1 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
2 * Copyright (c) 2016 Facebook
3 * Copyright (c) 2018 Covalent IO, Inc. http://covalent.io
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of version 2 of the GNU General Public
7 * License as published by the Free Software Foundation.
9 * This program is distributed in the hope that it will be useful, but
10 * WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
12 * General Public License for more details.
14 #include <uapi/linux/btf.h>
15 #include <linux/kernel.h>
16 #include <linux/types.h>
17 #include <linux/slab.h>
18 #include <linux/bpf.h>
19 #include <linux/btf.h>
20 #include <linux/bpf_verifier.h>
21 #include <linux/filter.h>
22 #include <net/netlink.h>
23 #include <linux/file.h>
24 #include <linux/vmalloc.h>
25 #include <linux/stringify.h>
26 #include <linux/bsearch.h>
27 #include <linux/sort.h>
28 #include <linux/perf_event.h>
29 #include <linux/ctype.h>
33 static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
34 #define BPF_PROG_TYPE(_id, _name) \
35 [_id] = & _name ## _verifier_ops,
36 #define BPF_MAP_TYPE(_id, _ops)
37 #include <linux/bpf_types.h>
42 /* bpf_check() is a static code analyzer that walks eBPF program
43 * instruction by instruction and updates register/stack state.
44 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
46 * The first pass is depth-first-search to check that the program is a DAG.
47 * It rejects the following programs:
48 * - larger than BPF_MAXINSNS insns
49 * - if loop is present (detected via back-edge)
50 * - unreachable insns exist (shouldn't be a forest. program = one function)
51 * - out of bounds or malformed jumps
52 * The second pass is all possible path descent from the 1st insn.
53 * Since it's analyzing all pathes through the program, the length of the
54 * analysis is limited to 64k insn, which may be hit even if total number of
55 * insn is less then 4K, but there are too many branches that change stack/regs.
56 * Number of 'branches to be analyzed' is limited to 1k
58 * On entry to each instruction, each register has a type, and the instruction
59 * changes the types of the registers depending on instruction semantics.
60 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
63 * All registers are 64-bit.
64 * R0 - return register
65 * R1-R5 argument passing registers
66 * R6-R9 callee saved registers
67 * R10 - frame pointer read-only
69 * At the start of BPF program the register R1 contains a pointer to bpf_context
70 * and has type PTR_TO_CTX.
72 * Verifier tracks arithmetic operations on pointers in case:
73 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
74 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
75 * 1st insn copies R10 (which has FRAME_PTR) type into R1
76 * and 2nd arithmetic instruction is pattern matched to recognize
77 * that it wants to construct a pointer to some element within stack.
78 * So after 2nd insn, the register R1 has type PTR_TO_STACK
79 * (and -20 constant is saved for further stack bounds checking).
80 * Meaning that this reg is a pointer to stack plus known immediate constant.
82 * Most of the time the registers have SCALAR_VALUE type, which
83 * means the register has some value, but it's not a valid pointer.
84 * (like pointer plus pointer becomes SCALAR_VALUE type)
86 * When verifier sees load or store instructions the type of base register
87 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK, PTR_TO_SOCKET. These are
88 * four pointer types recognized by check_mem_access() function.
90 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
91 * and the range of [ptr, ptr + map's value_size) is accessible.
93 * registers used to pass values to function calls are checked against
94 * function argument constraints.
96 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
97 * It means that the register type passed to this function must be
98 * PTR_TO_STACK and it will be used inside the function as
99 * 'pointer to map element key'
101 * For example the argument constraints for bpf_map_lookup_elem():
102 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
103 * .arg1_type = ARG_CONST_MAP_PTR,
104 * .arg2_type = ARG_PTR_TO_MAP_KEY,
106 * ret_type says that this function returns 'pointer to map elem value or null'
107 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
108 * 2nd argument should be a pointer to stack, which will be used inside
109 * the helper function as a pointer to map element key.
111 * On the kernel side the helper function looks like:
112 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
114 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
115 * void *key = (void *) (unsigned long) r2;
118 * here kernel can access 'key' and 'map' pointers safely, knowing that
119 * [key, key + map->key_size) bytes are valid and were initialized on
120 * the stack of eBPF program.
123 * Corresponding eBPF program may look like:
124 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
125 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
126 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
127 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
128 * here verifier looks at prototype of map_lookup_elem() and sees:
129 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
130 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
132 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
133 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
134 * and were initialized prior to this call.
135 * If it's ok, then verifier allows this BPF_CALL insn and looks at
136 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
137 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
138 * returns ether pointer to map value or NULL.
140 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
141 * insn, the register holding that pointer in the true branch changes state to
142 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
143 * branch. See check_cond_jmp_op().
145 * After the call R0 is set to return type of the function and registers R1-R5
146 * are set to NOT_INIT to indicate that they are no longer readable.
148 * The following reference types represent a potential reference to a kernel
149 * resource which, after first being allocated, must be checked and freed by
151 * - PTR_TO_SOCKET_OR_NULL, PTR_TO_SOCKET
153 * When the verifier sees a helper call return a reference type, it allocates a
154 * pointer id for the reference and stores it in the current function state.
155 * Similar to the way that PTR_TO_MAP_VALUE_OR_NULL is converted into
156 * PTR_TO_MAP_VALUE, PTR_TO_SOCKET_OR_NULL becomes PTR_TO_SOCKET when the type
157 * passes through a NULL-check conditional. For the branch wherein the state is
158 * changed to CONST_IMM, the verifier releases the reference.
160 * For each helper function that allocates a reference, such as
161 * bpf_sk_lookup_tcp(), there is a corresponding release function, such as
162 * bpf_sk_release(). When a reference type passes into the release function,
163 * the verifier also releases the reference. If any unchecked or unreleased
164 * reference remains at the end of the program, the verifier rejects it.
167 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
168 struct bpf_verifier_stack_elem {
169 /* verifer state is 'st'
170 * before processing instruction 'insn_idx'
171 * and after processing instruction 'prev_insn_idx'
173 struct bpf_verifier_state st;
176 struct bpf_verifier_stack_elem *next;
179 #define BPF_COMPLEXITY_LIMIT_INSNS 131072
180 #define BPF_COMPLEXITY_LIMIT_STACK 1024
181 #define BPF_COMPLEXITY_LIMIT_STATES 64
183 #define BPF_MAP_PTR_UNPRIV 1UL
184 #define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \
185 POISON_POINTER_DELTA))
186 #define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
188 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
190 return BPF_MAP_PTR(aux->map_state) == BPF_MAP_PTR_POISON;
193 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
195 return aux->map_state & BPF_MAP_PTR_UNPRIV;
198 static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
199 const struct bpf_map *map, bool unpriv)
201 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON & BPF_MAP_PTR_UNPRIV);
202 unpriv |= bpf_map_ptr_unpriv(aux);
203 aux->map_state = (unsigned long)map |
204 (unpriv ? BPF_MAP_PTR_UNPRIV : 0UL);
207 struct bpf_call_arg_meta {
208 struct bpf_map *map_ptr;
213 s64 msize_smax_value;
214 u64 msize_umax_value;
218 static DEFINE_MUTEX(bpf_verifier_lock);
220 static const struct bpf_line_info *
221 find_linfo(const struct bpf_verifier_env *env, u32 insn_off)
223 const struct bpf_line_info *linfo;
224 const struct bpf_prog *prog;
228 nr_linfo = prog->aux->nr_linfo;
230 if (!nr_linfo || insn_off >= prog->len)
233 linfo = prog->aux->linfo;
234 for (i = 1; i < nr_linfo; i++)
235 if (insn_off < linfo[i].insn_off)
238 return &linfo[i - 1];
241 void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
246 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
248 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
249 "verifier log line truncated - local buffer too short\n");
251 n = min(log->len_total - log->len_used - 1, n);
254 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
260 /* log_level controls verbosity level of eBPF verifier.
261 * bpf_verifier_log_write() is used to dump the verification trace to the log,
262 * so the user can figure out what's wrong with the program
264 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
265 const char *fmt, ...)
269 if (!bpf_verifier_log_needed(&env->log))
273 bpf_verifier_vlog(&env->log, fmt, args);
276 EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
278 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
280 struct bpf_verifier_env *env = private_data;
283 if (!bpf_verifier_log_needed(&env->log))
287 bpf_verifier_vlog(&env->log, fmt, args);
291 static const char *ltrim(const char *s)
299 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env,
301 const char *prefix_fmt, ...)
303 const struct bpf_line_info *linfo;
305 if (!bpf_verifier_log_needed(&env->log))
308 linfo = find_linfo(env, insn_off);
309 if (!linfo || linfo == env->prev_linfo)
315 va_start(args, prefix_fmt);
316 bpf_verifier_vlog(&env->log, prefix_fmt, args);
321 ltrim(btf_name_by_offset(env->prog->aux->btf,
324 env->prev_linfo = linfo;
327 static bool type_is_pkt_pointer(enum bpf_reg_type type)
329 return type == PTR_TO_PACKET ||
330 type == PTR_TO_PACKET_META;
333 static bool reg_type_may_be_null(enum bpf_reg_type type)
335 return type == PTR_TO_MAP_VALUE_OR_NULL ||
336 type == PTR_TO_SOCKET_OR_NULL;
339 static bool type_is_refcounted(enum bpf_reg_type type)
341 return type == PTR_TO_SOCKET;
344 static bool type_is_refcounted_or_null(enum bpf_reg_type type)
346 return type == PTR_TO_SOCKET || type == PTR_TO_SOCKET_OR_NULL;
349 static bool reg_is_refcounted(const struct bpf_reg_state *reg)
351 return type_is_refcounted(reg->type);
354 static bool reg_is_refcounted_or_null(const struct bpf_reg_state *reg)
356 return type_is_refcounted_or_null(reg->type);
359 static bool arg_type_is_refcounted(enum bpf_arg_type type)
361 return type == ARG_PTR_TO_SOCKET;
364 /* Determine whether the function releases some resources allocated by another
365 * function call. The first reference type argument will be assumed to be
366 * released by release_reference().
368 static bool is_release_function(enum bpf_func_id func_id)
370 return func_id == BPF_FUNC_sk_release;
373 /* string representation of 'enum bpf_reg_type' */
374 static const char * const reg_type_str[] = {
376 [SCALAR_VALUE] = "inv",
377 [PTR_TO_CTX] = "ctx",
378 [CONST_PTR_TO_MAP] = "map_ptr",
379 [PTR_TO_MAP_VALUE] = "map_value",
380 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
381 [PTR_TO_STACK] = "fp",
382 [PTR_TO_PACKET] = "pkt",
383 [PTR_TO_PACKET_META] = "pkt_meta",
384 [PTR_TO_PACKET_END] = "pkt_end",
385 [PTR_TO_FLOW_KEYS] = "flow_keys",
386 [PTR_TO_SOCKET] = "sock",
387 [PTR_TO_SOCKET_OR_NULL] = "sock_or_null",
390 static char slot_type_char[] = {
391 [STACK_INVALID] = '?',
397 static void print_liveness(struct bpf_verifier_env *env,
398 enum bpf_reg_liveness live)
400 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
402 if (live & REG_LIVE_READ)
404 if (live & REG_LIVE_WRITTEN)
406 if (live & REG_LIVE_DONE)
410 static struct bpf_func_state *func(struct bpf_verifier_env *env,
411 const struct bpf_reg_state *reg)
413 struct bpf_verifier_state *cur = env->cur_state;
415 return cur->frame[reg->frameno];
418 static void print_verifier_state(struct bpf_verifier_env *env,
419 const struct bpf_func_state *state)
421 const struct bpf_reg_state *reg;
426 verbose(env, " frame%d:", state->frameno);
427 for (i = 0; i < MAX_BPF_REG; i++) {
428 reg = &state->regs[i];
432 verbose(env, " R%d", i);
433 print_liveness(env, reg->live);
434 verbose(env, "=%s", reg_type_str[t]);
435 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
436 tnum_is_const(reg->var_off)) {
437 /* reg->off should be 0 for SCALAR_VALUE */
438 verbose(env, "%lld", reg->var_off.value + reg->off);
439 if (t == PTR_TO_STACK)
440 verbose(env, ",call_%d", func(env, reg)->callsite);
442 verbose(env, "(id=%d", reg->id);
443 if (t != SCALAR_VALUE)
444 verbose(env, ",off=%d", reg->off);
445 if (type_is_pkt_pointer(t))
446 verbose(env, ",r=%d", reg->range);
447 else if (t == CONST_PTR_TO_MAP ||
448 t == PTR_TO_MAP_VALUE ||
449 t == PTR_TO_MAP_VALUE_OR_NULL)
450 verbose(env, ",ks=%d,vs=%d",
451 reg->map_ptr->key_size,
452 reg->map_ptr->value_size);
453 if (tnum_is_const(reg->var_off)) {
454 /* Typically an immediate SCALAR_VALUE, but
455 * could be a pointer whose offset is too big
458 verbose(env, ",imm=%llx", reg->var_off.value);
460 if (reg->smin_value != reg->umin_value &&
461 reg->smin_value != S64_MIN)
462 verbose(env, ",smin_value=%lld",
463 (long long)reg->smin_value);
464 if (reg->smax_value != reg->umax_value &&
465 reg->smax_value != S64_MAX)
466 verbose(env, ",smax_value=%lld",
467 (long long)reg->smax_value);
468 if (reg->umin_value != 0)
469 verbose(env, ",umin_value=%llu",
470 (unsigned long long)reg->umin_value);
471 if (reg->umax_value != U64_MAX)
472 verbose(env, ",umax_value=%llu",
473 (unsigned long long)reg->umax_value);
474 if (!tnum_is_unknown(reg->var_off)) {
477 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
478 verbose(env, ",var_off=%s", tn_buf);
484 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
485 char types_buf[BPF_REG_SIZE + 1];
489 for (j = 0; j < BPF_REG_SIZE; j++) {
490 if (state->stack[i].slot_type[j] != STACK_INVALID)
492 types_buf[j] = slot_type_char[
493 state->stack[i].slot_type[j]];
495 types_buf[BPF_REG_SIZE] = 0;
498 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
499 print_liveness(env, state->stack[i].spilled_ptr.live);
500 if (state->stack[i].slot_type[0] == STACK_SPILL)
502 reg_type_str[state->stack[i].spilled_ptr.type]);
504 verbose(env, "=%s", types_buf);
506 if (state->acquired_refs && state->refs[0].id) {
507 verbose(env, " refs=%d", state->refs[0].id);
508 for (i = 1; i < state->acquired_refs; i++)
509 if (state->refs[i].id)
510 verbose(env, ",%d", state->refs[i].id);
515 #define COPY_STATE_FN(NAME, COUNT, FIELD, SIZE) \
516 static int copy_##NAME##_state(struct bpf_func_state *dst, \
517 const struct bpf_func_state *src) \
521 if (WARN_ON_ONCE(dst->COUNT < src->COUNT)) { \
522 /* internal bug, make state invalid to reject the program */ \
523 memset(dst, 0, sizeof(*dst)); \
526 memcpy(dst->FIELD, src->FIELD, \
527 sizeof(*src->FIELD) * (src->COUNT / SIZE)); \
530 /* copy_reference_state() */
531 COPY_STATE_FN(reference, acquired_refs, refs, 1)
532 /* copy_stack_state() */
533 COPY_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
536 #define REALLOC_STATE_FN(NAME, COUNT, FIELD, SIZE) \
537 static int realloc_##NAME##_state(struct bpf_func_state *state, int size, \
540 u32 old_size = state->COUNT; \
541 struct bpf_##NAME##_state *new_##FIELD; \
542 int slot = size / SIZE; \
544 if (size <= old_size || !size) { \
547 state->COUNT = slot * SIZE; \
548 if (!size && old_size) { \
549 kfree(state->FIELD); \
550 state->FIELD = NULL; \
554 new_##FIELD = kmalloc_array(slot, sizeof(struct bpf_##NAME##_state), \
560 memcpy(new_##FIELD, state->FIELD, \
561 sizeof(*new_##FIELD) * (old_size / SIZE)); \
562 memset(new_##FIELD + old_size / SIZE, 0, \
563 sizeof(*new_##FIELD) * (size - old_size) / SIZE); \
565 state->COUNT = slot * SIZE; \
566 kfree(state->FIELD); \
567 state->FIELD = new_##FIELD; \
570 /* realloc_reference_state() */
571 REALLOC_STATE_FN(reference, acquired_refs, refs, 1)
572 /* realloc_stack_state() */
573 REALLOC_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
574 #undef REALLOC_STATE_FN
576 /* do_check() starts with zero-sized stack in struct bpf_verifier_state to
577 * make it consume minimal amount of memory. check_stack_write() access from
578 * the program calls into realloc_func_state() to grow the stack size.
579 * Note there is a non-zero 'parent' pointer inside bpf_verifier_state
580 * which realloc_stack_state() copies over. It points to previous
581 * bpf_verifier_state which is never reallocated.
583 static int realloc_func_state(struct bpf_func_state *state, int stack_size,
584 int refs_size, bool copy_old)
586 int err = realloc_reference_state(state, refs_size, copy_old);
589 return realloc_stack_state(state, stack_size, copy_old);
592 /* Acquire a pointer id from the env and update the state->refs to include
593 * this new pointer reference.
594 * On success, returns a valid pointer id to associate with the register
595 * On failure, returns a negative errno.
597 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
599 struct bpf_func_state *state = cur_func(env);
600 int new_ofs = state->acquired_refs;
603 err = realloc_reference_state(state, state->acquired_refs + 1, true);
607 state->refs[new_ofs].id = id;
608 state->refs[new_ofs].insn_idx = insn_idx;
613 /* release function corresponding to acquire_reference_state(). Idempotent. */
614 static int __release_reference_state(struct bpf_func_state *state, int ptr_id)
621 last_idx = state->acquired_refs - 1;
622 for (i = 0; i < state->acquired_refs; i++) {
623 if (state->refs[i].id == ptr_id) {
624 if (last_idx && i != last_idx)
625 memcpy(&state->refs[i], &state->refs[last_idx],
626 sizeof(*state->refs));
627 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
628 state->acquired_refs--;
635 /* variation on the above for cases where we expect that there must be an
636 * outstanding reference for the specified ptr_id.
638 static int release_reference_state(struct bpf_verifier_env *env, int ptr_id)
640 struct bpf_func_state *state = cur_func(env);
643 err = __release_reference_state(state, ptr_id);
644 if (WARN_ON_ONCE(err != 0))
645 verbose(env, "verifier internal error: can't release reference\n");
649 static int transfer_reference_state(struct bpf_func_state *dst,
650 struct bpf_func_state *src)
652 int err = realloc_reference_state(dst, src->acquired_refs, false);
655 err = copy_reference_state(dst, src);
661 static void free_func_state(struct bpf_func_state *state)
670 static void free_verifier_state(struct bpf_verifier_state *state,
675 for (i = 0; i <= state->curframe; i++) {
676 free_func_state(state->frame[i]);
677 state->frame[i] = NULL;
683 /* copy verifier state from src to dst growing dst stack space
684 * when necessary to accommodate larger src stack
686 static int copy_func_state(struct bpf_func_state *dst,
687 const struct bpf_func_state *src)
691 err = realloc_func_state(dst, src->allocated_stack, src->acquired_refs,
695 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
696 err = copy_reference_state(dst, src);
699 return copy_stack_state(dst, src);
702 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
703 const struct bpf_verifier_state *src)
705 struct bpf_func_state *dst;
708 /* if dst has more stack frames then src frame, free them */
709 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
710 free_func_state(dst_state->frame[i]);
711 dst_state->frame[i] = NULL;
713 dst_state->curframe = src->curframe;
714 for (i = 0; i <= src->curframe; i++) {
715 dst = dst_state->frame[i];
717 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
720 dst_state->frame[i] = dst;
722 err = copy_func_state(dst, src->frame[i]);
729 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
732 struct bpf_verifier_state *cur = env->cur_state;
733 struct bpf_verifier_stack_elem *elem, *head = env->head;
736 if (env->head == NULL)
740 err = copy_verifier_state(cur, &head->st);
745 *insn_idx = head->insn_idx;
747 *prev_insn_idx = head->prev_insn_idx;
749 free_verifier_state(&head->st, false);
756 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
757 int insn_idx, int prev_insn_idx)
759 struct bpf_verifier_state *cur = env->cur_state;
760 struct bpf_verifier_stack_elem *elem;
763 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
767 elem->insn_idx = insn_idx;
768 elem->prev_insn_idx = prev_insn_idx;
769 elem->next = env->head;
772 err = copy_verifier_state(&elem->st, cur);
775 if (env->stack_size > BPF_COMPLEXITY_LIMIT_STACK) {
776 verbose(env, "BPF program is too complex\n");
781 free_verifier_state(env->cur_state, true);
782 env->cur_state = NULL;
783 /* pop all elements and return */
784 while (!pop_stack(env, NULL, NULL));
788 #define CALLER_SAVED_REGS 6
789 static const int caller_saved[CALLER_SAVED_REGS] = {
790 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
793 static void __mark_reg_not_init(struct bpf_reg_state *reg);
795 /* Mark the unknown part of a register (variable offset or scalar value) as
796 * known to have the value @imm.
798 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
800 /* Clear id, off, and union(map_ptr, range) */
801 memset(((u8 *)reg) + sizeof(reg->type), 0,
802 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
803 reg->var_off = tnum_const(imm);
804 reg->smin_value = (s64)imm;
805 reg->smax_value = (s64)imm;
806 reg->umin_value = imm;
807 reg->umax_value = imm;
810 /* Mark the 'variable offset' part of a register as zero. This should be
811 * used only on registers holding a pointer type.
813 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
815 __mark_reg_known(reg, 0);
818 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
820 __mark_reg_known(reg, 0);
821 reg->type = SCALAR_VALUE;
824 static void mark_reg_known_zero(struct bpf_verifier_env *env,
825 struct bpf_reg_state *regs, u32 regno)
827 if (WARN_ON(regno >= MAX_BPF_REG)) {
828 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
829 /* Something bad happened, let's kill all regs */
830 for (regno = 0; regno < MAX_BPF_REG; regno++)
831 __mark_reg_not_init(regs + regno);
834 __mark_reg_known_zero(regs + regno);
837 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
839 return type_is_pkt_pointer(reg->type);
842 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
844 return reg_is_pkt_pointer(reg) ||
845 reg->type == PTR_TO_PACKET_END;
848 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
849 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
850 enum bpf_reg_type which)
852 /* The register can already have a range from prior markings.
853 * This is fine as long as it hasn't been advanced from its
856 return reg->type == which &&
859 tnum_equals_const(reg->var_off, 0);
862 /* Attempts to improve min/max values based on var_off information */
863 static void __update_reg_bounds(struct bpf_reg_state *reg)
865 /* min signed is max(sign bit) | min(other bits) */
866 reg->smin_value = max_t(s64, reg->smin_value,
867 reg->var_off.value | (reg->var_off.mask & S64_MIN));
868 /* max signed is min(sign bit) | max(other bits) */
869 reg->smax_value = min_t(s64, reg->smax_value,
870 reg->var_off.value | (reg->var_off.mask & S64_MAX));
871 reg->umin_value = max(reg->umin_value, reg->var_off.value);
872 reg->umax_value = min(reg->umax_value,
873 reg->var_off.value | reg->var_off.mask);
876 /* Uses signed min/max values to inform unsigned, and vice-versa */
877 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
879 /* Learn sign from signed bounds.
