1 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
2 * Copyright (c) 2016 Facebook
3 * Copyright (c) 2018 Covalent IO, Inc. http://covalent.io
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of version 2 of the GNU General Public
7 * License as published by the Free Software Foundation.
9 * This program is distributed in the hope that it will be useful, but
10 * WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
12 * General Public License for more details.
14 #include <linux/kernel.h>
15 #include <linux/types.h>
16 #include <linux/slab.h>
17 #include <linux/bpf.h>
18 #include <linux/bpf_verifier.h>
19 #include <linux/filter.h>
20 #include <net/netlink.h>
21 #include <linux/file.h>
22 #include <linux/vmalloc.h>
23 #include <linux/stringify.h>
24 #include <linux/bsearch.h>
25 #include <linux/sort.h>
26 #include <linux/perf_event.h>
30 static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
31 #define BPF_PROG_TYPE(_id, _name) \
32 [_id] = & _name ## _verifier_ops,
33 #define BPF_MAP_TYPE(_id, _ops)
34 #include <linux/bpf_types.h>
39 /* bpf_check() is a static code analyzer that walks eBPF program
40 * instruction by instruction and updates register/stack state.
41 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
43 * The first pass is depth-first-search to check that the program is a DAG.
44 * It rejects the following programs:
45 * - larger than BPF_MAXINSNS insns
46 * - if loop is present (detected via back-edge)
47 * - unreachable insns exist (shouldn't be a forest. program = one function)
48 * - out of bounds or malformed jumps
49 * The second pass is all possible path descent from the 1st insn.
50 * Since it's analyzing all pathes through the program, the length of the
51 * analysis is limited to 64k insn, which may be hit even if total number of
52 * insn is less then 4K, but there are too many branches that change stack/regs.
53 * Number of 'branches to be analyzed' is limited to 1k
55 * On entry to each instruction, each register has a type, and the instruction
56 * changes the types of the registers depending on instruction semantics.
57 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
60 * All registers are 64-bit.
61 * R0 - return register
62 * R1-R5 argument passing registers
63 * R6-R9 callee saved registers
64 * R10 - frame pointer read-only
66 * At the start of BPF program the register R1 contains a pointer to bpf_context
67 * and has type PTR_TO_CTX.
69 * Verifier tracks arithmetic operations on pointers in case:
70 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
71 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
72 * 1st insn copies R10 (which has FRAME_PTR) type into R1
73 * and 2nd arithmetic instruction is pattern matched to recognize
74 * that it wants to construct a pointer to some element within stack.
75 * So after 2nd insn, the register R1 has type PTR_TO_STACK
76 * (and -20 constant is saved for further stack bounds checking).
77 * Meaning that this reg is a pointer to stack plus known immediate constant.
79 * Most of the time the registers have SCALAR_VALUE type, which
80 * means the register has some value, but it's not a valid pointer.
81 * (like pointer plus pointer becomes SCALAR_VALUE type)
83 * When verifier sees load or store instructions the type of base register
84 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK, PTR_TO_SOCKET. These are
85 * four pointer types recognized by check_mem_access() function.
87 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
88 * and the range of [ptr, ptr + map's value_size) is accessible.
90 * registers used to pass values to function calls are checked against
91 * function argument constraints.
93 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
94 * It means that the register type passed to this function must be
95 * PTR_TO_STACK and it will be used inside the function as
96 * 'pointer to map element key'
98 * For example the argument constraints for bpf_map_lookup_elem():
99 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
100 * .arg1_type = ARG_CONST_MAP_PTR,
101 * .arg2_type = ARG_PTR_TO_MAP_KEY,
103 * ret_type says that this function returns 'pointer to map elem value or null'
104 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
105 * 2nd argument should be a pointer to stack, which will be used inside
106 * the helper function as a pointer to map element key.
108 * On the kernel side the helper function looks like:
109 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
111 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
112 * void *key = (void *) (unsigned long) r2;
115 * here kernel can access 'key' and 'map' pointers safely, knowing that
116 * [key, key + map->key_size) bytes are valid and were initialized on
117 * the stack of eBPF program.
120 * Corresponding eBPF program may look like:
121 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
122 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
123 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
124 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
125 * here verifier looks at prototype of map_lookup_elem() and sees:
126 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
127 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
129 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
130 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
131 * and were initialized prior to this call.
132 * If it's ok, then verifier allows this BPF_CALL insn and looks at
133 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
134 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
135 * returns ether pointer to map value or NULL.
137 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
138 * insn, the register holding that pointer in the true branch changes state to
139 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
140 * branch. See check_cond_jmp_op().
142 * After the call R0 is set to return type of the function and registers R1-R5
143 * are set to NOT_INIT to indicate that they are no longer readable.
145 * The following reference types represent a potential reference to a kernel
146 * resource which, after first being allocated, must be checked and freed by
148 * - PTR_TO_SOCKET_OR_NULL, PTR_TO_SOCKET
150 * When the verifier sees a helper call return a reference type, it allocates a
151 * pointer id for the reference and stores it in the current function state.
152 * Similar to the way that PTR_TO_MAP_VALUE_OR_NULL is converted into
153 * PTR_TO_MAP_VALUE, PTR_TO_SOCKET_OR_NULL becomes PTR_TO_SOCKET when the type
154 * passes through a NULL-check conditional. For the branch wherein the state is
155 * changed to CONST_IMM, the verifier releases the reference.
157 * For each helper function that allocates a reference, such as
158 * bpf_sk_lookup_tcp(), there is a corresponding release function, such as
159 * bpf_sk_release(). When a reference type passes into the release function,
160 * the verifier also releases the reference. If any unchecked or unreleased
161 * reference remains at the end of the program, the verifier rejects it.
164 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
165 struct bpf_verifier_stack_elem {
166 /* verifer state is 'st'
167 * before processing instruction 'insn_idx'
168 * and after processing instruction 'prev_insn_idx'
170 struct bpf_verifier_state st;
173 struct bpf_verifier_stack_elem *next;
176 #define BPF_COMPLEXITY_LIMIT_INSNS 131072
177 #define BPF_COMPLEXITY_LIMIT_STACK 1024
179 #define BPF_MAP_PTR_UNPRIV 1UL
180 #define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \
181 POISON_POINTER_DELTA))
182 #define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
184 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
186 return BPF_MAP_PTR(aux->map_state) == BPF_MAP_PTR_POISON;
189 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
191 return aux->map_state & BPF_MAP_PTR_UNPRIV;
194 static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
195 const struct bpf_map *map, bool unpriv)
197 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON & BPF_MAP_PTR_UNPRIV);
198 unpriv |= bpf_map_ptr_unpriv(aux);
199 aux->map_state = (unsigned long)map |
200 (unpriv ? BPF_MAP_PTR_UNPRIV : 0UL);
203 struct bpf_call_arg_meta {
204 struct bpf_map *map_ptr;
209 s64 msize_smax_value;
210 u64 msize_umax_value;
214 static DEFINE_MUTEX(bpf_verifier_lock);
216 void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
221 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
223 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
224 "verifier log line truncated - local buffer too short\n");
226 n = min(log->len_total - log->len_used - 1, n);
229 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
235 /* log_level controls verbosity level of eBPF verifier.
236 * bpf_verifier_log_write() is used to dump the verification trace to the log,
237 * so the user can figure out what's wrong with the program
239 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
240 const char *fmt, ...)
244 if (!bpf_verifier_log_needed(&env->log))
248 bpf_verifier_vlog(&env->log, fmt, args);
251 EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
253 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
255 struct bpf_verifier_env *env = private_data;
258 if (!bpf_verifier_log_needed(&env->log))
262 bpf_verifier_vlog(&env->log, fmt, args);
266 static bool type_is_pkt_pointer(enum bpf_reg_type type)
268 return type == PTR_TO_PACKET ||
269 type == PTR_TO_PACKET_META;
272 static bool reg_type_may_be_null(enum bpf_reg_type type)
274 return type == PTR_TO_MAP_VALUE_OR_NULL ||
275 type == PTR_TO_SOCKET_OR_NULL;
278 static bool type_is_refcounted(enum bpf_reg_type type)
280 return type == PTR_TO_SOCKET;
283 static bool type_is_refcounted_or_null(enum bpf_reg_type type)
285 return type == PTR_TO_SOCKET || type == PTR_TO_SOCKET_OR_NULL;
288 static bool reg_is_refcounted(const struct bpf_reg_state *reg)
290 return type_is_refcounted(reg->type);
293 static bool reg_is_refcounted_or_null(const struct bpf_reg_state *reg)
295 return type_is_refcounted_or_null(reg->type);
298 static bool arg_type_is_refcounted(enum bpf_arg_type type)
300 return type == ARG_PTR_TO_SOCKET;
303 /* Determine whether the function releases some resources allocated by another
304 * function call. The first reference type argument will be assumed to be
305 * released by release_reference().
307 static bool is_release_function(enum bpf_func_id func_id)
309 return func_id == BPF_FUNC_sk_release;
312 /* string representation of 'enum bpf_reg_type' */
313 static const char * const reg_type_str[] = {
315 [SCALAR_VALUE] = "inv",
316 [PTR_TO_CTX] = "ctx",
317 [CONST_PTR_TO_MAP] = "map_ptr",
318 [PTR_TO_MAP_VALUE] = "map_value",
319 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
320 [PTR_TO_STACK] = "fp",
321 [PTR_TO_PACKET] = "pkt",
322 [PTR_TO_PACKET_META] = "pkt_meta",
323 [PTR_TO_PACKET_END] = "pkt_end",
324 [PTR_TO_FLOW_KEYS] = "flow_keys",
325 [PTR_TO_SOCKET] = "sock",
326 [PTR_TO_SOCKET_OR_NULL] = "sock_or_null",
329 static char slot_type_char[] = {
330 [STACK_INVALID] = '?',
336 static void print_liveness(struct bpf_verifier_env *env,
337 enum bpf_reg_liveness live)
339 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN))
341 if (live & REG_LIVE_READ)
343 if (live & REG_LIVE_WRITTEN)
347 static struct bpf_func_state *func(struct bpf_verifier_env *env,
348 const struct bpf_reg_state *reg)
350 struct bpf_verifier_state *cur = env->cur_state;
352 return cur->frame[reg->frameno];
355 static void print_verifier_state(struct bpf_verifier_env *env,
356 const struct bpf_func_state *state)
358 const struct bpf_reg_state *reg;
363 verbose(env, " frame%d:", state->frameno);
364 for (i = 0; i < MAX_BPF_REG; i++) {
365 reg = &state->regs[i];
369 verbose(env, " R%d", i);
370 print_liveness(env, reg->live);
371 verbose(env, "=%s", reg_type_str[t]);
372 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
373 tnum_is_const(reg->var_off)) {
374 /* reg->off should be 0 for SCALAR_VALUE */
375 verbose(env, "%lld", reg->var_off.value + reg->off);
376 if (t == PTR_TO_STACK)
377 verbose(env, ",call_%d", func(env, reg)->callsite);
379 verbose(env, "(id=%d", reg->id);
380 if (t != SCALAR_VALUE)
381 verbose(env, ",off=%d", reg->off);
382 if (type_is_pkt_pointer(t))
383 verbose(env, ",r=%d", reg->range);
384 else if (t == CONST_PTR_TO_MAP ||
385 t == PTR_TO_MAP_VALUE ||
386 t == PTR_TO_MAP_VALUE_OR_NULL)
387 verbose(env, ",ks=%d,vs=%d",
388 reg->map_ptr->key_size,
389 reg->map_ptr->value_size);
390 if (tnum_is_const(reg->var_off)) {
391 /* Typically an immediate SCALAR_VALUE, but
392 * could be a pointer whose offset is too big
395 verbose(env, ",imm=%llx", reg->var_off.value);
397 if (reg->smin_value != reg->umin_value &&
398 reg->smin_value != S64_MIN)
399 verbose(env, ",smin_value=%lld",
400 (long long)reg->smin_value);
401 if (reg->smax_value != reg->umax_value &&
402 reg->smax_value != S64_MAX)
403 verbose(env, ",smax_value=%lld",
404 (long long)reg->smax_value);
405 if (reg->umin_value != 0)
406 verbose(env, ",umin_value=%llu",
407 (unsigned long long)reg->umin_value);
408 if (reg->umax_value != U64_MAX)
409 verbose(env, ",umax_value=%llu",
410 (unsigned long long)reg->umax_value);
411 if (!tnum_is_unknown(reg->var_off)) {
414 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
415 verbose(env, ",var_off=%s", tn_buf);
421 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
422 char types_buf[BPF_REG_SIZE + 1];
426 for (j = 0; j < BPF_REG_SIZE; j++) {
427 if (state->stack[i].slot_type[j] != STACK_INVALID)
429 types_buf[j] = slot_type_char[
430 state->stack[i].slot_type[j]];
432 types_buf[BPF_REG_SIZE] = 0;
435 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
436 print_liveness(env, state->stack[i].spilled_ptr.live);
437 if (state->stack[i].slot_type[0] == STACK_SPILL)
439 reg_type_str[state->stack[i].spilled_ptr.type]);
441 verbose(env, "=%s", types_buf);
443 if (state->acquired_refs && state->refs[0].id) {
444 verbose(env, " refs=%d", state->refs[0].id);
445 for (i = 1; i < state->acquired_refs; i++)
446 if (state->refs[i].id)
447 verbose(env, ",%d", state->refs[i].id);
452 #define COPY_STATE_FN(NAME, COUNT, FIELD, SIZE) \
453 static int copy_##NAME##_state(struct bpf_func_state *dst, \
454 const struct bpf_func_state *src) \
458 if (WARN_ON_ONCE(dst->COUNT < src->COUNT)) { \
459 /* internal bug, make state invalid to reject the program */ \
460 memset(dst, 0, sizeof(*dst)); \
463 memcpy(dst->FIELD, src->FIELD, \
464 sizeof(*src->FIELD) * (src->COUNT / SIZE)); \
467 /* copy_reference_state() */
468 COPY_STATE_FN(reference, acquired_refs, refs, 1)
469 /* copy_stack_state() */
470 COPY_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
473 #define REALLOC_STATE_FN(NAME, COUNT, FIELD, SIZE) \
474 static int realloc_##NAME##_state(struct bpf_func_state *state, int size, \
477 u32 old_size = state->COUNT; \
478 struct bpf_##NAME##_state *new_##FIELD; \
479 int slot = size / SIZE; \
481 if (size <= old_size || !size) { \
484 state->COUNT = slot * SIZE; \
485 if (!size && old_size) { \
486 kfree(state->FIELD); \
487 state->FIELD = NULL; \
491 new_##FIELD = kmalloc_array(slot, sizeof(struct bpf_##NAME##_state), \
497 memcpy(new_##FIELD, state->FIELD, \
498 sizeof(*new_##FIELD) * (old_size / SIZE)); \
499 memset(new_##FIELD + old_size / SIZE, 0, \
500 sizeof(*new_##FIELD) * (size - old_size) / SIZE); \
502 state->COUNT = slot * SIZE; \
503 kfree(state->FIELD); \
504 state->FIELD = new_##FIELD; \
507 /* realloc_reference_state() */
508 REALLOC_STATE_FN(reference, acquired_refs, refs, 1)
509 /* realloc_stack_state() */
510 REALLOC_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
511 #undef REALLOC_STATE_FN
513 /* do_check() starts with zero-sized stack in struct bpf_verifier_state to
514 * make it consume minimal amount of memory. check_stack_write() access from
515 * the program calls into realloc_func_state() to grow the stack size.
516 * Note there is a non-zero 'parent' pointer inside bpf_verifier_state
517 * which realloc_stack_state() copies over. It points to previous
518 * bpf_verifier_state which is never reallocated.
520 static int realloc_func_state(struct bpf_func_state *state, int stack_size,
521 int refs_size, bool copy_old)
523 int err = realloc_reference_state(state, refs_size, copy_old);
526 return realloc_stack_state(state, stack_size, copy_old);
529 /* Acquire a pointer id from the env and update the state->refs to include
530 * this new pointer reference.
531 * On success, returns a valid pointer id to associate with the register
532 * On failure, returns a negative errno.
534 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
536 struct bpf_func_state *state = cur_func(env);
537 int new_ofs = state->acquired_refs;
540 err = realloc_reference_state(state, state->acquired_refs + 1, true);
544 state->refs[new_ofs].id = id;
545 state->refs[new_ofs].insn_idx = insn_idx;
550 /* release function corresponding to acquire_reference_state(). Idempotent. */
551 static int __release_reference_state(struct bpf_func_state *state, int ptr_id)
558 last_idx = state->acquired_refs - 1;
559 for (i = 0; i < state->acquired_refs; i++) {
560 if (state->refs[i].id == ptr_id) {
561 if (last_idx && i != last_idx)
562 memcpy(&state->refs[i], &state->refs[last_idx],
563 sizeof(*state->refs));
564 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
565 state->acquired_refs--;
572 /* variation on the above for cases where we expect that there must be an
573 * outstanding reference for the specified ptr_id.
575 static int release_reference_state(struct bpf_verifier_env *env, int ptr_id)
577 struct bpf_func_state *state = cur_func(env);
580 err = __release_reference_state(state, ptr_id);
581 if (WARN_ON_ONCE(err != 0))
582 verbose(env, "verifier internal error: can't release reference\n");
586 static int transfer_reference_state(struct bpf_func_state *dst,
587 struct bpf_func_state *src)
589 int err = realloc_reference_state(dst, src->acquired_refs, false);
592 err = copy_reference_state(dst, src);
598 static void free_func_state(struct bpf_func_state *state)
607 static void free_verifier_state(struct bpf_verifier_state *state,
612 for (i = 0; i <= state->curframe; i++) {
613 free_func_state(state->frame[i]);
614 state->frame[i] = NULL;
620 /* copy verifier state from src to dst growing dst stack space
621 * when necessary to accommodate larger src stack
623 static int copy_func_state(struct bpf_func_state *dst,
624 const struct bpf_func_state *src)
628 err = realloc_func_state(dst, src->allocated_stack, src->acquired_refs,
632 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
633 err = copy_reference_state(dst, src);
636 return copy_stack_state(dst, src);
639 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
640 const struct bpf_verifier_state *src)
642 struct bpf_func_state *dst;
645 /* if dst has more stack frames then src frame, free them */
646 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
647 free_func_state(dst_state->frame[i]);
648 dst_state->frame[i] = NULL;
650 dst_state->curframe = src->curframe;
651 for (i = 0; i <= src->curframe; i++) {
652 dst = dst_state->frame[i];
654 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
657 dst_state->frame[i] = dst;
659 err = copy_func_state(dst, src->frame[i]);
666 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
669 struct bpf_verifier_state *cur = env->cur_state;
670 struct bpf_verifier_stack_elem *elem, *head = env->head;
673 if (env->head == NULL)
677 err = copy_verifier_state(cur, &head->st);
682 *insn_idx = head->insn_idx;
684 *prev_insn_idx = head->prev_insn_idx;
686 free_verifier_state(&head->st, false);
693 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
694 int insn_idx, int prev_insn_idx)
696 struct bpf_verifier_state *cur = env->cur_state;
697 struct bpf_verifier_stack_elem *elem;
700 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
704 elem->insn_idx = insn_idx;
705 elem->prev_insn_idx = prev_insn_idx;
706 elem->next = env->head;
709 err = copy_verifier_state(&elem->st, cur);
712 if (env->stack_size > BPF_COMPLEXITY_LIMIT_STACK) {
713 verbose(env, "BPF program is too complex\n");
718 free_verifier_state(env->cur_state, true);
719 env->cur_state = NULL;
720 /* pop all elements and return */
721 while (!pop_stack(env, NULL, NULL));
725 #define CALLER_SAVED_REGS 6
726 static const int caller_saved[CALLER_SAVED_REGS] = {
727 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
730 static void __mark_reg_not_init(struct bpf_reg_state *reg);
732 /* Mark the unknown part of a register (variable offset or scalar value) as
733 * known to have the value @imm.
735 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
737 /* Clear id, off, and union(map_ptr, range) */
738 memset(((u8 *)reg) + sizeof(reg->type), 0,
739 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
740 reg->var_off = tnum_const(imm);
741 reg->smin_value = (s64)imm;
742 reg->smax_value = (s64)imm;
743 reg->umin_value = imm;
744 reg->umax_value = imm;
747 /* Mark the 'variable offset' part of a register as zero. This should be
748 * used only on registers holding a pointer type.
750 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
752 __mark_reg_known(reg, 0);
755 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
757 __mark_reg_known(reg, 0);
758 reg->type = SCALAR_VALUE;
761 static void mark_reg_known_zero(struct bpf_verifier_env *env,
762 struct bpf_reg_state *regs, u32 regno)
764 if (WARN_ON(regno >= MAX_BPF_REG)) {
765 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
766 /* Something bad happened, let's kill all regs */
767 for (regno = 0; regno < MAX_BPF_REG; regno++)
768 __mark_reg_not_init(regs + regno);
771 __mark_reg_known_zero(regs + regno);
774 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
776 return type_is_pkt_pointer(reg->type);
779 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
781 return reg_is_pkt_pointer(reg) ||
782 reg->type == PTR_TO_PACKET_END;
785 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
786 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
787 enum bpf_reg_type which)
789 /* The register can already have a range from prior markings.
