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;
219 static DEFINE_MUTEX(bpf_verifier_lock);
221 static const struct bpf_line_info *
222 find_linfo(const struct bpf_verifier_env *env, u32 insn_off)
224 const struct bpf_line_info *linfo;
225 const struct bpf_prog *prog;
229 nr_linfo = prog->aux->nr_linfo;
231 if (!nr_linfo || insn_off >= prog->len)
234 linfo = prog->aux->linfo;
235 for (i = 1; i < nr_linfo; i++)
236 if (insn_off < linfo[i].insn_off)
239 return &linfo[i - 1];
242 void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
247 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
249 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
250 "verifier log line truncated - local buffer too short\n");
252 n = min(log->len_total - log->len_used - 1, n);
255 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
261 /* log_level controls verbosity level of eBPF verifier.
262 * bpf_verifier_log_write() is used to dump the verification trace to the log,
263 * so the user can figure out what's wrong with the program
265 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
266 const char *fmt, ...)
270 if (!bpf_verifier_log_needed(&env->log))
274 bpf_verifier_vlog(&env->log, fmt, args);
277 EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
279 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
281 struct bpf_verifier_env *env = private_data;
284 if (!bpf_verifier_log_needed(&env->log))
288 bpf_verifier_vlog(&env->log, fmt, args);
292 static const char *ltrim(const char *s)
300 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env,
302 const char *prefix_fmt, ...)
304 const struct bpf_line_info *linfo;
306 if (!bpf_verifier_log_needed(&env->log))
309 linfo = find_linfo(env, insn_off);
310 if (!linfo || linfo == env->prev_linfo)
316 va_start(args, prefix_fmt);
317 bpf_verifier_vlog(&env->log, prefix_fmt, args);
322 ltrim(btf_name_by_offset(env->prog->aux->btf,
325 env->prev_linfo = linfo;
328 static bool type_is_pkt_pointer(enum bpf_reg_type type)
330 return type == PTR_TO_PACKET ||
331 type == PTR_TO_PACKET_META;
334 static bool type_is_sk_pointer(enum bpf_reg_type type)
336 return type == PTR_TO_SOCKET ||
337 type == PTR_TO_SOCK_COMMON ||
338 type == PTR_TO_TCP_SOCK;
341 static bool reg_type_may_be_null(enum bpf_reg_type type)
343 return type == PTR_TO_MAP_VALUE_OR_NULL ||
344 type == PTR_TO_SOCKET_OR_NULL ||
345 type == PTR_TO_SOCK_COMMON_OR_NULL ||
346 type == PTR_TO_TCP_SOCK_OR_NULL;
349 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
351 return reg->type == PTR_TO_MAP_VALUE &&
352 map_value_has_spin_lock(reg->map_ptr);
355 static bool arg_type_may_be_refcounted(enum bpf_arg_type type)
357 return type == ARG_PTR_TO_SOCK_COMMON;
360 /* Determine whether the function releases some resources allocated by another
361 * function call. The first reference type argument will be assumed to be
362 * released by release_reference().
364 static bool is_release_function(enum bpf_func_id func_id)
366 return func_id == BPF_FUNC_sk_release;
369 static bool is_acquire_function(enum bpf_func_id func_id)
371 return func_id == BPF_FUNC_sk_lookup_tcp ||
372 func_id == BPF_FUNC_sk_lookup_udp ||
373 func_id == BPF_FUNC_skc_lookup_tcp;
376 static bool is_ptr_cast_function(enum bpf_func_id func_id)
378 return func_id == BPF_FUNC_tcp_sock ||
379 func_id == BPF_FUNC_sk_fullsock;
382 /* string representation of 'enum bpf_reg_type' */
383 static const char * const reg_type_str[] = {
385 [SCALAR_VALUE] = "inv",
386 [PTR_TO_CTX] = "ctx",
387 [CONST_PTR_TO_MAP] = "map_ptr",
388 [PTR_TO_MAP_VALUE] = "map_value",
389 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
390 [PTR_TO_STACK] = "fp",
391 [PTR_TO_PACKET] = "pkt",
392 [PTR_TO_PACKET_META] = "pkt_meta",
393 [PTR_TO_PACKET_END] = "pkt_end",
394 [PTR_TO_FLOW_KEYS] = "flow_keys",
395 [PTR_TO_SOCKET] = "sock",
396 [PTR_TO_SOCKET_OR_NULL] = "sock_or_null",
397 [PTR_TO_SOCK_COMMON] = "sock_common",
398 [PTR_TO_SOCK_COMMON_OR_NULL] = "sock_common_or_null",
399 [PTR_TO_TCP_SOCK] = "tcp_sock",
400 [PTR_TO_TCP_SOCK_OR_NULL] = "tcp_sock_or_null",
403 static char slot_type_char[] = {
404 [STACK_INVALID] = '?',
410 static void print_liveness(struct bpf_verifier_env *env,
411 enum bpf_reg_liveness live)
413 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
415 if (live & REG_LIVE_READ)
417 if (live & REG_LIVE_WRITTEN)
419 if (live & REG_LIVE_DONE)
423 static struct bpf_func_state *func(struct bpf_verifier_env *env,
424 const struct bpf_reg_state *reg)
426 struct bpf_verifier_state *cur = env->cur_state;
428 return cur->frame[reg->frameno];
431 static void print_verifier_state(struct bpf_verifier_env *env,
432 const struct bpf_func_state *state)
434 const struct bpf_reg_state *reg;
439 verbose(env, " frame%d:", state->frameno);
440 for (i = 0; i < MAX_BPF_REG; i++) {
441 reg = &state->regs[i];
445 verbose(env, " R%d", i);
446 print_liveness(env, reg->live);
447 verbose(env, "=%s", reg_type_str[t]);
448 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
449 tnum_is_const(reg->var_off)) {
450 /* reg->off should be 0 for SCALAR_VALUE */
451 verbose(env, "%lld", reg->var_off.value + reg->off);
452 if (t == PTR_TO_STACK)
453 verbose(env, ",call_%d", func(env, reg)->callsite);
455 verbose(env, "(id=%d ref_obj_id=%d", reg->id,
457 if (t != SCALAR_VALUE)
458 verbose(env, ",off=%d", reg->off);
459 if (type_is_pkt_pointer(t))
460 verbose(env, ",r=%d", reg->range);
461 else if (t == CONST_PTR_TO_MAP ||
462 t == PTR_TO_MAP_VALUE ||
463 t == PTR_TO_MAP_VALUE_OR_NULL)
464 verbose(env, ",ks=%d,vs=%d",
465 reg->map_ptr->key_size,
466 reg->map_ptr->value_size);
467 if (tnum_is_const(reg->var_off)) {
468 /* Typically an immediate SCALAR_VALUE, but
469 * could be a pointer whose offset is too big
472 verbose(env, ",imm=%llx", reg->var_off.value);
474 if (reg->smin_value != reg->umin_value &&
475 reg->smin_value != S64_MIN)
476 verbose(env, ",smin_value=%lld",
477 (long long)reg->smin_value);
478 if (reg->smax_value != reg->umax_value &&
479 reg->smax_value != S64_MAX)
480 verbose(env, ",smax_value=%lld",
481 (long long)reg->smax_value);
482 if (reg->umin_value != 0)
483 verbose(env, ",umin_value=%llu",
484 (unsigned long long)reg->umin_value);
485 if (reg->umax_value != U64_MAX)
486 verbose(env, ",umax_value=%llu",
487 (unsigned long long)reg->umax_value);
488 if (!tnum_is_unknown(reg->var_off)) {
491 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
492 verbose(env, ",var_off=%s", tn_buf);
498 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
499 char types_buf[BPF_REG_SIZE + 1];
503 for (j = 0; j < BPF_REG_SIZE; j++) {
504 if (state->stack[i].slot_type[j] != STACK_INVALID)
506 types_buf[j] = slot_type_char[
507 state->stack[i].slot_type[j]];
509 types_buf[BPF_REG_SIZE] = 0;
512 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
513 print_liveness(env, state->stack[i].spilled_ptr.live);
514 if (state->stack[i].slot_type[0] == STACK_SPILL)
516 reg_type_str[state->stack[i].spilled_ptr.type]);
518 verbose(env, "=%s", types_buf);
520 if (state->acquired_refs && state->refs[0].id) {
521 verbose(env, " refs=%d", state->refs[0].id);
522 for (i = 1; i < state->acquired_refs; i++)
523 if (state->refs[i].id)
524 verbose(env, ",%d", state->refs[i].id);
529 #define COPY_STATE_FN(NAME, COUNT, FIELD, SIZE) \
530 static int copy_##NAME##_state(struct bpf_func_state *dst, \
531 const struct bpf_func_state *src) \
535 if (WARN_ON_ONCE(dst->COUNT < src->COUNT)) { \
536 /* internal bug, make state invalid to reject the program */ \
537 memset(dst, 0, sizeof(*dst)); \
540 memcpy(dst->FIELD, src->FIELD, \
541 sizeof(*src->FIELD) * (src->COUNT / SIZE)); \
544 /* copy_reference_state() */
545 COPY_STATE_FN(reference, acquired_refs, refs, 1)
546 /* copy_stack_state() */
547 COPY_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
550 #define REALLOC_STATE_FN(NAME, COUNT, FIELD, SIZE) \
551 static int realloc_##NAME##_state(struct bpf_func_state *state, int size, \
554 u32 old_size = state->COUNT; \
555 struct bpf_##NAME##_state *new_##FIELD; \
556 int slot = size / SIZE; \
558 if (size <= old_size || !size) { \
561 state->COUNT = slot * SIZE; \
562 if (!size && old_size) { \
563 kfree(state->FIELD); \
564 state->FIELD = NULL; \
568 new_##FIELD = kmalloc_array(slot, sizeof(struct bpf_##NAME##_state), \
574 memcpy(new_##FIELD, state->FIELD, \
575 sizeof(*new_##FIELD) * (old_size / SIZE)); \
576 memset(new_##FIELD + old_size / SIZE, 0, \
577 sizeof(*new_##FIELD) * (size - old_size) / SIZE); \
579 state->COUNT = slot * SIZE; \
580 kfree(state->FIELD); \
581 state->FIELD = new_##FIELD; \
584 /* realloc_reference_state() */
585 REALLOC_STATE_FN(reference, acquired_refs, refs, 1)
586 /* realloc_stack_state() */
587 REALLOC_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
588 #undef REALLOC_STATE_FN
590 /* do_check() starts with zero-sized stack in struct bpf_verifier_state to
591 * make it consume minimal amount of memory. check_stack_write() access from
592 * the program calls into realloc_func_state() to grow the stack size.
593 * Note there is a non-zero 'parent' pointer inside bpf_verifier_state
594 * which realloc_stack_state() copies over. It points to previous
595 * bpf_verifier_state which is never reallocated.
597 static int realloc_func_state(struct bpf_func_state *state, int stack_size,
598 int refs_size, bool copy_old)
600 int err = realloc_reference_state(state, refs_size, copy_old);
603 return realloc_stack_state(state, stack_size, copy_old);
606 /* Acquire a pointer id from the env and update the state->refs to include
607 * this new pointer reference.
608 * On success, returns a valid pointer id to associate with the register
609 * On failure, returns a negative errno.
611 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
613 struct bpf_func_state *state = cur_func(env);
614 int new_ofs = state->acquired_refs;
617 err = realloc_reference_state(state, state->acquired_refs + 1, true);
621 state->refs[new_ofs].id = id;
622 state->refs[new_ofs].insn_idx = insn_idx;
627 /* release function corresponding to acquire_reference_state(). Idempotent. */
628 static int release_reference_state(struct bpf_func_state *state, int ptr_id)
632 last_idx = state->acquired_refs - 1;
633 for (i = 0; i < state->acquired_refs; i++) {
634 if (state->refs[i].id == ptr_id) {
635 if (last_idx && i != last_idx)
636 memcpy(&state->refs[i], &state->refs[last_idx],
637 sizeof(*state->refs));
638 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
639 state->acquired_refs--;
646 static int transfer_reference_state(struct bpf_func_state *dst,
647 struct bpf_func_state *src)
649 int err = realloc_reference_state(dst, src->acquired_refs, false);
652 err = copy_reference_state(dst, src);
658 static void free_func_state(struct bpf_func_state *state)
667 static void free_verifier_state(struct bpf_verifier_state *state,
672 for (i = 0; i <= state->curframe; i++) {
673 free_func_state(state->frame[i]);
674 state->frame[i] = NULL;
680 /* copy verifier state from src to dst growing dst stack space
681 * when necessary to accommodate larger src stack
683 static int copy_func_state(struct bpf_func_state *dst,
684 const struct bpf_func_state *src)
688 err = realloc_func_state(dst, src->allocated_stack, src->acquired_refs,
692 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
693 err = copy_reference_state(dst, src);
696 return copy_stack_state(dst, src);
699 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
700 const struct bpf_verifier_state *src)
702 struct bpf_func_state *dst;
705 /* if dst has more stack frames then src frame, free them */
706 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
707 free_func_state(dst_state->frame[i]);
708 dst_state->frame[i] = NULL;
710 dst_state->speculative = src->speculative;
711 dst_state->curframe = src->curframe;
712 dst_state->active_spin_lock = src->active_spin_lock;
713 for (i = 0; i <= src->curframe; i++) {
714 dst = dst_state->frame[i];
716 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
719 dst_state->frame[i] = dst;
721 err = copy_func_state(dst, src->frame[i]);
728 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
731 struct bpf_verifier_state *cur = env->cur_state;
732 struct bpf_verifier_stack_elem *elem, *head = env->head;
735 if (env->head == NULL)
739 err = copy_verifier_state(cur, &head->st);
744 *insn_idx = head->insn_idx;
746 *prev_insn_idx = head->prev_insn_idx;
748 free_verifier_state(&head->st, false);
755 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
756 int insn_idx, int prev_insn_idx,
759 struct bpf_verifier_state *cur = env->cur_state;
760 struct bpf_verifier_stack_elem *elem;
763 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
767 elem->insn_idx = insn_idx;
768 elem->prev_insn_idx = prev_insn_idx;
769 elem->next = env->head;
772 err = copy_verifier_state(&elem->st, cur);
775 elem->st.speculative |= speculative;
776 if (env->stack_size > BPF_COMPLEXITY_LIMIT_STACK) {
777 verbose(env, "BPF program is too complex\n");
782 free_verifier_state(env->cur_state, true);
783 env->cur_state = NULL;
784 /* pop all elements and return */
785 while (!pop_stack(env, NULL, NULL));
789 #define CALLER_SAVED_REGS 6
790 static const int caller_saved[CALLER_SAVED_REGS] = {
791 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
794 static void __mark_reg_not_init(struct bpf_reg_state *reg);
796 /* Mark the unknown part of a register (variable offset or scalar value) as
797 * known to have the value @imm.
799 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
801 /* Clear id, off, and union(map_ptr, range) */
802 memset(((u8 *)reg) + sizeof(reg->type), 0,
803 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
804 reg->var_off = tnum_const(imm);
805 reg->smin_value = (s64)imm;
806 reg->smax_value = (s64)imm;
807 reg->umin_value = imm;
808 reg->umax_value = imm;
811 /* Mark the 'variable offset' part of a register as zero. This should be
812 * used only on registers holding a pointer type.
814 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
816 __mark_reg_known(reg, 0);
819 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
821 __mark_reg_known(reg, 0);
822 reg->type = SCALAR_VALUE;
825 static void mark_reg_known_zero(struct bpf_verifier_env *env,
826 struct bpf_reg_state *regs, u32 regno)
828 if (WARN_ON(regno >= MAX_BPF_REG)) {
829 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
830 /* Something bad happened, let's kill all regs */
831 for (regno = 0; regno < MAX_BPF_REG; regno++)
832 __mark_reg_not_init(regs + regno);
835 __mark_reg_known_zero(regs + regno);
838 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
840 return type_is_pkt_pointer(reg->type);
843 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
845 return reg_is_pkt_pointer(reg) ||
846 reg->type == PTR_TO_PACKET_END;
849 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
850 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
851 enum bpf_reg_type which)
853 /* The register can already have a range from prior markings.
854 * This is fine as long as it hasn't been advanced from its
857 return reg->type == which &&
860 tnum_equals_const(reg->var_off, 0);
863 /* Attempts to improve min/max values based on var_off information */
864 static void __update_reg_bounds(struct bpf_reg_state *reg)
866 /* min signed is max(sign bit) | min(other bits) */
867 reg->smin_value = max_t(s64, reg->smin_value,
868 reg->var_off.value | (reg->var_off.mask & S64_MIN));
869 /* max signed is min(sign bit) | max(other bits) */
870 reg->smax_value = min_t(s64, reg->smax_value,
871 reg->var_off.value | (reg->var_off.mask & S64_MAX));
872 reg->umin_value = max(reg->umin_value, reg->var_off.value);
873 reg->umax_value = min(reg->umax_value,
874 reg->var_off.value | reg->var_off.mask);
877 /* Uses signed min/max values to inform unsigned, and vice-versa */
878 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
880 /* Learn sign from signed bounds.
881 * If we cannot cross the sign boundary, then signed and unsigned bounds
882 * are the same, so combine. This works even in the negative case, e.g.
883 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
885 if (reg->smin_value >= 0 || reg->smax_value < 0) {
886 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
888 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
892 /* Learn sign from unsigned bounds. Signed bounds cross the sign
893 * boundary, so we must be careful.
895 if ((s64)reg->umax_value >= 0) {
896 /* Positive. We can't learn anything from the smin, but smax
897 * is positive, hence safe.
899 reg->smin_value = reg->umin_value;
900 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
902 } else if ((s64)reg->umin_value < 0) {
903 /* Negative. We can't learn anything from the smax, but smin
904 * is negative, hence safe.
906 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
908 reg->smax_value = reg->umax_value;
912 /* Attempts to improve var_off based on unsigned min/max information */
913 static void __reg_bound_offset(struct bpf_reg_state *reg)
915 reg->var_off = tnum_intersect(reg->var_off,
916 tnum_range(reg->umin_value,
920 /* Reset the min/max bounds of a register */
921 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
923 reg->smin_value = S64_MIN;
924 reg->smax_value = S64_MAX;
926 reg->umax_value = U64_MAX;
929 /* Mark a register as having a completely unknown (scalar) value. */
930 static void __mark_reg_unknown(struct bpf_reg_state *reg)
933 * Clear type, id, off, and union(map_ptr, range) and
934 * padding between 'type' and union
936 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
937 reg->type = SCALAR_VALUE;
938 reg->var_off = tnum_unknown;
940 __mark_reg_unbounded(reg);
943 static void mark_reg_unknown(struct bpf_verifier_env *env,
944 struct bpf_reg_state *regs, u32 regno)
946 if (WARN_ON(regno >= MAX_BPF_REG)) {
947 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
948 /* Something bad happened, let's kill all regs except FP */
949 for (regno = 0; regno < BPF_REG_FP; regno++)
950 __mark_reg_not_init(regs + regno);
953 __mark_reg_unknown(regs + regno);
956 static void __mark_reg_not_init(struct bpf_reg_state *reg)
958 __mark_reg_unknown(reg);
959 reg->type = NOT_INIT;
962 static void mark_reg_not_init(struct bpf_verifier_env *env,
963 struct bpf_reg_state *regs, u32 regno)
965 if (WARN_ON(regno >= MAX_BPF_REG)) {
966 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
967 /* Something bad happened, let's kill all regs except FP */
968 for (regno = 0; regno < BPF_REG_FP; regno++)
969 __mark_reg_not_init(regs + regno);
972 __mark_reg_not_init(regs + regno);
975 static void init_reg_state(struct bpf_verifier_env *env,
976 struct bpf_func_state *state)
978 struct bpf_reg_state *regs = state->regs;
981 for (i = 0; i < MAX_BPF_REG; i++) {
982 mark_reg_not_init(env, regs, i);
983 regs[i].live = REG_LIVE_NONE;
984 regs[i].parent = NULL;
988 regs[BPF_REG_FP].type = PTR_TO_STACK;
989 mark_reg_known_zero(env, regs, BPF_REG_FP);
990 regs[BPF_REG_FP].frameno = state->frameno;
992 /* 1st arg to a function */
993 regs[BPF_REG_1].type = PTR_TO_CTX;
994 mark_reg_known_zero(env, regs, BPF_REG_1);
997 #define BPF_MAIN_FUNC (-1)
998 static void init_func_state(struct bpf_verifier_env *env,
999 struct bpf_func_state *state,
1000 int callsite, int frameno, int subprogno)
1002 state->callsite = callsite;
1003 state->frameno = frameno;
1004 state->subprogno = subprogno;
1005 init_reg_state(env, state);
1009 SRC_OP, /* register is used as source operand */
1010 DST_OP, /* register is used as destination operand */
1011 DST_OP_NO_MARK /* same as above, check only, don't mark */
1014 static int cmp_subprogs(const void *a, const void *b)
1016 return ((struct bpf_subprog_info *)a)->start -
1017 ((struct bpf_subprog_info *)b)->start;
1020 static int find_subprog(struct bpf_verifier_env *env, int off)
1022 struct bpf_subprog_info *p;
1024 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
1025 sizeof(env->subprog_info[0]), cmp_subprogs);
1028 return p - env->subprog_info;
1032 static int add_subprog(struct bpf_verifier_env *env, int off)
1034 int insn_cnt = env->prog->len;
1037 if (off >= insn_cnt || off < 0) {
1038 verbose(env, "call to invalid destination\n");
1041 ret = find_subprog(env, off);
1044 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
1045 verbose(env, "too many subprograms\n");
1048 env->subprog_info[env->subprog_cnt++].start = off;
1049 sort(env->subprog_info, env->subprog_cnt,
1050 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
1054 static int check_subprogs(struct bpf_verifier_env *env)
1056 int i, ret, subprog_start, subprog_end, off, cur_subprog = 0;
1057 struct bpf_subprog_info *subprog = env->subprog_info;
1058 struct bpf_insn *insn = env->prog->insnsi;
1059 int insn_cnt = env->prog->len;
1061 /* Add entry function. */
1062 ret = add_subprog(env, 0);
1066 /* determine subprog starts. The end is one before the next starts */
1067 for (i = 0; i < insn_cnt; i++) {
1068 if (insn[i].code != (BPF_JMP | BPF_CALL))
1070 if (insn[i].src_reg != BPF_PSEUDO_CALL)
1072 if (!env->allow_ptr_leaks) {
1073 verbose(env, "function calls to other bpf functions are allowed for root only\n");
1076 ret = add_subprog(env, i + insn[i].imm + 1);
1081 /* Add a fake 'exit' subprog which could simplify subprog iteration
1082 * logic. 'subprog_cnt' should not be increased.
1084 subprog[env->subprog_cnt].start = insn_cnt;
1086 if (env->log.level > 1)
1087 for (i = 0; i < env->subprog_cnt; i++)
1088 verbose(env, "func#%d @%d\n", i, subprog[i].start);
1090 /* now check that all jumps are within the same subprog */
1091 subprog_start = subprog[cur_subprog].start;
1092 subprog_end = subprog[cur_subprog + 1].start;
1093 for (i = 0; i < insn_cnt; i++) {
1094 u8 code = insn[i].code;
1096 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
1098 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
1100 off = i + insn[i].off + 1;
1101 if (off < subprog_start || off >= subprog_end) {
1102 verbose(env, "jump out of range from insn %d to %d\n", i, off);
1106 if (i == subprog_end - 1) {
1107 /* to avoid fall-through from one subprog into another
1108 * the last insn of the subprog should be either exit
1109 * or unconditional jump back
1111 if (code != (BPF_JMP | BPF_EXIT) &&
1112 code != (BPF_JMP | BPF_JA)) {
1113 verbose(env, "last insn is not an exit or jmp\n");
1116 subprog_start = subprog_end;
1118 if (cur_subprog < env->subprog_cnt)
1119 subprog_end = subprog[cur_subprog + 1].start;
1125 /* Parentage chain of this register (or stack slot) should take care of all
1126 * issues like callee-saved registers, stack slot allocation time, etc.
