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_STACK 1024
180 #define BPF_COMPLEXITY_LIMIT_STATES 64
182 #define BPF_MAP_PTR_UNPRIV 1UL
183 #define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \
184 POISON_POINTER_DELTA))
185 #define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
187 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
189 return BPF_MAP_PTR(aux->map_state) == BPF_MAP_PTR_POISON;
192 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
194 return aux->map_state & BPF_MAP_PTR_UNPRIV;
197 static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
198 const struct bpf_map *map, bool unpriv)
200 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON & BPF_MAP_PTR_UNPRIV);
201 unpriv |= bpf_map_ptr_unpriv(aux);
202 aux->map_state = (unsigned long)map |
203 (unpriv ? BPF_MAP_PTR_UNPRIV : 0UL);
206 struct bpf_call_arg_meta {
207 struct bpf_map *map_ptr;
212 s64 msize_smax_value;
213 u64 msize_umax_value;
218 static DEFINE_MUTEX(bpf_verifier_lock);
220 static const struct bpf_line_info *
221 find_linfo(const struct bpf_verifier_env *env, u32 insn_off)
223 const struct bpf_line_info *linfo;
224 const struct bpf_prog *prog;
228 nr_linfo = prog->aux->nr_linfo;
230 if (!nr_linfo || insn_off >= prog->len)
233 linfo = prog->aux->linfo;
234 for (i = 1; i < nr_linfo; i++)
235 if (insn_off < linfo[i].insn_off)
238 return &linfo[i - 1];
241 void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
246 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
248 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
249 "verifier log line truncated - local buffer too short\n");
251 n = min(log->len_total - log->len_used - 1, n);
254 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
260 /* log_level controls verbosity level of eBPF verifier.
261 * bpf_verifier_log_write() is used to dump the verification trace to the log,
262 * so the user can figure out what's wrong with the program
264 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
265 const char *fmt, ...)
269 if (!bpf_verifier_log_needed(&env->log))
273 bpf_verifier_vlog(&env->log, fmt, args);
276 EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
278 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
280 struct bpf_verifier_env *env = private_data;
283 if (!bpf_verifier_log_needed(&env->log))
287 bpf_verifier_vlog(&env->log, fmt, args);
291 static const char *ltrim(const char *s)
299 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env,
301 const char *prefix_fmt, ...)
303 const struct bpf_line_info *linfo;
305 if (!bpf_verifier_log_needed(&env->log))
308 linfo = find_linfo(env, insn_off);
309 if (!linfo || linfo == env->prev_linfo)
315 va_start(args, prefix_fmt);
316 bpf_verifier_vlog(&env->log, prefix_fmt, args);
321 ltrim(btf_name_by_offset(env->prog->aux->btf,
324 env->prev_linfo = linfo;
327 static bool type_is_pkt_pointer(enum bpf_reg_type type)
329 return type == PTR_TO_PACKET ||
330 type == PTR_TO_PACKET_META;
333 static bool type_is_sk_pointer(enum bpf_reg_type type)
335 return type == PTR_TO_SOCKET ||
336 type == PTR_TO_SOCK_COMMON ||
337 type == PTR_TO_TCP_SOCK;
340 static bool reg_type_may_be_null(enum bpf_reg_type type)
342 return type == PTR_TO_MAP_VALUE_OR_NULL ||
343 type == PTR_TO_SOCKET_OR_NULL ||
344 type == PTR_TO_SOCK_COMMON_OR_NULL ||
345 type == PTR_TO_TCP_SOCK_OR_NULL;
348 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
350 return reg->type == PTR_TO_MAP_VALUE &&
351 map_value_has_spin_lock(reg->map_ptr);
354 static bool reg_type_may_be_refcounted_or_null(enum bpf_reg_type type)
356 return type == PTR_TO_SOCKET ||
357 type == PTR_TO_SOCKET_OR_NULL ||
358 type == PTR_TO_TCP_SOCK ||
359 type == PTR_TO_TCP_SOCK_OR_NULL;
362 static bool arg_type_may_be_refcounted(enum bpf_arg_type type)
364 return type == ARG_PTR_TO_SOCK_COMMON;
367 /* Determine whether the function releases some resources allocated by another
368 * function call. The first reference type argument will be assumed to be
369 * released by release_reference().
371 static bool is_release_function(enum bpf_func_id func_id)
373 return func_id == BPF_FUNC_sk_release;
376 static bool is_acquire_function(enum bpf_func_id func_id)
378 return func_id == BPF_FUNC_sk_lookup_tcp ||
379 func_id == BPF_FUNC_sk_lookup_udp ||
380 func_id == BPF_FUNC_skc_lookup_tcp;
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 & BPF_LOG_LEVEL2)
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 */
1144 /* if read wasn't screened by an earlier write ... */
1145 if (writes && state->live & REG_LIVE_WRITTEN)
1147 if (parent->live & REG_LIVE_DONE) {
1148 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
1149 reg_type_str[parent->type],
1150 parent->var_off.value, parent->off);
1153 if (parent->live & REG_LIVE_READ)
1154 /* The parentage chain never changes and
1155 * this parent was already marked as LIVE_READ.
1156 * There is no need to keep walking the chain again and
1157 * keep re-marking all parents as LIVE_READ.
1158 * This case happens when the same register is read
1159 * multiple times without writes into it in-between.
1162 /* ... then we depend on parent's value */
1163 parent->live |= REG_LIVE_READ;
1165 parent = state->parent;
1170 if (env->longest_mark_read_walk < cnt)
1171 env->longest_mark_read_walk = cnt;
1175 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
1176 enum reg_arg_type t)
1178 struct bpf_verifier_state *vstate = env->cur_state;
1179 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1180 struct bpf_reg_state *regs = state->regs;
1182 if (regno >= MAX_BPF_REG) {
1183 verbose(env, "R%d is invalid\n", regno);
1188 /* check whether register used as source operand can be read */
1189 if (regs[regno].type == NOT_INIT) {
1190 verbose(env, "R%d !read_ok\n", regno);
1193 /* We don't need to worry about FP liveness because it's read-only */
1194 if (regno != BPF_REG_FP)
1195 return mark_reg_read(env, ®s[regno],
1196 regs[regno].parent);
1198 /* check whether register used as dest operand can be written to */
1199 if (regno == BPF_REG_FP) {
1200 verbose(env, "frame pointer is read only\n");
1203 regs[regno].live |= REG_LIVE_WRITTEN;
1205 mark_reg_unknown(env, regs, regno);
1210 static bool is_spillable_regtype(enum bpf_reg_type type)
1213 case PTR_TO_MAP_VALUE:
1214 case PTR_TO_MAP_VALUE_OR_NULL:
1218 case PTR_TO_PACKET_META:
1219 case PTR_TO_PACKET_END:
1220 case PTR_TO_FLOW_KEYS:
1221 case CONST_PTR_TO_MAP:
1223 case PTR_TO_SOCKET_OR_NULL:
1224 case PTR_TO_SOCK_COMMON:
1225 case PTR_TO_SOCK_COMMON_OR_NULL:
1226 case PTR_TO_TCP_SOCK:
1227 case PTR_TO_TCP_SOCK_OR_NULL:
1234 /* Does this register contain a constant zero? */
1235 static bool register_is_null(struct bpf_reg_state *reg)
1237 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
1240 /* check_stack_read/write functions track spill/fill of registers,
1241 * stack boundary and alignment are checked in check_mem_access()
1243 static int check_stack_write(struct bpf_verifier_env *env,
1244 struct bpf_func_state *state, /* func where register points to */
1245 int off, int size, int value_regno, int insn_idx)
1247 struct bpf_func_state *cur; /* state of the current function */
1248 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
1249 enum bpf_reg_type type;
1251 err = realloc_func_state(state, round_up(slot + 1, BPF_REG_SIZE),
1252 state->acquired_refs, true);
1255 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
1256 * so it's aligned access and [off, off + size) are within stack limits
1258 if (!env->allow_ptr_leaks &&
1259 state->stack[spi].slot_type[0] == STACK_SPILL &&
1260 size != BPF_REG_SIZE) {
1261 verbose(env, "attempt to corrupt spilled pointer on stack\n");
1265 cur = env->cur_state->frame[env->cur_state->curframe];
1266 if (value_regno >= 0 &&
1267 is_spillable_regtype((type = cur->regs[value_regno].type))) {
1269 /* register containing pointer is being spilled into stack */
1270 if (size != BPF_REG_SIZE) {
1271 verbose(env, "invalid size of register spill\n");
1275 if (state != cur && type == PTR_TO_STACK) {
1276 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
1280 /* save register state */
1281 state->stack[spi].spilled_ptr = cur->regs[value_regno];
1282 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1284 for (i = 0; i < BPF_REG_SIZE; i++) {
1285 if (state->stack[spi].slot_type[i] == STACK_MISC &&
1286 !env->allow_ptr_leaks) {
1287 int *poff = &env->insn_aux_data[insn_idx].sanitize_stack_off;
1288 int soff = (-spi - 1) * BPF_REG_SIZE;
1290 /* detected reuse of integer stack slot with a pointer
1291 * which means either llvm is reusing stack slot or
1292 * an attacker is trying to exploit CVE-2018-3639
1293 * (speculative store bypass)
1294 * Have to sanitize that slot with preemptive
1297 if (*poff && *poff != soff) {
1298 /* disallow programs where single insn stores
1299 * into two different stack slots, since verifier
1300 * cannot sanitize them
1303 "insn %d cannot access two stack slots fp%d and fp%d",
1304 insn_idx, *poff, soff);
1309 state->stack[spi].slot_type[i] = STACK_SPILL;
1312 u8 type = STACK_MISC;
1314 /* regular write of data into stack destroys any spilled ptr */
1315 state->stack[spi].spilled_ptr.type = NOT_INIT;
1316 /* Mark slots as STACK_MISC if they belonged to spilled ptr. */
1317 if (state->stack[spi].slot_type[0] == STACK_SPILL)
1318 for (i = 0; i < BPF_REG_SIZE; i++)
1319 state->stack[spi].slot_type[i] = STACK_MISC;
1321 /* only mark the slot as written if all 8 bytes were written
1322 * otherwise read propagation may incorrectly stop too soon
1323 * when stack slots are partially written.
1324 * This heuristic means that read propagation will be
1325 * conservative, since it will add reg_live_read marks
1326 * to stack slots all the way to first state when programs
1327 * writes+reads less than 8 bytes
1329 if (size == BPF_REG_SIZE)
1330 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1332 /* when we zero initialize stack slots mark them as such */
1333 if (value_regno >= 0 &&
1334 register_is_null(&cur->regs[value_regno]))
1337 /* Mark slots affected by this stack write. */
1338 for (i = 0; i < size; i++)
1339 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
1345 static int check_stack_read(struct bpf_verifier_env *env,
1346 struct bpf_func_state *reg_state /* func where register points to */,
1347 int off, int size, int value_regno)
1349 struct bpf_verifier_state *vstate = env->cur_state;
1350 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1351 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
1354 if (reg_state->allocated_stack <= slot) {
1355 verbose(env, "invalid read from stack off %d+0 size %d\n",
1359 stype = reg_state->stack[spi].slot_type;
1361 if (stype[0] == STACK_SPILL) {
1362 if (size != BPF_REG_SIZE) {
1363 verbose(env, "invalid size of register spill\n");
1366 for (i = 1; i < BPF_REG_SIZE; i++) {
1367 if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) {
1368 verbose(env, "corrupted spill memory\n");
1373 if (value_regno >= 0) {
1374 /* restore register state from stack */
1375 state->regs[value_regno] = reg_state->stack[spi].spilled_ptr;
1376 /* mark reg as written since spilled pointer state likely
1377 * has its liveness marks cleared by is_state_visited()
1378 * which resets stack/reg liveness for state transitions
1380 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
1382 mark_reg_read(env, ®_state->stack[spi].spilled_ptr,
1383 reg_state->stack[spi].spilled_ptr.parent);
1388 for (i = 0; i < size; i++) {
1389 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_MISC)
1391 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_ZERO) {
1395 verbose(env, "invalid read from stack off %d+%d size %d\n",
1399 mark_reg_read(env, ®_state->stack[spi].spilled_ptr,
1400 reg_state->stack[spi].spilled_ptr.parent);
1401 if (value_regno >= 0) {
1402 if (zeros == size) {
1403 /* any size read into register is zero extended,
1404 * so the whole register == const_zero
1406 __mark_reg_const_zero(&state->regs[value_regno]);
1408 /* have read misc data from the stack */
1409 mark_reg_unknown(env, state->regs, value_regno);
1411 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
1417 static int check_stack_access(struct bpf_verifier_env *env,
1418 const struct bpf_reg_state *reg,
1421 /* Stack accesses must be at a fixed offset, so that we
1422 * can determine what type of data were returned. See
1423 * check_stack_read().
1425 if (!tnum_is_const(reg->var_off)) {
1428 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1429 verbose(env, "variable stack access var_off=%s off=%d size=%d",
1434 if (off >= 0 || off < -MAX_BPF_STACK) {
1435 verbose(env, "invalid stack off=%d size=%d\n", off, size);
1442 /* check read/write into map element returned by bpf_map_lookup_elem() */
1443 static int __check_map_access(struct bpf_verifier_env *env, u32 regno, int off,
1444 int size, bool zero_size_allowed)
1446 struct bpf_reg_state *regs = cur_regs(env);
1447 struct bpf_map *map = regs[regno].map_ptr;
1449 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
1450 off + size > map->value_size) {
1451 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
1452 map->value_size, off, size);
1458 /* check read/write into a map element with possible variable offset */
1459 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
1460 int off, int size, bool zero_size_allowed)
1462 struct bpf_verifier_state *vstate = env->cur_state;
1463 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1464 struct bpf_reg_state *reg = &state->regs[regno];
1467 /* We may have adjusted the register to this map value, so we
1468 * need to try adding each of min_value and max_value to off
1469 * to make sure our theoretical access will be safe.
1471 if (env->log.level & BPF_LOG_LEVEL)
1472 print_verifier_state(env, state);
1474 /* The minimum value is only important with signed
1475 * comparisons where we can't assume the floor of a
1476 * value is 0. If we are using signed variables for our
1477 * index'es we need to make sure that whatever we use
1478 * will have a set floor within our range.
1480 if (reg->smin_value < 0 &&
1481 (reg->smin_value == S64_MIN ||
1482 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
1483 reg->smin_value + off < 0)) {
1484 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1488 err = __check_map_access(env, regno, reg->smin_value + off, size,
1491 verbose(env, "R%d min value is outside of the array range\n",
1496 /* If we haven't set a max value then we need to bail since we can't be
1497 * sure we won't do bad things.
1498 * If reg->umax_value + off could overflow, treat that as unbounded too.
1500 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
1501 verbose(env, "R%d unbounded memory access, make sure to bounds check any array access into a map\n",
1505 err = __check_map_access(env, regno, reg->umax_value + off, size,
1508 verbose(env, "R%d max value is outside of the array range\n",
1511 if (map_value_has_spin_lock(reg->map_ptr)) {
1512 u32 lock = reg->map_ptr->spin_lock_off;
1514 /* if any part of struct bpf_spin_lock can be touched by
1515 * load/store reject this program.
1516 * To check that [x1, x2) overlaps with [y1, y2)
1517 * it is sufficient to check x1 < y2 && y1 < x2.
1519 if (reg->smin_value + off < lock + sizeof(struct bpf_spin_lock) &&
1520 lock < reg->umax_value + off + size) {
1521 verbose(env, "bpf_spin_lock cannot be accessed directly by load/store\n");
1528 #define MAX_PACKET_OFF 0xffff
1530 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
1531 const struct bpf_call_arg_meta *meta,
1532 enum bpf_access_type t)
1534 switch (env->prog->type) {
1535 /* Program types only with direct read access go here! */
1536 case BPF_PROG_TYPE_LWT_IN:
1537 case BPF_PROG_TYPE_LWT_OUT:
1538 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
1539 case BPF_PROG_TYPE_SK_REUSEPORT:
1540 case BPF_PROG_TYPE_FLOW_DISSECTOR:
1541 case BPF_PROG_TYPE_CGROUP_SKB:
1546 /* Program types with direct read + write access go here! */
1547 case BPF_PROG_TYPE_SCHED_CLS:
1548 case BPF_PROG_TYPE_SCHED_ACT:
1549 case BPF_PROG_TYPE_XDP:
1550 case BPF_PROG_TYPE_LWT_XMIT:
1551 case BPF_PROG_TYPE_SK_SKB:
1552 case BPF_PROG_TYPE_SK_MSG:
1554 return meta->pkt_access;
1556 env->seen_direct_write = true;
1563 static int __check_packet_access(struct bpf_verifier_env *env, u32 regno,
1564 int off, int size, bool zero_size_allowed)
1566 struct bpf_reg_state *regs = cur_regs(env);
1567 struct bpf_reg_state *reg = ®s[regno];
1569 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
1570 (u64)off + size > reg->range) {
1571 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
1572 off, size, regno, reg->id, reg->off, reg->range);
1578 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
1579 int size, bool zero_size_allowed)
1581 struct bpf_reg_state *regs = cur_regs(env);
1582 struct bpf_reg_state *reg = ®s[regno];
1585 /* We may have added a variable offset to the packet pointer; but any
1586 * reg->range we have comes after that. We are only checking the fixed
1590 /* We don't allow negative numbers, because we aren't tracking enough
1591 * detail to prove they're safe.
