1 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
2 * Copyright (c) 2016 Facebook
3 * Copyright (c) 2018 Covalent IO, Inc. http://covalent.io
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of version 2 of the GNU General Public
7 * License as published by the Free Software Foundation.
9 * This program is distributed in the hope that it will be useful, but
10 * WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
12 * General Public License for more details.
14 #include <uapi/linux/btf.h>
15 #include <linux/kernel.h>
16 #include <linux/types.h>
17 #include <linux/slab.h>
18 #include <linux/bpf.h>
19 #include <linux/btf.h>
20 #include <linux/bpf_verifier.h>
21 #include <linux/filter.h>
22 #include <net/netlink.h>
23 #include <linux/file.h>
24 #include <linux/vmalloc.h>
25 #include <linux/stringify.h>
26 #include <linux/bsearch.h>
27 #include <linux/sort.h>
28 #include <linux/perf_event.h>
29 #include <linux/ctype.h>
33 static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
34 #define BPF_PROG_TYPE(_id, _name) \
35 [_id] = & _name ## _verifier_ops,
36 #define BPF_MAP_TYPE(_id, _ops)
37 #include <linux/bpf_types.h>
42 /* bpf_check() is a static code analyzer that walks eBPF program
43 * instruction by instruction and updates register/stack state.
44 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
46 * The first pass is depth-first-search to check that the program is a DAG.
47 * It rejects the following programs:
48 * - larger than BPF_MAXINSNS insns
49 * - if loop is present (detected via back-edge)
50 * - unreachable insns exist (shouldn't be a forest. program = one function)
51 * - out of bounds or malformed jumps
52 * The second pass is all possible path descent from the 1st insn.
53 * Since it's analyzing all pathes through the program, the length of the
54 * analysis is limited to 64k insn, which may be hit even if total number of
55 * insn is less then 4K, but there are too many branches that change stack/regs.
56 * Number of 'branches to be analyzed' is limited to 1k
58 * On entry to each instruction, each register has a type, and the instruction
59 * changes the types of the registers depending on instruction semantics.
60 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
63 * All registers are 64-bit.
64 * R0 - return register
65 * R1-R5 argument passing registers
66 * R6-R9 callee saved registers
67 * R10 - frame pointer read-only
69 * At the start of BPF program the register R1 contains a pointer to bpf_context
70 * and has type PTR_TO_CTX.
72 * Verifier tracks arithmetic operations on pointers in case:
73 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
74 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
75 * 1st insn copies R10 (which has FRAME_PTR) type into R1
76 * and 2nd arithmetic instruction is pattern matched to recognize
77 * that it wants to construct a pointer to some element within stack.
78 * So after 2nd insn, the register R1 has type PTR_TO_STACK
79 * (and -20 constant is saved for further stack bounds checking).
80 * Meaning that this reg is a pointer to stack plus known immediate constant.
82 * Most of the time the registers have SCALAR_VALUE type, which
83 * means the register has some value, but it's not a valid pointer.
84 * (like pointer plus pointer becomes SCALAR_VALUE type)
86 * When verifier sees load or store instructions the type of base register
87 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK, PTR_TO_SOCKET. These are
88 * four pointer types recognized by check_mem_access() function.
90 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
91 * and the range of [ptr, ptr + map's value_size) is accessible.
93 * registers used to pass values to function calls are checked against
94 * function argument constraints.
96 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
97 * It means that the register type passed to this function must be
98 * PTR_TO_STACK and it will be used inside the function as
99 * 'pointer to map element key'
101 * For example the argument constraints for bpf_map_lookup_elem():
102 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
103 * .arg1_type = ARG_CONST_MAP_PTR,
104 * .arg2_type = ARG_PTR_TO_MAP_KEY,
106 * ret_type says that this function returns 'pointer to map elem value or null'
107 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
108 * 2nd argument should be a pointer to stack, which will be used inside
109 * the helper function as a pointer to map element key.
111 * On the kernel side the helper function looks like:
112 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
114 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
115 * void *key = (void *) (unsigned long) r2;
118 * here kernel can access 'key' and 'map' pointers safely, knowing that
119 * [key, key + map->key_size) bytes are valid and were initialized on
120 * the stack of eBPF program.
123 * Corresponding eBPF program may look like:
124 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
125 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
126 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
127 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
128 * here verifier looks at prototype of map_lookup_elem() and sees:
129 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
130 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
132 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
133 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
134 * and were initialized prior to this call.
135 * If it's ok, then verifier allows this BPF_CALL insn and looks at
136 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
137 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
138 * returns ether pointer to map value or NULL.
140 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
141 * insn, the register holding that pointer in the true branch changes state to
142 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
143 * branch. See check_cond_jmp_op().
145 * After the call R0 is set to return type of the function and registers R1-R5
146 * are set to NOT_INIT to indicate that they are no longer readable.
148 * The following reference types represent a potential reference to a kernel
149 * resource which, after first being allocated, must be checked and freed by
151 * - PTR_TO_SOCKET_OR_NULL, PTR_TO_SOCKET
153 * When the verifier sees a helper call return a reference type, it allocates a
154 * pointer id for the reference and stores it in the current function state.
155 * Similar to the way that PTR_TO_MAP_VALUE_OR_NULL is converted into
156 * PTR_TO_MAP_VALUE, PTR_TO_SOCKET_OR_NULL becomes PTR_TO_SOCKET when the type
157 * passes through a NULL-check conditional. For the branch wherein the state is
158 * changed to CONST_IMM, the verifier releases the reference.
160 * For each helper function that allocates a reference, such as
161 * bpf_sk_lookup_tcp(), there is a corresponding release function, such as
162 * bpf_sk_release(). When a reference type passes into the release function,
163 * the verifier also releases the reference. If any unchecked or unreleased
164 * reference remains at the end of the program, the verifier rejects it.
167 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
168 struct bpf_verifier_stack_elem {
169 /* verifer state is 'st'
170 * before processing instruction 'insn_idx'
171 * and after processing instruction 'prev_insn_idx'
173 struct bpf_verifier_state st;
176 struct bpf_verifier_stack_elem *next;
179 #define BPF_COMPLEXITY_LIMIT_INSNS 131072
180 #define BPF_COMPLEXITY_LIMIT_STACK 1024
181 #define BPF_COMPLEXITY_LIMIT_STATES 64
183 #define BPF_MAP_PTR_UNPRIV 1UL
184 #define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \
185 POISON_POINTER_DELTA))
186 #define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
188 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
190 return BPF_MAP_PTR(aux->map_state) == BPF_MAP_PTR_POISON;
193 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
195 return aux->map_state & BPF_MAP_PTR_UNPRIV;
198 static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
199 const struct bpf_map *map, bool unpriv)
201 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON & BPF_MAP_PTR_UNPRIV);
202 unpriv |= bpf_map_ptr_unpriv(aux);
203 aux->map_state = (unsigned long)map |
204 (unpriv ? BPF_MAP_PTR_UNPRIV : 0UL);
207 struct bpf_call_arg_meta {
208 struct bpf_map *map_ptr;
213 s64 msize_smax_value;
214 u64 msize_umax_value;
218 static DEFINE_MUTEX(bpf_verifier_lock);
220 static const struct bpf_line_info *
221 find_linfo(const struct bpf_verifier_env *env, u32 insn_off)
223 const struct bpf_line_info *linfo;
224 const struct bpf_prog *prog;
228 nr_linfo = prog->aux->nr_linfo;
230 if (!nr_linfo || insn_off >= prog->len)
233 linfo = prog->aux->linfo;
234 for (i = 1; i < nr_linfo; i++)
235 if (insn_off < linfo[i].insn_off)
238 return &linfo[i - 1];
241 void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
246 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
248 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
249 "verifier log line truncated - local buffer too short\n");
251 n = min(log->len_total - log->len_used - 1, n);
254 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
260 /* log_level controls verbosity level of eBPF verifier.
261 * bpf_verifier_log_write() is used to dump the verification trace to the log,
262 * so the user can figure out what's wrong with the program
264 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
265 const char *fmt, ...)
269 if (!bpf_verifier_log_needed(&env->log))
273 bpf_verifier_vlog(&env->log, fmt, args);
276 EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
278 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
280 struct bpf_verifier_env *env = private_data;
283 if (!bpf_verifier_log_needed(&env->log))
287 bpf_verifier_vlog(&env->log, fmt, args);
291 static const char *ltrim(const char *s)
299 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env,
301 const char *prefix_fmt, ...)
303 const struct bpf_line_info *linfo;
305 if (!bpf_verifier_log_needed(&env->log))
308 linfo = find_linfo(env, insn_off);
309 if (!linfo || linfo == env->prev_linfo)
315 va_start(args, prefix_fmt);
316 bpf_verifier_vlog(&env->log, prefix_fmt, args);
321 ltrim(btf_name_by_offset(env->prog->aux->btf,
324 env->prev_linfo = linfo;
327 static bool type_is_pkt_pointer(enum bpf_reg_type type)
329 return type == PTR_TO_PACKET ||
330 type == PTR_TO_PACKET_META;
333 static bool reg_type_may_be_null(enum bpf_reg_type type)
335 return type == PTR_TO_MAP_VALUE_OR_NULL ||
336 type == PTR_TO_SOCKET_OR_NULL;
339 static bool type_is_refcounted(enum bpf_reg_type type)
341 return type == PTR_TO_SOCKET;
344 static bool type_is_refcounted_or_null(enum bpf_reg_type type)
346 return type == PTR_TO_SOCKET || type == PTR_TO_SOCKET_OR_NULL;
349 static bool reg_is_refcounted(const struct bpf_reg_state *reg)
351 return type_is_refcounted(reg->type);
354 static bool reg_is_refcounted_or_null(const struct bpf_reg_state *reg)
356 return type_is_refcounted_or_null(reg->type);
359 static bool arg_type_is_refcounted(enum bpf_arg_type type)
361 return type == ARG_PTR_TO_SOCKET;
364 /* Determine whether the function releases some resources allocated by another
365 * function call. The first reference type argument will be assumed to be
366 * released by release_reference().
368 static bool is_release_function(enum bpf_func_id func_id)
370 return func_id == BPF_FUNC_sk_release;
373 /* string representation of 'enum bpf_reg_type' */
374 static const char * const reg_type_str[] = {
376 [SCALAR_VALUE] = "inv",
377 [PTR_TO_CTX] = "ctx",
378 [CONST_PTR_TO_MAP] = "map_ptr",
379 [PTR_TO_MAP_VALUE] = "map_value",
380 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
381 [PTR_TO_STACK] = "fp",
382 [PTR_TO_PACKET] = "pkt",
383 [PTR_TO_PACKET_META] = "pkt_meta",
384 [PTR_TO_PACKET_END] = "pkt_end",
385 [PTR_TO_FLOW_KEYS] = "flow_keys",
386 [PTR_TO_SOCKET] = "sock",
387 [PTR_TO_SOCKET_OR_NULL] = "sock_or_null",
390 static char slot_type_char[] = {
391 [STACK_INVALID] = '?',
397 static void print_liveness(struct bpf_verifier_env *env,
398 enum bpf_reg_liveness live)
400 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
402 if (live & REG_LIVE_READ)
404 if (live & REG_LIVE_WRITTEN)
406 if (live & REG_LIVE_DONE)
410 static struct bpf_func_state *func(struct bpf_verifier_env *env,
411 const struct bpf_reg_state *reg)
413 struct bpf_verifier_state *cur = env->cur_state;
415 return cur->frame[reg->frameno];
418 static void print_verifier_state(struct bpf_verifier_env *env,
419 const struct bpf_func_state *state)
421 const struct bpf_reg_state *reg;
426 verbose(env, " frame%d:", state->frameno);
427 for (i = 0; i < MAX_BPF_REG; i++) {
428 reg = &state->regs[i];
432 verbose(env, " R%d", i);
433 print_liveness(env, reg->live);
434 verbose(env, "=%s", reg_type_str[t]);
435 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
436 tnum_is_const(reg->var_off)) {
437 /* reg->off should be 0 for SCALAR_VALUE */
438 verbose(env, "%lld", reg->var_off.value + reg->off);
439 if (t == PTR_TO_STACK)
440 verbose(env, ",call_%d", func(env, reg)->callsite);
442 verbose(env, "(id=%d", reg->id);
443 if (t != SCALAR_VALUE)
444 verbose(env, ",off=%d", reg->off);
445 if (type_is_pkt_pointer(t))
446 verbose(env, ",r=%d", reg->range);
447 else if (t == CONST_PTR_TO_MAP ||
448 t == PTR_TO_MAP_VALUE ||
449 t == PTR_TO_MAP_VALUE_OR_NULL)
450 verbose(env, ",ks=%d,vs=%d",
451 reg->map_ptr->key_size,
452 reg->map_ptr->value_size);
453 if (tnum_is_const(reg->var_off)) {
454 /* Typically an immediate SCALAR_VALUE, but
455 * could be a pointer whose offset is too big
458 verbose(env, ",imm=%llx", reg->var_off.value);
460 if (reg->smin_value != reg->umin_value &&
461 reg->smin_value != S64_MIN)
462 verbose(env, ",smin_value=%lld",
463 (long long)reg->smin_value);
464 if (reg->smax_value != reg->umax_value &&
465 reg->smax_value != S64_MAX)
466 verbose(env, ",smax_value=%lld",
467 (long long)reg->smax_value);
468 if (reg->umin_value != 0)
469 verbose(env, ",umin_value=%llu",
470 (unsigned long long)reg->umin_value);
471 if (reg->umax_value != U64_MAX)
472 verbose(env, ",umax_value=%llu",
473 (unsigned long long)reg->umax_value);
474 if (!tnum_is_unknown(reg->var_off)) {
477 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
478 verbose(env, ",var_off=%s", tn_buf);
484 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
485 char types_buf[BPF_REG_SIZE + 1];
489 for (j = 0; j < BPF_REG_SIZE; j++) {
490 if (state->stack[i].slot_type[j] != STACK_INVALID)
492 types_buf[j] = slot_type_char[
493 state->stack[i].slot_type[j]];
495 types_buf[BPF_REG_SIZE] = 0;
498 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
499 print_liveness(env, state->stack[i].spilled_ptr.live);
500 if (state->stack[i].slot_type[0] == STACK_SPILL)
502 reg_type_str[state->stack[i].spilled_ptr.type]);
504 verbose(env, "=%s", types_buf);
506 if (state->acquired_refs && state->refs[0].id) {
507 verbose(env, " refs=%d", state->refs[0].id);
508 for (i = 1; i < state->acquired_refs; i++)
509 if (state->refs[i].id)
510 verbose(env, ",%d", state->refs[i].id);
515 #define COPY_STATE_FN(NAME, COUNT, FIELD, SIZE) \
516 static int copy_##NAME##_state(struct bpf_func_state *dst, \
517 const struct bpf_func_state *src) \
521 if (WARN_ON_ONCE(dst->COUNT < src->COUNT)) { \
522 /* internal bug, make state invalid to reject the program */ \
523 memset(dst, 0, sizeof(*dst)); \
526 memcpy(dst->FIELD, src->FIELD, \
527 sizeof(*src->FIELD) * (src->COUNT / SIZE)); \
530 /* copy_reference_state() */
531 COPY_STATE_FN(reference, acquired_refs, refs, 1)
532 /* copy_stack_state() */
533 COPY_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
536 #define REALLOC_STATE_FN(NAME, COUNT, FIELD, SIZE) \
537 static int realloc_##NAME##_state(struct bpf_func_state *state, int size, \
540 u32 old_size = state->COUNT; \
541 struct bpf_##NAME##_state *new_##FIELD; \
542 int slot = size / SIZE; \
544 if (size <= old_size || !size) { \
547 state->COUNT = slot * SIZE; \
548 if (!size && old_size) { \
549 kfree(state->FIELD); \
550 state->FIELD = NULL; \
554 new_##FIELD = kmalloc_array(slot, sizeof(struct bpf_##NAME##_state), \
560 memcpy(new_##FIELD, state->FIELD, \
561 sizeof(*new_##FIELD) * (old_size / SIZE)); \
562 memset(new_##FIELD + old_size / SIZE, 0, \
563 sizeof(*new_##FIELD) * (size - old_size) / SIZE); \
565 state->COUNT = slot * SIZE; \
566 kfree(state->FIELD); \
567 state->FIELD = new_##FIELD; \
570 /* realloc_reference_state() */
571 REALLOC_STATE_FN(reference, acquired_refs, refs, 1)
572 /* realloc_stack_state() */
573 REALLOC_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
574 #undef REALLOC_STATE_FN
576 /* do_check() starts with zero-sized stack in struct bpf_verifier_state to
577 * make it consume minimal amount of memory. check_stack_write() access from
578 * the program calls into realloc_func_state() to grow the stack size.
579 * Note there is a non-zero 'parent' pointer inside bpf_verifier_state
580 * which realloc_stack_state() copies over. It points to previous
581 * bpf_verifier_state which is never reallocated.
583 static int realloc_func_state(struct bpf_func_state *state, int stack_size,
584 int refs_size, bool copy_old)
586 int err = realloc_reference_state(state, refs_size, copy_old);
589 return realloc_stack_state(state, stack_size, copy_old);
592 /* Acquire a pointer id from the env and update the state->refs to include
593 * this new pointer reference.
594 * On success, returns a valid pointer id to associate with the register
595 * On failure, returns a negative errno.
597 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
599 struct bpf_func_state *state = cur_func(env);
600 int new_ofs = state->acquired_refs;
603 err = realloc_reference_state(state, state->acquired_refs + 1, true);
607 state->refs[new_ofs].id = id;
608 state->refs[new_ofs].insn_idx = insn_idx;
613 /* release function corresponding to acquire_reference_state(). Idempotent. */
614 static int __release_reference_state(struct bpf_func_state *state, int ptr_id)
621 last_idx = state->acquired_refs - 1;
622 for (i = 0; i < state->acquired_refs; i++) {
623 if (state->refs[i].id == ptr_id) {
624 if (last_idx && i != last_idx)
625 memcpy(&state->refs[i], &state->refs[last_idx],
626 sizeof(*state->refs));
627 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
628 state->acquired_refs--;
635 /* variation on the above for cases where we expect that there must be an
636 * outstanding reference for the specified ptr_id.
638 static int release_reference_state(struct bpf_verifier_env *env, int ptr_id)
640 struct bpf_func_state *state = cur_func(env);
643 err = __release_reference_state(state, ptr_id);
644 if (WARN_ON_ONCE(err != 0))
645 verbose(env, "verifier internal error: can't release reference\n");
649 static int transfer_reference_state(struct bpf_func_state *dst,
650 struct bpf_func_state *src)
652 int err = realloc_reference_state(dst, src->acquired_refs, false);
655 err = copy_reference_state(dst, src);
661 static void free_func_state(struct bpf_func_state *state)
670 static void free_verifier_state(struct bpf_verifier_state *state,
675 for (i = 0; i <= state->curframe; i++) {
676 free_func_state(state->frame[i]);
677 state->frame[i] = NULL;
683 /* copy verifier state from src to dst growing dst stack space
684 * when necessary to accommodate larger src stack
686 static int copy_func_state(struct bpf_func_state *dst,
687 const struct bpf_func_state *src)
691 err = realloc_func_state(dst, src->allocated_stack, src->acquired_refs,
695 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
696 err = copy_reference_state(dst, src);
699 return copy_stack_state(dst, src);
702 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
703 const struct bpf_verifier_state *src)
705 struct bpf_func_state *dst;
708 /* if dst has more stack frames then src frame, free them */
709 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
710 free_func_state(dst_state->frame[i]);
711 dst_state->frame[i] = NULL;
713 dst_state->speculative = src->speculative;
714 dst_state->curframe = src->curframe;
715 for (i = 0; i <= src->curframe; i++) {
716 dst = dst_state->frame[i];
718 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
721 dst_state->frame[i] = dst;
723 err = copy_func_state(dst, src->frame[i]);
730 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
733 struct bpf_verifier_state *cur = env->cur_state;
734 struct bpf_verifier_stack_elem *elem, *head = env->head;
737 if (env->head == NULL)
741 err = copy_verifier_state(cur, &head->st);
746 *insn_idx = head->insn_idx;
748 *prev_insn_idx = head->prev_insn_idx;
750 free_verifier_state(&head->st, false);
757 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
758 int insn_idx, int prev_insn_idx,
761 struct bpf_verifier_state *cur = env->cur_state;
762 struct bpf_verifier_stack_elem *elem;
765 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
769 elem->insn_idx = insn_idx;
770 elem->prev_insn_idx = prev_insn_idx;
771 elem->next = env->head;
774 err = copy_verifier_state(&elem->st, cur);
777 elem->st.speculative |= speculative;
778 if (env->stack_size > BPF_COMPLEXITY_LIMIT_STACK) {
779 verbose(env, "BPF program is too complex\n");
784 free_verifier_state(env->cur_state, true);
785 env->cur_state = NULL;
786 /* pop all elements and return */
787 while (!pop_stack(env, NULL, NULL));
791 #define CALLER_SAVED_REGS 6
792 static const int caller_saved[CALLER_SAVED_REGS] = {
793 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
796 static void __mark_reg_not_init(struct bpf_reg_state *reg);
798 /* Mark the unknown part of a register (variable offset or scalar value) as
799 * known to have the value @imm.
801 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
803 /* Clear id, off, and union(map_ptr, range) */
804 memset(((u8 *)reg) + sizeof(reg->type), 0,
805 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
806 reg->var_off = tnum_const(imm);
807 reg->smin_value = (s64)imm;
808 reg->smax_value = (s64)imm;
809 reg->umin_value = imm;
810 reg->umax_value = imm;
813 /* Mark the 'variable offset' part of a register as zero. This should be
814 * used only on registers holding a pointer type.
816 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
818 __mark_reg_known(reg, 0);
821 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
823 __mark_reg_known(reg, 0);
824 reg->type = SCALAR_VALUE;
827 static void mark_reg_known_zero(struct bpf_verifier_env *env,
828 struct bpf_reg_state *regs, u32 regno)
830 if (WARN_ON(regno >= MAX_BPF_REG)) {
831 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
832 /* Something bad happened, let's kill all regs */
833 for (regno = 0; regno < MAX_BPF_REG; regno++)
834 __mark_reg_not_init(regs + regno);
837 __mark_reg_known_zero(regs + regno);
840 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
842 return type_is_pkt_pointer(reg->type);
845 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
847 return reg_is_pkt_pointer(reg) ||
848 reg->type == PTR_TO_PACKET_END;
851 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
852 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
853 enum bpf_reg_type which)
855 /* The register can already have a range from prior markings.
856 * This is fine as long as it hasn't been advanced from its
859 return reg->type == which &&
862 tnum_equals_const(reg->var_off, 0);
865 /* Attempts to improve min/max values based on var_off information */
866 static void __update_reg_bounds(struct bpf_reg_state *reg)
868 /* min signed is max(sign bit) | min(other bits) */
869 reg->smin_value = max_t(s64, reg->smin_value,
870 reg->var_off.value | (reg->var_off.mask & S64_MIN));
871 /* max signed is min(sign bit) | max(other bits) */
872 reg->smax_value = min_t(s64, reg->smax_value,
873 reg->var_off.value | (reg->var_off.mask & S64_MAX));
874 reg->umin_value = max(reg->umin_value, reg->var_off.value);
875 reg->umax_value = min(reg->umax_value,
876 reg->var_off.value | reg->var_off.mask);
879 /* Uses signed min/max values to inform unsigned, and vice-versa */
880 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
882 /* Learn sign from signed bounds.
883 * If we cannot cross the sign boundary, then signed and unsigned bounds
884 * are the same, so combine. This works even in the negative case, e.g.
885 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
887 if (reg->smin_value >= 0 || reg->smax_value < 0) {
888 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
890 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
894 /* Learn sign from unsigned bounds. Signed bounds cross the sign
895 * boundary, so we must be careful.
897 if ((s64)reg->umax_value >= 0) {
898 /* Positive. We can't learn anything from the smin, but smax
899 * is positive, hence safe.
