080	insn_buf[2] = *insn;
23081	cnt = 3;
23082
23083	new_prog = bpf_patch_insn_data(env, off: i + delta, patch: insn_buf, len: cnt);
23084	if (!new_prog)
23085	return -ENOMEM;
23086
23087	delta += cnt - 1;
23088	env->prog = prog = new_prog;
23089	insn = new_prog->insnsi + i + delta;
23090	goto patch_call_imm;
23091	}
23092
23093	if (is_storage_get_function(func_id: insn->imm)) {
23094	if (env->insn_aux_data[i + delta].non_sleepable)
23095	insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_ATOMIC);
23096	else
23097	insn_buf[0] = BPF_MOV64_IMM(BPF_REG_5, (__force __s32)GFP_KERNEL);
23098	insn_buf[1] = *insn;
23099	cnt = 2;
23100
23101	new_prog = bpf_patch_insn_data(env, off: i + delta, patch: insn_buf, len: cnt);
23102	if (!new_prog)
23103	return -ENOMEM;
23104
23105	delta += cnt - 1;
23106	env->prog = prog = new_prog;
23107	insn = new_prog->insnsi + i + delta;
23108	goto patch_call_imm;
23109	}
23110
23111	/* bpf_per_cpu_ptr() and bpf_this_cpu_ptr() */
23112	if (env->insn_aux_data[i + delta].call_with_percpu_alloc_ptr) {
23113	/* patch with 'r1 = *(u64 *)(r1 + 0)' since for percpu data,
23114	* bpf_mem_alloc() returns a ptr to the percpu data ptr.
23115	*/
23116	insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_1, BPF_REG_1, 0);
23117	insn_buf[1] = *insn;
23118	cnt = 2;
23119
23120	new_prog = bpf_patch_insn_data(env, off: i + delta, patch: insn_buf, len: cnt);
23121	if (!new_prog)
23122	return -ENOMEM;
23123
23124	delta += cnt - 1;
23125	env->prog = prog = new_prog;
23126	insn = new_prog->insnsi + i + delta;
23127	goto patch_call_imm;
23128	}
23129
23130	/* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
23131	* and other inlining handlers are currently limited to 64 bit
23132	* only.
23133	*/
23134	if (prog->jit_requested && BITS_PER_LONG == 64 &&
23135	(insn->imm == BPF_FUNC_map_lookup_elem ||
23136	insn->imm == BPF_FUNC_map_update_elem ||
23137	insn->imm == BPF_FUNC_map_delete_elem ||
23138	insn->imm == BPF_FUNC_map_push_elem ||
23139	insn->imm == BPF_FUNC_map_pop_elem ||
23140	insn->imm == BPF_FUNC_map_peek_elem ||
23141	insn->imm == BPF_FUNC_redirect_map ||
23142	insn->imm == BPF_FUNC_for_each_map_elem ||
23143	insn->imm == BPF_FUNC_map_lookup_percpu_elem)) {
23144	aux = &env->insn_aux_data[i + delta];
23145	if (bpf_map_ptr_poisoned(aux))
23146	goto patch_call_imm;
23147
23148	map_ptr = aux->map_ptr_state.map_ptr;
23149	ops = map_ptr->ops;
23150	if (insn->imm == BPF_FUNC_map_lookup_elem &&
23151	ops->map_gen_lookup) {
23152	cnt = ops->map_gen_lookup(map_ptr, insn_buf);
23153	if (cnt == -EOPNOTSUPP)
23154	goto patch_map_ops_generic;
23155	if (cnt <= 0 || cnt >= INSN_BUF_SIZE) {
23156	verifier_bug(env, "%d insns generated for map lookup", cnt);
23157	return -EFAULT;
23158	}
23159
23160	new_prog = bpf_patch_insn_data(env, off: i + delta,
23161	patch: insn_buf, len: cnt);
23162	if (!new_prog)
23163	return -ENOMEM;
23164
23165	delta += cnt - 1;
23166	env->prog = prog = new_prog;
23167	insn = new_prog->insnsi + i + delta;
23168	goto next_insn;
23169	}
23170
23171	BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
23172	(void *(*)(struct bpf_map *map, void *key))NULL));
23173	BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
23174	(long (*)(struct bpf_map *map, void *key))NULL));
23175	BUILD_BUG_ON(!__same_type(ops->map_update_elem,
23176	(long (*)(struct bpf_map *map, void *key, void *value,
23177	u64 flags))NULL));
23178	BUILD_BUG_ON(!__same_type(ops->map_push_elem,
23179	(long (*)(struct bpf_map *map, void *value,
23180	u64 flags))NULL));
23181	BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
23182	(long (*)(struct bpf_map *map, void *value))NULL));
23183	BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
23184	(long (*)(struct bpf_map *map, void *value))NULL));
23185	BUILD_BUG_ON(!__same_type(ops->map_redirect,
23186	(long (*)(struct bpf_map *map, u64 index, u64 flags))NULL));
23187	BUILD_BUG_ON(!__same_type(ops->map_for_each_callback,
23188	(long (*)(struct bpf_map *map,
23189	bpf_callback_t callback_fn,
23190	void *callback_ctx,
23191	u64 flags))NULL));
23192	BUILD_BUG_ON(!__same_type(ops->map_lookup_percpu_elem,
23193	(void *(*)(struct bpf_map *map, void *key, u32 cpu))NULL));
23194
23195	patch_map_ops_generic:
23196	switch (insn->imm) {
23197	case BPF_FUNC_map_lookup_elem:
23198	insn->imm = BPF_CALL_IMM(ops->map_lookup_elem);
23199	goto next_insn;
23200	case BPF_FUNC_map_update_elem:
23201	insn->imm = BPF_CALL_IMM(ops->map_update_elem);
23202	goto next_insn;
23203	case BPF_FUNC_map_delete_elem:
23204	insn->imm = BPF_CALL_IMM(ops->map_delete_elem);
23205	goto next_insn;
23206	case BPF_FUNC_map_push_elem:
23207	insn->imm = BPF_CALL_IMM(ops->map_push_elem);
23208	goto next_insn;
23209	case BPF_FUNC_map_pop_elem:
23210	insn->imm = BPF_CALL_IMM(ops->map_pop_elem);
23211	goto next_insn;
23212	case BPF_FUNC_map_peek_elem:
23213	insn->imm = BPF_CALL_IMM(ops->map_peek_elem);
23214	goto next_insn;
23215	case BPF_FUNC_redirect_map:
23216	insn->imm = BPF_CALL_IMM(ops->map_redirect);
23217	goto next_insn;
23218	case BPF_FUNC_for_each_map_elem:
23219	insn->imm = BPF_CALL_IMM(ops->map_for_each_callback);
23220	goto next_insn;
23221	case BPF_FUNC_map_lookup_percpu_elem:
23222	insn->imm = BPF_CALL_IMM(ops->map_lookup_percpu_elem);
23223	goto next_insn;
23224	}
23225
23226	goto patch_call_imm;
23227	}
23228
23229	/* Implement bpf_jiffies64 inline. */
23230	if (prog->jit_requested && BITS_PER_LONG == 64 &&
23231	insn->imm == BPF_FUNC_jiffies64) {
23232	struct bpf_insn ld_jiffies_addr[2] = {
23233	BPF_LD_IMM64(BPF_REG_0,
23234	(unsigned long)&jiffies),
23235	};
23236
23237	insn_buf[0] = ld_jiffies_addr[0];
23238	insn_buf[1] = ld_jiffies_addr[1];
23239	insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
23240	BPF_REG_0, 0);
23241	cnt = 3;
23242
23243	new_prog = bpf_patch_insn_data(env, off: i + delta, patch: insn_buf,
23244	len: cnt);
23245	if (!new_prog)
23246	return -ENOMEM;
23247
23248	delta += cnt - 1;
23249	env->prog = prog = new_prog;
23250	insn = new_prog->insnsi + i + delta;
23251	goto next_insn;
23252	}
23253
23254	#if defined(CONFIG_X86_64) && !defined(CONFIG_UML)
23255	/* Implement bpf_get_smp_processor_id() inline. */
23256	if (insn->imm == BPF_FUNC_get_smp_processor_id &&
23257	verifier_inlines_helper_call(env, imm: insn->imm)) {
23258	/* BPF_FUNC_get_smp_processor_id inlining is an
23259	* optimization, so if cpu_number is ever
23260	* changed in some incompatible and hard to support
23261	* way, it's fine to back out this inlining logic
23262	*/
23263	#ifdef CONFIG_SMP
23264	insn_buf[0] = BPF_MOV64_IMM(BPF_REG_0, (u32)(unsigned long)&cpu_number);
23265	insn_buf[1] = BPF_MOV64_PERCPU_REG(BPF_REG_0, BPF_REG_0);
23266	insn_buf[2] = BPF_LDX_MEM(BPF_W, BPF_REG_0, BPF_REG_0, 0);
23267	cnt = 3;
23268	#else
23269	insn_buf[0] = BPF_ALU32_REG(BPF_XOR, BPF_REG_0, BPF_REG_0);
23270	cnt = 1;
23271	#endif
23272	new_prog = bpf_patch_insn_data(env, off: i + delta, patch: insn_buf, len: cnt);
23273	if (!new_prog)
23274	return -ENOMEM;
23275
23276	delta += cnt - 1;
23277	env->prog = prog = new_prog;
23278	insn = new_prog->insnsi + i + delta;
23279	goto next_insn;
23280	}
23281	#endif
23282	/* Implement bpf_get_func_arg inline. */
23283	if (prog_type == BPF_PROG_TYPE_TRACING &&
23284	insn->imm == BPF_FUNC_get_func_arg) {
23285	/* Load nr_args from ctx - 8 */
23286	insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
23287	insn_buf[1] = BPF_JMP32_REG(BPF_JGE, BPF_REG_2, BPF_REG_0, 6);
23288	insn_buf[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_2, 3);
23289	insn_buf[3] = BPF_ALU64_REG(BPF_ADD, BPF_REG_2, BPF_REG_1);
23290	insn_buf[4] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_2, 0);
23291	insn_buf[5] = BPF_STX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
23292	insn_buf[6] = BPF_MOV64_IMM(BPF_REG_0, 0);
23293	insn_buf[7] = BPF_JMP_A(1);
23294	insn_buf[8] = BPF_MOV64_IMM(BPF_REG_0, -EINVAL);
23295	cnt = 9;
23296
23297	new_prog = bpf_patch_insn_data(env, off: i + delta, patch: insn_buf, len: cnt);
23298	if (!new_prog)
23299	return -ENOMEM;
23300
23301	delta += cnt - 1;
23302	env->prog = prog = new_prog;
23303	insn = new_prog->insnsi + i + delta;
23304	goto next_insn;
23305	}
23306
23307	/* Implement bpf_get_func_ret inline. */
23308	if (prog_type == BPF_PROG_TYPE_TRACING &&
23309	insn->imm == BPF_FUNC_get_func_ret) {
23310	if (eatype == BPF_TRACE_FEXIT ||
23311	eatype == BPF_MODIFY_RETURN) {
23312	/* Load nr_args from ctx - 8 */
23313	insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
23314	insn_buf[1] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_0, 3);
23315	insn_buf[2] = BPF_ALU64_REG(BPF_ADD, BPF_REG_0, BPF_REG_1);
23316	insn_buf[3] = BPF_LDX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
23317	insn_buf[4] = BPF_STX_MEM(BPF_DW, BPF_REG_2, BPF_REG_3, 0);
23318	insn_buf[5] = BPF_MOV64_IMM(BPF_REG_0, 0);
23319	cnt = 6;
23320	} else {
23321	insn_buf[0] = BPF_MOV64_IMM(BPF_REG_0, -EOPNOTSUPP);
23322	cnt = 1;
23323	}
23324
23325	new_prog = bpf_patch_insn_data(env, off: i + delta, patch: insn_buf, len: cnt);
23326	if (!new_prog)
23327	return -ENOMEM;
23328
23329	delta += cnt - 1;
23330	env->prog = prog = new_prog;
23331	insn = new_prog->insnsi + i + delta;
23332	goto next_insn;
23333	}
23334
23335	/* Implement get_func_arg_cnt inline. */
23336	if (prog_type == BPF_PROG_TYPE_TRACING &&
23337	insn->imm == BPF_FUNC_get_func_arg_cnt) {
23338	/* Load nr_args from ctx - 8 */
23339	insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
23340
23341	new_prog = bpf_patch_insn_data(env, off: i + delta, patch: insn_buf, len: 1);
23342	if (!new_prog)
23343	return -ENOMEM;
23344
23345	env->prog = prog = new_prog;
23346	insn = new_prog->insnsi + i + delta;
23347	goto next_insn;
23348	}
23349
23350	/* Implement bpf_get_func_ip inline. */
23351	if (prog_type == BPF_PROG_TYPE_TRACING &&
23352	insn->imm == BPF_FUNC_get_func_ip) {
23353	/* Load IP address from ctx - 16 */
23354	insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -16);
23355
23356	new_prog = bpf_patch_insn_data(env, off: i + delta, patch: insn_buf, len: 1);
23357	if (!new_prog)
23358	return -ENOMEM;
23359
23360	env->prog = prog = new_prog;
23361	insn = new_prog->insnsi + i + delta;
23362	goto next_insn;
23363	}
23364
23365	/* Implement bpf_get_branch_snapshot inline. */
23366	if (IS_ENABLED(CONFIG_PERF_EVENTS) &&
23367	prog->jit_requested && BITS_PER_LONG == 64 &&
23368	insn->imm == BPF_FUNC_get_branch_snapshot) {
23369	/* We are dealing with the following func protos:
23370	* u64 bpf_get_branch_snapshot(void *buf, u32 size, u64 flags);
23371	* int perf_snapshot_branch_stack(struct perf_branch_entry *entries, u32 cnt);
23372	*/
23373	const u32 br_entry_size = sizeof(struct perf_branch_entry);
23374
23375	/* struct perf_branch_entry is part of UAPI and is
23376	* used as an array element, so extremely unlikely to
23377	* ever grow or shrink
23378	*/
23379	BUILD_BUG_ON(br_entry_size != 24);
23380
23381	/* if (unlikely(flags)) return -EINVAL */
23382	insn_buf[0] = BPF_JMP_IMM(BPF_JNE, BPF_REG_3, 0, 7);
23383
23384	/* Transform size (bytes) into number of entries (cnt = size / 24).
