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 /* Program types only with direct read access go here! */
1391 case BPF_PROG_TYPE_LWT_IN:
1392 case BPF_PROG_TYPE_LWT_OUT:
1393 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
1394 case BPF_PROG_TYPE_SK_REUSEPORT:
1395 case BPF_PROG_TYPE_FLOW_DISSECTOR:
1396 case BPF_PROG_TYPE_CGROUP_SKB:
1401 /* Program types with direct read + write access go here! */
1402 case BPF_PROG_TYPE_SCHED_CLS:
1403 case BPF_PROG_TYPE_SCHED_ACT:
1404 case BPF_PROG_TYPE_XDP:
1405 case BPF_PROG_TYPE_LWT_XMIT:
1406 case BPF_PROG_TYPE_SK_SKB:
1407 case BPF_PROG_TYPE_SK_MSG:
1409 return meta->pkt_access;
1411 env->seen_direct_write = true;
1418 static int __check_packet_access(struct bpf_verifier_env *env, u32 regno,
1419 int off, int size, bool zero_size_allowed)
1421 struct bpf_reg_state *regs = cur_regs(env);
1422 struct bpf_reg_state *reg = ®s[regno];
1424 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
1425 (u64)off + size > reg->range) {
1426 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
1427 off, size, regno, reg->id, reg->off, reg->range);
1433 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
1434 int size, bool zero_size_allowed)
1436 struct bpf_reg_state *regs = cur_regs(env);
1437 struct bpf_reg_state *reg = ®s[regno];
1440 /* We may have added a variable offset to the packet pointer; but any
1441 * reg->range we have comes after that. We are only checking the fixed
1445 /* We don't allow negative numbers, because we aren't tracking enough
1446 * detail to prove they're safe.
1448 if (reg->smin_value < 0) {
1449 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1453 err = __check_packet_access(env, regno, off, size, zero_size_allowed);
1455 verbose(env, "R%d offset is outside of the packet\n", regno);
1461 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
1462 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
1463 enum bpf_access_type t, enum bpf_reg_type *reg_type)
1465 struct bpf_insn_access_aux info = {
1466 .reg_type = *reg_type,
1469 if (env->ops->is_valid_access &&
1470 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
1471 /* A non zero info.ctx_field_size indicates that this field is a
1472 * candidate for later verifier transformation to load the whole
1473 * field and then apply a mask when accessed with a narrower
1474 * access than actual ctx access size. A zero info.ctx_field_size
1475 * will only allow for whole field access and rejects any other
1476 * type of narrower access.
1478 *reg_type = info.reg_type;
1480 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
1481 /* remember the offset of last byte accessed in ctx */
1482 if (env->prog->aux->max_ctx_offset < off + size)
1483 env->prog->aux->max_ctx_offset = off + size;
1487 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
1491 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
1494 if (size < 0 || off < 0 ||
1495 (u64)off + size > sizeof(struct bpf_flow_keys)) {
1496 verbose(env, "invalid access to flow keys off=%d size=%d\n",
1503 static int check_sock_access(struct bpf_verifier_env *env, u32 regno, int off,
1504 int size, enum bpf_access_type t)
1506 struct bpf_reg_state *regs = cur_regs(env);
1507 struct bpf_reg_state *reg = ®s[regno];
1508 struct bpf_insn_access_aux info;
1510 if (reg->smin_value < 0) {
1511 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1516 if (!bpf_sock_is_valid_access(off, size, t, &info)) {
1517 verbose(env, "invalid bpf_sock access off=%d size=%d\n",
1525 static bool __is_pointer_value(bool allow_ptr_leaks,
1526 const struct bpf_reg_state *reg)
1528 if (allow_ptr_leaks)
1531 return reg->type != SCALAR_VALUE;
1534 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
1536 return cur_regs(env) + regno;
1539 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
1541 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
1544 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
1546 const struct bpf_reg_state *reg = reg_state(env, regno);
1548 return reg->type == PTR_TO_CTX ||
1549 reg->type == PTR_TO_SOCKET;
1552 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
1554 const struct bpf_reg_state *reg = reg_state(env, regno);
1556 return type_is_pkt_pointer(reg->type);
1559 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
1561 const struct bpf_reg_state *reg = reg_state(env, regno);
1563 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
1564 return reg->type == PTR_TO_FLOW_KEYS;
1567 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
1568 const struct bpf_reg_state *reg,
1569 int off, int size, bool strict)
1571 struct tnum reg_off;
1574 /* Byte size accesses are always allowed. */
1575 if (!strict || size == 1)
1578 /* For platforms that do not have a Kconfig enabling
1579 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
1580 * NET_IP_ALIGN is universally set to '2'. And on platforms
1581 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
1582 * to this code only in strict mode where we want to emulate
1583 * the NET_IP_ALIGN==2 checking. Therefore use an
1584 * unconditional IP align value of '2'.
1588 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
1589 if (!tnum_is_aligned(reg_off, size)) {
1592 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1594 "misaligned packet access off %d+%s+%d+%d size %d\n",
1595 ip_align, tn_buf, reg->off, off, size);
1602 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
1603 const struct bpf_reg_state *reg,
1604 const char *pointer_desc,
1605 int off, int size, bool strict)
1607 struct tnum reg_off;
1609 /* Byte size accesses are always allowed. */
1610 if (!strict || size == 1)
1613 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
1614 if (!tnum_is_aligned(reg_off, size)) {
1617 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1618 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
1619 pointer_desc, tn_buf, reg->off, off, size);
1626 static int check_ptr_alignment(struct bpf_verifier_env *env,
1627 const struct bpf_reg_state *reg, int off,
1628 int size, bool strict_alignment_once)
1630 bool strict = env->strict_alignment || strict_alignment_once;
1631 const char *pointer_desc = "";
1633 switch (reg->type) {
1635 case PTR_TO_PACKET_META:
1636 /* Special case, because of NET_IP_ALIGN. Given metadata sits
1637 * right in front, treat it the very same way.
1639 return check_pkt_ptr_alignment(env, reg, off, size, strict);
1640 case PTR_TO_FLOW_KEYS:
1641 pointer_desc = "flow keys ";
1643 case PTR_TO_MAP_VALUE:
1644 pointer_desc = "value ";
1647 pointer_desc = "context ";
1650 pointer_desc = "stack ";
1651 /* The stack spill tracking logic in check_stack_write()
1652 * and check_stack_read() relies on stack accesses being
1658 pointer_desc = "sock ";
1663 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
1667 static int update_stack_depth(struct bpf_verifier_env *env,
1668 const struct bpf_func_state *func,
1671 u16 stack = env->subprog_info[func->subprogno].stack_depth;
1676 /* update known max for given subprogram */
1677 env->subprog_info[func->subprogno].stack_depth = -off;
1681 /* starting from main bpf function walk all instructions of the function
1682 * and recursively walk all callees that given function can call.
1683 * Ignore jump and exit insns.
1684 * Since recursion is prevented by check_cfg() this algorithm
1685 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
1687 static int check_max_stack_depth(struct bpf_verifier_env *env)
1689 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
1690 struct bpf_subprog_info *subprog = env->subprog_info;
1691 struct bpf_insn *insn = env->prog->insnsi;
1692 int ret_insn[MAX_CALL_FRAMES];
1693 int ret_prog[MAX_CALL_FRAMES];
1696 /* round up to 32-bytes, since this is granularity
1697 * of interpreter stack size
1699 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
1700 if (depth > MAX_BPF_STACK) {
1701 verbose(env, "combined stack size of %d calls is %d. Too large\n",
1706 subprog_end = subprog[idx + 1].start;
1707 for (; i < subprog_end; i++) {
1708 if (insn[i].code != (BPF_JMP | BPF_CALL))
1710 if (insn[i].src_reg != BPF_PSEUDO_CALL)
1712 /* remember insn and function to return to */
1713 ret_insn[frame] = i + 1;
1714 ret_prog[frame] = idx;
1716 /* find the callee */
1717 i = i + insn[i].imm + 1;
1718 idx = find_subprog(env, i);
1720 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
1725 if (frame >= MAX_CALL_FRAMES) {
1726 WARN_ONCE(1, "verifier bug. Call stack is too deep\n");
1731 /* end of for() loop means the last insn of the 'subprog'
1732 * was reached. Doesn't matter whether it was JA or EXIT
1736 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
1738 i = ret_insn[frame];
1739 idx = ret_prog[frame];
1743 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
1744 static int get_callee_stack_depth(struct bpf_verifier_env *env,
1745 const struct bpf_insn *insn, int idx)
1747 int start = idx + insn->imm + 1, subprog;
1749 subprog = find_subprog(env, start);
1751 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
1755 return env->subprog_info[subprog].stack_depth;
1759 static int check_ctx_reg(struct bpf_verifier_env *env,
1760 const struct bpf_reg_state *reg, int regno)
1762 /* Access to ctx or passing it to a helper is only allowed in
1763 * its original, unmodified form.
1767 verbose(env, "dereference of modified ctx ptr R%d off=%d disallowed\n",
1772 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
1775 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1776 verbose(env, "variable ctx access var_off=%s disallowed\n", tn_buf);
1783 /* truncate register to smaller size (in bytes)
1784 * must be called with size < BPF_REG_SIZE
1786 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
1790 /* clear high bits in bit representation */
1791 reg->var_off = tnum_cast(reg->var_off, size);
1793 /* fix arithmetic bounds */
1794 mask = ((u64)1 << (size * 8)) - 1;
1795 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
1796 reg->umin_value &= mask;
1797 reg->umax_value &= mask;
1799 reg->umin_value = 0;
1800 reg->umax_value = mask;
1802 reg->smin_value = reg->umin_value;
1803 reg->smax_value = reg->umax_value;
1806 /* check whether memory at (regno + off) is accessible for t = (read | write)
1807 * if t==write, value_regno is a register which value is stored into memory
1808 * if t==read, value_regno is a register which will receive the value from memory
1809 * if t==write && value_regno==-1, some unknown value is stored into memory
1810 * if t==read && value_regno==-1, don't care what we read from memory
1812 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
1813 int off, int bpf_size, enum bpf_access_type t,
1814 int value_regno, bool strict_alignment_once)
1816 struct bpf_reg_state *regs = cur_regs(env);
1817 struct bpf_reg_state *reg = regs + regno;
1818 struct bpf_func_state *state;
1821 size = bpf_size_to_bytes(bpf_size);
1825 /* alignment checks will add in reg->off themselves */
1826 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
1830 /* for access checks, reg->off is just part of off */
1833 if (reg->type == PTR_TO_MAP_VALUE) {
1834 if (t == BPF_WRITE && value_regno >= 0 &&
1835 is_pointer_value(env, value_regno)) {
1836 verbose(env, "R%d leaks addr into map\n", value_regno);
1840 err = check_map_access(env, regno, off, size, false);
1841 if (!err && t == BPF_READ && value_regno >= 0)
1842 mark_reg_unknown(env, regs, value_regno);
1844 } else if (reg->type == PTR_TO_CTX) {
1845 enum bpf_reg_type reg_type = SCALAR_VALUE;
1847 if (t == BPF_WRITE && value_regno >= 0 &&
1848 is_pointer_value(env, value_regno)) {
1849 verbose(env, "R%d leaks addr into ctx\n", value_regno);
1853 err = check_ctx_reg(env, reg, regno);
1857 err = check_ctx_access(env, insn_idx, off, size, t, ®_type);
1858 if (!err && t == BPF_READ && value_regno >= 0) {
1859 /* ctx access returns either a scalar, or a
1860 * PTR_TO_PACKET[_META,_END]. In the latter
1861 * case, we know the offset is zero.
1863 if (reg_type == SCALAR_VALUE)
1864 mark_reg_unknown(env, regs, value_regno);
1866 mark_reg_known_zero(env, regs,
1868 regs[value_regno].type = reg_type;
1871 } else if (reg->type == PTR_TO_STACK) {
1872 /* stack accesses must be at a fixed offset, so that we can
1873 * determine what type of data were returned.
1874 * See check_stack_read().
1876 if (!tnum_is_const(reg->var_off)) {
1879 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1880 verbose(env, "variable stack access var_off=%s off=%d size=%d",
1884 off += reg->var_off.value;
1885 if (off >= 0 || off < -MAX_BPF_STACK) {
1886 verbose(env, "invalid stack off=%d size=%d\n", off,
1891 state = func(env, reg);
1892 err = update_stack_depth(env, state, off);
1897 err = check_stack_write(env, state, off, size,
1898 value_regno, insn_idx);
1900 err = check_stack_read(env, state, off, size,
1902 } else if (reg_is_pkt_pointer(reg)) {
1903 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
1904 verbose(env, "cannot write into packet\n");
1907 if (t == BPF_WRITE && value_regno >= 0 &&
1908 is_pointer_value(env, value_regno)) {
1909 verbose(env, "R%d leaks addr into packet\n",
1913 err = check_packet_access(env, regno, off, size, false);
1914 if (!err && t == BPF_READ && value_regno >= 0)
1915 mark_reg_unknown(env, regs, value_regno);
1916 } else if (reg->type == PTR_TO_FLOW_KEYS) {
1917 if (t == BPF_WRITE && value_regno >= 0 &&
1918 is_pointer_value(env, value_regno)) {
1919 verbose(env, "R%d leaks addr into flow keys\n",
1924 err = check_flow_keys_access(env, off, size);
1925 if (!err && t == BPF_READ && value_regno >= 0)
1926 mark_reg_unknown(env, regs, value_regno);
1927 } else if (reg->type == PTR_TO_SOCKET) {
1928 if (t == BPF_WRITE) {
1929 verbose(env, "cannot write into socket\n");
1932 err = check_sock_access(env, regno, off, size, t);
1933 if (!err && value_regno >= 0)
1934 mark_reg_unknown(env, regs, value_regno);
1936 verbose(env, "R%d invalid mem access '%s'\n", regno,
1937 reg_type_str[reg->type]);
1941 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
1942 regs[value_regno].type == SCALAR_VALUE) {
1943 /* b/h/w load zero-extends, mark upper bits as known 0 */
1944 coerce_reg_to_size(®s[value_regno], size);
1949 static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
1953 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
1955 verbose(env, "BPF_XADD uses reserved fields\n");
1959 /* check src1 operand */
1960 err = check_reg_arg(env, insn->src_reg, SRC_OP);
1964 /* check src2 operand */
1965 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
1969 if (is_pointer_value(env, insn->src_reg)) {
1970 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
1974 if (is_ctx_reg(env, insn->dst_reg) ||
1975 is_pkt_reg(env, insn->dst_reg) ||
1976 is_flow_key_reg(env, insn->dst_reg)) {
1977 verbose(env, "BPF_XADD stores into R%d %s is not allowed\n",
1979 reg_type_str[reg_state(env, insn->dst_reg)->type]);
1983 /* check whether atomic_add can read the memory */
1984 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1985 BPF_SIZE(insn->code), BPF_READ, -1, true);
1989 /* check whether atomic_add can write into the same memory */
1990 return check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1991 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
1994 /* when register 'regno' is passed into function that will read 'access_size'
1995 * bytes from that pointer, make sure that it's within stack boundary
1996 * and all elements of stack are initialized.
1997 * Unlike most pointer bounds-checking functions, this one doesn't take an
1998 * 'off' argument, so it has to add in reg->off itself.