901 reg->smin_value = reg->umin_value;
902 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
904 } else if ((s64)reg->umin_value < 0) {
905 /* Negative. We can't learn anything from the smax, but smin
906 * is negative, hence safe.
908 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
910 reg->smax_value = reg->umax_value;
914 /* Attempts to improve var_off based on unsigned min/max information */
915 static void __reg_bound_offset(struct bpf_reg_state *reg)
917 reg->var_off = tnum_intersect(reg->var_off,
918 tnum_range(reg->umin_value,
922 /* Reset the min/max bounds of a register */
923 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
925 reg->smin_value = S64_MIN;
926 reg->smax_value = S64_MAX;
928 reg->umax_value = U64_MAX;
931 /* Mark a register as having a completely unknown (scalar) value. */
932 static void __mark_reg_unknown(struct bpf_reg_state *reg)
935 * Clear type, id, off, and union(map_ptr, range) and
936 * padding between 'type' and union
938 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
939 reg->type = SCALAR_VALUE;
940 reg->var_off = tnum_unknown;
942 __mark_reg_unbounded(reg);
945 static void mark_reg_unknown(struct bpf_verifier_env *env,
946 struct bpf_reg_state *regs, u32 regno)
948 if (WARN_ON(regno >= MAX_BPF_REG)) {
949 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
950 /* Something bad happened, let's kill all regs except FP */
951 for (regno = 0; regno < BPF_REG_FP; regno++)
952 __mark_reg_not_init(regs + regno);
955 __mark_reg_unknown(regs + regno);
958 static void __mark_reg_not_init(struct bpf_reg_state *reg)
960 __mark_reg_unknown(reg);
961 reg->type = NOT_INIT;
964 static void mark_reg_not_init(struct bpf_verifier_env *env,
965 struct bpf_reg_state *regs, u32 regno)
967 if (WARN_ON(regno >= MAX_BPF_REG)) {
968 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
969 /* Something bad happened, let's kill all regs except FP */
970 for (regno = 0; regno < BPF_REG_FP; regno++)
971 __mark_reg_not_init(regs + regno);
974 __mark_reg_not_init(regs + regno);
977 static void init_reg_state(struct bpf_verifier_env *env,
978 struct bpf_func_state *state)
980 struct bpf_reg_state *regs = state->regs;
983 for (i = 0; i < MAX_BPF_REG; i++) {
984 mark_reg_not_init(env, regs, i);
985 regs[i].live = REG_LIVE_NONE;
986 regs[i].parent = NULL;
990 regs[BPF_REG_FP].type = PTR_TO_STACK;
991 mark_reg_known_zero(env, regs, BPF_REG_FP);
992 regs[BPF_REG_FP].frameno = state->frameno;
994 /* 1st arg to a function */
995 regs[BPF_REG_1].type = PTR_TO_CTX;
996 mark_reg_known_zero(env, regs, BPF_REG_1);
999 #define BPF_MAIN_FUNC (-1)
1000 static void init_func_state(struct bpf_verifier_env *env,
1001 struct bpf_func_state *state,
1002 int callsite, int frameno, int subprogno)
1004 state->callsite = callsite;
1005 state->frameno = frameno;
1006 state->subprogno = subprogno;
1007 init_reg_state(env, state);
1011 SRC_OP, /* register is used as source operand */
1012 DST_OP, /* register is used as destination operand */
1013 DST_OP_NO_MARK /* same as above, check only, don't mark */
1016 static int cmp_subprogs(const void *a, const void *b)
1018 return ((struct bpf_subprog_info *)a)->start -
1019 ((struct bpf_subprog_info *)b)->start;
1022 static int find_subprog(struct bpf_verifier_env *env, int off)
1024 struct bpf_subprog_info *p;
1026 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
1027 sizeof(env->subprog_info[0]), cmp_subprogs);
1030 return p - env->subprog_info;
1034 static int add_subprog(struct bpf_verifier_env *env, int off)
1036 int insn_cnt = env->prog->len;
1039 if (off >= insn_cnt || off < 0) {
1040 verbose(env, "call to invalid destination\n");
1043 ret = find_subprog(env, off);
1046 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
1047 verbose(env, "too many subprograms\n");
1050 env->subprog_info[env->subprog_cnt++].start = off;
1051 sort(env->subprog_info, env->subprog_cnt,
1052 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
1056 static int check_subprogs(struct bpf_verifier_env *env)
1058 int i, ret, subprog_start, subprog_end, off, cur_subprog = 0;
1059 struct bpf_subprog_info *subprog = env->subprog_info;
1060 struct bpf_insn *insn = env->prog->insnsi;
1061 int insn_cnt = env->prog->len;
1063 /* Add entry function. */
1064 ret = add_subprog(env, 0);
1068 /* determine subprog starts. The end is one before the next starts */
1069 for (i = 0; i < insn_cnt; i++) {
1070 if (insn[i].code != (BPF_JMP | BPF_CALL))
1072 if (insn[i].src_reg != BPF_PSEUDO_CALL)
1074 if (!env->allow_ptr_leaks) {
1075 verbose(env, "function calls to other bpf functions are allowed for root only\n");
1078 ret = add_subprog(env, i + insn[i].imm + 1);
1083 /* Add a fake 'exit' subprog which could simplify subprog iteration
1084 * logic. 'subprog_cnt' should not be increased.
1086 subprog[env->subprog_cnt].start = insn_cnt;
1088 if (env->log.level > 1)
1089 for (i = 0; i < env->subprog_cnt; i++)
1090 verbose(env, "func#%d @%d\n", i, subprog[i].start);
1092 /* now check that all jumps are within the same subprog */
1093 subprog_start = subprog[cur_subprog].start;
1094 subprog_end = subprog[cur_subprog + 1].start;
1095 for (i = 0; i < insn_cnt; i++) {
1096 u8 code = insn[i].code;
1098 if (BPF_CLASS(code) != BPF_JMP)
1100 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
1102 off = i + insn[i].off + 1;
1103 if (off < subprog_start || off >= subprog_end) {
1104 verbose(env, "jump out of range from insn %d to %d\n", i, off);
1108 if (i == subprog_end - 1) {
1109 /* to avoid fall-through from one subprog into another
1110 * the last insn of the subprog should be either exit
1111 * or unconditional jump back
1113 if (code != (BPF_JMP | BPF_EXIT) &&
1114 code != (BPF_JMP | BPF_JA)) {
1115 verbose(env, "last insn is not an exit or jmp\n");
1118 subprog_start = subprog_end;
1120 if (cur_subprog < env->subprog_cnt)
1121 subprog_end = subprog[cur_subprog + 1].start;
1127 /* Parentage chain of this register (or stack slot) should take care of all
1128 * issues like callee-saved registers, stack slot allocation time, etc.
1130 static int mark_reg_read(struct bpf_verifier_env *env,
1131 const struct bpf_reg_state *state,
1132 struct bpf_reg_state *parent)
1134 bool writes = parent == state->parent; /* Observe write marks */
1137 /* if read wasn't screened by an earlier write ... */
1138 if (writes && state->live & REG_LIVE_WRITTEN)
1140 if (parent->live & REG_LIVE_DONE) {
1141 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
1142 reg_type_str[parent->type],
1143 parent->var_off.value, parent->off);
1146 /* ... then we depend on parent's value */
1147 parent->live |= REG_LIVE_READ;
1149 parent = state->parent;
1155 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
1156 enum reg_arg_type t)
1158 struct bpf_verifier_state *vstate = env->cur_state;
1159 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1160 struct bpf_reg_state *regs = state->regs;
1162 if (regno >= MAX_BPF_REG) {
1163 verbose(env, "R%d is invalid\n", regno);
1168 /* check whether register used as source operand can be read */
1169 if (regs[regno].type == NOT_INIT) {
1170 verbose(env, "R%d !read_ok\n", regno);
1173 /* We don't need to worry about FP liveness because it's read-only */
1174 if (regno != BPF_REG_FP)
1175 return mark_reg_read(env, ®s[regno],
1176 regs[regno].parent);
1178 /* check whether register used as dest operand can be written to */
1179 if (regno == BPF_REG_FP) {
1180 verbose(env, "frame pointer is read only\n");
1183 regs[regno].live |= REG_LIVE_WRITTEN;
1185 mark_reg_unknown(env, regs, regno);
1190 static bool is_spillable_regtype(enum bpf_reg_type type)
1193 case PTR_TO_MAP_VALUE:
1194 case PTR_TO_MAP_VALUE_OR_NULL:
1198 case PTR_TO_PACKET_META:
1199 case PTR_TO_PACKET_END:
1200 case PTR_TO_FLOW_KEYS:
1201 case CONST_PTR_TO_MAP:
1203 case PTR_TO_SOCKET_OR_NULL:
1210 /* Does this register contain a constant zero? */
1211 static bool register_is_null(struct bpf_reg_state *reg)
1213 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
1216 /* check_stack_read/write functions track spill/fill of registers,
1217 * stack boundary and alignment are checked in check_mem_access()
1219 static int check_stack_write(struct bpf_verifier_env *env,
1220 struct bpf_func_state *state, /* func where register points to */
1221 int off, int size, int value_regno, int insn_idx)
1223 struct bpf_func_state *cur; /* state of the current function */
1224 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
1225 enum bpf_reg_type type;
1227 err = realloc_func_state(state, round_up(slot + 1, BPF_REG_SIZE),
1228 state->acquired_refs, true);
1231 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
1232 * so it's aligned access and [off, off + size) are within stack limits
1234 if (!env->allow_ptr_leaks &&
1235 state->stack[spi].slot_type[0] == STACK_SPILL &&
1236 size != BPF_REG_SIZE) {
1237 verbose(env, "attempt to corrupt spilled pointer on stack\n");
1241 cur = env->cur_state->frame[env->cur_state->curframe];
1242 if (value_regno >= 0 &&
1243 is_spillable_regtype((type = cur->regs[value_regno].type))) {
1245 /* register containing pointer is being spilled into stack */
1246 if (size != BPF_REG_SIZE) {
1247 verbose(env, "invalid size of register spill\n");
1251 if (state != cur && type == PTR_TO_STACK) {
1252 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
1256 /* save register state */
1257 state->stack[spi].spilled_ptr = cur->regs[value_regno];
1258 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1260 for (i = 0; i < BPF_REG_SIZE; i++) {
1261 if (state->stack[spi].slot_type[i] == STACK_MISC &&
1262 !env->allow_ptr_leaks) {
1263 int *poff = &env->insn_aux_data[insn_idx].sanitize_stack_off;
1264 int soff = (-spi - 1) * BPF_REG_SIZE;
1266 /* detected reuse of integer stack slot with a pointer
1267 * which means either llvm is reusing stack slot or
1268 * an attacker is trying to exploit CVE-2018-3639
1269 * (speculative store bypass)
1270 * Have to sanitize that slot with preemptive
1273 if (*poff && *poff != soff) {
1274 /* disallow programs where single insn stores
1275 * into two different stack slots, since verifier
1276 * cannot sanitize them
1279 "insn %d cannot access two stack slots fp%d and fp%d",
1280 insn_idx, *poff, soff);
1285 state->stack[spi].slot_type[i] = STACK_SPILL;
1288 u8 type = STACK_MISC;
1290 /* regular write of data into stack destroys any spilled ptr */
1291 state->stack[spi].spilled_ptr.type = NOT_INIT;
1292 /* Mark slots as STACK_MISC if they belonged to spilled ptr. */
1293 if (state->stack[spi].slot_type[0] == STACK_SPILL)
1294 for (i = 0; i < BPF_REG_SIZE; i++)
1295 state->stack[spi].slot_type[i] = STACK_MISC;
1297 /* only mark the slot as written if all 8 bytes were written
1298 * otherwise read propagation may incorrectly stop too soon
1299 * when stack slots are partially written.
1300 * This heuristic means that read propagation will be
1301 * conservative, since it will add reg_live_read marks
1302 * to stack slots all the way to first state when programs
1303 * writes+reads less than 8 bytes
1305 if (size == BPF_REG_SIZE)
1306 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1308 /* when we zero initialize stack slots mark them as such */
1309 if (value_regno >= 0 &&
1310 register_is_null(&cur->regs[value_regno]))
1313 /* Mark slots affected by this stack write. */
1314 for (i = 0; i < size; i++)
1315 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
1321 static int check_stack_read(struct bpf_verifier_env *env,
1322 struct bpf_func_state *reg_state /* func where register points to */,
1323 int off, int size, int value_regno)
1325 struct bpf_verifier_state *vstate = env->cur_state;
1326 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1327 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
1330 if (reg_state->allocated_stack <= slot) {
1331 verbose(env, "invalid read from stack off %d+0 size %d\n",
1335 stype = reg_state->stack[spi].slot_type;
1337 if (stype[0] == STACK_SPILL) {
1338 if (size != BPF_REG_SIZE) {
1339 verbose(env, "invalid size of register spill\n");
1342 for (i = 1; i < BPF_REG_SIZE; i++) {
1343 if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) {
1344 verbose(env, "corrupted spill memory\n");
1349 if (value_regno >= 0) {
1350 /* restore register state from stack */
1351 state->regs[value_regno] = reg_state->stack[spi].spilled_ptr;
1352 /* mark reg as written since spilled pointer state likely
1353 * has its liveness marks cleared by is_state_visited()
1354 * which resets stack/reg liveness for state transitions
1356 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
1358 mark_reg_read(env, ®_state->stack[spi].spilled_ptr,
1359 reg_state->stack[spi].spilled_ptr.parent);
1364 for (i = 0; i < size; i++) {
1365 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_MISC)
1367 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_ZERO) {
1371 verbose(env, "invalid read from stack off %d+%d size %d\n",
1375 mark_reg_read(env, ®_state->stack[spi].spilled_ptr,
1376 reg_state->stack[spi].spilled_ptr.parent);
1377 if (value_regno >= 0) {
1378 if (zeros == size) {
1379 /* any size read into register is zero extended,
1380 * so the whole register == const_zero
1382 __mark_reg_const_zero(&state->regs[value_regno]);
1384 /* have read misc data from the stack */
1385 mark_reg_unknown(env, state->regs, value_regno);
1387 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
1393 static int check_stack_access(struct bpf_verifier_env *env,
1394 const struct bpf_reg_state *reg,
1397 /* Stack accesses must be at a fixed offset, so that we
1398 * can determine what type of data were returned. See
1399 * check_stack_read().
1401 if (!tnum_is_const(reg->var_off)) {
1404 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1405 verbose(env, "variable stack access var_off=%s off=%d size=%d",
1410 if (off >= 0 || off < -MAX_BPF_STACK) {
1411 verbose(env, "invalid stack off=%d size=%d\n", off, size);
1418 /* check read/write into map element returned by bpf_map_lookup_elem() */
1419 static int __check_map_access(struct bpf_verifier_env *env, u32 regno, int off,
1420 int size, bool zero_size_allowed)
1422 struct bpf_reg_state *regs = cur_regs(env);
1423 struct bpf_map *map = regs[regno].map_ptr;
1425 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
1426 off + size > map->value_size) {
1427 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
1428 map->value_size, off, size);
1434 /* check read/write into a map element with possible variable offset */
1435 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
1436 int off, int size, bool zero_size_allowed)
1438 struct bpf_verifier_state *vstate = env->cur_state;
1439 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1440 struct bpf_reg_state *reg = &state->regs[regno];
1443 /* We may have adjusted the register to this map value, so we
1444 * need to try adding each of min_value and max_value to off
1445 * to make sure our theoretical access will be safe.
1448 print_verifier_state(env, state);
1450 /* The minimum value is only important with signed
1451 * comparisons where we can't assume the floor of a
1452 * value is 0. If we are using signed variables for our
1453 * index'es we need to make sure that whatever we use
1454 * will have a set floor within our range.
1456 if (reg->smin_value < 0 &&
1457 (reg->smin_value == S64_MIN ||
1458 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
1459 reg->smin_value + off < 0)) {
1460 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1464 err = __check_map_access(env, regno, reg->smin_value + off, size,
1467 verbose(env, "R%d min value is outside of the array range\n",
1472 /* If we haven't set a max value then we need to bail since we can't be
1473 * sure we won't do bad things.
1474 * If reg->umax_value + off could overflow, treat that as unbounded too.
1476 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
1477 verbose(env, "R%d unbounded memory access, make sure to bounds check any array access into a map\n",
1481 err = __check_map_access(env, regno, reg->umax_value + off, size,
1484 verbose(env, "R%d max value is outside of the array range\n",
1489 #define MAX_PACKET_OFF 0xffff
1491 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
1492 const struct bpf_call_arg_meta *meta,
1493 enum bpf_access_type t)
1495 switch (env->prog->type) {
1496 /* Program types only with direct read access go here! */
1497 case BPF_PROG_TYPE_LWT_IN:
1498 case BPF_PROG_TYPE_LWT_OUT:
1499 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
1500 case BPF_PROG_TYPE_SK_REUSEPORT:
1501 case BPF_PROG_TYPE_FLOW_DISSECTOR:
1502 case BPF_PROG_TYPE_CGROUP_SKB:
1507 /* Program types with direct read + write access go here! */
1508 case BPF_PROG_TYPE_SCHED_CLS:
1509 case BPF_PROG_TYPE_SCHED_ACT:
1510 case BPF_PROG_TYPE_XDP:
1511 case BPF_PROG_TYPE_LWT_XMIT:
1512 case BPF_PROG_TYPE_SK_SKB:
1513 case BPF_PROG_TYPE_SK_MSG:
1515 return meta->pkt_access;
1517 env->seen_direct_write = true;
1524 static int __check_packet_access(struct bpf_verifier_env *env, u32 regno,
1525 int off, int size, bool zero_size_allowed)
1527 struct bpf_reg_state *regs = cur_regs(env);
1528 struct bpf_reg_state *reg = ®s[regno];
1530 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
1531 (u64)off + size > reg->range) {
1532 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
1533 off, size, regno, reg->id, reg->off, reg->range);
1539 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
1540 int size, bool zero_size_allowed)
1542 struct bpf_reg_state *regs = cur_regs(env);
1543 struct bpf_reg_state *reg = ®s[regno];
1546 /* We may have added a variable offset to the packet pointer; but any
1547 * reg->range we have comes after that. We are only checking the fixed
1551 /* We don't allow negative numbers, because we aren't tracking enough
1552 * detail to prove they're safe.
1554 if (reg->smin_value < 0) {
1555 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1559 err = __check_packet_access(env, regno, off, size, zero_size_allowed);
1561 verbose(env, "R%d offset is outside of the packet\n", regno);
1565 /* __check_packet_access has made sure "off + size - 1" is within u16.
1566 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
1567 * otherwise find_good_pkt_pointers would have refused to set range info
1568 * that __check_packet_access would have rejected this pkt access.
1569 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
1571 env->prog->aux->max_pkt_offset =
1572 max_t(u32, env->prog->aux->max_pkt_offset,
1573 off + reg->umax_value + size - 1);
1578 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
1579 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
1580 enum bpf_access_type t, enum bpf_reg_type *reg_type)
1582 struct bpf_insn_access_aux info = {
1583 .reg_type = *reg_type,
1586 if (env->ops->is_valid_access &&
1587 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
1588 /* A non zero info.ctx_field_size indicates that this field is a
1589 * candidate for later verifier transformation to load the whole
1590 * field and then apply a mask when accessed with a narrower
1591 * access than actual ctx access size. A zero info.ctx_field_size
1592 * will only allow for whole field access and rejects any other
1593 * type of narrower access.
1595 *reg_type = info.reg_type;
1597 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
1598 /* remember the offset of last byte accessed in ctx */
1599 if (env->prog->aux->max_ctx_offset < off + size)
1600 env->prog->aux->max_ctx_offset = off + size;
1604 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
1608 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
1611 if (size < 0 || off < 0 ||
1612 (u64)off + size > sizeof(struct bpf_flow_keys)) {
1613 verbose(env, "invalid access to flow keys off=%d size=%d\n",
1620 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
1621 u32 regno, int off, int size,
1622 enum bpf_access_type t)
1624 struct bpf_reg_state *regs = cur_regs(env);
1625 struct bpf_reg_state *reg = ®s[regno];
1626 struct bpf_insn_access_aux info = {};
1628 if (reg->smin_value < 0) {
1629 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1634 if (!bpf_sock_is_valid_access(off, size, t, &info)) {
1635 verbose(env, "invalid bpf_sock access off=%d size=%d\n",
1640 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
1645 static bool __is_pointer_value(bool allow_ptr_leaks,
1646 const struct bpf_reg_state *reg)
1648 if (allow_ptr_leaks)
1651 return reg->type != SCALAR_VALUE;
1654 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
1656 return cur_regs(env) + regno;
1659 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
1661 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
1664 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
1666 const struct bpf_reg_state *reg = reg_state(env, regno);
1668 return reg->type == PTR_TO_CTX ||
1669 reg->type == PTR_TO_SOCKET;
1672 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
1674 const struct bpf_reg_state *reg = reg_state(env, regno);
1676 return type_is_pkt_pointer(reg->type);
1679 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
1681 const struct bpf_reg_state *reg = reg_state(env, regno);
1683 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
1684 return reg->type == PTR_TO_FLOW_KEYS;
1687 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
1688 const struct bpf_reg_state *reg,
1689 int off, int size, bool strict)
1691 struct tnum reg_off;
1694 /* Byte size accesses are always allowed. */
1695 if (!strict || size == 1)
1698 /* For platforms that do not have a Kconfig enabling
1699 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
1700 * NET_IP_ALIGN is universally set to '2'. And on platforms
1701 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
1702 * to this code only in strict mode where we want to emulate
1703 * the NET_IP_ALIGN==2 checking. Therefore use an
1704 * unconditional IP align value of '2'.
1708 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
1709 if (!tnum_is_aligned(reg_off, size)) {
1712 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1714 "misaligned packet access off %d+%s+%d+%d size %d\n",
1715 ip_align, tn_buf, reg->off, off, size);
1722 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
1723 const struct bpf_reg_state *reg,
1724 const char *pointer_desc,
1725 int off, int size, bool strict)
1727 struct tnum reg_off;
1729 /* Byte size accesses are always allowed. */
1730 if (!strict || size == 1)
1733 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
1734 if (!tnum_is_aligned(reg_off, size)) {
1737 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1738 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
1739 pointer_desc, tn_buf, reg->off, off, size);
1746 static int check_ptr_alignment(struct bpf_verifier_env *env,
1747 const struct bpf_reg_state *reg, int off,
1748 int size, bool strict_alignment_once)
1750 bool strict = env->strict_alignment || strict_alignment_once;
1751 const char *pointer_desc = "";
1753 switch (reg->type) {
1755 case PTR_TO_PACKET_META:
1756 /* Special case, because of NET_IP_ALIGN. Given metadata sits
1757 * right in front, treat it the very same way.
1759 return check_pkt_ptr_alignment(env, reg, off, size, strict);
1760 case PTR_TO_FLOW_KEYS:
1761 pointer_desc = "flow keys ";
1763 case PTR_TO_MAP_VALUE:
1764 pointer_desc = "value ";
1767 pointer_desc = "context ";
1770 pointer_desc = "stack ";
1771 /* The stack spill tracking logic in check_stack_write()
1772 * and check_stack_read() relies on stack accesses being
1778 pointer_desc = "sock ";
1783 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
1787 static int update_stack_depth(struct bpf_verifier_env *env,
1788 const struct bpf_func_state *func,
1791 u16 stack = env->subprog_info[func->subprogno].stack_depth;
1796 /* update known max for given subprogram */
1797 env->subprog_info[func->subprogno].stack_depth = -off;
1801 /* starting from main bpf function walk all instructions of the function
1802 * and recursively walk all callees that given function can call.
1803 * Ignore jump and exit insns.