1128 static int mark_reg_read(struct bpf_verifier_env *env,
1129 const struct bpf_reg_state *state,
1130 struct bpf_reg_state *parent)
1132 bool writes = parent == state->parent; /* Observe write marks */
1135 /* if read wasn't screened by an earlier write ... */
1136 if (writes && state->live & REG_LIVE_WRITTEN)
1138 if (parent->live & REG_LIVE_DONE) {
1139 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
1140 reg_type_str[parent->type],
1141 parent->var_off.value, parent->off);
1144 /* ... then we depend on parent's value */
1145 parent->live |= REG_LIVE_READ;
1147 parent = state->parent;
1153 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
1154 enum reg_arg_type t)
1156 struct bpf_verifier_state *vstate = env->cur_state;
1157 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1158 struct bpf_reg_state *regs = state->regs;
1160 if (regno >= MAX_BPF_REG) {
1161 verbose(env, "R%d is invalid\n", regno);
1166 /* check whether register used as source operand can be read */
1167 if (regs[regno].type == NOT_INIT) {
1168 verbose(env, "R%d !read_ok\n", regno);
1171 /* We don't need to worry about FP liveness because it's read-only */
1172 if (regno != BPF_REG_FP)
1173 return mark_reg_read(env, ®s[regno],
1174 regs[regno].parent);
1176 /* check whether register used as dest operand can be written to */
1177 if (regno == BPF_REG_FP) {
1178 verbose(env, "frame pointer is read only\n");
1181 regs[regno].live |= REG_LIVE_WRITTEN;
1183 mark_reg_unknown(env, regs, regno);
1188 static bool is_spillable_regtype(enum bpf_reg_type type)
1191 case PTR_TO_MAP_VALUE:
1192 case PTR_TO_MAP_VALUE_OR_NULL:
1196 case PTR_TO_PACKET_META:
1197 case PTR_TO_PACKET_END:
1198 case PTR_TO_FLOW_KEYS:
1199 case CONST_PTR_TO_MAP:
1201 case PTR_TO_SOCKET_OR_NULL:
1202 case PTR_TO_SOCK_COMMON:
1203 case PTR_TO_SOCK_COMMON_OR_NULL:
1204 case PTR_TO_TCP_SOCK:
1205 case PTR_TO_TCP_SOCK_OR_NULL:
1212 /* Does this register contain a constant zero? */
1213 static bool register_is_null(struct bpf_reg_state *reg)
1215 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
1218 /* check_stack_read/write functions track spill/fill of registers,
1219 * stack boundary and alignment are checked in check_mem_access()
1221 static int check_stack_write(struct bpf_verifier_env *env,
1222 struct bpf_func_state *state, /* func where register points to */
1223 int off, int size, int value_regno, int insn_idx)
1225 struct bpf_func_state *cur; /* state of the current function */
1226 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
1227 enum bpf_reg_type type;
1229 err = realloc_func_state(state, round_up(slot + 1, BPF_REG_SIZE),
1230 state->acquired_refs, true);
1233 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
1234 * so it's aligned access and [off, off + size) are within stack limits
1236 if (!env->allow_ptr_leaks &&
1237 state->stack[spi].slot_type[0] == STACK_SPILL &&
1238 size != BPF_REG_SIZE) {
1239 verbose(env, "attempt to corrupt spilled pointer on stack\n");
1243 cur = env->cur_state->frame[env->cur_state->curframe];
1244 if (value_regno >= 0 &&
1245 is_spillable_regtype((type = cur->regs[value_regno].type))) {
1247 /* register containing pointer is being spilled into stack */
1248 if (size != BPF_REG_SIZE) {
1249 verbose(env, "invalid size of register spill\n");
1253 if (state != cur && type == PTR_TO_STACK) {
1254 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
1258 /* save register state */
1259 state->stack[spi].spilled_ptr = cur->regs[value_regno];
1260 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1262 for (i = 0; i < BPF_REG_SIZE; i++) {
1263 if (state->stack[spi].slot_type[i] == STACK_MISC &&
1264 !env->allow_ptr_leaks) {
1265 int *poff = &env->insn_aux_data[insn_idx].sanitize_stack_off;
1266 int soff = (-spi - 1) * BPF_REG_SIZE;
1268 /* detected reuse of integer stack slot with a pointer
1269 * which means either llvm is reusing stack slot or
1270 * an attacker is trying to exploit CVE-2018-3639
1271 * (speculative store bypass)
1272 * Have to sanitize that slot with preemptive
1275 if (*poff && *poff != soff) {
1276 /* disallow programs where single insn stores
1277 * into two different stack slots, since verifier
1278 * cannot sanitize them
1281 "insn %d cannot access two stack slots fp%d and fp%d",
1282 insn_idx, *poff, soff);
1287 state->stack[spi].slot_type[i] = STACK_SPILL;
1290 u8 type = STACK_MISC;
1292 /* regular write of data into stack destroys any spilled ptr */
1293 state->stack[spi].spilled_ptr.type = NOT_INIT;
1294 /* Mark slots as STACK_MISC if they belonged to spilled ptr. */
1295 if (state->stack[spi].slot_type[0] == STACK_SPILL)
1296 for (i = 0; i < BPF_REG_SIZE; i++)
1297 state->stack[spi].slot_type[i] = STACK_MISC;
1299 /* only mark the slot as written if all 8 bytes were written
1300 * otherwise read propagation may incorrectly stop too soon
1301 * when stack slots are partially written.
1302 * This heuristic means that read propagation will be
1303 * conservative, since it will add reg_live_read marks
1304 * to stack slots all the way to first state when programs
1305 * writes+reads less than 8 bytes
1307 if (size == BPF_REG_SIZE)
1308 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1310 /* when we zero initialize stack slots mark them as such */
1311 if (value_regno >= 0 &&
1312 register_is_null(&cur->regs[value_regno]))
1315 /* Mark slots affected by this stack write. */
1316 for (i = 0; i < size; i++)
1317 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
1323 static int check_stack_read(struct bpf_verifier_env *env,
1324 struct bpf_func_state *reg_state /* func where register points to */,
1325 int off, int size, int value_regno)
1327 struct bpf_verifier_state *vstate = env->cur_state;
1328 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1329 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
1332 if (reg_state->allocated_stack <= slot) {
1333 verbose(env, "invalid read from stack off %d+0 size %d\n",
1337 stype = reg_state->stack[spi].slot_type;
1339 if (stype[0] == STACK_SPILL) {
1340 if (size != BPF_REG_SIZE) {
1341 verbose(env, "invalid size of register spill\n");
1344 for (i = 1; i < BPF_REG_SIZE; i++) {
1345 if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) {
1346 verbose(env, "corrupted spill memory\n");
1351 if (value_regno >= 0) {
1352 /* restore register state from stack */
1353 state->regs[value_regno] = reg_state->stack[spi].spilled_ptr;
1354 /* mark reg as written since spilled pointer state likely
1355 * has its liveness marks cleared by is_state_visited()
1356 * which resets stack/reg liveness for state transitions
1358 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
1360 mark_reg_read(env, ®_state->stack[spi].spilled_ptr,
1361 reg_state->stack[spi].spilled_ptr.parent);
1366 for (i = 0; i < size; i++) {
1367 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_MISC)
1369 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_ZERO) {
1373 verbose(env, "invalid read from stack off %d+%d size %d\n",
1377 mark_reg_read(env, ®_state->stack[spi].spilled_ptr,
1378 reg_state->stack[spi].spilled_ptr.parent);
1379 if (value_regno >= 0) {
1380 if (zeros == size) {
1381 /* any size read into register is zero extended,
1382 * so the whole register == const_zero
1384 __mark_reg_const_zero(&state->regs[value_regno]);
1386 /* have read misc data from the stack */
1387 mark_reg_unknown(env, state->regs, value_regno);
1389 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
1395 static int check_stack_access(struct bpf_verifier_env *env,
1396 const struct bpf_reg_state *reg,
1399 /* Stack accesses must be at a fixed offset, so that we
1400 * can determine what type of data were returned. See
1401 * check_stack_read().
1403 if (!tnum_is_const(reg->var_off)) {
1406 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1407 verbose(env, "variable stack access var_off=%s off=%d size=%d",
1412 if (off >= 0 || off < -MAX_BPF_STACK) {
1413 verbose(env, "invalid stack off=%d size=%d\n", off, size);
1420 /* check read/write into map element returned by bpf_map_lookup_elem() */
1421 static int __check_map_access(struct bpf_verifier_env *env, u32 regno, int off,
1422 int size, bool zero_size_allowed)
1424 struct bpf_reg_state *regs = cur_regs(env);
1425 struct bpf_map *map = regs[regno].map_ptr;
1427 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
1428 off + size > map->value_size) {
1429 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
1430 map->value_size, off, size);
1436 /* check read/write into a map element with possible variable offset */
1437 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
1438 int off, int size, bool zero_size_allowed)
1440 struct bpf_verifier_state *vstate = env->cur_state;
1441 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1442 struct bpf_reg_state *reg = &state->regs[regno];
1445 /* We may have adjusted the register to this map value, so we
1446 * need to try adding each of min_value and max_value to off
1447 * to make sure our theoretical access will be safe.
1450 print_verifier_state(env, state);
1452 /* The minimum value is only important with signed
1453 * comparisons where we can't assume the floor of a
1454 * value is 0. If we are using signed variables for our
1455 * index'es we need to make sure that whatever we use
1456 * will have a set floor within our range.
1458 if (reg->smin_value < 0 &&
1459 (reg->smin_value == S64_MIN ||
1460 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
1461 reg->smin_value + off < 0)) {
1462 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1466 err = __check_map_access(env, regno, reg->smin_value + off, size,
1469 verbose(env, "R%d min value is outside of the array range\n",
1474 /* If we haven't set a max value then we need to bail since we can't be
1475 * sure we won't do bad things.
1476 * If reg->umax_value + off could overflow, treat that as unbounded too.
1478 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
1479 verbose(env, "R%d unbounded memory access, make sure to bounds check any array access into a map\n",
1483 err = __check_map_access(env, regno, reg->umax_value + off, size,
1486 verbose(env, "R%d max value is outside of the array range\n",
1489 if (map_value_has_spin_lock(reg->map_ptr)) {
1490 u32 lock = reg->map_ptr->spin_lock_off;
1492 /* if any part of struct bpf_spin_lock can be touched by
1493 * load/store reject this program.
1494 * To check that [x1, x2) overlaps with [y1, y2)
1495 * it is sufficient to check x1 < y2 && y1 < x2.
1497 if (reg->smin_value + off < lock + sizeof(struct bpf_spin_lock) &&
1498 lock < reg->umax_value + off + size) {
1499 verbose(env, "bpf_spin_lock cannot be accessed directly by load/store\n");
1506 #define MAX_PACKET_OFF 0xffff
1508 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
1509 const struct bpf_call_arg_meta *meta,
1510 enum bpf_access_type t)
1512 switch (env->prog->type) {
1513 /* Program types only with direct read access go here! */
1514 case BPF_PROG_TYPE_LWT_IN:
1515 case BPF_PROG_TYPE_LWT_OUT:
1516 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
1517 case BPF_PROG_TYPE_SK_REUSEPORT:
1518 case BPF_PROG_TYPE_FLOW_DISSECTOR:
1519 case BPF_PROG_TYPE_CGROUP_SKB:
1524 /* Program types with direct read + write access go here! */
1525 case BPF_PROG_TYPE_SCHED_CLS:
1526 case BPF_PROG_TYPE_SCHED_ACT:
1527 case BPF_PROG_TYPE_XDP:
1528 case BPF_PROG_TYPE_LWT_XMIT:
1529 case BPF_PROG_TYPE_SK_SKB:
1530 case BPF_PROG_TYPE_SK_MSG:
1532 return meta->pkt_access;
1534 env->seen_direct_write = true;
1541 static int __check_packet_access(struct bpf_verifier_env *env, u32 regno,
1542 int off, int size, bool zero_size_allowed)
1544 struct bpf_reg_state *regs = cur_regs(env);
1545 struct bpf_reg_state *reg = ®s[regno];
1547 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
1548 (u64)off + size > reg->range) {
1549 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
1550 off, size, regno, reg->id, reg->off, reg->range);
1556 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
1557 int size, bool zero_size_allowed)
1559 struct bpf_reg_state *regs = cur_regs(env);
1560 struct bpf_reg_state *reg = ®s[regno];
1563 /* We may have added a variable offset to the packet pointer; but any
1564 * reg->range we have comes after that. We are only checking the fixed
1568 /* We don't allow negative numbers, because we aren't tracking enough
1569 * detail to prove they're safe.
1571 if (reg->smin_value < 0) {
1572 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1576 err = __check_packet_access(env, regno, off, size, zero_size_allowed);
1578 verbose(env, "R%d offset is outside of the packet\n", regno);
1582 /* __check_packet_access has made sure "off + size - 1" is within u16.
1583 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
1584 * otherwise find_good_pkt_pointers would have refused to set range info
1585 * that __check_packet_access would have rejected this pkt access.
1586 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
1588 env->prog->aux->max_pkt_offset =
1589 max_t(u32, env->prog->aux->max_pkt_offset,
1590 off + reg->umax_value + size - 1);
1595 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
1596 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
1597 enum bpf_access_type t, enum bpf_reg_type *reg_type)
1599 struct bpf_insn_access_aux info = {
1600 .reg_type = *reg_type,
1603 if (env->ops->is_valid_access &&
1604 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
1605 /* A non zero info.ctx_field_size indicates that this field is a
1606 * candidate for later verifier transformation to load the whole
1607 * field and then apply a mask when accessed with a narrower
1608 * access than actual ctx access size. A zero info.ctx_field_size
1609 * will only allow for whole field access and rejects any other
1610 * type of narrower access.
1612 *reg_type = info.reg_type;
1614 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
1615 /* remember the offset of last byte accessed in ctx */
1616 if (env->prog->aux->max_ctx_offset < off + size)
1617 env->prog->aux->max_ctx_offset = off + size;
1621 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
1625 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
1628 if (size < 0 || off < 0 ||
1629 (u64)off + size > sizeof(struct bpf_flow_keys)) {
1630 verbose(env, "invalid access to flow keys off=%d size=%d\n",
1637 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
1638 u32 regno, int off, int size,
1639 enum bpf_access_type t)
1641 struct bpf_reg_state *regs = cur_regs(env);
1642 struct bpf_reg_state *reg = ®s[regno];
1643 struct bpf_insn_access_aux info = {};
1646 if (reg->smin_value < 0) {
1647 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1652 switch (reg->type) {
1653 case PTR_TO_SOCK_COMMON:
1654 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
1657 valid = bpf_sock_is_valid_access(off, size, t, &info);
1659 case PTR_TO_TCP_SOCK:
1660 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
1668 env->insn_aux_data[insn_idx].ctx_field_size =
1669 info.ctx_field_size;
1673 verbose(env, "R%d invalid %s access off=%d size=%d\n",
1674 regno, reg_type_str[reg->type], off, size);
1679 static bool __is_pointer_value(bool allow_ptr_leaks,
1680 const struct bpf_reg_state *reg)
1682 if (allow_ptr_leaks)
1685 return reg->type != SCALAR_VALUE;
1688 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
1690 return cur_regs(env) + regno;
1693 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
1695 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
1698 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
1700 const struct bpf_reg_state *reg = reg_state(env, regno);
1702 return reg->type == PTR_TO_CTX;
1705 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
1707 const struct bpf_reg_state *reg = reg_state(env, regno);
1709 return type_is_sk_pointer(reg->type);
1712 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
1714 const struct bpf_reg_state *reg = reg_state(env, regno);
1716 return type_is_pkt_pointer(reg->type);
1719 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
1721 const struct bpf_reg_state *reg = reg_state(env, regno);
1723 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
1724 return reg->type == PTR_TO_FLOW_KEYS;
1727 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
1728 const struct bpf_reg_state *reg,
1729 int off, int size, bool strict)
1731 struct tnum reg_off;
1734 /* Byte size accesses are always allowed. */
1735 if (!strict || size == 1)
1738 /* For platforms that do not have a Kconfig enabling
1739 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
1740 * NET_IP_ALIGN is universally set to '2'. And on platforms
1741 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
1742 * to this code only in strict mode where we want to emulate
1743 * the NET_IP_ALIGN==2 checking. Therefore use an
1744 * unconditional IP align value of '2'.
1748 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
1749 if (!tnum_is_aligned(reg_off, size)) {
1752 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1754 "misaligned packet access off %d+%s+%d+%d size %d\n",
1755 ip_align, tn_buf, reg->off, off, size);
1762 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
1763 const struct bpf_reg_state *reg,
1764 const char *pointer_desc,
1765 int off, int size, bool strict)
1767 struct tnum reg_off;
1769 /* Byte size accesses are always allowed. */
1770 if (!strict || size == 1)
1773 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
1774 if (!tnum_is_aligned(reg_off, size)) {
1777 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1778 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
1779 pointer_desc, tn_buf, reg->off, off, size);
1786 static int check_ptr_alignment(struct bpf_verifier_env *env,
1787 const struct bpf_reg_state *reg, int off,
1788 int size, bool strict_alignment_once)
1790 bool strict = env->strict_alignment || strict_alignment_once;
1791 const char *pointer_desc = "";
1793 switch (reg->type) {
1795 case PTR_TO_PACKET_META:
1796 /* Special case, because of NET_IP_ALIGN. Given metadata sits
1797 * right in front, treat it the very same way.
1799 return check_pkt_ptr_alignment(env, reg, off, size, strict);
1800 case PTR_TO_FLOW_KEYS:
1801 pointer_desc = "flow keys ";
1803 case PTR_TO_MAP_VALUE:
1804 pointer_desc = "value ";
1807 pointer_desc = "context ";
1810 pointer_desc = "stack ";
1811 /* The stack spill tracking logic in check_stack_write()
1812 * and check_stack_read() relies on stack accesses being
1818 pointer_desc = "sock ";
1820 case PTR_TO_SOCK_COMMON:
1821 pointer_desc = "sock_common ";
1823 case PTR_TO_TCP_SOCK:
1824 pointer_desc = "tcp_sock ";
1829 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
1833 static int update_stack_depth(struct bpf_verifier_env *env,
1834 const struct bpf_func_state *func,
1837 u16 stack = env->subprog_info[func->subprogno].stack_depth;
1842 /* update known max for given subprogram */
1843 env->subprog_info[func->subprogno].stack_depth = -off;
1847 /* starting from main bpf function walk all instructions of the function
1848 * and recursively walk all callees that given function can call.
1849 * Ignore jump and exit insns.
1850 * Since recursion is prevented by check_cfg() this algorithm
1851 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
1853 static int check_max_stack_depth(struct bpf_verifier_env *env)
1855 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
1856 struct bpf_subprog_info *subprog = env->subprog_info;
1857 struct bpf_insn *insn = env->prog->insnsi;
1858 int ret_insn[MAX_CALL_FRAMES];
1859 int ret_prog[MAX_CALL_FRAMES];
1862 /* round up to 32-bytes, since this is granularity
1863 * of interpreter stack size
1865 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
1866 if (depth > MAX_BPF_STACK) {
1867 verbose(env, "combined stack size of %d calls is %d. Too large\n",
1872 subprog_end = subprog[idx + 1].start;
1873 for (; i < subprog_end; i++) {
1874 if (insn[i].code != (BPF_JMP | BPF_CALL))
1876 if (insn[i].src_reg != BPF_PSEUDO_CALL)
1878 /* remember insn and function to return to */
1879 ret_insn[frame] = i + 1;
1880 ret_prog[frame] = idx;
1882 /* find the callee */
1883 i = i + insn[i].imm + 1;
1884 idx = find_subprog(env, i);
1886 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
1891 if (frame >= MAX_CALL_FRAMES) {
1892 WARN_ONCE(1, "verifier bug. Call stack is too deep\n");
1897 /* end of for() loop means the last insn of the 'subprog'
1898 * was reached. Doesn't matter whether it was JA or EXIT
1902 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
1904 i = ret_insn[frame];
1905 idx = ret_prog[frame];
1909 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
1910 static int get_callee_stack_depth(struct bpf_verifier_env *env,
1911 const struct bpf_insn *insn, int idx)
1913 int start = idx + insn->imm + 1, subprog;
1915 subprog = find_subprog(env, start);
1917 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
1921 return env->subprog_info[subprog].stack_depth;
1925 static int check_ctx_reg(struct bpf_verifier_env *env,
1926 const struct bpf_reg_state *reg, int regno)
1928 /* Access to ctx or passing it to a helper is only allowed in
1929 * its original, unmodified form.
1933 verbose(env, "dereference of modified ctx ptr R%d off=%d disallowed\n",
1938 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
1941 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1942 verbose(env, "variable ctx access var_off=%s disallowed\n", tn_buf);
1949 /* truncate register to smaller size (in bytes)
1950 * must be called with size < BPF_REG_SIZE
1952 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
1956 /* clear high bits in bit representation */
1957 reg->var_off = tnum_cast(reg->var_off, size);
1959 /* fix arithmetic bounds */
1960 mask = ((u64)1 << (size * 8)) - 1;
1961 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
1962 reg->umin_value &= mask;
1963 reg->umax_value &= mask;
1965 reg->umin_value = 0;
1966 reg->umax_value = mask;
1968 reg->smin_value = reg->umin_value;
1969 reg->smax_value = reg->umax_value;
1972 /* check whether memory at (regno + off) is accessible for t = (read | write)
1973 * if t==write, value_regno is a register which value is stored into memory
1974 * if t==read, value_regno is a register which will receive the value from memory
1975 * if t==write && value_regno==-1, some unknown value is stored into memory
1976 * if t==read && value_regno==-1, don't care what we read from memory
1978 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
1979 int off, int bpf_size, enum bpf_access_type t,
1980 int value_regno, bool strict_alignment_once)
1982 struct bpf_reg_state *regs = cur_regs(env);
1983 struct bpf_reg_state *reg = regs + regno;
1984 struct bpf_func_state *state;
1987 size = bpf_size_to_bytes(bpf_size);
1991 /* alignment checks will add in reg->off themselves */
1992 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
1996 /* for access checks, reg->off is just part of off */
1999 if (reg->type == PTR_TO_MAP_VALUE) {
2000 if (t == BPF_WRITE && value_regno >= 0 &&
2001 is_pointer_value(env, value_regno)) {
2002 verbose(env, "R%d leaks addr into map\n", value_regno);
2006 err = check_map_access(env, regno, off, size, false);
2007 if (!err && t == BPF_READ && value_regno >= 0)
2008 mark_reg_unknown(env, regs, value_regno);
2010 } else if (reg->type == PTR_TO_CTX) {
2011 enum bpf_reg_type reg_type = SCALAR_VALUE;
2013 if (t == BPF_WRITE && value_regno >= 0 &&
2014 is_pointer_value(env, value_regno)) {
2015 verbose(env, "R%d leaks addr into ctx\n", value_regno);
2019 err = check_ctx_reg(env, reg, regno);
2023 err = check_ctx_access(env, insn_idx, off, size, t, ®_type);
2024 if (!err && t == BPF_READ && value_regno >= 0) {
2025 /* ctx access returns either a scalar, or a
2026 * PTR_TO_PACKET[_META,_END]. In the latter
2027 * case, we know the offset is zero.
2029 if (reg_type == SCALAR_VALUE) {
2030 mark_reg_unknown(env, regs, value_regno);
2032 mark_reg_known_zero(env, regs,
2034 if (reg_type_may_be_null(reg_type))
2035 regs[value_regno].id = ++env->id_gen;
2037 regs[value_regno].type = reg_type;
2040 } else if (reg->type == PTR_TO_STACK) {
2041 off += reg->var_off.value;
2042 err = check_stack_access(env, reg, off, size);
2046 state = func(env, reg);
2047 err = update_stack_depth(env, state, off);
2052 err = check_stack_write(env, state, off, size,
2053 value_regno, insn_idx);
2055 err = check_stack_read(env, state, off, size,
2057 } else if (reg_is_pkt_pointer(reg)) {
2058 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
2059 verbose(env, "cannot write into packet\n");
2062 if (t == BPF_WRITE && value_regno >= 0 &&
2063 is_pointer_value(env, value_regno)) {
2064 verbose(env, "R%d leaks addr into packet\n",
2068 err = check_packet_access(env, regno, off, size, false);
2069 if (!err && t == BPF_READ && value_regno >= 0)
2070 mark_reg_unknown(env, regs, value_regno);
2071 } else if (reg->type == PTR_TO_FLOW_KEYS) {
2072 if (t == BPF_WRITE && value_regno >= 0 &&
2073 is_pointer_value(env, value_regno)) {
2074 verbose(env, "R%d leaks addr into flow keys\n",
2079 err = check_flow_keys_access(env, off, size);
2080 if (!err && t == BPF_READ && value_regno >= 0)
2081 mark_reg_unknown(env, regs, value_regno);
2082 } else if (type_is_sk_pointer(reg->type)) {
2083 if (t == BPF_WRITE) {
2084 verbose(env, "R%d cannot write into %s\n",
2085 regno, reg_type_str[reg->type]);
2088 err = check_sock_access(env, insn_idx, regno, off, size, t);
2089 if (!err && value_regno >= 0)
2090 mark_reg_unknown(env, regs, value_regno);
2092 verbose(env, "R%d invalid mem access '%s'\n", regno,
2093 reg_type_str[reg->type]);
2097 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
2098 regs[value_regno].type == SCALAR_VALUE) {
2099 /* b/h/w load zero-extends, mark upper bits as known 0 */
2100 coerce_reg_to_size(®s[value_regno], size);
2105 static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
2109 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
2111 verbose(env, "BPF_XADD uses reserved fields\n");
2115 /* check src1 operand */
2116 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2120 /* check src2 operand */
2121 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2125 if (is_pointer_value(env, insn->src_reg)) {
2126 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
2130 if (is_ctx_reg(env, insn->dst_reg) ||
2131 is_pkt_reg(env, insn->dst_reg) ||
2132 is_flow_key_reg(env, insn->dst_reg) ||
2133 is_sk_reg(env, insn->dst_reg)) {
2134 verbose(env, "BPF_XADD stores into R%d %s is not allowed\n",
2136 reg_type_str[reg_state(env, insn->dst_reg)->type]);
2140 /* check whether atomic_add can read the memory */
2141 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
2142 BPF_SIZE(insn->code), BPF_READ, -1, true);
2146 /* check whether atomic_add can write into the same memory */
2147 return check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
2148 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
2151 /* when register 'regno' is passed into function that will read 'access_size'
2152 * bytes from that pointer, make sure that it's within stack boundary
2153 * and all elements of stack are initialized.
2154 * Unlike most pointer bounds-checking functions, this one doesn't take an
2155 * 'off' argument, so it has to add in reg->off itself.