880 * If we cannot cross the sign boundary, then signed and unsigned bounds
881 * are the same, so combine. This works even in the negative case, e.g.
882 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
884 if (reg->smin_value >= 0 || reg->smax_value < 0) {
885 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
887 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
891 /* Learn sign from unsigned bounds. Signed bounds cross the sign
892 * boundary, so we must be careful.
894 if ((s64)reg->umax_value >= 0) {
895 /* Positive. We can't learn anything from the smin, but smax
896 * is positive, hence safe.
898 reg->smin_value = reg->umin_value;
899 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
901 } else if ((s64)reg->umin_value < 0) {
902 /* Negative. We can't learn anything from the smax, but smin
903 * is negative, hence safe.
905 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
907 reg->smax_value = reg->umax_value;
911 /* Attempts to improve var_off based on unsigned min/max information */
912 static void __reg_bound_offset(struct bpf_reg_state *reg)
914 reg->var_off = tnum_intersect(reg->var_off,
915 tnum_range(reg->umin_value,
919 /* Reset the min/max bounds of a register */
920 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
922 reg->smin_value = S64_MIN;
923 reg->smax_value = S64_MAX;
925 reg->umax_value = U64_MAX;
928 /* Mark a register as having a completely unknown (scalar) value. */
929 static void __mark_reg_unknown(struct bpf_reg_state *reg)
932 * Clear type, id, off, and union(map_ptr, range) and
933 * padding between 'type' and union
935 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
936 reg->type = SCALAR_VALUE;
937 reg->var_off = tnum_unknown;
939 __mark_reg_unbounded(reg);
942 static void mark_reg_unknown(struct bpf_verifier_env *env,
943 struct bpf_reg_state *regs, u32 regno)
945 if (WARN_ON(regno >= MAX_BPF_REG)) {
946 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
947 /* Something bad happened, let's kill all regs except FP */
948 for (regno = 0; regno < BPF_REG_FP; regno++)
949 __mark_reg_not_init(regs + regno);
952 __mark_reg_unknown(regs + regno);
955 static void __mark_reg_not_init(struct bpf_reg_state *reg)
957 __mark_reg_unknown(reg);
958 reg->type = NOT_INIT;
961 static void mark_reg_not_init(struct bpf_verifier_env *env,
962 struct bpf_reg_state *regs, u32 regno)
964 if (WARN_ON(regno >= MAX_BPF_REG)) {
965 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
966 /* Something bad happened, let's kill all regs except FP */
967 for (regno = 0; regno < BPF_REG_FP; regno++)
968 __mark_reg_not_init(regs + regno);
971 __mark_reg_not_init(regs + regno);
974 static void init_reg_state(struct bpf_verifier_env *env,
975 struct bpf_func_state *state)
977 struct bpf_reg_state *regs = state->regs;
980 for (i = 0; i < MAX_BPF_REG; i++) {
981 mark_reg_not_init(env, regs, i);
982 regs[i].live = REG_LIVE_NONE;
983 regs[i].parent = NULL;
987 regs[BPF_REG_FP].type = PTR_TO_STACK;
988 mark_reg_known_zero(env, regs, BPF_REG_FP);
989 regs[BPF_REG_FP].frameno = state->frameno;
991 /* 1st arg to a function */
992 regs[BPF_REG_1].type = PTR_TO_CTX;
993 mark_reg_known_zero(env, regs, BPF_REG_1);
996 #define BPF_MAIN_FUNC (-1)
997 static void init_func_state(struct bpf_verifier_env *env,
998 struct bpf_func_state *state,
999 int callsite, int frameno, int subprogno)
1001 state->callsite = callsite;
1002 state->frameno = frameno;
1003 state->subprogno = subprogno;
1004 init_reg_state(env, state);
1008 SRC_OP, /* register is used as source operand */
1009 DST_OP, /* register is used as destination operand */
1010 DST_OP_NO_MARK /* same as above, check only, don't mark */
1013 static int cmp_subprogs(const void *a, const void *b)
1015 return ((struct bpf_subprog_info *)a)->start -
1016 ((struct bpf_subprog_info *)b)->start;
1019 static int find_subprog(struct bpf_verifier_env *env, int off)
1021 struct bpf_subprog_info *p;
1023 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
1024 sizeof(env->subprog_info[0]), cmp_subprogs);
1027 return p - env->subprog_info;
1031 static int add_subprog(struct bpf_verifier_env *env, int off)
1033 int insn_cnt = env->prog->len;
1036 if (off >= insn_cnt || off < 0) {
1037 verbose(env, "call to invalid destination\n");
1040 ret = find_subprog(env, off);
1043 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
1044 verbose(env, "too many subprograms\n");
1047 env->subprog_info[env->subprog_cnt++].start = off;
1048 sort(env->subprog_info, env->subprog_cnt,
1049 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
1053 static int check_subprogs(struct bpf_verifier_env *env)
1055 int i, ret, subprog_start, subprog_end, off, cur_subprog = 0;
1056 struct bpf_subprog_info *subprog = env->subprog_info;
1057 struct bpf_insn *insn = env->prog->insnsi;
1058 int insn_cnt = env->prog->len;
1060 /* Add entry function. */
1061 ret = add_subprog(env, 0);
1065 /* determine subprog starts. The end is one before the next starts */
1066 for (i = 0; i < insn_cnt; i++) {
1067 if (insn[i].code != (BPF_JMP | BPF_CALL))
1069 if (insn[i].src_reg != BPF_PSEUDO_CALL)
1071 if (!env->allow_ptr_leaks) {
1072 verbose(env, "function calls to other bpf functions are allowed for root only\n");
1075 ret = add_subprog(env, i + insn[i].imm + 1);
1080 /* Add a fake 'exit' subprog which could simplify subprog iteration
1081 * logic. 'subprog_cnt' should not be increased.
1083 subprog[env->subprog_cnt].start = insn_cnt;
1085 if (env->log.level > 1)
1086 for (i = 0; i < env->subprog_cnt; i++)
1087 verbose(env, "func#%d @%d\n", i, subprog[i].start);
1089 /* now check that all jumps are within the same subprog */
1090 subprog_start = subprog[cur_subprog].start;
1091 subprog_end = subprog[cur_subprog + 1].start;
1092 for (i = 0; i < insn_cnt; i++) {
1093 u8 code = insn[i].code;
1095 if (BPF_CLASS(code) != BPF_JMP)
1097 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
1099 off = i + insn[i].off + 1;
1100 if (off < subprog_start || off >= subprog_end) {
1101 verbose(env, "jump out of range from insn %d to %d\n", i, off);
1105 if (i == subprog_end - 1) {
1106 /* to avoid fall-through from one subprog into another
1107 * the last insn of the subprog should be either exit
1108 * or unconditional jump back
1110 if (code != (BPF_JMP | BPF_EXIT) &&
1111 code != (BPF_JMP | BPF_JA)) {
1112 verbose(env, "last insn is not an exit or jmp\n");
1115 subprog_start = subprog_end;
1117 if (cur_subprog < env->subprog_cnt)
1118 subprog_end = subprog[cur_subprog + 1].start;
1124 /* Parentage chain of this register (or stack slot) should take care of all
1125 * issues like callee-saved registers, stack slot allocation time, etc.
1127 static int mark_reg_read(struct bpf_verifier_env *env,
1128 const struct bpf_reg_state *state,
1129 struct bpf_reg_state *parent)
1131 bool writes = parent == state->parent; /* Observe write marks */
1134 /* if read wasn't screened by an earlier write ... */
1135 if (writes && state->live & REG_LIVE_WRITTEN)
1137 if (parent->live & REG_LIVE_DONE) {
1138 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
1139 reg_type_str[parent->type],
1140 parent->var_off.value, parent->off);
1143 /* ... then we depend on parent's value */
1144 parent->live |= REG_LIVE_READ;
1146 parent = state->parent;
1152 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
1153 enum reg_arg_type t)
1155 struct bpf_verifier_state *vstate = env->cur_state;
1156 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1157 struct bpf_reg_state *regs = state->regs;
1159 if (regno >= MAX_BPF_REG) {
1160 verbose(env, "R%d is invalid\n", regno);
1165 /* check whether register used as source operand can be read */
1166 if (regs[regno].type == NOT_INIT) {
1167 verbose(env, "R%d !read_ok\n", regno);
1170 /* We don't need to worry about FP liveness because it's read-only */
1171 if (regno != BPF_REG_FP)
1172 return mark_reg_read(env, ®s[regno],
1173 regs[regno].parent);
1175 /* check whether register used as dest operand can be written to */
1176 if (regno == BPF_REG_FP) {
1177 verbose(env, "frame pointer is read only\n");
1180 regs[regno].live |= REG_LIVE_WRITTEN;
1182 mark_reg_unknown(env, regs, regno);
1187 static bool is_spillable_regtype(enum bpf_reg_type type)
1190 case PTR_TO_MAP_VALUE:
1191 case PTR_TO_MAP_VALUE_OR_NULL:
1195 case PTR_TO_PACKET_META:
1196 case PTR_TO_PACKET_END:
1197 case PTR_TO_FLOW_KEYS:
1198 case CONST_PTR_TO_MAP:
1200 case PTR_TO_SOCKET_OR_NULL:
1207 /* Does this register contain a constant zero? */
1208 static bool register_is_null(struct bpf_reg_state *reg)
1210 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
1213 /* check_stack_read/write functions track spill/fill of registers,
1214 * stack boundary and alignment are checked in check_mem_access()
1216 static int check_stack_write(struct bpf_verifier_env *env,
1217 struct bpf_func_state *state, /* func where register points to */
1218 int off, int size, int value_regno, int insn_idx)
1220 struct bpf_func_state *cur; /* state of the current function */
1221 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
1222 enum bpf_reg_type type;
1224 err = realloc_func_state(state, round_up(slot + 1, BPF_REG_SIZE),
1225 state->acquired_refs, true);
1228 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
1229 * so it's aligned access and [off, off + size) are within stack limits
1231 if (!env->allow_ptr_leaks &&
1232 state->stack[spi].slot_type[0] == STACK_SPILL &&
1233 size != BPF_REG_SIZE) {
1234 verbose(env, "attempt to corrupt spilled pointer on stack\n");
1238 cur = env->cur_state->frame[env->cur_state->curframe];
1239 if (value_regno >= 0 &&
1240 is_spillable_regtype((type = cur->regs[value_regno].type))) {
1242 /* register containing pointer is being spilled into stack */
1243 if (size != BPF_REG_SIZE) {
1244 verbose(env, "invalid size of register spill\n");
1248 if (state != cur && type == PTR_TO_STACK) {
1249 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
1253 /* save register state */
1254 state->stack[spi].spilled_ptr = cur->regs[value_regno];
1255 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1257 for (i = 0; i < BPF_REG_SIZE; i++) {
1258 if (state->stack[spi].slot_type[i] == STACK_MISC &&
1259 !env->allow_ptr_leaks) {
1260 int *poff = &env->insn_aux_data[insn_idx].sanitize_stack_off;
1261 int soff = (-spi - 1) * BPF_REG_SIZE;
1263 /* detected reuse of integer stack slot with a pointer
1264 * which means either llvm is reusing stack slot or
1265 * an attacker is trying to exploit CVE-2018-3639
1266 * (speculative store bypass)
1267 * Have to sanitize that slot with preemptive
1270 if (*poff && *poff != soff) {
1271 /* disallow programs where single insn stores
1272 * into two different stack slots, since verifier
1273 * cannot sanitize them
1276 "insn %d cannot access two stack slots fp%d and fp%d",
1277 insn_idx, *poff, soff);
1282 state->stack[spi].slot_type[i] = STACK_SPILL;
1285 u8 type = STACK_MISC;
1287 /* regular write of data into stack destroys any spilled ptr */
1288 state->stack[spi].spilled_ptr.type = NOT_INIT;
1289 /* Mark slots as STACK_MISC if they belonged to spilled ptr. */
1290 if (state->stack[spi].slot_type[0] == STACK_SPILL)
1291 for (i = 0; i < BPF_REG_SIZE; i++)
1292 state->stack[spi].slot_type[i] = STACK_MISC;
1294 /* only mark the slot as written if all 8 bytes were written
1295 * otherwise read propagation may incorrectly stop too soon
1296 * when stack slots are partially written.
1297 * This heuristic means that read propagation will be
1298 * conservative, since it will add reg_live_read marks
1299 * to stack slots all the way to first state when programs
1300 * writes+reads less than 8 bytes
1302 if (size == BPF_REG_SIZE)
1303 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1305 /* when we zero initialize stack slots mark them as such */
1306 if (value_regno >= 0 &&
1307 register_is_null(&cur->regs[value_regno]))
1310 /* Mark slots affected by this stack write. */
1311 for (i = 0; i < size; i++)
1312 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
1318 static int check_stack_read(struct bpf_verifier_env *env,
1319 struct bpf_func_state *reg_state /* func where register points to */,
1320 int off, int size, int value_regno)
1322 struct bpf_verifier_state *vstate = env->cur_state;
1323 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1324 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
1327 if (reg_state->allocated_stack <= slot) {
1328 verbose(env, "invalid read from stack off %d+0 size %d\n",
1332 stype = reg_state->stack[spi].slot_type;
1334 if (stype[0] == STACK_SPILL) {
1335 if (size != BPF_REG_SIZE) {
1336 verbose(env, "invalid size of register spill\n");
1339 for (i = 1; i < BPF_REG_SIZE; i++) {
1340 if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) {
1341 verbose(env, "corrupted spill memory\n");
1346 if (value_regno >= 0) {
1347 /* restore register state from stack */
1348 state->regs[value_regno] = reg_state->stack[spi].spilled_ptr;
1349 /* mark reg as written since spilled pointer state likely
1350 * has its liveness marks cleared by is_state_visited()
1351 * which resets stack/reg liveness for state transitions
1353 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
1355 mark_reg_read(env, ®_state->stack[spi].spilled_ptr,
1356 reg_state->stack[spi].spilled_ptr.parent);
1361 for (i = 0; i < size; i++) {
1362 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_MISC)
1364 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_ZERO) {
1368 verbose(env, "invalid read from stack off %d+%d size %d\n",
1372 mark_reg_read(env, ®_state->stack[spi].spilled_ptr,
1373 reg_state->stack[spi].spilled_ptr.parent);
1374 if (value_regno >= 0) {
1375 if (zeros == size) {
1376 /* any size read into register is zero extended,
1377 * so the whole register == const_zero
1379 __mark_reg_const_zero(&state->regs[value_regno]);
1381 /* have read misc data from the stack */
1382 mark_reg_unknown(env, state->regs, value_regno);
1384 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
1390 /* check read/write into map element returned by bpf_map_lookup_elem() */
1391 static int __check_map_access(struct bpf_verifier_env *env, u32 regno, int off,
1392 int size, bool zero_size_allowed)
1394 struct bpf_reg_state *regs = cur_regs(env);
1395 struct bpf_map *map = regs[regno].map_ptr;
1397 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
1398 off + size > map->value_size) {
1399 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
1400 map->value_size, off, size);
1406 /* check read/write into a map element with possible variable offset */
1407 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
1408 int off, int size, bool zero_size_allowed)
1410 struct bpf_verifier_state *vstate = env->cur_state;
1411 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1412 struct bpf_reg_state *reg = &state->regs[regno];
1415 /* We may have adjusted the register to this map value, so we
1416 * need to try adding each of min_value and max_value to off
1417 * to make sure our theoretical access will be safe.
1420 print_verifier_state(env, state);
1421 /* The minimum value is only important with signed
1422 * comparisons where we can't assume the floor of a
1423 * value is 0. If we are using signed variables for our
1424 * index'es we need to make sure that whatever we use
1425 * will have a set floor within our range.
1427 if (reg->smin_value < 0) {
1428 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1432 err = __check_map_access(env, regno, reg->smin_value + off, size,
1435 verbose(env, "R%d min value is outside of the array range\n",
1440 /* If we haven't set a max value then we need to bail since we can't be
1441 * sure we won't do bad things.
1442 * If reg->umax_value + off could overflow, treat that as unbounded too.
1444 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
1445 verbose(env, "R%d unbounded memory access, make sure to bounds check any array access into a map\n",
1449 err = __check_map_access(env, regno, reg->umax_value + off, size,
1452 verbose(env, "R%d max value is outside of the array range\n",
1457 #define MAX_PACKET_OFF 0xffff
1459 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
1460 const struct bpf_call_arg_meta *meta,
1461 enum bpf_access_type t)
1463 switch (env->prog->type) {
1464 /* Program types only with direct read access go here! */
1465 case BPF_PROG_TYPE_LWT_IN:
1466 case BPF_PROG_TYPE_LWT_OUT:
1467 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
1468 case BPF_PROG_TYPE_SK_REUSEPORT:
1469 case BPF_PROG_TYPE_FLOW_DISSECTOR:
1470 case BPF_PROG_TYPE_CGROUP_SKB:
1475 /* Program types with direct read + write access go here! */
1476 case BPF_PROG_TYPE_SCHED_CLS:
1477 case BPF_PROG_TYPE_SCHED_ACT:
1478 case BPF_PROG_TYPE_XDP:
1479 case BPF_PROG_TYPE_LWT_XMIT:
1480 case BPF_PROG_TYPE_SK_SKB:
1481 case BPF_PROG_TYPE_SK_MSG:
1483 return meta->pkt_access;
1485 env->seen_direct_write = true;
1492 static int __check_packet_access(struct bpf_verifier_env *env, u32 regno,
1493 int off, int size, bool zero_size_allowed)
1495 struct bpf_reg_state *regs = cur_regs(env);
1496 struct bpf_reg_state *reg = ®s[regno];
1498 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
1499 (u64)off + size > reg->range) {
1500 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
1501 off, size, regno, reg->id, reg->off, reg->range);
1507 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
1508 int size, bool zero_size_allowed)
1510 struct bpf_reg_state *regs = cur_regs(env);
1511 struct bpf_reg_state *reg = ®s[regno];
1514 /* We may have added a variable offset to the packet pointer; but any
1515 * reg->range we have comes after that. We are only checking the fixed
1519 /* We don't allow negative numbers, because we aren't tracking enough
1520 * detail to prove they're safe.
1522 if (reg->smin_value < 0) {
1523 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1527 err = __check_packet_access(env, regno, off, size, zero_size_allowed);
1529 verbose(env, "R%d offset is outside of the packet\n", regno);
1533 /* __check_packet_access has made sure "off + size - 1" is within u16.
1534 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
1535 * otherwise find_good_pkt_pointers would have refused to set range info
1536 * that __check_packet_access would have rejected this pkt access.
1537 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
1539 env->prog->aux->max_pkt_offset =
1540 max_t(u32, env->prog->aux->max_pkt_offset,
1541 off + reg->umax_value + size - 1);
1546 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
1547 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
1548 enum bpf_access_type t, enum bpf_reg_type *reg_type)
1550 struct bpf_insn_access_aux info = {
1551 .reg_type = *reg_type,
1554 if (env->ops->is_valid_access &&
1555 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
1556 /* A non zero info.ctx_field_size indicates that this field is a
1557 * candidate for later verifier transformation to load the whole
1558 * field and then apply a mask when accessed with a narrower
1559 * access than actual ctx access size. A zero info.ctx_field_size
1560 * will only allow for whole field access and rejects any other
1561 * type of narrower access.
1563 *reg_type = info.reg_type;
1565 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
1566 /* remember the offset of last byte accessed in ctx */
1567 if (env->prog->aux->max_ctx_offset < off + size)
1568 env->prog->aux->max_ctx_offset = off + size;
1572 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
1576 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
1579 if (size < 0 || off < 0 ||
1580 (u64)off + size > sizeof(struct bpf_flow_keys)) {
1581 verbose(env, "invalid access to flow keys off=%d size=%d\n",
1588 static int check_sock_access(struct bpf_verifier_env *env, u32 regno, int off,
1589 int size, enum bpf_access_type t)
1591 struct bpf_reg_state *regs = cur_regs(env);
1592 struct bpf_reg_state *reg = ®s[regno];
1593 struct bpf_insn_access_aux info;
1595 if (reg->smin_value < 0) {
1596 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1601 if (!bpf_sock_is_valid_access(off, size, t, &info)) {
1602 verbose(env, "invalid bpf_sock access off=%d size=%d\n",
1610 static bool __is_pointer_value(bool allow_ptr_leaks,
1611 const struct bpf_reg_state *reg)
1613 if (allow_ptr_leaks)
1616 return reg->type != SCALAR_VALUE;
1619 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
1621 return cur_regs(env) + regno;
1624 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
1626 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
1629 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
1631 const struct bpf_reg_state *reg = reg_state(env, regno);
1633 return reg->type == PTR_TO_CTX ||
1634 reg->type == PTR_TO_SOCKET;
1637 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
1639 const struct bpf_reg_state *reg = reg_state(env, regno);
1641 return type_is_pkt_pointer(reg->type);
1644 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
1646 const struct bpf_reg_state *reg = reg_state(env, regno);
1648 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
1649 return reg->type == PTR_TO_FLOW_KEYS;
1652 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
1653 const struct bpf_reg_state *reg,
1654 int off, int size, bool strict)
1656 struct tnum reg_off;
1659 /* Byte size accesses are always allowed. */
1660 if (!strict || size == 1)
1663 /* For platforms that do not have a Kconfig enabling
1664 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
1665 * NET_IP_ALIGN is universally set to '2'. And on platforms
1666 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
1667 * to this code only in strict mode where we want to emulate
1668 * the NET_IP_ALIGN==2 checking. Therefore use an
1669 * unconditional IP align value of '2'.
1673 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
1674 if (!tnum_is_aligned(reg_off, size)) {
1677 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1679 "misaligned packet access off %d+%s+%d+%d size %d\n",
1680 ip_align, tn_buf, reg->off, off, size);
1687 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
1688 const struct bpf_reg_state *reg,
1689 const char *pointer_desc,
1690 int off, int size, bool strict)
1692 struct tnum reg_off;
1694 /* Byte size accesses are always allowed. */
1695 if (!strict || size == 1)
1698 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
1699 if (!tnum_is_aligned(reg_off, size)) {
1702 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1703 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
1704 pointer_desc, tn_buf, reg->off, off, size);
1711 static int check_ptr_alignment(struct bpf_verifier_env *env,
1712 const struct bpf_reg_state *reg, int off,
1713 int size, bool strict_alignment_once)
1715 bool strict = env->strict_alignment || strict_alignment_once;
1716 const char *pointer_desc = "";
1718 switch (reg->type) {
1720 case PTR_TO_PACKET_META:
1721 /* Special case, because of NET_IP_ALIGN. Given metadata sits
1722 * right in front, treat it the very same way.