790 * This is fine as long as it hasn't been advanced from its
793 return reg->type == which &&
796 tnum_equals_const(reg->var_off, 0);
799 /* Attempts to improve min/max values based on var_off information */
800 static void __update_reg_bounds(struct bpf_reg_state *reg)
802 /* min signed is max(sign bit) | min(other bits) */
803 reg->smin_value = max_t(s64, reg->smin_value,
804 reg->var_off.value | (reg->var_off.mask & S64_MIN));
805 /* max signed is min(sign bit) | max(other bits) */
806 reg->smax_value = min_t(s64, reg->smax_value,
807 reg->var_off.value | (reg->var_off.mask & S64_MAX));
808 reg->umin_value = max(reg->umin_value, reg->var_off.value);
809 reg->umax_value = min(reg->umax_value,
810 reg->var_off.value | reg->var_off.mask);
813 /* Uses signed min/max values to inform unsigned, and vice-versa */
814 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
816 /* Learn sign from signed bounds.
817 * If we cannot cross the sign boundary, then signed and unsigned bounds
818 * are the same, so combine. This works even in the negative case, e.g.
819 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
821 if (reg->smin_value >= 0 || reg->smax_value < 0) {
822 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
824 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
828 /* Learn sign from unsigned bounds. Signed bounds cross the sign
829 * boundary, so we must be careful.
831 if ((s64)reg->umax_value >= 0) {
832 /* Positive. We can't learn anything from the smin, but smax
833 * is positive, hence safe.
835 reg->smin_value = reg->umin_value;
836 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
838 } else if ((s64)reg->umin_value < 0) {
839 /* Negative. We can't learn anything from the smax, but smin
840 * is negative, hence safe.
842 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
844 reg->smax_value = reg->umax_value;
848 /* Attempts to improve var_off based on unsigned min/max information */
849 static void __reg_bound_offset(struct bpf_reg_state *reg)
851 reg->var_off = tnum_intersect(reg->var_off,
852 tnum_range(reg->umin_value,
856 /* Reset the min/max bounds of a register */
857 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
859 reg->smin_value = S64_MIN;
860 reg->smax_value = S64_MAX;
862 reg->umax_value = U64_MAX;
865 /* Mark a register as having a completely unknown (scalar) value. */
866 static void __mark_reg_unknown(struct bpf_reg_state *reg)
869 * Clear type, id, off, and union(map_ptr, range) and
870 * padding between 'type' and union
872 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
873 reg->type = SCALAR_VALUE;
874 reg->var_off = tnum_unknown;
876 __mark_reg_unbounded(reg);
879 static void mark_reg_unknown(struct bpf_verifier_env *env,
880 struct bpf_reg_state *regs, u32 regno)
882 if (WARN_ON(regno >= MAX_BPF_REG)) {
883 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
884 /* Something bad happened, let's kill all regs except FP */
885 for (regno = 0; regno < BPF_REG_FP; regno++)
886 __mark_reg_not_init(regs + regno);
889 __mark_reg_unknown(regs + regno);
892 static void __mark_reg_not_init(struct bpf_reg_state *reg)
894 __mark_reg_unknown(reg);
895 reg->type = NOT_INIT;
898 static void mark_reg_not_init(struct bpf_verifier_env *env,
899 struct bpf_reg_state *regs, u32 regno)
901 if (WARN_ON(regno >= MAX_BPF_REG)) {
902 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
903 /* Something bad happened, let's kill all regs except FP */
904 for (regno = 0; regno < BPF_REG_FP; regno++)
905 __mark_reg_not_init(regs + regno);
908 __mark_reg_not_init(regs + regno);
911 static void init_reg_state(struct bpf_verifier_env *env,
912 struct bpf_func_state *state)
914 struct bpf_reg_state *regs = state->regs;
917 for (i = 0; i < MAX_BPF_REG; i++) {
918 mark_reg_not_init(env, regs, i);
919 regs[i].live = REG_LIVE_NONE;
920 regs[i].parent = NULL;
924 regs[BPF_REG_FP].type = PTR_TO_STACK;
925 mark_reg_known_zero(env, regs, BPF_REG_FP);
926 regs[BPF_REG_FP].frameno = state->frameno;
928 /* 1st arg to a function */
929 regs[BPF_REG_1].type = PTR_TO_CTX;
930 mark_reg_known_zero(env, regs, BPF_REG_1);
933 #define BPF_MAIN_FUNC (-1)
934 static void init_func_state(struct bpf_verifier_env *env,
935 struct bpf_func_state *state,
936 int callsite, int frameno, int subprogno)
938 state->callsite = callsite;
939 state->frameno = frameno;
940 state->subprogno = subprogno;
941 init_reg_state(env, state);
945 SRC_OP, /* register is used as source operand */
946 DST_OP, /* register is used as destination operand */
947 DST_OP_NO_MARK /* same as above, check only, don't mark */
950 static int cmp_subprogs(const void *a, const void *b)
952 return ((struct bpf_subprog_info *)a)->start -
953 ((struct bpf_subprog_info *)b)->start;
956 static int find_subprog(struct bpf_verifier_env *env, int off)
958 struct bpf_subprog_info *p;
960 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
961 sizeof(env->subprog_info[0]), cmp_subprogs);
964 return p - env->subprog_info;
968 static int add_subprog(struct bpf_verifier_env *env, int off)
970 int insn_cnt = env->prog->len;
973 if (off >= insn_cnt || off < 0) {
974 verbose(env, "call to invalid destination\n");
977 ret = find_subprog(env, off);
980 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
981 verbose(env, "too many subprograms\n");
984 env->subprog_info[env->subprog_cnt++].start = off;
985 sort(env->subprog_info, env->subprog_cnt,
986 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
990 static int check_subprogs(struct bpf_verifier_env *env)
992 int i, ret, subprog_start, subprog_end, off, cur_subprog = 0;
993 struct bpf_subprog_info *subprog = env->subprog_info;
994 struct bpf_insn *insn = env->prog->insnsi;
995 int insn_cnt = env->prog->len;
997 /* Add entry function. */
998 ret = add_subprog(env, 0);
1002 /* determine subprog starts. The end is one before the next starts */
1003 for (i = 0; i < insn_cnt; i++) {
1004 if (insn[i].code != (BPF_JMP | BPF_CALL))
1006 if (insn[i].src_reg != BPF_PSEUDO_CALL)
1008 if (!env->allow_ptr_leaks) {
1009 verbose(env, "function calls to other bpf functions are allowed for root only\n");
1012 ret = add_subprog(env, i + insn[i].imm + 1);
1017 /* Add a fake 'exit' subprog which could simplify subprog iteration
1018 * logic. 'subprog_cnt' should not be increased.
1020 subprog[env->subprog_cnt].start = insn_cnt;
1022 if (env->log.level > 1)
1023 for (i = 0; i < env->subprog_cnt; i++)
1024 verbose(env, "func#%d @%d\n", i, subprog[i].start);
1026 /* now check that all jumps are within the same subprog */
1027 subprog_start = subprog[cur_subprog].start;
1028 subprog_end = subprog[cur_subprog + 1].start;
1029 for (i = 0; i < insn_cnt; i++) {
1030 u8 code = insn[i].code;
1032 if (BPF_CLASS(code) != BPF_JMP)
1034 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
1036 off = i + insn[i].off + 1;
1037 if (off < subprog_start || off >= subprog_end) {
1038 verbose(env, "jump out of range from insn %d to %d\n", i, off);
1042 if (i == subprog_end - 1) {
1043 /* to avoid fall-through from one subprog into another
1044 * the last insn of the subprog should be either exit
1045 * or unconditional jump back
1047 if (code != (BPF_JMP | BPF_EXIT) &&
1048 code != (BPF_JMP | BPF_JA)) {
1049 verbose(env, "last insn is not an exit or jmp\n");
1052 subprog_start = subprog_end;
1054 if (cur_subprog < env->subprog_cnt)
1055 subprog_end = subprog[cur_subprog + 1].start;
1061 /* Parentage chain of this register (or stack slot) should take care of all
1062 * issues like callee-saved registers, stack slot allocation time, etc.
1064 static int mark_reg_read(struct bpf_verifier_env *env,
1065 const struct bpf_reg_state *state,
1066 struct bpf_reg_state *parent)
1068 bool writes = parent == state->parent; /* Observe write marks */
1071 /* if read wasn't screened by an earlier write ... */
1072 if (writes && state->live & REG_LIVE_WRITTEN)
1074 /* ... then we depend on parent's value */
1075 parent->live |= REG_LIVE_READ;
1077 parent = state->parent;
1083 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
1084 enum reg_arg_type t)
1086 struct bpf_verifier_state *vstate = env->cur_state;
1087 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1088 struct bpf_reg_state *regs = state->regs;
1090 if (regno >= MAX_BPF_REG) {
1091 verbose(env, "R%d is invalid\n", regno);
1096 /* check whether register used as source operand can be read */
1097 if (regs[regno].type == NOT_INIT) {
1098 verbose(env, "R%d !read_ok\n", regno);
1101 /* We don't need to worry about FP liveness because it's read-only */
1102 if (regno != BPF_REG_FP)
1103 return mark_reg_read(env, ®s[regno],
1104 regs[regno].parent);
1106 /* check whether register used as dest operand can be written to */
1107 if (regno == BPF_REG_FP) {
1108 verbose(env, "frame pointer is read only\n");
1111 regs[regno].live |= REG_LIVE_WRITTEN;
1113 mark_reg_unknown(env, regs, regno);
1118 static bool is_spillable_regtype(enum bpf_reg_type type)
1121 case PTR_TO_MAP_VALUE:
1122 case PTR_TO_MAP_VALUE_OR_NULL:
1126 case PTR_TO_PACKET_META:
1127 case PTR_TO_PACKET_END:
1128 case PTR_TO_FLOW_KEYS:
1129 case CONST_PTR_TO_MAP:
1131 case PTR_TO_SOCKET_OR_NULL:
1138 /* Does this register contain a constant zero? */
1139 static bool register_is_null(struct bpf_reg_state *reg)
1141 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
1144 /* check_stack_read/write functions track spill/fill of registers,
1145 * stack boundary and alignment are checked in check_mem_access()
1147 static int check_stack_write(struct bpf_verifier_env *env,
1148 struct bpf_func_state *state, /* func where register points to */
1149 int off, int size, int value_regno, int insn_idx)
1151 struct bpf_func_state *cur; /* state of the current function */
1152 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
1153 enum bpf_reg_type type;
1155 err = realloc_func_state(state, round_up(slot + 1, BPF_REG_SIZE),
1156 state->acquired_refs, true);
1159 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
1160 * so it's aligned access and [off, off + size) are within stack limits
1162 if (!env->allow_ptr_leaks &&
1163 state->stack[spi].slot_type[0] == STACK_SPILL &&
1164 size != BPF_REG_SIZE) {
1165 verbose(env, "attempt to corrupt spilled pointer on stack\n");
1169 cur = env->cur_state->frame[env->cur_state->curframe];
1170 if (value_regno >= 0 &&
1171 is_spillable_regtype((type = cur->regs[value_regno].type))) {
1173 /* register containing pointer is being spilled into stack */
1174 if (size != BPF_REG_SIZE) {
1175 verbose(env, "invalid size of register spill\n");
1179 if (state != cur && type == PTR_TO_STACK) {
1180 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
1184 /* save register state */
1185 state->stack[spi].spilled_ptr = cur->regs[value_regno];
1186 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1188 for (i = 0; i < BPF_REG_SIZE; i++) {
1189 if (state->stack[spi].slot_type[i] == STACK_MISC &&
1190 !env->allow_ptr_leaks) {
1191 int *poff = &env->insn_aux_data[insn_idx].sanitize_stack_off;
1192 int soff = (-spi - 1) * BPF_REG_SIZE;
1194 /* detected reuse of integer stack slot with a pointer
1195 * which means either llvm is reusing stack slot or
1196 * an attacker is trying to exploit CVE-2018-3639
1197 * (speculative store bypass)
1198 * Have to sanitize that slot with preemptive
1201 if (*poff && *poff != soff) {
1202 /* disallow programs where single insn stores
1203 * into two different stack slots, since verifier
1204 * cannot sanitize them
1207 "insn %d cannot access two stack slots fp%d and fp%d",
1208 insn_idx, *poff, soff);
1213 state->stack[spi].slot_type[i] = STACK_SPILL;
1216 u8 type = STACK_MISC;
1218 /* regular write of data into stack destroys any spilled ptr */
1219 state->stack[spi].spilled_ptr.type = NOT_INIT;
1221 /* only mark the slot as written if all 8 bytes were written
1222 * otherwise read propagation may incorrectly stop too soon
1223 * when stack slots are partially written.
1224 * This heuristic means that read propagation will be
1225 * conservative, since it will add reg_live_read marks
1226 * to stack slots all the way to first state when programs
1227 * writes+reads less than 8 bytes
1229 if (size == BPF_REG_SIZE)
1230 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1232 /* when we zero initialize stack slots mark them as such */
1233 if (value_regno >= 0 &&
1234 register_is_null(&cur->regs[value_regno]))
1237 for (i = 0; i < size; i++)
1238 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
1244 static int check_stack_read(struct bpf_verifier_env *env,
1245 struct bpf_func_state *reg_state /* func where register points to */,
1246 int off, int size, int value_regno)
1248 struct bpf_verifier_state *vstate = env->cur_state;
1249 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1250 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
1253 if (reg_state->allocated_stack <= slot) {
1254 verbose(env, "invalid read from stack off %d+0 size %d\n",
1258 stype = reg_state->stack[spi].slot_type;
1260 if (stype[0] == STACK_SPILL) {
1261 if (size != BPF_REG_SIZE) {
1262 verbose(env, "invalid size of register spill\n");
1265 for (i = 1; i < BPF_REG_SIZE; i++) {
1266 if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) {
1267 verbose(env, "corrupted spill memory\n");
1272 if (value_regno >= 0) {
1273 /* restore register state from stack */
1274 state->regs[value_regno] = reg_state->stack[spi].spilled_ptr;
1275 /* mark reg as written since spilled pointer state likely
1276 * has its liveness marks cleared by is_state_visited()
1277 * which resets stack/reg liveness for state transitions
1279 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
1281 mark_reg_read(env, ®_state->stack[spi].spilled_ptr,
1282 reg_state->stack[spi].spilled_ptr.parent);
1287 for (i = 0; i < size; i++) {
1288 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_MISC)
1290 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_ZERO) {
1294 verbose(env, "invalid read from stack off %d+%d size %d\n",
1298 mark_reg_read(env, ®_state->stack[spi].spilled_ptr,
1299 reg_state->stack[spi].spilled_ptr.parent);
1300 if (value_regno >= 0) {
1301 if (zeros == size) {
1302 /* any size read into register is zero extended,
1303 * so the whole register == const_zero
1305 __mark_reg_const_zero(&state->regs[value_regno]);
1307 /* have read misc data from the stack */
1308 mark_reg_unknown(env, state->regs, value_regno);
1310 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
1316 /* check read/write into map element returned by bpf_map_lookup_elem() */
1317 static int __check_map_access(struct bpf_verifier_env *env, u32 regno, int off,
1318 int size, bool zero_size_allowed)
1320 struct bpf_reg_state *regs = cur_regs(env);
1321 struct bpf_map *map = regs[regno].map_ptr;
1323 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
1324 off + size > map->value_size) {
1325 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
1326 map->value_size, off, size);
1332 /* check read/write into a map element with possible variable offset */
1333 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
1334 int off, int size, bool zero_size_allowed)
1336 struct bpf_verifier_state *vstate = env->cur_state;
1337 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1338 struct bpf_reg_state *reg = &state->regs[regno];
1341 /* We may have adjusted the register to this map value, so we
1342 * need to try adding each of min_value and max_value to off
1343 * to make sure our theoretical access will be safe.
1346 print_verifier_state(env, state);
1347 /* The minimum value is only important with signed
1348 * comparisons where we can't assume the floor of a
1349 * value is 0. If we are using signed variables for our
1350 * index'es we need to make sure that whatever we use
1351 * will have a set floor within our range.
1353 if (reg->smin_value < 0) {
1354 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1358 err = __check_map_access(env, regno, reg->smin_value + off, size,
1361 verbose(env, "R%d min value is outside of the array range\n",
1366 /* If we haven't set a max value then we need to bail since we can't be
1367 * sure we won't do bad things.
1368 * If reg->umax_value + off could overflow, treat that as unbounded too.
1370 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
1371 verbose(env, "R%d unbounded memory access, make sure to bounds check any array access into a map\n",
1375 err = __check_map_access(env, regno, reg->umax_value + off, size,
1378 verbose(env, "R%d max value is outside of the array range\n",
1383 #define MAX_PACKET_OFF 0xffff
1385 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
1386 const struct bpf_call_arg_meta *meta,
1387 enum bpf_access_type t)
1389 switch (env->prog->type) {
1390 case BPF_PROG_TYPE_LWT_IN:
1391 case BPF_PROG_TYPE_LWT_OUT:
1392 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
1393 case BPF_PROG_TYPE_SK_REUSEPORT:
1394 /* dst_input() and dst_output() can't write for now */
1398 case BPF_PROG_TYPE_SCHED_CLS:
1399 case BPF_PROG_TYPE_SCHED_ACT:
1400 case BPF_PROG_TYPE_XDP:
1401 case BPF_PROG_TYPE_LWT_XMIT:
1402 case BPF_PROG_TYPE_SK_SKB:
1403 case BPF_PROG_TYPE_SK_MSG:
1404 case BPF_PROG_TYPE_FLOW_DISSECTOR:
1406 return meta->pkt_access;
1408 env->seen_direct_write = true;
1415 static int __check_packet_access(struct bpf_verifier_env *env, u32 regno,
1416 int off, int size, bool zero_size_allowed)
1418 struct bpf_reg_state *regs = cur_regs(env);
1419 struct bpf_reg_state *reg = ®s[regno];
1421 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
1422 (u64)off + size > reg->range) {
1423 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
1424 off, size, regno, reg->id, reg->off, reg->range);
1430 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
1431 int size, bool zero_size_allowed)
1433 struct bpf_reg_state *regs = cur_regs(env);
1434 struct bpf_reg_state *reg = ®s[regno];
1437 /* We may have added a variable offset to the packet pointer; but any
1438 * reg->range we have comes after that. We are only checking the fixed
1442 /* We don't allow negative numbers, because we aren't tracking enough
1443 * detail to prove they're safe.
1445 if (reg->smin_value < 0) {
1446 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1450 err = __check_packet_access(env, regno, off, size, zero_size_allowed);
1452 verbose(env, "R%d offset is outside of the packet\n", regno);
1458 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
1459 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
1460 enum bpf_access_type t, enum bpf_reg_type *reg_type)
1462 struct bpf_insn_access_aux info = {
1463 .reg_type = *reg_type,
1466 if (env->ops->is_valid_access &&
1467 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
1468 /* A non zero info.ctx_field_size indicates that this field is a
1469 * candidate for later verifier transformation to load the whole
1470 * field and then apply a mask when accessed with a narrower
1471 * access than actual ctx access size. A zero info.ctx_field_size
1472 * will only allow for whole field access and rejects any other
1473 * type of narrower access.
1475 *reg_type = info.reg_type;
1477 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
1478 /* remember the offset of last byte accessed in ctx */
1479 if (env->prog->aux->max_ctx_offset < off + size)
1480 env->prog->aux->max_ctx_offset = off + size;
1484 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
1488 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
1491 if (size < 0 || off < 0 ||
1492 (u64)off + size > sizeof(struct bpf_flow_keys)) {
1493 verbose(env, "invalid access to flow keys off=%d size=%d\n",
1500 static int check_sock_access(struct bpf_verifier_env *env, u32 regno, int off,
1501 int size, enum bpf_access_type t)
1503 struct bpf_reg_state *regs = cur_regs(env);
1504 struct bpf_reg_state *reg = ®s[regno];
1505 struct bpf_insn_access_aux info;
1507 if (reg->smin_value < 0) {
1508 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1513 if (!bpf_sock_is_valid_access(off, size, t, &info)) {
1514 verbose(env, "invalid bpf_sock access off=%d size=%d\n",
1522 static bool __is_pointer_value(bool allow_ptr_leaks,
1523 const struct bpf_reg_state *reg)
1525 if (allow_ptr_leaks)
1528 return reg->type != SCALAR_VALUE;
1531 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
1533 return cur_regs(env) + regno;
1536 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
1538 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
1541 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
1543 const struct bpf_reg_state *reg = reg_state(env, regno);
1545 return reg->type == PTR_TO_CTX ||
1546 reg->type == PTR_TO_SOCKET;
1549 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
1551 const struct bpf_reg_state *reg = reg_state(env, regno);
1553 return type_is_pkt_pointer(reg->type);
1556 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
1558 const struct bpf_reg_state *reg = reg_state(env, regno);
1560 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
1561 return reg->type == PTR_TO_FLOW_KEYS;
1564 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
1565 const struct bpf_reg_state *reg,
1566 int off, int size, bool strict)
1568 struct tnum reg_off;
1571 /* Byte size accesses are always allowed. */
1572 if (!strict || size == 1)
1575 /* For platforms that do not have a Kconfig enabling
1576 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
1577 * NET_IP_ALIGN is universally set to '2'. And on platforms
1578 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
1579 * to this code only in strict mode where we want to emulate
1580 * the NET_IP_ALIGN==2 checking. Therefore use an
1581 * unconditional IP align value of '2'.