2000 static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
2001 int access_size, bool zero_size_allowed,
2002 struct bpf_call_arg_meta *meta)
2004 struct bpf_reg_state *reg = reg_state(env, regno);
2005 struct bpf_func_state *state = func(env, reg);
2006 int off, i, slot, spi;
2008 if (reg->type != PTR_TO_STACK) {
2009 /* Allow zero-byte read from NULL, regardless of pointer type */
2010 if (zero_size_allowed && access_size == 0 &&
2011 register_is_null(reg))
2014 verbose(env, "R%d type=%s expected=%s\n", regno,
2015 reg_type_str[reg->type],
2016 reg_type_str[PTR_TO_STACK]);
2020 /* Only allow fixed-offset stack reads */
2021 if (!tnum_is_const(reg->var_off)) {
2024 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2025 verbose(env, "invalid variable stack read R%d var_off=%s\n",
2029 off = reg->off + reg->var_off.value;
2030 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
2031 access_size < 0 || (access_size == 0 && !zero_size_allowed)) {
2032 verbose(env, "invalid stack type R%d off=%d access_size=%d\n",
2033 regno, off, access_size);
2037 if (meta && meta->raw_mode) {
2038 meta->access_size = access_size;
2039 meta->regno = regno;
2043 for (i = 0; i < access_size; i++) {
2046 slot = -(off + i) - 1;
2047 spi = slot / BPF_REG_SIZE;
2048 if (state->allocated_stack <= slot)
2050 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
2051 if (*stype == STACK_MISC)
2053 if (*stype == STACK_ZERO) {
2054 /* helper can write anything into the stack */
2055 *stype = STACK_MISC;
2059 verbose(env, "invalid indirect read from stack off %d+%d size %d\n",
2060 off, i, access_size);
2063 /* reading any byte out of 8-byte 'spill_slot' will cause
2064 * the whole slot to be marked as 'read'
2066 mark_reg_read(env, &state->stack[spi].spilled_ptr,
2067 state->stack[spi].spilled_ptr.parent);
2069 return update_stack_depth(env, state, off);
2072 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
2073 int access_size, bool zero_size_allowed,
2074 struct bpf_call_arg_meta *meta)
2076 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
2078 switch (reg->type) {
2080 case PTR_TO_PACKET_META:
2081 return check_packet_access(env, regno, reg->off, access_size,
2083 case PTR_TO_MAP_VALUE:
2084 return check_map_access(env, regno, reg->off, access_size,
2086 default: /* scalar_value|ptr_to_stack or invalid ptr */
2087 return check_stack_boundary(env, regno, access_size,
2088 zero_size_allowed, meta);
2092 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
2094 return type == ARG_PTR_TO_MEM ||
2095 type == ARG_PTR_TO_MEM_OR_NULL ||
2096 type == ARG_PTR_TO_UNINIT_MEM;
2099 static bool arg_type_is_mem_size(enum bpf_arg_type type)
2101 return type == ARG_CONST_SIZE ||
2102 type == ARG_CONST_SIZE_OR_ZERO;
2105 static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
2106 enum bpf_arg_type arg_type,
2107 struct bpf_call_arg_meta *meta)
2109 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
2110 enum bpf_reg_type expected_type, type = reg->type;
2113 if (arg_type == ARG_DONTCARE)
2116 err = check_reg_arg(env, regno, SRC_OP);
2120 if (arg_type == ARG_ANYTHING) {
2121 if (is_pointer_value(env, regno)) {
2122 verbose(env, "R%d leaks addr into helper function\n",
2129 if (type_is_pkt_pointer(type) &&
2130 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
2131 verbose(env, "helper access to the packet is not allowed\n");
2135 if (arg_type == ARG_PTR_TO_MAP_KEY ||
2136 arg_type == ARG_PTR_TO_MAP_VALUE ||
2137 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
2138 expected_type = PTR_TO_STACK;
2139 if (!type_is_pkt_pointer(type) && type != PTR_TO_MAP_VALUE &&
2140 type != expected_type)
2142 } else if (arg_type == ARG_CONST_SIZE ||
2143 arg_type == ARG_CONST_SIZE_OR_ZERO) {
2144 expected_type = SCALAR_VALUE;
2145 if (type != expected_type)
2147 } else if (arg_type == ARG_CONST_MAP_PTR) {
2148 expected_type = CONST_PTR_TO_MAP;
2149 if (type != expected_type)
2151 } else if (arg_type == ARG_PTR_TO_CTX) {
2152 expected_type = PTR_TO_CTX;
2153 if (type != expected_type)
2155 err = check_ctx_reg(env, reg, regno);
2158 } else if (arg_type == ARG_PTR_TO_SOCKET) {
2159 expected_type = PTR_TO_SOCKET;
2160 if (type != expected_type)
2162 if (meta->ptr_id || !reg->id) {
2163 verbose(env, "verifier internal error: mismatched references meta=%d, reg=%d\n",
2164 meta->ptr_id, reg->id);
2167 meta->ptr_id = reg->id;
2168 } else if (arg_type_is_mem_ptr(arg_type)) {
2169 expected_type = PTR_TO_STACK;
2170 /* One exception here. In case function allows for NULL to be
2171 * passed in as argument, it's a SCALAR_VALUE type. Final test
2172 * happens during stack boundary checking.
2174 if (register_is_null(reg) &&
2175 arg_type == ARG_PTR_TO_MEM_OR_NULL)
2176 /* final test in check_stack_boundary() */;
2177 else if (!type_is_pkt_pointer(type) &&
2178 type != PTR_TO_MAP_VALUE &&
2179 type != expected_type)
2181 meta->raw_mode = arg_type == ARG_PTR_TO_UNINIT_MEM;
2183 verbose(env, "unsupported arg_type %d\n", arg_type);
2187 if (arg_type == ARG_CONST_MAP_PTR) {
2188 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
2189 meta->map_ptr = reg->map_ptr;
2190 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
2191 /* bpf_map_xxx(..., map_ptr, ..., key) call:
2192 * check that [key, key + map->key_size) are within
2193 * stack limits and initialized
2195 if (!meta->map_ptr) {
2196 /* in function declaration map_ptr must come before
2197 * map_key, so that it's verified and known before
2198 * we have to check map_key here. Otherwise it means
2199 * that kernel subsystem misconfigured verifier
2201 verbose(env, "invalid map_ptr to access map->key\n");
2204 err = check_helper_mem_access(env, regno,
2205 meta->map_ptr->key_size, false,
2207 } else if (arg_type == ARG_PTR_TO_MAP_VALUE ||
2208 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
2209 /* bpf_map_xxx(..., map_ptr, ..., value) call:
2210 * check [value, value + map->value_size) validity
2212 if (!meta->map_ptr) {
2213 /* kernel subsystem misconfigured verifier */
2214 verbose(env, "invalid map_ptr to access map->value\n");
2217 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE);
2218 err = check_helper_mem_access(env, regno,
2219 meta->map_ptr->value_size, false,
2221 } else if (arg_type_is_mem_size(arg_type)) {
2222 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
2224 /* remember the mem_size which may be used later
2225 * to refine return values.
2227 meta->msize_smax_value = reg->smax_value;
2228 meta->msize_umax_value = reg->umax_value;
2230 /* The register is SCALAR_VALUE; the access check
2231 * happens using its boundaries.
2233 if (!tnum_is_const(reg->var_off))
2234 /* For unprivileged variable accesses, disable raw
2235 * mode so that the program is required to
2236 * initialize all the memory that the helper could
2237 * just partially fill up.
2241 if (reg->smin_value < 0) {
2242 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
2247 if (reg->umin_value == 0) {
2248 err = check_helper_mem_access(env, regno - 1, 0,
2255 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
2256 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
2260 err = check_helper_mem_access(env, regno - 1,
2262 zero_size_allowed, meta);
2267 verbose(env, "R%d type=%s expected=%s\n", regno,
2268 reg_type_str[type], reg_type_str[expected_type]);
2272 static int check_map_func_compatibility(struct bpf_verifier_env *env,
2273 struct bpf_map *map, int func_id)
2278 /* We need a two way check, first is from map perspective ... */
2279 switch (map->map_type) {
2280 case BPF_MAP_TYPE_PROG_ARRAY:
2281 if (func_id != BPF_FUNC_tail_call)
2284 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
2285 if (func_id != BPF_FUNC_perf_event_read &&
2286 func_id != BPF_FUNC_perf_event_output &&
2287 func_id != BPF_FUNC_perf_event_read_value)
2290 case BPF_MAP_TYPE_STACK_TRACE:
2291 if (func_id != BPF_FUNC_get_stackid)
2294 case BPF_MAP_TYPE_CGROUP_ARRAY:
2295 if (func_id != BPF_FUNC_skb_under_cgroup &&
2296 func_id != BPF_FUNC_current_task_under_cgroup)
2299 case BPF_MAP_TYPE_CGROUP_STORAGE:
2300 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
2301 if (func_id != BPF_FUNC_get_local_storage)
2304 /* devmap returns a pointer to a live net_device ifindex that we cannot
2305 * allow to be modified from bpf side. So do not allow lookup elements
2308 case BPF_MAP_TYPE_DEVMAP:
2309 if (func_id != BPF_FUNC_redirect_map)
2312 /* Restrict bpf side of cpumap and xskmap, open when use-cases
2315 case BPF_MAP_TYPE_CPUMAP:
2316 case BPF_MAP_TYPE_XSKMAP:
2317 if (func_id != BPF_FUNC_redirect_map)
2320 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
2321 case BPF_MAP_TYPE_HASH_OF_MAPS:
2322 if (func_id != BPF_FUNC_map_lookup_elem)
2325 case BPF_MAP_TYPE_SOCKMAP:
2326 if (func_id != BPF_FUNC_sk_redirect_map &&
2327 func_id != BPF_FUNC_sock_map_update &&
2328 func_id != BPF_FUNC_map_delete_elem &&
2329 func_id != BPF_FUNC_msg_redirect_map)
2332 case BPF_MAP_TYPE_SOCKHASH:
2333 if (func_id != BPF_FUNC_sk_redirect_hash &&
2334 func_id != BPF_FUNC_sock_hash_update &&
2335 func_id != BPF_FUNC_map_delete_elem &&
2336 func_id != BPF_FUNC_msg_redirect_hash)
2339 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
2340 if (func_id != BPF_FUNC_sk_select_reuseport)
2343 case BPF_MAP_TYPE_QUEUE:
2344 case BPF_MAP_TYPE_STACK:
2345 if (func_id != BPF_FUNC_map_peek_elem &&
2346 func_id != BPF_FUNC_map_pop_elem &&
2347 func_id != BPF_FUNC_map_push_elem)
2354 /* ... and second from the function itself. */
2356 case BPF_FUNC_tail_call:
2357 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
2359 if (env->subprog_cnt > 1) {
2360 verbose(env, "tail_calls are not allowed in programs with bpf-to-bpf calls\n");
2364 case BPF_FUNC_perf_event_read:
2365 case BPF_FUNC_perf_event_output:
2366 case BPF_FUNC_perf_event_read_value:
2367 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
2370 case BPF_FUNC_get_stackid:
2371 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
2374 case BPF_FUNC_current_task_under_cgroup:
2375 case BPF_FUNC_skb_under_cgroup:
2376 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
2379 case BPF_FUNC_redirect_map:
2380 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
2381 map->map_type != BPF_MAP_TYPE_CPUMAP &&
2382 map->map_type != BPF_MAP_TYPE_XSKMAP)
2385 case BPF_FUNC_sk_redirect_map:
2386 case BPF_FUNC_msg_redirect_map:
2387 case BPF_FUNC_sock_map_update:
2388 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
2391 case BPF_FUNC_sk_redirect_hash:
2392 case BPF_FUNC_msg_redirect_hash:
2393 case BPF_FUNC_sock_hash_update:
2394 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
2397 case BPF_FUNC_get_local_storage:
2398 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
2399 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
2402 case BPF_FUNC_sk_select_reuseport:
2403 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY)
2406 case BPF_FUNC_map_peek_elem:
2407 case BPF_FUNC_map_pop_elem:
2408 case BPF_FUNC_map_push_elem:
2409 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
2410 map->map_type != BPF_MAP_TYPE_STACK)
2419 verbose(env, "cannot pass map_type %d into func %s#%d\n",
2420 map->map_type, func_id_name(func_id), func_id);
2424 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
2428 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
2430 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
2432 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
2434 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
2436 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
2439 /* We only support one arg being in raw mode at the moment,
2440 * which is sufficient for the helper functions we have
2446 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
2447 enum bpf_arg_type arg_next)
2449 return (arg_type_is_mem_ptr(arg_curr) &&
2450 !arg_type_is_mem_size(arg_next)) ||
2451 (!arg_type_is_mem_ptr(arg_curr) &&
2452 arg_type_is_mem_size(arg_next));
2455 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
2457 /* bpf_xxx(..., buf, len) call will access 'len'
2458 * bytes from memory 'buf'. Both arg types need
2459 * to be paired, so make sure there's no buggy
2460 * helper function specification.
2462 if (arg_type_is_mem_size(fn->arg1_type) ||
2463 arg_type_is_mem_ptr(fn->arg5_type) ||
2464 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
2465 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
2466 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
2467 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
2473 static bool check_refcount_ok(const struct bpf_func_proto *fn)
2477 if (arg_type_is_refcounted(fn->arg1_type))
2479 if (arg_type_is_refcounted(fn->arg2_type))
2481 if (arg_type_is_refcounted(fn->arg3_type))
2483 if (arg_type_is_refcounted(fn->arg4_type))
2485 if (arg_type_is_refcounted(fn->arg5_type))
2488 /* We only support one arg being unreferenced at the moment,
2489 * which is sufficient for the helper functions we have right now.
2494 static int check_func_proto(const struct bpf_func_proto *fn)
2496 return check_raw_mode_ok(fn) &&
2497 check_arg_pair_ok(fn) &&
2498 check_refcount_ok(fn) ? 0 : -EINVAL;
2501 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
2502 * are now invalid, so turn them into unknown SCALAR_VALUE.
2504 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
2505 struct bpf_func_state *state)
2507 struct bpf_reg_state *regs = state->regs, *reg;
2510 for (i = 0; i < MAX_BPF_REG; i++)
2511 if (reg_is_pkt_pointer_any(®s[i]))
2512 mark_reg_unknown(env, regs, i);
2514 bpf_for_each_spilled_reg(i, state, reg) {
2517 if (reg_is_pkt_pointer_any(reg))
2518 __mark_reg_unknown(reg);
2522 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
2524 struct bpf_verifier_state *vstate = env->cur_state;
2527 for (i = 0; i <= vstate->curframe; i++)
2528 __clear_all_pkt_pointers(env, vstate->frame[i]);
2531 static void release_reg_references(struct bpf_verifier_env *env,
2532 struct bpf_func_state *state, int id)
2534 struct bpf_reg_state *regs = state->regs, *reg;
2537 for (i = 0; i < MAX_BPF_REG; i++)
2538 if (regs[i].id == id)
2539 mark_reg_unknown(env, regs, i);
2541 bpf_for_each_spilled_reg(i, state, reg) {
2544 if (reg_is_refcounted(reg) && reg->id == id)
2545 __mark_reg_unknown(reg);
2549 /* The pointer with the specified id has released its reference to kernel
2550 * resources. Identify all copies of the same pointer and clear the reference.
2552 static int release_reference(struct bpf_verifier_env *env,
2553 struct bpf_call_arg_meta *meta)
2555 struct bpf_verifier_state *vstate = env->cur_state;
2558 for (i = 0; i <= vstate->curframe; i++)
2559 release_reg_references(env, vstate->frame[i], meta->ptr_id);
2561 return release_reference_state(env, meta->ptr_id);
2564 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
2567 struct bpf_verifier_state *state = env->cur_state;
2568 struct bpf_func_state *caller, *callee;
2569 int i, err, subprog, target_insn;
2571 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
2572 verbose(env, "the call stack of %d frames is too deep\n",
2573 state->curframe + 2);
2577 target_insn = *insn_idx + insn->imm;
2578 subprog = find_subprog(env, target_insn + 1);
2580 verbose(env, "verifier bug. No program starts at insn %d\n",
2585 caller = state->frame[state->curframe];
2586 if (state->frame[state->curframe + 1]) {
2587 verbose(env, "verifier bug. Frame %d already allocated\n",
2588 state->curframe + 1);
2592 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
2595 state->frame[state->curframe + 1] = callee;
2597 /* callee cannot access r0, r6 - r9 for reading and has to write
2598 * into its own stack before reading from it.