1724 return check_pkt_ptr_alignment(env, reg, off, size, strict);
1725 case PTR_TO_FLOW_KEYS:
1726 pointer_desc = "flow keys ";
1728 case PTR_TO_MAP_VALUE:
1729 pointer_desc = "value ";
1732 pointer_desc = "context ";
1735 pointer_desc = "stack ";
1736 /* The stack spill tracking logic in check_stack_write()
1737 * and check_stack_read() relies on stack accesses being
1743 pointer_desc = "sock ";
1748 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
1752 static int update_stack_depth(struct bpf_verifier_env *env,
1753 const struct bpf_func_state *func,
1756 u16 stack = env->subprog_info[func->subprogno].stack_depth;
1761 /* update known max for given subprogram */
1762 env->subprog_info[func->subprogno].stack_depth = -off;
1766 /* starting from main bpf function walk all instructions of the function
1767 * and recursively walk all callees that given function can call.
1768 * Ignore jump and exit insns.
1769 * Since recursion is prevented by check_cfg() this algorithm
1770 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
1772 static int check_max_stack_depth(struct bpf_verifier_env *env)
1774 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
1775 struct bpf_subprog_info *subprog = env->subprog_info;
1776 struct bpf_insn *insn = env->prog->insnsi;
1777 int ret_insn[MAX_CALL_FRAMES];
1778 int ret_prog[MAX_CALL_FRAMES];
1781 /* round up to 32-bytes, since this is granularity
1782 * of interpreter stack size
1784 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
1785 if (depth > MAX_BPF_STACK) {
1786 verbose(env, "combined stack size of %d calls is %d. Too large\n",
1791 subprog_end = subprog[idx + 1].start;
1792 for (; i < subprog_end; i++) {
1793 if (insn[i].code != (BPF_JMP | BPF_CALL))
1795 if (insn[i].src_reg != BPF_PSEUDO_CALL)
1797 /* remember insn and function to return to */
1798 ret_insn[frame] = i + 1;
1799 ret_prog[frame] = idx;
1801 /* find the callee */
1802 i = i + insn[i].imm + 1;
1803 idx = find_subprog(env, i);
1805 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
1810 if (frame >= MAX_CALL_FRAMES) {
1811 WARN_ONCE(1, "verifier bug. Call stack is too deep\n");
1816 /* end of for() loop means the last insn of the 'subprog'
1817 * was reached. Doesn't matter whether it was JA or EXIT
1821 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
1823 i = ret_insn[frame];
1824 idx = ret_prog[frame];
1828 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
1829 static int get_callee_stack_depth(struct bpf_verifier_env *env,
1830 const struct bpf_insn *insn, int idx)
1832 int start = idx + insn->imm + 1, subprog;
1834 subprog = find_subprog(env, start);
1836 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
1840 return env->subprog_info[subprog].stack_depth;
1844 static int check_ctx_reg(struct bpf_verifier_env *env,
1845 const struct bpf_reg_state *reg, int regno)
1847 /* Access to ctx or passing it to a helper is only allowed in
1848 * its original, unmodified form.
1852 verbose(env, "dereference of modified ctx ptr R%d off=%d disallowed\n",
1857 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
1860 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1861 verbose(env, "variable ctx access var_off=%s disallowed\n", tn_buf);
1868 /* truncate register to smaller size (in bytes)
1869 * must be called with size < BPF_REG_SIZE
1871 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
1875 /* clear high bits in bit representation */
1876 reg->var_off = tnum_cast(reg->var_off, size);
1878 /* fix arithmetic bounds */
1879 mask = ((u64)1 << (size * 8)) - 1;
1880 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
1881 reg->umin_value &= mask;
1882 reg->umax_value &= mask;
1884 reg->umin_value = 0;
1885 reg->umax_value = mask;
1887 reg->smin_value = reg->umin_value;
1888 reg->smax_value = reg->umax_value;
1891 /* check whether memory at (regno + off) is accessible for t = (read | write)
1892 * if t==write, value_regno is a register which value is stored into memory
1893 * if t==read, value_regno is a register which will receive the value from memory
1894 * if t==write && value_regno==-1, some unknown value is stored into memory
1895 * if t==read && value_regno==-1, don't care what we read from memory
1897 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
1898 int off, int bpf_size, enum bpf_access_type t,
1899 int value_regno, bool strict_alignment_once)
1901 struct bpf_reg_state *regs = cur_regs(env);
1902 struct bpf_reg_state *reg = regs + regno;
1903 struct bpf_func_state *state;
1906 size = bpf_size_to_bytes(bpf_size);
1910 /* alignment checks will add in reg->off themselves */
1911 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
1915 /* for access checks, reg->off is just part of off */
1918 if (reg->type == PTR_TO_MAP_VALUE) {
1919 if (t == BPF_WRITE && value_regno >= 0 &&
1920 is_pointer_value(env, value_regno)) {
1921 verbose(env, "R%d leaks addr into map\n", value_regno);
1925 err = check_map_access(env, regno, off, size, false);
1926 if (!err && t == BPF_READ && value_regno >= 0)
1927 mark_reg_unknown(env, regs, value_regno);
1929 } else if (reg->type == PTR_TO_CTX) {
1930 enum bpf_reg_type reg_type = SCALAR_VALUE;
1932 if (t == BPF_WRITE && value_regno >= 0 &&
1933 is_pointer_value(env, value_regno)) {
1934 verbose(env, "R%d leaks addr into ctx\n", value_regno);
1938 err = check_ctx_reg(env, reg, regno);
1942 err = check_ctx_access(env, insn_idx, off, size, t, ®_type);
1943 if (!err && t == BPF_READ && value_regno >= 0) {
1944 /* ctx access returns either a scalar, or a
1945 * PTR_TO_PACKET[_META,_END]. In the latter
1946 * case, we know the offset is zero.
1948 if (reg_type == SCALAR_VALUE)
1949 mark_reg_unknown(env, regs, value_regno);
1951 mark_reg_known_zero(env, regs,
1953 regs[value_regno].type = reg_type;
1956 } else if (reg->type == PTR_TO_STACK) {
1957 /* stack accesses must be at a fixed offset, so that we can
1958 * determine what type of data were returned.
1959 * See check_stack_read().
1961 if (!tnum_is_const(reg->var_off)) {
1964 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1965 verbose(env, "variable stack access var_off=%s off=%d size=%d",
1969 off += reg->var_off.value;
1970 if (off >= 0 || off < -MAX_BPF_STACK) {
1971 verbose(env, "invalid stack off=%d size=%d\n", off,
1976 state = func(env, reg);
1977 err = update_stack_depth(env, state, off);
1982 err = check_stack_write(env, state, off, size,
1983 value_regno, insn_idx);
1985 err = check_stack_read(env, state, off, size,
1987 } else if (reg_is_pkt_pointer(reg)) {
1988 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
1989 verbose(env, "cannot write into packet\n");
1992 if (t == BPF_WRITE && value_regno >= 0 &&
1993 is_pointer_value(env, value_regno)) {
1994 verbose(env, "R%d leaks addr into packet\n",
1998 err = check_packet_access(env, regno, off, size, false);
1999 if (!err && t == BPF_READ && value_regno >= 0)
2000 mark_reg_unknown(env, regs, value_regno);
2001 } else if (reg->type == PTR_TO_FLOW_KEYS) {
2002 if (t == BPF_WRITE && value_regno >= 0 &&
2003 is_pointer_value(env, value_regno)) {
2004 verbose(env, "R%d leaks addr into flow keys\n",
2009 err = check_flow_keys_access(env, off, size);
2010 if (!err && t == BPF_READ && value_regno >= 0)
2011 mark_reg_unknown(env, regs, value_regno);
2012 } else if (reg->type == PTR_TO_SOCKET) {
2013 if (t == BPF_WRITE) {
2014 verbose(env, "cannot write into socket\n");
2017 err = check_sock_access(env, regno, off, size, t);
2018 if (!err && value_regno >= 0)
2019 mark_reg_unknown(env, regs, value_regno);
2021 verbose(env, "R%d invalid mem access '%s'\n", regno,
2022 reg_type_str[reg->type]);
2026 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
2027 regs[value_regno].type == SCALAR_VALUE) {
2028 /* b/h/w load zero-extends, mark upper bits as known 0 */
2029 coerce_reg_to_size(®s[value_regno], size);
2034 static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
2038 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
2040 verbose(env, "BPF_XADD uses reserved fields\n");
2044 /* check src1 operand */
2045 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2049 /* check src2 operand */
2050 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2054 if (is_pointer_value(env, insn->src_reg)) {
2055 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
2059 if (is_ctx_reg(env, insn->dst_reg) ||
2060 is_pkt_reg(env, insn->dst_reg) ||
2061 is_flow_key_reg(env, insn->dst_reg)) {
2062 verbose(env, "BPF_XADD stores into R%d %s is not allowed\n",
2064 reg_type_str[reg_state(env, insn->dst_reg)->type]);
2068 /* check whether atomic_add can read the memory */
2069 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
2070 BPF_SIZE(insn->code), BPF_READ, -1, true);
2074 /* check whether atomic_add can write into the same memory */
2075 return check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
2076 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
2079 /* when register 'regno' is passed into function that will read 'access_size'
2080 * bytes from that pointer, make sure that it's within stack boundary
2081 * and all elements of stack are initialized.
2082 * Unlike most pointer bounds-checking functions, this one doesn't take an
2083 * 'off' argument, so it has to add in reg->off itself.
2085 static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
2086 int access_size, bool zero_size_allowed,
2087 struct bpf_call_arg_meta *meta)
2089 struct bpf_reg_state *reg = reg_state(env, regno);
2090 struct bpf_func_state *state = func(env, reg);
2091 int off, i, slot, spi;
2093 if (reg->type != PTR_TO_STACK) {
2094 /* Allow zero-byte read from NULL, regardless of pointer type */
2095 if (zero_size_allowed && access_size == 0 &&
2096 register_is_null(reg))
2099 verbose(env, "R%d type=%s expected=%s\n", regno,
2100 reg_type_str[reg->type],
2101 reg_type_str[PTR_TO_STACK]);
2105 /* Only allow fixed-offset stack reads */
2106 if (!tnum_is_const(reg->var_off)) {
2109 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2110 verbose(env, "invalid variable stack read R%d var_off=%s\n",
2114 off = reg->off + reg->var_off.value;
2115 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
2116 access_size < 0 || (access_size == 0 && !zero_size_allowed)) {
2117 verbose(env, "invalid stack type R%d off=%d access_size=%d\n",
2118 regno, off, access_size);
2122 if (meta && meta->raw_mode) {
2123 meta->access_size = access_size;
2124 meta->regno = regno;
2128 for (i = 0; i < access_size; i++) {
2131 slot = -(off + i) - 1;
2132 spi = slot / BPF_REG_SIZE;
2133 if (state->allocated_stack <= slot)
2135 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
2136 if (*stype == STACK_MISC)
2138 if (*stype == STACK_ZERO) {
2139 /* helper can write anything into the stack */
2140 *stype = STACK_MISC;
2144 verbose(env, "invalid indirect read from stack off %d+%d size %d\n",
2145 off, i, access_size);
2148 /* reading any byte out of 8-byte 'spill_slot' will cause
2149 * the whole slot to be marked as 'read'
2151 mark_reg_read(env, &state->stack[spi].spilled_ptr,
2152 state->stack[spi].spilled_ptr.parent);
2154 return update_stack_depth(env, state, off);
2157 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
2158 int access_size, bool zero_size_allowed,
2159 struct bpf_call_arg_meta *meta)
2161 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
2163 switch (reg->type) {
2165 case PTR_TO_PACKET_META:
2166 return check_packet_access(env, regno, reg->off, access_size,
2168 case PTR_TO_MAP_VALUE:
2169 return check_map_access(env, regno, reg->off, access_size,
2171 default: /* scalar_value|ptr_to_stack or invalid ptr */
2172 return check_stack_boundary(env, regno, access_size,
2173 zero_size_allowed, meta);
2177 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
2179 return type == ARG_PTR_TO_MEM ||
2180 type == ARG_PTR_TO_MEM_OR_NULL ||
2181 type == ARG_PTR_TO_UNINIT_MEM;
2184 static bool arg_type_is_mem_size(enum bpf_arg_type type)
2186 return type == ARG_CONST_SIZE ||
2187 type == ARG_CONST_SIZE_OR_ZERO;
2190 static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
2191 enum bpf_arg_type arg_type,
2192 struct bpf_call_arg_meta *meta)
2194 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
2195 enum bpf_reg_type expected_type, type = reg->type;
2198 if (arg_type == ARG_DONTCARE)
2201 err = check_reg_arg(env, regno, SRC_OP);
2205 if (arg_type == ARG_ANYTHING) {
2206 if (is_pointer_value(env, regno)) {
2207 verbose(env, "R%d leaks addr into helper function\n",
2214 if (type_is_pkt_pointer(type) &&
2215 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
2216 verbose(env, "helper access to the packet is not allowed\n");
2220 if (arg_type == ARG_PTR_TO_MAP_KEY ||
2221 arg_type == ARG_PTR_TO_MAP_VALUE ||
2222 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
2223 expected_type = PTR_TO_STACK;
2224 if (!type_is_pkt_pointer(type) && type != PTR_TO_MAP_VALUE &&
2225 type != expected_type)
2227 } else if (arg_type == ARG_CONST_SIZE ||
2228 arg_type == ARG_CONST_SIZE_OR_ZERO) {
2229 expected_type = SCALAR_VALUE;
2230 if (type != expected_type)
2232 } else if (arg_type == ARG_CONST_MAP_PTR) {
2233 expected_type = CONST_PTR_TO_MAP;
2234 if (type != expected_type)
2236 } else if (arg_type == ARG_PTR_TO_CTX) {
2237 expected_type = PTR_TO_CTX;
2238 if (type != expected_type)
2240 err = check_ctx_reg(env, reg, regno);
2243 } else if (arg_type == ARG_PTR_TO_SOCKET) {
2244 expected_type = PTR_TO_SOCKET;
2245 if (type != expected_type)
2247 if (meta->ptr_id || !reg->id) {
2248 verbose(env, "verifier internal error: mismatched references meta=%d, reg=%d\n",
2249 meta->ptr_id, reg->id);
2252 meta->ptr_id = reg->id;
2253 } else if (arg_type_is_mem_ptr(arg_type)) {
2254 expected_type = PTR_TO_STACK;
2255 /* One exception here. In case function allows for NULL to be
2256 * passed in as argument, it's a SCALAR_VALUE type. Final test
2257 * happens during stack boundary checking.
2259 if (register_is_null(reg) &&
2260 arg_type == ARG_PTR_TO_MEM_OR_NULL)
2261 /* final test in check_stack_boundary() */;
2262 else if (!type_is_pkt_pointer(type) &&
2263 type != PTR_TO_MAP_VALUE &&
2264 type != expected_type)
2266 meta->raw_mode = arg_type == ARG_PTR_TO_UNINIT_MEM;
2268 verbose(env, "unsupported arg_type %d\n", arg_type);
2272 if (arg_type == ARG_CONST_MAP_PTR) {
2273 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
2274 meta->map_ptr = reg->map_ptr;
2275 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
2276 /* bpf_map_xxx(..., map_ptr, ..., key) call:
2277 * check that [key, key + map->key_size) are within
2278 * stack limits and initialized
2280 if (!meta->map_ptr) {
2281 /* in function declaration map_ptr must come before
2282 * map_key, so that it's verified and known before
2283 * we have to check map_key here. Otherwise it means
2284 * that kernel subsystem misconfigured verifier
2286 verbose(env, "invalid map_ptr to access map->key\n");
2289 err = check_helper_mem_access(env, regno,
2290 meta->map_ptr->key_size, false,
2292 } else if (arg_type == ARG_PTR_TO_MAP_VALUE ||
2293 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
2294 /* bpf_map_xxx(..., map_ptr, ..., value) call:
2295 * check [value, value + map->value_size) validity
2297 if (!meta->map_ptr) {
2298 /* kernel subsystem misconfigured verifier */
2299 verbose(env, "invalid map_ptr to access map->value\n");
2302 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE);
2303 err = check_helper_mem_access(env, regno,
2304 meta->map_ptr->value_size, false,
2306 } else if (arg_type_is_mem_size(arg_type)) {
2307 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
2309 /* remember the mem_size which may be used later
2310 * to refine return values.
2312 meta->msize_smax_value = reg->smax_value;
2313 meta->msize_umax_value = reg->umax_value;
2315 /* The register is SCALAR_VALUE; the access check
2316 * happens using its boundaries.
2318 if (!tnum_is_const(reg->var_off))
2319 /* For unprivileged variable accesses, disable raw
2320 * mode so that the program is required to
2321 * initialize all the memory that the helper could
2322 * just partially fill up.
2326 if (reg->smin_value < 0) {
2327 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
2332 if (reg->umin_value == 0) {
2333 err = check_helper_mem_access(env, regno - 1, 0,
2340 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
2341 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
2345 err = check_helper_mem_access(env, regno - 1,
2347 zero_size_allowed, meta);
2352 verbose(env, "R%d type=%s expected=%s\n", regno,
2353 reg_type_str[type], reg_type_str[expected_type]);
2357 static int check_map_func_compatibility(struct bpf_verifier_env *env,
2358 struct bpf_map *map, int func_id)
2363 /* We need a two way check, first is from map perspective ... */
2364 switch (map->map_type) {
2365 case BPF_MAP_TYPE_PROG_ARRAY:
2366 if (func_id != BPF_FUNC_tail_call)
2369 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
2370 if (func_id != BPF_FUNC_perf_event_read &&
2371 func_id != BPF_FUNC_perf_event_output &&
2372 func_id != BPF_FUNC_perf_event_read_value)
2375 case BPF_MAP_TYPE_STACK_TRACE:
2376 if (func_id != BPF_FUNC_get_stackid)
2379 case BPF_MAP_TYPE_CGROUP_ARRAY:
2380 if (func_id != BPF_FUNC_skb_under_cgroup &&
2381 func_id != BPF_FUNC_current_task_under_cgroup)
2384 case BPF_MAP_TYPE_CGROUP_STORAGE:
2385 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
2386 if (func_id != BPF_FUNC_get_local_storage)
2389 /* devmap returns a pointer to a live net_device ifindex that we cannot
2390 * allow to be modified from bpf side. So do not allow lookup elements
2393 case BPF_MAP_TYPE_DEVMAP:
2394 if (func_id != BPF_FUNC_redirect_map)
2397 /* Restrict bpf side of cpumap and xskmap, open when use-cases
2400 case BPF_MAP_TYPE_CPUMAP:
2401 case BPF_MAP_TYPE_XSKMAP:
2402 if (func_id != BPF_FUNC_redirect_map)
2405 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
2406 case BPF_MAP_TYPE_HASH_OF_MAPS:
2407 if (func_id != BPF_FUNC_map_lookup_elem)
2410 case BPF_MAP_TYPE_SOCKMAP:
2411 if (func_id != BPF_FUNC_sk_redirect_map &&
2412 func_id != BPF_FUNC_sock_map_update &&
2413 func_id != BPF_FUNC_map_delete_elem &&
2414 func_id != BPF_FUNC_msg_redirect_map)
2417 case BPF_MAP_TYPE_SOCKHASH:
2418 if (func_id != BPF_FUNC_sk_redirect_hash &&
2419 func_id != BPF_FUNC_sock_hash_update &&
2420 func_id != BPF_FUNC_map_delete_elem &&
2421 func_id != BPF_FUNC_msg_redirect_hash)
2424 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
2425 if (func_id != BPF_FUNC_sk_select_reuseport)
2428 case BPF_MAP_TYPE_QUEUE:
2429 case BPF_MAP_TYPE_STACK:
2430 if (func_id != BPF_FUNC_map_peek_elem &&
2431 func_id != BPF_FUNC_map_pop_elem &&
2432 func_id != BPF_FUNC_map_push_elem)
2439 /* ... and second from the function itself. */
2441 case BPF_FUNC_tail_call:
2442 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
2444 if (env->subprog_cnt > 1) {
2445 verbose(env, "tail_calls are not allowed in programs with bpf-to-bpf calls\n");
2449 case BPF_FUNC_perf_event_read:
2450 case BPF_FUNC_perf_event_output:
2451 case BPF_FUNC_perf_event_read_value:
2452 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
2455 case BPF_FUNC_get_stackid:
2456 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
2459 case BPF_FUNC_current_task_under_cgroup:
2460 case BPF_FUNC_skb_under_cgroup:
2461 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
2464 case BPF_FUNC_redirect_map:
2465 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
2466 map->map_type != BPF_MAP_TYPE_CPUMAP &&
2467 map->map_type != BPF_MAP_TYPE_XSKMAP)
2470 case BPF_FUNC_sk_redirect_map:
2471 case BPF_FUNC_msg_redirect_map:
2472 case BPF_FUNC_sock_map_update:
2473 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
2476 case BPF_FUNC_sk_redirect_hash:
2477 case BPF_FUNC_msg_redirect_hash:
2478 case BPF_FUNC_sock_hash_update:
2479 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
2482 case BPF_FUNC_get_local_storage:
2483 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
2484 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
2487 case BPF_FUNC_sk_select_reuseport:
2488 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY)
2491 case BPF_FUNC_map_peek_elem:
2492 case BPF_FUNC_map_pop_elem:
2493 case BPF_FUNC_map_push_elem:
2494 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
2495 map->map_type != BPF_MAP_TYPE_STACK)
2504 verbose(env, "cannot pass map_type %d into func %s#%d\n",
2505 map->map_type, func_id_name(func_id), func_id);
2509 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
2513 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
2515 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
2517 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
2519 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
2521 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
2524 /* We only support one arg being in raw mode at the moment,
2525 * which is sufficient for the helper functions we have
2531 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
2532 enum bpf_arg_type arg_next)
2534 return (arg_type_is_mem_ptr(arg_curr) &&
2535 !arg_type_is_mem_size(arg_next)) ||
2536 (!arg_type_is_mem_ptr(arg_curr) &&
2537 arg_type_is_mem_size(arg_next));
2540 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
2542 /* bpf_xxx(..., buf, len) call will access 'len'
2543 * bytes from memory 'buf'. Both arg types need
2544 * to be paired, so make sure there's no buggy
2545 * helper function specification.
2547 if (arg_type_is_mem_size(fn->arg1_type) ||
2548 arg_type_is_mem_ptr(fn->arg5_type) ||
2549 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
2550 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
2551 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
2552 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
2558 static bool check_refcount_ok(const struct bpf_func_proto *fn)
2562 if (arg_type_is_refcounted(fn->arg1_type))
2564 if (arg_type_is_refcounted(fn->arg2_type))
2566 if (arg_type_is_refcounted(fn->arg3_type))
2568 if (arg_type_is_refcounted(fn->arg4_type))
2570 if (arg_type_is_refcounted(fn->arg5_type))
2573 /* We only support one arg being unreferenced at the moment,
2574 * which is sufficient for the helper functions we have right now.
2579 static int check_func_proto(const struct bpf_func_proto *fn)
2581 return check_raw_mode_ok(fn) &&
2582 check_arg_pair_ok(fn) &&
2583 check_refcount_ok(fn) ? 0 : -EINVAL;
2586 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
2587 * are now invalid, so turn them into unknown SCALAR_VALUE.