1593 if (reg->smin_value < 0) {
1594 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1598 err = __check_packet_access(env, regno, off, size, zero_size_allowed);
1600 verbose(env, "R%d offset is outside of the packet\n", regno);
1604 /* __check_packet_access has made sure "off + size - 1" is within u16.
1605 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
1606 * otherwise find_good_pkt_pointers would have refused to set range info
1607 * that __check_packet_access would have rejected this pkt access.
1608 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
1610 env->prog->aux->max_pkt_offset =
1611 max_t(u32, env->prog->aux->max_pkt_offset,
1612 off + reg->umax_value + size - 1);
1617 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
1618 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
1619 enum bpf_access_type t, enum bpf_reg_type *reg_type)
1621 struct bpf_insn_access_aux info = {
1622 .reg_type = *reg_type,
1625 if (env->ops->is_valid_access &&
1626 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
1627 /* A non zero info.ctx_field_size indicates that this field is a
1628 * candidate for later verifier transformation to load the whole
1629 * field and then apply a mask when accessed with a narrower
1630 * access than actual ctx access size. A zero info.ctx_field_size
1631 * will only allow for whole field access and rejects any other
1632 * type of narrower access.
1634 *reg_type = info.reg_type;
1636 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
1637 /* remember the offset of last byte accessed in ctx */
1638 if (env->prog->aux->max_ctx_offset < off + size)
1639 env->prog->aux->max_ctx_offset = off + size;
1643 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
1647 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
1650 if (size < 0 || off < 0 ||
1651 (u64)off + size > sizeof(struct bpf_flow_keys)) {
1652 verbose(env, "invalid access to flow keys off=%d size=%d\n",
1659 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
1660 u32 regno, int off, int size,
1661 enum bpf_access_type t)
1663 struct bpf_reg_state *regs = cur_regs(env);
1664 struct bpf_reg_state *reg = ®s[regno];
1665 struct bpf_insn_access_aux info = {};
1668 if (reg->smin_value < 0) {
1669 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
1674 switch (reg->type) {
1675 case PTR_TO_SOCK_COMMON:
1676 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
1679 valid = bpf_sock_is_valid_access(off, size, t, &info);
1681 case PTR_TO_TCP_SOCK:
1682 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
1690 env->insn_aux_data[insn_idx].ctx_field_size =
1691 info.ctx_field_size;
1695 verbose(env, "R%d invalid %s access off=%d size=%d\n",
1696 regno, reg_type_str[reg->type], off, size);
1701 static bool __is_pointer_value(bool allow_ptr_leaks,
1702 const struct bpf_reg_state *reg)
1704 if (allow_ptr_leaks)
1707 return reg->type != SCALAR_VALUE;
1710 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
1712 return cur_regs(env) + regno;
1715 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
1717 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
1720 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
1722 const struct bpf_reg_state *reg = reg_state(env, regno);
1724 return reg->type == PTR_TO_CTX;
1727 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
1729 const struct bpf_reg_state *reg = reg_state(env, regno);
1731 return type_is_sk_pointer(reg->type);
1734 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
1736 const struct bpf_reg_state *reg = reg_state(env, regno);
1738 return type_is_pkt_pointer(reg->type);
1741 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
1743 const struct bpf_reg_state *reg = reg_state(env, regno);
1745 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
1746 return reg->type == PTR_TO_FLOW_KEYS;
1749 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
1750 const struct bpf_reg_state *reg,
1751 int off, int size, bool strict)
1753 struct tnum reg_off;
1756 /* Byte size accesses are always allowed. */
1757 if (!strict || size == 1)
1760 /* For platforms that do not have a Kconfig enabling
1761 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
1762 * NET_IP_ALIGN is universally set to '2'. And on platforms
1763 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
1764 * to this code only in strict mode where we want to emulate
1765 * the NET_IP_ALIGN==2 checking. Therefore use an
1766 * unconditional IP align value of '2'.
1770 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
1771 if (!tnum_is_aligned(reg_off, size)) {
1774 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1776 "misaligned packet access off %d+%s+%d+%d size %d\n",
1777 ip_align, tn_buf, reg->off, off, size);
1784 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
1785 const struct bpf_reg_state *reg,
1786 const char *pointer_desc,
1787 int off, int size, bool strict)
1789 struct tnum reg_off;
1791 /* Byte size accesses are always allowed. */
1792 if (!strict || size == 1)
1795 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
1796 if (!tnum_is_aligned(reg_off, size)) {
1799 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1800 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
1801 pointer_desc, tn_buf, reg->off, off, size);
1808 static int check_ptr_alignment(struct bpf_verifier_env *env,
1809 const struct bpf_reg_state *reg, int off,
1810 int size, bool strict_alignment_once)
1812 bool strict = env->strict_alignment || strict_alignment_once;
1813 const char *pointer_desc = "";
1815 switch (reg->type) {
1817 case PTR_TO_PACKET_META:
1818 /* Special case, because of NET_IP_ALIGN. Given metadata sits
1819 * right in front, treat it the very same way.
1821 return check_pkt_ptr_alignment(env, reg, off, size, strict);
1822 case PTR_TO_FLOW_KEYS:
1823 pointer_desc = "flow keys ";
1825 case PTR_TO_MAP_VALUE:
1826 pointer_desc = "value ";
1829 pointer_desc = "context ";
1832 pointer_desc = "stack ";
1833 /* The stack spill tracking logic in check_stack_write()
1834 * and check_stack_read() relies on stack accesses being
1840 pointer_desc = "sock ";
1842 case PTR_TO_SOCK_COMMON:
1843 pointer_desc = "sock_common ";
1845 case PTR_TO_TCP_SOCK:
1846 pointer_desc = "tcp_sock ";
1851 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
1855 static int update_stack_depth(struct bpf_verifier_env *env,
1856 const struct bpf_func_state *func,
1859 u16 stack = env->subprog_info[func->subprogno].stack_depth;
1864 /* update known max for given subprogram */
1865 env->subprog_info[func->subprogno].stack_depth = -off;
1869 /* starting from main bpf function walk all instructions of the function
1870 * and recursively walk all callees that given function can call.
1871 * Ignore jump and exit insns.
1872 * Since recursion is prevented by check_cfg() this algorithm
1873 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
1875 static int check_max_stack_depth(struct bpf_verifier_env *env)
1877 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
1878 struct bpf_subprog_info *subprog = env->subprog_info;
1879 struct bpf_insn *insn = env->prog->insnsi;
1880 int ret_insn[MAX_CALL_FRAMES];
1881 int ret_prog[MAX_CALL_FRAMES];
1884 /* round up to 32-bytes, since this is granularity
1885 * of interpreter stack size
1887 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
1888 if (depth > MAX_BPF_STACK) {
1889 verbose(env, "combined stack size of %d calls is %d. Too large\n",
1894 subprog_end = subprog[idx + 1].start;
1895 for (; i < subprog_end; i++) {
1896 if (insn[i].code != (BPF_JMP | BPF_CALL))
1898 if (insn[i].src_reg != BPF_PSEUDO_CALL)
1900 /* remember insn and function to return to */
1901 ret_insn[frame] = i + 1;
1902 ret_prog[frame] = idx;
1904 /* find the callee */
1905 i = i + insn[i].imm + 1;
1906 idx = find_subprog(env, i);
1908 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
1913 if (frame >= MAX_CALL_FRAMES) {
1914 WARN_ONCE(1, "verifier bug. Call stack is too deep\n");
1919 /* end of for() loop means the last insn of the 'subprog'
1920 * was reached. Doesn't matter whether it was JA or EXIT
1924 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
1926 i = ret_insn[frame];
1927 idx = ret_prog[frame];
1931 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
1932 static int get_callee_stack_depth(struct bpf_verifier_env *env,
1933 const struct bpf_insn *insn, int idx)
1935 int start = idx + insn->imm + 1, subprog;
1937 subprog = find_subprog(env, start);
1939 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
1943 return env->subprog_info[subprog].stack_depth;
1947 static int check_ctx_reg(struct bpf_verifier_env *env,
1948 const struct bpf_reg_state *reg, int regno)
1950 /* Access to ctx or passing it to a helper is only allowed in
1951 * its original, unmodified form.
1955 verbose(env, "dereference of modified ctx ptr R%d off=%d disallowed\n",
1960 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
1963 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
1964 verbose(env, "variable ctx access var_off=%s disallowed\n", tn_buf);
1971 /* truncate register to smaller size (in bytes)
1972 * must be called with size < BPF_REG_SIZE
1974 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
1978 /* clear high bits in bit representation */
1979 reg->var_off = tnum_cast(reg->var_off, size);
1981 /* fix arithmetic bounds */
1982 mask = ((u64)1 << (size * 8)) - 1;
1983 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
1984 reg->umin_value &= mask;
1985 reg->umax_value &= mask;
1987 reg->umin_value = 0;
1988 reg->umax_value = mask;
1990 reg->smin_value = reg->umin_value;
1991 reg->smax_value = reg->umax_value;
1994 /* check whether memory at (regno + off) is accessible for t = (read | write)
1995 * if t==write, value_regno is a register which value is stored into memory
1996 * if t==read, value_regno is a register which will receive the value from memory
1997 * if t==write && value_regno==-1, some unknown value is stored into memory
1998 * if t==read && value_regno==-1, don't care what we read from memory
2000 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
2001 int off, int bpf_size, enum bpf_access_type t,
2002 int value_regno, bool strict_alignment_once)
2004 struct bpf_reg_state *regs = cur_regs(env);
2005 struct bpf_reg_state *reg = regs + regno;
2006 struct bpf_func_state *state;
2009 size = bpf_size_to_bytes(bpf_size);
2013 /* alignment checks will add in reg->off themselves */
2014 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
2018 /* for access checks, reg->off is just part of off */
2021 if (reg->type == PTR_TO_MAP_VALUE) {
2022 if (t == BPF_WRITE && value_regno >= 0 &&
2023 is_pointer_value(env, value_regno)) {
2024 verbose(env, "R%d leaks addr into map\n", value_regno);
2028 err = check_map_access(env, regno, off, size, false);
2029 if (!err && t == BPF_READ && value_regno >= 0)
2030 mark_reg_unknown(env, regs, value_regno);
2032 } else if (reg->type == PTR_TO_CTX) {
2033 enum bpf_reg_type reg_type = SCALAR_VALUE;
2035 if (t == BPF_WRITE && value_regno >= 0 &&
2036 is_pointer_value(env, value_regno)) {
2037 verbose(env, "R%d leaks addr into ctx\n", value_regno);
2041 err = check_ctx_reg(env, reg, regno);
2045 err = check_ctx_access(env, insn_idx, off, size, t, ®_type);
2046 if (!err && t == BPF_READ && value_regno >= 0) {
2047 /* ctx access returns either a scalar, or a
2048 * PTR_TO_PACKET[_META,_END]. In the latter
2049 * case, we know the offset is zero.
2051 if (reg_type == SCALAR_VALUE) {
2052 mark_reg_unknown(env, regs, value_regno);
2054 mark_reg_known_zero(env, regs,
2056 if (reg_type_may_be_null(reg_type))
2057 regs[value_regno].id = ++env->id_gen;
2059 regs[value_regno].type = reg_type;
2062 } else if (reg->type == PTR_TO_STACK) {
2063 off += reg->var_off.value;
2064 err = check_stack_access(env, reg, off, size);
2068 state = func(env, reg);
2069 err = update_stack_depth(env, state, off);
2074 err = check_stack_write(env, state, off, size,
2075 value_regno, insn_idx);
2077 err = check_stack_read(env, state, off, size,
2079 } else if (reg_is_pkt_pointer(reg)) {
2080 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
2081 verbose(env, "cannot write into packet\n");
2084 if (t == BPF_WRITE && value_regno >= 0 &&
2085 is_pointer_value(env, value_regno)) {
2086 verbose(env, "R%d leaks addr into packet\n",
2090 err = check_packet_access(env, regno, off, size, false);
2091 if (!err && t == BPF_READ && value_regno >= 0)
2092 mark_reg_unknown(env, regs, value_regno);
2093 } else if (reg->type == PTR_TO_FLOW_KEYS) {
2094 if (t == BPF_WRITE && value_regno >= 0 &&
2095 is_pointer_value(env, value_regno)) {
2096 verbose(env, "R%d leaks addr into flow keys\n",
2101 err = check_flow_keys_access(env, off, size);
2102 if (!err && t == BPF_READ && value_regno >= 0)
2103 mark_reg_unknown(env, regs, value_regno);
2104 } else if (type_is_sk_pointer(reg->type)) {
2105 if (t == BPF_WRITE) {
2106 verbose(env, "R%d cannot write into %s\n",
2107 regno, reg_type_str[reg->type]);
2110 err = check_sock_access(env, insn_idx, regno, off, size, t);
2111 if (!err && value_regno >= 0)
2112 mark_reg_unknown(env, regs, value_regno);
2114 verbose(env, "R%d invalid mem access '%s'\n", regno,
2115 reg_type_str[reg->type]);
2119 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
2120 regs[value_regno].type == SCALAR_VALUE) {
2121 /* b/h/w load zero-extends, mark upper bits as known 0 */
2122 coerce_reg_to_size(®s[value_regno], size);
2127 static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
2131 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
2133 verbose(env, "BPF_XADD uses reserved fields\n");
2137 /* check src1 operand */
2138 err = check_reg_arg(env, insn->src_reg, SRC_OP);
2142 /* check src2 operand */
2143 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
2147 if (is_pointer_value(env, insn->src_reg)) {
2148 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
2152 if (is_ctx_reg(env, insn->dst_reg) ||
2153 is_pkt_reg(env, insn->dst_reg) ||
2154 is_flow_key_reg(env, insn->dst_reg) ||
2155 is_sk_reg(env, insn->dst_reg)) {
2156 verbose(env, "BPF_XADD stores into R%d %s is not allowed\n",
2158 reg_type_str[reg_state(env, insn->dst_reg)->type]);
2162 /* check whether atomic_add can read the memory */
2163 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
2164 BPF_SIZE(insn->code), BPF_READ, -1, true);
2168 /* check whether atomic_add can write into the same memory */
2169 return check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
2170 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
2173 static int __check_stack_boundary(struct bpf_verifier_env *env, u32 regno,
2174 int off, int access_size,
2175 bool zero_size_allowed)
2177 struct bpf_reg_state *reg = reg_state(env, regno);
2179 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
2180 access_size < 0 || (access_size == 0 && !zero_size_allowed)) {
2181 if (tnum_is_const(reg->var_off)) {
2182 verbose(env, "invalid stack type R%d off=%d access_size=%d\n",
2183 regno, off, access_size);
2187 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2188 verbose(env, "invalid stack type R%d var_off=%s access_size=%d\n",
2189 regno, tn_buf, access_size);
2196 /* when register 'regno' is passed into function that will read 'access_size'
2197 * bytes from that pointer, make sure that it's within stack boundary
2198 * and all elements of stack are initialized.
2199 * Unlike most pointer bounds-checking functions, this one doesn't take an
2200 * 'off' argument, so it has to add in reg->off itself.
2202 static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
2203 int access_size, bool zero_size_allowed,
2204 struct bpf_call_arg_meta *meta)
2206 struct bpf_reg_state *reg = reg_state(env, regno);
2207 struct bpf_func_state *state = func(env, reg);
2208 int err, min_off, max_off, i, slot, spi;
2210 if (reg->type != PTR_TO_STACK) {
2211 /* Allow zero-byte read from NULL, regardless of pointer type */
2212 if (zero_size_allowed && access_size == 0 &&
2213 register_is_null(reg))
2216 verbose(env, "R%d type=%s expected=%s\n", regno,
2217 reg_type_str[reg->type],
2218 reg_type_str[PTR_TO_STACK]);
2222 if (tnum_is_const(reg->var_off)) {
2223 min_off = max_off = reg->var_off.value + reg->off;
2224 err = __check_stack_boundary(env, regno, min_off, access_size,
2229 /* Only initialized buffer on stack is allowed to be accessed
2230 * with variable offset. With uninitialized buffer it's hard to
2231 * guarantee that whole memory is marked as initialized on
2232 * helper return since specific bounds are unknown what may
2233 * cause uninitialized stack leaking.