23385	* But to avoid expensive division instruction, we implement
23386	* divide-by-3 through multiplication, followed by further
23387	* division by 8 through 3-bit right shift.
23388	* Refer to book "Hacker's Delight, 2nd ed." by Henry S. Warren, Jr.,
23389	* p. 227, chapter "Unsigned Division by 3" for details and proofs.
23390	*
23391	* N / 3 <=> M * N / 2^33, where M = (2^33 + 1) / 3 = 0xaaaaaaab.
23392	*/
23393	insn_buf[1] = BPF_MOV32_IMM(BPF_REG_0, 0xaaaaaaab);
23394	insn_buf[2] = BPF_ALU64_REG(BPF_MUL, BPF_REG_2, BPF_REG_0);
23395	insn_buf[3] = BPF_ALU64_IMM(BPF_RSH, BPF_REG_2, 36);
23396
23397	/* call perf_snapshot_branch_stack implementation */
23398	insn_buf[4] = BPF_EMIT_CALL(static_call_query(perf_snapshot_branch_stack));
23399	/* if (entry_cnt == 0) return -ENOENT */
23400	insn_buf[5] = BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, 4);
23401	/* return entry_cnt * sizeof(struct perf_branch_entry) */
23402	insn_buf[6] = BPF_ALU32_IMM(BPF_MUL, BPF_REG_0, br_entry_size);
23403	insn_buf[7] = BPF_JMP_A(3);
23404	/* return -EINVAL; */
23405	insn_buf[8] = BPF_MOV64_IMM(BPF_REG_0, -EINVAL);
23406	insn_buf[9] = BPF_JMP_A(1);
23407	/* return -ENOENT; */
23408	insn_buf[10] = BPF_MOV64_IMM(BPF_REG_0, -ENOENT);
23409	cnt = 11;
23410
23411	new_prog = bpf_patch_insn_data(env, off: i + delta, patch: insn_buf, len: cnt);
23412	if (!new_prog)
23413	return -ENOMEM;
23414
23415	delta += cnt - 1;
23416	env->prog = prog = new_prog;
23417	insn = new_prog->insnsi + i + delta;
23418	goto next_insn;
23419	}
23420
23421	/* Implement bpf_kptr_xchg inline */
23422	if (prog->jit_requested && BITS_PER_LONG == 64 &&
23423	insn->imm == BPF_FUNC_kptr_xchg &&
23424	bpf_jit_supports_ptr_xchg()) {
23425	insn_buf[0] = BPF_MOV64_REG(BPF_REG_0, BPF_REG_2);
23426	insn_buf[1] = BPF_ATOMIC_OP(BPF_DW, BPF_XCHG, BPF_REG_1, BPF_REG_0, 0);
23427	cnt = 2;
23428
23429	new_prog = bpf_patch_insn_data(env, off: i + delta, patch: insn_buf, len: cnt);
23430	if (!new_prog)
23431	return -ENOMEM;
23432
23433	delta += cnt - 1;
23434	env->prog = prog = new_prog;
23435	insn = new_prog->insnsi + i + delta;
23436	goto next_insn;
23437	}
23438	patch_call_imm:
23439	fn = env->ops->get_func_proto(insn->imm, env->prog);
23440	/* all functions that have prototype and verifier allowed
23441	* programs to call them, must be real in-kernel functions
23442	*/
23443	if (!fn->func) {
23444	verifier_bug(env,
23445	"not inlined functions %s#%d is missing func",
23446	func_id_name(insn->imm), insn->imm);
23447	return -EFAULT;
23448	}
23449	insn->imm = fn->func - __bpf_call_base;
23450	next_insn:
23451	if (subprogs[cur_subprog + 1].start == i + delta + 1) {
23452	subprogs[cur_subprog].stack_depth += stack_depth_extra;
23453	subprogs[cur_subprog].stack_extra = stack_depth_extra;
23454
23455	stack_depth = subprogs[cur_subprog].stack_depth;
23456	if (stack_depth > MAX_BPF_STACK && !prog->jit_requested) {
23457	verbose(private_data: env, fmt: "stack size %d(extra %d) is too large\n",
23458	stack_depth, stack_depth_extra);
23459	return -EINVAL;
23460	}
23461	cur_subprog++;
23462	stack_depth = subprogs[cur_subprog].stack_depth;
23463	stack_depth_extra = 0;
23464	}
23465	i++;
23466	insn++;
23467	}
23468
23469	env->prog->aux->stack_depth = subprogs[0].stack_depth;
23470	for (i = 0; i < env->subprog_cnt; i++) {
23471	int delta = bpf_jit_supports_timed_may_goto() ? 2 : 1;
23472	int subprog_start = subprogs[i].start;
23473	int stack_slots = subprogs[i].stack_extra / 8;
23474	int slots = delta, cnt = 0;
23475
23476	if (!stack_slots)
23477	continue;
23478	/* We need two slots in case timed may_goto is supported. */
23479	if (stack_slots > slots) {
23480	verifier_bug(env, "stack_slots supports may_goto only");
23481	return -EFAULT;
23482	}
23483
23484	stack_depth = subprogs[i].stack_depth;
23485	if (bpf_jit_supports_timed_may_goto()) {
23486	insn_buf[cnt++] = BPF_ST_MEM(BPF_DW, BPF_REG_FP, -stack_depth,
23487	BPF_MAX_TIMED_LOOPS);
23488	insn_buf[cnt++] = BPF_ST_MEM(BPF_DW, BPF_REG_FP, -stack_depth + 8, 0);
23489	} else {
23490	/* Add ST insn to subprog prologue to init extra stack */
23491	insn_buf[cnt++] = BPF_ST_MEM(BPF_DW, BPF_REG_FP, -stack_depth,
23492	BPF_MAX_LOOPS);
23493	}
23494	/* Copy first actual insn to preserve it */
23495	insn_buf[cnt++] = env->prog->insnsi[subprog_start];
23496
23497	new_prog = bpf_patch_insn_data(env, off: subprog_start, patch: insn_buf, len: cnt);
23498	if (!new_prog)
23499	return -ENOMEM;
23500	env->prog = prog = new_prog;
23501	/*
23502	* If may_goto is a first insn of a prog there could be a jmp
23503	* insn that points to it, hence adjust all such jmps to point
23504	* to insn after BPF_ST that inits may_goto count.
23505	* Adjustment will succeed because bpf_patch_insn_data() didn't fail.
23506	*/
23507	WARN_ON(adjust_jmp_off(env->prog, subprog_start, delta));
23508	}
23509
23510	/* Since poke tab is now finalized, publish aux to tracker. */
23511	for (i = 0; i < prog->aux->size_poke_tab; i++) {
23512	map_ptr = prog->aux->poke_tab[i].tail_call.map;
23513	if (!map_ptr->ops->map_poke_track ||
23514	!map_ptr->ops->map_poke_untrack ||
23515	!map_ptr->ops->map_poke_run) {
23516	verifier_bug(env, "poke tab is misconfigured");
23517	return -EFAULT;
23518	}
23519
23520	ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
23521	if (ret < 0) {
23522	verbose(private_data: env, fmt: "tracking tail call prog failed\n");
23523	return ret;
23524	}
23525	}
23526
23527	ret = sort_kfunc_descs_by_imm_off(env);
23528	if (ret)
23529	return ret;
23530
23531	return 0;
23532	}
23533
23534	static struct bpf_prog *inline_bpf_loop(struct bpf_verifier_env *env,
23535	int position,
23536	s32 stack_base,
23537	u32 callback_subprogno,
23538	u32 *total_cnt)
23539	{
23540	s32 r6_offset = stack_base + 0 * BPF_REG_SIZE;
23541	s32 r7_offset = stack_base + 1 * BPF_REG_SIZE;
23542	s32 r8_offset = stack_base + 2 * BPF_REG_SIZE;
23543	int reg_loop_max = BPF_REG_6;
23544	int reg_loop_cnt = BPF_REG_7;
23545	int reg_loop_ctx = BPF_REG_8;
23546
23547	struct bpf_insn *insn_buf = env->insn_buf;
23548	struct bpf_prog *new_prog;
23549	u32 callback_start;
23550	u32 call_insn_offset;
23551	s32 callback_offset;
23552	u32 cnt = 0;
23553
23554	/* This represents an inlined version of bpf_iter.c:bpf_loop,
23555	* be careful to modify this code in sync.
23556	*/
23557
23558	/* Return error and jump to the end of the patch if
23559	* expected number of iterations is too big.
