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 <linux/kernel.h>
15 #include <linux/types.h>
16 #include <linux/slab.h>
17 #include <linux/bpf.h>
18 #include <linux/bpf_verifier.h>
19 #include <linux/filter.h>
20 #include <net/netlink.h>
21 #include <linux/file.h>
22 #include <linux/vmalloc.h>
23 #include <linux/stringify.h>
24 #include <linux/bsearch.h>
25 #include <linux/sort.h>
26 #include <linux/perf_event.h>
30 static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
31 #define BPF_PROG_TYPE(_id, _name) \
32 [_id] = & _name ## _verifier_ops,
33 #define BPF_MAP_TYPE(_id, _ops)
34 #include <linux/bpf_types.h>
39 /* bpf_check() is a static code analyzer that walks eBPF program
40 * instruction by instruction and updates register/stack state.
41 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
43 * The first pass is depth-first-search to check that the program is a DAG.
44 * It rejects the following programs:
45 * - larger than BPF_MAXINSNS insns
46 * - if loop is present (detected via back-edge)
47 * - unreachable insns exist (shouldn't be a forest. program = one function)
48 * - out of bounds or malformed jumps
49 * The second pass is all possible path descent from the 1st insn.
50 * Since it's analyzing all pathes through the program, the length of the
51 * analysis is limited to 64k insn, which may be hit even if total number of
52 * insn is less then 4K, but there are too many branches that change stack/regs.
53 * Number of 'branches to be analyzed' is limited to 1k
55 * On entry to each instruction, each register has a type, and the instruction
56 * changes the types of the registers depending on instruction semantics.
57 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
60 * All registers are 64-bit.
61 * R0 - return register
62 * R1-R5 argument passing registers
63 * R6-R9 callee saved registers
64 * R10 - frame pointer read-only
66 * At the start of BPF program the register R1 contains a pointer to bpf_context
67 * and has type PTR_TO_CTX.
69 * Verifier tracks arithmetic operations on pointers in case:
70 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
71 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
72 * 1st insn copies R10 (which has FRAME_PTR) type into R1
73 * and 2nd arithmetic instruction is pattern matched to recognize
74 * that it wants to construct a pointer to some element within stack.
75 * So after 2nd insn, the register R1 has type PTR_TO_STACK
76 * (and -20 constant is saved for further stack bounds checking).
77 * Meaning that this reg is a pointer to stack plus known immediate constant.
79 * Most of the time the registers have SCALAR_VALUE type, which
80 * means the register has some value, but it's not a valid pointer.
81 * (like pointer plus pointer becomes SCALAR_VALUE type)
83 * When verifier sees load or store instructions the type of base register
84 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK, PTR_TO_SOCKET. These are
85 * four pointer types recognized by check_mem_access() function.
87 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
88 * and the range of [ptr, ptr + map's value_size) is accessible.
90 * registers used to pass values to function calls are checked against
91 * function argument constraints.
93 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
94 * It means that the register type passed to this function must be
95 * PTR_TO_STACK and it will be used inside the function as
96 * 'pointer to map element key'
98 * For example the argument constraints for bpf_map_lookup_elem():
99 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
100 * .arg1_type = ARG_CONST_MAP_PTR,
101 * .arg2_type = ARG_PTR_TO_MAP_KEY,
103 * ret_type says that this function returns 'pointer to map elem value or null'
104 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
105 * 2nd argument should be a pointer to stack, which will be used inside
106 * the helper function as a pointer to map element key.
108 * On the kernel side the helper function looks like:
109 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
111 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
112 * void *key = (void *) (unsigned long) r2;
115 * here kernel can access 'key' and 'map' pointers safely, knowing that
116 * [key, key + map->key_size) bytes are valid and were initialized on
117 * the stack of eBPF program.
120 * Corresponding eBPF program may look like:
121 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
122 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
123 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
124 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
125 * here verifier looks at prototype of map_lookup_elem() and sees:
126 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
127 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
129 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
130 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
131 * and were initialized prior to this call.
132 * If it's ok, then verifier allows this BPF_CALL insn and looks at
133 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
134 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
135 * returns ether pointer to map value or NULL.
137 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
138 * insn, the register holding that pointer in the true branch changes state to
139 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
140 * branch. See check_cond_jmp_op().
142 * After the call R0 is set to return type of the function and registers R1-R5
143 * are set to NOT_INIT to indicate that they are no longer readable.
145 * The following reference types represent a potential reference to a kernel
146 * resource which, after first being allocated, must be checked and freed by
148 * - PTR_TO_SOCKET_OR_NULL, PTR_TO_SOCKET
150 * When the verifier sees a helper call return a reference type, it allocates a
151 * pointer id for the reference and stores it in the current function state.
152 * Similar to the way that PTR_TO_MAP_VALUE_OR_NULL is converted into
153 * PTR_TO_MAP_VALUE, PTR_TO_SOCKET_OR_NULL becomes PTR_TO_SOCKET when the type
154 * passes through a NULL-check conditional. For the branch wherein the state is
155 * changed to CONST_IMM, the verifier releases the reference.
157 * For each helper function that allocates a reference, such as
158 * bpf_sk_lookup_tcp(), there is a corresponding release function, such as
159 * bpf_sk_release(). When a reference type passes into the release function,
160 * the verifier also releases the reference. If any unchecked or unreleased
161 * reference remains at the end of the program, the verifier rejects it.
164 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
165 struct bpf_verifier_stack_elem {
166 /* verifer state is 'st'
167 * before processing instruction 'insn_idx'
168 * and after processing instruction 'prev_insn_idx'
170 struct bpf_verifier_state st;
173 struct bpf_verifier_stack_elem *next;
176 #define BPF_COMPLEXITY_LIMIT_INSNS 131072
177 #define BPF_COMPLEXITY_LIMIT_STACK 1024
179 #define BPF_MAP_PTR_UNPRIV 1UL
180 #define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \
181 POISON_POINTER_DELTA))
182 #define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
184 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
186 return BPF_MAP_PTR(aux->map_state) == BPF_MAP_PTR_POISON;
189 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
191 return aux->map_state & BPF_MAP_PTR_UNPRIV;
194 static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
195 const struct bpf_map *map, bool unpriv)
197 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON & BPF_MAP_PTR_UNPRIV);
198 unpriv |= bpf_map_ptr_unpriv(aux);
199 aux->map_state = (unsigned long)map |
200 (unpriv ? BPF_MAP_PTR_UNPRIV : 0UL);
203 struct bpf_call_arg_meta {
204 struct bpf_map *map_ptr;
209 s64 msize_smax_value;
210 u64 msize_umax_value;
214 static DEFINE_MUTEX(bpf_verifier_lock);
216 void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
221 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
223 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
224 "verifier log line truncated - local buffer too short\n");
226 n = min(log->len_total - log->len_used - 1, n);
229 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
235 /* log_level controls verbosity level of eBPF verifier.
236 * bpf_verifier_log_write() is used to dump the verification trace to the log,
237 * so the user can figure out what's wrong with the program
239 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
240 const char *fmt, ...)
244 if (!bpf_verifier_log_needed(&env->log))
248 bpf_verifier_vlog(&env->log, fmt, args);
251 EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
253 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
255 struct bpf_verifier_env *env = private_data;
258 if (!bpf_verifier_log_needed(&env->log))
262 bpf_verifier_vlog(&env->log, fmt, args);
266 static bool type_is_pkt_pointer(enum bpf_reg_type type)
268 return type == PTR_TO_PACKET ||
269 type == PTR_TO_PACKET_META;
272 static bool reg_type_may_be_null(enum bpf_reg_type type)
274 return type == PTR_TO_MAP_VALUE_OR_NULL ||
275 type == PTR_TO_SOCKET_OR_NULL;
278 static bool type_is_refcounted(enum bpf_reg_type type)
280 return type == PTR_TO_SOCKET;
283 static bool type_is_refcounted_or_null(enum bpf_reg_type type)
285 return type == PTR_TO_SOCKET || type == PTR_TO_SOCKET_OR_NULL;
288 static bool reg_is_refcounted(const struct bpf_reg_state *reg)
290 return type_is_refcounted(reg->type);
293 static bool reg_is_refcounted_or_null(const struct bpf_reg_state *reg)
295 return type_is_refcounted_or_null(reg->type);
298 static bool arg_type_is_refcounted(enum bpf_arg_type type)
300 return type == ARG_PTR_TO_SOCKET;
303 /* Determine whether the function releases some resources allocated by another
304 * function call. The first reference type argument will be assumed to be
305 * released by release_reference().
307 static bool is_release_function(enum bpf_func_id func_id)
309 return func_id == BPF_FUNC_sk_release;
312 /* string representation of 'enum bpf_reg_type' */
313 static const char * const reg_type_str[] = {
315 [SCALAR_VALUE] = "inv",
316 [PTR_TO_CTX] = "ctx",
317 [CONST_PTR_TO_MAP] = "map_ptr",
318 [PTR_TO_MAP_VALUE] = "map_value",
319 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
320 [PTR_TO_STACK] = "fp",
321 [PTR_TO_PACKET] = "pkt",
322 [PTR_TO_PACKET_META] = "pkt_meta",
323 [PTR_TO_PACKET_END] = "pkt_end",
324 [PTR_TO_FLOW_KEYS] = "flow_keys",
325 [PTR_TO_SOCKET] = "sock",
326 [PTR_TO_SOCKET_OR_NULL] = "sock_or_null",
329 static char slot_type_char[] = {
330 [STACK_INVALID] = '?',
336 static void print_liveness(struct bpf_verifier_env *env,
337 enum bpf_reg_liveness live)
339 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN))
341 if (live & REG_LIVE_READ)
343 if (live & REG_LIVE_WRITTEN)
347 static struct bpf_func_state *func(struct bpf_verifier_env *env,
348 const struct bpf_reg_state *reg)
350 struct bpf_verifier_state *cur = env->cur_state;
352 return cur->frame[reg->frameno];
355 static void print_verifier_state(struct bpf_verifier_env *env,
356 const struct bpf_func_state *state)
358 const struct bpf_reg_state *reg;
363 verbose(env, " frame%d:", state->frameno);
364 for (i = 0; i < MAX_BPF_REG; i++) {
365 reg = &state->regs[i];
369 verbose(env, " R%d", i);
370 print_liveness(env, reg->live);
371 verbose(env, "=%s", reg_type_str[t]);
372 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
373 tnum_is_const(reg->var_off)) {
374 /* reg->off should be 0 for SCALAR_VALUE */
375 verbose(env, "%lld", reg->var_off.value + reg->off);
376 if (t == PTR_TO_STACK)
377 verbose(env, ",call_%d", func(env, reg)->callsite);
379 verbose(env, "(id=%d", reg->id);
380 if (t != SCALAR_VALUE)
381 verbose(env, ",off=%d", reg->off);
382 if (type_is_pkt_pointer(t))
383 verbose(env, ",r=%d", reg->range);
384 else if (t == CONST_PTR_TO_MAP ||
385 t == PTR_TO_MAP_VALUE ||
386 t == PTR_TO_MAP_VALUE_OR_NULL)
387 verbose(env, ",ks=%d,vs=%d",
388 reg->map_ptr->key_size,
389 reg->map_ptr->value_size);
390 if (tnum_is_const(reg->var_off)) {
391 /* Typically an immediate SCALAR_VALUE, but
392 * could be a pointer whose offset is too big
395 verbose(env, ",imm=%llx", reg->var_off.value);
397 if (reg->smin_value != reg->umin_value &&
398 reg->smin_value != S64_MIN)
399 verbose(env, ",smin_value=%lld",
400 (long long)reg->smin_value);
401 if (reg->smax_value != reg->umax_value &&
402 reg->smax_value != S64_MAX)
403 verbose(env, ",smax_value=%lld",
404 (long long)reg->smax_value);
405 if (reg->umin_value != 0)
406 verbose(env, ",umin_value=%llu",
407 (unsigned long long)reg->umin_value);
408 if (reg->umax_value != U64_MAX)
409 verbose(env, ",umax_value=%llu",
410 (unsigned long long)reg->umax_value);
411 if (!tnum_is_unknown(reg->var_off)) {
414 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
415 verbose(env, ",var_off=%s", tn_buf);
421 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
422 char types_buf[BPF_REG_SIZE + 1];
426 for (j = 0; j < BPF_REG_SIZE; j++) {
427 if (state->stack[i].slot_type[j] != STACK_INVALID)
429 types_buf[j] = slot_type_char[
430 state->stack[i].slot_type[j]];
432 types_buf[BPF_REG_SIZE] = 0;
435 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
436 print_liveness(env, state->stack[i].spilled_ptr.live);
437 if (state->stack[i].slot_type[0] == STACK_SPILL)
439 reg_type_str[state->stack[i].spilled_ptr.type]);
441 verbose(env, "=%s", types_buf);
443 if (state->acquired_refs && state->refs[0].id) {
444 verbose(env, " refs=%d", state->refs[0].id);
445 for (i = 1; i < state->acquired_refs; i++)
446 if (state->refs[i].id)
447 verbose(env, ",%d", state->refs[i].id);
452 #define COPY_STATE_FN(NAME, COUNT, FIELD, SIZE) \
453 static int copy_##NAME##_state(struct bpf_func_state *dst, \
454 const struct bpf_func_state *src) \
458 if (WARN_ON_ONCE(dst->COUNT < src->COUNT)) { \
459 /* internal bug, make state invalid to reject the program */ \
460 memset(dst, 0, sizeof(*dst)); \
463 memcpy(dst->FIELD, src->FIELD, \
464 sizeof(*src->FIELD) * (src->COUNT / SIZE)); \
467 /* copy_reference_state() */
468 COPY_STATE_FN(reference, acquired_refs, refs, 1)
469 /* copy_stack_state() */
470 COPY_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
473 #define REALLOC_STATE_FN(NAME, COUNT, FIELD, SIZE) \
474 static int realloc_##NAME##_state(struct bpf_func_state *state, int size, \
477 u32 old_size = state->COUNT; \
478 struct bpf_##NAME##_state *new_##FIELD; \
479 int slot = size / SIZE; \
481 if (size <= old_size || !size) { \
484 state->COUNT = slot * SIZE; \
485 if (!size && old_size) { \
486 kfree(state->FIELD); \
487 state->FIELD = NULL; \
491 new_##FIELD = kmalloc_array(slot, sizeof(struct bpf_##NAME##_state), \
497 memcpy(new_##FIELD, state->FIELD, \
498 sizeof(*new_##FIELD) * (old_size / SIZE)); \
499 memset(new_##FIELD + old_size / SIZE, 0, \
500 sizeof(*new_##FIELD) * (size - old_size) / SIZE); \
502 state->COUNT = slot * SIZE; \
503 kfree(state->FIELD); \
504 state->FIELD = new_##FIELD; \
507 /* realloc_reference_state() */
508 REALLOC_STATE_FN(reference, acquired_refs, refs, 1)
509 /* realloc_stack_state() */
510 REALLOC_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
511 #undef REALLOC_STATE_FN
513 /* do_check() starts with zero-sized stack in struct bpf_verifier_state to
514 * make it consume minimal amount of memory. check_stack_write() access from
515 * the program calls into realloc_func_state() to grow the stack size.
516 * Note there is a non-zero 'parent' pointer inside bpf_verifier_state
517 * which realloc_stack_state() copies over. It points to previous
518 * bpf_verifier_state which is never reallocated.
520 static int realloc_func_state(struct bpf_func_state *state, int stack_size,
521 int refs_size, bool copy_old)
523 int err = realloc_reference_state(state, refs_size, copy_old);
526 return realloc_stack_state(state, stack_size, copy_old);
529 /* Acquire a pointer id from the env and update the state->refs to include
530 * this new pointer reference.
531 * On success, returns a valid pointer id to associate with the register
532 * On failure, returns a negative errno.
534 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
536 struct bpf_func_state *state = cur_func(env);
537 int new_ofs = state->acquired_refs;
540 err = realloc_reference_state(state, state->acquired_refs + 1, true);
544 state->refs[new_ofs].id = id;
545 state->refs[new_ofs].insn_idx = insn_idx;
550 /* release function corresponding to acquire_reference_state(). Idempotent. */
551 static int __release_reference_state(struct bpf_func_state *state, int ptr_id)
558 last_idx = state->acquired_refs - 1;
559 for (i = 0; i < state->acquired_refs; i++) {
560 if (state->refs[i].id == ptr_id) {
561 if (last_idx && i != last_idx)
562 memcpy(&state->refs[i], &state->refs[last_idx],
563 sizeof(*state->refs));
564 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
565 state->acquired_refs--;
572 /* variation on the above for cases where we expect that there must be an
573 * outstanding reference for the specified ptr_id.
575 static int release_reference_state(struct bpf_verifier_env *env, int ptr_id)
577 struct bpf_func_state *state = cur_func(env);
580 err = __release_reference_state(state, ptr_id);
581 if (WARN_ON_ONCE(err != 0))
582 verbose(env, "verifier internal error: can't release reference\n");
586 static int transfer_reference_state(struct bpf_func_state *dst,
587 struct bpf_func_state *src)
589 int err = realloc_reference_state(dst, src->acquired_refs, false);
592 err = copy_reference_state(dst, src);
598 static void free_func_state(struct bpf_func_state *state)
607 static void free_verifier_state(struct bpf_verifier_state *state,
612 for (i = 0; i <= state->curframe; i++) {
613 free_func_state(state->frame[i]);
614 state->frame[i] = NULL;
620 /* copy verifier state from src to dst growing dst stack space
621 * when necessary to accommodate larger src stack
623 static int copy_func_state(struct bpf_func_state *dst,
624 const struct bpf_func_state *src)
628 err = realloc_func_state(dst, src->allocated_stack, src->acquired_refs,
632 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
633 err = copy_reference_state(dst, src);
636 return copy_stack_state(dst, src);
639 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
640 const struct bpf_verifier_state *src)
642 struct bpf_func_state *dst;
645 /* if dst has more stack frames then src frame, free them */
646 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
647 free_func_state(dst_state->frame[i]);
648 dst_state->frame[i] = NULL;
650 dst_state->curframe = src->curframe;
651 for (i = 0; i <= src->curframe; i++) {
652 dst = dst_state->frame[i];
654 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
657 dst_state->frame[i] = dst;
659 err = copy_func_state(dst, src->frame[i]);
666 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
669 struct bpf_verifier_state *cur = env->cur_state;
670 struct bpf_verifier_stack_elem *elem, *head = env->head;
673 if (env->head == NULL)
677 err = copy_verifier_state(cur, &head->st);
682 *insn_idx = head->insn_idx;
684 *prev_insn_idx = head->prev_insn_idx;
686 free_verifier_state(&head->st, false);
693 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
694 int insn_idx, int prev_insn_idx)
696 struct bpf_verifier_state *cur = env->cur_state;
697 struct bpf_verifier_stack_elem *elem;
700 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
704 elem->insn_idx = insn_idx;
705 elem->prev_insn_idx = prev_insn_idx;
706 elem->next = env->head;
709 err = copy_verifier_state(&elem->st, cur);
712 if (env->stack_size > BPF_COMPLEXITY_LIMIT_STACK) {
713 verbose(env, "BPF program is too complex\n");
718 free_verifier_state(env->cur_state, true);
719 env->cur_state = NULL;
720 /* pop all elements and return */
721 while (!pop_stack(env, NULL, NULL));
725 #define CALLER_SAVED_REGS 6
726 static const int caller_saved[CALLER_SAVED_REGS] = {
727 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
730 static void __mark_reg_not_init(struct bpf_reg_state *reg);
732 /* Mark the unknown part of a register (variable offset or scalar value) as
733 * known to have the value @imm.
735 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
737 /* Clear id, off, and union(map_ptr, range) */
738 memset(((u8 *)reg) + sizeof(reg->type), 0,
739 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
740 reg->var_off = tnum_const(imm);
741 reg->smin_value = (s64)imm;
742 reg->smax_value = (s64)imm;
743 reg->umin_value = imm;
744 reg->umax_value = imm;
747 /* Mark the 'variable offset' part of a register as zero. This should be
748 * used only on registers holding a pointer type.
750 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
752 __mark_reg_known(reg, 0);
755 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
757 __mark_reg_known(reg, 0);
758 reg->type = SCALAR_VALUE;
761 static void mark_reg_known_zero(struct bpf_verifier_env *env,
762 struct bpf_reg_state *regs, u32 regno)
764 if (WARN_ON(regno >= MAX_BPF_REG)) {
765 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
766 /* Something bad happened, let's kill all regs */
767 for (regno = 0; regno < MAX_BPF_REG; regno++)
768 __mark_reg_not_init(regs + regno);
771 __mark_reg_known_zero(regs + regno);
774 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
776 return type_is_pkt_pointer(reg->type);
779 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
781 return reg_is_pkt_pointer(reg) ||
782 reg->type == PTR_TO_PACKET_END;
785 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
786 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
787 enum bpf_reg_type which)
789 /* The register can already have a range from prior markings.