2235 if (meta && meta->raw_mode)
2238 min_off = reg->smin_value + reg->off;
2239 max_off = reg->umax_value + reg->off;
2240 err = __check_stack_boundary(env, regno, min_off, access_size,
2244 err = __check_stack_boundary(env, regno, max_off, access_size,
2250 if (meta && meta->raw_mode) {
2251 meta->access_size = access_size;
2252 meta->regno = regno;
2256 for (i = min_off; i < max_off + access_size; i++) {
2260 spi = slot / BPF_REG_SIZE;
2261 if (state->allocated_stack <= slot)
2263 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
2264 if (*stype == STACK_MISC)
2266 if (*stype == STACK_ZERO) {
2267 /* helper can write anything into the stack */
2268 *stype = STACK_MISC;
2272 if (tnum_is_const(reg->var_off)) {
2273 verbose(env, "invalid indirect read from stack off %d+%d size %d\n",
2274 min_off, i - min_off, access_size);
2278 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2279 verbose(env, "invalid indirect read from stack var_off %s+%d size %d\n",
2280 tn_buf, i - min_off, access_size);
2284 /* reading any byte out of 8-byte 'spill_slot' will cause
2285 * the whole slot to be marked as 'read'
2287 mark_reg_read(env, &state->stack[spi].spilled_ptr,
2288 state->stack[spi].spilled_ptr.parent);
2290 return update_stack_depth(env, state, min_off);
2293 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
2294 int access_size, bool zero_size_allowed,
2295 struct bpf_call_arg_meta *meta)
2297 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
2299 switch (reg->type) {
2301 case PTR_TO_PACKET_META:
2302 return check_packet_access(env, regno, reg->off, access_size,
2304 case PTR_TO_MAP_VALUE:
2305 return check_map_access(env, regno, reg->off, access_size,
2307 default: /* scalar_value|ptr_to_stack or invalid ptr */
2308 return check_stack_boundary(env, regno, access_size,
2309 zero_size_allowed, meta);
2313 /* Implementation details:
2314 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL
2315 * Two bpf_map_lookups (even with the same key) will have different reg->id.
2316 * For traditional PTR_TO_MAP_VALUE the verifier clears reg->id after
2317 * value_or_null->value transition, since the verifier only cares about
2318 * the range of access to valid map value pointer and doesn't care about actual
2319 * address of the map element.
2320 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
2321 * reg->id > 0 after value_or_null->value transition. By doing so
2322 * two bpf_map_lookups will be considered two different pointers that
2323 * point to different bpf_spin_locks.
2324 * The verifier allows taking only one bpf_spin_lock at a time to avoid
2326 * Since only one bpf_spin_lock is allowed the checks are simpler than
2327 * reg_is_refcounted() logic. The verifier needs to remember only
2328 * one spin_lock instead of array of acquired_refs.
2329 * cur_state->active_spin_lock remembers which map value element got locked
2330 * and clears it after bpf_spin_unlock.
2332 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
2335 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
2336 struct bpf_verifier_state *cur = env->cur_state;
2337 bool is_const = tnum_is_const(reg->var_off);
2338 struct bpf_map *map = reg->map_ptr;
2339 u64 val = reg->var_off.value;
2341 if (reg->type != PTR_TO_MAP_VALUE) {
2342 verbose(env, "R%d is not a pointer to map_value\n", regno);
2347 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
2353 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
2357 if (!map_value_has_spin_lock(map)) {
2358 if (map->spin_lock_off == -E2BIG)
2360 "map '%s' has more than one 'struct bpf_spin_lock'\n",
2362 else if (map->spin_lock_off == -ENOENT)
2364 "map '%s' doesn't have 'struct bpf_spin_lock'\n",
2368 "map '%s' is not a struct type or bpf_spin_lock is mangled\n",
2372 if (map->spin_lock_off != val + reg->off) {
2373 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock'\n",
2378 if (cur->active_spin_lock) {
2380 "Locking two bpf_spin_locks are not allowed\n");
2383 cur->active_spin_lock = reg->id;
2385 if (!cur->active_spin_lock) {
2386 verbose(env, "bpf_spin_unlock without taking a lock\n");
2389 if (cur->active_spin_lock != reg->id) {
2390 verbose(env, "bpf_spin_unlock of different lock\n");
2393 cur->active_spin_lock = 0;
2398 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
2400 return type == ARG_PTR_TO_MEM ||
2401 type == ARG_PTR_TO_MEM_OR_NULL ||
2402 type == ARG_PTR_TO_UNINIT_MEM;
2405 static bool arg_type_is_mem_size(enum bpf_arg_type type)
2407 return type == ARG_CONST_SIZE ||
2408 type == ARG_CONST_SIZE_OR_ZERO;
2411 static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
2412 enum bpf_arg_type arg_type,
2413 struct bpf_call_arg_meta *meta)
2415 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
2416 enum bpf_reg_type expected_type, type = reg->type;
2419 if (arg_type == ARG_DONTCARE)
2422 err = check_reg_arg(env, regno, SRC_OP);
2426 if (arg_type == ARG_ANYTHING) {
2427 if (is_pointer_value(env, regno)) {
2428 verbose(env, "R%d leaks addr into helper function\n",
2435 if (type_is_pkt_pointer(type) &&
2436 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
2437 verbose(env, "helper access to the packet is not allowed\n");
2441 if (arg_type == ARG_PTR_TO_MAP_KEY ||
2442 arg_type == ARG_PTR_TO_MAP_VALUE ||
2443 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
2444 expected_type = PTR_TO_STACK;
2445 if (!type_is_pkt_pointer(type) && type != PTR_TO_MAP_VALUE &&
2446 type != expected_type)
2448 } else if (arg_type == ARG_CONST_SIZE ||
2449 arg_type == ARG_CONST_SIZE_OR_ZERO) {
2450 expected_type = SCALAR_VALUE;
2451 if (type != expected_type)
2453 } else if (arg_type == ARG_CONST_MAP_PTR) {
2454 expected_type = CONST_PTR_TO_MAP;
2455 if (type != expected_type)
2457 } else if (arg_type == ARG_PTR_TO_CTX) {
2458 expected_type = PTR_TO_CTX;
2459 if (type != expected_type)
2461 err = check_ctx_reg(env, reg, regno);
2464 } else if (arg_type == ARG_PTR_TO_SOCK_COMMON) {
2465 expected_type = PTR_TO_SOCK_COMMON;
2466 /* Any sk pointer can be ARG_PTR_TO_SOCK_COMMON */
2467 if (!type_is_sk_pointer(type))
2469 if (reg->ref_obj_id) {
2470 if (meta->ref_obj_id) {
2471 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
2472 regno, reg->ref_obj_id,
2476 meta->ref_obj_id = reg->ref_obj_id;
2478 } else if (arg_type == ARG_PTR_TO_SPIN_LOCK) {
2479 if (meta->func_id == BPF_FUNC_spin_lock) {
2480 if (process_spin_lock(env, regno, true))
2482 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
2483 if (process_spin_lock(env, regno, false))
2486 verbose(env, "verifier internal error\n");
2489 } else if (arg_type_is_mem_ptr(arg_type)) {
2490 expected_type = PTR_TO_STACK;
2491 /* One exception here. In case function allows for NULL to be
2492 * passed in as argument, it's a SCALAR_VALUE type. Final test
2493 * happens during stack boundary checking.
2495 if (register_is_null(reg) &&
2496 arg_type == ARG_PTR_TO_MEM_OR_NULL)
2497 /* final test in check_stack_boundary() */;
2498 else if (!type_is_pkt_pointer(type) &&
2499 type != PTR_TO_MAP_VALUE &&
2500 type != expected_type)
2502 meta->raw_mode = arg_type == ARG_PTR_TO_UNINIT_MEM;
2504 verbose(env, "unsupported arg_type %d\n", arg_type);
2508 if (arg_type == ARG_CONST_MAP_PTR) {
2509 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
2510 meta->map_ptr = reg->map_ptr;
2511 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
2512 /* bpf_map_xxx(..., map_ptr, ..., key) call:
2513 * check that [key, key + map->key_size) are within
2514 * stack limits and initialized
2516 if (!meta->map_ptr) {
2517 /* in function declaration map_ptr must come before
2518 * map_key, so that it's verified and known before
2519 * we have to check map_key here. Otherwise it means
2520 * that kernel subsystem misconfigured verifier
2522 verbose(env, "invalid map_ptr to access map->key\n");
2525 err = check_helper_mem_access(env, regno,
2526 meta->map_ptr->key_size, false,
2528 } else if (arg_type == ARG_PTR_TO_MAP_VALUE ||
2529 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
2530 /* bpf_map_xxx(..., map_ptr, ..., value) call:
2531 * check [value, value + map->value_size) validity
2533 if (!meta->map_ptr) {
2534 /* kernel subsystem misconfigured verifier */
2535 verbose(env, "invalid map_ptr to access map->value\n");
2538 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE);
2539 err = check_helper_mem_access(env, regno,
2540 meta->map_ptr->value_size, false,
2542 } else if (arg_type_is_mem_size(arg_type)) {
2543 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
2545 /* remember the mem_size which may be used later
2546 * to refine return values.
2548 meta->msize_smax_value = reg->smax_value;
2549 meta->msize_umax_value = reg->umax_value;
2551 /* The register is SCALAR_VALUE; the access check
2552 * happens using its boundaries.
2554 if (!tnum_is_const(reg->var_off))
2555 /* For unprivileged variable accesses, disable raw
2556 * mode so that the program is required to
2557 * initialize all the memory that the helper could
2558 * just partially fill up.
2562 if (reg->smin_value < 0) {
2563 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
2568 if (reg->umin_value == 0) {
2569 err = check_helper_mem_access(env, regno - 1, 0,
2576 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
2577 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
2581 err = check_helper_mem_access(env, regno - 1,
2583 zero_size_allowed, meta);
2588 verbose(env, "R%d type=%s expected=%s\n", regno,
2589 reg_type_str[type], reg_type_str[expected_type]);
2593 static int check_map_func_compatibility(struct bpf_verifier_env *env,
2594 struct bpf_map *map, int func_id)
2599 /* We need a two way check, first is from map perspective ... */
2600 switch (map->map_type) {
2601 case BPF_MAP_TYPE_PROG_ARRAY:
2602 if (func_id != BPF_FUNC_tail_call)
2605 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
2606 if (func_id != BPF_FUNC_perf_event_read &&
2607 func_id != BPF_FUNC_perf_event_output &&
2608 func_id != BPF_FUNC_perf_event_read_value)
2611 case BPF_MAP_TYPE_STACK_TRACE:
2612 if (func_id != BPF_FUNC_get_stackid)
2615 case BPF_MAP_TYPE_CGROUP_ARRAY:
2616 if (func_id != BPF_FUNC_skb_under_cgroup &&
2617 func_id != BPF_FUNC_current_task_under_cgroup)
2620 case BPF_MAP_TYPE_CGROUP_STORAGE:
2621 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
2622 if (func_id != BPF_FUNC_get_local_storage)
2625 /* devmap returns a pointer to a live net_device ifindex that we cannot
2626 * allow to be modified from bpf side. So do not allow lookup elements
2629 case BPF_MAP_TYPE_DEVMAP:
2630 if (func_id != BPF_FUNC_redirect_map)
2633 /* Restrict bpf side of cpumap and xskmap, open when use-cases
2636 case BPF_MAP_TYPE_CPUMAP:
2637 case BPF_MAP_TYPE_XSKMAP:
2638 if (func_id != BPF_FUNC_redirect_map)
2641 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
2642 case BPF_MAP_TYPE_HASH_OF_MAPS:
2643 if (func_id != BPF_FUNC_map_lookup_elem)
2646 case BPF_MAP_TYPE_SOCKMAP:
2647 if (func_id != BPF_FUNC_sk_redirect_map &&
2648 func_id != BPF_FUNC_sock_map_update &&
2649 func_id != BPF_FUNC_map_delete_elem &&
2650 func_id != BPF_FUNC_msg_redirect_map)
2653 case BPF_MAP_TYPE_SOCKHASH:
2654 if (func_id != BPF_FUNC_sk_redirect_hash &&
2655 func_id != BPF_FUNC_sock_hash_update &&
2656 func_id != BPF_FUNC_map_delete_elem &&
2657 func_id != BPF_FUNC_msg_redirect_hash)
2660 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
2661 if (func_id != BPF_FUNC_sk_select_reuseport)
2664 case BPF_MAP_TYPE_QUEUE:
2665 case BPF_MAP_TYPE_STACK:
2666 if (func_id != BPF_FUNC_map_peek_elem &&
2667 func_id != BPF_FUNC_map_pop_elem &&
2668 func_id != BPF_FUNC_map_push_elem)
2675 /* ... and second from the function itself. */
2677 case BPF_FUNC_tail_call:
2678 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
2680 if (env->subprog_cnt > 1) {
2681 verbose(env, "tail_calls are not allowed in programs with bpf-to-bpf calls\n");
2685 case BPF_FUNC_perf_event_read:
2686 case BPF_FUNC_perf_event_output:
2687 case BPF_FUNC_perf_event_read_value:
2688 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
2691 case BPF_FUNC_get_stackid:
2692 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
2695 case BPF_FUNC_current_task_under_cgroup:
2696 case BPF_FUNC_skb_under_cgroup:
2697 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
2700 case BPF_FUNC_redirect_map:
2701 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
2702 map->map_type != BPF_MAP_TYPE_CPUMAP &&
2703 map->map_type != BPF_MAP_TYPE_XSKMAP)
2706 case BPF_FUNC_sk_redirect_map:
2707 case BPF_FUNC_msg_redirect_map:
2708 case BPF_FUNC_sock_map_update:
2709 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
2712 case BPF_FUNC_sk_redirect_hash:
2713 case BPF_FUNC_msg_redirect_hash:
2714 case BPF_FUNC_sock_hash_update:
2715 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
2718 case BPF_FUNC_get_local_storage:
2719 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
2720 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
2723 case BPF_FUNC_sk_select_reuseport:
2724 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY)
2727 case BPF_FUNC_map_peek_elem:
2728 case BPF_FUNC_map_pop_elem:
2729 case BPF_FUNC_map_push_elem:
2730 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
2731 map->map_type != BPF_MAP_TYPE_STACK)
2740 verbose(env, "cannot pass map_type %d into func %s#%d\n",
2741 map->map_type, func_id_name(func_id), func_id);
2745 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
2749 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
2751 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
2753 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
2755 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
2757 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
2760 /* We only support one arg being in raw mode at the moment,
2761 * which is sufficient for the helper functions we have
2767 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
2768 enum bpf_arg_type arg_next)
2770 return (arg_type_is_mem_ptr(arg_curr) &&
2771 !arg_type_is_mem_size(arg_next)) ||
2772 (!arg_type_is_mem_ptr(arg_curr) &&
2773 arg_type_is_mem_size(arg_next));
2776 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
2778 /* bpf_xxx(..., buf, len) call will access 'len'
2779 * bytes from memory 'buf'. Both arg types need
2780 * to be paired, so make sure there's no buggy
2781 * helper function specification.
2783 if (arg_type_is_mem_size(fn->arg1_type) ||
2784 arg_type_is_mem_ptr(fn->arg5_type) ||
2785 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
2786 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
2787 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
2788 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
2794 static bool check_refcount_ok(const struct bpf_func_proto *fn, int func_id)
2798 if (arg_type_may_be_refcounted(fn->arg1_type))
2800 if (arg_type_may_be_refcounted(fn->arg2_type))
2802 if (arg_type_may_be_refcounted(fn->arg3_type))
2804 if (arg_type_may_be_refcounted(fn->arg4_type))
2806 if (arg_type_may_be_refcounted(fn->arg5_type))
2809 /* A reference acquiring function cannot acquire
2810 * another refcounted ptr.
2812 if (is_acquire_function(func_id) && count)
2815 /* We only support one arg being unreferenced at the moment,
2816 * which is sufficient for the helper functions we have right now.
2821 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
2823 return check_raw_mode_ok(fn) &&
2824 check_arg_pair_ok(fn) &&
2825 check_refcount_ok(fn, func_id) ? 0 : -EINVAL;
2828 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
2829 * are now invalid, so turn them into unknown SCALAR_VALUE.
2831 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
2832 struct bpf_func_state *state)
2834 struct bpf_reg_state *regs = state->regs, *reg;
2837 for (i = 0; i < MAX_BPF_REG; i++)
2838 if (reg_is_pkt_pointer_any(®s[i]))
2839 mark_reg_unknown(env, regs, i);
2841 bpf_for_each_spilled_reg(i, state, reg) {
2844 if (reg_is_pkt_pointer_any(reg))
2845 __mark_reg_unknown(reg);
2849 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
2851 struct bpf_verifier_state *vstate = env->cur_state;
2854 for (i = 0; i <= vstate->curframe; i++)
2855 __clear_all_pkt_pointers(env, vstate->frame[i]);
2858 static void release_reg_references(struct bpf_verifier_env *env,
2859 struct bpf_func_state *state,
2862 struct bpf_reg_state *regs = state->regs, *reg;
2865 for (i = 0; i < MAX_BPF_REG; i++)
2866 if (regs[i].ref_obj_id == ref_obj_id)
2867 mark_reg_unknown(env, regs, i);
2869 bpf_for_each_spilled_reg(i, state, reg) {
2872 if (reg->ref_obj_id == ref_obj_id)
2873 __mark_reg_unknown(reg);
2877 /* The pointer with the specified id has released its reference to kernel
2878 * resources. Identify all copies of the same pointer and clear the reference.