1585 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
1586 if (!tnum_is_aligned(reg_off, size)) {
1589 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1591 "misaligned packet access off %d+%s+%d+%d size %d\n",
1592 ip_align, tn_buf, reg->off, off, size);
1599 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
1600 const struct bpf_reg_state *reg,
1601 const char *pointer_desc,
1602 int off, int size, bool strict)
1604 struct tnum reg_off;
1606 /* Byte size accesses are always allowed. */
1607 if (!strict || size == 1)
1610 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
1611 if (!tnum_is_aligned(reg_off, size)) {
1614 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1615 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
1616 pointer_desc, tn_buf, reg->off, off, size);
1623 static int check_ptr_alignment(struct bpf_verifier_env *env,
1624 const struct bpf_reg_state *reg, int off,
1625 int size, bool strict_alignment_once)
1627 bool strict = env->strict_alignment || strict_alignment_once;
1628 const char *pointer_desc = "";
1630 switch (reg->type) {
1632 case PTR_TO_PACKET_META:
1633 /* Special case, because of NET_IP_ALIGN. Given metadata sits
1634 * right in front, treat it the very same way.
1636 return check_pkt_ptr_alignment(env, reg, off, size, strict);
1637 case PTR_TO_FLOW_KEYS:
1638 pointer_desc = "flow keys ";
1640 case PTR_TO_MAP_VALUE:
1641 pointer_desc = "value ";
1644 pointer_desc = "context ";
1647 pointer_desc = "stack ";
1648 /* The stack spill tracking logic in check_stack_write()
1649 * and check_stack_read() relies on stack accesses being
1655 pointer_desc = "sock ";
1660 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
1664 static int update_stack_depth(struct bpf_verifier_env *env,
1665 const struct bpf_func_state *func,
1668 u16 stack = env->subprog_info[func->subprogno].stack_depth;
1673 /* update known max for given subprogram */
1674 env->subprog_info[func->subprogno].stack_depth = -off;
1678 /* starting from main bpf function walk all instructions of the function
1679 * and recursively walk all callees that given function can call.
1680 * Ignore jump and exit insns.
1681 * Since recursion is prevented by check_cfg() this algorithm
1682 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
1684 static int check_max_stack_depth(struct bpf_verifier_env *env)
1686 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
1687 struct bpf_subprog_info *subprog = env->subprog_info;
1688 struct bpf_insn *insn = env->prog->insnsi;
1689 int ret_insn[MAX_CALL_FRAMES];
1690 int ret_prog[MAX_CALL_FRAMES];
1693 /* round up to 32-bytes, since this is granularity
1694 * of interpreter stack size
1696 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
1697 if (depth > MAX_BPF_STACK) {
1698 verbose(env, "combined stack size of %d calls is %d. Too large\n",
1703 subprog_end = subprog[idx + 1].start;
1704 for (; i < subprog_end; i++) {
1705 if (insn[i].code != (BPF_JMP | BPF_CALL))
1707 if (insn[i].src_reg != BPF_PSEUDO_CALL)
1709 /* remember insn and function to return to */
1710 ret_insn[frame] = i + 1;
1711 ret_prog[frame] = idx;
1713 /* find the callee */
1714 i = i + insn[i].imm + 1;
1715 idx = find_subprog(env, i);
1717 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
1722 if (frame >= MAX_CALL_FRAMES) {
1723 WARN_ONCE(1, "verifier bug. Call stack is too deep\n");
1728 /* end of for() loop means the last insn of the 'subprog'
1729 * was reached. Doesn't matter whether it was JA or EXIT
1733 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
1735 i = ret_insn[frame];
1736 idx = ret_prog[frame];
1740 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
1741 static int get_callee_stack_depth(struct bpf_verifier_env *env,
1742 const struct bpf_insn *insn, int idx)
1744 int start = idx + insn->imm + 1, subprog;
1746 subprog = find_subprog(env, start);
1748 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
1752 return env->subprog_info[subprog].stack_depth;
1756 static int check_ctx_reg(struct bpf_verifier_env *env,
1757 const struct bpf_reg_state *reg, int regno)
1759 /* Access to ctx or passing it to a helper is only allowed in
1760 * its original, unmodified form.
1764 verbose(env, "dereference of modified ctx ptr R%d off=%d disallowed\n",
1769 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
1772 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1773 verbose(env, "variable ctx access var_off=%s disallowed\n", tn_buf);
1780 /* truncate register to smaller size (in bytes)
1781 * must be called with size < BPF_REG_SIZE
1783 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
1787 /* clear high bits in bit representation */
1788 reg->var_off = tnum_cast(reg->var_off, size);
1790 /* fix arithmetic bounds */
1791 mask = ((u64)1 << (size * 8)) - 1;
1792 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
1793 reg->umin_value &= mask;
1794 reg->umax_value &= mask;
1796 reg->umin_value = 0;
1797 reg->umax_value = mask;
1799 reg->smin_value = reg->umin_value;
1800 reg->smax_value = reg->umax_value;
1803 /* check whether memory at (regno + off) is accessible for t = (read | write)
1804 * if t==write, value_regno is a register which value is stored into memory
1805 * if t==read, value_regno is a register which will receive the value from memory
1806 * if t==write && value_regno==-1, some unknown value is stored into memory
1807 * if t==read && value_regno==-1, don't care what we read from memory
1809 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
1810 int off, int bpf_size, enum bpf_access_type t,
1811 int value_regno, bool strict_alignment_once)
1813 struct bpf_reg_state *regs = cur_regs(env);
1814 struct bpf_reg_state *reg = regs + regno;
1815 struct bpf_func_state *state;
1818 size = bpf_size_to_bytes(bpf_size);
1822 /* alignment checks will add in reg->off themselves */
1823 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
1827 /* for access checks, reg->off is just part of off */
1830 if (reg->type == PTR_TO_MAP_VALUE) {
1831 if (t == BPF_WRITE && value_regno >= 0 &&
1832 is_pointer_value(env, value_regno)) {
1833 verbose(env, "R%d leaks addr into map\n", value_regno);
1837 err = check_map_access(env, regno, off, size, false);
1838 if (!err && t == BPF_READ && value_regno >= 0)
1839 mark_reg_unknown(env, regs, value_regno);
1841 } else if (reg->type == PTR_TO_CTX) {
1842 enum bpf_reg_type reg_type = SCALAR_VALUE;
1844 if (t == BPF_WRITE && value_regno >= 0 &&
1845 is_pointer_value(env, value_regno)) {
1846 verbose(env, "R%d leaks addr into ctx\n", value_regno);
1850 err = check_ctx_reg(env, reg, regno);
1854 err = check_ctx_access(env, insn_idx, off, size, t, ®_type);
1855 if (!err && t == BPF_READ && value_regno >= 0) {
1856 /* ctx access returns either a scalar, or a
1857 * PTR_TO_PACKET[_META,_END]. In the latter
1858 * case, we know the offset is zero.
1860 if (reg_type == SCALAR_VALUE)
1861 mark_reg_unknown(env, regs, value_regno);
1863 mark_reg_known_zero(env, regs,
1865 regs[value_regno].type = reg_type;
1868 } else if (reg->type == PTR_TO_STACK) {
1869 /* stack accesses must be at a fixed offset, so that we can
1870 * determine what type of data were returned.
1871 * See check_stack_read().
1873 if (!tnum_is_const(reg->var_off)) {
1876 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1877 verbose(env, "variable stack access var_off=%s off=%d size=%d",
1881 off += reg->var_off.value;
1882 if (off >= 0 || off < -MAX_BPF_STACK) {
1883 verbose(env, "invalid stack off=%d size=%d\n", off,
1888 state = func(env, reg);
1889 err = update_stack_depth(env, state, off);
1894 err = check_stack_write(env, state, off, size,
1895 value_regno, insn_idx);
1897 err = check_stack_read(env, state, off, size,
1899 } else if (reg_is_pkt_pointer(reg)) {
1900 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
1901 verbose(env, "cannot write into packet\n");
1904 if (t == BPF_WRITE && value_regno >= 0 &&
1905 is_pointer_value(env, value_regno)) {
1906 verbose(env, "R%d leaks addr into packet\n",
1910 err = check_packet_access(env, regno, off, size, false);
1911 if (!err && t == BPF_READ && value_regno >= 0)
1912 mark_reg_unknown(env, regs, value_regno);
1913 } else if (reg->type == PTR_TO_FLOW_KEYS) {
1914 if (t == BPF_WRITE && value_regno >= 0 &&
1915 is_pointer_value(env, value_regno)) {
1916 verbose(env, "R%d leaks addr into flow keys\n",
1921 err = check_flow_keys_access(env, off, size);
1922 if (!err && t == BPF_READ && value_regno >= 0)
1923 mark_reg_unknown(env, regs, value_regno);
1924 } else if (reg->type == PTR_TO_SOCKET) {
1925 if (t == BPF_WRITE) {
1926 verbose(env, "cannot write into socket\n");
1929 err = check_sock_access(env, regno, off, size, t);
1930 if (!err && value_regno >= 0)
1931 mark_reg_unknown(env, regs, value_regno);
1933 verbose(env, "R%d invalid mem access '%s'\n", regno,
1934 reg_type_str[reg->type]);
1938 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
1939 regs[value_regno].type == SCALAR_VALUE) {
1940 /* b/h/w load zero-extends, mark upper bits as known 0 */
1941 coerce_reg_to_size(®s[value_regno], size);
1946 static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
1950 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
1952 verbose(env, "BPF_XADD uses reserved fields\n");
1956 /* check src1 operand */
1957 err = check_reg_arg(env, insn->src_reg, SRC_OP);
1961 /* check src2 operand */
1962 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
1966 if (is_pointer_value(env, insn->src_reg)) {
1967 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
1971 if (is_ctx_reg(env, insn->dst_reg) ||
1972 is_pkt_reg(env, insn->dst_reg) ||
1973 is_flow_key_reg(env, insn->dst_reg)) {
1974 verbose(env, "BPF_XADD stores into R%d %s is not allowed\n",
1976 reg_type_str[reg_state(env, insn->dst_reg)->type]);
1980 /* check whether atomic_add can read the memory */
1981 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1982 BPF_SIZE(insn->code), BPF_READ, -1, true);
1986 /* check whether atomic_add can write into the same memory */
1987 return check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
1988 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
1991 /* when register 'regno' is passed into function that will read 'access_size'
1992 * bytes from that pointer, make sure that it's within stack boundary
1993 * and all elements of stack are initialized.
1994 * Unlike most pointer bounds-checking functions, this one doesn't take an
1995 * 'off' argument, so it has to add in reg->off itself.
1997 static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
1998 int access_size, bool zero_size_allowed,
1999 struct bpf_call_arg_meta *meta)
2001 struct bpf_reg_state *reg = reg_state(env, regno);
2002 struct bpf_func_state *state = func(env, reg);
2003 int off, i, slot, spi;
2005 if (reg->type != PTR_TO_STACK) {
2006 /* Allow zero-byte read from NULL, regardless of pointer type */
2007 if (zero_size_allowed && access_size == 0 &&
2008 register_is_null(reg))
2011 verbose(env, "R%d type=%s expected=%s\n", regno,
2012 reg_type_str[reg->type],
2013 reg_type_str[PTR_TO_STACK]);
2017 /* Only allow fixed-offset stack reads */
2018 if (!tnum_is_const(reg->var_off)) {
2021 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2022 verbose(env, "invalid variable stack read R%d var_off=%s\n",
2026 off = reg->off + reg->var_off.value;
2027 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
2028 access_size < 0 || (access_size == 0 && !zero_size_allowed)) {
2029 verbose(env, "invalid stack type R%d off=%d access_size=%d\n",
2030 regno, off, access_size);
2034 if (meta && meta->raw_mode) {
2035 meta->access_size = access_size;
2036 meta->regno = regno;
2040 for (i = 0; i < access_size; i++) {
2043 slot = -(off + i) - 1;
2044 spi = slot / BPF_REG_SIZE;
2045 if (state->allocated_stack <= slot)
2047 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
2048 if (*stype == STACK_MISC)
2050 if (*stype == STACK_ZERO) {
2051 /* helper can write anything into the stack */
2052 *stype = STACK_MISC;
2056 verbose(env, "invalid indirect read from stack off %d+%d size %d\n",
2057 off, i, access_size);
2060 /* reading any byte out of 8-byte 'spill_slot' will cause
2061 * the whole slot to be marked as 'read'
2063 mark_reg_read(env, &state->stack[spi].spilled_ptr,
2064 state->stack[spi].spilled_ptr.parent);
2066 return update_stack_depth(env, state, off);
2069 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
2070 int access_size, bool zero_size_allowed,
2071 struct bpf_call_arg_meta *meta)
2073 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
2075 switch (reg->type) {
2077 case PTR_TO_PACKET_META:
2078 return check_packet_access(env, regno, reg->off, access_size,
2080 case PTR_TO_MAP_VALUE:
2081 return check_map_access(env, regno, reg->off, access_size,
2083 default: /* scalar_value|ptr_to_stack or invalid ptr */
2084 return check_stack_boundary(env, regno, access_size,
2085 zero_size_allowed, meta);
2089 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
2091 return type == ARG_PTR_TO_MEM ||
2092 type == ARG_PTR_TO_MEM_OR_NULL ||
2093 type == ARG_PTR_TO_UNINIT_MEM;
2096 static bool arg_type_is_mem_size(enum bpf_arg_type type)
2098 return type == ARG_CONST_SIZE ||
2099 type == ARG_CONST_SIZE_OR_ZERO;
2102 static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
2103 enum bpf_arg_type arg_type,
2104 struct bpf_call_arg_meta *meta)
2106 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
2107 enum bpf_reg_type expected_type, type = reg->type;
2110 if (arg_type == ARG_DONTCARE)
2113 err = check_reg_arg(env, regno, SRC_OP);
2117 if (arg_type == ARG_ANYTHING) {
2118 if (is_pointer_value(env, regno)) {
2119 verbose(env, "R%d leaks addr into helper function\n",
2126 if (type_is_pkt_pointer(type) &&
2127 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
2128 verbose(env, "helper access to the packet is not allowed\n");
2132 if (arg_type == ARG_PTR_TO_MAP_KEY ||
2133 arg_type == ARG_PTR_TO_MAP_VALUE ||
2134 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
2135 expected_type = PTR_TO_STACK;
2136 if (!type_is_pkt_pointer(type) && type != PTR_TO_MAP_VALUE &&
2137 type != expected_type)
2139 } else if (arg_type == ARG_CONST_SIZE ||
2140 arg_type == ARG_CONST_SIZE_OR_ZERO) {
2141 expected_type = SCALAR_VALUE;
2142 if (type != expected_type)
2144 } else if (arg_type == ARG_CONST_MAP_PTR) {
2145 expected_type = CONST_PTR_TO_MAP;
2146 if (type != expected_type)
2148 } else if (arg_type == ARG_PTR_TO_CTX) {
2149 expected_type = PTR_TO_CTX;
2150 if (type != expected_type)
2152 err = check_ctx_reg(env, reg, regno);
2155 } else if (arg_type == ARG_PTR_TO_SOCKET) {
2156 expected_type = PTR_TO_SOCKET;
2157 if (type != expected_type)
2159 if (meta->ptr_id || !reg->id) {
2160 verbose(env, "verifier internal error: mismatched references meta=%d, reg=%d\n",
2161 meta->ptr_id, reg->id);
2164 meta->ptr_id = reg->id;
2165 } else if (arg_type_is_mem_ptr(arg_type)) {
2166 expected_type = PTR_TO_STACK;
2167 /* One exception here. In case function allows for NULL to be
2168 * passed in as argument, it's a SCALAR_VALUE type. Final test
2169 * happens during stack boundary checking.
2171 if (register_is_null(reg) &&
2172 arg_type == ARG_PTR_TO_MEM_OR_NULL)
2173 /* final test in check_stack_boundary() */;
2174 else if (!type_is_pkt_pointer(type) &&
2175 type != PTR_TO_MAP_VALUE &&
2176 type != expected_type)
2178 meta->raw_mode = arg_type == ARG_PTR_TO_UNINIT_MEM;
2180 verbose(env, "unsupported arg_type %d\n", arg_type);
2184 if (arg_type == ARG_CONST_MAP_PTR) {
2185 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
2186 meta->map_ptr = reg->map_ptr;
2187 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
2188 /* bpf_map_xxx(..., map_ptr, ..., key) call:
2189 * check that [key, key + map->key_size) are within
2190 * stack limits and initialized
2192 if (!meta->map_ptr) {
2193 /* in function declaration map_ptr must come before
2194 * map_key, so that it's verified and known before
2195 * we have to check map_key here. Otherwise it means
2196 * that kernel subsystem misconfigured verifier
2198 verbose(env, "invalid map_ptr to access map->key\n");
2201 err = check_helper_mem_access(env, regno,
2202 meta->map_ptr->key_size, false,
2204 } else if (arg_type == ARG_PTR_TO_MAP_VALUE ||
2205 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
2206 /* bpf_map_xxx(..., map_ptr, ..., value) call:
2207 * check [value, value + map->value_size) validity
2209 if (!meta->map_ptr) {
2210 /* kernel subsystem misconfigured verifier */
2211 verbose(env, "invalid map_ptr to access map->value\n");
2214 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE);
2215 err = check_helper_mem_access(env, regno,
2216 meta->map_ptr->value_size, false,
2218 } else if (arg_type_is_mem_size(arg_type)) {
2219 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
2221 /* remember the mem_size which may be used later
2222 * to refine return values.
2224 meta->msize_smax_value = reg->smax_value;
2225 meta->msize_umax_value = reg->umax_value;
2227 /* The register is SCALAR_VALUE; the access check
2228 * happens using its boundaries.
2230 if (!tnum_is_const(reg->var_off))
2231 /* For unprivileged variable accesses, disable raw
2232 * mode so that the program is required to
2233 * initialize all the memory that the helper could
2234 * just partially fill up.
2238 if (reg->smin_value < 0) {
2239 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
2244 if (reg->umin_value == 0) {
2245 err = check_helper_mem_access(env, regno - 1, 0,
2252 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
2253 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
2257 err = check_helper_mem_access(env, regno - 1,
2259 zero_size_allowed, meta);
2264 verbose(env, "R%d type=%s expected=%s\n", regno,
2265 reg_type_str[type], reg_type_str[expected_type]);
2269 static int check_map_func_compatibility(struct bpf_verifier_env *env,
2270 struct bpf_map *map, int func_id)
2275 /* We need a two way check, first is from map perspective ... */
2276 switch (map->map_type) {
2277 case BPF_MAP_TYPE_PROG_ARRAY:
2278 if (func_id != BPF_FUNC_tail_call)
2281 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
2282 if (func_id != BPF_FUNC_perf_event_read &&
2283 func_id != BPF_FUNC_perf_event_output &&
2284 func_id != BPF_FUNC_perf_event_read_value)
2287 case BPF_MAP_TYPE_STACK_TRACE:
2288 if (func_id != BPF_FUNC_get_stackid)
2291 case BPF_MAP_TYPE_CGROUP_ARRAY:
2292 if (func_id != BPF_FUNC_skb_under_cgroup &&
2293 func_id != BPF_FUNC_current_task_under_cgroup)
2296 case BPF_MAP_TYPE_CGROUP_STORAGE:
2297 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
2298 if (func_id != BPF_FUNC_get_local_storage)
2301 /* devmap returns a pointer to a live net_device ifindex that we cannot
2302 * allow to be modified from bpf side. So do not allow lookup elements
2305 case BPF_MAP_TYPE_DEVMAP:
2306 if (func_id != BPF_FUNC_redirect_map)
2309 /* Restrict bpf side of cpumap and xskmap, open when use-cases
2312 case BPF_MAP_TYPE_CPUMAP:
2313 case BPF_MAP_TYPE_XSKMAP:
2314 if (func_id != BPF_FUNC_redirect_map)
2317 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
2318 case BPF_MAP_TYPE_HASH_OF_MAPS:
2319 if (func_id != BPF_FUNC_map_lookup_elem)
2322 case BPF_MAP_TYPE_SOCKMAP:
2323 if (func_id != BPF_FUNC_sk_redirect_map &&
2324 func_id != BPF_FUNC_sock_map_update &&
2325 func_id != BPF_FUNC_map_delete_elem &&
2326 func_id != BPF_FUNC_msg_redirect_map)
2329 case BPF_MAP_TYPE_SOCKHASH:
2330 if (func_id != BPF_FUNC_sk_redirect_hash &&
2331 func_id != BPF_FUNC_sock_hash_update &&
2332 func_id != BPF_FUNC_map_delete_elem &&
2333 func_id != BPF_FUNC_msg_redirect_hash)
2336 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
2337 if (func_id != BPF_FUNC_sk_select_reuseport)
2340 case BPF_MAP_TYPE_QUEUE:
2341 case BPF_MAP_TYPE_STACK:
2342 if (func_id != BPF_FUNC_map_peek_elem &&
2343 func_id != BPF_FUNC_map_pop_elem &&
2344 func_id != BPF_FUNC_map_push_elem)
2351 /* ... and second from the function itself. */
2353 case BPF_FUNC_tail_call:
2354 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
2356 if (env->subprog_cnt > 1) {
2357 verbose(env, "tail_calls are not allowed in programs with bpf-to-bpf calls\n");
2361 case BPF_FUNC_perf_event_read:
2362 case BPF_FUNC_perf_event_output:
2363 case BPF_FUNC_perf_event_read_value:
2364 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
2367 case BPF_FUNC_get_stackid:
2368 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
2371 case BPF_FUNC_current_task_under_cgroup:
2372 case BPF_FUNC_skb_under_cgroup:
2373 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
2376 case BPF_FUNC_redirect_map:
2377 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
2378 map->map_type != BPF_MAP_TYPE_CPUMAP &&
2379 map->map_type != BPF_MAP_TYPE_XSKMAP)
2382 case BPF_FUNC_sk_redirect_map:
2383 case BPF_FUNC_msg_redirect_map:
2384 case BPF_FUNC_sock_map_update:
2385 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
2388 case BPF_FUNC_sk_redirect_hash:
2389 case BPF_FUNC_msg_redirect_hash:
2390 case BPF_FUNC_sock_hash_update:
2391 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
2394 case BPF_FUNC_get_local_storage:
2395 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
2396 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
2399 case BPF_FUNC_sk_select_reuseport:
2400 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY)
2403 case BPF_FUNC_map_peek_elem:
2404 case BPF_FUNC_map_pop_elem:
2405 case BPF_FUNC_map_push_elem:
2406 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
2407 map->map_type != BPF_MAP_TYPE_STACK)
2416 verbose(env, "cannot pass map_type %d into func %s#%d\n",
2417 map->map_type, func_id_name(func_id), func_id);
2421 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
2425 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
2427 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
2429 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
2431 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
2433 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
2436 /* We only support one arg being in raw mode at the moment,
2437 * which is sufficient for the helper functions we have
2443 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
2444 enum bpf_arg_type arg_next)
2446 return (arg_type_is_mem_ptr(arg_curr) &&
2447 !arg_type_is_mem_size(arg_next)) ||
2448 (!arg_type_is_mem_ptr(arg_curr) &&
2449 arg_type_is_mem_size(arg_next));
2452 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
2454 /* bpf_xxx(..., buf, len) call will access 'len'
2455 * bytes from memory 'buf'. Both arg types need
2456 * to be paired, so make sure there's no buggy
2457 * helper function specification.