2599 * callee can read/write into caller's stack
2601 init_func_state(env, callee,
2602 /* remember the callsite, it will be used by bpf_exit */
2603 *insn_idx /* callsite */,
2604 state->curframe + 1 /* frameno within this callchain */,
2605 subprog /* subprog number within this prog */);
2607 /* Transfer references to the callee */
2608 err = transfer_reference_state(callee, caller);
2612 /* copy r1 - r5 args that callee can access. The copy includes parent
2613 * pointers, which connects us up to the liveness chain
2615 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
2616 callee->regs[i] = caller->regs[i];
2618 /* after the call registers r0 - r5 were scratched */
2619 for (i = 0; i < CALLER_SAVED_REGS; i++) {
2620 mark_reg_not_init(env, caller->regs, caller_saved[i]);
2621 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
2624 /* only increment it after check_reg_arg() finished */
2627 /* and go analyze first insn of the callee */
2628 *insn_idx = target_insn;
2630 if (env->log.level) {
2631 verbose(env, "caller:\n");
2632 print_verifier_state(env, caller);
2633 verbose(env, "callee:\n");
2634 print_verifier_state(env, callee);
2639 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
2641 struct bpf_verifier_state *state = env->cur_state;
2642 struct bpf_func_state *caller, *callee;
2643 struct bpf_reg_state *r0;
2646 callee = state->frame[state->curframe];
2647 r0 = &callee->regs[BPF_REG_0];
2648 if (r0->type == PTR_TO_STACK) {
2649 /* technically it's ok to return caller's stack pointer
2650 * (or caller's caller's pointer) back to the caller,
2651 * since these pointers are valid. Only current stack
2652 * pointer will be invalid as soon as function exits,
2653 * but let's be conservative
2655 verbose(env, "cannot return stack pointer to the caller\n");
2660 caller = state->frame[state->curframe];
2661 /* return to the caller whatever r0 had in the callee */
2662 caller->regs[BPF_REG_0] = *r0;
2664 /* Transfer references to the caller */
2665 err = transfer_reference_state(caller, callee);
2669 *insn_idx = callee->callsite + 1;
2670 if (env->log.level) {
2671 verbose(env, "returning from callee:\n");
2672 print_verifier_state(env, callee);
2673 verbose(env, "to caller at %d:\n", *insn_idx);
2674 print_verifier_state(env, caller);
2676 /* clear everything in the callee */
2677 free_func_state(callee);
2678 state->frame[state->curframe + 1] = NULL;
2682 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
2684 struct bpf_call_arg_meta *meta)
2686 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
2688 if (ret_type != RET_INTEGER ||
2689 (func_id != BPF_FUNC_get_stack &&
2690 func_id != BPF_FUNC_probe_read_str))
2693 ret_reg->smax_value = meta->msize_smax_value;
2694 ret_reg->umax_value = meta->msize_umax_value;
2695 __reg_deduce_bounds(ret_reg);
2696 __reg_bound_offset(ret_reg);
2700 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
2701 int func_id, int insn_idx)
2703 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
2705 if (func_id != BPF_FUNC_tail_call &&
2706 func_id != BPF_FUNC_map_lookup_elem &&
2707 func_id != BPF_FUNC_map_update_elem &&
2708 func_id != BPF_FUNC_map_delete_elem &&
2709 func_id != BPF_FUNC_map_push_elem &&
2710 func_id != BPF_FUNC_map_pop_elem &&
2711 func_id != BPF_FUNC_map_peek_elem)
2714 if (meta->map_ptr == NULL) {
2715 verbose(env, "kernel subsystem misconfigured verifier\n");
2719 if (!BPF_MAP_PTR(aux->map_state))
2720 bpf_map_ptr_store(aux, meta->map_ptr,
2721 meta->map_ptr->unpriv_array);
2722 else if (BPF_MAP_PTR(aux->map_state) != meta->map_ptr)
2723 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
2724 meta->map_ptr->unpriv_array);
2728 static int check_reference_leak(struct bpf_verifier_env *env)
2730 struct bpf_func_state *state = cur_func(env);
2733 for (i = 0; i < state->acquired_refs; i++) {
2734 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
2735 state->refs[i].id, state->refs[i].insn_idx);
2737 return state->acquired_refs ? -EINVAL : 0;
2740 static int check_helper_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
2742 const struct bpf_func_proto *fn = NULL;
2743 struct bpf_reg_state *regs;
2744 struct bpf_call_arg_meta meta;
2748 /* find function prototype */
2749 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
2750 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
2755 if (env->ops->get_func_proto)
2756 fn = env->ops->get_func_proto(func_id, env->prog);
2758 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
2763 /* eBPF programs must be GPL compatible to use GPL-ed functions */
2764 if (!env->prog->gpl_compatible && fn->gpl_only) {
2765 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
2769 /* With LD_ABS/IND some JITs save/restore skb from r1. */
2770 changes_data = bpf_helper_changes_pkt_data(fn->func);
2771 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
2772 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
2773 func_id_name(func_id), func_id);
2777 memset(&meta, 0, sizeof(meta));
2778 meta.pkt_access = fn->pkt_access;
2780 err = check_func_proto(fn);
2782 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
2783 func_id_name(func_id), func_id);
2788 err = check_func_arg(env, BPF_REG_1, fn->arg1_type, &meta);
2791 err = check_func_arg(env, BPF_REG_2, fn->arg2_type, &meta);
2794 err = check_func_arg(env, BPF_REG_3, fn->arg3_type, &meta);
2797 err = check_func_arg(env, BPF_REG_4, fn->arg4_type, &meta);
2800 err = check_func_arg(env, BPF_REG_5, fn->arg5_type, &meta);
2804 err = record_func_map(env, &meta, func_id, insn_idx);
2808 /* Mark slots with STACK_MISC in case of raw mode, stack offset
2809 * is inferred from register state.
2811 for (i = 0; i < meta.access_size; i++) {
2812 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
2813 BPF_WRITE, -1, false);
2818 if (func_id == BPF_FUNC_tail_call) {
2819 err = check_reference_leak(env);
2821 verbose(env, "tail_call would lead to reference leak\n");
2824 } else if (is_release_function(func_id)) {
2825 err = release_reference(env, &meta);
2830 regs = cur_regs(env);
2832 /* check that flags argument in get_local_storage(map, flags) is 0,
2833 * this is required because get_local_storage() can't return an error.
2835 if (func_id == BPF_FUNC_get_local_storage &&
2836 !register_is_null(®s[BPF_REG_2])) {
2837 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
2841 /* reset caller saved regs */
2842 for (i = 0; i < CALLER_SAVED_REGS; i++) {
2843 mark_reg_not_init(env, regs, caller_saved[i]);
2844 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
2847 /* update return register (already marked as written above) */
2848 if (fn->ret_type == RET_INTEGER) {
2849 /* sets type to SCALAR_VALUE */
2850 mark_reg_unknown(env, regs, BPF_REG_0);
2851 } else if (fn->ret_type == RET_VOID) {
2852 regs[BPF_REG_0].type = NOT_INIT;
2853 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL ||
2854 fn->ret_type == RET_PTR_TO_MAP_VALUE) {
2855 if (fn->ret_type == RET_PTR_TO_MAP_VALUE)
2856 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE;
2858 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
2859 /* There is no offset yet applied, variable or fixed */
2860 mark_reg_known_zero(env, regs, BPF_REG_0);
2861 /* remember map_ptr, so that check_map_access()
2862 * can check 'value_size' boundary of memory access
2863 * to map element returned from bpf_map_lookup_elem()
2865 if (meta.map_ptr == NULL) {
2867 "kernel subsystem misconfigured verifier\n");
2870 regs[BPF_REG_0].map_ptr = meta.map_ptr;
2871 regs[BPF_REG_0].id = ++env->id_gen;
2872 } else if (fn->ret_type == RET_PTR_TO_SOCKET_OR_NULL) {
2873 int id = acquire_reference_state(env, insn_idx);
2876 mark_reg_known_zero(env, regs, BPF_REG_0);
2877 regs[BPF_REG_0].type = PTR_TO_SOCKET_OR_NULL;
2878 regs[BPF_REG_0].id = id;
2880 verbose(env, "unknown return type %d of func %s#%d\n",
2881 fn->ret_type, func_id_name(func_id), func_id);
2885 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
2887 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
2891 if (func_id == BPF_FUNC_get_stack && !env->prog->has_callchain_buf) {
2892 const char *err_str;
2894 #ifdef CONFIG_PERF_EVENTS
2895 err = get_callchain_buffers(sysctl_perf_event_max_stack);
2896 err_str = "cannot get callchain buffer for func %s#%d\n";
2899 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
2902 verbose(env, err_str, func_id_name(func_id), func_id);
2906 env->prog->has_callchain_buf = true;
2910 clear_all_pkt_pointers(env);
2914 static bool signed_add_overflows(s64 a, s64 b)
2916 /* Do the add in u64, where overflow is well-defined */
2917 s64 res = (s64)((u64)a + (u64)b);
2924 static bool signed_sub_overflows(s64 a, s64 b)
2926 /* Do the sub in u64, where overflow is well-defined */
2927 s64 res = (s64)((u64)a - (u64)b);
2934 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
2935 const struct bpf_reg_state *reg,
2936 enum bpf_reg_type type)
2938 bool known = tnum_is_const(reg->var_off);
2939 s64 val = reg->var_off.value;
2940 s64 smin = reg->smin_value;
2942 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
2943 verbose(env, "math between %s pointer and %lld is not allowed\n",
2944 reg_type_str[type], val);
2948 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
2949 verbose(env, "%s pointer offset %d is not allowed\n",
2950 reg_type_str[type], reg->off);
2954 if (smin == S64_MIN) {
2955 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
2956 reg_type_str[type]);
2960 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
2961 verbose(env, "value %lld makes %s pointer be out of bounds\n",
2962 smin, reg_type_str[type]);
2969 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
2970 * Caller should also handle BPF_MOV case separately.
2971 * If we return -EACCES, caller may want to try again treating pointer as a
2972 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
2974 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
2975 struct bpf_insn *insn,
2976 const struct bpf_reg_state *ptr_reg,
2977 const struct bpf_reg_state *off_reg)
2979 struct bpf_verifier_state *vstate = env->cur_state;
2980 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2981 struct bpf_reg_state *regs = state->regs, *dst_reg;
2982 bool known = tnum_is_const(off_reg->var_off);
2983 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
2984 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
2985 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
2986 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
2987 u8 opcode = BPF_OP(insn->code);
2988 u32 dst = insn->dst_reg;
2990 dst_reg = ®s[dst];
2992 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
2993 smin_val > smax_val || umin_val > umax_val) {
2994 /* Taint dst register if offset had invalid bounds derived from
2995 * e.g. dead branches.
2997 __mark_reg_unknown(dst_reg);
3001 if (BPF_CLASS(insn->code) != BPF_ALU64) {
3002 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
3004 "R%d 32-bit pointer arithmetic prohibited\n",
3009 switch (ptr_reg->type) {
3010 case PTR_TO_MAP_VALUE_OR_NULL:
3011 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
3012 dst, reg_type_str[ptr_reg->type]);
3014 case CONST_PTR_TO_MAP:
3015 case PTR_TO_PACKET_END:
3017 case PTR_TO_SOCKET_OR_NULL:
3018 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
3019 dst, reg_type_str[ptr_reg->type]);
3025 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
3026 * The id may be overwritten later if we create a new variable offset.
3028 dst_reg->type = ptr_reg->type;
3029 dst_reg->id = ptr_reg->id;
3031 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
3032 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
3037 /* We can take a fixed offset as long as it doesn't overflow
3038 * the s32 'off' field
3040 if (known && (ptr_reg->off + smin_val ==
3041 (s64)(s32)(ptr_reg->off + smin_val))) {
3042 /* pointer += K. Accumulate it into fixed offset */
3043 dst_reg->smin_value = smin_ptr;
3044 dst_reg->smax_value = smax_ptr;
3045 dst_reg->umin_value = umin_ptr;
3046 dst_reg->umax_value = umax_ptr;
3047 dst_reg->var_off = ptr_reg->var_off;
3048 dst_reg->off = ptr_reg->off + smin_val;
3049 dst_reg->range = ptr_reg->range;
3052 /* A new variable offset is created. Note that off_reg->off
3053 * == 0, since it's a scalar.
3054 * dst_reg gets the pointer type and since some positive
3055 * integer value was added to the pointer, give it a new 'id'
3056 * if it's a PTR_TO_PACKET.
3057 * this creates a new 'base' pointer, off_reg (variable) gets
3058 * added into the variable offset, and we copy the fixed offset
3061 if (signed_add_overflows(smin_ptr, smin_val) ||
3062 signed_add_overflows(smax_ptr, smax_val)) {
3063 dst_reg->smin_value = S64_MIN;
3064 dst_reg->smax_value = S64_MAX;
3066 dst_reg->smin_value = smin_ptr + smin_val;
3067 dst_reg->smax_value = smax_ptr + smax_val;
3069 if (umin_ptr + umin_val < umin_ptr ||
3070 umax_ptr + umax_val < umax_ptr) {
3071 dst_reg->umin_value = 0;
3072 dst_reg->umax_value = U64_MAX;
3074 dst_reg->umin_value = umin_ptr + umin_val;
3075 dst_reg->umax_value = umax_ptr + umax_val;
3077 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
3078 dst_reg->off = ptr_reg->off;
3079 if (reg_is_pkt_pointer(ptr_reg)) {
3080 dst_reg->id = ++env->id_gen;
3081 /* something was added to pkt_ptr, set range to zero */
3086 if (dst_reg == off_reg) {
3087 /* scalar -= pointer. Creates an unknown scalar */
3088 verbose(env, "R%d tried to subtract pointer from scalar\n",
3092 /* We don't allow subtraction from FP, because (according to
3093 * test_verifier.c test "invalid fp arithmetic", JITs might not
3094 * be able to deal with it.
3096 if (ptr_reg->type == PTR_TO_STACK) {
3097 verbose(env, "R%d subtraction from stack pointer prohibited\n",
3101 if (known && (ptr_reg->off - smin_val ==
3102 (s64)(s32)(ptr_reg->off - smin_val))) {
3103 /* pointer -= K. Subtract it from fixed offset */
3104 dst_reg->smin_value = smin_ptr;
3105 dst_reg->smax_value = smax_ptr;
3106 dst_reg->umin_value = umin_ptr;
3107 dst_reg->umax_value = umax_ptr;
3108 dst_reg->var_off = ptr_reg->var_off;
3109 dst_reg->id = ptr_reg->id;
3110 dst_reg->off = ptr_reg->off - smin_val;
3111 dst_reg->range = ptr_reg->range;
3114 /* A new variable offset is created. If the subtrahend is known
3115 * nonnegative, then any reg->range we had before is still good.
3117 if (signed_sub_overflows(smin_ptr, smax_val) ||
3118 signed_sub_overflows(smax_ptr, smin_val)) {
3119 /* Overflow possible, we know nothing */
3120 dst_reg->smin_value = S64_MIN;
3121 dst_reg->smax_value = S64_MAX;
3123 dst_reg->smin_value = smin_ptr - smax_val;
3124 dst_reg->smax_value = smax_ptr - smin_val;
3126 if (umin_ptr < umax_val) {
3127 /* Overflow possible, we know nothing */
3128 dst_reg->umin_value = 0;
3129 dst_reg->umax_value = U64_MAX;
3131 /* Cannot overflow (as long as bounds are consistent) */
3132 dst_reg->umin_value = umin_ptr - umax_val;
3133 dst_reg->umax_value = umax_ptr - umin_val;
3135 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
3136 dst_reg->off = ptr_reg->off;
3137 if (reg_is_pkt_pointer(ptr_reg)) {
3138 dst_reg->id = ++env->id_gen;
3139 /* something was added to pkt_ptr, set range to zero */
3147 /* bitwise ops on pointers are troublesome, prohibit. */
3148 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
3149 dst, bpf_alu_string[opcode >> 4]);
3152 /* other operators (e.g. MUL,LSH) produce non-pointer results */
3153 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
3154 dst, bpf_alu_string[opcode >> 4]);
3158 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
3161 __update_reg_bounds(dst_reg);
3162 __reg_deduce_bounds(dst_reg);
3163 __reg_bound_offset(dst_reg);
3167 /* WARNING: This function does calculations on 64-bit values, but the actual
3168 * execution may occur on 32-bit values. Therefore, things like bitshifts
3169 * need extra checks in the 32-bit case.
3171 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
3172 struct bpf_insn *insn,
3173 struct bpf_reg_state *dst_reg,
3174 struct bpf_reg_state src_reg)
3176 struct bpf_reg_state *regs = cur_regs(env);
3177 u8 opcode = BPF_OP(insn->code);
3178 bool src_known, dst_known;
3179 s64 smin_val, smax_val;
3180 u64 umin_val, umax_val;
3181 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
3183 if (insn_bitness == 32) {
3184 /* Relevant for 32-bit RSH: Information can propagate towards
3185 * LSB, so it isn't sufficient to only truncate the output to
3188 coerce_reg_to_size(dst_reg, 4);
3189 coerce_reg_to_size(&src_reg, 4);
3192 smin_val = src_reg.smin_value;
3193 smax_val = src_reg.smax_value;
3194 umin_val = src_reg.umin_value;
3195 umax_val = src_reg.umax_value;
3196 src_known = tnum_is_const(src_reg.var_off);
3197 dst_known = tnum_is_const(dst_reg->var_off);
3199 if ((src_known && (smin_val != smax_val || umin_val != umax_val)) ||
3200 smin_val > smax_val || umin_val > umax_val) {
3201 /* Taint dst register if offset had invalid bounds derived from
3202 * e.g. dead branches.
