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 reg_type_may_be_refcounted_or_null(enum bpf_reg_type type)
357 return type == PTR_TO_SOCKET ||
358 type == PTR_TO_SOCKET_OR_NULL ||
359 type == PTR_TO_TCP_SOCK ||
360 type == PTR_TO_TCP_SOCK_OR_NULL;
363 static bool arg_type_may_be_refcounted(enum bpf_arg_type type)
365 return type == ARG_PTR_TO_SOCK_COMMON;
368 /* Determine whether the function releases some resources allocated by another
369 * function call. The first reference type argument will be assumed to be
370 * released by release_reference().
372 static bool is_release_function(enum bpf_func_id func_id)
374 return func_id == BPF_FUNC_sk_release;
377 static bool is_acquire_function(enum bpf_func_id func_id)
379 return func_id == BPF_FUNC_sk_lookup_tcp ||
380 func_id == BPF_FUNC_sk_lookup_udp;
383 static bool is_ptr_cast_function(enum bpf_func_id func_id)
385 return func_id == BPF_FUNC_tcp_sock ||
386 func_id == BPF_FUNC_sk_fullsock;
389 /* string representation of 'enum bpf_reg_type' */
390 static const char * const reg_type_str[] = {
392 [SCALAR_VALUE] = "inv",
393 [PTR_TO_CTX] = "ctx",
394 [CONST_PTR_TO_MAP] = "map_ptr",
395 [PTR_TO_MAP_VALUE] = "map_value",
396 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
397 [PTR_TO_STACK] = "fp",
398 [PTR_TO_PACKET] = "pkt",
399 [PTR_TO_PACKET_META] = "pkt_meta",
400 [PTR_TO_PACKET_END] = "pkt_end",
401 [PTR_TO_FLOW_KEYS] = "flow_keys",
402 [PTR_TO_SOCKET] = "sock",
403 [PTR_TO_SOCKET_OR_NULL] = "sock_or_null",
404 [PTR_TO_SOCK_COMMON] = "sock_common",
405 [PTR_TO_SOCK_COMMON_OR_NULL] = "sock_common_or_null",
406 [PTR_TO_TCP_SOCK] = "tcp_sock",
407 [PTR_TO_TCP_SOCK_OR_NULL] = "tcp_sock_or_null",
410 static char slot_type_char[] = {
411 [STACK_INVALID] = '?',
417 static void print_liveness(struct bpf_verifier_env *env,
418 enum bpf_reg_liveness live)
420 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
422 if (live & REG_LIVE_READ)
424 if (live & REG_LIVE_WRITTEN)
426 if (live & REG_LIVE_DONE)
430 static struct bpf_func_state *func(struct bpf_verifier_env *env,
431 const struct bpf_reg_state *reg)
433 struct bpf_verifier_state *cur = env->cur_state;
435 return cur->frame[reg->frameno];
438 static void print_verifier_state(struct bpf_verifier_env *env,
439 const struct bpf_func_state *state)
441 const struct bpf_reg_state *reg;
446 verbose(env, " frame%d:", state->frameno);
447 for (i = 0; i < MAX_BPF_REG; i++) {
448 reg = &state->regs[i];
452 verbose(env, " R%d", i);
453 print_liveness(env, reg->live);
454 verbose(env, "=%s", reg_type_str[t]);
455 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
456 tnum_is_const(reg->var_off)) {
457 /* reg->off should be 0 for SCALAR_VALUE */
458 verbose(env, "%lld", reg->var_off.value + reg->off);
459 if (t == PTR_TO_STACK)
460 verbose(env, ",call_%d", func(env, reg)->callsite);
462 verbose(env, "(id=%d", reg->id);
463 if (reg_type_may_be_refcounted_or_null(t))
464 verbose(env, ",ref_obj_id=%d", reg->ref_obj_id);
465 if (t != SCALAR_VALUE)
466 verbose(env, ",off=%d", reg->off);
467 if (type_is_pkt_pointer(t))
468 verbose(env, ",r=%d", reg->range);
469 else if (t == CONST_PTR_TO_MAP ||
470 t == PTR_TO_MAP_VALUE ||
471 t == PTR_TO_MAP_VALUE_OR_NULL)
472 verbose(env, ",ks=%d,vs=%d",
473 reg->map_ptr->key_size,
474 reg->map_ptr->value_size);
475 if (tnum_is_const(reg->var_off)) {
476 /* Typically an immediate SCALAR_VALUE, but
477 * could be a pointer whose offset is too big
480 verbose(env, ",imm=%llx", reg->var_off.value);
482 if (reg->smin_value != reg->umin_value &&
483 reg->smin_value != S64_MIN)
484 verbose(env, ",smin_value=%lld",
485 (long long)reg->smin_value);
486 if (reg->smax_value != reg->umax_value &&
487 reg->smax_value != S64_MAX)
488 verbose(env, ",smax_value=%lld",
489 (long long)reg->smax_value);
490 if (reg->umin_value != 0)
491 verbose(env, ",umin_value=%llu",
492 (unsigned long long)reg->umin_value);
493 if (reg->umax_value != U64_MAX)
494 verbose(env, ",umax_value=%llu",
495 (unsigned long long)reg->umax_value);
496 if (!tnum_is_unknown(reg->var_off)) {
499 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
500 verbose(env, ",var_off=%s", tn_buf);
506 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
507 char types_buf[BPF_REG_SIZE + 1];
511 for (j = 0; j < BPF_REG_SIZE; j++) {
512 if (state->stack[i].slot_type[j] != STACK_INVALID)
514 types_buf[j] = slot_type_char[
515 state->stack[i].slot_type[j]];
517 types_buf[BPF_REG_SIZE] = 0;
520 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
521 print_liveness(env, state->stack[i].spilled_ptr.live);
522 if (state->stack[i].slot_type[0] == STACK_SPILL)
524 reg_type_str[state->stack[i].spilled_ptr.type]);
526 verbose(env, "=%s", types_buf);
528 if (state->acquired_refs && state->refs[0].id) {
529 verbose(env, " refs=%d", state->refs[0].id);
530 for (i = 1; i < state->acquired_refs; i++)
531 if (state->refs[i].id)
532 verbose(env, ",%d", state->refs[i].id);
537 #define COPY_STATE_FN(NAME, COUNT, FIELD, SIZE) \
538 static int copy_##NAME##_state(struct bpf_func_state *dst, \
539 const struct bpf_func_state *src) \
543 if (WARN_ON_ONCE(dst->COUNT < src->COUNT)) { \
544 /* internal bug, make state invalid to reject the program */ \
545 memset(dst, 0, sizeof(*dst)); \
548 memcpy(dst->FIELD, src->FIELD, \
549 sizeof(*src->FIELD) * (src->COUNT / SIZE)); \
552 /* copy_reference_state() */
553 COPY_STATE_FN(reference, acquired_refs, refs, 1)
554 /* copy_stack_state() */
555 COPY_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
558 #define REALLOC_STATE_FN(NAME, COUNT, FIELD, SIZE) \
559 static int realloc_##NAME##_state(struct bpf_func_state *state, int size, \
562 u32 old_size = state->COUNT; \
563 struct bpf_##NAME##_state *new_##FIELD; \
564 int slot = size / SIZE; \
566 if (size <= old_size || !size) { \
569 state->COUNT = slot * SIZE; \
570 if (!size && old_size) { \
571 kfree(state->FIELD); \
572 state->FIELD = NULL; \
576 new_##FIELD = kmalloc_array(slot, sizeof(struct bpf_##NAME##_state), \
582 memcpy(new_##FIELD, state->FIELD, \
583 sizeof(*new_##FIELD) * (old_size / SIZE)); \
584 memset(new_##FIELD + old_size / SIZE, 0, \
585 sizeof(*new_##FIELD) * (size - old_size) / SIZE); \
587 state->COUNT = slot * SIZE; \
588 kfree(state->FIELD); \
589 state->FIELD = new_##FIELD; \
592 /* realloc_reference_state() */
593 REALLOC_STATE_FN(reference, acquired_refs, refs, 1)
594 /* realloc_stack_state() */
595 REALLOC_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
596 #undef REALLOC_STATE_FN
598 /* do_check() starts with zero-sized stack in struct bpf_verifier_state to
599 * make it consume minimal amount of memory. check_stack_write() access from
600 * the program calls into realloc_func_state() to grow the stack size.
601 * Note there is a non-zero 'parent' pointer inside bpf_verifier_state
602 * which realloc_stack_state() copies over. It points to previous
603 * bpf_verifier_state which is never reallocated.
605 static int realloc_func_state(struct bpf_func_state *state, int stack_size,
606 int refs_size, bool copy_old)
608 int err = realloc_reference_state(state, refs_size, copy_old);
611 return realloc_stack_state(state, stack_size, copy_old);
614 /* Acquire a pointer id from the env and update the state->refs to include
615 * this new pointer reference.
616 * On success, returns a valid pointer id to associate with the register
617 * On failure, returns a negative errno.
619 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
621 struct bpf_func_state *state = cur_func(env);
622 int new_ofs = state->acquired_refs;
625 err = realloc_reference_state(state, state->acquired_refs + 1, true);
629 state->refs[new_ofs].id = id;
630 state->refs[new_ofs].insn_idx = insn_idx;
635 /* release function corresponding to acquire_reference_state(). Idempotent. */
636 static int release_reference_state(struct bpf_func_state *state, int ptr_id)
640 last_idx = state->acquired_refs - 1;
641 for (i = 0; i < state->acquired_refs; i++) {
642 if (state->refs[i].id == ptr_id) {
643 if (last_idx && i != last_idx)
644 memcpy(&state->refs[i], &state->refs[last_idx],
645 sizeof(*state->refs));
646 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
647 state->acquired_refs--;
654 static int transfer_reference_state(struct bpf_func_state *dst,
655 struct bpf_func_state *src)
657 int err = realloc_reference_state(dst, src->acquired_refs, false);
660 err = copy_reference_state(dst, src);
666 static void free_func_state(struct bpf_func_state *state)
675 static void free_verifier_state(struct bpf_verifier_state *state,
680 for (i = 0; i <= state->curframe; i++) {
681 free_func_state(state->frame[i]);
682 state->frame[i] = NULL;
688 /* copy verifier state from src to dst growing dst stack space
689 * when necessary to accommodate larger src stack
691 static int copy_func_state(struct bpf_func_state *dst,
692 const struct bpf_func_state *src)
696 err = realloc_func_state(dst, src->allocated_stack, src->acquired_refs,
700 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
701 err = copy_reference_state(dst, src);
704 return copy_stack_state(dst, src);
707 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
708 const struct bpf_verifier_state *src)
710 struct bpf_func_state *dst;
713 /* if dst has more stack frames then src frame, free them */
714 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
715 free_func_state(dst_state->frame[i]);
716 dst_state->frame[i] = NULL;
718 dst_state->speculative = src->speculative;
719 dst_state->curframe = src->curframe;
720 dst_state->active_spin_lock = src->active_spin_lock;
721 for (i = 0; i <= src->curframe; i++) {
722 dst = dst_state->frame[i];
724 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
727 dst_state->frame[i] = dst;
729 err = copy_func_state(dst, src->frame[i]);
736 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
739 struct bpf_verifier_state *cur = env->cur_state;
740 struct bpf_verifier_stack_elem *elem, *head = env->head;
743 if (env->head == NULL)
747 err = copy_verifier_state(cur, &head->st);
752 *insn_idx = head->insn_idx;
754 *prev_insn_idx = head->prev_insn_idx;
756 free_verifier_state(&head->st, false);
763 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
764 int insn_idx, int prev_insn_idx,
767 struct bpf_verifier_state *cur = env->cur_state;
768 struct bpf_verifier_stack_elem *elem;
771 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
775 elem->insn_idx = insn_idx;
776 elem->prev_insn_idx = prev_insn_idx;
777 elem->next = env->head;
780 err = copy_verifier_state(&elem->st, cur);
783 elem->st.speculative |= speculative;
784 if (env->stack_size > BPF_COMPLEXITY_LIMIT_STACK) {
785 verbose(env, "BPF program is too complex\n");
790 free_verifier_state(env->cur_state, true);
791 env->cur_state = NULL;
792 /* pop all elements and return */
793 while (!pop_stack(env, NULL, NULL));
797 #define CALLER_SAVED_REGS 6
798 static const int caller_saved[CALLER_SAVED_REGS] = {
799 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
802 static void __mark_reg_not_init(struct bpf_reg_state *reg);
804 /* Mark the unknown part of a register (variable offset or scalar value) as
805 * known to have the value @imm.
807 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
809 /* Clear id, off, and union(map_ptr, range) */
810 memset(((u8 *)reg) + sizeof(reg->type), 0,
811 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
812 reg->var_off = tnum_const(imm);
813 reg->smin_value = (s64)imm;
814 reg->smax_value = (s64)imm;
815 reg->umin_value = imm;
816 reg->umax_value = imm;
819 /* Mark the 'variable offset' part of a register as zero. This should be
820 * used only on registers holding a pointer type.
822 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
824 __mark_reg_known(reg, 0);
827 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
829 __mark_reg_known(reg, 0);
830 reg->type = SCALAR_VALUE;
833 static void mark_reg_known_zero(struct bpf_verifier_env *env,
834 struct bpf_reg_state *regs, u32 regno)
836 if (WARN_ON(regno >= MAX_BPF_REG)) {
837 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
838 /* Something bad happened, let's kill all regs */
839 for (regno = 0; regno < MAX_BPF_REG; regno++)
840 __mark_reg_not_init(regs + regno);
843 __mark_reg_known_zero(regs + regno);
846 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
848 return type_is_pkt_pointer(reg->type);
851 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
853 return reg_is_pkt_pointer(reg) ||
854 reg->type == PTR_TO_PACKET_END;
857 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
858 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
859 enum bpf_reg_type which)
861 /* The register can already have a range from prior markings.
862 * This is fine as long as it hasn't been advanced from its
865 return reg->type == which &&
868 tnum_equals_const(reg->var_off, 0);
871 /* Attempts to improve min/max values based on var_off information */
872 static void __update_reg_bounds(struct bpf_reg_state *reg)
874 /* min signed is max(sign bit) | min(other bits) */
875 reg->smin_value = max_t(s64, reg->smin_value,
876 reg->var_off.value | (reg->var_off.mask & S64_MIN));
877 /* max signed is min(sign bit) | max(other bits) */
878 reg->smax_value = min_t(s64, reg->smax_value,
879 reg->var_off.value | (reg->var_off.mask & S64_MAX));
880 reg->umin_value = max(reg->umin_value, reg->var_off.value);
881 reg->umax_value = min(reg->umax_value,
882 reg->var_off.value | reg->var_off.mask);
885 /* Uses signed min/max values to inform unsigned, and vice-versa */
886 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
888 /* Learn sign from signed bounds.
889 * If we cannot cross the sign boundary, then signed and unsigned bounds
890 * are the same, so combine. This works even in the negative case, e.g.
891 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
893 if (reg->smin_value >= 0 || reg->smax_value < 0) {
894 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
896 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
900 /* Learn sign from unsigned bounds. Signed bounds cross the sign
901 * boundary, so we must be careful.
903 if ((s64)reg->umax_value >= 0) {
904 /* Positive. We can't learn anything from the smin, but smax
905 * is positive, hence safe.
907 reg->smin_value = reg->umin_value;
908 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
910 } else if ((s64)reg->umin_value < 0) {
911 /* Negative. We can't learn anything from the smax, but smin
912 * is negative, hence safe.
914 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
916 reg->smax_value = reg->umax_value;
920 /* Attempts to improve var_off based on unsigned min/max information */
921 static void __reg_bound_offset(struct bpf_reg_state *reg)
923 reg->var_off = tnum_intersect(reg->var_off,
924 tnum_range(reg->umin_value,
928 /* Reset the min/max bounds of a register */
929 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
931 reg->smin_value = S64_MIN;
932 reg->smax_value = S64_MAX;
934 reg->umax_value = U64_MAX;
937 /* Mark a register as having a completely unknown (scalar) value. */
938 static void __mark_reg_unknown(struct bpf_reg_state *reg)
941 * Clear type, id, off, and union(map_ptr, range) and
942 * padding between 'type' and union
944 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
945 reg->type = SCALAR_VALUE;
946 reg->var_off = tnum_unknown;
948 __mark_reg_unbounded(reg);
951 static void mark_reg_unknown(struct bpf_verifier_env *env,
952 struct bpf_reg_state *regs, u32 regno)
954 if (WARN_ON(regno >= MAX_BPF_REG)) {
955 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
956 /* Something bad happened, let's kill all regs except FP */
957 for (regno = 0; regno < BPF_REG_FP; regno++)
958 __mark_reg_not_init(regs + regno);
961 __mark_reg_unknown(regs + regno);
964 static void __mark_reg_not_init(struct bpf_reg_state *reg)
966 __mark_reg_unknown(reg);
967 reg->type = NOT_INIT;
970 static void mark_reg_not_init(struct bpf_verifier_env *env,
971 struct bpf_reg_state *regs, u32 regno)
973 if (WARN_ON(regno >= MAX_BPF_REG)) {
974 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
975 /* Something bad happened, let's kill all regs except FP */
976 for (regno = 0; regno < BPF_REG_FP; regno++)
977 __mark_reg_not_init(regs + regno);
980 __mark_reg_not_init(regs + regno);
983 static void init_reg_state(struct bpf_verifier_env *env,
984 struct bpf_func_state *state)
986 struct bpf_reg_state *regs = state->regs;
989 for (i = 0; i < MAX_BPF_REG; i++) {
990 mark_reg_not_init(env, regs, i);
991 regs[i].live = REG_LIVE_NONE;
992 regs[i].parent = NULL;
996 regs[BPF_REG_FP].type = PTR_TO_STACK;
997 mark_reg_known_zero(env, regs, BPF_REG_FP);
998 regs[BPF_REG_FP].frameno = state->frameno;
1000 /* 1st arg to a function */
1001 regs[BPF_REG_1].type = PTR_TO_CTX;
1002 mark_reg_known_zero(env, regs, BPF_REG_1);
1005 #define BPF_MAIN_FUNC (-1)
1006 static void init_func_state(struct bpf_verifier_env *env,
1007 struct bpf_func_state *state,
1008 int callsite, int frameno, int subprogno)
1010 state->callsite = callsite;
1011 state->frameno = frameno;
1012 state->subprogno = subprogno;
1013 init_reg_state(env, state);
1017 SRC_OP, /* register is used as source operand */
1018 DST_OP, /* register is used as destination operand */
1019 DST_OP_NO_MARK /* same as above, check only, don't mark */
1022 static int cmp_subprogs(const void *a, const void *b)
1024 return ((struct bpf_subprog_info *)a)->start -
1025 ((struct bpf_subprog_info *)b)->start;
1028 static int find_subprog(struct bpf_verifier_env *env, int off)
1030 struct bpf_subprog_info *p;
1032 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
1033 sizeof(env->subprog_info[0]), cmp_subprogs);
1036 return p - env->subprog_info;
1040 static int add_subprog(struct bpf_verifier_env *env, int off)
1042 int insn_cnt = env->prog->len;
1045 if (off >= insn_cnt || off < 0) {
1046 verbose(env, "call to invalid destination\n");
1049 ret = find_subprog(env, off);
1052 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
1053 verbose(env, "too many subprograms\n");
1056 env->subprog_info[env->subprog_cnt++].start = off;
1057 sort(env->subprog_info, env->subprog_cnt,
1058 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
1062 static int check_subprogs(struct bpf_verifier_env *env)
1064 int i, ret, subprog_start, subprog_end, off, cur_subprog = 0;
1065 struct bpf_subprog_info *subprog = env->subprog_info;
1066 struct bpf_insn *insn = env->prog->insnsi;
1067 int insn_cnt = env->prog->len;
1069 /* Add entry function. */
1070 ret = add_subprog(env, 0);
1074 /* determine subprog starts. The end is one before the next starts */
1075 for (i = 0; i < insn_cnt; i++) {
1076 if (insn[i].code != (BPF_JMP | BPF_CALL))
1078 if (insn[i].src_reg != BPF_PSEUDO_CALL)
1080 if (!env->allow_ptr_leaks) {
1081 verbose(env, "function calls to other bpf functions are allowed for root only\n");
1084 ret = add_subprog(env, i + insn[i].imm + 1);
1089 /* Add a fake 'exit' subprog which could simplify subprog iteration
1090 * logic. 'subprog_cnt' should not be increased.
1092 subprog[env->subprog_cnt].start = insn_cnt;
1094 if (env->log.level > 1)
1095 for (i = 0; i < env->subprog_cnt; i++)
1096 verbose(env, "func#%d @%d\n", i, subprog[i].start);
1098 /* now check that all jumps are within the same subprog */
1099 subprog_start = subprog[cur_subprog].start;
1100 subprog_end = subprog[cur_subprog + 1].start;
1101 for (i = 0; i < insn_cnt; i++) {
1102 u8 code = insn[i].code;
1104 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
1106 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
1108 off = i + insn[i].off + 1;
1109 if (off < subprog_start || off >= subprog_end) {
1110 verbose(env, "jump out of range from insn %d to %d\n", i, off);
1114 if (i == subprog_end - 1) {
1115 /* to avoid fall-through from one subprog into another
1116 * the last insn of the subprog should be either exit
1117 * or unconditional jump back
1119 if (code != (BPF_JMP | BPF_EXIT) &&
1120 code != (BPF_JMP | BPF_JA)) {
1121 verbose(env, "last insn is not an exit or jmp\n");
1124 subprog_start = subprog_end;
1126 if (cur_subprog < env->subprog_cnt)
1127 subprog_end = subprog[cur_subprog + 1].start;
1133 /* Parentage chain of this register (or stack slot) should take care of all
1134 * issues like callee-saved registers, stack slot allocation time, etc.
1136 static int mark_reg_read(struct bpf_verifier_env *env,
1137 const struct bpf_reg_state *state,
1138 struct bpf_reg_state *parent)
1140 bool writes = parent == state->parent; /* Observe write marks */
1143 /* if read wasn't screened by an earlier write ... */
1144 if (writes && state->live & REG_LIVE_WRITTEN)
1146 if (parent->live & REG_LIVE_DONE) {
1147 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
1148 reg_type_str[parent->type],
1149 parent->var_off.value, parent->off);
1152 /* ... then we depend on parent's value */
1153 parent->live |= REG_LIVE_READ;
1155 parent = state->parent;
1161 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
1162 enum reg_arg_type t)
1164 struct bpf_verifier_state *vstate = env->cur_state;
1165 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1166 struct bpf_reg_state *regs = state->regs;
1168 if (regno >= MAX_BPF_REG) {
1169 verbose(env, "R%d is invalid\n", regno);
1174 /* check whether register used as source operand can be read */
1175 if (regs[regno].type == NOT_INIT) {
1176 verbose(env, "R%d !read_ok\n", regno);
1179 /* We don't need to worry about FP liveness because it's read-only */
1180 if (regno != BPF_REG_FP)
1181 return mark_reg_read(env, ®s[regno],
1182 regs[regno].parent);
1184 /* check whether register used as dest operand can be written to */
1185 if (regno == BPF_REG_FP) {
1186 verbose(env, "frame pointer is read only\n");
1189 regs[regno].live |= REG_LIVE_WRITTEN;
1191 mark_reg_unknown(env, regs, regno);
1196 static bool is_spillable_regtype(enum bpf_reg_type type)
1199 case PTR_TO_MAP_VALUE:
1200 case PTR_TO_MAP_VALUE_OR_NULL:
1204 case PTR_TO_PACKET_META:
1205 case PTR_TO_PACKET_END:
1206 case PTR_TO_FLOW_KEYS:
1207 case CONST_PTR_TO_MAP:
1209 case PTR_TO_SOCKET_OR_NULL:
1210 case PTR_TO_SOCK_COMMON:
1211 case PTR_TO_SOCK_COMMON_OR_NULL:
1212 case PTR_TO_TCP_SOCK:
1213 case PTR_TO_TCP_SOCK_OR_NULL:
1220 /* Does this register contain a constant zero? */
1221 static bool register_is_null(struct bpf_reg_state *reg)
1223 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
1226 /* check_stack_read/write functions track spill/fill of registers,
1227 * stack boundary and alignment are checked in check_mem_access()
1229 static int check_stack_write(struct bpf_verifier_env *env,
1230 struct bpf_func_state *state, /* func where register points to */
1231 int off, int size, int value_regno, int insn_idx)
1233 struct bpf_func_state *cur; /* state of the current function */
1234 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
1235 enum bpf_reg_type type;
1237 err = realloc_func_state(state, round_up(slot + 1, BPF_REG_SIZE),
1238 state->acquired_refs, true);
1241 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
1242 * so it's aligned access and [off, off + size) are within stack limits
1244 if (!env->allow_ptr_leaks &&
1245 state->stack[spi].slot_type[0] == STACK_SPILL &&
1246 size != BPF_REG_SIZE) {
1247 verbose(env, "attempt to corrupt spilled pointer on stack\n");
1251 cur = env->cur_state->frame[env->cur_state->curframe];
1252 if (value_regno >= 0 &&
1253 is_spillable_regtype((type = cur->regs[value_regno].type))) {
1255 /* register containing pointer is being spilled into stack */
1256 if (size != BPF_REG_SIZE) {
1257 verbose(env, "invalid size of register spill\n");
1261 if (state != cur && type == PTR_TO_STACK) {
1262 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
1266 /* save register state */
1267 state->stack[spi].spilled_ptr = cur->regs[value_regno];
1268 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1270 for (i = 0; i < BPF_REG_SIZE; i++) {
1271 if (state->stack[spi].slot_type[i] == STACK_MISC &&
1272 !env->allow_ptr_leaks) {
1273 int *poff = &env->insn_aux_data[insn_idx].sanitize_stack_off;
1274 int soff = (-spi - 1) * BPF_REG_SIZE;
1276 /* detected reuse of integer stack slot with a pointer
1277 * which means either llvm is reusing stack slot or
1278 * an attacker is trying to exploit CVE-2018-3639
1279 * (speculative store bypass)
1280 * Have to sanitize that slot with preemptive
1283 if (*poff && *poff != soff) {
1284 /* disallow programs where single insn stores
1285 * into two different stack slots, since verifier
1286 * cannot sanitize them
1289 "insn %d cannot access two stack slots fp%d and fp%d",
1290 insn_idx, *poff, soff);
1295 state->stack[spi].slot_type[i] = STACK_SPILL;
1298 u8 type = STACK_MISC;
1300 /* regular write of data into stack destroys any spilled ptr */
1301 state->stack[spi].spilled_ptr.type = NOT_INIT;
1302 /* Mark slots as STACK_MISC if they belonged to spilled ptr. */
1303 if (state->stack[spi].slot_type[0] == STACK_SPILL)
1304 for (i = 0; i < BPF_REG_SIZE; i++)
1305 state->stack[spi].slot_type[i] = STACK_MISC;
1307 /* only mark the slot as written if all 8 bytes were written
1308 * otherwise read propagation may incorrectly stop too soon
1309 * when stack slots are partially written.