2157 static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
2158 int access_size, bool zero_size_allowed,
2159 struct bpf_call_arg_meta *meta)
2161 struct bpf_reg_state *reg = reg_state(env, regno);
2162 struct bpf_func_state *state = func(env, reg);
2163 int off, i, slot, spi;
2165 if (reg->type != PTR_TO_STACK) {
2166 /* Allow zero-byte read from NULL, regardless of pointer type */
2167 if (zero_size_allowed && access_size == 0 &&
2168 register_is_null(reg))
2171 verbose(env, "R%d type=%s expected=%s\n", regno,
2172 reg_type_str[reg->type],
2173 reg_type_str[PTR_TO_STACK]);
2177 /* Only allow fixed-offset stack reads */
2178 if (!tnum_is_const(reg->var_off)) {
2181 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2182 verbose(env, "invalid variable stack read R%d var_off=%s\n",
2186 off = reg->off + reg->var_off.value;
2187 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
2188 access_size < 0 || (access_size == 0 && !zero_size_allowed)) {
2189 verbose(env, "invalid stack type R%d off=%d access_size=%d\n",
2190 regno, off, access_size);
2194 if (meta && meta->raw_mode) {
2195 meta->access_size = access_size;
2196 meta->regno = regno;
2200 for (i = 0; i < access_size; i++) {
2203 slot = -(off + i) - 1;
2204 spi = slot / BPF_REG_SIZE;
2205 if (state->allocated_stack <= slot)
2207 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
2208 if (*stype == STACK_MISC)
2210 if (*stype == STACK_ZERO) {
2211 /* helper can write anything into the stack */
2212 *stype = STACK_MISC;
2216 verbose(env, "invalid indirect read from stack off %d+%d size %d\n",
2217 off, i, access_size);
2220 /* reading any byte out of 8-byte 'spill_slot' will cause
2221 * the whole slot to be marked as 'read'
2223 mark_reg_read(env, &state->stack[spi].spilled_ptr,
2224 state->stack[spi].spilled_ptr.parent);
2226 return update_stack_depth(env, state, off);
2229 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
2230 int access_size, bool zero_size_allowed,
2231 struct bpf_call_arg_meta *meta)
2233 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
2235 switch (reg->type) {
2237 case PTR_TO_PACKET_META:
2238 return check_packet_access(env, regno, reg->off, access_size,
2240 case PTR_TO_MAP_VALUE:
2241 return check_map_access(env, regno, reg->off, access_size,
2243 default: /* scalar_value|ptr_to_stack or invalid ptr */
2244 return check_stack_boundary(env, regno, access_size,
2245 zero_size_allowed, meta);
2249 /* Implementation details:
2250 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL
2251 * Two bpf_map_lookups (even with the same key) will have different reg->id.
2252 * For traditional PTR_TO_MAP_VALUE the verifier clears reg->id after
2253 * value_or_null->value transition, since the verifier only cares about
2254 * the range of access to valid map value pointer and doesn't care about actual
2255 * address of the map element.
2256 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
2257 * reg->id > 0 after value_or_null->value transition. By doing so
2258 * two bpf_map_lookups will be considered two different pointers that
2259 * point to different bpf_spin_locks.
2260 * The verifier allows taking only one bpf_spin_lock at a time to avoid
2262 * Since only one bpf_spin_lock is allowed the checks are simpler than
2263 * reg_is_refcounted() logic. The verifier needs to remember only
2264 * one spin_lock instead of array of acquired_refs.
2265 * cur_state->active_spin_lock remembers which map value element got locked
2266 * and clears it after bpf_spin_unlock.
2268 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
2271 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
2272 struct bpf_verifier_state *cur = env->cur_state;
2273 bool is_const = tnum_is_const(reg->var_off);
2274 struct bpf_map *map = reg->map_ptr;
2275 u64 val = reg->var_off.value;
2277 if (reg->type != PTR_TO_MAP_VALUE) {
2278 verbose(env, "R%d is not a pointer to map_value\n", regno);
2283 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
2289 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
2293 if (!map_value_has_spin_lock(map)) {
2294 if (map->spin_lock_off == -E2BIG)
2296 "map '%s' has more than one 'struct bpf_spin_lock'\n",
2298 else if (map->spin_lock_off == -ENOENT)
2300 "map '%s' doesn't have 'struct bpf_spin_lock'\n",
2304 "map '%s' is not a struct type or bpf_spin_lock is mangled\n",
2308 if (map->spin_lock_off != val + reg->off) {
2309 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock'\n",
2314 if (cur->active_spin_lock) {
2316 "Locking two bpf_spin_locks are not allowed\n");
2319 cur->active_spin_lock = reg->id;
2321 if (!cur->active_spin_lock) {
2322 verbose(env, "bpf_spin_unlock without taking a lock\n");
2325 if (cur->active_spin_lock != reg->id) {
2326 verbose(env, "bpf_spin_unlock of different lock\n");
2329 cur->active_spin_lock = 0;
2334 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
2336 return type == ARG_PTR_TO_MEM ||
2337 type == ARG_PTR_TO_MEM_OR_NULL ||
2338 type == ARG_PTR_TO_UNINIT_MEM;
2341 static bool arg_type_is_mem_size(enum bpf_arg_type type)
2343 return type == ARG_CONST_SIZE ||
2344 type == ARG_CONST_SIZE_OR_ZERO;
2347 static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
2348 enum bpf_arg_type arg_type,
2349 struct bpf_call_arg_meta *meta)
2351 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
2352 enum bpf_reg_type expected_type, type = reg->type;
2355 if (arg_type == ARG_DONTCARE)
2358 err = check_reg_arg(env, regno, SRC_OP);
2362 if (arg_type == ARG_ANYTHING) {
2363 if (is_pointer_value(env, regno)) {
2364 verbose(env, "R%d leaks addr into helper function\n",
2371 if (type_is_pkt_pointer(type) &&
2372 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
2373 verbose(env, "helper access to the packet is not allowed\n");
2377 if (arg_type == ARG_PTR_TO_MAP_KEY ||
2378 arg_type == ARG_PTR_TO_MAP_VALUE ||
2379 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
2380 expected_type = PTR_TO_STACK;
2381 if (!type_is_pkt_pointer(type) && type != PTR_TO_MAP_VALUE &&
2382 type != expected_type)
2384 } else if (arg_type == ARG_CONST_SIZE ||
2385 arg_type == ARG_CONST_SIZE_OR_ZERO) {
2386 expected_type = SCALAR_VALUE;
2387 if (type != expected_type)
2389 } else if (arg_type == ARG_CONST_MAP_PTR) {
2390 expected_type = CONST_PTR_TO_MAP;
2391 if (type != expected_type)
2393 } else if (arg_type == ARG_PTR_TO_CTX) {
2394 expected_type = PTR_TO_CTX;
2395 if (type != expected_type)
2397 err = check_ctx_reg(env, reg, regno);
2400 } else if (arg_type == ARG_PTR_TO_SOCK_COMMON) {
2401 expected_type = PTR_TO_SOCK_COMMON;
2402 /* Any sk pointer can be ARG_PTR_TO_SOCK_COMMON */
2403 if (!type_is_sk_pointer(type))
2405 if (reg->ref_obj_id) {
2406 if (meta->ref_obj_id) {
2407 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
2408 regno, reg->ref_obj_id,
2412 meta->ref_obj_id = reg->ref_obj_id;
2414 } else if (arg_type == ARG_PTR_TO_SPIN_LOCK) {
2415 if (meta->func_id == BPF_FUNC_spin_lock) {
2416 if (process_spin_lock(env, regno, true))
2418 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
2419 if (process_spin_lock(env, regno, false))
2422 verbose(env, "verifier internal error\n");
2425 } else if (arg_type_is_mem_ptr(arg_type)) {
2426 expected_type = PTR_TO_STACK;
2427 /* One exception here. In case function allows for NULL to be
2428 * passed in as argument, it's a SCALAR_VALUE type. Final test
2429 * happens during stack boundary checking.
2431 if (register_is_null(reg) &&
2432 arg_type == ARG_PTR_TO_MEM_OR_NULL)
2433 /* final test in check_stack_boundary() */;
2434 else if (!type_is_pkt_pointer(type) &&
2435 type != PTR_TO_MAP_VALUE &&
2436 type != expected_type)
2438 meta->raw_mode = arg_type == ARG_PTR_TO_UNINIT_MEM;
2440 verbose(env, "unsupported arg_type %d\n", arg_type);
2444 if (arg_type == ARG_CONST_MAP_PTR) {
2445 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
2446 meta->map_ptr = reg->map_ptr;
2447 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
2448 /* bpf_map_xxx(..., map_ptr, ..., key) call:
2449 * check that [key, key + map->key_size) are within
2450 * stack limits and initialized
2452 if (!meta->map_ptr) {
2453 /* in function declaration map_ptr must come before
2454 * map_key, so that it's verified and known before
2455 * we have to check map_key here. Otherwise it means
2456 * that kernel subsystem misconfigured verifier
2458 verbose(env, "invalid map_ptr to access map->key\n");
2461 err = check_helper_mem_access(env, regno,
2462 meta->map_ptr->key_size, false,
2464 } else if (arg_type == ARG_PTR_TO_MAP_VALUE ||
2465 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
2466 /* bpf_map_xxx(..., map_ptr, ..., value) call:
2467 * check [value, value + map->value_size) validity
2469 if (!meta->map_ptr) {
2470 /* kernel subsystem misconfigured verifier */
2471 verbose(env, "invalid map_ptr to access map->value\n");
2474 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE);
2475 err = check_helper_mem_access(env, regno,
2476 meta->map_ptr->value_size, false,
2478 } else if (arg_type_is_mem_size(arg_type)) {
2479 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
2481 /* remember the mem_size which may be used later
2482 * to refine return values.
2484 meta->msize_smax_value = reg->smax_value;
2485 meta->msize_umax_value = reg->umax_value;
2487 /* The register is SCALAR_VALUE; the access check
2488 * happens using its boundaries.
2490 if (!tnum_is_const(reg->var_off))
2491 /* For unprivileged variable accesses, disable raw
2492 * mode so that the program is required to
2493 * initialize all the memory that the helper could
2494 * just partially fill up.
2498 if (reg->smin_value < 0) {
2499 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
2504 if (reg->umin_value == 0) {
2505 err = check_helper_mem_access(env, regno - 1, 0,
2512 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
2513 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
2517 err = check_helper_mem_access(env, regno - 1,
2519 zero_size_allowed, meta);
2524 verbose(env, "R%d type=%s expected=%s\n", regno,
2525 reg_type_str[type], reg_type_str[expected_type]);
2529 static int check_map_func_compatibility(struct bpf_verifier_env *env,
2530 struct bpf_map *map, int func_id)
2535 /* We need a two way check, first is from map perspective ... */
2536 switch (map->map_type) {
2537 case BPF_MAP_TYPE_PROG_ARRAY:
2538 if (func_id != BPF_FUNC_tail_call)
2541 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
2542 if (func_id != BPF_FUNC_perf_event_read &&
2543 func_id != BPF_FUNC_perf_event_output &&
2544 func_id != BPF_FUNC_perf_event_read_value)
2547 case BPF_MAP_TYPE_STACK_TRACE:
2548 if (func_id != BPF_FUNC_get_stackid)
2551 case BPF_MAP_TYPE_CGROUP_ARRAY:
2552 if (func_id != BPF_FUNC_skb_under_cgroup &&
2553 func_id != BPF_FUNC_current_task_under_cgroup)
2556 case BPF_MAP_TYPE_CGROUP_STORAGE:
2557 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
2558 if (func_id != BPF_FUNC_get_local_storage)
2561 /* devmap returns a pointer to a live net_device ifindex that we cannot
2562 * allow to be modified from bpf side. So do not allow lookup elements
2565 case BPF_MAP_TYPE_DEVMAP:
2566 if (func_id != BPF_FUNC_redirect_map)
2569 /* Restrict bpf side of cpumap and xskmap, open when use-cases
2572 case BPF_MAP_TYPE_CPUMAP:
2573 case BPF_MAP_TYPE_XSKMAP:
2574 if (func_id != BPF_FUNC_redirect_map)
2577 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
2578 case BPF_MAP_TYPE_HASH_OF_MAPS:
2579 if (func_id != BPF_FUNC_map_lookup_elem)
2582 case BPF_MAP_TYPE_SOCKMAP:
2583 if (func_id != BPF_FUNC_sk_redirect_map &&
2584 func_id != BPF_FUNC_sock_map_update &&
2585 func_id != BPF_FUNC_map_delete_elem &&
2586 func_id != BPF_FUNC_msg_redirect_map)
2589 case BPF_MAP_TYPE_SOCKHASH:
2590 if (func_id != BPF_FUNC_sk_redirect_hash &&
2591 func_id != BPF_FUNC_sock_hash_update &&
2592 func_id != BPF_FUNC_map_delete_elem &&
2593 func_id != BPF_FUNC_msg_redirect_hash)
2596 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
2597 if (func_id != BPF_FUNC_sk_select_reuseport)
2600 case BPF_MAP_TYPE_QUEUE:
2601 case BPF_MAP_TYPE_STACK:
2602 if (func_id != BPF_FUNC_map_peek_elem &&
2603 func_id != BPF_FUNC_map_pop_elem &&
2604 func_id != BPF_FUNC_map_push_elem)
2611 /* ... and second from the function itself. */
2613 case BPF_FUNC_tail_call:
2614 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
2616 if (env->subprog_cnt > 1) {
2617 verbose(env, "tail_calls are not allowed in programs with bpf-to-bpf calls\n");
2621 case BPF_FUNC_perf_event_read:
2622 case BPF_FUNC_perf_event_output:
2623 case BPF_FUNC_perf_event_read_value:
2624 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
2627 case BPF_FUNC_get_stackid:
2628 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
2631 case BPF_FUNC_current_task_under_cgroup:
2632 case BPF_FUNC_skb_under_cgroup:
2633 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
2636 case BPF_FUNC_redirect_map:
2637 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
2638 map->map_type != BPF_MAP_TYPE_CPUMAP &&
2639 map->map_type != BPF_MAP_TYPE_XSKMAP)
2642 case BPF_FUNC_sk_redirect_map:
2643 case BPF_FUNC_msg_redirect_map:
2644 case BPF_FUNC_sock_map_update:
2645 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
2648 case BPF_FUNC_sk_redirect_hash:
2649 case BPF_FUNC_msg_redirect_hash:
2650 case BPF_FUNC_sock_hash_update:
2651 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
2654 case BPF_FUNC_get_local_storage:
2655 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
2656 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
2659 case BPF_FUNC_sk_select_reuseport:
2660 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY)
2663 case BPF_FUNC_map_peek_elem:
2664 case BPF_FUNC_map_pop_elem:
2665 case BPF_FUNC_map_push_elem:
2666 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
2667 map->map_type != BPF_MAP_TYPE_STACK)
2676 verbose(env, "cannot pass map_type %d into func %s#%d\n",
2677 map->map_type, func_id_name(func_id), func_id);
2681 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
2685 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
2687 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
2689 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
2691 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
2693 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
2696 /* We only support one arg being in raw mode at the moment,
2697 * which is sufficient for the helper functions we have
2703 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
2704 enum bpf_arg_type arg_next)
2706 return (arg_type_is_mem_ptr(arg_curr) &&
2707 !arg_type_is_mem_size(arg_next)) ||
2708 (!arg_type_is_mem_ptr(arg_curr) &&
2709 arg_type_is_mem_size(arg_next));
2712 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
2714 /* bpf_xxx(..., buf, len) call will access 'len'
2715 * bytes from memory 'buf'. Both arg types need
2716 * to be paired, so make sure there's no buggy
2717 * helper function specification.
2719 if (arg_type_is_mem_size(fn->arg1_type) ||
2720 arg_type_is_mem_ptr(fn->arg5_type) ||
2721 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
2722 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
2723 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
2724 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
2730 static bool check_refcount_ok(const struct bpf_func_proto *fn, int func_id)
2734 if (arg_type_may_be_refcounted(fn->arg1_type))
2736 if (arg_type_may_be_refcounted(fn->arg2_type))
2738 if (arg_type_may_be_refcounted(fn->arg3_type))
2740 if (arg_type_may_be_refcounted(fn->arg4_type))
2742 if (arg_type_may_be_refcounted(fn->arg5_type))
2745 /* A reference acquiring function cannot acquire
2746 * another refcounted ptr.
2748 if (is_acquire_function(func_id) && count)
2751 /* We only support one arg being unreferenced at the moment,
2752 * which is sufficient for the helper functions we have right now.
2757 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
2759 return check_raw_mode_ok(fn) &&
2760 check_arg_pair_ok(fn) &&
2761 check_refcount_ok(fn, func_id) ? 0 : -EINVAL;
2764 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
2765 * are now invalid, so turn them into unknown SCALAR_VALUE.
2767 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
2768 struct bpf_func_state *state)
2770 struct bpf_reg_state *regs = state->regs, *reg;
2773 for (i = 0; i < MAX_BPF_REG; i++)
2774 if (reg_is_pkt_pointer_any(®s[i]))
2775 mark_reg_unknown(env, regs, i);
2777 bpf_for_each_spilled_reg(i, state, reg) {
2780 if (reg_is_pkt_pointer_any(reg))
2781 __mark_reg_unknown(reg);
2785 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
2787 struct bpf_verifier_state *vstate = env->cur_state;
2790 for (i = 0; i <= vstate->curframe; i++)
2791 __clear_all_pkt_pointers(env, vstate->frame[i]);
2794 static void release_reg_references(struct bpf_verifier_env *env,
2795 struct bpf_func_state *state,
2798 struct bpf_reg_state *regs = state->regs, *reg;
2801 for (i = 0; i < MAX_BPF_REG; i++)
2802 if (regs[i].ref_obj_id == ref_obj_id)
2803 mark_reg_unknown(env, regs, i);
2805 bpf_for_each_spilled_reg(i, state, reg) {
2808 if (reg->ref_obj_id == ref_obj_id)
2809 __mark_reg_unknown(reg);
2813 /* The pointer with the specified id has released its reference to kernel
2814 * resources. Identify all copies of the same pointer and clear the reference.
2816 static int release_reference(struct bpf_verifier_env *env,
2819 struct bpf_verifier_state *vstate = env->cur_state;
2823 err = release_reference_state(cur_func(env), ref_obj_id);
2827 for (i = 0; i <= vstate->curframe; i++)
2828 release_reg_references(env, vstate->frame[i], ref_obj_id);
2833 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
2836 struct bpf_verifier_state *state = env->cur_state;
2837 struct bpf_func_state *caller, *callee;
2838 int i, err, subprog, target_insn;
2840 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
2841 verbose(env, "the call stack of %d frames is too deep\n",
2842 state->curframe + 2);
2846 target_insn = *insn_idx + insn->imm;
2847 subprog = find_subprog(env, target_insn + 1);
2849 verbose(env, "verifier bug. No program starts at insn %d\n",
2854 caller = state->frame[state->curframe];
2855 if (state->frame[state->curframe + 1]) {
2856 verbose(env, "verifier bug. Frame %d already allocated\n",
2857 state->curframe + 1);
2861 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
2864 state->frame[state->curframe + 1] = callee;
2866 /* callee cannot access r0, r6 - r9 for reading and has to write
2867 * into its own stack before reading from it.
2868 * callee can read/write into caller's stack
2870 init_func_state(env, callee,
2871 /* remember the callsite, it will be used by bpf_exit */
2872 *insn_idx /* callsite */,
2873 state->curframe + 1 /* frameno within this callchain */,
2874 subprog /* subprog number within this prog */);
2876 /* Transfer references to the callee */
2877 err = transfer_reference_state(callee, caller);
2881 /* copy r1 - r5 args that callee can access. The copy includes parent
2882 * pointers, which connects us up to the liveness chain
2884 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
2885 callee->regs[i] = caller->regs[i];
2887 /* after the call registers r0 - r5 were scratched */
2888 for (i = 0; i < CALLER_SAVED_REGS; i++) {
2889 mark_reg_not_init(env, caller->regs, caller_saved[i]);
2890 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
2893 /* only increment it after check_reg_arg() finished */
2896 /* and go analyze first insn of the callee */
2897 *insn_idx = target_insn;
2899 if (env->log.level) {
2900 verbose(env, "caller:\n");
2901 print_verifier_state(env, caller);
2902 verbose(env, "callee:\n");
2903 print_verifier_state(env, callee);
2908 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
2910 struct bpf_verifier_state *state = env->cur_state;
2911 struct bpf_func_state *caller, *callee;
2912 struct bpf_reg_state *r0;
2915 callee = state->frame[state->curframe];
2916 r0 = &callee->regs[BPF_REG_0];
2917 if (r0->type == PTR_TO_STACK) {
2918 /* technically it's ok to return caller's stack pointer
2919 * (or caller's caller's pointer) back to the caller,
2920 * since these pointers are valid. Only current stack
2921 * pointer will be invalid as soon as function exits,
2922 * but let's be conservative
2924 verbose(env, "cannot return stack pointer to the caller\n");
2929 caller = state->frame[state->curframe];
2930 /* return to the caller whatever r0 had in the callee */
2931 caller->regs[BPF_REG_0] = *r0;
2933 /* Transfer references to the caller */
2934 err = transfer_reference_state(caller, callee);
2938 *insn_idx = callee->callsite + 1;
2939 if (env->log.level) {
2940 verbose(env, "returning from callee:\n");
2941 print_verifier_state(env, callee);
2942 verbose(env, "to caller at %d:\n", *insn_idx);
2943 print_verifier_state(env, caller);
2945 /* clear everything in the callee */
2946 free_func_state(callee);
2947 state->frame[state->curframe + 1] = NULL;
2951 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
2953 struct bpf_call_arg_meta *meta)
2955 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
2957 if (ret_type != RET_INTEGER ||
2958 (func_id != BPF_FUNC_get_stack &&
2959 func_id != BPF_FUNC_probe_read_str))
2962 ret_reg->smax_value = meta->msize_smax_value;
2963 ret_reg->umax_value = meta->msize_umax_value;
2964 __reg_deduce_bounds(ret_reg);
2965 __reg_bound_offset(ret_reg);
2969 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
2970 int func_id, int insn_idx)
2972 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
2974 if (func_id != BPF_FUNC_tail_call &&
2975 func_id != BPF_FUNC_map_lookup_elem &&
2976 func_id != BPF_FUNC_map_update_elem &&
2977 func_id != BPF_FUNC_map_delete_elem &&
2978 func_id != BPF_FUNC_map_push_elem &&
2979 func_id != BPF_FUNC_map_pop_elem &&
2980 func_id != BPF_FUNC_map_peek_elem)
2983 if (meta->map_ptr == NULL) {
2984 verbose(env, "kernel subsystem misconfigured verifier\n");
2988 if (!BPF_MAP_PTR(aux->map_state))
2989 bpf_map_ptr_store(aux, meta->map_ptr,
2990 meta->map_ptr->unpriv_array);
2991 else if (BPF_MAP_PTR(aux->map_state) != meta->map_ptr)
2992 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
2993 meta->map_ptr->unpriv_array);
2997 static int check_reference_leak(struct bpf_verifier_env *env)
2999 struct bpf_func_state *state = cur_func(env);
3002 for (i = 0; i < state->acquired_refs; i++) {
3003 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
3004 state->refs[i].id, state->refs[i].insn_idx);
3006 return state->acquired_refs ? -EINVAL : 0;
3009 static int check_helper_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
3011 const struct bpf_func_proto *fn = NULL;
3012 struct bpf_reg_state *regs;
3013 struct bpf_call_arg_meta meta;
3017 /* find function prototype */
3018 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
3019 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
3024 if (env->ops->get_func_proto)
3025 fn = env->ops->get_func_proto(func_id, env->prog);
3027 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
3032 /* eBPF programs must be GPL compatible to use GPL-ed functions */
3033 if (!env->prog->gpl_compatible && fn->gpl_only) {
3034 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
3038 /* With LD_ABS/IND some JITs save/restore skb from r1. */
3039 changes_data = bpf_helper_changes_pkt_data(fn->func);
3040 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
3041 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
3042 func_id_name(func_id), func_id);
3046 memset(&meta, 0, sizeof(meta));
3047 meta.pkt_access = fn->pkt_access;
3049 err = check_func_proto(fn, func_id);
3051 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
3052 func_id_name(func_id), func_id);
3056 meta.func_id = func_id;
3058 err = check_func_arg(env, BPF_REG_1, fn->arg1_type, &meta);
3061 err = check_func_arg(env, BPF_REG_2, fn->arg2_type, &meta);
3064 err = check_func_arg(env, BPF_REG_3, fn->arg3_type, &meta);
3067 err = check_func_arg(env, BPF_REG_4, fn->arg4_type, &meta);
3070 err = check_func_arg(env, BPF_REG_5, fn->arg5_type, &meta);
3074 err = record_func_map(env, &meta, func_id, insn_idx);
3078 /* Mark slots with STACK_MISC in case of raw mode, stack offset
3079 * is inferred from register state.
3081 for (i = 0; i < meta.access_size; i++) {
3082 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
3083 BPF_WRITE, -1, false);
3088 if (func_id == BPF_FUNC_tail_call) {
3089 err = check_reference_leak(env);
3091 verbose(env, "tail_call would lead to reference leak\n");
3094 } else if (is_release_function(func_id)) {
3095 err = release_reference(env, meta.ref_obj_id);
3097 verbose(env, "func %s#%d reference has not been acquired before\n",
3098 func_id_name(func_id), func_id);
3103 regs = cur_regs(env);
3105 /* check that flags argument in get_local_storage(map, flags) is 0,
3106 * this is required because get_local_storage() can't return an error.