2880 static int release_reference(struct bpf_verifier_env *env,
2883 struct bpf_verifier_state *vstate = env->cur_state;
2887 err = release_reference_state(cur_func(env), ref_obj_id);
2891 for (i = 0; i <= vstate->curframe; i++)
2892 release_reg_references(env, vstate->frame[i], ref_obj_id);
2897 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
2900 struct bpf_verifier_state *state = env->cur_state;
2901 struct bpf_func_state *caller, *callee;
2902 int i, err, subprog, target_insn;
2904 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
2905 verbose(env, "the call stack of %d frames is too deep\n",
2906 state->curframe + 2);
2910 target_insn = *insn_idx + insn->imm;
2911 subprog = find_subprog(env, target_insn + 1);
2913 verbose(env, "verifier bug. No program starts at insn %d\n",
2918 caller = state->frame[state->curframe];
2919 if (state->frame[state->curframe + 1]) {
2920 verbose(env, "verifier bug. Frame %d already allocated\n",
2921 state->curframe + 1);
2925 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
2928 state->frame[state->curframe + 1] = callee;
2930 /* callee cannot access r0, r6 - r9 for reading and has to write
2931 * into its own stack before reading from it.
2932 * callee can read/write into caller's stack
2934 init_func_state(env, callee,
2935 /* remember the callsite, it will be used by bpf_exit */
2936 *insn_idx /* callsite */,
2937 state->curframe + 1 /* frameno within this callchain */,
2938 subprog /* subprog number within this prog */);
2940 /* Transfer references to the callee */
2941 err = transfer_reference_state(callee, caller);
2945 /* copy r1 - r5 args that callee can access. The copy includes parent
2946 * pointers, which connects us up to the liveness chain
2948 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
2949 callee->regs[i] = caller->regs[i];
2951 /* after the call registers r0 - r5 were scratched */
2952 for (i = 0; i < CALLER_SAVED_REGS; i++) {
2953 mark_reg_not_init(env, caller->regs, caller_saved[i]);
2954 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
2957 /* only increment it after check_reg_arg() finished */
2960 /* and go analyze first insn of the callee */
2961 *insn_idx = target_insn;
2963 if (env->log.level & BPF_LOG_LEVEL) {
2964 verbose(env, "caller:\n");
2965 print_verifier_state(env, caller);
2966 verbose(env, "callee:\n");
2967 print_verifier_state(env, callee);
2972 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
2974 struct bpf_verifier_state *state = env->cur_state;
2975 struct bpf_func_state *caller, *callee;
2976 struct bpf_reg_state *r0;
2979 callee = state->frame[state->curframe];
2980 r0 = &callee->regs[BPF_REG_0];
2981 if (r0->type == PTR_TO_STACK) {
2982 /* technically it's ok to return caller's stack pointer
2983 * (or caller's caller's pointer) back to the caller,
2984 * since these pointers are valid. Only current stack
2985 * pointer will be invalid as soon as function exits,
2986 * but let's be conservative
2988 verbose(env, "cannot return stack pointer to the caller\n");
2993 caller = state->frame[state->curframe];
2994 /* return to the caller whatever r0 had in the callee */
2995 caller->regs[BPF_REG_0] = *r0;
2997 /* Transfer references to the caller */
2998 err = transfer_reference_state(caller, callee);
3002 *insn_idx = callee->callsite + 1;
3003 if (env->log.level & BPF_LOG_LEVEL) {
3004 verbose(env, "returning from callee:\n");
3005 print_verifier_state(env, callee);
3006 verbose(env, "to caller at %d:\n", *insn_idx);
3007 print_verifier_state(env, caller);
3009 /* clear everything in the callee */
3010 free_func_state(callee);
3011 state->frame[state->curframe + 1] = NULL;
3015 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
3017 struct bpf_call_arg_meta *meta)
3019 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
3021 if (ret_type != RET_INTEGER ||
3022 (func_id != BPF_FUNC_get_stack &&
3023 func_id != BPF_FUNC_probe_read_str))
3026 ret_reg->smax_value = meta->msize_smax_value;
3027 ret_reg->umax_value = meta->msize_umax_value;
3028 __reg_deduce_bounds(ret_reg);
3029 __reg_bound_offset(ret_reg);
3033 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
3034 int func_id, int insn_idx)
3036 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
3038 if (func_id != BPF_FUNC_tail_call &&
3039 func_id != BPF_FUNC_map_lookup_elem &&
3040 func_id != BPF_FUNC_map_update_elem &&
3041 func_id != BPF_FUNC_map_delete_elem &&
3042 func_id != BPF_FUNC_map_push_elem &&
3043 func_id != BPF_FUNC_map_pop_elem &&
3044 func_id != BPF_FUNC_map_peek_elem)
3047 if (meta->map_ptr == NULL) {
3048 verbose(env, "kernel subsystem misconfigured verifier\n");
3052 if (!BPF_MAP_PTR(aux->map_state))
3053 bpf_map_ptr_store(aux, meta->map_ptr,
3054 meta->map_ptr->unpriv_array);
3055 else if (BPF_MAP_PTR(aux->map_state) != meta->map_ptr)
3056 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
3057 meta->map_ptr->unpriv_array);
3061 static int check_reference_leak(struct bpf_verifier_env *env)
3063 struct bpf_func_state *state = cur_func(env);
3066 for (i = 0; i < state->acquired_refs; i++) {
3067 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
3068 state->refs[i].id, state->refs[i].insn_idx);
3070 return state->acquired_refs ? -EINVAL : 0;
3073 static int check_helper_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
3075 const struct bpf_func_proto *fn = NULL;
3076 struct bpf_reg_state *regs;
3077 struct bpf_call_arg_meta meta;
3081 /* find function prototype */
3082 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
3083 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
3088 if (env->ops->get_func_proto)
3089 fn = env->ops->get_func_proto(func_id, env->prog);
3091 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
3096 /* eBPF programs must be GPL compatible to use GPL-ed functions */
3097 if (!env->prog->gpl_compatible && fn->gpl_only) {
3098 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
3102 /* With LD_ABS/IND some JITs save/restore skb from r1. */
3103 changes_data = bpf_helper_changes_pkt_data(fn->func);
3104 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
3105 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
3106 func_id_name(func_id), func_id);
3110 memset(&meta, 0, sizeof(meta));
3111 meta.pkt_access = fn->pkt_access;
3113 err = check_func_proto(fn, func_id);
3115 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
3116 func_id_name(func_id), func_id);
3120 meta.func_id = func_id;
3122 err = check_func_arg(env, BPF_REG_1, fn->arg1_type, &meta);
3125 err = check_func_arg(env, BPF_REG_2, fn->arg2_type, &meta);
3128 err = check_func_arg(env, BPF_REG_3, fn->arg3_type, &meta);
3131 err = check_func_arg(env, BPF_REG_4, fn->arg4_type, &meta);
3134 err = check_func_arg(env, BPF_REG_5, fn->arg5_type, &meta);
3138 err = record_func_map(env, &meta, func_id, insn_idx);
3142 /* Mark slots with STACK_MISC in case of raw mode, stack offset
3143 * is inferred from register state.
3145 for (i = 0; i < meta.access_size; i++) {
3146 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
3147 BPF_WRITE, -1, false);
3152 if (func_id == BPF_FUNC_tail_call) {
3153 err = check_reference_leak(env);
3155 verbose(env, "tail_call would lead to reference leak\n");
3158 } else if (is_release_function(func_id)) {
3159 err = release_reference(env, meta.ref_obj_id);
3161 verbose(env, "func %s#%d reference has not been acquired before\n",
3162 func_id_name(func_id), func_id);
3167 regs = cur_regs(env);
3169 /* check that flags argument in get_local_storage(map, flags) is 0,
3170 * this is required because get_local_storage() can't return an error.
3172 if (func_id == BPF_FUNC_get_local_storage &&
3173 !register_is_null(®s[BPF_REG_2])) {
3174 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
3178 /* reset caller saved regs */
3179 for (i = 0; i < CALLER_SAVED_REGS; i++) {
3180 mark_reg_not_init(env, regs, caller_saved[i]);
3181 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
3184 /* update return register (already marked as written above) */
3185 if (fn->ret_type == RET_INTEGER) {
3186 /* sets type to SCALAR_VALUE */
3187 mark_reg_unknown(env, regs, BPF_REG_0);
3188 } else if (fn->ret_type == RET_VOID) {
3189 regs[BPF_REG_0].type = NOT_INIT;
3190 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL ||
3191 fn->ret_type == RET_PTR_TO_MAP_VALUE) {
3192 /* There is no offset yet applied, variable or fixed */
3193 mark_reg_known_zero(env, regs, BPF_REG_0);
3194 /* remember map_ptr, so that check_map_access()
3195 * can check 'value_size' boundary of memory access
3196 * to map element returned from bpf_map_lookup_elem()
3198 if (meta.map_ptr == NULL) {
3200 "kernel subsystem misconfigured verifier\n");
3203 regs[BPF_REG_0].map_ptr = meta.map_ptr;
3204 if (fn->ret_type == RET_PTR_TO_MAP_VALUE) {
3205 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE;
3206 if (map_value_has_spin_lock(meta.map_ptr))
3207 regs[BPF_REG_0].id = ++env->id_gen;
3209 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
3210 regs[BPF_REG_0].id = ++env->id_gen;
3212 } else if (fn->ret_type == RET_PTR_TO_SOCKET_OR_NULL) {
3213 mark_reg_known_zero(env, regs, BPF_REG_0);
3214 regs[BPF_REG_0].type = PTR_TO_SOCKET_OR_NULL;
3215 regs[BPF_REG_0].id = ++env->id_gen;
3216 } else if (fn->ret_type == RET_PTR_TO_SOCK_COMMON_OR_NULL) {
3217 mark_reg_known_zero(env, regs, BPF_REG_0);
3218 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON_OR_NULL;
3219 regs[BPF_REG_0].id = ++env->id_gen;
3220 } else if (fn->ret_type == RET_PTR_TO_TCP_SOCK_OR_NULL) {
3221 mark_reg_known_zero(env, regs, BPF_REG_0);
3222 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK_OR_NULL;
3223 regs[BPF_REG_0].id = ++env->id_gen;
3225 verbose(env, "unknown return type %d of func %s#%d\n",
3226 fn->ret_type, func_id_name(func_id), func_id);
3230 if (is_ptr_cast_function(func_id)) {
3231 /* For release_reference() */
3232 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
3233 } else if (is_acquire_function(func_id)) {
3234 int id = acquire_reference_state(env, insn_idx);
3238 /* For mark_ptr_or_null_reg() */
3239 regs[BPF_REG_0].id = id;
3240 /* For release_reference() */
3241 regs[BPF_REG_0].ref_obj_id = id;
3244 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
3246 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
3250 if (func_id == BPF_FUNC_get_stack && !env->prog->has_callchain_buf) {
3251 const char *err_str;
3253 #ifdef CONFIG_PERF_EVENTS
3254 err = get_callchain_buffers(sysctl_perf_event_max_stack);
3255 err_str = "cannot get callchain buffer for func %s#%d\n";
3258 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
3261 verbose(env, err_str, func_id_name(func_id), func_id);
3265 env->prog->has_callchain_buf = true;
3269 clear_all_pkt_pointers(env);
3273 static bool signed_add_overflows(s64 a, s64 b)
3275 /* Do the add in u64, where overflow is well-defined */
3276 s64 res = (s64)((u64)a + (u64)b);
3283 static bool signed_sub_overflows(s64 a, s64 b)
3285 /* Do the sub in u64, where overflow is well-defined */
3286 s64 res = (s64)((u64)a - (u64)b);
3293 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
3294 const struct bpf_reg_state *reg,
3295 enum bpf_reg_type type)
3297 bool known = tnum_is_const(reg->var_off);
3298 s64 val = reg->var_off.value;
3299 s64 smin = reg->smin_value;
3301 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
3302 verbose(env, "math between %s pointer and %lld is not allowed\n",
3303 reg_type_str[type], val);
3307 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
3308 verbose(env, "%s pointer offset %d is not allowed\n",
3309 reg_type_str[type], reg->off);
3313 if (smin == S64_MIN) {
3314 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
3315 reg_type_str[type]);
3319 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
3320 verbose(env, "value %lld makes %s pointer be out of bounds\n",
3321 smin, reg_type_str[type]);
3328 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
3330 return &env->insn_aux_data[env->insn_idx];
3333 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
3334 u32 *ptr_limit, u8 opcode, bool off_is_neg)
3336 bool mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
3337 (opcode == BPF_SUB && !off_is_neg);
3340 switch (ptr_reg->type) {
3342 off = ptr_reg->off + ptr_reg->var_off.value;
3344 *ptr_limit = MAX_BPF_STACK + off;
3348 case PTR_TO_MAP_VALUE:
3350 *ptr_limit = ptr_reg->umax_value + ptr_reg->off;
3352 off = ptr_reg->smin_value + ptr_reg->off;
3353 *ptr_limit = ptr_reg->map_ptr->value_size - off;
3361 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
3362 const struct bpf_insn *insn)
3364 return env->allow_ptr_leaks || BPF_SRC(insn->code) == BPF_K;
3367 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
3368 u32 alu_state, u32 alu_limit)
3370 /* If we arrived here from different branches with different
3371 * state or limits to sanitize, then this won't work.
3373 if (aux->alu_state &&
3374 (aux->alu_state != alu_state ||
3375 aux->alu_limit != alu_limit))
3378 /* Corresponding fixup done in fixup_bpf_calls(). */
3379 aux->alu_state = alu_state;
3380 aux->alu_limit = alu_limit;
3384 static int sanitize_val_alu(struct bpf_verifier_env *env,
3385 struct bpf_insn *insn)
3387 struct bpf_insn_aux_data *aux = cur_aux(env);
3389 if (can_skip_alu_sanitation(env, insn))
3392 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
3395 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
3396 struct bpf_insn *insn,
3397 const struct bpf_reg_state *ptr_reg,
3398 struct bpf_reg_state *dst_reg,
3401 struct bpf_verifier_state *vstate = env->cur_state;
3402 struct bpf_insn_aux_data *aux = cur_aux(env);
3403 bool ptr_is_dst_reg = ptr_reg == dst_reg;
3404 u8 opcode = BPF_OP(insn->code);
3405 u32 alu_state, alu_limit;
3406 struct bpf_reg_state tmp;
3409 if (can_skip_alu_sanitation(env, insn))
3412 /* We already marked aux for masking from non-speculative
3413 * paths, thus we got here in the first place. We only care
3414 * to explore bad access from here.
3416 if (vstate->speculative)
3419 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
3420 alu_state |= ptr_is_dst_reg ?
3421 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
3423 if (retrieve_ptr_limit(ptr_reg, &alu_limit, opcode, off_is_neg))
3425 if (update_alu_sanitation_state(aux, alu_state, alu_limit))
3428 /* Simulate and find potential out-of-bounds access under
3429 * speculative execution from truncation as a result of
3430 * masking when off was not within expected range. If off
3431 * sits in dst, then we temporarily need to move ptr there
3432 * to simulate dst (== 0) +/-= ptr. Needed, for example,
3433 * for cases where we use K-based arithmetic in one direction
3434 * and truncated reg-based in the other in order to explore
3437 if (!ptr_is_dst_reg) {
3439 *dst_reg = *ptr_reg;
3441 ret = push_stack(env, env->insn_idx + 1, env->insn_idx, true);
3442 if (!ptr_is_dst_reg && ret)
3444 return !ret ? -EFAULT : 0;
3447 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
3448 * Caller should also handle BPF_MOV case separately.
3449 * If we return -EACCES, caller may want to try again treating pointer as a
3450 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
3452 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
3453 struct bpf_insn *insn,
3454 const struct bpf_reg_state *ptr_reg,
3455 const struct bpf_reg_state *off_reg)
3457 struct bpf_verifier_state *vstate = env->cur_state;
3458 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3459 struct bpf_reg_state *regs = state->regs, *dst_reg;
3460 bool known = tnum_is_const(off_reg->var_off);
3461 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
3462 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
3463 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
3464 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
3465 u32 dst = insn->dst_reg, src = insn->src_reg;
3466 u8 opcode = BPF_OP(insn->code);
3469 dst_reg = ®s[dst];
3471 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
3472 smin_val > smax_val || umin_val > umax_val) {
3473 /* Taint dst register if offset had invalid bounds derived from
3474 * e.g. dead branches.