1804 * Since recursion is prevented by check_cfg() this algorithm
1805 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
1807 static int check_max_stack_depth(struct bpf_verifier_env *env)
1809 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
1810 struct bpf_subprog_info *subprog = env->subprog_info;
1811 struct bpf_insn *insn = env->prog->insnsi;
1812 int ret_insn[MAX_CALL_FRAMES];
1813 int ret_prog[MAX_CALL_FRAMES];
1816 /* round up to 32-bytes, since this is granularity
1817 * of interpreter stack size
1819 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
1820 if (depth > MAX_BPF_STACK) {
1821 verbose(env, "combined stack size of %d calls is %d. Too large\n",
1826 subprog_end = subprog[idx + 1].start;
1827 for (; i < subprog_end; i++) {
1828 if (insn[i].code != (BPF_JMP | BPF_CALL))
1830 if (insn[i].src_reg != BPF_PSEUDO_CALL)
1832 /* remember insn and function to return to */
1833 ret_insn[frame] = i + 1;
1834 ret_prog[frame] = idx;
1836 /* find the callee */
1837 i = i + insn[i].imm + 1;
1838 idx = find_subprog(env, i);
1840 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
1845 if (frame >= MAX_CALL_FRAMES) {
1846 WARN_ONCE(1, "verifier bug. Call stack is too deep\n");
1851 /* end of for() loop means the last insn of the 'subprog'
1852 * was reached. Doesn't matter whether it was JA or EXIT
1856 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
1858 i = ret_insn[frame];
1859 idx = ret_prog[frame];
1863 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
1864 static int get_callee_stack_depth(struct bpf_verifier_env *env,
1865 const struct bpf_insn *insn, int idx)
1867 int start = idx + insn->imm + 1, subprog;
1869 subprog = find_subprog(env, start);
1871 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
1875 return env->subprog_info[subprog].stack_depth;
1879 static int check_ctx_reg(struct bpf_verifier_env *env,
1880 const struct bpf_reg_state *reg, int regno)
1882 /* Access to ctx or passing it to a helper is only allowed in
1883 * its original, unmodified form.
1887 verbose(env, "dereference of modified ctx ptr R%d off=%d disallowed\n",
1892 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
1895 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1896 verbose(env, "variable ctx access var_off=%s disallowed\n", tn_buf);
1903 /* truncate register to smaller size (in bytes)
1904 * must be called with size < BPF_REG_SIZE
1906 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
1910 /* clear high bits in bit representation */
1911 reg->var_off = tnum_cast(reg->var_off, size);
1913 /* fix arithmetic bounds */
1914 mask = ((u64)1 << (size * 8)) - 1;
1915 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
1916 reg->umin_value &= mask;
1917 reg->umax_value &= mask;
1919 reg->umin_value = 0;
1920 reg->umax_value = mask;
1922 reg->smin_value = reg->umin_value;
1923 reg->smax_value = reg->umax_value;
1926 /* check whether memory at (regno + off) is accessible for t = (read | write)
1927 * if t==write, value_regno is a register which value is stored into memory
1928 * if t==read, value_regno is a register which will receive the value from memory
1929 * if t==write && value_regno==-1, some unknown value is stored into memory
1930 * if t==read && value_regno==-1, don't care what we read from memory
1932 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
1933 int off, int bpf_size, enum bpf_access_type t,
1934 int value_regno, bool strict_alignment_once)
1936 struct bpf_reg_state *regs = cur_regs(env);
1937 struct bpf_reg_state *reg = regs + regno;
1938 struct bpf_func_state *state;
1941 size = bpf_size_to_bytes(bpf_size);
1945 /* alignment checks will add in reg->off themselves */
1946 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
1950 /* for access checks, reg->off is just part of off */
1953 if (reg->type == PTR_TO_MAP_VALUE) {
1954 if (t == BPF_WRITE && value_regno >= 0 &&
1955 is_pointer_value(env, value_regno)) {
1956 verbose(env, "R%d leaks addr into map\n", value_regno);
1960 err = check_map_access(env, regno, off, size, false);
1961 if (!err && t == BPF_READ && value_regno >= 0)
1962 mark_reg_unknown(env, regs, value_regno);
1964 } else if (reg->type == PTR_TO_CTX) {
1965 enum bpf_reg_type reg_type = SCALAR_VALUE;
1967 if (t == BPF_WRITE && value_regno >= 0 &&
1968 is_pointer_value(env, value_regno)) {
1969 verbose(env, "R%d leaks addr into ctx\n", value_regno);
1973 err = check_ctx_reg(env, reg, regno);
1977 err = check_ctx_access(env, insn_idx, off, size, t, ®_type);
1978 if (!err && t == BPF_READ && value_regno >= 0) {
1979 /* ctx access returns either a scalar, or a
1980 * PTR_TO_PACKET[_META,_END]. In the latter
1981 * case, we know the offset is zero.
1983 if (reg_type == SCALAR_VALUE)
1984 mark_reg_unknown(env, regs, value_regno);
1986 mark_reg_known_zero(env, regs,
1988 regs[value_regno].type = reg_type;
1991 } else if (reg->type == PTR_TO_STACK) {
1992 off += reg->var_off.value;
1993 err = check_stack_access(env, reg, off, size);
1997 state = func(env, reg);
1998 err = update_stack_depth(env, state, off);
2003 err = check_stack_write(env, state, off, size,
2004 value_regno, insn_idx);
2006 err = check_stack_read(env, state, off, size,
2008 } else if (reg_is_pkt_pointer(reg)) {
2009 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
2010 verbose(env, "cannot write into packet\n");
2013 if (t == BPF_WRITE && value_regno >= 0 &&
2014 is_pointer_value(env, value_regno)) {
2015 verbose(env, "R%d leaks addr into packet\n",
2019 err = check_packet_access(env, regno, off, size, false);
2020 if (!err && t == BPF_READ && value_regno >= 0)
2021 mark_reg_unknown(env, regs, value_regno);
2022 } else if (reg->type == PTR_TO_FLOW_KEYS) {
2023 if (t == BPF_WRITE && value_regno >= 0 &&
2024 is_pointer_value(env, value_regno)) {
2025 verbose(env, "R%d leaks addr into flow keys\n",
2030 err = check_flow_keys_access(env, off, size);
2031 if (!err && t == BPF_READ && value_regno >= 0)
2032 mark_reg_unknown(env, regs, value_regno);
2033 } else if (reg->type == PTR_TO_SOCKET) {
2034 if (t == BPF_WRITE) {
2035 verbose(env, "cannot write into socket\n");
2038 err = check_sock_access(env, insn_idx, regno, off, size, t);
2039 if (!err && value_regno >= 0)
2040 mark_reg_unknown(env, regs, value_regno);
2042 verbose(env, "R%d invalid mem access '%s'\n", regno,
2043 reg_type_str[reg->type]);
2047 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
2048 regs[value_regno].type == SCALAR_VALUE) {
2049 /* b/h/w load zero-extends, mark upper bits as known 0 */
2050 coerce_reg_to_size(®s[value_regno], size);
2055 static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
2059 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
2061 verbose(env, "BPF_XADD uses reserved fields\n");
2065 /* check src1 operand */
2066 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2070 /* check src2 operand */
2071 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2075 if (is_pointer_value(env, insn->src_reg)) {
2076 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
2080 if (is_ctx_reg(env, insn->dst_reg) ||
2081 is_pkt_reg(env, insn->dst_reg) ||
2082 is_flow_key_reg(env, insn->dst_reg)) {
2083 verbose(env, "BPF_XADD stores into R%d %s is not allowed\n",
2085 reg_type_str[reg_state(env, insn->dst_reg)->type]);
2089 /* check whether atomic_add can read the memory */
2090 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
2091 BPF_SIZE(insn->code), BPF_READ, -1, true);
2095 /* check whether atomic_add can write into the same memory */
2096 return check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
2097 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
2100 /* when register 'regno' is passed into function that will read 'access_size'
2101 * bytes from that pointer, make sure that it's within stack boundary
2102 * and all elements of stack are initialized.
2103 * Unlike most pointer bounds-checking functions, this one doesn't take an
2104 * 'off' argument, so it has to add in reg->off itself.
2106 static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
2107 int access_size, bool zero_size_allowed,
2108 struct bpf_call_arg_meta *meta)
2110 struct bpf_reg_state *reg = reg_state(env, regno);
2111 struct bpf_func_state *state = func(env, reg);
2112 int off, i, slot, spi;
2114 if (reg->type != PTR_TO_STACK) {
2115 /* Allow zero-byte read from NULL, regardless of pointer type */
2116 if (zero_size_allowed && access_size == 0 &&
2117 register_is_null(reg))
2120 verbose(env, "R%d type=%s expected=%s\n", regno,
2121 reg_type_str[reg->type],
2122 reg_type_str[PTR_TO_STACK]);
2126 /* Only allow fixed-offset stack reads */
2127 if (!tnum_is_const(reg->var_off)) {
2130 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2131 verbose(env, "invalid variable stack read R%d var_off=%s\n",
2135 off = reg->off + reg->var_off.value;
2136 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
2137 access_size < 0 || (access_size == 0 && !zero_size_allowed)) {
2138 verbose(env, "invalid stack type R%d off=%d access_size=%d\n",
2139 regno, off, access_size);
2143 if (meta && meta->raw_mode) {
2144 meta->access_size = access_size;
2145 meta->regno = regno;
2149 for (i = 0; i < access_size; i++) {
2152 slot = -(off + i) - 1;
2153 spi = slot / BPF_REG_SIZE;
2154 if (state->allocated_stack <= slot)
2156 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
2157 if (*stype == STACK_MISC)
2159 if (*stype == STACK_ZERO) {
2160 /* helper can write anything into the stack */
2161 *stype = STACK_MISC;
2165 verbose(env, "invalid indirect read from stack off %d+%d size %d\n",
2166 off, i, access_size);
2169 /* reading any byte out of 8-byte 'spill_slot' will cause
2170 * the whole slot to be marked as 'read'
2172 mark_reg_read(env, &state->stack[spi].spilled_ptr,
2173 state->stack[spi].spilled_ptr.parent);
2175 return update_stack_depth(env, state, off);
2178 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
2179 int access_size, bool zero_size_allowed,
2180 struct bpf_call_arg_meta *meta)
2182 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
2184 switch (reg->type) {
2186 case PTR_TO_PACKET_META:
2187 return check_packet_access(env, regno, reg->off, access_size,
2189 case PTR_TO_MAP_VALUE:
2190 return check_map_access(env, regno, reg->off, access_size,
2192 default: /* scalar_value|ptr_to_stack or invalid ptr */
2193 return check_stack_boundary(env, regno, access_size,
2194 zero_size_allowed, meta);
2198 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
2200 return type == ARG_PTR_TO_MEM ||
2201 type == ARG_PTR_TO_MEM_OR_NULL ||
2202 type == ARG_PTR_TO_UNINIT_MEM;
2205 static bool arg_type_is_mem_size(enum bpf_arg_type type)
2207 return type == ARG_CONST_SIZE ||
2208 type == ARG_CONST_SIZE_OR_ZERO;
2211 static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
2212 enum bpf_arg_type arg_type,
2213 struct bpf_call_arg_meta *meta)
2215 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
2216 enum bpf_reg_type expected_type, type = reg->type;
2219 if (arg_type == ARG_DONTCARE)
2222 err = check_reg_arg(env, regno, SRC_OP);
2226 if (arg_type == ARG_ANYTHING) {
2227 if (is_pointer_value(env, regno)) {
2228 verbose(env, "R%d leaks addr into helper function\n",
2235 if (type_is_pkt_pointer(type) &&
2236 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
2237 verbose(env, "helper access to the packet is not allowed\n");
2241 if (arg_type == ARG_PTR_TO_MAP_KEY ||
2242 arg_type == ARG_PTR_TO_MAP_VALUE ||
2243 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
2244 expected_type = PTR_TO_STACK;
2245 if (!type_is_pkt_pointer(type) && type != PTR_TO_MAP_VALUE &&
2246 type != expected_type)
2248 } else if (arg_type == ARG_CONST_SIZE ||
2249 arg_type == ARG_CONST_SIZE_OR_ZERO) {
2250 expected_type = SCALAR_VALUE;
2251 if (type != expected_type)
2253 } else if (arg_type == ARG_CONST_MAP_PTR) {
2254 expected_type = CONST_PTR_TO_MAP;
2255 if (type != expected_type)
2257 } else if (arg_type == ARG_PTR_TO_CTX) {
2258 expected_type = PTR_TO_CTX;
2259 if (type != expected_type)
2261 err = check_ctx_reg(env, reg, regno);
2264 } else if (arg_type == ARG_PTR_TO_SOCKET) {
2265 expected_type = PTR_TO_SOCKET;
2266 if (type != expected_type)
2268 if (meta->ptr_id || !reg->id) {
2269 verbose(env, "verifier internal error: mismatched references meta=%d, reg=%d\n",
2270 meta->ptr_id, reg->id);
2273 meta->ptr_id = reg->id;
2274 } else if (arg_type_is_mem_ptr(arg_type)) {
2275 expected_type = PTR_TO_STACK;
2276 /* One exception here. In case function allows for NULL to be
2277 * passed in as argument, it's a SCALAR_VALUE type. Final test
2278 * happens during stack boundary checking.
2280 if (register_is_null(reg) &&
2281 arg_type == ARG_PTR_TO_MEM_OR_NULL)
2282 /* final test in check_stack_boundary() */;
2283 else if (!type_is_pkt_pointer(type) &&
2284 type != PTR_TO_MAP_VALUE &&
2285 type != expected_type)
2287 meta->raw_mode = arg_type == ARG_PTR_TO_UNINIT_MEM;
2289 verbose(env, "unsupported arg_type %d\n", arg_type);
2293 if (arg_type == ARG_CONST_MAP_PTR) {
2294 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
2295 meta->map_ptr = reg->map_ptr;
2296 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
2297 /* bpf_map_xxx(..., map_ptr, ..., key) call:
2298 * check that [key, key + map->key_size) are within
2299 * stack limits and initialized
2301 if (!meta->map_ptr) {
2302 /* in function declaration map_ptr must come before
2303 * map_key, so that it's verified and known before
2304 * we have to check map_key here. Otherwise it means
2305 * that kernel subsystem misconfigured verifier
2307 verbose(env, "invalid map_ptr to access map->key\n");
2310 err = check_helper_mem_access(env, regno,
2311 meta->map_ptr->key_size, false,
2313 } else if (arg_type == ARG_PTR_TO_MAP_VALUE ||
2314 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
2315 /* bpf_map_xxx(..., map_ptr, ..., value) call:
2316 * check [value, value + map->value_size) validity
2318 if (!meta->map_ptr) {
2319 /* kernel subsystem misconfigured verifier */
2320 verbose(env, "invalid map_ptr to access map->value\n");
2323 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE);
2324 err = check_helper_mem_access(env, regno,
2325 meta->map_ptr->value_size, false,
2327 } else if (arg_type_is_mem_size(arg_type)) {
2328 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
2330 /* remember the mem_size which may be used later
2331 * to refine return values.
2333 meta->msize_smax_value = reg->smax_value;
2334 meta->msize_umax_value = reg->umax_value;
2336 /* The register is SCALAR_VALUE; the access check
2337 * happens using its boundaries.
2339 if (!tnum_is_const(reg->var_off))
2340 /* For unprivileged variable accesses, disable raw
2341 * mode so that the program is required to
2342 * initialize all the memory that the helper could
2343 * just partially fill up.
2347 if (reg->smin_value < 0) {
2348 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
2353 if (reg->umin_value == 0) {
2354 err = check_helper_mem_access(env, regno - 1, 0,
2361 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
2362 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
2366 err = check_helper_mem_access(env, regno - 1,
2368 zero_size_allowed, meta);
2373 verbose(env, "R%d type=%s expected=%s\n", regno,
2374 reg_type_str[type], reg_type_str[expected_type]);
2378 static int check_map_func_compatibility(struct bpf_verifier_env *env,
2379 struct bpf_map *map, int func_id)
2384 /* We need a two way check, first is from map perspective ... */
2385 switch (map->map_type) {
2386 case BPF_MAP_TYPE_PROG_ARRAY:
2387 if (func_id != BPF_FUNC_tail_call)
2390 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
2391 if (func_id != BPF_FUNC_perf_event_read &&
2392 func_id != BPF_FUNC_perf_event_output &&
2393 func_id != BPF_FUNC_perf_event_read_value)
2396 case BPF_MAP_TYPE_STACK_TRACE:
2397 if (func_id != BPF_FUNC_get_stackid)
2400 case BPF_MAP_TYPE_CGROUP_ARRAY:
2401 if (func_id != BPF_FUNC_skb_under_cgroup &&
2402 func_id != BPF_FUNC_current_task_under_cgroup)
2405 case BPF_MAP_TYPE_CGROUP_STORAGE:
2406 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
2407 if (func_id != BPF_FUNC_get_local_storage)
2410 /* devmap returns a pointer to a live net_device ifindex that we cannot
2411 * allow to be modified from bpf side. So do not allow lookup elements
2414 case BPF_MAP_TYPE_DEVMAP:
2415 if (func_id != BPF_FUNC_redirect_map)
2418 /* Restrict bpf side of cpumap and xskmap, open when use-cases
2421 case BPF_MAP_TYPE_CPUMAP:
2422 case BPF_MAP_TYPE_XSKMAP:
2423 if (func_id != BPF_FUNC_redirect_map)
2426 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
2427 case BPF_MAP_TYPE_HASH_OF_MAPS:
2428 if (func_id != BPF_FUNC_map_lookup_elem)
2431 case BPF_MAP_TYPE_SOCKMAP:
2432 if (func_id != BPF_FUNC_sk_redirect_map &&
2433 func_id != BPF_FUNC_sock_map_update &&
2434 func_id != BPF_FUNC_map_delete_elem &&
2435 func_id != BPF_FUNC_msg_redirect_map)
2438 case BPF_MAP_TYPE_SOCKHASH:
2439 if (func_id != BPF_FUNC_sk_redirect_hash &&
2440 func_id != BPF_FUNC_sock_hash_update &&
2441 func_id != BPF_FUNC_map_delete_elem &&
2442 func_id != BPF_FUNC_msg_redirect_hash)
2445 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
2446 if (func_id != BPF_FUNC_sk_select_reuseport)
2449 case BPF_MAP_TYPE_QUEUE:
2450 case BPF_MAP_TYPE_STACK:
2451 if (func_id != BPF_FUNC_map_peek_elem &&
2452 func_id != BPF_FUNC_map_pop_elem &&
2453 func_id != BPF_FUNC_map_push_elem)
2460 /* ... and second from the function itself. */
2462 case BPF_FUNC_tail_call:
2463 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
2465 if (env->subprog_cnt > 1) {
2466 verbose(env, "tail_calls are not allowed in programs with bpf-to-bpf calls\n");
2470 case BPF_FUNC_perf_event_read:
2471 case BPF_FUNC_perf_event_output:
2472 case BPF_FUNC_perf_event_read_value:
2473 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
2476 case BPF_FUNC_get_stackid:
2477 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
2480 case BPF_FUNC_current_task_under_cgroup:
2481 case BPF_FUNC_skb_under_cgroup:
2482 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
2485 case BPF_FUNC_redirect_map:
2486 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
2487 map->map_type != BPF_MAP_TYPE_CPUMAP &&
2488 map->map_type != BPF_MAP_TYPE_XSKMAP)
2491 case BPF_FUNC_sk_redirect_map:
2492 case BPF_FUNC_msg_redirect_map:
2493 case BPF_FUNC_sock_map_update:
2494 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
2497 case BPF_FUNC_sk_redirect_hash:
2498 case BPF_FUNC_msg_redirect_hash:
2499 case BPF_FUNC_sock_hash_update:
2500 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
2503 case BPF_FUNC_get_local_storage:
2504 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
2505 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
2508 case BPF_FUNC_sk_select_reuseport:
2509 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY)
2512 case BPF_FUNC_map_peek_elem:
2513 case BPF_FUNC_map_pop_elem:
2514 case BPF_FUNC_map_push_elem:
2515 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
2516 map->map_type != BPF_MAP_TYPE_STACK)
2525 verbose(env, "cannot pass map_type %d into func %s#%d\n",
2526 map->map_type, func_id_name(func_id), func_id);
2530 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
2534 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
2536 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
2538 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
2540 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
2542 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
2545 /* We only support one arg being in raw mode at the moment,
2546 * which is sufficient for the helper functions we have
2552 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
2553 enum bpf_arg_type arg_next)
2555 return (arg_type_is_mem_ptr(arg_curr) &&
2556 !arg_type_is_mem_size(arg_next)) ||
2557 (!arg_type_is_mem_ptr(arg_curr) &&
2558 arg_type_is_mem_size(arg_next));
2561 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
2563 /* bpf_xxx(..., buf, len) call will access 'len'
2564 * bytes from memory 'buf'. Both arg types need
2565 * to be paired, so make sure there's no buggy
2566 * helper function specification.
2568 if (arg_type_is_mem_size(fn->arg1_type) ||
2569 arg_type_is_mem_ptr(fn->arg5_type) ||
2570 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
2571 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
2572 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
2573 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
2579 static bool check_refcount_ok(const struct bpf_func_proto *fn)
2583 if (arg_type_is_refcounted(fn->arg1_type))
2585 if (arg_type_is_refcounted(fn->arg2_type))
2587 if (arg_type_is_refcounted(fn->arg3_type))
2589 if (arg_type_is_refcounted(fn->arg4_type))
2591 if (arg_type_is_refcounted(fn->arg5_type))
2594 /* We only support one arg being unreferenced at the moment,
2595 * which is sufficient for the helper functions we have right now.
2600 static int check_func_proto(const struct bpf_func_proto *fn)
2602 return check_raw_mode_ok(fn) &&
2603 check_arg_pair_ok(fn) &&
2604 check_refcount_ok(fn) ? 0 : -EINVAL;
2607 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
2608 * are now invalid, so turn them into unknown SCALAR_VALUE.
2610 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
2611 struct bpf_func_state *state)
2613 struct bpf_reg_state *regs = state->regs, *reg;
2616 for (i = 0; i < MAX_BPF_REG; i++)
2617 if (reg_is_pkt_pointer_any(®s[i]))
2618 mark_reg_unknown(env, regs, i);
2620 bpf_for_each_spilled_reg(i, state, reg) {
2623 if (reg_is_pkt_pointer_any(reg))
2624 __mark_reg_unknown(reg);
2628 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
2630 struct bpf_verifier_state *vstate = env->cur_state;
2633 for (i = 0; i <= vstate->curframe; i++)
2634 __clear_all_pkt_pointers(env, vstate->frame[i]);
2637 static void release_reg_references(struct bpf_verifier_env *env,
2638 struct bpf_func_state *state, int id)
2640 struct bpf_reg_state *regs = state->regs, *reg;
2643 for (i = 0; i < MAX_BPF_REG; i++)
2644 if (regs[i].id == id)
2645 mark_reg_unknown(env, regs, i);
2647 bpf_for_each_spilled_reg(i, state, reg) {
2650 if (reg_is_refcounted(reg) && reg->id == id)
2651 __mark_reg_unknown(reg);
2655 /* The pointer with the specified id has released its reference to kernel
2656 * resources. Identify all copies of the same pointer and clear the reference.
2658 static int release_reference(struct bpf_verifier_env *env,
2659 struct bpf_call_arg_meta *meta)
2661 struct bpf_verifier_state *vstate = env->cur_state;
2664 for (i = 0; i <= vstate->curframe; i++)
2665 release_reg_references(env, vstate->frame[i], meta->ptr_id);
2667 return release_reference_state(env, meta->ptr_id);
2670 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
2673 struct bpf_verifier_state *state = env->cur_state;
2674 struct bpf_func_state *caller, *callee;
2675 int i, err, subprog, target_insn;
2677 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
2678 verbose(env, "the call stack of %d frames is too deep\n",
2679 state->curframe + 2);
2683 target_insn = *insn_idx + insn->imm;
2684 subprog = find_subprog(env, target_insn + 1);
2686 verbose(env, "verifier bug. No program starts at insn %d\n",
2691 caller = state->frame[state->curframe];
2692 if (state->frame[state->curframe + 1]) {
2693 verbose(env, "verifier bug. Frame %d already allocated\n",
2694 state->curframe + 1);
2698 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
2701 state->frame[state->curframe + 1] = callee;
2703 /* callee cannot access r0, r6 - r9 for reading and has to write
2704 * into its own stack before reading from it.