1310 * This heuristic means that read propagation will be
1311 * conservative, since it will add reg_live_read marks
1312 * to stack slots all the way to first state when programs
1313 * writes+reads less than 8 bytes
1315 if (size == BPF_REG_SIZE)
1316 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1318 /* when we zero initialize stack slots mark them as such */
1319 if (value_regno >= 0 &&
1320 register_is_null(&cur->regs[value_regno]))
1323 /* Mark slots affected by this stack write. */
1324 for (i = 0; i < size; i++)
1325 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
1331 static int check_stack_read(struct bpf_verifier_env *env,
1332 struct bpf_func_state *reg_state /* func where register points to */,
1333 int off, int size, int value_regno)
1335 struct bpf_verifier_state *vstate = env->cur_state;
1336 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1337 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
1340 if (reg_state->allocated_stack <= slot) {
1341 verbose(env, "invalid read from stack off %d+0 size %d\n",
1345 stype = reg_state->stack[spi].slot_type;
1347 if (stype[0] == STACK_SPILL) {
1348 if (size != BPF_REG_SIZE) {
1349 verbose(env, "invalid size of register spill\n");
1352 for (i = 1; i < BPF_REG_SIZE; i++) {
1353 if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) {
1354 verbose(env, "corrupted spill memory\n");
1359 if (value_regno >= 0) {
1360 /* restore register state from stack */
1361 state->regs[value_regno] = reg_state->stack[spi].spilled_ptr;
1362 /* mark reg as written since spilled pointer state likely
1363 * has its liveness marks cleared by is_state_visited()
1364 * which resets stack/reg liveness for state transitions
1366 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
1368 mark_reg_read(env, ®_state->stack[spi].spilled_ptr,
1369 reg_state->stack[spi].spilled_ptr.parent);
1374 for (i = 0; i < size; i++) {
1375 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_MISC)
1377 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_ZERO) {
1381 verbose(env, "invalid read from stack off %d+%d size %d\n",
1385 mark_reg_read(env, ®_state->stack[spi].spilled_ptr,
1386 reg_state->stack[spi].spilled_ptr.parent);
1387 if (value_regno >= 0) {
1388 if (zeros == size) {
1389 /* any size read into register is zero extended,
1390 * so the whole register == const_zero
1392 __mark_reg_const_zero(&state->regs[value_regno]);
1394 /* have read misc data from the stack */
1395 mark_reg_unknown(env, state->regs, value_regno);
1397 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
1403 static int check_stack_access(struct bpf_verifier_env *env,
1404 const struct bpf_reg_state *reg,
1407 /* Stack accesses must be at a fixed offset, so that we
1408 * can determine what type of data were returned. See
1409 * check_stack_read().
1411 if (!tnum_is_const(reg->var_off)) {
1414 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1415 verbose(env, "variable stack access var_off=%s off=%d size=%d",
1420 if (off >= 0 || off < -MAX_BPF_STACK) {
1421 verbose(env, "invalid stack off=%d size=%d\n", off, size);
1428 /* check read/write into map element returned by bpf_map_lookup_elem() */
1429 static int __check_map_access(struct bpf_verifier_env *env, u32 regno, int off,
1430 int size, bool zero_size_allowed)
1432 struct bpf_reg_state *regs = cur_regs(env);
1433 struct bpf_map *map = regs[regno].map_ptr;
1435 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
1436 off + size > map->value_size) {
1437 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
1438 map->value_size, off, size);
1444 /* check read/write into a map element with possible variable offset */
1445 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
1446 int off, int size, bool zero_size_allowed)
1448 struct bpf_verifier_state *vstate = env->cur_state;
1449 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1450 struct bpf_reg_state *reg = &state->regs[regno];
1453 /* We may have adjusted the register to this map value, so we
1454 * need to try adding each of min_value and max_value to off
1455 * to make sure our theoretical access will be safe.
1458 print_verifier_state(env, state);
1460 /* The minimum value is only important with signed
1461 * comparisons where we can't assume the floor of a
1462 * value is 0. If we are using signed variables for our
1463 * index'es we need to make sure that whatever we use
1464 * will have a set floor within our range.
1466 if (reg->smin_value < 0 &&
1467 (reg->smin_value == S64_MIN ||
1468 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
1469 reg->smin_value + off < 0)) {
1470 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1474 err = __check_map_access(env, regno, reg->smin_value + off, size,
1477 verbose(env, "R%d min value is outside of the array range\n",
1482 /* If we haven't set a max value then we need to bail since we can't be
1483 * sure we won't do bad things.
1484 * If reg->umax_value + off could overflow, treat that as unbounded too.
1486 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
1487 verbose(env, "R%d unbounded memory access, make sure to bounds check any array access into a map\n",
1491 err = __check_map_access(env, regno, reg->umax_value + off, size,
1494 verbose(env, "R%d max value is outside of the array range\n",
1497 if (map_value_has_spin_lock(reg->map_ptr)) {
1498 u32 lock = reg->map_ptr->spin_lock_off;
1500 /* if any part of struct bpf_spin_lock can be touched by
1501 * load/store reject this program.
1502 * To check that [x1, x2) overlaps with [y1, y2)
1503 * it is sufficient to check x1 < y2 && y1 < x2.
1505 if (reg->smin_value + off < lock + sizeof(struct bpf_spin_lock) &&
1506 lock < reg->umax_value + off + size) {
1507 verbose(env, "bpf_spin_lock cannot be accessed directly by load/store\n");
1514 #define MAX_PACKET_OFF 0xffff
1516 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
1517 const struct bpf_call_arg_meta *meta,
1518 enum bpf_access_type t)
1520 switch (env->prog->type) {
1521 /* Program types only with direct read access go here! */
1522 case BPF_PROG_TYPE_LWT_IN:
1523 case BPF_PROG_TYPE_LWT_OUT:
1524 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
1525 case BPF_PROG_TYPE_SK_REUSEPORT:
1526 case BPF_PROG_TYPE_FLOW_DISSECTOR:
1527 case BPF_PROG_TYPE_CGROUP_SKB:
1532 /* Program types with direct read + write access go here! */
1533 case BPF_PROG_TYPE_SCHED_CLS:
1534 case BPF_PROG_TYPE_SCHED_ACT:
1535 case BPF_PROG_TYPE_XDP:
1536 case BPF_PROG_TYPE_LWT_XMIT:
1537 case BPF_PROG_TYPE_SK_SKB:
1538 case BPF_PROG_TYPE_SK_MSG:
1540 return meta->pkt_access;
1542 env->seen_direct_write = true;
1549 static int __check_packet_access(struct bpf_verifier_env *env, u32 regno,
1550 int off, int size, bool zero_size_allowed)
1552 struct bpf_reg_state *regs = cur_regs(env);
1553 struct bpf_reg_state *reg = ®s[regno];
1555 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
1556 (u64)off + size > reg->range) {
1557 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
1558 off, size, regno, reg->id, reg->off, reg->range);
1564 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
1565 int size, bool zero_size_allowed)
1567 struct bpf_reg_state *regs = cur_regs(env);
1568 struct bpf_reg_state *reg = ®s[regno];
1571 /* We may have added a variable offset to the packet pointer; but any
1572 * reg->range we have comes after that. We are only checking the fixed
1576 /* We don't allow negative numbers, because we aren't tracking enough
1577 * detail to prove they're safe.
1579 if (reg->smin_value < 0) {
1580 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1584 err = __check_packet_access(env, regno, off, size, zero_size_allowed);
1586 verbose(env, "R%d offset is outside of the packet\n", regno);
1590 /* __check_packet_access has made sure "off + size - 1" is within u16.
1591 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
1592 * otherwise find_good_pkt_pointers would have refused to set range info
1593 * that __check_packet_access would have rejected this pkt access.
1594 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
1596 env->prog->aux->max_pkt_offset =
1597 max_t(u32, env->prog->aux->max_pkt_offset,
1598 off + reg->umax_value + size - 1);
1603 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
1604 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
1605 enum bpf_access_type t, enum bpf_reg_type *reg_type)
1607 struct bpf_insn_access_aux info = {
1608 .reg_type = *reg_type,
1611 if (env->ops->is_valid_access &&
1612 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
1613 /* A non zero info.ctx_field_size indicates that this field is a
1614 * candidate for later verifier transformation to load the whole
1615 * field and then apply a mask when accessed with a narrower
1616 * access than actual ctx access size. A zero info.ctx_field_size
1617 * will only allow for whole field access and rejects any other
1618 * type of narrower access.
1620 *reg_type = info.reg_type;
1622 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
1623 /* remember the offset of last byte accessed in ctx */
1624 if (env->prog->aux->max_ctx_offset < off + size)
1625 env->prog->aux->max_ctx_offset = off + size;
1629 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
1633 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
1636 if (size < 0 || off < 0 ||
1637 (u64)off + size > sizeof(struct bpf_flow_keys)) {
1638 verbose(env, "invalid access to flow keys off=%d size=%d\n",
1645 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
1646 u32 regno, int off, int size,
1647 enum bpf_access_type t)
1649 struct bpf_reg_state *regs = cur_regs(env);
1650 struct bpf_reg_state *reg = ®s[regno];
1651 struct bpf_insn_access_aux info = {};
1654 if (reg->smin_value < 0) {
1655 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1660 switch (reg->type) {
1661 case PTR_TO_SOCK_COMMON:
1662 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
1665 valid = bpf_sock_is_valid_access(off, size, t, &info);
1667 case PTR_TO_TCP_SOCK:
1668 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
1676 env->insn_aux_data[insn_idx].ctx_field_size =
1677 info.ctx_field_size;
1681 verbose(env, "R%d invalid %s access off=%d size=%d\n",
1682 regno, reg_type_str[reg->type], off, size);
1687 static bool __is_pointer_value(bool allow_ptr_leaks,
1688 const struct bpf_reg_state *reg)
1690 if (allow_ptr_leaks)
1693 return reg->type != SCALAR_VALUE;
1696 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
1698 return cur_regs(env) + regno;
1701 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
1703 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
1706 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
1708 const struct bpf_reg_state *reg = reg_state(env, regno);
1710 return reg->type == PTR_TO_CTX;
1713 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
1715 const struct bpf_reg_state *reg = reg_state(env, regno);
1717 return type_is_sk_pointer(reg->type);
1720 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
1722 const struct bpf_reg_state *reg = reg_state(env, regno);
1724 return type_is_pkt_pointer(reg->type);
1727 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
1729 const struct bpf_reg_state *reg = reg_state(env, regno);
1731 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
1732 return reg->type == PTR_TO_FLOW_KEYS;
1735 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
1736 const struct bpf_reg_state *reg,
1737 int off, int size, bool strict)
1739 struct tnum reg_off;
1742 /* Byte size accesses are always allowed. */
1743 if (!strict || size == 1)
1746 /* For platforms that do not have a Kconfig enabling
1747 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
1748 * NET_IP_ALIGN is universally set to '2'. And on platforms
1749 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
1750 * to this code only in strict mode where we want to emulate
1751 * the NET_IP_ALIGN==2 checking. Therefore use an
1752 * unconditional IP align value of '2'.
1756 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
1757 if (!tnum_is_aligned(reg_off, size)) {
1760 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1762 "misaligned packet access off %d+%s+%d+%d size %d\n",
1763 ip_align, tn_buf, reg->off, off, size);
1770 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
1771 const struct bpf_reg_state *reg,
1772 const char *pointer_desc,
1773 int off, int size, bool strict)
1775 struct tnum reg_off;
1777 /* Byte size accesses are always allowed. */
1778 if (!strict || size == 1)
1781 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
1782 if (!tnum_is_aligned(reg_off, size)) {
1785 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1786 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
1787 pointer_desc, tn_buf, reg->off, off, size);
1794 static int check_ptr_alignment(struct bpf_verifier_env *env,
1795 const struct bpf_reg_state *reg, int off,
1796 int size, bool strict_alignment_once)
1798 bool strict = env->strict_alignment || strict_alignment_once;
1799 const char *pointer_desc = "";
1801 switch (reg->type) {
1803 case PTR_TO_PACKET_META:
1804 /* Special case, because of NET_IP_ALIGN. Given metadata sits
1805 * right in front, treat it the very same way.
1807 return check_pkt_ptr_alignment(env, reg, off, size, strict);
1808 case PTR_TO_FLOW_KEYS:
1809 pointer_desc = "flow keys ";
1811 case PTR_TO_MAP_VALUE:
1812 pointer_desc = "value ";
1815 pointer_desc = "context ";
1818 pointer_desc = "stack ";
1819 /* The stack spill tracking logic in check_stack_write()
1820 * and check_stack_read() relies on stack accesses being
1826 pointer_desc = "sock ";
1828 case PTR_TO_SOCK_COMMON:
1829 pointer_desc = "sock_common ";
1831 case PTR_TO_TCP_SOCK:
1832 pointer_desc = "tcp_sock ";
1837 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
1841 static int update_stack_depth(struct bpf_verifier_env *env,
1842 const struct bpf_func_state *func,
1845 u16 stack = env->subprog_info[func->subprogno].stack_depth;
1850 /* update known max for given subprogram */
1851 env->subprog_info[func->subprogno].stack_depth = -off;
1855 /* starting from main bpf function walk all instructions of the function
1856 * and recursively walk all callees that given function can call.
1857 * Ignore jump and exit insns.
1858 * Since recursion is prevented by check_cfg() this algorithm
1859 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
1861 static int check_max_stack_depth(struct bpf_verifier_env *env)
1863 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
1864 struct bpf_subprog_info *subprog = env->subprog_info;
1865 struct bpf_insn *insn = env->prog->insnsi;
1866 int ret_insn[MAX_CALL_FRAMES];
1867 int ret_prog[MAX_CALL_FRAMES];
1870 /* round up to 32-bytes, since this is granularity
1871 * of interpreter stack size
1873 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
1874 if (depth > MAX_BPF_STACK) {
1875 verbose(env, "combined stack size of %d calls is %d. Too large\n",
1880 subprog_end = subprog[idx + 1].start;
1881 for (; i < subprog_end; i++) {
1882 if (insn[i].code != (BPF_JMP | BPF_CALL))
1884 if (insn[i].src_reg != BPF_PSEUDO_CALL)
1886 /* remember insn and function to return to */
1887 ret_insn[frame] = i + 1;
1888 ret_prog[frame] = idx;
1890 /* find the callee */
1891 i = i + insn[i].imm + 1;
1892 idx = find_subprog(env, i);
1894 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
1899 if (frame >= MAX_CALL_FRAMES) {
1900 verbose(env, "the call stack of %d frames is too deep !\n",
1906 /* end of for() loop means the last insn of the 'subprog'
1907 * was reached. Doesn't matter whether it was JA or EXIT
1911 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
1913 i = ret_insn[frame];
1914 idx = ret_prog[frame];
1918 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
1919 static int get_callee_stack_depth(struct bpf_verifier_env *env,
1920 const struct bpf_insn *insn, int idx)
1922 int start = idx + insn->imm + 1, subprog;
1924 subprog = find_subprog(env, start);
1926 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
1930 return env->subprog_info[subprog].stack_depth;
1934 static int check_ctx_reg(struct bpf_verifier_env *env,
1935 const struct bpf_reg_state *reg, int regno)
1937 /* Access to ctx or passing it to a helper is only allowed in
1938 * its original, unmodified form.
1942 verbose(env, "dereference of modified ctx ptr R%d off=%d disallowed\n",
1947 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
1950 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1951 verbose(env, "variable ctx access var_off=%s disallowed\n", tn_buf);
1958 /* truncate register to smaller size (in bytes)
1959 * must be called with size < BPF_REG_SIZE
1961 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
1965 /* clear high bits in bit representation */
1966 reg->var_off = tnum_cast(reg->var_off, size);
1968 /* fix arithmetic bounds */
1969 mask = ((u64)1 << (size * 8)) - 1;
1970 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
1971 reg->umin_value &= mask;
1972 reg->umax_value &= mask;
1974 reg->umin_value = 0;
1975 reg->umax_value = mask;
1977 reg->smin_value = reg->umin_value;
1978 reg->smax_value = reg->umax_value;
1981 /* check whether memory at (regno + off) is accessible for t = (read | write)
1982 * if t==write, value_regno is a register which value is stored into memory
1983 * if t==read, value_regno is a register which will receive the value from memory
1984 * if t==write && value_regno==-1, some unknown value is stored into memory
1985 * if t==read && value_regno==-1, don't care what we read from memory
1987 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
1988 int off, int bpf_size, enum bpf_access_type t,
1989 int value_regno, bool strict_alignment_once)
1991 struct bpf_reg_state *regs = cur_regs(env);
1992 struct bpf_reg_state *reg = regs + regno;
1993 struct bpf_func_state *state;
1996 size = bpf_size_to_bytes(bpf_size);
2000 /* alignment checks will add in reg->off themselves */
2001 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
2005 /* for access checks, reg->off is just part of off */
2008 if (reg->type == PTR_TO_MAP_VALUE) {
2009 if (t == BPF_WRITE && value_regno >= 0 &&
2010 is_pointer_value(env, value_regno)) {
2011 verbose(env, "R%d leaks addr into map\n", value_regno);
2015 err = check_map_access(env, regno, off, size, false);
2016 if (!err && t == BPF_READ && value_regno >= 0)
2017 mark_reg_unknown(env, regs, value_regno);
2019 } else if (reg->type == PTR_TO_CTX) {
2020 enum bpf_reg_type reg_type = SCALAR_VALUE;
2022 if (t == BPF_WRITE && value_regno >= 0 &&
2023 is_pointer_value(env, value_regno)) {
2024 verbose(env, "R%d leaks addr into ctx\n", value_regno);
2028 err = check_ctx_reg(env, reg, regno);
2032 err = check_ctx_access(env, insn_idx, off, size, t, ®_type);
2033 if (!err && t == BPF_READ && value_regno >= 0) {
2034 /* ctx access returns either a scalar, or a
2035 * PTR_TO_PACKET[_META,_END]. In the latter
2036 * case, we know the offset is zero.
2038 if (reg_type == SCALAR_VALUE) {
2039 mark_reg_unknown(env, regs, value_regno);
2041 mark_reg_known_zero(env, regs,
2043 if (reg_type_may_be_null(reg_type))
2044 regs[value_regno].id = ++env->id_gen;
2046 regs[value_regno].type = reg_type;
2049 } else if (reg->type == PTR_TO_STACK) {
2050 off += reg->var_off.value;
2051 err = check_stack_access(env, reg, off, size);
2055 state = func(env, reg);
2056 err = update_stack_depth(env, state, off);
2061 err = check_stack_write(env, state, off, size,
2062 value_regno, insn_idx);
2064 err = check_stack_read(env, state, off, size,
2066 } else if (reg_is_pkt_pointer(reg)) {
2067 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
2068 verbose(env, "cannot write into packet\n");
2071 if (t == BPF_WRITE && value_regno >= 0 &&
2072 is_pointer_value(env, value_regno)) {
2073 verbose(env, "R%d leaks addr into packet\n",
2077 err = check_packet_access(env, regno, off, size, false);
2078 if (!err && t == BPF_READ && value_regno >= 0)
2079 mark_reg_unknown(env, regs, value_regno);
2080 } else if (reg->type == PTR_TO_FLOW_KEYS) {
2081 if (t == BPF_WRITE && value_regno >= 0 &&
2082 is_pointer_value(env, value_regno)) {
2083 verbose(env, "R%d leaks addr into flow keys\n",
2088 err = check_flow_keys_access(env, off, size);
2089 if (!err && t == BPF_READ && value_regno >= 0)
2090 mark_reg_unknown(env, regs, value_regno);
2091 } else if (type_is_sk_pointer(reg->type)) {
2092 if (t == BPF_WRITE) {
2093 verbose(env, "R%d cannot write into %s\n",
2094 regno, reg_type_str[reg->type]);
2097 err = check_sock_access(env, insn_idx, regno, off, size, t);
2098 if (!err && value_regno >= 0)
2099 mark_reg_unknown(env, regs, value_regno);
2101 verbose(env, "R%d invalid mem access '%s'\n", regno,
2102 reg_type_str[reg->type]);
2106 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
2107 regs[value_regno].type == SCALAR_VALUE) {
2108 /* b/h/w load zero-extends, mark upper bits as known 0 */
2109 coerce_reg_to_size(®s[value_regno], size);
2114 static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
2118 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
2120 verbose(env, "BPF_XADD uses reserved fields\n");
2124 /* check src1 operand */
2125 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2129 /* check src2 operand */
2130 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2134 if (is_pointer_value(env, insn->src_reg)) {
2135 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
2139 if (is_ctx_reg(env, insn->dst_reg) ||
2140 is_pkt_reg(env, insn->dst_reg) ||
2141 is_flow_key_reg(env, insn->dst_reg) ||
2142 is_sk_reg(env, insn->dst_reg)) {
2143 verbose(env, "BPF_XADD stores into R%d %s is not allowed\n",
2145 reg_type_str[reg_state(env, insn->dst_reg)->type]);
2149 /* check whether atomic_add can read the memory */
2150 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
2151 BPF_SIZE(insn->code), BPF_READ, -1, true);
2155 /* check whether atomic_add can write into the same memory */
2156 return check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
2157 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
2160 /* when register 'regno' is passed into function that will read 'access_size'
2161 * bytes from that pointer, make sure that it's within stack boundary
2162 * and all elements of stack are initialized.
2163 * Unlike most pointer bounds-checking functions, this one doesn't take an
2164 * 'off' argument, so it has to add in reg->off itself.
2166 static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
2167 int access_size, bool zero_size_allowed,
2168 struct bpf_call_arg_meta *meta)
2170 struct bpf_reg_state *reg = reg_state(env, regno);
2171 struct bpf_func_state *state = func(env, reg);
2172 int off, i, slot, spi;
2174 if (reg->type != PTR_TO_STACK) {
2175 /* Allow zero-byte read from NULL, regardless of pointer type */
2176 if (zero_size_allowed && access_size == 0 &&
2177 register_is_null(reg))
2180 verbose(env, "R%d type=%s expected=%s\n", regno,
2181 reg_type_str[reg->type],
2182 reg_type_str[PTR_TO_STACK]);
2186 /* Only allow fixed-offset stack reads */
2187 if (!tnum_is_const(reg->var_off)) {
2190 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2191 verbose(env, "invalid variable stack read R%d var_off=%s\n",
2195 off = reg->off + reg->var_off.value;
2196 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
2197 access_size < 0 || (access_size == 0 && !zero_size_allowed)) {
2198 verbose(env, "invalid stack type R%d off=%d access_size=%d\n",
2199 regno, off, access_size);
2203 if (meta && meta->raw_mode) {
2204 meta->access_size = access_size;
2205 meta->regno = regno;
2209 for (i = 0; i < access_size; i++) {
2212 slot = -(off + i) - 1;
2213 spi = slot / BPF_REG_SIZE;
2214 if (state->allocated_stack <= slot)
2216 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
2217 if (*stype == STACK_MISC)
2219 if (*stype == STACK_ZERO) {
2220 /* helper can write anything into the stack */
2221 *stype = STACK_MISC;
2225 verbose(env, "invalid indirect read from stack off %d+%d size %d\n",
2226 off, i, access_size);
2229 /* reading any byte out of 8-byte 'spill_slot' will cause
2230 * the whole slot to be marked as 'read'
2232 mark_reg_read(env, &state->stack[spi].spilled_ptr,
2233 state->stack[spi].spilled_ptr.parent);
2235 return update_stack_depth(env, state, off);
2238 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
2239 int access_size, bool zero_size_allowed,
2240 struct bpf_call_arg_meta *meta)
2242 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
2244 switch (reg->type) {
2246 case PTR_TO_PACKET_META:
2247 return check_packet_access(env, regno, reg->off, access_size,
2249 case PTR_TO_MAP_VALUE:
2250 return check_map_access(env, regno, reg->off, access_size,
2252 default: /* scalar_value|ptr_to_stack or invalid ptr */
2253 return check_stack_boundary(env, regno, access_size,
2254 zero_size_allowed, meta);
2258 /* Implementation details:
2259 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL
2260 * Two bpf_map_lookups (even with the same key) will have different reg->id.