2459 if (arg_type_is_mem_size(fn->arg1_type) ||
2460 arg_type_is_mem_ptr(fn->arg5_type) ||
2461 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
2462 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
2463 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
2464 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
2470 static bool check_refcount_ok(const struct bpf_func_proto *fn)
2474 if (arg_type_is_refcounted(fn->arg1_type))
2476 if (arg_type_is_refcounted(fn->arg2_type))
2478 if (arg_type_is_refcounted(fn->arg3_type))
2480 if (arg_type_is_refcounted(fn->arg4_type))
2482 if (arg_type_is_refcounted(fn->arg5_type))
2485 /* We only support one arg being unreferenced at the moment,
2486 * which is sufficient for the helper functions we have right now.
2491 static int check_func_proto(const struct bpf_func_proto *fn)
2493 return check_raw_mode_ok(fn) &&
2494 check_arg_pair_ok(fn) &&
2495 check_refcount_ok(fn) ? 0 : -EINVAL;
2498 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
2499 * are now invalid, so turn them into unknown SCALAR_VALUE.
2501 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
2502 struct bpf_func_state *state)
2504 struct bpf_reg_state *regs = state->regs, *reg;
2507 for (i = 0; i < MAX_BPF_REG; i++)
2508 if (reg_is_pkt_pointer_any(®s[i]))
2509 mark_reg_unknown(env, regs, i);
2511 bpf_for_each_spilled_reg(i, state, reg) {
2514 if (reg_is_pkt_pointer_any(reg))
2515 __mark_reg_unknown(reg);
2519 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
2521 struct bpf_verifier_state *vstate = env->cur_state;
2524 for (i = 0; i <= vstate->curframe; i++)
2525 __clear_all_pkt_pointers(env, vstate->frame[i]);
2528 static void release_reg_references(struct bpf_verifier_env *env,
2529 struct bpf_func_state *state, int id)
2531 struct bpf_reg_state *regs = state->regs, *reg;
2534 for (i = 0; i < MAX_BPF_REG; i++)
2535 if (regs[i].id == id)
2536 mark_reg_unknown(env, regs, i);
2538 bpf_for_each_spilled_reg(i, state, reg) {
2541 if (reg_is_refcounted(reg) && reg->id == id)
2542 __mark_reg_unknown(reg);
2546 /* The pointer with the specified id has released its reference to kernel
2547 * resources. Identify all copies of the same pointer and clear the reference.
2549 static int release_reference(struct bpf_verifier_env *env,
2550 struct bpf_call_arg_meta *meta)
2552 struct bpf_verifier_state *vstate = env->cur_state;
2555 for (i = 0; i <= vstate->curframe; i++)
2556 release_reg_references(env, vstate->frame[i], meta->ptr_id);
2558 return release_reference_state(env, meta->ptr_id);
2561 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
2564 struct bpf_verifier_state *state = env->cur_state;
2565 struct bpf_func_state *caller, *callee;
2566 int i, err, subprog, target_insn;
2568 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
2569 verbose(env, "the call stack of %d frames is too deep\n",
2570 state->curframe + 2);
2574 target_insn = *insn_idx + insn->imm;
2575 subprog = find_subprog(env, target_insn + 1);
2577 verbose(env, "verifier bug. No program starts at insn %d\n",
2582 caller = state->frame[state->curframe];
2583 if (state->frame[state->curframe + 1]) {
2584 verbose(env, "verifier bug. Frame %d already allocated\n",
2585 state->curframe + 1);
2589 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
2592 state->frame[state->curframe + 1] = callee;
2594 /* callee cannot access r0, r6 - r9 for reading and has to write
2595 * into its own stack before reading from it.
2596 * callee can read/write into caller's stack
2598 init_func_state(env, callee,
2599 /* remember the callsite, it will be used by bpf_exit */
2600 *insn_idx /* callsite */,
2601 state->curframe + 1 /* frameno within this callchain */,
2602 subprog /* subprog number within this prog */);
2604 /* Transfer references to the callee */
2605 err = transfer_reference_state(callee, caller);
2609 /* copy r1 - r5 args that callee can access. The copy includes parent
2610 * pointers, which connects us up to the liveness chain
2612 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
2613 callee->regs[i] = caller->regs[i];
2615 /* after the call registers r0 - r5 were scratched */
2616 for (i = 0; i < CALLER_SAVED_REGS; i++) {
2617 mark_reg_not_init(env, caller->regs, caller_saved[i]);
2618 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
2621 /* only increment it after check_reg_arg() finished */
2624 /* and go analyze first insn of the callee */
2625 *insn_idx = target_insn;
2627 if (env->log.level) {
2628 verbose(env, "caller:\n");
2629 print_verifier_state(env, caller);
2630 verbose(env, "callee:\n");
2631 print_verifier_state(env, callee);
2636 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
2638 struct bpf_verifier_state *state = env->cur_state;
2639 struct bpf_func_state *caller, *callee;
2640 struct bpf_reg_state *r0;
2643 callee = state->frame[state->curframe];
2644 r0 = &callee->regs[BPF_REG_0];
2645 if (r0->type == PTR_TO_STACK) {
2646 /* technically it's ok to return caller's stack pointer
2647 * (or caller's caller's pointer) back to the caller,
2648 * since these pointers are valid. Only current stack
2649 * pointer will be invalid as soon as function exits,
2650 * but let's be conservative
2652 verbose(env, "cannot return stack pointer to the caller\n");
2657 caller = state->frame[state->curframe];
2658 /* return to the caller whatever r0 had in the callee */
2659 caller->regs[BPF_REG_0] = *r0;
2661 /* Transfer references to the caller */
2662 err = transfer_reference_state(caller, callee);
2666 *insn_idx = callee->callsite + 1;
2667 if (env->log.level) {
2668 verbose(env, "returning from callee:\n");
2669 print_verifier_state(env, callee);
2670 verbose(env, "to caller at %d:\n", *insn_idx);
2671 print_verifier_state(env, caller);
2673 /* clear everything in the callee */
2674 free_func_state(callee);
2675 state->frame[state->curframe + 1] = NULL;
2679 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
2681 struct bpf_call_arg_meta *meta)
2683 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
2685 if (ret_type != RET_INTEGER ||
2686 (func_id != BPF_FUNC_get_stack &&
2687 func_id != BPF_FUNC_probe_read_str))
2690 ret_reg->smax_value = meta->msize_smax_value;
2691 ret_reg->umax_value = meta->msize_umax_value;
2692 __reg_deduce_bounds(ret_reg);
2693 __reg_bound_offset(ret_reg);
2697 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
2698 int func_id, int insn_idx)
2700 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
2702 if (func_id != BPF_FUNC_tail_call &&
2703 func_id != BPF_FUNC_map_lookup_elem &&
2704 func_id != BPF_FUNC_map_update_elem &&
2705 func_id != BPF_FUNC_map_delete_elem &&
2706 func_id != BPF_FUNC_map_push_elem &&
2707 func_id != BPF_FUNC_map_pop_elem &&
2708 func_id != BPF_FUNC_map_peek_elem)
2711 if (meta->map_ptr == NULL) {
2712 verbose(env, "kernel subsystem misconfigured verifier\n");
2716 if (!BPF_MAP_PTR(aux->map_state))
2717 bpf_map_ptr_store(aux, meta->map_ptr,
2718 meta->map_ptr->unpriv_array);
2719 else if (BPF_MAP_PTR(aux->map_state) != meta->map_ptr)
2720 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
2721 meta->map_ptr->unpriv_array);
2725 static int check_reference_leak(struct bpf_verifier_env *env)
2727 struct bpf_func_state *state = cur_func(env);
2730 for (i = 0; i < state->acquired_refs; i++) {
2731 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
2732 state->refs[i].id, state->refs[i].insn_idx);
2734 return state->acquired_refs ? -EINVAL : 0;
2737 static int check_helper_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
2739 const struct bpf_func_proto *fn = NULL;
2740 struct bpf_reg_state *regs;
2741 struct bpf_call_arg_meta meta;
2745 /* find function prototype */
2746 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
2747 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
2752 if (env->ops->get_func_proto)
2753 fn = env->ops->get_func_proto(func_id, env->prog);
2755 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
2760 /* eBPF programs must be GPL compatible to use GPL-ed functions */
2761 if (!env->prog->gpl_compatible && fn->gpl_only) {
2762 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
2766 /* With LD_ABS/IND some JITs save/restore skb from r1. */
2767 changes_data = bpf_helper_changes_pkt_data(fn->func);
2768 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
2769 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
2770 func_id_name(func_id), func_id);
2774 memset(&meta, 0, sizeof(meta));
2775 meta.pkt_access = fn->pkt_access;
2777 err = check_func_proto(fn);
2779 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
2780 func_id_name(func_id), func_id);
2785 err = check_func_arg(env, BPF_REG_1, fn->arg1_type, &meta);
2788 err = check_func_arg(env, BPF_REG_2, fn->arg2_type, &meta);
2791 err = check_func_arg(env, BPF_REG_3, fn->arg3_type, &meta);
2794 err = check_func_arg(env, BPF_REG_4, fn->arg4_type, &meta);
2797 err = check_func_arg(env, BPF_REG_5, fn->arg5_type, &meta);
2801 err = record_func_map(env, &meta, func_id, insn_idx);
2805 /* Mark slots with STACK_MISC in case of raw mode, stack offset
2806 * is inferred from register state.
2808 for (i = 0; i < meta.access_size; i++) {
2809 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
2810 BPF_WRITE, -1, false);
2815 if (func_id == BPF_FUNC_tail_call) {
2816 err = check_reference_leak(env);
2818 verbose(env, "tail_call would lead to reference leak\n");
2821 } else if (is_release_function(func_id)) {
2822 err = release_reference(env, &meta);
2827 regs = cur_regs(env);
2829 /* check that flags argument in get_local_storage(map, flags) is 0,
2830 * this is required because get_local_storage() can't return an error.
2832 if (func_id == BPF_FUNC_get_local_storage &&
2833 !register_is_null(®s[BPF_REG_2])) {
2834 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
2838 /* reset caller saved regs */
2839 for (i = 0; i < CALLER_SAVED_REGS; i++) {
2840 mark_reg_not_init(env, regs, caller_saved[i]);
2841 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
2844 /* update return register (already marked as written above) */
2845 if (fn->ret_type == RET_INTEGER) {
2846 /* sets type to SCALAR_VALUE */
2847 mark_reg_unknown(env, regs, BPF_REG_0);
2848 } else if (fn->ret_type == RET_VOID) {
2849 regs[BPF_REG_0].type = NOT_INIT;
2850 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL ||
2851 fn->ret_type == RET_PTR_TO_MAP_VALUE) {
2852 if (fn->ret_type == RET_PTR_TO_MAP_VALUE)
2853 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE;
2855 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
2856 /* There is no offset yet applied, variable or fixed */
2857 mark_reg_known_zero(env, regs, BPF_REG_0);
2858 /* remember map_ptr, so that check_map_access()
2859 * can check 'value_size' boundary of memory access
2860 * to map element returned from bpf_map_lookup_elem()
2862 if (meta.map_ptr == NULL) {
2864 "kernel subsystem misconfigured verifier\n");
2867 regs[BPF_REG_0].map_ptr = meta.map_ptr;
2868 regs[BPF_REG_0].id = ++env->id_gen;
2869 } else if (fn->ret_type == RET_PTR_TO_SOCKET_OR_NULL) {
2870 int id = acquire_reference_state(env, insn_idx);
2873 mark_reg_known_zero(env, regs, BPF_REG_0);
2874 regs[BPF_REG_0].type = PTR_TO_SOCKET_OR_NULL;
2875 regs[BPF_REG_0].id = id;
2877 verbose(env, "unknown return type %d of func %s#%d\n",
2878 fn->ret_type, func_id_name(func_id), func_id);
2882 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
2884 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
2888 if (func_id == BPF_FUNC_get_stack && !env->prog->has_callchain_buf) {
2889 const char *err_str;
2891 #ifdef CONFIG_PERF_EVENTS
2892 err = get_callchain_buffers(sysctl_perf_event_max_stack);
2893 err_str = "cannot get callchain buffer for func %s#%d\n";
2896 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
2899 verbose(env, err_str, func_id_name(func_id), func_id);
2903 env->prog->has_callchain_buf = true;
2907 clear_all_pkt_pointers(env);
2911 static bool signed_add_overflows(s64 a, s64 b)
2913 /* Do the add in u64, where overflow is well-defined */
2914 s64 res = (s64)((u64)a + (u64)b);
2921 static bool signed_sub_overflows(s64 a, s64 b)
2923 /* Do the sub in u64, where overflow is well-defined */
2924 s64 res = (s64)((u64)a - (u64)b);
2931 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
2932 const struct bpf_reg_state *reg,
2933 enum bpf_reg_type type)
2935 bool known = tnum_is_const(reg->var_off);
2936 s64 val = reg->var_off.value;
2937 s64 smin = reg->smin_value;
2939 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
2940 verbose(env, "math between %s pointer and %lld is not allowed\n",
2941 reg_type_str[type], val);
2945 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
2946 verbose(env, "%s pointer offset %d is not allowed\n",
2947 reg_type_str[type], reg->off);
2951 if (smin == S64_MIN) {
2952 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
2953 reg_type_str[type]);
2957 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
2958 verbose(env, "value %lld makes %s pointer be out of bounds\n",
2959 smin, reg_type_str[type]);
2966 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
2967 * Caller should also handle BPF_MOV case separately.
2968 * If we return -EACCES, caller may want to try again treating pointer as a
2969 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
2971 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
2972 struct bpf_insn *insn,
2973 const struct bpf_reg_state *ptr_reg,
2974 const struct bpf_reg_state *off_reg)
2976 struct bpf_verifier_state *vstate = env->cur_state;
2977 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2978 struct bpf_reg_state *regs = state->regs, *dst_reg;
2979 bool known = tnum_is_const(off_reg->var_off);
2980 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
2981 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
2982 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
2983 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
2984 u8 opcode = BPF_OP(insn->code);
2985 u32 dst = insn->dst_reg;
2987 dst_reg = ®s[dst];
2989 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
2990 smin_val > smax_val || umin_val > umax_val) {
2991 /* Taint dst register if offset had invalid bounds derived from
2992 * e.g. dead branches.
2994 __mark_reg_unknown(dst_reg);
2998 if (BPF_CLASS(insn->code) != BPF_ALU64) {
2999 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
3001 "R%d 32-bit pointer arithmetic prohibited\n",
3006 switch (ptr_reg->type) {
3007 case PTR_TO_MAP_VALUE_OR_NULL:
3008 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
3009 dst, reg_type_str[ptr_reg->type]);
3011 case CONST_PTR_TO_MAP:
3012 case PTR_TO_PACKET_END:
3014 case PTR_TO_SOCKET_OR_NULL:
3015 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
3016 dst, reg_type_str[ptr_reg->type]);
3022 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
3023 * The id may be overwritten later if we create a new variable offset.
3025 dst_reg->type = ptr_reg->type;
3026 dst_reg->id = ptr_reg->id;
3028 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
3029 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
3034 /* We can take a fixed offset as long as it doesn't overflow
3035 * the s32 'off' field
3037 if (known && (ptr_reg->off + smin_val ==
3038 (s64)(s32)(ptr_reg->off + smin_val))) {
3039 /* pointer += K. Accumulate it into fixed offset */
3040 dst_reg->smin_value = smin_ptr;
3041 dst_reg->smax_value = smax_ptr;
3042 dst_reg->umin_value = umin_ptr;
3043 dst_reg->umax_value = umax_ptr;
3044 dst_reg->var_off = ptr_reg->var_off;
3045 dst_reg->off = ptr_reg->off + smin_val;
3046 dst_reg->range = ptr_reg->range;
3049 /* A new variable offset is created. Note that off_reg->off
3050 * == 0, since it's a scalar.
3051 * dst_reg gets the pointer type and since some positive
3052 * integer value was added to the pointer, give it a new 'id'
3053 * if it's a PTR_TO_PACKET.
3054 * this creates a new 'base' pointer, off_reg (variable) gets
3055 * added into the variable offset, and we copy the fixed offset
3058 if (signed_add_overflows(smin_ptr, smin_val) ||
3059 signed_add_overflows(smax_ptr, smax_val)) {
3060 dst_reg->smin_value = S64_MIN;
3061 dst_reg->smax_value = S64_MAX;
3063 dst_reg->smin_value = smin_ptr + smin_val;
3064 dst_reg->smax_value = smax_ptr + smax_val;
3066 if (umin_ptr + umin_val < umin_ptr ||
3067 umax_ptr + umax_val < umax_ptr) {
3068 dst_reg->umin_value = 0;
3069 dst_reg->umax_value = U64_MAX;
3071 dst_reg->umin_value = umin_ptr + umin_val;
3072 dst_reg->umax_value = umax_ptr + umax_val;
3074 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
3075 dst_reg->off = ptr_reg->off;
3076 if (reg_is_pkt_pointer(ptr_reg)) {
3077 dst_reg->id = ++env->id_gen;
3078 /* something was added to pkt_ptr, set range to zero */
3083 if (dst_reg == off_reg) {
3084 /* scalar -= pointer. Creates an unknown scalar */
3085 verbose(env, "R%d tried to subtract pointer from scalar\n",
3089 /* We don't allow subtraction from FP, because (according to
3090 * test_verifier.c test "invalid fp arithmetic", JITs might not
3091 * be able to deal with it.
3093 if (ptr_reg->type == PTR_TO_STACK) {
3094 verbose(env, "R%d subtraction from stack pointer prohibited\n",
3098 if (known && (ptr_reg->off - smin_val ==
3099 (s64)(s32)(ptr_reg->off - smin_val))) {
3100 /* pointer -= K. Subtract it from fixed offset */
3101 dst_reg->smin_value = smin_ptr;
3102 dst_reg->smax_value = smax_ptr;
3103 dst_reg->umin_value = umin_ptr;
3104 dst_reg->umax_value = umax_ptr;
3105 dst_reg->var_off = ptr_reg->var_off;
3106 dst_reg->id = ptr_reg->id;
3107 dst_reg->off = ptr_reg->off - smin_val;
3108 dst_reg->range = ptr_reg->range;
3111 /* A new variable offset is created. If the subtrahend is known
3112 * nonnegative, then any reg->range we had before is still good.
3114 if (signed_sub_overflows(smin_ptr, smax_val) ||
3115 signed_sub_overflows(smax_ptr, smin_val)) {
3116 /* Overflow possible, we know nothing */
3117 dst_reg->smin_value = S64_MIN;
3118 dst_reg->smax_value = S64_MAX;
3120 dst_reg->smin_value = smin_ptr - smax_val;
3121 dst_reg->smax_value = smax_ptr - smin_val;
3123 if (umin_ptr < umax_val) {
3124 /* Overflow possible, we know nothing */
3125 dst_reg->umin_value = 0;
3126 dst_reg->umax_value = U64_MAX;
3128 /* Cannot overflow (as long as bounds are consistent) */
3129 dst_reg->umin_value = umin_ptr - umax_val;
3130 dst_reg->umax_value = umax_ptr - umin_val;
3132 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
3133 dst_reg->off = ptr_reg->off;
3134 if (reg_is_pkt_pointer(ptr_reg)) {
3135 dst_reg->id = ++env->id_gen;
3136 /* something was added to pkt_ptr, set range to zero */
3144 /* bitwise ops on pointers are troublesome, prohibit. */
3145 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
3146 dst, bpf_alu_string[opcode >> 4]);
3149 /* other operators (e.g. MUL,LSH) produce non-pointer results */
3150 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
3151 dst, bpf_alu_string[opcode >> 4]);
3155 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
3158 __update_reg_bounds(dst_reg);
3159 __reg_deduce_bounds(dst_reg);
3160 __reg_bound_offset(dst_reg);
3164 /* WARNING: This function does calculations on 64-bit values, but the actual
3165 * execution may occur on 32-bit values. Therefore, things like bitshifts
3166 * need extra checks in the 32-bit case.