790 * This is fine as long as it hasn't been advanced from its
793 return reg->type == which &&
796 tnum_equals_const(reg->var_off, 0);
799 /* Attempts to improve min/max values based on var_off information */
800 static void __update_reg_bounds(struct bpf_reg_state *reg)
802 /* min signed is max(sign bit) | min(other bits) */
803 reg->smin_value = max_t(s64, reg->smin_value,
804 reg->var_off.value | (reg->var_off.mask & S64_MIN));
805 /* max signed is min(sign bit) | max(other bits) */
806 reg->smax_value = min_t(s64, reg->smax_value,
807 reg->var_off.value | (reg->var_off.mask & S64_MAX));
808 reg->umin_value = max(reg->umin_value, reg->var_off.value);
809 reg->umax_value = min(reg->umax_value,
810 reg->var_off.value | reg->var_off.mask);
813 /* Uses signed min/max values to inform unsigned, and vice-versa */
814 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
816 /* Learn sign from signed bounds.
817 * If we cannot cross the sign boundary, then signed and unsigned bounds
818 * are the same, so combine. This works even in the negative case, e.g.
819 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
821 if (reg->smin_value >= 0 || reg->smax_value < 0) {
822 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
824 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
828 /* Learn sign from unsigned bounds. Signed bounds cross the sign
829 * boundary, so we must be careful.
831 if ((s64)reg->umax_value >= 0) {
832 /* Positive. We can't learn anything from the smin, but smax
833 * is positive, hence safe.
835 reg->smin_value = reg->umin_value;
836 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
838 } else if ((s64)reg->umin_value < 0) {
839 /* Negative. We can't learn anything from the smax, but smin
840 * is negative, hence safe.
842 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
844 reg->smax_value = reg->umax_value;
848 /* Attempts to improve var_off based on unsigned min/max information */
849 static void __reg_bound_offset(struct bpf_reg_state *reg)
851 reg->var_off = tnum_intersect(reg->var_off,
852 tnum_range(reg->umin_value,
856 /* Reset the min/max bounds of a register */
857 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
859 reg->smin_value = S64_MIN;
860 reg->smax_value = S64_MAX;
862 reg->umax_value = U64_MAX;
865 /* Mark a register as having a completely unknown (scalar) value. */
866 static void __mark_reg_unknown(struct bpf_reg_state *reg)
869 * Clear type, id, off, and union(map_ptr, range) and
870 * padding between 'type' and union
872 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
873 reg->type = SCALAR_VALUE;
874 reg->var_off = tnum_unknown;
876 __mark_reg_unbounded(reg);
879 static void mark_reg_unknown(struct bpf_verifier_env *env,
880 struct bpf_reg_state *regs, u32 regno)
882 if (WARN_ON(regno >= MAX_BPF_REG)) {
883 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
884 /* Something bad happened, let's kill all regs except FP */
885 for (regno = 0; regno < BPF_REG_FP; regno++)
886 __mark_reg_not_init(regs + regno);
889 __mark_reg_unknown(regs + regno);
892 static void __mark_reg_not_init(struct bpf_reg_state *reg)
894 __mark_reg_unknown(reg);
895 reg->type = NOT_INIT;
898 static void mark_reg_not_init(struct bpf_verifier_env *env,
899 struct bpf_reg_state *regs, u32 regno)
901 if (WARN_ON(regno >= MAX_BPF_REG)) {
902 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
903 /* Something bad happened, let's kill all regs except FP */
904 for (regno = 0; regno < BPF_REG_FP; regno++)
905 __mark_reg_not_init(regs + regno);
908 __mark_reg_not_init(regs + regno);
911 static void init_reg_state(struct bpf_verifier_env *env,
912 struct bpf_func_state *state)
914 struct bpf_reg_state *regs = state->regs;
917 for (i = 0; i < MAX_BPF_REG; i++) {
918 mark_reg_not_init(env, regs, i);
919 regs[i].live = REG_LIVE_NONE;
920 regs[i].parent = NULL;
924 regs[BPF_REG_FP].type = PTR_TO_STACK;
925 mark_reg_known_zero(env, regs, BPF_REG_FP);
926 regs[BPF_REG_FP].frameno = state->frameno;
928 /* 1st arg to a function */
929 regs[BPF_REG_1].type = PTR_TO_CTX;
930 mark_reg_known_zero(env, regs, BPF_REG_1);
933 #define BPF_MAIN_FUNC (-1)
934 static void init_func_state(struct bpf_verifier_env *env,
935 struct bpf_func_state *state,
936 int callsite, int frameno, int subprogno)
938 state->callsite = callsite;
939 state->frameno = frameno;
940 state->subprogno = subprogno;
941 init_reg_state(env, state);
945 SRC_OP, /* register is used as source operand */
946 DST_OP, /* register is used as destination operand */
947 DST_OP_NO_MARK /* same as above, check only, don't mark */
950 static int cmp_subprogs(const void *a, const void *b)
952 return ((struct bpf_subprog_info *)a)->start -
953 ((struct bpf_subprog_info *)b)->start;
956 static int find_subprog(struct bpf_verifier_env *env, int off)
958 struct bpf_subprog_info *p;
960 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
961 sizeof(env->subprog_info[0]), cmp_subprogs);
964 return p - env->subprog_info;
968 static int add_subprog(struct bpf_verifier_env *env, int off)
970 int insn_cnt = env->prog->len;
973 if (off >= insn_cnt || off < 0) {
974 verbose(env, "call to invalid destination\n");
977 ret = find_subprog(env, off);
980 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
981 verbose(env, "too many subprograms\n");
984 env->subprog_info[env->subprog_cnt++].start = off;
985 sort(env->subprog_info, env->subprog_cnt,
986 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
990 static int check_subprogs(struct bpf_verifier_env *env)
992 int i, ret, subprog_start, subprog_end, off, cur_subprog = 0;
993 struct bpf_subprog_info *subprog = env->subprog_info;
994 struct bpf_insn *insn = env->prog->insnsi;
995 int insn_cnt = env->prog->len;
997 /* Add entry function. */
998 ret = add_subprog(env, 0);
1002 /* determine subprog starts. The end is one before the next starts */
1003 for (i = 0; i < insn_cnt; i++) {
1004 if (insn[i].code != (BPF_JMP | BPF_CALL))
1006 if (insn[i].src_reg != BPF_PSEUDO_CALL)
1008 if (!env->allow_ptr_leaks) {
1009 verbose(env, "function calls to other bpf functions are allowed for root only\n");
1012 ret = add_subprog(env, i + insn[i].imm + 1);
1017 /* Add a fake 'exit' subprog which could simplify subprog iteration
1018 * logic. 'subprog_cnt' should not be increased.
1020 subprog[env->subprog_cnt].start = insn_cnt;
1022 if (env->log.level > 1)
1023 for (i = 0; i < env->subprog_cnt; i++)
1024 verbose(env, "func#%d @%d\n", i, subprog[i].start);
1026 /* now check that all jumps are within the same subprog */
1027 subprog_start = subprog[cur_subprog].start;
1028 subprog_end = subprog[cur_subprog + 1].start;
1029 for (i = 0; i < insn_cnt; i++) {
1030 u8 code = insn[i].code;
1032 if (BPF_CLASS(code) != BPF_JMP)
1034 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
1036 off = i + insn[i].off + 1;
1037 if (off < subprog_start || off >= subprog_end) {
1038 verbose(env, "jump out of range from insn %d to %d\n", i, off);
1042 if (i == subprog_end - 1) {
1043 /* to avoid fall-through from one subprog into another
1044 * the last insn of the subprog should be either exit
1045 * or unconditional jump back
1047 if (code != (BPF_JMP | BPF_EXIT) &&
1048 code != (BPF_JMP | BPF_JA)) {
1049 verbose(env, "last insn is not an exit or jmp\n");
1052 subprog_start = subprog_end;
1054 if (cur_subprog < env->subprog_cnt)
1055 subprog_end = subprog[cur_subprog + 1].start;
1061 /* Parentage chain of this register (or stack slot) should take care of all
1062 * issues like callee-saved registers, stack slot allocation time, etc.
1064 static int mark_reg_read(struct bpf_verifier_env *env,
1065 const struct bpf_reg_state *state,
1066 struct bpf_reg_state *parent)
1068 bool writes = parent == state->parent; /* Observe write marks */
1071 /* if read wasn't screened by an earlier write ... */
1072 if (writes && state->live & REG_LIVE_WRITTEN)
1074 /* ... then we depend on parent's value */
1075 parent->live |= REG_LIVE_READ;
1077 parent = state->parent;
1083 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
1084 enum reg_arg_type t)
1086 struct bpf_verifier_state *vstate = env->cur_state;
1087 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1088 struct bpf_reg_state *regs = state->regs;
1090 if (regno >= MAX_BPF_REG) {
1091 verbose(env, "R%d is invalid\n", regno);
1096 /* check whether register used as source operand can be read */
1097 if (regs[regno].type == NOT_INIT) {
1098 verbose(env, "R%d !read_ok\n", regno);
1101 /* We don't need to worry about FP liveness because it's read-only */
1102 if (regno != BPF_REG_FP)
1103 return mark_reg_read(env, ®s[regno],
1104 regs[regno].parent);
1106 /* check whether register used as dest operand can be written to */
1107 if (regno == BPF_REG_FP) {
1108 verbose(env, "frame pointer is read only\n");
1111 regs[regno].live |= REG_LIVE_WRITTEN;
1113 mark_reg_unknown(env, regs, regno);
1118 static bool is_spillable_regtype(enum bpf_reg_type type)
1121 case PTR_TO_MAP_VALUE:
1122 case PTR_TO_MAP_VALUE_OR_NULL:
1126 case PTR_TO_PACKET_META:
1127 case PTR_TO_PACKET_END:
1128 case PTR_TO_FLOW_KEYS:
1129 case CONST_PTR_TO_MAP:
1131 case PTR_TO_SOCKET_OR_NULL:
1138 /* Does this register contain a constant zero? */
1139 static bool register_is_null(struct bpf_reg_state *reg)
1141 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
1144 /* check_stack_read/write functions track spill/fill of registers,
1145 * stack boundary and alignment are checked in check_mem_access()
1147 static int check_stack_write(struct bpf_verifier_env *env,
1148 struct bpf_func_state *state, /* func where register points to */
1149 int off, int size, int value_regno, int insn_idx)
1151 struct bpf_func_state *cur; /* state of the current function */
1152 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
1153 enum bpf_reg_type type;
1155 err = realloc_func_state(state, round_up(slot + 1, BPF_REG_SIZE),
1156 state->acquired_refs, true);
1159 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
1160 * so it's aligned access and [off, off + size) are within stack limits
1162 if (!env->allow_ptr_leaks &&
1163 state->stack[spi].slot_type[0] == STACK_SPILL &&
1164 size != BPF_REG_SIZE) {
1165 verbose(env, "attempt to corrupt spilled pointer on stack\n");
1169 cur = env->cur_state->frame[env->cur_state->curframe];
1170 if (value_regno >= 0 &&
1171 is_spillable_regtype((type = cur->regs[value_regno].type))) {
1173 /* register containing pointer is being spilled into stack */
1174 if (size != BPF_REG_SIZE) {
1175 verbose(env, "invalid size of register spill\n");
1179 if (state != cur && type == PTR_TO_STACK) {
1180 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
1184 /* save register state */
1185 state->stack[spi].spilled_ptr = cur->regs[value_regno];
1186 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1188 for (i = 0; i < BPF_REG_SIZE; i++) {
1189 if (state->stack[spi].slot_type[i] == STACK_MISC &&
1190 !env->allow_ptr_leaks) {
1191 int *poff = &env->insn_aux_data[insn_idx].sanitize_stack_off;
1192 int soff = (-spi - 1) * BPF_REG_SIZE;
1194 /* detected reuse of integer stack slot with a pointer
1195 * which means either llvm is reusing stack slot or
1196 * an attacker is trying to exploit CVE-2018-3639
1197 * (speculative store bypass)
1198 * Have to sanitize that slot with preemptive
1201 if (*poff && *poff != soff) {
1202 /* disallow programs where single insn stores
1203 * into two different stack slots, since verifier
1204 * cannot sanitize them
1207 "insn %d cannot access two stack slots fp%d and fp%d",
1208 insn_idx, *poff, soff);
1213 state->stack[spi].slot_type[i] = STACK_SPILL;
1216 u8 type = STACK_MISC;
1218 /* regular write of data into stack destroys any spilled ptr */
1219 state->stack[spi].spilled_ptr.type = NOT_INIT;
1221 /* only mark the slot as written if all 8 bytes were written
1222 * otherwise read propagation may incorrectly stop too soon
1223 * when stack slots are partially written.
1224 * This heuristic means that read propagation will be
1225 * conservative, since it will add reg_live_read marks
1226 * to stack slots all the way to first state when programs
1227 * writes+reads less than 8 bytes
1229 if (size == BPF_REG_SIZE)
1230 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1232 /* when we zero initialize stack slots mark them as such */
1233 if (value_regno >= 0 &&
1234 register_is_null(&cur->regs[value_regno]))
1237 for (i = 0; i < size; i++)
1238 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
1244 static int check_stack_read(struct bpf_verifier_env *env,
1245 struct bpf_func_state *reg_state /* func where register points to */,
1246 int off, int size, int value_regno)
1248 struct bpf_verifier_state *vstate = env->cur_state;
1249 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1250 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
1253 if (reg_state->allocated_stack <= slot) {
1254 verbose(env, "invalid read from stack off %d+0 size %d\n",
1258 stype = reg_state->stack[spi].slot_type;
1260 if (stype[0] == STACK_SPILL) {
1261 if (size != BPF_REG_SIZE) {
1262 verbose(env, "invalid size of register spill\n");
1265 for (i = 1; i < BPF_REG_SIZE; i++) {
1266 if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) {
1267 verbose(env, "corrupted spill memory\n");
1272 if (value_regno >= 0) {
1273 /* restore register state from stack */
1274 state->regs[value_regno] = reg_state->stack[spi].spilled_ptr;
1275 /* mark reg as written since spilled pointer state likely
1276 * has its liveness marks cleared by is_state_visited()
1277 * which resets stack/reg liveness for state transitions
1279 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
1281 mark_reg_read(env, ®_state->stack[spi].spilled_ptr,
1282 reg_state->stack[spi].spilled_ptr.parent);
1287 for (i = 0; i < size; i++) {
1288 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_MISC)
1290 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_ZERO) {
1294 verbose(env, "invalid read from stack off %d+%d size %d\n",
1298 mark_reg_read(env, ®_state->stack[spi].spilled_ptr,
1299 reg_state->stack[spi].spilled_ptr.parent);
1300 if (value_regno >= 0) {
1301 if (zeros == size) {
1302 /* any size read into register is zero extended,
1303 * so the whole register == const_zero
1305 __mark_reg_const_zero(&state->regs[value_regno]);
1307 /* have read misc data from the stack */
1308 mark_reg_unknown(env, state->regs, value_regno);
1310 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
1316 /* check read/write into map element returned by bpf_map_lookup_elem() */
1317 static int __check_map_access(struct bpf_verifier_env *env, u32 regno, int off,
1318 int size, bool zero_size_allowed)
1320 struct bpf_reg_state *regs = cur_regs(env);
1321 struct bpf_map *map = regs[regno].map_ptr;
1323 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
1324 off + size > map->value_size) {
1325 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
1326 map->value_size, off, size);
1332 /* check read/write into a map element with possible variable offset */
1333 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
1334 int off, int size, bool zero_size_allowed)
1336 struct bpf_verifier_state *vstate = env->cur_state;
1337 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1338 struct bpf_reg_state *reg = &state->regs[regno];
1341 /* We may have adjusted the register to this map value, so we
1342 * need to try adding each of min_value and max_value to off
1343 * to make sure our theoretical access will be safe.
1346 print_verifier_state(env, state);
1347 /* The minimum value is only important with signed
1348 * comparisons where we can't assume the floor of a
1349 * value is 0. If we are using signed variables for our
1350 * index'es we need to make sure that whatever we use
1351 * will have a set floor within our range.
1353 if (reg->smin_value < 0) {
1354 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1358 err = __check_map_access(env, regno, reg->smin_value + off, size,
1361 verbose(env, "R%d min value is outside of the array range\n",
1366 /* If we haven't set a max value then we need to bail since we can't be
1367 * sure we won't do bad things.
1368 * If reg->umax_value + off could overflow, treat that as unbounded too.
1370 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
1371 verbose(env, "R%d unbounded memory access, make sure to bounds check any array access into a map\n",
1375 err = __check_map_access(env, regno, reg->umax_value + off, size,
1378 verbose(env, "R%d max value is outside of the array range\n",
1383 #define MAX_PACKET_OFF 0xffff
1385 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
1386 const struct bpf_call_arg_meta *meta,
1387 enum bpf_access_type t)
1389 switch (env->prog->type) {
1390 case BPF_PROG_TYPE_LWT_IN:
1391 case BPF_PROG_TYPE_LWT_OUT:
1392 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
1393 case BPF_PROG_TYPE_SK_REUSEPORT:
1394 /* dst_input() and dst_output() can't write for now */
1398 case BPF_PROG_TYPE_SCHED_CLS:
1399 case BPF_PROG_TYPE_SCHED_ACT:
1400 case BPF_PROG_TYPE_XDP:
1401 case BPF_PROG_TYPE_LWT_XMIT:
1402 case BPF_PROG_TYPE_SK_SKB:
1403 case BPF_PROG_TYPE_SK_MSG:
1404 case BPF_PROG_TYPE_FLOW_DISSECTOR:
1406 return meta->pkt_access;
1408 env->seen_direct_write = true;
1415 static int __check_packet_access(struct bpf_verifier_env *env, u32 regno,
1416 int off, int size, bool zero_size_allowed)
1418 struct bpf_reg_state *regs = cur_regs(env);
1419 struct bpf_reg_state *reg = ®s[regno];
1421 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
1422 (u64)off + size > reg->range) {
1423 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
1424 off, size, regno, reg->id, reg->off, reg->range);
1430 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
1431 int size, bool zero_size_allowed)
1433 struct bpf_reg_state *regs = cur_regs(env);
1434 struct bpf_reg_state *reg = ®s[regno];
1437 /* We may have added a variable offset to the packet pointer; but any
1438 * reg->range we have comes after that. We are only checking the fixed
1442 /* We don't allow negative numbers, because we aren't tracking enough
1443 * detail to prove they're safe.
1445 if (reg->smin_value < 0) {
1446 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1450 err = __check_packet_access(env, regno, off, size, zero_size_allowed);
1452 verbose(env, "R%d offset is outside of the packet\n", regno);
1458 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
1459 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
1460 enum bpf_access_type t, enum bpf_reg_type *reg_type)
1462 struct bpf_insn_access_aux info = {
1463 .reg_type = *reg_type,
1466 if (env->ops->is_valid_access &&
1467 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
1468 /* A non zero info.ctx_field_size indicates that this field is a
1469 * candidate for later verifier transformation to load the whole
1470 * field and then apply a mask when accessed with a narrower
1471 * access than actual ctx access size. A zero info.ctx_field_size
1472 * will only allow for whole field access and rejects any other
1473 * type of narrower access.
1475 *reg_type = info.reg_type;
1477 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
1478 /* remember the offset of last byte accessed in ctx */
1479 if (env->prog->aux->max_ctx_offset < off + size)
1480 env->prog->aux->max_ctx_offset = off + size;
1484 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
1488 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
1491 if (size < 0 || off < 0 ||
1492 (u64)off + size > sizeof(struct bpf_flow_keys)) {
1493 verbose(env, "invalid access to flow keys off=%d size=%d\n",
1500 static int check_sock_access(struct bpf_verifier_env *env, u32 regno, int off,
1501 int size, enum bpf_access_type t)
1503 struct bpf_reg_state *regs = cur_regs(env);
1504 struct bpf_reg_state *reg = ®s[regno];
1505 struct bpf_insn_access_aux info;
1507 if (reg->smin_value < 0) {
1508 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1513 if (!bpf_sock_is_valid_access(off, size, t, &info)) {
1514 verbose(env, "invalid bpf_sock access off=%d size=%d\n",
1522 static bool __is_pointer_value(bool allow_ptr_leaks,
1523 const struct bpf_reg_state *reg)
1525 if (allow_ptr_leaks)
1528 return reg->type != SCALAR_VALUE;
1531 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
1533 return __is_pointer_value(env->allow_ptr_leaks, cur_regs(env) + regno);
1536 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
1538 const struct bpf_reg_state *reg = cur_regs(env) + regno;
1540 return reg->type == PTR_TO_CTX ||
1541 reg->type == PTR_TO_SOCKET;
1544 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
1546 const struct bpf_reg_state *reg = cur_regs(env) + regno;
1548 return type_is_pkt_pointer(reg->type);
1551 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
1552 const struct bpf_reg_state *reg,
1553 int off, int size, bool strict)
1555 struct tnum reg_off;
1558 /* Byte size accesses are always allowed. */
1559 if (!strict || size == 1)
1562 /* For platforms that do not have a Kconfig enabling
1563 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
1564 * NET_IP_ALIGN is universally set to '2'. And on platforms
1565 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
1566 * to this code only in strict mode where we want to emulate
1567 * the NET_IP_ALIGN==2 checking. Therefore use an
1568 * unconditional IP align value of '2'.