3204 __mark_reg_unknown(dst_reg);
3209 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
3210 __mark_reg_unknown(dst_reg);
3216 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
3217 signed_add_overflows(dst_reg->smax_value, smax_val)) {
3218 dst_reg->smin_value = S64_MIN;
3219 dst_reg->smax_value = S64_MAX;
3221 dst_reg->smin_value += smin_val;
3222 dst_reg->smax_value += smax_val;
3224 if (dst_reg->umin_value + umin_val < umin_val ||
3225 dst_reg->umax_value + umax_val < umax_val) {
3226 dst_reg->umin_value = 0;
3227 dst_reg->umax_value = U64_MAX;
3229 dst_reg->umin_value += umin_val;
3230 dst_reg->umax_value += umax_val;
3232 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
3235 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
3236 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
3237 /* Overflow possible, we know nothing */
3238 dst_reg->smin_value = S64_MIN;
3239 dst_reg->smax_value = S64_MAX;
3241 dst_reg->smin_value -= smax_val;
3242 dst_reg->smax_value -= smin_val;
3244 if (dst_reg->umin_value < umax_val) {
3245 /* Overflow possible, we know nothing */
3246 dst_reg->umin_value = 0;
3247 dst_reg->umax_value = U64_MAX;
3249 /* Cannot overflow (as long as bounds are consistent) */
3250 dst_reg->umin_value -= umax_val;
3251 dst_reg->umax_value -= umin_val;
3253 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
3256 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
3257 if (smin_val < 0 || dst_reg->smin_value < 0) {
3258 /* Ain't nobody got time to multiply that sign */
3259 __mark_reg_unbounded(dst_reg);
3260 __update_reg_bounds(dst_reg);
3263 /* Both values are positive, so we can work with unsigned and
3264 * copy the result to signed (unless it exceeds S64_MAX).
3266 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
3267 /* Potential overflow, we know nothing */
3268 __mark_reg_unbounded(dst_reg);
3269 /* (except what we can learn from the var_off) */
3270 __update_reg_bounds(dst_reg);
3273 dst_reg->umin_value *= umin_val;
3274 dst_reg->umax_value *= umax_val;
3275 if (dst_reg->umax_value > S64_MAX) {
3276 /* Overflow possible, we know nothing */
3277 dst_reg->smin_value = S64_MIN;
3278 dst_reg->smax_value = S64_MAX;
3280 dst_reg->smin_value = dst_reg->umin_value;
3281 dst_reg->smax_value = dst_reg->umax_value;
3285 if (src_known && dst_known) {
3286 __mark_reg_known(dst_reg, dst_reg->var_off.value &
3287 src_reg.var_off.value);
3290 /* We get our minimum from the var_off, since that's inherently
3291 * bitwise. Our maximum is the minimum of the operands' maxima.
3293 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
3294 dst_reg->umin_value = dst_reg->var_off.value;
3295 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
3296 if (dst_reg->smin_value < 0 || smin_val < 0) {
3297 /* Lose signed bounds when ANDing negative numbers,
3298 * ain't nobody got time for that.
3300 dst_reg->smin_value = S64_MIN;
3301 dst_reg->smax_value = S64_MAX;
3303 /* ANDing two positives gives a positive, so safe to
3304 * cast result into s64.
3306 dst_reg->smin_value = dst_reg->umin_value;
3307 dst_reg->smax_value = dst_reg->umax_value;
3309 /* We may learn something more from the var_off */
3310 __update_reg_bounds(dst_reg);
3313 if (src_known && dst_known) {
3314 __mark_reg_known(dst_reg, dst_reg->var_off.value |
3315 src_reg.var_off.value);
3318 /* We get our maximum from the var_off, and our minimum is the
3319 * maximum of the operands' minima
3321 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
3322 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
3323 dst_reg->umax_value = dst_reg->var_off.value |
3324 dst_reg->var_off.mask;
3325 if (dst_reg->smin_value < 0 || smin_val < 0) {
3326 /* Lose signed bounds when ORing negative numbers,
3327 * ain't nobody got time for that.
3329 dst_reg->smin_value = S64_MIN;
3330 dst_reg->smax_value = S64_MAX;
3332 /* ORing two positives gives a positive, so safe to
3333 * cast result into s64.
3335 dst_reg->smin_value = dst_reg->umin_value;
3336 dst_reg->smax_value = dst_reg->umax_value;
3338 /* We may learn something more from the var_off */
3339 __update_reg_bounds(dst_reg);
3342 if (umax_val >= insn_bitness) {
3343 /* Shifts greater than 31 or 63 are undefined.
3344 * This includes shifts by a negative number.
3346 mark_reg_unknown(env, regs, insn->dst_reg);
3349 /* We lose all sign bit information (except what we can pick
3352 dst_reg->smin_value = S64_MIN;
3353 dst_reg->smax_value = S64_MAX;
3354 /* If we might shift our top bit out, then we know nothing */
3355 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
3356 dst_reg->umin_value = 0;
3357 dst_reg->umax_value = U64_MAX;
3359 dst_reg->umin_value <<= umin_val;
3360 dst_reg->umax_value <<= umax_val;
3362 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
3363 /* We may learn something more from the var_off */
3364 __update_reg_bounds(dst_reg);
3367 if (umax_val >= insn_bitness) {
3368 /* Shifts greater than 31 or 63 are undefined.
3369 * This includes shifts by a negative number.
3371 mark_reg_unknown(env, regs, insn->dst_reg);
3374 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
3375 * be negative, then either:
3376 * 1) src_reg might be zero, so the sign bit of the result is
3377 * unknown, so we lose our signed bounds
3378 * 2) it's known negative, thus the unsigned bounds capture the
3380 * 3) the signed bounds cross zero, so they tell us nothing
3382 * If the value in dst_reg is known nonnegative, then again the
3383 * unsigned bounts capture the signed bounds.
3384 * Thus, in all cases it suffices to blow away our signed bounds
3385 * and rely on inferring new ones from the unsigned bounds and
3386 * var_off of the result.
3388 dst_reg->smin_value = S64_MIN;
3389 dst_reg->smax_value = S64_MAX;
3390 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
3391 dst_reg->umin_value >>= umax_val;
3392 dst_reg->umax_value >>= umin_val;
3393 /* We may learn something more from the var_off */
3394 __update_reg_bounds(dst_reg);
3397 if (umax_val >= insn_bitness) {
3398 /* Shifts greater than 31 or 63 are undefined.
3399 * This includes shifts by a negative number.
3401 mark_reg_unknown(env, regs, insn->dst_reg);
3405 /* Upon reaching here, src_known is true and
3406 * umax_val is equal to umin_val.
3408 dst_reg->smin_value >>= umin_val;
3409 dst_reg->smax_value >>= umin_val;
3410 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val);
3412 /* blow away the dst_reg umin_value/umax_value and rely on
3413 * dst_reg var_off to refine the result.
3415 dst_reg->umin_value = 0;
3416 dst_reg->umax_value = U64_MAX;
3417 __update_reg_bounds(dst_reg);
3420 mark_reg_unknown(env, regs, insn->dst_reg);
3424 if (BPF_CLASS(insn->code) != BPF_ALU64) {
3425 /* 32-bit ALU ops are (32,32)->32 */
3426 coerce_reg_to_size(dst_reg, 4);
3429 __reg_deduce_bounds(dst_reg);
3430 __reg_bound_offset(dst_reg);
3434 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
3437 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
3438 struct bpf_insn *insn)
3440 struct bpf_verifier_state *vstate = env->cur_state;
3441 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3442 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
3443 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
3444 u8 opcode = BPF_OP(insn->code);
3446 dst_reg = ®s[insn->dst_reg];
3448 if (dst_reg->type != SCALAR_VALUE)
3450 if (BPF_SRC(insn->code) == BPF_X) {
3451 src_reg = ®s[insn->src_reg];
3452 if (src_reg->type != SCALAR_VALUE) {
3453 if (dst_reg->type != SCALAR_VALUE) {
3454 /* Combining two pointers by any ALU op yields
3455 * an arbitrary scalar. Disallow all math except
3456 * pointer subtraction
3458 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
3459 mark_reg_unknown(env, regs, insn->dst_reg);
3462 verbose(env, "R%d pointer %s pointer prohibited\n",
3464 bpf_alu_string[opcode >> 4]);
3467 /* scalar += pointer
3468 * This is legal, but we have to reverse our
3469 * src/dest handling in computing the range
3471 return adjust_ptr_min_max_vals(env, insn,
3474 } else if (ptr_reg) {
3475 /* pointer += scalar */
3476 return adjust_ptr_min_max_vals(env, insn,
3480 /* Pretend the src is a reg with a known value, since we only
3481 * need to be able to read from this state.
3483 off_reg.type = SCALAR_VALUE;
3484 __mark_reg_known(&off_reg, insn->imm);
3486 if (ptr_reg) /* pointer += K */
3487 return adjust_ptr_min_max_vals(env, insn,
3491 /* Got here implies adding two SCALAR_VALUEs */
3492 if (WARN_ON_ONCE(ptr_reg)) {
3493 print_verifier_state(env, state);
3494 verbose(env, "verifier internal error: unexpected ptr_reg\n");
3497 if (WARN_ON(!src_reg)) {
3498 print_verifier_state(env, state);
3499 verbose(env, "verifier internal error: no src_reg\n");
3502 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
3505 /* check validity of 32-bit and 64-bit arithmetic operations */
3506 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
3508 struct bpf_reg_state *regs = cur_regs(env);
3509 u8 opcode = BPF_OP(insn->code);
3512 if (opcode == BPF_END || opcode == BPF_NEG) {
3513 if (opcode == BPF_NEG) {
3514 if (BPF_SRC(insn->code) != 0 ||
3515 insn->src_reg != BPF_REG_0 ||
3516 insn->off != 0 || insn->imm != 0) {
3517 verbose(env, "BPF_NEG uses reserved fields\n");
3521 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
3522 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
3523 BPF_CLASS(insn->code) == BPF_ALU64) {
3524 verbose(env, "BPF_END uses reserved fields\n");
3529 /* check src operand */
3530 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3534 if (is_pointer_value(env, insn->dst_reg)) {
3535 verbose(env, "R%d pointer arithmetic prohibited\n",
3540 /* check dest operand */
3541 err = check_reg_arg(env, insn->dst_reg, DST_OP);
3545 } else if (opcode == BPF_MOV) {
3547 if (BPF_SRC(insn->code) == BPF_X) {
3548 if (insn->imm != 0 || insn->off != 0) {
3549 verbose(env, "BPF_MOV uses reserved fields\n");
3553 /* check src operand */
3554 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3558 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
3559 verbose(env, "BPF_MOV uses reserved fields\n");
3564 /* check dest operand, mark as required later */
3565 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
3569 if (BPF_SRC(insn->code) == BPF_X) {
3570 if (BPF_CLASS(insn->code) == BPF_ALU64) {
3572 * copy register state to dest reg
3574 regs[insn->dst_reg] = regs[insn->src_reg];
3575 regs[insn->dst_reg].live |= REG_LIVE_WRITTEN;
3578 if (is_pointer_value(env, insn->src_reg)) {
3580 "R%d partial copy of pointer\n",
3584 mark_reg_unknown(env, regs, insn->dst_reg);
3585 coerce_reg_to_size(®s[insn->dst_reg], 4);
3589 * remember the value we stored into this reg
3591 /* clear any state __mark_reg_known doesn't set */
3592 mark_reg_unknown(env, regs, insn->dst_reg);
3593 regs[insn->dst_reg].type = SCALAR_VALUE;
3594 if (BPF_CLASS(insn->code) == BPF_ALU64) {
3595 __mark_reg_known(regs + insn->dst_reg,
3598 __mark_reg_known(regs + insn->dst_reg,
3603 } else if (opcode > BPF_END) {
3604 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
3607 } else { /* all other ALU ops: and, sub, xor, add, ... */
3609 if (BPF_SRC(insn->code) == BPF_X) {
3610 if (insn->imm != 0 || insn->off != 0) {
3611 verbose(env, "BPF_ALU uses reserved fields\n");
3614 /* check src1 operand */
3615 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3619 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
3620 verbose(env, "BPF_ALU uses reserved fields\n");
3625 /* check src2 operand */
3626 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3630 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
3631 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
3632 verbose(env, "div by zero\n");
3636 if (opcode == BPF_ARSH && BPF_CLASS(insn->code) != BPF_ALU64) {
3637 verbose(env, "BPF_ARSH not supported for 32 bit ALU\n");
3641 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
3642 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
3643 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
3645 if (insn->imm < 0 || insn->imm >= size) {
3646 verbose(env, "invalid shift %d\n", insn->imm);
3651 /* check dest operand */
3652 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
3656 return adjust_reg_min_max_vals(env, insn);
3662 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
3663 struct bpf_reg_state *dst_reg,
3664 enum bpf_reg_type type,
3665 bool range_right_open)
3667 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3668 struct bpf_reg_state *regs = state->regs, *reg;
3672 if (dst_reg->off < 0 ||
3673 (dst_reg->off == 0 && range_right_open))
3674 /* This doesn't give us any range */
3677 if (dst_reg->umax_value > MAX_PACKET_OFF ||
3678 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
3679 /* Risk of overflow. For instance, ptr + (1<<63) may be less
3680 * than pkt_end, but that's because it's also less than pkt.
3684 new_range = dst_reg->off;
3685 if (range_right_open)
3688 /* Examples for register markings:
3690 * pkt_data in dst register:
3694 * if (r2 > pkt_end) goto <handle exception>
3699 * if (r2 < pkt_end) goto <access okay>
3700 * <handle exception>
3703 * r2 == dst_reg, pkt_end == src_reg
3704 * r2=pkt(id=n,off=8,r=0)
3705 * r3=pkt(id=n,off=0,r=0)
3707 * pkt_data in src register:
3711 * if (pkt_end >= r2) goto <access okay>
3712 * <handle exception>
3716 * if (pkt_end <= r2) goto <handle exception>
3720 * pkt_end == dst_reg, r2 == src_reg
3721 * r2=pkt(id=n,off=8,r=0)
3722 * r3=pkt(id=n,off=0,r=0)
3724 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
3725 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
3726 * and [r3, r3 + 8-1) respectively is safe to access depending on
3730 /* If our ids match, then we must have the same max_value. And we
3731 * don't care about the other reg's fixed offset, since if it's too big
3732 * the range won't allow anything.
3733 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
3735 for (i = 0; i < MAX_BPF_REG; i++)
3736 if (regs[i].type == type && regs[i].id == dst_reg->id)
3737 /* keep the maximum range already checked */
3738 regs[i].range = max(regs[i].range, new_range);
3740 for (j = 0; j <= vstate->curframe; j++) {
3741 state = vstate->frame[j];
3742 bpf_for_each_spilled_reg(i, state, reg) {
3745 if (reg->type == type && reg->id == dst_reg->id)
3746 reg->range = max(reg->range, new_range);
3751 /* Adjusts the register min/max values in the case that the dst_reg is the
3752 * variable register that we are working on, and src_reg is a constant or we're
3753 * simply doing a BPF_K check.
3754 * In JEQ/JNE cases we also adjust the var_off values.
3756 static void reg_set_min_max(struct bpf_reg_state *true_reg,
3757 struct bpf_reg_state *false_reg, u64 val,
3760 /* If the dst_reg is a pointer, we can't learn anything about its
3761 * variable offset from the compare (unless src_reg were a pointer into
3762 * the same object, but we don't bother with that.
3763 * Since false_reg and true_reg have the same type by construction, we
3764 * only need to check one of them for pointerness.
3766 if (__is_pointer_value(false, false_reg))
3771 /* If this is false then we know nothing Jon Snow, but if it is
3772 * true then we know for sure.
3774 __mark_reg_known(true_reg, val);
3777 /* If this is true we know nothing Jon Snow, but if it is false
3778 * we know the value for sure;
3780 __mark_reg_known(false_reg, val);
3783 false_reg->umax_value = min(false_reg->umax_value, val);
3784 true_reg->umin_value = max(true_reg->umin_value, val + 1);
3787 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
3788 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
3791 false_reg->umin_value = max(false_reg->umin_value, val);
3792 true_reg->umax_value = min(true_reg->umax_value, val - 1);
3795 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
3796 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
3799 false_reg->umax_value = min(false_reg->umax_value, val - 1);
3800 true_reg->umin_value = max(true_reg->umin_value, val);
3803 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
3804 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
3807 false_reg->umin_value = max(false_reg->umin_value, val + 1);
3808 true_reg->umax_value = min(true_reg->umax_value, val);
3811 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
3812 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
3818 __reg_deduce_bounds(false_reg);
3819 __reg_deduce_bounds(true_reg);
3820 /* We might have learned some bits from the bounds. */
3821 __reg_bound_offset(false_reg);
3822 __reg_bound_offset(true_reg);
3823 /* Intersecting with the old var_off might have improved our bounds
3824 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
3825 * then new var_off is (0; 0x7f...fc) which improves our umax.
3827 __update_reg_bounds(false_reg);
3828 __update_reg_bounds(true_reg);
3831 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
3834 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
3835 struct bpf_reg_state *false_reg, u64 val,
3838 if (__is_pointer_value(false, false_reg))
3843 /* If this is false then we know nothing Jon Snow, but if it is
3844 * true then we know for sure.