23560	*/
23561	insn_buf[cnt++] = BPF_JMP_IMM(BPF_JLE, BPF_REG_1, BPF_MAX_LOOPS, 2);
23562	insn_buf[cnt++] = BPF_MOV32_IMM(BPF_REG_0, -E2BIG);
23563	insn_buf[cnt++] = BPF_JMP_IMM(BPF_JA, 0, 0, 16);
23564	/* spill R6, R7, R8 to use these as loop vars */
23565	insn_buf[cnt++] = BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_6, r6_offset);
23566	insn_buf[cnt++] = BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_7, r7_offset);
23567	insn_buf[cnt++] = BPF_STX_MEM(BPF_DW, BPF_REG_10, BPF_REG_8, r8_offset);
23568	/* initialize loop vars */
23569	insn_buf[cnt++] = BPF_MOV64_REG(reg_loop_max, BPF_REG_1);
23570	insn_buf[cnt++] = BPF_MOV32_IMM(reg_loop_cnt, 0);
23571	insn_buf[cnt++] = BPF_MOV64_REG(reg_loop_ctx, BPF_REG_3);
23572	/* loop header,
23573	* if reg_loop_cnt >= reg_loop_max skip the loop body
23574	*/
23575	insn_buf[cnt++] = BPF_JMP_REG(BPF_JGE, reg_loop_cnt, reg_loop_max, 5);
23576	/* callback call,
23577	* correct callback offset would be set after patching
23578	*/
23579	insn_buf[cnt++] = BPF_MOV64_REG(BPF_REG_1, reg_loop_cnt);
23580	insn_buf[cnt++] = BPF_MOV64_REG(BPF_REG_2, reg_loop_ctx);
23581	insn_buf[cnt++] = BPF_CALL_REL(0);
23582	/* increment loop counter */
23583	insn_buf[cnt++] = BPF_ALU64_IMM(BPF_ADD, reg_loop_cnt, 1);
23584	/* jump to loop header if callback returned 0 */
23585	insn_buf[cnt++] = BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, -6);
23586	/* return value of bpf_loop,
23587	* set R0 to the number of iterations
23588	*/
23589	insn_buf[cnt++] = BPF_MOV64_REG(BPF_REG_0, reg_loop_cnt);
23590	/* restore original values of R6, R7, R8 */
23591	insn_buf[cnt++] = BPF_LDX_MEM(BPF_DW, BPF_REG_6, BPF_REG_10, r6_offset);
23592	insn_buf[cnt++] = BPF_LDX_MEM(BPF_DW, BPF_REG_7, BPF_REG_10, r7_offset);
23593	insn_buf[cnt++] = BPF_LDX_MEM(BPF_DW, BPF_REG_8, BPF_REG_10, r8_offset);
23594
23595	*total_cnt = cnt;
23596	new_prog = bpf_patch_insn_data(env, off: position, patch: insn_buf, len: cnt);
23597	if (!new_prog)
23598	return new_prog;
23599
23600	/* callback start is known only after patching */
23601	callback_start = env->subprog_info[callback_subprogno].start;
23602	/* Note: insn_buf[12] is an offset of BPF_CALL_REL instruction */
23603	call_insn_offset = position + 12;
23604	callback_offset = callback_start - call_insn_offset - 1;
23605	new_prog->insnsi[call_insn_offset].imm = callback_offset;
23606
23607	return new_prog;
23608	}
23609
23610	static bool is_bpf_loop_call(struct bpf_insn *insn)
23611	{
23612	return insn->code == (BPF_JMP | BPF_CALL) &&
23613	insn->src_reg == 0 &&
23614	insn->imm == BPF_FUNC_loop;
23615	}
23616
23617	/* For all sub-programs in the program (including main) check
23618	* insn_aux_data to see if there are bpf_loop calls that require
23619	* inlining. If such calls are found the calls are replaced with a
23620	* sequence of instructions produced by `inline_bpf_loop` function and
23621	* subprog stack_depth is increased by the size of 3 registers.
23622	* This stack space is used to spill values of the R6, R7, R8. These
23623	* registers are used to store the loop bound, counter and context
23624	* variables.
23625	*/
23626	static int optimize_bpf_loop(struct bpf_verifier_env *env)
23627	{
23628	struct bpf_subprog_info *subprogs = env->subprog_info;
23629	int i, cur_subprog = 0, cnt, delta = 0;
23630	struct bpf_insn *insn = env->prog->insnsi;
23631	int insn_cnt = env->prog->len;
23632	u16 stack_depth = subprogs[cur_subprog].stack_depth;
23633	u16 stack_depth_roundup = round_up(stack_depth, 8) - stack_depth;
23634	u16 stack_depth_extra = 0;
23635
23636	for (i = 0; i < insn_cnt; i++, insn++) {
23637	struct bpf_loop_inline_state *inline_state =
23638	&env->insn_aux_data[i + delta].loop_inline_state;
23639
23640	if (is_bpf_loop_call(insn) && inline_state->fit_for_inline) {
23641	struct bpf_prog *new_prog;
23642
23643	stack_depth_extra = BPF_REG_SIZE * 3 + stack_depth_roundup;
23644	new_prog = inline_bpf_loop(env,
23645	position: i + delta,
23646	stack_base: -(stack_depth + stack_depth_extra),
23647	callback_subprogno: inline_state->callback_subprogno,
23648	total_cnt: &cnt);
23649	if (!new_prog)
23650	return -ENOMEM;
23651
23652	delta += cnt - 1;
23653	env->prog = new_prog;
23654	insn = new_prog->insnsi + i + delta;
23655	}
23656
23657	if (subprogs[cur_subprog + 1].start == i + delta + 1) {
23658	subprogs[cur_subprog].stack_depth += stack_depth_extra;
23659	cur_subprog++;
23660	stack_depth = subprogs[cur_subprog].stack_depth;
23661	stack_depth_roundup = round_up(stack_depth, 8) - stack_depth;
23662	stack_depth_extra = 0;
23663	}
23664	}
23665
23666	env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
23667
23668	return 0;
23669	}
23670
23671	/* Remove unnecessary spill/fill pairs, members of fastcall pattern,
23672	* adjust subprograms stack depth when possible.
23673	*/
23674	static int remove_fastcall_spills_fills(struct bpf_verifier_env *env)
23675	{
23676	struct bpf_subprog_info *subprog = env->subprog_info;
23677	struct bpf_insn_aux_data *aux = env->insn_aux_data;
23678	struct bpf_insn *insn = env->prog->insnsi;
23679	int insn_cnt = env->prog->len;
23680	u32 spills_num;
23681	bool modified = false;
23682	int i, j;
23683
23684	for (i = 0; i < insn_cnt; i++, insn++) {
23685	if (aux[i].fastcall_spills_num > 0) {
23686	spills_num = aux[i].fastcall_spills_num;
23687	/* NOPs would be removed by opt_remove_nops() */
23688	for (j = 1; j <= spills_num; ++j) {
23689	*(insn - j) = NOP;
23690	*(insn + j) = NOP;
23691	}
23692	modified = true;
23693	}
23694	if ((subprog + 1)->start == i + 1) {
23695	if (modified && !subprog->keep_fastcall_stack)
23696	subprog->stack_depth = -subprog->fastcall_stack_off;
23697	subprog++;
23698	modified = false;
23699	}
23700	}
23701
23702	return 0;
23703	}
23704
23705	static void free_states(struct bpf_verifier_env *env)
23706	{
23707	struct bpf_verifier_state_list *sl;
23708	struct list_head *head, *pos, *tmp;
23709	struct bpf_scc_info *info;
23710	int i, j;
23711
23712	free_verifier_state(state: env->cur_state, free_self: true);
23713	env->cur_state = NULL;
23714	while (!pop_stack(env, NULL, NULL, pop_log: false));
23715
23716	list_for_each_safe(pos, tmp, &env->free_list) {
23717	sl = container_of(pos, struct bpf_verifier_state_list, node);
23718	free_verifier_state(state: &sl->state, free_self: false);
23719	kfree(objp: sl);
23720	}
23721	INIT_LIST_HEAD(list: &env->free_list);
23722
23723	for (i = 0; i < env->scc_cnt; ++i) {
23724	info = env->scc_info[i];
23725	if (!info)
23726	continue;
23727	for (j = 0; j < info->num_visits; j++)
23728	free_backedges(visit: &info->visits[j]);
23729	kvfree(addr: info);
23730	env->scc_info[i] = NULL;
23731	}
23732
23733	if (!env->explored_states)
23734	return;
23735
23736	for (i = 0; i < state_htab_size(env); i++) {
23737	head = &env->explored_states[i];
23738
23739	list_for_each_safe(pos, tmp, head) {
23740	sl = container_of(pos, struct bpf_verifier_state_list, node);
23741	free_verifier_state(state: &sl->state, free_self: false);
23742	kfree(objp: sl);
23743	}
23744	INIT_LIST_HEAD(list: &env->explored_states[i]);
23745	}
23746	}
23747
23748	static int do_check_common(struct bpf_verifier_env *env, int subprog)
23749	{
23750	bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
23751	struct bpf_subprog_info *sub = subprog_info(env, subprog);
23752	struct bpf_prog_aux *aux = env->prog->aux;
23753	struct bpf_verifier_state *state;
23754	struct bpf_reg_state *regs;
23755	int ret, i;
23756
23757	env->prev_linfo = NULL;
23758	env->pass_cnt++;
23759
23760	state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL_ACCOUNT);
23761	if (!state)
23762	return -ENOMEM;
23763	state->curframe = 0;
23764	state->speculative = false;
23765	state->branches = 1;
23766	state->in_sleepable = env->prog->sleepable;
23767	state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL_ACCOUNT);
23768	if (!state->frame[0]) {
23769	kfree(objp: state);
23770	return -ENOMEM;
23771	}
23772	env->cur_state = state;
23773	init_func_state(env, state: state->frame[0],
23774	BPF_MAIN_FUNC /* callsite */,
23775	frameno: 0 /* frameno */,
23776	subprogno: subprog);
23777	state->first_insn_idx = env->subprog_info[subprog].start;
23778	state->last_insn_idx = -1;
23779
23780	regs = state->frame[state->curframe]->regs;
23781	if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
23782	const char *sub_name = subprog_name(env, subprog);
23783	struct bpf_subprog_arg_info *arg;
23784	struct bpf_reg_state *reg;
23785
23786	if (env->log.level & BPF_LOG_LEVEL)
23787	verbose(private_data: env, fmt: "Validating %s() func#%d...\n", sub_name, subprog);
23788	ret = btf_prepare_func_args(env, subprog);
23789	if (ret)
23790	goto out;
23791
23792	if (subprog_is_exc_cb(env, subprog)) {
23793	state->frame[0]->in_exception_callback_fn = true;
23794	/* We have already ensured that the callback returns an integer, just
23795	* like all global subprogs. We need to determine it only has a single
23796	* scalar argument.
23797	*/
23798	if (sub->arg_cnt != 1 || sub->args[0].arg_type != ARG_ANYTHING) {
23799	verbose(private_data: env, fmt: "exception cb only supports single integer argument\n");
23800	ret = -EINVAL;
23801	goto out;
23802	}
23803	}
23804	for (i = BPF_REG_1; i <= sub->arg_cnt; i++) {
23805	arg = &sub->args[i - BPF_REG_1];
23806	reg = &regs[i];
23807
23808	if (arg->arg_type == ARG_PTR_TO_CTX) {
23809	reg->type = PTR_TO_CTX;
23810	mark_reg_known_zero(env, regs, regno: i);
23811	} else if (arg->arg_type == ARG_ANYTHING) {
23812	reg->type = SCALAR_VALUE;
23813	mark_reg_unknown(env, regs, regno: i);
23814	} else if (arg->arg_type == (ARG_PTR_TO_DYNPTR | MEM_RDONLY)) {
23815	/* assume unspecial LOCAL dynptr type */
23816	__mark_dynptr_reg(reg, type: BPF_DYNPTR_TYPE_LOCAL, first_slot: true, dynptr_id: ++env->id_gen);
23817	} else if (base_type(type: arg->arg_type) == ARG_PTR_TO_MEM) {
23818	reg->type = PTR_TO_MEM;
23819	reg->type |= arg->arg_type &
23820	(PTR_MAYBE_NULL | PTR_UNTRUSTED | MEM_RDONLY);
23821	mark_reg_known_zero(env, regs, regno: i);
23822	reg->mem_size = arg->mem_size;
23823	if (arg->arg_type & PTR_MAYBE_NULL)
23824	reg->id = ++env->id_gen;
23825	} else if (base_type(type: arg->arg_type) == ARG_PTR_TO_BTF_ID) {
23826	reg->type = PTR_TO_BTF_ID;
23827	if (arg->arg_type & PTR_MAYBE_NULL)
23828	reg->type |= PTR_MAYBE_NULL;
23829	if (arg->arg_type & PTR_UNTRUSTED)
23830	reg->type |= PTR_UNTRUSTED;
23831	if (arg->arg_type & PTR_TRUSTED)
23832	reg->type |= PTR_TRUSTED;
23833	mark_reg_known_zero(env, regs, regno: i);
23834	reg->btf = bpf_get_btf_vmlinux(); /* can't fail at this point */
23835	reg->btf_id = arg->btf_id;
23836	reg->id = ++env->id_gen;
23837	} else if (base_type(type: arg->arg_type) == ARG_PTR_TO_ARENA) {
23838	/* caller can pass either PTR_TO_ARENA or SCALAR */
23839	mark_reg_unknown(env, regs, regno: i);
23840	} else {
23841	verifier_bug(env, "unhandled arg#%d type %d",
23842	i - BPF_REG_1, arg->arg_type);
23843	ret = -EFAULT;
23844	goto out;
23845	}
23846	}
23847	} else {
23848	/* if main BPF program has associated BTF info, validate that
23849	* it's matching expected signature, and otherwise mark BTF
23850	* info for main program as unreliable
23851	*/
23852	if (env->prog->aux->func_info_aux) {
23853	ret = btf_prepare_func_args(env, subprog: 0);
23854	if (ret || sub->arg_cnt != 1 || sub->args[0].arg_type != ARG_PTR_TO_CTX)
23855	env->prog->aux->func_info_aux[0].unreliable = true;
23856	}
23857
23858	/* 1st arg to a function */
23859	regs[BPF_REG_1].type = PTR_TO_CTX;
23860	mark_reg_known_zero(env, regs, regno: BPF_REG_1);
23861	}
23862
23863	/* Acquire references for struct_ops program arguments tagged with "__ref" */
23864	if (!subprog && env->prog->type == BPF_PROG_TYPE_STRUCT_OPS) {
23865	for (i = 0; i < aux->ctx_arg_info_size; i++)
23866	aux->ctx_arg_info[i].ref_obj_id = aux->ctx_arg_info[i].refcounted ?