2261 * For traditional PTR_TO_MAP_VALUE the verifier clears reg->id after
2262 * value_or_null->value transition, since the verifier only cares about
2263 * the range of access to valid map value pointer and doesn't care about actual
2264 * address of the map element.
2265 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
2266 * reg->id > 0 after value_or_null->value transition. By doing so
2267 * two bpf_map_lookups will be considered two different pointers that
2268 * point to different bpf_spin_locks.
2269 * The verifier allows taking only one bpf_spin_lock at a time to avoid
2271 * Since only one bpf_spin_lock is allowed the checks are simpler than
2272 * reg_is_refcounted() logic. The verifier needs to remember only
2273 * one spin_lock instead of array of acquired_refs.
2274 * cur_state->active_spin_lock remembers which map value element got locked
2275 * and clears it after bpf_spin_unlock.
2277 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
2280 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
2281 struct bpf_verifier_state *cur = env->cur_state;
2282 bool is_const = tnum_is_const(reg->var_off);
2283 struct bpf_map *map = reg->map_ptr;
2284 u64 val = reg->var_off.value;
2286 if (reg->type != PTR_TO_MAP_VALUE) {
2287 verbose(env, "R%d is not a pointer to map_value\n", regno);
2292 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
2298 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
2302 if (!map_value_has_spin_lock(map)) {
2303 if (map->spin_lock_off == -E2BIG)
2305 "map '%s' has more than one 'struct bpf_spin_lock'\n",
2307 else if (map->spin_lock_off == -ENOENT)
2309 "map '%s' doesn't have 'struct bpf_spin_lock'\n",
2313 "map '%s' is not a struct type or bpf_spin_lock is mangled\n",
2317 if (map->spin_lock_off != val + reg->off) {
2318 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock'\n",
2323 if (cur->active_spin_lock) {
2325 "Locking two bpf_spin_locks are not allowed\n");
2328 cur->active_spin_lock = reg->id;
2330 if (!cur->active_spin_lock) {
2331 verbose(env, "bpf_spin_unlock without taking a lock\n");
2334 if (cur->active_spin_lock != reg->id) {
2335 verbose(env, "bpf_spin_unlock of different lock\n");
2338 cur->active_spin_lock = 0;
2343 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
2345 return type == ARG_PTR_TO_MEM ||
2346 type == ARG_PTR_TO_MEM_OR_NULL ||
2347 type == ARG_PTR_TO_UNINIT_MEM;
2350 static bool arg_type_is_mem_size(enum bpf_arg_type type)
2352 return type == ARG_CONST_SIZE ||
2353 type == ARG_CONST_SIZE_OR_ZERO;
2356 static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
2357 enum bpf_arg_type arg_type,
2358 struct bpf_call_arg_meta *meta)
2360 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
2361 enum bpf_reg_type expected_type, type = reg->type;
2364 if (arg_type == ARG_DONTCARE)
2367 err = check_reg_arg(env, regno, SRC_OP);
2371 if (arg_type == ARG_ANYTHING) {
2372 if (is_pointer_value(env, regno)) {
2373 verbose(env, "R%d leaks addr into helper function\n",
2380 if (type_is_pkt_pointer(type) &&
2381 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
2382 verbose(env, "helper access to the packet is not allowed\n");
2386 if (arg_type == ARG_PTR_TO_MAP_KEY ||
2387 arg_type == ARG_PTR_TO_MAP_VALUE ||
2388 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
2389 expected_type = PTR_TO_STACK;
2390 if (!type_is_pkt_pointer(type) && type != PTR_TO_MAP_VALUE &&
2391 type != expected_type)
2393 } else if (arg_type == ARG_CONST_SIZE ||
2394 arg_type == ARG_CONST_SIZE_OR_ZERO) {
2395 expected_type = SCALAR_VALUE;
2396 if (type != expected_type)
2398 } else if (arg_type == ARG_CONST_MAP_PTR) {
2399 expected_type = CONST_PTR_TO_MAP;
2400 if (type != expected_type)
2402 } else if (arg_type == ARG_PTR_TO_CTX) {
2403 expected_type = PTR_TO_CTX;
2404 if (type != expected_type)
2406 err = check_ctx_reg(env, reg, regno);
2409 } else if (arg_type == ARG_PTR_TO_SOCK_COMMON) {
2410 expected_type = PTR_TO_SOCK_COMMON;
2411 /* Any sk pointer can be ARG_PTR_TO_SOCK_COMMON */
2412 if (!type_is_sk_pointer(type))
2414 if (reg->ref_obj_id) {
2415 if (meta->ref_obj_id) {
2416 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
2417 regno, reg->ref_obj_id,
2421 meta->ref_obj_id = reg->ref_obj_id;
2423 } else if (arg_type == ARG_PTR_TO_SPIN_LOCK) {
2424 if (meta->func_id == BPF_FUNC_spin_lock) {
2425 if (process_spin_lock(env, regno, true))
2427 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
2428 if (process_spin_lock(env, regno, false))
2431 verbose(env, "verifier internal error\n");
2434 } else if (arg_type_is_mem_ptr(arg_type)) {
2435 expected_type = PTR_TO_STACK;
2436 /* One exception here. In case function allows for NULL to be
2437 * passed in as argument, it's a SCALAR_VALUE type. Final test
2438 * happens during stack boundary checking.
2440 if (register_is_null(reg) &&
2441 arg_type == ARG_PTR_TO_MEM_OR_NULL)
2442 /* final test in check_stack_boundary() */;
2443 else if (!type_is_pkt_pointer(type) &&
2444 type != PTR_TO_MAP_VALUE &&
2445 type != expected_type)
2447 meta->raw_mode = arg_type == ARG_PTR_TO_UNINIT_MEM;
2449 verbose(env, "unsupported arg_type %d\n", arg_type);
2453 if (arg_type == ARG_CONST_MAP_PTR) {
2454 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
2455 meta->map_ptr = reg->map_ptr;
2456 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
2457 /* bpf_map_xxx(..., map_ptr, ..., key) call:
2458 * check that [key, key + map->key_size) are within
2459 * stack limits and initialized
2461 if (!meta->map_ptr) {
2462 /* in function declaration map_ptr must come before
2463 * map_key, so that it's verified and known before
2464 * we have to check map_key here. Otherwise it means
2465 * that kernel subsystem misconfigured verifier
2467 verbose(env, "invalid map_ptr to access map->key\n");
2470 err = check_helper_mem_access(env, regno,
2471 meta->map_ptr->key_size, false,
2473 } else if (arg_type == ARG_PTR_TO_MAP_VALUE ||
2474 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
2475 /* bpf_map_xxx(..., map_ptr, ..., value) call:
2476 * check [value, value + map->value_size) validity
2478 if (!meta->map_ptr) {
2479 /* kernel subsystem misconfigured verifier */
2480 verbose(env, "invalid map_ptr to access map->value\n");
2483 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE);
2484 err = check_helper_mem_access(env, regno,
2485 meta->map_ptr->value_size, false,
2487 } else if (arg_type_is_mem_size(arg_type)) {
2488 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
2490 /* remember the mem_size which may be used later
2491 * to refine return values.
2493 meta->msize_smax_value = reg->smax_value;
2494 meta->msize_umax_value = reg->umax_value;
2496 /* The register is SCALAR_VALUE; the access check
2497 * happens using its boundaries.
2499 if (!tnum_is_const(reg->var_off))
2500 /* For unprivileged variable accesses, disable raw
2501 * mode so that the program is required to
2502 * initialize all the memory that the helper could
2503 * just partially fill up.
2507 if (reg->smin_value < 0) {
2508 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
2513 if (reg->umin_value == 0) {
2514 err = check_helper_mem_access(env, regno - 1, 0,
2521 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
2522 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
2526 err = check_helper_mem_access(env, regno - 1,
2528 zero_size_allowed, meta);
2533 verbose(env, "R%d type=%s expected=%s\n", regno,
2534 reg_type_str[type], reg_type_str[expected_type]);
2538 static int check_map_func_compatibility(struct bpf_verifier_env *env,
2539 struct bpf_map *map, int func_id)
2544 /* We need a two way check, first is from map perspective ... */
2545 switch (map->map_type) {
2546 case BPF_MAP_TYPE_PROG_ARRAY:
2547 if (func_id != BPF_FUNC_tail_call)
2550 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
2551 if (func_id != BPF_FUNC_perf_event_read &&
2552 func_id != BPF_FUNC_perf_event_output &&
2553 func_id != BPF_FUNC_perf_event_read_value)
2556 case BPF_MAP_TYPE_STACK_TRACE:
2557 if (func_id != BPF_FUNC_get_stackid)
2560 case BPF_MAP_TYPE_CGROUP_ARRAY:
2561 if (func_id != BPF_FUNC_skb_under_cgroup &&
2562 func_id != BPF_FUNC_current_task_under_cgroup)
2565 case BPF_MAP_TYPE_CGROUP_STORAGE:
2566 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
2567 if (func_id != BPF_FUNC_get_local_storage)
2570 /* devmap returns a pointer to a live net_device ifindex that we cannot
2571 * allow to be modified from bpf side. So do not allow lookup elements
2574 case BPF_MAP_TYPE_DEVMAP:
2575 if (func_id != BPF_FUNC_redirect_map)
2578 /* Restrict bpf side of cpumap and xskmap, open when use-cases
2581 case BPF_MAP_TYPE_CPUMAP:
2582 case BPF_MAP_TYPE_XSKMAP:
2583 if (func_id != BPF_FUNC_redirect_map)
2586 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
2587 case BPF_MAP_TYPE_HASH_OF_MAPS:
2588 if (func_id != BPF_FUNC_map_lookup_elem)
2591 case BPF_MAP_TYPE_SOCKMAP:
2592 if (func_id != BPF_FUNC_sk_redirect_map &&
2593 func_id != BPF_FUNC_sock_map_update &&
2594 func_id != BPF_FUNC_map_delete_elem &&
2595 func_id != BPF_FUNC_msg_redirect_map)
2598 case BPF_MAP_TYPE_SOCKHASH:
2599 if (func_id != BPF_FUNC_sk_redirect_hash &&
2600 func_id != BPF_FUNC_sock_hash_update &&
2601 func_id != BPF_FUNC_map_delete_elem &&
2602 func_id != BPF_FUNC_msg_redirect_hash)
2605 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
2606 if (func_id != BPF_FUNC_sk_select_reuseport)
2609 case BPF_MAP_TYPE_QUEUE:
2610 case BPF_MAP_TYPE_STACK:
2611 if (func_id != BPF_FUNC_map_peek_elem &&
2612 func_id != BPF_FUNC_map_pop_elem &&
2613 func_id != BPF_FUNC_map_push_elem)
2620 /* ... and second from the function itself. */
2622 case BPF_FUNC_tail_call:
2623 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
2625 if (env->subprog_cnt > 1) {
2626 verbose(env, "tail_calls are not allowed in programs with bpf-to-bpf calls\n");
2630 case BPF_FUNC_perf_event_read:
2631 case BPF_FUNC_perf_event_output:
2632 case BPF_FUNC_perf_event_read_value:
2633 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
2636 case BPF_FUNC_get_stackid:
2637 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
2640 case BPF_FUNC_current_task_under_cgroup:
2641 case BPF_FUNC_skb_under_cgroup:
2642 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
2645 case BPF_FUNC_redirect_map:
2646 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
2647 map->map_type != BPF_MAP_TYPE_CPUMAP &&
2648 map->map_type != BPF_MAP_TYPE_XSKMAP)
2651 case BPF_FUNC_sk_redirect_map:
2652 case BPF_FUNC_msg_redirect_map:
2653 case BPF_FUNC_sock_map_update:
2654 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
2657 case BPF_FUNC_sk_redirect_hash:
2658 case BPF_FUNC_msg_redirect_hash:
2659 case BPF_FUNC_sock_hash_update:
2660 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
2663 case BPF_FUNC_get_local_storage:
2664 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
2665 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
2668 case BPF_FUNC_sk_select_reuseport:
2669 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY)
2672 case BPF_FUNC_map_peek_elem:
2673 case BPF_FUNC_map_pop_elem:
2674 case BPF_FUNC_map_push_elem:
2675 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
2676 map->map_type != BPF_MAP_TYPE_STACK)
2685 verbose(env, "cannot pass map_type %d into func %s#%d\n",
2686 map->map_type, func_id_name(func_id), func_id);
2690 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
2694 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
2696 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
2698 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
2700 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
2702 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
2705 /* We only support one arg being in raw mode at the moment,
2706 * which is sufficient for the helper functions we have
2712 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
2713 enum bpf_arg_type arg_next)
2715 return (arg_type_is_mem_ptr(arg_curr) &&
2716 !arg_type_is_mem_size(arg_next)) ||
2717 (!arg_type_is_mem_ptr(arg_curr) &&
2718 arg_type_is_mem_size(arg_next));
2721 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
2723 /* bpf_xxx(..., buf, len) call will access 'len'
2724 * bytes from memory 'buf'. Both arg types need
2725 * to be paired, so make sure there's no buggy
2726 * helper function specification.
2728 if (arg_type_is_mem_size(fn->arg1_type) ||
2729 arg_type_is_mem_ptr(fn->arg5_type) ||
2730 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
2731 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
2732 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
2733 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
2739 static bool check_refcount_ok(const struct bpf_func_proto *fn, int func_id)
2743 if (arg_type_may_be_refcounted(fn->arg1_type))
2745 if (arg_type_may_be_refcounted(fn->arg2_type))
2747 if (arg_type_may_be_refcounted(fn->arg3_type))
2749 if (arg_type_may_be_refcounted(fn->arg4_type))
2751 if (arg_type_may_be_refcounted(fn->arg5_type))
2754 /* A reference acquiring function cannot acquire
2755 * another refcounted ptr.
2757 if (is_acquire_function(func_id) && count)
2760 /* We only support one arg being unreferenced at the moment,
2761 * which is sufficient for the helper functions we have right now.
2766 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
2768 return check_raw_mode_ok(fn) &&
2769 check_arg_pair_ok(fn) &&
2770 check_refcount_ok(fn, func_id) ? 0 : -EINVAL;
2773 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
2774 * are now invalid, so turn them into unknown SCALAR_VALUE.
2776 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
2777 struct bpf_func_state *state)
2779 struct bpf_reg_state *regs = state->regs, *reg;
2782 for (i = 0; i < MAX_BPF_REG; i++)
2783 if (reg_is_pkt_pointer_any(®s[i]))
2784 mark_reg_unknown(env, regs, i);
2786 bpf_for_each_spilled_reg(i, state, reg) {
2789 if (reg_is_pkt_pointer_any(reg))
2790 __mark_reg_unknown(reg);
2794 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
2796 struct bpf_verifier_state *vstate = env->cur_state;
2799 for (i = 0; i <= vstate->curframe; i++)
2800 __clear_all_pkt_pointers(env, vstate->frame[i]);
2803 static void release_reg_references(struct bpf_verifier_env *env,
2804 struct bpf_func_state *state,
2807 struct bpf_reg_state *regs = state->regs, *reg;
2810 for (i = 0; i < MAX_BPF_REG; i++)
2811 if (regs[i].ref_obj_id == ref_obj_id)
2812 mark_reg_unknown(env, regs, i);
2814 bpf_for_each_spilled_reg(i, state, reg) {
2817 if (reg->ref_obj_id == ref_obj_id)
2818 __mark_reg_unknown(reg);
2822 /* The pointer with the specified id has released its reference to kernel
2823 * resources. Identify all copies of the same pointer and clear the reference.
2825 static int release_reference(struct bpf_verifier_env *env,
2828 struct bpf_verifier_state *vstate = env->cur_state;
2832 err = release_reference_state(cur_func(env), ref_obj_id);
2836 for (i = 0; i <= vstate->curframe; i++)
2837 release_reg_references(env, vstate->frame[i], ref_obj_id);
2842 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
2845 struct bpf_verifier_state *state = env->cur_state;
2846 struct bpf_func_state *caller, *callee;
2847 int i, err, subprog, target_insn;
2849 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
2850 verbose(env, "the call stack of %d frames is too deep\n",
2851 state->curframe + 2);
2855 target_insn = *insn_idx + insn->imm;
2856 subprog = find_subprog(env, target_insn + 1);
2858 verbose(env, "verifier bug. No program starts at insn %d\n",
2863 caller = state->frame[state->curframe];
2864 if (state->frame[state->curframe + 1]) {
2865 verbose(env, "verifier bug. Frame %d already allocated\n",
2866 state->curframe + 1);
2870 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
2873 state->frame[state->curframe + 1] = callee;
2875 /* callee cannot access r0, r6 - r9 for reading and has to write
2876 * into its own stack before reading from it.
2877 * callee can read/write into caller's stack
2879 init_func_state(env, callee,
2880 /* remember the callsite, it will be used by bpf_exit */
2881 *insn_idx /* callsite */,
2882 state->curframe + 1 /* frameno within this callchain */,
2883 subprog /* subprog number within this prog */);
2885 /* Transfer references to the callee */
2886 err = transfer_reference_state(callee, caller);
2890 /* copy r1 - r5 args that callee can access. The copy includes parent
2891 * pointers, which connects us up to the liveness chain
2893 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
2894 callee->regs[i] = caller->regs[i];
2896 /* after the call registers r0 - r5 were scratched */
2897 for (i = 0; i < CALLER_SAVED_REGS; i++) {
2898 mark_reg_not_init(env, caller->regs, caller_saved[i]);
2899 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
2902 /* only increment it after check_reg_arg() finished */
2905 /* and go analyze first insn of the callee */
2906 *insn_idx = target_insn;
2908 if (env->log.level) {
2909 verbose(env, "caller:\n");
2910 print_verifier_state(env, caller);
2911 verbose(env, "callee:\n");
2912 print_verifier_state(env, callee);
2917 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
2919 struct bpf_verifier_state *state = env->cur_state;
2920 struct bpf_func_state *caller, *callee;
2921 struct bpf_reg_state *r0;
2924 callee = state->frame[state->curframe];
2925 r0 = &callee->regs[BPF_REG_0];
2926 if (r0->type == PTR_TO_STACK) {
2927 /* technically it's ok to return caller's stack pointer
2928 * (or caller's caller's pointer) back to the caller,
2929 * since these pointers are valid. Only current stack
2930 * pointer will be invalid as soon as function exits,
2931 * but let's be conservative
2933 verbose(env, "cannot return stack pointer to the caller\n");
2938 caller = state->frame[state->curframe];
2939 /* return to the caller whatever r0 had in the callee */
2940 caller->regs[BPF_REG_0] = *r0;
2942 /* Transfer references to the caller */
2943 err = transfer_reference_state(caller, callee);
2947 *insn_idx = callee->callsite + 1;
2948 if (env->log.level) {
2949 verbose(env, "returning from callee:\n");
2950 print_verifier_state(env, callee);
2951 verbose(env, "to caller at %d:\n", *insn_idx);
2952 print_verifier_state(env, caller);
2954 /* clear everything in the callee */
2955 free_func_state(callee);
2956 state->frame[state->curframe + 1] = NULL;
2960 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
2962 struct bpf_call_arg_meta *meta)
2964 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
2966 if (ret_type != RET_INTEGER ||
2967 (func_id != BPF_FUNC_get_stack &&
2968 func_id != BPF_FUNC_probe_read_str))
2971 ret_reg->smax_value = meta->msize_smax_value;
2972 ret_reg->umax_value = meta->msize_umax_value;
2973 __reg_deduce_bounds(ret_reg);
2974 __reg_bound_offset(ret_reg);
2978 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
2979 int func_id, int insn_idx)
2981 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
2983 if (func_id != BPF_FUNC_tail_call &&
2984 func_id != BPF_FUNC_map_lookup_elem &&
2985 func_id != BPF_FUNC_map_update_elem &&
2986 func_id != BPF_FUNC_map_delete_elem &&
2987 func_id != BPF_FUNC_map_push_elem &&
2988 func_id != BPF_FUNC_map_pop_elem &&
2989 func_id != BPF_FUNC_map_peek_elem)
2992 if (meta->map_ptr == NULL) {
2993 verbose(env, "kernel subsystem misconfigured verifier\n");
2997 if (!BPF_MAP_PTR(aux->map_state))
2998 bpf_map_ptr_store(aux, meta->map_ptr,
2999 meta->map_ptr->unpriv_array);
3000 else if (BPF_MAP_PTR(aux->map_state) != meta->map_ptr)
3001 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
3002 meta->map_ptr->unpriv_array);
3006 static int check_reference_leak(struct bpf_verifier_env *env)
3008 struct bpf_func_state *state = cur_func(env);
3011 for (i = 0; i < state->acquired_refs; i++) {
3012 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
3013 state->refs[i].id, state->refs[i].insn_idx);
3015 return state->acquired_refs ? -EINVAL : 0;
3018 static int check_helper_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
3020 const struct bpf_func_proto *fn = NULL;
3021 struct bpf_reg_state *regs;
3022 struct bpf_call_arg_meta meta;
3026 /* find function prototype */
3027 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
3028 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
3033 if (env->ops->get_func_proto)
3034 fn = env->ops->get_func_proto(func_id, env->prog);
3036 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
3041 /* eBPF programs must be GPL compatible to use GPL-ed functions */
3042 if (!env->prog->gpl_compatible && fn->gpl_only) {
3043 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
3047 /* With LD_ABS/IND some JITs save/restore skb from r1. */
3048 changes_data = bpf_helper_changes_pkt_data(fn->func);
3049 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
3050 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
3051 func_id_name(func_id), func_id);
3055 memset(&meta, 0, sizeof(meta));
3056 meta.pkt_access = fn->pkt_access;
3058 err = check_func_proto(fn, func_id);
3060 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
3061 func_id_name(func_id), func_id);
3065 meta.func_id = func_id;
3067 err = check_func_arg(env, BPF_REG_1, fn->arg1_type, &meta);
3070 err = check_func_arg(env, BPF_REG_2, fn->arg2_type, &meta);
3073 err = check_func_arg(env, BPF_REG_3, fn->arg3_type, &meta);
3076 err = check_func_arg(env, BPF_REG_4, fn->arg4_type, &meta);
3079 err = check_func_arg(env, BPF_REG_5, fn->arg5_type, &meta);
3083 err = record_func_map(env, &meta, func_id, insn_idx);
3087 /* Mark slots with STACK_MISC in case of raw mode, stack offset
3088 * is inferred from register state.
3090 for (i = 0; i < meta.access_size; i++) {
3091 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
3092 BPF_WRITE, -1, false);
3097 if (func_id == BPF_FUNC_tail_call) {
3098 err = check_reference_leak(env);
3100 verbose(env, "tail_call would lead to reference leak\n");
3103 } else if (is_release_function(func_id)) {
3104 err = release_reference(env, meta.ref_obj_id);
3106 verbose(env, "func %s#%d reference has not been acquired before\n",
3107 func_id_name(func_id), func_id);
3112 regs = cur_regs(env);
3114 /* check that flags argument in get_local_storage(map, flags) is 0,
3115 * this is required because get_local_storage() can't return an error.