3476 __mark_reg_unknown(dst_reg);
3480 if (BPF_CLASS(insn->code) != BPF_ALU64) {
3481 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
3483 "R%d 32-bit pointer arithmetic prohibited\n",
3488 switch (ptr_reg->type) {
3489 case PTR_TO_MAP_VALUE_OR_NULL:
3490 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
3491 dst, reg_type_str[ptr_reg->type]);
3493 case CONST_PTR_TO_MAP:
3494 case PTR_TO_PACKET_END:
3496 case PTR_TO_SOCKET_OR_NULL:
3497 case PTR_TO_SOCK_COMMON:
3498 case PTR_TO_SOCK_COMMON_OR_NULL:
3499 case PTR_TO_TCP_SOCK:
3500 case PTR_TO_TCP_SOCK_OR_NULL:
3501 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
3502 dst, reg_type_str[ptr_reg->type]);
3504 case PTR_TO_MAP_VALUE:
3505 if (!env->allow_ptr_leaks && !known && (smin_val < 0) != (smax_val < 0)) {
3506 verbose(env, "R%d has unknown scalar with mixed signed bounds, pointer arithmetic with it prohibited for !root\n",
3507 off_reg == dst_reg ? dst : src);
3515 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
3516 * The id may be overwritten later if we create a new variable offset.
3518 dst_reg->type = ptr_reg->type;
3519 dst_reg->id = ptr_reg->id;
3521 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
3522 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
3527 ret = sanitize_ptr_alu(env, insn, ptr_reg, dst_reg, smin_val < 0);
3529 verbose(env, "R%d tried to add from different maps or paths\n", dst);
3532 /* We can take a fixed offset as long as it doesn't overflow
3533 * the s32 'off' field
3535 if (known && (ptr_reg->off + smin_val ==
3536 (s64)(s32)(ptr_reg->off + smin_val))) {
3537 /* pointer += K. Accumulate it into fixed offset */
3538 dst_reg->smin_value = smin_ptr;
3539 dst_reg->smax_value = smax_ptr;
3540 dst_reg->umin_value = umin_ptr;
3541 dst_reg->umax_value = umax_ptr;
3542 dst_reg->var_off = ptr_reg->var_off;
3543 dst_reg->off = ptr_reg->off + smin_val;
3544 dst_reg->raw = ptr_reg->raw;
3547 /* A new variable offset is created. Note that off_reg->off
3548 * == 0, since it's a scalar.
3549 * dst_reg gets the pointer type and since some positive
3550 * integer value was added to the pointer, give it a new 'id'
3551 * if it's a PTR_TO_PACKET.
3552 * this creates a new 'base' pointer, off_reg (variable) gets
3553 * added into the variable offset, and we copy the fixed offset
3556 if (signed_add_overflows(smin_ptr, smin_val) ||
3557 signed_add_overflows(smax_ptr, smax_val)) {
3558 dst_reg->smin_value = S64_MIN;
3559 dst_reg->smax_value = S64_MAX;
3561 dst_reg->smin_value = smin_ptr + smin_val;
3562 dst_reg->smax_value = smax_ptr + smax_val;
3564 if (umin_ptr + umin_val < umin_ptr ||
3565 umax_ptr + umax_val < umax_ptr) {
3566 dst_reg->umin_value = 0;
3567 dst_reg->umax_value = U64_MAX;
3569 dst_reg->umin_value = umin_ptr + umin_val;
3570 dst_reg->umax_value = umax_ptr + umax_val;
3572 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
3573 dst_reg->off = ptr_reg->off;
3574 dst_reg->raw = ptr_reg->raw;
3575 if (reg_is_pkt_pointer(ptr_reg)) {
3576 dst_reg->id = ++env->id_gen;
3577 /* something was added to pkt_ptr, set range to zero */
3582 ret = sanitize_ptr_alu(env, insn, ptr_reg, dst_reg, smin_val < 0);
3584 verbose(env, "R%d tried to sub from different maps or paths\n", dst);
3587 if (dst_reg == off_reg) {
3588 /* scalar -= pointer. Creates an unknown scalar */
3589 verbose(env, "R%d tried to subtract pointer from scalar\n",
3593 /* We don't allow subtraction from FP, because (according to
3594 * test_verifier.c test "invalid fp arithmetic", JITs might not
3595 * be able to deal with it.
3597 if (ptr_reg->type == PTR_TO_STACK) {
3598 verbose(env, "R%d subtraction from stack pointer prohibited\n",
3602 if (known && (ptr_reg->off - smin_val ==
3603 (s64)(s32)(ptr_reg->off - smin_val))) {
3604 /* pointer -= K. Subtract it from fixed offset */
3605 dst_reg->smin_value = smin_ptr;
3606 dst_reg->smax_value = smax_ptr;
3607 dst_reg->umin_value = umin_ptr;
3608 dst_reg->umax_value = umax_ptr;
3609 dst_reg->var_off = ptr_reg->var_off;
3610 dst_reg->id = ptr_reg->id;
3611 dst_reg->off = ptr_reg->off - smin_val;
3612 dst_reg->raw = ptr_reg->raw;
3615 /* A new variable offset is created. If the subtrahend is known
3616 * nonnegative, then any reg->range we had before is still good.
3618 if (signed_sub_overflows(smin_ptr, smax_val) ||
3619 signed_sub_overflows(smax_ptr, smin_val)) {
3620 /* Overflow possible, we know nothing */
3621 dst_reg->smin_value = S64_MIN;
3622 dst_reg->smax_value = S64_MAX;
3624 dst_reg->smin_value = smin_ptr - smax_val;
3625 dst_reg->smax_value = smax_ptr - smin_val;
3627 if (umin_ptr < umax_val) {
3628 /* Overflow possible, we know nothing */
3629 dst_reg->umin_value = 0;
3630 dst_reg->umax_value = U64_MAX;
3632 /* Cannot overflow (as long as bounds are consistent) */
3633 dst_reg->umin_value = umin_ptr - umax_val;
3634 dst_reg->umax_value = umax_ptr - umin_val;
3636 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
3637 dst_reg->off = ptr_reg->off;
3638 dst_reg->raw = ptr_reg->raw;
3639 if (reg_is_pkt_pointer(ptr_reg)) {
3640 dst_reg->id = ++env->id_gen;
3641 /* something was added to pkt_ptr, set range to zero */
3649 /* bitwise ops on pointers are troublesome, prohibit. */
3650 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
3651 dst, bpf_alu_string[opcode >> 4]);
3654 /* other operators (e.g. MUL,LSH) produce non-pointer results */
3655 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
3656 dst, bpf_alu_string[opcode >> 4]);
3660 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
3663 __update_reg_bounds(dst_reg);
3664 __reg_deduce_bounds(dst_reg);
3665 __reg_bound_offset(dst_reg);
3667 /* For unprivileged we require that resulting offset must be in bounds
3668 * in order to be able to sanitize access later on.
3670 if (!env->allow_ptr_leaks) {
3671 if (dst_reg->type == PTR_TO_MAP_VALUE &&
3672 check_map_access(env, dst, dst_reg->off, 1, false)) {
3673 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
3674 "prohibited for !root\n", dst);
3676 } else if (dst_reg->type == PTR_TO_STACK &&
3677 check_stack_access(env, dst_reg, dst_reg->off +
3678 dst_reg->var_off.value, 1)) {
3679 verbose(env, "R%d stack pointer arithmetic goes out of range, "
3680 "prohibited for !root\n", dst);
3688 /* WARNING: This function does calculations on 64-bit values, but the actual
3689 * execution may occur on 32-bit values. Therefore, things like bitshifts
3690 * need extra checks in the 32-bit case.
3692 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
3693 struct bpf_insn *insn,
3694 struct bpf_reg_state *dst_reg,
3695 struct bpf_reg_state src_reg)
3697 struct bpf_reg_state *regs = cur_regs(env);
3698 u8 opcode = BPF_OP(insn->code);
3699 bool src_known, dst_known;
3700 s64 smin_val, smax_val;
3701 u64 umin_val, umax_val;
3702 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
3703 u32 dst = insn->dst_reg;
3706 if (insn_bitness == 32) {
3707 /* Relevant for 32-bit RSH: Information can propagate towards
3708 * LSB, so it isn't sufficient to only truncate the output to
3711 coerce_reg_to_size(dst_reg, 4);
3712 coerce_reg_to_size(&src_reg, 4);
3715 smin_val = src_reg.smin_value;
3716 smax_val = src_reg.smax_value;
3717 umin_val = src_reg.umin_value;
3718 umax_val = src_reg.umax_value;
3719 src_known = tnum_is_const(src_reg.var_off);
3720 dst_known = tnum_is_const(dst_reg->var_off);
3722 if ((src_known && (smin_val != smax_val || umin_val != umax_val)) ||
3723 smin_val > smax_val || umin_val > umax_val) {
3724 /* Taint dst register if offset had invalid bounds derived from
3725 * e.g. dead branches.
3727 __mark_reg_unknown(dst_reg);
3732 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
3733 __mark_reg_unknown(dst_reg);
3739 ret = sanitize_val_alu(env, insn);
3741 verbose(env, "R%d tried to add from different pointers or scalars\n", dst);
3744 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
3745 signed_add_overflows(dst_reg->smax_value, smax_val)) {
3746 dst_reg->smin_value = S64_MIN;
3747 dst_reg->smax_value = S64_MAX;
3749 dst_reg->smin_value += smin_val;
3750 dst_reg->smax_value += smax_val;
3752 if (dst_reg->umin_value + umin_val < umin_val ||
3753 dst_reg->umax_value + umax_val < umax_val) {
3754 dst_reg->umin_value = 0;
3755 dst_reg->umax_value = U64_MAX;
3757 dst_reg->umin_value += umin_val;
3758 dst_reg->umax_value += umax_val;
3760 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
3763 ret = sanitize_val_alu(env, insn);
3765 verbose(env, "R%d tried to sub from different pointers or scalars\n", dst);
3768 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
3769 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
3770 /* Overflow possible, we know nothing */
3771 dst_reg->smin_value = S64_MIN;
3772 dst_reg->smax_value = S64_MAX;
3774 dst_reg->smin_value -= smax_val;
3775 dst_reg->smax_value -= smin_val;
3777 if (dst_reg->umin_value < umax_val) {
3778 /* Overflow possible, we know nothing */
3779 dst_reg->umin_value = 0;
3780 dst_reg->umax_value = U64_MAX;
3782 /* Cannot overflow (as long as bounds are consistent) */
3783 dst_reg->umin_value -= umax_val;
3784 dst_reg->umax_value -= umin_val;
3786 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
3789 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
3790 if (smin_val < 0 || dst_reg->smin_value < 0) {
3791 /* Ain't nobody got time to multiply that sign */
3792 __mark_reg_unbounded(dst_reg);
3793 __update_reg_bounds(dst_reg);
3796 /* Both values are positive, so we can work with unsigned and
3797 * copy the result to signed (unless it exceeds S64_MAX).
3799 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
3800 /* Potential overflow, we know nothing */
3801 __mark_reg_unbounded(dst_reg);
3802 /* (except what we can learn from the var_off) */
3803 __update_reg_bounds(dst_reg);
3806 dst_reg->umin_value *= umin_val;
3807 dst_reg->umax_value *= umax_val;
3808 if (dst_reg->umax_value > S64_MAX) {
3809 /* Overflow possible, we know nothing */
3810 dst_reg->smin_value = S64_MIN;
3811 dst_reg->smax_value = S64_MAX;
3813 dst_reg->smin_value = dst_reg->umin_value;
3814 dst_reg->smax_value = dst_reg->umax_value;
3818 if (src_known && dst_known) {
3819 __mark_reg_known(dst_reg, dst_reg->var_off.value &
3820 src_reg.var_off.value);
3823 /* We get our minimum from the var_off, since that's inherently
3824 * bitwise. Our maximum is the minimum of the operands' maxima.
3826 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
3827 dst_reg->umin_value = dst_reg->var_off.value;
3828 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
3829 if (dst_reg->smin_value < 0 || smin_val < 0) {
3830 /* Lose signed bounds when ANDing negative numbers,
3831 * ain't nobody got time for that.
3833 dst_reg->smin_value = S64_MIN;
3834 dst_reg->smax_value = S64_MAX;
3836 /* ANDing two positives gives a positive, so safe to
3837 * cast result into s64.
3839 dst_reg->smin_value = dst_reg->umin_value;
3840 dst_reg->smax_value = dst_reg->umax_value;
3842 /* We may learn something more from the var_off */
3843 __update_reg_bounds(dst_reg);
3846 if (src_known && dst_known) {
3847 __mark_reg_known(dst_reg, dst_reg->var_off.value |
3848 src_reg.var_off.value);
3851 /* We get our maximum from the var_off, and our minimum is the
3852 * maximum of the operands' minima
3854 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
3855 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
3856 dst_reg->umax_value = dst_reg->var_off.value |
3857 dst_reg->var_off.mask;
3858 if (dst_reg->smin_value < 0 || smin_val < 0) {
3859 /* Lose signed bounds when ORing negative numbers,
3860 * ain't nobody got time for that.
3862 dst_reg->smin_value = S64_MIN;
3863 dst_reg->smax_value = S64_MAX;
3865 /* ORing two positives gives a positive, so safe to
3866 * cast result into s64.
3868 dst_reg->smin_value = dst_reg->umin_value;
3869 dst_reg->smax_value = dst_reg->umax_value;
3871 /* We may learn something more from the var_off */
3872 __update_reg_bounds(dst_reg);
3875 if (umax_val >= insn_bitness) {
3876 /* Shifts greater than 31 or 63 are undefined.
3877 * This includes shifts by a negative number.
3879 mark_reg_unknown(env, regs, insn->dst_reg);
3882 /* We lose all sign bit information (except what we can pick
3885 dst_reg->smin_value = S64_MIN;
3886 dst_reg->smax_value = S64_MAX;
3887 /* If we might shift our top bit out, then we know nothing */
3888 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
3889 dst_reg->umin_value = 0;
3890 dst_reg->umax_value = U64_MAX;
3892 dst_reg->umin_value <<= umin_val;
3893 dst_reg->umax_value <<= umax_val;
3895 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
3896 /* We may learn something more from the var_off */
3897 __update_reg_bounds(dst_reg);
3900 if (umax_val >= insn_bitness) {
3901 /* Shifts greater than 31 or 63 are undefined.
3902 * This includes shifts by a negative number.
3904 mark_reg_unknown(env, regs, insn->dst_reg);
3907 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
3908 * be negative, then either:
3909 * 1) src_reg might be zero, so the sign bit of the result is
3910 * unknown, so we lose our signed bounds
3911 * 2) it's known negative, thus the unsigned bounds capture the
3913 * 3) the signed bounds cross zero, so they tell us nothing
3915 * If the value in dst_reg is known nonnegative, then again the
3916 * unsigned bounts capture the signed bounds.
3917 * Thus, in all cases it suffices to blow away our signed bounds
3918 * and rely on inferring new ones from the unsigned bounds and
3919 * var_off of the result.
3921 dst_reg->smin_value = S64_MIN;
3922 dst_reg->smax_value = S64_MAX;
3923 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
3924 dst_reg->umin_value >>= umax_val;
3925 dst_reg->umax_value >>= umin_val;
3926 /* We may learn something more from the var_off */
3927 __update_reg_bounds(dst_reg);
3930 if (umax_val >= insn_bitness) {
3931 /* Shifts greater than 31 or 63 are undefined.
3932 * This includes shifts by a negative number.
3934 mark_reg_unknown(env, regs, insn->dst_reg);
3938 /* Upon reaching here, src_known is true and
3939 * umax_val is equal to umin_val.
3941 dst_reg->smin_value >>= umin_val;
3942 dst_reg->smax_value >>= umin_val;
3943 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val);
3945 /* blow away the dst_reg umin_value/umax_value and rely on
3946 * dst_reg var_off to refine the result.
3948 dst_reg->umin_value = 0;
3949 dst_reg->umax_value = U64_MAX;
3950 __update_reg_bounds(dst_reg);
3953 mark_reg_unknown(env, regs, insn->dst_reg);
3957 if (BPF_CLASS(insn->code) != BPF_ALU64) {
3958 /* 32-bit ALU ops are (32,32)->32 */
3959 coerce_reg_to_size(dst_reg, 4);
3962 __reg_deduce_bounds(dst_reg);
3963 __reg_bound_offset(dst_reg);
3967 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
3970 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
3971 struct bpf_insn *insn)
3973 struct bpf_verifier_state *vstate = env->cur_state;
3974 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3975 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
3976 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
3977 u8 opcode = BPF_OP(insn->code);
3979 dst_reg = ®s[insn->dst_reg];
3981 if (dst_reg->type != SCALAR_VALUE)
3983 if (BPF_SRC(insn->code) == BPF_X) {
3984 src_reg = ®s[insn->src_reg];
3985 if (src_reg->type != SCALAR_VALUE) {
3986 if (dst_reg->type != SCALAR_VALUE) {
3987 /* Combining two pointers by any ALU op yields
3988 * an arbitrary scalar. Disallow all math except
3989 * pointer subtraction
3991 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
3992 mark_reg_unknown(env, regs, insn->dst_reg);
3995 verbose(env, "R%d pointer %s pointer prohibited\n",
3997 bpf_alu_string[opcode >> 4]);
4000 /* scalar += pointer
4001 * This is legal, but we have to reverse our
4002 * src/dest handling in computing the range
4004 return adjust_ptr_min_max_vals(env, insn,
4007 } else if (ptr_reg) {
4008 /* pointer += scalar */
4009 return adjust_ptr_min_max_vals(env, insn,
4013 /* Pretend the src is a reg with a known value, since we only
4014 * need to be able to read from this state.