23867	acquire_reference(env, insn_idx: 0) : 0;
23868	}
23869
23870	ret = do_check(env);
23871	out:
23872	if (!ret && pop_log)
23873	bpf_vlog_reset(log: &env->log, new_pos: 0);
23874	free_states(env);
23875	return ret;
23876	}
23877
23878	/* Lazily verify all global functions based on their BTF, if they are called
23879	* from main BPF program or any of subprograms transitively.
23880	* BPF global subprogs called from dead code are not validated.
23881	* All callable global functions must pass verification.
23882	* Otherwise the whole program is rejected.
23883	* Consider:
23884	* int bar(int);
23885	* int foo(int f)
23886	* {
23887	* return bar(f);
23888	* }
23889	* int bar(int b)
23890	* {
23891	* ...
23892	* }
23893	* foo() will be verified first for R1=any_scalar_value. During verification it
23894	* will be assumed that bar() already verified successfully and call to bar()
23895	* from foo() will be checked for type match only. Later bar() will be verified
23896	* independently to check that it's safe for R1=any_scalar_value.
23897	*/
23898	static int do_check_subprogs(struct bpf_verifier_env *env)
23899	{
23900	struct bpf_prog_aux *aux = env->prog->aux;
23901	struct bpf_func_info_aux *sub_aux;
23902	int i, ret, new_cnt;
23903
23904	if (!aux->func_info)
23905	return 0;
23906
23907	/* exception callback is presumed to be always called */
23908	if (env->exception_callback_subprog)
23909	subprog_aux(env, subprog: env->exception_callback_subprog)->called = true;
23910
23911	again:
23912	new_cnt = 0;
23913	for (i = 1; i < env->subprog_cnt; i++) {
23914	if (!subprog_is_global(env, subprog: i))
23915	continue;
23916
23917	sub_aux = subprog_aux(env, subprog: i);
23918	if (!sub_aux->called || sub_aux->verified)
23919	continue;
23920
23921	env->insn_idx = env->subprog_info[i].start;
23922	WARN_ON_ONCE(env->insn_idx == 0);
23923	ret = do_check_common(env, subprog: i);
23924	if (ret) {
23925	return ret;
23926	} else if (env->log.level & BPF_LOG_LEVEL) {
23927	verbose(private_data: env, fmt: "Func#%d ('%s') is safe for any args that match its prototype\n",
23928	i, subprog_name(env, subprog: i));
23929	}
23930
23931	/* We verified new global subprog, it might have called some
23932	* more global subprogs that we haven't verified yet, so we
23933	* need to do another pass over subprogs to verify those.
23934	*/
23935	sub_aux->verified = true;
23936	new_cnt++;
23937	}
23938
23939	/* We can't loop forever as we verify at least one global subprog on
23940	* each pass.
23941	*/
23942	if (new_cnt)
23943	goto again;
23944
23945	return 0;
23946	}
23947
23948	static int do_check_main(struct bpf_verifier_env *env)
23949	{
23950	int ret;
23951
23952	env->insn_idx = 0;
23953	ret = do_check_common(env, subprog: 0);
23954	if (!ret)
23955	env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
23956	return ret;
23957	}
23958
23959
23960	static void print_verification_stats(struct bpf_verifier_env *env)
23961	{
23962	int i;
23963
23964	if (env->log.level & BPF_LOG_STATS) {
23965	verbose(private_data: env, fmt: "verification time %lld usec\n",
23966	div_u64(dividend: env->verification_time, divisor: 1000));
23967	verbose(private_data: env, fmt: "stack depth ");
23968	for (i = 0; i < env->subprog_cnt; i++) {
23969	u32 depth = env->subprog_info[i].stack_depth;
23970
23971	verbose(private_data: env, fmt: "%d", depth);
23972	if (i + 1 < env->subprog_cnt)
23973	verbose(private_data: env, fmt: "+");
23974	}
23975	verbose(private_data: env, fmt: "\n");
23976	}
23977	verbose(private_data: env, fmt: "processed %d insns (limit %d) max_states_per_insn %d "
23978	"total_states %d peak_states %d mark_read %d\n",
23979	env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
23980	env->max_states_per_insn, env->total_states,
23981	env->peak_states, env->longest_mark_read_walk);
23982	}
23983
23984	int bpf_prog_ctx_arg_info_init(struct bpf_prog *prog,
23985	const struct bpf_ctx_arg_aux *info, u32 cnt)
23986	{
23987	prog->aux->ctx_arg_info = kmemdup_array(src: info, count: cnt, element_size: sizeof(*info), GFP_KERNEL_ACCOUNT);
23988	prog->aux->ctx_arg_info_size = cnt;
23989
23990	return prog->aux->ctx_arg_info ? 0 : -ENOMEM;
23991	}
23992
23993	static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
23994	{
23995	const struct btf_type *t, *func_proto;
23996	const struct bpf_struct_ops_desc *st_ops_desc;
23997	const struct bpf_struct_ops *st_ops;
23998	const struct btf_member *member;
23999	struct bpf_prog *prog = env->prog;
24000	bool has_refcounted_arg = false;
24001	u32 btf_id, member_idx, member_off;
24002	struct btf *btf;
24003	const char *mname;
24004	int i, err;
24005
24006	if (!prog->gpl_compatible) {
24007	verbose(private_data: env, fmt: "struct ops programs must have a GPL compatible license\n");
24008	return -EINVAL;
24009	}
24010
24011	if (!prog->aux->attach_btf_id)
24012	return -ENOTSUPP;
24013
24014	btf = prog->aux->attach_btf;
24015	if (btf_is_module(btf)) {
24016	/* Make sure st_ops is valid through the lifetime of env */
24017	env->attach_btf_mod = btf_try_get_module(btf);
24018	if (!env->attach_btf_mod) {
24019	verbose(private_data: env, fmt: "struct_ops module %s is not found\n",
24020	btf_get_name(btf));
24021	return -ENOTSUPP;
24022	}
24023	}
24024
24025	btf_id = prog->aux->attach_btf_id;
24026	st_ops_desc = bpf_struct_ops_find(btf, type_id: btf_id);
24027	if (!st_ops_desc) {
24028	verbose(private_data: env, fmt: "attach_btf_id %u is not a supported struct\n",
24029	btf_id);
24030	return -ENOTSUPP;
24031	}
24032	st_ops = st_ops_desc->st_ops;
24033
24034	t = st_ops_desc->type;
24035	member_idx = prog->expected_attach_type;
24036	if (member_idx >= btf_type_vlen(t)) {
24037	verbose(private_data: env, fmt: "attach to invalid member idx %u of struct %s\n",
24038	member_idx, st_ops->name);
24039	return -EINVAL;
24040	}
24041
24042	member = &btf_type_member(t)[member_idx];
24043	mname = btf_name_by_offset(btf, offset: member->name_off);
24044	func_proto = btf_type_resolve_func_ptr(btf, id: member->type,
24045	NULL);
24046	if (!func_proto) {
24047	verbose(private_data: env, fmt: "attach to invalid member %s(@idx %u) of struct %s\n",
24048	mname, member_idx, st_ops->name);
24049	return -EINVAL;
24050	}
24051
24052	member_off = __btf_member_bit_offset(struct_type: t, member) / 8;
24053	err = bpf_struct_ops_supported(st_ops, moff: member_off);
24054	if (err) {
24055	verbose(private_data: env, fmt: "attach to unsupported member %s of struct %s\n",
24056	mname, st_ops->name);
24057	return err;
24058	}
24059
24060	if (st_ops->check_member) {
24061	err = st_ops->check_member(t, member, prog);
24062
24063	if (err) {
24064	verbose(private_data: env, fmt: "attach to unsupported member %s of struct %s\n",
24065	mname, st_ops->name);
24066	return err;
24067	}
24068	}
24069
24070	if (prog->aux->priv_stack_requested && !bpf_jit_supports_private_stack()) {
24071	verbose(private_data: env, fmt: "Private stack not supported by jit\n");
24072	return -EACCES;
24073	}
24074
24075	for (i = 0; i < st_ops_desc->arg_info[member_idx].cnt; i++) {
24076	if (st_ops_desc->arg_info[member_idx].info->refcounted) {
24077	has_refcounted_arg = true;
24078	break;
24079	}
24080	}
24081
24082	/* Tail call is not allowed for programs with refcounted arguments since we
24083	* cannot guarantee that valid refcounted kptrs will be passed to the callee.
24084	*/
24085	for (i = 0; i < env->subprog_cnt; i++) {
24086	if (has_refcounted_arg && env->subprog_info[i].has_tail_call) {
24087	verbose(private_data: env, fmt: "program with __ref argument cannot tail call\n");
24088	return -EINVAL;
24089	}
24090	}
24091
24092	prog->aux->st_ops = st_ops;
24093	prog->aux->attach_st_ops_member_off = member_off;
24094
24095	prog->aux->attach_func_proto = func_proto;
24096	prog->aux->attach_func_name = mname;
24097	env->ops = st_ops->verifier_ops;
24098
24099	return bpf_prog_ctx_arg_info_init(prog, info: st_ops_desc->arg_info[member_idx].info,
24100	cnt: st_ops_desc->arg_info[member_idx].cnt);
24101	}
24102	#define SECURITY_PREFIX "security_"
24103
24104	static int check_attach_modify_return(unsigned long addr, const char *func_name)
24105	{
24106	if (within_error_injection_list(addr) ||
24107	!strncmp(SECURITY_PREFIX, func_name, sizeof(SECURITY_PREFIX) - 1))
24108	return 0;
24109
24110	return -EINVAL;
24111	}
24112
24113	/* list of non-sleepable functions that are otherwise on
24114	* ALLOW_ERROR_INJECTION list
24115	*/
24116	BTF_SET_START(btf_non_sleepable_error_inject)
24117	/* Three functions below can be called from sleepable and non-sleepable context.
24118	* Assume non-sleepable from bpf safety point of view.