1572 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
1573 if (!tnum_is_aligned(reg_off, size)) {
1576 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1578 "misaligned packet access off %d+%s+%d+%d size %d\n",
1579 ip_align, tn_buf, reg->off, off, size);
1586 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
1587 const struct bpf_reg_state *reg,
1588 const char *pointer_desc,
1589 int off, int size, bool strict)
1591 struct tnum reg_off;
1593 /* Byte size accesses are always allowed. */
1594 if (!strict || size == 1)
1597 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
1598 if (!tnum_is_aligned(reg_off, size)) {
1601 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1602 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
1603 pointer_desc, tn_buf, reg->off, off, size);
1610 static int check_ptr_alignment(struct bpf_verifier_env *env,
1611 const struct bpf_reg_state *reg, int off,
1612 int size, bool strict_alignment_once)
1614 bool strict = env->strict_alignment || strict_alignment_once;
1615 const char *pointer_desc = "";
1617 switch (reg->type) {
1619 case PTR_TO_PACKET_META:
1620 /* Special case, because of NET_IP_ALIGN. Given metadata sits
1621 * right in front, treat it the very same way.
1623 return check_pkt_ptr_alignment(env, reg, off, size, strict);
1624 case PTR_TO_FLOW_KEYS:
1625 pointer_desc = "flow keys ";
1627 case PTR_TO_MAP_VALUE:
1628 pointer_desc = "value ";
1631 pointer_desc = "context ";
1634 pointer_desc = "stack ";
1635 /* The stack spill tracking logic in check_stack_write()
1636 * and check_stack_read() relies on stack accesses being
1642 pointer_desc = "sock ";
1647 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
1651 static int update_stack_depth(struct bpf_verifier_env *env,
1652 const struct bpf_func_state *func,
1655 u16 stack = env->subprog_info[func->subprogno].stack_depth;
1660 /* update known max for given subprogram */
1661 env->subprog_info[func->subprogno].stack_depth = -off;
1665 /* starting from main bpf function walk all instructions of the function
1666 * and recursively walk all callees that given function can call.
1667 * Ignore jump and exit insns.
1668 * Since recursion is prevented by check_cfg() this algorithm
1669 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
1671 static int check_max_stack_depth(struct bpf_verifier_env *env)
1673 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
1674 struct bpf_subprog_info *subprog = env->subprog_info;
1675 struct bpf_insn *insn = env->prog->insnsi;
1676 int ret_insn[MAX_CALL_FRAMES];
1677 int ret_prog[MAX_CALL_FRAMES];
1680 /* round up to 32-bytes, since this is granularity
1681 * of interpreter stack size
1683 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
1684 if (depth > MAX_BPF_STACK) {
1685 verbose(env, "combined stack size of %d calls is %d. Too large\n",
1690 subprog_end = subprog[idx + 1].start;
1691 for (; i < subprog_end; i++) {
1692 if (insn[i].code != (BPF_JMP | BPF_CALL))
1694 if (insn[i].src_reg != BPF_PSEUDO_CALL)
1696 /* remember insn and function to return to */
1697 ret_insn[frame] = i + 1;
1698 ret_prog[frame] = idx;
1700 /* find the callee */
1701 i = i + insn[i].imm + 1;
1702 idx = find_subprog(env, i);
1704 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
1709 if (frame >= MAX_CALL_FRAMES) {
1710 WARN_ONCE(1, "verifier bug. Call stack is too deep\n");
1715 /* end of for() loop means the last insn of the 'subprog'
1716 * was reached. Doesn't matter whether it was JA or EXIT
1720 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
1722 i = ret_insn[frame];
1723 idx = ret_prog[frame];
1727 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
1728 static int get_callee_stack_depth(struct bpf_verifier_env *env,
1729 const struct bpf_insn *insn, int idx)
1731 int start = idx + insn->imm + 1, subprog;
1733 subprog = find_subprog(env, start);
1735 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
1739 return env->subprog_info[subprog].stack_depth;
1743 static int check_ctx_reg(struct bpf_verifier_env *env,
1744 const struct bpf_reg_state *reg, int regno)
1746 /* Access to ctx or passing it to a helper is only allowed in
1747 * its original, unmodified form.
1751 verbose(env, "dereference of modified ctx ptr R%d off=%d disallowed\n",
1756 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
1759 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1760 verbose(env, "variable ctx access var_off=%s disallowed\n", tn_buf);
1767 /* truncate register to smaller size (in bytes)
1768 * must be called with size < BPF_REG_SIZE
1770 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
1774 /* clear high bits in bit representation */
1775 reg->var_off = tnum_cast(reg->var_off, size);
1777 /* fix arithmetic bounds */
1778 mask = ((u64)1 << (size * 8)) - 1;
1779 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
1780 reg->umin_value &= mask;
1781 reg->umax_value &= mask;
1783 reg->umin_value = 0;
1784 reg->umax_value = mask;
1786 reg->smin_value = reg->umin_value;
1787 reg->smax_value = reg->umax_value;
1790 /* check whether memory at (regno + off) is accessible for t = (read | write)
1791 * if t==write, value_regno is a register which value is stored into memory
1792 * if t==read, value_regno is a register which will receive the value from memory
1793 * if t==write && value_regno==-1, some unknown value is stored into memory
1794 * if t==read && value_regno==-1, don't care what we read from memory
1796 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
1797 int off, int bpf_size, enum bpf_access_type t,
1798 int value_regno, bool strict_alignment_once)
1800 struct bpf_reg_state *regs = cur_regs(env);
1801 struct bpf_reg_state *reg = regs + regno;
1802 struct bpf_func_state *state;
1805 size = bpf_size_to_bytes(bpf_size);
1809 /* alignment checks will add in reg->off themselves */
1810 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
1814 /* for access checks, reg->off is just part of off */
1817 if (reg->type == PTR_TO_MAP_VALUE) {
1818 if (t == BPF_WRITE && value_regno >= 0 &&
1819 is_pointer_value(env, value_regno)) {
1820 verbose(env, "R%d leaks addr into map\n", value_regno);
1824 err = check_map_access(env, regno, off, size, false);
1825 if (!err && t == BPF_READ && value_regno >= 0)
1826 mark_reg_unknown(env, regs, value_regno);
1828 } else if (reg->type == PTR_TO_CTX) {
1829 enum bpf_reg_type reg_type = SCALAR_VALUE;
1831 if (t == BPF_WRITE && value_regno >= 0 &&
1832 is_pointer_value(env, value_regno)) {
1833 verbose(env, "R%d leaks addr into ctx\n", value_regno);
1837 err = check_ctx_reg(env, reg, regno);
1841 err = check_ctx_access(env, insn_idx, off, size, t, ®_type);
1842 if (!err && t == BPF_READ && value_regno >= 0) {
1843 /* ctx access returns either a scalar, or a
1844 * PTR_TO_PACKET[_META,_END]. In the latter
1845 * case, we know the offset is zero.
1847 if (reg_type == SCALAR_VALUE)
1848 mark_reg_unknown(env, regs, value_regno);
1850 mark_reg_known_zero(env, regs,
1852 regs[value_regno].type = reg_type;
1855 } else if (reg->type == PTR_TO_STACK) {
1856 /* stack accesses must be at a fixed offset, so that we can
1857 * determine what type of data were returned.
1858 * See check_stack_read().
1860 if (!tnum_is_const(reg->var_off)) {
1863 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1864 verbose(env, "variable stack access var_off=%s off=%d size=%d",
1868 off += reg->var_off.value;
1869 if (off >= 0 || off < -MAX_BPF_STACK) {
1870 verbose(env, "invalid stack off=%d size=%d\n", off,
1875 state = func(env, reg);
1876 err = update_stack_depth(env, state, off);
1881 err = check_stack_write(env, state, off, size,
1882 value_regno, insn_idx);
1884 err = check_stack_read(env, state, off, size,
1886 } else if (reg_is_pkt_pointer(reg)) {
1887 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
1888 verbose(env, "cannot write into packet\n");
1891 if (t == BPF_WRITE && value_regno >= 0 &&
1892 is_pointer_value(env, value_regno)) {
1893 verbose(env, "R%d leaks addr into packet\n",
1897 err = check_packet_access(env, regno, off, size, false);
1898 if (!err && t == BPF_READ && value_regno >= 0)
1899 mark_reg_unknown(env, regs, value_regno);
1900 } else if (reg->type == PTR_TO_FLOW_KEYS) {
1901 if (t == BPF_WRITE && value_regno >= 0 &&
1902 is_pointer_value(env, value_regno)) {
1903 verbose(env, "R%d leaks addr into flow keys\n",
1908 err = check_flow_keys_access(env, off, size);
1909 if (!err && t == BPF_READ && value_regno >= 0)
1910 mark_reg_unknown(env, regs, value_regno);
1911 } else if (reg->type == PTR_TO_SOCKET) {
1912 if (t == BPF_WRITE) {
1913 verbose(env, "cannot write into socket\n");
1916 err = check_sock_access(env, regno, off, size, t);
1917 if (!err && value_regno >= 0)
1918 mark_reg_unknown(env, regs, value_regno);
1920 verbose(env, "R%d invalid mem access '%s'\n", regno,
1921 reg_type_str[reg->type]);
1925 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
1926 regs[value_regno].type == SCALAR_VALUE) {
1927 /* b/h/w load zero-extends, mark upper bits as known 0 */
1928 coerce_reg_to_size(®s[value_regno], size);
1933 static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
1937 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
1939 verbose(env, "BPF_XADD uses reserved fields\n");
1943 /* check src1 operand */
1944 err = check_reg_arg(env, insn->src_reg, SRC_OP);
1948 /* check src2 operand */
1949 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
1953 if (is_pointer_value(env, insn->src_reg)) {
1954 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
1958 if (is_ctx_reg(env, insn->dst_reg) ||
1959 is_pkt_reg(env, insn->dst_reg)) {
1960 verbose(env, "BPF_XADD stores into R%d %s is not allowed\n",
1961 insn->dst_reg, reg_type_str[insn->dst_reg]);
1965 /* check whether atomic_add can read the memory */
1966 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1967 BPF_SIZE(insn->code), BPF_READ, -1, true);
1971 /* check whether atomic_add can write into the same memory */
1972 return check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1973 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
1976 /* when register 'regno' is passed into function that will read 'access_size'
1977 * bytes from that pointer, make sure that it's within stack boundary
1978 * and all elements of stack are initialized.
1979 * Unlike most pointer bounds-checking functions, this one doesn't take an
1980 * 'off' argument, so it has to add in reg->off itself.
1982 static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
1983 int access_size, bool zero_size_allowed,
1984 struct bpf_call_arg_meta *meta)
1986 struct bpf_reg_state *reg = cur_regs(env) + regno;
1987 struct bpf_func_state *state = func(env, reg);
1988 int off, i, slot, spi;
1990 if (reg->type != PTR_TO_STACK) {
1991 /* Allow zero-byte read from NULL, regardless of pointer type */
1992 if (zero_size_allowed && access_size == 0 &&
1993 register_is_null(reg))
1996 verbose(env, "R%d type=%s expected=%s\n", regno,
1997 reg_type_str[reg->type],
1998 reg_type_str[PTR_TO_STACK]);
2002 /* Only allow fixed-offset stack reads */
2003 if (!tnum_is_const(reg->var_off)) {
2006 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2007 verbose(env, "invalid variable stack read R%d var_off=%s\n",
2011 off = reg->off + reg->var_off.value;
2012 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
2013 access_size < 0 || (access_size == 0 && !zero_size_allowed)) {
2014 verbose(env, "invalid stack type R%d off=%d access_size=%d\n",
2015 regno, off, access_size);
2019 if (meta && meta->raw_mode) {
2020 meta->access_size = access_size;
2021 meta->regno = regno;
2025 for (i = 0; i < access_size; i++) {
2028 slot = -(off + i) - 1;
2029 spi = slot / BPF_REG_SIZE;
2030 if (state->allocated_stack <= slot)
2032 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
2033 if (*stype == STACK_MISC)
2035 if (*stype == STACK_ZERO) {
2036 /* helper can write anything into the stack */
2037 *stype = STACK_MISC;
2041 verbose(env, "invalid indirect read from stack off %d+%d size %d\n",
2042 off, i, access_size);
2045 /* reading any byte out of 8-byte 'spill_slot' will cause
2046 * the whole slot to be marked as 'read'
2048 mark_reg_read(env, &state->stack[spi].spilled_ptr,
2049 state->stack[spi].spilled_ptr.parent);
2051 return update_stack_depth(env, state, off);
2054 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
2055 int access_size, bool zero_size_allowed,
2056 struct bpf_call_arg_meta *meta)
2058 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
2060 switch (reg->type) {
2062 case PTR_TO_PACKET_META:
2063 return check_packet_access(env, regno, reg->off, access_size,
2065 case PTR_TO_FLOW_KEYS:
2066 return check_flow_keys_access(env, reg->off, access_size);
2067 case PTR_TO_MAP_VALUE:
2068 return check_map_access(env, regno, reg->off, access_size,
2070 default: /* scalar_value|ptr_to_stack or invalid ptr */
2071 return check_stack_boundary(env, regno, access_size,
2072 zero_size_allowed, meta);
2076 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
2078 return type == ARG_PTR_TO_MEM ||
2079 type == ARG_PTR_TO_MEM_OR_NULL ||
2080 type == ARG_PTR_TO_UNINIT_MEM;
2083 static bool arg_type_is_mem_size(enum bpf_arg_type type)
2085 return type == ARG_CONST_SIZE ||
2086 type == ARG_CONST_SIZE_OR_ZERO;
2089 static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
2090 enum bpf_arg_type arg_type,
2091 struct bpf_call_arg_meta *meta)
2093 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
2094 enum bpf_reg_type expected_type, type = reg->type;
2097 if (arg_type == ARG_DONTCARE)
2100 err = check_reg_arg(env, regno, SRC_OP);
2104 if (arg_type == ARG_ANYTHING) {
2105 if (is_pointer_value(env, regno)) {
2106 verbose(env, "R%d leaks addr into helper function\n",
2113 if (type_is_pkt_pointer(type) &&
2114 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
2115 verbose(env, "helper access to the packet is not allowed\n");
2119 if (arg_type == ARG_PTR_TO_MAP_KEY ||
2120 arg_type == ARG_PTR_TO_MAP_VALUE ||
2121 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
2122 expected_type = PTR_TO_STACK;
2123 if (!type_is_pkt_pointer(type) && type != PTR_TO_MAP_VALUE &&
2124 type != expected_type)
2126 } else if (arg_type == ARG_CONST_SIZE ||
2127 arg_type == ARG_CONST_SIZE_OR_ZERO) {
2128 expected_type = SCALAR_VALUE;
2129 if (type != expected_type)
2131 } else if (arg_type == ARG_CONST_MAP_PTR) {
2132 expected_type = CONST_PTR_TO_MAP;
2133 if (type != expected_type)
2135 } else if (arg_type == ARG_PTR_TO_CTX) {
2136 expected_type = PTR_TO_CTX;
2137 if (type != expected_type)
2139 err = check_ctx_reg(env, reg, regno);
2142 } else if (arg_type == ARG_PTR_TO_SOCKET) {
2143 expected_type = PTR_TO_SOCKET;
2144 if (type != expected_type)
2146 if (meta->ptr_id || !reg->id) {
2147 verbose(env, "verifier internal error: mismatched references meta=%d, reg=%d\n",
2148 meta->ptr_id, reg->id);
2151 meta->ptr_id = reg->id;
2152 } else if (arg_type_is_mem_ptr(arg_type)) {
2153 expected_type = PTR_TO_STACK;
2154 /* One exception here. In case function allows for NULL to be
2155 * passed in as argument, it's a SCALAR_VALUE type. Final test
2156 * happens during stack boundary checking.
2158 if (register_is_null(reg) &&
2159 arg_type == ARG_PTR_TO_MEM_OR_NULL)
2160 /* final test in check_stack_boundary() */;
2161 else if (!type_is_pkt_pointer(type) &&
2162 type != PTR_TO_MAP_VALUE &&
2163 type != expected_type)
2165 meta->raw_mode = arg_type == ARG_PTR_TO_UNINIT_MEM;
2167 verbose(env, "unsupported arg_type %d\n", arg_type);
2171 if (arg_type == ARG_CONST_MAP_PTR) {
2172 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
2173 meta->map_ptr = reg->map_ptr;
2174 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
2175 /* bpf_map_xxx(..., map_ptr, ..., key) call:
2176 * check that [key, key + map->key_size) are within
2177 * stack limits and initialized
2179 if (!meta->map_ptr) {
2180 /* in function declaration map_ptr must come before
2181 * map_key, so that it's verified and known before
2182 * we have to check map_key here. Otherwise it means
2183 * that kernel subsystem misconfigured verifier
2185 verbose(env, "invalid map_ptr to access map->key\n");
2188 err = check_helper_mem_access(env, regno,
2189 meta->map_ptr->key_size, false,
2191 } else if (arg_type == ARG_PTR_TO_MAP_VALUE ||
2192 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
2193 /* bpf_map_xxx(..., map_ptr, ..., value) call:
2194 * check [value, value + map->value_size) validity
2196 if (!meta->map_ptr) {
2197 /* kernel subsystem misconfigured verifier */
2198 verbose(env, "invalid map_ptr to access map->value\n");
2201 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE);
2202 err = check_helper_mem_access(env, regno,
2203 meta->map_ptr->value_size, false,
2205 } else if (arg_type_is_mem_size(arg_type)) {
2206 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
2208 /* remember the mem_size which may be used later
2209 * to refine return values.
2211 meta->msize_smax_value = reg->smax_value;
2212 meta->msize_umax_value = reg->umax_value;
2214 /* The register is SCALAR_VALUE; the access check
2215 * happens using its boundaries.
2217 if (!tnum_is_const(reg->var_off))
2218 /* For unprivileged variable accesses, disable raw
2219 * mode so that the program is required to
2220 * initialize all the memory that the helper could
2221 * just partially fill up.
2225 if (reg->smin_value < 0) {
2226 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
2231 if (reg->umin_value == 0) {
2232 err = check_helper_mem_access(env, regno - 1, 0,
2239 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
2240 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
2244 err = check_helper_mem_access(env, regno - 1,
2246 zero_size_allowed, meta);
2251 verbose(env, "R%d type=%s expected=%s\n", regno,
2252 reg_type_str[type], reg_type_str[expected_type]);
2256 static int check_map_func_compatibility(struct bpf_verifier_env *env,
2257 struct bpf_map *map, int func_id)
2262 /* We need a two way check, first is from map perspective ... */
2263 switch (map->map_type) {
2264 case BPF_MAP_TYPE_PROG_ARRAY:
2265 if (func_id != BPF_FUNC_tail_call)
2268 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
2269 if (func_id != BPF_FUNC_perf_event_read &&
2270 func_id != BPF_FUNC_perf_event_output &&
2271 func_id != BPF_FUNC_perf_event_read_value)
2274 case BPF_MAP_TYPE_STACK_TRACE:
2275 if (func_id != BPF_FUNC_get_stackid)
2278 case BPF_MAP_TYPE_CGROUP_ARRAY:
2279 if (func_id != BPF_FUNC_skb_under_cgroup &&
2280 func_id != BPF_FUNC_current_task_under_cgroup)
2283 case BPF_MAP_TYPE_CGROUP_STORAGE:
2284 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
2285 if (func_id != BPF_FUNC_get_local_storage)
2288 /* devmap returns a pointer to a live net_device ifindex that we cannot
2289 * allow to be modified from bpf side. So do not allow lookup elements
2292 case BPF_MAP_TYPE_DEVMAP:
2293 if (func_id != BPF_FUNC_redirect_map)
2296 /* Restrict bpf side of cpumap and xskmap, open when use-cases
2299 case BPF_MAP_TYPE_CPUMAP:
2300 case BPF_MAP_TYPE_XSKMAP:
2301 if (func_id != BPF_FUNC_redirect_map)
2304 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
2305 case BPF_MAP_TYPE_HASH_OF_MAPS:
2306 if (func_id != BPF_FUNC_map_lookup_elem)
2309 case BPF_MAP_TYPE_SOCKMAP:
2310 if (func_id != BPF_FUNC_sk_redirect_map &&
2311 func_id != BPF_FUNC_sock_map_update &&
2312 func_id != BPF_FUNC_map_delete_elem &&
2313 func_id != BPF_FUNC_msg_redirect_map)
2316 case BPF_MAP_TYPE_SOCKHASH:
2317 if (func_id != BPF_FUNC_sk_redirect_hash &&
2318 func_id != BPF_FUNC_sock_hash_update &&
2319 func_id != BPF_FUNC_map_delete_elem &&
2320 func_id != BPF_FUNC_msg_redirect_hash)
2323 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
2324 if (func_id != BPF_FUNC_sk_select_reuseport)
2327 case BPF_MAP_TYPE_QUEUE:
2328 case BPF_MAP_TYPE_STACK:
2329 if (func_id != BPF_FUNC_map_peek_elem &&
2330 func_id != BPF_FUNC_map_pop_elem &&
2331 func_id != BPF_FUNC_map_push_elem)
2338 /* ... and second from the function itself. */
2340 case BPF_FUNC_tail_call:
2341 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
2343 if (env->subprog_cnt > 1) {
2344 verbose(env, "tail_calls are not allowed in programs with bpf-to-bpf calls\n");
2348 case BPF_FUNC_perf_event_read:
2349 case BPF_FUNC_perf_event_output:
2350 case BPF_FUNC_perf_event_read_value:
2351 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
2354 case BPF_FUNC_get_stackid:
2355 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
2358 case BPF_FUNC_current_task_under_cgroup:
2359 case BPF_FUNC_skb_under_cgroup:
2360 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
2363 case BPF_FUNC_redirect_map:
2364 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
2365 map->map_type != BPF_MAP_TYPE_CPUMAP &&
2366 map->map_type != BPF_MAP_TYPE_XSKMAP)
2369 case BPF_FUNC_sk_redirect_map:
2370 case BPF_FUNC_msg_redirect_map:
2371 case BPF_FUNC_sock_map_update:
2372 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
2375 case BPF_FUNC_sk_redirect_hash:
2376 case BPF_FUNC_msg_redirect_hash:
2377 case BPF_FUNC_sock_hash_update:
2378 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
2381 case BPF_FUNC_get_local_storage:
2382 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
2383 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
2386 case BPF_FUNC_sk_select_reuseport:
2387 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY)
2390 case BPF_FUNC_map_peek_elem:
2391 case BPF_FUNC_map_pop_elem:
2392 case BPF_FUNC_map_push_elem:
2393 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
2394 map->map_type != BPF_MAP_TYPE_STACK)
2403 verbose(env, "cannot pass map_type %d into func %s#%d\n",
2404 map->map_type, func_id_name(func_id), func_id);
2408 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
2412 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
2414 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
2416 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
2418 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
2420 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
2423 /* We only support one arg being in raw mode at the moment,
2424 * which is sufficient for the helper functions we have
2430 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
2431 enum bpf_arg_type arg_next)
2433 return (arg_type_is_mem_ptr(arg_curr) &&
2434 !arg_type_is_mem_size(arg_next)) ||
2435 (!arg_type_is_mem_ptr(arg_curr) &&
2436 arg_type_is_mem_size(arg_next));
2439 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
2441 /* bpf_xxx(..., buf, len) call will access 'len'
2442 * bytes from memory 'buf'. Both arg types need
2443 * to be paired, so make sure there's no buggy
2444 * helper function specification.