2589 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
2590 struct bpf_func_state *state)
2592 struct bpf_reg_state *regs = state->regs, *reg;
2595 for (i = 0; i < MAX_BPF_REG; i++)
2596 if (reg_is_pkt_pointer_any(®s[i]))
2597 mark_reg_unknown(env, regs, i);
2599 bpf_for_each_spilled_reg(i, state, reg) {
2602 if (reg_is_pkt_pointer_any(reg))
2603 __mark_reg_unknown(reg);
2607 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
2609 struct bpf_verifier_state *vstate = env->cur_state;
2612 for (i = 0; i <= vstate->curframe; i++)
2613 __clear_all_pkt_pointers(env, vstate->frame[i]);
2616 static void release_reg_references(struct bpf_verifier_env *env,
2617 struct bpf_func_state *state, int id)
2619 struct bpf_reg_state *regs = state->regs, *reg;
2622 for (i = 0; i < MAX_BPF_REG; i++)
2623 if (regs[i].id == id)
2624 mark_reg_unknown(env, regs, i);
2626 bpf_for_each_spilled_reg(i, state, reg) {
2629 if (reg_is_refcounted(reg) && reg->id == id)
2630 __mark_reg_unknown(reg);
2634 /* The pointer with the specified id has released its reference to kernel
2635 * resources. Identify all copies of the same pointer and clear the reference.
2637 static int release_reference(struct bpf_verifier_env *env,
2638 struct bpf_call_arg_meta *meta)
2640 struct bpf_verifier_state *vstate = env->cur_state;
2643 for (i = 0; i <= vstate->curframe; i++)
2644 release_reg_references(env, vstate->frame[i], meta->ptr_id);
2646 return release_reference_state(env, meta->ptr_id);
2649 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
2652 struct bpf_verifier_state *state = env->cur_state;
2653 struct bpf_func_state *caller, *callee;
2654 int i, err, subprog, target_insn;
2656 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
2657 verbose(env, "the call stack of %d frames is too deep\n",
2658 state->curframe + 2);
2662 target_insn = *insn_idx + insn->imm;
2663 subprog = find_subprog(env, target_insn + 1);
2665 verbose(env, "verifier bug. No program starts at insn %d\n",
2670 caller = state->frame[state->curframe];
2671 if (state->frame[state->curframe + 1]) {
2672 verbose(env, "verifier bug. Frame %d already allocated\n",
2673 state->curframe + 1);
2677 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
2680 state->frame[state->curframe + 1] = callee;
2682 /* callee cannot access r0, r6 - r9 for reading and has to write
2683 * into its own stack before reading from it.
2684 * callee can read/write into caller's stack
2686 init_func_state(env, callee,
2687 /* remember the callsite, it will be used by bpf_exit */
2688 *insn_idx /* callsite */,
2689 state->curframe + 1 /* frameno within this callchain */,
2690 subprog /* subprog number within this prog */);
2692 /* Transfer references to the callee */
2693 err = transfer_reference_state(callee, caller);
2697 /* copy r1 - r5 args that callee can access. The copy includes parent
2698 * pointers, which connects us up to the liveness chain
2700 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
2701 callee->regs[i] = caller->regs[i];
2703 /* after the call registers r0 - r5 were scratched */
2704 for (i = 0; i < CALLER_SAVED_REGS; i++) {
2705 mark_reg_not_init(env, caller->regs, caller_saved[i]);
2706 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
2709 /* only increment it after check_reg_arg() finished */
2712 /* and go analyze first insn of the callee */
2713 *insn_idx = target_insn;
2715 if (env->log.level) {
2716 verbose(env, "caller:\n");
2717 print_verifier_state(env, caller);
2718 verbose(env, "callee:\n");
2719 print_verifier_state(env, callee);
2724 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
2726 struct bpf_verifier_state *state = env->cur_state;
2727 struct bpf_func_state *caller, *callee;
2728 struct bpf_reg_state *r0;
2731 callee = state->frame[state->curframe];
2732 r0 = &callee->regs[BPF_REG_0];
2733 if (r0->type == PTR_TO_STACK) {
2734 /* technically it's ok to return caller's stack pointer
2735 * (or caller's caller's pointer) back to the caller,
2736 * since these pointers are valid. Only current stack
2737 * pointer will be invalid as soon as function exits,
2738 * but let's be conservative
2740 verbose(env, "cannot return stack pointer to the caller\n");
2745 caller = state->frame[state->curframe];
2746 /* return to the caller whatever r0 had in the callee */
2747 caller->regs[BPF_REG_0] = *r0;
2749 /* Transfer references to the caller */
2750 err = transfer_reference_state(caller, callee);
2754 *insn_idx = callee->callsite + 1;
2755 if (env->log.level) {
2756 verbose(env, "returning from callee:\n");
2757 print_verifier_state(env, callee);
2758 verbose(env, "to caller at %d:\n", *insn_idx);
2759 print_verifier_state(env, caller);
2761 /* clear everything in the callee */
2762 free_func_state(callee);
2763 state->frame[state->curframe + 1] = NULL;
2767 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
2769 struct bpf_call_arg_meta *meta)
2771 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
2773 if (ret_type != RET_INTEGER ||
2774 (func_id != BPF_FUNC_get_stack &&
2775 func_id != BPF_FUNC_probe_read_str))
2778 ret_reg->smax_value = meta->msize_smax_value;
2779 ret_reg->umax_value = meta->msize_umax_value;
2780 __reg_deduce_bounds(ret_reg);
2781 __reg_bound_offset(ret_reg);
2785 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
2786 int func_id, int insn_idx)
2788 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
2790 if (func_id != BPF_FUNC_tail_call &&
2791 func_id != BPF_FUNC_map_lookup_elem &&
2792 func_id != BPF_FUNC_map_update_elem &&
2793 func_id != BPF_FUNC_map_delete_elem &&
2794 func_id != BPF_FUNC_map_push_elem &&
2795 func_id != BPF_FUNC_map_pop_elem &&
2796 func_id != BPF_FUNC_map_peek_elem)
2799 if (meta->map_ptr == NULL) {
2800 verbose(env, "kernel subsystem misconfigured verifier\n");
2804 if (!BPF_MAP_PTR(aux->map_state))
2805 bpf_map_ptr_store(aux, meta->map_ptr,
2806 meta->map_ptr->unpriv_array);
2807 else if (BPF_MAP_PTR(aux->map_state) != meta->map_ptr)
2808 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
2809 meta->map_ptr->unpriv_array);
2813 static int check_reference_leak(struct bpf_verifier_env *env)
2815 struct bpf_func_state *state = cur_func(env);
2818 for (i = 0; i < state->acquired_refs; i++) {
2819 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
2820 state->refs[i].id, state->refs[i].insn_idx);
2822 return state->acquired_refs ? -EINVAL : 0;
2825 static int check_helper_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
2827 const struct bpf_func_proto *fn = NULL;
2828 struct bpf_reg_state *regs;
2829 struct bpf_call_arg_meta meta;
2833 /* find function prototype */
2834 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
2835 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
2840 if (env->ops->get_func_proto)
2841 fn = env->ops->get_func_proto(func_id, env->prog);
2843 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
2848 /* eBPF programs must be GPL compatible to use GPL-ed functions */
2849 if (!env->prog->gpl_compatible && fn->gpl_only) {
2850 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
2854 /* With LD_ABS/IND some JITs save/restore skb from r1. */
2855 changes_data = bpf_helper_changes_pkt_data(fn->func);
2856 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
2857 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
2858 func_id_name(func_id), func_id);
2862 memset(&meta, 0, sizeof(meta));
2863 meta.pkt_access = fn->pkt_access;
2865 err = check_func_proto(fn);
2867 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
2868 func_id_name(func_id), func_id);
2873 err = check_func_arg(env, BPF_REG_1, fn->arg1_type, &meta);
2876 err = check_func_arg(env, BPF_REG_2, fn->arg2_type, &meta);
2879 err = check_func_arg(env, BPF_REG_3, fn->arg3_type, &meta);
2882 err = check_func_arg(env, BPF_REG_4, fn->arg4_type, &meta);
2885 err = check_func_arg(env, BPF_REG_5, fn->arg5_type, &meta);
2889 err = record_func_map(env, &meta, func_id, insn_idx);
2893 /* Mark slots with STACK_MISC in case of raw mode, stack offset
2894 * is inferred from register state.
2896 for (i = 0; i < meta.access_size; i++) {
2897 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
2898 BPF_WRITE, -1, false);
2903 if (func_id == BPF_FUNC_tail_call) {
2904 err = check_reference_leak(env);
2906 verbose(env, "tail_call would lead to reference leak\n");
2909 } else if (is_release_function(func_id)) {
2910 err = release_reference(env, &meta);
2915 regs = cur_regs(env);
2917 /* check that flags argument in get_local_storage(map, flags) is 0,
2918 * this is required because get_local_storage() can't return an error.
2920 if (func_id == BPF_FUNC_get_local_storage &&
2921 !register_is_null(®s[BPF_REG_2])) {
2922 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
2926 /* reset caller saved regs */
2927 for (i = 0; i < CALLER_SAVED_REGS; i++) {
2928 mark_reg_not_init(env, regs, caller_saved[i]);
2929 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
2932 /* update return register (already marked as written above) */
2933 if (fn->ret_type == RET_INTEGER) {
2934 /* sets type to SCALAR_VALUE */
2935 mark_reg_unknown(env, regs, BPF_REG_0);
2936 } else if (fn->ret_type == RET_VOID) {
2937 regs[BPF_REG_0].type = NOT_INIT;
2938 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL ||
2939 fn->ret_type == RET_PTR_TO_MAP_VALUE) {
2940 /* There is no offset yet applied, variable or fixed */
2941 mark_reg_known_zero(env, regs, BPF_REG_0);
2942 /* remember map_ptr, so that check_map_access()
2943 * can check 'value_size' boundary of memory access
2944 * to map element returned from bpf_map_lookup_elem()
2946 if (meta.map_ptr == NULL) {
2948 "kernel subsystem misconfigured verifier\n");
2951 regs[BPF_REG_0].map_ptr = meta.map_ptr;
2952 if (fn->ret_type == RET_PTR_TO_MAP_VALUE) {
2953 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE;
2955 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
2956 regs[BPF_REG_0].id = ++env->id_gen;
2958 } else if (fn->ret_type == RET_PTR_TO_SOCKET_OR_NULL) {
2959 int id = acquire_reference_state(env, insn_idx);
2962 mark_reg_known_zero(env, regs, BPF_REG_0);
2963 regs[BPF_REG_0].type = PTR_TO_SOCKET_OR_NULL;
2964 regs[BPF_REG_0].id = id;
2966 verbose(env, "unknown return type %d of func %s#%d\n",
2967 fn->ret_type, func_id_name(func_id), func_id);
2971 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
2973 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
2977 if (func_id == BPF_FUNC_get_stack && !env->prog->has_callchain_buf) {
2978 const char *err_str;
2980 #ifdef CONFIG_PERF_EVENTS
2981 err = get_callchain_buffers(sysctl_perf_event_max_stack);
2982 err_str = "cannot get callchain buffer for func %s#%d\n";
2985 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
2988 verbose(env, err_str, func_id_name(func_id), func_id);
2992 env->prog->has_callchain_buf = true;
2996 clear_all_pkt_pointers(env);
3000 static bool signed_add_overflows(s64 a, s64 b)
3002 /* Do the add in u64, where overflow is well-defined */
3003 s64 res = (s64)((u64)a + (u64)b);
3010 static bool signed_sub_overflows(s64 a, s64 b)
3012 /* Do the sub in u64, where overflow is well-defined */
3013 s64 res = (s64)((u64)a - (u64)b);
3020 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
3021 const struct bpf_reg_state *reg,
3022 enum bpf_reg_type type)
3024 bool known = tnum_is_const(reg->var_off);
3025 s64 val = reg->var_off.value;
3026 s64 smin = reg->smin_value;
3028 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
3029 verbose(env, "math between %s pointer and %lld is not allowed\n",
3030 reg_type_str[type], val);
3034 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
3035 verbose(env, "%s pointer offset %d is not allowed\n",
3036 reg_type_str[type], reg->off);
3040 if (smin == S64_MIN) {
3041 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
3042 reg_type_str[type]);
3046 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
3047 verbose(env, "value %lld makes %s pointer be out of bounds\n",
3048 smin, reg_type_str[type]);
3055 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
3056 * Caller should also handle BPF_MOV case separately.
3057 * If we return -EACCES, caller may want to try again treating pointer as a
3058 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
3060 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
3061 struct bpf_insn *insn,
3062 const struct bpf_reg_state *ptr_reg,
3063 const struct bpf_reg_state *off_reg)
3065 struct bpf_verifier_state *vstate = env->cur_state;
3066 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3067 struct bpf_reg_state *regs = state->regs, *dst_reg;
3068 bool known = tnum_is_const(off_reg->var_off);
3069 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
3070 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
3071 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
3072 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
3073 u8 opcode = BPF_OP(insn->code);
3074 u32 dst = insn->dst_reg;
3076 dst_reg = ®s[dst];
3078 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
3079 smin_val > smax_val || umin_val > umax_val) {
3080 /* Taint dst register if offset had invalid bounds derived from
3081 * e.g. dead branches.
3083 __mark_reg_unknown(dst_reg);
3087 if (BPF_CLASS(insn->code) != BPF_ALU64) {
3088 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
3090 "R%d 32-bit pointer arithmetic prohibited\n",
3095 switch (ptr_reg->type) {
3096 case PTR_TO_MAP_VALUE_OR_NULL:
3097 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
3098 dst, reg_type_str[ptr_reg->type]);
3100 case CONST_PTR_TO_MAP:
3101 case PTR_TO_PACKET_END:
3103 case PTR_TO_SOCKET_OR_NULL:
3104 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
3105 dst, reg_type_str[ptr_reg->type]);
3111 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
3112 * The id may be overwritten later if we create a new variable offset.
3114 dst_reg->type = ptr_reg->type;
3115 dst_reg->id = ptr_reg->id;
3117 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
3118 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
3123 /* We can take a fixed offset as long as it doesn't overflow
3124 * the s32 'off' field
3126 if (known && (ptr_reg->off + smin_val ==
3127 (s64)(s32)(ptr_reg->off + smin_val))) {
3128 /* pointer += K. Accumulate it into fixed offset */
3129 dst_reg->smin_value = smin_ptr;
3130 dst_reg->smax_value = smax_ptr;
3131 dst_reg->umin_value = umin_ptr;
3132 dst_reg->umax_value = umax_ptr;
3133 dst_reg->var_off = ptr_reg->var_off;
3134 dst_reg->off = ptr_reg->off + smin_val;
3135 dst_reg->raw = ptr_reg->raw;
3138 /* A new variable offset is created. Note that off_reg->off
3139 * == 0, since it's a scalar.
3140 * dst_reg gets the pointer type and since some positive
3141 * integer value was added to the pointer, give it a new 'id'
3142 * if it's a PTR_TO_PACKET.
3143 * this creates a new 'base' pointer, off_reg (variable) gets
3144 * added into the variable offset, and we copy the fixed offset
3147 if (signed_add_overflows(smin_ptr, smin_val) ||
3148 signed_add_overflows(smax_ptr, smax_val)) {
3149 dst_reg->smin_value = S64_MIN;
3150 dst_reg->smax_value = S64_MAX;
3152 dst_reg->smin_value = smin_ptr + smin_val;
3153 dst_reg->smax_value = smax_ptr + smax_val;
3155 if (umin_ptr + umin_val < umin_ptr ||
3156 umax_ptr + umax_val < umax_ptr) {
3157 dst_reg->umin_value = 0;
3158 dst_reg->umax_value = U64_MAX;
3160 dst_reg->umin_value = umin_ptr + umin_val;
3161 dst_reg->umax_value = umax_ptr + umax_val;
3163 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
3164 dst_reg->off = ptr_reg->off;
3165 dst_reg->raw = ptr_reg->raw;
3166 if (reg_is_pkt_pointer(ptr_reg)) {
3167 dst_reg->id = ++env->id_gen;
3168 /* something was added to pkt_ptr, set range to zero */
3173 if (dst_reg == off_reg) {
3174 /* scalar -= pointer. Creates an unknown scalar */
3175 verbose(env, "R%d tried to subtract pointer from scalar\n",
3179 /* We don't allow subtraction from FP, because (according to
3180 * test_verifier.c test "invalid fp arithmetic", JITs might not
3181 * be able to deal with it.
3183 if (ptr_reg->type == PTR_TO_STACK) {
3184 verbose(env, "R%d subtraction from stack pointer prohibited\n",
3188 if (known && (ptr_reg->off - smin_val ==
3189 (s64)(s32)(ptr_reg->off - smin_val))) {
3190 /* pointer -= K. Subtract it from fixed offset */
3191 dst_reg->smin_value = smin_ptr;
3192 dst_reg->smax_value = smax_ptr;
3193 dst_reg->umin_value = umin_ptr;
3194 dst_reg->umax_value = umax_ptr;
3195 dst_reg->var_off = ptr_reg->var_off;
3196 dst_reg->id = ptr_reg->id;
3197 dst_reg->off = ptr_reg->off - smin_val;
3198 dst_reg->raw = ptr_reg->raw;
3201 /* A new variable offset is created. If the subtrahend is known
3202 * nonnegative, then any reg->range we had before is still good.
3204 if (signed_sub_overflows(smin_ptr, smax_val) ||
3205 signed_sub_overflows(smax_ptr, smin_val)) {
3206 /* Overflow possible, we know nothing */
3207 dst_reg->smin_value = S64_MIN;
3208 dst_reg->smax_value = S64_MAX;
3210 dst_reg->smin_value = smin_ptr - smax_val;
3211 dst_reg->smax_value = smax_ptr - smin_val;
3213 if (umin_ptr < umax_val) {
3214 /* Overflow possible, we know nothing */
3215 dst_reg->umin_value = 0;
3216 dst_reg->umax_value = U64_MAX;
3218 /* Cannot overflow (as long as bounds are consistent) */
3219 dst_reg->umin_value = umin_ptr - umax_val;
3220 dst_reg->umax_value = umax_ptr - umin_val;
3222 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
3223 dst_reg->off = ptr_reg->off;
3224 dst_reg->raw = ptr_reg->raw;
3225 if (reg_is_pkt_pointer(ptr_reg)) {
3226 dst_reg->id = ++env->id_gen;
3227 /* something was added to pkt_ptr, set range to zero */
3235 /* bitwise ops on pointers are troublesome, prohibit. */
3236 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
3237 dst, bpf_alu_string[opcode >> 4]);
3240 /* other operators (e.g. MUL,LSH) produce non-pointer results */
3241 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
3242 dst, bpf_alu_string[opcode >> 4]);
3246 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
3249 __update_reg_bounds(dst_reg);
3250 __reg_deduce_bounds(dst_reg);
3251 __reg_bound_offset(dst_reg);
3253 /* For unprivileged we require that resulting offset must be in bounds
3254 * in order to be able to sanitize access later on.
3256 if (!env->allow_ptr_leaks && dst_reg->type == PTR_TO_MAP_VALUE &&
3257 check_map_access(env, dst, dst_reg->off, 1, false)) {
3258 verbose(env, "R%d pointer arithmetic of map value goes out of range, prohibited for !root\n",
3266 /* WARNING: This function does calculations on 64-bit values, but the actual
3267 * execution may occur on 32-bit values. Therefore, things like bitshifts
3268 * need extra checks in the 32-bit case.
3270 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
3271 struct bpf_insn *insn,
3272 struct bpf_reg_state *dst_reg,
3273 struct bpf_reg_state src_reg)
3275 struct bpf_reg_state *regs = cur_regs(env);
3276 u8 opcode = BPF_OP(insn->code);
3277 bool src_known, dst_known;
3278 s64 smin_val, smax_val;
3279 u64 umin_val, umax_val;
3280 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
3282 if (insn_bitness == 32) {
3283 /* Relevant for 32-bit RSH: Information can propagate towards
3284 * LSB, so it isn't sufficient to only truncate the output to
3287 coerce_reg_to_size(dst_reg, 4);
3288 coerce_reg_to_size(&src_reg, 4);
3291 smin_val = src_reg.smin_value;
3292 smax_val = src_reg.smax_value;
3293 umin_val = src_reg.umin_value;
3294 umax_val = src_reg.umax_value;
3295 src_known = tnum_is_const(src_reg.var_off);
3296 dst_known = tnum_is_const(dst_reg->var_off);
3298 if ((src_known && (smin_val != smax_val || umin_val != umax_val)) ||
3299 smin_val > smax_val || umin_val > umax_val) {
3300 /* Taint dst register if offset had invalid bounds derived from
3301 * e.g. dead branches.
3303 __mark_reg_unknown(dst_reg);
3308 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
3309 __mark_reg_unknown(dst_reg);
3315 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
3316 signed_add_overflows(dst_reg->smax_value, smax_val)) {
3317 dst_reg->smin_value = S64_MIN;
3318 dst_reg->smax_value = S64_MAX;
3320 dst_reg->smin_value += smin_val;
3321 dst_reg->smax_value += smax_val;
3323 if (dst_reg->umin_value + umin_val < umin_val ||
3324 dst_reg->umax_value + umax_val < umax_val) {
3325 dst_reg->umin_value = 0;
3326 dst_reg->umax_value = U64_MAX;
3328 dst_reg->umin_value += umin_val;
3329 dst_reg->umax_value += umax_val;
3331 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
3334 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
3335 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
3336 /* Overflow possible, we know nothing */
3337 dst_reg->smin_value = S64_MIN;
3338 dst_reg->smax_value = S64_MAX;
3340 dst_reg->smin_value -= smax_val;
3341 dst_reg->smax_value -= smin_val;
3343 if (dst_reg->umin_value < umax_val) {
3344 /* Overflow possible, we know nothing */
3345 dst_reg->umin_value = 0;
3346 dst_reg->umax_value = U64_MAX;
3348 /* Cannot overflow (as long as bounds are consistent) */
3349 dst_reg->umin_value -= umax_val;
3350 dst_reg->umax_value -= umin_val;
3352 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
3355 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
3356 if (smin_val < 0 || dst_reg->smin_value < 0) {
3357 /* Ain't nobody got time to multiply that sign */
3358 __mark_reg_unbounded(dst_reg);
3359 __update_reg_bounds(dst_reg);
3362 /* Both values are positive, so we can work with unsigned and
3363 * copy the result to signed (unless it exceeds S64_MAX).
3365 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
3366 /* Potential overflow, we know nothing */
3367 __mark_reg_unbounded(dst_reg);
3368 /* (except what we can learn from the var_off) */
3369 __update_reg_bounds(dst_reg);
3372 dst_reg->umin_value *= umin_val;
3373 dst_reg->umax_value *= umax_val;
3374 if (dst_reg->umax_value > S64_MAX) {
3375 /* Overflow possible, we know nothing */
3376 dst_reg->smin_value = S64_MIN;
3377 dst_reg->smax_value = S64_MAX;
3379 dst_reg->smin_value = dst_reg->umin_value;
3380 dst_reg->smax_value = dst_reg->umax_value;
3384 if (src_known && dst_known) {
3385 __mark_reg_known(dst_reg, dst_reg->var_off.value &
3386 src_reg.var_off.value);
3389 /* We get our minimum from the var_off, since that's inherently
3390 * bitwise. Our maximum is the minimum of the operands' maxima.
3392 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
3393 dst_reg->umin_value = dst_reg->var_off.value;
3394 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
3395 if (dst_reg->smin_value < 0 || smin_val < 0) {
3396 /* Lose signed bounds when ANDing negative numbers,
3397 * ain't nobody got time for that.