2705 * callee can read/write into caller's stack
2707 init_func_state(env, callee,
2708 /* remember the callsite, it will be used by bpf_exit */
2709 *insn_idx /* callsite */,
2710 state->curframe + 1 /* frameno within this callchain */,
2711 subprog /* subprog number within this prog */);
2713 /* Transfer references to the callee */
2714 err = transfer_reference_state(callee, caller);
2718 /* copy r1 - r5 args that callee can access. The copy includes parent
2719 * pointers, which connects us up to the liveness chain
2721 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
2722 callee->regs[i] = caller->regs[i];
2724 /* after the call registers r0 - r5 were scratched */
2725 for (i = 0; i < CALLER_SAVED_REGS; i++) {
2726 mark_reg_not_init(env, caller->regs, caller_saved[i]);
2727 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
2730 /* only increment it after check_reg_arg() finished */
2733 /* and go analyze first insn of the callee */
2734 *insn_idx = target_insn;
2736 if (env->log.level) {
2737 verbose(env, "caller:\n");
2738 print_verifier_state(env, caller);
2739 verbose(env, "callee:\n");
2740 print_verifier_state(env, callee);
2745 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
2747 struct bpf_verifier_state *state = env->cur_state;
2748 struct bpf_func_state *caller, *callee;
2749 struct bpf_reg_state *r0;
2752 callee = state->frame[state->curframe];
2753 r0 = &callee->regs[BPF_REG_0];
2754 if (r0->type == PTR_TO_STACK) {
2755 /* technically it's ok to return caller's stack pointer
2756 * (or caller's caller's pointer) back to the caller,
2757 * since these pointers are valid. Only current stack
2758 * pointer will be invalid as soon as function exits,
2759 * but let's be conservative
2761 verbose(env, "cannot return stack pointer to the caller\n");
2766 caller = state->frame[state->curframe];
2767 /* return to the caller whatever r0 had in the callee */
2768 caller->regs[BPF_REG_0] = *r0;
2770 /* Transfer references to the caller */
2771 err = transfer_reference_state(caller, callee);
2775 *insn_idx = callee->callsite + 1;
2776 if (env->log.level) {
2777 verbose(env, "returning from callee:\n");
2778 print_verifier_state(env, callee);
2779 verbose(env, "to caller at %d:\n", *insn_idx);
2780 print_verifier_state(env, caller);
2782 /* clear everything in the callee */
2783 free_func_state(callee);
2784 state->frame[state->curframe + 1] = NULL;
2788 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
2790 struct bpf_call_arg_meta *meta)
2792 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
2794 if (ret_type != RET_INTEGER ||
2795 (func_id != BPF_FUNC_get_stack &&
2796 func_id != BPF_FUNC_probe_read_str))
2799 ret_reg->smax_value = meta->msize_smax_value;
2800 ret_reg->umax_value = meta->msize_umax_value;
2801 __reg_deduce_bounds(ret_reg);
2802 __reg_bound_offset(ret_reg);
2806 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
2807 int func_id, int insn_idx)
2809 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
2811 if (func_id != BPF_FUNC_tail_call &&
2812 func_id != BPF_FUNC_map_lookup_elem &&
2813 func_id != BPF_FUNC_map_update_elem &&
2814 func_id != BPF_FUNC_map_delete_elem &&
2815 func_id != BPF_FUNC_map_push_elem &&
2816 func_id != BPF_FUNC_map_pop_elem &&
2817 func_id != BPF_FUNC_map_peek_elem)
2820 if (meta->map_ptr == NULL) {
2821 verbose(env, "kernel subsystem misconfigured verifier\n");
2825 if (!BPF_MAP_PTR(aux->map_state))
2826 bpf_map_ptr_store(aux, meta->map_ptr,
2827 meta->map_ptr->unpriv_array);
2828 else if (BPF_MAP_PTR(aux->map_state) != meta->map_ptr)
2829 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
2830 meta->map_ptr->unpriv_array);
2834 static int check_reference_leak(struct bpf_verifier_env *env)
2836 struct bpf_func_state *state = cur_func(env);
2839 for (i = 0; i < state->acquired_refs; i++) {
2840 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
2841 state->refs[i].id, state->refs[i].insn_idx);
2843 return state->acquired_refs ? -EINVAL : 0;
2846 static int check_helper_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
2848 const struct bpf_func_proto *fn = NULL;
2849 struct bpf_reg_state *regs;
2850 struct bpf_call_arg_meta meta;
2854 /* find function prototype */
2855 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
2856 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
2861 if (env->ops->get_func_proto)
2862 fn = env->ops->get_func_proto(func_id, env->prog);
2864 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
2869 /* eBPF programs must be GPL compatible to use GPL-ed functions */
2870 if (!env->prog->gpl_compatible && fn->gpl_only) {
2871 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
2875 /* With LD_ABS/IND some JITs save/restore skb from r1. */
2876 changes_data = bpf_helper_changes_pkt_data(fn->func);
2877 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
2878 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
2879 func_id_name(func_id), func_id);
2883 memset(&meta, 0, sizeof(meta));
2884 meta.pkt_access = fn->pkt_access;
2886 err = check_func_proto(fn);
2888 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
2889 func_id_name(func_id), func_id);
2894 err = check_func_arg(env, BPF_REG_1, fn->arg1_type, &meta);
2897 err = check_func_arg(env, BPF_REG_2, fn->arg2_type, &meta);
2900 err = check_func_arg(env, BPF_REG_3, fn->arg3_type, &meta);
2903 err = check_func_arg(env, BPF_REG_4, fn->arg4_type, &meta);
2906 err = check_func_arg(env, BPF_REG_5, fn->arg5_type, &meta);
2910 err = record_func_map(env, &meta, func_id, insn_idx);
2914 /* Mark slots with STACK_MISC in case of raw mode, stack offset
2915 * is inferred from register state.
2917 for (i = 0; i < meta.access_size; i++) {
2918 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
2919 BPF_WRITE, -1, false);
2924 if (func_id == BPF_FUNC_tail_call) {
2925 err = check_reference_leak(env);
2927 verbose(env, "tail_call would lead to reference leak\n");
2930 } else if (is_release_function(func_id)) {
2931 err = release_reference(env, &meta);
2936 regs = cur_regs(env);
2938 /* check that flags argument in get_local_storage(map, flags) is 0,
2939 * this is required because get_local_storage() can't return an error.
2941 if (func_id == BPF_FUNC_get_local_storage &&
2942 !register_is_null(®s[BPF_REG_2])) {
2943 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
2947 /* reset caller saved regs */
2948 for (i = 0; i < CALLER_SAVED_REGS; i++) {
2949 mark_reg_not_init(env, regs, caller_saved[i]);
2950 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
2953 /* update return register (already marked as written above) */
2954 if (fn->ret_type == RET_INTEGER) {
2955 /* sets type to SCALAR_VALUE */
2956 mark_reg_unknown(env, regs, BPF_REG_0);
2957 } else if (fn->ret_type == RET_VOID) {
2958 regs[BPF_REG_0].type = NOT_INIT;
2959 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL ||
2960 fn->ret_type == RET_PTR_TO_MAP_VALUE) {
2961 /* There is no offset yet applied, variable or fixed */
2962 mark_reg_known_zero(env, regs, BPF_REG_0);
2963 /* remember map_ptr, so that check_map_access()
2964 * can check 'value_size' boundary of memory access
2965 * to map element returned from bpf_map_lookup_elem()
2967 if (meta.map_ptr == NULL) {
2969 "kernel subsystem misconfigured verifier\n");
2972 regs[BPF_REG_0].map_ptr = meta.map_ptr;
2973 if (fn->ret_type == RET_PTR_TO_MAP_VALUE) {
2974 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE;
2976 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
2977 regs[BPF_REG_0].id = ++env->id_gen;
2979 } else if (fn->ret_type == RET_PTR_TO_SOCKET_OR_NULL) {
2980 int id = acquire_reference_state(env, insn_idx);
2983 mark_reg_known_zero(env, regs, BPF_REG_0);
2984 regs[BPF_REG_0].type = PTR_TO_SOCKET_OR_NULL;
2985 regs[BPF_REG_0].id = id;
2987 verbose(env, "unknown return type %d of func %s#%d\n",
2988 fn->ret_type, func_id_name(func_id), func_id);
2992 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
2994 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
2998 if (func_id == BPF_FUNC_get_stack && !env->prog->has_callchain_buf) {
2999 const char *err_str;
3001 #ifdef CONFIG_PERF_EVENTS
3002 err = get_callchain_buffers(sysctl_perf_event_max_stack);
3003 err_str = "cannot get callchain buffer for func %s#%d\n";
3006 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
3009 verbose(env, err_str, func_id_name(func_id), func_id);
3013 env->prog->has_callchain_buf = true;
3017 clear_all_pkt_pointers(env);
3021 static bool signed_add_overflows(s64 a, s64 b)
3023 /* Do the add in u64, where overflow is well-defined */
3024 s64 res = (s64)((u64)a + (u64)b);
3031 static bool signed_sub_overflows(s64 a, s64 b)
3033 /* Do the sub in u64, where overflow is well-defined */
3034 s64 res = (s64)((u64)a - (u64)b);
3041 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
3042 const struct bpf_reg_state *reg,
3043 enum bpf_reg_type type)
3045 bool known = tnum_is_const(reg->var_off);
3046 s64 val = reg->var_off.value;
3047 s64 smin = reg->smin_value;
3049 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
3050 verbose(env, "math between %s pointer and %lld is not allowed\n",
3051 reg_type_str[type], val);
3055 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
3056 verbose(env, "%s pointer offset %d is not allowed\n",
3057 reg_type_str[type], reg->off);
3061 if (smin == S64_MIN) {
3062 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
3063 reg_type_str[type]);
3067 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
3068 verbose(env, "value %lld makes %s pointer be out of bounds\n",
3069 smin, reg_type_str[type]);
3076 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
3078 return &env->insn_aux_data[env->insn_idx];
3081 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
3082 u32 *ptr_limit, u8 opcode, bool off_is_neg)
3084 bool mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
3085 (opcode == BPF_SUB && !off_is_neg);
3088 switch (ptr_reg->type) {
3090 off = ptr_reg->off + ptr_reg->var_off.value;
3092 *ptr_limit = MAX_BPF_STACK + off;
3096 case PTR_TO_MAP_VALUE:
3098 *ptr_limit = ptr_reg->umax_value + ptr_reg->off;
3100 off = ptr_reg->smin_value + ptr_reg->off;
3101 *ptr_limit = ptr_reg->map_ptr->value_size - off;
3109 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
3110 const struct bpf_insn *insn)
3112 return env->allow_ptr_leaks || BPF_SRC(insn->code) == BPF_K;
3115 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
3116 u32 alu_state, u32 alu_limit)
3118 /* If we arrived here from different branches with different
3119 * state or limits to sanitize, then this won't work.
3121 if (aux->alu_state &&
3122 (aux->alu_state != alu_state ||
3123 aux->alu_limit != alu_limit))
3126 /* Corresponding fixup done in fixup_bpf_calls(). */
3127 aux->alu_state = alu_state;
3128 aux->alu_limit = alu_limit;
3132 static int sanitize_val_alu(struct bpf_verifier_env *env,
3133 struct bpf_insn *insn)
3135 struct bpf_insn_aux_data *aux = cur_aux(env);
3137 if (can_skip_alu_sanitation(env, insn))
3140 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
3143 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
3144 struct bpf_insn *insn,
3145 const struct bpf_reg_state *ptr_reg,
3146 struct bpf_reg_state *dst_reg,
3149 struct bpf_verifier_state *vstate = env->cur_state;
3150 struct bpf_insn_aux_data *aux = cur_aux(env);
3151 bool ptr_is_dst_reg = ptr_reg == dst_reg;
3152 u8 opcode = BPF_OP(insn->code);
3153 u32 alu_state, alu_limit;
3154 struct bpf_reg_state tmp;
3157 if (can_skip_alu_sanitation(env, insn))
3160 /* We already marked aux for masking from non-speculative
3161 * paths, thus we got here in the first place. We only care
3162 * to explore bad access from here.
3164 if (vstate->speculative)
3167 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
3168 alu_state |= ptr_is_dst_reg ?
3169 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
3171 if (retrieve_ptr_limit(ptr_reg, &alu_limit, opcode, off_is_neg))
3173 if (update_alu_sanitation_state(aux, alu_state, alu_limit))
3176 /* Simulate and find potential out-of-bounds access under
3177 * speculative execution from truncation as a result of
3178 * masking when off was not within expected range. If off
3179 * sits in dst, then we temporarily need to move ptr there
3180 * to simulate dst (== 0) +/-= ptr. Needed, for example,
3181 * for cases where we use K-based arithmetic in one direction
3182 * and truncated reg-based in the other in order to explore
3185 if (!ptr_is_dst_reg) {
3187 *dst_reg = *ptr_reg;
3189 ret = push_stack(env, env->insn_idx + 1, env->insn_idx, true);
3190 if (!ptr_is_dst_reg)
3192 return !ret ? -EFAULT : 0;
3195 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
3196 * Caller should also handle BPF_MOV case separately.
3197 * If we return -EACCES, caller may want to try again treating pointer as a
3198 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
3200 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
3201 struct bpf_insn *insn,
3202 const struct bpf_reg_state *ptr_reg,
3203 const struct bpf_reg_state *off_reg)
3205 struct bpf_verifier_state *vstate = env->cur_state;
3206 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3207 struct bpf_reg_state *regs = state->regs, *dst_reg;
3208 bool known = tnum_is_const(off_reg->var_off);
3209 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
3210 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
3211 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
3212 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
3213 u32 dst = insn->dst_reg, src = insn->src_reg;
3214 u8 opcode = BPF_OP(insn->code);
3217 dst_reg = ®s[dst];
3219 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
3220 smin_val > smax_val || umin_val > umax_val) {
3221 /* Taint dst register if offset had invalid bounds derived from
3222 * e.g. dead branches.
3224 __mark_reg_unknown(dst_reg);
3228 if (BPF_CLASS(insn->code) != BPF_ALU64) {
3229 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
3231 "R%d 32-bit pointer arithmetic prohibited\n",
3236 switch (ptr_reg->type) {
3237 case PTR_TO_MAP_VALUE_OR_NULL:
3238 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
3239 dst, reg_type_str[ptr_reg->type]);
3241 case CONST_PTR_TO_MAP:
3242 case PTR_TO_PACKET_END:
3244 case PTR_TO_SOCKET_OR_NULL:
3245 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
3246 dst, reg_type_str[ptr_reg->type]);
3248 case PTR_TO_MAP_VALUE:
3249 if (!env->allow_ptr_leaks && !known && (smin_val < 0) != (smax_val < 0)) {
3250 verbose(env, "R%d has unknown scalar with mixed signed bounds, pointer arithmetic with it prohibited for !root\n",
3251 off_reg == dst_reg ? dst : src);
3259 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
3260 * The id may be overwritten later if we create a new variable offset.
3262 dst_reg->type = ptr_reg->type;
3263 dst_reg->id = ptr_reg->id;
3265 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
3266 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
3271 ret = sanitize_ptr_alu(env, insn, ptr_reg, dst_reg, smin_val < 0);
3273 verbose(env, "R%d tried to add from different maps or paths\n", dst);
3276 /* We can take a fixed offset as long as it doesn't overflow
3277 * the s32 'off' field
3279 if (known && (ptr_reg->off + smin_val ==
3280 (s64)(s32)(ptr_reg->off + smin_val))) {
3281 /* pointer += K. Accumulate it into fixed offset */
3282 dst_reg->smin_value = smin_ptr;
3283 dst_reg->smax_value = smax_ptr;
3284 dst_reg->umin_value = umin_ptr;
3285 dst_reg->umax_value = umax_ptr;
3286 dst_reg->var_off = ptr_reg->var_off;
3287 dst_reg->off = ptr_reg->off + smin_val;
3288 dst_reg->raw = ptr_reg->raw;
3291 /* A new variable offset is created. Note that off_reg->off
3292 * == 0, since it's a scalar.
3293 * dst_reg gets the pointer type and since some positive
3294 * integer value was added to the pointer, give it a new 'id'
3295 * if it's a PTR_TO_PACKET.
3296 * this creates a new 'base' pointer, off_reg (variable) gets
3297 * added into the variable offset, and we copy the fixed offset
3300 if (signed_add_overflows(smin_ptr, smin_val) ||
3301 signed_add_overflows(smax_ptr, smax_val)) {
3302 dst_reg->smin_value = S64_MIN;
3303 dst_reg->smax_value = S64_MAX;
3305 dst_reg->smin_value = smin_ptr + smin_val;
3306 dst_reg->smax_value = smax_ptr + smax_val;
3308 if (umin_ptr + umin_val < umin_ptr ||
3309 umax_ptr + umax_val < umax_ptr) {
3310 dst_reg->umin_value = 0;
3311 dst_reg->umax_value = U64_MAX;
3313 dst_reg->umin_value = umin_ptr + umin_val;
3314 dst_reg->umax_value = umax_ptr + umax_val;
3316 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
3317 dst_reg->off = ptr_reg->off;
3318 dst_reg->raw = ptr_reg->raw;
3319 if (reg_is_pkt_pointer(ptr_reg)) {
3320 dst_reg->id = ++env->id_gen;
3321 /* something was added to pkt_ptr, set range to zero */
3326 ret = sanitize_ptr_alu(env, insn, ptr_reg, dst_reg, smin_val < 0);
3328 verbose(env, "R%d tried to sub from different maps or paths\n", dst);
3331 if (dst_reg == off_reg) {
3332 /* scalar -= pointer. Creates an unknown scalar */
3333 verbose(env, "R%d tried to subtract pointer from scalar\n",
3337 /* We don't allow subtraction from FP, because (according to
3338 * test_verifier.c test "invalid fp arithmetic", JITs might not
3339 * be able to deal with it.
3341 if (ptr_reg->type == PTR_TO_STACK) {
3342 verbose(env, "R%d subtraction from stack pointer prohibited\n",
3346 if (known && (ptr_reg->off - smin_val ==
3347 (s64)(s32)(ptr_reg->off - smin_val))) {
3348 /* pointer -= K. Subtract it from fixed offset */
3349 dst_reg->smin_value = smin_ptr;
3350 dst_reg->smax_value = smax_ptr;
3351 dst_reg->umin_value = umin_ptr;
3352 dst_reg->umax_value = umax_ptr;
3353 dst_reg->var_off = ptr_reg->var_off;
3354 dst_reg->id = ptr_reg->id;
3355 dst_reg->off = ptr_reg->off - smin_val;
3356 dst_reg->raw = ptr_reg->raw;
3359 /* A new variable offset is created. If the subtrahend is known
3360 * nonnegative, then any reg->range we had before is still good.
3362 if (signed_sub_overflows(smin_ptr, smax_val) ||
3363 signed_sub_overflows(smax_ptr, smin_val)) {
3364 /* Overflow possible, we know nothing */
3365 dst_reg->smin_value = S64_MIN;
3366 dst_reg->smax_value = S64_MAX;
3368 dst_reg->smin_value = smin_ptr - smax_val;
3369 dst_reg->smax_value = smax_ptr - smin_val;
3371 if (umin_ptr < umax_val) {
3372 /* Overflow possible, we know nothing */
3373 dst_reg->umin_value = 0;
3374 dst_reg->umax_value = U64_MAX;
3376 /* Cannot overflow (as long as bounds are consistent) */
3377 dst_reg->umin_value = umin_ptr - umax_val;
3378 dst_reg->umax_value = umax_ptr - umin_val;
3380 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
3381 dst_reg->off = ptr_reg->off;
3382 dst_reg->raw = ptr_reg->raw;
3383 if (reg_is_pkt_pointer(ptr_reg)) {
3384 dst_reg->id = ++env->id_gen;
3385 /* something was added to pkt_ptr, set range to zero */
3393 /* bitwise ops on pointers are troublesome, prohibit. */
3394 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
3395 dst, bpf_alu_string[opcode >> 4]);
3398 /* other operators (e.g. MUL,LSH) produce non-pointer results */
3399 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
3400 dst, bpf_alu_string[opcode >> 4]);
3404 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
3407 __update_reg_bounds(dst_reg);
3408 __reg_deduce_bounds(dst_reg);
3409 __reg_bound_offset(dst_reg);
3411 /* For unprivileged we require that resulting offset must be in bounds
3412 * in order to be able to sanitize access later on.
3414 if (!env->allow_ptr_leaks) {
3415 if (dst_reg->type == PTR_TO_MAP_VALUE &&
3416 check_map_access(env, dst, dst_reg->off, 1, false)) {
3417 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
3418 "prohibited for !root\n", dst);
3420 } else if (dst_reg->type == PTR_TO_STACK &&
3421 check_stack_access(env, dst_reg, dst_reg->off +
3422 dst_reg->var_off.value, 1)) {
3423 verbose(env, "R%d stack pointer arithmetic goes out of range, "
3424 "prohibited for !root\n", dst);
3432 /* WARNING: This function does calculations on 64-bit values, but the actual
3433 * execution may occur on 32-bit values. Therefore, things like bitshifts
3434 * need extra checks in the 32-bit case.
3436 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
3437 struct bpf_insn *insn,
3438 struct bpf_reg_state *dst_reg,
3439 struct bpf_reg_state src_reg)
3441 struct bpf_reg_state *regs = cur_regs(env);
3442 u8 opcode = BPF_OP(insn->code);
3443 bool src_known, dst_known;
3444 s64 smin_val, smax_val;
3445 u64 umin_val, umax_val;
3446 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
3447 u32 dst = insn->dst_reg;
3450 if (insn_bitness == 32) {
3451 /* Relevant for 32-bit RSH: Information can propagate towards
3452 * LSB, so it isn't sufficient to only truncate the output to
3455 coerce_reg_to_size(dst_reg, 4);
3456 coerce_reg_to_size(&src_reg, 4);
3459 smin_val = src_reg.smin_value;
3460 smax_val = src_reg.smax_value;
3461 umin_val = src_reg.umin_value;
3462 umax_val = src_reg.umax_value;
3463 src_known = tnum_is_const(src_reg.var_off);
3464 dst_known = tnum_is_const(dst_reg->var_off);
3466 if ((src_known && (smin_val != smax_val || umin_val != umax_val)) ||
3467 smin_val > smax_val || umin_val > umax_val) {
3468 /* Taint dst register if offset had invalid bounds derived from
3469 * e.g. dead branches.
3471 __mark_reg_unknown(dst_reg);
3476 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
3477 __mark_reg_unknown(dst_reg);
3483 ret = sanitize_val_alu(env, insn);
3485 verbose(env, "R%d tried to add from different pointers or scalars\n", dst);
3488 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
3489 signed_add_overflows(dst_reg->smax_value, smax_val)) {
3490 dst_reg->smin_value = S64_MIN;
3491 dst_reg->smax_value = S64_MAX;
3493 dst_reg->smin_value += smin_val;
3494 dst_reg->smax_value += smax_val;
3496 if (dst_reg->umin_value + umin_val < umin_val ||
3497 dst_reg->umax_value + umax_val < umax_val) {
3498 dst_reg->umin_value = 0;
3499 dst_reg->umax_value = U64_MAX;
3501 dst_reg->umin_value += umin_val;
3502 dst_reg->umax_value += umax_val;
3504 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
3507 ret = sanitize_val_alu(env, insn);
3509 verbose(env, "R%d tried to sub from different pointers or scalars\n", dst);
3512 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
3513 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
3514 /* Overflow possible, we know nothing */
3515 dst_reg->smin_value = S64_MIN;
3516 dst_reg->smax_value = S64_MAX;
3518 dst_reg->smin_value -= smax_val;
3519 dst_reg->smax_value -= smin_val;
3521 if (dst_reg->umin_value < umax_val) {
3522 /* Overflow possible, we know nothing */
3523 dst_reg->umin_value = 0;
3524 dst_reg->umax_value = U64_MAX;
3526 /* Cannot overflow (as long as bounds are consistent) */
3527 dst_reg->umin_value -= umax_val;
3528 dst_reg->umax_value -= umin_val;
3530 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
3533 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
3534 if (smin_val < 0 || dst_reg->smin_value < 0) {
3535 /* Ain't nobody got time to multiply that sign */
3536 __mark_reg_unbounded(dst_reg);
3537 __update_reg_bounds(dst_reg);
3540 /* Both values are positive, so we can work with unsigned and
3541 * copy the result to signed (unless it exceeds S64_MAX).