3846 __mark_reg_known(true_reg, val);
3849 /* If this is true we know nothing Jon Snow, but if it is false
3850 * we know the value for sure;
3852 __mark_reg_known(false_reg, val);
3855 true_reg->umax_value = min(true_reg->umax_value, val - 1);
3856 false_reg->umin_value = max(false_reg->umin_value, val);
3859 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
3860 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
3863 true_reg->umin_value = max(true_reg->umin_value, val + 1);
3864 false_reg->umax_value = min(false_reg->umax_value, val);
3867 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
3868 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
3871 true_reg->umax_value = min(true_reg->umax_value, val);
3872 false_reg->umin_value = max(false_reg->umin_value, val + 1);
3875 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
3876 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
3879 true_reg->umin_value = max(true_reg->umin_value, val);
3880 false_reg->umax_value = min(false_reg->umax_value, val - 1);
3883 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
3884 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
3890 __reg_deduce_bounds(false_reg);
3891 __reg_deduce_bounds(true_reg);
3892 /* We might have learned some bits from the bounds. */
3893 __reg_bound_offset(false_reg);
3894 __reg_bound_offset(true_reg);
3895 /* Intersecting with the old var_off might have improved our bounds
3896 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
3897 * then new var_off is (0; 0x7f...fc) which improves our umax.
3899 __update_reg_bounds(false_reg);
3900 __update_reg_bounds(true_reg);
3903 /* Regs are known to be equal, so intersect their min/max/var_off */
3904 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
3905 struct bpf_reg_state *dst_reg)
3907 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
3908 dst_reg->umin_value);
3909 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
3910 dst_reg->umax_value);
3911 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
3912 dst_reg->smin_value);
3913 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
3914 dst_reg->smax_value);
3915 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
3917 /* We might have learned new bounds from the var_off. */
3918 __update_reg_bounds(src_reg);
3919 __update_reg_bounds(dst_reg);
3920 /* We might have learned something about the sign bit. */
3921 __reg_deduce_bounds(src_reg);
3922 __reg_deduce_bounds(dst_reg);
3923 /* We might have learned some bits from the bounds. */
3924 __reg_bound_offset(src_reg);
3925 __reg_bound_offset(dst_reg);
3926 /* Intersecting with the old var_off might have improved our bounds
3927 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
3928 * then new var_off is (0; 0x7f...fc) which improves our umax.
3930 __update_reg_bounds(src_reg);
3931 __update_reg_bounds(dst_reg);
3934 static void reg_combine_min_max(struct bpf_reg_state *true_src,
3935 struct bpf_reg_state *true_dst,
3936 struct bpf_reg_state *false_src,
3937 struct bpf_reg_state *false_dst,
3942 __reg_combine_min_max(true_src, true_dst);
3945 __reg_combine_min_max(false_src, false_dst);
3950 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
3951 struct bpf_reg_state *reg, u32 id,
3954 if (reg_type_may_be_null(reg->type) && reg->id == id) {
3955 /* Old offset (both fixed and variable parts) should
3956 * have been known-zero, because we don't allow pointer
3957 * arithmetic on pointers that might be NULL.
3959 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
3960 !tnum_equals_const(reg->var_off, 0) ||
3962 __mark_reg_known_zero(reg);
3966 reg->type = SCALAR_VALUE;
3967 } else if (reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
3968 if (reg->map_ptr->inner_map_meta) {
3969 reg->type = CONST_PTR_TO_MAP;
3970 reg->map_ptr = reg->map_ptr->inner_map_meta;
3972 reg->type = PTR_TO_MAP_VALUE;
3974 } else if (reg->type == PTR_TO_SOCKET_OR_NULL) {
3975 reg->type = PTR_TO_SOCKET;
3977 if (is_null || !reg_is_refcounted(reg)) {
3978 /* We don't need id from this point onwards anymore,
3979 * thus we should better reset it, so that state
3980 * pruning has chances to take effect.
3987 /* The logic is similar to find_good_pkt_pointers(), both could eventually
3988 * be folded together at some point.
3990 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
3993 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3994 struct bpf_reg_state *reg, *regs = state->regs;
3995 u32 id = regs[regno].id;
3998 if (reg_is_refcounted_or_null(®s[regno]) && is_null)
3999 __release_reference_state(state, id);
4001 for (i = 0; i < MAX_BPF_REG; i++)
4002 mark_ptr_or_null_reg(state, ®s[i], id, is_null);
4004 for (j = 0; j <= vstate->curframe; j++) {
4005 state = vstate->frame[j];
4006 bpf_for_each_spilled_reg(i, state, reg) {
4009 mark_ptr_or_null_reg(state, reg, id, is_null);
4014 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
4015 struct bpf_reg_state *dst_reg,
4016 struct bpf_reg_state *src_reg,
4017 struct bpf_verifier_state *this_branch,
4018 struct bpf_verifier_state *other_branch)
4020 if (BPF_SRC(insn->code) != BPF_X)
4023 switch (BPF_OP(insn->code)) {
4025 if ((dst_reg->type == PTR_TO_PACKET &&
4026 src_reg->type == PTR_TO_PACKET_END) ||
4027 (dst_reg->type == PTR_TO_PACKET_META &&
4028 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
4029 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
4030 find_good_pkt_pointers(this_branch, dst_reg,
4031 dst_reg->type, false);
4032 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
4033 src_reg->type == PTR_TO_PACKET) ||
4034 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
4035 src_reg->type == PTR_TO_PACKET_META)) {
4036 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
4037 find_good_pkt_pointers(other_branch, src_reg,
4038 src_reg->type, true);
4044 if ((dst_reg->type == PTR_TO_PACKET &&
4045 src_reg->type == PTR_TO_PACKET_END) ||
4046 (dst_reg->type == PTR_TO_PACKET_META &&
4047 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
4048 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
4049 find_good_pkt_pointers(other_branch, dst_reg,
4050 dst_reg->type, true);
4051 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
4052 src_reg->type == PTR_TO_PACKET) ||
4053 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
4054 src_reg->type == PTR_TO_PACKET_META)) {
4055 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
4056 find_good_pkt_pointers(this_branch, src_reg,
4057 src_reg->type, false);
4063 if ((dst_reg->type == PTR_TO_PACKET &&
4064 src_reg->type == PTR_TO_PACKET_END) ||
4065 (dst_reg->type == PTR_TO_PACKET_META &&
4066 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
4067 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
4068 find_good_pkt_pointers(this_branch, dst_reg,
4069 dst_reg->type, true);
4070 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
4071 src_reg->type == PTR_TO_PACKET) ||
4072 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
4073 src_reg->type == PTR_TO_PACKET_META)) {
4074 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
4075 find_good_pkt_pointers(other_branch, src_reg,
4076 src_reg->type, false);
4082 if ((dst_reg->type == PTR_TO_PACKET &&
4083 src_reg->type == PTR_TO_PACKET_END) ||
4084 (dst_reg->type == PTR_TO_PACKET_META &&
4085 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
4086 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
4087 find_good_pkt_pointers(other_branch, dst_reg,
4088 dst_reg->type, false);
4089 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
4090 src_reg->type == PTR_TO_PACKET) ||
4091 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
4092 src_reg->type == PTR_TO_PACKET_META)) {
4093 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
4094 find_good_pkt_pointers(this_branch, src_reg,
4095 src_reg->type, true);
4107 static int check_cond_jmp_op(struct bpf_verifier_env *env,
4108 struct bpf_insn *insn, int *insn_idx)
4110 struct bpf_verifier_state *this_branch = env->cur_state;
4111 struct bpf_verifier_state *other_branch;
4112 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
4113 struct bpf_reg_state *dst_reg, *other_branch_regs;
4114 u8 opcode = BPF_OP(insn->code);
4117 if (opcode > BPF_JSLE) {
4118 verbose(env, "invalid BPF_JMP opcode %x\n", opcode);
4122 if (BPF_SRC(insn->code) == BPF_X) {
4123 if (insn->imm != 0) {
4124 verbose(env, "BPF_JMP uses reserved fields\n");
4128 /* check src1 operand */
4129 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4133 if (is_pointer_value(env, insn->src_reg)) {
4134 verbose(env, "R%d pointer comparison prohibited\n",
4139 if (insn->src_reg != BPF_REG_0) {
4140 verbose(env, "BPF_JMP uses reserved fields\n");
4145 /* check src2 operand */
4146 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4150 dst_reg = ®s[insn->dst_reg];
4152 /* detect if R == 0 where R was initialized to zero earlier */
4153 if (BPF_SRC(insn->code) == BPF_K &&
4154 (opcode == BPF_JEQ || opcode == BPF_JNE) &&
4155 dst_reg->type == SCALAR_VALUE &&
4156 tnum_is_const(dst_reg->var_off)) {
4157 if ((opcode == BPF_JEQ && dst_reg->var_off.value == insn->imm) ||
4158 (opcode == BPF_JNE && dst_reg->var_off.value != insn->imm)) {
4159 /* if (imm == imm) goto pc+off;
4160 * only follow the goto, ignore fall-through
4162 *insn_idx += insn->off;
4165 /* if (imm != imm) goto pc+off;
4166 * only follow fall-through branch, since
4167 * that's where the program will go
4173 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx);
4176 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
4178 /* detect if we are comparing against a constant value so we can adjust
4179 * our min/max values for our dst register.
4180 * this is only legit if both are scalars (or pointers to the same
4181 * object, I suppose, but we don't support that right now), because
4182 * otherwise the different base pointers mean the offsets aren't
4185 if (BPF_SRC(insn->code) == BPF_X) {
4186 if (dst_reg->type == SCALAR_VALUE &&
4187 regs[insn->src_reg].type == SCALAR_VALUE) {
4188 if (tnum_is_const(regs[insn->src_reg].var_off))
4189 reg_set_min_max(&other_branch_regs[insn->dst_reg],
4190 dst_reg, regs[insn->src_reg].var_off.value,
4192 else if (tnum_is_const(dst_reg->var_off))
4193 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
4194 ®s[insn->src_reg],
4195 dst_reg->var_off.value, opcode);
4196 else if (opcode == BPF_JEQ || opcode == BPF_JNE)
4197 /* Comparing for equality, we can combine knowledge */
4198 reg_combine_min_max(&other_branch_regs[insn->src_reg],
4199 &other_branch_regs[insn->dst_reg],
4200 ®s[insn->src_reg],
4201 ®s[insn->dst_reg], opcode);
4203 } else if (dst_reg->type == SCALAR_VALUE) {
4204 reg_set_min_max(&other_branch_regs[insn->dst_reg],
4205 dst_reg, insn->imm, opcode);
4208 /* detect if R == 0 where R is returned from bpf_map_lookup_elem() */
4209 if (BPF_SRC(insn->code) == BPF_K &&
4210 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
4211 reg_type_may_be_null(dst_reg->type)) {
4212 /* Mark all identical registers in each branch as either
4213 * safe or unknown depending R == 0 or R != 0 conditional.
4215 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
4217 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
4219 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
4220 this_branch, other_branch) &&
4221 is_pointer_value(env, insn->dst_reg)) {
4222 verbose(env, "R%d pointer comparison prohibited\n",
4227 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
4231 /* return the map pointer stored inside BPF_LD_IMM64 instruction */
4232 static struct bpf_map *ld_imm64_to_map_ptr(struct bpf_insn *insn)
4234 u64 imm64 = ((u64) (u32) insn[0].imm) | ((u64) (u32) insn[1].imm) << 32;
4236 return (struct bpf_map *) (unsigned long) imm64;
4239 /* verify BPF_LD_IMM64 instruction */
4240 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
4242 struct bpf_reg_state *regs = cur_regs(env);
4245 if (BPF_SIZE(insn->code) != BPF_DW) {
4246 verbose(env, "invalid BPF_LD_IMM insn\n");
4249 if (insn->off != 0) {
4250 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
4254 err = check_reg_arg(env, insn->dst_reg, DST_OP);
4258 if (insn->src_reg == 0) {
4259 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
4261 regs[insn->dst_reg].type = SCALAR_VALUE;
4262 __mark_reg_known(®s[insn->dst_reg], imm);
4266 /* replace_map_fd_with_map_ptr() should have caught bad ld_imm64 */
4267 BUG_ON(insn->src_reg != BPF_PSEUDO_MAP_FD);
4269 regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
4270 regs[insn->dst_reg].map_ptr = ld_imm64_to_map_ptr(insn);
4274 static bool may_access_skb(enum bpf_prog_type type)
4277 case BPF_PROG_TYPE_SOCKET_FILTER:
4278 case BPF_PROG_TYPE_SCHED_CLS:
4279 case BPF_PROG_TYPE_SCHED_ACT:
4286 /* verify safety of LD_ABS|LD_IND instructions:
4287 * - they can only appear in the programs where ctx == skb
4288 * - since they are wrappers of function calls, they scratch R1-R5 registers,
4289 * preserve R6-R9, and store return value into R0
4292 * ctx == skb == R6 == CTX
4295 * SRC == any register
4296 * IMM == 32-bit immediate
4299 * R0 - 8/16/32-bit skb data converted to cpu endianness
4301 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
4303 struct bpf_reg_state *regs = cur_regs(env);
4304 u8 mode = BPF_MODE(insn->code);
4307 if (!may_access_skb(env->prog->type)) {
4308 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
4312 if (!env->ops->gen_ld_abs) {
4313 verbose(env, "bpf verifier is misconfigured\n");
4317 if (env->subprog_cnt > 1) {
4318 /* when program has LD_ABS insn JITs and interpreter assume
4319 * that r1 == ctx == skb which is not the case for callees
4320 * that can have arbitrary arguments. It's problematic
4321 * for main prog as well since JITs would need to analyze
4322 * all functions in order to make proper register save/restore
4323 * decisions in the main prog. Hence disallow LD_ABS with calls
4325 verbose(env, "BPF_LD_[ABS|IND] instructions cannot be mixed with bpf-to-bpf calls\n");
4329 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
4330 BPF_SIZE(insn->code) == BPF_DW ||
4331 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
4332 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
4336 /* check whether implicit source operand (register R6) is readable */
4337 err = check_reg_arg(env, BPF_REG_6, SRC_OP);
4341 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
4342 * gen_ld_abs() may terminate the program at runtime, leading to
4345 err = check_reference_leak(env);
4347 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
4351 if (regs[BPF_REG_6].type != PTR_TO_CTX) {
4353 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
4357 if (mode == BPF_IND) {
4358 /* check explicit source operand */
4359 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4364 /* reset caller saved regs to unreadable */
4365 for (i = 0; i < CALLER_SAVED_REGS; i++) {
4366 mark_reg_not_init(env, regs, caller_saved[i]);
4367 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
4370 /* mark destination R0 register as readable, since it contains
4371 * the value fetched from the packet.
4372 * Already marked as written above.