3399 dst_reg->smin_value = S64_MIN;
3400 dst_reg->smax_value = S64_MAX;
3402 /* ANDing two positives gives a positive, so safe to
3403 * cast result into s64.
3405 dst_reg->smin_value = dst_reg->umin_value;
3406 dst_reg->smax_value = dst_reg->umax_value;
3408 /* We may learn something more from the var_off */
3409 __update_reg_bounds(dst_reg);
3412 if (src_known && dst_known) {
3413 __mark_reg_known(dst_reg, dst_reg->var_off.value |
3414 src_reg.var_off.value);
3417 /* We get our maximum from the var_off, and our minimum is the
3418 * maximum of the operands' minima
3420 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
3421 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
3422 dst_reg->umax_value = dst_reg->var_off.value |
3423 dst_reg->var_off.mask;
3424 if (dst_reg->smin_value < 0 || smin_val < 0) {
3425 /* Lose signed bounds when ORing negative numbers,
3426 * ain't nobody got time for that.
3428 dst_reg->smin_value = S64_MIN;
3429 dst_reg->smax_value = S64_MAX;
3431 /* ORing two positives gives a positive, so safe to
3432 * cast result into s64.
3434 dst_reg->smin_value = dst_reg->umin_value;
3435 dst_reg->smax_value = dst_reg->umax_value;
3437 /* We may learn something more from the var_off */
3438 __update_reg_bounds(dst_reg);
3441 if (umax_val >= insn_bitness) {
3442 /* Shifts greater than 31 or 63 are undefined.
3443 * This includes shifts by a negative number.
3445 mark_reg_unknown(env, regs, insn->dst_reg);
3448 /* We lose all sign bit information (except what we can pick
3451 dst_reg->smin_value = S64_MIN;
3452 dst_reg->smax_value = S64_MAX;
3453 /* If we might shift our top bit out, then we know nothing */
3454 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
3455 dst_reg->umin_value = 0;
3456 dst_reg->umax_value = U64_MAX;
3458 dst_reg->umin_value <<= umin_val;
3459 dst_reg->umax_value <<= umax_val;
3461 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
3462 /* We may learn something more from the var_off */
3463 __update_reg_bounds(dst_reg);
3466 if (umax_val >= insn_bitness) {
3467 /* Shifts greater than 31 or 63 are undefined.
3468 * This includes shifts by a negative number.
3470 mark_reg_unknown(env, regs, insn->dst_reg);
3473 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
3474 * be negative, then either:
3475 * 1) src_reg might be zero, so the sign bit of the result is
3476 * unknown, so we lose our signed bounds
3477 * 2) it's known negative, thus the unsigned bounds capture the
3479 * 3) the signed bounds cross zero, so they tell us nothing
3481 * If the value in dst_reg is known nonnegative, then again the
3482 * unsigned bounts capture the signed bounds.
3483 * Thus, in all cases it suffices to blow away our signed bounds
3484 * and rely on inferring new ones from the unsigned bounds and
3485 * var_off of the result.
3487 dst_reg->smin_value = S64_MIN;
3488 dst_reg->smax_value = S64_MAX;
3489 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
3490 dst_reg->umin_value >>= umax_val;
3491 dst_reg->umax_value >>= umin_val;
3492 /* We may learn something more from the var_off */
3493 __update_reg_bounds(dst_reg);
3496 if (umax_val >= insn_bitness) {
3497 /* Shifts greater than 31 or 63 are undefined.
3498 * This includes shifts by a negative number.
3500 mark_reg_unknown(env, regs, insn->dst_reg);
3504 /* Upon reaching here, src_known is true and
3505 * umax_val is equal to umin_val.
3507 dst_reg->smin_value >>= umin_val;
3508 dst_reg->smax_value >>= umin_val;
3509 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val);
3511 /* blow away the dst_reg umin_value/umax_value and rely on
3512 * dst_reg var_off to refine the result.
3514 dst_reg->umin_value = 0;
3515 dst_reg->umax_value = U64_MAX;
3516 __update_reg_bounds(dst_reg);
3519 mark_reg_unknown(env, regs, insn->dst_reg);
3523 if (BPF_CLASS(insn->code) != BPF_ALU64) {
3524 /* 32-bit ALU ops are (32,32)->32 */
3525 coerce_reg_to_size(dst_reg, 4);
3528 __reg_deduce_bounds(dst_reg);
3529 __reg_bound_offset(dst_reg);
3533 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
3536 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
3537 struct bpf_insn *insn)
3539 struct bpf_verifier_state *vstate = env->cur_state;
3540 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3541 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
3542 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
3543 u8 opcode = BPF_OP(insn->code);
3545 dst_reg = ®s[insn->dst_reg];
3547 if (dst_reg->type != SCALAR_VALUE)
3549 if (BPF_SRC(insn->code) == BPF_X) {
3550 src_reg = ®s[insn->src_reg];
3551 if (src_reg->type != SCALAR_VALUE) {
3552 if (dst_reg->type != SCALAR_VALUE) {
3553 /* Combining two pointers by any ALU op yields
3554 * an arbitrary scalar. Disallow all math except
3555 * pointer subtraction
3557 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
3558 mark_reg_unknown(env, regs, insn->dst_reg);
3561 verbose(env, "R%d pointer %s pointer prohibited\n",
3563 bpf_alu_string[opcode >> 4]);
3566 /* scalar += pointer
3567 * This is legal, but we have to reverse our
3568 * src/dest handling in computing the range
3570 return adjust_ptr_min_max_vals(env, insn,
3573 } else if (ptr_reg) {
3574 /* pointer += scalar */
3575 return adjust_ptr_min_max_vals(env, insn,
3579 /* Pretend the src is a reg with a known value, since we only
3580 * need to be able to read from this state.
3582 off_reg.type = SCALAR_VALUE;
3583 __mark_reg_known(&off_reg, insn->imm);
3585 if (ptr_reg) /* pointer += K */
3586 return adjust_ptr_min_max_vals(env, insn,
3590 /* Got here implies adding two SCALAR_VALUEs */
3591 if (WARN_ON_ONCE(ptr_reg)) {
3592 print_verifier_state(env, state);
3593 verbose(env, "verifier internal error: unexpected ptr_reg\n");
3596 if (WARN_ON(!src_reg)) {
3597 print_verifier_state(env, state);
3598 verbose(env, "verifier internal error: no src_reg\n");
3601 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
3604 /* check validity of 32-bit and 64-bit arithmetic operations */
3605 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
3607 struct bpf_reg_state *regs = cur_regs(env);
3608 u8 opcode = BPF_OP(insn->code);
3611 if (opcode == BPF_END || opcode == BPF_NEG) {
3612 if (opcode == BPF_NEG) {
3613 if (BPF_SRC(insn->code) != 0 ||
3614 insn->src_reg != BPF_REG_0 ||
3615 insn->off != 0 || insn->imm != 0) {
3616 verbose(env, "BPF_NEG uses reserved fields\n");
3620 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
3621 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
3622 BPF_CLASS(insn->code) == BPF_ALU64) {
3623 verbose(env, "BPF_END uses reserved fields\n");
3628 /* check src operand */
3629 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3633 if (is_pointer_value(env, insn->dst_reg)) {
3634 verbose(env, "R%d pointer arithmetic prohibited\n",
3639 /* check dest operand */
3640 err = check_reg_arg(env, insn->dst_reg, DST_OP);
3644 } else if (opcode == BPF_MOV) {
3646 if (BPF_SRC(insn->code) == BPF_X) {
3647 if (insn->imm != 0 || insn->off != 0) {
3648 verbose(env, "BPF_MOV uses reserved fields\n");
3652 /* check src operand */
3653 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3657 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
3658 verbose(env, "BPF_MOV uses reserved fields\n");
3663 /* check dest operand, mark as required later */
3664 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
3668 if (BPF_SRC(insn->code) == BPF_X) {
3669 struct bpf_reg_state *src_reg = regs + insn->src_reg;
3670 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
3672 if (BPF_CLASS(insn->code) == BPF_ALU64) {
3674 * copy register state to dest reg
3676 *dst_reg = *src_reg;
3677 dst_reg->live |= REG_LIVE_WRITTEN;
3680 if (is_pointer_value(env, insn->src_reg)) {
3682 "R%d partial copy of pointer\n",
3685 } else if (src_reg->type == SCALAR_VALUE) {
3686 *dst_reg = *src_reg;
3687 dst_reg->live |= REG_LIVE_WRITTEN;
3689 mark_reg_unknown(env, regs,
3692 coerce_reg_to_size(dst_reg, 4);
3696 * remember the value we stored into this reg
3698 /* clear any state __mark_reg_known doesn't set */
3699 mark_reg_unknown(env, regs, insn->dst_reg);
3700 regs[insn->dst_reg].type = SCALAR_VALUE;
3701 if (BPF_CLASS(insn->code) == BPF_ALU64) {
3702 __mark_reg_known(regs + insn->dst_reg,
3705 __mark_reg_known(regs + insn->dst_reg,
3710 } else if (opcode > BPF_END) {
3711 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
3714 } else { /* all other ALU ops: and, sub, xor, add, ... */
3716 if (BPF_SRC(insn->code) == BPF_X) {
3717 if (insn->imm != 0 || insn->off != 0) {
3718 verbose(env, "BPF_ALU uses reserved fields\n");
3721 /* check src1 operand */
3722 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3726 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
3727 verbose(env, "BPF_ALU uses reserved fields\n");
3732 /* check src2 operand */
3733 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3737 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
3738 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
3739 verbose(env, "div by zero\n");
3743 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
3744 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
3745 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
3747 if (insn->imm < 0 || insn->imm >= size) {
3748 verbose(env, "invalid shift %d\n", insn->imm);
3753 /* check dest operand */
3754 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
3758 return adjust_reg_min_max_vals(env, insn);
3764 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
3765 struct bpf_reg_state *dst_reg,
3766 enum bpf_reg_type type,
3767 bool range_right_open)
3769 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3770 struct bpf_reg_state *regs = state->regs, *reg;
3774 if (dst_reg->off < 0 ||
3775 (dst_reg->off == 0 && range_right_open))
3776 /* This doesn't give us any range */
3779 if (dst_reg->umax_value > MAX_PACKET_OFF ||
3780 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
3781 /* Risk of overflow. For instance, ptr + (1<<63) may be less
3782 * than pkt_end, but that's because it's also less than pkt.
3786 new_range = dst_reg->off;
3787 if (range_right_open)
3790 /* Examples for register markings:
3792 * pkt_data in dst register:
3796 * if (r2 > pkt_end) goto <handle exception>
3801 * if (r2 < pkt_end) goto <access okay>
3802 * <handle exception>
3805 * r2 == dst_reg, pkt_end == src_reg
3806 * r2=pkt(id=n,off=8,r=0)
3807 * r3=pkt(id=n,off=0,r=0)
3809 * pkt_data in src register:
3813 * if (pkt_end >= r2) goto <access okay>
3814 * <handle exception>
3818 * if (pkt_end <= r2) goto <handle exception>
3822 * pkt_end == dst_reg, r2 == src_reg
3823 * r2=pkt(id=n,off=8,r=0)
3824 * r3=pkt(id=n,off=0,r=0)
3826 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
3827 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
3828 * and [r3, r3 + 8-1) respectively is safe to access depending on
3832 /* If our ids match, then we must have the same max_value. And we
3833 * don't care about the other reg's fixed offset, since if it's too big
3834 * the range won't allow anything.
3835 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
3837 for (i = 0; i < MAX_BPF_REG; i++)
3838 if (regs[i].type == type && regs[i].id == dst_reg->id)
3839 /* keep the maximum range already checked */
3840 regs[i].range = max(regs[i].range, new_range);
3842 for (j = 0; j <= vstate->curframe; j++) {
3843 state = vstate->frame[j];
3844 bpf_for_each_spilled_reg(i, state, reg) {
3847 if (reg->type == type && reg->id == dst_reg->id)
3848 reg->range = max(reg->range, new_range);
3853 /* compute branch direction of the expression "if (reg opcode val) goto target;"
3855 * 1 - branch will be taken and "goto target" will be executed
3856 * 0 - branch will not be taken and fall-through to next insn
3857 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value range [0,10]
3859 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
3861 if (__is_pointer_value(false, reg))
3866 if (tnum_is_const(reg->var_off))
3867 return !!tnum_equals_const(reg->var_off, val);
3870 if (tnum_is_const(reg->var_off))
3871 return !tnum_equals_const(reg->var_off, val);
3874 if ((~reg->var_off.mask & reg->var_off.value) & val)
3876 if (!((reg->var_off.mask | reg->var_off.value) & val))
3880 if (reg->umin_value > val)
3882 else if (reg->umax_value <= val)
3886 if (reg->smin_value > (s64)val)
3888 else if (reg->smax_value < (s64)val)
3892 if (reg->umax_value < val)
3894 else if (reg->umin_value >= val)
3898 if (reg->smax_value < (s64)val)
3900 else if (reg->smin_value >= (s64)val)
3904 if (reg->umin_value >= val)
3906 else if (reg->umax_value < val)
3910 if (reg->smin_value >= (s64)val)
3912 else if (reg->smax_value < (s64)val)
3916 if (reg->umax_value <= val)
3918 else if (reg->umin_value > val)
3922 if (reg->smax_value <= (s64)val)
3924 else if (reg->smin_value > (s64)val)
3932 /* Adjusts the register min/max values in the case that the dst_reg is the
3933 * variable register that we are working on, and src_reg is a constant or we're
3934 * simply doing a BPF_K check.
3935 * In JEQ/JNE cases we also adjust the var_off values.
3937 static void reg_set_min_max(struct bpf_reg_state *true_reg,
3938 struct bpf_reg_state *false_reg, u64 val,
3941 /* If the dst_reg is a pointer, we can't learn anything about its
3942 * variable offset from the compare (unless src_reg were a pointer into
3943 * the same object, but we don't bother with that.
3944 * Since false_reg and true_reg have the same type by construction, we
3945 * only need to check one of them for pointerness.
3947 if (__is_pointer_value(false, false_reg))
3952 /* If this is false then we know nothing Jon Snow, but if it is
3953 * true then we know for sure.
3955 __mark_reg_known(true_reg, val);
3958 /* If this is true we know nothing Jon Snow, but if it is false
3959 * we know the value for sure;
3961 __mark_reg_known(false_reg, val);
3964 false_reg->var_off = tnum_and(false_reg->var_off,
3966 if (is_power_of_2(val))
3967 true_reg->var_off = tnum_or(true_reg->var_off,
3971 false_reg->umax_value = min(false_reg->umax_value, val);
3972 true_reg->umin_value = max(true_reg->umin_value, val + 1);
3975 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
3976 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
3979 false_reg->umin_value = max(false_reg->umin_value, val);
3980 true_reg->umax_value = min(true_reg->umax_value, val - 1);
3983 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
3984 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
3987 false_reg->umax_value = min(false_reg->umax_value, val - 1);
3988 true_reg->umin_value = max(true_reg->umin_value, val);
3991 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
3992 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
3995 false_reg->umin_value = max(false_reg->umin_value, val + 1);
3996 true_reg->umax_value = min(true_reg->umax_value, val);
3999 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
4000 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
4006 __reg_deduce_bounds(false_reg);
4007 __reg_deduce_bounds(true_reg);
4008 /* We might have learned some bits from the bounds. */
4009 __reg_bound_offset(false_reg);
4010 __reg_bound_offset(true_reg);
4011 /* Intersecting with the old var_off might have improved our bounds
4012 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
4013 * then new var_off is (0; 0x7f...fc) which improves our umax.
4015 __update_reg_bounds(false_reg);
4016 __update_reg_bounds(true_reg);
4019 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
4022 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
4023 struct bpf_reg_state *false_reg, u64 val,
4026 if (__is_pointer_value(false, false_reg))
4031 /* If this is false then we know nothing Jon Snow, but if it is
4032 * true then we know for sure.
4034 __mark_reg_known(true_reg, val);
4037 /* If this is true we know nothing Jon Snow, but if it is false
4038 * we know the value for sure;
4040 __mark_reg_known(false_reg, val);
4043 false_reg->var_off = tnum_and(false_reg->var_off,
4045 if (is_power_of_2(val))
4046 true_reg->var_off = tnum_or(true_reg->var_off,
4050 true_reg->umax_value = min(true_reg->umax_value, val - 1);
4051 false_reg->umin_value = max(false_reg->umin_value, val);
4054 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
4055 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
4058 true_reg->umin_value = max(true_reg->umin_value, val + 1);
4059 false_reg->umax_value = min(false_reg->umax_value, val);
4062 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
4063 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
4066 true_reg->umax_value = min(true_reg->umax_value, val);
4067 false_reg->umin_value = max(false_reg->umin_value, val + 1);
4070 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
4071 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
4074 true_reg->umin_value = max(true_reg->umin_value, val);
4075 false_reg->umax_value = min(false_reg->umax_value, val - 1);
4078 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
4079 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
4085 __reg_deduce_bounds(false_reg);
4086 __reg_deduce_bounds(true_reg);
4087 /* We might have learned some bits from the bounds. */
4088 __reg_bound_offset(false_reg);
4089 __reg_bound_offset(true_reg);
4090 /* Intersecting with the old var_off might have improved our bounds
4091 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
4092 * then new var_off is (0; 0x7f...fc) which improves our umax.
4094 __update_reg_bounds(false_reg);
4095 __update_reg_bounds(true_reg);
4098 /* Regs are known to be equal, so intersect their min/max/var_off */
4099 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
4100 struct bpf_reg_state *dst_reg)
4102 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
4103 dst_reg->umin_value);
4104 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
4105 dst_reg->umax_value);
4106 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
4107 dst_reg->smin_value);
4108 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
4109 dst_reg->smax_value);
4110 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
4112 /* We might have learned new bounds from the var_off. */
4113 __update_reg_bounds(src_reg);
4114 __update_reg_bounds(dst_reg);
4115 /* We might have learned something about the sign bit. */
4116 __reg_deduce_bounds(src_reg);
4117 __reg_deduce_bounds(dst_reg);
4118 /* We might have learned some bits from the bounds. */
4119 __reg_bound_offset(src_reg);
4120 __reg_bound_offset(dst_reg);
4121 /* Intersecting with the old var_off might have improved our bounds
4122 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
4123 * then new var_off is (0; 0x7f...fc) which improves our umax.
4125 __update_reg_bounds(src_reg);
4126 __update_reg_bounds(dst_reg);
4129 static void reg_combine_min_max(struct bpf_reg_state *true_src,
4130 struct bpf_reg_state *true_dst,
4131 struct bpf_reg_state *false_src,
4132 struct bpf_reg_state *false_dst,
4137 __reg_combine_min_max(true_src, true_dst);
4140 __reg_combine_min_max(false_src, false_dst);
4145 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
4146 struct bpf_reg_state *reg, u32 id,
4149 if (reg_type_may_be_null(reg->type) && reg->id == id) {
4150 /* Old offset (both fixed and variable parts) should
4151 * have been known-zero, because we don't allow pointer
4152 * arithmetic on pointers that might be NULL.
4154 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
4155 !tnum_equals_const(reg->var_off, 0) ||
4157 __mark_reg_known_zero(reg);
4161 reg->type = SCALAR_VALUE;
4162 } else if (reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
4163 if (reg->map_ptr->inner_map_meta) {
4164 reg->type = CONST_PTR_TO_MAP;
4165 reg->map_ptr = reg->map_ptr->inner_map_meta;
4167 reg->type = PTR_TO_MAP_VALUE;
4169 } else if (reg->type == PTR_TO_SOCKET_OR_NULL) {
4170 reg->type = PTR_TO_SOCKET;
4172 if (is_null || !reg_is_refcounted(reg)) {
4173 /* We don't need id from this point onwards anymore,
4174 * thus we should better reset it, so that state
4175 * pruning has chances to take effect.
4182 /* The logic is similar to find_good_pkt_pointers(), both could eventually
4183 * be folded together at some point.
4185 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
4188 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4189 struct bpf_reg_state *reg, *regs = state->regs;
4190 u32 id = regs[regno].id;
4193 if (reg_is_refcounted_or_null(®s[regno]) && is_null)
4194 __release_reference_state(state, id);
4196 for (i = 0; i < MAX_BPF_REG; i++)
4197 mark_ptr_or_null_reg(state, ®s[i], id, is_null);
4199 for (j = 0; j <= vstate->curframe; j++) {
4200 state = vstate->frame[j];
4201 bpf_for_each_spilled_reg(i, state, reg) {
4204 mark_ptr_or_null_reg(state, reg, id, is_null);
4209 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
4210 struct bpf_reg_state *dst_reg,
4211 struct bpf_reg_state *src_reg,
4212 struct bpf_verifier_state *this_branch,
4213 struct bpf_verifier_state *other_branch)
4215 if (BPF_SRC(insn->code) != BPF_X)
4218 switch (BPF_OP(insn->code)) {
4220 if ((dst_reg->type == PTR_TO_PACKET &&
4221 src_reg->type == PTR_TO_PACKET_END) ||
4222 (dst_reg->type == PTR_TO_PACKET_META &&
4223 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
4224 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
4225 find_good_pkt_pointers(this_branch, dst_reg,
4226 dst_reg->type, false);
4227 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
4228 src_reg->type == PTR_TO_PACKET) ||
4229 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
4230 src_reg->type == PTR_TO_PACKET_META)) {
4231 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
4232 find_good_pkt_pointers(other_branch, src_reg,
4233 src_reg->type, true);
4239 if ((dst_reg->type == PTR_TO_PACKET &&
4240 src_reg->type == PTR_TO_PACKET_END) ||
4241 (dst_reg->type == PTR_TO_PACKET_META &&
4242 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
4243 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
4244 find_good_pkt_pointers(other_branch, dst_reg,
4245 dst_reg->type, true);
4246 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
4247 src_reg->type == PTR_TO_PACKET) ||
4248 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
4249 src_reg->type == PTR_TO_PACKET_META)) {
4250 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
4251 find_good_pkt_pointers(this_branch, src_reg,
4252 src_reg->type, false);
4258 if ((dst_reg->type == PTR_TO_PACKET &&
4259 src_reg->type == PTR_TO_PACKET_END) ||
4260 (dst_reg->type == PTR_TO_PACKET_META &&
4261 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
4262 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
4263 find_good_pkt_pointers(this_branch, dst_reg,
4264 dst_reg->type, true);
4265 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
4266 src_reg->type == PTR_TO_PACKET) ||
4267 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
4268 src_reg->type == PTR_TO_PACKET_META)) {
4269 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
4270 find_good_pkt_pointers(other_branch, src_reg,
4271 src_reg->type, false);
4277 if ((dst_reg->type == PTR_TO_PACKET &&
4278 src_reg->type == PTR_TO_PACKET_END) ||
4279 (dst_reg->type == PTR_TO_PACKET_META &&
4280 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
4281 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
4282 find_good_pkt_pointers(other_branch, dst_reg,
4283 dst_reg->type, false);
4284 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
4285 src_reg->type == PTR_TO_PACKET) ||
4286 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
4287 src_reg->type == PTR_TO_PACKET_META)) {
4288 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
4289 find_good_pkt_pointers(this_branch, src_reg,
4290 src_reg->type, true);
4302 static int check_cond_jmp_op(struct bpf_verifier_env *env,
4303 struct bpf_insn *insn, int *insn_idx)
4305 struct bpf_verifier_state *this_branch = env->cur_state;
4306 struct bpf_verifier_state *other_branch;
4307 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
4308 struct bpf_reg_state *dst_reg, *other_branch_regs;
4309 u8 opcode = BPF_OP(insn->code);
4312 if (opcode > BPF_JSLE) {
4313 verbose(env, "invalid BPF_JMP opcode %x\n", opcode);
4317 if (BPF_SRC(insn->code) == BPF_X) {
4318 if (insn->imm != 0) {
4319 verbose(env, "BPF_JMP uses reserved fields\n");
4323 /* check src1 operand */
4324 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4328 if (is_pointer_value(env, insn->src_reg)) {
4329 verbose(env, "R%d pointer comparison prohibited\n",
4334 if (insn->src_reg != BPF_REG_0) {
4335 verbose(env, "BPF_JMP uses reserved fields\n");
4340 /* check src2 operand */
4341 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4345 dst_reg = ®s[insn->dst_reg];
4347 if (BPF_SRC(insn->code) == BPF_K) {
4348 int pred = is_branch_taken(dst_reg, insn->imm, opcode);
4351 /* only follow the goto, ignore fall-through */
4352 *insn_idx += insn->off;
4354 } else if (pred == 0) {
4355 /* only follow fall-through branch, since
4356 * that's where the program will go
4362 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx);
4365 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
4367 /* detect if we are comparing against a constant value so we can adjust
4368 * our min/max values for our dst register.