4016 off_reg.type = SCALAR_VALUE;
4017 __mark_reg_known(&off_reg, insn->imm);
4019 if (ptr_reg) /* pointer += K */
4020 return adjust_ptr_min_max_vals(env, insn,
4024 /* Got here implies adding two SCALAR_VALUEs */
4025 if (WARN_ON_ONCE(ptr_reg)) {
4026 print_verifier_state(env, state);
4027 verbose(env, "verifier internal error: unexpected ptr_reg\n");
4030 if (WARN_ON(!src_reg)) {
4031 print_verifier_state(env, state);
4032 verbose(env, "verifier internal error: no src_reg\n");
4035 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
4038 /* check validity of 32-bit and 64-bit arithmetic operations */
4039 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
4041 struct bpf_reg_state *regs = cur_regs(env);
4042 u8 opcode = BPF_OP(insn->code);
4045 if (opcode == BPF_END || opcode == BPF_NEG) {
4046 if (opcode == BPF_NEG) {
4047 if (BPF_SRC(insn->code) != 0 ||
4048 insn->src_reg != BPF_REG_0 ||
4049 insn->off != 0 || insn->imm != 0) {
4050 verbose(env, "BPF_NEG uses reserved fields\n");
4054 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
4055 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
4056 BPF_CLASS(insn->code) == BPF_ALU64) {
4057 verbose(env, "BPF_END uses reserved fields\n");
4062 /* check src operand */
4063 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4067 if (is_pointer_value(env, insn->dst_reg)) {
4068 verbose(env, "R%d pointer arithmetic prohibited\n",
4073 /* check dest operand */
4074 err = check_reg_arg(env, insn->dst_reg, DST_OP);
4078 } else if (opcode == BPF_MOV) {
4080 if (BPF_SRC(insn->code) == BPF_X) {
4081 if (insn->imm != 0 || insn->off != 0) {
4082 verbose(env, "BPF_MOV uses reserved fields\n");
4086 /* check src operand */
4087 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4091 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
4092 verbose(env, "BPF_MOV uses reserved fields\n");
4097 /* check dest operand, mark as required later */
4098 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
4102 if (BPF_SRC(insn->code) == BPF_X) {
4103 struct bpf_reg_state *src_reg = regs + insn->src_reg;
4104 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
4106 if (BPF_CLASS(insn->code) == BPF_ALU64) {
4108 * copy register state to dest reg
4110 *dst_reg = *src_reg;
4111 dst_reg->live |= REG_LIVE_WRITTEN;
4114 if (is_pointer_value(env, insn->src_reg)) {
4116 "R%d partial copy of pointer\n",
4119 } else if (src_reg->type == SCALAR_VALUE) {
4120 *dst_reg = *src_reg;
4121 dst_reg->live |= REG_LIVE_WRITTEN;
4123 mark_reg_unknown(env, regs,
4126 coerce_reg_to_size(dst_reg, 4);
4130 * remember the value we stored into this reg
4132 /* clear any state __mark_reg_known doesn't set */
4133 mark_reg_unknown(env, regs, insn->dst_reg);
4134 regs[insn->dst_reg].type = SCALAR_VALUE;
4135 if (BPF_CLASS(insn->code) == BPF_ALU64) {
4136 __mark_reg_known(regs + insn->dst_reg,
4139 __mark_reg_known(regs + insn->dst_reg,
4144 } else if (opcode > BPF_END) {
4145 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
4148 } else { /* all other ALU ops: and, sub, xor, add, ... */
4150 if (BPF_SRC(insn->code) == BPF_X) {
4151 if (insn->imm != 0 || insn->off != 0) {
4152 verbose(env, "BPF_ALU uses reserved fields\n");
4155 /* check src1 operand */
4156 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4160 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
4161 verbose(env, "BPF_ALU uses reserved fields\n");
4166 /* check src2 operand */
4167 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4171 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
4172 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
4173 verbose(env, "div by zero\n");
4177 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
4178 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
4179 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
4181 if (insn->imm < 0 || insn->imm >= size) {
4182 verbose(env, "invalid shift %d\n", insn->imm);
4187 /* check dest operand */
4188 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
4192 return adjust_reg_min_max_vals(env, insn);
4198 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
4199 struct bpf_reg_state *dst_reg,
4200 enum bpf_reg_type type,
4201 bool range_right_open)
4203 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4204 struct bpf_reg_state *regs = state->regs, *reg;
4208 if (dst_reg->off < 0 ||
4209 (dst_reg->off == 0 && range_right_open))
4210 /* This doesn't give us any range */
4213 if (dst_reg->umax_value > MAX_PACKET_OFF ||
4214 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
4215 /* Risk of overflow. For instance, ptr + (1<<63) may be less
4216 * than pkt_end, but that's because it's also less than pkt.
4220 new_range = dst_reg->off;
4221 if (range_right_open)
4224 /* Examples for register markings:
4226 * pkt_data in dst register:
4230 * if (r2 > pkt_end) goto <handle exception>
4235 * if (r2 < pkt_end) goto <access okay>
4236 * <handle exception>
4239 * r2 == dst_reg, pkt_end == src_reg
4240 * r2=pkt(id=n,off=8,r=0)
4241 * r3=pkt(id=n,off=0,r=0)
4243 * pkt_data in src register:
4247 * if (pkt_end >= r2) goto <access okay>
4248 * <handle exception>
4252 * if (pkt_end <= r2) goto <handle exception>
4256 * pkt_end == dst_reg, r2 == src_reg
4257 * r2=pkt(id=n,off=8,r=0)
4258 * r3=pkt(id=n,off=0,r=0)
4260 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
4261 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
4262 * and [r3, r3 + 8-1) respectively is safe to access depending on
4266 /* If our ids match, then we must have the same max_value. And we
4267 * don't care about the other reg's fixed offset, since if it's too big
4268 * the range won't allow anything.
4269 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
4271 for (i = 0; i < MAX_BPF_REG; i++)
4272 if (regs[i].type == type && regs[i].id == dst_reg->id)
4273 /* keep the maximum range already checked */
4274 regs[i].range = max(regs[i].range, new_range);
4276 for (j = 0; j <= vstate->curframe; j++) {
4277 state = vstate->frame[j];
4278 bpf_for_each_spilled_reg(i, state, reg) {
4281 if (reg->type == type && reg->id == dst_reg->id)
4282 reg->range = max(reg->range, new_range);
4287 /* compute branch direction of the expression "if (reg opcode val) goto target;"
4289 * 1 - branch will be taken and "goto target" will be executed
4290 * 0 - branch will not be taken and fall-through to next insn
4291 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value range [0,10]
4293 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
4296 struct bpf_reg_state reg_lo;
4299 if (__is_pointer_value(false, reg))
4305 /* For JMP32, only low 32 bits are compared, coerce_reg_to_size
4306 * could truncate high bits and update umin/umax according to
4307 * information of low bits.
4309 coerce_reg_to_size(reg, 4);
4310 /* smin/smax need special handling. For example, after coerce,
4311 * if smin_value is 0x00000000ffffffffLL, the value is -1 when
4312 * used as operand to JMP32. It is a negative number from s32's
4313 * point of view, while it is a positive number when seen as
4314 * s64. The smin/smax are kept as s64, therefore, when used with
4315 * JMP32, they need to be transformed into s32, then sign
4316 * extended back to s64.
4318 * Also, smin/smax were copied from umin/umax. If umin/umax has
4319 * different sign bit, then min/max relationship doesn't
4320 * maintain after casting into s32, for this case, set smin/smax
4323 if ((reg->umax_value ^ reg->umin_value) &
4325 reg->smin_value = S32_MIN;
4326 reg->smax_value = S32_MAX;
4328 reg->smin_value = (s64)(s32)reg->smin_value;
4329 reg->smax_value = (s64)(s32)reg->smax_value;
4332 sval = (s64)(s32)val;
4339 if (tnum_is_const(reg->var_off))
4340 return !!tnum_equals_const(reg->var_off, val);
4343 if (tnum_is_const(reg->var_off))
4344 return !tnum_equals_const(reg->var_off, val);
4347 if ((~reg->var_off.mask & reg->var_off.value) & val)
4349 if (!((reg->var_off.mask | reg->var_off.value) & val))
4353 if (reg->umin_value > val)
4355 else if (reg->umax_value <= val)
4359 if (reg->smin_value > sval)
4361 else if (reg->smax_value < sval)
4365 if (reg->umax_value < val)
4367 else if (reg->umin_value >= val)
4371 if (reg->smax_value < sval)
4373 else if (reg->smin_value >= sval)
4377 if (reg->umin_value >= val)
4379 else if (reg->umax_value < val)
4383 if (reg->smin_value >= sval)
4385 else if (reg->smax_value < sval)
4389 if (reg->umax_value <= val)
4391 else if (reg->umin_value > val)
4395 if (reg->smax_value <= sval)
4397 else if (reg->smin_value > sval)
4405 /* Generate min value of the high 32-bit from TNUM info. */
4406 static u64 gen_hi_min(struct tnum var)
4408 return var.value & ~0xffffffffULL;
4411 /* Generate max value of the high 32-bit from TNUM info. */
4412 static u64 gen_hi_max(struct tnum var)
4414 return (var.value | var.mask) & ~0xffffffffULL;
4417 /* Return true if VAL is compared with a s64 sign extended from s32, and they
4418 * are with the same signedness.
4420 static bool cmp_val_with_extended_s64(s64 sval, struct bpf_reg_state *reg)
4422 return ((s32)sval >= 0 &&
4423 reg->smin_value >= 0 && reg->smax_value <= S32_MAX) ||
4425 reg->smax_value <= 0 && reg->smin_value >= S32_MIN);
4428 /* Adjusts the register min/max values in the case that the dst_reg is the
4429 * variable register that we are working on, and src_reg is a constant or we're
4430 * simply doing a BPF_K check.
4431 * In JEQ/JNE cases we also adjust the var_off values.
4433 static void reg_set_min_max(struct bpf_reg_state *true_reg,
4434 struct bpf_reg_state *false_reg, u64 val,
4435 u8 opcode, bool is_jmp32)
4439 /* If the dst_reg is a pointer, we can't learn anything about its
4440 * variable offset from the compare (unless src_reg were a pointer into
4441 * the same object, but we don't bother with that.
4442 * Since false_reg and true_reg have the same type by construction, we
4443 * only need to check one of them for pointerness.
4445 if (__is_pointer_value(false, false_reg))
4448 val = is_jmp32 ? (u32)val : val;
4449 sval = is_jmp32 ? (s64)(s32)val : (s64)val;
4455 struct bpf_reg_state *reg =
4456 opcode == BPF_JEQ ? true_reg : false_reg;
4458 /* For BPF_JEQ, if this is false we know nothing Jon Snow, but
4459 * if it is true we know the value for sure. Likewise for
4463 u64 old_v = reg->var_off.value;
4464 u64 hi_mask = ~0xffffffffULL;
4466 reg->var_off.value = (old_v & hi_mask) | val;
4467 reg->var_off.mask &= hi_mask;
4469 __mark_reg_known(reg, val);
4474 false_reg->var_off = tnum_and(false_reg->var_off,
4476 if (is_power_of_2(val))
4477 true_reg->var_off = tnum_or(true_reg->var_off,
4483 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
4484 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
4487 false_umax += gen_hi_max(false_reg->var_off);
4488 true_umin += gen_hi_min(true_reg->var_off);
4490 false_reg->umax_value = min(false_reg->umax_value, false_umax);
4491 true_reg->umin_value = max(true_reg->umin_value, true_umin);
4497 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
4498 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
4500 /* If the full s64 was not sign-extended from s32 then don't
4501 * deduct further info.
4503 if (is_jmp32 && !cmp_val_with_extended_s64(sval, false_reg))
4505 false_reg->smax_value = min(false_reg->smax_value, false_smax);
4506 true_reg->smin_value = max(true_reg->smin_value, true_smin);
4512 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
4513 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
4516 false_umin += gen_hi_min(false_reg->var_off);
4517 true_umax += gen_hi_max(true_reg->var_off);
4519 false_reg->umin_value = max(false_reg->umin_value, false_umin);
4520 true_reg->umax_value = min(true_reg->umax_value, true_umax);
4526 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
4527 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
4529 if (is_jmp32 && !cmp_val_with_extended_s64(sval, false_reg))
4531 false_reg->smin_value = max(false_reg->smin_value, false_smin);
4532 true_reg->smax_value = min(true_reg->smax_value, true_smax);
4539 __reg_deduce_bounds(false_reg);
4540 __reg_deduce_bounds(true_reg);
4541 /* We might have learned some bits from the bounds. */
4542 __reg_bound_offset(false_reg);
4543 __reg_bound_offset(true_reg);
4544 /* Intersecting with the old var_off might have improved our bounds
4545 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
4546 * then new var_off is (0; 0x7f...fc) which improves our umax.
4548 __update_reg_bounds(false_reg);
4549 __update_reg_bounds(true_reg);
4552 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
4555 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
4556 struct bpf_reg_state *false_reg, u64 val,
4557 u8 opcode, bool is_jmp32)
4561 if (__is_pointer_value(false, false_reg))
4564 val = is_jmp32 ? (u32)val : val;
4565 sval = is_jmp32 ? (s64)(s32)val : (s64)val;
4571 struct bpf_reg_state *reg =
4572 opcode == BPF_JEQ ? true_reg : false_reg;
4575 u64 old_v = reg->var_off.value;
4576 u64 hi_mask = ~0xffffffffULL;
4578 reg->var_off.value = (old_v & hi_mask) | val;
4579 reg->var_off.mask &= hi_mask;
4581 __mark_reg_known(reg, val);
4586 false_reg->var_off = tnum_and(false_reg->var_off,
4588 if (is_power_of_2(val))
4589 true_reg->var_off = tnum_or(true_reg->var_off,
4595 u64 false_umin = opcode == BPF_JGT ? val : val + 1;
4596 u64 true_umax = opcode == BPF_JGT ? val - 1 : val;
4599 false_umin += gen_hi_min(false_reg->var_off);
4600 true_umax += gen_hi_max(true_reg->var_off);
4602 false_reg->umin_value = max(false_reg->umin_value, false_umin);
4603 true_reg->umax_value = min(true_reg->umax_value, true_umax);
4609 s64 false_smin = opcode == BPF_JSGT ? sval : sval + 1;
4610 s64 true_smax = opcode == BPF_JSGT ? sval - 1 : sval;
4612 if (is_jmp32 && !cmp_val_with_extended_s64(sval, false_reg))
4614 false_reg->smin_value = max(false_reg->smin_value, false_smin);
4615 true_reg->smax_value = min(true_reg->smax_value, true_smax);
4621 u64 false_umax = opcode == BPF_JLT ? val : val - 1;
4622 u64 true_umin = opcode == BPF_JLT ? val + 1 : val;
4625 false_umax += gen_hi_max(false_reg->var_off);
4626 true_umin += gen_hi_min(true_reg->var_off);
4628 false_reg->umax_value = min(false_reg->umax_value, false_umax);
4629 true_reg->umin_value = max(true_reg->umin_value, true_umin);
4635 s64 false_smax = opcode == BPF_JSLT ? sval : sval - 1;
4636 s64 true_smin = opcode == BPF_JSLT ? sval + 1 : sval;
4638 if (is_jmp32 && !cmp_val_with_extended_s64(sval, false_reg))
4640 false_reg->smax_value = min(false_reg->smax_value, false_smax);
4641 true_reg->smin_value = max(true_reg->smin_value, true_smin);
4648 __reg_deduce_bounds(false_reg);
4649 __reg_deduce_bounds(true_reg);
4650 /* We might have learned some bits from the bounds. */
4651 __reg_bound_offset(false_reg);
4652 __reg_bound_offset(true_reg);
4653 /* Intersecting with the old var_off might have improved our bounds
4654 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
4655 * then new var_off is (0; 0x7f...fc) which improves our umax.