4374 mark_reg_unknown(env, regs, BPF_REG_0);
4378 static int check_return_code(struct bpf_verifier_env *env)
4380 struct bpf_reg_state *reg;
4381 struct tnum range = tnum_range(0, 1);
4383 switch (env->prog->type) {
4384 case BPF_PROG_TYPE_CGROUP_SKB:
4385 case BPF_PROG_TYPE_CGROUP_SOCK:
4386 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
4387 case BPF_PROG_TYPE_SOCK_OPS:
4388 case BPF_PROG_TYPE_CGROUP_DEVICE:
4394 reg = cur_regs(env) + BPF_REG_0;
4395 if (reg->type != SCALAR_VALUE) {
4396 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
4397 reg_type_str[reg->type]);
4401 if (!tnum_in(range, reg->var_off)) {
4402 verbose(env, "At program exit the register R0 ");
4403 if (!tnum_is_unknown(reg->var_off)) {
4406 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4407 verbose(env, "has value %s", tn_buf);
4409 verbose(env, "has unknown scalar value");
4411 verbose(env, " should have been 0 or 1\n");
4417 /* non-recursive DFS pseudo code
4418 * 1 procedure DFS-iterative(G,v):
4419 * 2 label v as discovered
4420 * 3 let S be a stack
4422 * 5 while S is not empty
4424 * 7 if t is what we're looking for:
4426 * 9 for all edges e in G.adjacentEdges(t) do
4427 * 10 if edge e is already labelled
4428 * 11 continue with the next edge
4429 * 12 w <- G.adjacentVertex(t,e)
4430 * 13 if vertex w is not discovered and not explored
4431 * 14 label e as tree-edge
4432 * 15 label w as discovered
4435 * 18 else if vertex w is discovered
4436 * 19 label e as back-edge
4438 * 21 // vertex w is explored
4439 * 22 label e as forward- or cross-edge
4440 * 23 label t as explored
4445 * 0x11 - discovered and fall-through edge labelled
4446 * 0x12 - discovered and fall-through and branch edges labelled
4457 #define STATE_LIST_MARK ((struct bpf_verifier_state_list *) -1L)
4459 static int *insn_stack; /* stack of insns to process */
4460 static int cur_stack; /* current stack index */
4461 static int *insn_state;
4463 /* t, w, e - match pseudo-code above:
4464 * t - index of current instruction
4465 * w - next instruction
4468 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env)
4470 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
4473 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
4476 if (w < 0 || w >= env->prog->len) {
4477 verbose(env, "jump out of range from insn %d to %d\n", t, w);
4482 /* mark branch target for state pruning */
4483 env->explored_states[w] = STATE_LIST_MARK;
4485 if (insn_state[w] == 0) {
4487 insn_state[t] = DISCOVERED | e;
4488 insn_state[w] = DISCOVERED;
4489 if (cur_stack >= env->prog->len)
4491 insn_stack[cur_stack++] = w;
4493 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
4494 verbose(env, "back-edge from insn %d to %d\n", t, w);
4496 } else if (insn_state[w] == EXPLORED) {
4497 /* forward- or cross-edge */
4498 insn_state[t] = DISCOVERED | e;
4500 verbose(env, "insn state internal bug\n");
4506 /* non-recursive depth-first-search to detect loops in BPF program
4507 * loop == back-edge in directed graph
4509 static int check_cfg(struct bpf_verifier_env *env)
4511 struct bpf_insn *insns = env->prog->insnsi;
4512 int insn_cnt = env->prog->len;
4516 ret = check_subprogs(env);
4520 insn_state = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
4524 insn_stack = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
4530 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
4531 insn_stack[0] = 0; /* 0 is the first instruction */
4537 t = insn_stack[cur_stack - 1];
4539 if (BPF_CLASS(insns[t].code) == BPF_JMP) {
4540 u8 opcode = BPF_OP(insns[t].code);
4542 if (opcode == BPF_EXIT) {
4544 } else if (opcode == BPF_CALL) {
4545 ret = push_insn(t, t + 1, FALLTHROUGH, env);
4550 if (t + 1 < insn_cnt)
4551 env->explored_states[t + 1] = STATE_LIST_MARK;
4552 if (insns[t].src_reg == BPF_PSEUDO_CALL) {
4553 env->explored_states[t] = STATE_LIST_MARK;
4554 ret = push_insn(t, t + insns[t].imm + 1, BRANCH, env);
4560 } else if (opcode == BPF_JA) {
4561 if (BPF_SRC(insns[t].code) != BPF_K) {
4565 /* unconditional jump with single edge */
4566 ret = push_insn(t, t + insns[t].off + 1,
4572 /* tell verifier to check for equivalent states
4573 * after every call and jump
4575 if (t + 1 < insn_cnt)
4576 env->explored_states[t + 1] = STATE_LIST_MARK;
4578 /* conditional jump with two edges */
4579 env->explored_states[t] = STATE_LIST_MARK;
4580 ret = push_insn(t, t + 1, FALLTHROUGH, env);
4586 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env);
4593 /* all other non-branch instructions with single
4596 ret = push_insn(t, t + 1, FALLTHROUGH, env);
4604 insn_state[t] = EXPLORED;
4605 if (cur_stack-- <= 0) {
4606 verbose(env, "pop stack internal bug\n");
4613 for (i = 0; i < insn_cnt; i++) {
4614 if (insn_state[i] != EXPLORED) {
4615 verbose(env, "unreachable insn %d\n", i);
4620 ret = 0; /* cfg looks good */
4628 /* check %cur's range satisfies %old's */
4629 static bool range_within(struct bpf_reg_state *old,
4630 struct bpf_reg_state *cur)
4632 return old->umin_value <= cur->umin_value &&
4633 old->umax_value >= cur->umax_value &&
4634 old->smin_value <= cur->smin_value &&
4635 old->smax_value >= cur->smax_value;
4638 /* Maximum number of register states that can exist at once */
4639 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
4645 /* If in the old state two registers had the same id, then they need to have
4646 * the same id in the new state as well. But that id could be different from
4647 * the old state, so we need to track the mapping from old to new ids.
4648 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
4649 * regs with old id 5 must also have new id 9 for the new state to be safe. But
4650 * regs with a different old id could still have new id 9, we don't care about
4652 * So we look through our idmap to see if this old id has been seen before. If
4653 * so, we require the new id to match; otherwise, we add the id pair to the map.
4655 static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
4659 for (i = 0; i < ID_MAP_SIZE; i++) {
4660 if (!idmap[i].old) {
4661 /* Reached an empty slot; haven't seen this id before */
4662 idmap[i].old = old_id;
4663 idmap[i].cur = cur_id;
4666 if (idmap[i].old == old_id)
4667 return idmap[i].cur == cur_id;
4669 /* We ran out of idmap slots, which should be impossible */
4674 /* Returns true if (rold safe implies rcur safe) */
4675 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
4676 struct idpair *idmap)
4680 if (!(rold->live & REG_LIVE_READ))
4681 /* explored state didn't use this */
4684 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
4686 if (rold->type == PTR_TO_STACK)
4687 /* two stack pointers are equal only if they're pointing to
4688 * the same stack frame, since fp-8 in foo != fp-8 in bar
4690 return equal && rold->frameno == rcur->frameno;
4695 if (rold->type == NOT_INIT)
4696 /* explored state can't have used this */
4698 if (rcur->type == NOT_INIT)
4700 switch (rold->type) {
4702 if (rcur->type == SCALAR_VALUE) {
4703 /* new val must satisfy old val knowledge */
4704 return range_within(rold, rcur) &&
4705 tnum_in(rold->var_off, rcur->var_off);
4707 /* We're trying to use a pointer in place of a scalar.
4708 * Even if the scalar was unbounded, this could lead to
4709 * pointer leaks because scalars are allowed to leak
4710 * while pointers are not. We could make this safe in
4711 * special cases if root is calling us, but it's
4712 * probably not worth the hassle.
4716 case PTR_TO_MAP_VALUE:
4717 /* If the new min/max/var_off satisfy the old ones and
4718 * everything else matches, we are OK.
4719 * We don't care about the 'id' value, because nothing
4720 * uses it for PTR_TO_MAP_VALUE (only for ..._OR_NULL)
4722 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
4723 range_within(rold, rcur) &&
4724 tnum_in(rold->var_off, rcur->var_off);
4725 case PTR_TO_MAP_VALUE_OR_NULL:
4726 /* a PTR_TO_MAP_VALUE could be safe to use as a
4727 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
4728 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
4729 * checked, doing so could have affected others with the same
4730 * id, and we can't check for that because we lost the id when
4731 * we converted to a PTR_TO_MAP_VALUE.
4733 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
4735 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
4737 /* Check our ids match any regs they're supposed to */
4738 return check_ids(rold->id, rcur->id, idmap);
4739 case PTR_TO_PACKET_META:
4741 if (rcur->type != rold->type)
4743 /* We must have at least as much range as the old ptr
4744 * did, so that any accesses which were safe before are
4745 * still safe. This is true even if old range < old off,
4746 * since someone could have accessed through (ptr - k), or
4747 * even done ptr -= k in a register, to get a safe access.
4749 if (rold->range > rcur->range)
4751 /* If the offsets don't match, we can't trust our alignment;
4752 * nor can we be sure that we won't fall out of range.
4754 if (rold->off != rcur->off)
4756 /* id relations must be preserved */
4757 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
4759 /* new val must satisfy old val knowledge */
4760 return range_within(rold, rcur) &&
4761 tnum_in(rold->var_off, rcur->var_off);
4763 case CONST_PTR_TO_MAP:
4764 case PTR_TO_PACKET_END:
4765 case PTR_TO_FLOW_KEYS:
4767 case PTR_TO_SOCKET_OR_NULL:
4768 /* Only valid matches are exact, which memcmp() above
4769 * would have accepted
4772 /* Don't know what's going on, just say it's not safe */
4776 /* Shouldn't get here; if we do, say it's not safe */
4781 static bool stacksafe(struct bpf_func_state *old,
4782 struct bpf_func_state *cur,
4783 struct idpair *idmap)
4787 /* if explored stack has more populated slots than current stack
4788 * such stacks are not equivalent
4790 if (old->allocated_stack > cur->allocated_stack)
4793 /* walk slots of the explored stack and ignore any additional
4794 * slots in the current stack, since explored(safe) state
4797 for (i = 0; i < old->allocated_stack; i++) {
4798 spi = i / BPF_REG_SIZE;
4800 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ))
4801 /* explored state didn't use this */
4804 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
4806 /* if old state was safe with misc data in the stack
4807 * it will be safe with zero-initialized stack.
4808 * The opposite is not true
4810 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
4811 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
4813 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
4814 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
4815 /* Ex: old explored (safe) state has STACK_SPILL in
4816 * this stack slot, but current has has STACK_MISC ->
4817 * this verifier states are not equivalent,
4818 * return false to continue verification of this path
4821 if (i % BPF_REG_SIZE)
4823 if (old->stack[spi].slot_type[0] != STACK_SPILL)
4825 if (!regsafe(&old->stack[spi].spilled_ptr,
4826 &cur->stack[spi].spilled_ptr,
4828 /* when explored and current stack slot are both storing
4829 * spilled registers, check that stored pointers types
4830 * are the same as well.
4831 * Ex: explored safe path could have stored
4832 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
4833 * but current path has stored:
4834 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
4835 * such verifier states are not equivalent.
4836 * return false to continue verification of this path
4843 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
4845 if (old->acquired_refs != cur->acquired_refs)
4847 return !memcmp(old->refs, cur->refs,
4848 sizeof(*old->refs) * old->acquired_refs);
4851 /* compare two verifier states
4853 * all states stored in state_list are known to be valid, since
4854 * verifier reached 'bpf_exit' instruction through them
4856 * this function is called when verifier exploring different branches of
4857 * execution popped from the state stack. If it sees an old state that has
4858 * more strict register state and more strict stack state then this execution
4859 * branch doesn't need to be explored further, since verifier already
4860 * concluded that more strict state leads to valid finish.
4862 * Therefore two states are equivalent if register state is more conservative
4863 * and explored stack state is more conservative than the current one.
4866 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
4867 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
4869 * In other words if current stack state (one being explored) has more
4870 * valid slots than old one that already passed validation, it means
4871 * the verifier can stop exploring and conclude that current state is valid too
4873 * Similarly with registers. If explored state has register type as invalid
4874 * whereas register type in current state is meaningful, it means that
4875 * the current state will reach 'bpf_exit' instruction safely
4877 static bool func_states_equal(struct bpf_func_state *old,
4878 struct bpf_func_state *cur)
4880 struct idpair *idmap;
4884 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
4885 /* If we failed to allocate the idmap, just say it's not safe */
4889 for (i = 0; i < MAX_BPF_REG; i++) {
4890 if (!regsafe(&old->regs[i], &cur->regs[i], idmap))
4894 if (!stacksafe(old, cur, idmap))
4897 if (!refsafe(old, cur))
4905 static bool states_equal(struct bpf_verifier_env *env,
4906 struct bpf_verifier_state *old,
4907 struct bpf_verifier_state *cur)
4911 if (old->curframe != cur->curframe)
4914 /* for states to be equal callsites have to be the same
4915 * and all frame states need to be equivalent
4917 for (i = 0; i <= old->curframe; i++) {
4918 if (old->frame[i]->callsite != cur->frame[i]->callsite)
4920 if (!func_states_equal(old->frame[i], cur->frame[i]))
4926 /* A write screens off any subsequent reads; but write marks come from the
4927 * straight-line code between a state and its parent. When we arrive at an
4928 * equivalent state (jump target or such) we didn't arrive by the straight-line
4929 * code, so read marks in the state must propagate to the parent regardless
4930 * of the state's write marks. That's what 'parent == state->parent' comparison
4931 * in mark_reg_read() is for.
4933 static int propagate_liveness(struct bpf_verifier_env *env,
4934 const struct bpf_verifier_state *vstate,
4935 struct bpf_verifier_state *vparent)
4937 int i, frame, err = 0;
4938 struct bpf_func_state *state, *parent;
4940 if (vparent->curframe != vstate->curframe) {
4941 WARN(1, "propagate_live: parent frame %d current frame %d\n",
4942 vparent->curframe, vstate->curframe);
4945 /* Propagate read liveness of registers... */
4946 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
4947 /* We don't need to worry about FP liveness because it's read-only */
4948 for (i = 0; i < BPF_REG_FP; i++) {
4949 if (vparent->frame[vparent->curframe]->regs[i].live & REG_LIVE_READ)
4951 if (vstate->frame[vstate->curframe]->regs[i].live & REG_LIVE_READ) {
4952 err = mark_reg_read(env, &vstate->frame[vstate->curframe]->regs[i],
4953 &vparent->frame[vstate->curframe]->regs[i]);
4959 /* ... and stack slots */
4960 for (frame = 0; frame <= vstate->curframe; frame++) {
4961 state = vstate->frame[frame];
4962 parent = vparent->frame[frame];
4963 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
4964 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
4965 if (parent->stack[i].spilled_ptr.live & REG_LIVE_READ)
4967 if (state->stack[i].spilled_ptr.live & REG_LIVE_READ)
4968 mark_reg_read(env, &state->stack[i].spilled_ptr,
4969 &parent->stack[i].spilled_ptr);
4975 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
4977 struct bpf_verifier_state_list *new_sl;
4978 struct bpf_verifier_state_list *sl;
4979 struct bpf_verifier_state *cur = env->cur_state, *new;
4982 sl = env->explored_states[insn_idx];
4984 /* this 'insn_idx' instruction wasn't marked, so we will not
4985 * be doing state search here
4989 while (sl != STATE_LIST_MARK) {
4990 if (states_equal(env, &sl->state, cur)) {
4991 /* reached equivalent register/stack state,
4993 * Registers read by the continuation are read by us.
4994 * If we have any write marks in env->cur_state, they
4995 * will prevent corresponding reads in the continuation
4996 * from reaching our parent (an explored_state). Our
4997 * own state will get the read marks recorded, but
4998 * they'll be immediately forgotten as we're pruning
4999 * this state and will pop a new one.
5001 err = propagate_liveness(env, &sl->state, cur);
5009 /* there were no equivalent states, remember current one.
5010 * technically the current state is not proven to be safe yet,
5011 * but it will either reach outer most bpf_exit (which means it's safe)
5012 * or it will be rejected. Since there are no loops, we won't be
5013 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
5014 * again on the way to bpf_exit
5016 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
5020 /* add new state to the head of linked list */
5021 new = &new_sl->state;
5022 err = copy_verifier_state(new, cur);
5024 free_verifier_state(new, false);
5028 new_sl->next = env->explored_states[insn_idx];
5029 env->explored_states[insn_idx] = new_sl;
5030 /* connect new state to parentage chain */
5031 for (i = 0; i < BPF_REG_FP; i++)
5032 cur_regs(env)[i].parent = &new->frame[new->curframe]->regs[i];
5033 /* clear write marks in current state: the writes we did are not writes
5034 * our child did, so they don't screen off its reads from us.
5035 * (There are no read marks in current state, because reads always mark
5036 * their parent and current state never has children yet. Only
5037 * explored_states can get read marks.)