4369 * this is only legit if both are scalars (or pointers to the same
4370 * object, I suppose, but we don't support that right now), because
4371 * otherwise the different base pointers mean the offsets aren't
4374 if (BPF_SRC(insn->code) == BPF_X) {
4375 if (dst_reg->type == SCALAR_VALUE &&
4376 regs[insn->src_reg].type == SCALAR_VALUE) {
4377 if (tnum_is_const(regs[insn->src_reg].var_off))
4378 reg_set_min_max(&other_branch_regs[insn->dst_reg],
4379 dst_reg, regs[insn->src_reg].var_off.value,
4381 else if (tnum_is_const(dst_reg->var_off))
4382 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
4383 ®s[insn->src_reg],
4384 dst_reg->var_off.value, opcode);
4385 else if (opcode == BPF_JEQ || opcode == BPF_JNE)
4386 /* Comparing for equality, we can combine knowledge */
4387 reg_combine_min_max(&other_branch_regs[insn->src_reg],
4388 &other_branch_regs[insn->dst_reg],
4389 ®s[insn->src_reg],
4390 ®s[insn->dst_reg], opcode);
4392 } else if (dst_reg->type == SCALAR_VALUE) {
4393 reg_set_min_max(&other_branch_regs[insn->dst_reg],
4394 dst_reg, insn->imm, opcode);
4397 /* detect if R == 0 where R is returned from bpf_map_lookup_elem() */
4398 if (BPF_SRC(insn->code) == BPF_K &&
4399 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
4400 reg_type_may_be_null(dst_reg->type)) {
4401 /* Mark all identical registers in each branch as either
4402 * safe or unknown depending R == 0 or R != 0 conditional.
4404 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
4406 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
4408 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
4409 this_branch, other_branch) &&
4410 is_pointer_value(env, insn->dst_reg)) {
4411 verbose(env, "R%d pointer comparison prohibited\n",
4416 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
4420 /* return the map pointer stored inside BPF_LD_IMM64 instruction */
4421 static struct bpf_map *ld_imm64_to_map_ptr(struct bpf_insn *insn)
4423 u64 imm64 = ((u64) (u32) insn[0].imm) | ((u64) (u32) insn[1].imm) << 32;
4425 return (struct bpf_map *) (unsigned long) imm64;
4428 /* verify BPF_LD_IMM64 instruction */
4429 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
4431 struct bpf_reg_state *regs = cur_regs(env);
4434 if (BPF_SIZE(insn->code) != BPF_DW) {
4435 verbose(env, "invalid BPF_LD_IMM insn\n");
4438 if (insn->off != 0) {
4439 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
4443 err = check_reg_arg(env, insn->dst_reg, DST_OP);
4447 if (insn->src_reg == 0) {
4448 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
4450 regs[insn->dst_reg].type = SCALAR_VALUE;
4451 __mark_reg_known(®s[insn->dst_reg], imm);
4455 /* replace_map_fd_with_map_ptr() should have caught bad ld_imm64 */
4456 BUG_ON(insn->src_reg != BPF_PSEUDO_MAP_FD);
4458 regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
4459 regs[insn->dst_reg].map_ptr = ld_imm64_to_map_ptr(insn);
4463 static bool may_access_skb(enum bpf_prog_type type)
4466 case BPF_PROG_TYPE_SOCKET_FILTER:
4467 case BPF_PROG_TYPE_SCHED_CLS:
4468 case BPF_PROG_TYPE_SCHED_ACT:
4475 /* verify safety of LD_ABS|LD_IND instructions:
4476 * - they can only appear in the programs where ctx == skb
4477 * - since they are wrappers of function calls, they scratch R1-R5 registers,
4478 * preserve R6-R9, and store return value into R0
4481 * ctx == skb == R6 == CTX
4484 * SRC == any register
4485 * IMM == 32-bit immediate
4488 * R0 - 8/16/32-bit skb data converted to cpu endianness
4490 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
4492 struct bpf_reg_state *regs = cur_regs(env);
4493 u8 mode = BPF_MODE(insn->code);
4496 if (!may_access_skb(env->prog->type)) {
4497 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
4501 if (!env->ops->gen_ld_abs) {
4502 verbose(env, "bpf verifier is misconfigured\n");
4506 if (env->subprog_cnt > 1) {
4507 /* when program has LD_ABS insn JITs and interpreter assume
4508 * that r1 == ctx == skb which is not the case for callees
4509 * that can have arbitrary arguments. It's problematic
4510 * for main prog as well since JITs would need to analyze
4511 * all functions in order to make proper register save/restore
4512 * decisions in the main prog. Hence disallow LD_ABS with calls
4514 verbose(env, "BPF_LD_[ABS|IND] instructions cannot be mixed with bpf-to-bpf calls\n");
4518 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
4519 BPF_SIZE(insn->code) == BPF_DW ||
4520 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
4521 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
4525 /* check whether implicit source operand (register R6) is readable */
4526 err = check_reg_arg(env, BPF_REG_6, SRC_OP);
4530 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
4531 * gen_ld_abs() may terminate the program at runtime, leading to
4534 err = check_reference_leak(env);
4536 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
4540 if (regs[BPF_REG_6].type != PTR_TO_CTX) {
4542 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
4546 if (mode == BPF_IND) {
4547 /* check explicit source operand */
4548 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4553 /* reset caller saved regs to unreadable */
4554 for (i = 0; i < CALLER_SAVED_REGS; i++) {
4555 mark_reg_not_init(env, regs, caller_saved[i]);
4556 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
4559 /* mark destination R0 register as readable, since it contains
4560 * the value fetched from the packet.
4561 * Already marked as written above.
4563 mark_reg_unknown(env, regs, BPF_REG_0);
4567 static int check_return_code(struct bpf_verifier_env *env)
4569 struct bpf_reg_state *reg;
4570 struct tnum range = tnum_range(0, 1);
4572 switch (env->prog->type) {
4573 case BPF_PROG_TYPE_CGROUP_SKB:
4574 case BPF_PROG_TYPE_CGROUP_SOCK:
4575 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
4576 case BPF_PROG_TYPE_SOCK_OPS:
4577 case BPF_PROG_TYPE_CGROUP_DEVICE:
4583 reg = cur_regs(env) + BPF_REG_0;
4584 if (reg->type != SCALAR_VALUE) {
4585 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
4586 reg_type_str[reg->type]);
4590 if (!tnum_in(range, reg->var_off)) {
4591 verbose(env, "At program exit the register R0 ");
4592 if (!tnum_is_unknown(reg->var_off)) {
4595 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4596 verbose(env, "has value %s", tn_buf);
4598 verbose(env, "has unknown scalar value");
4600 verbose(env, " should have been 0 or 1\n");
4606 /* non-recursive DFS pseudo code
4607 * 1 procedure DFS-iterative(G,v):
4608 * 2 label v as discovered
4609 * 3 let S be a stack
4611 * 5 while S is not empty
4613 * 7 if t is what we're looking for:
4615 * 9 for all edges e in G.adjacentEdges(t) do
4616 * 10 if edge e is already labelled
4617 * 11 continue with the next edge
4618 * 12 w <- G.adjacentVertex(t,e)
4619 * 13 if vertex w is not discovered and not explored
4620 * 14 label e as tree-edge
4621 * 15 label w as discovered
4624 * 18 else if vertex w is discovered
4625 * 19 label e as back-edge
4627 * 21 // vertex w is explored
4628 * 22 label e as forward- or cross-edge
4629 * 23 label t as explored
4634 * 0x11 - discovered and fall-through edge labelled
4635 * 0x12 - discovered and fall-through and branch edges labelled
4646 #define STATE_LIST_MARK ((struct bpf_verifier_state_list *) -1L)
4648 static int *insn_stack; /* stack of insns to process */
4649 static int cur_stack; /* current stack index */
4650 static int *insn_state;
4652 /* t, w, e - match pseudo-code above:
4653 * t - index of current instruction
4654 * w - next instruction
4657 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env)
4659 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
4662 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
4665 if (w < 0 || w >= env->prog->len) {
4666 verbose_linfo(env, t, "%d: ", t);
4667 verbose(env, "jump out of range from insn %d to %d\n", t, w);
4672 /* mark branch target for state pruning */
4673 env->explored_states[w] = STATE_LIST_MARK;
4675 if (insn_state[w] == 0) {
4677 insn_state[t] = DISCOVERED | e;
4678 insn_state[w] = DISCOVERED;
4679 if (cur_stack >= env->prog->len)
4681 insn_stack[cur_stack++] = w;
4683 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
4684 verbose_linfo(env, t, "%d: ", t);
4685 verbose_linfo(env, w, "%d: ", w);
4686 verbose(env, "back-edge from insn %d to %d\n", t, w);
4688 } else if (insn_state[w] == EXPLORED) {
4689 /* forward- or cross-edge */
4690 insn_state[t] = DISCOVERED | e;
4692 verbose(env, "insn state internal bug\n");
4698 /* non-recursive depth-first-search to detect loops in BPF program
4699 * loop == back-edge in directed graph
4701 static int check_cfg(struct bpf_verifier_env *env)
4703 struct bpf_insn *insns = env->prog->insnsi;
4704 int insn_cnt = env->prog->len;
4708 insn_state = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
4712 insn_stack = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
4718 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
4719 insn_stack[0] = 0; /* 0 is the first instruction */
4725 t = insn_stack[cur_stack - 1];
4727 if (BPF_CLASS(insns[t].code) == BPF_JMP) {
4728 u8 opcode = BPF_OP(insns[t].code);
4730 if (opcode == BPF_EXIT) {
4732 } else if (opcode == BPF_CALL) {
4733 ret = push_insn(t, t + 1, FALLTHROUGH, env);
4738 if (t + 1 < insn_cnt)
4739 env->explored_states[t + 1] = STATE_LIST_MARK;
4740 if (insns[t].src_reg == BPF_PSEUDO_CALL) {
4741 env->explored_states[t] = STATE_LIST_MARK;
4742 ret = push_insn(t, t + insns[t].imm + 1, BRANCH, env);
4748 } else if (opcode == BPF_JA) {
4749 if (BPF_SRC(insns[t].code) != BPF_K) {
4753 /* unconditional jump with single edge */
4754 ret = push_insn(t, t + insns[t].off + 1,
4760 /* tell verifier to check for equivalent states
4761 * after every call and jump
4763 if (t + 1 < insn_cnt)
4764 env->explored_states[t + 1] = STATE_LIST_MARK;
4766 /* conditional jump with two edges */
4767 env->explored_states[t] = STATE_LIST_MARK;
4768 ret = push_insn(t, t + 1, FALLTHROUGH, env);
4774 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env);
4781 /* all other non-branch instructions with single
4784 ret = push_insn(t, t + 1, FALLTHROUGH, env);
4792 insn_state[t] = EXPLORED;
4793 if (cur_stack-- <= 0) {
4794 verbose(env, "pop stack internal bug\n");
4801 for (i = 0; i < insn_cnt; i++) {
4802 if (insn_state[i] != EXPLORED) {
4803 verbose(env, "unreachable insn %d\n", i);
4808 ret = 0; /* cfg looks good */
4816 /* The minimum supported BTF func info size */
4817 #define MIN_BPF_FUNCINFO_SIZE 8
4818 #define MAX_FUNCINFO_REC_SIZE 252
4820 static int check_btf_func(struct bpf_verifier_env *env,
4821 const union bpf_attr *attr,
4822 union bpf_attr __user *uattr)
4824 u32 i, nfuncs, urec_size, min_size, prev_offset;
4825 u32 krec_size = sizeof(struct bpf_func_info);
4826 struct bpf_func_info *krecord;
4827 const struct btf_type *type;
4828 struct bpf_prog *prog;
4829 const struct btf *btf;
4830 void __user *urecord;
4833 nfuncs = attr->func_info_cnt;
4837 if (nfuncs != env->subprog_cnt) {
4838 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
4842 urec_size = attr->func_info_rec_size;
4843 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
4844 urec_size > MAX_FUNCINFO_REC_SIZE ||
4845 urec_size % sizeof(u32)) {
4846 verbose(env, "invalid func info rec size %u\n", urec_size);
4851 btf = prog->aux->btf;
4853 urecord = u64_to_user_ptr(attr->func_info);
4854 min_size = min_t(u32, krec_size, urec_size);
4856 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
4860 for (i = 0; i < nfuncs; i++) {
4861 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
4863 if (ret == -E2BIG) {
4864 verbose(env, "nonzero tailing record in func info");
4865 /* set the size kernel expects so loader can zero
4866 * out the rest of the record.
4868 if (put_user(min_size, &uattr->func_info_rec_size))
4874 if (copy_from_user(&krecord[i], urecord, min_size)) {
4879 /* check insn_off */
4881 if (krecord[i].insn_off) {
4883 "nonzero insn_off %u for the first func info record",
4884 krecord[i].insn_off);
4888 } else if (krecord[i].insn_off <= prev_offset) {
4890 "same or smaller insn offset (%u) than previous func info record (%u)",
4891 krecord[i].insn_off, prev_offset);
4896 if (env->subprog_info[i].start != krecord[i].insn_off) {
4897 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
4903 type = btf_type_by_id(btf, krecord[i].type_id);
4904 if (!type || BTF_INFO_KIND(type->info) != BTF_KIND_FUNC) {
4905 verbose(env, "invalid type id %d in func info",
4906 krecord[i].type_id);
4911 prev_offset = krecord[i].insn_off;
4912 urecord += urec_size;
4915 prog->aux->func_info = krecord;
4916 prog->aux->func_info_cnt = nfuncs;
4924 static void adjust_btf_func(struct bpf_verifier_env *env)
4928 if (!env->prog->aux->func_info)
4931 for (i = 0; i < env->subprog_cnt; i++)
4932 env->prog->aux->func_info[i].insn_off = env->subprog_info[i].start;
4935 #define MIN_BPF_LINEINFO_SIZE (offsetof(struct bpf_line_info, line_col) + \
4936 sizeof(((struct bpf_line_info *)(0))->line_col))
4937 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
4939 static int check_btf_line(struct bpf_verifier_env *env,
4940 const union bpf_attr *attr,
4941 union bpf_attr __user *uattr)
4943 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
4944 struct bpf_subprog_info *sub;
4945 struct bpf_line_info *linfo;
4946 struct bpf_prog *prog;
4947 const struct btf *btf;
4948 void __user *ulinfo;
4951 nr_linfo = attr->line_info_cnt;
4955 rec_size = attr->line_info_rec_size;
4956 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
4957 rec_size > MAX_LINEINFO_REC_SIZE ||
4958 rec_size & (sizeof(u32) - 1))
4961 /* Need to zero it in case the userspace may
4962 * pass in a smaller bpf_line_info object.
4964 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
4965 GFP_KERNEL | __GFP_NOWARN);
4970 btf = prog->aux->btf;
4973 sub = env->subprog_info;
4974 ulinfo = u64_to_user_ptr(attr->line_info);
4975 expected_size = sizeof(struct bpf_line_info);
4976 ncopy = min_t(u32, expected_size, rec_size);
4977 for (i = 0; i < nr_linfo; i++) {
4978 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
4980 if (err == -E2BIG) {
4981 verbose(env, "nonzero tailing record in line_info");
4982 if (put_user(expected_size,
4983 &uattr->line_info_rec_size))
4989 if (copy_from_user(&linfo[i], ulinfo, ncopy)) {
4995 * Check insn_off to ensure
4996 * 1) strictly increasing AND
4997 * 2) bounded by prog->len
4999 * The linfo[0].insn_off == 0 check logically falls into
5000 * the later "missing bpf_line_info for func..." case
5001 * because the first linfo[0].insn_off must be the
5002 * first sub also and the first sub must have
5003 * subprog_info[0].start == 0.
5005 if ((i && linfo[i].insn_off <= prev_offset) ||
5006 linfo[i].insn_off >= prog->len) {
5007 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
5008 i, linfo[i].insn_off, prev_offset,
5014 if (!prog->insnsi[linfo[i].insn_off].code) {
5016 "Invalid insn code at line_info[%u].insn_off\n",
5022 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
5023 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
5024 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
5029 if (s != env->subprog_cnt) {
5030 if (linfo[i].insn_off == sub[s].start) {
5031 sub[s].linfo_idx = i;
5033 } else if (sub[s].start < linfo[i].insn_off) {
5034 verbose(env, "missing bpf_line_info for func#%u\n", s);
5040 prev_offset = linfo[i].insn_off;
5044 if (s != env->subprog_cnt) {
5045 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
5046 env->subprog_cnt - s, s);
5051 prog->aux->linfo = linfo;
5052 prog->aux->nr_linfo = nr_linfo;
5061 static int check_btf_info(struct bpf_verifier_env *env,
5062 const union bpf_attr *attr,
5063 union bpf_attr __user *uattr)
5068 if (!attr->func_info_cnt && !attr->line_info_cnt)
5071 btf = btf_get_by_fd(attr->prog_btf_fd);
5073 return PTR_ERR(btf);
5074 env->prog->aux->btf = btf;
5076 err = check_btf_func(env, attr, uattr);
5080 err = check_btf_line(env, attr, uattr);
5087 /* check %cur's range satisfies %old's */
5088 static bool range_within(struct bpf_reg_state *old,
5089 struct bpf_reg_state *cur)
5091 return old->umin_value <= cur->umin_value &&
5092 old->umax_value >= cur->umax_value &&
5093 old->smin_value <= cur->smin_value &&
5094 old->smax_value >= cur->smax_value;
5097 /* Maximum number of register states that can exist at once */
5098 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
5104 /* If in the old state two registers had the same id, then they need to have
5105 * the same id in the new state as well. But that id could be different from
5106 * the old state, so we need to track the mapping from old to new ids.
5107 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
5108 * regs with old id 5 must also have new id 9 for the new state to be safe. But
5109 * regs with a different old id could still have new id 9, we don't care about
5111 * So we look through our idmap to see if this old id has been seen before. If
5112 * so, we require the new id to match; otherwise, we add the id pair to the map.
5114 static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
5118 for (i = 0; i < ID_MAP_SIZE; i++) {
5119 if (!idmap[i].old) {
5120 /* Reached an empty slot; haven't seen this id before */
5121 idmap[i].old = old_id;
5122 idmap[i].cur = cur_id;
5125 if (idmap[i].old == old_id)
5126 return idmap[i].cur == cur_id;
5128 /* We ran out of idmap slots, which should be impossible */
5133 static void clean_func_state(struct bpf_verifier_env *env,
5134 struct bpf_func_state *st)
5136 enum bpf_reg_liveness live;
5139 for (i = 0; i < BPF_REG_FP; i++) {
5140 live = st->regs[i].live;
5141 /* liveness must not touch this register anymore */
5142 st->regs[i].live |= REG_LIVE_DONE;
5143 if (!(live & REG_LIVE_READ))
5144 /* since the register is unused, clear its state
5145 * to make further comparison simpler
5147 __mark_reg_not_init(&st->regs[i]);
5150 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
5151 live = st->stack[i].spilled_ptr.live;
5152 /* liveness must not touch this stack slot anymore */
5153 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
5154 if (!(live & REG_LIVE_READ)) {
5155 __mark_reg_not_init(&st->stack[i].spilled_ptr);
5156 for (j = 0; j < BPF_REG_SIZE; j++)
5157 st->stack[i].slot_type[j] = STACK_INVALID;
5162 static void clean_verifier_state(struct bpf_verifier_env *env,
5163 struct bpf_verifier_state *st)
5167 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
5168 /* all regs in this state in all frames were already marked */
5171 for (i = 0; i <= st->curframe; i++)
5172 clean_func_state(env, st->frame[i]);
5175 /* the parentage chains form a tree.
5176 * the verifier states are added to state lists at given insn and
5177 * pushed into state stack for future exploration.
5178 * when the verifier reaches bpf_exit insn some of the verifer states
5179 * stored in the state lists have their final liveness state already,
5180 * but a lot of states will get revised from liveness point of view when
5181 * the verifier explores other branches.
5184 * 2: if r1 == 100 goto pc+1
5187 * when the verifier reaches exit insn the register r0 in the state list of
5188 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
5189 * of insn 2 and goes exploring further. At the insn 4 it will walk the
5190 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
5192 * Since the verifier pushes the branch states as it sees them while exploring
5193 * the program the condition of walking the branch instruction for the second
5194 * time means that all states below this branch were already explored and
5195 * their final liveness markes are already propagated.
5196 * Hence when the verifier completes the search of state list in is_state_visited()
5197 * we can call this clean_live_states() function to mark all liveness states
5198 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
5200 * This function also clears the registers and stack for states that !READ
5201 * to simplify state merging.
5203 * Important note here that walking the same branch instruction in the callee
5204 * doesn't meant that the states are DONE. The verifier has to compare
5207 static void clean_live_states(struct bpf_verifier_env *env, int insn,
5208 struct bpf_verifier_state *cur)
5210 struct bpf_verifier_state_list *sl;
5213 sl = env->explored_states[insn];
5217 while (sl != STATE_LIST_MARK) {
5218 if (sl->state.curframe != cur->curframe)
5220 for (i = 0; i <= cur->curframe; i++)
5221 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
5223 clean_verifier_state(env, &sl->state);
5229 /* Returns true if (rold safe implies rcur safe) */
5230 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
5231 struct idpair *idmap)
5235 if (!(rold->live & REG_LIVE_READ))
5236 /* explored state didn't use this */
5239 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
5241 if (rold->type == PTR_TO_STACK)
5242 /* two stack pointers are equal only if they're pointing to
5243 * the same stack frame, since fp-8 in foo != fp-8 in bar
5245 return equal && rold->frameno == rcur->frameno;
5250 if (rold->type == NOT_INIT)
5251 /* explored state can't have used this */
5253 if (rcur->type == NOT_INIT)
5255 switch (rold->type) {
5257 if (rcur->type == SCALAR_VALUE) {
5258 /* new val must satisfy old val knowledge */
5259 return range_within(rold, rcur) &&
5260 tnum_in(rold->var_off, rcur->var_off);
5262 /* We're trying to use a pointer in place of a scalar.