3108 if (func_id == BPF_FUNC_get_local_storage &&
3109 !register_is_null(®s[BPF_REG_2])) {
3110 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
3114 /* reset caller saved regs */
3115 for (i = 0; i < CALLER_SAVED_REGS; i++) {
3116 mark_reg_not_init(env, regs, caller_saved[i]);
3117 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
3120 /* update return register (already marked as written above) */
3121 if (fn->ret_type == RET_INTEGER) {
3122 /* sets type to SCALAR_VALUE */
3123 mark_reg_unknown(env, regs, BPF_REG_0);
3124 } else if (fn->ret_type == RET_VOID) {
3125 regs[BPF_REG_0].type = NOT_INIT;
3126 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL ||
3127 fn->ret_type == RET_PTR_TO_MAP_VALUE) {
3128 /* There is no offset yet applied, variable or fixed */
3129 mark_reg_known_zero(env, regs, BPF_REG_0);
3130 /* remember map_ptr, so that check_map_access()
3131 * can check 'value_size' boundary of memory access
3132 * to map element returned from bpf_map_lookup_elem()
3134 if (meta.map_ptr == NULL) {
3136 "kernel subsystem misconfigured verifier\n");
3139 regs[BPF_REG_0].map_ptr = meta.map_ptr;
3140 if (fn->ret_type == RET_PTR_TO_MAP_VALUE) {
3141 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE;
3142 if (map_value_has_spin_lock(meta.map_ptr))
3143 regs[BPF_REG_0].id = ++env->id_gen;
3145 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
3146 regs[BPF_REG_0].id = ++env->id_gen;
3148 } else if (fn->ret_type == RET_PTR_TO_SOCKET_OR_NULL) {
3149 mark_reg_known_zero(env, regs, BPF_REG_0);
3150 regs[BPF_REG_0].type = PTR_TO_SOCKET_OR_NULL;
3151 regs[BPF_REG_0].id = ++env->id_gen;
3152 } else if (fn->ret_type == RET_PTR_TO_SOCK_COMMON_OR_NULL) {
3153 mark_reg_known_zero(env, regs, BPF_REG_0);
3154 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON_OR_NULL;
3155 regs[BPF_REG_0].id = ++env->id_gen;
3156 } else if (fn->ret_type == RET_PTR_TO_TCP_SOCK_OR_NULL) {
3157 mark_reg_known_zero(env, regs, BPF_REG_0);
3158 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK_OR_NULL;
3159 regs[BPF_REG_0].id = ++env->id_gen;
3161 verbose(env, "unknown return type %d of func %s#%d\n",
3162 fn->ret_type, func_id_name(func_id), func_id);
3166 if (is_ptr_cast_function(func_id)) {
3167 /* For release_reference() */
3168 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
3169 } else if (is_acquire_function(func_id)) {
3170 int id = acquire_reference_state(env, insn_idx);
3174 /* For mark_ptr_or_null_reg() */
3175 regs[BPF_REG_0].id = id;
3176 /* For release_reference() */
3177 regs[BPF_REG_0].ref_obj_id = id;
3180 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
3182 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
3186 if (func_id == BPF_FUNC_get_stack && !env->prog->has_callchain_buf) {
3187 const char *err_str;
3189 #ifdef CONFIG_PERF_EVENTS
3190 err = get_callchain_buffers(sysctl_perf_event_max_stack);
3191 err_str = "cannot get callchain buffer for func %s#%d\n";
3194 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
3197 verbose(env, err_str, func_id_name(func_id), func_id);
3201 env->prog->has_callchain_buf = true;
3205 clear_all_pkt_pointers(env);
3209 static bool signed_add_overflows(s64 a, s64 b)
3211 /* Do the add in u64, where overflow is well-defined */
3212 s64 res = (s64)((u64)a + (u64)b);
3219 static bool signed_sub_overflows(s64 a, s64 b)
3221 /* Do the sub in u64, where overflow is well-defined */
3222 s64 res = (s64)((u64)a - (u64)b);
3229 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
3230 const struct bpf_reg_state *reg,
3231 enum bpf_reg_type type)
3233 bool known = tnum_is_const(reg->var_off);
3234 s64 val = reg->var_off.value;
3235 s64 smin = reg->smin_value;
3237 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
3238 verbose(env, "math between %s pointer and %lld is not allowed\n",
3239 reg_type_str[type], val);
3243 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
3244 verbose(env, "%s pointer offset %d is not allowed\n",
3245 reg_type_str[type], reg->off);
3249 if (smin == S64_MIN) {
3250 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
3251 reg_type_str[type]);
3255 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
3256 verbose(env, "value %lld makes %s pointer be out of bounds\n",
3257 smin, reg_type_str[type]);
3264 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
3266 return &env->insn_aux_data[env->insn_idx];
3269 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
3270 u32 *ptr_limit, u8 opcode, bool off_is_neg)
3272 bool mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
3273 (opcode == BPF_SUB && !off_is_neg);
3276 switch (ptr_reg->type) {
3278 off = ptr_reg->off + ptr_reg->var_off.value;
3280 *ptr_limit = MAX_BPF_STACK + off;
3284 case PTR_TO_MAP_VALUE:
3286 *ptr_limit = ptr_reg->umax_value + ptr_reg->off;
3288 off = ptr_reg->smin_value + ptr_reg->off;
3289 *ptr_limit = ptr_reg->map_ptr->value_size - off;
3297 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
3298 const struct bpf_insn *insn)
3300 return env->allow_ptr_leaks || BPF_SRC(insn->code) == BPF_K;
3303 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
3304 u32 alu_state, u32 alu_limit)
3306 /* If we arrived here from different branches with different
3307 * state or limits to sanitize, then this won't work.
3309 if (aux->alu_state &&
3310 (aux->alu_state != alu_state ||
3311 aux->alu_limit != alu_limit))
3314 /* Corresponding fixup done in fixup_bpf_calls(). */
3315 aux->alu_state = alu_state;
3316 aux->alu_limit = alu_limit;
3320 static int sanitize_val_alu(struct bpf_verifier_env *env,
3321 struct bpf_insn *insn)
3323 struct bpf_insn_aux_data *aux = cur_aux(env);
3325 if (can_skip_alu_sanitation(env, insn))
3328 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
3331 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
3332 struct bpf_insn *insn,
3333 const struct bpf_reg_state *ptr_reg,
3334 struct bpf_reg_state *dst_reg,
3337 struct bpf_verifier_state *vstate = env->cur_state;
3338 struct bpf_insn_aux_data *aux = cur_aux(env);
3339 bool ptr_is_dst_reg = ptr_reg == dst_reg;
3340 u8 opcode = BPF_OP(insn->code);
3341 u32 alu_state, alu_limit;
3342 struct bpf_reg_state tmp;
3345 if (can_skip_alu_sanitation(env, insn))
3348 /* We already marked aux for masking from non-speculative
3349 * paths, thus we got here in the first place. We only care
3350 * to explore bad access from here.
3352 if (vstate->speculative)
3355 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
3356 alu_state |= ptr_is_dst_reg ?
3357 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
3359 if (retrieve_ptr_limit(ptr_reg, &alu_limit, opcode, off_is_neg))
3361 if (update_alu_sanitation_state(aux, alu_state, alu_limit))
3364 /* Simulate and find potential out-of-bounds access under
3365 * speculative execution from truncation as a result of
3366 * masking when off was not within expected range. If off
3367 * sits in dst, then we temporarily need to move ptr there
3368 * to simulate dst (== 0) +/-= ptr. Needed, for example,
3369 * for cases where we use K-based arithmetic in one direction
3370 * and truncated reg-based in the other in order to explore
3373 if (!ptr_is_dst_reg) {
3375 *dst_reg = *ptr_reg;
3377 ret = push_stack(env, env->insn_idx + 1, env->insn_idx, true);
3378 if (!ptr_is_dst_reg)
3380 return !ret ? -EFAULT : 0;
3383 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
3384 * Caller should also handle BPF_MOV case separately.
3385 * If we return -EACCES, caller may want to try again treating pointer as a
3386 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
3388 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
3389 struct bpf_insn *insn,
3390 const struct bpf_reg_state *ptr_reg,
3391 const struct bpf_reg_state *off_reg)
3393 struct bpf_verifier_state *vstate = env->cur_state;
3394 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3395 struct bpf_reg_state *regs = state->regs, *dst_reg;
3396 bool known = tnum_is_const(off_reg->var_off);
3397 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
3398 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
3399 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
3400 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
3401 u32 dst = insn->dst_reg, src = insn->src_reg;
3402 u8 opcode = BPF_OP(insn->code);
3405 dst_reg = ®s[dst];
3407 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
3408 smin_val > smax_val || umin_val > umax_val) {
3409 /* Taint dst register if offset had invalid bounds derived from
3410 * e.g. dead branches.
3412 __mark_reg_unknown(dst_reg);
3416 if (BPF_CLASS(insn->code) != BPF_ALU64) {
3417 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
3419 "R%d 32-bit pointer arithmetic prohibited\n",
3424 switch (ptr_reg->type) {
3425 case PTR_TO_MAP_VALUE_OR_NULL:
3426 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
3427 dst, reg_type_str[ptr_reg->type]);
3429 case CONST_PTR_TO_MAP:
3430 case PTR_TO_PACKET_END:
3432 case PTR_TO_SOCKET_OR_NULL:
3433 case PTR_TO_SOCK_COMMON:
3434 case PTR_TO_SOCK_COMMON_OR_NULL:
3435 case PTR_TO_TCP_SOCK:
3436 case PTR_TO_TCP_SOCK_OR_NULL:
3437 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
3438 dst, reg_type_str[ptr_reg->type]);
3440 case PTR_TO_MAP_VALUE:
3441 if (!env->allow_ptr_leaks && !known && (smin_val < 0) != (smax_val < 0)) {
3442 verbose(env, "R%d has unknown scalar with mixed signed bounds, pointer arithmetic with it prohibited for !root\n",
3443 off_reg == dst_reg ? dst : src);
3451 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
3452 * The id may be overwritten later if we create a new variable offset.
3454 dst_reg->type = ptr_reg->type;
3455 dst_reg->id = ptr_reg->id;
3457 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
3458 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
3463 ret = sanitize_ptr_alu(env, insn, ptr_reg, dst_reg, smin_val < 0);
3465 verbose(env, "R%d tried to add from different maps or paths\n", dst);
3468 /* We can take a fixed offset as long as it doesn't overflow
3469 * the s32 'off' field
3471 if (known && (ptr_reg->off + smin_val ==
3472 (s64)(s32)(ptr_reg->off + smin_val))) {
3473 /* pointer += K. Accumulate it into fixed offset */
3474 dst_reg->smin_value = smin_ptr;
3475 dst_reg->smax_value = smax_ptr;
3476 dst_reg->umin_value = umin_ptr;
3477 dst_reg->umax_value = umax_ptr;
3478 dst_reg->var_off = ptr_reg->var_off;
3479 dst_reg->off = ptr_reg->off + smin_val;
3480 dst_reg->raw = ptr_reg->raw;
3483 /* A new variable offset is created. Note that off_reg->off
3484 * == 0, since it's a scalar.
3485 * dst_reg gets the pointer type and since some positive
3486 * integer value was added to the pointer, give it a new 'id'
3487 * if it's a PTR_TO_PACKET.
3488 * this creates a new 'base' pointer, off_reg (variable) gets
3489 * added into the variable offset, and we copy the fixed offset
3492 if (signed_add_overflows(smin_ptr, smin_val) ||
3493 signed_add_overflows(smax_ptr, smax_val)) {
3494 dst_reg->smin_value = S64_MIN;
3495 dst_reg->smax_value = S64_MAX;
3497 dst_reg->smin_value = smin_ptr + smin_val;
3498 dst_reg->smax_value = smax_ptr + smax_val;
3500 if (umin_ptr + umin_val < umin_ptr ||
3501 umax_ptr + umax_val < umax_ptr) {
3502 dst_reg->umin_value = 0;
3503 dst_reg->umax_value = U64_MAX;
3505 dst_reg->umin_value = umin_ptr + umin_val;
3506 dst_reg->umax_value = umax_ptr + umax_val;
3508 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
3509 dst_reg->off = ptr_reg->off;
3510 dst_reg->raw = ptr_reg->raw;
3511 if (reg_is_pkt_pointer(ptr_reg)) {
3512 dst_reg->id = ++env->id_gen;
3513 /* something was added to pkt_ptr, set range to zero */
3518 ret = sanitize_ptr_alu(env, insn, ptr_reg, dst_reg, smin_val < 0);
3520 verbose(env, "R%d tried to sub from different maps or paths\n", dst);
3523 if (dst_reg == off_reg) {
3524 /* scalar -= pointer. Creates an unknown scalar */
3525 verbose(env, "R%d tried to subtract pointer from scalar\n",
3529 /* We don't allow subtraction from FP, because (according to
3530 * test_verifier.c test "invalid fp arithmetic", JITs might not
3531 * be able to deal with it.
3533 if (ptr_reg->type == PTR_TO_STACK) {
3534 verbose(env, "R%d subtraction from stack pointer prohibited\n",
3538 if (known && (ptr_reg->off - smin_val ==
3539 (s64)(s32)(ptr_reg->off - smin_val))) {
3540 /* pointer -= K. Subtract it from fixed offset */
3541 dst_reg->smin_value = smin_ptr;
3542 dst_reg->smax_value = smax_ptr;
3543 dst_reg->umin_value = umin_ptr;
3544 dst_reg->umax_value = umax_ptr;
3545 dst_reg->var_off = ptr_reg->var_off;
3546 dst_reg->id = ptr_reg->id;
3547 dst_reg->off = ptr_reg->off - smin_val;
3548 dst_reg->raw = ptr_reg->raw;
3551 /* A new variable offset is created. If the subtrahend is known
3552 * nonnegative, then any reg->range we had before is still good.
3554 if (signed_sub_overflows(smin_ptr, smax_val) ||
3555 signed_sub_overflows(smax_ptr, smin_val)) {
3556 /* Overflow possible, we know nothing */
3557 dst_reg->smin_value = S64_MIN;
3558 dst_reg->smax_value = S64_MAX;
3560 dst_reg->smin_value = smin_ptr - smax_val;
3561 dst_reg->smax_value = smax_ptr - smin_val;
3563 if (umin_ptr < umax_val) {
3564 /* Overflow possible, we know nothing */
3565 dst_reg->umin_value = 0;
3566 dst_reg->umax_value = U64_MAX;
3568 /* Cannot overflow (as long as bounds are consistent) */
3569 dst_reg->umin_value = umin_ptr - umax_val;
3570 dst_reg->umax_value = umax_ptr - umin_val;
3572 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
3573 dst_reg->off = ptr_reg->off;
3574 dst_reg->raw = ptr_reg->raw;
3575 if (reg_is_pkt_pointer(ptr_reg)) {
3576 dst_reg->id = ++env->id_gen;
3577 /* something was added to pkt_ptr, set range to zero */
3585 /* bitwise ops on pointers are troublesome, prohibit. */
3586 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
3587 dst, bpf_alu_string[opcode >> 4]);
3590 /* other operators (e.g. MUL,LSH) produce non-pointer results */
3591 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
3592 dst, bpf_alu_string[opcode >> 4]);
3596 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
3599 __update_reg_bounds(dst_reg);
3600 __reg_deduce_bounds(dst_reg);
3601 __reg_bound_offset(dst_reg);
3603 /* For unprivileged we require that resulting offset must be in bounds
3604 * in order to be able to sanitize access later on.
3606 if (!env->allow_ptr_leaks) {
3607 if (dst_reg->type == PTR_TO_MAP_VALUE &&
3608 check_map_access(env, dst, dst_reg->off, 1, false)) {
3609 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
3610 "prohibited for !root\n", dst);
3612 } else if (dst_reg->type == PTR_TO_STACK &&
3613 check_stack_access(env, dst_reg, dst_reg->off +
3614 dst_reg->var_off.value, 1)) {
3615 verbose(env, "R%d stack pointer arithmetic goes out of range, "
3616 "prohibited for !root\n", dst);
3624 /* WARNING: This function does calculations on 64-bit values, but the actual
3625 * execution may occur on 32-bit values. Therefore, things like bitshifts
3626 * need extra checks in the 32-bit case.
3628 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
3629 struct bpf_insn *insn,
3630 struct bpf_reg_state *dst_reg,
3631 struct bpf_reg_state src_reg)
3633 struct bpf_reg_state *regs = cur_regs(env);
3634 u8 opcode = BPF_OP(insn->code);
3635 bool src_known, dst_known;
3636 s64 smin_val, smax_val;
3637 u64 umin_val, umax_val;
3638 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
3639 u32 dst = insn->dst_reg;
3642 if (insn_bitness == 32) {
3643 /* Relevant for 32-bit RSH: Information can propagate towards
3644 * LSB, so it isn't sufficient to only truncate the output to
3647 coerce_reg_to_size(dst_reg, 4);
3648 coerce_reg_to_size(&src_reg, 4);
3651 smin_val = src_reg.smin_value;
3652 smax_val = src_reg.smax_value;
3653 umin_val = src_reg.umin_value;
3654 umax_val = src_reg.umax_value;
3655 src_known = tnum_is_const(src_reg.var_off);
3656 dst_known = tnum_is_const(dst_reg->var_off);
3658 if ((src_known && (smin_val != smax_val || umin_val != umax_val)) ||
3659 smin_val > smax_val || umin_val > umax_val) {
3660 /* Taint dst register if offset had invalid bounds derived from
3661 * e.g. dead branches.
3663 __mark_reg_unknown(dst_reg);
3668 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
3669 __mark_reg_unknown(dst_reg);
3675 ret = sanitize_val_alu(env, insn);
3677 verbose(env, "R%d tried to add from different pointers or scalars\n", dst);
3680 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
3681 signed_add_overflows(dst_reg->smax_value, smax_val)) {
3682 dst_reg->smin_value = S64_MIN;
3683 dst_reg->smax_value = S64_MAX;
3685 dst_reg->smin_value += smin_val;
3686 dst_reg->smax_value += smax_val;
3688 if (dst_reg->umin_value + umin_val < umin_val ||
3689 dst_reg->umax_value + umax_val < umax_val) {
3690 dst_reg->umin_value = 0;
3691 dst_reg->umax_value = U64_MAX;
3693 dst_reg->umin_value += umin_val;
3694 dst_reg->umax_value += umax_val;
3696 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
3699 ret = sanitize_val_alu(env, insn);
3701 verbose(env, "R%d tried to sub from different pointers or scalars\n", dst);
3704 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
3705 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
3706 /* Overflow possible, we know nothing */
3707 dst_reg->smin_value = S64_MIN;
3708 dst_reg->smax_value = S64_MAX;
3710 dst_reg->smin_value -= smax_val;
3711 dst_reg->smax_value -= smin_val;
3713 if (dst_reg->umin_value < umax_val) {
3714 /* Overflow possible, we know nothing */
3715 dst_reg->umin_value = 0;
3716 dst_reg->umax_value = U64_MAX;
3718 /* Cannot overflow (as long as bounds are consistent) */
3719 dst_reg->umin_value -= umax_val;
3720 dst_reg->umax_value -= umin_val;
3722 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
3725 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
3726 if (smin_val < 0 || dst_reg->smin_value < 0) {
3727 /* Ain't nobody got time to multiply that sign */
3728 __mark_reg_unbounded(dst_reg);
3729 __update_reg_bounds(dst_reg);
3732 /* Both values are positive, so we can work with unsigned and
3733 * copy the result to signed (unless it exceeds S64_MAX).
3735 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
3736 /* Potential overflow, we know nothing */
3737 __mark_reg_unbounded(dst_reg);
3738 /* (except what we can learn from the var_off) */
3739 __update_reg_bounds(dst_reg);
3742 dst_reg->umin_value *= umin_val;
3743 dst_reg->umax_value *= umax_val;
3744 if (dst_reg->umax_value > S64_MAX) {
3745 /* Overflow possible, we know nothing */
3746 dst_reg->smin_value = S64_MIN;
3747 dst_reg->smax_value = S64_MAX;
3749 dst_reg->smin_value = dst_reg->umin_value;
3750 dst_reg->smax_value = dst_reg->umax_value;
3754 if (src_known && dst_known) {
3755 __mark_reg_known(dst_reg, dst_reg->var_off.value &
3756 src_reg.var_off.value);
3759 /* We get our minimum from the var_off, since that's inherently
3760 * bitwise. Our maximum is the minimum of the operands' maxima.
3762 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
3763 dst_reg->umin_value = dst_reg->var_off.value;
3764 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
3765 if (dst_reg->smin_value < 0 || smin_val < 0) {
3766 /* Lose signed bounds when ANDing negative numbers,
3767 * ain't nobody got time for that.
3769 dst_reg->smin_value = S64_MIN;
3770 dst_reg->smax_value = S64_MAX;
3772 /* ANDing two positives gives a positive, so safe to
3773 * cast result into s64.
3775 dst_reg->smin_value = dst_reg->umin_value;
3776 dst_reg->smax_value = dst_reg->umax_value;
3778 /* We may learn something more from the var_off */
3779 __update_reg_bounds(dst_reg);
3782 if (src_known && dst_known) {
3783 __mark_reg_known(dst_reg, dst_reg->var_off.value |
3784 src_reg.var_off.value);
3787 /* We get our maximum from the var_off, and our minimum is the
3788 * maximum of the operands' minima
3790 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
3791 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
3792 dst_reg->umax_value = dst_reg->var_off.value |
3793 dst_reg->var_off.mask;
3794 if (dst_reg->smin_value < 0 || smin_val < 0) {
3795 /* Lose signed bounds when ORing negative numbers,
3796 * ain't nobody got time for that.
3798 dst_reg->smin_value = S64_MIN;
3799 dst_reg->smax_value = S64_MAX;
3801 /* ORing two positives gives a positive, so safe to
3802 * cast result into s64.
3804 dst_reg->smin_value = dst_reg->umin_value;
3805 dst_reg->smax_value = dst_reg->umax_value;
3807 /* We may learn something more from the var_off */
3808 __update_reg_bounds(dst_reg);
3811 if (umax_val >= insn_bitness) {
3812 /* Shifts greater than 31 or 63 are undefined.
3813 * This includes shifts by a negative number.
3815 mark_reg_unknown(env, regs, insn->dst_reg);
3818 /* We lose all sign bit information (except what we can pick
3821 dst_reg->smin_value = S64_MIN;
3822 dst_reg->smax_value = S64_MAX;
3823 /* If we might shift our top bit out, then we know nothing */
3824 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
3825 dst_reg->umin_value = 0;
3826 dst_reg->umax_value = U64_MAX;
3828 dst_reg->umin_value <<= umin_val;
3829 dst_reg->umax_value <<= umax_val;
3831 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
3832 /* We may learn something more from the var_off */
3833 __update_reg_bounds(dst_reg);
3836 if (umax_val >= insn_bitness) {
3837 /* Shifts greater than 31 or 63 are undefined.
3838 * This includes shifts by a negative number.
3840 mark_reg_unknown(env, regs, insn->dst_reg);
3843 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
3844 * be negative, then either:
3845 * 1) src_reg might be zero, so the sign bit of the result is
3846 * unknown, so we lose our signed bounds
3847 * 2) it's known negative, thus the unsigned bounds capture the
3849 * 3) the signed bounds cross zero, so they tell us nothing
3851 * If the value in dst_reg is known nonnegative, then again the
3852 * unsigned bounts capture the signed bounds.
3853 * Thus, in all cases it suffices to blow away our signed bounds
3854 * and rely on inferring new ones from the unsigned bounds and
3855 * var_off of the result.
3857 dst_reg->smin_value = S64_MIN;
3858 dst_reg->smax_value = S64_MAX;
3859 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
3860 dst_reg->umin_value >>= umax_val;
3861 dst_reg->umax_value >>= umin_val;
3862 /* We may learn something more from the var_off */
3863 __update_reg_bounds(dst_reg);
3866 if (umax_val >= insn_bitness) {
3867 /* Shifts greater than 31 or 63 are undefined.
3868 * This includes shifts by a negative number.
3870 mark_reg_unknown(env, regs, insn->dst_reg);
3874 /* Upon reaching here, src_known is true and
3875 * umax_val is equal to umin_val.
3877 dst_reg->smin_value >>= umin_val;
3878 dst_reg->smax_value >>= umin_val;
3879 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val);
3881 /* blow away the dst_reg umin_value/umax_value and rely on
3882 * dst_reg var_off to refine the result.
3884 dst_reg->umin_value = 0;
3885 dst_reg->umax_value = U64_MAX;
3886 __update_reg_bounds(dst_reg);
3889 mark_reg_unknown(env, regs, insn->dst_reg);
3893 if (BPF_CLASS(insn->code) != BPF_ALU64) {
3894 /* 32-bit ALU ops are (32,32)->32 */
3895 coerce_reg_to_size(dst_reg, 4);
3898 __reg_deduce_bounds(dst_reg);
3899 __reg_bound_offset(dst_reg);
3903 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
3906 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
3907 struct bpf_insn *insn)
3909 struct bpf_verifier_state *vstate = env->cur_state;
3910 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3911 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
3912 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
3913 u8 opcode = BPF_OP(insn->code);
3915 dst_reg = ®s[insn->dst_reg];
3917 if (dst_reg->type != SCALAR_VALUE)
3919 if (BPF_SRC(insn->code) == BPF_X) {
3920 src_reg = ®s[insn->src_reg];
3921 if (src_reg->type != SCALAR_VALUE) {
3922 if (dst_reg->type != SCALAR_VALUE) {
3923 /* Combining two pointers by any ALU op yields
3924 * an arbitrary scalar. Disallow all math except
3925 * pointer subtraction
3927 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
3928 mark_reg_unknown(env, regs, insn->dst_reg);
3931 verbose(env, "R%d pointer %s pointer prohibited\n",
3933 bpf_alu_string[opcode >> 4]);
3936 /* scalar += pointer
3937 * This is legal, but we have to reverse our
3938 * src/dest handling in computing the range
3940 return adjust_ptr_min_max_vals(env, insn,
3943 } else if (ptr_reg) {
3944 /* pointer += scalar */
3945 return adjust_ptr_min_max_vals(env, insn,
3949 /* Pretend the src is a reg with a known value, since we only
3950 * need to be able to read from this state.