4657 __update_reg_bounds(false_reg);
4658 __update_reg_bounds(true_reg);
4661 /* Regs are known to be equal, so intersect their min/max/var_off */
4662 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
4663 struct bpf_reg_state *dst_reg)
4665 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
4666 dst_reg->umin_value);
4667 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
4668 dst_reg->umax_value);
4669 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
4670 dst_reg->smin_value);
4671 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
4672 dst_reg->smax_value);
4673 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
4675 /* We might have learned new bounds from the var_off. */
4676 __update_reg_bounds(src_reg);
4677 __update_reg_bounds(dst_reg);
4678 /* We might have learned something about the sign bit. */
4679 __reg_deduce_bounds(src_reg);
4680 __reg_deduce_bounds(dst_reg);
4681 /* We might have learned some bits from the bounds. */
4682 __reg_bound_offset(src_reg);
4683 __reg_bound_offset(dst_reg);
4684 /* Intersecting with the old var_off might have improved our bounds
4685 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
4686 * then new var_off is (0; 0x7f...fc) which improves our umax.
4688 __update_reg_bounds(src_reg);
4689 __update_reg_bounds(dst_reg);
4692 static void reg_combine_min_max(struct bpf_reg_state *true_src,
4693 struct bpf_reg_state *true_dst,
4694 struct bpf_reg_state *false_src,
4695 struct bpf_reg_state *false_dst,
4700 __reg_combine_min_max(true_src, true_dst);
4703 __reg_combine_min_max(false_src, false_dst);
4708 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
4709 struct bpf_reg_state *reg, u32 id,
4712 if (reg_type_may_be_null(reg->type) && reg->id == id) {
4713 /* Old offset (both fixed and variable parts) should
4714 * have been known-zero, because we don't allow pointer
4715 * arithmetic on pointers that might be NULL.
4717 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
4718 !tnum_equals_const(reg->var_off, 0) ||
4720 __mark_reg_known_zero(reg);
4724 reg->type = SCALAR_VALUE;
4725 } else if (reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
4726 if (reg->map_ptr->inner_map_meta) {
4727 reg->type = CONST_PTR_TO_MAP;
4728 reg->map_ptr = reg->map_ptr->inner_map_meta;
4730 reg->type = PTR_TO_MAP_VALUE;
4732 } else if (reg->type == PTR_TO_SOCKET_OR_NULL) {
4733 reg->type = PTR_TO_SOCKET;
4734 } else if (reg->type == PTR_TO_SOCK_COMMON_OR_NULL) {
4735 reg->type = PTR_TO_SOCK_COMMON;
4736 } else if (reg->type == PTR_TO_TCP_SOCK_OR_NULL) {
4737 reg->type = PTR_TO_TCP_SOCK;
4740 /* We don't need id and ref_obj_id from this point
4741 * onwards anymore, thus we should better reset it,
4742 * so that state pruning has chances to take effect.
4745 reg->ref_obj_id = 0;
4746 } else if (!reg_may_point_to_spin_lock(reg)) {
4747 /* For not-NULL ptr, reg->ref_obj_id will be reset
4748 * in release_reg_references().
4750 * reg->id is still used by spin_lock ptr. Other
4751 * than spin_lock ptr type, reg->id can be reset.
4758 /* The logic is similar to find_good_pkt_pointers(), both could eventually
4759 * be folded together at some point.
4761 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
4764 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4765 struct bpf_reg_state *reg, *regs = state->regs;
4766 u32 ref_obj_id = regs[regno].ref_obj_id;
4767 u32 id = regs[regno].id;
4770 if (ref_obj_id && ref_obj_id == id && is_null)
4771 /* regs[regno] is in the " == NULL" branch.
4772 * No one could have freed the reference state before
4773 * doing the NULL check.
4775 WARN_ON_ONCE(release_reference_state(state, id));
4777 for (i = 0; i < MAX_BPF_REG; i++)
4778 mark_ptr_or_null_reg(state, ®s[i], id, is_null);
4780 for (j = 0; j <= vstate->curframe; j++) {
4781 state = vstate->frame[j];
4782 bpf_for_each_spilled_reg(i, state, reg) {
4785 mark_ptr_or_null_reg(state, reg, id, is_null);
4790 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
4791 struct bpf_reg_state *dst_reg,
4792 struct bpf_reg_state *src_reg,
4793 struct bpf_verifier_state *this_branch,
4794 struct bpf_verifier_state *other_branch)
4796 if (BPF_SRC(insn->code) != BPF_X)
4799 /* Pointers are always 64-bit. */
4800 if (BPF_CLASS(insn->code) == BPF_JMP32)
4803 switch (BPF_OP(insn->code)) {
4805 if ((dst_reg->type == PTR_TO_PACKET &&
4806 src_reg->type == PTR_TO_PACKET_END) ||
4807 (dst_reg->type == PTR_TO_PACKET_META &&
4808 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
4809 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
4810 find_good_pkt_pointers(this_branch, dst_reg,
4811 dst_reg->type, false);
4812 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
4813 src_reg->type == PTR_TO_PACKET) ||
4814 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
4815 src_reg->type == PTR_TO_PACKET_META)) {
4816 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
4817 find_good_pkt_pointers(other_branch, src_reg,
4818 src_reg->type, true);
4824 if ((dst_reg->type == PTR_TO_PACKET &&
4825 src_reg->type == PTR_TO_PACKET_END) ||
4826 (dst_reg->type == PTR_TO_PACKET_META &&
4827 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
4828 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
4829 find_good_pkt_pointers(other_branch, dst_reg,
4830 dst_reg->type, true);
4831 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
4832 src_reg->type == PTR_TO_PACKET) ||
4833 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
4834 src_reg->type == PTR_TO_PACKET_META)) {
4835 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
4836 find_good_pkt_pointers(this_branch, src_reg,
4837 src_reg->type, false);
4843 if ((dst_reg->type == PTR_TO_PACKET &&
4844 src_reg->type == PTR_TO_PACKET_END) ||
4845 (dst_reg->type == PTR_TO_PACKET_META &&
4846 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
4847 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
4848 find_good_pkt_pointers(this_branch, dst_reg,
4849 dst_reg->type, true);
4850 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
4851 src_reg->type == PTR_TO_PACKET) ||
4852 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
4853 src_reg->type == PTR_TO_PACKET_META)) {
4854 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
4855 find_good_pkt_pointers(other_branch, src_reg,
4856 src_reg->type, false);
4862 if ((dst_reg->type == PTR_TO_PACKET &&
4863 src_reg->type == PTR_TO_PACKET_END) ||
4864 (dst_reg->type == PTR_TO_PACKET_META &&
4865 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
4866 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
4867 find_good_pkt_pointers(other_branch, dst_reg,
4868 dst_reg->type, false);
4869 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
4870 src_reg->type == PTR_TO_PACKET) ||
4871 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
4872 src_reg->type == PTR_TO_PACKET_META)) {
4873 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
4874 find_good_pkt_pointers(this_branch, src_reg,
4875 src_reg->type, true);
4887 static int check_cond_jmp_op(struct bpf_verifier_env *env,
4888 struct bpf_insn *insn, int *insn_idx)
4890 struct bpf_verifier_state *this_branch = env->cur_state;
4891 struct bpf_verifier_state *other_branch;
4892 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
4893 struct bpf_reg_state *dst_reg, *other_branch_regs;
4894 u8 opcode = BPF_OP(insn->code);
4898 /* Only conditional jumps are expected to reach here. */
4899 if (opcode == BPF_JA || opcode > BPF_JSLE) {
4900 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
4904 if (BPF_SRC(insn->code) == BPF_X) {
4905 if (insn->imm != 0) {
4906 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
4910 /* check src1 operand */
4911 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4915 if (is_pointer_value(env, insn->src_reg)) {
4916 verbose(env, "R%d pointer comparison prohibited\n",
4921 if (insn->src_reg != BPF_REG_0) {
4922 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
4927 /* check src2 operand */
4928 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4932 dst_reg = ®s[insn->dst_reg];
4933 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
4935 if (BPF_SRC(insn->code) == BPF_K) {
4936 int pred = is_branch_taken(dst_reg, insn->imm, opcode,
4940 /* only follow the goto, ignore fall-through */
4941 *insn_idx += insn->off;
4943 } else if (pred == 0) {
4944 /* only follow fall-through branch, since
4945 * that's where the program will go
4951 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
4955 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
4957 /* detect if we are comparing against a constant value so we can adjust
4958 * our min/max values for our dst register.
4959 * this is only legit if both are scalars (or pointers to the same
4960 * object, I suppose, but we don't support that right now), because
4961 * otherwise the different base pointers mean the offsets aren't
4964 if (BPF_SRC(insn->code) == BPF_X) {
4965 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
4966 struct bpf_reg_state lo_reg0 = *dst_reg;
4967 struct bpf_reg_state lo_reg1 = *src_reg;
4968 struct bpf_reg_state *src_lo, *dst_lo;
4972 coerce_reg_to_size(dst_lo, 4);
4973 coerce_reg_to_size(src_lo, 4);
4975 if (dst_reg->type == SCALAR_VALUE &&
4976 src_reg->type == SCALAR_VALUE) {
4977 if (tnum_is_const(src_reg->var_off) ||
4978 (is_jmp32 && tnum_is_const(src_lo->var_off)))
4979 reg_set_min_max(&other_branch_regs[insn->dst_reg],
4982 ? src_lo->var_off.value
4983 : src_reg->var_off.value,
4985 else if (tnum_is_const(dst_reg->var_off) ||
4986 (is_jmp32 && tnum_is_const(dst_lo->var_off)))
4987 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
4990 ? dst_lo->var_off.value
4991 : dst_reg->var_off.value,
4993 else if (!is_jmp32 &&
4994 (opcode == BPF_JEQ || opcode == BPF_JNE))
4995 /* Comparing for equality, we can combine knowledge */
4996 reg_combine_min_max(&other_branch_regs[insn->src_reg],
4997 &other_branch_regs[insn->dst_reg],
4998 src_reg, dst_reg, opcode);
5000 } else if (dst_reg->type == SCALAR_VALUE) {
5001 reg_set_min_max(&other_branch_regs[insn->dst_reg],
5002 dst_reg, insn->imm, opcode, is_jmp32);
5005 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
5006 * NOTE: these optimizations below are related with pointer comparison
5007 * which will never be JMP32.
5009 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
5010 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
5011 reg_type_may_be_null(dst_reg->type)) {
5012 /* Mark all identical registers in each branch as either
5013 * safe or unknown depending R == 0 or R != 0 conditional.
5015 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
5017 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
5019 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
5020 this_branch, other_branch) &&
5021 is_pointer_value(env, insn->dst_reg)) {
5022 verbose(env, "R%d pointer comparison prohibited\n",
5026 if (env->log.level & BPF_LOG_LEVEL)
5027 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
5031 /* return the map pointer stored inside BPF_LD_IMM64 instruction */
5032 static struct bpf_map *ld_imm64_to_map_ptr(struct bpf_insn *insn)
5034 u64 imm64 = ((u64) (u32) insn[0].imm) | ((u64) (u32) insn[1].imm) << 32;
5036 return (struct bpf_map *) (unsigned long) imm64;
5039 /* verify BPF_LD_IMM64 instruction */
5040 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
5042 struct bpf_reg_state *regs = cur_regs(env);
5045 if (BPF_SIZE(insn->code) != BPF_DW) {
5046 verbose(env, "invalid BPF_LD_IMM insn\n");
5049 if (insn->off != 0) {
5050 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
5054 err = check_reg_arg(env, insn->dst_reg, DST_OP);
5058 if (insn->src_reg == 0) {
5059 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
5061 regs[insn->dst_reg].type = SCALAR_VALUE;
5062 __mark_reg_known(®s[insn->dst_reg], imm);
5066 /* replace_map_fd_with_map_ptr() should have caught bad ld_imm64 */
5067 BUG_ON(insn->src_reg != BPF_PSEUDO_MAP_FD);
5069 regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
5070 regs[insn->dst_reg].map_ptr = ld_imm64_to_map_ptr(insn);
5074 static bool may_access_skb(enum bpf_prog_type type)
5077 case BPF_PROG_TYPE_SOCKET_FILTER:
5078 case BPF_PROG_TYPE_SCHED_CLS:
5079 case BPF_PROG_TYPE_SCHED_ACT:
5086 /* verify safety of LD_ABS|LD_IND instructions:
5087 * - they can only appear in the programs where ctx == skb
5088 * - since they are wrappers of function calls, they scratch R1-R5 registers,
5089 * preserve R6-R9, and store return value into R0
5092 * ctx == skb == R6 == CTX
5095 * SRC == any register
5096 * IMM == 32-bit immediate
5099 * R0 - 8/16/32-bit skb data converted to cpu endianness
5101 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
5103 struct bpf_reg_state *regs = cur_regs(env);
5104 u8 mode = BPF_MODE(insn->code);
5107 if (!may_access_skb(env->prog->type)) {
5108 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
5112 if (!env->ops->gen_ld_abs) {
5113 verbose(env, "bpf verifier is misconfigured\n");
5117 if (env->subprog_cnt > 1) {
5118 /* when program has LD_ABS insn JITs and interpreter assume
5119 * that r1 == ctx == skb which is not the case for callees
5120 * that can have arbitrary arguments. It's problematic
5121 * for main prog as well since JITs would need to analyze
5122 * all functions in order to make proper register save/restore
5123 * decisions in the main prog. Hence disallow LD_ABS with calls
5125 verbose(env, "BPF_LD_[ABS|IND] instructions cannot be mixed with bpf-to-bpf calls\n");
5129 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
5130 BPF_SIZE(insn->code) == BPF_DW ||
5131 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
5132 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
5136 /* check whether implicit source operand (register R6) is readable */
5137 err = check_reg_arg(env, BPF_REG_6, SRC_OP);
5141 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
5142 * gen_ld_abs() may terminate the program at runtime, leading to
5145 err = check_reference_leak(env);
5147 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
5151 if (env->cur_state->active_spin_lock) {
5152 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
5156 if (regs[BPF_REG_6].type != PTR_TO_CTX) {
5158 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
5162 if (mode == BPF_IND) {
5163 /* check explicit source operand */
5164 err = check_reg_arg(env, insn->src_reg, SRC_OP);
5169 /* reset caller saved regs to unreadable */
5170 for (i = 0; i < CALLER_SAVED_REGS; i++) {
5171 mark_reg_not_init(env, regs, caller_saved[i]);
5172 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
5175 /* mark destination R0 register as readable, since it contains
5176 * the value fetched from the packet.
5177 * Already marked as written above.