3543 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
3544 /* Potential overflow, we know nothing */
3545 __mark_reg_unbounded(dst_reg);
3546 /* (except what we can learn from the var_off) */
3547 __update_reg_bounds(dst_reg);
3550 dst_reg->umin_value *= umin_val;
3551 dst_reg->umax_value *= umax_val;
3552 if (dst_reg->umax_value > S64_MAX) {
3553 /* Overflow possible, we know nothing */
3554 dst_reg->smin_value = S64_MIN;
3555 dst_reg->smax_value = S64_MAX;
3557 dst_reg->smin_value = dst_reg->umin_value;
3558 dst_reg->smax_value = dst_reg->umax_value;
3562 if (src_known && dst_known) {
3563 __mark_reg_known(dst_reg, dst_reg->var_off.value &
3564 src_reg.var_off.value);
3567 /* We get our minimum from the var_off, since that's inherently
3568 * bitwise. Our maximum is the minimum of the operands' maxima.
3570 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
3571 dst_reg->umin_value = dst_reg->var_off.value;
3572 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
3573 if (dst_reg->smin_value < 0 || smin_val < 0) {
3574 /* Lose signed bounds when ANDing negative numbers,
3575 * ain't nobody got time for that.
3577 dst_reg->smin_value = S64_MIN;
3578 dst_reg->smax_value = S64_MAX;
3580 /* ANDing two positives gives a positive, so safe to
3581 * cast result into s64.
3583 dst_reg->smin_value = dst_reg->umin_value;
3584 dst_reg->smax_value = dst_reg->umax_value;
3586 /* We may learn something more from the var_off */
3587 __update_reg_bounds(dst_reg);
3590 if (src_known && dst_known) {
3591 __mark_reg_known(dst_reg, dst_reg->var_off.value |
3592 src_reg.var_off.value);
3595 /* We get our maximum from the var_off, and our minimum is the
3596 * maximum of the operands' minima
3598 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
3599 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
3600 dst_reg->umax_value = dst_reg->var_off.value |
3601 dst_reg->var_off.mask;
3602 if (dst_reg->smin_value < 0 || smin_val < 0) {
3603 /* Lose signed bounds when ORing negative numbers,
3604 * ain't nobody got time for that.
3606 dst_reg->smin_value = S64_MIN;
3607 dst_reg->smax_value = S64_MAX;
3609 /* ORing two positives gives a positive, so safe to
3610 * cast result into s64.
3612 dst_reg->smin_value = dst_reg->umin_value;
3613 dst_reg->smax_value = dst_reg->umax_value;
3615 /* We may learn something more from the var_off */
3616 __update_reg_bounds(dst_reg);
3619 if (umax_val >= insn_bitness) {
3620 /* Shifts greater than 31 or 63 are undefined.
3621 * This includes shifts by a negative number.
3623 mark_reg_unknown(env, regs, insn->dst_reg);
3626 /* We lose all sign bit information (except what we can pick
3629 dst_reg->smin_value = S64_MIN;
3630 dst_reg->smax_value = S64_MAX;
3631 /* If we might shift our top bit out, then we know nothing */
3632 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
3633 dst_reg->umin_value = 0;
3634 dst_reg->umax_value = U64_MAX;
3636 dst_reg->umin_value <<= umin_val;
3637 dst_reg->umax_value <<= umax_val;
3639 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
3640 /* We may learn something more from the var_off */
3641 __update_reg_bounds(dst_reg);
3644 if (umax_val >= insn_bitness) {
3645 /* Shifts greater than 31 or 63 are undefined.
3646 * This includes shifts by a negative number.
3648 mark_reg_unknown(env, regs, insn->dst_reg);
3651 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
3652 * be negative, then either:
3653 * 1) src_reg might be zero, so the sign bit of the result is
3654 * unknown, so we lose our signed bounds
3655 * 2) it's known negative, thus the unsigned bounds capture the
3657 * 3) the signed bounds cross zero, so they tell us nothing
3659 * If the value in dst_reg is known nonnegative, then again the
3660 * unsigned bounts capture the signed bounds.
3661 * Thus, in all cases it suffices to blow away our signed bounds
3662 * and rely on inferring new ones from the unsigned bounds and
3663 * var_off of the result.
3665 dst_reg->smin_value = S64_MIN;
3666 dst_reg->smax_value = S64_MAX;
3667 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
3668 dst_reg->umin_value >>= umax_val;
3669 dst_reg->umax_value >>= umin_val;
3670 /* We may learn something more from the var_off */
3671 __update_reg_bounds(dst_reg);
3674 if (umax_val >= insn_bitness) {
3675 /* Shifts greater than 31 or 63 are undefined.
3676 * This includes shifts by a negative number.
3678 mark_reg_unknown(env, regs, insn->dst_reg);
3682 /* Upon reaching here, src_known is true and
3683 * umax_val is equal to umin_val.
3685 dst_reg->smin_value >>= umin_val;
3686 dst_reg->smax_value >>= umin_val;
3687 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val);
3689 /* blow away the dst_reg umin_value/umax_value and rely on
3690 * dst_reg var_off to refine the result.
3692 dst_reg->umin_value = 0;
3693 dst_reg->umax_value = U64_MAX;
3694 __update_reg_bounds(dst_reg);
3697 mark_reg_unknown(env, regs, insn->dst_reg);
3701 if (BPF_CLASS(insn->code) != BPF_ALU64) {
3702 /* 32-bit ALU ops are (32,32)->32 */
3703 coerce_reg_to_size(dst_reg, 4);
3706 __reg_deduce_bounds(dst_reg);
3707 __reg_bound_offset(dst_reg);
3711 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
3714 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
3715 struct bpf_insn *insn)
3717 struct bpf_verifier_state *vstate = env->cur_state;
3718 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3719 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
3720 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
3721 u8 opcode = BPF_OP(insn->code);
3723 dst_reg = ®s[insn->dst_reg];
3725 if (dst_reg->type != SCALAR_VALUE)
3727 if (BPF_SRC(insn->code) == BPF_X) {
3728 src_reg = ®s[insn->src_reg];
3729 if (src_reg->type != SCALAR_VALUE) {
3730 if (dst_reg->type != SCALAR_VALUE) {
3731 /* Combining two pointers by any ALU op yields
3732 * an arbitrary scalar. Disallow all math except
3733 * pointer subtraction
3735 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
3736 mark_reg_unknown(env, regs, insn->dst_reg);
3739 verbose(env, "R%d pointer %s pointer prohibited\n",
3741 bpf_alu_string[opcode >> 4]);
3744 /* scalar += pointer
3745 * This is legal, but we have to reverse our
3746 * src/dest handling in computing the range
3748 return adjust_ptr_min_max_vals(env, insn,
3751 } else if (ptr_reg) {
3752 /* pointer += scalar */
3753 return adjust_ptr_min_max_vals(env, insn,
3757 /* Pretend the src is a reg with a known value, since we only
3758 * need to be able to read from this state.
3760 off_reg.type = SCALAR_VALUE;
3761 __mark_reg_known(&off_reg, insn->imm);
3763 if (ptr_reg) /* pointer += K */
3764 return adjust_ptr_min_max_vals(env, insn,
3768 /* Got here implies adding two SCALAR_VALUEs */
3769 if (WARN_ON_ONCE(ptr_reg)) {
3770 print_verifier_state(env, state);
3771 verbose(env, "verifier internal error: unexpected ptr_reg\n");
3774 if (WARN_ON(!src_reg)) {
3775 print_verifier_state(env, state);
3776 verbose(env, "verifier internal error: no src_reg\n");
3779 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
3782 /* check validity of 32-bit and 64-bit arithmetic operations */
3783 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
3785 struct bpf_reg_state *regs = cur_regs(env);
3786 u8 opcode = BPF_OP(insn->code);
3789 if (opcode == BPF_END || opcode == BPF_NEG) {
3790 if (opcode == BPF_NEG) {
3791 if (BPF_SRC(insn->code) != 0 ||
3792 insn->src_reg != BPF_REG_0 ||
3793 insn->off != 0 || insn->imm != 0) {
3794 verbose(env, "BPF_NEG uses reserved fields\n");
3798 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
3799 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
3800 BPF_CLASS(insn->code) == BPF_ALU64) {
3801 verbose(env, "BPF_END uses reserved fields\n");
3806 /* check src operand */
3807 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3811 if (is_pointer_value(env, insn->dst_reg)) {
3812 verbose(env, "R%d pointer arithmetic prohibited\n",
3817 /* check dest operand */
3818 err = check_reg_arg(env, insn->dst_reg, DST_OP);
3822 } else if (opcode == BPF_MOV) {
3824 if (BPF_SRC(insn->code) == BPF_X) {
3825 if (insn->imm != 0 || insn->off != 0) {
3826 verbose(env, "BPF_MOV uses reserved fields\n");
3830 /* check src operand */
3831 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3835 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
3836 verbose(env, "BPF_MOV uses reserved fields\n");
3841 /* check dest operand, mark as required later */
3842 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
3846 if (BPF_SRC(insn->code) == BPF_X) {
3847 struct bpf_reg_state *src_reg = regs + insn->src_reg;
3848 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
3850 if (BPF_CLASS(insn->code) == BPF_ALU64) {
3852 * copy register state to dest reg
3854 *dst_reg = *src_reg;
3855 dst_reg->live |= REG_LIVE_WRITTEN;
3858 if (is_pointer_value(env, insn->src_reg)) {
3860 "R%d partial copy of pointer\n",
3863 } else if (src_reg->type == SCALAR_VALUE) {
3864 *dst_reg = *src_reg;
3865 dst_reg->live |= REG_LIVE_WRITTEN;
3867 mark_reg_unknown(env, regs,
3870 coerce_reg_to_size(dst_reg, 4);
3874 * remember the value we stored into this reg
3876 /* clear any state __mark_reg_known doesn't set */
3877 mark_reg_unknown(env, regs, insn->dst_reg);
3878 regs[insn->dst_reg].type = SCALAR_VALUE;
3879 if (BPF_CLASS(insn->code) == BPF_ALU64) {
3880 __mark_reg_known(regs + insn->dst_reg,
3883 __mark_reg_known(regs + insn->dst_reg,
3888 } else if (opcode > BPF_END) {
3889 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
3892 } else { /* all other ALU ops: and, sub, xor, add, ... */
3894 if (BPF_SRC(insn->code) == BPF_X) {
3895 if (insn->imm != 0 || insn->off != 0) {
3896 verbose(env, "BPF_ALU uses reserved fields\n");
3899 /* check src1 operand */
3900 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3904 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
3905 verbose(env, "BPF_ALU uses reserved fields\n");
3910 /* check src2 operand */
3911 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3915 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
3916 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
3917 verbose(env, "div by zero\n");
3921 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
3922 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
3923 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
3925 if (insn->imm < 0 || insn->imm >= size) {
3926 verbose(env, "invalid shift %d\n", insn->imm);
3931 /* check dest operand */
3932 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
3936 return adjust_reg_min_max_vals(env, insn);
3942 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
3943 struct bpf_reg_state *dst_reg,
3944 enum bpf_reg_type type,
3945 bool range_right_open)
3947 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3948 struct bpf_reg_state *regs = state->regs, *reg;
3952 if (dst_reg->off < 0 ||
3953 (dst_reg->off == 0 && range_right_open))
3954 /* This doesn't give us any range */
3957 if (dst_reg->umax_value > MAX_PACKET_OFF ||
3958 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
3959 /* Risk of overflow. For instance, ptr + (1<<63) may be less
3960 * than pkt_end, but that's because it's also less than pkt.
3964 new_range = dst_reg->off;
3965 if (range_right_open)
3968 /* Examples for register markings:
3970 * pkt_data in dst register:
3974 * if (r2 > pkt_end) goto <handle exception>
3979 * if (r2 < pkt_end) goto <access okay>
3980 * <handle exception>
3983 * r2 == dst_reg, pkt_end == src_reg
3984 * r2=pkt(id=n,off=8,r=0)
3985 * r3=pkt(id=n,off=0,r=0)
3987 * pkt_data in src register:
3991 * if (pkt_end >= r2) goto <access okay>
3992 * <handle exception>
3996 * if (pkt_end <= r2) goto <handle exception>
4000 * pkt_end == dst_reg, r2 == src_reg
4001 * r2=pkt(id=n,off=8,r=0)
4002 * r3=pkt(id=n,off=0,r=0)
4004 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
4005 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
4006 * and [r3, r3 + 8-1) respectively is safe to access depending on
4010 /* If our ids match, then we must have the same max_value. And we
4011 * don't care about the other reg's fixed offset, since if it's too big
4012 * the range won't allow anything.
4013 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
4015 for (i = 0; i < MAX_BPF_REG; i++)
4016 if (regs[i].type == type && regs[i].id == dst_reg->id)
4017 /* keep the maximum range already checked */
4018 regs[i].range = max(regs[i].range, new_range);
4020 for (j = 0; j <= vstate->curframe; j++) {
4021 state = vstate->frame[j];
4022 bpf_for_each_spilled_reg(i, state, reg) {
4025 if (reg->type == type && reg->id == dst_reg->id)
4026 reg->range = max(reg->range, new_range);
4031 /* compute branch direction of the expression "if (reg opcode val) goto target;"
4033 * 1 - branch will be taken and "goto target" will be executed
4034 * 0 - branch will not be taken and fall-through to next insn
4035 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value range [0,10]
4037 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
4039 if (__is_pointer_value(false, reg))
4044 if (tnum_is_const(reg->var_off))
4045 return !!tnum_equals_const(reg->var_off, val);
4048 if (tnum_is_const(reg->var_off))
4049 return !tnum_equals_const(reg->var_off, val);
4052 if ((~reg->var_off.mask & reg->var_off.value) & val)
4054 if (!((reg->var_off.mask | reg->var_off.value) & val))
4058 if (reg->umin_value > val)
4060 else if (reg->umax_value <= val)
4064 if (reg->smin_value > (s64)val)
4066 else if (reg->smax_value < (s64)val)
4070 if (reg->umax_value < val)
4072 else if (reg->umin_value >= val)
4076 if (reg->smax_value < (s64)val)
4078 else if (reg->smin_value >= (s64)val)
4082 if (reg->umin_value >= val)
4084 else if (reg->umax_value < val)
4088 if (reg->smin_value >= (s64)val)
4090 else if (reg->smax_value < (s64)val)
4094 if (reg->umax_value <= val)
4096 else if (reg->umin_value > val)
4100 if (reg->smax_value <= (s64)val)
4102 else if (reg->smin_value > (s64)val)
4110 /* Adjusts the register min/max values in the case that the dst_reg is the
4111 * variable register that we are working on, and src_reg is a constant or we're
4112 * simply doing a BPF_K check.
4113 * In JEQ/JNE cases we also adjust the var_off values.
4115 static void reg_set_min_max(struct bpf_reg_state *true_reg,
4116 struct bpf_reg_state *false_reg, u64 val,
4119 /* If the dst_reg is a pointer, we can't learn anything about its
4120 * variable offset from the compare (unless src_reg were a pointer into
4121 * the same object, but we don't bother with that.
4122 * Since false_reg and true_reg have the same type by construction, we
4123 * only need to check one of them for pointerness.
4125 if (__is_pointer_value(false, false_reg))
4130 /* If this is false then we know nothing Jon Snow, but if it is
4131 * true then we know for sure.
4133 __mark_reg_known(true_reg, val);
4136 /* If this is true we know nothing Jon Snow, but if it is false
4137 * we know the value for sure;
4139 __mark_reg_known(false_reg, val);
4142 false_reg->var_off = tnum_and(false_reg->var_off,
4144 if (is_power_of_2(val))
4145 true_reg->var_off = tnum_or(true_reg->var_off,
4149 false_reg->umax_value = min(false_reg->umax_value, val);
4150 true_reg->umin_value = max(true_reg->umin_value, val + 1);
4153 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
4154 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
4157 false_reg->umin_value = max(false_reg->umin_value, val);
4158 true_reg->umax_value = min(true_reg->umax_value, val - 1);
4161 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
4162 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
4165 false_reg->umax_value = min(false_reg->umax_value, val - 1);
4166 true_reg->umin_value = max(true_reg->umin_value, val);
4169 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
4170 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
4173 false_reg->umin_value = max(false_reg->umin_value, val + 1);
4174 true_reg->umax_value = min(true_reg->umax_value, val);
4177 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
4178 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
4184 __reg_deduce_bounds(false_reg);
4185 __reg_deduce_bounds(true_reg);
4186 /* We might have learned some bits from the bounds. */
4187 __reg_bound_offset(false_reg);
4188 __reg_bound_offset(true_reg);
4189 /* Intersecting with the old var_off might have improved our bounds
4190 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
4191 * then new var_off is (0; 0x7f...fc) which improves our umax.
4193 __update_reg_bounds(false_reg);
4194 __update_reg_bounds(true_reg);
4197 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
4200 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
4201 struct bpf_reg_state *false_reg, u64 val,
4204 if (__is_pointer_value(false, false_reg))
4209 /* If this is false then we know nothing Jon Snow, but if it is
4210 * true then we know for sure.
4212 __mark_reg_known(true_reg, val);
4215 /* If this is true we know nothing Jon Snow, but if it is false
4216 * we know the value for sure;
4218 __mark_reg_known(false_reg, val);
4221 false_reg->var_off = tnum_and(false_reg->var_off,
4223 if (is_power_of_2(val))
4224 true_reg->var_off = tnum_or(true_reg->var_off,
4228 true_reg->umax_value = min(true_reg->umax_value, val - 1);
4229 false_reg->umin_value = max(false_reg->umin_value, val);
4232 true_reg->smax_value = min_t(s64, true_reg->smax_value, val - 1);
4233 false_reg->smin_value = max_t(s64, false_reg->smin_value, val);
4236 true_reg->umin_value = max(true_reg->umin_value, val + 1);
4237 false_reg->umax_value = min(false_reg->umax_value, val);
4240 true_reg->smin_value = max_t(s64, true_reg->smin_value, val + 1);
4241 false_reg->smax_value = min_t(s64, false_reg->smax_value, val);
4244 true_reg->umax_value = min(true_reg->umax_value, val);
4245 false_reg->umin_value = max(false_reg->umin_value, val + 1);
4248 true_reg->smax_value = min_t(s64, true_reg->smax_value, val);
4249 false_reg->smin_value = max_t(s64, false_reg->smin_value, val + 1);
4252 true_reg->umin_value = max(true_reg->umin_value, val);
4253 false_reg->umax_value = min(false_reg->umax_value, val - 1);
4256 true_reg->smin_value = max_t(s64, true_reg->smin_value, val);
4257 false_reg->smax_value = min_t(s64, false_reg->smax_value, val - 1);
4263 __reg_deduce_bounds(false_reg);
4264 __reg_deduce_bounds(true_reg);
4265 /* We might have learned some bits from the bounds. */
4266 __reg_bound_offset(false_reg);
4267 __reg_bound_offset(true_reg);
4268 /* Intersecting with the old var_off might have improved our bounds
4269 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
4270 * then new var_off is (0; 0x7f...fc) which improves our umax.
4272 __update_reg_bounds(false_reg);
4273 __update_reg_bounds(true_reg);
4276 /* Regs are known to be equal, so intersect their min/max/var_off */
4277 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
4278 struct bpf_reg_state *dst_reg)
4280 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
4281 dst_reg->umin_value);
4282 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
4283 dst_reg->umax_value);
4284 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
4285 dst_reg->smin_value);
4286 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
4287 dst_reg->smax_value);
4288 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
4290 /* We might have learned new bounds from the var_off. */
4291 __update_reg_bounds(src_reg);
4292 __update_reg_bounds(dst_reg);
4293 /* We might have learned something about the sign bit. */
4294 __reg_deduce_bounds(src_reg);
4295 __reg_deduce_bounds(dst_reg);
4296 /* We might have learned some bits from the bounds. */
4297 __reg_bound_offset(src_reg);
4298 __reg_bound_offset(dst_reg);
4299 /* Intersecting with the old var_off might have improved our bounds
4300 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
4301 * then new var_off is (0; 0x7f...fc) which improves our umax.
4303 __update_reg_bounds(src_reg);
4304 __update_reg_bounds(dst_reg);
4307 static void reg_combine_min_max(struct bpf_reg_state *true_src,
4308 struct bpf_reg_state *true_dst,
4309 struct bpf_reg_state *false_src,
4310 struct bpf_reg_state *false_dst,
4315 __reg_combine_min_max(true_src, true_dst);
4318 __reg_combine_min_max(false_src, false_dst);
4323 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
4324 struct bpf_reg_state *reg, u32 id,
4327 if (reg_type_may_be_null(reg->type) && reg->id == id) {
4328 /* Old offset (both fixed and variable parts) should
4329 * have been known-zero, because we don't allow pointer
4330 * arithmetic on pointers that might be NULL.
4332 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
4333 !tnum_equals_const(reg->var_off, 0) ||
4335 __mark_reg_known_zero(reg);
4339 reg->type = SCALAR_VALUE;
4340 } else if (reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
4341 if (reg->map_ptr->inner_map_meta) {
4342 reg->type = CONST_PTR_TO_MAP;
4343 reg->map_ptr = reg->map_ptr->inner_map_meta;
4345 reg->type = PTR_TO_MAP_VALUE;
4347 } else if (reg->type == PTR_TO_SOCKET_OR_NULL) {
4348 reg->type = PTR_TO_SOCKET;
4350 if (is_null || !reg_is_refcounted(reg)) {
4351 /* We don't need id from this point onwards anymore,
4352 * thus we should better reset it, so that state
4353 * pruning has chances to take effect.
4360 /* The logic is similar to find_good_pkt_pointers(), both could eventually
4361 * be folded together at some point.