3952 off_reg.type = SCALAR_VALUE;
3953 __mark_reg_known(&off_reg, insn->imm);
3955 if (ptr_reg) /* pointer += K */
3956 return adjust_ptr_min_max_vals(env, insn,
3960 /* Got here implies adding two SCALAR_VALUEs */
3961 if (WARN_ON_ONCE(ptr_reg)) {
3962 print_verifier_state(env, state);
3963 verbose(env, "verifier internal error: unexpected ptr_reg\n");
3966 if (WARN_ON(!src_reg)) {
3967 print_verifier_state(env, state);
3968 verbose(env, "verifier internal error: no src_reg\n");
3971 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
3974 /* check validity of 32-bit and 64-bit arithmetic operations */
3975 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
3977 struct bpf_reg_state *regs = cur_regs(env);
3978 u8 opcode = BPF_OP(insn->code);
3981 if (opcode == BPF_END || opcode == BPF_NEG) {
3982 if (opcode == BPF_NEG) {
3983 if (BPF_SRC(insn->code) != 0 ||
3984 insn->src_reg != BPF_REG_0 ||
3985 insn->off != 0 || insn->imm != 0) {
3986 verbose(env, "BPF_NEG uses reserved fields\n");
3990 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
3991 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
3992 BPF_CLASS(insn->code) == BPF_ALU64) {
3993 verbose(env, "BPF_END uses reserved fields\n");
3998 /* check src operand */
3999 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4003 if (is_pointer_value(env, insn->dst_reg)) {
4004 verbose(env, "R%d pointer arithmetic prohibited\n",
4009 /* check dest operand */
4010 err = check_reg_arg(env, insn->dst_reg, DST_OP);
4014 } else if (opcode == BPF_MOV) {
4016 if (BPF_SRC(insn->code) == BPF_X) {
4017 if (insn->imm != 0 || insn->off != 0) {
4018 verbose(env, "BPF_MOV uses reserved fields\n");
4022 /* check src operand */
4023 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4027 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
4028 verbose(env, "BPF_MOV uses reserved fields\n");
4033 /* check dest operand, mark as required later */
4034 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
4038 if (BPF_SRC(insn->code) == BPF_X) {
4039 struct bpf_reg_state *src_reg = regs + insn->src_reg;
4040 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
4042 if (BPF_CLASS(insn->code) == BPF_ALU64) {
4044 * copy register state to dest reg
4046 *dst_reg = *src_reg;
4047 dst_reg->live |= REG_LIVE_WRITTEN;
4050 if (is_pointer_value(env, insn->src_reg)) {
4052 "R%d partial copy of pointer\n",
4055 } else if (src_reg->type == SCALAR_VALUE) {
4056 *dst_reg = *src_reg;
4057 dst_reg->live |= REG_LIVE_WRITTEN;
4059 mark_reg_unknown(env, regs,
4062 coerce_reg_to_size(dst_reg, 4);
4066 * remember the value we stored into this reg
4068 /* clear any state __mark_reg_known doesn't set */
4069 mark_reg_unknown(env, regs, insn->dst_reg);
4070 regs[insn->dst_reg].type = SCALAR_VALUE;
4071 if (BPF_CLASS(insn->code) == BPF_ALU64) {
4072 __mark_reg_known(regs + insn->dst_reg,
4075 __mark_reg_known(regs + insn->dst_reg,
4080 } else if (opcode > BPF_END) {
4081 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
4084 } else { /* all other ALU ops: and, sub, xor, add, ... */
4086 if (BPF_SRC(insn->code) == BPF_X) {
4087 if (insn->imm != 0 || insn->off != 0) {
4088 verbose(env, "BPF_ALU uses reserved fields\n");
4091 /* check src1 operand */
4092 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4096 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
4097 verbose(env, "BPF_ALU uses reserved fields\n");
4102 /* check src2 operand */
4103 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4107 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
4108 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
4109 verbose(env, "div by zero\n");
4113 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
4114 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
4115 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
4117 if (insn->imm < 0 || insn->imm >= size) {
4118 verbose(env, "invalid shift %d\n", insn->imm);
4123 /* check dest operand */
4124 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
4128 return adjust_reg_min_max_vals(env, insn);
4134 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
4135 struct bpf_reg_state *dst_reg,
4136 enum bpf_reg_type type,
4137 bool range_right_open)
4139 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4140 struct bpf_reg_state *regs = state->regs, *reg;
4144 if (dst_reg->off < 0 ||
4145 (dst_reg->off == 0 && range_right_open))
4146 /* This doesn't give us any range */
4149 if (dst_reg->umax_value > MAX_PACKET_OFF ||
4150 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
4151 /* Risk of overflow. For instance, ptr + (1<<63) may be less
4152 * than pkt_end, but that's because it's also less than pkt.
4156 new_range = dst_reg->off;
4157 if (range_right_open)
4160 /* Examples for register markings:
4162 * pkt_data in dst register:
4166 * if (r2 > pkt_end) goto <handle exception>
4171 * if (r2 < pkt_end) goto <access okay>
4172 * <handle exception>
4175 * r2 == dst_reg, pkt_end == src_reg
4176 * r2=pkt(id=n,off=8,r=0)
4177 * r3=pkt(id=n,off=0,r=0)
4179 * pkt_data in src register:
4183 * if (pkt_end >= r2) goto <access okay>
4184 * <handle exception>
4188 * if (pkt_end <= r2) goto <handle exception>
4192 * pkt_end == dst_reg, r2 == src_reg
4193 * r2=pkt(id=n,off=8,r=0)
4194 * r3=pkt(id=n,off=0,r=0)
4196 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
4197 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
4198 * and [r3, r3 + 8-1) respectively is safe to access depending on
4202 /* If our ids match, then we must have the same max_value. And we
4203 * don't care about the other reg's fixed offset, since if it's too big
4204 * the range won't allow anything.
4205 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
4207 for (i = 0; i < MAX_BPF_REG; i++)
4208 if (regs[i].type == type && regs[i].id == dst_reg->id)
4209 /* keep the maximum range already checked */
4210 regs[i].range = max(regs[i].range, new_range);
4212 for (j = 0; j <= vstate->curframe; j++) {
4213 state = vstate->frame[j];
4214 bpf_for_each_spilled_reg(i, state, reg) {
4217 if (reg->type == type && reg->id == dst_reg->id)
4218 reg->range = max(reg->range, new_range);
4223 /* compute branch direction of the expression "if (reg opcode val) goto target;"
4225 * 1 - branch will be taken and "goto target" will be executed
4226 * 0 - branch will not be taken and fall-through to next insn
4227 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value range [0,10]
4229 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
4232 struct bpf_reg_state reg_lo;
4235 if (__is_pointer_value(false, reg))
4241 /* For JMP32, only low 32 bits are compared, coerce_reg_to_size
4242 * could truncate high bits and update umin/umax according to
4243 * information of low bits.
4245 coerce_reg_to_size(reg, 4);
4246 /* smin/smax need special handling. For example, after coerce,
4247 * if smin_value is 0x00000000ffffffffLL, the value is -1 when
4248 * used as operand to JMP32. It is a negative number from s32's
4249 * point of view, while it is a positive number when seen as
4250 * s64. The smin/smax are kept as s64, therefore, when used with
4251 * JMP32, they need to be transformed into s32, then sign
4252 * extended back to s64.
4254 * Also, smin/smax were copied from umin/umax. If umin/umax has
4255 * different sign bit, then min/max relationship doesn't
4256 * maintain after casting into s32, for this case, set smin/smax
4259 if ((reg->umax_value ^ reg->umin_value) &
4261 reg->smin_value = S32_MIN;
4262 reg->smax_value = S32_MAX;
4264 reg->smin_value = (s64)(s32)reg->smin_value;
4265 reg->smax_value = (s64)(s32)reg->smax_value;
4268 sval = (s64)(s32)val;
4275 if (tnum_is_const(reg->var_off))
4276 return !!tnum_equals_const(reg->var_off, val);
4279 if (tnum_is_const(reg->var_off))
4280 return !tnum_equals_const(reg->var_off, val);
4283 if ((~reg->var_off.mask & reg->var_off.value) & val)
4285 if (!((reg->var_off.mask | reg->var_off.value) & val))
4289 if (reg->umin_value > val)
4291 else if (reg->umax_value <= val)
4295 if (reg->smin_value > sval)
4297 else if (reg->smax_value < sval)
4301 if (reg->umax_value < val)
4303 else if (reg->umin_value >= val)
4307 if (reg->smax_value < sval)
4309 else if (reg->smin_value >= sval)
4313 if (reg->umin_value >= val)
4315 else if (reg->umax_value < val)
4319 if (reg->smin_value >= sval)
4321 else if (reg->smax_value < sval)
4325 if (reg->umax_value <= val)
4327 else if (reg->umin_value > val)
4331 if (reg->smax_value <= sval)
4333 else if (reg->smin_value > sval)
4341 /* Generate min value of the high 32-bit from TNUM info. */
4342 static u64 gen_hi_min(struct tnum var)
4344 return var.value & ~0xffffffffULL;
4347 /* Generate max value of the high 32-bit from TNUM info. */
4348 static u64 gen_hi_max(struct tnum var)
4350 return (var.value | var.mask) & ~0xffffffffULL;
4353 /* Return true if VAL is compared with a s64 sign extended from s32, and they
4354 * are with the same signedness.
4356 static bool cmp_val_with_extended_s64(s64 sval, struct bpf_reg_state *reg)
4358 return ((s32)sval >= 0 &&
4359 reg->smin_value >= 0 && reg->smax_value <= S32_MAX) ||
4361 reg->smax_value <= 0 && reg->smin_value >= S32_MIN);
4364 /* Adjusts the register min/max values in the case that the dst_reg is the
4365 * variable register that we are working on, and src_reg is a constant or we're
4366 * simply doing a BPF_K check.
4367 * In JEQ/JNE cases we also adjust the var_off values.
4369 static void reg_set_min_max(struct bpf_reg_state *true_reg,
4370 struct bpf_reg_state *false_reg, u64 val,
4371 u8 opcode, bool is_jmp32)
4375 /* If the dst_reg is a pointer, we can't learn anything about its
4376 * variable offset from the compare (unless src_reg were a pointer into
4377 * the same object, but we don't bother with that.
4378 * Since false_reg and true_reg have the same type by construction, we
4379 * only need to check one of them for pointerness.
4381 if (__is_pointer_value(false, false_reg))
4384 val = is_jmp32 ? (u32)val : val;
4385 sval = is_jmp32 ? (s64)(s32)val : (s64)val;
4391 struct bpf_reg_state *reg =
4392 opcode == BPF_JEQ ? true_reg : false_reg;
4394 /* For BPF_JEQ, if this is false we know nothing Jon Snow, but
4395 * if it is true we know the value for sure. Likewise for
4399 u64 old_v = reg->var_off.value;
4400 u64 hi_mask = ~0xffffffffULL;
4402 reg->var_off.value = (old_v & hi_mask) | val;
4403 reg->var_off.mask &= hi_mask;
4405 __mark_reg_known(reg, val);
4410 false_reg->var_off = tnum_and(false_reg->var_off,
4412 if (is_power_of_2(val))
4413 true_reg->var_off = tnum_or(true_reg->var_off,
4419 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
4420 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
4423 false_umax += gen_hi_max(false_reg->var_off);
4424 true_umin += gen_hi_min(true_reg->var_off);
4426 false_reg->umax_value = min(false_reg->umax_value, false_umax);
4427 true_reg->umin_value = max(true_reg->umin_value, true_umin);
4433 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
4434 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
4436 /* If the full s64 was not sign-extended from s32 then don't
4437 * deduct further info.
4439 if (is_jmp32 && !cmp_val_with_extended_s64(sval, false_reg))
4441 false_reg->smax_value = min(false_reg->smax_value, false_smax);
4442 true_reg->smin_value = max(true_reg->smin_value, true_smin);
4448 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
4449 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
4452 false_umin += gen_hi_min(false_reg->var_off);
4453 true_umax += gen_hi_max(true_reg->var_off);
4455 false_reg->umin_value = max(false_reg->umin_value, false_umin);
4456 true_reg->umax_value = min(true_reg->umax_value, true_umax);
4462 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
4463 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
4465 if (is_jmp32 && !cmp_val_with_extended_s64(sval, false_reg))
4467 false_reg->smin_value = max(false_reg->smin_value, false_smin);
4468 true_reg->smax_value = min(true_reg->smax_value, true_smax);
4475 __reg_deduce_bounds(false_reg);
4476 __reg_deduce_bounds(true_reg);
4477 /* We might have learned some bits from the bounds. */
4478 __reg_bound_offset(false_reg);
4479 __reg_bound_offset(true_reg);
4480 /* Intersecting with the old var_off might have improved our bounds
4481 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
4482 * then new var_off is (0; 0x7f...fc) which improves our umax.
4484 __update_reg_bounds(false_reg);
4485 __update_reg_bounds(true_reg);
4488 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
4491 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
4492 struct bpf_reg_state *false_reg, u64 val,
4493 u8 opcode, bool is_jmp32)
4497 if (__is_pointer_value(false, false_reg))
4500 val = is_jmp32 ? (u32)val : val;
4501 sval = is_jmp32 ? (s64)(s32)val : (s64)val;
4507 struct bpf_reg_state *reg =
4508 opcode == BPF_JEQ ? true_reg : false_reg;
4511 u64 old_v = reg->var_off.value;
4512 u64 hi_mask = ~0xffffffffULL;
4514 reg->var_off.value = (old_v & hi_mask) | val;
4515 reg->var_off.mask &= hi_mask;
4517 __mark_reg_known(reg, val);
4522 false_reg->var_off = tnum_and(false_reg->var_off,
4524 if (is_power_of_2(val))
4525 true_reg->var_off = tnum_or(true_reg->var_off,
4531 u64 false_umin = opcode == BPF_JGT ? val : val + 1;
4532 u64 true_umax = opcode == BPF_JGT ? val - 1 : val;
4535 false_umin += gen_hi_min(false_reg->var_off);
4536 true_umax += gen_hi_max(true_reg->var_off);
4538 false_reg->umin_value = max(false_reg->umin_value, false_umin);
4539 true_reg->umax_value = min(true_reg->umax_value, true_umax);
4545 s64 false_smin = opcode == BPF_JSGT ? sval : sval + 1;
4546 s64 true_smax = opcode == BPF_JSGT ? sval - 1 : sval;
4548 if (is_jmp32 && !cmp_val_with_extended_s64(sval, false_reg))
4550 false_reg->smin_value = max(false_reg->smin_value, false_smin);
4551 true_reg->smax_value = min(true_reg->smax_value, true_smax);
4557 u64 false_umax = opcode == BPF_JLT ? val : val - 1;
4558 u64 true_umin = opcode == BPF_JLT ? val + 1 : val;
4561 false_umax += gen_hi_max(false_reg->var_off);
4562 true_umin += gen_hi_min(true_reg->var_off);
4564 false_reg->umax_value = min(false_reg->umax_value, false_umax);
4565 true_reg->umin_value = max(true_reg->umin_value, true_umin);
4571 s64 false_smax = opcode == BPF_JSLT ? sval : sval - 1;
4572 s64 true_smin = opcode == BPF_JSLT ? sval + 1 : sval;
4574 if (is_jmp32 && !cmp_val_with_extended_s64(sval, false_reg))
4576 false_reg->smax_value = min(false_reg->smax_value, false_smax);
4577 true_reg->smin_value = max(true_reg->smin_value, true_smin);
4584 __reg_deduce_bounds(false_reg);
4585 __reg_deduce_bounds(true_reg);
4586 /* We might have learned some bits from the bounds. */
4587 __reg_bound_offset(false_reg);
4588 __reg_bound_offset(true_reg);
4589 /* Intersecting with the old var_off might have improved our bounds
4590 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
4591 * then new var_off is (0; 0x7f...fc) which improves our umax.
4593 __update_reg_bounds(false_reg);
4594 __update_reg_bounds(true_reg);
4597 /* Regs are known to be equal, so intersect their min/max/var_off */
4598 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
4599 struct bpf_reg_state *dst_reg)
4601 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
4602 dst_reg->umin_value);
4603 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
4604 dst_reg->umax_value);
4605 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
4606 dst_reg->smin_value);
4607 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
4608 dst_reg->smax_value);
4609 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
4611 /* We might have learned new bounds from the var_off. */
4612 __update_reg_bounds(src_reg);
4613 __update_reg_bounds(dst_reg);
4614 /* We might have learned something about the sign bit. */
4615 __reg_deduce_bounds(src_reg);
4616 __reg_deduce_bounds(dst_reg);
4617 /* We might have learned some bits from the bounds. */
4618 __reg_bound_offset(src_reg);
4619 __reg_bound_offset(dst_reg);
4620 /* Intersecting with the old var_off might have improved our bounds
4621 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
4622 * then new var_off is (0; 0x7f...fc) which improves our umax.
4624 __update_reg_bounds(src_reg);
4625 __update_reg_bounds(dst_reg);
4628 static void reg_combine_min_max(struct bpf_reg_state *true_src,
4629 struct bpf_reg_state *true_dst,
4630 struct bpf_reg_state *false_src,
4631 struct bpf_reg_state *false_dst,
4636 __reg_combine_min_max(true_src, true_dst);
4639 __reg_combine_min_max(false_src, false_dst);
4644 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
4645 struct bpf_reg_state *reg, u32 id,
4648 if (reg_type_may_be_null(reg->type) && reg->id == id) {
4649 /* Old offset (both fixed and variable parts) should
4650 * have been known-zero, because we don't allow pointer
4651 * arithmetic on pointers that might be NULL.
4653 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
4654 !tnum_equals_const(reg->var_off, 0) ||
4656 __mark_reg_known_zero(reg);
4660 reg->type = SCALAR_VALUE;
4661 } else if (reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
4662 if (reg->map_ptr->inner_map_meta) {
4663 reg->type = CONST_PTR_TO_MAP;
4664 reg->map_ptr = reg->map_ptr->inner_map_meta;
4666 reg->type = PTR_TO_MAP_VALUE;
4668 } else if (reg->type == PTR_TO_SOCKET_OR_NULL) {
4669 reg->type = PTR_TO_SOCKET;
4670 } else if (reg->type == PTR_TO_SOCK_COMMON_OR_NULL) {
4671 reg->type = PTR_TO_SOCK_COMMON;
4672 } else if (reg->type == PTR_TO_TCP_SOCK_OR_NULL) {
4673 reg->type = PTR_TO_TCP_SOCK;
4676 /* We don't need id and ref_obj_id from this point
4677 * onwards anymore, thus we should better reset it,
4678 * so that state pruning has chances to take effect.
4681 reg->ref_obj_id = 0;
4682 } else if (!reg_may_point_to_spin_lock(reg)) {
4683 /* For not-NULL ptr, reg->ref_obj_id will be reset
4684 * in release_reg_references().
4686 * reg->id is still used by spin_lock ptr. Other
4687 * than spin_lock ptr type, reg->id can be reset.
4694 /* The logic is similar to find_good_pkt_pointers(), both could eventually
4695 * be folded together at some point.
4697 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
4700 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4701 struct bpf_reg_state *reg, *regs = state->regs;
4702 u32 ref_obj_id = regs[regno].ref_obj_id;
4703 u32 id = regs[regno].id;
4706 if (ref_obj_id && ref_obj_id == id && is_null)
4707 /* regs[regno] is in the " == NULL" branch.
4708 * No one could have freed the reference state before
4709 * doing the NULL check.
4711 WARN_ON_ONCE(release_reference_state(state, id));
4713 for (i = 0; i < MAX_BPF_REG; i++)
4714 mark_ptr_or_null_reg(state, ®s[i], id, is_null);
4716 for (j = 0; j <= vstate->curframe; j++) {
4717 state = vstate->frame[j];
4718 bpf_for_each_spilled_reg(i, state, reg) {
4721 mark_ptr_or_null_reg(state, reg, id, is_null);
4726 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
4727 struct bpf_reg_state *dst_reg,
4728 struct bpf_reg_state *src_reg,
4729 struct bpf_verifier_state *this_branch,
4730 struct bpf_verifier_state *other_branch)
4732 if (BPF_SRC(insn->code) != BPF_X)
4735 /* Pointers are always 64-bit. */
4736 if (BPF_CLASS(insn->code) == BPF_JMP32)
4739 switch (BPF_OP(insn->code)) {
4741 if ((dst_reg->type == PTR_TO_PACKET &&
4742 src_reg->type == PTR_TO_PACKET_END) ||
4743 (dst_reg->type == PTR_TO_PACKET_META &&
4744 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
4745 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
4746 find_good_pkt_pointers(this_branch, dst_reg,
4747 dst_reg->type, false);
4748 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
4749 src_reg->type == PTR_TO_PACKET) ||
4750 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
4751 src_reg->type == PTR_TO_PACKET_META)) {
4752 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
4753 find_good_pkt_pointers(other_branch, src_reg,
4754 src_reg->type, true);
4760 if ((dst_reg->type == PTR_TO_PACKET &&
4761 src_reg->type == PTR_TO_PACKET_END) ||
4762 (dst_reg->type == PTR_TO_PACKET_META &&
4763 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
4764 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
4765 find_good_pkt_pointers(other_branch, dst_reg,
4766 dst_reg->type, true);
4767 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
4768 src_reg->type == PTR_TO_PACKET) ||
4769 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
4770 src_reg->type == PTR_TO_PACKET_META)) {
4771 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
4772 find_good_pkt_pointers(this_branch, src_reg,
4773 src_reg->type, false);
4779 if ((dst_reg->type == PTR_TO_PACKET &&
4780 src_reg->type == PTR_TO_PACKET_END) ||
4781 (dst_reg->type == PTR_TO_PACKET_META &&
4782 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
4783 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
4784 find_good_pkt_pointers(this_branch, dst_reg,
4785 dst_reg->type, true);
4786 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
4787 src_reg->type == PTR_TO_PACKET) ||
4788 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
4789 src_reg->type == PTR_TO_PACKET_META)) {
4790 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
4791 find_good_pkt_pointers(other_branch, src_reg,
4792 src_reg->type, false);
4798 if ((dst_reg->type == PTR_TO_PACKET &&
4799 src_reg->type == PTR_TO_PACKET_END) ||
4800 (dst_reg->type == PTR_TO_PACKET_META &&
4801 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
4802 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
4803 find_good_pkt_pointers(other_branch, dst_reg,
4804 dst_reg->type, false);
4805 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
4806 src_reg->type == PTR_TO_PACKET) ||
4807 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
4808 src_reg->type == PTR_TO_PACKET_META)) {
4809 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
4810 find_good_pkt_pointers(this_branch, src_reg,
4811 src_reg->type, true);
4823 static int check_cond_jmp_op(struct bpf_verifier_env *env,
4824 struct bpf_insn *insn, int *insn_idx)
4826 struct bpf_verifier_state *this_branch = env->cur_state;
4827 struct bpf_verifier_state *other_branch;
4828 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
4829 struct bpf_reg_state *dst_reg, *other_branch_regs;
4830 u8 opcode = BPF_OP(insn->code);
4834 /* Only conditional jumps are expected to reach here. */
4835 if (opcode == BPF_JA || opcode > BPF_JSLE) {
4836 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
4840 if (BPF_SRC(insn->code) == BPF_X) {
4841 if (insn->imm != 0) {
4842 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
4846 /* check src1 operand */
4847 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4851 if (is_pointer_value(env, insn->src_reg)) {
4852 verbose(env, "R%d pointer comparison prohibited\n",
4857 if (insn->src_reg != BPF_REG_0) {
4858 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
4863 /* check src2 operand */
4864 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4868 dst_reg = ®s[insn->dst_reg];
4869 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
4871 if (BPF_SRC(insn->code) == BPF_K) {
4872 int pred = is_branch_taken(dst_reg, insn->imm, opcode,
4876 /* only follow the goto, ignore fall-through */
4877 *insn_idx += insn->off;
4879 } else if (pred == 0) {
4880 /* only follow fall-through branch, since
4881 * that's where the program will go
4887 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
4891 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
4893 /* detect if we are comparing against a constant value so we can adjust
4894 * our min/max values for our dst register.