24119	*/
24120	BTF_ID(func, __filemap_add_folio)
24121	#ifdef CONFIG_FAIL_PAGE_ALLOC
24122	BTF_ID(func, should_fail_alloc_page)
24123	#endif
24124	#ifdef CONFIG_FAILSLAB
24125	BTF_ID(func, should_failslab)
24126	#endif
24127	BTF_SET_END(btf_non_sleepable_error_inject)
24128
24129	static int check_non_sleepable_error_inject(u32 btf_id)
24130	{
24131	return btf_id_set_contains(set: &btf_non_sleepable_error_inject, id: btf_id);
24132	}
24133
24134	int bpf_check_attach_target(struct bpf_verifier_log *log,
24135	const struct bpf_prog *prog,
24136	const struct bpf_prog *tgt_prog,
24137	u32 btf_id,
24138	struct bpf_attach_target_info *tgt_info)
24139	{
24140	bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
24141	bool prog_tracing = prog->type == BPF_PROG_TYPE_TRACING;
24142	char trace_symbol[KSYM_SYMBOL_LEN];
24143	const char prefix[] = "btf_trace_";
24144	struct bpf_raw_event_map *btp;
24145	int ret = 0, subprog = -1, i;
24146	const struct btf_type *t;
24147	bool conservative = true;
24148	const char *tname, *fname;
24149	struct btf *btf;
24150	long addr = 0;
24151	struct module *mod = NULL;
24152
24153	if (!btf_id) {
24154	bpf_log(log, fmt: "Tracing programs must provide btf_id\n");
24155	return -EINVAL;
24156	}
24157	btf = tgt_prog ? tgt_prog->aux->btf : prog->aux->attach_btf;
24158	if (!btf) {
24159	bpf_log(log,
24160	fmt: "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
24161	return -EINVAL;
24162	}
24163	t = btf_type_by_id(btf, type_id: btf_id);
24164	if (!t) {
24165	bpf_log(log, fmt: "attach_btf_id %u is invalid\n", btf_id);
24166	return -EINVAL;
24167	}
24168	tname = btf_name_by_offset(btf, offset: t->name_off);
24169	if (!tname) {
24170	bpf_log(log, fmt: "attach_btf_id %u doesn't have a name\n", btf_id);
24171	return -EINVAL;
24172	}
24173	if (tgt_prog) {
24174	struct bpf_prog_aux *aux = tgt_prog->aux;
24175	bool tgt_changes_pkt_data;
24176	bool tgt_might_sleep;
24177
24178	if (bpf_prog_is_dev_bound(aux: prog->aux) &&
24179	!bpf_prog_dev_bound_match(lhs: prog, rhs: tgt_prog)) {
24180	bpf_log(log, fmt: "Target program bound device mismatch");
24181	return -EINVAL;
24182	}
24183
24184	for (i = 0; i < aux->func_info_cnt; i++)
24185	if (aux->func_info[i].type_id == btf_id) {
24186	subprog = i;
24187	break;
24188	}
24189	if (subprog == -1) {
24190	bpf_log(log, fmt: "Subprog %s doesn't exist\n", tname);
24191	return -EINVAL;
24192	}
24193	if (aux->func && aux->func[subprog]->aux->exception_cb) {
24194	bpf_log(log,
24195	fmt: "%s programs cannot attach to exception callback\n",
24196	prog_extension ? "Extension" : "FENTRY/FEXIT");
24197	return -EINVAL;
24198	}
24199	conservative = aux->func_info_aux[subprog].unreliable;
24200	if (prog_extension) {
24201	if (conservative) {
24202	bpf_log(log,
24203	fmt: "Cannot replace static functions\n");
24204	return -EINVAL;
24205	}
24206	if (!prog->jit_requested) {
24207	bpf_log(log,
24208	fmt: "Extension programs should be JITed\n");
24209	return -EINVAL;
24210	}
24211	tgt_changes_pkt_data = aux->func
24212	? aux->func[subprog]->aux->changes_pkt_data
24213	: aux->changes_pkt_data;
24214	if (prog->aux->changes_pkt_data && !tgt_changes_pkt_data) {
24215	bpf_log(log,
24216	fmt: "Extension program changes packet data, while original does not\n");
24217	return -EINVAL;
24218	}
24219
24220	tgt_might_sleep = aux->func
24221	? aux->func[subprog]->aux->might_sleep
24222	: aux->might_sleep;
24223	if (prog->aux->might_sleep && !tgt_might_sleep) {
24224	bpf_log(log,
24225	fmt: "Extension program may sleep, while original does not\n");
24226	return -EINVAL;
24227	}
24228	}
24229	if (!tgt_prog->jited) {
24230	bpf_log(log, fmt: "Can attach to only JITed progs\n");
24231	return -EINVAL;
24232	}
24233	if (prog_tracing) {
24234	if (aux->attach_tracing_prog) {
24235	/*
24236	* Target program is an fentry/fexit which is already attached
24237	* to another tracing program. More levels of nesting
24238	* attachment are not allowed.
24239	*/
24240	bpf_log(log, fmt: "Cannot nest tracing program attach more than once\n");
24241	return -EINVAL;
24242	}
24243	} else if (tgt_prog->type == prog->type) {
24244	/*
24245	* To avoid potential call chain cycles, prevent attaching of a
24246	* program extension to another extension. It's ok to attach
24247	* fentry/fexit to extension program.
24248	*/
24249	bpf_log(log, fmt: "Cannot recursively attach\n");
24250	return -EINVAL;
24251	}
24252	if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
24253	prog_extension &&
24254	(tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
24255	tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
24256	/* Program extensions can extend all program types
24257	* except fentry/fexit. The reason is the following.
24258	* The fentry/fexit programs are used for performance
24259	* analysis, stats and can be attached to any program
24260	* type. When extension program is replacing XDP function
24261	* it is necessary to allow performance analysis of all
24262	* functions. Both original XDP program and its program
24263	* extension. Hence attaching fentry/fexit to
24264	* BPF_PROG_TYPE_EXT is allowed. If extending of
24265	* fentry/fexit was allowed it would be possible to create
24266	* long call chain fentry->extension->fentry->extension
24267	* beyond reasonable stack size. Hence extending fentry
24268	* is not allowed.
24269	*/
24270	bpf_log(log, fmt: "Cannot extend fentry/fexit\n");
24271	return -EINVAL;
24272	}
24273	} else {
24274	if (prog_extension) {
24275	bpf_log(log, fmt: "Cannot replace kernel functions\n");
24276	return -EINVAL;
24277	}
24278	}
24279
24280	switch (prog->expected_attach_type) {
24281	case BPF_TRACE_RAW_TP:
24282	if (tgt_prog) {
24283	bpf_log(log,
24284	fmt: "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
24285	return -EINVAL;
24286	}
24287	if (!btf_type_is_typedef(t)) {
24288	bpf_log(log, fmt: "attach_btf_id %u is not a typedef\n",
24289	btf_id);
24290	return -EINVAL;
24291	}
24292	if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
24293	bpf_log(log, fmt: "attach_btf_id %u points to wrong type name %s\n",
24294	btf_id, tname);
24295	return -EINVAL;
24296	}
24297	tname += sizeof(prefix) - 1;
24298
24299	/* The func_proto of "btf_trace_##tname" is generated from typedef without argument
24300	* names. Thus using bpf_raw_event_map to get argument names.
24301	*/
24302	btp = bpf_get_raw_tracepoint(name: tname);
24303	if (!btp)
24304	return -EINVAL;
24305	fname = kallsyms_lookup(addr: (unsigned long)btp->bpf_func, NULL, NULL, NULL,
24306	namebuf: trace_symbol);
24307	bpf_put_raw_tracepoint(btp);
24308
24309	if (fname)
24310	ret = btf_find_by_name_kind(btf, name: fname, kind: BTF_KIND_FUNC);
24311
24312	if (!fname || ret < 0) {
24313	bpf_log(log, fmt: "Cannot find btf of tracepoint template, fall back to %s%s.\n",
24314	prefix, tname);
24315	t = btf_type_by_id(btf, type_id: t->type);
24316	if (!btf_type_is_ptr(t))
24317	/* should never happen in valid vmlinux build */
24318	return -EINVAL;
24319	} else {
24320	t = btf_type_by_id(btf, type_id: ret);
24321	if (!btf_type_is_func(t))
24322	/* should never happen in valid vmlinux build */
24323	return -EINVAL;
24324	}
24325
24326	t = btf_type_by_id(btf, type_id: t->type);
24327	if (!btf_type_is_func_proto(t))
24328	/* should never happen in valid vmlinux build */
24329	return -EINVAL;
24330
24331	break;
24332	case BPF_TRACE_ITER:
24333	if (!btf_type_is_func(t)) {
24334	bpf_log(log, fmt: "attach_btf_id %u is not a function\n",
24335	btf_id);
24336	return -EINVAL;
24337	}
24338	t = btf_type_by_id(btf, type_id: t->type);
24339	if (!btf_type_is_func_proto(t))
24340	return -EINVAL;
24341	ret = btf_distill_func_proto(log, btf, func_proto: t, func_name: tname, m: &tgt_info->fmodel);
24342	if (ret)
24343	return ret;
24344	break;
24345	default:
24346	if (!prog_extension)
24347	return -EINVAL;
24348	fallthrough;
24349	case BPF_MODIFY_RETURN:
24350	case BPF_LSM_MAC:
24351	case BPF_LSM_CGROUP:
24352	case BPF_TRACE_FENTRY:
24353	case BPF_TRACE_FEXIT:
24354	if (!btf_type_is_func(t)) {
24355	bpf_log(log, fmt: "attach_btf_id %u is not a function\n",
24356	btf_id);
24357	return -EINVAL;
24358	}
24359	if (prog_extension &&
24360	btf_check_type_match(log, prog, btf, t))
24361	return -EINVAL;
24362	t = btf_type_by_id(btf, type_id: t->type);
24363	if (!btf_type_is_func_proto(t))
24364	return -EINVAL;
24365
24366	if ((prog->aux->saved_dst_prog_type || prog->aux->saved_dst_attach_type) &&
24367	(!tgt_prog || prog->aux->saved_dst_prog_type != tgt_prog->type ||
24368	prog->aux->saved_dst_attach_type != tgt_prog->expected_attach_type))
24369	return -EINVAL;
24370
24371	if (tgt_prog && conservative)
24372	t = NULL;
24373
24374	ret = btf_distill_func_proto(log, btf, func_proto: t, func_name: tname, m: &tgt_info->fmodel);
24375	if (ret < 0)
24376	return ret;
24377
24378	if (tgt_prog) {
24379	if (subprog == 0)
24380	addr = (long) tgt_prog->bpf_func;
24381	else
24382	addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
24383	} else {
24384	if (btf_is_module(btf)) {
24385	mod = btf_try_get_module(btf);
24386	if (mod)
24387	addr = find_kallsyms_symbol_value(mod, name: tname);
24388	else
24389	addr = 0;
24390	} else {
24391	addr = kallsyms_lookup_name(name: tname);
24392	}
24393	if (!addr) {
24394	module_put(module: mod);
24395	bpf_log(log,
24396	fmt: "The address of function %s cannot be found\n",
24397	tname);
24398	return -ENOENT;
24399	}
24400	}
24401
24402	if (prog->sleepable) {
24403	ret = -EINVAL;
24404	switch (prog->type) {
24405	case BPF_PROG_TYPE_TRACING:
24406
24407	/* fentry/fexit/fmod_ret progs can be sleepable if they are
24408	* attached to ALLOW_ERROR_INJECTION and are not in denylist.
24409	*/
24410	if (!check_non_sleepable_error_inject(btf_id) &&
24411	within_error_injection_list(addr))
24412	ret = 0;
24413	/* fentry/fexit/fmod_ret progs can also be sleepable if they are
24414	* in the fmodret id set with the KF_SLEEPABLE flag.
24415	*/
24416	else {
24417	u32 *flags = btf_kfunc_is_modify_return(btf, kfunc_btf_id: btf_id,
24418	prog);
24419
24420	if (flags && (*flags & KF_SLEEPABLE))
24421	ret = 0;
24422	}
24423	break;
24424	case BPF_PROG_TYPE_LSM:
24425	/* LSM progs check that they are attached to bpf_lsm_*() funcs.
24426	* Only some of them are sleepable.