4363 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
4366 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4367 struct bpf_reg_state *reg, *regs = state->regs;
4368 u32 id = regs[regno].id;
4371 if (reg_is_refcounted_or_null(®s[regno]) && is_null)
4372 __release_reference_state(state, id);
4374 for (i = 0; i < MAX_BPF_REG; i++)
4375 mark_ptr_or_null_reg(state, ®s[i], id, is_null);
4377 for (j = 0; j <= vstate->curframe; j++) {
4378 state = vstate->frame[j];
4379 bpf_for_each_spilled_reg(i, state, reg) {
4382 mark_ptr_or_null_reg(state, reg, id, is_null);
4387 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
4388 struct bpf_reg_state *dst_reg,
4389 struct bpf_reg_state *src_reg,
4390 struct bpf_verifier_state *this_branch,
4391 struct bpf_verifier_state *other_branch)
4393 if (BPF_SRC(insn->code) != BPF_X)
4396 switch (BPF_OP(insn->code)) {
4398 if ((dst_reg->type == PTR_TO_PACKET &&
4399 src_reg->type == PTR_TO_PACKET_END) ||
4400 (dst_reg->type == PTR_TO_PACKET_META &&
4401 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
4402 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
4403 find_good_pkt_pointers(this_branch, dst_reg,
4404 dst_reg->type, false);
4405 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
4406 src_reg->type == PTR_TO_PACKET) ||
4407 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
4408 src_reg->type == PTR_TO_PACKET_META)) {
4409 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
4410 find_good_pkt_pointers(other_branch, src_reg,
4411 src_reg->type, true);
4417 if ((dst_reg->type == PTR_TO_PACKET &&
4418 src_reg->type == PTR_TO_PACKET_END) ||
4419 (dst_reg->type == PTR_TO_PACKET_META &&
4420 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
4421 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
4422 find_good_pkt_pointers(other_branch, dst_reg,
4423 dst_reg->type, true);
4424 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
4425 src_reg->type == PTR_TO_PACKET) ||
4426 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
4427 src_reg->type == PTR_TO_PACKET_META)) {
4428 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
4429 find_good_pkt_pointers(this_branch, src_reg,
4430 src_reg->type, false);
4436 if ((dst_reg->type == PTR_TO_PACKET &&
4437 src_reg->type == PTR_TO_PACKET_END) ||
4438 (dst_reg->type == PTR_TO_PACKET_META &&
4439 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
4440 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
4441 find_good_pkt_pointers(this_branch, dst_reg,
4442 dst_reg->type, true);
4443 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
4444 src_reg->type == PTR_TO_PACKET) ||
4445 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
4446 src_reg->type == PTR_TO_PACKET_META)) {
4447 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
4448 find_good_pkt_pointers(other_branch, src_reg,
4449 src_reg->type, false);
4455 if ((dst_reg->type == PTR_TO_PACKET &&
4456 src_reg->type == PTR_TO_PACKET_END) ||
4457 (dst_reg->type == PTR_TO_PACKET_META &&
4458 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
4459 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
4460 find_good_pkt_pointers(other_branch, dst_reg,
4461 dst_reg->type, false);
4462 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
4463 src_reg->type == PTR_TO_PACKET) ||
4464 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
4465 src_reg->type == PTR_TO_PACKET_META)) {
4466 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
4467 find_good_pkt_pointers(this_branch, src_reg,
4468 src_reg->type, true);
4480 static int check_cond_jmp_op(struct bpf_verifier_env *env,
4481 struct bpf_insn *insn, int *insn_idx)
4483 struct bpf_verifier_state *this_branch = env->cur_state;
4484 struct bpf_verifier_state *other_branch;
4485 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
4486 struct bpf_reg_state *dst_reg, *other_branch_regs;
4487 u8 opcode = BPF_OP(insn->code);
4490 if (opcode > BPF_JSLE) {
4491 verbose(env, "invalid BPF_JMP opcode %x\n", opcode);
4495 if (BPF_SRC(insn->code) == BPF_X) {
4496 if (insn->imm != 0) {
4497 verbose(env, "BPF_JMP uses reserved fields\n");
4501 /* check src1 operand */
4502 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4506 if (is_pointer_value(env, insn->src_reg)) {
4507 verbose(env, "R%d pointer comparison prohibited\n",
4512 if (insn->src_reg != BPF_REG_0) {
4513 verbose(env, "BPF_JMP uses reserved fields\n");
4518 /* check src2 operand */
4519 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4523 dst_reg = ®s[insn->dst_reg];
4525 if (BPF_SRC(insn->code) == BPF_K) {
4526 int pred = is_branch_taken(dst_reg, insn->imm, opcode);
4529 /* only follow the goto, ignore fall-through */
4530 *insn_idx += insn->off;
4532 } else if (pred == 0) {
4533 /* only follow fall-through branch, since
4534 * that's where the program will go
4540 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
4544 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
4546 /* detect if we are comparing against a constant value so we can adjust
4547 * our min/max values for our dst register.
4548 * this is only legit if both are scalars (or pointers to the same
4549 * object, I suppose, but we don't support that right now), because
4550 * otherwise the different base pointers mean the offsets aren't
4553 if (BPF_SRC(insn->code) == BPF_X) {
4554 if (dst_reg->type == SCALAR_VALUE &&
4555 regs[insn->src_reg].type == SCALAR_VALUE) {
4556 if (tnum_is_const(regs[insn->src_reg].var_off))
4557 reg_set_min_max(&other_branch_regs[insn->dst_reg],
4558 dst_reg, regs[insn->src_reg].var_off.value,
4560 else if (tnum_is_const(dst_reg->var_off))
4561 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
4562 ®s[insn->src_reg],
4563 dst_reg->var_off.value, opcode);
4564 else if (opcode == BPF_JEQ || opcode == BPF_JNE)
4565 /* Comparing for equality, we can combine knowledge */
4566 reg_combine_min_max(&other_branch_regs[insn->src_reg],
4567 &other_branch_regs[insn->dst_reg],
4568 ®s[insn->src_reg],
4569 ®s[insn->dst_reg], opcode);
4571 } else if (dst_reg->type == SCALAR_VALUE) {
4572 reg_set_min_max(&other_branch_regs[insn->dst_reg],
4573 dst_reg, insn->imm, opcode);
4576 /* detect if R == 0 where R is returned from bpf_map_lookup_elem() */
4577 if (BPF_SRC(insn->code) == BPF_K &&
4578 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
4579 reg_type_may_be_null(dst_reg->type)) {
4580 /* Mark all identical registers in each branch as either
4581 * safe or unknown depending R == 0 or R != 0 conditional.
4583 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
4585 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
4587 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
4588 this_branch, other_branch) &&
4589 is_pointer_value(env, insn->dst_reg)) {
4590 verbose(env, "R%d pointer comparison prohibited\n",
4595 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
4599 /* return the map pointer stored inside BPF_LD_IMM64 instruction */
4600 static struct bpf_map *ld_imm64_to_map_ptr(struct bpf_insn *insn)
4602 u64 imm64 = ((u64) (u32) insn[0].imm) | ((u64) (u32) insn[1].imm) << 32;
4604 return (struct bpf_map *) (unsigned long) imm64;
4607 /* verify BPF_LD_IMM64 instruction */
4608 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
4610 struct bpf_reg_state *regs = cur_regs(env);
4613 if (BPF_SIZE(insn->code) != BPF_DW) {
4614 verbose(env, "invalid BPF_LD_IMM insn\n");
4617 if (insn->off != 0) {
4618 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
4622 err = check_reg_arg(env, insn->dst_reg, DST_OP);
4626 if (insn->src_reg == 0) {
4627 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
4629 regs[insn->dst_reg].type = SCALAR_VALUE;
4630 __mark_reg_known(®s[insn->dst_reg], imm);
4634 /* replace_map_fd_with_map_ptr() should have caught bad ld_imm64 */
4635 BUG_ON(insn->src_reg != BPF_PSEUDO_MAP_FD);
4637 regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
4638 regs[insn->dst_reg].map_ptr = ld_imm64_to_map_ptr(insn);
4642 static bool may_access_skb(enum bpf_prog_type type)
4645 case BPF_PROG_TYPE_SOCKET_FILTER:
4646 case BPF_PROG_TYPE_SCHED_CLS:
4647 case BPF_PROG_TYPE_SCHED_ACT:
4654 /* verify safety of LD_ABS|LD_IND instructions:
4655 * - they can only appear in the programs where ctx == skb
4656 * - since they are wrappers of function calls, they scratch R1-R5 registers,
4657 * preserve R6-R9, and store return value into R0
4660 * ctx == skb == R6 == CTX
4663 * SRC == any register
4664 * IMM == 32-bit immediate
4667 * R0 - 8/16/32-bit skb data converted to cpu endianness
4669 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
4671 struct bpf_reg_state *regs = cur_regs(env);
4672 u8 mode = BPF_MODE(insn->code);
4675 if (!may_access_skb(env->prog->type)) {
4676 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
4680 if (!env->ops->gen_ld_abs) {
4681 verbose(env, "bpf verifier is misconfigured\n");
4685 if (env->subprog_cnt > 1) {
4686 /* when program has LD_ABS insn JITs and interpreter assume
4687 * that r1 == ctx == skb which is not the case for callees
4688 * that can have arbitrary arguments. It's problematic
4689 * for main prog as well since JITs would need to analyze
4690 * all functions in order to make proper register save/restore
4691 * decisions in the main prog. Hence disallow LD_ABS with calls
4693 verbose(env, "BPF_LD_[ABS|IND] instructions cannot be mixed with bpf-to-bpf calls\n");
4697 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
4698 BPF_SIZE(insn->code) == BPF_DW ||
4699 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
4700 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
4704 /* check whether implicit source operand (register R6) is readable */
4705 err = check_reg_arg(env, BPF_REG_6, SRC_OP);
4709 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
4710 * gen_ld_abs() may terminate the program at runtime, leading to
4713 err = check_reference_leak(env);
4715 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
4719 if (regs[BPF_REG_6].type != PTR_TO_CTX) {
4721 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
4725 if (mode == BPF_IND) {
4726 /* check explicit source operand */
4727 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4732 /* reset caller saved regs to unreadable */
4733 for (i = 0; i < CALLER_SAVED_REGS; i++) {
4734 mark_reg_not_init(env, regs, caller_saved[i]);
4735 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
4738 /* mark destination R0 register as readable, since it contains
4739 * the value fetched from the packet.
4740 * Already marked as written above.
4742 mark_reg_unknown(env, regs, BPF_REG_0);
4746 static int check_return_code(struct bpf_verifier_env *env)
4748 struct bpf_reg_state *reg;
4749 struct tnum range = tnum_range(0, 1);
4751 switch (env->prog->type) {
4752 case BPF_PROG_TYPE_CGROUP_SKB:
4753 case BPF_PROG_TYPE_CGROUP_SOCK:
4754 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
4755 case BPF_PROG_TYPE_SOCK_OPS:
4756 case BPF_PROG_TYPE_CGROUP_DEVICE:
4762 reg = cur_regs(env) + BPF_REG_0;
4763 if (reg->type != SCALAR_VALUE) {
4764 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
4765 reg_type_str[reg->type]);
4769 if (!tnum_in(range, reg->var_off)) {
4770 verbose(env, "At program exit the register R0 ");
4771 if (!tnum_is_unknown(reg->var_off)) {
4774 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4775 verbose(env, "has value %s", tn_buf);
4777 verbose(env, "has unknown scalar value");
4779 verbose(env, " should have been 0 or 1\n");
4785 /* non-recursive DFS pseudo code
4786 * 1 procedure DFS-iterative(G,v):
4787 * 2 label v as discovered
4788 * 3 let S be a stack
4790 * 5 while S is not empty
4792 * 7 if t is what we're looking for:
4794 * 9 for all edges e in G.adjacentEdges(t) do
4795 * 10 if edge e is already labelled
4796 * 11 continue with the next edge
4797 * 12 w <- G.adjacentVertex(t,e)
4798 * 13 if vertex w is not discovered and not explored
4799 * 14 label e as tree-edge
4800 * 15 label w as discovered
4803 * 18 else if vertex w is discovered
4804 * 19 label e as back-edge
4806 * 21 // vertex w is explored
4807 * 22 label e as forward- or cross-edge
4808 * 23 label t as explored
4813 * 0x11 - discovered and fall-through edge labelled
4814 * 0x12 - discovered and fall-through and branch edges labelled
4825 #define STATE_LIST_MARK ((struct bpf_verifier_state_list *) -1L)
4827 static int *insn_stack; /* stack of insns to process */
4828 static int cur_stack; /* current stack index */
4829 static int *insn_state;
4831 /* t, w, e - match pseudo-code above:
4832 * t - index of current instruction
4833 * w - next instruction
4836 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env)
4838 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
4841 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
4844 if (w < 0 || w >= env->prog->len) {
4845 verbose_linfo(env, t, "%d: ", t);
4846 verbose(env, "jump out of range from insn %d to %d\n", t, w);
4851 /* mark branch target for state pruning */
4852 env->explored_states[w] = STATE_LIST_MARK;
4854 if (insn_state[w] == 0) {
4856 insn_state[t] = DISCOVERED | e;
4857 insn_state[w] = DISCOVERED;
4858 if (cur_stack >= env->prog->len)
4860 insn_stack[cur_stack++] = w;
4862 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
4863 verbose_linfo(env, t, "%d: ", t);
4864 verbose_linfo(env, w, "%d: ", w);
4865 verbose(env, "back-edge from insn %d to %d\n", t, w);
4867 } else if (insn_state[w] == EXPLORED) {
4868 /* forward- or cross-edge */
4869 insn_state[t] = DISCOVERED | e;
4871 verbose(env, "insn state internal bug\n");
4877 /* non-recursive depth-first-search to detect loops in BPF program
4878 * loop == back-edge in directed graph
4880 static int check_cfg(struct bpf_verifier_env *env)
4882 struct bpf_insn *insns = env->prog->insnsi;
4883 int insn_cnt = env->prog->len;
4887 insn_state = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
4891 insn_stack = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
4897 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
4898 insn_stack[0] = 0; /* 0 is the first instruction */
4904 t = insn_stack[cur_stack - 1];
4906 if (BPF_CLASS(insns[t].code) == BPF_JMP) {
4907 u8 opcode = BPF_OP(insns[t].code);
4909 if (opcode == BPF_EXIT) {
4911 } else if (opcode == BPF_CALL) {
4912 ret = push_insn(t, t + 1, FALLTHROUGH, env);
4917 if (t + 1 < insn_cnt)
4918 env->explored_states[t + 1] = STATE_LIST_MARK;
4919 if (insns[t].src_reg == BPF_PSEUDO_CALL) {
4920 env->explored_states[t] = STATE_LIST_MARK;
4921 ret = push_insn(t, t + insns[t].imm + 1, BRANCH, env);
4927 } else if (opcode == BPF_JA) {
4928 if (BPF_SRC(insns[t].code) != BPF_K) {
4932 /* unconditional jump with single edge */
4933 ret = push_insn(t, t + insns[t].off + 1,
4939 /* tell verifier to check for equivalent states
4940 * after every call and jump
4942 if (t + 1 < insn_cnt)
4943 env->explored_states[t + 1] = STATE_LIST_MARK;
4945 /* conditional jump with two edges */
4946 env->explored_states[t] = STATE_LIST_MARK;
4947 ret = push_insn(t, t + 1, FALLTHROUGH, env);
4953 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env);
4960 /* all other non-branch instructions with single
4963 ret = push_insn(t, t + 1, FALLTHROUGH, env);
4971 insn_state[t] = EXPLORED;
4972 if (cur_stack-- <= 0) {
4973 verbose(env, "pop stack internal bug\n");
4980 for (i = 0; i < insn_cnt; i++) {
4981 if (insn_state[i] != EXPLORED) {
4982 verbose(env, "unreachable insn %d\n", i);
4987 ret = 0; /* cfg looks good */
4995 /* The minimum supported BTF func info size */
4996 #define MIN_BPF_FUNCINFO_SIZE 8
4997 #define MAX_FUNCINFO_REC_SIZE 252
4999 static int check_btf_func(struct bpf_verifier_env *env,
5000 const union bpf_attr *attr,
5001 union bpf_attr __user *uattr)
5003 u32 i, nfuncs, urec_size, min_size, prev_offset;
5004 u32 krec_size = sizeof(struct bpf_func_info);
5005 struct bpf_func_info *krecord;
5006 const struct btf_type *type;
5007 struct bpf_prog *prog;
5008 const struct btf *btf;
5009 void __user *urecord;
5012 nfuncs = attr->func_info_cnt;
5016 if (nfuncs != env->subprog_cnt) {
5017 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
5021 urec_size = attr->func_info_rec_size;
5022 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
5023 urec_size > MAX_FUNCINFO_REC_SIZE ||
5024 urec_size % sizeof(u32)) {
5025 verbose(env, "invalid func info rec size %u\n", urec_size);
5030 btf = prog->aux->btf;
5032 urecord = u64_to_user_ptr(attr->func_info);
5033 min_size = min_t(u32, krec_size, urec_size);
5035 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
5039 for (i = 0; i < nfuncs; i++) {
5040 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
5042 if (ret == -E2BIG) {
5043 verbose(env, "nonzero tailing record in func info");
5044 /* set the size kernel expects so loader can zero
5045 * out the rest of the record.
5047 if (put_user(min_size, &uattr->func_info_rec_size))
5053 if (copy_from_user(&krecord[i], urecord, min_size)) {
5058 /* check insn_off */
5060 if (krecord[i].insn_off) {
5062 "nonzero insn_off %u for the first func info record",
5063 krecord[i].insn_off);
5067 } else if (krecord[i].insn_off <= prev_offset) {
5069 "same or smaller insn offset (%u) than previous func info record (%u)",
5070 krecord[i].insn_off, prev_offset);
5075 if (env->subprog_info[i].start != krecord[i].insn_off) {
5076 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
5082 type = btf_type_by_id(btf, krecord[i].type_id);
5083 if (!type || BTF_INFO_KIND(type->info) != BTF_KIND_FUNC) {
5084 verbose(env, "invalid type id %d in func info",
5085 krecord[i].type_id);
5090 prev_offset = krecord[i].insn_off;
5091 urecord += urec_size;
5094 prog->aux->func_info = krecord;
5095 prog->aux->func_info_cnt = nfuncs;
5103 static void adjust_btf_func(struct bpf_verifier_env *env)
5107 if (!env->prog->aux->func_info)
5110 for (i = 0; i < env->subprog_cnt; i++)
5111 env->prog->aux->func_info[i].insn_off = env->subprog_info[i].start;
5114 #define MIN_BPF_LINEINFO_SIZE (offsetof(struct bpf_line_info, line_col) + \
5115 sizeof(((struct bpf_line_info *)(0))->line_col))
5116 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
5118 static int check_btf_line(struct bpf_verifier_env *env,
5119 const union bpf_attr *attr,
5120 union bpf_attr __user *uattr)
5122 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
5123 struct bpf_subprog_info *sub;
5124 struct bpf_line_info *linfo;
5125 struct bpf_prog *prog;
5126 const struct btf *btf;
5127 void __user *ulinfo;
5130 nr_linfo = attr->line_info_cnt;
5134 rec_size = attr->line_info_rec_size;
5135 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
5136 rec_size > MAX_LINEINFO_REC_SIZE ||
5137 rec_size & (sizeof(u32) - 1))
5140 /* Need to zero it in case the userspace may
5141 * pass in a smaller bpf_line_info object.
5143 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
5144 GFP_KERNEL | __GFP_NOWARN);
5149 btf = prog->aux->btf;
5152 sub = env->subprog_info;
5153 ulinfo = u64_to_user_ptr(attr->line_info);
5154 expected_size = sizeof(struct bpf_line_info);
5155 ncopy = min_t(u32, expected_size, rec_size);
5156 for (i = 0; i < nr_linfo; i++) {
5157 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
5159 if (err == -E2BIG) {
5160 verbose(env, "nonzero tailing record in line_info");
5161 if (put_user(expected_size,
5162 &uattr->line_info_rec_size))
5168 if (copy_from_user(&linfo[i], ulinfo, ncopy)) {
5174 * Check insn_off to ensure
5175 * 1) strictly increasing AND
5176 * 2) bounded by prog->len
5178 * The linfo[0].insn_off == 0 check logically falls into
5179 * the later "missing bpf_line_info for func..." case
5180 * because the first linfo[0].insn_off must be the
5181 * first sub also and the first sub must have
5182 * subprog_info[0].start == 0.
5184 if ((i && linfo[i].insn_off <= prev_offset) ||
5185 linfo[i].insn_off >= prog->len) {
5186 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
5187 i, linfo[i].insn_off, prev_offset,
5193 if (!prog->insnsi[linfo[i].insn_off].code) {
5195 "Invalid insn code at line_info[%u].insn_off\n",
5201 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
5202 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
5203 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
5208 if (s != env->subprog_cnt) {
5209 if (linfo[i].insn_off == sub[s].start) {
5210 sub[s].linfo_idx = i;
5212 } else if (sub[s].start < linfo[i].insn_off) {
5213 verbose(env, "missing bpf_line_info for func#%u\n", s);
5219 prev_offset = linfo[i].insn_off;
5223 if (s != env->subprog_cnt) {
5224 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
5225 env->subprog_cnt - s, s);
5230 prog->aux->linfo = linfo;
5231 prog->aux->nr_linfo = nr_linfo;
5240 static int check_btf_info(struct bpf_verifier_env *env,
5241 const union bpf_attr *attr,
5242 union bpf_attr __user *uattr)
5247 if (!attr->func_info_cnt && !attr->line_info_cnt)
5250 btf = btf_get_by_fd(attr->prog_btf_fd);
5252 return PTR_ERR(btf);
5253 env->prog->aux->btf = btf;
5255 err = check_btf_func(env, attr, uattr);
5259 err = check_btf_line(env, attr, uattr);
5266 /* check %cur's range satisfies %old's */
5267 static bool range_within(struct bpf_reg_state *old,
5268 struct bpf_reg_state *cur)
5270 return old->umin_value <= cur->umin_value &&
5271 old->umax_value >= cur->umax_value &&
5272 old->smin_value <= cur->smin_value &&
5273 old->smax_value >= cur->smax_value;
5276 /* Maximum number of register states that can exist at once */
5277 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
5283 /* If in the old state two registers had the same id, then they need to have
5284 * the same id in the new state as well. But that id could be different from
5285 * the old state, so we need to track the mapping from old to new ids.
5286 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
5287 * regs with old id 5 must also have new id 9 for the new state to be safe. But
5288 * regs with a different old id could still have new id 9, we don't care about
5290 * So we look through our idmap to see if this old id has been seen before. If
5291 * so, we require the new id to match; otherwise, we add the id pair to the map.
5293 static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
5297 for (i = 0; i < ID_MAP_SIZE; i++) {
5298 if (!idmap[i].old) {
5299 /* Reached an empty slot; haven't seen this id before */
5300 idmap[i].old = old_id;
5301 idmap[i].cur = cur_id;
5304 if (idmap[i].old == old_id)
5305 return idmap[i].cur == cur_id;
5307 /* We ran out of idmap slots, which should be impossible */
5312 static void clean_func_state(struct bpf_verifier_env *env,
5313 struct bpf_func_state *st)
5315 enum bpf_reg_liveness live;
5318 for (i = 0; i < BPF_REG_FP; i++) {
5319 live = st->regs[i].live;
5320 /* liveness must not touch this register anymore */
5321 st->regs[i].live |= REG_LIVE_DONE;
5322 if (!(live & REG_LIVE_READ))
5323 /* since the register is unused, clear its state
5324 * to make further comparison simpler
5326 __mark_reg_not_init(&st->regs[i]);
5329 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
5330 live = st->stack[i].spilled_ptr.live;
5331 /* liveness must not touch this stack slot anymore */
5332 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
5333 if (!(live & REG_LIVE_READ)) {
5334 __mark_reg_not_init(&st->stack[i].spilled_ptr);
5335 for (j = 0; j < BPF_REG_SIZE; j++)
5336 st->stack[i].slot_type[j] = STACK_INVALID;
5341 static void clean_verifier_state(struct bpf_verifier_env *env,
5342 struct bpf_verifier_state *st)
5346 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
5347 /* all regs in this state in all frames were already marked */
5350 for (i = 0; i <= st->curframe; i++)
5351 clean_func_state(env, st->frame[i]);
5354 /* the parentage chains form a tree.
5355 * the verifier states are added to state lists at given insn and
5356 * pushed into state stack for future exploration.
5357 * when the verifier reaches bpf_exit insn some of the verifer states
5358 * stored in the state lists have their final liveness state already,
5359 * but a lot of states will get revised from liveness point of view when
5360 * the verifier explores other branches.
5363 * 2: if r1 == 100 goto pc+1
5366 * when the verifier reaches exit insn the register r0 in the state list of
5367 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
5368 * of insn 2 and goes exploring further. At the insn 4 it will walk the
5369 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
5371 * Since the verifier pushes the branch states as it sees them while exploring
5372 * the program the condition of walking the branch instruction for the second
5373 * time means that all states below this branch were already explored and
5374 * their final liveness markes are already propagated.
5375 * Hence when the verifier completes the search of state list in is_state_visited()
5376 * we can call this clean_live_states() function to mark all liveness states
5377 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
5379 * This function also clears the registers and stack for states that !READ
5380 * to simplify state merging.