5263 * Even if the scalar was unbounded, this could lead to
5264 * pointer leaks because scalars are allowed to leak
5265 * while pointers are not. We could make this safe in
5266 * special cases if root is calling us, but it's
5267 * probably not worth the hassle.
5271 case PTR_TO_MAP_VALUE:
5272 /* If the new min/max/var_off satisfy the old ones and
5273 * everything else matches, we are OK.
5274 * We don't care about the 'id' value, because nothing
5275 * uses it for PTR_TO_MAP_VALUE (only for ..._OR_NULL)
5277 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
5278 range_within(rold, rcur) &&
5279 tnum_in(rold->var_off, rcur->var_off);
5280 case PTR_TO_MAP_VALUE_OR_NULL:
5281 /* a PTR_TO_MAP_VALUE could be safe to use as a
5282 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
5283 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
5284 * checked, doing so could have affected others with the same
5285 * id, and we can't check for that because we lost the id when
5286 * we converted to a PTR_TO_MAP_VALUE.
5288 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
5290 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
5292 /* Check our ids match any regs they're supposed to */
5293 return check_ids(rold->id, rcur->id, idmap);
5294 case PTR_TO_PACKET_META:
5296 if (rcur->type != rold->type)
5298 /* We must have at least as much range as the old ptr
5299 * did, so that any accesses which were safe before are
5300 * still safe. This is true even if old range < old off,
5301 * since someone could have accessed through (ptr - k), or
5302 * even done ptr -= k in a register, to get a safe access.
5304 if (rold->range > rcur->range)
5306 /* If the offsets don't match, we can't trust our alignment;
5307 * nor can we be sure that we won't fall out of range.
5309 if (rold->off != rcur->off)
5311 /* id relations must be preserved */
5312 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
5314 /* new val must satisfy old val knowledge */
5315 return range_within(rold, rcur) &&
5316 tnum_in(rold->var_off, rcur->var_off);
5318 case CONST_PTR_TO_MAP:
5319 case PTR_TO_PACKET_END:
5320 case PTR_TO_FLOW_KEYS:
5322 case PTR_TO_SOCKET_OR_NULL:
5323 /* Only valid matches are exact, which memcmp() above
5324 * would have accepted
5327 /* Don't know what's going on, just say it's not safe */
5331 /* Shouldn't get here; if we do, say it's not safe */
5336 static bool stacksafe(struct bpf_func_state *old,
5337 struct bpf_func_state *cur,
5338 struct idpair *idmap)
5342 /* walk slots of the explored stack and ignore any additional
5343 * slots in the current stack, since explored(safe) state
5346 for (i = 0; i < old->allocated_stack; i++) {
5347 spi = i / BPF_REG_SIZE;
5349 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
5350 i += BPF_REG_SIZE - 1;
5351 /* explored state didn't use this */
5355 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
5358 /* explored stack has more populated slots than current stack
5359 * and these slots were used
5361 if (i >= cur->allocated_stack)
5364 /* if old state was safe with misc data in the stack
5365 * it will be safe with zero-initialized stack.
5366 * The opposite is not true
5368 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
5369 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
5371 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
5372 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
5373 /* Ex: old explored (safe) state has STACK_SPILL in
5374 * this stack slot, but current has has STACK_MISC ->
5375 * this verifier states are not equivalent,
5376 * return false to continue verification of this path
5379 if (i % BPF_REG_SIZE)
5381 if (old->stack[spi].slot_type[0] != STACK_SPILL)
5383 if (!regsafe(&old->stack[spi].spilled_ptr,
5384 &cur->stack[spi].spilled_ptr,
5386 /* when explored and current stack slot are both storing
5387 * spilled registers, check that stored pointers types
5388 * are the same as well.
5389 * Ex: explored safe path could have stored
5390 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
5391 * but current path has stored:
5392 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
5393 * such verifier states are not equivalent.
5394 * return false to continue verification of this path
5401 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
5403 if (old->acquired_refs != cur->acquired_refs)
5405 return !memcmp(old->refs, cur->refs,
5406 sizeof(*old->refs) * old->acquired_refs);
5409 /* compare two verifier states
5411 * all states stored in state_list are known to be valid, since
5412 * verifier reached 'bpf_exit' instruction through them
5414 * this function is called when verifier exploring different branches of
5415 * execution popped from the state stack. If it sees an old state that has
5416 * more strict register state and more strict stack state then this execution
5417 * branch doesn't need to be explored further, since verifier already
5418 * concluded that more strict state leads to valid finish.
5420 * Therefore two states are equivalent if register state is more conservative
5421 * and explored stack state is more conservative than the current one.
5424 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
5425 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
5427 * In other words if current stack state (one being explored) has more
5428 * valid slots than old one that already passed validation, it means
5429 * the verifier can stop exploring and conclude that current state is valid too
5431 * Similarly with registers. If explored state has register type as invalid
5432 * whereas register type in current state is meaningful, it means that
5433 * the current state will reach 'bpf_exit' instruction safely
5435 static bool func_states_equal(struct bpf_func_state *old,
5436 struct bpf_func_state *cur)
5438 struct idpair *idmap;
5442 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
5443 /* If we failed to allocate the idmap, just say it's not safe */
5447 for (i = 0; i < MAX_BPF_REG; i++) {
5448 if (!regsafe(&old->regs[i], &cur->regs[i], idmap))
5452 if (!stacksafe(old, cur, idmap))
5455 if (!refsafe(old, cur))
5463 static bool states_equal(struct bpf_verifier_env *env,
5464 struct bpf_verifier_state *old,
5465 struct bpf_verifier_state *cur)
5469 if (old->curframe != cur->curframe)
5472 /* for states to be equal callsites have to be the same
5473 * and all frame states need to be equivalent
5475 for (i = 0; i <= old->curframe; i++) {
5476 if (old->frame[i]->callsite != cur->frame[i]->callsite)
5478 if (!func_states_equal(old->frame[i], cur->frame[i]))
5484 /* A write screens off any subsequent reads; but write marks come from the
5485 * straight-line code between a state and its parent. When we arrive at an
5486 * equivalent state (jump target or such) we didn't arrive by the straight-line
5487 * code, so read marks in the state must propagate to the parent regardless
5488 * of the state's write marks. That's what 'parent == state->parent' comparison
5489 * in mark_reg_read() is for.
5491 static int propagate_liveness(struct bpf_verifier_env *env,
5492 const struct bpf_verifier_state *vstate,
5493 struct bpf_verifier_state *vparent)
5495 int i, frame, err = 0;
5496 struct bpf_func_state *state, *parent;
5498 if (vparent->curframe != vstate->curframe) {
5499 WARN(1, "propagate_live: parent frame %d current frame %d\n",
5500 vparent->curframe, vstate->curframe);
5503 /* Propagate read liveness of registers... */
5504 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
5505 /* We don't need to worry about FP liveness because it's read-only */
5506 for (i = 0; i < BPF_REG_FP; i++) {
5507 if (vparent->frame[vparent->curframe]->regs[i].live & REG_LIVE_READ)
5509 if (vstate->frame[vstate->curframe]->regs[i].live & REG_LIVE_READ) {
5510 err = mark_reg_read(env, &vstate->frame[vstate->curframe]->regs[i],
5511 &vparent->frame[vstate->curframe]->regs[i]);
5517 /* ... and stack slots */
5518 for (frame = 0; frame <= vstate->curframe; frame++) {
5519 state = vstate->frame[frame];
5520 parent = vparent->frame[frame];
5521 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
5522 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
5523 if (parent->stack[i].spilled_ptr.live & REG_LIVE_READ)
5525 if (state->stack[i].spilled_ptr.live & REG_LIVE_READ)
5526 mark_reg_read(env, &state->stack[i].spilled_ptr,
5527 &parent->stack[i].spilled_ptr);
5533 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
5535 struct bpf_verifier_state_list *new_sl;
5536 struct bpf_verifier_state_list *sl;
5537 struct bpf_verifier_state *cur = env->cur_state, *new;
5538 int i, j, err, states_cnt = 0;
5540 sl = env->explored_states[insn_idx];
5542 /* this 'insn_idx' instruction wasn't marked, so we will not
5543 * be doing state search here
5547 clean_live_states(env, insn_idx, cur);
5549 while (sl != STATE_LIST_MARK) {
5550 if (states_equal(env, &sl->state, cur)) {
5551 /* reached equivalent register/stack state,
5553 * Registers read by the continuation are read by us.
5554 * If we have any write marks in env->cur_state, they
5555 * will prevent corresponding reads in the continuation
5556 * from reaching our parent (an explored_state). Our
5557 * own state will get the read marks recorded, but
5558 * they'll be immediately forgotten as we're pruning
5559 * this state and will pop a new one.
5561 err = propagate_liveness(env, &sl->state, cur);
5570 if (!env->allow_ptr_leaks && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
5573 /* there were no equivalent states, remember current one.
5574 * technically the current state is not proven to be safe yet,
5575 * but it will either reach outer most bpf_exit (which means it's safe)
5576 * or it will be rejected. Since there are no loops, we won't be
5577 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
5578 * again on the way to bpf_exit
5580 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
5584 /* add new state to the head of linked list */
5585 new = &new_sl->state;
5586 err = copy_verifier_state(new, cur);
5588 free_verifier_state(new, false);
5592 new_sl->next = env->explored_states[insn_idx];
5593 env->explored_states[insn_idx] = new_sl;
5594 /* connect new state to parentage chain. Current frame needs all
5595 * registers connected. Only r6 - r9 of the callers are alive (pushed
5596 * to the stack implicitly by JITs) so in callers' frames connect just
5597 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
5598 * the state of the call instruction (with WRITTEN set), and r0 comes
5599 * from callee with its full parentage chain, anyway.
5601 for (j = 0; j <= cur->curframe; j++)
5602 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
5603 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
5604 /* clear write marks in current state: the writes we did are not writes
5605 * our child did, so they don't screen off its reads from us.
5606 * (There are no read marks in current state, because reads always mark
5607 * their parent and current state never has children yet. Only
5608 * explored_states can get read marks.)
5610 for (i = 0; i < BPF_REG_FP; i++)
5611 cur->frame[cur->curframe]->regs[i].live = REG_LIVE_NONE;
5613 /* all stack frames are accessible from callee, clear them all */
5614 for (j = 0; j <= cur->curframe; j++) {
5615 struct bpf_func_state *frame = cur->frame[j];
5616 struct bpf_func_state *newframe = new->frame[j];
5618 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
5619 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
5620 frame->stack[i].spilled_ptr.parent =
5621 &newframe->stack[i].spilled_ptr;
5627 /* Return true if it's OK to have the same insn return a different type. */
5628 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
5633 case PTR_TO_SOCKET_OR_NULL:
5640 /* If an instruction was previously used with particular pointer types, then we
5641 * need to be careful to avoid cases such as the below, where it may be ok
5642 * for one branch accessing the pointer, but not ok for the other branch:
5647 * R1 = some_other_valid_ptr;
5650 * R2 = *(u32 *)(R1 + 0);
5652 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
5654 return src != prev && (!reg_type_mismatch_ok(src) ||
5655 !reg_type_mismatch_ok(prev));
5658 static int do_check(struct bpf_verifier_env *env)
5660 struct bpf_verifier_state *state;
5661 struct bpf_insn *insns = env->prog->insnsi;
5662 struct bpf_reg_state *regs;
5663 int insn_cnt = env->prog->len, i;
5664 int insn_processed = 0;
5665 bool do_print_state = false;
5667 env->prev_linfo = NULL;
5669 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
5672 state->curframe = 0;
5673 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
5674 if (!state->frame[0]) {
5678 env->cur_state = state;
5679 init_func_state(env, state->frame[0],
5680 BPF_MAIN_FUNC /* callsite */,
5682 0 /* subprogno, zero == main subprog */);
5685 struct bpf_insn *insn;
5689 if (env->insn_idx >= insn_cnt) {
5690 verbose(env, "invalid insn idx %d insn_cnt %d\n",
5691 env->insn_idx, insn_cnt);
5695 insn = &insns[env->insn_idx];
5696 class = BPF_CLASS(insn->code);
5698 if (++insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
5700 "BPF program is too large. Processed %d insn\n",
5705 err = is_state_visited(env, env->insn_idx);
5709 /* found equivalent state, can prune the search */
5710 if (env->log.level) {
5712 verbose(env, "\nfrom %d to %d: safe\n",
5713 env->prev_insn_idx, env->insn_idx);
5715 verbose(env, "%d: safe\n", env->insn_idx);
5717 goto process_bpf_exit;
5720 if (signal_pending(current))
5726 if (env->log.level > 1 || (env->log.level && do_print_state)) {
5727 if (env->log.level > 1)
5728 verbose(env, "%d:", env->insn_idx);
5730 verbose(env, "\nfrom %d to %d:",
5731 env->prev_insn_idx, env->insn_idx);
5732 print_verifier_state(env, state->frame[state->curframe]);
5733 do_print_state = false;
5736 if (env->log.level) {
5737 const struct bpf_insn_cbs cbs = {
5738 .cb_print = verbose,
5739 .private_data = env,
5742 verbose_linfo(env, env->insn_idx, "; ");
5743 verbose(env, "%d: ", env->insn_idx);
5744 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
5747 if (bpf_prog_is_dev_bound(env->prog->aux)) {
5748 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
5749 env->prev_insn_idx);
5754 regs = cur_regs(env);
5755 env->insn_aux_data[env->insn_idx].seen = true;
5757 if (class == BPF_ALU || class == BPF_ALU64) {
5758 err = check_alu_op(env, insn);
5762 } else if (class == BPF_LDX) {
5763 enum bpf_reg_type *prev_src_type, src_reg_type;
5765 /* check for reserved fields is already done */
5767 /* check src operand */
5768 err = check_reg_arg(env, insn->src_reg, SRC_OP);
5772 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
5776 src_reg_type = regs[insn->src_reg].type;
5778 /* check that memory (src_reg + off) is readable,
5779 * the state of dst_reg will be updated by this func
5781 err = check_mem_access(env, env->insn_idx, insn->src_reg,
5782 insn->off, BPF_SIZE(insn->code),
5783 BPF_READ, insn->dst_reg, false);
5787 prev_src_type = &env->insn_aux_data[env->insn_idx].ptr_type;
5789 if (*prev_src_type == NOT_INIT) {
5791 * dst_reg = *(u32 *)(src_reg + off)
5792 * save type to validate intersecting paths
5794 *prev_src_type = src_reg_type;
5796 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
5797 /* ABuser program is trying to use the same insn
5798 * dst_reg = *(u32*) (src_reg + off)
5799 * with different pointer types:
5800 * src_reg == ctx in one branch and
5801 * src_reg == stack|map in some other branch.
5804 verbose(env, "same insn cannot be used with different pointers\n");
5808 } else if (class == BPF_STX) {
5809 enum bpf_reg_type *prev_dst_type, dst_reg_type;
5811 if (BPF_MODE(insn->code) == BPF_XADD) {
5812 err = check_xadd(env, env->insn_idx, insn);
5819 /* check src1 operand */
5820 err = check_reg_arg(env, insn->src_reg, SRC_OP);
5823 /* check src2 operand */
5824 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
5828 dst_reg_type = regs[insn->dst_reg].type;
5830 /* check that memory (dst_reg + off) is writeable */
5831 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
5832 insn->off, BPF_SIZE(insn->code),
5833 BPF_WRITE, insn->src_reg, false);
5837 prev_dst_type = &env->insn_aux_data[env->insn_idx].ptr_type;
5839 if (*prev_dst_type == NOT_INIT) {
5840 *prev_dst_type = dst_reg_type;
5841 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
5842 verbose(env, "same insn cannot be used with different pointers\n");
5846 } else if (class == BPF_ST) {
5847 if (BPF_MODE(insn->code) != BPF_MEM ||
5848 insn->src_reg != BPF_REG_0) {
5849 verbose(env, "BPF_ST uses reserved fields\n");
5852 /* check src operand */
5853 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
5857 if (is_ctx_reg(env, insn->dst_reg)) {
5858 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
5860 reg_type_str[reg_state(env, insn->dst_reg)->type]);
5864 /* check that memory (dst_reg + off) is writeable */
5865 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
5866 insn->off, BPF_SIZE(insn->code),
5867 BPF_WRITE, -1, false);
5871 } else if (class == BPF_JMP) {
5872 u8 opcode = BPF_OP(insn->code);
5874 if (opcode == BPF_CALL) {
5875 if (BPF_SRC(insn->code) != BPF_K ||
5877 (insn->src_reg != BPF_REG_0 &&
5878 insn->src_reg != BPF_PSEUDO_CALL) ||
5879 insn->dst_reg != BPF_REG_0) {
5880 verbose(env, "BPF_CALL uses reserved fields\n");
5884 if (insn->src_reg == BPF_PSEUDO_CALL)
5885 err = check_func_call(env, insn, &env->insn_idx);
5887 err = check_helper_call(env, insn->imm, env->insn_idx);
5891 } else if (opcode == BPF_JA) {
5892 if (BPF_SRC(insn->code) != BPF_K ||
5894 insn->src_reg != BPF_REG_0 ||
5895 insn->dst_reg != BPF_REG_0) {
5896 verbose(env, "BPF_JA uses reserved fields\n");
5900 env->insn_idx += insn->off + 1;
5903 } else if (opcode == BPF_EXIT) {
5904 if (BPF_SRC(insn->code) != BPF_K ||
5906 insn->src_reg != BPF_REG_0 ||
5907 insn->dst_reg != BPF_REG_0) {
5908 verbose(env, "BPF_EXIT uses reserved fields\n");
5912 if (state->curframe) {
5913 /* exit from nested function */
5914 env->prev_insn_idx = env->insn_idx;
5915 err = prepare_func_exit(env, &env->insn_idx);
5918 do_print_state = true;
5922 err = check_reference_leak(env);
5926 /* eBPF calling convetion is such that R0 is used
5927 * to return the value from eBPF program.
5928 * Make sure that it's readable at this time
5929 * of bpf_exit, which means that program wrote
5930 * something into it earlier
5932 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
5936 if (is_pointer_value(env, BPF_REG_0)) {
5937 verbose(env, "R0 leaks addr as return value\n");
5941 err = check_return_code(env);
5945 err = pop_stack(env, &env->prev_insn_idx,
5952 do_print_state = true;
5956 err = check_cond_jmp_op(env, insn, &env->insn_idx);
5960 } else if (class == BPF_LD) {
5961 u8 mode = BPF_MODE(insn->code);
5963 if (mode == BPF_ABS || mode == BPF_IND) {
5964 err = check_ld_abs(env, insn);
5968 } else if (mode == BPF_IMM) {
5969 err = check_ld_imm(env, insn);
5974 env->insn_aux_data[env->insn_idx].seen = true;
5976 verbose(env, "invalid BPF_LD mode\n");
5980 verbose(env, "unknown insn class %d\n", class);
5987 verbose(env, "processed %d insns (limit %d), stack depth ",
5988 insn_processed, BPF_COMPLEXITY_LIMIT_INSNS);
5989 for (i = 0; i < env->subprog_cnt; i++) {
5990 u32 depth = env->subprog_info[i].stack_depth;
5992 verbose(env, "%d", depth);
5993 if (i + 1 < env->subprog_cnt)
5997 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
6001 static int check_map_prealloc(struct bpf_map *map)
6003 return (map->map_type != BPF_MAP_TYPE_HASH &&
6004 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
6005 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
6006 !(map->map_flags & BPF_F_NO_PREALLOC);
6009 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
6010 struct bpf_map *map,
6011 struct bpf_prog *prog)
6014 /* Make sure that BPF_PROG_TYPE_PERF_EVENT programs only use
6015 * preallocated hash maps, since doing memory allocation
6016 * in overflow_handler can crash depending on where nmi got
6019 if (prog->type == BPF_PROG_TYPE_PERF_EVENT) {
6020 if (!check_map_prealloc(map)) {
6021 verbose(env, "perf_event programs can only use preallocated hash map\n");
6024 if (map->inner_map_meta &&
6025 !check_map_prealloc(map->inner_map_meta)) {
6026 verbose(env, "perf_event programs can only use preallocated inner hash map\n");
6031 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
6032 !bpf_offload_prog_map_match(prog, map)) {
6033 verbose(env, "offload device mismatch between prog and map\n");
6040 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
6042 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
6043 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
6046 /* look for pseudo eBPF instructions that access map FDs and
6047 * replace them with actual map pointers
6049 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env *env)
6051 struct bpf_insn *insn = env->prog->insnsi;
6052 int insn_cnt = env->prog->len;
6055 err = bpf_prog_calc_tag(env->prog);
6059 for (i = 0; i < insn_cnt; i++, insn++) {
6060 if (BPF_CLASS(insn->code) == BPF_LDX &&
6061 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
6062 verbose(env, "BPF_LDX uses reserved fields\n");
6066 if (BPF_CLASS(insn->code) == BPF_STX &&
6067 ((BPF_MODE(insn->code) != BPF_MEM &&
6068 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
6069 verbose(env, "BPF_STX uses reserved fields\n");
6073 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
6074 struct bpf_map *map;
6077 if (i == insn_cnt - 1 || insn[1].code != 0 ||
6078 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
6080 verbose(env, "invalid bpf_ld_imm64 insn\n");
6084 if (insn->src_reg == 0)
6085 /* valid generic load 64-bit imm */
6088 if (insn->src_reg != BPF_PSEUDO_MAP_FD) {
6090 "unrecognized bpf_ld_imm64 insn\n");
6094 f = fdget(insn->imm);
6095 map = __bpf_map_get(f);
6097 verbose(env, "fd %d is not pointing to valid bpf_map\n",
6099 return PTR_ERR(map);
6102 err = check_map_prog_compatibility(env, map, env->prog);
6108 /* store map pointer inside BPF_LD_IMM64 instruction */
6109 insn[0].imm = (u32) (unsigned long) map;
6110 insn[1].imm = ((u64) (unsigned long) map) >> 32;
6112 /* check whether we recorded this map already */
6113 for (j = 0; j < env->used_map_cnt; j++)
6114 if (env->used_maps[j] == map) {
6119 if (env->used_map_cnt >= MAX_USED_MAPS) {
6124 /* hold the map. If the program is rejected by verifier,
6125 * the map will be released by release_maps() or it
6126 * will be used by the valid program until it's unloaded
6127 * and all maps are released in free_used_maps()
6129 map = bpf_map_inc(map, false);
6132 return PTR_ERR(map);
6134 env->used_maps[env->used_map_cnt++] = map;
6136 if (bpf_map_is_cgroup_storage(map) &&
6137 bpf_cgroup_storage_assign(env->prog, map)) {
6138 verbose(env, "only one cgroup storage of each type is allowed\n");
6150 /* Basic sanity check before we invest more work here. */
6151 if (!bpf_opcode_in_insntable(insn->code)) {
6152 verbose(env, "unknown opcode %02x\n", insn->code);
6157 /* now all pseudo BPF_LD_IMM64 instructions load valid
6158 * 'struct bpf_map *' into a register instead of user map_fd.