4895 * this is only legit if both are scalars (or pointers to the same
4896 * object, I suppose, but we don't support that right now), because
4897 * otherwise the different base pointers mean the offsets aren't
4900 if (BPF_SRC(insn->code) == BPF_X) {
4901 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
4902 struct bpf_reg_state lo_reg0 = *dst_reg;
4903 struct bpf_reg_state lo_reg1 = *src_reg;
4904 struct bpf_reg_state *src_lo, *dst_lo;
4908 coerce_reg_to_size(dst_lo, 4);
4909 coerce_reg_to_size(src_lo, 4);
4911 if (dst_reg->type == SCALAR_VALUE &&
4912 src_reg->type == SCALAR_VALUE) {
4913 if (tnum_is_const(src_reg->var_off) ||
4914 (is_jmp32 && tnum_is_const(src_lo->var_off)))
4915 reg_set_min_max(&other_branch_regs[insn->dst_reg],
4918 ? src_lo->var_off.value
4919 : src_reg->var_off.value,
4921 else if (tnum_is_const(dst_reg->var_off) ||
4922 (is_jmp32 && tnum_is_const(dst_lo->var_off)))
4923 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
4926 ? dst_lo->var_off.value
4927 : dst_reg->var_off.value,
4929 else if (!is_jmp32 &&
4930 (opcode == BPF_JEQ || opcode == BPF_JNE))
4931 /* Comparing for equality, we can combine knowledge */
4932 reg_combine_min_max(&other_branch_regs[insn->src_reg],
4933 &other_branch_regs[insn->dst_reg],
4934 src_reg, dst_reg, opcode);
4936 } else if (dst_reg->type == SCALAR_VALUE) {
4937 reg_set_min_max(&other_branch_regs[insn->dst_reg],
4938 dst_reg, insn->imm, opcode, is_jmp32);
4941 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
4942 * NOTE: these optimizations below are related with pointer comparison
4943 * which will never be JMP32.
4945 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
4946 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
4947 reg_type_may_be_null(dst_reg->type)) {
4948 /* Mark all identical registers in each branch as either
4949 * safe or unknown depending R == 0 or R != 0 conditional.
4951 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
4953 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
4955 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
4956 this_branch, other_branch) &&
4957 is_pointer_value(env, insn->dst_reg)) {
4958 verbose(env, "R%d pointer comparison prohibited\n",
4963 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
4967 /* return the map pointer stored inside BPF_LD_IMM64 instruction */
4968 static struct bpf_map *ld_imm64_to_map_ptr(struct bpf_insn *insn)
4970 u64 imm64 = ((u64) (u32) insn[0].imm) | ((u64) (u32) insn[1].imm) << 32;
4972 return (struct bpf_map *) (unsigned long) imm64;
4975 /* verify BPF_LD_IMM64 instruction */
4976 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
4978 struct bpf_reg_state *regs = cur_regs(env);
4981 if (BPF_SIZE(insn->code) != BPF_DW) {
4982 verbose(env, "invalid BPF_LD_IMM insn\n");
4985 if (insn->off != 0) {
4986 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
4990 err = check_reg_arg(env, insn->dst_reg, DST_OP);
4994 if (insn->src_reg == 0) {
4995 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
4997 regs[insn->dst_reg].type = SCALAR_VALUE;
4998 __mark_reg_known(®s[insn->dst_reg], imm);
5002 /* replace_map_fd_with_map_ptr() should have caught bad ld_imm64 */
5003 BUG_ON(insn->src_reg != BPF_PSEUDO_MAP_FD);
5005 regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
5006 regs[insn->dst_reg].map_ptr = ld_imm64_to_map_ptr(insn);
5010 static bool may_access_skb(enum bpf_prog_type type)
5013 case BPF_PROG_TYPE_SOCKET_FILTER:
5014 case BPF_PROG_TYPE_SCHED_CLS:
5015 case BPF_PROG_TYPE_SCHED_ACT:
5022 /* verify safety of LD_ABS|LD_IND instructions:
5023 * - they can only appear in the programs where ctx == skb
5024 * - since they are wrappers of function calls, they scratch R1-R5 registers,
5025 * preserve R6-R9, and store return value into R0
5028 * ctx == skb == R6 == CTX
5031 * SRC == any register
5032 * IMM == 32-bit immediate
5035 * R0 - 8/16/32-bit skb data converted to cpu endianness
5037 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
5039 struct bpf_reg_state *regs = cur_regs(env);
5040 u8 mode = BPF_MODE(insn->code);
5043 if (!may_access_skb(env->prog->type)) {
5044 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
5048 if (!env->ops->gen_ld_abs) {
5049 verbose(env, "bpf verifier is misconfigured\n");
5053 if (env->subprog_cnt > 1) {
5054 /* when program has LD_ABS insn JITs and interpreter assume
5055 * that r1 == ctx == skb which is not the case for callees
5056 * that can have arbitrary arguments. It's problematic
5057 * for main prog as well since JITs would need to analyze
5058 * all functions in order to make proper register save/restore
5059 * decisions in the main prog. Hence disallow LD_ABS with calls
5061 verbose(env, "BPF_LD_[ABS|IND] instructions cannot be mixed with bpf-to-bpf calls\n");
5065 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
5066 BPF_SIZE(insn->code) == BPF_DW ||
5067 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
5068 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
5072 /* check whether implicit source operand (register R6) is readable */
5073 err = check_reg_arg(env, BPF_REG_6, SRC_OP);
5077 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
5078 * gen_ld_abs() may terminate the program at runtime, leading to
5081 err = check_reference_leak(env);
5083 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
5087 if (env->cur_state->active_spin_lock) {
5088 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
5092 if (regs[BPF_REG_6].type != PTR_TO_CTX) {
5094 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
5098 if (mode == BPF_IND) {
5099 /* check explicit source operand */
5100 err = check_reg_arg(env, insn->src_reg, SRC_OP);
5105 /* reset caller saved regs to unreadable */
5106 for (i = 0; i < CALLER_SAVED_REGS; i++) {
5107 mark_reg_not_init(env, regs, caller_saved[i]);
5108 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
5111 /* mark destination R0 register as readable, since it contains
5112 * the value fetched from the packet.
5113 * Already marked as written above.
5115 mark_reg_unknown(env, regs, BPF_REG_0);
5119 static int check_return_code(struct bpf_verifier_env *env)
5121 struct bpf_reg_state *reg;
5122 struct tnum range = tnum_range(0, 1);
5124 switch (env->prog->type) {
5125 case BPF_PROG_TYPE_CGROUP_SKB:
5126 case BPF_PROG_TYPE_CGROUP_SOCK:
5127 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
5128 case BPF_PROG_TYPE_SOCK_OPS:
5129 case BPF_PROG_TYPE_CGROUP_DEVICE:
5135 reg = cur_regs(env) + BPF_REG_0;
5136 if (reg->type != SCALAR_VALUE) {
5137 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
5138 reg_type_str[reg->type]);
5142 if (!tnum_in(range, reg->var_off)) {
5143 verbose(env, "At program exit the register R0 ");
5144 if (!tnum_is_unknown(reg->var_off)) {
5147 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5148 verbose(env, "has value %s", tn_buf);
5150 verbose(env, "has unknown scalar value");
5152 verbose(env, " should have been 0 or 1\n");
5158 /* non-recursive DFS pseudo code
5159 * 1 procedure DFS-iterative(G,v):
5160 * 2 label v as discovered
5161 * 3 let S be a stack
5163 * 5 while S is not empty
5165 * 7 if t is what we're looking for:
5167 * 9 for all edges e in G.adjacentEdges(t) do
5168 * 10 if edge e is already labelled
5169 * 11 continue with the next edge
5170 * 12 w <- G.adjacentVertex(t,e)
5171 * 13 if vertex w is not discovered and not explored
5172 * 14 label e as tree-edge
5173 * 15 label w as discovered
5176 * 18 else if vertex w is discovered
5177 * 19 label e as back-edge
5179 * 21 // vertex w is explored
5180 * 22 label e as forward- or cross-edge
5181 * 23 label t as explored
5186 * 0x11 - discovered and fall-through edge labelled
5187 * 0x12 - discovered and fall-through and branch edges labelled
5198 #define STATE_LIST_MARK ((struct bpf_verifier_state_list *) -1L)
5200 static int *insn_stack; /* stack of insns to process */
5201 static int cur_stack; /* current stack index */
5202 static int *insn_state;
5204 /* t, w, e - match pseudo-code above:
5205 * t - index of current instruction
5206 * w - next instruction
5209 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env)
5211 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
5214 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
5217 if (w < 0 || w >= env->prog->len) {
5218 verbose_linfo(env, t, "%d: ", t);
5219 verbose(env, "jump out of range from insn %d to %d\n", t, w);
5224 /* mark branch target for state pruning */
5225 env->explored_states[w] = STATE_LIST_MARK;
5227 if (insn_state[w] == 0) {
5229 insn_state[t] = DISCOVERED | e;
5230 insn_state[w] = DISCOVERED;
5231 if (cur_stack >= env->prog->len)
5233 insn_stack[cur_stack++] = w;
5235 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
5236 verbose_linfo(env, t, "%d: ", t);
5237 verbose_linfo(env, w, "%d: ", w);
5238 verbose(env, "back-edge from insn %d to %d\n", t, w);
5240 } else if (insn_state[w] == EXPLORED) {
5241 /* forward- or cross-edge */
5242 insn_state[t] = DISCOVERED | e;
5244 verbose(env, "insn state internal bug\n");
5250 /* non-recursive depth-first-search to detect loops in BPF program
5251 * loop == back-edge in directed graph
5253 static int check_cfg(struct bpf_verifier_env *env)
5255 struct bpf_insn *insns = env->prog->insnsi;
5256 int insn_cnt = env->prog->len;
5260 insn_state = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
5264 insn_stack = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
5270 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
5271 insn_stack[0] = 0; /* 0 is the first instruction */
5277 t = insn_stack[cur_stack - 1];
5279 if (BPF_CLASS(insns[t].code) == BPF_JMP ||
5280 BPF_CLASS(insns[t].code) == BPF_JMP32) {
5281 u8 opcode = BPF_OP(insns[t].code);
5283 if (opcode == BPF_EXIT) {
5285 } else if (opcode == BPF_CALL) {
5286 ret = push_insn(t, t + 1, FALLTHROUGH, env);
5291 if (t + 1 < insn_cnt)
5292 env->explored_states[t + 1] = STATE_LIST_MARK;
5293 if (insns[t].src_reg == BPF_PSEUDO_CALL) {
5294 env->explored_states[t] = STATE_LIST_MARK;
5295 ret = push_insn(t, t + insns[t].imm + 1, BRANCH, env);
5301 } else if (opcode == BPF_JA) {
5302 if (BPF_SRC(insns[t].code) != BPF_K) {
5306 /* unconditional jump with single edge */
5307 ret = push_insn(t, t + insns[t].off + 1,
5313 /* tell verifier to check for equivalent states
5314 * after every call and jump
5316 if (t + 1 < insn_cnt)
5317 env->explored_states[t + 1] = STATE_LIST_MARK;
5319 /* conditional jump with two edges */
5320 env->explored_states[t] = STATE_LIST_MARK;
5321 ret = push_insn(t, t + 1, FALLTHROUGH, env);
5327 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env);
5334 /* all other non-branch instructions with single
5337 ret = push_insn(t, t + 1, FALLTHROUGH, env);
5345 insn_state[t] = EXPLORED;
5346 if (cur_stack-- <= 0) {
5347 verbose(env, "pop stack internal bug\n");
5354 for (i = 0; i < insn_cnt; i++) {
5355 if (insn_state[i] != EXPLORED) {
5356 verbose(env, "unreachable insn %d\n", i);
5361 ret = 0; /* cfg looks good */
5369 /* The minimum supported BTF func info size */
5370 #define MIN_BPF_FUNCINFO_SIZE 8
5371 #define MAX_FUNCINFO_REC_SIZE 252
5373 static int check_btf_func(struct bpf_verifier_env *env,
5374 const union bpf_attr *attr,
5375 union bpf_attr __user *uattr)
5377 u32 i, nfuncs, urec_size, min_size;
5378 u32 krec_size = sizeof(struct bpf_func_info);
5379 struct bpf_func_info *krecord;
5380 const struct btf_type *type;
5381 struct bpf_prog *prog;
5382 const struct btf *btf;
5383 void __user *urecord;
5384 u32 prev_offset = 0;
5387 nfuncs = attr->func_info_cnt;
5391 if (nfuncs != env->subprog_cnt) {
5392 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
5396 urec_size = attr->func_info_rec_size;
5397 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
5398 urec_size > MAX_FUNCINFO_REC_SIZE ||
5399 urec_size % sizeof(u32)) {
5400 verbose(env, "invalid func info rec size %u\n", urec_size);
5405 btf = prog->aux->btf;
5407 urecord = u64_to_user_ptr(attr->func_info);
5408 min_size = min_t(u32, krec_size, urec_size);
5410 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
5414 for (i = 0; i < nfuncs; i++) {
5415 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
5417 if (ret == -E2BIG) {
5418 verbose(env, "nonzero tailing record in func info");
5419 /* set the size kernel expects so loader can zero
5420 * out the rest of the record.
5422 if (put_user(min_size, &uattr->func_info_rec_size))
5428 if (copy_from_user(&krecord[i], urecord, min_size)) {
5433 /* check insn_off */
5435 if (krecord[i].insn_off) {
5437 "nonzero insn_off %u for the first func info record",
5438 krecord[i].insn_off);
5442 } else if (krecord[i].insn_off <= prev_offset) {
5444 "same or smaller insn offset (%u) than previous func info record (%u)",
5445 krecord[i].insn_off, prev_offset);
5450 if (env->subprog_info[i].start != krecord[i].insn_off) {
5451 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
5457 type = btf_type_by_id(btf, krecord[i].type_id);
5458 if (!type || BTF_INFO_KIND(type->info) != BTF_KIND_FUNC) {
5459 verbose(env, "invalid type id %d in func info",
5460 krecord[i].type_id);
5465 prev_offset = krecord[i].insn_off;
5466 urecord += urec_size;
5469 prog->aux->func_info = krecord;
5470 prog->aux->func_info_cnt = nfuncs;
5478 static void adjust_btf_func(struct bpf_verifier_env *env)
5482 if (!env->prog->aux->func_info)
5485 for (i = 0; i < env->subprog_cnt; i++)
5486 env->prog->aux->func_info[i].insn_off = env->subprog_info[i].start;
5489 #define MIN_BPF_LINEINFO_SIZE (offsetof(struct bpf_line_info, line_col) + \
5490 sizeof(((struct bpf_line_info *)(0))->line_col))
5491 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
5493 static int check_btf_line(struct bpf_verifier_env *env,
5494 const union bpf_attr *attr,
5495 union bpf_attr __user *uattr)
5497 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
5498 struct bpf_subprog_info *sub;
5499 struct bpf_line_info *linfo;
5500 struct bpf_prog *prog;
5501 const struct btf *btf;
5502 void __user *ulinfo;
5505 nr_linfo = attr->line_info_cnt;
5509 rec_size = attr->line_info_rec_size;
5510 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
5511 rec_size > MAX_LINEINFO_REC_SIZE ||
5512 rec_size & (sizeof(u32) - 1))
5515 /* Need to zero it in case the userspace may
5516 * pass in a smaller bpf_line_info object.
5518 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
5519 GFP_KERNEL | __GFP_NOWARN);
5524 btf = prog->aux->btf;
5527 sub = env->subprog_info;
5528 ulinfo = u64_to_user_ptr(attr->line_info);
5529 expected_size = sizeof(struct bpf_line_info);
5530 ncopy = min_t(u32, expected_size, rec_size);
5531 for (i = 0; i < nr_linfo; i++) {
5532 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
5534 if (err == -E2BIG) {
5535 verbose(env, "nonzero tailing record in line_info");
5536 if (put_user(expected_size,
5537 &uattr->line_info_rec_size))
5543 if (copy_from_user(&linfo[i], ulinfo, ncopy)) {
5549 * Check insn_off to ensure
5550 * 1) strictly increasing AND
5551 * 2) bounded by prog->len
5553 * The linfo[0].insn_off == 0 check logically falls into
5554 * the later "missing bpf_line_info for func..." case
5555 * because the first linfo[0].insn_off must be the
5556 * first sub also and the first sub must have
5557 * subprog_info[0].start == 0.
5559 if ((i && linfo[i].insn_off <= prev_offset) ||
5560 linfo[i].insn_off >= prog->len) {
5561 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
5562 i, linfo[i].insn_off, prev_offset,
5568 if (!prog->insnsi[linfo[i].insn_off].code) {
5570 "Invalid insn code at line_info[%u].insn_off\n",
5576 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
5577 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
5578 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
5583 if (s != env->subprog_cnt) {
5584 if (linfo[i].insn_off == sub[s].start) {
5585 sub[s].linfo_idx = i;
5587 } else if (sub[s].start < linfo[i].insn_off) {
5588 verbose(env, "missing bpf_line_info for func#%u\n", s);
5594 prev_offset = linfo[i].insn_off;
5598 if (s != env->subprog_cnt) {
5599 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
5600 env->subprog_cnt - s, s);
5605 prog->aux->linfo = linfo;
5606 prog->aux->nr_linfo = nr_linfo;
5615 static int check_btf_info(struct bpf_verifier_env *env,
5616 const union bpf_attr *attr,
5617 union bpf_attr __user *uattr)
5622 if (!attr->func_info_cnt && !attr->line_info_cnt)
5625 btf = btf_get_by_fd(attr->prog_btf_fd);
5627 return PTR_ERR(btf);
5628 env->prog->aux->btf = btf;
5630 err = check_btf_func(env, attr, uattr);
5634 err = check_btf_line(env, attr, uattr);
5641 /* check %cur's range satisfies %old's */
5642 static bool range_within(struct bpf_reg_state *old,
5643 struct bpf_reg_state *cur)
5645 return old->umin_value <= cur->umin_value &&
5646 old->umax_value >= cur->umax_value &&
5647 old->smin_value <= cur->smin_value &&
5648 old->smax_value >= cur->smax_value;
5651 /* Maximum number of register states that can exist at once */
5652 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
5658 /* If in the old state two registers had the same id, then they need to have
5659 * the same id in the new state as well. But that id could be different from
5660 * the old state, so we need to track the mapping from old to new ids.
5661 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
5662 * regs with old id 5 must also have new id 9 for the new state to be safe. But
5663 * regs with a different old id could still have new id 9, we don't care about
5665 * So we look through our idmap to see if this old id has been seen before. If
5666 * so, we require the new id to match; otherwise, we add the id pair to the map.
5668 static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
5672 for (i = 0; i < ID_MAP_SIZE; i++) {
5673 if (!idmap[i].old) {
5674 /* Reached an empty slot; haven't seen this id before */
5675 idmap[i].old = old_id;
5676 idmap[i].cur = cur_id;
5679 if (idmap[i].old == old_id)
5680 return idmap[i].cur == cur_id;
5682 /* We ran out of idmap slots, which should be impossible */
5687 static void clean_func_state(struct bpf_verifier_env *env,
5688 struct bpf_func_state *st)
5690 enum bpf_reg_liveness live;
5693 for (i = 0; i < BPF_REG_FP; i++) {
5694 live = st->regs[i].live;
5695 /* liveness must not touch this register anymore */
5696 st->regs[i].live |= REG_LIVE_DONE;
5697 if (!(live & REG_LIVE_READ))
5698 /* since the register is unused, clear its state
5699 * to make further comparison simpler
5701 __mark_reg_not_init(&st->regs[i]);
5704 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
5705 live = st->stack[i].spilled_ptr.live;
5706 /* liveness must not touch this stack slot anymore */
5707 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
5708 if (!(live & REG_LIVE_READ)) {
5709 __mark_reg_not_init(&st->stack[i].spilled_ptr);
5710 for (j = 0; j < BPF_REG_SIZE; j++)
5711 st->stack[i].slot_type[j] = STACK_INVALID;
5716 static void clean_verifier_state(struct bpf_verifier_env *env,
5717 struct bpf_verifier_state *st)
5721 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
5722 /* all regs in this state in all frames were already marked */
5725 for (i = 0; i <= st->curframe; i++)
5726 clean_func_state(env, st->frame[i]);
5729 /* the parentage chains form a tree.
5730 * the verifier states are added to state lists at given insn and
5731 * pushed into state stack for future exploration.
5732 * when the verifier reaches bpf_exit insn some of the verifer states
5733 * stored in the state lists have their final liveness state already,
5734 * but a lot of states will get revised from liveness point of view when
5735 * the verifier explores other branches.
5738 * 2: if r1 == 100 goto pc+1
5741 * when the verifier reaches exit insn the register r0 in the state list of
5742 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
5743 * of insn 2 and goes exploring further. At the insn 4 it will walk the
5744 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
5746 * Since the verifier pushes the branch states as it sees them while exploring
5747 * the program the condition of walking the branch instruction for the second
5748 * time means that all states below this branch were already explored and
5749 * their final liveness markes are already propagated.
5750 * Hence when the verifier completes the search of state list in is_state_visited()
5751 * we can call this clean_live_states() function to mark all liveness states
5752 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
5754 * This function also clears the registers and stack for states that !READ
5755 * to simplify state merging.
5757 * Important note here that walking the same branch instruction in the callee
5758 * doesn't meant that the states are DONE. The verifier has to compare
5761 static void clean_live_states(struct bpf_verifier_env *env, int insn,
5762 struct bpf_verifier_state *cur)
5764 struct bpf_verifier_state_list *sl;
5767 sl = env->explored_states[insn];
5771 while (sl != STATE_LIST_MARK) {
5772 if (sl->state.curframe != cur->curframe)
5774 for (i = 0; i <= cur->curframe; i++)
5775 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
5777 clean_verifier_state(env, &sl->state);
5783 /* Returns true if (rold safe implies rcur safe) */
5784 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
5785 struct idpair *idmap)
5789 if (!(rold->live & REG_LIVE_READ))
5790 /* explored state didn't use this */
5793 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
5795 if (rold->type == PTR_TO_STACK)
5796 /* two stack pointers are equal only if they're pointing to
5797 * the same stack frame, since fp-8 in foo != fp-8 in bar
5799 return equal && rold->frameno == rcur->frameno;
5804 if (rold->type == NOT_INIT)
5805 /* explored state can't have used this */
5807 if (rcur->type == NOT_INIT)
5809 switch (rold->type) {
5811 if (rcur->type == SCALAR_VALUE) {
5812 /* new val must satisfy old val knowledge */
5813 return range_within(rold, rcur) &&
5814 tnum_in(rold->var_off, rcur->var_off);
5816 /* We're trying to use a pointer in place of a scalar.
5817 * Even if the scalar was unbounded, this could lead to
5818 * pointer leaks because scalars are allowed to leak
5819 * while pointers are not. We could make this safe in
5820 * special cases if root is calling us, but it's
5821 * probably not worth the hassle.
5825 case PTR_TO_MAP_VALUE:
5826 /* If the new min/max/var_off satisfy the old ones and
5827 * everything else matches, we are OK.
5828 * 'id' is not compared, since it's only used for maps with
5829 * bpf_spin_lock inside map element and in such cases if
5830 * the rest of the prog is valid for one map element then
5831 * it's valid for all map elements regardless of the key
5832 * used in bpf_map_lookup()
5834 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
5835 range_within(rold, rcur) &&
5836 tnum_in(rold->var_off, rcur->var_off);
5837 case PTR_TO_MAP_VALUE_OR_NULL:
5838 /* a PTR_TO_MAP_VALUE could be safe to use as a
5839 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
5840 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
5841 * checked, doing so could have affected others with the same
5842 * id, and we can't check for that because we lost the id when
5843 * we converted to a PTR_TO_MAP_VALUE.
5845 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
5847 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
5849 /* Check our ids match any regs they're supposed to */
5850 return check_ids(rold->id, rcur->id, idmap);
5851 case PTR_TO_PACKET_META:
5853 if (rcur->type != rold->type)
5855 /* We must have at least as much range as the old ptr
5856 * did, so that any accesses which were safe before are
5857 * still safe. This is true even if old range < old off,
5858 * since someone could have accessed through (ptr - k), or
5859 * even done ptr -= k in a register, to get a safe access.
5861 if (rold->range > rcur->range)
5863 /* If the offsets don't match, we can't trust our alignment;
5864 * nor can we be sure that we won't fall out of range.
5866 if (rold->off != rcur->off)
5868 /* id relations must be preserved */
5869 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
5871 /* new val must satisfy old val knowledge */
5872 return range_within(rold, rcur) &&
5873 tnum_in(rold->var_off, rcur->var_off);
5875 case CONST_PTR_TO_MAP:
5876 case PTR_TO_PACKET_END:
5877 case PTR_TO_FLOW_KEYS:
5879 case PTR_TO_SOCKET_OR_NULL:
5880 case PTR_TO_SOCK_COMMON:
5881 case PTR_TO_SOCK_COMMON_OR_NULL:
5882 case PTR_TO_TCP_SOCK:
5883 case PTR_TO_TCP_SOCK_OR_NULL:
5884 /* Only valid matches are exact, which memcmp() above
5885 * would have accepted
5888 /* Don't know what's going on, just say it's not safe */
5892 /* Shouldn't get here; if we do, say it's not safe */
5897 static bool stacksafe(struct bpf_func_state *old,
5898 struct bpf_func_state *cur,
5899 struct idpair *idmap)
5903 /* walk slots of the explored stack and ignore any additional
5904 * slots in the current stack, since explored(safe) state
5907 for (i = 0; i < old->allocated_stack; i++) {
5908 spi = i / BPF_REG_SIZE;
5910 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
5911 i += BPF_REG_SIZE - 1;
5912 /* explored state didn't use this */
5916 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
5919 /* explored stack has more populated slots than current stack
5920 * and these slots were used
5922 if (i >= cur->allocated_stack)
5925 /* if old state was safe with misc data in the stack
5926 * it will be safe with zero-initialized stack.
5927 * The opposite is not true
5929 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
5930 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
5932 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
5933 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
5934 /* Ex: old explored (safe) state has STACK_SPILL in
5935 * this stack slot, but current has has STACK_MISC ->
5936 * this verifier states are not equivalent,
5937 * return false to continue verification of this path
5940 if (i % BPF_REG_SIZE)
5942 if (old->stack[spi].slot_type[0] != STACK_SPILL)
5944 if (!regsafe(&old->stack[spi].spilled_ptr,
5945 &cur->stack[spi].spilled_ptr,
5947 /* when explored and current stack slot are both storing
5948 * spilled registers, check that stored pointers types
5949 * are the same as well.