2446 if (arg_type_is_mem_size(fn->arg1_type) ||
2447 arg_type_is_mem_ptr(fn->arg5_type) ||
2448 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
2449 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
2450 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
2451 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
2457 static bool check_refcount_ok(const struct bpf_func_proto *fn)
2461 if (arg_type_is_refcounted(fn->arg1_type))
2463 if (arg_type_is_refcounted(fn->arg2_type))
2465 if (arg_type_is_refcounted(fn->arg3_type))
2467 if (arg_type_is_refcounted(fn->arg4_type))
2469 if (arg_type_is_refcounted(fn->arg5_type))
2472 /* We only support one arg being unreferenced at the moment,
2473 * which is sufficient for the helper functions we have right now.
2478 static int check_func_proto(const struct bpf_func_proto *fn)
2480 return check_raw_mode_ok(fn) &&
2481 check_arg_pair_ok(fn) &&
2482 check_refcount_ok(fn) ? 0 : -EINVAL;
2485 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
2486 * are now invalid, so turn them into unknown SCALAR_VALUE.
2488 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
2489 struct bpf_func_state *state)
2491 struct bpf_reg_state *regs = state->regs, *reg;
2494 for (i = 0; i < MAX_BPF_REG; i++)
2495 if (reg_is_pkt_pointer_any(®s[i]))
2496 mark_reg_unknown(env, regs, i);
2498 bpf_for_each_spilled_reg(i, state, reg) {
2501 if (reg_is_pkt_pointer_any(reg))
2502 __mark_reg_unknown(reg);
2506 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
2508 struct bpf_verifier_state *vstate = env->cur_state;
2511 for (i = 0; i <= vstate->curframe; i++)
2512 __clear_all_pkt_pointers(env, vstate->frame[i]);
2515 static void release_reg_references(struct bpf_verifier_env *env,
2516 struct bpf_func_state *state, int id)
2518 struct bpf_reg_state *regs = state->regs, *reg;
2521 for (i = 0; i < MAX_BPF_REG; i++)
2522 if (regs[i].id == id)
2523 mark_reg_unknown(env, regs, i);
2525 bpf_for_each_spilled_reg(i, state, reg) {
2528 if (reg_is_refcounted(reg) && reg->id == id)
2529 __mark_reg_unknown(reg);
2533 /* The pointer with the specified id has released its reference to kernel
2534 * resources. Identify all copies of the same pointer and clear the reference.
2536 static int release_reference(struct bpf_verifier_env *env,
2537 struct bpf_call_arg_meta *meta)
2539 struct bpf_verifier_state *vstate = env->cur_state;
2542 for (i = 0; i <= vstate->curframe; i++)
2543 release_reg_references(env, vstate->frame[i], meta->ptr_id);
2545 return release_reference_state(env, meta->ptr_id);
2548 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
2551 struct bpf_verifier_state *state = env->cur_state;
2552 struct bpf_func_state *caller, *callee;
2553 int i, err, subprog, target_insn;
2555 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
2556 verbose(env, "the call stack of %d frames is too deep\n",
2557 state->curframe + 2);
2561 target_insn = *insn_idx + insn->imm;
2562 subprog = find_subprog(env, target_insn + 1);
2564 verbose(env, "verifier bug. No program starts at insn %d\n",
2569 caller = state->frame[state->curframe];
2570 if (state->frame[state->curframe + 1]) {
2571 verbose(env, "verifier bug. Frame %d already allocated\n",
2572 state->curframe + 1);
2576 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
2579 state->frame[state->curframe + 1] = callee;
2581 /* callee cannot access r0, r6 - r9 for reading and has to write
2582 * into its own stack before reading from it.
2583 * callee can read/write into caller's stack
2585 init_func_state(env, callee,
2586 /* remember the callsite, it will be used by bpf_exit */
2587 *insn_idx /* callsite */,
2588 state->curframe + 1 /* frameno within this callchain */,
2589 subprog /* subprog number within this prog */);
2591 /* Transfer references to the callee */
2592 err = transfer_reference_state(callee, caller);
2596 /* copy r1 - r5 args that callee can access. The copy includes parent
2597 * pointers, which connects us up to the liveness chain
2599 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
2600 callee->regs[i] = caller->regs[i];
2602 /* after the call registers r0 - r5 were scratched */
2603 for (i = 0; i < CALLER_SAVED_REGS; i++) {
2604 mark_reg_not_init(env, caller->regs, caller_saved[i]);
2605 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
2608 /* only increment it after check_reg_arg() finished */
2611 /* and go analyze first insn of the callee */
2612 *insn_idx = target_insn;
2614 if (env->log.level) {
2615 verbose(env, "caller:\n");
2616 print_verifier_state(env, caller);
2617 verbose(env, "callee:\n");
2618 print_verifier_state(env, callee);
2623 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
2625 struct bpf_verifier_state *state = env->cur_state;
2626 struct bpf_func_state *caller, *callee;
2627 struct bpf_reg_state *r0;
2630 callee = state->frame[state->curframe];
2631 r0 = &callee->regs[BPF_REG_0];
2632 if (r0->type == PTR_TO_STACK) {
2633 /* technically it's ok to return caller's stack pointer
2634 * (or caller's caller's pointer) back to the caller,
2635 * since these pointers are valid. Only current stack
2636 * pointer will be invalid as soon as function exits,
2637 * but let's be conservative
2639 verbose(env, "cannot return stack pointer to the caller\n");
2644 caller = state->frame[state->curframe];
2645 /* return to the caller whatever r0 had in the callee */
2646 caller->regs[BPF_REG_0] = *r0;
2648 /* Transfer references to the caller */
2649 err = transfer_reference_state(caller, callee);
2653 *insn_idx = callee->callsite + 1;
2654 if (env->log.level) {
2655 verbose(env, "returning from callee:\n");
2656 print_verifier_state(env, callee);
2657 verbose(env, "to caller at %d:\n", *insn_idx);
2658 print_verifier_state(env, caller);
2660 /* clear everything in the callee */
2661 free_func_state(callee);
2662 state->frame[state->curframe + 1] = NULL;
2666 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
2668 struct bpf_call_arg_meta *meta)
2670 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
2672 if (ret_type != RET_INTEGER ||
2673 (func_id != BPF_FUNC_get_stack &&
2674 func_id != BPF_FUNC_probe_read_str))
2677 ret_reg->smax_value = meta->msize_smax_value;
2678 ret_reg->umax_value = meta->msize_umax_value;
2679 __reg_deduce_bounds(ret_reg);
2680 __reg_bound_offset(ret_reg);
2684 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
2685 int func_id, int insn_idx)
2687 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
2689 if (func_id != BPF_FUNC_tail_call &&
2690 func_id != BPF_FUNC_map_lookup_elem &&
2691 func_id != BPF_FUNC_map_update_elem &&
2692 func_id != BPF_FUNC_map_delete_elem &&
2693 func_id != BPF_FUNC_map_push_elem &&
2694 func_id != BPF_FUNC_map_pop_elem &&
2695 func_id != BPF_FUNC_map_peek_elem)
2698 if (meta->map_ptr == NULL) {
2699 verbose(env, "kernel subsystem misconfigured verifier\n");
2703 if (!BPF_MAP_PTR(aux->map_state))
2704 bpf_map_ptr_store(aux, meta->map_ptr,
2705 meta->map_ptr->unpriv_array);
2706 else if (BPF_MAP_PTR(aux->map_state) != meta->map_ptr)
2707 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
2708 meta->map_ptr->unpriv_array);
2712 static int check_reference_leak(struct bpf_verifier_env *env)
2714 struct bpf_func_state *state = cur_func(env);
2717 for (i = 0; i < state->acquired_refs; i++) {
2718 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
2719 state->refs[i].id, state->refs[i].insn_idx);
2721 return state->acquired_refs ? -EINVAL : 0;
2724 static int check_helper_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
2726 const struct bpf_func_proto *fn = NULL;
2727 struct bpf_reg_state *regs;
2728 struct bpf_call_arg_meta meta;
2732 /* find function prototype */
2733 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
2734 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
2739 if (env->ops->get_func_proto)
2740 fn = env->ops->get_func_proto(func_id, env->prog);
2742 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
2747 /* eBPF programs must be GPL compatible to use GPL-ed functions */
2748 if (!env->prog->gpl_compatible && fn->gpl_only) {
2749 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
2753 /* With LD_ABS/IND some JITs save/restore skb from r1. */
2754 changes_data = bpf_helper_changes_pkt_data(fn->func);
2755 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
2756 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
2757 func_id_name(func_id), func_id);
2761 memset(&meta, 0, sizeof(meta));
2762 meta.pkt_access = fn->pkt_access;
2764 err = check_func_proto(fn);
2766 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
2767 func_id_name(func_id), func_id);
2772 err = check_func_arg(env, BPF_REG_1, fn->arg1_type, &meta);
2775 err = check_func_arg(env, BPF_REG_2, fn->arg2_type, &meta);
2778 err = check_func_arg(env, BPF_REG_3, fn->arg3_type, &meta);
2781 err = check_func_arg(env, BPF_REG_4, fn->arg4_type, &meta);
2784 err = check_func_arg(env, BPF_REG_5, fn->arg5_type, &meta);
2788 err = record_func_map(env, &meta, func_id, insn_idx);
2792 /* Mark slots with STACK_MISC in case of raw mode, stack offset
2793 * is inferred from register state.
2795 for (i = 0; i < meta.access_size; i++) {
2796 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
2797 BPF_WRITE, -1, false);
2802 if (func_id == BPF_FUNC_tail_call) {
2803 err = check_reference_leak(env);
2805 verbose(env, "tail_call would lead to reference leak\n");
2808 } else if (is_release_function(func_id)) {
2809 err = release_reference(env, &meta);
2814 regs = cur_regs(env);
2816 /* check that flags argument in get_local_storage(map, flags) is 0,
2817 * this is required because get_local_storage() can't return an error.
2819 if (func_id == BPF_FUNC_get_local_storage &&
2820 !register_is_null(®s[BPF_REG_2])) {
2821 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
2825 /* reset caller saved regs */
2826 for (i = 0; i < CALLER_SAVED_REGS; i++) {
2827 mark_reg_not_init(env, regs, caller_saved[i]);
2828 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
2831 /* update return register (already marked as written above) */
2832 if (fn->ret_type == RET_INTEGER) {
2833 /* sets type to SCALAR_VALUE */
2834 mark_reg_unknown(env, regs, BPF_REG_0);
2835 } else if (fn->ret_type == RET_VOID) {
2836 regs[BPF_REG_0].type = NOT_INIT;
2837 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL ||
2838 fn->ret_type == RET_PTR_TO_MAP_VALUE) {
2839 if (fn->ret_type == RET_PTR_TO_MAP_VALUE)
2840 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE;
2842 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
2843 /* There is no offset yet applied, variable or fixed */
2844 mark_reg_known_zero(env, regs, BPF_REG_0);
2845 /* remember map_ptr, so that check_map_access()
2846 * can check 'value_size' boundary of memory access
2847 * to map element returned from bpf_map_lookup_elem()
2849 if (meta.map_ptr == NULL) {
2851 "kernel subsystem misconfigured verifier\n");
2854 regs[BPF_REG_0].map_ptr = meta.map_ptr;
2855 regs[BPF_REG_0].id = ++env->id_gen;
2856 } else if (fn->ret_type == RET_PTR_TO_SOCKET_OR_NULL) {
2857 int id = acquire_reference_state(env, insn_idx);
2860 mark_reg_known_zero(env, regs, BPF_REG_0);
2861 regs[BPF_REG_0].type = PTR_TO_SOCKET_OR_NULL;
2862 regs[BPF_REG_0].id = id;
2864 verbose(env, "unknown return type %d of func %s#%d\n",
2865 fn->ret_type, func_id_name(func_id), func_id);
2869 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
2871 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
2875 if (func_id == BPF_FUNC_get_stack && !env->prog->has_callchain_buf) {
2876 const char *err_str;
2878 #ifdef CONFIG_PERF_EVENTS
2879 err = get_callchain_buffers(sysctl_perf_event_max_stack);
2880 err_str = "cannot get callchain buffer for func %s#%d\n";
2883 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
2886 verbose(env, err_str, func_id_name(func_id), func_id);
2890 env->prog->has_callchain_buf = true;
2894 clear_all_pkt_pointers(env);
2898 static bool signed_add_overflows(s64 a, s64 b)
2900 /* Do the add in u64, where overflow is well-defined */
2901 s64 res = (s64)((u64)a + (u64)b);
2908 static bool signed_sub_overflows(s64 a, s64 b)
2910 /* Do the sub in u64, where overflow is well-defined */
2911 s64 res = (s64)((u64)a - (u64)b);
2918 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
2919 const struct bpf_reg_state *reg,
2920 enum bpf_reg_type type)
2922 bool known = tnum_is_const(reg->var_off);
2923 s64 val = reg->var_off.value;
2924 s64 smin = reg->smin_value;
2926 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
2927 verbose(env, "math between %s pointer and %lld is not allowed\n",
2928 reg_type_str[type], val);
2932 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
2933 verbose(env, "%s pointer offset %d is not allowed\n",
2934 reg_type_str[type], reg->off);
2938 if (smin == S64_MIN) {
2939 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
2940 reg_type_str[type]);
2944 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
2945 verbose(env, "value %lld makes %s pointer be out of bounds\n",
2946 smin, reg_type_str[type]);
2953 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
2954 * Caller should also handle BPF_MOV case separately.
2955 * If we return -EACCES, caller may want to try again treating pointer as a
2956 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
2958 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
2959 struct bpf_insn *insn,
2960 const struct bpf_reg_state *ptr_reg,
2961 const struct bpf_reg_state *off_reg)
2963 struct bpf_verifier_state *vstate = env->cur_state;
2964 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2965 struct bpf_reg_state *regs = state->regs, *dst_reg;
2966 bool known = tnum_is_const(off_reg->var_off);
2967 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
2968 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
2969 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
2970 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
2971 u8 opcode = BPF_OP(insn->code);
2972 u32 dst = insn->dst_reg;
2974 dst_reg = ®s[dst];
2976 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
2977 smin_val > smax_val || umin_val > umax_val) {
2978 /* Taint dst register if offset had invalid bounds derived from
2979 * e.g. dead branches.
2981 __mark_reg_unknown(dst_reg);
2985 if (BPF_CLASS(insn->code) != BPF_ALU64) {
2986 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
2988 "R%d 32-bit pointer arithmetic prohibited\n",
2993 switch (ptr_reg->type) {
2994 case PTR_TO_MAP_VALUE_OR_NULL:
2995 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
2996 dst, reg_type_str[ptr_reg->type]);
2998 case CONST_PTR_TO_MAP:
2999 case PTR_TO_PACKET_END:
3001 case PTR_TO_SOCKET_OR_NULL:
3002 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
3003 dst, reg_type_str[ptr_reg->type]);
3009 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
3010 * The id may be overwritten later if we create a new variable offset.
3012 dst_reg->type = ptr_reg->type;
3013 dst_reg->id = ptr_reg->id;
3015 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
3016 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
3021 /* We can take a fixed offset as long as it doesn't overflow
3022 * the s32 'off' field
3024 if (known && (ptr_reg->off + smin_val ==
3025 (s64)(s32)(ptr_reg->off + smin_val))) {
3026 /* pointer += K. Accumulate it into fixed offset */
3027 dst_reg->smin_value = smin_ptr;
3028 dst_reg->smax_value = smax_ptr;
3029 dst_reg->umin_value = umin_ptr;
3030 dst_reg->umax_value = umax_ptr;
3031 dst_reg->var_off = ptr_reg->var_off;
3032 dst_reg->off = ptr_reg->off + smin_val;
3033 dst_reg->range = ptr_reg->range;
3036 /* A new variable offset is created. Note that off_reg->off
3037 * == 0, since it's a scalar.
3038 * dst_reg gets the pointer type and since some positive
3039 * integer value was added to the pointer, give it a new 'id'
3040 * if it's a PTR_TO_PACKET.
3041 * this creates a new 'base' pointer, off_reg (variable) gets
3042 * added into the variable offset, and we copy the fixed offset
3045 if (signed_add_overflows(smin_ptr, smin_val) ||
3046 signed_add_overflows(smax_ptr, smax_val)) {
3047 dst_reg->smin_value = S64_MIN;
3048 dst_reg->smax_value = S64_MAX;
3050 dst_reg->smin_value = smin_ptr + smin_val;
3051 dst_reg->smax_value = smax_ptr + smax_val;
3053 if (umin_ptr + umin_val < umin_ptr ||
3054 umax_ptr + umax_val < umax_ptr) {
3055 dst_reg->umin_value = 0;
3056 dst_reg->umax_value = U64_MAX;
3058 dst_reg->umin_value = umin_ptr + umin_val;
3059 dst_reg->umax_value = umax_ptr + umax_val;
3061 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
3062 dst_reg->off = ptr_reg->off;
3063 if (reg_is_pkt_pointer(ptr_reg)) {
3064 dst_reg->id = ++env->id_gen;
3065 /* something was added to pkt_ptr, set range to zero */
3070 if (dst_reg == off_reg) {
3071 /* scalar -= pointer. Creates an unknown scalar */
3072 verbose(env, "R%d tried to subtract pointer from scalar\n",
3076 /* We don't allow subtraction from FP, because (according to
3077 * test_verifier.c test "invalid fp arithmetic", JITs might not
3078 * be able to deal with it.
3080 if (ptr_reg->type == PTR_TO_STACK) {
3081 verbose(env, "R%d subtraction from stack pointer prohibited\n",
3085 if (known && (ptr_reg->off - smin_val ==
3086 (s64)(s32)(ptr_reg->off - smin_val))) {
3087 /* pointer -= K. Subtract it from fixed offset */
3088 dst_reg->smin_value = smin_ptr;
3089 dst_reg->smax_value = smax_ptr;
3090 dst_reg->umin_value = umin_ptr;
3091 dst_reg->umax_value = umax_ptr;
3092 dst_reg->var_off = ptr_reg->var_off;
3093 dst_reg->id = ptr_reg->id;
3094 dst_reg->off = ptr_reg->off - smin_val;
3095 dst_reg->range = ptr_reg->range;
3098 /* A new variable offset is created. If the subtrahend is known
3099 * nonnegative, then any reg->range we had before is still good.
3101 if (signed_sub_overflows(smin_ptr, smax_val) ||
3102 signed_sub_overflows(smax_ptr, smin_val)) {
3103 /* Overflow possible, we know nothing */
3104 dst_reg->smin_value = S64_MIN;
3105 dst_reg->smax_value = S64_MAX;
3107 dst_reg->smin_value = smin_ptr - smax_val;
3108 dst_reg->smax_value = smax_ptr - smin_val;
3110 if (umin_ptr < umax_val) {
3111 /* Overflow possible, we know nothing */
3112 dst_reg->umin_value = 0;
3113 dst_reg->umax_value = U64_MAX;
3115 /* Cannot overflow (as long as bounds are consistent) */
3116 dst_reg->umin_value = umin_ptr - umax_val;
3117 dst_reg->umax_value = umax_ptr - umin_val;
3119 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
3120 dst_reg->off = ptr_reg->off;
3121 if (reg_is_pkt_pointer(ptr_reg)) {
3122 dst_reg->id = ++env->id_gen;
3123 /* something was added to pkt_ptr, set range to zero */
3131 /* bitwise ops on pointers are troublesome, prohibit. */
3132 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
3133 dst, bpf_alu_string[opcode >> 4]);
3136 /* other operators (e.g. MUL,LSH) produce non-pointer results */
3137 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
3138 dst, bpf_alu_string[opcode >> 4]);
3142 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
3145 __update_reg_bounds(dst_reg);
3146 __reg_deduce_bounds(dst_reg);
3147 __reg_bound_offset(dst_reg);
3151 /* WARNING: This function does calculations on 64-bit values, but the actual
3152 * execution may occur on 32-bit values. Therefore, things like bitshifts
3153 * need extra checks in the 32-bit case.