6159 * These pointers will be used later by verifier to validate map access.
6164 /* drop refcnt of maps used by the rejected program */
6165 static void release_maps(struct bpf_verifier_env *env)
6167 enum bpf_cgroup_storage_type stype;
6170 for_each_cgroup_storage_type(stype) {
6171 if (!env->prog->aux->cgroup_storage[stype])
6173 bpf_cgroup_storage_release(env->prog,
6174 env->prog->aux->cgroup_storage[stype]);
6177 for (i = 0; i < env->used_map_cnt; i++)
6178 bpf_map_put(env->used_maps[i]);
6181 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
6182 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
6184 struct bpf_insn *insn = env->prog->insnsi;
6185 int insn_cnt = env->prog->len;
6188 for (i = 0; i < insn_cnt; i++, insn++)
6189 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
6193 /* single env->prog->insni[off] instruction was replaced with the range
6194 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
6195 * [0, off) and [off, end) to new locations, so the patched range stays zero
6197 static int adjust_insn_aux_data(struct bpf_verifier_env *env, u32 prog_len,
6200 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
6205 new_data = vzalloc(array_size(prog_len,
6206 sizeof(struct bpf_insn_aux_data)));
6209 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
6210 memcpy(new_data + off + cnt - 1, old_data + off,
6211 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
6212 for (i = off; i < off + cnt - 1; i++)
6213 new_data[i].seen = true;
6214 env->insn_aux_data = new_data;
6219 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
6225 /* NOTE: fake 'exit' subprog should be updated as well. */
6226 for (i = 0; i <= env->subprog_cnt; i++) {
6227 if (env->subprog_info[i].start <= off)
6229 env->subprog_info[i].start += len - 1;
6233 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
6234 const struct bpf_insn *patch, u32 len)
6236 struct bpf_prog *new_prog;
6238 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
6241 if (adjust_insn_aux_data(env, new_prog->len, off, len))
6243 adjust_subprog_starts(env, off, len);
6247 /* The verifier does more data flow analysis than llvm and will not
6248 * explore branches that are dead at run time. Malicious programs can
6249 * have dead code too. Therefore replace all dead at-run-time code
6252 * Just nops are not optimal, e.g. if they would sit at the end of the
6253 * program and through another bug we would manage to jump there, then
6254 * we'd execute beyond program memory otherwise. Returning exception
6255 * code also wouldn't work since we can have subprogs where the dead
6256 * code could be located.
6258 static void sanitize_dead_code(struct bpf_verifier_env *env)
6260 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
6261 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
6262 struct bpf_insn *insn = env->prog->insnsi;
6263 const int insn_cnt = env->prog->len;
6266 for (i = 0; i < insn_cnt; i++) {
6267 if (aux_data[i].seen)
6269 memcpy(insn + i, &trap, sizeof(trap));
6273 /* convert load instructions that access fields of a context type into a
6274 * sequence of instructions that access fields of the underlying structure:
6275 * struct __sk_buff -> struct sk_buff
6276 * struct bpf_sock_ops -> struct sock
6278 static int convert_ctx_accesses(struct bpf_verifier_env *env)
6280 const struct bpf_verifier_ops *ops = env->ops;
6281 int i, cnt, size, ctx_field_size, delta = 0;
6282 const int insn_cnt = env->prog->len;
6283 struct bpf_insn insn_buf[16], *insn;
6284 u32 target_size, size_default, off;
6285 struct bpf_prog *new_prog;
6286 enum bpf_access_type type;
6287 bool is_narrower_load;
6289 if (ops->gen_prologue || env->seen_direct_write) {
6290 if (!ops->gen_prologue) {
6291 verbose(env, "bpf verifier is misconfigured\n");
6294 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
6296 if (cnt >= ARRAY_SIZE(insn_buf)) {
6297 verbose(env, "bpf verifier is misconfigured\n");
6300 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
6304 env->prog = new_prog;
6309 if (bpf_prog_is_dev_bound(env->prog->aux))
6312 insn = env->prog->insnsi + delta;
6314 for (i = 0; i < insn_cnt; i++, insn++) {
6315 bpf_convert_ctx_access_t convert_ctx_access;
6317 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
6318 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
6319 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
6320 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
6322 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
6323 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
6324 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
6325 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
6330 if (type == BPF_WRITE &&
6331 env->insn_aux_data[i + delta].sanitize_stack_off) {
6332 struct bpf_insn patch[] = {
6333 /* Sanitize suspicious stack slot with zero.
6334 * There are no memory dependencies for this store,
6335 * since it's only using frame pointer and immediate
6338 BPF_ST_MEM(BPF_DW, BPF_REG_FP,
6339 env->insn_aux_data[i + delta].sanitize_stack_off,
6341 /* the original STX instruction will immediately
6342 * overwrite the same stack slot with appropriate value
6347 cnt = ARRAY_SIZE(patch);
6348 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
6353 env->prog = new_prog;
6354 insn = new_prog->insnsi + i + delta;
6358 switch (env->insn_aux_data[i + delta].ptr_type) {
6360 if (!ops->convert_ctx_access)
6362 convert_ctx_access = ops->convert_ctx_access;
6365 convert_ctx_access = bpf_sock_convert_ctx_access;
6371 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
6372 size = BPF_LDST_BYTES(insn);
6374 /* If the read access is a narrower load of the field,
6375 * convert to a 4/8-byte load, to minimum program type specific
6376 * convert_ctx_access changes. If conversion is successful,
6377 * we will apply proper mask to the result.
6379 is_narrower_load = size < ctx_field_size;
6380 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
6382 if (is_narrower_load) {
6385 if (type == BPF_WRITE) {
6386 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
6391 if (ctx_field_size == 4)
6393 else if (ctx_field_size == 8)
6396 insn->off = off & ~(size_default - 1);
6397 insn->code = BPF_LDX | BPF_MEM | size_code;
6401 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
6403 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
6404 (ctx_field_size && !target_size)) {
6405 verbose(env, "bpf verifier is misconfigured\n");
6409 if (is_narrower_load && size < target_size) {
6410 u8 shift = (off & (size_default - 1)) * 8;
6412 if (ctx_field_size <= 4) {
6414 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
6417 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
6418 (1 << size * 8) - 1);
6421 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
6424 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
6425 (1 << size * 8) - 1);
6429 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
6435 /* keep walking new program and skip insns we just inserted */
6436 env->prog = new_prog;
6437 insn = new_prog->insnsi + i + delta;
6443 static int jit_subprogs(struct bpf_verifier_env *env)
6445 struct bpf_prog *prog = env->prog, **func, *tmp;
6446 int i, j, subprog_start, subprog_end = 0, len, subprog;
6447 struct bpf_insn *insn;
6451 if (env->subprog_cnt <= 1)
6454 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
6455 if (insn->code != (BPF_JMP | BPF_CALL) ||
6456 insn->src_reg != BPF_PSEUDO_CALL)
6458 /* Upon error here we cannot fall back to interpreter but
6459 * need a hard reject of the program. Thus -EFAULT is
6460 * propagated in any case.
6462 subprog = find_subprog(env, i + insn->imm + 1);
6464 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
6468 /* temporarily remember subprog id inside insn instead of
6469 * aux_data, since next loop will split up all insns into funcs
6471 insn->off = subprog;
6472 /* remember original imm in case JIT fails and fallback
6473 * to interpreter will be needed
6475 env->insn_aux_data[i].call_imm = insn->imm;
6476 /* point imm to __bpf_call_base+1 from JITs point of view */
6480 err = bpf_prog_alloc_jited_linfo(prog);
6485 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
6489 for (i = 0; i < env->subprog_cnt; i++) {
6490 subprog_start = subprog_end;
6491 subprog_end = env->subprog_info[i + 1].start;
6493 len = subprog_end - subprog_start;
6494 func[i] = bpf_prog_alloc(bpf_prog_size(len), GFP_USER);
6497 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
6498 len * sizeof(struct bpf_insn));
6499 func[i]->type = prog->type;
6501 if (bpf_prog_calc_tag(func[i]))
6503 func[i]->is_func = 1;
6504 func[i]->aux->func_idx = i;
6505 /* the btf and func_info will be freed only at prog->aux */
6506 func[i]->aux->btf = prog->aux->btf;
6507 func[i]->aux->func_info = prog->aux->func_info;
6509 /* Use bpf_prog_F_tag to indicate functions in stack traces.
6510 * Long term would need debug info to populate names
6512 func[i]->aux->name[0] = 'F';
6513 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
6514 func[i]->jit_requested = 1;
6515 func[i]->aux->linfo = prog->aux->linfo;
6516 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
6517 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
6518 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
6519 func[i] = bpf_int_jit_compile(func[i]);
6520 if (!func[i]->jited) {
6526 /* at this point all bpf functions were successfully JITed
6527 * now populate all bpf_calls with correct addresses and
6528 * run last pass of JIT
6530 for (i = 0; i < env->subprog_cnt; i++) {
6531 insn = func[i]->insnsi;
6532 for (j = 0; j < func[i]->len; j++, insn++) {
6533 if (insn->code != (BPF_JMP | BPF_CALL) ||
6534 insn->src_reg != BPF_PSEUDO_CALL)
6536 subprog = insn->off;
6537 insn->imm = (u64 (*)(u64, u64, u64, u64, u64))
6538 func[subprog]->bpf_func -
6542 /* we use the aux data to keep a list of the start addresses
6543 * of the JITed images for each function in the program
6545 * for some architectures, such as powerpc64, the imm field
6546 * might not be large enough to hold the offset of the start
6547 * address of the callee's JITed image from __bpf_call_base
6549 * in such cases, we can lookup the start address of a callee
6550 * by using its subprog id, available from the off field of
6551 * the call instruction, as an index for this list
6553 func[i]->aux->func = func;
6554 func[i]->aux->func_cnt = env->subprog_cnt;
6556 for (i = 0; i < env->subprog_cnt; i++) {
6557 old_bpf_func = func[i]->bpf_func;
6558 tmp = bpf_int_jit_compile(func[i]);
6559 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
6560 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
6567 /* finally lock prog and jit images for all functions and
6570 for (i = 0; i < env->subprog_cnt; i++) {
6571 bpf_prog_lock_ro(func[i]);
6572 bpf_prog_kallsyms_add(func[i]);
6575 /* Last step: make now unused interpreter insns from main
6576 * prog consistent for later dump requests, so they can
6577 * later look the same as if they were interpreted only.
6579 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
6580 if (insn->code != (BPF_JMP | BPF_CALL) ||
6581 insn->src_reg != BPF_PSEUDO_CALL)
6583 insn->off = env->insn_aux_data[i].call_imm;
6584 subprog = find_subprog(env, i + insn->off + 1);
6585 insn->imm = subprog;
6589 prog->bpf_func = func[0]->bpf_func;
6590 prog->aux->func = func;
6591 prog->aux->func_cnt = env->subprog_cnt;
6592 bpf_prog_free_unused_jited_linfo(prog);
6595 for (i = 0; i < env->subprog_cnt; i++)
6597 bpf_jit_free(func[i]);
6600 /* cleanup main prog to be interpreted */
6601 prog->jit_requested = 0;
6602 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
6603 if (insn->code != (BPF_JMP | BPF_CALL) ||
6604 insn->src_reg != BPF_PSEUDO_CALL)
6607 insn->imm = env->insn_aux_data[i].call_imm;
6609 bpf_prog_free_jited_linfo(prog);
6613 static int fixup_call_args(struct bpf_verifier_env *env)
6615 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
6616 struct bpf_prog *prog = env->prog;
6617 struct bpf_insn *insn = prog->insnsi;
6622 if (env->prog->jit_requested &&
6623 !bpf_prog_is_dev_bound(env->prog->aux)) {
6624 err = jit_subprogs(env);
6630 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
6631 for (i = 0; i < prog->len; i++, insn++) {
6632 if (insn->code != (BPF_JMP | BPF_CALL) ||
6633 insn->src_reg != BPF_PSEUDO_CALL)
6635 depth = get_callee_stack_depth(env, insn, i);
6638 bpf_patch_call_args(insn, depth);
6645 /* fixup insn->imm field of bpf_call instructions
6646 * and inline eligible helpers as explicit sequence of BPF instructions
6648 * this function is called after eBPF program passed verification
6650 static int fixup_bpf_calls(struct bpf_verifier_env *env)
6652 struct bpf_prog *prog = env->prog;
6653 struct bpf_insn *insn = prog->insnsi;
6654 const struct bpf_func_proto *fn;
6655 const int insn_cnt = prog->len;
6656 const struct bpf_map_ops *ops;
6657 struct bpf_insn_aux_data *aux;
6658 struct bpf_insn insn_buf[16];
6659 struct bpf_prog *new_prog;
6660 struct bpf_map *map_ptr;
6661 int i, cnt, delta = 0;
6663 for (i = 0; i < insn_cnt; i++, insn++) {
6664 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
6665 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
6666 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
6667 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
6668 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
6669 struct bpf_insn mask_and_div[] = {
6670 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
6672 BPF_JMP_IMM(BPF_JNE, insn->src_reg, 0, 2),
6673 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
6674 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
6677 struct bpf_insn mask_and_mod[] = {
6678 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
6679 /* Rx mod 0 -> Rx */
6680 BPF_JMP_IMM(BPF_JEQ, insn->src_reg, 0, 1),
6683 struct bpf_insn *patchlet;
6685 if (insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
6686 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
6687 patchlet = mask_and_div + (is64 ? 1 : 0);
6688 cnt = ARRAY_SIZE(mask_and_div) - (is64 ? 1 : 0);
6690 patchlet = mask_and_mod + (is64 ? 1 : 0);
6691 cnt = ARRAY_SIZE(mask_and_mod) - (is64 ? 1 : 0);
6694 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
6699 env->prog = prog = new_prog;
6700 insn = new_prog->insnsi + i + delta;
6704 if (BPF_CLASS(insn->code) == BPF_LD &&
6705 (BPF_MODE(insn->code) == BPF_ABS ||
6706 BPF_MODE(insn->code) == BPF_IND)) {
6707 cnt = env->ops->gen_ld_abs(insn, insn_buf);
6708 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
6709 verbose(env, "bpf verifier is misconfigured\n");
6713 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
6718 env->prog = prog = new_prog;
6719 insn = new_prog->insnsi + i + delta;
6723 if (insn->code != (BPF_JMP | BPF_CALL))
6725 if (insn->src_reg == BPF_PSEUDO_CALL)
6728 if (insn->imm == BPF_FUNC_get_route_realm)
6729 prog->dst_needed = 1;
6730 if (insn->imm == BPF_FUNC_get_prandom_u32)
6731 bpf_user_rnd_init_once();
6732 if (insn->imm == BPF_FUNC_override_return)
6733 prog->kprobe_override = 1;
6734 if (insn->imm == BPF_FUNC_tail_call) {
6735 /* If we tail call into other programs, we
6736 * cannot make any assumptions since they can
6737 * be replaced dynamically during runtime in
6738 * the program array.
6740 prog->cb_access = 1;
6741 env->prog->aux->stack_depth = MAX_BPF_STACK;
6742 env->prog->aux->max_pkt_offset = MAX_PACKET_OFF;
6744 /* mark bpf_tail_call as different opcode to avoid
6745 * conditional branch in the interpeter for every normal
6746 * call and to prevent accidental JITing by JIT compiler
6747 * that doesn't support bpf_tail_call yet
6750 insn->code = BPF_JMP | BPF_TAIL_CALL;
6752 aux = &env->insn_aux_data[i + delta];
6753 if (!bpf_map_ptr_unpriv(aux))
6756 /* instead of changing every JIT dealing with tail_call
6757 * emit two extra insns:
6758 * if (index >= max_entries) goto out;
6759 * index &= array->index_mask;
6760 * to avoid out-of-bounds cpu speculation
6762 if (bpf_map_ptr_poisoned(aux)) {
6763 verbose(env, "tail_call abusing map_ptr\n");
6767 map_ptr = BPF_MAP_PTR(aux->map_state);
6768 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
6769 map_ptr->max_entries, 2);
6770 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
6771 container_of(map_ptr,
6774 insn_buf[2] = *insn;
6776 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
6781 env->prog = prog = new_prog;
6782 insn = new_prog->insnsi + i + delta;
6786 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
6787 * and other inlining handlers are currently limited to 64 bit
6790 if (prog->jit_requested && BITS_PER_LONG == 64 &&
6791 (insn->imm == BPF_FUNC_map_lookup_elem ||
6792 insn->imm == BPF_FUNC_map_update_elem ||
6793 insn->imm == BPF_FUNC_map_delete_elem ||
6794 insn->imm == BPF_FUNC_map_push_elem ||
6795 insn->imm == BPF_FUNC_map_pop_elem ||
6796 insn->imm == BPF_FUNC_map_peek_elem)) {
6797 aux = &env->insn_aux_data[i + delta];
6798 if (bpf_map_ptr_poisoned(aux))
6799 goto patch_call_imm;
6801 map_ptr = BPF_MAP_PTR(aux->map_state);
6803 if (insn->imm == BPF_FUNC_map_lookup_elem &&
6804 ops->map_gen_lookup) {
6805 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
6806 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
6807 verbose(env, "bpf verifier is misconfigured\n");
6811 new_prog = bpf_patch_insn_data(env, i + delta,
6817 env->prog = prog = new_prog;
6818 insn = new_prog->insnsi + i + delta;
6822 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
6823 (void *(*)(struct bpf_map *map, void *key))NULL));
6824 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
6825 (int (*)(struct bpf_map *map, void *key))NULL));
6826 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
6827 (int (*)(struct bpf_map *map, void *key, void *value,
6829 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
6830 (int (*)(struct bpf_map *map, void *value,
6832 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
6833 (int (*)(struct bpf_map *map, void *value))NULL));
6834 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
6835 (int (*)(struct bpf_map *map, void *value))NULL));
6837 switch (insn->imm) {
6838 case BPF_FUNC_map_lookup_elem:
6839 insn->imm = BPF_CAST_CALL(ops->map_lookup_elem) -
6842 case BPF_FUNC_map_update_elem:
6843 insn->imm = BPF_CAST_CALL(ops->map_update_elem) -
6846 case BPF_FUNC_map_delete_elem:
6847 insn->imm = BPF_CAST_CALL(ops->map_delete_elem) -
6850 case BPF_FUNC_map_push_elem:
6851 insn->imm = BPF_CAST_CALL(ops->map_push_elem) -
6854 case BPF_FUNC_map_pop_elem:
6855 insn->imm = BPF_CAST_CALL(ops->map_pop_elem) -
6858 case BPF_FUNC_map_peek_elem:
6859 insn->imm = BPF_CAST_CALL(ops->map_peek_elem) -
6864 goto patch_call_imm;
6868 fn = env->ops->get_func_proto(insn->imm, env->prog);
6869 /* all functions that have prototype and verifier allowed
6870 * programs to call them, must be real in-kernel functions
6874 "kernel subsystem misconfigured func %s#%d\n",
6875 func_id_name(insn->imm), insn->imm);
6878 insn->imm = fn->func - __bpf_call_base;
6884 static void free_states(struct bpf_verifier_env *env)
6886 struct bpf_verifier_state_list *sl, *sln;
6889 if (!env->explored_states)
6892 for (i = 0; i < env->prog->len; i++) {
6893 sl = env->explored_states[i];
6896 while (sl != STATE_LIST_MARK) {
6898 free_verifier_state(&sl->state, false);
6904 kfree(env->explored_states);
6907 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr,
6908 union bpf_attr __user *uattr)
6910 struct bpf_verifier_env *env;
6911 struct bpf_verifier_log *log;
6914 /* no program is valid */
6915 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
6918 /* 'struct bpf_verifier_env' can be global, but since it's not small,
6919 * allocate/free it every time bpf_check() is called
6921 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
6926 env->insn_aux_data =
6927 vzalloc(array_size(sizeof(struct bpf_insn_aux_data),
6930 if (!env->insn_aux_data)
6933 env->ops = bpf_verifier_ops[env->prog->type];
6935 /* grab the mutex to protect few globals used by verifier */
6936 mutex_lock(&bpf_verifier_lock);
6938 if (attr->log_level || attr->log_buf || attr->log_size) {
6939 /* user requested verbose verifier output
6940 * and supplied buffer to store the verification trace
6942 log->level = attr->log_level;
6943 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
6944 log->len_total = attr->log_size;
6947 /* log attributes have to be sane */
6948 if (log->len_total < 128 || log->len_total > UINT_MAX >> 8 ||
6949 !log->level || !log->ubuf)
6953 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
6954 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
6955 env->strict_alignment = true;
6956 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
6957 env->strict_alignment = false;
6959 ret = replace_map_fd_with_map_ptr(env);
6961 goto skip_full_check;
6963 if (bpf_prog_is_dev_bound(env->prog->aux)) {
6964 ret = bpf_prog_offload_verifier_prep(env->prog);
6966 goto skip_full_check;
6969 env->explored_states = kcalloc(env->prog->len,
6970 sizeof(struct bpf_verifier_state_list *),
6973 if (!env->explored_states)
6974 goto skip_full_check;
6976 env->allow_ptr_leaks = capable(CAP_SYS_ADMIN);
6978 ret = check_subprogs(env);
6980 goto skip_full_check;
6982 ret = check_btf_info(env, attr, uattr);
6984 goto skip_full_check;
6986 ret = check_cfg(env);
6988 goto skip_full_check;
6990 ret = do_check(env);
6991 if (env->cur_state) {
6992 free_verifier_state(env->cur_state, true);
6993 env->cur_state = NULL;
6996 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
6997 ret = bpf_prog_offload_finalize(env);
7000 while (!pop_stack(env, NULL, NULL));
7004 ret = check_max_stack_depth(env);
7006 /* instruction rewrites happen after this point */
7008 sanitize_dead_code(env);
7011 /* program is valid, convert *(u32*)(ctx + off) accesses */
7012 ret = convert_ctx_accesses(env);
7015 ret = fixup_bpf_calls(env);
7018 ret = fixup_call_args(env);
7020 if (log->level && bpf_verifier_log_full(log))
7022 if (log->level && !log->ubuf) {
7024 goto err_release_maps;
7027 if (ret == 0 && env->used_map_cnt) {
7028 /* if program passed verifier, update used_maps in bpf_prog_info */
7029 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
7030 sizeof(env->used_maps[0]),
7033 if (!env->prog->aux->used_maps) {
7035 goto err_release_maps;
7038 memcpy(env->prog->aux->used_maps, env->used_maps,
7039 sizeof(env->used_maps[0]) * env->used_map_cnt);
7040 env->prog->aux->used_map_cnt = env->used_map_cnt;
7042 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
7043 * bpf_ld_imm64 instructions
7045 convert_pseudo_ld_imm64(env);
7049 adjust_btf_func(env);
7052 if (!env->prog->aux->used_maps)
7053 /* if we didn't copy map pointers into bpf_prog_info, release
7054 * them now. Otherwise free_used_maps() will release them.
7059 mutex_unlock(&bpf_verifier_lock);
7060 vfree(env->insn_aux_data);