3168 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
3169 struct bpf_insn *insn,
3170 struct bpf_reg_state *dst_reg,
3171 struct bpf_reg_state src_reg)
3173 struct bpf_reg_state *regs = cur_regs(env);
3174 u8 opcode = BPF_OP(insn->code);
3175 bool src_known, dst_known;
3176 s64 smin_val, smax_val;
3177 u64 umin_val, umax_val;
3178 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
3180 if (insn_bitness == 32) {
3181 /* Relevant for 32-bit RSH: Information can propagate towards
3182 * LSB, so it isn't sufficient to only truncate the output to
3185 coerce_reg_to_size(dst_reg, 4);
3186 coerce_reg_to_size(&src_reg, 4);
3189 smin_val = src_reg.smin_value;
3190 smax_val = src_reg.smax_value;
3191 umin_val = src_reg.umin_value;
3192 umax_val = src_reg.umax_value;
3193 src_known = tnum_is_const(src_reg.var_off);
3194 dst_known = tnum_is_const(dst_reg->var_off);
3196 if ((src_known && (smin_val != smax_val || umin_val != umax_val)) ||
3197 smin_val > smax_val || umin_val > umax_val) {
3198 /* Taint dst register if offset had invalid bounds derived from
3199 * e.g. dead branches.
3201 __mark_reg_unknown(dst_reg);
3206 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
3207 __mark_reg_unknown(dst_reg);
3213 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
3214 signed_add_overflows(dst_reg->smax_value, smax_val)) {
3215 dst_reg->smin_value = S64_MIN;
3216 dst_reg->smax_value = S64_MAX;
3218 dst_reg->smin_value += smin_val;
3219 dst_reg->smax_value += smax_val;
3221 if (dst_reg->umin_value + umin_val < umin_val ||
3222 dst_reg->umax_value + umax_val < umax_val) {
3223 dst_reg->umin_value = 0;
3224 dst_reg->umax_value = U64_MAX;
3226 dst_reg->umin_value += umin_val;
3227 dst_reg->umax_value += umax_val;
3229 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
3232 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
3233 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
3234 /* Overflow possible, we know nothing */
3235 dst_reg->smin_value = S64_MIN;
3236 dst_reg->smax_value = S64_MAX;
3238 dst_reg->smin_value -= smax_val;
3239 dst_reg->smax_value -= smin_val;
3241 if (dst_reg->umin_value < umax_val) {
3242 /* Overflow possible, we know nothing */
3243 dst_reg->umin_value = 0;
3244 dst_reg->umax_value = U64_MAX;
3246 /* Cannot overflow (as long as bounds are consistent) */
3247 dst_reg->umin_value -= umax_val;
3248 dst_reg->umax_value -= umin_val;
3250 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
3253 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
3254 if (smin_val < 0 || dst_reg->smin_value < 0) {
3255 /* Ain't nobody got time to multiply that sign */
3256 __mark_reg_unbounded(dst_reg);
3257 __update_reg_bounds(dst_reg);
3260 /* Both values are positive, so we can work with unsigned and
3261 * copy the result to signed (unless it exceeds S64_MAX).
3263 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
3264 /* Potential overflow, we know nothing */
3265 __mark_reg_unbounded(dst_reg);
3266 /* (except what we can learn from the var_off) */
3267 __update_reg_bounds(dst_reg);
3270 dst_reg->umin_value *= umin_val;
3271 dst_reg->umax_value *= umax_val;
3272 if (dst_reg->umax_value > S64_MAX) {
3273 /* Overflow possible, we know nothing */
3274 dst_reg->smin_value = S64_MIN;
3275 dst_reg->smax_value = S64_MAX;
3277 dst_reg->smin_value = dst_reg->umin_value;
3278 dst_reg->smax_value = dst_reg->umax_value;
3282 if (src_known && dst_known) {
3283 __mark_reg_known(dst_reg, dst_reg->var_off.value &
3284 src_reg.var_off.value);
3287 /* We get our minimum from the var_off, since that's inherently
3288 * bitwise. Our maximum is the minimum of the operands' maxima.
3290 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
3291 dst_reg->umin_value = dst_reg->var_off.value;
3292 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
3293 if (dst_reg->smin_value < 0 || smin_val < 0) {
3294 /* Lose signed bounds when ANDing negative numbers,
3295 * ain't nobody got time for that.
3297 dst_reg->smin_value = S64_MIN;
3298 dst_reg->smax_value = S64_MAX;
3300 /* ANDing two positives gives a positive, so safe to
3301 * cast result into s64.
3303 dst_reg->smin_value = dst_reg->umin_value;
3304 dst_reg->smax_value = dst_reg->umax_value;
3306 /* We may learn something more from the var_off */
3307 __update_reg_bounds(dst_reg);
3310 if (src_known && dst_known) {
3311 __mark_reg_known(dst_reg, dst_reg->var_off.value |
3312 src_reg.var_off.value);
3315 /* We get our maximum from the var_off, and our minimum is the
3316 * maximum of the operands' minima
3318 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
3319 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
3320 dst_reg->umax_value = dst_reg->var_off.value |
3321 dst_reg->var_off.mask;
3322 if (dst_reg->smin_value < 0 || smin_val < 0) {
3323 /* Lose signed bounds when ORing negative numbers,
3324 * ain't nobody got time for that.
3326 dst_reg->smin_value = S64_MIN;
3327 dst_reg->smax_value = S64_MAX;
3329 /* ORing two positives gives a positive, so safe to
3330 * cast result into s64.
3332 dst_reg->smin_value = dst_reg->umin_value;
3333 dst_reg->smax_value = dst_reg->umax_value;
3335 /* We may learn something more from the var_off */
3336 __update_reg_bounds(dst_reg);
3339 if (umax_val >= insn_bitness) {
3340 /* Shifts greater than 31 or 63 are undefined.
3341 * This includes shifts by a negative number.
3343 mark_reg_unknown(env, regs, insn->dst_reg);
3346 /* We lose all sign bit information (except what we can pick
3349 dst_reg->smin_value = S64_MIN;
3350 dst_reg->smax_value = S64_MAX;
3351 /* If we might shift our top bit out, then we know nothing */
3352 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
3353 dst_reg->umin_value = 0;
3354 dst_reg->umax_value = U64_MAX;
3356 dst_reg->umin_value <<= umin_val;
3357 dst_reg->umax_value <<= umax_val;
3359 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
3360 /* We may learn something more from the var_off */
3361 __update_reg_bounds(dst_reg);
3364 if (umax_val >= insn_bitness) {
3365 /* Shifts greater than 31 or 63 are undefined.
3366 * This includes shifts by a negative number.
3368 mark_reg_unknown(env, regs, insn->dst_reg);
3371 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
3372 * be negative, then either:
3373 * 1) src_reg might be zero, so the sign bit of the result is
3374 * unknown, so we lose our signed bounds
3375 * 2) it's known negative, thus the unsigned bounds capture the
3377 * 3) the signed bounds cross zero, so they tell us nothing
3379 * If the value in dst_reg is known nonnegative, then again the
3380 * unsigned bounts capture the signed bounds.
3381 * Thus, in all cases it suffices to blow away our signed bounds
3382 * and rely on inferring new ones from the unsigned bounds and
3383 * var_off of the result.
3385 dst_reg->smin_value = S64_MIN;
3386 dst_reg->smax_value = S64_MAX;
3387 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
3388 dst_reg->umin_value >>= umax_val;
3389 dst_reg->umax_value >>= umin_val;
3390 /* We may learn something more from the var_off */
3391 __update_reg_bounds(dst_reg);
3394 if (umax_val >= insn_bitness) {
3395 /* Shifts greater than 31 or 63 are undefined.
3396 * This includes shifts by a negative number.
3398 mark_reg_unknown(env, regs, insn->dst_reg);
3402 /* Upon reaching here, src_known is true and
3403 * umax_val is equal to umin_val.
3405 dst_reg->smin_value >>= umin_val;
3406 dst_reg->smax_value >>= umin_val;
3407 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val);
3409 /* blow away the dst_reg umin_value/umax_value and rely on
3410 * dst_reg var_off to refine the result.
3412 dst_reg->umin_value = 0;
3413 dst_reg->umax_value = U64_MAX;
3414 __update_reg_bounds(dst_reg);
3417 mark_reg_unknown(env, regs, insn->dst_reg);
3421 if (BPF_CLASS(insn->code) != BPF_ALU64) {
3422 /* 32-bit ALU ops are (32,32)->32 */
3423 coerce_reg_to_size(dst_reg, 4);
3426 __reg_deduce_bounds(dst_reg);
3427 __reg_bound_offset(dst_reg);
3431 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
3434 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
3435 struct bpf_insn *insn)
3437 struct bpf_verifier_state *vstate = env->cur_state;
3438 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3439 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
3440 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
3441 u8 opcode = BPF_OP(insn->code);
3443 dst_reg = ®s[insn->dst_reg];
3445 if (dst_reg->type != SCALAR_VALUE)
3447 if (BPF_SRC(insn->code) == BPF_X) {
3448 src_reg = ®s[insn->src_reg];
3449 if (src_reg->type != SCALAR_VALUE) {
3450 if (dst_reg->type != SCALAR_VALUE) {
3451 /* Combining two pointers by any ALU op yields
3452 * an arbitrary scalar. Disallow all math except
3453 * pointer subtraction
3455 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
3456 mark_reg_unknown(env, regs, insn->dst_reg);
3459 verbose(env, "R%d pointer %s pointer prohibited\n",
3461 bpf_alu_string[opcode >> 4]);
3464 /* scalar += pointer
3465 * This is legal, but we have to reverse our
3466 * src/dest handling in computing the range
3468 return adjust_ptr_min_max_vals(env, insn,
3471 } else if (ptr_reg) {
3472 /* pointer += scalar */
3473 return adjust_ptr_min_max_vals(env, insn,
3477 /* Pretend the src is a reg with a known value, since we only
3478 * need to be able to read from this state.
3480 off_reg.type = SCALAR_VALUE;
3481 __mark_reg_known(&off_reg, insn->imm);
3483 if (ptr_reg) /* pointer += K */
3484 return adjust_ptr_min_max_vals(env, insn,
3488 /* Got here implies adding two SCALAR_VALUEs */
3489 if (WARN_ON_ONCE(ptr_reg)) {
3490 print_verifier_state(env, state);
3491 verbose(env, "verifier internal error: unexpected ptr_reg\n");
3494 if (WARN_ON(!src_reg)) {
3495 print_verifier_state(env, state);
3496 verbose(env, "verifier internal error: no src_reg\n");
3499 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
3502 /* check validity of 32-bit and 64-bit arithmetic operations */
3503 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
3505 struct bpf_reg_state *regs = cur_regs(env);
3506 u8 opcode = BPF_OP(insn->code);
3509 if (opcode == BPF_END || opcode == BPF_NEG) {
3510 if (opcode == BPF_NEG) {
3511 if (BPF_SRC(insn->code) != 0 ||
3512 insn->src_reg != BPF_REG_0 ||
3513 insn->off != 0 || insn->imm != 0) {
3514 verbose(env, "BPF_NEG uses reserved fields\n");
3518 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
3519 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
3520 BPF_CLASS(insn->code) == BPF_ALU64) {
3521 verbose(env, "BPF_END uses reserved fields\n");
3526 /* check src operand */
3527 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3531 if (is_pointer_value(env, insn->dst_reg)) {
3532 verbose(env, "R%d pointer arithmetic prohibited\n",
3537 /* check dest operand */
3538 err = check_reg_arg(env, insn->dst_reg, DST_OP);
3542 } else if (opcode == BPF_MOV) {
3544 if (BPF_SRC(insn->code) == BPF_X) {
3545 if (insn->imm != 0 || insn->off != 0) {
3546 verbose(env, "BPF_MOV uses reserved fields\n");
3550 /* check src operand */
3551 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3555 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
3556 verbose(env, "BPF_MOV uses reserved fields\n");
3561 /* check dest operand, mark as required later */
3562 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
3566 if (BPF_SRC(insn->code) == BPF_X) {
3567 if (BPF_CLASS(insn->code) == BPF_ALU64) {
3569 * copy register state to dest reg
3571 regs[insn->dst_reg] = regs[insn->src_reg];
3572 regs[insn->dst_reg].live |= REG_LIVE_WRITTEN;
3575 if (is_pointer_value(env, insn->src_reg)) {
3577 "R%d partial copy of pointer\n",
3581 mark_reg_unknown(env, regs, insn->dst_reg);
3582 coerce_reg_to_size(®s[insn->dst_reg], 4);
3586 * remember the value we stored into this reg
3588 /* clear any state __mark_reg_known doesn't set */
3589 mark_reg_unknown(env, regs, insn->dst_reg);
3590 regs[insn->dst_reg].type = SCALAR_VALUE;
3591 if (BPF_CLASS(insn->code) == BPF_ALU64) {
3592 __mark_reg_known(regs + insn->dst_reg,
3595 __mark_reg_known(regs + insn->dst_reg,
3600 } else if (opcode > BPF_END) {
3601 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
3604 } else { /* all other ALU ops: and, sub, xor, add, ... */
3606 if (BPF_SRC(insn->code) == BPF_X) {
3607 if (insn->imm != 0 || insn->off != 0) {
3608 verbose(env, "BPF_ALU uses reserved fields\n");
3611 /* check src1 operand */
3612 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3616 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
3617 verbose(env, "BPF_ALU uses reserved fields\n");
3622 /* check src2 operand */
3623 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3627 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
3628 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
3629 verbose(env, "div by zero\n");
3633 if (opcode == BPF_ARSH && BPF_CLASS(insn->code) != BPF_ALU64) {
3634 verbose(env, "BPF_ARSH not supported for 32 bit ALU\n");
3638 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
3639 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
3640 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
3642 if (insn->imm < 0 || insn->imm >= size) {
3643 verbose(env, "invalid shift %d\n", insn->imm);
3648 /* check dest operand */
3649 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
3653 return adjust_reg_min_max_vals(env, insn);
3659 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
3660 struct bpf_reg_state *dst_reg,
3661 enum bpf_reg_type type,
3662 bool range_right_open)
3664 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3665 struct bpf_reg_state *regs = state->regs, *reg;
3669 if (dst_reg->off < 0 ||
3670 (dst_reg->off == 0 && range_right_open))
3671 /* This doesn't give us any range */
3674 if (dst_reg->umax_value > MAX_PACKET_OFF ||
3675 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
3676 /* Risk of overflow. For instance, ptr + (1<<63) may be less
3677 * than pkt_end, but that's because it's also less than pkt.
3681 new_range = dst_reg->off;
3682 if (range_right_open)
3685 /* Examples for register markings:
3687 * pkt_data in dst register:
3691 * if (r2 > pkt_end) goto <handle exception>
3696 * if (r2 < pkt_end) goto <access okay>
3697 * <handle exception>
3700 * r2 == dst_reg, pkt_end == src_reg
3701 * r2=pkt(id=n,off=8,r=0)
3702 * r3=pkt(id=n,off=0,r=0)
3704 * pkt_data in src register:
3708 * if (pkt_end >= r2) goto <access okay>
3709 * <handle exception>
3713 * if (pkt_end <= r2) goto <handle exception>
3717 * pkt_end == dst_reg, r2 == src_reg
3718 * r2=pkt(id=n,off=8,r=0)
3719 * r3=pkt(id=n,off=0,r=0)
3721 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
3722 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
3723 * and [r3, r3 + 8-1) respectively is safe to access depending on
3727 /* If our ids match, then we must have the same max_value. And we
3728 * don't care about the other reg's fixed offset, since if it's too big
3729 * the range won't allow anything.
3730 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
3732 for (i = 0; i < MAX_BPF_REG; i++)
3733 if (regs[i].type == type && regs[i].id == dst_reg->id)
3734 /* keep the maximum range already checked */
3735 regs[i].range = max(regs[i].range, new_range);
3737 for (j = 0; j <= vstate->curframe; j++) {
3738 state = vstate->frame[j];
3739 bpf_for_each_spilled_reg(i, state, reg) {
3742 if (reg->type == type && reg->id == dst_reg->id)
3743 reg->range = max(reg->range, new_range);
3748 /* Adjusts the register min/max values in the case that the dst_reg is the
3749 * variable register that we are working on, and src_reg is a constant or we're
3750 * simply doing a BPF_K check.
3751 * In JEQ/JNE cases we also adjust the var_off values.
3753 static void reg_set_min_max(struct bpf_reg_state *true_reg,
3754 struct bpf_reg_state *false_reg, u64 val,
3757 /* If the dst_reg is a pointer, we can't learn anything about its
3758 * variable offset from the compare (unless src_reg were a pointer into
3759 * the same object, but we don't bother with that.
3760 * Since false_reg and true_reg have the same type by construction, we
3761 * only need to check one of them for pointerness.
3763 if (__is_pointer_value(false, false_reg))
3768 /* If this is false then we know nothing Jon Snow, but if it is
3769 * true then we know for sure.
3771 __mark_reg_known(true_reg, val);
3774 /* If this is true we know nothing Jon Snow, but if it is false
3775 * we know the value for sure;
3777 __mark_reg_known(false_reg, val);
3780 false_reg->umax_value = min(false_reg->umax_value, val);
3781 true_reg->umin_value = max(true_reg->umin_value, val + 1);
3784 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
3785 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
3788 false_reg->umin_value = max(false_reg->umin_value, val);
3789 true_reg->umax_value = min(true_reg->umax_value, val - 1);
3792 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
3793 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
3796 false_reg->umax_value = min(false_reg->umax_value, val - 1);
3797 true_reg->umin_value = max(true_reg->umin_value, val);
3800 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
3801 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
3804 false_reg->umin_value = max(false_reg->umin_value, val + 1);
3805 true_reg->umax_value = min(true_reg->umax_value, val);
3808 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
3809 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
3815 __reg_deduce_bounds(false_reg);
3816 __reg_deduce_bounds(true_reg);
3817 /* We might have learned some bits from the bounds. */
3818 __reg_bound_offset(false_reg);
3819 __reg_bound_offset(true_reg);
3820 /* Intersecting with the old var_off might have improved our bounds
3821 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
3822 * then new var_off is (0; 0x7f...fc) which improves our umax.
3824 __update_reg_bounds(false_reg);
3825 __update_reg_bounds(true_reg);
3828 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
3831 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
3832 struct bpf_reg_state *false_reg, u64 val,
3835 if (__is_pointer_value(false, false_reg))
3840 /* If this is false then we know nothing Jon Snow, but if it is
3841 * true then we know for sure.
3843 __mark_reg_known(true_reg, val);
3846 /* If this is true we know nothing Jon Snow, but if it is false
3847 * we know the value for sure;
3849 __mark_reg_known(false_reg, val);
3852 true_reg->umax_value = min(true_reg->umax_value, val - 1);
3853 false_reg->umin_value = max(false_reg->umin_value, val);
3856 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
3857 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
3860 true_reg->umin_value = max(true_reg->umin_value, val + 1);
3861 false_reg->umax_value = min(false_reg->umax_value, val);
3864 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
3865 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
3868 true_reg->umax_value = min(true_reg->umax_value, val);
3869 false_reg->umin_value = max(false_reg->umin_value, val + 1);
3872 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
3873 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
3876 true_reg->umin_value = max(true_reg->umin_value, val);
3877 false_reg->umax_value = min(false_reg->umax_value, val - 1);
3880 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
3881 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
3887 __reg_deduce_bounds(false_reg);
3888 __reg_deduce_bounds(true_reg);
3889 /* We might have learned some bits from the bounds. */
3890 __reg_bound_offset(false_reg);
3891 __reg_bound_offset(true_reg);
3892 /* Intersecting with the old var_off might have improved our bounds
3893 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
3894 * then new var_off is (0; 0x7f...fc) which improves our umax.
3896 __update_reg_bounds(false_reg);
3897 __update_reg_bounds(true_reg);
3900 /* Regs are known to be equal, so intersect their min/max/var_off */
3901 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
3902 struct bpf_reg_state *dst_reg)
3904 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
3905 dst_reg->umin_value);
3906 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
3907 dst_reg->umax_value);
3908 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
3909 dst_reg->smin_value);
3910 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
3911 dst_reg->smax_value);
3912 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
3914 /* We might have learned new bounds from the var_off. */
3915 __update_reg_bounds(src_reg);
3916 __update_reg_bounds(dst_reg);
3917 /* We might have learned something about the sign bit. */
3918 __reg_deduce_bounds(src_reg);
3919 __reg_deduce_bounds(dst_reg);
3920 /* We might have learned some bits from the bounds. */
3921 __reg_bound_offset(src_reg);
3922 __reg_bound_offset(dst_reg);
3923 /* Intersecting with the old var_off might have improved our bounds
3924 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
3925 * then new var_off is (0; 0x7f...fc) which improves our umax.