3155 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
3156 struct bpf_insn *insn,
3157 struct bpf_reg_state *dst_reg,
3158 struct bpf_reg_state src_reg)
3160 struct bpf_reg_state *regs = cur_regs(env);
3161 u8 opcode = BPF_OP(insn->code);
3162 bool src_known, dst_known;
3163 s64 smin_val, smax_val;
3164 u64 umin_val, umax_val;
3165 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
3167 if (insn_bitness == 32) {
3168 /* Relevant for 32-bit RSH: Information can propagate towards
3169 * LSB, so it isn't sufficient to only truncate the output to
3172 coerce_reg_to_size(dst_reg, 4);
3173 coerce_reg_to_size(&src_reg, 4);
3176 smin_val = src_reg.smin_value;
3177 smax_val = src_reg.smax_value;
3178 umin_val = src_reg.umin_value;
3179 umax_val = src_reg.umax_value;
3180 src_known = tnum_is_const(src_reg.var_off);
3181 dst_known = tnum_is_const(dst_reg->var_off);
3183 if ((src_known && (smin_val != smax_val || umin_val != umax_val)) ||
3184 smin_val > smax_val || umin_val > umax_val) {
3185 /* Taint dst register if offset had invalid bounds derived from
3186 * e.g. dead branches.
3188 __mark_reg_unknown(dst_reg);
3193 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
3194 __mark_reg_unknown(dst_reg);
3200 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
3201 signed_add_overflows(dst_reg->smax_value, smax_val)) {
3202 dst_reg->smin_value = S64_MIN;
3203 dst_reg->smax_value = S64_MAX;
3205 dst_reg->smin_value += smin_val;
3206 dst_reg->smax_value += smax_val;
3208 if (dst_reg->umin_value + umin_val < umin_val ||
3209 dst_reg->umax_value + umax_val < umax_val) {
3210 dst_reg->umin_value = 0;
3211 dst_reg->umax_value = U64_MAX;
3213 dst_reg->umin_value += umin_val;
3214 dst_reg->umax_value += umax_val;
3216 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
3219 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
3220 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
3221 /* Overflow possible, we know nothing */
3222 dst_reg->smin_value = S64_MIN;
3223 dst_reg->smax_value = S64_MAX;
3225 dst_reg->smin_value -= smax_val;
3226 dst_reg->smax_value -= smin_val;
3228 if (dst_reg->umin_value < umax_val) {
3229 /* Overflow possible, we know nothing */
3230 dst_reg->umin_value = 0;
3231 dst_reg->umax_value = U64_MAX;
3233 /* Cannot overflow (as long as bounds are consistent) */
3234 dst_reg->umin_value -= umax_val;
3235 dst_reg->umax_value -= umin_val;
3237 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
3240 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
3241 if (smin_val < 0 || dst_reg->smin_value < 0) {
3242 /* Ain't nobody got time to multiply that sign */
3243 __mark_reg_unbounded(dst_reg);
3244 __update_reg_bounds(dst_reg);
3247 /* Both values are positive, so we can work with unsigned and
3248 * copy the result to signed (unless it exceeds S64_MAX).
3250 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
3251 /* Potential overflow, we know nothing */
3252 __mark_reg_unbounded(dst_reg);
3253 /* (except what we can learn from the var_off) */
3254 __update_reg_bounds(dst_reg);
3257 dst_reg->umin_value *= umin_val;
3258 dst_reg->umax_value *= umax_val;
3259 if (dst_reg->umax_value > S64_MAX) {
3260 /* Overflow possible, we know nothing */
3261 dst_reg->smin_value = S64_MIN;
3262 dst_reg->smax_value = S64_MAX;
3264 dst_reg->smin_value = dst_reg->umin_value;
3265 dst_reg->smax_value = dst_reg->umax_value;
3269 if (src_known && dst_known) {
3270 __mark_reg_known(dst_reg, dst_reg->var_off.value &
3271 src_reg.var_off.value);
3274 /* We get our minimum from the var_off, since that's inherently
3275 * bitwise. Our maximum is the minimum of the operands' maxima.
3277 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
3278 dst_reg->umin_value = dst_reg->var_off.value;
3279 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
3280 if (dst_reg->smin_value < 0 || smin_val < 0) {
3281 /* Lose signed bounds when ANDing negative numbers,
3282 * ain't nobody got time for that.
3284 dst_reg->smin_value = S64_MIN;
3285 dst_reg->smax_value = S64_MAX;
3287 /* ANDing two positives gives a positive, so safe to
3288 * cast result into s64.
3290 dst_reg->smin_value = dst_reg->umin_value;
3291 dst_reg->smax_value = dst_reg->umax_value;
3293 /* We may learn something more from the var_off */
3294 __update_reg_bounds(dst_reg);
3297 if (src_known && dst_known) {
3298 __mark_reg_known(dst_reg, dst_reg->var_off.value |
3299 src_reg.var_off.value);
3302 /* We get our maximum from the var_off, and our minimum is the
3303 * maximum of the operands' minima
3305 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
3306 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
3307 dst_reg->umax_value = dst_reg->var_off.value |
3308 dst_reg->var_off.mask;
3309 if (dst_reg->smin_value < 0 || smin_val < 0) {
3310 /* Lose signed bounds when ORing negative numbers,
3311 * ain't nobody got time for that.
3313 dst_reg->smin_value = S64_MIN;
3314 dst_reg->smax_value = S64_MAX;
3316 /* ORing two positives gives a positive, so safe to
3317 * cast result into s64.
3319 dst_reg->smin_value = dst_reg->umin_value;
3320 dst_reg->smax_value = dst_reg->umax_value;
3322 /* We may learn something more from the var_off */
3323 __update_reg_bounds(dst_reg);
3326 if (umax_val >= insn_bitness) {
3327 /* Shifts greater than 31 or 63 are undefined.
3328 * This includes shifts by a negative number.
3330 mark_reg_unknown(env, regs, insn->dst_reg);
3333 /* We lose all sign bit information (except what we can pick
3336 dst_reg->smin_value = S64_MIN;
3337 dst_reg->smax_value = S64_MAX;
3338 /* If we might shift our top bit out, then we know nothing */
3339 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
3340 dst_reg->umin_value = 0;
3341 dst_reg->umax_value = U64_MAX;
3343 dst_reg->umin_value <<= umin_val;
3344 dst_reg->umax_value <<= umax_val;
3346 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
3347 /* We may learn something more from the var_off */
3348 __update_reg_bounds(dst_reg);
3351 if (umax_val >= insn_bitness) {
3352 /* Shifts greater than 31 or 63 are undefined.
3353 * This includes shifts by a negative number.
3355 mark_reg_unknown(env, regs, insn->dst_reg);
3358 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
3359 * be negative, then either:
3360 * 1) src_reg might be zero, so the sign bit of the result is
3361 * unknown, so we lose our signed bounds
3362 * 2) it's known negative, thus the unsigned bounds capture the
3364 * 3) the signed bounds cross zero, so they tell us nothing
3366 * If the value in dst_reg is known nonnegative, then again the
3367 * unsigned bounts capture the signed bounds.
3368 * Thus, in all cases it suffices to blow away our signed bounds
3369 * and rely on inferring new ones from the unsigned bounds and
3370 * var_off of the result.
3372 dst_reg->smin_value = S64_MIN;
3373 dst_reg->smax_value = S64_MAX;
3374 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
3375 dst_reg->umin_value >>= umax_val;
3376 dst_reg->umax_value >>= umin_val;
3377 /* We may learn something more from the var_off */
3378 __update_reg_bounds(dst_reg);
3381 if (umax_val >= insn_bitness) {
3382 /* Shifts greater than 31 or 63 are undefined.
3383 * This includes shifts by a negative number.
3385 mark_reg_unknown(env, regs, insn->dst_reg);
3389 /* Upon reaching here, src_known is true and
3390 * umax_val is equal to umin_val.
3392 dst_reg->smin_value >>= umin_val;
3393 dst_reg->smax_value >>= umin_val;
3394 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val);
3396 /* blow away the dst_reg umin_value/umax_value and rely on
3397 * dst_reg var_off to refine the result.
3399 dst_reg->umin_value = 0;
3400 dst_reg->umax_value = U64_MAX;
3401 __update_reg_bounds(dst_reg);
3404 mark_reg_unknown(env, regs, insn->dst_reg);
3408 if (BPF_CLASS(insn->code) != BPF_ALU64) {
3409 /* 32-bit ALU ops are (32,32)->32 */
3410 coerce_reg_to_size(dst_reg, 4);
3413 __reg_deduce_bounds(dst_reg);
3414 __reg_bound_offset(dst_reg);
3418 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
3421 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
3422 struct bpf_insn *insn)
3424 struct bpf_verifier_state *vstate = env->cur_state;
3425 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3426 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
3427 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
3428 u8 opcode = BPF_OP(insn->code);
3430 dst_reg = ®s[insn->dst_reg];
3432 if (dst_reg->type != SCALAR_VALUE)
3434 if (BPF_SRC(insn->code) == BPF_X) {
3435 src_reg = ®s[insn->src_reg];
3436 if (src_reg->type != SCALAR_VALUE) {
3437 if (dst_reg->type != SCALAR_VALUE) {
3438 /* Combining two pointers by any ALU op yields
3439 * an arbitrary scalar. Disallow all math except
3440 * pointer subtraction
3442 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
3443 mark_reg_unknown(env, regs, insn->dst_reg);
3446 verbose(env, "R%d pointer %s pointer prohibited\n",
3448 bpf_alu_string[opcode >> 4]);
3451 /* scalar += pointer
3452 * This is legal, but we have to reverse our
3453 * src/dest handling in computing the range
3455 return adjust_ptr_min_max_vals(env, insn,
3458 } else if (ptr_reg) {
3459 /* pointer += scalar */
3460 return adjust_ptr_min_max_vals(env, insn,
3464 /* Pretend the src is a reg with a known value, since we only
3465 * need to be able to read from this state.
3467 off_reg.type = SCALAR_VALUE;
3468 __mark_reg_known(&off_reg, insn->imm);
3470 if (ptr_reg) /* pointer += K */
3471 return adjust_ptr_min_max_vals(env, insn,
3475 /* Got here implies adding two SCALAR_VALUEs */
3476 if (WARN_ON_ONCE(ptr_reg)) {
3477 print_verifier_state(env, state);
3478 verbose(env, "verifier internal error: unexpected ptr_reg\n");
3481 if (WARN_ON(!src_reg)) {
3482 print_verifier_state(env, state);
3483 verbose(env, "verifier internal error: no src_reg\n");
3486 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
3489 /* check validity of 32-bit and 64-bit arithmetic operations */
3490 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
3492 struct bpf_reg_state *regs = cur_regs(env);
3493 u8 opcode = BPF_OP(insn->code);
3496 if (opcode == BPF_END || opcode == BPF_NEG) {
3497 if (opcode == BPF_NEG) {
3498 if (BPF_SRC(insn->code) != 0 ||
3499 insn->src_reg != BPF_REG_0 ||
3500 insn->off != 0 || insn->imm != 0) {
3501 verbose(env, "BPF_NEG uses reserved fields\n");
3505 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
3506 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
3507 BPF_CLASS(insn->code) == BPF_ALU64) {
3508 verbose(env, "BPF_END uses reserved fields\n");
3513 /* check src operand */
3514 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3518 if (is_pointer_value(env, insn->dst_reg)) {
3519 verbose(env, "R%d pointer arithmetic prohibited\n",
3524 /* check dest operand */
3525 err = check_reg_arg(env, insn->dst_reg, DST_OP);
3529 } else if (opcode == BPF_MOV) {
3531 if (BPF_SRC(insn->code) == BPF_X) {
3532 if (insn->imm != 0 || insn->off != 0) {
3533 verbose(env, "BPF_MOV uses reserved fields\n");
3537 /* check src operand */
3538 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3542 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
3543 verbose(env, "BPF_MOV uses reserved fields\n");
3548 /* check dest operand, mark as required later */
3549 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
3553 if (BPF_SRC(insn->code) == BPF_X) {
3554 if (BPF_CLASS(insn->code) == BPF_ALU64) {
3556 * copy register state to dest reg
3558 regs[insn->dst_reg] = regs[insn->src_reg];
3559 regs[insn->dst_reg].live |= REG_LIVE_WRITTEN;
3562 if (is_pointer_value(env, insn->src_reg)) {
3564 "R%d partial copy of pointer\n",
3568 mark_reg_unknown(env, regs, insn->dst_reg);
3569 coerce_reg_to_size(®s[insn->dst_reg], 4);
3573 * remember the value we stored into this reg
3575 /* clear any state __mark_reg_known doesn't set */
3576 mark_reg_unknown(env, regs, insn->dst_reg);
3577 regs[insn->dst_reg].type = SCALAR_VALUE;
3578 if (BPF_CLASS(insn->code) == BPF_ALU64) {
3579 __mark_reg_known(regs + insn->dst_reg,
3582 __mark_reg_known(regs + insn->dst_reg,
3587 } else if (opcode > BPF_END) {
3588 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
3591 } else { /* all other ALU ops: and, sub, xor, add, ... */
3593 if (BPF_SRC(insn->code) == BPF_X) {
3594 if (insn->imm != 0 || insn->off != 0) {
3595 verbose(env, "BPF_ALU uses reserved fields\n");
3598 /* check src1 operand */
3599 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3603 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
3604 verbose(env, "BPF_ALU uses reserved fields\n");
3609 /* check src2 operand */
3610 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3614 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
3615 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
3616 verbose(env, "div by zero\n");
3620 if (opcode == BPF_ARSH && BPF_CLASS(insn->code) != BPF_ALU64) {
3621 verbose(env, "BPF_ARSH not supported for 32 bit ALU\n");
3625 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
3626 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
3627 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
3629 if (insn->imm < 0 || insn->imm >= size) {
3630 verbose(env, "invalid shift %d\n", insn->imm);
3635 /* check dest operand */
3636 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
3640 return adjust_reg_min_max_vals(env, insn);
3646 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
3647 struct bpf_reg_state *dst_reg,
3648 enum bpf_reg_type type,
3649 bool range_right_open)
3651 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3652 struct bpf_reg_state *regs = state->regs, *reg;
3656 if (dst_reg->off < 0 ||
3657 (dst_reg->off == 0 && range_right_open))
3658 /* This doesn't give us any range */
3661 if (dst_reg->umax_value > MAX_PACKET_OFF ||
3662 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
3663 /* Risk of overflow. For instance, ptr + (1<<63) may be less
3664 * than pkt_end, but that's because it's also less than pkt.
3668 new_range = dst_reg->off;
3669 if (range_right_open)
3672 /* Examples for register markings:
3674 * pkt_data in dst register:
3678 * if (r2 > pkt_end) goto <handle exception>
3683 * if (r2 < pkt_end) goto <access okay>
3684 * <handle exception>
3687 * r2 == dst_reg, pkt_end == src_reg
3688 * r2=pkt(id=n,off=8,r=0)
3689 * r3=pkt(id=n,off=0,r=0)
3691 * pkt_data in src register:
3695 * if (pkt_end >= r2) goto <access okay>
3696 * <handle exception>
3700 * if (pkt_end <= r2) goto <handle exception>
3704 * pkt_end == dst_reg, r2 == src_reg
3705 * r2=pkt(id=n,off=8,r=0)
3706 * r3=pkt(id=n,off=0,r=0)
3708 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
3709 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
3710 * and [r3, r3 + 8-1) respectively is safe to access depending on
3714 /* If our ids match, then we must have the same max_value. And we
3715 * don't care about the other reg's fixed offset, since if it's too big
3716 * the range won't allow anything.
3717 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
3719 for (i = 0; i < MAX_BPF_REG; i++)
3720 if (regs[i].type == type && regs[i].id == dst_reg->id)
3721 /* keep the maximum range already checked */
3722 regs[i].range = max(regs[i].range, new_range);
3724 for (j = 0; j <= vstate->curframe; j++) {
3725 state = vstate->frame[j];
3726 bpf_for_each_spilled_reg(i, state, reg) {
3729 if (reg->type == type && reg->id == dst_reg->id)
3730 reg->range = max(reg->range, new_range);
3735 /* Adjusts the register min/max values in the case that the dst_reg is the
3736 * variable register that we are working on, and src_reg is a constant or we're
3737 * simply doing a BPF_K check.
3738 * In JEQ/JNE cases we also adjust the var_off values.
3740 static void reg_set_min_max(struct bpf_reg_state *true_reg,
3741 struct bpf_reg_state *false_reg, u64 val,
3744 /* If the dst_reg is a pointer, we can't learn anything about its
3745 * variable offset from the compare (unless src_reg were a pointer into
3746 * the same object, but we don't bother with that.
3747 * Since false_reg and true_reg have the same type by construction, we
3748 * only need to check one of them for pointerness.
3750 if (__is_pointer_value(false, false_reg))
3755 /* If this is false then we know nothing Jon Snow, but if it is
3756 * true then we know for sure.
3758 __mark_reg_known(true_reg, val);
3761 /* If this is true we know nothing Jon Snow, but if it is false
3762 * we know the value for sure;
3764 __mark_reg_known(false_reg, val);
3767 false_reg->umax_value = min(false_reg->umax_value, val);
3768 true_reg->umin_value = max(true_reg->umin_value, val + 1);
3771 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
3772 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
3775 false_reg->umin_value = max(false_reg->umin_value, val);
3776 true_reg->umax_value = min(true_reg->umax_value, val - 1);
3779 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
3780 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
3783 false_reg->umax_value = min(false_reg->umax_value, val - 1);
3784 true_reg->umin_value = max(true_reg->umin_value, val);
3787 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
3788 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
3791 false_reg->umin_value = max(false_reg->umin_value, val + 1);
3792 true_reg->umax_value = min(true_reg->umax_value, val);
3795 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
3796 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
3802 __reg_deduce_bounds(false_reg);
3803 __reg_deduce_bounds(true_reg);
3804 /* We might have learned some bits from the bounds. */
3805 __reg_bound_offset(false_reg);
3806 __reg_bound_offset(true_reg);
3807 /* Intersecting with the old var_off might have improved our bounds
3808 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
3809 * then new var_off is (0; 0x7f...fc) which improves our umax.
3811 __update_reg_bounds(false_reg);
3812 __update_reg_bounds(true_reg);
3815 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
3818 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
3819 struct bpf_reg_state *false_reg, u64 val,
3822 if (__is_pointer_value(false, false_reg))
3827 /* If this is false then we know nothing Jon Snow, but if it is
3828 * true then we know for sure.
3830 __mark_reg_known(true_reg, val);
3833 /* If this is true we know nothing Jon Snow, but if it is false
3834 * we know the value for sure;
3836 __mark_reg_known(false_reg, val);
3839 true_reg->umax_value = min(true_reg->umax_value, val - 1);
3840 false_reg->umin_value = max(false_reg->umin_value, val);
3843 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
3844 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
3847 true_reg->umin_value = max(true_reg->umin_value, val + 1);
3848 false_reg->umax_value = min(false_reg->umax_value, val);
3851 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
3852 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
3855 true_reg->umax_value = min(true_reg->umax_value, val);
3856 false_reg->umin_value = max(false_reg->umin_value, val + 1);
3859 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
3860 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
3863 true_reg->umin_value = max(true_reg->umin_value, val);
3864 false_reg->umax_value = min(false_reg->umax_value, val - 1);
3867 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
3868 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
3874 __reg_deduce_bounds(false_reg);
3875 __reg_deduce_bounds(true_reg);
3876 /* We might have learned some bits from the bounds. */
3877 __reg_bound_offset(false_reg);
3878 __reg_bound_offset(true_reg);
3879 /* Intersecting with the old var_off might have improved our bounds
3880 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
3881 * then new var_off is (0; 0x7f...fc) which improves our umax.
3883 __update_reg_bounds(false_reg);
3884 __update_reg_bounds(true_reg);
3887 /* Regs are known to be equal, so intersect their min/max/var_off */
3888 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
3889 struct bpf_reg_state *dst_reg)
3891 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
3892 dst_reg->umin_value);
3893 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
3894 dst_reg->umax_value);
3895 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
3896 dst_reg->smin_value);
3897 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
3898 dst_reg->smax_value);
3899 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
3901 /* We might have learned new bounds from the var_off. */
3902 __update_reg_bounds(src_reg);
3903 __update_reg_bounds(dst_reg);
3904 /* We might have learned something about the sign bit. */
3905 __reg_deduce_bounds(src_reg);
3906 __reg_deduce_bounds(dst_reg);
3907 /* We might have learned some bits from the bounds. */
3908 __reg_bound_offset(src_reg);
3909 __reg_bound_offset(dst_reg);
3910 /* Intersecting with the old var_off might have improved our bounds
3911 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
3912 * then new var_off is (0; 0x7f...fc) which improves our umax.