24427	*/
24428	if (bpf_lsm_is_sleepable_hook(btf_id))
24429	ret = 0;
24430	break;
24431	default:
24432	break;
24433	}
24434	if (ret) {
24435	module_put(module: mod);
24436	bpf_log(log, fmt: "%s is not sleepable\n", tname);
24437	return ret;
24438	}
24439	} else if (prog->expected_attach_type == BPF_MODIFY_RETURN) {
24440	if (tgt_prog) {
24441	module_put(module: mod);
24442	bpf_log(log, fmt: "can't modify return codes of BPF programs\n");
24443	return -EINVAL;
24444	}
24445	ret = -EINVAL;
24446	if (btf_kfunc_is_modify_return(btf, kfunc_btf_id: btf_id, prog) ||
24447	!check_attach_modify_return(addr, func_name: tname))
24448	ret = 0;
24449	if (ret) {
24450	module_put(module: mod);
24451	bpf_log(log, fmt: "%s() is not modifiable\n", tname);
24452	return ret;
24453	}
24454	}
24455
24456	break;
24457	}
24458	tgt_info->tgt_addr = addr;
24459	tgt_info->tgt_name = tname;
24460	tgt_info->tgt_type = t;
24461	tgt_info->tgt_mod = mod;
24462	return 0;
24463	}
24464
24465	BTF_SET_START(btf_id_deny)
24466	BTF_ID_UNUSED
24467	#ifdef CONFIG_SMP
24468	BTF_ID(func, ___migrate_enable)
24469	BTF_ID(func, migrate_disable)
24470	BTF_ID(func, migrate_enable)
24471	#endif
24472	#if !defined CONFIG_PREEMPT_RCU && !defined CONFIG_TINY_RCU
24473	BTF_ID(func, rcu_read_unlock_strict)
24474	#endif
24475	#if defined(CONFIG_DEBUG_PREEMPT) || defined(CONFIG_TRACE_PREEMPT_TOGGLE)
24476	BTF_ID(func, preempt_count_add)
24477	BTF_ID(func, preempt_count_sub)
24478	#endif
24479	#ifdef CONFIG_PREEMPT_RCU
24480	BTF_ID(func, __rcu_read_lock)
24481	BTF_ID(func, __rcu_read_unlock)
24482	#endif
24483	BTF_SET_END(btf_id_deny)
24484
24485	/* fexit and fmod_ret can't be used to attach to __noreturn functions.
24486	* Currently, we must manually list all __noreturn functions here. Once a more
24487	* robust solution is implemented, this workaround can be removed.
24488	*/
24489	BTF_SET_START(noreturn_deny)
24490	#ifdef CONFIG_IA32_EMULATION
24491	BTF_ID(func, __ia32_sys_exit)
24492	BTF_ID(func, __ia32_sys_exit_group)
24493	#endif
24494	#ifdef CONFIG_KUNIT
24495	BTF_ID(func, __kunit_abort)
24496	BTF_ID(func, kunit_try_catch_throw)
24497	#endif
24498	#ifdef CONFIG_MODULES
24499	BTF_ID(func, __module_put_and_kthread_exit)
24500	#endif
24501	#ifdef CONFIG_X86_64
24502	BTF_ID(func, __x64_sys_exit)
24503	BTF_ID(func, __x64_sys_exit_group)
24504	#endif
24505	BTF_ID(func, do_exit)
24506	BTF_ID(func, do_group_exit)
24507	BTF_ID(func, kthread_complete_and_exit)
24508	BTF_ID(func, kthread_exit)
24509	BTF_ID(func, make_task_dead)
24510	BTF_SET_END(noreturn_deny)
24511
24512	static bool can_be_sleepable(struct bpf_prog *prog)
24513	{
24514	if (prog->type == BPF_PROG_TYPE_TRACING) {
24515	switch (prog->expected_attach_type) {
24516	case BPF_TRACE_FENTRY:
24517	case BPF_TRACE_FEXIT:
24518	case BPF_MODIFY_RETURN:
24519	case BPF_TRACE_ITER:
24520	return true;
24521	default:
24522	return false;
24523	}
24524	}
24525	return prog->type == BPF_PROG_TYPE_LSM ||
24526	prog->type == BPF_PROG_TYPE_KPROBE /* only for uprobes */ ||
24527	prog->type == BPF_PROG_TYPE_STRUCT_OPS;
24528	}
24529
24530	static int check_attach_btf_id(struct bpf_verifier_env *env)
24531	{
24532	struct bpf_prog *prog = env->prog;
24533	struct bpf_prog *tgt_prog = prog->aux->dst_prog;
24534	struct bpf_attach_target_info tgt_info = {};
24535	u32 btf_id = prog->aux->attach_btf_id;
24536	struct bpf_trampoline *tr;
24537	int ret;
24538	u64 key;
24539
24540	if (prog->type == BPF_PROG_TYPE_SYSCALL) {
24541	if (prog->sleepable)
24542	/* attach_btf_id checked to be zero already */
24543	return 0;
24544	verbose(private_data: env, fmt: "Syscall programs can only be sleepable\n");
24545	return -EINVAL;
24546	}
24547
24548	if (prog->sleepable && !can_be_sleepable(prog)) {
24549	verbose(private_data: env, fmt: "Only fentry/fexit/fmod_ret, lsm, iter, uprobe, and struct_ops programs can be sleepable\n");
24550	return -EINVAL;
24551	}
24552
24553	if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
24554	return check_struct_ops_btf_id(env);
24555
24556	if (prog->type != BPF_PROG_TYPE_TRACING &&
24557	prog->type != BPF_PROG_TYPE_LSM &&
24558	prog->type != BPF_PROG_TYPE_EXT)
24559	return 0;
24560
24561	ret = bpf_check_attach_target(log: &env->log, prog, tgt_prog, btf_id, tgt_info: &tgt_info);
24562	if (ret)
24563	return ret;
24564
24565	if (tgt_prog && prog->type == BPF_PROG_TYPE_EXT) {
24566	/* to make freplace equivalent to their targets, they need to
24567	* inherit env->ops and expected_attach_type for the rest of the
24568	* verification
24569	*/
24570	env->ops = bpf_verifier_ops[tgt_prog->type];
24571	prog->expected_attach_type = tgt_prog->expected_attach_type;
24572	}
24573
24574	/* store info about the attachment target that will be used later */
24575	prog->aux->attach_func_proto = tgt_info.tgt_type;
24576	prog->aux->attach_func_name = tgt_info.tgt_name;
24577	prog->aux->mod = tgt_info.tgt_mod;
24578
24579	if (tgt_prog) {
24580	prog->aux->saved_dst_prog_type = tgt_prog->type;
24581	prog->aux->saved_dst_attach_type = tgt_prog->expected_attach_type;
24582	}
24583
24584	if (prog->expected_attach_type == BPF_TRACE_RAW_TP) {
24585	prog->aux->attach_btf_trace = true;
24586	return 0;
24587	} else if (prog->expected_attach_type == BPF_TRACE_ITER) {
24588	return bpf_iter_prog_supported(prog);
24589	}
24590
24591	if (prog->type == BPF_PROG_TYPE_LSM) {
24592	ret = bpf_lsm_verify_prog(vlog: &env->log, prog);
24593	if (ret < 0)
24594	return ret;
24595	} else if (prog->type == BPF_PROG_TYPE_TRACING &&
24596	btf_id_set_contains(set: &btf_id_deny, id: btf_id)) {
24597	verbose(private_data: env, fmt: "Attaching tracing programs to function '%s' is rejected.\n",
24598	tgt_info.tgt_name);
24599	return -EINVAL;
24600	} else if ((prog->expected_attach_type == BPF_TRACE_FEXIT ||
24601	prog->expected_attach_type == BPF_MODIFY_RETURN) &&
24602	btf_id_set_contains(set: &noreturn_deny, id: btf_id)) {
24603	verbose(private_data: env, fmt: "Attaching fexit/fmod_ret to __noreturn function '%s' is rejected.\n",
24604	tgt_info.tgt_name);
24605	return -EINVAL;
24606	}
24607
24608	key = bpf_trampoline_compute_key(tgt_prog, btf: prog->aux->attach_btf, btf_id);
24609	tr = bpf_trampoline_get(key, tgt_info: &tgt_info);
24610	if (!tr)
24611	return -ENOMEM;
24612
24613	if (tgt_prog && tgt_prog->aux->tail_call_reachable)
24614	tr->flags = BPF_TRAMP_F_TAIL_CALL_CTX;
24615
24616	prog->aux->dst_trampoline = tr;
24617	return 0;
24618	}
24619
24620	struct btf *bpf_get_btf_vmlinux(void)
24621	{
24622	if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
24623	mutex_lock(&bpf_verifier_lock);
24624	if (!btf_vmlinux)
24625	btf_vmlinux = btf_parse_vmlinux();
24626	mutex_unlock(lock: &bpf_verifier_lock);
24627	}
24628	return btf_vmlinux;
24629	}
24630
24631	/*
24632	* The add_fd_from_fd_array() is executed only if fd_array_cnt is non-zero. In
24633	* this case expect that every file descriptor in the array is either a map or
24634	* a BTF. Everything else is considered to be trash.
24635	*/
24636	static int add_fd_from_fd_array(struct bpf_verifier_env *env, int fd)
24637	{
24638	struct bpf_map *map;
24639	struct btf *btf;
24640	CLASS(fd, f)(fd);
24641	int err;
24642
24643	map = __bpf_map_get(f);
24644	if (!IS_ERR(ptr: map)) {
24645	err = __add_used_map(env, map);
24646	if (err < 0)
24647	return err;
24648	return 0;
24649	}
24650
24651	btf = __btf_get_by_fd(f);
24652	if (!IS_ERR(ptr: btf)) {
24653	err = __add_used_btf(env, btf);
24654	if (err < 0)
24655	return err;
24656	return 0;
24657	}
24658
24659	verbose(private_data: env, fmt: "fd %d is not pointing to valid bpf_map or btf\n", fd);
24660	return PTR_ERR(ptr: map);
24661	}
24662
24663	static int process_fd_array(struct bpf_verifier_env *env, union bpf_attr *attr, bpfptr_t uattr)
24664	{
24665	size_t size = sizeof(int);
24666	int ret;
24667	int fd;
24668	u32 i;
24669
24670	env->fd_array = make_bpfptr(addr: attr->fd_array, is_kernel: uattr.is_kernel);
24671
24672	/*
24673	* The only difference between old (no fd_array_cnt is given) and new
24674	* APIs is that in the latter case the fd_array is expected to be
24675	* continuous and is scanned for map fds right away
24676	*/
24677	if (!attr->fd_array_cnt)
24678	return 0;
24679
24680	/* Check for integer overflow */
24681	if (attr->fd_array_cnt >= (U32_MAX / size)) {
24682	verbose(private_data: env, fmt: "fd_array_cnt is too big (%u)\n", attr->fd_array_cnt);
24683	return -EINVAL;
24684	}
24685
24686	for (i = 0; i < attr->fd_array_cnt; i++) {
24687	if (copy_from_bpfptr_offset(dst: &fd, src: env->fd_array, offset: i * size, size))
24688	return -EFAULT;
24689
24690	ret = add_fd_from_fd_array(env, fd);
24691	if (ret)
24692	return ret;
24693	}
24694
24695	return 0;
24696	}
24697
24698	/* Each field is a register bitmask */
24699	struct insn_live_regs {
24700	u16 use; /* registers read by instruction */
24701	u16 def; /* registers written by instruction */
24702	u16 in; /* registers that may be alive before instruction */
24703	u16 out; /* registers that may be alive after instruction */
24704	};
24705
24706	/* Bitmask with 1s for all caller saved registers */
24707	#define ALL_CALLER_SAVED_REGS ((1u << CALLER_SAVED_REGS) - 1)
24708
24709	/* Compute info->{use,def} fields for the instruction */
24710	static void compute_insn_live_regs(struct bpf_verifier_env *env,
24711	struct bpf_insn *insn,
24712	struct insn_live_regs *info)
24713	{
24714	struct call_summary cs;
24715	u8 class = BPF_CLASS(insn->code);
24716	u8 code = BPF_OP(insn->code);
24717	u8 mode = BPF_MODE(insn->code);
24718	u16 src = BIT(insn->src_reg);
24719	u16 dst = BIT(insn->dst_reg);
24720	u16 r0 = BIT(0);
24721	u16 def = 0;
24722	u16 use = 0xffff;
24723
24724	switch (class) {
24725	case BPF_LD:
24726	switch (mode) {
24727	case BPF_IMM:
24728	if (BPF_SIZE(insn->code) == BPF_DW) {
24729	def = dst;
24730	use = 0;
24731	}
24732	break;
24733	case BPF_LD | BPF_ABS:
24734	case BPF_LD | BPF_IND:
24735	/* stick with defaults */
24736	break;
24737	}
24738	break;
24739	case BPF_LDX:
24740	switch (mode) {
24741	case BPF_MEM:
24742	case BPF_MEMSX:
24743	def = dst;
24744	use = src;
24745	break;
24746	}
24747	break;
24748	case BPF_ST:
24749	switch (mode) {
24750	case BPF_MEM:
24751	def = 0;
24752	use = dst;
24753	break;
24754	}
24755	break;
24756	case BPF_STX:
24757	switch (mode) {
24758	case BPF_MEM:
24759	def = 0;
24760	use = dst | src;
24761	break;
24762	case BPF_ATOMIC:
24763	switch (insn->imm) {
24764	case BPF_CMPXCHG:
24765	use = r0 | dst | src;
24766	def = r0;
24767	break;
24768	case BPF_LOAD_ACQ:
24769	def = dst;
24770	use = src;
24771	break;
24772	case BPF_STORE_REL:
24773	def = 0;
24774	use = dst | src;
24775	break;
24776	default:
24777	use = dst | src;
24778	if (insn->imm & BPF_FETCH)
24779	def = src;
24780	else
24781	def = 0;
24782	}
24783	break;
24784	}
24785	break;
24786	case BPF_ALU:
24787	case BPF_ALU64:
24788	switch (code) {
24789	case BPF_END:
24790	use = dst;
24791	def = dst;
24792	break;
24793	case BPF_MOV:
24794	def = dst;
24795	if (BPF_SRC(insn->code) == BPF_K)
24796	use = 0;
24797	else
24798	use = src;
24799	break;
24800	default:
24801	def = dst;
24802	if (BPF_SRC(insn->code) == BPF_K)
24803	use = dst;
24804	else
24805	use = dst | src;
24806	}
24807	break;
24808	case BPF_JMP:
24809	case BPF_JMP32:
24810	switch (code) {
24811	case BPF_JA:
24812	case BPF_JCOND:
24813	def = 0;
24814	use = 0;
24815	break;
24816	case BPF_EXIT:
24817	def = 0;
24818	use = r0;
24819	break;
24820	case BPF_CALL:
24821	def = ALL_CALLER_SAVED_REGS;
24822	use = def & ~BIT(BPF_REG_0);
24823	if (get_call_summary(env, call: insn, cs: &cs))
24824	use = GENMASK(cs.num_params, 1);
24825	break;
24826	default:
24827	def = 0;
24828	if (BPF_SRC(insn->code) == BPF_K)
24829	use = dst;
24830	else
24831	use = dst | src;
24832	}
24833	break;
24834	}
24835
24836	info->def = def;
24837	info->use = use;
24838	}
24839
24840	/* Compute may-live registers after each instruction in the program.