3927 __update_reg_bounds(src_reg);
3928 __update_reg_bounds(dst_reg);
3931 static void reg_combine_min_max(struct bpf_reg_state *true_src,
3932 struct bpf_reg_state *true_dst,
3933 struct bpf_reg_state *false_src,
3934 struct bpf_reg_state *false_dst,
3939 __reg_combine_min_max(true_src, true_dst);
3942 __reg_combine_min_max(false_src, false_dst);
3947 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
3948 struct bpf_reg_state *reg, u32 id,
3951 if (reg_type_may_be_null(reg->type) && reg->id == id) {
3952 /* Old offset (both fixed and variable parts) should
3953 * have been known-zero, because we don't allow pointer
3954 * arithmetic on pointers that might be NULL.
3956 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
3957 !tnum_equals_const(reg->var_off, 0) ||
3959 __mark_reg_known_zero(reg);
3963 reg->type = SCALAR_VALUE;
3964 } else if (reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
3965 if (reg->map_ptr->inner_map_meta) {
3966 reg->type = CONST_PTR_TO_MAP;
3967 reg->map_ptr = reg->map_ptr->inner_map_meta;
3969 reg->type = PTR_TO_MAP_VALUE;
3971 } else if (reg->type == PTR_TO_SOCKET_OR_NULL) {
3972 reg->type = PTR_TO_SOCKET;
3974 if (is_null || !reg_is_refcounted(reg)) {
3975 /* We don't need id from this point onwards anymore,
3976 * thus we should better reset it, so that state
3977 * pruning has chances to take effect.
3984 /* The logic is similar to find_good_pkt_pointers(), both could eventually
3985 * be folded together at some point.
3987 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
3990 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3991 struct bpf_reg_state *reg, *regs = state->regs;
3992 u32 id = regs[regno].id;
3995 if (reg_is_refcounted_or_null(®s[regno]) && is_null)
3996 __release_reference_state(state, id);
3998 for (i = 0; i < MAX_BPF_REG; i++)
3999 mark_ptr_or_null_reg(state, ®s[i], id, is_null);
4001 for (j = 0; j <= vstate->curframe; j++) {
4002 state = vstate->frame[j];
4003 bpf_for_each_spilled_reg(i, state, reg) {
4006 mark_ptr_or_null_reg(state, reg, id, is_null);
4011 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
4012 struct bpf_reg_state *dst_reg,
4013 struct bpf_reg_state *src_reg,
4014 struct bpf_verifier_state *this_branch,
4015 struct bpf_verifier_state *other_branch)
4017 if (BPF_SRC(insn->code) != BPF_X)
4020 switch (BPF_OP(insn->code)) {
4022 if ((dst_reg->type == PTR_TO_PACKET &&
4023 src_reg->type == PTR_TO_PACKET_END) ||
4024 (dst_reg->type == PTR_TO_PACKET_META &&
4025 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
4026 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
4027 find_good_pkt_pointers(this_branch, dst_reg,
4028 dst_reg->type, false);
4029 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
4030 src_reg->type == PTR_TO_PACKET) ||
4031 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
4032 src_reg->type == PTR_TO_PACKET_META)) {
4033 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
4034 find_good_pkt_pointers(other_branch, src_reg,
4035 src_reg->type, true);
4041 if ((dst_reg->type == PTR_TO_PACKET &&
4042 src_reg->type == PTR_TO_PACKET_END) ||
4043 (dst_reg->type == PTR_TO_PACKET_META &&
4044 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
4045 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
4046 find_good_pkt_pointers(other_branch, dst_reg,
4047 dst_reg->type, true);
4048 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
4049 src_reg->type == PTR_TO_PACKET) ||
4050 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
4051 src_reg->type == PTR_TO_PACKET_META)) {
4052 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
4053 find_good_pkt_pointers(this_branch, src_reg,
4054 src_reg->type, false);
4060 if ((dst_reg->type == PTR_TO_PACKET &&
4061 src_reg->type == PTR_TO_PACKET_END) ||
4062 (dst_reg->type == PTR_TO_PACKET_META &&
4063 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
4064 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
4065 find_good_pkt_pointers(this_branch, dst_reg,
4066 dst_reg->type, true);
4067 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
4068 src_reg->type == PTR_TO_PACKET) ||
4069 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
4070 src_reg->type == PTR_TO_PACKET_META)) {
4071 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
4072 find_good_pkt_pointers(other_branch, src_reg,
4073 src_reg->type, false);
4079 if ((dst_reg->type == PTR_TO_PACKET &&
4080 src_reg->type == PTR_TO_PACKET_END) ||
4081 (dst_reg->type == PTR_TO_PACKET_META &&
4082 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
4083 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
4084 find_good_pkt_pointers(other_branch, dst_reg,
4085 dst_reg->type, false);
4086 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
4087 src_reg->type == PTR_TO_PACKET) ||
4088 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
4089 src_reg->type == PTR_TO_PACKET_META)) {
4090 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
4091 find_good_pkt_pointers(this_branch, src_reg,
4092 src_reg->type, true);
4104 static int check_cond_jmp_op(struct bpf_verifier_env *env,
4105 struct bpf_insn *insn, int *insn_idx)
4107 struct bpf_verifier_state *this_branch = env->cur_state;
4108 struct bpf_verifier_state *other_branch;
4109 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
4110 struct bpf_reg_state *dst_reg, *other_branch_regs;
4111 u8 opcode = BPF_OP(insn->code);
4114 if (opcode > BPF_JSLE) {
4115 verbose(env, "invalid BPF_JMP opcode %x\n", opcode);
4119 if (BPF_SRC(insn->code) == BPF_X) {
4120 if (insn->imm != 0) {
4121 verbose(env, "BPF_JMP uses reserved fields\n");
4125 /* check src1 operand */
4126 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4130 if (is_pointer_value(env, insn->src_reg)) {
4131 verbose(env, "R%d pointer comparison prohibited\n",
4136 if (insn->src_reg != BPF_REG_0) {
4137 verbose(env, "BPF_JMP uses reserved fields\n");
4142 /* check src2 operand */
4143 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4147 dst_reg = ®s[insn->dst_reg];
4149 /* detect if R == 0 where R was initialized to zero earlier */
4150 if (BPF_SRC(insn->code) == BPF_K &&
4151 (opcode == BPF_JEQ || opcode == BPF_JNE) &&
4152 dst_reg->type == SCALAR_VALUE &&
4153 tnum_is_const(dst_reg->var_off)) {
4154 if ((opcode == BPF_JEQ && dst_reg->var_off.value == insn->imm) ||
4155 (opcode == BPF_JNE && dst_reg->var_off.value != insn->imm)) {
4156 /* if (imm == imm) goto pc+off;
4157 * only follow the goto, ignore fall-through
4159 *insn_idx += insn->off;
4162 /* if (imm != imm) goto pc+off;
4163 * only follow fall-through branch, since
4164 * that's where the program will go
4170 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx);
4173 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
4175 /* detect if we are comparing against a constant value so we can adjust
4176 * our min/max values for our dst register.
4177 * this is only legit if both are scalars (or pointers to the same
4178 * object, I suppose, but we don't support that right now), because
4179 * otherwise the different base pointers mean the offsets aren't
4182 if (BPF_SRC(insn->code) == BPF_X) {
4183 if (dst_reg->type == SCALAR_VALUE &&
4184 regs[insn->src_reg].type == SCALAR_VALUE) {
4185 if (tnum_is_const(regs[insn->src_reg].var_off))
4186 reg_set_min_max(&other_branch_regs[insn->dst_reg],
4187 dst_reg, regs[insn->src_reg].var_off.value,
4189 else if (tnum_is_const(dst_reg->var_off))
4190 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
4191 ®s[insn->src_reg],
4192 dst_reg->var_off.value, opcode);
4193 else if (opcode == BPF_JEQ || opcode == BPF_JNE)
4194 /* Comparing for equality, we can combine knowledge */
4195 reg_combine_min_max(&other_branch_regs[insn->src_reg],
4196 &other_branch_regs[insn->dst_reg],
4197 ®s[insn->src_reg],
4198 ®s[insn->dst_reg], opcode);
4200 } else if (dst_reg->type == SCALAR_VALUE) {
4201 reg_set_min_max(&other_branch_regs[insn->dst_reg],
4202 dst_reg, insn->imm, opcode);
4205 /* detect if R == 0 where R is returned from bpf_map_lookup_elem() */
4206 if (BPF_SRC(insn->code) == BPF_K &&
4207 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
4208 reg_type_may_be_null(dst_reg->type)) {
4209 /* Mark all identical registers in each branch as either
4210 * safe or unknown depending R == 0 or R != 0 conditional.
4212 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
4214 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
4216 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
4217 this_branch, other_branch) &&
4218 is_pointer_value(env, insn->dst_reg)) {
4219 verbose(env, "R%d pointer comparison prohibited\n",
4224 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
4228 /* return the map pointer stored inside BPF_LD_IMM64 instruction */
4229 static struct bpf_map *ld_imm64_to_map_ptr(struct bpf_insn *insn)
4231 u64 imm64 = ((u64) (u32) insn[0].imm) | ((u64) (u32) insn[1].imm) << 32;
4233 return (struct bpf_map *) (unsigned long) imm64;
4236 /* verify BPF_LD_IMM64 instruction */
4237 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
4239 struct bpf_reg_state *regs = cur_regs(env);
4242 if (BPF_SIZE(insn->code) != BPF_DW) {
4243 verbose(env, "invalid BPF_LD_IMM insn\n");
4246 if (insn->off != 0) {
4247 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
4251 err = check_reg_arg(env, insn->dst_reg, DST_OP);
4255 if (insn->src_reg == 0) {
4256 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
4258 regs[insn->dst_reg].type = SCALAR_VALUE;
4259 __mark_reg_known(®s[insn->dst_reg], imm);
4263 /* replace_map_fd_with_map_ptr() should have caught bad ld_imm64 */
4264 BUG_ON(insn->src_reg != BPF_PSEUDO_MAP_FD);
4266 regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
4267 regs[insn->dst_reg].map_ptr = ld_imm64_to_map_ptr(insn);
4271 static bool may_access_skb(enum bpf_prog_type type)
4274 case BPF_PROG_TYPE_SOCKET_FILTER:
4275 case BPF_PROG_TYPE_SCHED_CLS:
4276 case BPF_PROG_TYPE_SCHED_ACT:
4283 /* verify safety of LD_ABS|LD_IND instructions:
4284 * - they can only appear in the programs where ctx == skb
4285 * - since they are wrappers of function calls, they scratch R1-R5 registers,
4286 * preserve R6-R9, and store return value into R0
4289 * ctx == skb == R6 == CTX
4292 * SRC == any register
4293 * IMM == 32-bit immediate
4296 * R0 - 8/16/32-bit skb data converted to cpu endianness
4298 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
4300 struct bpf_reg_state *regs = cur_regs(env);
4301 u8 mode = BPF_MODE(insn->code);
4304 if (!may_access_skb(env->prog->type)) {
4305 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
4309 if (!env->ops->gen_ld_abs) {
4310 verbose(env, "bpf verifier is misconfigured\n");
4314 if (env->subprog_cnt > 1) {
4315 /* when program has LD_ABS insn JITs and interpreter assume
4316 * that r1 == ctx == skb which is not the case for callees
4317 * that can have arbitrary arguments. It's problematic
4318 * for main prog as well since JITs would need to analyze
4319 * all functions in order to make proper register save/restore
4320 * decisions in the main prog. Hence disallow LD_ABS with calls
4322 verbose(env, "BPF_LD_[ABS|IND] instructions cannot be mixed with bpf-to-bpf calls\n");
4326 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
4327 BPF_SIZE(insn->code) == BPF_DW ||
4328 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
4329 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
4333 /* check whether implicit source operand (register R6) is readable */
4334 err = check_reg_arg(env, BPF_REG_6, SRC_OP);
4338 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
4339 * gen_ld_abs() may terminate the program at runtime, leading to
4342 err = check_reference_leak(env);
4344 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
4348 if (regs[BPF_REG_6].type != PTR_TO_CTX) {
4350 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
4354 if (mode == BPF_IND) {
4355 /* check explicit source operand */
4356 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4361 /* reset caller saved regs to unreadable */
4362 for (i = 0; i < CALLER_SAVED_REGS; i++) {
4363 mark_reg_not_init(env, regs, caller_saved[i]);
4364 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
4367 /* mark destination R0 register as readable, since it contains
4368 * the value fetched from the packet.
4369 * Already marked as written above.
4371 mark_reg_unknown(env, regs, BPF_REG_0);
4375 static int check_return_code(struct bpf_verifier_env *env)
4377 struct bpf_reg_state *reg;
4378 struct tnum range = tnum_range(0, 1);
4380 switch (env->prog->type) {
4381 case BPF_PROG_TYPE_CGROUP_SKB:
4382 case BPF_PROG_TYPE_CGROUP_SOCK:
4383 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
4384 case BPF_PROG_TYPE_SOCK_OPS:
4385 case BPF_PROG_TYPE_CGROUP_DEVICE:
4391 reg = cur_regs(env) + BPF_REG_0;
4392 if (reg->type != SCALAR_VALUE) {
4393 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
4394 reg_type_str[reg->type]);
4398 if (!tnum_in(range, reg->var_off)) {
4399 verbose(env, "At program exit the register R0 ");
4400 if (!tnum_is_unknown(reg->var_off)) {
4403 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4404 verbose(env, "has value %s", tn_buf);
4406 verbose(env, "has unknown scalar value");
4408 verbose(env, " should have been 0 or 1\n");
4414 /* non-recursive DFS pseudo code
4415 * 1 procedure DFS-iterative(G,v):
4416 * 2 label v as discovered
4417 * 3 let S be a stack
4419 * 5 while S is not empty
4421 * 7 if t is what we're looking for:
4423 * 9 for all edges e in G.adjacentEdges(t) do
4424 * 10 if edge e is already labelled
4425 * 11 continue with the next edge
4426 * 12 w <- G.adjacentVertex(t,e)
4427 * 13 if vertex w is not discovered and not explored
4428 * 14 label e as tree-edge
4429 * 15 label w as discovered
4432 * 18 else if vertex w is discovered
4433 * 19 label e as back-edge
4435 * 21 // vertex w is explored
4436 * 22 label e as forward- or cross-edge
4437 * 23 label t as explored
4442 * 0x11 - discovered and fall-through edge labelled
4443 * 0x12 - discovered and fall-through and branch edges labelled
4454 #define STATE_LIST_MARK ((struct bpf_verifier_state_list *) -1L)
4456 static int *insn_stack; /* stack of insns to process */
4457 static int cur_stack; /* current stack index */
4458 static int *insn_state;
4460 /* t, w, e - match pseudo-code above:
4461 * t - index of current instruction
4462 * w - next instruction
4465 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env)
4467 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
4470 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
4473 if (w < 0 || w >= env->prog->len) {
4474 verbose(env, "jump out of range from insn %d to %d\n", t, w);
4479 /* mark branch target for state pruning */
4480 env->explored_states[w] = STATE_LIST_MARK;
4482 if (insn_state[w] == 0) {
4484 insn_state[t] = DISCOVERED | e;
4485 insn_state[w] = DISCOVERED;
4486 if (cur_stack >= env->prog->len)
4488 insn_stack[cur_stack++] = w;
4490 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
4491 verbose(env, "back-edge from insn %d to %d\n", t, w);
4493 } else if (insn_state[w] == EXPLORED) {
4494 /* forward- or cross-edge */
4495 insn_state[t] = DISCOVERED | e;
4497 verbose(env, "insn state internal bug\n");
4503 /* non-recursive depth-first-search to detect loops in BPF program
4504 * loop == back-edge in directed graph
4506 static int check_cfg(struct bpf_verifier_env *env)
4508 struct bpf_insn *insns = env->prog->insnsi;
4509 int insn_cnt = env->prog->len;
4513 ret = check_subprogs(env);
4517 insn_state = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
4521 insn_stack = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
4527 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
4528 insn_stack[0] = 0; /* 0 is the first instruction */
4534 t = insn_stack[cur_stack - 1];
4536 if (BPF_CLASS(insns[t].code) == BPF_JMP) {
4537 u8 opcode = BPF_OP(insns[t].code);
4539 if (opcode == BPF_EXIT) {
4541 } else if (opcode == BPF_CALL) {
4542 ret = push_insn(t, t + 1, FALLTHROUGH, env);
4547 if (t + 1 < insn_cnt)
4548 env->explored_states[t + 1] = STATE_LIST_MARK;
4549 if (insns[t].src_reg == BPF_PSEUDO_CALL) {
4550 env->explored_states[t] = STATE_LIST_MARK;
4551 ret = push_insn(t, t + insns[t].imm + 1, BRANCH, env);
4557 } else if (opcode == BPF_JA) {
4558 if (BPF_SRC(insns[t].code) != BPF_K) {
4562 /* unconditional jump with single edge */
4563 ret = push_insn(t, t + insns[t].off + 1,
4569 /* tell verifier to check for equivalent states
4570 * after every call and jump
4572 if (t + 1 < insn_cnt)
4573 env->explored_states[t + 1] = STATE_LIST_MARK;
4575 /* conditional jump with two edges */
4576 env->explored_states[t] = STATE_LIST_MARK;
4577 ret = push_insn(t, t + 1, FALLTHROUGH, env);
4583 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env);
4590 /* all other non-branch instructions with single
4593 ret = push_insn(t, t + 1, FALLTHROUGH, env);
4601 insn_state[t] = EXPLORED;
4602 if (cur_stack-- <= 0) {
4603 verbose(env, "pop stack internal bug\n");
4610 for (i = 0; i < insn_cnt; i++) {
4611 if (insn_state[i] != EXPLORED) {
4612 verbose(env, "unreachable insn %d\n", i);
4617 ret = 0; /* cfg looks good */
4625 /* check %cur's range satisfies %old's */
4626 static bool range_within(struct bpf_reg_state *old,
4627 struct bpf_reg_state *cur)
4629 return old->umin_value <= cur->umin_value &&
4630 old->umax_value >= cur->umax_value &&
4631 old->smin_value <= cur->smin_value &&
4632 old->smax_value >= cur->smax_value;
4635 /* Maximum number of register states that can exist at once */
4636 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
4642 /* If in the old state two registers had the same id, then they need to have
4643 * the same id in the new state as well. But that id could be different from
4644 * the old state, so we need to track the mapping from old to new ids.
4645 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
4646 * regs with old id 5 must also have new id 9 for the new state to be safe. But
4647 * regs with a different old id could still have new id 9, we don't care about
4649 * So we look through our idmap to see if this old id has been seen before. If
4650 * so, we require the new id to match; otherwise, we add the id pair to the map.
4652 static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
4656 for (i = 0; i < ID_MAP_SIZE; i++) {
4657 if (!idmap[i].old) {
4658 /* Reached an empty slot; haven't seen this id before */
4659 idmap[i].old = old_id;
4660 idmap[i].cur = cur_id;
4663 if (idmap[i].old == old_id)
4664 return idmap[i].cur == cur_id;
4666 /* We ran out of idmap slots, which should be impossible */
4671 /* Returns true if (rold safe implies rcur safe) */
4672 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
4673 struct idpair *idmap)
4677 if (!(rold->live & REG_LIVE_READ))
4678 /* explored state didn't use this */
4681 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
4683 if (rold->type == PTR_TO_STACK)
4684 /* two stack pointers are equal only if they're pointing to
4685 * the same stack frame, since fp-8 in foo != fp-8 in bar
4687 return equal && rold->frameno == rcur->frameno;
4692 if (rold->type == NOT_INIT)
4693 /* explored state can't have used this */
4695 if (rcur->type == NOT_INIT)
4697 switch (rold->type) {
4699 if (rcur->type == SCALAR_VALUE) {
4700 /* new val must satisfy old val knowledge */
4701 return range_within(rold, rcur) &&
4702 tnum_in(rold->var_off, rcur->var_off);
4704 /* We're trying to use a pointer in place of a scalar.
4705 * Even if the scalar was unbounded, this could lead to
4706 * pointer leaks because scalars are allowed to leak
4707 * while pointers are not. We could make this safe in
4708 * special cases if root is calling us, but it's
4709 * probably not worth the hassle.
4713 case PTR_TO_MAP_VALUE:
4714 /* If the new min/max/var_off satisfy the old ones and
4715 * everything else matches, we are OK.
4716 * We don't care about the 'id' value, because nothing
4717 * uses it for PTR_TO_MAP_VALUE (only for ..._OR_NULL)
4719 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
4720 range_within(rold, rcur) &&
4721 tnum_in(rold->var_off, rcur->var_off);
4722 case PTR_TO_MAP_VALUE_OR_NULL:
4723 /* a PTR_TO_MAP_VALUE could be safe to use as a
4724 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
4725 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
4726 * checked, doing so could have affected others with the same
4727 * id, and we can't check for that because we lost the id when
4728 * we converted to a PTR_TO_MAP_VALUE.
4730 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
4732 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
4734 /* Check our ids match any regs they're supposed to */
4735 return check_ids(rold->id, rcur->id, idmap);
4736 case PTR_TO_PACKET_META:
4738 if (rcur->type != rold->type)
4740 /* We must have at least as much range as the old ptr
4741 * did, so that any accesses which were safe before are
4742 * still safe. This is true even if old range < old off,
4743 * since someone could have accessed through (ptr - k), or
4744 * even done ptr -= k in a register, to get a safe access.