5950 * Ex: explored safe path could have stored
5951 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
5952 * but current path has stored:
5953 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
5954 * such verifier states are not equivalent.
5955 * return false to continue verification of this path
5962 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
5964 if (old->acquired_refs != cur->acquired_refs)
5966 return !memcmp(old->refs, cur->refs,
5967 sizeof(*old->refs) * old->acquired_refs);
5970 /* compare two verifier states
5972 * all states stored in state_list are known to be valid, since
5973 * verifier reached 'bpf_exit' instruction through them
5975 * this function is called when verifier exploring different branches of
5976 * execution popped from the state stack. If it sees an old state that has
5977 * more strict register state and more strict stack state then this execution
5978 * branch doesn't need to be explored further, since verifier already
5979 * concluded that more strict state leads to valid finish.
5981 * Therefore two states are equivalent if register state is more conservative
5982 * and explored stack state is more conservative than the current one.
5985 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
5986 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
5988 * In other words if current stack state (one being explored) has more
5989 * valid slots than old one that already passed validation, it means
5990 * the verifier can stop exploring and conclude that current state is valid too
5992 * Similarly with registers. If explored state has register type as invalid
5993 * whereas register type in current state is meaningful, it means that
5994 * the current state will reach 'bpf_exit' instruction safely
5996 static bool func_states_equal(struct bpf_func_state *old,
5997 struct bpf_func_state *cur)
5999 struct idpair *idmap;
6003 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
6004 /* If we failed to allocate the idmap, just say it's not safe */
6008 for (i = 0; i < MAX_BPF_REG; i++) {
6009 if (!regsafe(&old->regs[i], &cur->regs[i], idmap))
6013 if (!stacksafe(old, cur, idmap))
6016 if (!refsafe(old, cur))
6024 static bool states_equal(struct bpf_verifier_env *env,
6025 struct bpf_verifier_state *old,
6026 struct bpf_verifier_state *cur)
6030 if (old->curframe != cur->curframe)
6033 /* Verification state from speculative execution simulation
6034 * must never prune a non-speculative execution one.
6036 if (old->speculative && !cur->speculative)
6039 if (old->active_spin_lock != cur->active_spin_lock)
6042 /* for states to be equal callsites have to be the same
6043 * and all frame states need to be equivalent
6045 for (i = 0; i <= old->curframe; i++) {
6046 if (old->frame[i]->callsite != cur->frame[i]->callsite)
6048 if (!func_states_equal(old->frame[i], cur->frame[i]))
6054 /* A write screens off any subsequent reads; but write marks come from the
6055 * straight-line code between a state and its parent. When we arrive at an
6056 * equivalent state (jump target or such) we didn't arrive by the straight-line
6057 * code, so read marks in the state must propagate to the parent regardless
6058 * of the state's write marks. That's what 'parent == state->parent' comparison
6059 * in mark_reg_read() is for.
6061 static int propagate_liveness(struct bpf_verifier_env *env,
6062 const struct bpf_verifier_state *vstate,
6063 struct bpf_verifier_state *vparent)
6065 int i, frame, err = 0;
6066 struct bpf_func_state *state, *parent;
6068 if (vparent->curframe != vstate->curframe) {
6069 WARN(1, "propagate_live: parent frame %d current frame %d\n",
6070 vparent->curframe, vstate->curframe);
6073 /* Propagate read liveness of registers... */
6074 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
6075 /* We don't need to worry about FP liveness because it's read-only */
6076 for (i = 0; i < BPF_REG_FP; i++) {
6077 if (vparent->frame[vparent->curframe]->regs[i].live & REG_LIVE_READ)
6079 if (vstate->frame[vstate->curframe]->regs[i].live & REG_LIVE_READ) {
6080 err = mark_reg_read(env, &vstate->frame[vstate->curframe]->regs[i],
6081 &vparent->frame[vstate->curframe]->regs[i]);
6087 /* ... and stack slots */
6088 for (frame = 0; frame <= vstate->curframe; frame++) {
6089 state = vstate->frame[frame];
6090 parent = vparent->frame[frame];
6091 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
6092 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
6093 if (parent->stack[i].spilled_ptr.live & REG_LIVE_READ)
6095 if (state->stack[i].spilled_ptr.live & REG_LIVE_READ)
6096 mark_reg_read(env, &state->stack[i].spilled_ptr,
6097 &parent->stack[i].spilled_ptr);
6103 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
6105 struct bpf_verifier_state_list *new_sl;
6106 struct bpf_verifier_state_list *sl;
6107 struct bpf_verifier_state *cur = env->cur_state, *new;
6108 int i, j, err, states_cnt = 0;
6110 sl = env->explored_states[insn_idx];
6112 /* this 'insn_idx' instruction wasn't marked, so we will not
6113 * be doing state search here
6117 clean_live_states(env, insn_idx, cur);
6119 while (sl != STATE_LIST_MARK) {
6120 if (states_equal(env, &sl->state, cur)) {
6121 /* reached equivalent register/stack state,
6123 * Registers read by the continuation are read by us.
6124 * If we have any write marks in env->cur_state, they
6125 * will prevent corresponding reads in the continuation
6126 * from reaching our parent (an explored_state). Our
6127 * own state will get the read marks recorded, but
6128 * they'll be immediately forgotten as we're pruning
6129 * this state and will pop a new one.
6131 err = propagate_liveness(env, &sl->state, cur);
6140 if (!env->allow_ptr_leaks && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
6143 /* there were no equivalent states, remember current one.
6144 * technically the current state is not proven to be safe yet,
6145 * but it will either reach outer most bpf_exit (which means it's safe)
6146 * or it will be rejected. Since there are no loops, we won't be
6147 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
6148 * again on the way to bpf_exit
6150 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
6154 /* add new state to the head of linked list */
6155 new = &new_sl->state;
6156 err = copy_verifier_state(new, cur);
6158 free_verifier_state(new, false);
6162 new_sl->next = env->explored_states[insn_idx];
6163 env->explored_states[insn_idx] = new_sl;
6164 /* connect new state to parentage chain. Current frame needs all
6165 * registers connected. Only r6 - r9 of the callers are alive (pushed
6166 * to the stack implicitly by JITs) so in callers' frames connect just
6167 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
6168 * the state of the call instruction (with WRITTEN set), and r0 comes
6169 * from callee with its full parentage chain, anyway.
6171 for (j = 0; j <= cur->curframe; j++)
6172 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
6173 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
6174 /* clear write marks in current state: the writes we did are not writes
6175 * our child did, so they don't screen off its reads from us.
6176 * (There are no read marks in current state, because reads always mark
6177 * their parent and current state never has children yet. Only
6178 * explored_states can get read marks.)
6180 for (i = 0; i < BPF_REG_FP; i++)
6181 cur->frame[cur->curframe]->regs[i].live = REG_LIVE_NONE;
6183 /* all stack frames are accessible from callee, clear them all */
6184 for (j = 0; j <= cur->curframe; j++) {
6185 struct bpf_func_state *frame = cur->frame[j];
6186 struct bpf_func_state *newframe = new->frame[j];
6188 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
6189 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
6190 frame->stack[i].spilled_ptr.parent =
6191 &newframe->stack[i].spilled_ptr;
6197 /* Return true if it's OK to have the same insn return a different type. */
6198 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
6203 case PTR_TO_SOCKET_OR_NULL:
6204 case PTR_TO_SOCK_COMMON:
6205 case PTR_TO_SOCK_COMMON_OR_NULL:
6206 case PTR_TO_TCP_SOCK:
6207 case PTR_TO_TCP_SOCK_OR_NULL:
6214 /* If an instruction was previously used with particular pointer types, then we
6215 * need to be careful to avoid cases such as the below, where it may be ok
6216 * for one branch accessing the pointer, but not ok for the other branch:
6221 * R1 = some_other_valid_ptr;
6224 * R2 = *(u32 *)(R1 + 0);
6226 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
6228 return src != prev && (!reg_type_mismatch_ok(src) ||
6229 !reg_type_mismatch_ok(prev));
6232 static int do_check(struct bpf_verifier_env *env)
6234 struct bpf_verifier_state *state;
6235 struct bpf_insn *insns = env->prog->insnsi;
6236 struct bpf_reg_state *regs;
6237 int insn_cnt = env->prog->len, i;
6238 int insn_processed = 0;
6239 bool do_print_state = false;
6241 env->prev_linfo = NULL;
6243 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
6246 state->curframe = 0;
6247 state->speculative = false;
6248 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
6249 if (!state->frame[0]) {
6253 env->cur_state = state;
6254 init_func_state(env, state->frame[0],
6255 BPF_MAIN_FUNC /* callsite */,
6257 0 /* subprogno, zero == main subprog */);
6260 struct bpf_insn *insn;
6264 if (env->insn_idx >= insn_cnt) {
6265 verbose(env, "invalid insn idx %d insn_cnt %d\n",
6266 env->insn_idx, insn_cnt);
6270 insn = &insns[env->insn_idx];
6271 class = BPF_CLASS(insn->code);
6273 if (++insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
6275 "BPF program is too large. Processed %d insn\n",
6280 err = is_state_visited(env, env->insn_idx);
6284 /* found equivalent state, can prune the search */
6285 if (env->log.level) {
6287 verbose(env, "\nfrom %d to %d%s: safe\n",
6288 env->prev_insn_idx, env->insn_idx,
6289 env->cur_state->speculative ?
6290 " (speculative execution)" : "");
6292 verbose(env, "%d: safe\n", env->insn_idx);
6294 goto process_bpf_exit;
6297 if (signal_pending(current))
6303 if (env->log.level > 1 || (env->log.level && do_print_state)) {
6304 if (env->log.level > 1)
6305 verbose(env, "%d:", env->insn_idx);
6307 verbose(env, "\nfrom %d to %d%s:",
6308 env->prev_insn_idx, env->insn_idx,
6309 env->cur_state->speculative ?
6310 " (speculative execution)" : "");
6311 print_verifier_state(env, state->frame[state->curframe]);
6312 do_print_state = false;
6315 if (env->log.level) {
6316 const struct bpf_insn_cbs cbs = {
6317 .cb_print = verbose,
6318 .private_data = env,
6321 verbose_linfo(env, env->insn_idx, "; ");
6322 verbose(env, "%d: ", env->insn_idx);
6323 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
6326 if (bpf_prog_is_dev_bound(env->prog->aux)) {
6327 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
6328 env->prev_insn_idx);
6333 regs = cur_regs(env);
6334 env->insn_aux_data[env->insn_idx].seen = true;
6336 if (class == BPF_ALU || class == BPF_ALU64) {
6337 err = check_alu_op(env, insn);
6341 } else if (class == BPF_LDX) {
6342 enum bpf_reg_type *prev_src_type, src_reg_type;
6344 /* check for reserved fields is already done */
6346 /* check src operand */
6347 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6351 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
6355 src_reg_type = regs[insn->src_reg].type;
6357 /* check that memory (src_reg + off) is readable,
6358 * the state of dst_reg will be updated by this func
6360 err = check_mem_access(env, env->insn_idx, insn->src_reg,
6361 insn->off, BPF_SIZE(insn->code),
6362 BPF_READ, insn->dst_reg, false);
6366 prev_src_type = &env->insn_aux_data[env->insn_idx].ptr_type;
6368 if (*prev_src_type == NOT_INIT) {
6370 * dst_reg = *(u32 *)(src_reg + off)
6371 * save type to validate intersecting paths
6373 *prev_src_type = src_reg_type;
6375 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
6376 /* ABuser program is trying to use the same insn
6377 * dst_reg = *(u32*) (src_reg + off)
6378 * with different pointer types:
6379 * src_reg == ctx in one branch and
6380 * src_reg == stack|map in some other branch.
6383 verbose(env, "same insn cannot be used with different pointers\n");
6387 } else if (class == BPF_STX) {
6388 enum bpf_reg_type *prev_dst_type, dst_reg_type;
6390 if (BPF_MODE(insn->code) == BPF_XADD) {
6391 err = check_xadd(env, env->insn_idx, insn);
6398 /* check src1 operand */
6399 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6402 /* check src2 operand */
6403 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
6407 dst_reg_type = regs[insn->dst_reg].type;
6409 /* check that memory (dst_reg + off) is writeable */
6410 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
6411 insn->off, BPF_SIZE(insn->code),
6412 BPF_WRITE, insn->src_reg, false);
6416 prev_dst_type = &env->insn_aux_data[env->insn_idx].ptr_type;
6418 if (*prev_dst_type == NOT_INIT) {
6419 *prev_dst_type = dst_reg_type;
6420 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
6421 verbose(env, "same insn cannot be used with different pointers\n");
6425 } else if (class == BPF_ST) {
6426 if (BPF_MODE(insn->code) != BPF_MEM ||
6427 insn->src_reg != BPF_REG_0) {
6428 verbose(env, "BPF_ST uses reserved fields\n");
6431 /* check src operand */
6432 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
6436 if (is_ctx_reg(env, insn->dst_reg)) {
6437 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
6439 reg_type_str[reg_state(env, insn->dst_reg)->type]);
6443 /* check that memory (dst_reg + off) is writeable */
6444 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
6445 insn->off, BPF_SIZE(insn->code),
6446 BPF_WRITE, -1, false);
6450 } else if (class == BPF_JMP || class == BPF_JMP32) {
6451 u8 opcode = BPF_OP(insn->code);
6453 if (opcode == BPF_CALL) {
6454 if (BPF_SRC(insn->code) != BPF_K ||
6456 (insn->src_reg != BPF_REG_0 &&
6457 insn->src_reg != BPF_PSEUDO_CALL) ||
6458 insn->dst_reg != BPF_REG_0 ||
6459 class == BPF_JMP32) {
6460 verbose(env, "BPF_CALL uses reserved fields\n");
6464 if (env->cur_state->active_spin_lock &&
6465 (insn->src_reg == BPF_PSEUDO_CALL ||
6466 insn->imm != BPF_FUNC_spin_unlock)) {
6467 verbose(env, "function calls are not allowed while holding a lock\n");
6470 if (insn->src_reg == BPF_PSEUDO_CALL)
6471 err = check_func_call(env, insn, &env->insn_idx);
6473 err = check_helper_call(env, insn->imm, env->insn_idx);
6477 } else if (opcode == BPF_JA) {
6478 if (BPF_SRC(insn->code) != BPF_K ||
6480 insn->src_reg != BPF_REG_0 ||
6481 insn->dst_reg != BPF_REG_0 ||
6482 class == BPF_JMP32) {
6483 verbose(env, "BPF_JA uses reserved fields\n");
6487 env->insn_idx += insn->off + 1;
6490 } else if (opcode == BPF_EXIT) {
6491 if (BPF_SRC(insn->code) != BPF_K ||
6493 insn->src_reg != BPF_REG_0 ||
6494 insn->dst_reg != BPF_REG_0 ||
6495 class == BPF_JMP32) {
6496 verbose(env, "BPF_EXIT uses reserved fields\n");
6500 if (env->cur_state->active_spin_lock) {
6501 verbose(env, "bpf_spin_unlock is missing\n");
6505 if (state->curframe) {
6506 /* exit from nested function */
6507 env->prev_insn_idx = env->insn_idx;
6508 err = prepare_func_exit(env, &env->insn_idx);
6511 do_print_state = true;
6515 err = check_reference_leak(env);
6519 /* eBPF calling convetion is such that R0 is used
6520 * to return the value from eBPF program.
6521 * Make sure that it's readable at this time
6522 * of bpf_exit, which means that program wrote
6523 * something into it earlier
6525 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
6529 if (is_pointer_value(env, BPF_REG_0)) {
6530 verbose(env, "R0 leaks addr as return value\n");
6534 err = check_return_code(env);
6538 err = pop_stack(env, &env->prev_insn_idx,
6545 do_print_state = true;
6549 err = check_cond_jmp_op(env, insn, &env->insn_idx);
6553 } else if (class == BPF_LD) {
6554 u8 mode = BPF_MODE(insn->code);
6556 if (mode == BPF_ABS || mode == BPF_IND) {
6557 err = check_ld_abs(env, insn);
6561 } else if (mode == BPF_IMM) {
6562 err = check_ld_imm(env, insn);
6567 env->insn_aux_data[env->insn_idx].seen = true;
6569 verbose(env, "invalid BPF_LD mode\n");
6573 verbose(env, "unknown insn class %d\n", class);
6580 verbose(env, "processed %d insns (limit %d), stack depth ",
6581 insn_processed, BPF_COMPLEXITY_LIMIT_INSNS);
6582 for (i = 0; i < env->subprog_cnt; i++) {
6583 u32 depth = env->subprog_info[i].stack_depth;
6585 verbose(env, "%d", depth);
6586 if (i + 1 < env->subprog_cnt)
6590 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
6594 static int check_map_prealloc(struct bpf_map *map)
6596 return (map->map_type != BPF_MAP_TYPE_HASH &&
6597 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
6598 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
6599 !(map->map_flags & BPF_F_NO_PREALLOC);
6602 static bool is_tracing_prog_type(enum bpf_prog_type type)
6605 case BPF_PROG_TYPE_KPROBE:
6606 case BPF_PROG_TYPE_TRACEPOINT:
6607 case BPF_PROG_TYPE_PERF_EVENT:
6608 case BPF_PROG_TYPE_RAW_TRACEPOINT:
6615 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
6616 struct bpf_map *map,
6617 struct bpf_prog *prog)
6620 /* Make sure that BPF_PROG_TYPE_PERF_EVENT programs only use
6621 * preallocated hash maps, since doing memory allocation
6622 * in overflow_handler can crash depending on where nmi got
6625 if (prog->type == BPF_PROG_TYPE_PERF_EVENT) {
6626 if (!check_map_prealloc(map)) {
6627 verbose(env, "perf_event programs can only use preallocated hash map\n");
6630 if (map->inner_map_meta &&
6631 !check_map_prealloc(map->inner_map_meta)) {
6632 verbose(env, "perf_event programs can only use preallocated inner hash map\n");
6637 if ((is_tracing_prog_type(prog->type) ||
6638 prog->type == BPF_PROG_TYPE_SOCKET_FILTER) &&
6639 map_value_has_spin_lock(map)) {
6640 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
6644 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
6645 !bpf_offload_prog_map_match(prog, map)) {
6646 verbose(env, "offload device mismatch between prog and map\n");
6653 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
6655 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
6656 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
6659 /* look for pseudo eBPF instructions that access map FDs and
6660 * replace them with actual map pointers
6662 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env *env)
6664 struct bpf_insn *insn = env->prog->insnsi;
6665 int insn_cnt = env->prog->len;
6668 err = bpf_prog_calc_tag(env->prog);
6672 for (i = 0; i < insn_cnt; i++, insn++) {
6673 if (BPF_CLASS(insn->code) == BPF_LDX &&
6674 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
6675 verbose(env, "BPF_LDX uses reserved fields\n");
6679 if (BPF_CLASS(insn->code) == BPF_STX &&
6680 ((BPF_MODE(insn->code) != BPF_MEM &&
6681 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
6682 verbose(env, "BPF_STX uses reserved fields\n");
6686 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
6687 struct bpf_map *map;
6690 if (i == insn_cnt - 1 || insn[1].code != 0 ||
6691 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
6693 verbose(env, "invalid bpf_ld_imm64 insn\n");
6697 if (insn->src_reg == 0)
6698 /* valid generic load 64-bit imm */
6701 if (insn[0].src_reg != BPF_PSEUDO_MAP_FD ||
6703 verbose(env, "unrecognized bpf_ld_imm64 insn\n");
6707 f = fdget(insn[0].imm);
6708 map = __bpf_map_get(f);
6710 verbose(env, "fd %d is not pointing to valid bpf_map\n",
6712 return PTR_ERR(map);
6715 err = check_map_prog_compatibility(env, map, env->prog);
6721 /* store map pointer inside BPF_LD_IMM64 instruction */
6722 insn[0].imm = (u32) (unsigned long) map;
6723 insn[1].imm = ((u64) (unsigned long) map) >> 32;
6725 /* check whether we recorded this map already */
6726 for (j = 0; j < env->used_map_cnt; j++)
6727 if (env->used_maps[j] == map) {
6732 if (env->used_map_cnt >= MAX_USED_MAPS) {
6737 /* hold the map. If the program is rejected by verifier,
6738 * the map will be released by release_maps() or it
6739 * will be used by the valid program until it's unloaded
6740 * and all maps are released in free_used_maps()
6742 map = bpf_map_inc(map, false);
6745 return PTR_ERR(map);
6747 env->used_maps[env->used_map_cnt++] = map;
6749 if (bpf_map_is_cgroup_storage(map) &&
6750 bpf_cgroup_storage_assign(env->prog, map)) {
6751 verbose(env, "only one cgroup storage of each type is allowed\n");
6763 /* Basic sanity check before we invest more work here. */
6764 if (!bpf_opcode_in_insntable(insn->code)) {
6765 verbose(env, "unknown opcode %02x\n", insn->code);
6770 /* now all pseudo BPF_LD_IMM64 instructions load valid
6771 * 'struct bpf_map *' into a register instead of user map_fd.
6772 * These pointers will be used later by verifier to validate map access.
6777 /* drop refcnt of maps used by the rejected program */
6778 static void release_maps(struct bpf_verifier_env *env)
6780 enum bpf_cgroup_storage_type stype;
6783 for_each_cgroup_storage_type(stype) {
6784 if (!env->prog->aux->cgroup_storage[stype])
6786 bpf_cgroup_storage_release(env->prog,
6787 env->prog->aux->cgroup_storage[stype]);
6790 for (i = 0; i < env->used_map_cnt; i++)
6791 bpf_map_put(env->used_maps[i]);
6794 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
6795 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
6797 struct bpf_insn *insn = env->prog->insnsi;
6798 int insn_cnt = env->prog->len;
6801 for (i = 0; i < insn_cnt; i++, insn++)
6802 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
6806 /* single env->prog->insni[off] instruction was replaced with the range
6807 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
6808 * [0, off) and [off, end) to new locations, so the patched range stays zero
6810 static int adjust_insn_aux_data(struct bpf_verifier_env *env, u32 prog_len,
6813 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
6818 new_data = vzalloc(array_size(prog_len,
6819 sizeof(struct bpf_insn_aux_data)));
6822 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
6823 memcpy(new_data + off + cnt - 1, old_data + off,
6824 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
6825 for (i = off; i < off + cnt - 1; i++)
6826 new_data[i].seen = true;
6827 env->insn_aux_data = new_data;
6832 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
6838 /* NOTE: fake 'exit' subprog should be updated as well. */
6839 for (i = 0; i <= env->subprog_cnt; i++) {
6840 if (env->subprog_info[i].start <= off)
6842 env->subprog_info[i].start += len - 1;
6846 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
6847 const struct bpf_insn *patch, u32 len)
6849 struct bpf_prog *new_prog;
6851 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
6854 if (adjust_insn_aux_data(env, new_prog->len, off, len))
6856 adjust_subprog_starts(env, off, len);
6860 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
6865 /* find first prog starting at or after off (first to remove) */
6866 for (i = 0; i < env->subprog_cnt; i++)
6867 if (env->subprog_info[i].start >= off)
6869 /* find first prog starting at or after off + cnt (first to stay) */
6870 for (j = i; j < env->subprog_cnt; j++)
6871 if (env->subprog_info[j].start >= off + cnt)
6873 /* if j doesn't start exactly at off + cnt, we are just removing
6874 * the front of previous prog
6876 if (env->subprog_info[j].start != off + cnt)
6880 struct bpf_prog_aux *aux = env->prog->aux;
6883 /* move fake 'exit' subprog as well */
6884 move = env->subprog_cnt + 1 - j;
6886 memmove(env->subprog_info + i,
6887 env->subprog_info + j,
6888 sizeof(*env->subprog_info) * move);
6889 env->subprog_cnt -= j - i;
6891 /* remove func_info */
6892 if (aux->func_info) {
6893 move = aux->func_info_cnt - j;
6895 memmove(aux->func_info + i,
6897 sizeof(*aux->func_info) * move);
6898 aux->func_info_cnt -= j - i;
6899 /* func_info->insn_off is set after all code rewrites,
6900 * in adjust_btf_func() - no need to adjust
6904 /* convert i from "first prog to remove" to "first to adjust" */
6905 if (env->subprog_info[i].start == off)
6909 /* update fake 'exit' subprog as well */
6910 for (; i <= env->subprog_cnt; i++)
6911 env->subprog_info[i].start -= cnt;
6916 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
6919 struct bpf_prog *prog = env->prog;
6920 u32 i, l_off, l_cnt, nr_linfo;
6921 struct bpf_line_info *linfo;
6923 nr_linfo = prog->aux->nr_linfo;
6927 linfo = prog->aux->linfo;
6929 /* find first line info to remove, count lines to be removed */
6930 for (i = 0; i < nr_linfo; i++)
6931 if (linfo[i].insn_off >= off)
6936 for (; i < nr_linfo; i++)
6937 if (linfo[i].insn_off < off + cnt)
6942 /* First live insn doesn't match first live linfo, it needs to "inherit"
6943 * last removed linfo. prog is already modified, so prog->len == off
6944 * means no live instructions after (tail of the program was removed).