24841	* The register is live after the instruction I if it is read by some
24842	* instruction S following I during program execution and is not
24843	* overwritten between I and S.
24844	*
24845	* Store result in env->insn_aux_data[i].live_regs.
24846	*/
24847	static int compute_live_registers(struct bpf_verifier_env *env)
24848	{
24849	struct bpf_insn_aux_data *insn_aux = env->insn_aux_data;
24850	struct bpf_insn *insns = env->prog->insnsi;
24851	struct insn_live_regs *state;
24852	int insn_cnt = env->prog->len;
24853	int err = 0, i, j;
24854	bool changed;
24855
24856	/* Use the following algorithm:
24857	* - define the following:
24858	* - I.use : a set of all registers read by instruction I;
24859	* - I.def : a set of all registers written by instruction I;
24860	* - I.in : a set of all registers that may be alive before I execution;
24861	* - I.out : a set of all registers that may be alive after I execution;
24862	* - insn_successors(I): a set of instructions S that might immediately
24863	* follow I for some program execution;
24864	* - associate separate empty sets 'I.in' and 'I.out' with each instruction;
24865	* - visit each instruction in a postorder and update
24866	* state[i].in, state[i].out as follows:
24867	*
24868	* state[i].out = U [state[s].in for S in insn_successors(i)]
24869	* state[i].in = (state[i].out / state[i].def) U state[i].use
24870	*
24871	* (where U stands for set union, / stands for set difference)
24872	* - repeat the computation while {in,out} fields changes for
24873	* any instruction.
24874	*/
24875	state = kvcalloc(insn_cnt, sizeof(*state), GFP_KERNEL_ACCOUNT);
24876	if (!state) {
24877	err = -ENOMEM;
24878	goto out;
24879	}
24880
24881	for (i = 0; i < insn_cnt; ++i)
24882	compute_insn_live_regs(env, insn: &insns[i], info: &state[i]);
24883
24884	changed = true;
24885	while (changed) {
24886	changed = false;
24887	for (i = 0; i < env->cfg.cur_postorder; ++i) {
24888	int insn_idx = env->cfg.insn_postorder[i];
24889	struct insn_live_regs *live = &state[insn_idx];
24890	struct bpf_iarray *succ;
24891	u16 new_out = 0;
24892	u16 new_in = 0;
24893
24894	succ = bpf_insn_successors(env, idx: insn_idx);
24895	for (int s = 0; s < succ->cnt; ++s)
24896	new_out |= state[succ->items[s]].in;
24897	new_in = (new_out & ~live->def) | live->use;
24898	if (new_out != live->out || new_in != live->in) {
24899	live->in = new_in;
24900	live->out = new_out;
24901	changed = true;
24902	}
24903	}
24904	}
24905
24906	for (i = 0; i < insn_cnt; ++i)
24907	insn_aux[i].live_regs_before = state[i].in;
24908
24909	if (env->log.level & BPF_LOG_LEVEL2) {
24910	verbose(private_data: env, fmt: "Live regs before insn:\n");
24911	for (i = 0; i < insn_cnt; ++i) {
24912	if (env->insn_aux_data[i].scc)
24913	verbose(private_data: env, fmt: "%3d ", env->insn_aux_data[i].scc);
24914	else
24915	verbose(private_data: env, fmt: " ");
24916	verbose(private_data: env, fmt: "%3d: ", i);
24917	for (j = BPF_REG_0; j < BPF_REG_10; ++j)
24918	if (insn_aux[i].live_regs_before & BIT(j))
24919	verbose(private_data: env, fmt: "%d", j);
24920	else
24921	verbose(private_data: env, fmt: ".");
24922	verbose(private_data: env, fmt: " ");
24923	verbose_insn(env, insn: &insns[i]);
24924	if (bpf_is_ldimm64(insn: &insns[i]))
24925	i++;
24926	}
24927	}
24928
24929	out:
24930	kvfree(addr: state);
24931	return err;
24932	}
24933
24934	/*
24935	* Compute strongly connected components (SCCs) on the CFG.
24936	* Assign an SCC number to each instruction, recorded in env->insn_aux[*].scc.
24937	* If instruction is a sole member of its SCC and there are no self edges,
24938	* assign it SCC number of zero.
24939	* Uses a non-recursive adaptation of Tarjan's algorithm for SCC computation.
24940	*/
24941	static int compute_scc(struct bpf_verifier_env *env)
24942	{
24943	const u32 NOT_ON_STACK = U32_MAX;
24944
24945	struct bpf_insn_aux_data *aux = env->insn_aux_data;
24946	const u32 insn_cnt = env->prog->len;
24947	int stack_sz, dfs_sz, err = 0;
24948	u32 *stack, *pre, *low, *dfs;
24949	u32 i, j, t, w;
24950	u32 next_preorder_num;
24951	u32 next_scc_id;
24952	bool assign_scc;
24953	struct bpf_iarray *succ;
24954
24955	next_preorder_num = 1;
24956	next_scc_id = 1;
24957	/*
24958	* - 'stack' accumulates vertices in DFS order, see invariant comment below;
24959	* - 'pre[t] == p' => preorder number of vertex 't' is 'p';
24960	* - 'low[t] == n' => smallest preorder number of the vertex reachable from 't' is 'n';
24961	* - 'dfs' DFS traversal stack, used to emulate explicit recursion.
24962	*/
24963	stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL_ACCOUNT);
24964	pre = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL_ACCOUNT);
24965	low = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL_ACCOUNT);
24966	dfs = kvcalloc(insn_cnt, sizeof(*dfs), GFP_KERNEL_ACCOUNT);
24967	if (!stack || !pre || !low || !dfs) {
24968	err = -ENOMEM;
24969	goto exit;
24970	}
24971	/*
24972	* References:
24973	* [1] R. Tarjan "Depth-First Search and Linear Graph Algorithms"
24974	* [2] D. J. Pearce "A Space-Efficient Algorithm for Finding Strongly Connected Components"
24975	*
24976	* The algorithm maintains the following invariant:
24977	* - suppose there is a path 'u' ~> 'v', such that 'pre[v] < pre[u]';
24978	* - then, vertex 'u' remains on stack while vertex 'v' is on stack.
24979	*
24980	* Consequently:
24981	* - If 'low[v] < pre[v]', there is a path from 'v' to some vertex 'u',
24982	* such that 'pre[u] == low[v]'; vertex 'u' is currently on the stack,
24983	* and thus there is an SCC (loop) containing both 'u' and 'v'.
24984	* - If 'low[v] == pre[v]', loops containing 'v' have been explored,
24985	* and 'v' can be considered the root of some SCC.
24986	*
24987	* Here is a pseudo-code for an explicitly recursive version of the algorithm:
24988	*
24989	* NOT_ON_STACK = insn_cnt + 1
24990	* pre = [0] * insn_cnt
24991	* low = [0] * insn_cnt
24992	* scc = [0] * insn_cnt
24993	* stack = []
24994	*
24995	* next_preorder_num = 1
24996	* next_scc_id = 1
24997	*
24998	* def recur(w):
24999	* nonlocal next_preorder_num
25000	* nonlocal next_scc_id
25001	*
25002	* pre[w] = next_preorder_num
25003	* low[w] = next_preorder_num
25004	* next_preorder_num += 1
25005	* stack.append(w)
25006	* for s in successors(w):
25007	* # Note: for classic algorithm the block below should look as:
25008	* #
25009	* # if pre[s] == 0:
25010	* # recur(s)
25011	* # low[w] = min(low[w], low[s])
25012	* # elif low[s] != NOT_ON_STACK:
25013	* # low[w] = min(low[w], pre[s])
25014	* #
25015	* # But replacing both 'min' instructions with 'low[w] = min(low[w], low[s])'
25016	* # does not break the invariant and makes itartive version of the algorithm
25017	* # simpler. See 'Algorithm #3' from [2].
25018	*
25019	* # 's' not yet visited
25020	* if pre[s] == 0:
25021	* recur(s)
25022	* # if 's' is on stack, pick lowest reachable preorder number from it;
25023	* # if 's' is not on stack 'low[s] == NOT_ON_STACK > low[w]',
25024	* # so 'min' would be a noop.
25025	* low[w] = min(low[w], low[s])
25026	*
25027	* if low[w] == pre[w]:
25028	* # 'w' is the root of an SCC, pop all vertices
25029	* # below 'w' on stack and assign same SCC to them.
25030	* while True:
25031	* t = stack.pop()
25032	* low[t] = NOT_ON_STACK
25033	* scc[t] = next_scc_id
25034	* if t == w:
25035	* break
25036	* next_scc_id += 1
25037	*
25038	* for i in range(0, insn_cnt):
25039	* if pre[i] == 0:
25040	* recur(i)
25041	*
25042	* Below implementation replaces explicit recursion with array 'dfs'.
25043	*/
25044	for (i = 0; i < insn_cnt; i++) {
25045	if (pre[i])
25046	continue;
25047	stack_sz = 0;
25048	dfs_sz = 1;
25049	dfs[0] = i;
25050	dfs_continue:
25051	while (dfs_sz) {
25052	w = dfs[dfs_sz - 1];
25053	if (pre[w] == 0) {
25054	low[w] = next_preorder_num;
25055	pre[w] = next_preorder_num;
25056	next_preorder_num++;
25057	stack[stack_sz++] = w;
25058	}
25059	/* Visit 'w' successors */
25060	succ = bpf_insn_successors(env, idx: w);
25061	for (j = 0; j < succ->cnt; ++j) {
25062	if (pre[succ->items[j]]) {
25063	low[w] = min(low[w], low[succ->items[j]]);
25064	} else {
25065	dfs[dfs_sz++] = succ->items[j];
25066	goto dfs_continue;
25067	}
25068	}
25069	/*
25070	* Preserve the invariant: if some vertex above in the stack
25071	* is reachable from 'w', keep 'w' on the stack.