3117 if (func_id == BPF_FUNC_get_local_storage &&
3118 !register_is_null(®s[BPF_REG_2])) {
3119 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
3123 /* reset caller saved regs */
3124 for (i = 0; i < CALLER_SAVED_REGS; i++) {
3125 mark_reg_not_init(env, regs, caller_saved[i]);
3126 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
3129 /* update return register (already marked as written above) */
3130 if (fn->ret_type == RET_INTEGER) {
3131 /* sets type to SCALAR_VALUE */
3132 mark_reg_unknown(env, regs, BPF_REG_0);
3133 } else if (fn->ret_type == RET_VOID) {
3134 regs[BPF_REG_0].type = NOT_INIT;
3135 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL ||
3136 fn->ret_type == RET_PTR_TO_MAP_VALUE) {
3137 /* There is no offset yet applied, variable or fixed */
3138 mark_reg_known_zero(env, regs, BPF_REG_0);
3139 /* remember map_ptr, so that check_map_access()
3140 * can check 'value_size' boundary of memory access
3141 * to map element returned from bpf_map_lookup_elem()
3143 if (meta.map_ptr == NULL) {
3145 "kernel subsystem misconfigured verifier\n");
3148 regs[BPF_REG_0].map_ptr = meta.map_ptr;
3149 if (fn->ret_type == RET_PTR_TO_MAP_VALUE) {
3150 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE;
3151 if (map_value_has_spin_lock(meta.map_ptr))
3152 regs[BPF_REG_0].id = ++env->id_gen;
3154 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
3155 regs[BPF_REG_0].id = ++env->id_gen;
3157 } else if (fn->ret_type == RET_PTR_TO_SOCKET_OR_NULL) {
3158 mark_reg_known_zero(env, regs, BPF_REG_0);
3159 regs[BPF_REG_0].type = PTR_TO_SOCKET_OR_NULL;
3160 if (is_acquire_function(func_id)) {
3161 int id = acquire_reference_state(env, insn_idx);
3165 /* For mark_ptr_or_null_reg() */
3166 regs[BPF_REG_0].id = id;
3167 /* For release_reference() */
3168 regs[BPF_REG_0].ref_obj_id = id;
3170 /* For mark_ptr_or_null_reg() */
3171 regs[BPF_REG_0].id = ++env->id_gen;
3173 } else if (fn->ret_type == RET_PTR_TO_TCP_SOCK_OR_NULL) {
3174 mark_reg_known_zero(env, regs, BPF_REG_0);
3175 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK_OR_NULL;
3176 regs[BPF_REG_0].id = ++env->id_gen;
3178 verbose(env, "unknown return type %d of func %s#%d\n",
3179 fn->ret_type, func_id_name(func_id), func_id);
3183 if (is_ptr_cast_function(func_id))
3184 /* For release_reference() */
3185 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
3187 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
3189 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
3193 if (func_id == BPF_FUNC_get_stack && !env->prog->has_callchain_buf) {
3194 const char *err_str;
3196 #ifdef CONFIG_PERF_EVENTS
3197 err = get_callchain_buffers(sysctl_perf_event_max_stack);
3198 err_str = "cannot get callchain buffer for func %s#%d\n";
3201 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
3204 verbose(env, err_str, func_id_name(func_id), func_id);
3208 env->prog->has_callchain_buf = true;
3212 clear_all_pkt_pointers(env);
3216 static bool signed_add_overflows(s64 a, s64 b)
3218 /* Do the add in u64, where overflow is well-defined */
3219 s64 res = (s64)((u64)a + (u64)b);
3226 static bool signed_sub_overflows(s64 a, s64 b)
3228 /* Do the sub in u64, where overflow is well-defined */
3229 s64 res = (s64)((u64)a - (u64)b);
3236 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
3237 const struct bpf_reg_state *reg,
3238 enum bpf_reg_type type)
3240 bool known = tnum_is_const(reg->var_off);
3241 s64 val = reg->var_off.value;
3242 s64 smin = reg->smin_value;
3244 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
3245 verbose(env, "math between %s pointer and %lld is not allowed\n",
3246 reg_type_str[type], val);
3250 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
3251 verbose(env, "%s pointer offset %d is not allowed\n",
3252 reg_type_str[type], reg->off);
3256 if (smin == S64_MIN) {
3257 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
3258 reg_type_str[type]);
3262 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
3263 verbose(env, "value %lld makes %s pointer be out of bounds\n",
3264 smin, reg_type_str[type]);
3271 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
3273 return &env->insn_aux_data[env->insn_idx];
3276 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
3277 u32 *ptr_limit, u8 opcode, bool off_is_neg)
3279 bool mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
3280 (opcode == BPF_SUB && !off_is_neg);
3283 switch (ptr_reg->type) {
3285 off = ptr_reg->off + ptr_reg->var_off.value;
3287 *ptr_limit = MAX_BPF_STACK + off;
3291 case PTR_TO_MAP_VALUE:
3293 *ptr_limit = ptr_reg->umax_value + ptr_reg->off;
3295 off = ptr_reg->smin_value + ptr_reg->off;
3296 *ptr_limit = ptr_reg->map_ptr->value_size - off;
3304 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
3305 const struct bpf_insn *insn)
3307 return env->allow_ptr_leaks || BPF_SRC(insn->code) == BPF_K;
3310 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
3311 u32 alu_state, u32 alu_limit)
3313 /* If we arrived here from different branches with different
3314 * state or limits to sanitize, then this won't work.
3316 if (aux->alu_state &&
3317 (aux->alu_state != alu_state ||
3318 aux->alu_limit != alu_limit))
3321 /* Corresponding fixup done in fixup_bpf_calls(). */
3322 aux->alu_state = alu_state;
3323 aux->alu_limit = alu_limit;
3327 static int sanitize_val_alu(struct bpf_verifier_env *env,
3328 struct bpf_insn *insn)
3330 struct bpf_insn_aux_data *aux = cur_aux(env);
3332 if (can_skip_alu_sanitation(env, insn))
3335 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
3338 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
3339 struct bpf_insn *insn,
3340 const struct bpf_reg_state *ptr_reg,
3341 struct bpf_reg_state *dst_reg,
3344 struct bpf_verifier_state *vstate = env->cur_state;
3345 struct bpf_insn_aux_data *aux = cur_aux(env);
3346 bool ptr_is_dst_reg = ptr_reg == dst_reg;
3347 u8 opcode = BPF_OP(insn->code);
3348 u32 alu_state, alu_limit;
3349 struct bpf_reg_state tmp;
3352 if (can_skip_alu_sanitation(env, insn))
3355 /* We already marked aux for masking from non-speculative
3356 * paths, thus we got here in the first place. We only care
3357 * to explore bad access from here.
3359 if (vstate->speculative)
3362 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
3363 alu_state |= ptr_is_dst_reg ?
3364 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
3366 if (retrieve_ptr_limit(ptr_reg, &alu_limit, opcode, off_is_neg))
3368 if (update_alu_sanitation_state(aux, alu_state, alu_limit))
3371 /* Simulate and find potential out-of-bounds access under
3372 * speculative execution from truncation as a result of
3373 * masking when off was not within expected range. If off
3374 * sits in dst, then we temporarily need to move ptr there
3375 * to simulate dst (== 0) +/-= ptr. Needed, for example,
3376 * for cases where we use K-based arithmetic in one direction
3377 * and truncated reg-based in the other in order to explore
3380 if (!ptr_is_dst_reg) {
3382 *dst_reg = *ptr_reg;
3384 ret = push_stack(env, env->insn_idx + 1, env->insn_idx, true);
3385 if (!ptr_is_dst_reg && ret)
3387 return !ret ? -EFAULT : 0;
3390 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
3391 * Caller should also handle BPF_MOV case separately.
3392 * If we return -EACCES, caller may want to try again treating pointer as a
3393 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
3395 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
3396 struct bpf_insn *insn,
3397 const struct bpf_reg_state *ptr_reg,
3398 const struct bpf_reg_state *off_reg)
3400 struct bpf_verifier_state *vstate = env->cur_state;
3401 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3402 struct bpf_reg_state *regs = state->regs, *dst_reg;
3403 bool known = tnum_is_const(off_reg->var_off);
3404 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
3405 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
3406 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
3407 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
3408 u32 dst = insn->dst_reg, src = insn->src_reg;
3409 u8 opcode = BPF_OP(insn->code);
3412 dst_reg = ®s[dst];
3414 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
3415 smin_val > smax_val || umin_val > umax_val) {
3416 /* Taint dst register if offset had invalid bounds derived from
3417 * e.g. dead branches.
3419 __mark_reg_unknown(dst_reg);
3423 if (BPF_CLASS(insn->code) != BPF_ALU64) {
3424 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
3426 "R%d 32-bit pointer arithmetic prohibited\n",
3431 switch (ptr_reg->type) {
3432 case PTR_TO_MAP_VALUE_OR_NULL:
3433 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
3434 dst, reg_type_str[ptr_reg->type]);
3436 case CONST_PTR_TO_MAP:
3437 case PTR_TO_PACKET_END:
3439 case PTR_TO_SOCKET_OR_NULL:
3440 case PTR_TO_SOCK_COMMON:
3441 case PTR_TO_SOCK_COMMON_OR_NULL:
3442 case PTR_TO_TCP_SOCK:
3443 case PTR_TO_TCP_SOCK_OR_NULL:
3444 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
3445 dst, reg_type_str[ptr_reg->type]);
3447 case PTR_TO_MAP_VALUE:
3448 if (!env->allow_ptr_leaks && !known && (smin_val < 0) != (smax_val < 0)) {
3449 verbose(env, "R%d has unknown scalar with mixed signed bounds, pointer arithmetic with it prohibited for !root\n",
3450 off_reg == dst_reg ? dst : src);
3458 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
3459 * The id may be overwritten later if we create a new variable offset.
3461 dst_reg->type = ptr_reg->type;
3462 dst_reg->id = ptr_reg->id;
3464 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
3465 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
3470 ret = sanitize_ptr_alu(env, insn, ptr_reg, dst_reg, smin_val < 0);
3472 verbose(env, "R%d tried to add from different maps or paths\n", dst);
3475 /* We can take a fixed offset as long as it doesn't overflow
3476 * the s32 'off' field
3478 if (known && (ptr_reg->off + smin_val ==
3479 (s64)(s32)(ptr_reg->off + smin_val))) {
3480 /* pointer += K. Accumulate it into fixed offset */
3481 dst_reg->smin_value = smin_ptr;
3482 dst_reg->smax_value = smax_ptr;
3483 dst_reg->umin_value = umin_ptr;
3484 dst_reg->umax_value = umax_ptr;
3485 dst_reg->var_off = ptr_reg->var_off;
3486 dst_reg->off = ptr_reg->off + smin_val;
3487 dst_reg->raw = ptr_reg->raw;
3490 /* A new variable offset is created. Note that off_reg->off
3491 * == 0, since it's a scalar.
3492 * dst_reg gets the pointer type and since some positive
3493 * integer value was added to the pointer, give it a new 'id'
3494 * if it's a PTR_TO_PACKET.
3495 * this creates a new 'base' pointer, off_reg (variable) gets
3496 * added into the variable offset, and we copy the fixed offset
3499 if (signed_add_overflows(smin_ptr, smin_val) ||
3500 signed_add_overflows(smax_ptr, smax_val)) {
3501 dst_reg->smin_value = S64_MIN;
3502 dst_reg->smax_value = S64_MAX;
3504 dst_reg->smin_value = smin_ptr + smin_val;
3505 dst_reg->smax_value = smax_ptr + smax_val;
3507 if (umin_ptr + umin_val < umin_ptr ||
3508 umax_ptr + umax_val < umax_ptr) {
3509 dst_reg->umin_value = 0;
3510 dst_reg->umax_value = U64_MAX;
3512 dst_reg->umin_value = umin_ptr + umin_val;
3513 dst_reg->umax_value = umax_ptr + umax_val;
3515 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
3516 dst_reg->off = ptr_reg->off;
3517 dst_reg->raw = ptr_reg->raw;
3518 if (reg_is_pkt_pointer(ptr_reg)) {
3519 dst_reg->id = ++env->id_gen;
3520 /* something was added to pkt_ptr, set range to zero */
3525 ret = sanitize_ptr_alu(env, insn, ptr_reg, dst_reg, smin_val < 0);
3527 verbose(env, "R%d tried to sub from different maps or paths\n", dst);
3530 if (dst_reg == off_reg) {
3531 /* scalar -= pointer. Creates an unknown scalar */
3532 verbose(env, "R%d tried to subtract pointer from scalar\n",
3536 /* We don't allow subtraction from FP, because (according to
3537 * test_verifier.c test "invalid fp arithmetic", JITs might not
3538 * be able to deal with it.
3540 if (ptr_reg->type == PTR_TO_STACK) {
3541 verbose(env, "R%d subtraction from stack pointer prohibited\n",
3545 if (known && (ptr_reg->off - smin_val ==
3546 (s64)(s32)(ptr_reg->off - smin_val))) {
3547 /* pointer -= K. Subtract it from fixed offset */
3548 dst_reg->smin_value = smin_ptr;
3549 dst_reg->smax_value = smax_ptr;
3550 dst_reg->umin_value = umin_ptr;
3551 dst_reg->umax_value = umax_ptr;
3552 dst_reg->var_off = ptr_reg->var_off;
3553 dst_reg->id = ptr_reg->id;
3554 dst_reg->off = ptr_reg->off - smin_val;
3555 dst_reg->raw = ptr_reg->raw;
3558 /* A new variable offset is created. If the subtrahend is known
3559 * nonnegative, then any reg->range we had before is still good.
3561 if (signed_sub_overflows(smin_ptr, smax_val) ||
3562 signed_sub_overflows(smax_ptr, smin_val)) {
3563 /* Overflow possible, we know nothing */
3564 dst_reg->smin_value = S64_MIN;
3565 dst_reg->smax_value = S64_MAX;
3567 dst_reg->smin_value = smin_ptr - smax_val;
3568 dst_reg->smax_value = smax_ptr - smin_val;
3570 if (umin_ptr < umax_val) {
3571 /* Overflow possible, we know nothing */
3572 dst_reg->umin_value = 0;
3573 dst_reg->umax_value = U64_MAX;
3575 /* Cannot overflow (as long as bounds are consistent) */
3576 dst_reg->umin_value = umin_ptr - umax_val;
3577 dst_reg->umax_value = umax_ptr - umin_val;
3579 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
3580 dst_reg->off = ptr_reg->off;
3581 dst_reg->raw = ptr_reg->raw;
3582 if (reg_is_pkt_pointer(ptr_reg)) {
3583 dst_reg->id = ++env->id_gen;
3584 /* something was added to pkt_ptr, set range to zero */
3592 /* bitwise ops on pointers are troublesome, prohibit. */
3593 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
3594 dst, bpf_alu_string[opcode >> 4]);
3597 /* other operators (e.g. MUL,LSH) produce non-pointer results */
3598 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
3599 dst, bpf_alu_string[opcode >> 4]);
3603 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
3606 __update_reg_bounds(dst_reg);
3607 __reg_deduce_bounds(dst_reg);
3608 __reg_bound_offset(dst_reg);
3610 /* For unprivileged we require that resulting offset must be in bounds
3611 * in order to be able to sanitize access later on.
3613 if (!env->allow_ptr_leaks) {
3614 if (dst_reg->type == PTR_TO_MAP_VALUE &&
3615 check_map_access(env, dst, dst_reg->off, 1, false)) {
3616 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
3617 "prohibited for !root\n", dst);
3619 } else if (dst_reg->type == PTR_TO_STACK &&
3620 check_stack_access(env, dst_reg, dst_reg->off +
3621 dst_reg->var_off.value, 1)) {
3622 verbose(env, "R%d stack pointer arithmetic goes out of range, "
3623 "prohibited for !root\n", dst);
3631 /* WARNING: This function does calculations on 64-bit values, but the actual
3632 * execution may occur on 32-bit values. Therefore, things like bitshifts
3633 * need extra checks in the 32-bit case.
3635 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
3636 struct bpf_insn *insn,
3637 struct bpf_reg_state *dst_reg,
3638 struct bpf_reg_state src_reg)
3640 struct bpf_reg_state *regs = cur_regs(env);
3641 u8 opcode = BPF_OP(insn->code);
3642 bool src_known, dst_known;
3643 s64 smin_val, smax_val;
3644 u64 umin_val, umax_val;
3645 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
3646 u32 dst = insn->dst_reg;
3649 if (insn_bitness == 32) {
3650 /* Relevant for 32-bit RSH: Information can propagate towards
3651 * LSB, so it isn't sufficient to only truncate the output to
3654 coerce_reg_to_size(dst_reg, 4);
3655 coerce_reg_to_size(&src_reg, 4);
3658 smin_val = src_reg.smin_value;
3659 smax_val = src_reg.smax_value;
3660 umin_val = src_reg.umin_value;
3661 umax_val = src_reg.umax_value;
3662 src_known = tnum_is_const(src_reg.var_off);
3663 dst_known = tnum_is_const(dst_reg->var_off);
3665 if ((src_known && (smin_val != smax_val || umin_val != umax_val)) ||
3666 smin_val > smax_val || umin_val > umax_val) {
3667 /* Taint dst register if offset had invalid bounds derived from
3668 * e.g. dead branches.
3670 __mark_reg_unknown(dst_reg);
3675 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
3676 __mark_reg_unknown(dst_reg);
3682 ret = sanitize_val_alu(env, insn);
3684 verbose(env, "R%d tried to add from different pointers or scalars\n", dst);
3687 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
3688 signed_add_overflows(dst_reg->smax_value, smax_val)) {
3689 dst_reg->smin_value = S64_MIN;
3690 dst_reg->smax_value = S64_MAX;
3692 dst_reg->smin_value += smin_val;
3693 dst_reg->smax_value += smax_val;
3695 if (dst_reg->umin_value + umin_val < umin_val ||
3696 dst_reg->umax_value + umax_val < umax_val) {
3697 dst_reg->umin_value = 0;
3698 dst_reg->umax_value = U64_MAX;
3700 dst_reg->umin_value += umin_val;
3701 dst_reg->umax_value += umax_val;
3703 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
3706 ret = sanitize_val_alu(env, insn);
3708 verbose(env, "R%d tried to sub from different pointers or scalars\n", dst);
3711 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
3712 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
3713 /* Overflow possible, we know nothing */
3714 dst_reg->smin_value = S64_MIN;
3715 dst_reg->smax_value = S64_MAX;
3717 dst_reg->smin_value -= smax_val;
3718 dst_reg->smax_value -= smin_val;
3720 if (dst_reg->umin_value < umax_val) {
3721 /* Overflow possible, we know nothing */
3722 dst_reg->umin_value = 0;
3723 dst_reg->umax_value = U64_MAX;
3725 /* Cannot overflow (as long as bounds are consistent) */
3726 dst_reg->umin_value -= umax_val;
3727 dst_reg->umax_value -= umin_val;
3729 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
3732 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
3733 if (smin_val < 0 || dst_reg->smin_value < 0) {
3734 /* Ain't nobody got time to multiply that sign */
3735 __mark_reg_unbounded(dst_reg);
3736 __update_reg_bounds(dst_reg);
3739 /* Both values are positive, so we can work with unsigned and
3740 * copy the result to signed (unless it exceeds S64_MAX).
3742 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
3743 /* Potential overflow, we know nothing */
3744 __mark_reg_unbounded(dst_reg);
3745 /* (except what we can learn from the var_off) */
3746 __update_reg_bounds(dst_reg);
3749 dst_reg->umin_value *= umin_val;
3750 dst_reg->umax_value *= umax_val;
3751 if (dst_reg->umax_value > S64_MAX) {
3752 /* Overflow possible, we know nothing */
3753 dst_reg->smin_value = S64_MIN;
3754 dst_reg->smax_value = S64_MAX;
3756 dst_reg->smin_value = dst_reg->umin_value;
3757 dst_reg->smax_value = dst_reg->umax_value;
3761 if (src_known && dst_known) {
3762 __mark_reg_known(dst_reg, dst_reg->var_off.value &
3763 src_reg.var_off.value);
3766 /* We get our minimum from the var_off, since that's inherently
3767 * bitwise. Our maximum is the minimum of the operands' maxima.
3769 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
3770 dst_reg->umin_value = dst_reg->var_off.value;
3771 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
3772 if (dst_reg->smin_value < 0 || smin_val < 0) {
3773 /* Lose signed bounds when ANDing negative numbers,
3774 * ain't nobody got time for that.
3776 dst_reg->smin_value = S64_MIN;
3777 dst_reg->smax_value = S64_MAX;
3779 /* ANDing two positives gives a positive, so safe to
3780 * cast result into s64.
3782 dst_reg->smin_value = dst_reg->umin_value;
3783 dst_reg->smax_value = dst_reg->umax_value;
3785 /* We may learn something more from the var_off */
3786 __update_reg_bounds(dst_reg);
3789 if (src_known && dst_known) {
3790 __mark_reg_known(dst_reg, dst_reg->var_off.value |
3791 src_reg.var_off.value);
3794 /* We get our maximum from the var_off, and our minimum is the
3795 * maximum of the operands' minima
3797 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
3798 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
3799 dst_reg->umax_value = dst_reg->var_off.value |
3800 dst_reg->var_off.mask;
3801 if (dst_reg->smin_value < 0 || smin_val < 0) {
3802 /* Lose signed bounds when ORing negative numbers,
3803 * ain't nobody got time for that.
3805 dst_reg->smin_value = S64_MIN;
3806 dst_reg->smax_value = S64_MAX;
3808 /* ORing two positives gives a positive, so safe to
3809 * cast result into s64.
3811 dst_reg->smin_value = dst_reg->umin_value;
3812 dst_reg->smax_value = dst_reg->umax_value;
3814 /* We may learn something more from the var_off */
3815 __update_reg_bounds(dst_reg);
3818 if (umax_val >= insn_bitness) {
3819 /* Shifts greater than 31 or 63 are undefined.
3820 * This includes shifts by a negative number.
3822 mark_reg_unknown(env, regs, insn->dst_reg);
3825 /* We lose all sign bit information (except what we can pick
3828 dst_reg->smin_value = S64_MIN;
3829 dst_reg->smax_value = S64_MAX;
3830 /* If we might shift our top bit out, then we know nothing */
3831 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
3832 dst_reg->umin_value = 0;
3833 dst_reg->umax_value = U64_MAX;
3835 dst_reg->umin_value <<= umin_val;
3836 dst_reg->umax_value <<= umax_val;
3838 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
3839 /* We may learn something more from the var_off */
3840 __update_reg_bounds(dst_reg);
3843 if (umax_val >= insn_bitness) {
3844 /* Shifts greater than 31 or 63 are undefined.
3845 * This includes shifts by a negative number.
3847 mark_reg_unknown(env, regs, insn->dst_reg);
3850 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
3851 * be negative, then either:
3852 * 1) src_reg might be zero, so the sign bit of the result is
3853 * unknown, so we lose our signed bounds
3854 * 2) it's known negative, thus the unsigned bounds capture the
3856 * 3) the signed bounds cross zero, so they tell us nothing
3858 * If the value in dst_reg is known nonnegative, then again the
3859 * unsigned bounts capture the signed bounds.
3860 * Thus, in all cases it suffices to blow away our signed bounds
3861 * and rely on inferring new ones from the unsigned bounds and
3862 * var_off of the result.
3864 dst_reg->smin_value = S64_MIN;
3865 dst_reg->smax_value = S64_MAX;
3866 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
3867 dst_reg->umin_value >>= umax_val;
3868 dst_reg->umax_value >>= umin_val;
3869 /* We may learn something more from the var_off */
3870 __update_reg_bounds(dst_reg);
3873 if (umax_val >= insn_bitness) {
3874 /* Shifts greater than 31 or 63 are undefined.
3875 * This includes shifts by a negative number.
3877 mark_reg_unknown(env, regs, insn->dst_reg);
3881 /* Upon reaching here, src_known is true and
3882 * umax_val is equal to umin_val.
3884 dst_reg->smin_value >>= umin_val;
3885 dst_reg->smax_value >>= umin_val;
3886 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val);
3888 /* blow away the dst_reg umin_value/umax_value and rely on
3889 * dst_reg var_off to refine the result.
3891 dst_reg->umin_value = 0;
3892 dst_reg->umax_value = U64_MAX;
3893 __update_reg_bounds(dst_reg);
3896 mark_reg_unknown(env, regs, insn->dst_reg);
3900 if (BPF_CLASS(insn->code) != BPF_ALU64) {
3901 /* 32-bit ALU ops are (32,32)->32 */
3902 coerce_reg_to_size(dst_reg, 4);
3905 __reg_deduce_bounds(dst_reg);
3906 __reg_bound_offset(dst_reg);
3910 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
3913 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
3914 struct bpf_insn *insn)
3916 struct bpf_verifier_state *vstate = env->cur_state;
3917 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3918 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
3919 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
3920 u8 opcode = BPF_OP(insn->code);
3922 dst_reg = ®s[insn->dst_reg];
3924 if (dst_reg->type != SCALAR_VALUE)
3926 if (BPF_SRC(insn->code) == BPF_X) {
3927 src_reg = ®s[insn->src_reg];
3928 if (src_reg->type != SCALAR_VALUE) {
3929 if (dst_reg->type != SCALAR_VALUE) {
3930 /* Combining two pointers by any ALU op yields
3931 * an arbitrary scalar. Disallow all math except
3932 * pointer subtraction
3934 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
3935 mark_reg_unknown(env, regs, insn->dst_reg);
3938 verbose(env, "R%d pointer %s pointer prohibited\n",
3940 bpf_alu_string[opcode >> 4]);
3943 /* scalar += pointer
3944 * This is legal, but we have to reverse our
3945 * src/dest handling in computing the range
3947 return adjust_ptr_min_max_vals(env, insn,
3950 } else if (ptr_reg) {
3951 /* pointer += scalar */
3952 return adjust_ptr_min_max_vals(env, insn,
3956 /* Pretend the src is a reg with a known value, since we only
3957 * need to be able to read from this state.