5039 for (i = 0; i < BPF_REG_FP; i++)
5040 cur->frame[cur->curframe]->regs[i].live = REG_LIVE_NONE;
5042 /* all stack frames are accessible from callee, clear them all */
5043 for (j = 0; j <= cur->curframe; j++) {
5044 struct bpf_func_state *frame = cur->frame[j];
5045 struct bpf_func_state *newframe = new->frame[j];
5047 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
5048 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
5049 frame->stack[i].spilled_ptr.parent =
5050 &newframe->stack[i].spilled_ptr;
5056 /* Return true if it's OK to have the same insn return a different type. */
5057 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
5062 case PTR_TO_SOCKET_OR_NULL:
5069 /* If an instruction was previously used with particular pointer types, then we
5070 * need to be careful to avoid cases such as the below, where it may be ok
5071 * for one branch accessing the pointer, but not ok for the other branch:
5076 * R1 = some_other_valid_ptr;
5079 * R2 = *(u32 *)(R1 + 0);
5081 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
5083 return src != prev && (!reg_type_mismatch_ok(src) ||
5084 !reg_type_mismatch_ok(prev));
5087 static int do_check(struct bpf_verifier_env *env)
5089 struct bpf_verifier_state *state;
5090 struct bpf_insn *insns = env->prog->insnsi;
5091 struct bpf_reg_state *regs;
5092 int insn_cnt = env->prog->len, i;
5093 int insn_idx, prev_insn_idx = 0;
5094 int insn_processed = 0;
5095 bool do_print_state = false;
5097 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
5100 state->curframe = 0;
5101 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
5102 if (!state->frame[0]) {
5106 env->cur_state = state;
5107 init_func_state(env, state->frame[0],
5108 BPF_MAIN_FUNC /* callsite */,
5110 0 /* subprogno, zero == main subprog */);
5113 struct bpf_insn *insn;
5117 if (insn_idx >= insn_cnt) {
5118 verbose(env, "invalid insn idx %d insn_cnt %d\n",
5119 insn_idx, insn_cnt);
5123 insn = &insns[insn_idx];
5124 class = BPF_CLASS(insn->code);
5126 if (++insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
5128 "BPF program is too large. Processed %d insn\n",
5133 err = is_state_visited(env, insn_idx);
5137 /* found equivalent state, can prune the search */
5138 if (env->log.level) {
5140 verbose(env, "\nfrom %d to %d: safe\n",
5141 prev_insn_idx, insn_idx);
5143 verbose(env, "%d: safe\n", insn_idx);
5145 goto process_bpf_exit;
5151 if (env->log.level > 1 || (env->log.level && do_print_state)) {
5152 if (env->log.level > 1)
5153 verbose(env, "%d:", insn_idx);
5155 verbose(env, "\nfrom %d to %d:",
5156 prev_insn_idx, insn_idx);
5157 print_verifier_state(env, state->frame[state->curframe]);
5158 do_print_state = false;
5161 if (env->log.level) {
5162 const struct bpf_insn_cbs cbs = {
5163 .cb_print = verbose,
5164 .private_data = env,
5167 verbose(env, "%d: ", insn_idx);
5168 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
5171 if (bpf_prog_is_dev_bound(env->prog->aux)) {
5172 err = bpf_prog_offload_verify_insn(env, insn_idx,
5178 regs = cur_regs(env);
5179 env->insn_aux_data[insn_idx].seen = true;
5181 if (class == BPF_ALU || class == BPF_ALU64) {
5182 err = check_alu_op(env, insn);
5186 } else if (class == BPF_LDX) {
5187 enum bpf_reg_type *prev_src_type, src_reg_type;
5189 /* check for reserved fields is already done */
5191 /* check src operand */
5192 err = check_reg_arg(env, insn->src_reg, SRC_OP);
5196 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
5200 src_reg_type = regs[insn->src_reg].type;
5202 /* check that memory (src_reg + off) is readable,
5203 * the state of dst_reg will be updated by this func
5205 err = check_mem_access(env, insn_idx, insn->src_reg, insn->off,
5206 BPF_SIZE(insn->code), BPF_READ,
5207 insn->dst_reg, false);
5211 prev_src_type = &env->insn_aux_data[insn_idx].ptr_type;
5213 if (*prev_src_type == NOT_INIT) {
5215 * dst_reg = *(u32 *)(src_reg + off)
5216 * save type to validate intersecting paths
5218 *prev_src_type = src_reg_type;
5220 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
5221 /* ABuser program is trying to use the same insn
5222 * dst_reg = *(u32*) (src_reg + off)
5223 * with different pointer types:
5224 * src_reg == ctx in one branch and
5225 * src_reg == stack|map in some other branch.
5228 verbose(env, "same insn cannot be used with different pointers\n");
5232 } else if (class == BPF_STX) {
5233 enum bpf_reg_type *prev_dst_type, dst_reg_type;
5235 if (BPF_MODE(insn->code) == BPF_XADD) {
5236 err = check_xadd(env, insn_idx, insn);
5243 /* check src1 operand */
5244 err = check_reg_arg(env, insn->src_reg, SRC_OP);
5247 /* check src2 operand */
5248 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
5252 dst_reg_type = regs[insn->dst_reg].type;
5254 /* check that memory (dst_reg + off) is writeable */
5255 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
5256 BPF_SIZE(insn->code), BPF_WRITE,
5257 insn->src_reg, false);
5261 prev_dst_type = &env->insn_aux_data[insn_idx].ptr_type;
5263 if (*prev_dst_type == NOT_INIT) {
5264 *prev_dst_type = dst_reg_type;
5265 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
5266 verbose(env, "same insn cannot be used with different pointers\n");
5270 } else if (class == BPF_ST) {
5271 if (BPF_MODE(insn->code) != BPF_MEM ||
5272 insn->src_reg != BPF_REG_0) {
5273 verbose(env, "BPF_ST uses reserved fields\n");
5276 /* check src operand */
5277 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
5281 if (is_ctx_reg(env, insn->dst_reg)) {
5282 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
5284 reg_type_str[reg_state(env, insn->dst_reg)->type]);
5288 /* check that memory (dst_reg + off) is writeable */
5289 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
5290 BPF_SIZE(insn->code), BPF_WRITE,
5295 } else if (class == BPF_JMP) {
5296 u8 opcode = BPF_OP(insn->code);
5298 if (opcode == BPF_CALL) {
5299 if (BPF_SRC(insn->code) != BPF_K ||
5301 (insn->src_reg != BPF_REG_0 &&
5302 insn->src_reg != BPF_PSEUDO_CALL) ||
5303 insn->dst_reg != BPF_REG_0) {
5304 verbose(env, "BPF_CALL uses reserved fields\n");
5308 if (insn->src_reg == BPF_PSEUDO_CALL)
5309 err = check_func_call(env, insn, &insn_idx);
5311 err = check_helper_call(env, insn->imm, insn_idx);
5315 } else if (opcode == BPF_JA) {
5316 if (BPF_SRC(insn->code) != BPF_K ||
5318 insn->src_reg != BPF_REG_0 ||
5319 insn->dst_reg != BPF_REG_0) {
5320 verbose(env, "BPF_JA uses reserved fields\n");
5324 insn_idx += insn->off + 1;
5327 } else if (opcode == BPF_EXIT) {
5328 if (BPF_SRC(insn->code) != BPF_K ||
5330 insn->src_reg != BPF_REG_0 ||
5331 insn->dst_reg != BPF_REG_0) {
5332 verbose(env, "BPF_EXIT uses reserved fields\n");
5336 if (state->curframe) {
5337 /* exit from nested function */
5338 prev_insn_idx = insn_idx;
5339 err = prepare_func_exit(env, &insn_idx);
5342 do_print_state = true;
5346 err = check_reference_leak(env);
5350 /* eBPF calling convetion is such that R0 is used
5351 * to return the value from eBPF program.
5352 * Make sure that it's readable at this time
5353 * of bpf_exit, which means that program wrote
5354 * something into it earlier
5356 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
5360 if (is_pointer_value(env, BPF_REG_0)) {
5361 verbose(env, "R0 leaks addr as return value\n");
5365 err = check_return_code(env);
5369 err = pop_stack(env, &prev_insn_idx, &insn_idx);
5375 do_print_state = true;
5379 err = check_cond_jmp_op(env, insn, &insn_idx);
5383 } else if (class == BPF_LD) {
5384 u8 mode = BPF_MODE(insn->code);
5386 if (mode == BPF_ABS || mode == BPF_IND) {
5387 err = check_ld_abs(env, insn);
5391 } else if (mode == BPF_IMM) {
5392 err = check_ld_imm(env, insn);
5397 env->insn_aux_data[insn_idx].seen = true;
5399 verbose(env, "invalid BPF_LD mode\n");
5403 verbose(env, "unknown insn class %d\n", class);
5410 verbose(env, "processed %d insns (limit %d), stack depth ",
5411 insn_processed, BPF_COMPLEXITY_LIMIT_INSNS);
5412 for (i = 0; i < env->subprog_cnt; i++) {
5413 u32 depth = env->subprog_info[i].stack_depth;
5415 verbose(env, "%d", depth);
5416 if (i + 1 < env->subprog_cnt)
5420 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
5424 static int check_map_prealloc(struct bpf_map *map)
5426 return (map->map_type != BPF_MAP_TYPE_HASH &&
5427 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
5428 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
5429 !(map->map_flags & BPF_F_NO_PREALLOC);
5432 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
5433 struct bpf_map *map,
5434 struct bpf_prog *prog)
5437 /* Make sure that BPF_PROG_TYPE_PERF_EVENT programs only use
5438 * preallocated hash maps, since doing memory allocation
5439 * in overflow_handler can crash depending on where nmi got
5442 if (prog->type == BPF_PROG_TYPE_PERF_EVENT) {
5443 if (!check_map_prealloc(map)) {
5444 verbose(env, "perf_event programs can only use preallocated hash map\n");
5447 if (map->inner_map_meta &&
5448 !check_map_prealloc(map->inner_map_meta)) {
5449 verbose(env, "perf_event programs can only use preallocated inner hash map\n");
5454 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
5455 !bpf_offload_prog_map_match(prog, map)) {
5456 verbose(env, "offload device mismatch between prog and map\n");
5463 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
5465 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
5466 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
5469 /* look for pseudo eBPF instructions that access map FDs and
5470 * replace them with actual map pointers
5472 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env *env)
5474 struct bpf_insn *insn = env->prog->insnsi;
5475 int insn_cnt = env->prog->len;
5478 err = bpf_prog_calc_tag(env->prog);
5482 for (i = 0; i < insn_cnt; i++, insn++) {
5483 if (BPF_CLASS(insn->code) == BPF_LDX &&
5484 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
5485 verbose(env, "BPF_LDX uses reserved fields\n");
5489 if (BPF_CLASS(insn->code) == BPF_STX &&
5490 ((BPF_MODE(insn->code) != BPF_MEM &&
5491 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
5492 verbose(env, "BPF_STX uses reserved fields\n");
5496 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
5497 struct bpf_map *map;
5500 if (i == insn_cnt - 1 || insn[1].code != 0 ||
5501 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
5503 verbose(env, "invalid bpf_ld_imm64 insn\n");
5507 if (insn->src_reg == 0)
5508 /* valid generic load 64-bit imm */
5511 if (insn->src_reg != BPF_PSEUDO_MAP_FD) {
5513 "unrecognized bpf_ld_imm64 insn\n");
5517 f = fdget(insn->imm);
5518 map = __bpf_map_get(f);
5520 verbose(env, "fd %d is not pointing to valid bpf_map\n",
5522 return PTR_ERR(map);
5525 err = check_map_prog_compatibility(env, map, env->prog);
5531 /* store map pointer inside BPF_LD_IMM64 instruction */
5532 insn[0].imm = (u32) (unsigned long) map;
5533 insn[1].imm = ((u64) (unsigned long) map) >> 32;
5535 /* check whether we recorded this map already */
5536 for (j = 0; j < env->used_map_cnt; j++)
5537 if (env->used_maps[j] == map) {
5542 if (env->used_map_cnt >= MAX_USED_MAPS) {
5547 /* hold the map. If the program is rejected by verifier,
5548 * the map will be released by release_maps() or it
5549 * will be used by the valid program until it's unloaded
5550 * and all maps are released in free_used_maps()
5552 map = bpf_map_inc(map, false);
5555 return PTR_ERR(map);
5557 env->used_maps[env->used_map_cnt++] = map;
5559 if (bpf_map_is_cgroup_storage(map) &&
5560 bpf_cgroup_storage_assign(env->prog, map)) {
5561 verbose(env, "only one cgroup storage of each type is allowed\n");
5573 /* Basic sanity check before we invest more work here. */
5574 if (!bpf_opcode_in_insntable(insn->code)) {
5575 verbose(env, "unknown opcode %02x\n", insn->code);
5580 /* now all pseudo BPF_LD_IMM64 instructions load valid
5581 * 'struct bpf_map *' into a register instead of user map_fd.
5582 * These pointers will be used later by verifier to validate map access.
5587 /* drop refcnt of maps used by the rejected program */
5588 static void release_maps(struct bpf_verifier_env *env)
5590 enum bpf_cgroup_storage_type stype;
5593 for_each_cgroup_storage_type(stype) {
5594 if (!env->prog->aux->cgroup_storage[stype])
5596 bpf_cgroup_storage_release(env->prog,
5597 env->prog->aux->cgroup_storage[stype]);
5600 for (i = 0; i < env->used_map_cnt; i++)
5601 bpf_map_put(env->used_maps[i]);
5604 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
5605 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
5607 struct bpf_insn *insn = env->prog->insnsi;
5608 int insn_cnt = env->prog->len;
5611 for (i = 0; i < insn_cnt; i++, insn++)
5612 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
5616 /* single env->prog->insni[off] instruction was replaced with the range
5617 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
5618 * [0, off) and [off, end) to new locations, so the patched range stays zero
5620 static int adjust_insn_aux_data(struct bpf_verifier_env *env, u32 prog_len,
5623 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
5628 new_data = vzalloc(array_size(prog_len,
5629 sizeof(struct bpf_insn_aux_data)));
5632 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
5633 memcpy(new_data + off + cnt - 1, old_data + off,
5634 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
5635 for (i = off; i < off + cnt - 1; i++)
5636 new_data[i].seen = true;
5637 env->insn_aux_data = new_data;
5642 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
5648 /* NOTE: fake 'exit' subprog should be updated as well. */
5649 for (i = 0; i <= env->subprog_cnt; i++) {
5650 if (env->subprog_info[i].start < off)
5652 env->subprog_info[i].start += len - 1;
5656 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
5657 const struct bpf_insn *patch, u32 len)
5659 struct bpf_prog *new_prog;
5661 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
5664 if (adjust_insn_aux_data(env, new_prog->len, off, len))
5666 adjust_subprog_starts(env, off, len);
5670 /* The verifier does more data flow analysis than llvm and will not
5671 * explore branches that are dead at run time. Malicious programs can
5672 * have dead code too. Therefore replace all dead at-run-time code
5675 * Just nops are not optimal, e.g. if they would sit at the end of the
5676 * program and through another bug we would manage to jump there, then
5677 * we'd execute beyond program memory otherwise. Returning exception
5678 * code also wouldn't work since we can have subprogs where the dead
5679 * code could be located.
5681 static void sanitize_dead_code(struct bpf_verifier_env *env)
5683 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
5684 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
5685 struct bpf_insn *insn = env->prog->insnsi;
5686 const int insn_cnt = env->prog->len;
5689 for (i = 0; i < insn_cnt; i++) {
5690 if (aux_data[i].seen)
5692 memcpy(insn + i, &trap, sizeof(trap));
5696 /* convert load instructions that access fields of a context type into a
5697 * sequence of instructions that access fields of the underlying structure:
5698 * struct __sk_buff -> struct sk_buff
5699 * struct bpf_sock_ops -> struct sock
5701 static int convert_ctx_accesses(struct bpf_verifier_env *env)
5703 const struct bpf_verifier_ops *ops = env->ops;
5704 int i, cnt, size, ctx_field_size, delta = 0;
5705 const int insn_cnt = env->prog->len;
5706 struct bpf_insn insn_buf[16], *insn;
5707 struct bpf_prog *new_prog;
5708 enum bpf_access_type type;
5709 bool is_narrower_load;
5712 if (ops->gen_prologue || env->seen_direct_write) {
5713 if (!ops->gen_prologue) {
5714 verbose(env, "bpf verifier is misconfigured\n");
5717 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
5719 if (cnt >= ARRAY_SIZE(insn_buf)) {
5720 verbose(env, "bpf verifier is misconfigured\n");
5723 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
5727 env->prog = new_prog;
5732 if (bpf_prog_is_dev_bound(env->prog->aux))
5735 insn = env->prog->insnsi + delta;
5737 for (i = 0; i < insn_cnt; i++, insn++) {
5738 bpf_convert_ctx_access_t convert_ctx_access;
5740 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
5741 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
5742 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
5743 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
5745 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
5746 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
5747 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
5748 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
5753 if (type == BPF_WRITE &&
5754 env->insn_aux_data[i + delta].sanitize_stack_off) {
5755 struct bpf_insn patch[] = {
5756 /* Sanitize suspicious stack slot with zero.
5757 * There are no memory dependencies for this store,
5758 * since it's only using frame pointer and immediate
5761 BPF_ST_MEM(BPF_DW, BPF_REG_FP,
5762 env->insn_aux_data[i + delta].sanitize_stack_off,
5764 /* the original STX instruction will immediately
5765 * overwrite the same stack slot with appropriate value
5770 cnt = ARRAY_SIZE(patch);
5771 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
5776 env->prog = new_prog;
5777 insn = new_prog->insnsi + i + delta;
5781 switch (env->insn_aux_data[i + delta].ptr_type) {
5783 if (!ops->convert_ctx_access)
5785 convert_ctx_access = ops->convert_ctx_access;
5788 convert_ctx_access = bpf_sock_convert_ctx_access;
5794 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
5795 size = BPF_LDST_BYTES(insn);
5797 /* If the read access is a narrower load of the field,
5798 * convert to a 4/8-byte load, to minimum program type specific
5799 * convert_ctx_access changes. If conversion is successful,
5800 * we will apply proper mask to the result.