5382 * Important note here that walking the same branch instruction in the callee
5383 * doesn't meant that the states are DONE. The verifier has to compare
5386 static void clean_live_states(struct bpf_verifier_env *env, int insn,
5387 struct bpf_verifier_state *cur)
5389 struct bpf_verifier_state_list *sl;
5392 sl = env->explored_states[insn];
5396 while (sl != STATE_LIST_MARK) {
5397 if (sl->state.curframe != cur->curframe)
5399 for (i = 0; i <= cur->curframe; i++)
5400 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
5402 clean_verifier_state(env, &sl->state);
5408 /* Returns true if (rold safe implies rcur safe) */
5409 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
5410 struct idpair *idmap)
5414 if (!(rold->live & REG_LIVE_READ))
5415 /* explored state didn't use this */
5418 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
5420 if (rold->type == PTR_TO_STACK)
5421 /* two stack pointers are equal only if they're pointing to
5422 * the same stack frame, since fp-8 in foo != fp-8 in bar
5424 return equal && rold->frameno == rcur->frameno;
5429 if (rold->type == NOT_INIT)
5430 /* explored state can't have used this */
5432 if (rcur->type == NOT_INIT)
5434 switch (rold->type) {
5436 if (rcur->type == SCALAR_VALUE) {
5437 /* new val must satisfy old val knowledge */
5438 return range_within(rold, rcur) &&
5439 tnum_in(rold->var_off, rcur->var_off);
5441 /* We're trying to use a pointer in place of a scalar.
5442 * Even if the scalar was unbounded, this could lead to
5443 * pointer leaks because scalars are allowed to leak
5444 * while pointers are not. We could make this safe in
5445 * special cases if root is calling us, but it's
5446 * probably not worth the hassle.
5450 case PTR_TO_MAP_VALUE:
5451 /* If the new min/max/var_off satisfy the old ones and
5452 * everything else matches, we are OK.
5453 * We don't care about the 'id' value, because nothing
5454 * uses it for PTR_TO_MAP_VALUE (only for ..._OR_NULL)
5456 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
5457 range_within(rold, rcur) &&
5458 tnum_in(rold->var_off, rcur->var_off);
5459 case PTR_TO_MAP_VALUE_OR_NULL:
5460 /* a PTR_TO_MAP_VALUE could be safe to use as a
5461 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
5462 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
5463 * checked, doing so could have affected others with the same
5464 * id, and we can't check for that because we lost the id when
5465 * we converted to a PTR_TO_MAP_VALUE.
5467 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
5469 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
5471 /* Check our ids match any regs they're supposed to */
5472 return check_ids(rold->id, rcur->id, idmap);
5473 case PTR_TO_PACKET_META:
5475 if (rcur->type != rold->type)
5477 /* We must have at least as much range as the old ptr
5478 * did, so that any accesses which were safe before are
5479 * still safe. This is true even if old range < old off,
5480 * since someone could have accessed through (ptr - k), or
5481 * even done ptr -= k in a register, to get a safe access.
5483 if (rold->range > rcur->range)
5485 /* If the offsets don't match, we can't trust our alignment;
5486 * nor can we be sure that we won't fall out of range.
5488 if (rold->off != rcur->off)
5490 /* id relations must be preserved */
5491 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
5493 /* new val must satisfy old val knowledge */
5494 return range_within(rold, rcur) &&
5495 tnum_in(rold->var_off, rcur->var_off);
5497 case CONST_PTR_TO_MAP:
5498 case PTR_TO_PACKET_END:
5499 case PTR_TO_FLOW_KEYS:
5501 case PTR_TO_SOCKET_OR_NULL:
5502 /* Only valid matches are exact, which memcmp() above
5503 * would have accepted
5506 /* Don't know what's going on, just say it's not safe */
5510 /* Shouldn't get here; if we do, say it's not safe */
5515 static bool stacksafe(struct bpf_func_state *old,
5516 struct bpf_func_state *cur,
5517 struct idpair *idmap)
5521 /* walk slots of the explored stack and ignore any additional
5522 * slots in the current stack, since explored(safe) state
5525 for (i = 0; i < old->allocated_stack; i++) {
5526 spi = i / BPF_REG_SIZE;
5528 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
5529 i += BPF_REG_SIZE - 1;
5530 /* explored state didn't use this */
5534 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
5537 /* explored stack has more populated slots than current stack
5538 * and these slots were used
5540 if (i >= cur->allocated_stack)
5543 /* if old state was safe with misc data in the stack
5544 * it will be safe with zero-initialized stack.
5545 * The opposite is not true
5547 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
5548 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
5550 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
5551 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
5552 /* Ex: old explored (safe) state has STACK_SPILL in
5553 * this stack slot, but current has has STACK_MISC ->
5554 * this verifier states are not equivalent,
5555 * return false to continue verification of this path
5558 if (i % BPF_REG_SIZE)
5560 if (old->stack[spi].slot_type[0] != STACK_SPILL)
5562 if (!regsafe(&old->stack[spi].spilled_ptr,
5563 &cur->stack[spi].spilled_ptr,
5565 /* when explored and current stack slot are both storing
5566 * spilled registers, check that stored pointers types
5567 * are the same as well.
5568 * Ex: explored safe path could have stored
5569 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
5570 * but current path has stored:
5571 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
5572 * such verifier states are not equivalent.
5573 * return false to continue verification of this path
5580 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
5582 if (old->acquired_refs != cur->acquired_refs)
5584 return !memcmp(old->refs, cur->refs,
5585 sizeof(*old->refs) * old->acquired_refs);
5588 /* compare two verifier states
5590 * all states stored in state_list are known to be valid, since
5591 * verifier reached 'bpf_exit' instruction through them
5593 * this function is called when verifier exploring different branches of
5594 * execution popped from the state stack. If it sees an old state that has
5595 * more strict register state and more strict stack state then this execution
5596 * branch doesn't need to be explored further, since verifier already
5597 * concluded that more strict state leads to valid finish.
5599 * Therefore two states are equivalent if register state is more conservative
5600 * and explored stack state is more conservative than the current one.
5603 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
5604 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
5606 * In other words if current stack state (one being explored) has more
5607 * valid slots than old one that already passed validation, it means
5608 * the verifier can stop exploring and conclude that current state is valid too
5610 * Similarly with registers. If explored state has register type as invalid
5611 * whereas register type in current state is meaningful, it means that
5612 * the current state will reach 'bpf_exit' instruction safely
5614 static bool func_states_equal(struct bpf_func_state *old,
5615 struct bpf_func_state *cur)
5617 struct idpair *idmap;
5621 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
5622 /* If we failed to allocate the idmap, just say it's not safe */
5626 for (i = 0; i < MAX_BPF_REG; i++) {
5627 if (!regsafe(&old->regs[i], &cur->regs[i], idmap))
5631 if (!stacksafe(old, cur, idmap))
5634 if (!refsafe(old, cur))
5642 static bool states_equal(struct bpf_verifier_env *env,
5643 struct bpf_verifier_state *old,
5644 struct bpf_verifier_state *cur)
5648 if (old->curframe != cur->curframe)
5651 /* Verification state from speculative execution simulation
5652 * must never prune a non-speculative execution one.
5654 if (old->speculative && !cur->speculative)
5657 /* for states to be equal callsites have to be the same
5658 * and all frame states need to be equivalent
5660 for (i = 0; i <= old->curframe; i++) {
5661 if (old->frame[i]->callsite != cur->frame[i]->callsite)
5663 if (!func_states_equal(old->frame[i], cur->frame[i]))
5669 /* A write screens off any subsequent reads; but write marks come from the
5670 * straight-line code between a state and its parent. When we arrive at an
5671 * equivalent state (jump target or such) we didn't arrive by the straight-line
5672 * code, so read marks in the state must propagate to the parent regardless
5673 * of the state's write marks. That's what 'parent == state->parent' comparison
5674 * in mark_reg_read() is for.
5676 static int propagate_liveness(struct bpf_verifier_env *env,
5677 const struct bpf_verifier_state *vstate,
5678 struct bpf_verifier_state *vparent)
5680 int i, frame, err = 0;
5681 struct bpf_func_state *state, *parent;
5683 if (vparent->curframe != vstate->curframe) {
5684 WARN(1, "propagate_live: parent frame %d current frame %d\n",
5685 vparent->curframe, vstate->curframe);
5688 /* Propagate read liveness of registers... */
5689 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
5690 /* We don't need to worry about FP liveness because it's read-only */
5691 for (i = 0; i < BPF_REG_FP; i++) {
5692 if (vparent->frame[vparent->curframe]->regs[i].live & REG_LIVE_READ)
5694 if (vstate->frame[vstate->curframe]->regs[i].live & REG_LIVE_READ) {
5695 err = mark_reg_read(env, &vstate->frame[vstate->curframe]->regs[i],
5696 &vparent->frame[vstate->curframe]->regs[i]);
5702 /* ... and stack slots */
5703 for (frame = 0; frame <= vstate->curframe; frame++) {
5704 state = vstate->frame[frame];
5705 parent = vparent->frame[frame];
5706 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
5707 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
5708 if (parent->stack[i].spilled_ptr.live & REG_LIVE_READ)
5710 if (state->stack[i].spilled_ptr.live & REG_LIVE_READ)
5711 mark_reg_read(env, &state->stack[i].spilled_ptr,
5712 &parent->stack[i].spilled_ptr);
5718 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
5720 struct bpf_verifier_state_list *new_sl;
5721 struct bpf_verifier_state_list *sl;
5722 struct bpf_verifier_state *cur = env->cur_state, *new;
5723 int i, j, err, states_cnt = 0;
5725 sl = env->explored_states[insn_idx];
5727 /* this 'insn_idx' instruction wasn't marked, so we will not
5728 * be doing state search here
5732 clean_live_states(env, insn_idx, cur);
5734 while (sl != STATE_LIST_MARK) {
5735 if (states_equal(env, &sl->state, cur)) {
5736 /* reached equivalent register/stack state,
5738 * Registers read by the continuation are read by us.
5739 * If we have any write marks in env->cur_state, they
5740 * will prevent corresponding reads in the continuation
5741 * from reaching our parent (an explored_state). Our
5742 * own state will get the read marks recorded, but
5743 * they'll be immediately forgotten as we're pruning
5744 * this state and will pop a new one.
5746 err = propagate_liveness(env, &sl->state, cur);
5755 if (!env->allow_ptr_leaks && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
5758 /* there were no equivalent states, remember current one.
5759 * technically the current state is not proven to be safe yet,
5760 * but it will either reach outer most bpf_exit (which means it's safe)
5761 * or it will be rejected. Since there are no loops, we won't be
5762 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
5763 * again on the way to bpf_exit
5765 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
5769 /* add new state to the head of linked list */
5770 new = &new_sl->state;
5771 err = copy_verifier_state(new, cur);
5773 free_verifier_state(new, false);
5777 new_sl->next = env->explored_states[insn_idx];
5778 env->explored_states[insn_idx] = new_sl;
5779 /* connect new state to parentage chain. Current frame needs all
5780 * registers connected. Only r6 - r9 of the callers are alive (pushed
5781 * to the stack implicitly by JITs) so in callers' frames connect just
5782 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
5783 * the state of the call instruction (with WRITTEN set), and r0 comes
5784 * from callee with its full parentage chain, anyway.
5786 for (j = 0; j <= cur->curframe; j++)
5787 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
5788 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
5789 /* clear write marks in current state: the writes we did are not writes
5790 * our child did, so they don't screen off its reads from us.
5791 * (There are no read marks in current state, because reads always mark
5792 * their parent and current state never has children yet. Only
5793 * explored_states can get read marks.)
5795 for (i = 0; i < BPF_REG_FP; i++)
5796 cur->frame[cur->curframe]->regs[i].live = REG_LIVE_NONE;
5798 /* all stack frames are accessible from callee, clear them all */
5799 for (j = 0; j <= cur->curframe; j++) {
5800 struct bpf_func_state *frame = cur->frame[j];
5801 struct bpf_func_state *newframe = new->frame[j];
5803 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
5804 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
5805 frame->stack[i].spilled_ptr.parent =
5806 &newframe->stack[i].spilled_ptr;
5812 /* Return true if it's OK to have the same insn return a different type. */
5813 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
5818 case PTR_TO_SOCKET_OR_NULL:
5825 /* If an instruction was previously used with particular pointer types, then we
5826 * need to be careful to avoid cases such as the below, where it may be ok
5827 * for one branch accessing the pointer, but not ok for the other branch:
5832 * R1 = some_other_valid_ptr;
5835 * R2 = *(u32 *)(R1 + 0);
5837 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
5839 return src != prev && (!reg_type_mismatch_ok(src) ||
5840 !reg_type_mismatch_ok(prev));
5843 static int do_check(struct bpf_verifier_env *env)
5845 struct bpf_verifier_state *state;
5846 struct bpf_insn *insns = env->prog->insnsi;
5847 struct bpf_reg_state *regs;
5848 int insn_cnt = env->prog->len, i;
5849 int insn_processed = 0;
5850 bool do_print_state = false;
5852 env->prev_linfo = NULL;
5854 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
5857 state->curframe = 0;
5858 state->speculative = false;
5859 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
5860 if (!state->frame[0]) {
5864 env->cur_state = state;
5865 init_func_state(env, state->frame[0],
5866 BPF_MAIN_FUNC /* callsite */,
5868 0 /* subprogno, zero == main subprog */);
5871 struct bpf_insn *insn;
5875 if (env->insn_idx >= insn_cnt) {
5876 verbose(env, "invalid insn idx %d insn_cnt %d\n",
5877 env->insn_idx, insn_cnt);
5881 insn = &insns[env->insn_idx];
5882 class = BPF_CLASS(insn->code);
5884 if (++insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
5886 "BPF program is too large. Processed %d insn\n",
5891 err = is_state_visited(env, env->insn_idx);
5895 /* found equivalent state, can prune the search */
5896 if (env->log.level) {
5898 verbose(env, "\nfrom %d to %d%s: safe\n",
5899 env->prev_insn_idx, env->insn_idx,
5900 env->cur_state->speculative ?
5901 " (speculative execution)" : "");
5903 verbose(env, "%d: safe\n", env->insn_idx);
5905 goto process_bpf_exit;
5908 if (signal_pending(current))
5914 if (env->log.level > 1 || (env->log.level && do_print_state)) {
5915 if (env->log.level > 1)
5916 verbose(env, "%d:", env->insn_idx);
5918 verbose(env, "\nfrom %d to %d%s:",
5919 env->prev_insn_idx, env->insn_idx,
5920 env->cur_state->speculative ?
5921 " (speculative execution)" : "");
5922 print_verifier_state(env, state->frame[state->curframe]);
5923 do_print_state = false;
5926 if (env->log.level) {
5927 const struct bpf_insn_cbs cbs = {
5928 .cb_print = verbose,
5929 .private_data = env,
5932 verbose_linfo(env, env->insn_idx, "; ");
5933 verbose(env, "%d: ", env->insn_idx);
5934 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
5937 if (bpf_prog_is_dev_bound(env->prog->aux)) {
5938 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
5939 env->prev_insn_idx);
5944 regs = cur_regs(env);
5945 env->insn_aux_data[env->insn_idx].seen = true;
5947 if (class == BPF_ALU || class == BPF_ALU64) {
5948 err = check_alu_op(env, insn);
5952 } else if (class == BPF_LDX) {
5953 enum bpf_reg_type *prev_src_type, src_reg_type;
5955 /* check for reserved fields is already done */
5957 /* check src operand */
5958 err = check_reg_arg(env, insn->src_reg, SRC_OP);
5962 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
5966 src_reg_type = regs[insn->src_reg].type;
5968 /* check that memory (src_reg + off) is readable,
5969 * the state of dst_reg will be updated by this func
5971 err = check_mem_access(env, env->insn_idx, insn->src_reg,
5972 insn->off, BPF_SIZE(insn->code),
5973 BPF_READ, insn->dst_reg, false);
5977 prev_src_type = &env->insn_aux_data[env->insn_idx].ptr_type;
5979 if (*prev_src_type == NOT_INIT) {
5981 * dst_reg = *(u32 *)(src_reg + off)
5982 * save type to validate intersecting paths
5984 *prev_src_type = src_reg_type;
5986 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
5987 /* ABuser program is trying to use the same insn
5988 * dst_reg = *(u32*) (src_reg + off)
5989 * with different pointer types:
5990 * src_reg == ctx in one branch and
5991 * src_reg == stack|map in some other branch.
5994 verbose(env, "same insn cannot be used with different pointers\n");
5998 } else if (class == BPF_STX) {
5999 enum bpf_reg_type *prev_dst_type, dst_reg_type;
6001 if (BPF_MODE(insn->code) == BPF_XADD) {
6002 err = check_xadd(env, env->insn_idx, insn);
6009 /* check src1 operand */
6010 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6013 /* check src2 operand */
6014 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
6018 dst_reg_type = regs[insn->dst_reg].type;
6020 /* check that memory (dst_reg + off) is writeable */
6021 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
6022 insn->off, BPF_SIZE(insn->code),
6023 BPF_WRITE, insn->src_reg, false);
6027 prev_dst_type = &env->insn_aux_data[env->insn_idx].ptr_type;
6029 if (*prev_dst_type == NOT_INIT) {
6030 *prev_dst_type = dst_reg_type;
6031 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
6032 verbose(env, "same insn cannot be used with different pointers\n");
6036 } else if (class == BPF_ST) {
6037 if (BPF_MODE(insn->code) != BPF_MEM ||
6038 insn->src_reg != BPF_REG_0) {
6039 verbose(env, "BPF_ST uses reserved fields\n");
6042 /* check src operand */
6043 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
6047 if (is_ctx_reg(env, insn->dst_reg)) {
6048 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
6050 reg_type_str[reg_state(env, insn->dst_reg)->type]);
6054 /* check that memory (dst_reg + off) is writeable */
6055 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
6056 insn->off, BPF_SIZE(insn->code),
6057 BPF_WRITE, -1, false);
6061 } else if (class == BPF_JMP) {
6062 u8 opcode = BPF_OP(insn->code);
6064 if (opcode == BPF_CALL) {
6065 if (BPF_SRC(insn->code) != BPF_K ||
6067 (insn->src_reg != BPF_REG_0 &&
6068 insn->src_reg != BPF_PSEUDO_CALL) ||
6069 insn->dst_reg != BPF_REG_0) {
6070 verbose(env, "BPF_CALL uses reserved fields\n");
6074 if (insn->src_reg == BPF_PSEUDO_CALL)
6075 err = check_func_call(env, insn, &env->insn_idx);
6077 err = check_helper_call(env, insn->imm, env->insn_idx);
6081 } else if (opcode == BPF_JA) {
6082 if (BPF_SRC(insn->code) != BPF_K ||
6084 insn->src_reg != BPF_REG_0 ||
6085 insn->dst_reg != BPF_REG_0) {
6086 verbose(env, "BPF_JA uses reserved fields\n");
6090 env->insn_idx += insn->off + 1;
6093 } else if (opcode == BPF_EXIT) {
6094 if (BPF_SRC(insn->code) != BPF_K ||
6096 insn->src_reg != BPF_REG_0 ||
6097 insn->dst_reg != BPF_REG_0) {
6098 verbose(env, "BPF_EXIT uses reserved fields\n");
6102 if (state->curframe) {
6103 /* exit from nested function */
6104 env->prev_insn_idx = env->insn_idx;
6105 err = prepare_func_exit(env, &env->insn_idx);
6108 do_print_state = true;
6112 err = check_reference_leak(env);
6116 /* eBPF calling convetion is such that R0 is used
6117 * to return the value from eBPF program.
6118 * Make sure that it's readable at this time
6119 * of bpf_exit, which means that program wrote
6120 * something into it earlier
6122 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
6126 if (is_pointer_value(env, BPF_REG_0)) {
6127 verbose(env, "R0 leaks addr as return value\n");
6131 err = check_return_code(env);
6135 err = pop_stack(env, &env->prev_insn_idx,
6142 do_print_state = true;
6146 err = check_cond_jmp_op(env, insn, &env->insn_idx);
6150 } else if (class == BPF_LD) {
6151 u8 mode = BPF_MODE(insn->code);
6153 if (mode == BPF_ABS || mode == BPF_IND) {
6154 err = check_ld_abs(env, insn);
6158 } else if (mode == BPF_IMM) {
6159 err = check_ld_imm(env, insn);
6164 env->insn_aux_data[env->insn_idx].seen = true;
6166 verbose(env, "invalid BPF_LD mode\n");
6170 verbose(env, "unknown insn class %d\n", class);
6177 verbose(env, "processed %d insns (limit %d), stack depth ",
6178 insn_processed, BPF_COMPLEXITY_LIMIT_INSNS);
6179 for (i = 0; i < env->subprog_cnt; i++) {
6180 u32 depth = env->subprog_info[i].stack_depth;
6182 verbose(env, "%d", depth);
6183 if (i + 1 < env->subprog_cnt)
6187 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
6191 static int check_map_prealloc(struct bpf_map *map)
6193 return (map->map_type != BPF_MAP_TYPE_HASH &&
6194 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
6195 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
6196 !(map->map_flags & BPF_F_NO_PREALLOC);
6199 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
6200 struct bpf_map *map,
6201 struct bpf_prog *prog)
6204 /* Make sure that BPF_PROG_TYPE_PERF_EVENT programs only use
6205 * preallocated hash maps, since doing memory allocation
6206 * in overflow_handler can crash depending on where nmi got
6209 if (prog->type == BPF_PROG_TYPE_PERF_EVENT) {
6210 if (!check_map_prealloc(map)) {
6211 verbose(env, "perf_event programs can only use preallocated hash map\n");
6214 if (map->inner_map_meta &&
6215 !check_map_prealloc(map->inner_map_meta)) {
6216 verbose(env, "perf_event programs can only use preallocated inner hash map\n");
6221 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
6222 !bpf_offload_prog_map_match(prog, map)) {
6223 verbose(env, "offload device mismatch between prog and map\n");
6230 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
6232 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
6233 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
6236 /* look for pseudo eBPF instructions that access map FDs and
6237 * replace them with actual map pointers
6239 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env *env)
6241 struct bpf_insn *insn = env->prog->insnsi;
6242 int insn_cnt = env->prog->len;
6245 err = bpf_prog_calc_tag(env->prog);
6249 for (i = 0; i < insn_cnt; i++, insn++) {
6250 if (BPF_CLASS(insn->code) == BPF_LDX &&
6251 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
6252 verbose(env, "BPF_LDX uses reserved fields\n");
6256 if (BPF_CLASS(insn->code) == BPF_STX &&
6257 ((BPF_MODE(insn->code) != BPF_MEM &&
6258 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
6259 verbose(env, "BPF_STX uses reserved fields\n");
6263 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
6264 struct bpf_map *map;
6267 if (i == insn_cnt - 1 || insn[1].code != 0 ||
6268 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
6270 verbose(env, "invalid bpf_ld_imm64 insn\n");
6274 if (insn->src_reg == 0)
6275 /* valid generic load 64-bit imm */
6278 if (insn->src_reg != BPF_PSEUDO_MAP_FD) {
6280 "unrecognized bpf_ld_imm64 insn\n");
6284 f = fdget(insn->imm);
6285 map = __bpf_map_get(f);
6287 verbose(env, "fd %d is not pointing to valid bpf_map\n",
6289 return PTR_ERR(map);
6292 err = check_map_prog_compatibility(env, map, env->prog);
6298 /* store map pointer inside BPF_LD_IMM64 instruction */
6299 insn[0].imm = (u32) (unsigned long) map;
6300 insn[1].imm = ((u64) (unsigned long) map) >> 32;
6302 /* check whether we recorded this map already */
6303 for (j = 0; j < env->used_map_cnt; j++)
6304 if (env->used_maps[j] == map) {
6309 if (env->used_map_cnt >= MAX_USED_MAPS) {
6314 /* hold the map. If the program is rejected by verifier,
6315 * the map will be released by release_maps() or it
6316 * will be used by the valid program until it's unloaded
6317 * and all maps are released in free_used_maps()
6319 map = bpf_map_inc(map, false);
6322 return PTR_ERR(map);
6324 env->used_maps[env->used_map_cnt++] = map;
6326 if (bpf_map_is_cgroup_storage(map) &&
6327 bpf_cgroup_storage_assign(env->prog, map)) {
6328 verbose(env, "only one cgroup storage of each type is allowed\n");
6340 /* Basic sanity check before we invest more work here. */
6341 if (!bpf_opcode_in_insntable(insn->code)) {
6342 verbose(env, "unknown opcode %02x\n", insn->code);
6347 /* now all pseudo BPF_LD_IMM64 instructions load valid
6348 * 'struct bpf_map *' into a register instead of user map_fd.