3914 __update_reg_bounds(src_reg);
3915 __update_reg_bounds(dst_reg);
3918 static void reg_combine_min_max(struct bpf_reg_state *true_src,
3919 struct bpf_reg_state *true_dst,
3920 struct bpf_reg_state *false_src,
3921 struct bpf_reg_state *false_dst,
3926 __reg_combine_min_max(true_src, true_dst);
3929 __reg_combine_min_max(false_src, false_dst);
3934 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
3935 struct bpf_reg_state *reg, u32 id,
3938 if (reg_type_may_be_null(reg->type) && reg->id == id) {
3939 /* Old offset (both fixed and variable parts) should
3940 * have been known-zero, because we don't allow pointer
3941 * arithmetic on pointers that might be NULL.
3943 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
3944 !tnum_equals_const(reg->var_off, 0) ||
3946 __mark_reg_known_zero(reg);
3950 reg->type = SCALAR_VALUE;
3951 } else if (reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
3952 if (reg->map_ptr->inner_map_meta) {
3953 reg->type = CONST_PTR_TO_MAP;
3954 reg->map_ptr = reg->map_ptr->inner_map_meta;
3956 reg->type = PTR_TO_MAP_VALUE;
3958 } else if (reg->type == PTR_TO_SOCKET_OR_NULL) {
3959 reg->type = PTR_TO_SOCKET;
3961 if (is_null || !reg_is_refcounted(reg)) {
3962 /* We don't need id from this point onwards anymore,
3963 * thus we should better reset it, so that state
3964 * pruning has chances to take effect.
3971 /* The logic is similar to find_good_pkt_pointers(), both could eventually
3972 * be folded together at some point.
3974 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
3977 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3978 struct bpf_reg_state *reg, *regs = state->regs;
3979 u32 id = regs[regno].id;
3982 if (reg_is_refcounted_or_null(®s[regno]) && is_null)
3983 __release_reference_state(state, id);
3985 for (i = 0; i < MAX_BPF_REG; i++)
3986 mark_ptr_or_null_reg(state, ®s[i], id, is_null);
3988 for (j = 0; j <= vstate->curframe; j++) {
3989 state = vstate->frame[j];
3990 bpf_for_each_spilled_reg(i, state, reg) {
3993 mark_ptr_or_null_reg(state, reg, id, is_null);
3998 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
3999 struct bpf_reg_state *dst_reg,
4000 struct bpf_reg_state *src_reg,
4001 struct bpf_verifier_state *this_branch,
4002 struct bpf_verifier_state *other_branch)
4004 if (BPF_SRC(insn->code) != BPF_X)
4007 switch (BPF_OP(insn->code)) {
4009 if ((dst_reg->type == PTR_TO_PACKET &&
4010 src_reg->type == PTR_TO_PACKET_END) ||
4011 (dst_reg->type == PTR_TO_PACKET_META &&
4012 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
4013 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
4014 find_good_pkt_pointers(this_branch, dst_reg,
4015 dst_reg->type, false);
4016 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
4017 src_reg->type == PTR_TO_PACKET) ||
4018 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
4019 src_reg->type == PTR_TO_PACKET_META)) {
4020 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
4021 find_good_pkt_pointers(other_branch, src_reg,
4022 src_reg->type, true);
4028 if ((dst_reg->type == PTR_TO_PACKET &&
4029 src_reg->type == PTR_TO_PACKET_END) ||
4030 (dst_reg->type == PTR_TO_PACKET_META &&
4031 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
4032 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
4033 find_good_pkt_pointers(other_branch, dst_reg,
4034 dst_reg->type, true);
4035 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
4036 src_reg->type == PTR_TO_PACKET) ||
4037 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
4038 src_reg->type == PTR_TO_PACKET_META)) {
4039 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
4040 find_good_pkt_pointers(this_branch, src_reg,
4041 src_reg->type, false);
4047 if ((dst_reg->type == PTR_TO_PACKET &&
4048 src_reg->type == PTR_TO_PACKET_END) ||
4049 (dst_reg->type == PTR_TO_PACKET_META &&
4050 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
4051 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
4052 find_good_pkt_pointers(this_branch, dst_reg,
4053 dst_reg->type, true);
4054 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
4055 src_reg->type == PTR_TO_PACKET) ||
4056 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
4057 src_reg->type == PTR_TO_PACKET_META)) {
4058 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
4059 find_good_pkt_pointers(other_branch, src_reg,
4060 src_reg->type, false);
4066 if ((dst_reg->type == PTR_TO_PACKET &&
4067 src_reg->type == PTR_TO_PACKET_END) ||
4068 (dst_reg->type == PTR_TO_PACKET_META &&
4069 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
4070 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
4071 find_good_pkt_pointers(other_branch, dst_reg,
4072 dst_reg->type, false);
4073 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
4074 src_reg->type == PTR_TO_PACKET) ||
4075 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
4076 src_reg->type == PTR_TO_PACKET_META)) {
4077 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
4078 find_good_pkt_pointers(this_branch, src_reg,
4079 src_reg->type, true);
4091 static int check_cond_jmp_op(struct bpf_verifier_env *env,
4092 struct bpf_insn *insn, int *insn_idx)
4094 struct bpf_verifier_state *this_branch = env->cur_state;
4095 struct bpf_verifier_state *other_branch;
4096 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
4097 struct bpf_reg_state *dst_reg, *other_branch_regs;
4098 u8 opcode = BPF_OP(insn->code);
4101 if (opcode > BPF_JSLE) {
4102 verbose(env, "invalid BPF_JMP opcode %x\n", opcode);
4106 if (BPF_SRC(insn->code) == BPF_X) {
4107 if (insn->imm != 0) {
4108 verbose(env, "BPF_JMP uses reserved fields\n");
4112 /* check src1 operand */
4113 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4117 if (is_pointer_value(env, insn->src_reg)) {
4118 verbose(env, "R%d pointer comparison prohibited\n",
4123 if (insn->src_reg != BPF_REG_0) {
4124 verbose(env, "BPF_JMP uses reserved fields\n");
4129 /* check src2 operand */
4130 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4134 dst_reg = ®s[insn->dst_reg];
4136 /* detect if R == 0 where R was initialized to zero earlier */
4137 if (BPF_SRC(insn->code) == BPF_K &&
4138 (opcode == BPF_JEQ || opcode == BPF_JNE) &&
4139 dst_reg->type == SCALAR_VALUE &&
4140 tnum_is_const(dst_reg->var_off)) {
4141 if ((opcode == BPF_JEQ && dst_reg->var_off.value == insn->imm) ||
4142 (opcode == BPF_JNE && dst_reg->var_off.value != insn->imm)) {
4143 /* if (imm == imm) goto pc+off;
4144 * only follow the goto, ignore fall-through
4146 *insn_idx += insn->off;
4149 /* if (imm != imm) goto pc+off;
4150 * only follow fall-through branch, since
4151 * that's where the program will go
4157 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx);
4160 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
4162 /* detect if we are comparing against a constant value so we can adjust
4163 * our min/max values for our dst register.
4164 * this is only legit if both are scalars (or pointers to the same
4165 * object, I suppose, but we don't support that right now), because
4166 * otherwise the different base pointers mean the offsets aren't
4169 if (BPF_SRC(insn->code) == BPF_X) {
4170 if (dst_reg->type == SCALAR_VALUE &&
4171 regs[insn->src_reg].type == SCALAR_VALUE) {
4172 if (tnum_is_const(regs[insn->src_reg].var_off))
4173 reg_set_min_max(&other_branch_regs[insn->dst_reg],
4174 dst_reg, regs[insn->src_reg].var_off.value,
4176 else if (tnum_is_const(dst_reg->var_off))
4177 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
4178 ®s[insn->src_reg],
4179 dst_reg->var_off.value, opcode);
4180 else if (opcode == BPF_JEQ || opcode == BPF_JNE)
4181 /* Comparing for equality, we can combine knowledge */
4182 reg_combine_min_max(&other_branch_regs[insn->src_reg],
4183 &other_branch_regs[insn->dst_reg],
4184 ®s[insn->src_reg],
4185 ®s[insn->dst_reg], opcode);
4187 } else if (dst_reg->type == SCALAR_VALUE) {
4188 reg_set_min_max(&other_branch_regs[insn->dst_reg],
4189 dst_reg, insn->imm, opcode);
4192 /* detect if R == 0 where R is returned from bpf_map_lookup_elem() */
4193 if (BPF_SRC(insn->code) == BPF_K &&
4194 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
4195 reg_type_may_be_null(dst_reg->type)) {
4196 /* Mark all identical registers in each branch as either
4197 * safe or unknown depending R == 0 or R != 0 conditional.
4199 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
4201 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
4203 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
4204 this_branch, other_branch) &&
4205 is_pointer_value(env, insn->dst_reg)) {
4206 verbose(env, "R%d pointer comparison prohibited\n",
4211 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
4215 /* return the map pointer stored inside BPF_LD_IMM64 instruction */
4216 static struct bpf_map *ld_imm64_to_map_ptr(struct bpf_insn *insn)
4218 u64 imm64 = ((u64) (u32) insn[0].imm) | ((u64) (u32) insn[1].imm) << 32;
4220 return (struct bpf_map *) (unsigned long) imm64;
4223 /* verify BPF_LD_IMM64 instruction */
4224 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
4226 struct bpf_reg_state *regs = cur_regs(env);
4229 if (BPF_SIZE(insn->code) != BPF_DW) {
4230 verbose(env, "invalid BPF_LD_IMM insn\n");
4233 if (insn->off != 0) {
4234 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
4238 err = check_reg_arg(env, insn->dst_reg, DST_OP);
4242 if (insn->src_reg == 0) {
4243 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
4245 regs[insn->dst_reg].type = SCALAR_VALUE;
4246 __mark_reg_known(®s[insn->dst_reg], imm);
4250 /* replace_map_fd_with_map_ptr() should have caught bad ld_imm64 */
4251 BUG_ON(insn->src_reg != BPF_PSEUDO_MAP_FD);
4253 regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
4254 regs[insn->dst_reg].map_ptr = ld_imm64_to_map_ptr(insn);
4258 static bool may_access_skb(enum bpf_prog_type type)
4261 case BPF_PROG_TYPE_SOCKET_FILTER:
4262 case BPF_PROG_TYPE_SCHED_CLS:
4263 case BPF_PROG_TYPE_SCHED_ACT:
4270 /* verify safety of LD_ABS|LD_IND instructions:
4271 * - they can only appear in the programs where ctx == skb
4272 * - since they are wrappers of function calls, they scratch R1-R5 registers,
4273 * preserve R6-R9, and store return value into R0
4276 * ctx == skb == R6 == CTX
4279 * SRC == any register
4280 * IMM == 32-bit immediate
4283 * R0 - 8/16/32-bit skb data converted to cpu endianness
4285 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
4287 struct bpf_reg_state *regs = cur_regs(env);
4288 u8 mode = BPF_MODE(insn->code);
4291 if (!may_access_skb(env->prog->type)) {
4292 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
4296 if (!env->ops->gen_ld_abs) {
4297 verbose(env, "bpf verifier is misconfigured\n");
4301 if (env->subprog_cnt > 1) {
4302 /* when program has LD_ABS insn JITs and interpreter assume
4303 * that r1 == ctx == skb which is not the case for callees
4304 * that can have arbitrary arguments. It's problematic
4305 * for main prog as well since JITs would need to analyze
4306 * all functions in order to make proper register save/restore
4307 * decisions in the main prog. Hence disallow LD_ABS with calls
4309 verbose(env, "BPF_LD_[ABS|IND] instructions cannot be mixed with bpf-to-bpf calls\n");
4313 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
4314 BPF_SIZE(insn->code) == BPF_DW ||
4315 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
4316 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
4320 /* check whether implicit source operand (register R6) is readable */
4321 err = check_reg_arg(env, BPF_REG_6, SRC_OP);
4325 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
4326 * gen_ld_abs() may terminate the program at runtime, leading to
4329 err = check_reference_leak(env);
4331 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
4335 if (regs[BPF_REG_6].type != PTR_TO_CTX) {
4337 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
4341 if (mode == BPF_IND) {
4342 /* check explicit source operand */
4343 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4348 /* reset caller saved regs to unreadable */
4349 for (i = 0; i < CALLER_SAVED_REGS; i++) {
4350 mark_reg_not_init(env, regs, caller_saved[i]);
4351 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
4354 /* mark destination R0 register as readable, since it contains
4355 * the value fetched from the packet.
4356 * Already marked as written above.
4358 mark_reg_unknown(env, regs, BPF_REG_0);
4362 static int check_return_code(struct bpf_verifier_env *env)
4364 struct bpf_reg_state *reg;
4365 struct tnum range = tnum_range(0, 1);
4367 switch (env->prog->type) {
4368 case BPF_PROG_TYPE_CGROUP_SKB:
4369 case BPF_PROG_TYPE_CGROUP_SOCK:
4370 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
4371 case BPF_PROG_TYPE_SOCK_OPS:
4372 case BPF_PROG_TYPE_CGROUP_DEVICE:
4378 reg = cur_regs(env) + BPF_REG_0;
4379 if (reg->type != SCALAR_VALUE) {
4380 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
4381 reg_type_str[reg->type]);
4385 if (!tnum_in(range, reg->var_off)) {
4386 verbose(env, "At program exit the register R0 ");
4387 if (!tnum_is_unknown(reg->var_off)) {
4390 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4391 verbose(env, "has value %s", tn_buf);
4393 verbose(env, "has unknown scalar value");
4395 verbose(env, " should have been 0 or 1\n");
4401 /* non-recursive DFS pseudo code
4402 * 1 procedure DFS-iterative(G,v):
4403 * 2 label v as discovered
4404 * 3 let S be a stack
4406 * 5 while S is not empty
4408 * 7 if t is what we're looking for:
4410 * 9 for all edges e in G.adjacentEdges(t) do
4411 * 10 if edge e is already labelled
4412 * 11 continue with the next edge
4413 * 12 w <- G.adjacentVertex(t,e)
4414 * 13 if vertex w is not discovered and not explored
4415 * 14 label e as tree-edge
4416 * 15 label w as discovered
4419 * 18 else if vertex w is discovered
4420 * 19 label e as back-edge
4422 * 21 // vertex w is explored
4423 * 22 label e as forward- or cross-edge
4424 * 23 label t as explored
4429 * 0x11 - discovered and fall-through edge labelled
4430 * 0x12 - discovered and fall-through and branch edges labelled
4441 #define STATE_LIST_MARK ((struct bpf_verifier_state_list *) -1L)
4443 static int *insn_stack; /* stack of insns to process */
4444 static int cur_stack; /* current stack index */
4445 static int *insn_state;
4447 /* t, w, e - match pseudo-code above:
4448 * t - index of current instruction
4449 * w - next instruction
4452 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env)
4454 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
4457 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
4460 if (w < 0 || w >= env->prog->len) {
4461 verbose(env, "jump out of range from insn %d to %d\n", t, w);
4466 /* mark branch target for state pruning */
4467 env->explored_states[w] = STATE_LIST_MARK;
4469 if (insn_state[w] == 0) {
4471 insn_state[t] = DISCOVERED | e;
4472 insn_state[w] = DISCOVERED;
4473 if (cur_stack >= env->prog->len)
4475 insn_stack[cur_stack++] = w;
4477 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
4478 verbose(env, "back-edge from insn %d to %d\n", t, w);
4480 } else if (insn_state[w] == EXPLORED) {
4481 /* forward- or cross-edge */
4482 insn_state[t] = DISCOVERED | e;
4484 verbose(env, "insn state internal bug\n");
4490 /* non-recursive depth-first-search to detect loops in BPF program
4491 * loop == back-edge in directed graph
4493 static int check_cfg(struct bpf_verifier_env *env)
4495 struct bpf_insn *insns = env->prog->insnsi;
4496 int insn_cnt = env->prog->len;
4500 ret = check_subprogs(env);
4504 insn_state = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
4508 insn_stack = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
4514 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
4515 insn_stack[0] = 0; /* 0 is the first instruction */
4521 t = insn_stack[cur_stack - 1];
4523 if (BPF_CLASS(insns[t].code) == BPF_JMP) {
4524 u8 opcode = BPF_OP(insns[t].code);
4526 if (opcode == BPF_EXIT) {
4528 } else if (opcode == BPF_CALL) {
4529 ret = push_insn(t, t + 1, FALLTHROUGH, env);
4534 if (t + 1 < insn_cnt)
4535 env->explored_states[t + 1] = STATE_LIST_MARK;
4536 if (insns[t].src_reg == BPF_PSEUDO_CALL) {
4537 env->explored_states[t] = STATE_LIST_MARK;
4538 ret = push_insn(t, t + insns[t].imm + 1, BRANCH, env);
4544 } else if (opcode == BPF_JA) {
4545 if (BPF_SRC(insns[t].code) != BPF_K) {
4549 /* unconditional jump with single edge */
4550 ret = push_insn(t, t + insns[t].off + 1,
4556 /* tell verifier to check for equivalent states
4557 * after every call and jump
4559 if (t + 1 < insn_cnt)
4560 env->explored_states[t + 1] = STATE_LIST_MARK;
4562 /* conditional jump with two edges */
4563 env->explored_states[t] = STATE_LIST_MARK;
4564 ret = push_insn(t, t + 1, FALLTHROUGH, env);
4570 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env);
4577 /* all other non-branch instructions with single
4580 ret = push_insn(t, t + 1, FALLTHROUGH, env);
4588 insn_state[t] = EXPLORED;
4589 if (cur_stack-- <= 0) {
4590 verbose(env, "pop stack internal bug\n");
4597 for (i = 0; i < insn_cnt; i++) {
4598 if (insn_state[i] != EXPLORED) {
4599 verbose(env, "unreachable insn %d\n", i);
4604 ret = 0; /* cfg looks good */
4612 /* check %cur's range satisfies %old's */
4613 static bool range_within(struct bpf_reg_state *old,
4614 struct bpf_reg_state *cur)
4616 return old->umin_value <= cur->umin_value &&
4617 old->umax_value >= cur->umax_value &&
4618 old->smin_value <= cur->smin_value &&
4619 old->smax_value >= cur->smax_value;
4622 /* Maximum number of register states that can exist at once */
4623 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
4629 /* If in the old state two registers had the same id, then they need to have
4630 * the same id in the new state as well. But that id could be different from
4631 * the old state, so we need to track the mapping from old to new ids.
4632 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
4633 * regs with old id 5 must also have new id 9 for the new state to be safe. But
4634 * regs with a different old id could still have new id 9, we don't care about
4636 * So we look through our idmap to see if this old id has been seen before. If
4637 * so, we require the new id to match; otherwise, we add the id pair to the map.
4639 static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
4643 for (i = 0; i < ID_MAP_SIZE; i++) {
4644 if (!idmap[i].old) {
4645 /* Reached an empty slot; haven't seen this id before */
4646 idmap[i].old = old_id;
4647 idmap[i].cur = cur_id;
4650 if (idmap[i].old == old_id)
4651 return idmap[i].cur == cur_id;
4653 /* We ran out of idmap slots, which should be impossible */
4658 /* Returns true if (rold safe implies rcur safe) */
4659 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
4660 struct idpair *idmap)
4664 if (!(rold->live & REG_LIVE_READ))
4665 /* explored state didn't use this */
4668 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
4670 if (rold->type == PTR_TO_STACK)
4671 /* two stack pointers are equal only if they're pointing to
4672 * the same stack frame, since fp-8 in foo != fp-8 in bar
4674 return equal && rold->frameno == rcur->frameno;
4679 if (rold->type == NOT_INIT)
4680 /* explored state can't have used this */
4682 if (rcur->type == NOT_INIT)
4684 switch (rold->type) {
4686 if (rcur->type == SCALAR_VALUE) {
4687 /* new val must satisfy old val knowledge */
4688 return range_within(rold, rcur) &&
4689 tnum_in(rold->var_off, rcur->var_off);
4691 /* We're trying to use a pointer in place of a scalar.
4692 * Even if the scalar was unbounded, this could lead to
4693 * pointer leaks because scalars are allowed to leak
4694 * while pointers are not. We could make this safe in
4695 * special cases if root is calling us, but it's
4696 * probably not worth the hassle.
4700 case PTR_TO_MAP_VALUE:
4701 /* If the new min/max/var_off satisfy the old ones and
4702 * everything else matches, we are OK.
4703 * We don't care about the 'id' value, because nothing
4704 * uses it for PTR_TO_MAP_VALUE (only for ..._OR_NULL)
4706 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
4707 range_within(rold, rcur) &&
4708 tnum_in(rold->var_off, rcur->var_off);
4709 case PTR_TO_MAP_VALUE_OR_NULL:
4710 /* a PTR_TO_MAP_VALUE could be safe to use as a
4711 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
4712 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
4713 * checked, doing so could have affected others with the same
4714 * id, and we can't check for that because we lost the id when
4715 * we converted to a PTR_TO_MAP_VALUE.