4746 if (rold->range > rcur->range)
4748 /* If the offsets don't match, we can't trust our alignment;
4749 * nor can we be sure that we won't fall out of range.
4751 if (rold->off != rcur->off)
4753 /* id relations must be preserved */
4754 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
4756 /* new val must satisfy old val knowledge */
4757 return range_within(rold, rcur) &&
4758 tnum_in(rold->var_off, rcur->var_off);
4760 case CONST_PTR_TO_MAP:
4761 case PTR_TO_PACKET_END:
4762 case PTR_TO_FLOW_KEYS:
4764 case PTR_TO_SOCKET_OR_NULL:
4765 /* Only valid matches are exact, which memcmp() above
4766 * would have accepted
4769 /* Don't know what's going on, just say it's not safe */
4773 /* Shouldn't get here; if we do, say it's not safe */
4778 static bool stacksafe(struct bpf_func_state *old,
4779 struct bpf_func_state *cur,
4780 struct idpair *idmap)
4784 /* if explored stack has more populated slots than current stack
4785 * such stacks are not equivalent
4787 if (old->allocated_stack > cur->allocated_stack)
4790 /* walk slots of the explored stack and ignore any additional
4791 * slots in the current stack, since explored(safe) state
4794 for (i = 0; i < old->allocated_stack; i++) {
4795 spi = i / BPF_REG_SIZE;
4797 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ))
4798 /* explored state didn't use this */
4801 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
4803 /* if old state was safe with misc data in the stack
4804 * it will be safe with zero-initialized stack.
4805 * The opposite is not true
4807 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
4808 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
4810 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
4811 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
4812 /* Ex: old explored (safe) state has STACK_SPILL in
4813 * this stack slot, but current has has STACK_MISC ->
4814 * this verifier states are not equivalent,
4815 * return false to continue verification of this path
4818 if (i % BPF_REG_SIZE)
4820 if (old->stack[spi].slot_type[0] != STACK_SPILL)
4822 if (!regsafe(&old->stack[spi].spilled_ptr,
4823 &cur->stack[spi].spilled_ptr,
4825 /* when explored and current stack slot are both storing
4826 * spilled registers, check that stored pointers types
4827 * are the same as well.
4828 * Ex: explored safe path could have stored
4829 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
4830 * but current path has stored:
4831 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
4832 * such verifier states are not equivalent.
4833 * return false to continue verification of this path
4840 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
4842 if (old->acquired_refs != cur->acquired_refs)
4844 return !memcmp(old->refs, cur->refs,
4845 sizeof(*old->refs) * old->acquired_refs);
4848 /* compare two verifier states
4850 * all states stored in state_list are known to be valid, since
4851 * verifier reached 'bpf_exit' instruction through them
4853 * this function is called when verifier exploring different branches of
4854 * execution popped from the state stack. If it sees an old state that has
4855 * more strict register state and more strict stack state then this execution
4856 * branch doesn't need to be explored further, since verifier already
4857 * concluded that more strict state leads to valid finish.
4859 * Therefore two states are equivalent if register state is more conservative
4860 * and explored stack state is more conservative than the current one.
4863 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
4864 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
4866 * In other words if current stack state (one being explored) has more
4867 * valid slots than old one that already passed validation, it means
4868 * the verifier can stop exploring and conclude that current state is valid too
4870 * Similarly with registers. If explored state has register type as invalid
4871 * whereas register type in current state is meaningful, it means that
4872 * the current state will reach 'bpf_exit' instruction safely
4874 static bool func_states_equal(struct bpf_func_state *old,
4875 struct bpf_func_state *cur)
4877 struct idpair *idmap;
4881 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
4882 /* If we failed to allocate the idmap, just say it's not safe */
4886 for (i = 0; i < MAX_BPF_REG; i++) {
4887 if (!regsafe(&old->regs[i], &cur->regs[i], idmap))
4891 if (!stacksafe(old, cur, idmap))
4894 if (!refsafe(old, cur))
4902 static bool states_equal(struct bpf_verifier_env *env,
4903 struct bpf_verifier_state *old,
4904 struct bpf_verifier_state *cur)
4908 if (old->curframe != cur->curframe)
4911 /* for states to be equal callsites have to be the same
4912 * and all frame states need to be equivalent
4914 for (i = 0; i <= old->curframe; i++) {
4915 if (old->frame[i]->callsite != cur->frame[i]->callsite)
4917 if (!func_states_equal(old->frame[i], cur->frame[i]))
4923 /* A write screens off any subsequent reads; but write marks come from the
4924 * straight-line code between a state and its parent. When we arrive at an
4925 * equivalent state (jump target or such) we didn't arrive by the straight-line
4926 * code, so read marks in the state must propagate to the parent regardless
4927 * of the state's write marks. That's what 'parent == state->parent' comparison
4928 * in mark_reg_read() is for.
4930 static int propagate_liveness(struct bpf_verifier_env *env,
4931 const struct bpf_verifier_state *vstate,
4932 struct bpf_verifier_state *vparent)
4934 int i, frame, err = 0;
4935 struct bpf_func_state *state, *parent;
4937 if (vparent->curframe != vstate->curframe) {
4938 WARN(1, "propagate_live: parent frame %d current frame %d\n",
4939 vparent->curframe, vstate->curframe);
4942 /* Propagate read liveness of registers... */
4943 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
4944 /* We don't need to worry about FP liveness because it's read-only */
4945 for (i = 0; i < BPF_REG_FP; i++) {
4946 if (vparent->frame[vparent->curframe]->regs[i].live & REG_LIVE_READ)
4948 if (vstate->frame[vstate->curframe]->regs[i].live & REG_LIVE_READ) {
4949 err = mark_reg_read(env, &vstate->frame[vstate->curframe]->regs[i],
4950 &vparent->frame[vstate->curframe]->regs[i]);
4956 /* ... and stack slots */
4957 for (frame = 0; frame <= vstate->curframe; frame++) {
4958 state = vstate->frame[frame];
4959 parent = vparent->frame[frame];
4960 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
4961 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
4962 if (parent->stack[i].spilled_ptr.live & REG_LIVE_READ)
4964 if (state->stack[i].spilled_ptr.live & REG_LIVE_READ)
4965 mark_reg_read(env, &state->stack[i].spilled_ptr,
4966 &parent->stack[i].spilled_ptr);
4972 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
4974 struct bpf_verifier_state_list *new_sl;
4975 struct bpf_verifier_state_list *sl;
4976 struct bpf_verifier_state *cur = env->cur_state, *new;
4979 sl = env->explored_states[insn_idx];
4981 /* this 'insn_idx' instruction wasn't marked, so we will not
4982 * be doing state search here
4986 while (sl != STATE_LIST_MARK) {
4987 if (states_equal(env, &sl->state, cur)) {
4988 /* reached equivalent register/stack state,
4990 * Registers read by the continuation are read by us.
4991 * If we have any write marks in env->cur_state, they
4992 * will prevent corresponding reads in the continuation
4993 * from reaching our parent (an explored_state). Our
4994 * own state will get the read marks recorded, but
4995 * they'll be immediately forgotten as we're pruning
4996 * this state and will pop a new one.
4998 err = propagate_liveness(env, &sl->state, cur);
5006 /* there were no equivalent states, remember current one.
5007 * technically the current state is not proven to be safe yet,
5008 * but it will either reach outer most bpf_exit (which means it's safe)
5009 * or it will be rejected. Since there are no loops, we won't be
5010 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
5011 * again on the way to bpf_exit
5013 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
5017 /* add new state to the head of linked list */
5018 new = &new_sl->state;
5019 err = copy_verifier_state(new, cur);
5021 free_verifier_state(new, false);
5025 new_sl->next = env->explored_states[insn_idx];
5026 env->explored_states[insn_idx] = new_sl;
5027 /* connect new state to parentage chain */
5028 for (i = 0; i < BPF_REG_FP; i++)
5029 cur_regs(env)[i].parent = &new->frame[new->curframe]->regs[i];
5030 /* clear write marks in current state: the writes we did are not writes
5031 * our child did, so they don't screen off its reads from us.
5032 * (There are no read marks in current state, because reads always mark
5033 * their parent and current state never has children yet. Only
5034 * explored_states can get read marks.)
5036 for (i = 0; i < BPF_REG_FP; i++)
5037 cur->frame[cur->curframe]->regs[i].live = REG_LIVE_NONE;
5039 /* all stack frames are accessible from callee, clear them all */
5040 for (j = 0; j <= cur->curframe; j++) {
5041 struct bpf_func_state *frame = cur->frame[j];
5042 struct bpf_func_state *newframe = new->frame[j];
5044 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
5045 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
5046 frame->stack[i].spilled_ptr.parent =
5047 &newframe->stack[i].spilled_ptr;
5053 /* Return true if it's OK to have the same insn return a different type. */
5054 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
5059 case PTR_TO_SOCKET_OR_NULL:
5066 /* If an instruction was previously used with particular pointer types, then we
5067 * need to be careful to avoid cases such as the below, where it may be ok
5068 * for one branch accessing the pointer, but not ok for the other branch:
5073 * R1 = some_other_valid_ptr;
5076 * R2 = *(u32 *)(R1 + 0);
5078 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
5080 return src != prev && (!reg_type_mismatch_ok(src) ||
5081 !reg_type_mismatch_ok(prev));
5084 static int do_check(struct bpf_verifier_env *env)
5086 struct bpf_verifier_state *state;
5087 struct bpf_insn *insns = env->prog->insnsi;
5088 struct bpf_reg_state *regs;
5089 int insn_cnt = env->prog->len, i;
5090 int insn_idx, prev_insn_idx = 0;
5091 int insn_processed = 0;
5092 bool do_print_state = false;
5094 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
5097 state->curframe = 0;
5098 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
5099 if (!state->frame[0]) {
5103 env->cur_state = state;
5104 init_func_state(env, state->frame[0],
5105 BPF_MAIN_FUNC /* callsite */,
5107 0 /* subprogno, zero == main subprog */);
5110 struct bpf_insn *insn;
5114 if (insn_idx >= insn_cnt) {
5115 verbose(env, "invalid insn idx %d insn_cnt %d\n",
5116 insn_idx, insn_cnt);
5120 insn = &insns[insn_idx];
5121 class = BPF_CLASS(insn->code);
5123 if (++insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
5125 "BPF program is too large. Processed %d insn\n",
5130 err = is_state_visited(env, insn_idx);
5134 /* found equivalent state, can prune the search */
5135 if (env->log.level) {
5137 verbose(env, "\nfrom %d to %d: safe\n",
5138 prev_insn_idx, insn_idx);
5140 verbose(env, "%d: safe\n", insn_idx);
5142 goto process_bpf_exit;
5148 if (env->log.level > 1 || (env->log.level && do_print_state)) {
5149 if (env->log.level > 1)
5150 verbose(env, "%d:", insn_idx);
5152 verbose(env, "\nfrom %d to %d:",
5153 prev_insn_idx, insn_idx);
5154 print_verifier_state(env, state->frame[state->curframe]);
5155 do_print_state = false;
5158 if (env->log.level) {
5159 const struct bpf_insn_cbs cbs = {
5160 .cb_print = verbose,
5161 .private_data = env,
5164 verbose(env, "%d: ", insn_idx);
5165 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
5168 if (bpf_prog_is_dev_bound(env->prog->aux)) {
5169 err = bpf_prog_offload_verify_insn(env, insn_idx,
5175 regs = cur_regs(env);
5176 env->insn_aux_data[insn_idx].seen = true;
5178 if (class == BPF_ALU || class == BPF_ALU64) {
5179 err = check_alu_op(env, insn);
5183 } else if (class == BPF_LDX) {
5184 enum bpf_reg_type *prev_src_type, src_reg_type;
5186 /* check for reserved fields is already done */
5188 /* check src operand */
5189 err = check_reg_arg(env, insn->src_reg, SRC_OP);
5193 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
5197 src_reg_type = regs[insn->src_reg].type;
5199 /* check that memory (src_reg + off) is readable,
5200 * the state of dst_reg will be updated by this func
5202 err = check_mem_access(env, insn_idx, insn->src_reg, insn->off,
5203 BPF_SIZE(insn->code), BPF_READ,
5204 insn->dst_reg, false);
5208 prev_src_type = &env->insn_aux_data[insn_idx].ptr_type;
5210 if (*prev_src_type == NOT_INIT) {
5212 * dst_reg = *(u32 *)(src_reg + off)
5213 * save type to validate intersecting paths
5215 *prev_src_type = src_reg_type;
5217 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
5218 /* ABuser program is trying to use the same insn
5219 * dst_reg = *(u32*) (src_reg + off)
5220 * with different pointer types:
5221 * src_reg == ctx in one branch and
5222 * src_reg == stack|map in some other branch.
5225 verbose(env, "same insn cannot be used with different pointers\n");
5229 } else if (class == BPF_STX) {
5230 enum bpf_reg_type *prev_dst_type, dst_reg_type;
5232 if (BPF_MODE(insn->code) == BPF_XADD) {
5233 err = check_xadd(env, insn_idx, insn);
5240 /* check src1 operand */
5241 err = check_reg_arg(env, insn->src_reg, SRC_OP);
5244 /* check src2 operand */
5245 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
5249 dst_reg_type = regs[insn->dst_reg].type;
5251 /* check that memory (dst_reg + off) is writeable */
5252 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
5253 BPF_SIZE(insn->code), BPF_WRITE,
5254 insn->src_reg, false);
5258 prev_dst_type = &env->insn_aux_data[insn_idx].ptr_type;
5260 if (*prev_dst_type == NOT_INIT) {
5261 *prev_dst_type = dst_reg_type;
5262 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
5263 verbose(env, "same insn cannot be used with different pointers\n");
5267 } else if (class == BPF_ST) {
5268 if (BPF_MODE(insn->code) != BPF_MEM ||
5269 insn->src_reg != BPF_REG_0) {
5270 verbose(env, "BPF_ST uses reserved fields\n");
5273 /* check src operand */
5274 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
5278 if (is_ctx_reg(env, insn->dst_reg)) {
5279 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
5281 reg_type_str[reg_state(env, insn->dst_reg)->type]);
5285 /* check that memory (dst_reg + off) is writeable */
5286 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
5287 BPF_SIZE(insn->code), BPF_WRITE,
5292 } else if (class == BPF_JMP) {
5293 u8 opcode = BPF_OP(insn->code);
5295 if (opcode == BPF_CALL) {
5296 if (BPF_SRC(insn->code) != BPF_K ||
5298 (insn->src_reg != BPF_REG_0 &&
5299 insn->src_reg != BPF_PSEUDO_CALL) ||
5300 insn->dst_reg != BPF_REG_0) {
5301 verbose(env, "BPF_CALL uses reserved fields\n");
5305 if (insn->src_reg == BPF_PSEUDO_CALL)
5306 err = check_func_call(env, insn, &insn_idx);
5308 err = check_helper_call(env, insn->imm, insn_idx);
5312 } else if (opcode == BPF_JA) {
5313 if (BPF_SRC(insn->code) != BPF_K ||
5315 insn->src_reg != BPF_REG_0 ||
5316 insn->dst_reg != BPF_REG_0) {
5317 verbose(env, "BPF_JA uses reserved fields\n");
5321 insn_idx += insn->off + 1;
5324 } else if (opcode == BPF_EXIT) {
5325 if (BPF_SRC(insn->code) != BPF_K ||
5327 insn->src_reg != BPF_REG_0 ||
5328 insn->dst_reg != BPF_REG_0) {
5329 verbose(env, "BPF_EXIT uses reserved fields\n");
5333 if (state->curframe) {
5334 /* exit from nested function */
5335 prev_insn_idx = insn_idx;
5336 err = prepare_func_exit(env, &insn_idx);
5339 do_print_state = true;
5343 err = check_reference_leak(env);
5347 /* eBPF calling convetion is such that R0 is used
5348 * to return the value from eBPF program.
5349 * Make sure that it's readable at this time
5350 * of bpf_exit, which means that program wrote
5351 * something into it earlier
5353 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
5357 if (is_pointer_value(env, BPF_REG_0)) {
5358 verbose(env, "R0 leaks addr as return value\n");
5362 err = check_return_code(env);
5366 err = pop_stack(env, &prev_insn_idx, &insn_idx);
5372 do_print_state = true;
5376 err = check_cond_jmp_op(env, insn, &insn_idx);
5380 } else if (class == BPF_LD) {
5381 u8 mode = BPF_MODE(insn->code);
5383 if (mode == BPF_ABS || mode == BPF_IND) {
5384 err = check_ld_abs(env, insn);
5388 } else if (mode == BPF_IMM) {
5389 err = check_ld_imm(env, insn);
5394 env->insn_aux_data[insn_idx].seen = true;
5396 verbose(env, "invalid BPF_LD mode\n");
5400 verbose(env, "unknown insn class %d\n", class);
5407 verbose(env, "processed %d insns (limit %d), stack depth ",
5408 insn_processed, BPF_COMPLEXITY_LIMIT_INSNS);
5409 for (i = 0; i < env->subprog_cnt; i++) {
5410 u32 depth = env->subprog_info[i].stack_depth;
5412 verbose(env, "%d", depth);
5413 if (i + 1 < env->subprog_cnt)
5417 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
5421 static int check_map_prealloc(struct bpf_map *map)
5423 return (map->map_type != BPF_MAP_TYPE_HASH &&
5424 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
5425 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
5426 !(map->map_flags & BPF_F_NO_PREALLOC);
5429 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
5430 struct bpf_map *map,
5431 struct bpf_prog *prog)
5434 /* Make sure that BPF_PROG_TYPE_PERF_EVENT programs only use
5435 * preallocated hash maps, since doing memory allocation
5436 * in overflow_handler can crash depending on where nmi got
5439 if (prog->type == BPF_PROG_TYPE_PERF_EVENT) {
5440 if (!check_map_prealloc(map)) {
5441 verbose(env, "perf_event programs can only use preallocated hash map\n");
5444 if (map->inner_map_meta &&
5445 !check_map_prealloc(map->inner_map_meta)) {
5446 verbose(env, "perf_event programs can only use preallocated inner hash map\n");
5451 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
5452 !bpf_offload_prog_map_match(prog, map)) {
5453 verbose(env, "offload device mismatch between prog and map\n");
5460 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
5462 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
5463 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
5466 /* look for pseudo eBPF instructions that access map FDs and
5467 * replace them with actual map pointers
5469 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env *env)
5471 struct bpf_insn *insn = env->prog->insnsi;
5472 int insn_cnt = env->prog->len;
5475 err = bpf_prog_calc_tag(env->prog);
5479 for (i = 0; i < insn_cnt; i++, insn++) {
5480 if (BPF_CLASS(insn->code) == BPF_LDX &&
5481 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
5482 verbose(env, "BPF_LDX uses reserved fields\n");
5486 if (BPF_CLASS(insn->code) == BPF_STX &&
5487 ((BPF_MODE(insn->code) != BPF_MEM &&
5488 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
5489 verbose(env, "BPF_STX uses reserved fields\n");
5493 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
5494 struct bpf_map *map;
5497 if (i == insn_cnt - 1 || insn[1].code != 0 ||
5498 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
5500 verbose(env, "invalid bpf_ld_imm64 insn\n");
5504 if (insn->src_reg == 0)
5505 /* valid generic load 64-bit imm */
5508 if (insn->src_reg != BPF_PSEUDO_MAP_FD) {
5510 "unrecognized bpf_ld_imm64 insn\n");
5514 f = fdget(insn->imm);
5515 map = __bpf_map_get(f);
5517 verbose(env, "fd %d is not pointing to valid bpf_map\n",
5519 return PTR_ERR(map);
5522 err = check_map_prog_compatibility(env, map, env->prog);
5528 /* store map pointer inside BPF_LD_IMM64 instruction */
5529 insn[0].imm = (u32) (unsigned long) map;
5530 insn[1].imm = ((u64) (unsigned long) map) >> 32;
5532 /* check whether we recorded this map already */
5533 for (j = 0; j < env->used_map_cnt; j++)
5534 if (env->used_maps[j] == map) {
5539 if (env->used_map_cnt >= MAX_USED_MAPS) {
5544 /* hold the map. If the program is rejected by verifier,
5545 * the map will be released by release_maps() or it
5546 * will be used by the valid program until it's unloaded
5547 * and all maps are released in free_used_maps()
5549 map = bpf_map_inc(map, false);
5552 return PTR_ERR(map);
5554 env->used_maps[env->used_map_cnt++] = map;
5556 if (bpf_map_is_cgroup_storage(map) &&
5557 bpf_cgroup_storage_assign(env->prog, map)) {
5558 verbose(env, "only one cgroup storage of each type is allowed\n");
5570 /* Basic sanity check before we invest more work here. */
5571 if (!bpf_opcode_in_insntable(insn->code)) {
5572 verbose(env, "unknown opcode %02x\n", insn->code);
5577 /* now all pseudo BPF_LD_IMM64 instructions load valid
5578 * 'struct bpf_map *' into a register instead of user map_fd.
5579 * These pointers will be used later by verifier to validate map access.