6349 * These pointers will be used later by verifier to validate map access.
6354 /* drop refcnt of maps used by the rejected program */
6355 static void release_maps(struct bpf_verifier_env *env)
6357 enum bpf_cgroup_storage_type stype;
6360 for_each_cgroup_storage_type(stype) {
6361 if (!env->prog->aux->cgroup_storage[stype])
6363 bpf_cgroup_storage_release(env->prog,
6364 env->prog->aux->cgroup_storage[stype]);
6367 for (i = 0; i < env->used_map_cnt; i++)
6368 bpf_map_put(env->used_maps[i]);
6371 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
6372 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
6374 struct bpf_insn *insn = env->prog->insnsi;
6375 int insn_cnt = env->prog->len;
6378 for (i = 0; i < insn_cnt; i++, insn++)
6379 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
6383 /* single env->prog->insni[off] instruction was replaced with the range
6384 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
6385 * [0, off) and [off, end) to new locations, so the patched range stays zero
6387 static int adjust_insn_aux_data(struct bpf_verifier_env *env, u32 prog_len,
6390 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
6395 new_data = vzalloc(array_size(prog_len,
6396 sizeof(struct bpf_insn_aux_data)));
6399 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
6400 memcpy(new_data + off + cnt - 1, old_data + off,
6401 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
6402 for (i = off; i < off + cnt - 1; i++)
6403 new_data[i].seen = true;
6404 env->insn_aux_data = new_data;
6409 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
6415 /* NOTE: fake 'exit' subprog should be updated as well. */
6416 for (i = 0; i <= env->subprog_cnt; i++) {
6417 if (env->subprog_info[i].start <= off)
6419 env->subprog_info[i].start += len - 1;
6423 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
6424 const struct bpf_insn *patch, u32 len)
6426 struct bpf_prog *new_prog;
6428 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
6431 if (adjust_insn_aux_data(env, new_prog->len, off, len))
6433 adjust_subprog_starts(env, off, len);
6437 /* The verifier does more data flow analysis than llvm and will not
6438 * explore branches that are dead at run time. Malicious programs can
6439 * have dead code too. Therefore replace all dead at-run-time code
6442 * Just nops are not optimal, e.g. if they would sit at the end of the
6443 * program and through another bug we would manage to jump there, then
6444 * we'd execute beyond program memory otherwise. Returning exception
6445 * code also wouldn't work since we can have subprogs where the dead
6446 * code could be located.
6448 static void sanitize_dead_code(struct bpf_verifier_env *env)
6450 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
6451 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
6452 struct bpf_insn *insn = env->prog->insnsi;
6453 const int insn_cnt = env->prog->len;
6456 for (i = 0; i < insn_cnt; i++) {
6457 if (aux_data[i].seen)
6459 memcpy(insn + i, &trap, sizeof(trap));
6463 /* convert load instructions that access fields of a context type into a
6464 * sequence of instructions that access fields of the underlying structure:
6465 * struct __sk_buff -> struct sk_buff
6466 * struct bpf_sock_ops -> struct sock
6468 static int convert_ctx_accesses(struct bpf_verifier_env *env)
6470 const struct bpf_verifier_ops *ops = env->ops;
6471 int i, cnt, size, ctx_field_size, delta = 0;
6472 const int insn_cnt = env->prog->len;
6473 struct bpf_insn insn_buf[16], *insn;
6474 u32 target_size, size_default, off;
6475 struct bpf_prog *new_prog;
6476 enum bpf_access_type type;
6477 bool is_narrower_load;
6479 if (ops->gen_prologue || env->seen_direct_write) {
6480 if (!ops->gen_prologue) {
6481 verbose(env, "bpf verifier is misconfigured\n");
6484 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
6486 if (cnt >= ARRAY_SIZE(insn_buf)) {
6487 verbose(env, "bpf verifier is misconfigured\n");
6490 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
6494 env->prog = new_prog;
6499 if (bpf_prog_is_dev_bound(env->prog->aux))
6502 insn = env->prog->insnsi + delta;
6504 for (i = 0; i < insn_cnt; i++, insn++) {
6505 bpf_convert_ctx_access_t convert_ctx_access;
6507 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
6508 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
6509 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
6510 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
6512 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
6513 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
6514 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
6515 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
6520 if (type == BPF_WRITE &&
6521 env->insn_aux_data[i + delta].sanitize_stack_off) {
6522 struct bpf_insn patch[] = {
6523 /* Sanitize suspicious stack slot with zero.
6524 * There are no memory dependencies for this store,
6525 * since it's only using frame pointer and immediate
6528 BPF_ST_MEM(BPF_DW, BPF_REG_FP,
6529 env->insn_aux_data[i + delta].sanitize_stack_off,
6531 /* the original STX instruction will immediately
6532 * overwrite the same stack slot with appropriate value
6537 cnt = ARRAY_SIZE(patch);
6538 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
6543 env->prog = new_prog;
6544 insn = new_prog->insnsi + i + delta;
6548 switch (env->insn_aux_data[i + delta].ptr_type) {
6550 if (!ops->convert_ctx_access)
6552 convert_ctx_access = ops->convert_ctx_access;
6555 convert_ctx_access = bpf_sock_convert_ctx_access;
6561 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
6562 size = BPF_LDST_BYTES(insn);
6564 /* If the read access is a narrower load of the field,
6565 * convert to a 4/8-byte load, to minimum program type specific
6566 * convert_ctx_access changes. If conversion is successful,
6567 * we will apply proper mask to the result.
6569 is_narrower_load = size < ctx_field_size;
6570 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
6572 if (is_narrower_load) {
6575 if (type == BPF_WRITE) {
6576 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
6581 if (ctx_field_size == 4)
6583 else if (ctx_field_size == 8)
6586 insn->off = off & ~(size_default - 1);
6587 insn->code = BPF_LDX | BPF_MEM | size_code;
6591 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
6593 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
6594 (ctx_field_size && !target_size)) {
6595 verbose(env, "bpf verifier is misconfigured\n");
6599 if (is_narrower_load && size < target_size) {
6600 u8 shift = (off & (size_default - 1)) * 8;
6602 if (ctx_field_size <= 4) {
6604 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
6607 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
6608 (1 << size * 8) - 1);
6611 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
6614 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
6615 (1 << size * 8) - 1);
6619 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
6625 /* keep walking new program and skip insns we just inserted */
6626 env->prog = new_prog;
6627 insn = new_prog->insnsi + i + delta;
6633 static int jit_subprogs(struct bpf_verifier_env *env)
6635 struct bpf_prog *prog = env->prog, **func, *tmp;
6636 int i, j, subprog_start, subprog_end = 0, len, subprog;
6637 struct bpf_insn *insn;
6641 if (env->subprog_cnt <= 1)
6644 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
6645 if (insn->code != (BPF_JMP | BPF_CALL) ||
6646 insn->src_reg != BPF_PSEUDO_CALL)
6648 /* Upon error here we cannot fall back to interpreter but
6649 * need a hard reject of the program. Thus -EFAULT is
6650 * propagated in any case.
6652 subprog = find_subprog(env, i + insn->imm + 1);
6654 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
6658 /* temporarily remember subprog id inside insn instead of
6659 * aux_data, since next loop will split up all insns into funcs
6661 insn->off = subprog;
6662 /* remember original imm in case JIT fails and fallback
6663 * to interpreter will be needed
6665 env->insn_aux_data[i].call_imm = insn->imm;
6666 /* point imm to __bpf_call_base+1 from JITs point of view */
6670 err = bpf_prog_alloc_jited_linfo(prog);
6675 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
6679 for (i = 0; i < env->subprog_cnt; i++) {
6680 subprog_start = subprog_end;
6681 subprog_end = env->subprog_info[i + 1].start;
6683 len = subprog_end - subprog_start;
6684 func[i] = bpf_prog_alloc(bpf_prog_size(len), GFP_USER);
6687 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
6688 len * sizeof(struct bpf_insn));
6689 func[i]->type = prog->type;
6691 if (bpf_prog_calc_tag(func[i]))
6693 func[i]->is_func = 1;
6694 func[i]->aux->func_idx = i;
6695 /* the btf and func_info will be freed only at prog->aux */
6696 func[i]->aux->btf = prog->aux->btf;
6697 func[i]->aux->func_info = prog->aux->func_info;
6699 /* Use bpf_prog_F_tag to indicate functions in stack traces.
6700 * Long term would need debug info to populate names
6702 func[i]->aux->name[0] = 'F';
6703 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
6704 func[i]->jit_requested = 1;
6705 func[i]->aux->linfo = prog->aux->linfo;
6706 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
6707 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
6708 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
6709 func[i] = bpf_int_jit_compile(func[i]);
6710 if (!func[i]->jited) {
6716 /* at this point all bpf functions were successfully JITed
6717 * now populate all bpf_calls with correct addresses and
6718 * run last pass of JIT
6720 for (i = 0; i < env->subprog_cnt; i++) {
6721 insn = func[i]->insnsi;
6722 for (j = 0; j < func[i]->len; j++, insn++) {
6723 if (insn->code != (BPF_JMP | BPF_CALL) ||
6724 insn->src_reg != BPF_PSEUDO_CALL)
6726 subprog = insn->off;
6727 insn->imm = (u64 (*)(u64, u64, u64, u64, u64))
6728 func[subprog]->bpf_func -
6732 /* we use the aux data to keep a list of the start addresses
6733 * of the JITed images for each function in the program
6735 * for some architectures, such as powerpc64, the imm field
6736 * might not be large enough to hold the offset of the start
6737 * address of the callee's JITed image from __bpf_call_base
6739 * in such cases, we can lookup the start address of a callee
6740 * by using its subprog id, available from the off field of
6741 * the call instruction, as an index for this list
6743 func[i]->aux->func = func;
6744 func[i]->aux->func_cnt = env->subprog_cnt;
6746 for (i = 0; i < env->subprog_cnt; i++) {
6747 old_bpf_func = func[i]->bpf_func;
6748 tmp = bpf_int_jit_compile(func[i]);
6749 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
6750 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
6757 /* finally lock prog and jit images for all functions and
6760 for (i = 0; i < env->subprog_cnt; i++) {
6761 bpf_prog_lock_ro(func[i]);
6762 bpf_prog_kallsyms_add(func[i]);
6765 /* Last step: make now unused interpreter insns from main
6766 * prog consistent for later dump requests, so they can
6767 * later look the same as if they were interpreted only.
6769 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
6770 if (insn->code != (BPF_JMP | BPF_CALL) ||
6771 insn->src_reg != BPF_PSEUDO_CALL)
6773 insn->off = env->insn_aux_data[i].call_imm;
6774 subprog = find_subprog(env, i + insn->off + 1);
6775 insn->imm = subprog;
6779 prog->bpf_func = func[0]->bpf_func;
6780 prog->aux->func = func;
6781 prog->aux->func_cnt = env->subprog_cnt;
6782 bpf_prog_free_unused_jited_linfo(prog);
6785 for (i = 0; i < env->subprog_cnt; i++)
6787 bpf_jit_free(func[i]);
6790 /* cleanup main prog to be interpreted */
6791 prog->jit_requested = 0;
6792 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
6793 if (insn->code != (BPF_JMP | BPF_CALL) ||
6794 insn->src_reg != BPF_PSEUDO_CALL)
6797 insn->imm = env->insn_aux_data[i].call_imm;
6799 bpf_prog_free_jited_linfo(prog);
6803 static int fixup_call_args(struct bpf_verifier_env *env)
6805 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
6806 struct bpf_prog *prog = env->prog;
6807 struct bpf_insn *insn = prog->insnsi;
6812 if (env->prog->jit_requested &&
6813 !bpf_prog_is_dev_bound(env->prog->aux)) {
6814 err = jit_subprogs(env);
6820 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
6821 for (i = 0; i < prog->len; i++, insn++) {
6822 if (insn->code != (BPF_JMP | BPF_CALL) ||
6823 insn->src_reg != BPF_PSEUDO_CALL)
6825 depth = get_callee_stack_depth(env, insn, i);
6828 bpf_patch_call_args(insn, depth);
6835 /* fixup insn->imm field of bpf_call instructions
6836 * and inline eligible helpers as explicit sequence of BPF instructions
6838 * this function is called after eBPF program passed verification
6840 static int fixup_bpf_calls(struct bpf_verifier_env *env)
6842 struct bpf_prog *prog = env->prog;
6843 struct bpf_insn *insn = prog->insnsi;
6844 const struct bpf_func_proto *fn;
6845 const int insn_cnt = prog->len;
6846 const struct bpf_map_ops *ops;
6847 struct bpf_insn_aux_data *aux;
6848 struct bpf_insn insn_buf[16];
6849 struct bpf_prog *new_prog;
6850 struct bpf_map *map_ptr;
6851 int i, cnt, delta = 0;
6853 for (i = 0; i < insn_cnt; i++, insn++) {
6854 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
6855 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
6856 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
6857 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
6858 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
6859 struct bpf_insn mask_and_div[] = {
6860 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
6862 BPF_JMP_IMM(BPF_JNE, insn->src_reg, 0, 2),
6863 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
6864 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
6867 struct bpf_insn mask_and_mod[] = {
6868 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
6869 /* Rx mod 0 -> Rx */
6870 BPF_JMP_IMM(BPF_JEQ, insn->src_reg, 0, 1),
6873 struct bpf_insn *patchlet;
6875 if (insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
6876 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
6877 patchlet = mask_and_div + (is64 ? 1 : 0);
6878 cnt = ARRAY_SIZE(mask_and_div) - (is64 ? 1 : 0);
6880 patchlet = mask_and_mod + (is64 ? 1 : 0);
6881 cnt = ARRAY_SIZE(mask_and_mod) - (is64 ? 1 : 0);
6884 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
6889 env->prog = prog = new_prog;
6890 insn = new_prog->insnsi + i + delta;
6894 if (BPF_CLASS(insn->code) == BPF_LD &&
6895 (BPF_MODE(insn->code) == BPF_ABS ||
6896 BPF_MODE(insn->code) == BPF_IND)) {
6897 cnt = env->ops->gen_ld_abs(insn, insn_buf);
6898 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
6899 verbose(env, "bpf verifier is misconfigured\n");
6903 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
6908 env->prog = prog = new_prog;
6909 insn = new_prog->insnsi + i + delta;
6913 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
6914 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
6915 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
6916 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
6917 struct bpf_insn insn_buf[16];
6918 struct bpf_insn *patch = &insn_buf[0];
6922 aux = &env->insn_aux_data[i + delta];
6923 if (!aux->alu_state)
6926 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
6927 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
6928 BPF_ALU_SANITIZE_SRC;
6930 off_reg = issrc ? insn->src_reg : insn->dst_reg;
6932 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
6933 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit - 1);
6934 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
6935 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
6936 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
6937 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
6939 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX,
6941 insn->src_reg = BPF_REG_AX;
6943 *patch++ = BPF_ALU64_REG(BPF_AND, off_reg,
6947 insn->code = insn->code == code_add ?
6948 code_sub : code_add;
6951 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
6952 cnt = patch - insn_buf;
6954 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
6959 env->prog = prog = new_prog;
6960 insn = new_prog->insnsi + i + delta;
6964 if (insn->code != (BPF_JMP | BPF_CALL))
6966 if (insn->src_reg == BPF_PSEUDO_CALL)
6969 if (insn->imm == BPF_FUNC_get_route_realm)
6970 prog->dst_needed = 1;
6971 if (insn->imm == BPF_FUNC_get_prandom_u32)
6972 bpf_user_rnd_init_once();
6973 if (insn->imm == BPF_FUNC_override_return)
6974 prog->kprobe_override = 1;
6975 if (insn->imm == BPF_FUNC_tail_call) {
6976 /* If we tail call into other programs, we
6977 * cannot make any assumptions since they can
6978 * be replaced dynamically during runtime in
6979 * the program array.
6981 prog->cb_access = 1;
6982 env->prog->aux->stack_depth = MAX_BPF_STACK;
6983 env->prog->aux->max_pkt_offset = MAX_PACKET_OFF;
6985 /* mark bpf_tail_call as different opcode to avoid
6986 * conditional branch in the interpeter for every normal
6987 * call and to prevent accidental JITing by JIT compiler
6988 * that doesn't support bpf_tail_call yet
6991 insn->code = BPF_JMP | BPF_TAIL_CALL;
6993 aux = &env->insn_aux_data[i + delta];
6994 if (!bpf_map_ptr_unpriv(aux))
6997 /* instead of changing every JIT dealing with tail_call
6998 * emit two extra insns:
6999 * if (index >= max_entries) goto out;
7000 * index &= array->index_mask;
7001 * to avoid out-of-bounds cpu speculation
7003 if (bpf_map_ptr_poisoned(aux)) {
7004 verbose(env, "tail_call abusing map_ptr\n");
7008 map_ptr = BPF_MAP_PTR(aux->map_state);
7009 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
7010 map_ptr->max_entries, 2);
7011 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
7012 container_of(map_ptr,
7015 insn_buf[2] = *insn;
7017 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
7022 env->prog = prog = new_prog;
7023 insn = new_prog->insnsi + i + delta;
7027 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
7028 * and other inlining handlers are currently limited to 64 bit
7031 if (prog->jit_requested && BITS_PER_LONG == 64 &&
7032 (insn->imm == BPF_FUNC_map_lookup_elem ||
7033 insn->imm == BPF_FUNC_map_update_elem ||
7034 insn->imm == BPF_FUNC_map_delete_elem ||
7035 insn->imm == BPF_FUNC_map_push_elem ||
7036 insn->imm == BPF_FUNC_map_pop_elem ||
7037 insn->imm == BPF_FUNC_map_peek_elem)) {
7038 aux = &env->insn_aux_data[i + delta];
7039 if (bpf_map_ptr_poisoned(aux))
7040 goto patch_call_imm;
7042 map_ptr = BPF_MAP_PTR(aux->map_state);
7044 if (insn->imm == BPF_FUNC_map_lookup_elem &&
7045 ops->map_gen_lookup) {
7046 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
7047 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
7048 verbose(env, "bpf verifier is misconfigured\n");
7052 new_prog = bpf_patch_insn_data(env, i + delta,
7058 env->prog = prog = new_prog;
7059 insn = new_prog->insnsi + i + delta;
7063 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
7064 (void *(*)(struct bpf_map *map, void *key))NULL));
7065 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
7066 (int (*)(struct bpf_map *map, void *key))NULL));
7067 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
7068 (int (*)(struct bpf_map *map, void *key, void *value,
7070 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
7071 (int (*)(struct bpf_map *map, void *value,
7073 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
7074 (int (*)(struct bpf_map *map, void *value))NULL));
7075 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
7076 (int (*)(struct bpf_map *map, void *value))NULL));
7078 switch (insn->imm) {
7079 case BPF_FUNC_map_lookup_elem:
7080 insn->imm = BPF_CAST_CALL(ops->map_lookup_elem) -
7083 case BPF_FUNC_map_update_elem:
7084 insn->imm = BPF_CAST_CALL(ops->map_update_elem) -
7087 case BPF_FUNC_map_delete_elem:
7088 insn->imm = BPF_CAST_CALL(ops->map_delete_elem) -
7091 case BPF_FUNC_map_push_elem:
7092 insn->imm = BPF_CAST_CALL(ops->map_push_elem) -
7095 case BPF_FUNC_map_pop_elem:
7096 insn->imm = BPF_CAST_CALL(ops->map_pop_elem) -
7099 case BPF_FUNC_map_peek_elem:
7100 insn->imm = BPF_CAST_CALL(ops->map_peek_elem) -
7105 goto patch_call_imm;
7109 fn = env->ops->get_func_proto(insn->imm, env->prog);
7110 /* all functions that have prototype and verifier allowed
7111 * programs to call them, must be real in-kernel functions
7115 "kernel subsystem misconfigured func %s#%d\n",
7116 func_id_name(insn->imm), insn->imm);
7119 insn->imm = fn->func - __bpf_call_base;
7125 static void free_states(struct bpf_verifier_env *env)
7127 struct bpf_verifier_state_list *sl, *sln;
7130 if (!env->explored_states)
7133 for (i = 0; i < env->prog->len; i++) {
7134 sl = env->explored_states[i];
7137 while (sl != STATE_LIST_MARK) {
7139 free_verifier_state(&sl->state, false);
7145 kfree(env->explored_states);
7148 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr,
7149 union bpf_attr __user *uattr)
7151 struct bpf_verifier_env *env;
7152 struct bpf_verifier_log *log;
7155 /* no program is valid */
7156 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
7159 /* 'struct bpf_verifier_env' can be global, but since it's not small,
7160 * allocate/free it every time bpf_check() is called
7162 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
7167 env->insn_aux_data =
7168 vzalloc(array_size(sizeof(struct bpf_insn_aux_data),
7171 if (!env->insn_aux_data)
7174 env->ops = bpf_verifier_ops[env->prog->type];
7176 /* grab the mutex to protect few globals used by verifier */
7177 mutex_lock(&bpf_verifier_lock);
7179 if (attr->log_level || attr->log_buf || attr->log_size) {
7180 /* user requested verbose verifier output
7181 * and supplied buffer to store the verification trace
7183 log->level = attr->log_level;
7184 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
7185 log->len_total = attr->log_size;
7188 /* log attributes have to be sane */
7189 if (log->len_total < 128 || log->len_total > UINT_MAX >> 8 ||
7190 !log->level || !log->ubuf)
7194 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
7195 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
7196 env->strict_alignment = true;
7197 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
7198 env->strict_alignment = false;
7200 ret = replace_map_fd_with_map_ptr(env);
7202 goto skip_full_check;
7204 if (bpf_prog_is_dev_bound(env->prog->aux)) {
7205 ret = bpf_prog_offload_verifier_prep(env->prog);
7207 goto skip_full_check;
7210 env->explored_states = kcalloc(env->prog->len,
7211 sizeof(struct bpf_verifier_state_list *),
7214 if (!env->explored_states)
7215 goto skip_full_check;
7217 env->allow_ptr_leaks = capable(CAP_SYS_ADMIN);
7219 ret = check_subprogs(env);
7221 goto skip_full_check;
7223 ret = check_btf_info(env, attr, uattr);
7225 goto skip_full_check;
7227 ret = check_cfg(env);
7229 goto skip_full_check;
7231 ret = do_check(env);
7232 if (env->cur_state) {
7233 free_verifier_state(env->cur_state, true);
7234 env->cur_state = NULL;
7237 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
7238 ret = bpf_prog_offload_finalize(env);
7241 while (!pop_stack(env, NULL, NULL));
7245 ret = check_max_stack_depth(env);
7247 /* instruction rewrites happen after this point */
7249 sanitize_dead_code(env);
7252 /* program is valid, convert *(u32*)(ctx + off) accesses */
7253 ret = convert_ctx_accesses(env);
7256 ret = fixup_bpf_calls(env);
7259 ret = fixup_call_args(env);
7261 if (log->level && bpf_verifier_log_full(log))
7263 if (log->level && !log->ubuf) {
7265 goto err_release_maps;
7268 if (ret == 0 && env->used_map_cnt) {
7269 /* if program passed verifier, update used_maps in bpf_prog_info */
7270 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
7271 sizeof(env->used_maps[0]),
7274 if (!env->prog->aux->used_maps) {
7276 goto err_release_maps;
7279 memcpy(env->prog->aux->used_maps, env->used_maps,
7280 sizeof(env->used_maps[0]) * env->used_map_cnt);
7281 env->prog->aux->used_map_cnt = env->used_map_cnt;
7283 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
7284 * bpf_ld_imm64 instructions
7286 convert_pseudo_ld_imm64(env);
7290 adjust_btf_func(env);
7293 if (!env->prog->aux->used_maps)
7294 /* if we didn't copy map pointers into bpf_prog_info, release
7295 * them now. Otherwise free_used_maps() will release them.
7300 mutex_unlock(&bpf_verifier_lock);
7301 vfree(env->insn_aux_data);