4717 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
4719 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
4721 /* Check our ids match any regs they're supposed to */
4722 return check_ids(rold->id, rcur->id, idmap);
4723 case PTR_TO_PACKET_META:
4725 if (rcur->type != rold->type)
4727 /* We must have at least as much range as the old ptr
4728 * did, so that any accesses which were safe before are
4729 * still safe. This is true even if old range < old off,
4730 * since someone could have accessed through (ptr - k), or
4731 * even done ptr -= k in a register, to get a safe access.
4733 if (rold->range > rcur->range)
4735 /* If the offsets don't match, we can't trust our alignment;
4736 * nor can we be sure that we won't fall out of range.
4738 if (rold->off != rcur->off)
4740 /* id relations must be preserved */
4741 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
4743 /* new val must satisfy old val knowledge */
4744 return range_within(rold, rcur) &&
4745 tnum_in(rold->var_off, rcur->var_off);
4747 case CONST_PTR_TO_MAP:
4748 case PTR_TO_PACKET_END:
4749 case PTR_TO_FLOW_KEYS:
4751 case PTR_TO_SOCKET_OR_NULL:
4752 /* Only valid matches are exact, which memcmp() above
4753 * would have accepted
4756 /* Don't know what's going on, just say it's not safe */
4760 /* Shouldn't get here; if we do, say it's not safe */
4765 static bool stacksafe(struct bpf_func_state *old,
4766 struct bpf_func_state *cur,
4767 struct idpair *idmap)
4771 /* if explored stack has more populated slots than current stack
4772 * such stacks are not equivalent
4774 if (old->allocated_stack > cur->allocated_stack)
4777 /* walk slots of the explored stack and ignore any additional
4778 * slots in the current stack, since explored(safe) state
4781 for (i = 0; i < old->allocated_stack; i++) {
4782 spi = i / BPF_REG_SIZE;
4784 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ))
4785 /* explored state didn't use this */
4788 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
4790 /* if old state was safe with misc data in the stack
4791 * it will be safe with zero-initialized stack.
4792 * The opposite is not true
4794 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
4795 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
4797 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
4798 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
4799 /* Ex: old explored (safe) state has STACK_SPILL in
4800 * this stack slot, but current has has STACK_MISC ->
4801 * this verifier states are not equivalent,
4802 * return false to continue verification of this path
4805 if (i % BPF_REG_SIZE)
4807 if (old->stack[spi].slot_type[0] != STACK_SPILL)
4809 if (!regsafe(&old->stack[spi].spilled_ptr,
4810 &cur->stack[spi].spilled_ptr,
4812 /* when explored and current stack slot are both storing
4813 * spilled registers, check that stored pointers types
4814 * are the same as well.
4815 * Ex: explored safe path could have stored
4816 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
4817 * but current path has stored:
4818 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
4819 * such verifier states are not equivalent.
4820 * return false to continue verification of this path
4827 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
4829 if (old->acquired_refs != cur->acquired_refs)
4831 return !memcmp(old->refs, cur->refs,
4832 sizeof(*old->refs) * old->acquired_refs);
4835 /* compare two verifier states
4837 * all states stored in state_list are known to be valid, since
4838 * verifier reached 'bpf_exit' instruction through them
4840 * this function is called when verifier exploring different branches of
4841 * execution popped from the state stack. If it sees an old state that has
4842 * more strict register state and more strict stack state then this execution
4843 * branch doesn't need to be explored further, since verifier already
4844 * concluded that more strict state leads to valid finish.
4846 * Therefore two states are equivalent if register state is more conservative
4847 * and explored stack state is more conservative than the current one.
4850 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
4851 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
4853 * In other words if current stack state (one being explored) has more
4854 * valid slots than old one that already passed validation, it means
4855 * the verifier can stop exploring and conclude that current state is valid too
4857 * Similarly with registers. If explored state has register type as invalid
4858 * whereas register type in current state is meaningful, it means that
4859 * the current state will reach 'bpf_exit' instruction safely
4861 static bool func_states_equal(struct bpf_func_state *old,
4862 struct bpf_func_state *cur)
4864 struct idpair *idmap;
4868 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
4869 /* If we failed to allocate the idmap, just say it's not safe */
4873 for (i = 0; i < MAX_BPF_REG; i++) {
4874 if (!regsafe(&old->regs[i], &cur->regs[i], idmap))
4878 if (!stacksafe(old, cur, idmap))
4881 if (!refsafe(old, cur))
4889 static bool states_equal(struct bpf_verifier_env *env,
4890 struct bpf_verifier_state *old,
4891 struct bpf_verifier_state *cur)
4895 if (old->curframe != cur->curframe)
4898 /* for states to be equal callsites have to be the same
4899 * and all frame states need to be equivalent
4901 for (i = 0; i <= old->curframe; i++) {
4902 if (old->frame[i]->callsite != cur->frame[i]->callsite)
4904 if (!func_states_equal(old->frame[i], cur->frame[i]))
4910 /* A write screens off any subsequent reads; but write marks come from the
4911 * straight-line code between a state and its parent. When we arrive at an
4912 * equivalent state (jump target or such) we didn't arrive by the straight-line
4913 * code, so read marks in the state must propagate to the parent regardless
4914 * of the state's write marks. That's what 'parent == state->parent' comparison
4915 * in mark_reg_read() is for.
4917 static int propagate_liveness(struct bpf_verifier_env *env,
4918 const struct bpf_verifier_state *vstate,
4919 struct bpf_verifier_state *vparent)
4921 int i, frame, err = 0;
4922 struct bpf_func_state *state, *parent;
4924 if (vparent->curframe != vstate->curframe) {
4925 WARN(1, "propagate_live: parent frame %d current frame %d\n",
4926 vparent->curframe, vstate->curframe);
4929 /* Propagate read liveness of registers... */
4930 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
4931 /* We don't need to worry about FP liveness because it's read-only */
4932 for (i = 0; i < BPF_REG_FP; i++) {
4933 if (vparent->frame[vparent->curframe]->regs[i].live & REG_LIVE_READ)
4935 if (vstate->frame[vstate->curframe]->regs[i].live & REG_LIVE_READ) {
4936 err = mark_reg_read(env, &vstate->frame[vstate->curframe]->regs[i],
4937 &vparent->frame[vstate->curframe]->regs[i]);
4943 /* ... and stack slots */
4944 for (frame = 0; frame <= vstate->curframe; frame++) {
4945 state = vstate->frame[frame];
4946 parent = vparent->frame[frame];
4947 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
4948 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
4949 if (parent->stack[i].spilled_ptr.live & REG_LIVE_READ)
4951 if (state->stack[i].spilled_ptr.live & REG_LIVE_READ)
4952 mark_reg_read(env, &state->stack[i].spilled_ptr,
4953 &parent->stack[i].spilled_ptr);
4959 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
4961 struct bpf_verifier_state_list *new_sl;
4962 struct bpf_verifier_state_list *sl;
4963 struct bpf_verifier_state *cur = env->cur_state, *new;
4966 sl = env->explored_states[insn_idx];
4968 /* this 'insn_idx' instruction wasn't marked, so we will not
4969 * be doing state search here
4973 while (sl != STATE_LIST_MARK) {
4974 if (states_equal(env, &sl->state, cur)) {
4975 /* reached equivalent register/stack state,
4977 * Registers read by the continuation are read by us.
4978 * If we have any write marks in env->cur_state, they
4979 * will prevent corresponding reads in the continuation
4980 * from reaching our parent (an explored_state). Our
4981 * own state will get the read marks recorded, but
4982 * they'll be immediately forgotten as we're pruning
4983 * this state and will pop a new one.
4985 err = propagate_liveness(env, &sl->state, cur);
4993 /* there were no equivalent states, remember current one.
4994 * technically the current state is not proven to be safe yet,
4995 * but it will either reach outer most bpf_exit (which means it's safe)
4996 * or it will be rejected. Since there are no loops, we won't be
4997 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
4998 * again on the way to bpf_exit
5000 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
5004 /* add new state to the head of linked list */
5005 new = &new_sl->state;
5006 err = copy_verifier_state(new, cur);
5008 free_verifier_state(new, false);
5012 new_sl->next = env->explored_states[insn_idx];
5013 env->explored_states[insn_idx] = new_sl;
5014 /* connect new state to parentage chain */
5015 for (i = 0; i < BPF_REG_FP; i++)
5016 cur_regs(env)[i].parent = &new->frame[new->curframe]->regs[i];
5017 /* clear write marks in current state: the writes we did are not writes
5018 * our child did, so they don't screen off its reads from us.
5019 * (There are no read marks in current state, because reads always mark
5020 * their parent and current state never has children yet. Only
5021 * explored_states can get read marks.)
5023 for (i = 0; i < BPF_REG_FP; i++)
5024 cur->frame[cur->curframe]->regs[i].live = REG_LIVE_NONE;
5026 /* all stack frames are accessible from callee, clear them all */
5027 for (j = 0; j <= cur->curframe; j++) {
5028 struct bpf_func_state *frame = cur->frame[j];
5029 struct bpf_func_state *newframe = new->frame[j];
5031 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
5032 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
5033 frame->stack[i].spilled_ptr.parent =
5034 &newframe->stack[i].spilled_ptr;
5040 /* Return true if it's OK to have the same insn return a different type. */
5041 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
5046 case PTR_TO_SOCKET_OR_NULL:
5053 /* If an instruction was previously used with particular pointer types, then we
5054 * need to be careful to avoid cases such as the below, where it may be ok
5055 * for one branch accessing the pointer, but not ok for the other branch:
5060 * R1 = some_other_valid_ptr;
5063 * R2 = *(u32 *)(R1 + 0);
5065 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
5067 return src != prev && (!reg_type_mismatch_ok(src) ||
5068 !reg_type_mismatch_ok(prev));
5071 static int do_check(struct bpf_verifier_env *env)
5073 struct bpf_verifier_state *state;
5074 struct bpf_insn *insns = env->prog->insnsi;
5075 struct bpf_reg_state *regs;
5076 int insn_cnt = env->prog->len, i;
5077 int insn_idx, prev_insn_idx = 0;
5078 int insn_processed = 0;
5079 bool do_print_state = false;
5081 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
5084 state->curframe = 0;
5085 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
5086 if (!state->frame[0]) {
5090 env->cur_state = state;
5091 init_func_state(env, state->frame[0],
5092 BPF_MAIN_FUNC /* callsite */,
5094 0 /* subprogno, zero == main subprog */);
5097 struct bpf_insn *insn;
5101 if (insn_idx >= insn_cnt) {
5102 verbose(env, "invalid insn idx %d insn_cnt %d\n",
5103 insn_idx, insn_cnt);
5107 insn = &insns[insn_idx];
5108 class = BPF_CLASS(insn->code);
5110 if (++insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
5112 "BPF program is too large. Processed %d insn\n",
5117 err = is_state_visited(env, insn_idx);
5121 /* found equivalent state, can prune the search */
5122 if (env->log.level) {
5124 verbose(env, "\nfrom %d to %d: safe\n",
5125 prev_insn_idx, insn_idx);
5127 verbose(env, "%d: safe\n", insn_idx);
5129 goto process_bpf_exit;
5135 if (env->log.level > 1 || (env->log.level && do_print_state)) {
5136 if (env->log.level > 1)
5137 verbose(env, "%d:", insn_idx);
5139 verbose(env, "\nfrom %d to %d:",
5140 prev_insn_idx, insn_idx);
5141 print_verifier_state(env, state->frame[state->curframe]);
5142 do_print_state = false;
5145 if (env->log.level) {
5146 const struct bpf_insn_cbs cbs = {
5147 .cb_print = verbose,
5148 .private_data = env,
5151 verbose(env, "%d: ", insn_idx);
5152 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
5155 if (bpf_prog_is_dev_bound(env->prog->aux)) {
5156 err = bpf_prog_offload_verify_insn(env, insn_idx,
5162 regs = cur_regs(env);
5163 env->insn_aux_data[insn_idx].seen = true;
5165 if (class == BPF_ALU || class == BPF_ALU64) {
5166 err = check_alu_op(env, insn);
5170 } else if (class == BPF_LDX) {
5171 enum bpf_reg_type *prev_src_type, src_reg_type;
5173 /* check for reserved fields is already done */
5175 /* check src operand */
5176 err = check_reg_arg(env, insn->src_reg, SRC_OP);
5180 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
5184 src_reg_type = regs[insn->src_reg].type;
5186 /* check that memory (src_reg + off) is readable,
5187 * the state of dst_reg will be updated by this func
5189 err = check_mem_access(env, insn_idx, insn->src_reg, insn->off,
5190 BPF_SIZE(insn->code), BPF_READ,
5191 insn->dst_reg, false);
5195 prev_src_type = &env->insn_aux_data[insn_idx].ptr_type;
5197 if (*prev_src_type == NOT_INIT) {
5199 * dst_reg = *(u32 *)(src_reg + off)
5200 * save type to validate intersecting paths
5202 *prev_src_type = src_reg_type;
5204 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
5205 /* ABuser program is trying to use the same insn
5206 * dst_reg = *(u32*) (src_reg + off)
5207 * with different pointer types:
5208 * src_reg == ctx in one branch and
5209 * src_reg == stack|map in some other branch.
5212 verbose(env, "same insn cannot be used with different pointers\n");
5216 } else if (class == BPF_STX) {
5217 enum bpf_reg_type *prev_dst_type, dst_reg_type;
5219 if (BPF_MODE(insn->code) == BPF_XADD) {
5220 err = check_xadd(env, insn_idx, insn);
5227 /* check src1 operand */
5228 err = check_reg_arg(env, insn->src_reg, SRC_OP);
5231 /* check src2 operand */
5232 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
5236 dst_reg_type = regs[insn->dst_reg].type;
5238 /* check that memory (dst_reg + off) is writeable */
5239 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
5240 BPF_SIZE(insn->code), BPF_WRITE,
5241 insn->src_reg, false);
5245 prev_dst_type = &env->insn_aux_data[insn_idx].ptr_type;
5247 if (*prev_dst_type == NOT_INIT) {
5248 *prev_dst_type = dst_reg_type;
5249 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
5250 verbose(env, "same insn cannot be used with different pointers\n");
5254 } else if (class == BPF_ST) {
5255 if (BPF_MODE(insn->code) != BPF_MEM ||
5256 insn->src_reg != BPF_REG_0) {
5257 verbose(env, "BPF_ST uses reserved fields\n");
5260 /* check src operand */
5261 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
5265 if (is_ctx_reg(env, insn->dst_reg)) {
5266 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
5267 insn->dst_reg, reg_type_str[insn->dst_reg]);
5271 /* check that memory (dst_reg + off) is writeable */
5272 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
5273 BPF_SIZE(insn->code), BPF_WRITE,
5278 } else if (class == BPF_JMP) {
5279 u8 opcode = BPF_OP(insn->code);
5281 if (opcode == BPF_CALL) {
5282 if (BPF_SRC(insn->code) != BPF_K ||
5284 (insn->src_reg != BPF_REG_0 &&
5285 insn->src_reg != BPF_PSEUDO_CALL) ||
5286 insn->dst_reg != BPF_REG_0) {
5287 verbose(env, "BPF_CALL uses reserved fields\n");
5291 if (insn->src_reg == BPF_PSEUDO_CALL)
5292 err = check_func_call(env, insn, &insn_idx);
5294 err = check_helper_call(env, insn->imm, insn_idx);
5298 } else if (opcode == BPF_JA) {
5299 if (BPF_SRC(insn->code) != BPF_K ||
5301 insn->src_reg != BPF_REG_0 ||
5302 insn->dst_reg != BPF_REG_0) {
5303 verbose(env, "BPF_JA uses reserved fields\n");
5307 insn_idx += insn->off + 1;
5310 } else if (opcode == BPF_EXIT) {
5311 if (BPF_SRC(insn->code) != BPF_K ||
5313 insn->src_reg != BPF_REG_0 ||
5314 insn->dst_reg != BPF_REG_0) {
5315 verbose(env, "BPF_EXIT uses reserved fields\n");
5319 if (state->curframe) {
5320 /* exit from nested function */
5321 prev_insn_idx = insn_idx;
5322 err = prepare_func_exit(env, &insn_idx);
5325 do_print_state = true;
5329 err = check_reference_leak(env);
5333 /* eBPF calling convetion is such that R0 is used
5334 * to return the value from eBPF program.
5335 * Make sure that it's readable at this time
5336 * of bpf_exit, which means that program wrote
5337 * something into it earlier
5339 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
5343 if (is_pointer_value(env, BPF_REG_0)) {
5344 verbose(env, "R0 leaks addr as return value\n");
5348 err = check_return_code(env);
5352 err = pop_stack(env, &prev_insn_idx, &insn_idx);
5358 do_print_state = true;
5362 err = check_cond_jmp_op(env, insn, &insn_idx);
5366 } else if (class == BPF_LD) {
5367 u8 mode = BPF_MODE(insn->code);
5369 if (mode == BPF_ABS || mode == BPF_IND) {
5370 err = check_ld_abs(env, insn);
5374 } else if (mode == BPF_IMM) {
5375 err = check_ld_imm(env, insn);
5380 env->insn_aux_data[insn_idx].seen = true;
5382 verbose(env, "invalid BPF_LD mode\n");
5386 verbose(env, "unknown insn class %d\n", class);
5393 verbose(env, "processed %d insns (limit %d), stack depth ",
5394 insn_processed, BPF_COMPLEXITY_LIMIT_INSNS);
5395 for (i = 0; i < env->subprog_cnt; i++) {
5396 u32 depth = env->subprog_info[i].stack_depth;
5398 verbose(env, "%d", depth);
5399 if (i + 1 < env->subprog_cnt)
5403 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
5407 static int check_map_prealloc(struct bpf_map *map)
5409 return (map->map_type != BPF_MAP_TYPE_HASH &&
5410 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
5411 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
5412 !(map->map_flags & BPF_F_NO_PREALLOC);
5415 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
5416 struct bpf_map *map,
5417 struct bpf_prog *prog)
5420 /* Make sure that BPF_PROG_TYPE_PERF_EVENT programs only use
5421 * preallocated hash maps, since doing memory allocation
5422 * in overflow_handler can crash depending on where nmi got
5425 if (prog->type == BPF_PROG_TYPE_PERF_EVENT) {
5426 if (!check_map_prealloc(map)) {
5427 verbose(env, "perf_event programs can only use preallocated hash map\n");
5430 if (map->inner_map_meta &&
5431 !check_map_prealloc(map->inner_map_meta)) {
5432 verbose(env, "perf_event programs can only use preallocated inner hash map\n");
5437 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
5438 !bpf_offload_prog_map_match(prog, map)) {
5439 verbose(env, "offload device mismatch between prog and map\n");
5446 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
5448 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
5449 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
5452 /* look for pseudo eBPF instructions that access map FDs and
5453 * replace them with actual map pointers
5455 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env *env)
5457 struct bpf_insn *insn = env->prog->insnsi;
5458 int insn_cnt = env->prog->len;
5461 err = bpf_prog_calc_tag(env->prog);
5465 for (i = 0; i < insn_cnt; i++, insn++) {
5466 if (BPF_CLASS(insn->code) == BPF_LDX &&
5467 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
5468 verbose(env, "BPF_LDX uses reserved fields\n");
5472 if (BPF_CLASS(insn->code) == BPF_STX &&
5473 ((BPF_MODE(insn->code) != BPF_MEM &&
5474 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
5475 verbose(env, "BPF_STX uses reserved fields\n");
5479 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
5480 struct bpf_map *map;
5483 if (i == insn_cnt - 1 || insn[1].code != 0 ||
5484 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
5486 verbose(env, "invalid bpf_ld_imm64 insn\n");
5490 if (insn->src_reg == 0)
5491 /* valid generic load 64-bit imm */
5494 if (insn->src_reg != BPF_PSEUDO_MAP_FD) {
5496 "unrecognized bpf_ld_imm64 insn\n");
5500 f = fdget(insn->imm);
5501 map = __bpf_map_get(f);
5503 verbose(env, "fd %d is not pointing to valid bpf_map\n",
5505 return PTR_ERR(map);
5508 err = check_map_prog_compatibility(env, map, env->prog);
5514 /* store map pointer inside BPF_LD_IMM64 instruction */
5515 insn[0].imm = (u32) (unsigned long) map;
5516 insn[1].imm = ((u64) (unsigned long) map) >> 32;
5518 /* check whether we recorded this map already */
5519 for (j = 0; j < env->used_map_cnt; j++)
5520 if (env->used_maps[j] == map) {
5525 if (env->used_map_cnt >= MAX_USED_MAPS) {
5530 /* hold the map. If the program is rejected by verifier,
5531 * the map will be released by release_maps() or it
5532 * will be used by the valid program until it's unloaded
5533 * and all maps are released in free_used_maps()
5535 map = bpf_map_inc(map, false);
5538 return PTR_ERR(map);
5540 env->used_maps[env->used_map_cnt++] = map;
5542 if (bpf_map_is_cgroup_storage(map) &&
5543 bpf_cgroup_storage_assign(env->prog, map)) {
5544 verbose(env, "only one cgroup storage of each type is allowed\n");
5556 /* Basic sanity check before we invest more work here. */
5557 if (!bpf_opcode_in_insntable(insn->code)) {
5558 verbose(env, "unknown opcode %02x\n", insn->code);
5563 /* now all pseudo BPF_LD_IMM64 instructions load valid
5564 * 'struct bpf_map *' into a register instead of user map_fd.