3959 off_reg.type = SCALAR_VALUE;
3960 __mark_reg_known(&off_reg, insn->imm);
3962 if (ptr_reg) /* pointer += K */
3963 return adjust_ptr_min_max_vals(env, insn,
3967 /* Got here implies adding two SCALAR_VALUEs */
3968 if (WARN_ON_ONCE(ptr_reg)) {
3969 print_verifier_state(env, state);
3970 verbose(env, "verifier internal error: unexpected ptr_reg\n");
3973 if (WARN_ON(!src_reg)) {
3974 print_verifier_state(env, state);
3975 verbose(env, "verifier internal error: no src_reg\n");
3978 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
3981 /* check validity of 32-bit and 64-bit arithmetic operations */
3982 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
3984 struct bpf_reg_state *regs = cur_regs(env);
3985 u8 opcode = BPF_OP(insn->code);
3988 if (opcode == BPF_END || opcode == BPF_NEG) {
3989 if (opcode == BPF_NEG) {
3990 if (BPF_SRC(insn->code) != 0 ||
3991 insn->src_reg != BPF_REG_0 ||
3992 insn->off != 0 || insn->imm != 0) {
3993 verbose(env, "BPF_NEG uses reserved fields\n");
3997 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
3998 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
3999 BPF_CLASS(insn->code) == BPF_ALU64) {
4000 verbose(env, "BPF_END uses reserved fields\n");
4005 /* check src operand */
4006 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4010 if (is_pointer_value(env, insn->dst_reg)) {
4011 verbose(env, "R%d pointer arithmetic prohibited\n",
4016 /* check dest operand */
4017 err = check_reg_arg(env, insn->dst_reg, DST_OP);
4021 } else if (opcode == BPF_MOV) {
4023 if (BPF_SRC(insn->code) == BPF_X) {
4024 if (insn->imm != 0 || insn->off != 0) {
4025 verbose(env, "BPF_MOV uses reserved fields\n");
4029 /* check src operand */
4030 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4034 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
4035 verbose(env, "BPF_MOV uses reserved fields\n");
4040 /* check dest operand, mark as required later */
4041 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
4045 if (BPF_SRC(insn->code) == BPF_X) {
4046 struct bpf_reg_state *src_reg = regs + insn->src_reg;
4047 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
4049 if (BPF_CLASS(insn->code) == BPF_ALU64) {
4051 * copy register state to dest reg
4053 *dst_reg = *src_reg;
4054 dst_reg->live |= REG_LIVE_WRITTEN;
4057 if (is_pointer_value(env, insn->src_reg)) {
4059 "R%d partial copy of pointer\n",
4062 } else if (src_reg->type == SCALAR_VALUE) {
4063 *dst_reg = *src_reg;
4064 dst_reg->live |= REG_LIVE_WRITTEN;
4066 mark_reg_unknown(env, regs,
4069 coerce_reg_to_size(dst_reg, 4);
4073 * remember the value we stored into this reg
4075 /* clear any state __mark_reg_known doesn't set */
4076 mark_reg_unknown(env, regs, insn->dst_reg);
4077 regs[insn->dst_reg].type = SCALAR_VALUE;
4078 if (BPF_CLASS(insn->code) == BPF_ALU64) {
4079 __mark_reg_known(regs + insn->dst_reg,
4082 __mark_reg_known(regs + insn->dst_reg,
4087 } else if (opcode > BPF_END) {
4088 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
4091 } else { /* all other ALU ops: and, sub, xor, add, ... */
4093 if (BPF_SRC(insn->code) == BPF_X) {
4094 if (insn->imm != 0 || insn->off != 0) {
4095 verbose(env, "BPF_ALU uses reserved fields\n");
4098 /* check src1 operand */
4099 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4103 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
4104 verbose(env, "BPF_ALU uses reserved fields\n");
4109 /* check src2 operand */
4110 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4114 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
4115 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
4116 verbose(env, "div by zero\n");
4120 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
4121 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
4122 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
4124 if (insn->imm < 0 || insn->imm >= size) {
4125 verbose(env, "invalid shift %d\n", insn->imm);
4130 /* check dest operand */
4131 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
4135 return adjust_reg_min_max_vals(env, insn);
4141 static void __find_good_pkt_pointers(struct bpf_func_state *state,
4142 struct bpf_reg_state *dst_reg,
4143 enum bpf_reg_type type, u16 new_range)
4145 struct bpf_reg_state *reg;
4148 for (i = 0; i < MAX_BPF_REG; i++) {
4149 reg = &state->regs[i];
4150 if (reg->type == type && reg->id == dst_reg->id)
4151 /* keep the maximum range already checked */
4152 reg->range = max(reg->range, new_range);
4155 bpf_for_each_spilled_reg(i, state, reg) {
4158 if (reg->type == type && reg->id == dst_reg->id)
4159 reg->range = max(reg->range, new_range);
4163 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
4164 struct bpf_reg_state *dst_reg,
4165 enum bpf_reg_type type,
4166 bool range_right_open)
4171 if (dst_reg->off < 0 ||
4172 (dst_reg->off == 0 && range_right_open))
4173 /* This doesn't give us any range */
4176 if (dst_reg->umax_value > MAX_PACKET_OFF ||
4177 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
4178 /* Risk of overflow. For instance, ptr + (1<<63) may be less
4179 * than pkt_end, but that's because it's also less than pkt.
4183 new_range = dst_reg->off;
4184 if (range_right_open)
4187 /* Examples for register markings:
4189 * pkt_data in dst register:
4193 * if (r2 > pkt_end) goto <handle exception>
4198 * if (r2 < pkt_end) goto <access okay>
4199 * <handle exception>
4202 * r2 == dst_reg, pkt_end == src_reg
4203 * r2=pkt(id=n,off=8,r=0)
4204 * r3=pkt(id=n,off=0,r=0)
4206 * pkt_data in src register:
4210 * if (pkt_end >= r2) goto <access okay>
4211 * <handle exception>
4215 * if (pkt_end <= r2) goto <handle exception>
4219 * pkt_end == dst_reg, r2 == src_reg
4220 * r2=pkt(id=n,off=8,r=0)
4221 * r3=pkt(id=n,off=0,r=0)
4223 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
4224 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
4225 * and [r3, r3 + 8-1) respectively is safe to access depending on
4229 /* If our ids match, then we must have the same max_value. And we
4230 * don't care about the other reg's fixed offset, since if it's too big
4231 * the range won't allow anything.
4232 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
4234 for (i = 0; i <= vstate->curframe; i++)
4235 __find_good_pkt_pointers(vstate->frame[i], dst_reg, type,
4239 /* compute branch direction of the expression "if (reg opcode val) goto target;"
4241 * 1 - branch will be taken and "goto target" will be executed
4242 * 0 - branch will not be taken and fall-through to next insn
4243 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value range [0,10]
4245 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
4248 struct bpf_reg_state reg_lo;
4251 if (__is_pointer_value(false, reg))
4257 /* For JMP32, only low 32 bits are compared, coerce_reg_to_size
4258 * could truncate high bits and update umin/umax according to
4259 * information of low bits.
4261 coerce_reg_to_size(reg, 4);
4262 /* smin/smax need special handling. For example, after coerce,
4263 * if smin_value is 0x00000000ffffffffLL, the value is -1 when
4264 * used as operand to JMP32. It is a negative number from s32's
4265 * point of view, while it is a positive number when seen as
4266 * s64. The smin/smax are kept as s64, therefore, when used with
4267 * JMP32, they need to be transformed into s32, then sign
4268 * extended back to s64.
4270 * Also, smin/smax were copied from umin/umax. If umin/umax has
4271 * different sign bit, then min/max relationship doesn't
4272 * maintain after casting into s32, for this case, set smin/smax
4275 if ((reg->umax_value ^ reg->umin_value) &
4277 reg->smin_value = S32_MIN;
4278 reg->smax_value = S32_MAX;
4280 reg->smin_value = (s64)(s32)reg->smin_value;
4281 reg->smax_value = (s64)(s32)reg->smax_value;
4284 sval = (s64)(s32)val;
4291 if (tnum_is_const(reg->var_off))
4292 return !!tnum_equals_const(reg->var_off, val);
4295 if (tnum_is_const(reg->var_off))
4296 return !tnum_equals_const(reg->var_off, val);
4299 if ((~reg->var_off.mask & reg->var_off.value) & val)
4301 if (!((reg->var_off.mask | reg->var_off.value) & val))
4305 if (reg->umin_value > val)
4307 else if (reg->umax_value <= val)
4311 if (reg->smin_value > sval)
4313 else if (reg->smax_value < sval)
4317 if (reg->umax_value < val)
4319 else if (reg->umin_value >= val)
4323 if (reg->smax_value < sval)
4325 else if (reg->smin_value >= sval)
4329 if (reg->umin_value >= val)
4331 else if (reg->umax_value < val)
4335 if (reg->smin_value >= sval)
4337 else if (reg->smax_value < sval)
4341 if (reg->umax_value <= val)
4343 else if (reg->umin_value > val)
4347 if (reg->smax_value <= sval)
4349 else if (reg->smin_value > sval)
4357 /* Generate min value of the high 32-bit from TNUM info. */
4358 static u64 gen_hi_min(struct tnum var)
4360 return var.value & ~0xffffffffULL;
4363 /* Generate max value of the high 32-bit from TNUM info. */
4364 static u64 gen_hi_max(struct tnum var)
4366 return (var.value | var.mask) & ~0xffffffffULL;
4369 /* Return true if VAL is compared with a s64 sign extended from s32, and they
4370 * are with the same signedness.
4372 static bool cmp_val_with_extended_s64(s64 sval, struct bpf_reg_state *reg)
4374 return ((s32)sval >= 0 &&
4375 reg->smin_value >= 0 && reg->smax_value <= S32_MAX) ||
4377 reg->smax_value <= 0 && reg->smin_value >= S32_MIN);
4380 /* Adjusts the register min/max values in the case that the dst_reg is the
4381 * variable register that we are working on, and src_reg is a constant or we're
4382 * simply doing a BPF_K check.
4383 * In JEQ/JNE cases we also adjust the var_off values.
4385 static void reg_set_min_max(struct bpf_reg_state *true_reg,
4386 struct bpf_reg_state *false_reg, u64 val,
4387 u8 opcode, bool is_jmp32)
4391 /* If the dst_reg is a pointer, we can't learn anything about its
4392 * variable offset from the compare (unless src_reg were a pointer into
4393 * the same object, but we don't bother with that.
4394 * Since false_reg and true_reg have the same type by construction, we
4395 * only need to check one of them for pointerness.
4397 if (__is_pointer_value(false, false_reg))
4400 val = is_jmp32 ? (u32)val : val;
4401 sval = is_jmp32 ? (s64)(s32)val : (s64)val;
4407 struct bpf_reg_state *reg =
4408 opcode == BPF_JEQ ? true_reg : false_reg;
4410 /* For BPF_JEQ, if this is false we know nothing Jon Snow, but
4411 * if it is true we know the value for sure. Likewise for
4415 u64 old_v = reg->var_off.value;
4416 u64 hi_mask = ~0xffffffffULL;
4418 reg->var_off.value = (old_v & hi_mask) | val;
4419 reg->var_off.mask &= hi_mask;
4421 __mark_reg_known(reg, val);
4426 false_reg->var_off = tnum_and(false_reg->var_off,
4428 if (is_power_of_2(val))
4429 true_reg->var_off = tnum_or(true_reg->var_off,
4435 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
4436 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
4439 false_umax += gen_hi_max(false_reg->var_off);
4440 true_umin += gen_hi_min(true_reg->var_off);
4442 false_reg->umax_value = min(false_reg->umax_value, false_umax);
4443 true_reg->umin_value = max(true_reg->umin_value, true_umin);
4449 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
4450 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
4452 /* If the full s64 was not sign-extended from s32 then don't
4453 * deduct further info.
4455 if (is_jmp32 && !cmp_val_with_extended_s64(sval, false_reg))
4457 false_reg->smax_value = min(false_reg->smax_value, false_smax);
4458 true_reg->smin_value = max(true_reg->smin_value, true_smin);
4464 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
4465 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
4468 false_umin += gen_hi_min(false_reg->var_off);
4469 true_umax += gen_hi_max(true_reg->var_off);
4471 false_reg->umin_value = max(false_reg->umin_value, false_umin);
4472 true_reg->umax_value = min(true_reg->umax_value, true_umax);
4478 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
4479 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
4481 if (is_jmp32 && !cmp_val_with_extended_s64(sval, false_reg))
4483 false_reg->smin_value = max(false_reg->smin_value, false_smin);
4484 true_reg->smax_value = min(true_reg->smax_value, true_smax);
4491 __reg_deduce_bounds(false_reg);
4492 __reg_deduce_bounds(true_reg);
4493 /* We might have learned some bits from the bounds. */
4494 __reg_bound_offset(false_reg);
4495 __reg_bound_offset(true_reg);
4496 /* Intersecting with the old var_off might have improved our bounds
4497 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
4498 * then new var_off is (0; 0x7f...fc) which improves our umax.
4500 __update_reg_bounds(false_reg);
4501 __update_reg_bounds(true_reg);
4504 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
4507 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
4508 struct bpf_reg_state *false_reg, u64 val,
4509 u8 opcode, bool is_jmp32)
4513 if (__is_pointer_value(false, false_reg))
4516 val = is_jmp32 ? (u32)val : val;
4517 sval = is_jmp32 ? (s64)(s32)val : (s64)val;
4523 struct bpf_reg_state *reg =
4524 opcode == BPF_JEQ ? true_reg : false_reg;
4527 u64 old_v = reg->var_off.value;
4528 u64 hi_mask = ~0xffffffffULL;
4530 reg->var_off.value = (old_v & hi_mask) | val;
4531 reg->var_off.mask &= hi_mask;
4533 __mark_reg_known(reg, val);
4538 false_reg->var_off = tnum_and(false_reg->var_off,
4540 if (is_power_of_2(val))
4541 true_reg->var_off = tnum_or(true_reg->var_off,
4547 u64 false_umin = opcode == BPF_JGT ? val : val + 1;
4548 u64 true_umax = opcode == BPF_JGT ? val - 1 : val;
4551 false_umin += gen_hi_min(false_reg->var_off);
4552 true_umax += gen_hi_max(true_reg->var_off);
4554 false_reg->umin_value = max(false_reg->umin_value, false_umin);
4555 true_reg->umax_value = min(true_reg->umax_value, true_umax);
4561 s64 false_smin = opcode == BPF_JSGT ? sval : sval + 1;
4562 s64 true_smax = opcode == BPF_JSGT ? sval - 1 : sval;
4564 if (is_jmp32 && !cmp_val_with_extended_s64(sval, false_reg))
4566 false_reg->smin_value = max(false_reg->smin_value, false_smin);
4567 true_reg->smax_value = min(true_reg->smax_value, true_smax);
4573 u64 false_umax = opcode == BPF_JLT ? val : val - 1;
4574 u64 true_umin = opcode == BPF_JLT ? val + 1 : val;
4577 false_umax += gen_hi_max(false_reg->var_off);
4578 true_umin += gen_hi_min(true_reg->var_off);
4580 false_reg->umax_value = min(false_reg->umax_value, false_umax);
4581 true_reg->umin_value = max(true_reg->umin_value, true_umin);
4587 s64 false_smax = opcode == BPF_JSLT ? sval : sval - 1;
4588 s64 true_smin = opcode == BPF_JSLT ? sval + 1 : sval;
4590 if (is_jmp32 && !cmp_val_with_extended_s64(sval, false_reg))
4592 false_reg->smax_value = min(false_reg->smax_value, false_smax);
4593 true_reg->smin_value = max(true_reg->smin_value, true_smin);
4600 __reg_deduce_bounds(false_reg);
4601 __reg_deduce_bounds(true_reg);
4602 /* We might have learned some bits from the bounds. */
4603 __reg_bound_offset(false_reg);
4604 __reg_bound_offset(true_reg);
4605 /* Intersecting with the old var_off might have improved our bounds
4606 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
4607 * then new var_off is (0; 0x7f...fc) which improves our umax.
4609 __update_reg_bounds(false_reg);
4610 __update_reg_bounds(true_reg);
4613 /* Regs are known to be equal, so intersect their min/max/var_off */
4614 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
4615 struct bpf_reg_state *dst_reg)
4617 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
4618 dst_reg->umin_value);
4619 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
4620 dst_reg->umax_value);
4621 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
4622 dst_reg->smin_value);
4623 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
4624 dst_reg->smax_value);
4625 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
4627 /* We might have learned new bounds from the var_off. */
4628 __update_reg_bounds(src_reg);
4629 __update_reg_bounds(dst_reg);
4630 /* We might have learned something about the sign bit. */
4631 __reg_deduce_bounds(src_reg);
4632 __reg_deduce_bounds(dst_reg);
4633 /* We might have learned some bits from the bounds. */
4634 __reg_bound_offset(src_reg);
4635 __reg_bound_offset(dst_reg);
4636 /* Intersecting with the old var_off might have improved our bounds
4637 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
4638 * then new var_off is (0; 0x7f...fc) which improves our umax.
4640 __update_reg_bounds(src_reg);
4641 __update_reg_bounds(dst_reg);
4644 static void reg_combine_min_max(struct bpf_reg_state *true_src,
4645 struct bpf_reg_state *true_dst,
4646 struct bpf_reg_state *false_src,
4647 struct bpf_reg_state *false_dst,
4652 __reg_combine_min_max(true_src, true_dst);
4655 __reg_combine_min_max(false_src, false_dst);
4660 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
4661 struct bpf_reg_state *reg, u32 id,
4664 if (reg_type_may_be_null(reg->type) && reg->id == id) {
4665 /* Old offset (both fixed and variable parts) should
4666 * have been known-zero, because we don't allow pointer
4667 * arithmetic on pointers that might be NULL.
4669 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
4670 !tnum_equals_const(reg->var_off, 0) ||
4672 __mark_reg_known_zero(reg);
4676 reg->type = SCALAR_VALUE;
4677 } else if (reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
4678 if (reg->map_ptr->inner_map_meta) {
4679 reg->type = CONST_PTR_TO_MAP;
4680 reg->map_ptr = reg->map_ptr->inner_map_meta;
4682 reg->type = PTR_TO_MAP_VALUE;
4684 } else if (reg->type == PTR_TO_SOCKET_OR_NULL) {
4685 reg->type = PTR_TO_SOCKET;
4686 } else if (reg->type == PTR_TO_SOCK_COMMON_OR_NULL) {
4687 reg->type = PTR_TO_SOCK_COMMON;
4688 } else if (reg->type == PTR_TO_TCP_SOCK_OR_NULL) {
4689 reg->type = PTR_TO_TCP_SOCK;
4692 /* We don't need id and ref_obj_id from this point
4693 * onwards anymore, thus we should better reset it,
4694 * so that state pruning has chances to take effect.
4697 reg->ref_obj_id = 0;
4698 } else if (!reg_may_point_to_spin_lock(reg)) {
4699 /* For not-NULL ptr, reg->ref_obj_id will be reset
4700 * in release_reg_references().
4702 * reg->id is still used by spin_lock ptr. Other
4703 * than spin_lock ptr type, reg->id can be reset.
4710 static void __mark_ptr_or_null_regs(struct bpf_func_state *state, u32 id,
4713 struct bpf_reg_state *reg;
4716 for (i = 0; i < MAX_BPF_REG; i++)
4717 mark_ptr_or_null_reg(state, &state->regs[i], id, is_null);
4719 bpf_for_each_spilled_reg(i, state, reg) {
4722 mark_ptr_or_null_reg(state, reg, id, is_null);
4726 /* The logic is similar to find_good_pkt_pointers(), both could eventually
4727 * be folded together at some point.
4729 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
4732 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4733 struct bpf_reg_state *regs = state->regs;
4734 u32 ref_obj_id = regs[regno].ref_obj_id;
4735 u32 id = regs[regno].id;
4738 if (ref_obj_id && ref_obj_id == id && is_null)
4739 /* regs[regno] is in the " == NULL" branch.
4740 * No one could have freed the reference state before
4741 * doing the NULL check.
4743 WARN_ON_ONCE(release_reference_state(state, id));
4745 for (i = 0; i <= vstate->curframe; i++)
4746 __mark_ptr_or_null_regs(vstate->frame[i], id, is_null);
4749 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
4750 struct bpf_reg_state *dst_reg,
4751 struct bpf_reg_state *src_reg,
4752 struct bpf_verifier_state *this_branch,
4753 struct bpf_verifier_state *other_branch)
4755 if (BPF_SRC(insn->code) != BPF_X)
4758 /* Pointers are always 64-bit. */
4759 if (BPF_CLASS(insn->code) == BPF_JMP32)
4762 switch (BPF_OP(insn->code)) {
4764 if ((dst_reg->type == PTR_TO_PACKET &&
4765 src_reg->type == PTR_TO_PACKET_END) ||
4766 (dst_reg->type == PTR_TO_PACKET_META &&
4767 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
4768 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
4769 find_good_pkt_pointers(this_branch, dst_reg,
4770 dst_reg->type, false);
4771 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
4772 src_reg->type == PTR_TO_PACKET) ||
4773 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
4774 src_reg->type == PTR_TO_PACKET_META)) {
4775 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
4776 find_good_pkt_pointers(other_branch, src_reg,
4777 src_reg->type, true);
4783 if ((dst_reg->type == PTR_TO_PACKET &&
4784 src_reg->type == PTR_TO_PACKET_END) ||
4785 (dst_reg->type == PTR_TO_PACKET_META &&
4786 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
4787 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
4788 find_good_pkt_pointers(other_branch, dst_reg,
4789 dst_reg->type, true);
4790 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
4791 src_reg->type == PTR_TO_PACKET) ||
4792 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
4793 src_reg->type == PTR_TO_PACKET_META)) {
4794 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
4795 find_good_pkt_pointers(this_branch, src_reg,
4796 src_reg->type, false);
4802 if ((dst_reg->type == PTR_TO_PACKET &&
4803 src_reg->type == PTR_TO_PACKET_END) ||
4804 (dst_reg->type == PTR_TO_PACKET_META &&
4805 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
4806 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
4807 find_good_pkt_pointers(this_branch, dst_reg,
4808 dst_reg->type, true);
4809 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
4810 src_reg->type == PTR_TO_PACKET) ||
4811 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
4812 src_reg->type == PTR_TO_PACKET_META)) {
4813 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
4814 find_good_pkt_pointers(other_branch, src_reg,
4815 src_reg->type, false);
4821 if ((dst_reg->type == PTR_TO_PACKET &&
4822 src_reg->type == PTR_TO_PACKET_END) ||
4823 (dst_reg->type == PTR_TO_PACKET_META &&
4824 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
4825 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
4826 find_good_pkt_pointers(other_branch, dst_reg,
4827 dst_reg->type, false);
4828 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
4829 src_reg->type == PTR_TO_PACKET) ||
4830 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
4831 src_reg->type == PTR_TO_PACKET_META)) {
4832 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
4833 find_good_pkt_pointers(this_branch, src_reg,
4834 src_reg->type, true);
4846 static int check_cond_jmp_op(struct bpf_verifier_env *env,
4847 struct bpf_insn *insn, int *insn_idx)
4849 struct bpf_verifier_state *this_branch = env->cur_state;
4850 struct bpf_verifier_state *other_branch;
4851 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
4852 struct bpf_reg_state *dst_reg, *other_branch_regs;
4853 u8 opcode = BPF_OP(insn->code);
4857 /* Only conditional jumps are expected to reach here. */
4858 if (opcode == BPF_JA || opcode > BPF_JSLE) {
4859 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
4863 if (BPF_SRC(insn->code) == BPF_X) {
4864 if (insn->imm != 0) {
4865 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
4869 /* check src1 operand */
4870 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4874 if (is_pointer_value(env, insn->src_reg)) {
4875 verbose(env, "R%d pointer comparison prohibited\n",
4880 if (insn->src_reg != BPF_REG_0) {
4881 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
4886 /* check src2 operand */
4887 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4891 dst_reg = ®s[insn->dst_reg];
4892 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
4894 if (BPF_SRC(insn->code) == BPF_K) {
4895 int pred = is_branch_taken(dst_reg, insn->imm, opcode,
4899 /* only follow the goto, ignore fall-through */
4900 *insn_idx += insn->off;
4902 } else if (pred == 0) {
4903 /* only follow fall-through branch, since
4904 * that's where the program will go
4910 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
4914 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
4916 /* detect if we are comparing against a constant value so we can adjust
4917 * our min/max values for our dst register.
4918 * this is only legit if both are scalars (or pointers to the same
4919 * object, I suppose, but we don't support that right now), because
4920 * otherwise the different base pointers mean the offsets aren't
4923 if (BPF_SRC(insn->code) == BPF_X) {
4924 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
4925 struct bpf_reg_state lo_reg0 = *dst_reg;
4926 struct bpf_reg_state lo_reg1 = *src_reg;
4927 struct bpf_reg_state *src_lo, *dst_lo;
4931 coerce_reg_to_size(dst_lo, 4);
4932 coerce_reg_to_size(src_lo, 4);
4934 if (dst_reg->type == SCALAR_VALUE &&
4935 src_reg->type == SCALAR_VALUE) {
4936 if (tnum_is_const(src_reg->var_off) ||
4937 (is_jmp32 && tnum_is_const(src_lo->var_off)))
4938 reg_set_min_max(&other_branch_regs[insn->dst_reg],
4941 ? src_lo->var_off.value
4942 : src_reg->var_off.value,
4944 else if (tnum_is_const(dst_reg->var_off) ||
4945 (is_jmp32 && tnum_is_const(dst_lo->var_off)))
4946 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
4949 ? dst_lo->var_off.value
4950 : dst_reg->var_off.value,
4952 else if (!is_jmp32 &&
4953 (opcode == BPF_JEQ || opcode == BPF_JNE))
4954 /* Comparing for equality, we can combine knowledge */
4955 reg_combine_min_max(&other_branch_regs[insn->src_reg],
4956 &other_branch_regs[insn->dst_reg],
4957 src_reg, dst_reg, opcode);
4959 } else if (dst_reg->type == SCALAR_VALUE) {
4960 reg_set_min_max(&other_branch_regs[insn->dst_reg],
4961 dst_reg, insn->imm, opcode, is_jmp32);
4964 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
4965 * NOTE: these optimizations below are related with pointer comparison
4966 * which will never be JMP32.