25072	*/
25073	if (low[w] < pre[w]) {
25074	dfs_sz--;
25075	goto dfs_continue;
25076	}
25077	/*
25078	* Assign SCC number only if component has two or more elements,
25079	* or if component has a self reference.
25080	*/
25081	assign_scc = stack[stack_sz - 1] != w;
25082	for (j = 0; j < succ->cnt; ++j) {
25083	if (succ->items[j] == w) {
25084	assign_scc = true;
25085	break;
25086	}
25087	}
25088	/* Pop component elements from stack */
25089	do {
25090	t = stack[--stack_sz];
25091	low[t] = NOT_ON_STACK;
25092	if (assign_scc)
25093	aux[t].scc = next_scc_id;
25094	} while (t != w);
25095	if (assign_scc)
25096	next_scc_id++;
25097	dfs_sz--;
25098	}
25099	}
25100	env->scc_info = kvcalloc(next_scc_id, sizeof(*env->scc_info), GFP_KERNEL_ACCOUNT);
25101	if (!env->scc_info) {
25102	err = -ENOMEM;
25103	goto exit;
25104	}
25105	env->scc_cnt = next_scc_id;
25106	exit:
25107	kvfree(addr: stack);
25108	kvfree(addr: pre);
25109	kvfree(addr: low);
25110	kvfree(addr: dfs);
25111	return err;
25112	}
25113
25114	int bpf_check(struct bpf_prog **prog, union bpf_attr *attr, bpfptr_t uattr, __u32 uattr_size)
25115	{
25116	u64 start_time = ktime_get_ns();
25117	struct bpf_verifier_env *env;
25118	int i, len, ret = -EINVAL, err;
25119	u32 log_true_size;
25120	bool is_priv;
25121
25122	BTF_TYPE_EMIT(enum bpf_features);
25123
25124	/* no program is valid */
25125	if (ARRAY_SIZE(bpf_verifier_ops) == 0)
25126	return -EINVAL;
25127
25128	/* 'struct bpf_verifier_env' can be global, but since it's not small,
25129	* allocate/free it every time bpf_check() is called
25130	*/
25131	env = kvzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL_ACCOUNT);
25132	if (!env)
25133	return -ENOMEM;
25134
25135	env->bt.env = env;
25136
25137	len = (*prog)->len;
25138	env->insn_aux_data =
25139	vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
25140	ret = -ENOMEM;
25141	if (!env->insn_aux_data)
25142	goto err_free_env;
25143	for (i = 0; i < len; i++)
25144	env->insn_aux_data[i].orig_idx = i;
25145	env->succ = iarray_realloc(NULL, n_elem: 2);
25146	if (!env->succ)
25147	goto err_free_env;
25148	env->prog = *prog;
25149	env->ops = bpf_verifier_ops[env->prog->type];
25150
25151	env->allow_ptr_leaks = bpf_allow_ptr_leaks(token: env->prog->aux->token);
25152	env->allow_uninit_stack = bpf_allow_uninit_stack(token: env->prog->aux->token);
25153	env->bypass_spec_v1 = bpf_bypass_spec_v1(token: env->prog->aux->token);
25154	env->bypass_spec_v4 = bpf_bypass_spec_v4(token: env->prog->aux->token);
25155	env->bpf_capable = is_priv = bpf_token_capable(token: env->prog->aux->token, CAP_BPF);
25156
25157	bpf_get_btf_vmlinux();
25158
25159	/* grab the mutex to protect few globals used by verifier */
25160	if (!is_priv)
25161	mutex_lock(&bpf_verifier_lock);
25162
25163	/* user could have requested verbose verifier output
25164	* and supplied buffer to store the verification trace
25165	*/
25166	ret = bpf_vlog_init(log: &env->log, log_level: attr->log_level,
25167	log_buf: (char __user *) (unsigned long) attr->log_buf,
25168	log_size: attr->log_size);
25169	if (ret)
25170	goto err_unlock;
25171
25172	ret = process_fd_array(env, attr, uattr);
25173	if (ret)
25174	goto skip_full_check;
25175
25176	mark_verifier_state_clean(env);
25177
25178	if (IS_ERR(ptr: btf_vmlinux)) {
25179	/* Either gcc or pahole or kernel are broken. */
25180	verbose(private_data: env, fmt: "in-kernel BTF is malformed\n");
25181	ret = PTR_ERR(ptr: btf_vmlinux);
25182	goto skip_full_check;
25183	}
25184
25185	env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
25186	if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
25187	env->strict_alignment = true;
25188	if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
25189	env->strict_alignment = false;
25190
25191	if (is_priv)
25192	env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
25193	env->test_reg_invariants = attr->prog_flags & BPF_F_TEST_REG_INVARIANTS;
25194
25195	env->explored_states = kvcalloc(state_htab_size(env),
25196	sizeof(struct list_head),
25197	GFP_KERNEL_ACCOUNT);
25198	ret = -ENOMEM;
25199	if (!env->explored_states)
25200	goto skip_full_check;
25201
25202	for (i = 0; i < state_htab_size(env); i++)
25203	INIT_LIST_HEAD(list: &env->explored_states[i]);
25204	INIT_LIST_HEAD(list: &env->free_list);
25205
25206	ret = check_btf_info_early(env, attr, uattr);
25207	if (ret < 0)
25208	goto skip_full_check;
25209
25210	ret = add_subprog_and_kfunc(env);
25211	if (ret < 0)
25212	goto skip_full_check;
25213
25214	ret = check_subprogs(env);
25215	if (ret < 0)
25216	goto skip_full_check;
25217
25218	ret = check_btf_info(env, attr, uattr);
25219	if (ret < 0)
25220	goto skip_full_check;
25221
25222	ret = resolve_pseudo_ldimm64(env);
25223	if (ret < 0)
25224	goto skip_full_check;
25225
25226	if (bpf_prog_is_offloaded(aux: env->prog->aux)) {
25227	ret = bpf_prog_offload_verifier_prep(prog: env->prog);
25228	if (ret)
25229	goto skip_full_check;
25230	}
25231
25232	ret = check_cfg(env);
25233	if (ret < 0)
25234	goto skip_full_check;
25235
25236	ret = compute_postorder(env);
25237	if (ret < 0)
25238	goto skip_full_check;
25239
25240	ret = bpf_stack_liveness_init(env);
25241	if (ret)
25242	goto skip_full_check;
25243
25244	ret = check_attach_btf_id(env);
25245	if (ret)
25246	goto skip_full_check;
25247
25248	ret = compute_scc(env);
25249	if (ret < 0)
25250	goto skip_full_check;
25251
25252	ret = compute_live_registers(env);
25253	if (ret < 0)
25254	goto skip_full_check;
25255
25256	ret = mark_fastcall_patterns(env);
25257	if (ret < 0)
25258	goto skip_full_check;
25259
25260	ret = do_check_main(env);
25261	ret = ret ?: do_check_subprogs(env);
25262
25263	if (ret == 0 && bpf_prog_is_offloaded(aux: env->prog->aux))
25264	ret = bpf_prog_offload_finalize(env);
25265
25266	skip_full_check:
25267	kvfree(addr: env->explored_states);
25268
25269	/* might decrease stack depth, keep it before passes that
25270	* allocate additional slots.
25271	*/
25272	if (ret == 0)
25273	ret = remove_fastcall_spills_fills(env);
25274
25275	if (ret == 0)
25276	ret = check_max_stack_depth(env);
25277
25278	/* instruction rewrites happen after this point */
25279	if (ret == 0)
25280	ret = optimize_bpf_loop(env);
25281
25282	if (is_priv) {
25283	if (ret == 0)
25284	opt_hard_wire_dead_code_branches(env);
25285	if (ret == 0)
25286	ret = opt_remove_dead_code(env);
25287	if (ret == 0)
25288	ret = opt_remove_nops(env);
25289	} else {
25290	if (ret == 0)
25291	sanitize_dead_code(env);
25292	}
25293
25294	if (ret == 0)
25295	/* program is valid, convert *(u32*)(ctx + off) accesses */
25296	ret = convert_ctx_accesses(env);
25297
25298	if (ret == 0)
25299	ret = do_misc_fixups(env);
25300
25301	/* do 32-bit optimization after insn patching has done so those patched
25302	* insns could be handled correctly.
25303	*/
25304	if (ret == 0 && !bpf_prog_is_offloaded(aux: env->prog->aux)) {
25305	ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
25306	env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
25307	: false;
25308	}
25309
25310	if (ret == 0)
25311	ret = fixup_call_args(env);
25312
25313	env->verification_time = ktime_get_ns() - start_time;
25314	print_verification_stats(env);
25315	env->prog->aux->verified_insns = env->insn_processed;
25316
25317	/* preserve original error even if log finalization is successful */
25318	err = bpf_vlog_finalize(log: &env->log, log_size_actual: &log_true_size);
25319	if (err)
25320	ret = err;
25321
25322	if (uattr_size >= offsetofend(union bpf_attr, log_true_size) &&
25323	copy_to_bpfptr_offset(dst: uattr, offsetof(union bpf_attr, log_true_size),
25324	src: &log_true_size, size: sizeof(log_true_size))) {
25325	ret = -EFAULT;
25326	goto err_release_maps;
25327	}
25328
25329	if (ret)
25330	goto err_release_maps;
25331
25332	if (env->used_map_cnt) {
25333	/* if program passed verifier, update used_maps in bpf_prog_info */
25334	env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
25335	sizeof(env->used_maps[0]),
25336	GFP_KERNEL_ACCOUNT);
25337
25338	if (!env->prog->aux->used_maps) {
25339	ret = -ENOMEM;
25340	goto err_release_maps;
25341	}
25342
25343	memcpy(env->prog->aux->used_maps, env->used_maps,
25344	sizeof(env->used_maps[0]) * env->used_map_cnt);
25345	env->prog->aux->used_map_cnt = env->used_map_cnt;
25346	}
25347	if (env->used_btf_cnt) {
25348	/* if program passed verifier, update used_btfs in bpf_prog_aux */
25349	env->prog->aux->used_btfs = kmalloc_array(env->used_btf_cnt,
25350	sizeof(env->used_btfs[0]),
25351	GFP_KERNEL_ACCOUNT);
25352	if (!env->prog->aux->used_btfs) {
25353	ret = -ENOMEM;
25354	goto err_release_maps;
25355	}
25356
25357	memcpy(env->prog->aux->used_btfs, env->used_btfs,
25358	sizeof(env->used_btfs[0]) * env->used_btf_cnt);
25359	env->prog->aux->used_btf_cnt = env->used_btf_cnt;
25360	}
25361	if (env->used_map_cnt || env->used_btf_cnt) {
25362	/* program is valid. Convert pseudo bpf_ld_imm64 into generic
25363	* bpf_ld_imm64 instructions
25364	*/
25365	convert_pseudo_ld_imm64(env);
25366	}
25367
25368	adjust_btf_func(env);
25369
25370	err_release_maps:
25371	if (ret)
25372	release_insn_arrays(env);
25373	if (!env->prog->aux->used_maps)
25374	/* if we didn't copy map pointers into bpf_prog_info, release
25375	* them now. Otherwise free_used_maps() will release them.
25376	*/
25377	release_maps(env);
25378	if (!env->prog->aux->used_btfs)
25379	release_btfs(env);
25380
25381	/* extension progs temporarily inherit the attach_type of their targets
25382	for verification purposes, so set it back to zero before returning
25383	*/
25384	if (env->prog->type == BPF_PROG_TYPE_EXT)
25385	env->prog->expected_attach_type = 0;
25386
25387	*prog = env->prog;
25388
25389	module_put(module: env->attach_btf_mod);
25390	err_unlock:
25391	if (!is_priv)
25392	mutex_unlock(lock: &bpf_verifier_lock);
25393	clear_insn_aux_data(env, start: 0, len: env->prog->len);
25394	vfree(addr: env->insn_aux_data);
25395	err_free_env:
25396	bpf_stack_liveness_free(env);
25397	kvfree(addr: env->cfg.insn_postorder);
25398	kvfree(addr: env->scc_info);
25399	kvfree(addr: env->succ);
25400	kvfree(addr: env->gotox_tmp_buf);
25401	kvfree(addr: env);
25402	return ret;
25403	}
25404