5802 is_narrower_load = size < ctx_field_size;
5803 if (is_narrower_load) {
5804 u32 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
5805 u32 off = insn->off;
5808 if (type == BPF_WRITE) {
5809 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
5814 if (ctx_field_size == 4)
5816 else if (ctx_field_size == 8)
5819 insn->off = off & ~(size_default - 1);
5820 insn->code = BPF_LDX | BPF_MEM | size_code;
5824 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
5826 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
5827 (ctx_field_size && !target_size)) {
5828 verbose(env, "bpf verifier is misconfigured\n");
5832 if (is_narrower_load && size < target_size) {
5833 if (ctx_field_size <= 4)
5834 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
5835 (1 << size * 8) - 1);
5837 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
5838 (1 << size * 8) - 1);
5841 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
5847 /* keep walking new program and skip insns we just inserted */
5848 env->prog = new_prog;
5849 insn = new_prog->insnsi + i + delta;
5855 static int jit_subprogs(struct bpf_verifier_env *env)
5857 struct bpf_prog *prog = env->prog, **func, *tmp;
5858 int i, j, subprog_start, subprog_end = 0, len, subprog;
5859 struct bpf_insn *insn;
5863 if (env->subprog_cnt <= 1)
5866 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
5867 if (insn->code != (BPF_JMP | BPF_CALL) ||
5868 insn->src_reg != BPF_PSEUDO_CALL)
5870 /* Upon error here we cannot fall back to interpreter but
5871 * need a hard reject of the program. Thus -EFAULT is
5872 * propagated in any case.
5874 subprog = find_subprog(env, i + insn->imm + 1);
5876 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
5880 /* temporarily remember subprog id inside insn instead of
5881 * aux_data, since next loop will split up all insns into funcs
5883 insn->off = subprog;
5884 /* remember original imm in case JIT fails and fallback
5885 * to interpreter will be needed
5887 env->insn_aux_data[i].call_imm = insn->imm;
5888 /* point imm to __bpf_call_base+1 from JITs point of view */
5892 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
5896 for (i = 0; i < env->subprog_cnt; i++) {
5897 subprog_start = subprog_end;
5898 subprog_end = env->subprog_info[i + 1].start;
5900 len = subprog_end - subprog_start;
5901 func[i] = bpf_prog_alloc(bpf_prog_size(len), GFP_USER);
5904 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
5905 len * sizeof(struct bpf_insn));
5906 func[i]->type = prog->type;
5908 if (bpf_prog_calc_tag(func[i]))
5910 func[i]->is_func = 1;
5911 /* Use bpf_prog_F_tag to indicate functions in stack traces.
5912 * Long term would need debug info to populate names
5914 func[i]->aux->name[0] = 'F';
5915 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
5916 func[i]->jit_requested = 1;
5917 func[i] = bpf_int_jit_compile(func[i]);
5918 if (!func[i]->jited) {
5924 /* at this point all bpf functions were successfully JITed
5925 * now populate all bpf_calls with correct addresses and
5926 * run last pass of JIT
5928 for (i = 0; i < env->subprog_cnt; i++) {
5929 insn = func[i]->insnsi;
5930 for (j = 0; j < func[i]->len; j++, insn++) {
5931 if (insn->code != (BPF_JMP | BPF_CALL) ||
5932 insn->src_reg != BPF_PSEUDO_CALL)
5934 subprog = insn->off;
5935 insn->imm = (u64 (*)(u64, u64, u64, u64, u64))
5936 func[subprog]->bpf_func -
5940 /* we use the aux data to keep a list of the start addresses
5941 * of the JITed images for each function in the program
5943 * for some architectures, such as powerpc64, the imm field
5944 * might not be large enough to hold the offset of the start
5945 * address of the callee's JITed image from __bpf_call_base
5947 * in such cases, we can lookup the start address of a callee
5948 * by using its subprog id, available from the off field of
5949 * the call instruction, as an index for this list
5951 func[i]->aux->func = func;
5952 func[i]->aux->func_cnt = env->subprog_cnt;
5954 for (i = 0; i < env->subprog_cnt; i++) {
5955 old_bpf_func = func[i]->bpf_func;
5956 tmp = bpf_int_jit_compile(func[i]);
5957 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
5958 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
5965 /* finally lock prog and jit images for all functions and
5968 for (i = 0; i < env->subprog_cnt; i++) {
5969 bpf_prog_lock_ro(func[i]);
5970 bpf_prog_kallsyms_add(func[i]);
5973 /* Last step: make now unused interpreter insns from main
5974 * prog consistent for later dump requests, so they can
5975 * later look the same as if they were interpreted only.
5977 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
5978 if (insn->code != (BPF_JMP | BPF_CALL) ||
5979 insn->src_reg != BPF_PSEUDO_CALL)
5981 insn->off = env->insn_aux_data[i].call_imm;
5982 subprog = find_subprog(env, i + insn->off + 1);
5983 insn->imm = subprog;
5987 prog->bpf_func = func[0]->bpf_func;
5988 prog->aux->func = func;
5989 prog->aux->func_cnt = env->subprog_cnt;
5992 for (i = 0; i < env->subprog_cnt; i++)
5994 bpf_jit_free(func[i]);
5997 /* cleanup main prog to be interpreted */
5998 prog->jit_requested = 0;
5999 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
6000 if (insn->code != (BPF_JMP | BPF_CALL) ||
6001 insn->src_reg != BPF_PSEUDO_CALL)
6004 insn->imm = env->insn_aux_data[i].call_imm;
6009 static int fixup_call_args(struct bpf_verifier_env *env)
6011 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
6012 struct bpf_prog *prog = env->prog;
6013 struct bpf_insn *insn = prog->insnsi;
6018 if (env->prog->jit_requested &&
6019 !bpf_prog_is_dev_bound(env->prog->aux)) {
6020 err = jit_subprogs(env);
6026 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
6027 for (i = 0; i < prog->len; i++, insn++) {
6028 if (insn->code != (BPF_JMP | BPF_CALL) ||
6029 insn->src_reg != BPF_PSEUDO_CALL)
6031 depth = get_callee_stack_depth(env, insn, i);
6034 bpf_patch_call_args(insn, depth);
6041 /* fixup insn->imm field of bpf_call instructions
6042 * and inline eligible helpers as explicit sequence of BPF instructions
6044 * this function is called after eBPF program passed verification
6046 static int fixup_bpf_calls(struct bpf_verifier_env *env)
6048 struct bpf_prog *prog = env->prog;
6049 struct bpf_insn *insn = prog->insnsi;
6050 const struct bpf_func_proto *fn;
6051 const int insn_cnt = prog->len;
6052 const struct bpf_map_ops *ops;
6053 struct bpf_insn_aux_data *aux;
6054 struct bpf_insn insn_buf[16];
6055 struct bpf_prog *new_prog;
6056 struct bpf_map *map_ptr;
6057 int i, cnt, delta = 0;
6059 for (i = 0; i < insn_cnt; i++, insn++) {
6060 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
6061 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
6062 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
6063 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
6064 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
6065 struct bpf_insn mask_and_div[] = {
6066 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
6068 BPF_JMP_IMM(BPF_JNE, insn->src_reg, 0, 2),
6069 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
6070 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
6073 struct bpf_insn mask_and_mod[] = {
6074 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
6075 /* Rx mod 0 -> Rx */
6076 BPF_JMP_IMM(BPF_JEQ, insn->src_reg, 0, 1),
6079 struct bpf_insn *patchlet;
6081 if (insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
6082 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
6083 patchlet = mask_and_div + (is64 ? 1 : 0);
6084 cnt = ARRAY_SIZE(mask_and_div) - (is64 ? 1 : 0);
6086 patchlet = mask_and_mod + (is64 ? 1 : 0);
6087 cnt = ARRAY_SIZE(mask_and_mod) - (is64 ? 1 : 0);
6090 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
6095 env->prog = prog = new_prog;
6096 insn = new_prog->insnsi + i + delta;
6100 if (BPF_CLASS(insn->code) == BPF_LD &&
6101 (BPF_MODE(insn->code) == BPF_ABS ||
6102 BPF_MODE(insn->code) == BPF_IND)) {
6103 cnt = env->ops->gen_ld_abs(insn, insn_buf);
6104 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
6105 verbose(env, "bpf verifier is misconfigured\n");
6109 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
6114 env->prog = prog = new_prog;
6115 insn = new_prog->insnsi + i + delta;
6119 if (insn->code != (BPF_JMP | BPF_CALL))
6121 if (insn->src_reg == BPF_PSEUDO_CALL)
6124 if (insn->imm == BPF_FUNC_get_route_realm)
6125 prog->dst_needed = 1;
6126 if (insn->imm == BPF_FUNC_get_prandom_u32)
6127 bpf_user_rnd_init_once();
6128 if (insn->imm == BPF_FUNC_override_return)
6129 prog->kprobe_override = 1;
6130 if (insn->imm == BPF_FUNC_tail_call) {
6131 /* If we tail call into other programs, we
6132 * cannot make any assumptions since they can
6133 * be replaced dynamically during runtime in
6134 * the program array.
6136 prog->cb_access = 1;
6137 env->prog->aux->stack_depth = MAX_BPF_STACK;
6139 /* mark bpf_tail_call as different opcode to avoid
6140 * conditional branch in the interpeter for every normal
6141 * call and to prevent accidental JITing by JIT compiler
6142 * that doesn't support bpf_tail_call yet
6145 insn->code = BPF_JMP | BPF_TAIL_CALL;
6147 aux = &env->insn_aux_data[i + delta];
6148 if (!bpf_map_ptr_unpriv(aux))
6151 /* instead of changing every JIT dealing with tail_call
6152 * emit two extra insns:
6153 * if (index >= max_entries) goto out;
6154 * index &= array->index_mask;
6155 * to avoid out-of-bounds cpu speculation
6157 if (bpf_map_ptr_poisoned(aux)) {
6158 verbose(env, "tail_call abusing map_ptr\n");
6162 map_ptr = BPF_MAP_PTR(aux->map_state);
6163 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
6164 map_ptr->max_entries, 2);
6165 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
6166 container_of(map_ptr,
6169 insn_buf[2] = *insn;
6171 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
6176 env->prog = prog = new_prog;
6177 insn = new_prog->insnsi + i + delta;
6181 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
6182 * and other inlining handlers are currently limited to 64 bit
6185 if (prog->jit_requested && BITS_PER_LONG == 64 &&
6186 (insn->imm == BPF_FUNC_map_lookup_elem ||
6187 insn->imm == BPF_FUNC_map_update_elem ||
6188 insn->imm == BPF_FUNC_map_delete_elem ||
6189 insn->imm == BPF_FUNC_map_push_elem ||
6190 insn->imm == BPF_FUNC_map_pop_elem ||
6191 insn->imm == BPF_FUNC_map_peek_elem)) {
6192 aux = &env->insn_aux_data[i + delta];
6193 if (bpf_map_ptr_poisoned(aux))
6194 goto patch_call_imm;
6196 map_ptr = BPF_MAP_PTR(aux->map_state);
6198 if (insn->imm == BPF_FUNC_map_lookup_elem &&
6199 ops->map_gen_lookup) {
6200 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
6201 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
6202 verbose(env, "bpf verifier is misconfigured\n");
6206 new_prog = bpf_patch_insn_data(env, i + delta,
6212 env->prog = prog = new_prog;
6213 insn = new_prog->insnsi + i + delta;
6217 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
6218 (void *(*)(struct bpf_map *map, void *key))NULL));
6219 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
6220 (int (*)(struct bpf_map *map, void *key))NULL));
6221 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
6222 (int (*)(struct bpf_map *map, void *key, void *value,
6224 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
6225 (int (*)(struct bpf_map *map, void *value,
6227 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
6228 (int (*)(struct bpf_map *map, void *value))NULL));
6229 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
6230 (int (*)(struct bpf_map *map, void *value))NULL));
6232 switch (insn->imm) {
6233 case BPF_FUNC_map_lookup_elem:
6234 insn->imm = BPF_CAST_CALL(ops->map_lookup_elem) -
6237 case BPF_FUNC_map_update_elem:
6238 insn->imm = BPF_CAST_CALL(ops->map_update_elem) -
6241 case BPF_FUNC_map_delete_elem:
6242 insn->imm = BPF_CAST_CALL(ops->map_delete_elem) -
6245 case BPF_FUNC_map_push_elem:
6246 insn->imm = BPF_CAST_CALL(ops->map_push_elem) -
6249 case BPF_FUNC_map_pop_elem:
6250 insn->imm = BPF_CAST_CALL(ops->map_pop_elem) -
6253 case BPF_FUNC_map_peek_elem:
6254 insn->imm = BPF_CAST_CALL(ops->map_peek_elem) -
6259 goto patch_call_imm;
6263 fn = env->ops->get_func_proto(insn->imm, env->prog);
6264 /* all functions that have prototype and verifier allowed
6265 * programs to call them, must be real in-kernel functions
6269 "kernel subsystem misconfigured func %s#%d\n",
6270 func_id_name(insn->imm), insn->imm);
6273 insn->imm = fn->func - __bpf_call_base;
6279 static void free_states(struct bpf_verifier_env *env)
6281 struct bpf_verifier_state_list *sl, *sln;
6284 if (!env->explored_states)
6287 for (i = 0; i < env->prog->len; i++) {
6288 sl = env->explored_states[i];
6291 while (sl != STATE_LIST_MARK) {
6293 free_verifier_state(&sl->state, false);
6299 kfree(env->explored_states);
6302 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr)
6304 struct bpf_verifier_env *env;
6305 struct bpf_verifier_log *log;
6308 /* no program is valid */
6309 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
6312 /* 'struct bpf_verifier_env' can be global, but since it's not small,
6313 * allocate/free it every time bpf_check() is called
6315 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
6320 env->insn_aux_data =
6321 vzalloc(array_size(sizeof(struct bpf_insn_aux_data),
6324 if (!env->insn_aux_data)
6327 env->ops = bpf_verifier_ops[env->prog->type];
6329 /* grab the mutex to protect few globals used by verifier */
6330 mutex_lock(&bpf_verifier_lock);
6332 if (attr->log_level || attr->log_buf || attr->log_size) {
6333 /* user requested verbose verifier output
6334 * and supplied buffer to store the verification trace
6336 log->level = attr->log_level;
6337 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
6338 log->len_total = attr->log_size;
6341 /* log attributes have to be sane */
6342 if (log->len_total < 128 || log->len_total > UINT_MAX >> 8 ||
6343 !log->level || !log->ubuf)
6347 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
6348 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
6349 env->strict_alignment = true;
6351 ret = replace_map_fd_with_map_ptr(env);
6353 goto skip_full_check;
6355 if (bpf_prog_is_dev_bound(env->prog->aux)) {
6356 ret = bpf_prog_offload_verifier_prep(env);
6358 goto skip_full_check;
6361 env->explored_states = kcalloc(env->prog->len,
6362 sizeof(struct bpf_verifier_state_list *),
6365 if (!env->explored_states)
6366 goto skip_full_check;
6368 env->allow_ptr_leaks = capable(CAP_SYS_ADMIN);
6370 ret = check_cfg(env);
6372 goto skip_full_check;
6374 ret = do_check(env);
6375 if (env->cur_state) {
6376 free_verifier_state(env->cur_state, true);
6377 env->cur_state = NULL;
6380 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
6381 ret = bpf_prog_offload_finalize(env);
6384 while (!pop_stack(env, NULL, NULL));
6388 sanitize_dead_code(env);
6391 ret = check_max_stack_depth(env);
6394 /* program is valid, convert *(u32*)(ctx + off) accesses */
6395 ret = convert_ctx_accesses(env);
6398 ret = fixup_bpf_calls(env);
6401 ret = fixup_call_args(env);
6403 if (log->level && bpf_verifier_log_full(log))
6405 if (log->level && !log->ubuf) {
6407 goto err_release_maps;
6410 if (ret == 0 && env->used_map_cnt) {
6411 /* if program passed verifier, update used_maps in bpf_prog_info */
6412 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
6413 sizeof(env->used_maps[0]),
6416 if (!env->prog->aux->used_maps) {
6418 goto err_release_maps;
6421 memcpy(env->prog->aux->used_maps, env->used_maps,
6422 sizeof(env->used_maps[0]) * env->used_map_cnt);
6423 env->prog->aux->used_map_cnt = env->used_map_cnt;
6425 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
6426 * bpf_ld_imm64 instructions
6428 convert_pseudo_ld_imm64(env);
6432 if (!env->prog->aux->used_maps)
6433 /* if we didn't copy map pointers into bpf_prog_info, release
6434 * them now. Otherwise free_used_maps() will release them.
6439 mutex_unlock(&bpf_verifier_lock);
6440 vfree(env->insn_aux_data);