4968 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
4969 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
4970 reg_type_may_be_null(dst_reg->type)) {
4971 /* Mark all identical registers in each branch as either
4972 * safe or unknown depending R == 0 or R != 0 conditional.
4974 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
4976 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
4978 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
4979 this_branch, other_branch) &&
4980 is_pointer_value(env, insn->dst_reg)) {
4981 verbose(env, "R%d pointer comparison prohibited\n",
4986 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
4990 /* return the map pointer stored inside BPF_LD_IMM64 instruction */
4991 static struct bpf_map *ld_imm64_to_map_ptr(struct bpf_insn *insn)
4993 u64 imm64 = ((u64) (u32) insn[0].imm) | ((u64) (u32) insn[1].imm) << 32;
4995 return (struct bpf_map *) (unsigned long) imm64;
4998 /* verify BPF_LD_IMM64 instruction */
4999 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
5001 struct bpf_reg_state *regs = cur_regs(env);
5004 if (BPF_SIZE(insn->code) != BPF_DW) {
5005 verbose(env, "invalid BPF_LD_IMM insn\n");
5008 if (insn->off != 0) {
5009 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
5013 err = check_reg_arg(env, insn->dst_reg, DST_OP);
5017 if (insn->src_reg == 0) {
5018 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
5020 regs[insn->dst_reg].type = SCALAR_VALUE;
5021 __mark_reg_known(®s[insn->dst_reg], imm);
5025 /* replace_map_fd_with_map_ptr() should have caught bad ld_imm64 */
5026 BUG_ON(insn->src_reg != BPF_PSEUDO_MAP_FD);
5028 regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
5029 regs[insn->dst_reg].map_ptr = ld_imm64_to_map_ptr(insn);
5033 static bool may_access_skb(enum bpf_prog_type type)
5036 case BPF_PROG_TYPE_SOCKET_FILTER:
5037 case BPF_PROG_TYPE_SCHED_CLS:
5038 case BPF_PROG_TYPE_SCHED_ACT:
5045 /* verify safety of LD_ABS|LD_IND instructions:
5046 * - they can only appear in the programs where ctx == skb
5047 * - since they are wrappers of function calls, they scratch R1-R5 registers,
5048 * preserve R6-R9, and store return value into R0
5051 * ctx == skb == R6 == CTX
5054 * SRC == any register
5055 * IMM == 32-bit immediate
5058 * R0 - 8/16/32-bit skb data converted to cpu endianness
5060 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
5062 struct bpf_reg_state *regs = cur_regs(env);
5063 u8 mode = BPF_MODE(insn->code);
5066 if (!may_access_skb(env->prog->type)) {
5067 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
5071 if (!env->ops->gen_ld_abs) {
5072 verbose(env, "bpf verifier is misconfigured\n");
5076 if (env->subprog_cnt > 1) {
5077 /* when program has LD_ABS insn JITs and interpreter assume
5078 * that r1 == ctx == skb which is not the case for callees
5079 * that can have arbitrary arguments. It's problematic
5080 * for main prog as well since JITs would need to analyze
5081 * all functions in order to make proper register save/restore
5082 * decisions in the main prog. Hence disallow LD_ABS with calls
5084 verbose(env, "BPF_LD_[ABS|IND] instructions cannot be mixed with bpf-to-bpf calls\n");
5088 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
5089 BPF_SIZE(insn->code) == BPF_DW ||
5090 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
5091 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
5095 /* check whether implicit source operand (register R6) is readable */
5096 err = check_reg_arg(env, BPF_REG_6, SRC_OP);
5100 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
5101 * gen_ld_abs() may terminate the program at runtime, leading to
5104 err = check_reference_leak(env);
5106 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
5110 if (env->cur_state->active_spin_lock) {
5111 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
5115 if (regs[BPF_REG_6].type != PTR_TO_CTX) {
5117 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
5121 if (mode == BPF_IND) {
5122 /* check explicit source operand */
5123 err = check_reg_arg(env, insn->src_reg, SRC_OP);
5128 /* reset caller saved regs to unreadable */
5129 for (i = 0; i < CALLER_SAVED_REGS; i++) {
5130 mark_reg_not_init(env, regs, caller_saved[i]);
5131 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
5134 /* mark destination R0 register as readable, since it contains
5135 * the value fetched from the packet.
5136 * Already marked as written above.
5138 mark_reg_unknown(env, regs, BPF_REG_0);
5142 static int check_return_code(struct bpf_verifier_env *env)
5144 struct bpf_reg_state *reg;
5145 struct tnum range = tnum_range(0, 1);
5147 switch (env->prog->type) {
5148 case BPF_PROG_TYPE_CGROUP_SKB:
5149 case BPF_PROG_TYPE_CGROUP_SOCK:
5150 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
5151 case BPF_PROG_TYPE_SOCK_OPS:
5152 case BPF_PROG_TYPE_CGROUP_DEVICE:
5158 reg = cur_regs(env) + BPF_REG_0;
5159 if (reg->type != SCALAR_VALUE) {
5160 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
5161 reg_type_str[reg->type]);
5165 if (!tnum_in(range, reg->var_off)) {
5166 verbose(env, "At program exit the register R0 ");
5167 if (!tnum_is_unknown(reg->var_off)) {
5170 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5171 verbose(env, "has value %s", tn_buf);
5173 verbose(env, "has unknown scalar value");
5175 verbose(env, " should have been 0 or 1\n");
5181 /* non-recursive DFS pseudo code
5182 * 1 procedure DFS-iterative(G,v):
5183 * 2 label v as discovered
5184 * 3 let S be a stack
5186 * 5 while S is not empty
5188 * 7 if t is what we're looking for:
5190 * 9 for all edges e in G.adjacentEdges(t) do
5191 * 10 if edge e is already labelled
5192 * 11 continue with the next edge
5193 * 12 w <- G.adjacentVertex(t,e)
5194 * 13 if vertex w is not discovered and not explored
5195 * 14 label e as tree-edge
5196 * 15 label w as discovered
5199 * 18 else if vertex w is discovered
5200 * 19 label e as back-edge
5202 * 21 // vertex w is explored
5203 * 22 label e as forward- or cross-edge
5204 * 23 label t as explored
5209 * 0x11 - discovered and fall-through edge labelled
5210 * 0x12 - discovered and fall-through and branch edges labelled
5221 #define STATE_LIST_MARK ((struct bpf_verifier_state_list *) -1L)
5223 static int *insn_stack; /* stack of insns to process */
5224 static int cur_stack; /* current stack index */
5225 static int *insn_state;
5227 /* t, w, e - match pseudo-code above:
5228 * t - index of current instruction
5229 * w - next instruction
5232 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env)
5234 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
5237 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
5240 if (w < 0 || w >= env->prog->len) {
5241 verbose_linfo(env, t, "%d: ", t);
5242 verbose(env, "jump out of range from insn %d to %d\n", t, w);
5247 /* mark branch target for state pruning */
5248 env->explored_states[w] = STATE_LIST_MARK;
5250 if (insn_state[w] == 0) {
5252 insn_state[t] = DISCOVERED | e;
5253 insn_state[w] = DISCOVERED;
5254 if (cur_stack >= env->prog->len)
5256 insn_stack[cur_stack++] = w;
5258 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
5259 verbose_linfo(env, t, "%d: ", t);
5260 verbose_linfo(env, w, "%d: ", w);
5261 verbose(env, "back-edge from insn %d to %d\n", t, w);
5263 } else if (insn_state[w] == EXPLORED) {
5264 /* forward- or cross-edge */
5265 insn_state[t] = DISCOVERED | e;
5267 verbose(env, "insn state internal bug\n");
5273 /* non-recursive depth-first-search to detect loops in BPF program
5274 * loop == back-edge in directed graph
5276 static int check_cfg(struct bpf_verifier_env *env)
5278 struct bpf_insn *insns = env->prog->insnsi;
5279 int insn_cnt = env->prog->len;
5283 insn_state = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
5287 insn_stack = kcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
5293 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
5294 insn_stack[0] = 0; /* 0 is the first instruction */
5300 t = insn_stack[cur_stack - 1];
5302 if (BPF_CLASS(insns[t].code) == BPF_JMP ||
5303 BPF_CLASS(insns[t].code) == BPF_JMP32) {
5304 u8 opcode = BPF_OP(insns[t].code);
5306 if (opcode == BPF_EXIT) {
5308 } else if (opcode == BPF_CALL) {
5309 ret = push_insn(t, t + 1, FALLTHROUGH, env);
5314 if (t + 1 < insn_cnt)
5315 env->explored_states[t + 1] = STATE_LIST_MARK;
5316 if (insns[t].src_reg == BPF_PSEUDO_CALL) {
5317 env->explored_states[t] = STATE_LIST_MARK;
5318 ret = push_insn(t, t + insns[t].imm + 1, BRANCH, env);
5324 } else if (opcode == BPF_JA) {
5325 if (BPF_SRC(insns[t].code) != BPF_K) {
5329 /* unconditional jump with single edge */
5330 ret = push_insn(t, t + insns[t].off + 1,
5336 /* tell verifier to check for equivalent states
5337 * after every call and jump
5339 if (t + 1 < insn_cnt)
5340 env->explored_states[t + 1] = STATE_LIST_MARK;
5342 /* conditional jump with two edges */
5343 env->explored_states[t] = STATE_LIST_MARK;
5344 ret = push_insn(t, t + 1, FALLTHROUGH, env);
5350 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env);
5357 /* all other non-branch instructions with single
5360 ret = push_insn(t, t + 1, FALLTHROUGH, env);
5368 insn_state[t] = EXPLORED;
5369 if (cur_stack-- <= 0) {
5370 verbose(env, "pop stack internal bug\n");
5377 for (i = 0; i < insn_cnt; i++) {
5378 if (insn_state[i] != EXPLORED) {
5379 verbose(env, "unreachable insn %d\n", i);
5384 ret = 0; /* cfg looks good */
5392 /* The minimum supported BTF func info size */
5393 #define MIN_BPF_FUNCINFO_SIZE 8
5394 #define MAX_FUNCINFO_REC_SIZE 252
5396 static int check_btf_func(struct bpf_verifier_env *env,
5397 const union bpf_attr *attr,
5398 union bpf_attr __user *uattr)
5400 u32 i, nfuncs, urec_size, min_size;
5401 u32 krec_size = sizeof(struct bpf_func_info);
5402 struct bpf_func_info *krecord;
5403 const struct btf_type *type;
5404 struct bpf_prog *prog;
5405 const struct btf *btf;
5406 void __user *urecord;
5407 u32 prev_offset = 0;
5410 nfuncs = attr->func_info_cnt;
5414 if (nfuncs != env->subprog_cnt) {
5415 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
5419 urec_size = attr->func_info_rec_size;
5420 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
5421 urec_size > MAX_FUNCINFO_REC_SIZE ||
5422 urec_size % sizeof(u32)) {
5423 verbose(env, "invalid func info rec size %u\n", urec_size);
5428 btf = prog->aux->btf;
5430 urecord = u64_to_user_ptr(attr->func_info);
5431 min_size = min_t(u32, krec_size, urec_size);
5433 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
5437 for (i = 0; i < nfuncs; i++) {
5438 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
5440 if (ret == -E2BIG) {
5441 verbose(env, "nonzero tailing record in func info");
5442 /* set the size kernel expects so loader can zero
5443 * out the rest of the record.
5445 if (put_user(min_size, &uattr->func_info_rec_size))
5451 if (copy_from_user(&krecord[i], urecord, min_size)) {
5456 /* check insn_off */
5458 if (krecord[i].insn_off) {
5460 "nonzero insn_off %u for the first func info record",
5461 krecord[i].insn_off);
5465 } else if (krecord[i].insn_off <= prev_offset) {
5467 "same or smaller insn offset (%u) than previous func info record (%u)",
5468 krecord[i].insn_off, prev_offset);
5473 if (env->subprog_info[i].start != krecord[i].insn_off) {
5474 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
5480 type = btf_type_by_id(btf, krecord[i].type_id);
5481 if (!type || BTF_INFO_KIND(type->info) != BTF_KIND_FUNC) {
5482 verbose(env, "invalid type id %d in func info",
5483 krecord[i].type_id);
5488 prev_offset = krecord[i].insn_off;
5489 urecord += urec_size;
5492 prog->aux->func_info = krecord;
5493 prog->aux->func_info_cnt = nfuncs;
5501 static void adjust_btf_func(struct bpf_verifier_env *env)
5505 if (!env->prog->aux->func_info)
5508 for (i = 0; i < env->subprog_cnt; i++)
5509 env->prog->aux->func_info[i].insn_off = env->subprog_info[i].start;
5512 #define MIN_BPF_LINEINFO_SIZE (offsetof(struct bpf_line_info, line_col) + \
5513 sizeof(((struct bpf_line_info *)(0))->line_col))
5514 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
5516 static int check_btf_line(struct bpf_verifier_env *env,
5517 const union bpf_attr *attr,
5518 union bpf_attr __user *uattr)
5520 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
5521 struct bpf_subprog_info *sub;
5522 struct bpf_line_info *linfo;
5523 struct bpf_prog *prog;
5524 const struct btf *btf;
5525 void __user *ulinfo;
5528 nr_linfo = attr->line_info_cnt;
5532 rec_size = attr->line_info_rec_size;
5533 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
5534 rec_size > MAX_LINEINFO_REC_SIZE ||
5535 rec_size & (sizeof(u32) - 1))
5538 /* Need to zero it in case the userspace may
5539 * pass in a smaller bpf_line_info object.
5541 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
5542 GFP_KERNEL | __GFP_NOWARN);
5547 btf = prog->aux->btf;
5550 sub = env->subprog_info;
5551 ulinfo = u64_to_user_ptr(attr->line_info);
5552 expected_size = sizeof(struct bpf_line_info);
5553 ncopy = min_t(u32, expected_size, rec_size);
5554 for (i = 0; i < nr_linfo; i++) {
5555 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
5557 if (err == -E2BIG) {
5558 verbose(env, "nonzero tailing record in line_info");
5559 if (put_user(expected_size,
5560 &uattr->line_info_rec_size))
5566 if (copy_from_user(&linfo[i], ulinfo, ncopy)) {
5572 * Check insn_off to ensure
5573 * 1) strictly increasing AND
5574 * 2) bounded by prog->len
5576 * The linfo[0].insn_off == 0 check logically falls into
5577 * the later "missing bpf_line_info for func..." case
5578 * because the first linfo[0].insn_off must be the
5579 * first sub also and the first sub must have
5580 * subprog_info[0].start == 0.
5582 if ((i && linfo[i].insn_off <= prev_offset) ||
5583 linfo[i].insn_off >= prog->len) {
5584 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
5585 i, linfo[i].insn_off, prev_offset,
5591 if (!prog->insnsi[linfo[i].insn_off].code) {
5593 "Invalid insn code at line_info[%u].insn_off\n",
5599 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
5600 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
5601 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
5606 if (s != env->subprog_cnt) {
5607 if (linfo[i].insn_off == sub[s].start) {
5608 sub[s].linfo_idx = i;
5610 } else if (sub[s].start < linfo[i].insn_off) {
5611 verbose(env, "missing bpf_line_info for func#%u\n", s);
5617 prev_offset = linfo[i].insn_off;
5621 if (s != env->subprog_cnt) {
5622 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
5623 env->subprog_cnt - s, s);
5628 prog->aux->linfo = linfo;
5629 prog->aux->nr_linfo = nr_linfo;
5638 static int check_btf_info(struct bpf_verifier_env *env,
5639 const union bpf_attr *attr,
5640 union bpf_attr __user *uattr)
5645 if (!attr->func_info_cnt && !attr->line_info_cnt)
5648 btf = btf_get_by_fd(attr->prog_btf_fd);
5650 return PTR_ERR(btf);
5651 env->prog->aux->btf = btf;
5653 err = check_btf_func(env, attr, uattr);
5657 err = check_btf_line(env, attr, uattr);
5664 /* check %cur's range satisfies %old's */
5665 static bool range_within(struct bpf_reg_state *old,
5666 struct bpf_reg_state *cur)
5668 return old->umin_value <= cur->umin_value &&
5669 old->umax_value >= cur->umax_value &&
5670 old->smin_value <= cur->smin_value &&
5671 old->smax_value >= cur->smax_value;
5674 /* Maximum number of register states that can exist at once */
5675 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
5681 /* If in the old state two registers had the same id, then they need to have
5682 * the same id in the new state as well. But that id could be different from
5683 * the old state, so we need to track the mapping from old to new ids.
5684 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
5685 * regs with old id 5 must also have new id 9 for the new state to be safe. But
5686 * regs with a different old id could still have new id 9, we don't care about
5688 * So we look through our idmap to see if this old id has been seen before. If
5689 * so, we require the new id to match; otherwise, we add the id pair to the map.
5691 static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
5695 for (i = 0; i < ID_MAP_SIZE; i++) {
5696 if (!idmap[i].old) {
5697 /* Reached an empty slot; haven't seen this id before */
5698 idmap[i].old = old_id;
5699 idmap[i].cur = cur_id;
5702 if (idmap[i].old == old_id)
5703 return idmap[i].cur == cur_id;
5705 /* We ran out of idmap slots, which should be impossible */
5710 static void clean_func_state(struct bpf_verifier_env *env,
5711 struct bpf_func_state *st)
5713 enum bpf_reg_liveness live;
5716 for (i = 0; i < BPF_REG_FP; i++) {
5717 live = st->regs[i].live;
5718 /* liveness must not touch this register anymore */
5719 st->regs[i].live |= REG_LIVE_DONE;
5720 if (!(live & REG_LIVE_READ))
5721 /* since the register is unused, clear its state
5722 * to make further comparison simpler
5724 __mark_reg_not_init(&st->regs[i]);
5727 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
5728 live = st->stack[i].spilled_ptr.live;
5729 /* liveness must not touch this stack slot anymore */
5730 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
5731 if (!(live & REG_LIVE_READ)) {
5732 __mark_reg_not_init(&st->stack[i].spilled_ptr);
5733 for (j = 0; j < BPF_REG_SIZE; j++)
5734 st->stack[i].slot_type[j] = STACK_INVALID;
5739 static void clean_verifier_state(struct bpf_verifier_env *env,
5740 struct bpf_verifier_state *st)
5744 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
5745 /* all regs in this state in all frames were already marked */
5748 for (i = 0; i <= st->curframe; i++)
5749 clean_func_state(env, st->frame[i]);
5752 /* the parentage chains form a tree.
5753 * the verifier states are added to state lists at given insn and
5754 * pushed into state stack for future exploration.
5755 * when the verifier reaches bpf_exit insn some of the verifer states
5756 * stored in the state lists have their final liveness state already,
5757 * but a lot of states will get revised from liveness point of view when
5758 * the verifier explores other branches.
5761 * 2: if r1 == 100 goto pc+1
5764 * when the verifier reaches exit insn the register r0 in the state list of
5765 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
5766 * of insn 2 and goes exploring further. At the insn 4 it will walk the
5767 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
5769 * Since the verifier pushes the branch states as it sees them while exploring
5770 * the program the condition of walking the branch instruction for the second
5771 * time means that all states below this branch were already explored and
5772 * their final liveness markes are already propagated.
5773 * Hence when the verifier completes the search of state list in is_state_visited()
5774 * we can call this clean_live_states() function to mark all liveness states
5775 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
5777 * This function also clears the registers and stack for states that !READ
5778 * to simplify state merging.
5780 * Important note here that walking the same branch instruction in the callee
5781 * doesn't meant that the states are DONE. The verifier has to compare
5784 static void clean_live_states(struct bpf_verifier_env *env, int insn,
5785 struct bpf_verifier_state *cur)
5787 struct bpf_verifier_state_list *sl;
5790 sl = env->explored_states[insn];
5794 while (sl != STATE_LIST_MARK) {
5795 if (sl->state.curframe != cur->curframe)
5797 for (i = 0; i <= cur->curframe; i++)
5798 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
5800 clean_verifier_state(env, &sl->state);
5806 /* Returns true if (rold safe implies rcur safe) */
5807 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
5808 struct idpair *idmap)
5812 if (!(rold->live & REG_LIVE_READ))
5813 /* explored state didn't use this */
5816 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
5818 if (rold->type == PTR_TO_STACK)
5819 /* two stack pointers are equal only if they're pointing to
5820 * the same stack frame, since fp-8 in foo != fp-8 in bar
5822 return equal && rold->frameno == rcur->frameno;
5827 if (rold->type == NOT_INIT)
5828 /* explored state can't have used this */
5830 if (rcur->type == NOT_INIT)
5832 switch (rold->type) {
5834 if (rcur->type == SCALAR_VALUE) {
5835 /* new val must satisfy old val knowledge */
5836 return range_within(rold, rcur) &&
5837 tnum_in(rold->var_off, rcur->var_off);
5839 /* We're trying to use a pointer in place of a scalar.
5840 * Even if the scalar was unbounded, this could lead to
5841 * pointer leaks because scalars are allowed to leak
5842 * while pointers are not. We could make this safe in
5843 * special cases if root is calling us, but it's
5844 * probably not worth the hassle.
5848 case PTR_TO_MAP_VALUE:
5849 /* If the new min/max/var_off satisfy the old ones and
5850 * everything else matches, we are OK.
5851 * 'id' is not compared, since it's only used for maps with
5852 * bpf_spin_lock inside map element and in such cases if
5853 * the rest of the prog is valid for one map element then
5854 * it's valid for all map elements regardless of the key
5855 * used in bpf_map_lookup()
5857 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
5858 range_within(rold, rcur) &&
5859 tnum_in(rold->var_off, rcur->var_off);
5860 case PTR_TO_MAP_VALUE_OR_NULL:
5861 /* a PTR_TO_MAP_VALUE could be safe to use as a
5862 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
5863 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
5864 * checked, doing so could have affected others with the same
5865 * id, and we can't check for that because we lost the id when
5866 * we converted to a PTR_TO_MAP_VALUE.
5868 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
5870 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
5872 /* Check our ids match any regs they're supposed to */
5873 return check_ids(rold->id, rcur->id, idmap);
5874 case PTR_TO_PACKET_META:
5876 if (rcur->type != rold->type)
5878 /* We must have at least as much range as the old ptr
5879 * did, so that any accesses which were safe before are
5880 * still safe. This is true even if old range < old off,
5881 * since someone could have accessed through (ptr - k), or
5882 * even done ptr -= k in a register, to get a safe access.
5884 if (rold->range > rcur->range)
5886 /* If the offsets don't match, we can't trust our alignment;
5887 * nor can we be sure that we won't fall out of range.
5889 if (rold->off != rcur->off)
5891 /* id relations must be preserved */
5892 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
5894 /* new val must satisfy old val knowledge */
5895 return range_within(rold, rcur) &&
5896 tnum_in(rold->var_off, rcur->var_off);
5898 case CONST_PTR_TO_MAP:
5899 case PTR_TO_PACKET_END:
5900 case PTR_TO_FLOW_KEYS:
5902 case PTR_TO_SOCKET_OR_NULL:
5903 case PTR_TO_SOCK_COMMON:
5904 case PTR_TO_SOCK_COMMON_OR_NULL:
5905 case PTR_TO_TCP_SOCK:
5906 case PTR_TO_TCP_SOCK_OR_NULL:
5907 /* Only valid matches are exact, which memcmp() above
5908 * would have accepted
5911 /* Don't know what's going on, just say it's not safe */
5915 /* Shouldn't get here; if we do, say it's not safe */
5920 static bool stacksafe(struct bpf_func_state *old,
5921 struct bpf_func_state *cur,
5922 struct idpair *idmap)
5926 /* walk slots of the explored stack and ignore any additional
5927 * slots in the current stack, since explored(safe) state
5930 for (i = 0; i < old->allocated_stack; i++) {
5931 spi = i / BPF_REG_SIZE;
5933 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
5934 i += BPF_REG_SIZE - 1;
5935 /* explored state didn't use this */
5939 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
5942 /* explored stack has more populated slots than current stack
5943 * and these slots were used
5945 if (i >= cur->allocated_stack)
5948 /* if old state was safe with misc data in the stack
5949 * it will be safe with zero-initialized stack.
5950 * The opposite is not true
5952 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
5953 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
5955 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
5956 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
5957 /* Ex: old explored (safe) state has STACK_SPILL in
5958 * this stack slot, but current has has STACK_MISC ->
5959 * this verifier states are not equivalent,
5960 * return false to continue verification of this path
5963 if (i % BPF_REG_SIZE)
5965 if (old->stack[spi].slot_type[0] != STACK_SPILL)
5967 if (!regsafe(&old->stack[spi].spilled_ptr,
5968 &cur->stack[spi].spilled_ptr,
5970 /* when explored and current stack slot are both storing
5971 * spilled registers, check that stored pointers types
5972 * are the same as well.
5973 * Ex: explored safe path could have stored
5974 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
5975 * but current path has stored:
5976 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
5977 * such verifier states are not equivalent.