6946 if (prog->len != off && l_cnt &&
6947 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
6949 linfo[--i].insn_off = off + cnt;
6952 /* remove the line info which refer to the removed instructions */
6954 memmove(linfo + l_off, linfo + i,
6955 sizeof(*linfo) * (nr_linfo - i));
6957 prog->aux->nr_linfo -= l_cnt;
6958 nr_linfo = prog->aux->nr_linfo;
6961 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
6962 for (i = l_off; i < nr_linfo; i++)
6963 linfo[i].insn_off -= cnt;
6965 /* fix up all subprogs (incl. 'exit') which start >= off */
6966 for (i = 0; i <= env->subprog_cnt; i++)
6967 if (env->subprog_info[i].linfo_idx > l_off) {
6968 /* program may have started in the removed region but
6969 * may not be fully removed
6971 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
6972 env->subprog_info[i].linfo_idx -= l_cnt;
6974 env->subprog_info[i].linfo_idx = l_off;
6980 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
6982 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
6983 unsigned int orig_prog_len = env->prog->len;
6986 if (bpf_prog_is_dev_bound(env->prog->aux))
6987 bpf_prog_offload_remove_insns(env, off, cnt);
6989 err = bpf_remove_insns(env->prog, off, cnt);
6993 err = adjust_subprog_starts_after_remove(env, off, cnt);
6997 err = bpf_adj_linfo_after_remove(env, off, cnt);
7001 memmove(aux_data + off, aux_data + off + cnt,
7002 sizeof(*aux_data) * (orig_prog_len - off - cnt));
7007 /* The verifier does more data flow analysis than llvm and will not
7008 * explore branches that are dead at run time. Malicious programs can
7009 * have dead code too. Therefore replace all dead at-run-time code
7012 * Just nops are not optimal, e.g. if they would sit at the end of the
7013 * program and through another bug we would manage to jump there, then
7014 * we'd execute beyond program memory otherwise. Returning exception
7015 * code also wouldn't work since we can have subprogs where the dead
7016 * code could be located.
7018 static void sanitize_dead_code(struct bpf_verifier_env *env)
7020 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
7021 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
7022 struct bpf_insn *insn = env->prog->insnsi;
7023 const int insn_cnt = env->prog->len;
7026 for (i = 0; i < insn_cnt; i++) {
7027 if (aux_data[i].seen)
7029 memcpy(insn + i, &trap, sizeof(trap));
7033 static bool insn_is_cond_jump(u8 code)
7037 if (BPF_CLASS(code) == BPF_JMP32)
7040 if (BPF_CLASS(code) != BPF_JMP)
7044 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
7047 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
7049 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
7050 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
7051 struct bpf_insn *insn = env->prog->insnsi;
7052 const int insn_cnt = env->prog->len;
7055 for (i = 0; i < insn_cnt; i++, insn++) {
7056 if (!insn_is_cond_jump(insn->code))
7059 if (!aux_data[i + 1].seen)
7061 else if (!aux_data[i + 1 + insn->off].seen)
7066 if (bpf_prog_is_dev_bound(env->prog->aux))
7067 bpf_prog_offload_replace_insn(env, i, &ja);
7069 memcpy(insn, &ja, sizeof(ja));
7073 static int opt_remove_dead_code(struct bpf_verifier_env *env)
7075 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
7076 int insn_cnt = env->prog->len;
7079 for (i = 0; i < insn_cnt; i++) {
7083 while (i + j < insn_cnt && !aux_data[i + j].seen)
7088 err = verifier_remove_insns(env, i, j);
7091 insn_cnt = env->prog->len;
7097 static int opt_remove_nops(struct bpf_verifier_env *env)
7099 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
7100 struct bpf_insn *insn = env->prog->insnsi;
7101 int insn_cnt = env->prog->len;
7104 for (i = 0; i < insn_cnt; i++) {
7105 if (memcmp(&insn[i], &ja, sizeof(ja)))
7108 err = verifier_remove_insns(env, i, 1);
7118 /* convert load instructions that access fields of a context type into a
7119 * sequence of instructions that access fields of the underlying structure:
7120 * struct __sk_buff -> struct sk_buff
7121 * struct bpf_sock_ops -> struct sock
7123 static int convert_ctx_accesses(struct bpf_verifier_env *env)
7125 const struct bpf_verifier_ops *ops = env->ops;
7126 int i, cnt, size, ctx_field_size, delta = 0;
7127 const int insn_cnt = env->prog->len;
7128 struct bpf_insn insn_buf[16], *insn;
7129 u32 target_size, size_default, off;
7130 struct bpf_prog *new_prog;
7131 enum bpf_access_type type;
7132 bool is_narrower_load;
7134 if (ops->gen_prologue || env->seen_direct_write) {
7135 if (!ops->gen_prologue) {
7136 verbose(env, "bpf verifier is misconfigured\n");
7139 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
7141 if (cnt >= ARRAY_SIZE(insn_buf)) {
7142 verbose(env, "bpf verifier is misconfigured\n");
7145 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
7149 env->prog = new_prog;
7154 if (bpf_prog_is_dev_bound(env->prog->aux))
7157 insn = env->prog->insnsi + delta;
7159 for (i = 0; i < insn_cnt; i++, insn++) {
7160 bpf_convert_ctx_access_t convert_ctx_access;
7162 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
7163 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
7164 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
7165 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
7167 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
7168 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
7169 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
7170 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
7175 if (type == BPF_WRITE &&
7176 env->insn_aux_data[i + delta].sanitize_stack_off) {
7177 struct bpf_insn patch[] = {
7178 /* Sanitize suspicious stack slot with zero.
7179 * There are no memory dependencies for this store,
7180 * since it's only using frame pointer and immediate
7183 BPF_ST_MEM(BPF_DW, BPF_REG_FP,
7184 env->insn_aux_data[i + delta].sanitize_stack_off,
7186 /* the original STX instruction will immediately
7187 * overwrite the same stack slot with appropriate value
7192 cnt = ARRAY_SIZE(patch);
7193 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
7198 env->prog = new_prog;
7199 insn = new_prog->insnsi + i + delta;
7203 switch (env->insn_aux_data[i + delta].ptr_type) {
7205 if (!ops->convert_ctx_access)
7207 convert_ctx_access = ops->convert_ctx_access;
7210 case PTR_TO_SOCK_COMMON:
7211 convert_ctx_access = bpf_sock_convert_ctx_access;
7213 case PTR_TO_TCP_SOCK:
7214 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
7220 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
7221 size = BPF_LDST_BYTES(insn);
7223 /* If the read access is a narrower load of the field,
7224 * convert to a 4/8-byte load, to minimum program type specific
7225 * convert_ctx_access changes. If conversion is successful,
7226 * we will apply proper mask to the result.
7228 is_narrower_load = size < ctx_field_size;
7229 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
7231 if (is_narrower_load) {
7234 if (type == BPF_WRITE) {
7235 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
7240 if (ctx_field_size == 4)
7242 else if (ctx_field_size == 8)
7245 insn->off = off & ~(size_default - 1);
7246 insn->code = BPF_LDX | BPF_MEM | size_code;
7250 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
7252 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
7253 (ctx_field_size && !target_size)) {
7254 verbose(env, "bpf verifier is misconfigured\n");
7258 if (is_narrower_load && size < target_size) {
7259 u8 shift = (off & (size_default - 1)) * 8;
7261 if (ctx_field_size <= 4) {
7263 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
7266 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
7267 (1 << size * 8) - 1);
7270 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
7273 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
7274 (1 << size * 8) - 1);
7278 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
7284 /* keep walking new program and skip insns we just inserted */
7285 env->prog = new_prog;
7286 insn = new_prog->insnsi + i + delta;
7292 static int jit_subprogs(struct bpf_verifier_env *env)
7294 struct bpf_prog *prog = env->prog, **func, *tmp;
7295 int i, j, subprog_start, subprog_end = 0, len, subprog;
7296 struct bpf_insn *insn;
7300 if (env->subprog_cnt <= 1)
7303 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
7304 if (insn->code != (BPF_JMP | BPF_CALL) ||
7305 insn->src_reg != BPF_PSEUDO_CALL)
7307 /* Upon error here we cannot fall back to interpreter but
7308 * need a hard reject of the program. Thus -EFAULT is
7309 * propagated in any case.
7311 subprog = find_subprog(env, i + insn->imm + 1);
7313 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
7317 /* temporarily remember subprog id inside insn instead of
7318 * aux_data, since next loop will split up all insns into funcs
7320 insn->off = subprog;
7321 /* remember original imm in case JIT fails and fallback
7322 * to interpreter will be needed
7324 env->insn_aux_data[i].call_imm = insn->imm;
7325 /* point imm to __bpf_call_base+1 from JITs point of view */
7329 err = bpf_prog_alloc_jited_linfo(prog);
7334 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
7338 for (i = 0; i < env->subprog_cnt; i++) {
7339 subprog_start = subprog_end;
7340 subprog_end = env->subprog_info[i + 1].start;
7342 len = subprog_end - subprog_start;
7343 /* BPF_PROG_RUN doesn't call subprogs directly,
7344 * hence main prog stats include the runtime of subprogs.
7345 * subprogs don't have IDs and not reachable via prog_get_next_id
7346 * func[i]->aux->stats will never be accessed and stays NULL
7348 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
7351 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
7352 len * sizeof(struct bpf_insn));
7353 func[i]->type = prog->type;
7355 if (bpf_prog_calc_tag(func[i]))
7357 func[i]->is_func = 1;
7358 func[i]->aux->func_idx = i;
7359 /* the btf and func_info will be freed only at prog->aux */
7360 func[i]->aux->btf = prog->aux->btf;
7361 func[i]->aux->func_info = prog->aux->func_info;
7363 /* Use bpf_prog_F_tag to indicate functions in stack traces.
7364 * Long term would need debug info to populate names
7366 func[i]->aux->name[0] = 'F';
7367 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
7368 func[i]->jit_requested = 1;
7369 func[i]->aux->linfo = prog->aux->linfo;
7370 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
7371 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
7372 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
7373 func[i] = bpf_int_jit_compile(func[i]);
7374 if (!func[i]->jited) {
7380 /* at this point all bpf functions were successfully JITed
7381 * now populate all bpf_calls with correct addresses and
7382 * run last pass of JIT
7384 for (i = 0; i < env->subprog_cnt; i++) {
7385 insn = func[i]->insnsi;
7386 for (j = 0; j < func[i]->len; j++, insn++) {
7387 if (insn->code != (BPF_JMP | BPF_CALL) ||
7388 insn->src_reg != BPF_PSEUDO_CALL)
7390 subprog = insn->off;
7391 insn->imm = (u64 (*)(u64, u64, u64, u64, u64))
7392 func[subprog]->bpf_func -
7396 /* we use the aux data to keep a list of the start addresses
7397 * of the JITed images for each function in the program
7399 * for some architectures, such as powerpc64, the imm field
7400 * might not be large enough to hold the offset of the start
7401 * address of the callee's JITed image from __bpf_call_base
7403 * in such cases, we can lookup the start address of a callee
7404 * by using its subprog id, available from the off field of
7405 * the call instruction, as an index for this list
7407 func[i]->aux->func = func;
7408 func[i]->aux->func_cnt = env->subprog_cnt;
7410 for (i = 0; i < env->subprog_cnt; i++) {
7411 old_bpf_func = func[i]->bpf_func;
7412 tmp = bpf_int_jit_compile(func[i]);
7413 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
7414 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
7421 /* finally lock prog and jit images for all functions and
7424 for (i = 0; i < env->subprog_cnt; i++) {
7425 bpf_prog_lock_ro(func[i]);
7426 bpf_prog_kallsyms_add(func[i]);
7429 /* Last step: make now unused interpreter insns from main
7430 * prog consistent for later dump requests, so they can
7431 * later look the same as if they were interpreted only.
7433 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
7434 if (insn->code != (BPF_JMP | BPF_CALL) ||
7435 insn->src_reg != BPF_PSEUDO_CALL)
7437 insn->off = env->insn_aux_data[i].call_imm;
7438 subprog = find_subprog(env, i + insn->off + 1);
7439 insn->imm = subprog;
7443 prog->bpf_func = func[0]->bpf_func;
7444 prog->aux->func = func;
7445 prog->aux->func_cnt = env->subprog_cnt;
7446 bpf_prog_free_unused_jited_linfo(prog);
7449 for (i = 0; i < env->subprog_cnt; i++)
7451 bpf_jit_free(func[i]);
7454 /* cleanup main prog to be interpreted */
7455 prog->jit_requested = 0;
7456 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
7457 if (insn->code != (BPF_JMP | BPF_CALL) ||
7458 insn->src_reg != BPF_PSEUDO_CALL)
7461 insn->imm = env->insn_aux_data[i].call_imm;
7463 bpf_prog_free_jited_linfo(prog);
7467 static int fixup_call_args(struct bpf_verifier_env *env)
7469 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
7470 struct bpf_prog *prog = env->prog;
7471 struct bpf_insn *insn = prog->insnsi;
7476 if (env->prog->jit_requested &&
7477 !bpf_prog_is_dev_bound(env->prog->aux)) {
7478 err = jit_subprogs(env);
7484 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
7485 for (i = 0; i < prog->len; i++, insn++) {
7486 if (insn->code != (BPF_JMP | BPF_CALL) ||
7487 insn->src_reg != BPF_PSEUDO_CALL)
7489 depth = get_callee_stack_depth(env, insn, i);
7492 bpf_patch_call_args(insn, depth);
7499 /* fixup insn->imm field of bpf_call instructions
7500 * and inline eligible helpers as explicit sequence of BPF instructions
7502 * this function is called after eBPF program passed verification
7504 static int fixup_bpf_calls(struct bpf_verifier_env *env)
7506 struct bpf_prog *prog = env->prog;
7507 struct bpf_insn *insn = prog->insnsi;
7508 const struct bpf_func_proto *fn;
7509 const int insn_cnt = prog->len;
7510 const struct bpf_map_ops *ops;
7511 struct bpf_insn_aux_data *aux;
7512 struct bpf_insn insn_buf[16];
7513 struct bpf_prog *new_prog;
7514 struct bpf_map *map_ptr;
7515 int i, cnt, delta = 0;
7517 for (i = 0; i < insn_cnt; i++, insn++) {
7518 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
7519 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
7520 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
7521 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
7522 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
7523 struct bpf_insn mask_and_div[] = {
7524 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
7526 BPF_JMP_IMM(BPF_JNE, insn->src_reg, 0, 2),
7527 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
7528 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
7531 struct bpf_insn mask_and_mod[] = {
7532 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
7533 /* Rx mod 0 -> Rx */
7534 BPF_JMP_IMM(BPF_JEQ, insn->src_reg, 0, 1),
7537 struct bpf_insn *patchlet;
7539 if (insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
7540 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
7541 patchlet = mask_and_div + (is64 ? 1 : 0);
7542 cnt = ARRAY_SIZE(mask_and_div) - (is64 ? 1 : 0);
7544 patchlet = mask_and_mod + (is64 ? 1 : 0);
7545 cnt = ARRAY_SIZE(mask_and_mod) - (is64 ? 1 : 0);
7548 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
7553 env->prog = prog = new_prog;
7554 insn = new_prog->insnsi + i + delta;
7558 if (BPF_CLASS(insn->code) == BPF_LD &&
7559 (BPF_MODE(insn->code) == BPF_ABS ||
7560 BPF_MODE(insn->code) == BPF_IND)) {
7561 cnt = env->ops->gen_ld_abs(insn, insn_buf);
7562 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
7563 verbose(env, "bpf verifier is misconfigured\n");
7567 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
7572 env->prog = prog = new_prog;
7573 insn = new_prog->insnsi + i + delta;
7577 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
7578 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
7579 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
7580 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
7581 struct bpf_insn insn_buf[16];
7582 struct bpf_insn *patch = &insn_buf[0];
7586 aux = &env->insn_aux_data[i + delta];
7587 if (!aux->alu_state ||
7588 aux->alu_state == BPF_ALU_NON_POINTER)
7591 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
7592 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
7593 BPF_ALU_SANITIZE_SRC;
7595 off_reg = issrc ? insn->src_reg : insn->dst_reg;
7597 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
7598 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit - 1);
7599 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
7600 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
7601 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
7602 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
7604 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX,
7606 insn->src_reg = BPF_REG_AX;
7608 *patch++ = BPF_ALU64_REG(BPF_AND, off_reg,
7612 insn->code = insn->code == code_add ?
7613 code_sub : code_add;
7616 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
7617 cnt = patch - insn_buf;
7619 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
7624 env->prog = prog = new_prog;
7625 insn = new_prog->insnsi + i + delta;
7629 if (insn->code != (BPF_JMP | BPF_CALL))
7631 if (insn->src_reg == BPF_PSEUDO_CALL)
7634 if (insn->imm == BPF_FUNC_get_route_realm)
7635 prog->dst_needed = 1;
7636 if (insn->imm == BPF_FUNC_get_prandom_u32)
7637 bpf_user_rnd_init_once();
7638 if (insn->imm == BPF_FUNC_override_return)
7639 prog->kprobe_override = 1;
7640 if (insn->imm == BPF_FUNC_tail_call) {
7641 /* If we tail call into other programs, we
7642 * cannot make any assumptions since they can
7643 * be replaced dynamically during runtime in
7644 * the program array.
7646 prog->cb_access = 1;
7647 env->prog->aux->stack_depth = MAX_BPF_STACK;
7648 env->prog->aux->max_pkt_offset = MAX_PACKET_OFF;
7650 /* mark bpf_tail_call as different opcode to avoid
7651 * conditional branch in the interpeter for every normal
7652 * call and to prevent accidental JITing by JIT compiler
7653 * that doesn't support bpf_tail_call yet
7656 insn->code = BPF_JMP | BPF_TAIL_CALL;
7658 aux = &env->insn_aux_data[i + delta];
7659 if (!bpf_map_ptr_unpriv(aux))
7662 /* instead of changing every JIT dealing with tail_call
7663 * emit two extra insns:
7664 * if (index >= max_entries) goto out;
7665 * index &= array->index_mask;
7666 * to avoid out-of-bounds cpu speculation
7668 if (bpf_map_ptr_poisoned(aux)) {
7669 verbose(env, "tail_call abusing map_ptr\n");
7673 map_ptr = BPF_MAP_PTR(aux->map_state);
7674 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
7675 map_ptr->max_entries, 2);
7676 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
7677 container_of(map_ptr,
7680 insn_buf[2] = *insn;
7682 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
7687 env->prog = prog = new_prog;
7688 insn = new_prog->insnsi + i + delta;
7692 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
7693 * and other inlining handlers are currently limited to 64 bit
7696 if (prog->jit_requested && BITS_PER_LONG == 64 &&
7697 (insn->imm == BPF_FUNC_map_lookup_elem ||
7698 insn->imm == BPF_FUNC_map_update_elem ||
7699 insn->imm == BPF_FUNC_map_delete_elem ||
7700 insn->imm == BPF_FUNC_map_push_elem ||
7701 insn->imm == BPF_FUNC_map_pop_elem ||
7702 insn->imm == BPF_FUNC_map_peek_elem)) {
7703 aux = &env->insn_aux_data[i + delta];
7704 if (bpf_map_ptr_poisoned(aux))
7705 goto patch_call_imm;
7707 map_ptr = BPF_MAP_PTR(aux->map_state);
7709 if (insn->imm == BPF_FUNC_map_lookup_elem &&
7710 ops->map_gen_lookup) {
7711 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
7712 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
7713 verbose(env, "bpf verifier is misconfigured\n");
7717 new_prog = bpf_patch_insn_data(env, i + delta,
7723 env->prog = prog = new_prog;
7724 insn = new_prog->insnsi + i + delta;
7728 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
7729 (void *(*)(struct bpf_map *map, void *key))NULL));
7730 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
7731 (int (*)(struct bpf_map *map, void *key))NULL));
7732 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
7733 (int (*)(struct bpf_map *map, void *key, void *value,
7735 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
7736 (int (*)(struct bpf_map *map, void *value,
7738 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
7739 (int (*)(struct bpf_map *map, void *value))NULL));
7740 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
7741 (int (*)(struct bpf_map *map, void *value))NULL));
7743 switch (insn->imm) {
7744 case BPF_FUNC_map_lookup_elem:
7745 insn->imm = BPF_CAST_CALL(ops->map_lookup_elem) -
7748 case BPF_FUNC_map_update_elem:
7749 insn->imm = BPF_CAST_CALL(ops->map_update_elem) -
7752 case BPF_FUNC_map_delete_elem:
7753 insn->imm = BPF_CAST_CALL(ops->map_delete_elem) -
7756 case BPF_FUNC_map_push_elem:
7757 insn->imm = BPF_CAST_CALL(ops->map_push_elem) -
7760 case BPF_FUNC_map_pop_elem:
7761 insn->imm = BPF_CAST_CALL(ops->map_pop_elem) -
7764 case BPF_FUNC_map_peek_elem:
7765 insn->imm = BPF_CAST_CALL(ops->map_peek_elem) -
7770 goto patch_call_imm;
7774 fn = env->ops->get_func_proto(insn->imm, env->prog);
7775 /* all functions that have prototype and verifier allowed
7776 * programs to call them, must be real in-kernel functions
7780 "kernel subsystem misconfigured func %s#%d\n",
7781 func_id_name(insn->imm), insn->imm);
7784 insn->imm = fn->func - __bpf_call_base;
7790 static void free_states(struct bpf_verifier_env *env)
7792 struct bpf_verifier_state_list *sl, *sln;
7795 if (!env->explored_states)
7798 for (i = 0; i < env->prog->len; i++) {
7799 sl = env->explored_states[i];
7802 while (sl != STATE_LIST_MARK) {
7804 free_verifier_state(&sl->state, false);
7810 kfree(env->explored_states);
7813 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr,
7814 union bpf_attr __user *uattr)
7816 struct bpf_verifier_env *env;
7817 struct bpf_verifier_log *log;
7818 int i, len, ret = -EINVAL;
7821 /* no program is valid */
7822 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
7825 /* 'struct bpf_verifier_env' can be global, but since it's not small,
7826 * allocate/free it every time bpf_check() is called
7828 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
7834 env->insn_aux_data =
7835 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
7837 if (!env->insn_aux_data)
7839 for (i = 0; i < len; i++)
7840 env->insn_aux_data[i].orig_idx = i;
7842 env->ops = bpf_verifier_ops[env->prog->type];
7844 /* grab the mutex to protect few globals used by verifier */
7845 mutex_lock(&bpf_verifier_lock);
7847 if (attr->log_level || attr->log_buf || attr->log_size) {
7848 /* user requested verbose verifier output
7849 * and supplied buffer to store the verification trace
7851 log->level = attr->log_level;
7852 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
7853 log->len_total = attr->log_size;
7856 /* log attributes have to be sane */
7857 if (log->len_total < 128 || log->len_total > UINT_MAX >> 8 ||
7858 !log->level || !log->ubuf)
7862 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
7863 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
7864 env->strict_alignment = true;
7865 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
7866 env->strict_alignment = false;
7868 is_priv = capable(CAP_SYS_ADMIN);
7869 env->allow_ptr_leaks = is_priv;
7871 ret = replace_map_fd_with_map_ptr(env);
7873 goto skip_full_check;
7875 if (bpf_prog_is_dev_bound(env->prog->aux)) {
7876 ret = bpf_prog_offload_verifier_prep(env->prog);
7878 goto skip_full_check;
7881 env->explored_states = kcalloc(env->prog->len,
7882 sizeof(struct bpf_verifier_state_list *),
7885 if (!env->explored_states)
7886 goto skip_full_check;
7888 ret = check_subprogs(env);
7890 goto skip_full_check;
7892 ret = check_btf_info(env, attr, uattr);
7894 goto skip_full_check;
7896 ret = check_cfg(env);
7898 goto skip_full_check;
7900 ret = do_check(env);
7901 if (env->cur_state) {
7902 free_verifier_state(env->cur_state, true);
7903 env->cur_state = NULL;
7906 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
7907 ret = bpf_prog_offload_finalize(env);
7910 while (!pop_stack(env, NULL, NULL));
7914 ret = check_max_stack_depth(env);
7916 /* instruction rewrites happen after this point */
7919 opt_hard_wire_dead_code_branches(env);
7921 ret = opt_remove_dead_code(env);
7923 ret = opt_remove_nops(env);
7926 sanitize_dead_code(env);
7930 /* program is valid, convert *(u32*)(ctx + off) accesses */
7931 ret = convert_ctx_accesses(env);
7934 ret = fixup_bpf_calls(env);
7937 ret = fixup_call_args(env);
7939 if (log->level && bpf_verifier_log_full(log))
7941 if (log->level && !log->ubuf) {
7943 goto err_release_maps;
7946 if (ret == 0 && env->used_map_cnt) {
7947 /* if program passed verifier, update used_maps in bpf_prog_info */
7948 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
7949 sizeof(env->used_maps[0]),
7952 if (!env->prog->aux->used_maps) {
7954 goto err_release_maps;
7957 memcpy(env->prog->aux->used_maps, env->used_maps,
7958 sizeof(env->used_maps[0]) * env->used_map_cnt);
7959 env->prog->aux->used_map_cnt = env->used_map_cnt;
7961 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
7962 * bpf_ld_imm64 instructions
7964 convert_pseudo_ld_imm64(env);
7968 adjust_btf_func(env);
7971 if (!env->prog->aux->used_maps)
7972 /* if we didn't copy map pointers into bpf_prog_info, release
7973 * them now. Otherwise free_used_maps() will release them.
7978 mutex_unlock(&bpf_verifier_lock);
7979 vfree(env->insn_aux_data);