5584 /* drop refcnt of maps used by the rejected program */
5585 static void release_maps(struct bpf_verifier_env *env)
5587 enum bpf_cgroup_storage_type stype;
5590 for_each_cgroup_storage_type(stype) {
5591 if (!env->prog->aux->cgroup_storage[stype])
5593 bpf_cgroup_storage_release(env->prog,
5594 env->prog->aux->cgroup_storage[stype]);
5597 for (i = 0; i < env->used_map_cnt; i++)
5598 bpf_map_put(env->used_maps[i]);
5601 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
5602 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
5604 struct bpf_insn *insn = env->prog->insnsi;
5605 int insn_cnt = env->prog->len;
5608 for (i = 0; i < insn_cnt; i++, insn++)
5609 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
5613 /* single env->prog->insni[off] instruction was replaced with the range
5614 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
5615 * [0, off) and [off, end) to new locations, so the patched range stays zero
5617 static int adjust_insn_aux_data(struct bpf_verifier_env *env, u32 prog_len,
5620 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
5625 new_data = vzalloc(array_size(prog_len,
5626 sizeof(struct bpf_insn_aux_data)));
5629 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
5630 memcpy(new_data + off + cnt - 1, old_data + off,
5631 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
5632 for (i = off; i < off + cnt - 1; i++)
5633 new_data[i].seen = true;
5634 env->insn_aux_data = new_data;
5639 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
5645 /* NOTE: fake 'exit' subprog should be updated as well. */
5646 for (i = 0; i <= env->subprog_cnt; i++) {
5647 if (env->subprog_info[i].start < off)
5649 env->subprog_info[i].start += len - 1;
5653 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
5654 const struct bpf_insn *patch, u32 len)
5656 struct bpf_prog *new_prog;
5658 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
5661 if (adjust_insn_aux_data(env, new_prog->len, off, len))
5663 adjust_subprog_starts(env, off, len);
5667 /* The verifier does more data flow analysis than llvm and will not
5668 * explore branches that are dead at run time. Malicious programs can
5669 * have dead code too. Therefore replace all dead at-run-time code
5672 * Just nops are not optimal, e.g. if they would sit at the end of the
5673 * program and through another bug we would manage to jump there, then
5674 * we'd execute beyond program memory otherwise. Returning exception
5675 * code also wouldn't work since we can have subprogs where the dead
5676 * code could be located.
5678 static void sanitize_dead_code(struct bpf_verifier_env *env)
5680 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
5681 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
5682 struct bpf_insn *insn = env->prog->insnsi;
5683 const int insn_cnt = env->prog->len;
5686 for (i = 0; i < insn_cnt; i++) {
5687 if (aux_data[i].seen)
5689 memcpy(insn + i, &trap, sizeof(trap));
5693 /* convert load instructions that access fields of a context type into a
5694 * sequence of instructions that access fields of the underlying structure:
5695 * struct __sk_buff -> struct sk_buff
5696 * struct bpf_sock_ops -> struct sock
5698 static int convert_ctx_accesses(struct bpf_verifier_env *env)
5700 const struct bpf_verifier_ops *ops = env->ops;
5701 int i, cnt, size, ctx_field_size, delta = 0;
5702 const int insn_cnt = env->prog->len;
5703 struct bpf_insn insn_buf[16], *insn;
5704 struct bpf_prog *new_prog;
5705 enum bpf_access_type type;
5706 bool is_narrower_load;
5709 if (ops->gen_prologue) {
5710 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
5712 if (cnt >= ARRAY_SIZE(insn_buf)) {
5713 verbose(env, "bpf verifier is misconfigured\n");
5716 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
5720 env->prog = new_prog;
5725 if (bpf_prog_is_dev_bound(env->prog->aux))
5728 insn = env->prog->insnsi + delta;
5730 for (i = 0; i < insn_cnt; i++, insn++) {
5731 bpf_convert_ctx_access_t convert_ctx_access;
5733 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
5734 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
5735 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
5736 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
5738 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
5739 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
5740 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
5741 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
5746 if (type == BPF_WRITE &&
5747 env->insn_aux_data[i + delta].sanitize_stack_off) {
5748 struct bpf_insn patch[] = {
5749 /* Sanitize suspicious stack slot with zero.
5750 * There are no memory dependencies for this store,
5751 * since it's only using frame pointer and immediate
5754 BPF_ST_MEM(BPF_DW, BPF_REG_FP,
5755 env->insn_aux_data[i + delta].sanitize_stack_off,
5757 /* the original STX instruction will immediately
5758 * overwrite the same stack slot with appropriate value
5763 cnt = ARRAY_SIZE(patch);
5764 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
5769 env->prog = new_prog;
5770 insn = new_prog->insnsi + i + delta;
5774 switch (env->insn_aux_data[i + delta].ptr_type) {
5776 if (!ops->convert_ctx_access)
5778 convert_ctx_access = ops->convert_ctx_access;
5781 convert_ctx_access = bpf_sock_convert_ctx_access;
5787 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
5788 size = BPF_LDST_BYTES(insn);
5790 /* If the read access is a narrower load of the field,
5791 * convert to a 4/8-byte load, to minimum program type specific
5792 * convert_ctx_access changes. If conversion is successful,
5793 * we will apply proper mask to the result.
5795 is_narrower_load = size < ctx_field_size;
5796 if (is_narrower_load) {
5797 u32 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
5798 u32 off = insn->off;
5801 if (type == BPF_WRITE) {
5802 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
5807 if (ctx_field_size == 4)
5809 else if (ctx_field_size == 8)
5812 insn->off = off & ~(size_default - 1);
5813 insn->code = BPF_LDX | BPF_MEM | size_code;
5817 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
5819 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
5820 (ctx_field_size && !target_size)) {
5821 verbose(env, "bpf verifier is misconfigured\n");
5825 if (is_narrower_load && size < target_size) {
5826 if (ctx_field_size <= 4)
5827 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
5828 (1 << size * 8) - 1);
5830 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
5831 (1 << size * 8) - 1);
5834 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
5840 /* keep walking new program and skip insns we just inserted */
5841 env->prog = new_prog;
5842 insn = new_prog->insnsi + i + delta;
5848 static int jit_subprogs(struct bpf_verifier_env *env)
5850 struct bpf_prog *prog = env->prog, **func, *tmp;
5851 int i, j, subprog_start, subprog_end = 0, len, subprog;
5852 struct bpf_insn *insn;
5856 if (env->subprog_cnt <= 1)
5859 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
5860 if (insn->code != (BPF_JMP | BPF_CALL) ||
5861 insn->src_reg != BPF_PSEUDO_CALL)
5863 /* Upon error here we cannot fall back to interpreter but
5864 * need a hard reject of the program. Thus -EFAULT is
5865 * propagated in any case.
5867 subprog = find_subprog(env, i + insn->imm + 1);
5869 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
5873 /* temporarily remember subprog id inside insn instead of
5874 * aux_data, since next loop will split up all insns into funcs
5876 insn->off = subprog;
5877 /* remember original imm in case JIT fails and fallback
5878 * to interpreter will be needed
5880 env->insn_aux_data[i].call_imm = insn->imm;
5881 /* point imm to __bpf_call_base+1 from JITs point of view */
5885 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
5889 for (i = 0; i < env->subprog_cnt; i++) {
5890 subprog_start = subprog_end;
5891 subprog_end = env->subprog_info[i + 1].start;
5893 len = subprog_end - subprog_start;
5894 func[i] = bpf_prog_alloc(bpf_prog_size(len), GFP_USER);
5897 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
5898 len * sizeof(struct bpf_insn));
5899 func[i]->type = prog->type;
5901 if (bpf_prog_calc_tag(func[i]))
5903 func[i]->is_func = 1;
5904 /* Use bpf_prog_F_tag to indicate functions in stack traces.
5905 * Long term would need debug info to populate names
5907 func[i]->aux->name[0] = 'F';
5908 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
5909 func[i]->jit_requested = 1;
5910 func[i] = bpf_int_jit_compile(func[i]);
5911 if (!func[i]->jited) {
5917 /* at this point all bpf functions were successfully JITed
5918 * now populate all bpf_calls with correct addresses and
5919 * run last pass of JIT
5921 for (i = 0; i < env->subprog_cnt; i++) {
5922 insn = func[i]->insnsi;
5923 for (j = 0; j < func[i]->len; j++, insn++) {
5924 if (insn->code != (BPF_JMP | BPF_CALL) ||
5925 insn->src_reg != BPF_PSEUDO_CALL)
5927 subprog = insn->off;
5928 insn->imm = (u64 (*)(u64, u64, u64, u64, u64))
5929 func[subprog]->bpf_func -
5933 /* we use the aux data to keep a list of the start addresses
5934 * of the JITed images for each function in the program
5936 * for some architectures, such as powerpc64, the imm field
5937 * might not be large enough to hold the offset of the start
5938 * address of the callee's JITed image from __bpf_call_base
5940 * in such cases, we can lookup the start address of a callee
5941 * by using its subprog id, available from the off field of
5942 * the call instruction, as an index for this list
5944 func[i]->aux->func = func;
5945 func[i]->aux->func_cnt = env->subprog_cnt;
5947 for (i = 0; i < env->subprog_cnt; i++) {
5948 old_bpf_func = func[i]->bpf_func;
5949 tmp = bpf_int_jit_compile(func[i]);
5950 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
5951 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
5958 /* finally lock prog and jit images for all functions and
5961 for (i = 0; i < env->subprog_cnt; i++) {
5962 bpf_prog_lock_ro(func[i]);
5963 bpf_prog_kallsyms_add(func[i]);
5966 /* Last step: make now unused interpreter insns from main
5967 * prog consistent for later dump requests, so they can
5968 * later look the same as if they were interpreted only.
5970 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
5971 if (insn->code != (BPF_JMP | BPF_CALL) ||
5972 insn->src_reg != BPF_PSEUDO_CALL)
5974 insn->off = env->insn_aux_data[i].call_imm;
5975 subprog = find_subprog(env, i + insn->off + 1);
5976 insn->imm = subprog;
5980 prog->bpf_func = func[0]->bpf_func;
5981 prog->aux->func = func;
5982 prog->aux->func_cnt = env->subprog_cnt;
5985 for (i = 0; i < env->subprog_cnt; i++)
5987 bpf_jit_free(func[i]);
5990 /* cleanup main prog to be interpreted */
5991 prog->jit_requested = 0;
5992 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
5993 if (insn->code != (BPF_JMP | BPF_CALL) ||
5994 insn->src_reg != BPF_PSEUDO_CALL)
5997 insn->imm = env->insn_aux_data[i].call_imm;
6002 static int fixup_call_args(struct bpf_verifier_env *env)
6004 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
6005 struct bpf_prog *prog = env->prog;
6006 struct bpf_insn *insn = prog->insnsi;
6011 if (env->prog->jit_requested &&
6012 !bpf_prog_is_dev_bound(env->prog->aux)) {
6013 err = jit_subprogs(env);
6019 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
6020 for (i = 0; i < prog->len; i++, insn++) {
6021 if (insn->code != (BPF_JMP | BPF_CALL) ||
6022 insn->src_reg != BPF_PSEUDO_CALL)
6024 depth = get_callee_stack_depth(env, insn, i);
6027 bpf_patch_call_args(insn, depth);
6034 /* fixup insn->imm field of bpf_call instructions
6035 * and inline eligible helpers as explicit sequence of BPF instructions
6037 * this function is called after eBPF program passed verification
6039 static int fixup_bpf_calls(struct bpf_verifier_env *env)
6041 struct bpf_prog *prog = env->prog;
6042 struct bpf_insn *insn = prog->insnsi;
6043 const struct bpf_func_proto *fn;
6044 const int insn_cnt = prog->len;
6045 const struct bpf_map_ops *ops;
6046 struct bpf_insn_aux_data *aux;
6047 struct bpf_insn insn_buf[16];
6048 struct bpf_prog *new_prog;
6049 struct bpf_map *map_ptr;
6050 int i, cnt, delta = 0;
6052 for (i = 0; i < insn_cnt; i++, insn++) {
6053 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
6054 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
6055 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
6056 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
6057 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
6058 struct bpf_insn mask_and_div[] = {
6059 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
6061 BPF_JMP_IMM(BPF_JNE, insn->src_reg, 0, 2),
6062 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
6063 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
6066 struct bpf_insn mask_and_mod[] = {
6067 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
6068 /* Rx mod 0 -> Rx */
6069 BPF_JMP_IMM(BPF_JEQ, insn->src_reg, 0, 1),
6072 struct bpf_insn *patchlet;
6074 if (insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
6075 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
6076 patchlet = mask_and_div + (is64 ? 1 : 0);
6077 cnt = ARRAY_SIZE(mask_and_div) - (is64 ? 1 : 0);
6079 patchlet = mask_and_mod + (is64 ? 1 : 0);
6080 cnt = ARRAY_SIZE(mask_and_mod) - (is64 ? 1 : 0);
6083 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
6088 env->prog = prog = new_prog;
6089 insn = new_prog->insnsi + i + delta;
6093 if (BPF_CLASS(insn->code) == BPF_LD &&
6094 (BPF_MODE(insn->code) == BPF_ABS ||
6095 BPF_MODE(insn->code) == BPF_IND)) {
6096 cnt = env->ops->gen_ld_abs(insn, insn_buf);
6097 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
6098 verbose(env, "bpf verifier is misconfigured\n");
6102 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
6107 env->prog = prog = new_prog;
6108 insn = new_prog->insnsi + i + delta;
6112 if (insn->code != (BPF_JMP | BPF_CALL))
6114 if (insn->src_reg == BPF_PSEUDO_CALL)
6117 if (insn->imm == BPF_FUNC_get_route_realm)
6118 prog->dst_needed = 1;
6119 if (insn->imm == BPF_FUNC_get_prandom_u32)
6120 bpf_user_rnd_init_once();
6121 if (insn->imm == BPF_FUNC_override_return)
6122 prog->kprobe_override = 1;
6123 if (insn->imm == BPF_FUNC_tail_call) {
6124 /* If we tail call into other programs, we
6125 * cannot make any assumptions since they can
6126 * be replaced dynamically during runtime in
6127 * the program array.
6129 prog->cb_access = 1;
6130 env->prog->aux->stack_depth = MAX_BPF_STACK;
6132 /* mark bpf_tail_call as different opcode to avoid
6133 * conditional branch in the interpeter for every normal
6134 * call and to prevent accidental JITing by JIT compiler
6135 * that doesn't support bpf_tail_call yet
6138 insn->code = BPF_JMP | BPF_TAIL_CALL;
6140 aux = &env->insn_aux_data[i + delta];
6141 if (!bpf_map_ptr_unpriv(aux))
6144 /* instead of changing every JIT dealing with tail_call
6145 * emit two extra insns:
6146 * if (index >= max_entries) goto out;
6147 * index &= array->index_mask;
6148 * to avoid out-of-bounds cpu speculation
6150 if (bpf_map_ptr_poisoned(aux)) {
6151 verbose(env, "tail_call abusing map_ptr\n");
6155 map_ptr = BPF_MAP_PTR(aux->map_state);
6156 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
6157 map_ptr->max_entries, 2);
6158 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
6159 container_of(map_ptr,
6162 insn_buf[2] = *insn;
6164 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
6169 env->prog = prog = new_prog;
6170 insn = new_prog->insnsi + i + delta;
6174 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
6175 * and other inlining handlers are currently limited to 64 bit
6178 if (prog->jit_requested && BITS_PER_LONG == 64 &&
6179 (insn->imm == BPF_FUNC_map_lookup_elem ||
6180 insn->imm == BPF_FUNC_map_update_elem ||
6181 insn->imm == BPF_FUNC_map_delete_elem ||
6182 insn->imm == BPF_FUNC_map_push_elem ||
6183 insn->imm == BPF_FUNC_map_pop_elem ||
6184 insn->imm == BPF_FUNC_map_peek_elem)) {
6185 aux = &env->insn_aux_data[i + delta];
6186 if (bpf_map_ptr_poisoned(aux))
6187 goto patch_call_imm;
6189 map_ptr = BPF_MAP_PTR(aux->map_state);
6191 if (insn->imm == BPF_FUNC_map_lookup_elem &&
6192 ops->map_gen_lookup) {
6193 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
6194 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
6195 verbose(env, "bpf verifier is misconfigured\n");
6199 new_prog = bpf_patch_insn_data(env, i + delta,
6205 env->prog = prog = new_prog;
6206 insn = new_prog->insnsi + i + delta;
6210 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
6211 (void *(*)(struct bpf_map *map, void *key))NULL));
6212 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
6213 (int (*)(struct bpf_map *map, void *key))NULL));
6214 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
6215 (int (*)(struct bpf_map *map, void *key, void *value,
6217 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
6218 (int (*)(struct bpf_map *map, void *value,
6220 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
6221 (int (*)(struct bpf_map *map, void *value))NULL));
6222 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
6223 (int (*)(struct bpf_map *map, void *value))NULL));
6225 switch (insn->imm) {
6226 case BPF_FUNC_map_lookup_elem:
6227 insn->imm = BPF_CAST_CALL(ops->map_lookup_elem) -
6230 case BPF_FUNC_map_update_elem:
6231 insn->imm = BPF_CAST_CALL(ops->map_update_elem) -
6234 case BPF_FUNC_map_delete_elem:
6235 insn->imm = BPF_CAST_CALL(ops->map_delete_elem) -
6238 case BPF_FUNC_map_push_elem:
6239 insn->imm = BPF_CAST_CALL(ops->map_push_elem) -
6242 case BPF_FUNC_map_pop_elem:
6243 insn->imm = BPF_CAST_CALL(ops->map_pop_elem) -
6246 case BPF_FUNC_map_peek_elem:
6247 insn->imm = BPF_CAST_CALL(ops->map_peek_elem) -
6252 goto patch_call_imm;
6256 fn = env->ops->get_func_proto(insn->imm, env->prog);
6257 /* all functions that have prototype and verifier allowed
6258 * programs to call them, must be real in-kernel functions
6262 "kernel subsystem misconfigured func %s#%d\n",
6263 func_id_name(insn->imm), insn->imm);
6266 insn->imm = fn->func - __bpf_call_base;
6272 static void free_states(struct bpf_verifier_env *env)
6274 struct bpf_verifier_state_list *sl, *sln;
6277 if (!env->explored_states)
6280 for (i = 0; i < env->prog->len; i++) {
6281 sl = env->explored_states[i];
6284 while (sl != STATE_LIST_MARK) {
6286 free_verifier_state(&sl->state, false);
6292 kfree(env->explored_states);
6295 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr)
6297 struct bpf_verifier_env *env;
6298 struct bpf_verifier_log *log;
6301 /* no program is valid */
6302 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
6305 /* 'struct bpf_verifier_env' can be global, but since it's not small,
6306 * allocate/free it every time bpf_check() is called
6308 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
6313 env->insn_aux_data =
6314 vzalloc(array_size(sizeof(struct bpf_insn_aux_data),
6317 if (!env->insn_aux_data)
6320 env->ops = bpf_verifier_ops[env->prog->type];
6322 /* grab the mutex to protect few globals used by verifier */
6323 mutex_lock(&bpf_verifier_lock);
6325 if (attr->log_level || attr->log_buf || attr->log_size) {
6326 /* user requested verbose verifier output
6327 * and supplied buffer to store the verification trace
6329 log->level = attr->log_level;
6330 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
6331 log->len_total = attr->log_size;
6334 /* log attributes have to be sane */
6335 if (log->len_total < 128 || log->len_total > UINT_MAX >> 8 ||
6336 !log->level || !log->ubuf)
6340 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
6341 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
6342 env->strict_alignment = true;
6344 ret = replace_map_fd_with_map_ptr(env);
6346 goto skip_full_check;
6348 if (bpf_prog_is_dev_bound(env->prog->aux)) {
6349 ret = bpf_prog_offload_verifier_prep(env);
6351 goto skip_full_check;
6354 env->explored_states = kcalloc(env->prog->len,
6355 sizeof(struct bpf_verifier_state_list *),
6358 if (!env->explored_states)
6359 goto skip_full_check;
6361 env->allow_ptr_leaks = capable(CAP_SYS_ADMIN);
6363 ret = check_cfg(env);
6365 goto skip_full_check;
6367 ret = do_check(env);
6368 if (env->cur_state) {
6369 free_verifier_state(env->cur_state, true);
6370 env->cur_state = NULL;
6373 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
6374 ret = bpf_prog_offload_finalize(env);
6377 while (!pop_stack(env, NULL, NULL));
6381 sanitize_dead_code(env);
6384 ret = check_max_stack_depth(env);
6387 /* program is valid, convert *(u32*)(ctx + off) accesses */
6388 ret = convert_ctx_accesses(env);
6391 ret = fixup_bpf_calls(env);
6394 ret = fixup_call_args(env);
6396 if (log->level && bpf_verifier_log_full(log))
6398 if (log->level && !log->ubuf) {
6400 goto err_release_maps;
6403 if (ret == 0 && env->used_map_cnt) {
6404 /* if program passed verifier, update used_maps in bpf_prog_info */
6405 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
6406 sizeof(env->used_maps[0]),
6409 if (!env->prog->aux->used_maps) {
6411 goto err_release_maps;
6414 memcpy(env->prog->aux->used_maps, env->used_maps,
6415 sizeof(env->used_maps[0]) * env->used_map_cnt);
6416 env->prog->aux->used_map_cnt = env->used_map_cnt;
6418 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
6419 * bpf_ld_imm64 instructions
6421 convert_pseudo_ld_imm64(env);
6425 if (!env->prog->aux->used_maps)
6426 /* if we didn't copy map pointers into bpf_prog_info, release
6427 * them now. Otherwise free_used_maps() will release them.
6432 mutex_unlock(&bpf_verifier_lock);
6433 vfree(env->insn_aux_data);