5978 * return false to continue verification of this path
5985 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
5987 if (old->acquired_refs != cur->acquired_refs)
5989 return !memcmp(old->refs, cur->refs,
5990 sizeof(*old->refs) * old->acquired_refs);
5993 /* compare two verifier states
5995 * all states stored in state_list are known to be valid, since
5996 * verifier reached 'bpf_exit' instruction through them
5998 * this function is called when verifier exploring different branches of
5999 * execution popped from the state stack. If it sees an old state that has
6000 * more strict register state and more strict stack state then this execution
6001 * branch doesn't need to be explored further, since verifier already
6002 * concluded that more strict state leads to valid finish.
6004 * Therefore two states are equivalent if register state is more conservative
6005 * and explored stack state is more conservative than the current one.
6008 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
6009 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
6011 * In other words if current stack state (one being explored) has more
6012 * valid slots than old one that already passed validation, it means
6013 * the verifier can stop exploring and conclude that current state is valid too
6015 * Similarly with registers. If explored state has register type as invalid
6016 * whereas register type in current state is meaningful, it means that
6017 * the current state will reach 'bpf_exit' instruction safely
6019 static bool func_states_equal(struct bpf_func_state *old,
6020 struct bpf_func_state *cur)
6022 struct idpair *idmap;
6026 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
6027 /* If we failed to allocate the idmap, just say it's not safe */
6031 for (i = 0; i < MAX_BPF_REG; i++) {
6032 if (!regsafe(&old->regs[i], &cur->regs[i], idmap))
6036 if (!stacksafe(old, cur, idmap))
6039 if (!refsafe(old, cur))
6047 static bool states_equal(struct bpf_verifier_env *env,
6048 struct bpf_verifier_state *old,
6049 struct bpf_verifier_state *cur)
6053 if (old->curframe != cur->curframe)
6056 /* Verification state from speculative execution simulation
6057 * must never prune a non-speculative execution one.
6059 if (old->speculative && !cur->speculative)
6062 if (old->active_spin_lock != cur->active_spin_lock)
6065 /* for states to be equal callsites have to be the same
6066 * and all frame states need to be equivalent
6068 for (i = 0; i <= old->curframe; i++) {
6069 if (old->frame[i]->callsite != cur->frame[i]->callsite)
6071 if (!func_states_equal(old->frame[i], cur->frame[i]))
6077 /* A write screens off any subsequent reads; but write marks come from the
6078 * straight-line code between a state and its parent. When we arrive at an
6079 * equivalent state (jump target or such) we didn't arrive by the straight-line
6080 * code, so read marks in the state must propagate to the parent regardless
6081 * of the state's write marks. That's what 'parent == state->parent' comparison
6082 * in mark_reg_read() is for.
6084 static int propagate_liveness(struct bpf_verifier_env *env,
6085 const struct bpf_verifier_state *vstate,
6086 struct bpf_verifier_state *vparent)
6088 int i, frame, err = 0;
6089 struct bpf_func_state *state, *parent;
6091 if (vparent->curframe != vstate->curframe) {
6092 WARN(1, "propagate_live: parent frame %d current frame %d\n",
6093 vparent->curframe, vstate->curframe);
6096 /* Propagate read liveness of registers... */
6097 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
6098 for (frame = 0; frame <= vstate->curframe; frame++) {
6099 /* We don't need to worry about FP liveness, it's read-only */
6100 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
6101 if (vparent->frame[frame]->regs[i].live & REG_LIVE_READ)
6103 if (vstate->frame[frame]->regs[i].live & REG_LIVE_READ) {
6104 err = mark_reg_read(env, &vstate->frame[frame]->regs[i],
6105 &vparent->frame[frame]->regs[i]);
6112 /* ... and stack slots */
6113 for (frame = 0; frame <= vstate->curframe; frame++) {
6114 state = vstate->frame[frame];
6115 parent = vparent->frame[frame];
6116 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
6117 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
6118 if (parent->stack[i].spilled_ptr.live & REG_LIVE_READ)
6120 if (state->stack[i].spilled_ptr.live & REG_LIVE_READ)
6121 mark_reg_read(env, &state->stack[i].spilled_ptr,
6122 &parent->stack[i].spilled_ptr);
6128 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
6130 struct bpf_verifier_state_list *new_sl;
6131 struct bpf_verifier_state_list *sl;
6132 struct bpf_verifier_state *cur = env->cur_state, *new;
6133 int i, j, err, states_cnt = 0;
6135 sl = env->explored_states[insn_idx];
6137 /* this 'insn_idx' instruction wasn't marked, so we will not
6138 * be doing state search here
6142 clean_live_states(env, insn_idx, cur);
6144 while (sl != STATE_LIST_MARK) {
6145 if (states_equal(env, &sl->state, cur)) {
6146 /* reached equivalent register/stack state,
6148 * Registers read by the continuation are read by us.
6149 * If we have any write marks in env->cur_state, they
6150 * will prevent corresponding reads in the continuation
6151 * from reaching our parent (an explored_state). Our
6152 * own state will get the read marks recorded, but
6153 * they'll be immediately forgotten as we're pruning
6154 * this state and will pop a new one.
6156 err = propagate_liveness(env, &sl->state, cur);
6165 if (!env->allow_ptr_leaks && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
6168 /* there were no equivalent states, remember current one.
6169 * technically the current state is not proven to be safe yet,
6170 * but it will either reach outer most bpf_exit (which means it's safe)
6171 * or it will be rejected. Since there are no loops, we won't be
6172 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
6173 * again on the way to bpf_exit
6175 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
6179 /* add new state to the head of linked list */
6180 new = &new_sl->state;
6181 err = copy_verifier_state(new, cur);
6183 free_verifier_state(new, false);
6187 new_sl->next = env->explored_states[insn_idx];
6188 env->explored_states[insn_idx] = new_sl;
6189 /* connect new state to parentage chain. Current frame needs all
6190 * registers connected. Only r6 - r9 of the callers are alive (pushed
6191 * to the stack implicitly by JITs) so in callers' frames connect just
6192 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
6193 * the state of the call instruction (with WRITTEN set), and r0 comes
6194 * from callee with its full parentage chain, anyway.
6196 for (j = 0; j <= cur->curframe; j++)
6197 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
6198 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
6199 /* clear write marks in current state: the writes we did are not writes
6200 * our child did, so they don't screen off its reads from us.
6201 * (There are no read marks in current state, because reads always mark
6202 * their parent and current state never has children yet. Only
6203 * explored_states can get read marks.)
6205 for (i = 0; i < BPF_REG_FP; i++)
6206 cur->frame[cur->curframe]->regs[i].live = REG_LIVE_NONE;
6208 /* all stack frames are accessible from callee, clear them all */
6209 for (j = 0; j <= cur->curframe; j++) {
6210 struct bpf_func_state *frame = cur->frame[j];
6211 struct bpf_func_state *newframe = new->frame[j];
6213 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
6214 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
6215 frame->stack[i].spilled_ptr.parent =
6216 &newframe->stack[i].spilled_ptr;
6222 /* Return true if it's OK to have the same insn return a different type. */
6223 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
6228 case PTR_TO_SOCKET_OR_NULL:
6229 case PTR_TO_SOCK_COMMON:
6230 case PTR_TO_SOCK_COMMON_OR_NULL:
6231 case PTR_TO_TCP_SOCK:
6232 case PTR_TO_TCP_SOCK_OR_NULL:
6239 /* If an instruction was previously used with particular pointer types, then we
6240 * need to be careful to avoid cases such as the below, where it may be ok
6241 * for one branch accessing the pointer, but not ok for the other branch:
6246 * R1 = some_other_valid_ptr;
6249 * R2 = *(u32 *)(R1 + 0);
6251 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
6253 return src != prev && (!reg_type_mismatch_ok(src) ||
6254 !reg_type_mismatch_ok(prev));
6257 static int do_check(struct bpf_verifier_env *env)
6259 struct bpf_verifier_state *state;
6260 struct bpf_insn *insns = env->prog->insnsi;
6261 struct bpf_reg_state *regs;
6262 int insn_cnt = env->prog->len, i;
6263 int insn_processed = 0;
6264 bool do_print_state = false;
6266 env->prev_linfo = NULL;
6268 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
6271 state->curframe = 0;
6272 state->speculative = false;
6273 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
6274 if (!state->frame[0]) {
6278 env->cur_state = state;
6279 init_func_state(env, state->frame[0],
6280 BPF_MAIN_FUNC /* callsite */,
6282 0 /* subprogno, zero == main subprog */);
6285 struct bpf_insn *insn;
6289 if (env->insn_idx >= insn_cnt) {
6290 verbose(env, "invalid insn idx %d insn_cnt %d\n",
6291 env->insn_idx, insn_cnt);
6295 insn = &insns[env->insn_idx];
6296 class = BPF_CLASS(insn->code);
6298 if (++insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
6300 "BPF program is too large. Processed %d insn\n",
6305 err = is_state_visited(env, env->insn_idx);
6309 /* found equivalent state, can prune the search */
6310 if (env->log.level) {
6312 verbose(env, "\nfrom %d to %d%s: safe\n",
6313 env->prev_insn_idx, env->insn_idx,
6314 env->cur_state->speculative ?
6315 " (speculative execution)" : "");
6317 verbose(env, "%d: safe\n", env->insn_idx);
6319 goto process_bpf_exit;
6322 if (signal_pending(current))
6328 if (env->log.level > 1 || (env->log.level && do_print_state)) {
6329 if (env->log.level > 1)
6330 verbose(env, "%d:", env->insn_idx);
6332 verbose(env, "\nfrom %d to %d%s:",
6333 env->prev_insn_idx, env->insn_idx,
6334 env->cur_state->speculative ?
6335 " (speculative execution)" : "");
6336 print_verifier_state(env, state->frame[state->curframe]);
6337 do_print_state = false;
6340 if (env->log.level) {
6341 const struct bpf_insn_cbs cbs = {
6342 .cb_print = verbose,
6343 .private_data = env,
6346 verbose_linfo(env, env->insn_idx, "; ");
6347 verbose(env, "%d: ", env->insn_idx);
6348 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
6351 if (bpf_prog_is_dev_bound(env->prog->aux)) {
6352 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
6353 env->prev_insn_idx);
6358 regs = cur_regs(env);
6359 env->insn_aux_data[env->insn_idx].seen = true;
6361 if (class == BPF_ALU || class == BPF_ALU64) {
6362 err = check_alu_op(env, insn);
6366 } else if (class == BPF_LDX) {
6367 enum bpf_reg_type *prev_src_type, src_reg_type;
6369 /* check for reserved fields is already done */
6371 /* check src operand */
6372 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6376 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
6380 src_reg_type = regs[insn->src_reg].type;
6382 /* check that memory (src_reg + off) is readable,
6383 * the state of dst_reg will be updated by this func
6385 err = check_mem_access(env, env->insn_idx, insn->src_reg,
6386 insn->off, BPF_SIZE(insn->code),
6387 BPF_READ, insn->dst_reg, false);
6391 prev_src_type = &env->insn_aux_data[env->insn_idx].ptr_type;
6393 if (*prev_src_type == NOT_INIT) {
6395 * dst_reg = *(u32 *)(src_reg + off)
6396 * save type to validate intersecting paths
6398 *prev_src_type = src_reg_type;
6400 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
6401 /* ABuser program is trying to use the same insn
6402 * dst_reg = *(u32*) (src_reg + off)
6403 * with different pointer types:
6404 * src_reg == ctx in one branch and
6405 * src_reg == stack|map in some other branch.
6408 verbose(env, "same insn cannot be used with different pointers\n");
6412 } else if (class == BPF_STX) {
6413 enum bpf_reg_type *prev_dst_type, dst_reg_type;
6415 if (BPF_MODE(insn->code) == BPF_XADD) {
6416 err = check_xadd(env, env->insn_idx, insn);
6423 /* check src1 operand */
6424 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6427 /* check src2 operand */
6428 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
6432 dst_reg_type = regs[insn->dst_reg].type;
6434 /* check that memory (dst_reg + off) is writeable */
6435 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
6436 insn->off, BPF_SIZE(insn->code),
6437 BPF_WRITE, insn->src_reg, false);
6441 prev_dst_type = &env->insn_aux_data[env->insn_idx].ptr_type;
6443 if (*prev_dst_type == NOT_INIT) {
6444 *prev_dst_type = dst_reg_type;
6445 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
6446 verbose(env, "same insn cannot be used with different pointers\n");
6450 } else if (class == BPF_ST) {
6451 if (BPF_MODE(insn->code) != BPF_MEM ||
6452 insn->src_reg != BPF_REG_0) {
6453 verbose(env, "BPF_ST uses reserved fields\n");
6456 /* check src operand */
6457 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
6461 if (is_ctx_reg(env, insn->dst_reg)) {
6462 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
6464 reg_type_str[reg_state(env, insn->dst_reg)->type]);
6468 /* check that memory (dst_reg + off) is writeable */
6469 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
6470 insn->off, BPF_SIZE(insn->code),
6471 BPF_WRITE, -1, false);
6475 } else if (class == BPF_JMP || class == BPF_JMP32) {
6476 u8 opcode = BPF_OP(insn->code);
6478 if (opcode == BPF_CALL) {
6479 if (BPF_SRC(insn->code) != BPF_K ||
6481 (insn->src_reg != BPF_REG_0 &&
6482 insn->src_reg != BPF_PSEUDO_CALL) ||
6483 insn->dst_reg != BPF_REG_0 ||
6484 class == BPF_JMP32) {
6485 verbose(env, "BPF_CALL uses reserved fields\n");
6489 if (env->cur_state->active_spin_lock &&
6490 (insn->src_reg == BPF_PSEUDO_CALL ||
6491 insn->imm != BPF_FUNC_spin_unlock)) {
6492 verbose(env, "function calls are not allowed while holding a lock\n");
6495 if (insn->src_reg == BPF_PSEUDO_CALL)
6496 err = check_func_call(env, insn, &env->insn_idx);
6498 err = check_helper_call(env, insn->imm, env->insn_idx);
6502 } else if (opcode == BPF_JA) {
6503 if (BPF_SRC(insn->code) != BPF_K ||
6505 insn->src_reg != BPF_REG_0 ||
6506 insn->dst_reg != BPF_REG_0 ||
6507 class == BPF_JMP32) {
6508 verbose(env, "BPF_JA uses reserved fields\n");
6512 env->insn_idx += insn->off + 1;
6515 } else if (opcode == BPF_EXIT) {
6516 if (BPF_SRC(insn->code) != BPF_K ||
6518 insn->src_reg != BPF_REG_0 ||
6519 insn->dst_reg != BPF_REG_0 ||
6520 class == BPF_JMP32) {
6521 verbose(env, "BPF_EXIT uses reserved fields\n");
6525 if (env->cur_state->active_spin_lock) {
6526 verbose(env, "bpf_spin_unlock is missing\n");
6530 if (state->curframe) {
6531 /* exit from nested function */
6532 env->prev_insn_idx = env->insn_idx;
6533 err = prepare_func_exit(env, &env->insn_idx);
6536 do_print_state = true;
6540 err = check_reference_leak(env);
6544 /* eBPF calling convetion is such that R0 is used
6545 * to return the value from eBPF program.
6546 * Make sure that it's readable at this time
6547 * of bpf_exit, which means that program wrote
6548 * something into it earlier
6550 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
6554 if (is_pointer_value(env, BPF_REG_0)) {
6555 verbose(env, "R0 leaks addr as return value\n");
6559 err = check_return_code(env);
6563 err = pop_stack(env, &env->prev_insn_idx,
6570 do_print_state = true;
6574 err = check_cond_jmp_op(env, insn, &env->insn_idx);
6578 } else if (class == BPF_LD) {
6579 u8 mode = BPF_MODE(insn->code);
6581 if (mode == BPF_ABS || mode == BPF_IND) {
6582 err = check_ld_abs(env, insn);
6586 } else if (mode == BPF_IMM) {
6587 err = check_ld_imm(env, insn);
6592 env->insn_aux_data[env->insn_idx].seen = true;
6594 verbose(env, "invalid BPF_LD mode\n");
6598 verbose(env, "unknown insn class %d\n", class);
6605 verbose(env, "processed %d insns (limit %d), stack depth ",
6606 insn_processed, BPF_COMPLEXITY_LIMIT_INSNS);
6607 for (i = 0; i < env->subprog_cnt; i++) {
6608 u32 depth = env->subprog_info[i].stack_depth;
6610 verbose(env, "%d", depth);
6611 if (i + 1 < env->subprog_cnt)
6615 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
6619 static int check_map_prealloc(struct bpf_map *map)
6621 return (map->map_type != BPF_MAP_TYPE_HASH &&
6622 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
6623 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
6624 !(map->map_flags & BPF_F_NO_PREALLOC);
6627 static bool is_tracing_prog_type(enum bpf_prog_type type)
6630 case BPF_PROG_TYPE_KPROBE:
6631 case BPF_PROG_TYPE_TRACEPOINT:
6632 case BPF_PROG_TYPE_PERF_EVENT:
6633 case BPF_PROG_TYPE_RAW_TRACEPOINT:
6640 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
6641 struct bpf_map *map,
6642 struct bpf_prog *prog)
6645 /* Make sure that BPF_PROG_TYPE_PERF_EVENT programs only use
6646 * preallocated hash maps, since doing memory allocation
6647 * in overflow_handler can crash depending on where nmi got
6650 if (prog->type == BPF_PROG_TYPE_PERF_EVENT) {
6651 if (!check_map_prealloc(map)) {
6652 verbose(env, "perf_event programs can only use preallocated hash map\n");
6655 if (map->inner_map_meta &&
6656 !check_map_prealloc(map->inner_map_meta)) {
6657 verbose(env, "perf_event programs can only use preallocated inner hash map\n");
6662 if ((is_tracing_prog_type(prog->type) ||
6663 prog->type == BPF_PROG_TYPE_SOCKET_FILTER) &&
6664 map_value_has_spin_lock(map)) {
6665 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
6669 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
6670 !bpf_offload_prog_map_match(prog, map)) {
6671 verbose(env, "offload device mismatch between prog and map\n");
6678 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
6680 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
6681 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
6684 /* look for pseudo eBPF instructions that access map FDs and
6685 * replace them with actual map pointers
6687 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env *env)
6689 struct bpf_insn *insn = env->prog->insnsi;
6690 int insn_cnt = env->prog->len;
6693 err = bpf_prog_calc_tag(env->prog);
6697 for (i = 0; i < insn_cnt; i++, insn++) {
6698 if (BPF_CLASS(insn->code) == BPF_LDX &&
6699 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
6700 verbose(env, "BPF_LDX uses reserved fields\n");
6704 if (BPF_CLASS(insn->code) == BPF_STX &&
6705 ((BPF_MODE(insn->code) != BPF_MEM &&
6706 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
6707 verbose(env, "BPF_STX uses reserved fields\n");
6711 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
6712 struct bpf_map *map;
6715 if (i == insn_cnt - 1 || insn[1].code != 0 ||
6716 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
6718 verbose(env, "invalid bpf_ld_imm64 insn\n");
6722 if (insn->src_reg == 0)
6723 /* valid generic load 64-bit imm */
6726 if (insn[0].src_reg != BPF_PSEUDO_MAP_FD ||
6728 verbose(env, "unrecognized bpf_ld_imm64 insn\n");
6732 f = fdget(insn[0].imm);
6733 map = __bpf_map_get(f);
6735 verbose(env, "fd %d is not pointing to valid bpf_map\n",
6737 return PTR_ERR(map);
6740 err = check_map_prog_compatibility(env, map, env->prog);
6746 /* store map pointer inside BPF_LD_IMM64 instruction */
6747 insn[0].imm = (u32) (unsigned long) map;
6748 insn[1].imm = ((u64) (unsigned long) map) >> 32;
6750 /* check whether we recorded this map already */
6751 for (j = 0; j < env->used_map_cnt; j++)
6752 if (env->used_maps[j] == map) {
6757 if (env->used_map_cnt >= MAX_USED_MAPS) {
6762 /* hold the map. If the program is rejected by verifier,
6763 * the map will be released by release_maps() or it
6764 * will be used by the valid program until it's unloaded
6765 * and all maps are released in free_used_maps()
6767 map = bpf_map_inc(map, false);
6770 return PTR_ERR(map);
6772 env->used_maps[env->used_map_cnt++] = map;
6774 if (bpf_map_is_cgroup_storage(map) &&
6775 bpf_cgroup_storage_assign(env->prog, map)) {
6776 verbose(env, "only one cgroup storage of each type is allowed\n");
6788 /* Basic sanity check before we invest more work here. */
6789 if (!bpf_opcode_in_insntable(insn->code)) {
6790 verbose(env, "unknown opcode %02x\n", insn->code);
6795 /* now all pseudo BPF_LD_IMM64 instructions load valid
6796 * 'struct bpf_map *' into a register instead of user map_fd.
6797 * These pointers will be used later by verifier to validate map access.
6802 /* drop refcnt of maps used by the rejected program */
6803 static void release_maps(struct bpf_verifier_env *env)
6805 enum bpf_cgroup_storage_type stype;
6808 for_each_cgroup_storage_type(stype) {
6809 if (!env->prog->aux->cgroup_storage[stype])
6811 bpf_cgroup_storage_release(env->prog,
6812 env->prog->aux->cgroup_storage[stype]);
6815 for (i = 0; i < env->used_map_cnt; i++)
6816 bpf_map_put(env->used_maps[i]);
6819 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
6820 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
6822 struct bpf_insn *insn = env->prog->insnsi;
6823 int insn_cnt = env->prog->len;
6826 for (i = 0; i < insn_cnt; i++, insn++)
6827 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
6831 /* single env->prog->insni[off] instruction was replaced with the range
6832 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
6833 * [0, off) and [off, end) to new locations, so the patched range stays zero
6835 static int adjust_insn_aux_data(struct bpf_verifier_env *env, u32 prog_len,
6838 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
6843 new_data = vzalloc(array_size(prog_len,
6844 sizeof(struct bpf_insn_aux_data)));
6847 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
6848 memcpy(new_data + off + cnt - 1, old_data + off,
6849 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
6850 for (i = off; i < off + cnt - 1; i++)
6851 new_data[i].seen = true;
6852 env->insn_aux_data = new_data;
6857 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
6863 /* NOTE: fake 'exit' subprog should be updated as well. */
6864 for (i = 0; i <= env->subprog_cnt; i++) {
6865 if (env->subprog_info[i].start <= off)
6867 env->subprog_info[i].start += len - 1;
6871 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
6872 const struct bpf_insn *patch, u32 len)
6874 struct bpf_prog *new_prog;
6876 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
6879 if (adjust_insn_aux_data(env, new_prog->len, off, len))
6881 adjust_subprog_starts(env, off, len);
6885 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
6890 /* find first prog starting at or after off (first to remove) */
6891 for (i = 0; i < env->subprog_cnt; i++)
6892 if (env->subprog_info[i].start >= off)
6894 /* find first prog starting at or after off + cnt (first to stay) */
6895 for (j = i; j < env->subprog_cnt; j++)
6896 if (env->subprog_info[j].start >= off + cnt)
6898 /* if j doesn't start exactly at off + cnt, we are just removing
6899 * the front of previous prog
6901 if (env->subprog_info[j].start != off + cnt)
6905 struct bpf_prog_aux *aux = env->prog->aux;
6908 /* move fake 'exit' subprog as well */
6909 move = env->subprog_cnt + 1 - j;
6911 memmove(env->subprog_info + i,
6912 env->subprog_info + j,
6913 sizeof(*env->subprog_info) * move);
6914 env->subprog_cnt -= j - i;
6916 /* remove func_info */
6917 if (aux->func_info) {
6918 move = aux->func_info_cnt - j;
6920 memmove(aux->func_info + i,
6922 sizeof(*aux->func_info) * move);
6923 aux->func_info_cnt -= j - i;
6924 /* func_info->insn_off is set after all code rewrites,
6925 * in adjust_btf_func() - no need to adjust
6929 /* convert i from "first prog to remove" to "first to adjust" */
6930 if (env->subprog_info[i].start == off)
6934 /* update fake 'exit' subprog as well */
6935 for (; i <= env->subprog_cnt; i++)
6936 env->subprog_info[i].start -= cnt;
6941 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
6944 struct bpf_prog *prog = env->prog;
6945 u32 i, l_off, l_cnt, nr_linfo;
6946 struct bpf_line_info *linfo;
6948 nr_linfo = prog->aux->nr_linfo;
6952 linfo = prog->aux->linfo;
6954 /* find first line info to remove, count lines to be removed */
6955 for (i = 0; i < nr_linfo; i++)
6956 if (linfo[i].insn_off >= off)
6961 for (; i < nr_linfo; i++)
6962 if (linfo[i].insn_off < off + cnt)
6967 /* First live insn doesn't match first live linfo, it needs to "inherit"
6968 * last removed linfo. prog is already modified, so prog->len == off
6969 * means no live instructions after (tail of the program was removed).