5565 * These pointers will be used later by verifier to validate map access.
5570 /* drop refcnt of maps used by the rejected program */
5571 static void release_maps(struct bpf_verifier_env *env)
5573 enum bpf_cgroup_storage_type stype;
5576 for_each_cgroup_storage_type(stype) {
5577 if (!env->prog->aux->cgroup_storage[stype])
5579 bpf_cgroup_storage_release(env->prog,
5580 env->prog->aux->cgroup_storage[stype]);
5583 for (i = 0; i < env->used_map_cnt; i++)
5584 bpf_map_put(env->used_maps[i]);
5587 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
5588 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
5590 struct bpf_insn *insn = env->prog->insnsi;
5591 int insn_cnt = env->prog->len;
5594 for (i = 0; i < insn_cnt; i++, insn++)
5595 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
5599 /* single env->prog->insni[off] instruction was replaced with the range
5600 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
5601 * [0, off) and [off, end) to new locations, so the patched range stays zero
5603 static int adjust_insn_aux_data(struct bpf_verifier_env *env, u32 prog_len,
5606 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
5611 new_data = vzalloc(array_size(prog_len,
5612 sizeof(struct bpf_insn_aux_data)));
5615 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
5616 memcpy(new_data + off + cnt - 1, old_data + off,
5617 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
5618 for (i = off; i < off + cnt - 1; i++)
5619 new_data[i].seen = true;
5620 env->insn_aux_data = new_data;
5625 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
5631 /* NOTE: fake 'exit' subprog should be updated as well. */
5632 for (i = 0; i <= env->subprog_cnt; i++) {
5633 if (env->subprog_info[i].start < off)
5635 env->subprog_info[i].start += len - 1;
5639 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
5640 const struct bpf_insn *patch, u32 len)
5642 struct bpf_prog *new_prog;
5644 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
5647 if (adjust_insn_aux_data(env, new_prog->len, off, len))
5649 adjust_subprog_starts(env, off, len);
5653 /* The verifier does more data flow analysis than llvm and will not
5654 * explore branches that are dead at run time. Malicious programs can
5655 * have dead code too. Therefore replace all dead at-run-time code
5658 * Just nops are not optimal, e.g. if they would sit at the end of the
5659 * program and through another bug we would manage to jump there, then
5660 * we'd execute beyond program memory otherwise. Returning exception
5661 * code also wouldn't work since we can have subprogs where the dead
5662 * code could be located.
5664 static void sanitize_dead_code(struct bpf_verifier_env *env)
5666 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
5667 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
5668 struct bpf_insn *insn = env->prog->insnsi;
5669 const int insn_cnt = env->prog->len;
5672 for (i = 0; i < insn_cnt; i++) {
5673 if (aux_data[i].seen)
5675 memcpy(insn + i, &trap, sizeof(trap));
5679 /* convert load instructions that access fields of a context type into a
5680 * sequence of instructions that access fields of the underlying structure:
5681 * struct __sk_buff -> struct sk_buff
5682 * struct bpf_sock_ops -> struct sock
5684 static int convert_ctx_accesses(struct bpf_verifier_env *env)
5686 const struct bpf_verifier_ops *ops = env->ops;
5687 int i, cnt, size, ctx_field_size, delta = 0;
5688 const int insn_cnt = env->prog->len;
5689 struct bpf_insn insn_buf[16], *insn;
5690 struct bpf_prog *new_prog;
5691 enum bpf_access_type type;
5692 bool is_narrower_load;
5695 if (ops->gen_prologue) {
5696 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
5698 if (cnt >= ARRAY_SIZE(insn_buf)) {
5699 verbose(env, "bpf verifier is misconfigured\n");
5702 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
5706 env->prog = new_prog;
5711 if (bpf_prog_is_dev_bound(env->prog->aux))
5714 insn = env->prog->insnsi + delta;
5716 for (i = 0; i < insn_cnt; i++, insn++) {
5717 bpf_convert_ctx_access_t convert_ctx_access;
5719 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
5720 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
5721 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
5722 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
5724 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
5725 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
5726 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
5727 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
5732 if (type == BPF_WRITE &&
5733 env->insn_aux_data[i + delta].sanitize_stack_off) {
5734 struct bpf_insn patch[] = {
5735 /* Sanitize suspicious stack slot with zero.
5736 * There are no memory dependencies for this store,
5737 * since it's only using frame pointer and immediate
5740 BPF_ST_MEM(BPF_DW, BPF_REG_FP,
5741 env->insn_aux_data[i + delta].sanitize_stack_off,
5743 /* the original STX instruction will immediately
5744 * overwrite the same stack slot with appropriate value
5749 cnt = ARRAY_SIZE(patch);
5750 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
5755 env->prog = new_prog;
5756 insn = new_prog->insnsi + i + delta;
5760 switch (env->insn_aux_data[i + delta].ptr_type) {
5762 if (!ops->convert_ctx_access)
5764 convert_ctx_access = ops->convert_ctx_access;
5767 convert_ctx_access = bpf_sock_convert_ctx_access;
5773 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
5774 size = BPF_LDST_BYTES(insn);
5776 /* If the read access is a narrower load of the field,
5777 * convert to a 4/8-byte load, to minimum program type specific
5778 * convert_ctx_access changes. If conversion is successful,
5779 * we will apply proper mask to the result.
5781 is_narrower_load = size < ctx_field_size;
5782 if (is_narrower_load) {
5783 u32 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
5784 u32 off = insn->off;
5787 if (type == BPF_WRITE) {
5788 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
5793 if (ctx_field_size == 4)
5795 else if (ctx_field_size == 8)
5798 insn->off = off & ~(size_default - 1);
5799 insn->code = BPF_LDX | BPF_MEM | size_code;
5803 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
5805 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
5806 (ctx_field_size && !target_size)) {
5807 verbose(env, "bpf verifier is misconfigured\n");
5811 if (is_narrower_load && size < target_size) {
5812 if (ctx_field_size <= 4)
5813 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
5814 (1 << size * 8) - 1);
5816 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
5817 (1 << size * 8) - 1);
5820 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
5826 /* keep walking new program and skip insns we just inserted */
5827 env->prog = new_prog;
5828 insn = new_prog->insnsi + i + delta;
5834 static int jit_subprogs(struct bpf_verifier_env *env)
5836 struct bpf_prog *prog = env->prog, **func, *tmp;
5837 int i, j, subprog_start, subprog_end = 0, len, subprog;
5838 struct bpf_insn *insn;
5842 if (env->subprog_cnt <= 1)
5845 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
5846 if (insn->code != (BPF_JMP | BPF_CALL) ||
5847 insn->src_reg != BPF_PSEUDO_CALL)
5849 /* Upon error here we cannot fall back to interpreter but
5850 * need a hard reject of the program. Thus -EFAULT is
5851 * propagated in any case.
5853 subprog = find_subprog(env, i + insn->imm + 1);
5855 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
5859 /* temporarily remember subprog id inside insn instead of
5860 * aux_data, since next loop will split up all insns into funcs
5862 insn->off = subprog;
5863 /* remember original imm in case JIT fails and fallback
5864 * to interpreter will be needed
5866 env->insn_aux_data[i].call_imm = insn->imm;
5867 /* point imm to __bpf_call_base+1 from JITs point of view */
5871 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
5875 for (i = 0; i < env->subprog_cnt; i++) {
5876 subprog_start = subprog_end;
5877 subprog_end = env->subprog_info[i + 1].start;
5879 len = subprog_end - subprog_start;
5880 func[i] = bpf_prog_alloc(bpf_prog_size(len), GFP_USER);
5883 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
5884 len * sizeof(struct bpf_insn));
5885 func[i]->type = prog->type;
5887 if (bpf_prog_calc_tag(func[i]))
5889 func[i]->is_func = 1;
5890 /* Use bpf_prog_F_tag to indicate functions in stack traces.
5891 * Long term would need debug info to populate names
5893 func[i]->aux->name[0] = 'F';
5894 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
5895 func[i]->jit_requested = 1;
5896 func[i] = bpf_int_jit_compile(func[i]);
5897 if (!func[i]->jited) {
5903 /* at this point all bpf functions were successfully JITed
5904 * now populate all bpf_calls with correct addresses and
5905 * run last pass of JIT
5907 for (i = 0; i < env->subprog_cnt; i++) {
5908 insn = func[i]->insnsi;
5909 for (j = 0; j < func[i]->len; j++, insn++) {
5910 if (insn->code != (BPF_JMP | BPF_CALL) ||
5911 insn->src_reg != BPF_PSEUDO_CALL)
5913 subprog = insn->off;
5914 insn->imm = (u64 (*)(u64, u64, u64, u64, u64))
5915 func[subprog]->bpf_func -
5919 /* we use the aux data to keep a list of the start addresses
5920 * of the JITed images for each function in the program
5922 * for some architectures, such as powerpc64, the imm field
5923 * might not be large enough to hold the offset of the start
5924 * address of the callee's JITed image from __bpf_call_base
5926 * in such cases, we can lookup the start address of a callee
5927 * by using its subprog id, available from the off field of
5928 * the call instruction, as an index for this list
5930 func[i]->aux->func = func;
5931 func[i]->aux->func_cnt = env->subprog_cnt;
5933 for (i = 0; i < env->subprog_cnt; i++) {
5934 old_bpf_func = func[i]->bpf_func;
5935 tmp = bpf_int_jit_compile(func[i]);
5936 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
5937 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
5944 /* finally lock prog and jit images for all functions and
5947 for (i = 0; i < env->subprog_cnt; i++) {
5948 bpf_prog_lock_ro(func[i]);
5949 bpf_prog_kallsyms_add(func[i]);
5952 /* Last step: make now unused interpreter insns from main
5953 * prog consistent for later dump requests, so they can
5954 * later look the same as if they were interpreted only.
5956 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
5957 if (insn->code != (BPF_JMP | BPF_CALL) ||
5958 insn->src_reg != BPF_PSEUDO_CALL)
5960 insn->off = env->insn_aux_data[i].call_imm;
5961 subprog = find_subprog(env, i + insn->off + 1);
5962 insn->imm = subprog;
5966 prog->bpf_func = func[0]->bpf_func;
5967 prog->aux->func = func;
5968 prog->aux->func_cnt = env->subprog_cnt;
5971 for (i = 0; i < env->subprog_cnt; i++)
5973 bpf_jit_free(func[i]);
5976 /* cleanup main prog to be interpreted */
5977 prog->jit_requested = 0;
5978 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
5979 if (insn->code != (BPF_JMP | BPF_CALL) ||
5980 insn->src_reg != BPF_PSEUDO_CALL)
5983 insn->imm = env->insn_aux_data[i].call_imm;
5988 static int fixup_call_args(struct bpf_verifier_env *env)
5990 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
5991 struct bpf_prog *prog = env->prog;
5992 struct bpf_insn *insn = prog->insnsi;
5997 if (env->prog->jit_requested &&
5998 !bpf_prog_is_dev_bound(env->prog->aux)) {
5999 err = jit_subprogs(env);
6005 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
6006 for (i = 0; i < prog->len; i++, insn++) {
6007 if (insn->code != (BPF_JMP | BPF_CALL) ||
6008 insn->src_reg != BPF_PSEUDO_CALL)
6010 depth = get_callee_stack_depth(env, insn, i);
6013 bpf_patch_call_args(insn, depth);
6020 /* fixup insn->imm field of bpf_call instructions
6021 * and inline eligible helpers as explicit sequence of BPF instructions
6023 * this function is called after eBPF program passed verification
6025 static int fixup_bpf_calls(struct bpf_verifier_env *env)
6027 struct bpf_prog *prog = env->prog;
6028 struct bpf_insn *insn = prog->insnsi;
6029 const struct bpf_func_proto *fn;
6030 const int insn_cnt = prog->len;
6031 const struct bpf_map_ops *ops;
6032 struct bpf_insn_aux_data *aux;
6033 struct bpf_insn insn_buf[16];
6034 struct bpf_prog *new_prog;
6035 struct bpf_map *map_ptr;
6036 int i, cnt, delta = 0;
6038 for (i = 0; i < insn_cnt; i++, insn++) {
6039 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
6040 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
6041 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
6042 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
6043 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
6044 struct bpf_insn mask_and_div[] = {
6045 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
6047 BPF_JMP_IMM(BPF_JNE, insn->src_reg, 0, 2),
6048 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
6049 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
6052 struct bpf_insn mask_and_mod[] = {
6053 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
6054 /* Rx mod 0 -> Rx */
6055 BPF_JMP_IMM(BPF_JEQ, insn->src_reg, 0, 1),
6058 struct bpf_insn *patchlet;
6060 if (insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
6061 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
6062 patchlet = mask_and_div + (is64 ? 1 : 0);
6063 cnt = ARRAY_SIZE(mask_and_div) - (is64 ? 1 : 0);
6065 patchlet = mask_and_mod + (is64 ? 1 : 0);
6066 cnt = ARRAY_SIZE(mask_and_mod) - (is64 ? 1 : 0);
6069 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
6074 env->prog = prog = new_prog;
6075 insn = new_prog->insnsi + i + delta;
6079 if (BPF_CLASS(insn->code) == BPF_LD &&
6080 (BPF_MODE(insn->code) == BPF_ABS ||
6081 BPF_MODE(insn->code) == BPF_IND)) {
6082 cnt = env->ops->gen_ld_abs(insn, insn_buf);
6083 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
6084 verbose(env, "bpf verifier is misconfigured\n");
6088 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
6093 env->prog = prog = new_prog;
6094 insn = new_prog->insnsi + i + delta;
6098 if (insn->code != (BPF_JMP | BPF_CALL))
6100 if (insn->src_reg == BPF_PSEUDO_CALL)
6103 if (insn->imm == BPF_FUNC_get_route_realm)
6104 prog->dst_needed = 1;
6105 if (insn->imm == BPF_FUNC_get_prandom_u32)
6106 bpf_user_rnd_init_once();
6107 if (insn->imm == BPF_FUNC_override_return)
6108 prog->kprobe_override = 1;
6109 if (insn->imm == BPF_FUNC_tail_call) {
6110 /* If we tail call into other programs, we
6111 * cannot make any assumptions since they can
6112 * be replaced dynamically during runtime in
6113 * the program array.
6115 prog->cb_access = 1;
6116 env->prog->aux->stack_depth = MAX_BPF_STACK;
6118 /* mark bpf_tail_call as different opcode to avoid
6119 * conditional branch in the interpeter for every normal
6120 * call and to prevent accidental JITing by JIT compiler
6121 * that doesn't support bpf_tail_call yet
6124 insn->code = BPF_JMP | BPF_TAIL_CALL;
6126 aux = &env->insn_aux_data[i + delta];
6127 if (!bpf_map_ptr_unpriv(aux))
6130 /* instead of changing every JIT dealing with tail_call
6131 * emit two extra insns:
6132 * if (index >= max_entries) goto out;
6133 * index &= array->index_mask;
6134 * to avoid out-of-bounds cpu speculation
6136 if (bpf_map_ptr_poisoned(aux)) {
6137 verbose(env, "tail_call abusing map_ptr\n");
6141 map_ptr = BPF_MAP_PTR(aux->map_state);
6142 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
6143 map_ptr->max_entries, 2);
6144 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
6145 container_of(map_ptr,
6148 insn_buf[2] = *insn;
6150 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
6155 env->prog = prog = new_prog;
6156 insn = new_prog->insnsi + i + delta;
6160 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
6161 * and other inlining handlers are currently limited to 64 bit
6164 if (prog->jit_requested && BITS_PER_LONG == 64 &&
6165 (insn->imm == BPF_FUNC_map_lookup_elem ||
6166 insn->imm == BPF_FUNC_map_update_elem ||
6167 insn->imm == BPF_FUNC_map_delete_elem)) {
6168 aux = &env->insn_aux_data[i + delta];
6169 if (bpf_map_ptr_poisoned(aux))
6170 goto patch_call_imm;
6172 map_ptr = BPF_MAP_PTR(aux->map_state);
6174 if (insn->imm == BPF_FUNC_map_lookup_elem &&
6175 ops->map_gen_lookup) {
6176 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
6177 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
6178 verbose(env, "bpf verifier is misconfigured\n");
6182 new_prog = bpf_patch_insn_data(env, i + delta,
6188 env->prog = prog = new_prog;
6189 insn = new_prog->insnsi + i + delta;
6193 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
6194 (void *(*)(struct bpf_map *map, void *key))NULL));
6195 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
6196 (int (*)(struct bpf_map *map, void *key))NULL));
6197 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
6198 (int (*)(struct bpf_map *map, void *key, void *value,
6200 switch (insn->imm) {
6201 case BPF_FUNC_map_lookup_elem:
6202 insn->imm = BPF_CAST_CALL(ops->map_lookup_elem) -
6205 case BPF_FUNC_map_update_elem:
6206 insn->imm = BPF_CAST_CALL(ops->map_update_elem) -
6209 case BPF_FUNC_map_delete_elem:
6210 insn->imm = BPF_CAST_CALL(ops->map_delete_elem) -
6215 goto patch_call_imm;
6219 fn = env->ops->get_func_proto(insn->imm, env->prog);
6220 /* all functions that have prototype and verifier allowed
6221 * programs to call them, must be real in-kernel functions
6225 "kernel subsystem misconfigured func %s#%d\n",
6226 func_id_name(insn->imm), insn->imm);
6229 insn->imm = fn->func - __bpf_call_base;
6235 static void free_states(struct bpf_verifier_env *env)
6237 struct bpf_verifier_state_list *sl, *sln;
6240 if (!env->explored_states)
6243 for (i = 0; i < env->prog->len; i++) {
6244 sl = env->explored_states[i];
6247 while (sl != STATE_LIST_MARK) {
6249 free_verifier_state(&sl->state, false);
6255 kfree(env->explored_states);
6258 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr)
6260 struct bpf_verifier_env *env;
6261 struct bpf_verifier_log *log;
6264 /* no program is valid */
6265 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
6268 /* 'struct bpf_verifier_env' can be global, but since it's not small,
6269 * allocate/free it every time bpf_check() is called
6271 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
6276 env->insn_aux_data =
6277 vzalloc(array_size(sizeof(struct bpf_insn_aux_data),
6280 if (!env->insn_aux_data)
6283 env->ops = bpf_verifier_ops[env->prog->type];
6285 /* grab the mutex to protect few globals used by verifier */
6286 mutex_lock(&bpf_verifier_lock);
6288 if (attr->log_level || attr->log_buf || attr->log_size) {
6289 /* user requested verbose verifier output
6290 * and supplied buffer to store the verification trace
6292 log->level = attr->log_level;
6293 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
6294 log->len_total = attr->log_size;
6297 /* log attributes have to be sane */
6298 if (log->len_total < 128 || log->len_total > UINT_MAX >> 8 ||
6299 !log->level || !log->ubuf)
6303 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
6304 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
6305 env->strict_alignment = true;
6307 ret = replace_map_fd_with_map_ptr(env);
6309 goto skip_full_check;
6311 if (bpf_prog_is_dev_bound(env->prog->aux)) {
6312 ret = bpf_prog_offload_verifier_prep(env);
6314 goto skip_full_check;
6317 env->explored_states = kcalloc(env->prog->len,
6318 sizeof(struct bpf_verifier_state_list *),
6321 if (!env->explored_states)
6322 goto skip_full_check;
6324 env->allow_ptr_leaks = capable(CAP_SYS_ADMIN);
6326 ret = check_cfg(env);
6328 goto skip_full_check;
6330 ret = do_check(env);
6331 if (env->cur_state) {
6332 free_verifier_state(env->cur_state, true);
6333 env->cur_state = NULL;
6336 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
6337 ret = bpf_prog_offload_finalize(env);
6340 while (!pop_stack(env, NULL, NULL));
6344 sanitize_dead_code(env);
6347 ret = check_max_stack_depth(env);
6350 /* program is valid, convert *(u32*)(ctx + off) accesses */
6351 ret = convert_ctx_accesses(env);
6354 ret = fixup_bpf_calls(env);
6357 ret = fixup_call_args(env);
6359 if (log->level && bpf_verifier_log_full(log))
6361 if (log->level && !log->ubuf) {
6363 goto err_release_maps;
6366 if (ret == 0 && env->used_map_cnt) {
6367 /* if program passed verifier, update used_maps in bpf_prog_info */
6368 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
6369 sizeof(env->used_maps[0]),
6372 if (!env->prog->aux->used_maps) {
6374 goto err_release_maps;
6377 memcpy(env->prog->aux->used_maps, env->used_maps,
6378 sizeof(env->used_maps[0]) * env->used_map_cnt);
6379 env->prog->aux->used_map_cnt = env->used_map_cnt;
6381 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
6382 * bpf_ld_imm64 instructions
6384 convert_pseudo_ld_imm64(env);
6388 if (!env->prog->aux->used_maps)
6389 /* if we didn't copy map pointers into bpf_prog_info, release
6390 * them now. Otherwise free_used_maps() will release them.
6395 mutex_unlock(&bpf_verifier_lock);
6396 vfree(env->insn_aux_data);