6971 if (prog->len != off && l_cnt &&
6972 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
6974 linfo[--i].insn_off = off + cnt;
6977 /* remove the line info which refer to the removed instructions */
6979 memmove(linfo + l_off, linfo + i,
6980 sizeof(*linfo) * (nr_linfo - i));
6982 prog->aux->nr_linfo -= l_cnt;
6983 nr_linfo = prog->aux->nr_linfo;
6986 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
6987 for (i = l_off; i < nr_linfo; i++)
6988 linfo[i].insn_off -= cnt;
6990 /* fix up all subprogs (incl. 'exit') which start >= off */
6991 for (i = 0; i <= env->subprog_cnt; i++)
6992 if (env->subprog_info[i].linfo_idx > l_off) {
6993 /* program may have started in the removed region but
6994 * may not be fully removed
6996 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
6997 env->subprog_info[i].linfo_idx -= l_cnt;
6999 env->subprog_info[i].linfo_idx = l_off;
7005 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
7007 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
7008 unsigned int orig_prog_len = env->prog->len;
7011 if (bpf_prog_is_dev_bound(env->prog->aux))
7012 bpf_prog_offload_remove_insns(env, off, cnt);
7014 err = bpf_remove_insns(env->prog, off, cnt);
7018 err = adjust_subprog_starts_after_remove(env, off, cnt);
7022 err = bpf_adj_linfo_after_remove(env, off, cnt);
7026 memmove(aux_data + off, aux_data + off + cnt,
7027 sizeof(*aux_data) * (orig_prog_len - off - cnt));
7032 /* The verifier does more data flow analysis than llvm and will not
7033 * explore branches that are dead at run time. Malicious programs can
7034 * have dead code too. Therefore replace all dead at-run-time code
7037 * Just nops are not optimal, e.g. if they would sit at the end of the
7038 * program and through another bug we would manage to jump there, then
7039 * we'd execute beyond program memory otherwise. Returning exception
7040 * code also wouldn't work since we can have subprogs where the dead
7041 * code could be located.
7043 static void sanitize_dead_code(struct bpf_verifier_env *env)
7045 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
7046 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
7047 struct bpf_insn *insn = env->prog->insnsi;
7048 const int insn_cnt = env->prog->len;
7051 for (i = 0; i < insn_cnt; i++) {
7052 if (aux_data[i].seen)
7054 memcpy(insn + i, &trap, sizeof(trap));
7058 static bool insn_is_cond_jump(u8 code)
7062 if (BPF_CLASS(code) == BPF_JMP32)
7065 if (BPF_CLASS(code) != BPF_JMP)
7069 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
7072 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
7074 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
7075 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
7076 struct bpf_insn *insn = env->prog->insnsi;
7077 const int insn_cnt = env->prog->len;
7080 for (i = 0; i < insn_cnt; i++, insn++) {
7081 if (!insn_is_cond_jump(insn->code))
7084 if (!aux_data[i + 1].seen)
7086 else if (!aux_data[i + 1 + insn->off].seen)
7091 if (bpf_prog_is_dev_bound(env->prog->aux))
7092 bpf_prog_offload_replace_insn(env, i, &ja);
7094 memcpy(insn, &ja, sizeof(ja));
7098 static int opt_remove_dead_code(struct bpf_verifier_env *env)
7100 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
7101 int insn_cnt = env->prog->len;
7104 for (i = 0; i < insn_cnt; i++) {
7108 while (i + j < insn_cnt && !aux_data[i + j].seen)
7113 err = verifier_remove_insns(env, i, j);
7116 insn_cnt = env->prog->len;
7122 static int opt_remove_nops(struct bpf_verifier_env *env)
7124 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
7125 struct bpf_insn *insn = env->prog->insnsi;
7126 int insn_cnt = env->prog->len;
7129 for (i = 0; i < insn_cnt; i++) {
7130 if (memcmp(&insn[i], &ja, sizeof(ja)))
7133 err = verifier_remove_insns(env, i, 1);
7143 /* convert load instructions that access fields of a context type into a
7144 * sequence of instructions that access fields of the underlying structure:
7145 * struct __sk_buff -> struct sk_buff
7146 * struct bpf_sock_ops -> struct sock
7148 static int convert_ctx_accesses(struct bpf_verifier_env *env)
7150 const struct bpf_verifier_ops *ops = env->ops;
7151 int i, cnt, size, ctx_field_size, delta = 0;
7152 const int insn_cnt = env->prog->len;
7153 struct bpf_insn insn_buf[16], *insn;
7154 u32 target_size, size_default, off;
7155 struct bpf_prog *new_prog;
7156 enum bpf_access_type type;
7157 bool is_narrower_load;
7159 if (ops->gen_prologue || env->seen_direct_write) {
7160 if (!ops->gen_prologue) {
7161 verbose(env, "bpf verifier is misconfigured\n");
7164 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
7166 if (cnt >= ARRAY_SIZE(insn_buf)) {
7167 verbose(env, "bpf verifier is misconfigured\n");
7170 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
7174 env->prog = new_prog;
7179 if (bpf_prog_is_dev_bound(env->prog->aux))
7182 insn = env->prog->insnsi + delta;
7184 for (i = 0; i < insn_cnt; i++, insn++) {
7185 bpf_convert_ctx_access_t convert_ctx_access;
7187 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
7188 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
7189 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
7190 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
7192 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
7193 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
7194 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
7195 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
7200 if (type == BPF_WRITE &&
7201 env->insn_aux_data[i + delta].sanitize_stack_off) {
7202 struct bpf_insn patch[] = {
7203 /* Sanitize suspicious stack slot with zero.
7204 * There are no memory dependencies for this store,
7205 * since it's only using frame pointer and immediate
7208 BPF_ST_MEM(BPF_DW, BPF_REG_FP,
7209 env->insn_aux_data[i + delta].sanitize_stack_off,
7211 /* the original STX instruction will immediately
7212 * overwrite the same stack slot with appropriate value
7217 cnt = ARRAY_SIZE(patch);
7218 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
7223 env->prog = new_prog;
7224 insn = new_prog->insnsi + i + delta;
7228 switch (env->insn_aux_data[i + delta].ptr_type) {
7230 if (!ops->convert_ctx_access)
7232 convert_ctx_access = ops->convert_ctx_access;
7235 case PTR_TO_SOCK_COMMON:
7236 convert_ctx_access = bpf_sock_convert_ctx_access;
7238 case PTR_TO_TCP_SOCK:
7239 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
7245 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
7246 size = BPF_LDST_BYTES(insn);
7248 /* If the read access is a narrower load of the field,
7249 * convert to a 4/8-byte load, to minimum program type specific
7250 * convert_ctx_access changes. If conversion is successful,
7251 * we will apply proper mask to the result.
7253 is_narrower_load = size < ctx_field_size;
7254 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
7256 if (is_narrower_load) {
7259 if (type == BPF_WRITE) {
7260 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
7265 if (ctx_field_size == 4)
7267 else if (ctx_field_size == 8)
7270 insn->off = off & ~(size_default - 1);
7271 insn->code = BPF_LDX | BPF_MEM | size_code;
7275 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
7277 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
7278 (ctx_field_size && !target_size)) {
7279 verbose(env, "bpf verifier is misconfigured\n");
7283 if (is_narrower_load && size < target_size) {
7284 u8 shift = (off & (size_default - 1)) * 8;
7286 if (ctx_field_size <= 4) {
7288 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
7291 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
7292 (1 << size * 8) - 1);
7295 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
7298 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
7299 (1 << size * 8) - 1);
7303 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
7309 /* keep walking new program and skip insns we just inserted */
7310 env->prog = new_prog;
7311 insn = new_prog->insnsi + i + delta;
7317 static int jit_subprogs(struct bpf_verifier_env *env)
7319 struct bpf_prog *prog = env->prog, **func, *tmp;
7320 int i, j, subprog_start, subprog_end = 0, len, subprog;
7321 struct bpf_insn *insn;
7325 if (env->subprog_cnt <= 1)
7328 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
7329 if (insn->code != (BPF_JMP | BPF_CALL) ||
7330 insn->src_reg != BPF_PSEUDO_CALL)
7332 /* Upon error here we cannot fall back to interpreter but
7333 * need a hard reject of the program. Thus -EFAULT is
7334 * propagated in any case.
7336 subprog = find_subprog(env, i + insn->imm + 1);
7338 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
7342 /* temporarily remember subprog id inside insn instead of
7343 * aux_data, since next loop will split up all insns into funcs
7345 insn->off = subprog;
7346 /* remember original imm in case JIT fails and fallback
7347 * to interpreter will be needed
7349 env->insn_aux_data[i].call_imm = insn->imm;
7350 /* point imm to __bpf_call_base+1 from JITs point of view */
7354 err = bpf_prog_alloc_jited_linfo(prog);
7359 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
7363 for (i = 0; i < env->subprog_cnt; i++) {
7364 subprog_start = subprog_end;
7365 subprog_end = env->subprog_info[i + 1].start;
7367 len = subprog_end - subprog_start;
7368 /* BPF_PROG_RUN doesn't call subprogs directly,
7369 * hence main prog stats include the runtime of subprogs.
7370 * subprogs don't have IDs and not reachable via prog_get_next_id
7371 * func[i]->aux->stats will never be accessed and stays NULL
7373 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
7376 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
7377 len * sizeof(struct bpf_insn));
7378 func[i]->type = prog->type;
7380 if (bpf_prog_calc_tag(func[i]))
7382 func[i]->is_func = 1;
7383 func[i]->aux->func_idx = i;
7384 /* the btf and func_info will be freed only at prog->aux */
7385 func[i]->aux->btf = prog->aux->btf;
7386 func[i]->aux->func_info = prog->aux->func_info;
7388 /* Use bpf_prog_F_tag to indicate functions in stack traces.
7389 * Long term would need debug info to populate names
7391 func[i]->aux->name[0] = 'F';
7392 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
7393 func[i]->jit_requested = 1;
7394 func[i]->aux->linfo = prog->aux->linfo;
7395 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
7396 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
7397 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
7398 func[i] = bpf_int_jit_compile(func[i]);
7399 if (!func[i]->jited) {
7405 /* at this point all bpf functions were successfully JITed
7406 * now populate all bpf_calls with correct addresses and
7407 * run last pass of JIT
7409 for (i = 0; i < env->subprog_cnt; i++) {
7410 insn = func[i]->insnsi;
7411 for (j = 0; j < func[i]->len; j++, insn++) {
7412 if (insn->code != (BPF_JMP | BPF_CALL) ||
7413 insn->src_reg != BPF_PSEUDO_CALL)
7415 subprog = insn->off;
7416 insn->imm = (u64 (*)(u64, u64, u64, u64, u64))
7417 func[subprog]->bpf_func -
7421 /* we use the aux data to keep a list of the start addresses
7422 * of the JITed images for each function in the program
7424 * for some architectures, such as powerpc64, the imm field
7425 * might not be large enough to hold the offset of the start
7426 * address of the callee's JITed image from __bpf_call_base
7428 * in such cases, we can lookup the start address of a callee
7429 * by using its subprog id, available from the off field of
7430 * the call instruction, as an index for this list
7432 func[i]->aux->func = func;
7433 func[i]->aux->func_cnt = env->subprog_cnt;
7435 for (i = 0; i < env->subprog_cnt; i++) {
7436 old_bpf_func = func[i]->bpf_func;
7437 tmp = bpf_int_jit_compile(func[i]);
7438 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
7439 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
7446 /* finally lock prog and jit images for all functions and
7449 for (i = 0; i < env->subprog_cnt; i++) {
7450 bpf_prog_lock_ro(func[i]);
7451 bpf_prog_kallsyms_add(func[i]);
7454 /* Last step: make now unused interpreter insns from main
7455 * prog consistent for later dump requests, so they can
7456 * later look the same as if they were interpreted only.
7458 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
7459 if (insn->code != (BPF_JMP | BPF_CALL) ||
7460 insn->src_reg != BPF_PSEUDO_CALL)
7462 insn->off = env->insn_aux_data[i].call_imm;
7463 subprog = find_subprog(env, i + insn->off + 1);
7464 insn->imm = subprog;
7468 prog->bpf_func = func[0]->bpf_func;
7469 prog->aux->func = func;
7470 prog->aux->func_cnt = env->subprog_cnt;
7471 bpf_prog_free_unused_jited_linfo(prog);
7474 for (i = 0; i < env->subprog_cnt; i++)
7476 bpf_jit_free(func[i]);
7479 /* cleanup main prog to be interpreted */
7480 prog->jit_requested = 0;
7481 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
7482 if (insn->code != (BPF_JMP | BPF_CALL) ||
7483 insn->src_reg != BPF_PSEUDO_CALL)
7486 insn->imm = env->insn_aux_data[i].call_imm;
7488 bpf_prog_free_jited_linfo(prog);
7492 static int fixup_call_args(struct bpf_verifier_env *env)
7494 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
7495 struct bpf_prog *prog = env->prog;
7496 struct bpf_insn *insn = prog->insnsi;
7501 if (env->prog->jit_requested &&
7502 !bpf_prog_is_dev_bound(env->prog->aux)) {
7503 err = jit_subprogs(env);
7509 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
7510 for (i = 0; i < prog->len; i++, insn++) {
7511 if (insn->code != (BPF_JMP | BPF_CALL) ||
7512 insn->src_reg != BPF_PSEUDO_CALL)
7514 depth = get_callee_stack_depth(env, insn, i);
7517 bpf_patch_call_args(insn, depth);
7524 /* fixup insn->imm field of bpf_call instructions
7525 * and inline eligible helpers as explicit sequence of BPF instructions
7527 * this function is called after eBPF program passed verification
7529 static int fixup_bpf_calls(struct bpf_verifier_env *env)
7531 struct bpf_prog *prog = env->prog;
7532 struct bpf_insn *insn = prog->insnsi;
7533 const struct bpf_func_proto *fn;
7534 const int insn_cnt = prog->len;
7535 const struct bpf_map_ops *ops;
7536 struct bpf_insn_aux_data *aux;
7537 struct bpf_insn insn_buf[16];
7538 struct bpf_prog *new_prog;
7539 struct bpf_map *map_ptr;
7540 int i, cnt, delta = 0;
7542 for (i = 0; i < insn_cnt; i++, insn++) {
7543 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
7544 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
7545 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
7546 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
7547 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
7548 struct bpf_insn mask_and_div[] = {
7549 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
7551 BPF_JMP_IMM(BPF_JNE, insn->src_reg, 0, 2),
7552 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
7553 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
7556 struct bpf_insn mask_and_mod[] = {
7557 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
7558 /* Rx mod 0 -> Rx */
7559 BPF_JMP_IMM(BPF_JEQ, insn->src_reg, 0, 1),
7562 struct bpf_insn *patchlet;
7564 if (insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
7565 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
7566 patchlet = mask_and_div + (is64 ? 1 : 0);
7567 cnt = ARRAY_SIZE(mask_and_div) - (is64 ? 1 : 0);
7569 patchlet = mask_and_mod + (is64 ? 1 : 0);
7570 cnt = ARRAY_SIZE(mask_and_mod) - (is64 ? 1 : 0);
7573 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
7578 env->prog = prog = new_prog;
7579 insn = new_prog->insnsi + i + delta;
7583 if (BPF_CLASS(insn->code) == BPF_LD &&
7584 (BPF_MODE(insn->code) == BPF_ABS ||
7585 BPF_MODE(insn->code) == BPF_IND)) {
7586 cnt = env->ops->gen_ld_abs(insn, insn_buf);
7587 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
7588 verbose(env, "bpf verifier is misconfigured\n");
7592 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
7597 env->prog = prog = new_prog;
7598 insn = new_prog->insnsi + i + delta;
7602 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
7603 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
7604 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
7605 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
7606 struct bpf_insn insn_buf[16];
7607 struct bpf_insn *patch = &insn_buf[0];
7611 aux = &env->insn_aux_data[i + delta];
7612 if (!aux->alu_state ||
7613 aux->alu_state == BPF_ALU_NON_POINTER)
7616 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
7617 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
7618 BPF_ALU_SANITIZE_SRC;
7620 off_reg = issrc ? insn->src_reg : insn->dst_reg;
7622 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
7623 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit - 1);
7624 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
7625 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
7626 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
7627 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
7629 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX,
7631 insn->src_reg = BPF_REG_AX;
7633 *patch++ = BPF_ALU64_REG(BPF_AND, off_reg,
7637 insn->code = insn->code == code_add ?
7638 code_sub : code_add;
7641 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
7642 cnt = patch - insn_buf;
7644 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
7649 env->prog = prog = new_prog;
7650 insn = new_prog->insnsi + i + delta;
7654 if (insn->code != (BPF_JMP | BPF_CALL))
7656 if (insn->src_reg == BPF_PSEUDO_CALL)
7659 if (insn->imm == BPF_FUNC_get_route_realm)
7660 prog->dst_needed = 1;
7661 if (insn->imm == BPF_FUNC_get_prandom_u32)
7662 bpf_user_rnd_init_once();
7663 if (insn->imm == BPF_FUNC_override_return)
7664 prog->kprobe_override = 1;
7665 if (insn->imm == BPF_FUNC_tail_call) {
7666 /* If we tail call into other programs, we
7667 * cannot make any assumptions since they can
7668 * be replaced dynamically during runtime in
7669 * the program array.
7671 prog->cb_access = 1;
7672 env->prog->aux->stack_depth = MAX_BPF_STACK;
7673 env->prog->aux->max_pkt_offset = MAX_PACKET_OFF;
7675 /* mark bpf_tail_call as different opcode to avoid
7676 * conditional branch in the interpeter for every normal
7677 * call and to prevent accidental JITing by JIT compiler
7678 * that doesn't support bpf_tail_call yet
7681 insn->code = BPF_JMP | BPF_TAIL_CALL;
7683 aux = &env->insn_aux_data[i + delta];
7684 if (!bpf_map_ptr_unpriv(aux))
7687 /* instead of changing every JIT dealing with tail_call
7688 * emit two extra insns:
7689 * if (index >= max_entries) goto out;
7690 * index &= array->index_mask;
7691 * to avoid out-of-bounds cpu speculation
7693 if (bpf_map_ptr_poisoned(aux)) {
7694 verbose(env, "tail_call abusing map_ptr\n");
7698 map_ptr = BPF_MAP_PTR(aux->map_state);
7699 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
7700 map_ptr->max_entries, 2);
7701 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
7702 container_of(map_ptr,
7705 insn_buf[2] = *insn;
7707 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
7712 env->prog = prog = new_prog;
7713 insn = new_prog->insnsi + i + delta;
7717 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
7718 * and other inlining handlers are currently limited to 64 bit
7721 if (prog->jit_requested && BITS_PER_LONG == 64 &&
7722 (insn->imm == BPF_FUNC_map_lookup_elem ||
7723 insn->imm == BPF_FUNC_map_update_elem ||
7724 insn->imm == BPF_FUNC_map_delete_elem ||
7725 insn->imm == BPF_FUNC_map_push_elem ||
7726 insn->imm == BPF_FUNC_map_pop_elem ||
7727 insn->imm == BPF_FUNC_map_peek_elem)) {
7728 aux = &env->insn_aux_data[i + delta];
7729 if (bpf_map_ptr_poisoned(aux))
7730 goto patch_call_imm;
7732 map_ptr = BPF_MAP_PTR(aux->map_state);
7734 if (insn->imm == BPF_FUNC_map_lookup_elem &&
7735 ops->map_gen_lookup) {
7736 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
7737 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
7738 verbose(env, "bpf verifier is misconfigured\n");
7742 new_prog = bpf_patch_insn_data(env, i + delta,
7748 env->prog = prog = new_prog;
7749 insn = new_prog->insnsi + i + delta;
7753 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
7754 (void *(*)(struct bpf_map *map, void *key))NULL));
7755 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
7756 (int (*)(struct bpf_map *map, void *key))NULL));
7757 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
7758 (int (*)(struct bpf_map *map, void *key, void *value,
7760 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
7761 (int (*)(struct bpf_map *map, void *value,
7763 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
7764 (int (*)(struct bpf_map *map, void *value))NULL));
7765 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
7766 (int (*)(struct bpf_map *map, void *value))NULL));
7768 switch (insn->imm) {
7769 case BPF_FUNC_map_lookup_elem:
7770 insn->imm = BPF_CAST_CALL(ops->map_lookup_elem) -
7773 case BPF_FUNC_map_update_elem:
7774 insn->imm = BPF_CAST_CALL(ops->map_update_elem) -
7777 case BPF_FUNC_map_delete_elem:
7778 insn->imm = BPF_CAST_CALL(ops->map_delete_elem) -
7781 case BPF_FUNC_map_push_elem:
7782 insn->imm = BPF_CAST_CALL(ops->map_push_elem) -
7785 case BPF_FUNC_map_pop_elem:
7786 insn->imm = BPF_CAST_CALL(ops->map_pop_elem) -
7789 case BPF_FUNC_map_peek_elem:
7790 insn->imm = BPF_CAST_CALL(ops->map_peek_elem) -
7795 goto patch_call_imm;
7799 fn = env->ops->get_func_proto(insn->imm, env->prog);
7800 /* all functions that have prototype and verifier allowed
7801 * programs to call them, must be real in-kernel functions
7805 "kernel subsystem misconfigured func %s#%d\n",
7806 func_id_name(insn->imm), insn->imm);
7809 insn->imm = fn->func - __bpf_call_base;
7815 static void free_states(struct bpf_verifier_env *env)
7817 struct bpf_verifier_state_list *sl, *sln;
7820 if (!env->explored_states)
7823 for (i = 0; i < env->prog->len; i++) {
7824 sl = env->explored_states[i];
7827 while (sl != STATE_LIST_MARK) {
7829 free_verifier_state(&sl->state, false);
7835 kfree(env->explored_states);
7838 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr,
7839 union bpf_attr __user *uattr)
7841 struct bpf_verifier_env *env;
7842 struct bpf_verifier_log *log;
7843 int i, len, ret = -EINVAL;
7846 /* no program is valid */
7847 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
7850 /* 'struct bpf_verifier_env' can be global, but since it's not small,
7851 * allocate/free it every time bpf_check() is called
7853 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
7859 env->insn_aux_data =
7860 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
7862 if (!env->insn_aux_data)
7864 for (i = 0; i < len; i++)
7865 env->insn_aux_data[i].orig_idx = i;
7867 env->ops = bpf_verifier_ops[env->prog->type];
7869 /* grab the mutex to protect few globals used by verifier */
7870 mutex_lock(&bpf_verifier_lock);
7872 if (attr->log_level || attr->log_buf || attr->log_size) {
7873 /* user requested verbose verifier output
7874 * and supplied buffer to store the verification trace
7876 log->level = attr->log_level;
7877 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
7878 log->len_total = attr->log_size;
7881 /* log attributes have to be sane */
7882 if (log->len_total < 128 || log->len_total > UINT_MAX >> 8 ||
7883 !log->level || !log->ubuf)
7887 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
7888 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
7889 env->strict_alignment = true;
7890 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
7891 env->strict_alignment = false;
7893 is_priv = capable(CAP_SYS_ADMIN);
7894 env->allow_ptr_leaks = is_priv;
7896 ret = replace_map_fd_with_map_ptr(env);
7898 goto skip_full_check;
7900 if (bpf_prog_is_dev_bound(env->prog->aux)) {
7901 ret = bpf_prog_offload_verifier_prep(env->prog);
7903 goto skip_full_check;
7906 env->explored_states = kcalloc(env->prog->len,
7907 sizeof(struct bpf_verifier_state_list *),
7910 if (!env->explored_states)
7911 goto skip_full_check;
7913 ret = check_subprogs(env);
7915 goto skip_full_check;
7917 ret = check_btf_info(env, attr, uattr);
7919 goto skip_full_check;
7921 ret = check_cfg(env);
7923 goto skip_full_check;
7925 ret = do_check(env);
7926 if (env->cur_state) {
7927 free_verifier_state(env->cur_state, true);
7928 env->cur_state = NULL;
7931 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
7932 ret = bpf_prog_offload_finalize(env);
7935 while (!pop_stack(env, NULL, NULL));
7939 ret = check_max_stack_depth(env);
7941 /* instruction rewrites happen after this point */
7944 opt_hard_wire_dead_code_branches(env);
7946 ret = opt_remove_dead_code(env);
7948 ret = opt_remove_nops(env);
7951 sanitize_dead_code(env);
7955 /* program is valid, convert *(u32*)(ctx + off) accesses */
7956 ret = convert_ctx_accesses(env);
7959 ret = fixup_bpf_calls(env);
7962 ret = fixup_call_args(env);
7964 if (log->level && bpf_verifier_log_full(log))
7966 if (log->level && !log->ubuf) {
7968 goto err_release_maps;
7971 if (ret == 0 && env->used_map_cnt) {
7972 /* if program passed verifier, update used_maps in bpf_prog_info */
7973 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
7974 sizeof(env->used_maps[0]),
7977 if (!env->prog->aux->used_maps) {
7979 goto err_release_maps;
7982 memcpy(env->prog->aux->used_maps, env->used_maps,
7983 sizeof(env->used_maps[0]) * env->used_map_cnt);
7984 env->prog->aux->used_map_cnt = env->used_map_cnt;
7986 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
7987 * bpf_ld_imm64 instructions
7989 convert_pseudo_ld_imm64(env);
7993 adjust_btf_func(env);
7996 if (!env->prog->aux->used_maps)
7997 /* if we didn't copy map pointers into bpf_prog_info, release
7998 * them now. Otherwise free_used_maps() will release them.
8003 mutex_unlock(&bpf_verifier_lock);
8004 vfree(env->insn_aux_data);