5179 mark_reg_unknown(env, regs, BPF_REG_0);
5183 static int check_return_code(struct bpf_verifier_env *env)
5185 struct bpf_reg_state *reg;
5186 struct tnum range = tnum_range(0, 1);
5188 switch (env->prog->type) {
5189 case BPF_PROG_TYPE_CGROUP_SKB:
5190 case BPF_PROG_TYPE_CGROUP_SOCK:
5191 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
5192 case BPF_PROG_TYPE_SOCK_OPS:
5193 case BPF_PROG_TYPE_CGROUP_DEVICE:
5199 reg = cur_regs(env) + BPF_REG_0;
5200 if (reg->type != SCALAR_VALUE) {
5201 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
5202 reg_type_str[reg->type]);
5206 if (!tnum_in(range, reg->var_off)) {
5207 verbose(env, "At program exit the register R0 ");
5208 if (!tnum_is_unknown(reg->var_off)) {
5211 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
5212 verbose(env, "has value %s", tn_buf);
5214 verbose(env, "has unknown scalar value");
5216 verbose(env, " should have been 0 or 1\n");
5222 /* non-recursive DFS pseudo code
5223 * 1 procedure DFS-iterative(G,v):
5224 * 2 label v as discovered
5225 * 3 let S be a stack
5227 * 5 while S is not empty
5229 * 7 if t is what we're looking for:
5231 * 9 for all edges e in G.adjacentEdges(t) do
5232 * 10 if edge e is already labelled
5233 * 11 continue with the next edge
5234 * 12 w <- G.adjacentVertex(t,e)
5235 * 13 if vertex w is not discovered and not explored
5236 * 14 label e as tree-edge
5237 * 15 label w as discovered
5240 * 18 else if vertex w is discovered
5241 * 19 label e as back-edge
5243 * 21 // vertex w is explored
5244 * 22 label e as forward- or cross-edge
5245 * 23 label t as explored
5250 * 0x11 - discovered and fall-through edge labelled
5251 * 0x12 - discovered and fall-through and branch edges labelled
5262 #define STATE_LIST_MARK ((struct bpf_verifier_state_list *) -1L)
5264 static int *insn_stack; /* stack of insns to process */
5265 static int cur_stack; /* current stack index */
5266 static int *insn_state;
5268 /* t, w, e - match pseudo-code above:
5269 * t - index of current instruction
5270 * w - next instruction
5273 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env)
5275 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
5278 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
5281 if (w < 0 || w >= env->prog->len) {
5282 verbose_linfo(env, t, "%d: ", t);
5283 verbose(env, "jump out of range from insn %d to %d\n", t, w);
5288 /* mark branch target for state pruning */
5289 env->explored_states[w] = STATE_LIST_MARK;
5291 if (insn_state[w] == 0) {
5293 insn_state[t] = DISCOVERED | e;
5294 insn_state[w] = DISCOVERED;
5295 if (cur_stack >= env->prog->len)
5297 insn_stack[cur_stack++] = w;
5299 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
5300 verbose_linfo(env, t, "%d: ", t);
5301 verbose_linfo(env, w, "%d: ", w);
5302 verbose(env, "back-edge from insn %d to %d\n", t, w);
5304 } else if (insn_state[w] == EXPLORED) {
5305 /* forward- or cross-edge */
5306 insn_state[t] = DISCOVERED | e;
5308 verbose(env, "insn state internal bug\n");
5314 /* non-recursive depth-first-search to detect loops in BPF program
5315 * loop == back-edge in directed graph
5317 static int check_cfg(struct bpf_verifier_env *env)
5319 struct bpf_insn *insns = env->prog->insnsi;
5320 int insn_cnt = env->prog->len;
5324 insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
5328 insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
5334 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
5335 insn_stack[0] = 0; /* 0 is the first instruction */
5341 t = insn_stack[cur_stack - 1];
5343 if (BPF_CLASS(insns[t].code) == BPF_JMP ||
5344 BPF_CLASS(insns[t].code) == BPF_JMP32) {
5345 u8 opcode = BPF_OP(insns[t].code);
5347 if (opcode == BPF_EXIT) {
5349 } else if (opcode == BPF_CALL) {
5350 ret = push_insn(t, t + 1, FALLTHROUGH, env);
5355 if (t + 1 < insn_cnt)
5356 env->explored_states[t + 1] = STATE_LIST_MARK;
5357 if (insns[t].src_reg == BPF_PSEUDO_CALL) {
5358 env->explored_states[t] = STATE_LIST_MARK;
5359 ret = push_insn(t, t + insns[t].imm + 1, BRANCH, env);
5365 } else if (opcode == BPF_JA) {
5366 if (BPF_SRC(insns[t].code) != BPF_K) {
5370 /* unconditional jump with single edge */
5371 ret = push_insn(t, t + insns[t].off + 1,
5377 /* tell verifier to check for equivalent states
5378 * after every call and jump
5380 if (t + 1 < insn_cnt)
5381 env->explored_states[t + 1] = STATE_LIST_MARK;
5383 /* conditional jump with two edges */
5384 env->explored_states[t] = STATE_LIST_MARK;
5385 ret = push_insn(t, t + 1, FALLTHROUGH, env);
5391 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env);
5398 /* all other non-branch instructions with single
5401 ret = push_insn(t, t + 1, FALLTHROUGH, env);
5409 insn_state[t] = EXPLORED;
5410 if (cur_stack-- <= 0) {
5411 verbose(env, "pop stack internal bug\n");
5418 for (i = 0; i < insn_cnt; i++) {
5419 if (insn_state[i] != EXPLORED) {
5420 verbose(env, "unreachable insn %d\n", i);
5425 ret = 0; /* cfg looks good */
5433 /* The minimum supported BTF func info size */
5434 #define MIN_BPF_FUNCINFO_SIZE 8
5435 #define MAX_FUNCINFO_REC_SIZE 252
5437 static int check_btf_func(struct bpf_verifier_env *env,
5438 const union bpf_attr *attr,
5439 union bpf_attr __user *uattr)
5441 u32 i, nfuncs, urec_size, min_size;
5442 u32 krec_size = sizeof(struct bpf_func_info);
5443 struct bpf_func_info *krecord;
5444 const struct btf_type *type;
5445 struct bpf_prog *prog;
5446 const struct btf *btf;
5447 void __user *urecord;
5448 u32 prev_offset = 0;
5451 nfuncs = attr->func_info_cnt;
5455 if (nfuncs != env->subprog_cnt) {
5456 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
5460 urec_size = attr->func_info_rec_size;
5461 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
5462 urec_size > MAX_FUNCINFO_REC_SIZE ||
5463 urec_size % sizeof(u32)) {
5464 verbose(env, "invalid func info rec size %u\n", urec_size);
5469 btf = prog->aux->btf;
5471 urecord = u64_to_user_ptr(attr->func_info);
5472 min_size = min_t(u32, krec_size, urec_size);
5474 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
5478 for (i = 0; i < nfuncs; i++) {
5479 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
5481 if (ret == -E2BIG) {
5482 verbose(env, "nonzero tailing record in func info");
5483 /* set the size kernel expects so loader can zero
5484 * out the rest of the record.
5486 if (put_user(min_size, &uattr->func_info_rec_size))
5492 if (copy_from_user(&krecord[i], urecord, min_size)) {
5497 /* check insn_off */
5499 if (krecord[i].insn_off) {
5501 "nonzero insn_off %u for the first func info record",
5502 krecord[i].insn_off);
5506 } else if (krecord[i].insn_off <= prev_offset) {
5508 "same or smaller insn offset (%u) than previous func info record (%u)",
5509 krecord[i].insn_off, prev_offset);
5514 if (env->subprog_info[i].start != krecord[i].insn_off) {
5515 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
5521 type = btf_type_by_id(btf, krecord[i].type_id);
5522 if (!type || BTF_INFO_KIND(type->info) != BTF_KIND_FUNC) {
5523 verbose(env, "invalid type id %d in func info",
5524 krecord[i].type_id);
5529 prev_offset = krecord[i].insn_off;
5530 urecord += urec_size;
5533 prog->aux->func_info = krecord;
5534 prog->aux->func_info_cnt = nfuncs;
5542 static void adjust_btf_func(struct bpf_verifier_env *env)
5546 if (!env->prog->aux->func_info)
5549 for (i = 0; i < env->subprog_cnt; i++)
5550 env->prog->aux->func_info[i].insn_off = env->subprog_info[i].start;
5553 #define MIN_BPF_LINEINFO_SIZE (offsetof(struct bpf_line_info, line_col) + \
5554 sizeof(((struct bpf_line_info *)(0))->line_col))
5555 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
5557 static int check_btf_line(struct bpf_verifier_env *env,
5558 const union bpf_attr *attr,
5559 union bpf_attr __user *uattr)
5561 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
5562 struct bpf_subprog_info *sub;
5563 struct bpf_line_info *linfo;
5564 struct bpf_prog *prog;
5565 const struct btf *btf;
5566 void __user *ulinfo;
5569 nr_linfo = attr->line_info_cnt;
5573 rec_size = attr->line_info_rec_size;
5574 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
5575 rec_size > MAX_LINEINFO_REC_SIZE ||
5576 rec_size & (sizeof(u32) - 1))
5579 /* Need to zero it in case the userspace may
5580 * pass in a smaller bpf_line_info object.
5582 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
5583 GFP_KERNEL | __GFP_NOWARN);
5588 btf = prog->aux->btf;
5591 sub = env->subprog_info;
5592 ulinfo = u64_to_user_ptr(attr->line_info);
5593 expected_size = sizeof(struct bpf_line_info);
5594 ncopy = min_t(u32, expected_size, rec_size);
5595 for (i = 0; i < nr_linfo; i++) {
5596 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
5598 if (err == -E2BIG) {
5599 verbose(env, "nonzero tailing record in line_info");
5600 if (put_user(expected_size,
5601 &uattr->line_info_rec_size))
5607 if (copy_from_user(&linfo[i], ulinfo, ncopy)) {
5613 * Check insn_off to ensure
5614 * 1) strictly increasing AND
5615 * 2) bounded by prog->len
5617 * The linfo[0].insn_off == 0 check logically falls into
5618 * the later "missing bpf_line_info for func..." case
5619 * because the first linfo[0].insn_off must be the
5620 * first sub also and the first sub must have
5621 * subprog_info[0].start == 0.
5623 if ((i && linfo[i].insn_off <= prev_offset) ||
5624 linfo[i].insn_off >= prog->len) {
5625 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
5626 i, linfo[i].insn_off, prev_offset,
5632 if (!prog->insnsi[linfo[i].insn_off].code) {
5634 "Invalid insn code at line_info[%u].insn_off\n",
5640 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
5641 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
5642 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
5647 if (s != env->subprog_cnt) {
5648 if (linfo[i].insn_off == sub[s].start) {
5649 sub[s].linfo_idx = i;
5651 } else if (sub[s].start < linfo[i].insn_off) {
5652 verbose(env, "missing bpf_line_info for func#%u\n", s);
5658 prev_offset = linfo[i].insn_off;
5662 if (s != env->subprog_cnt) {
5663 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
5664 env->subprog_cnt - s, s);
5669 prog->aux->linfo = linfo;
5670 prog->aux->nr_linfo = nr_linfo;
5679 static int check_btf_info(struct bpf_verifier_env *env,
5680 const union bpf_attr *attr,
5681 union bpf_attr __user *uattr)
5686 if (!attr->func_info_cnt && !attr->line_info_cnt)
5689 btf = btf_get_by_fd(attr->prog_btf_fd);
5691 return PTR_ERR(btf);
5692 env->prog->aux->btf = btf;
5694 err = check_btf_func(env, attr, uattr);
5698 err = check_btf_line(env, attr, uattr);
5705 /* check %cur's range satisfies %old's */
5706 static bool range_within(struct bpf_reg_state *old,
5707 struct bpf_reg_state *cur)
5709 return old->umin_value <= cur->umin_value &&
5710 old->umax_value >= cur->umax_value &&
5711 old->smin_value <= cur->smin_value &&
5712 old->smax_value >= cur->smax_value;
5715 /* Maximum number of register states that can exist at once */
5716 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
5722 /* If in the old state two registers had the same id, then they need to have
5723 * the same id in the new state as well. But that id could be different from
5724 * the old state, so we need to track the mapping from old to new ids.
5725 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
5726 * regs with old id 5 must also have new id 9 for the new state to be safe. But
5727 * regs with a different old id could still have new id 9, we don't care about
5729 * So we look through our idmap to see if this old id has been seen before. If
5730 * so, we require the new id to match; otherwise, we add the id pair to the map.
5732 static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
5736 for (i = 0; i < ID_MAP_SIZE; i++) {
5737 if (!idmap[i].old) {
5738 /* Reached an empty slot; haven't seen this id before */
5739 idmap[i].old = old_id;
5740 idmap[i].cur = cur_id;
5743 if (idmap[i].old == old_id)
5744 return idmap[i].cur == cur_id;
5746 /* We ran out of idmap slots, which should be impossible */
5751 static void clean_func_state(struct bpf_verifier_env *env,
5752 struct bpf_func_state *st)
5754 enum bpf_reg_liveness live;
5757 for (i = 0; i < BPF_REG_FP; i++) {
5758 live = st->regs[i].live;
5759 /* liveness must not touch this register anymore */
5760 st->regs[i].live |= REG_LIVE_DONE;
5761 if (!(live & REG_LIVE_READ))
5762 /* since the register is unused, clear its state
5763 * to make further comparison simpler
5765 __mark_reg_not_init(&st->regs[i]);
5768 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
5769 live = st->stack[i].spilled_ptr.live;
5770 /* liveness must not touch this stack slot anymore */
5771 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
5772 if (!(live & REG_LIVE_READ)) {
5773 __mark_reg_not_init(&st->stack[i].spilled_ptr);
5774 for (j = 0; j < BPF_REG_SIZE; j++)
5775 st->stack[i].slot_type[j] = STACK_INVALID;
5780 static void clean_verifier_state(struct bpf_verifier_env *env,
5781 struct bpf_verifier_state *st)
5785 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
5786 /* all regs in this state in all frames were already marked */
5789 for (i = 0; i <= st->curframe; i++)
5790 clean_func_state(env, st->frame[i]);
5793 /* the parentage chains form a tree.
5794 * the verifier states are added to state lists at given insn and
5795 * pushed into state stack for future exploration.
5796 * when the verifier reaches bpf_exit insn some of the verifer states
5797 * stored in the state lists have their final liveness state already,
5798 * but a lot of states will get revised from liveness point of view when
5799 * the verifier explores other branches.
5802 * 2: if r1 == 100 goto pc+1
5805 * when the verifier reaches exit insn the register r0 in the state list of
5806 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
5807 * of insn 2 and goes exploring further. At the insn 4 it will walk the
5808 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
5810 * Since the verifier pushes the branch states as it sees them while exploring
5811 * the program the condition of walking the branch instruction for the second
5812 * time means that all states below this branch were already explored and
5813 * their final liveness markes are already propagated.
5814 * Hence when the verifier completes the search of state list in is_state_visited()
5815 * we can call this clean_live_states() function to mark all liveness states
5816 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
5818 * This function also clears the registers and stack for states that !READ
5819 * to simplify state merging.
5821 * Important note here that walking the same branch instruction in the callee
5822 * doesn't meant that the states are DONE. The verifier has to compare
5825 static void clean_live_states(struct bpf_verifier_env *env, int insn,
5826 struct bpf_verifier_state *cur)
5828 struct bpf_verifier_state_list *sl;
5831 sl = env->explored_states[insn];
5835 while (sl != STATE_LIST_MARK) {
5836 if (sl->state.curframe != cur->curframe)
5838 for (i = 0; i <= cur->curframe; i++)
5839 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
5841 clean_verifier_state(env, &sl->state);
5847 /* Returns true if (rold safe implies rcur safe) */
5848 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
5849 struct idpair *idmap)
5853 if (!(rold->live & REG_LIVE_READ))
5854 /* explored state didn't use this */
5857 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
5859 if (rold->type == PTR_TO_STACK)
5860 /* two stack pointers are equal only if they're pointing to
5861 * the same stack frame, since fp-8 in foo != fp-8 in bar
5863 return equal && rold->frameno == rcur->frameno;
5868 if (rold->type == NOT_INIT)
5869 /* explored state can't have used this */
5871 if (rcur->type == NOT_INIT)
5873 switch (rold->type) {
5875 if (rcur->type == SCALAR_VALUE) {
5876 /* new val must satisfy old val knowledge */
5877 return range_within(rold, rcur) &&
5878 tnum_in(rold->var_off, rcur->var_off);
5880 /* We're trying to use a pointer in place of a scalar.
5881 * Even if the scalar was unbounded, this could lead to
5882 * pointer leaks because scalars are allowed to leak
5883 * while pointers are not. We could make this safe in
5884 * special cases if root is calling us, but it's
5885 * probably not worth the hassle.
5889 case PTR_TO_MAP_VALUE:
5890 /* If the new min/max/var_off satisfy the old ones and
5891 * everything else matches, we are OK.
5892 * 'id' is not compared, since it's only used for maps with
5893 * bpf_spin_lock inside map element and in such cases if
5894 * the rest of the prog is valid for one map element then
5895 * it's valid for all map elements regardless of the key
5896 * used in bpf_map_lookup()
5898 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
5899 range_within(rold, rcur) &&
5900 tnum_in(rold->var_off, rcur->var_off);
5901 case PTR_TO_MAP_VALUE_OR_NULL:
5902 /* a PTR_TO_MAP_VALUE could be safe to use as a
5903 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
5904 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
5905 * checked, doing so could have affected others with the same
5906 * id, and we can't check for that because we lost the id when
5907 * we converted to a PTR_TO_MAP_VALUE.
5909 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
5911 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
5913 /* Check our ids match any regs they're supposed to */
5914 return check_ids(rold->id, rcur->id, idmap);
5915 case PTR_TO_PACKET_META:
5917 if (rcur->type != rold->type)
5919 /* We must have at least as much range as the old ptr
5920 * did, so that any accesses which were safe before are
5921 * still safe. This is true even if old range < old off,
5922 * since someone could have accessed through (ptr - k), or
5923 * even done ptr -= k in a register, to get a safe access.
5925 if (rold->range > rcur->range)
5927 /* If the offsets don't match, we can't trust our alignment;
5928 * nor can we be sure that we won't fall out of range.
5930 if (rold->off != rcur->off)
5932 /* id relations must be preserved */
5933 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
5935 /* new val must satisfy old val knowledge */
5936 return range_within(rold, rcur) &&
5937 tnum_in(rold->var_off, rcur->var_off);
5939 case CONST_PTR_TO_MAP:
5940 case PTR_TO_PACKET_END:
5941 case PTR_TO_FLOW_KEYS:
5943 case PTR_TO_SOCKET_OR_NULL:
5944 case PTR_TO_SOCK_COMMON:
5945 case PTR_TO_SOCK_COMMON_OR_NULL:
5946 case PTR_TO_TCP_SOCK:
5947 case PTR_TO_TCP_SOCK_OR_NULL:
5948 /* Only valid matches are exact, which memcmp() above
5949 * would have accepted
5952 /* Don't know what's going on, just say it's not safe */
5956 /* Shouldn't get here; if we do, say it's not safe */
5961 static bool stacksafe(struct bpf_func_state *old,
5962 struct bpf_func_state *cur,
5963 struct idpair *idmap)
5967 /* walk slots of the explored stack and ignore any additional
5968 * slots in the current stack, since explored(safe) state
5971 for (i = 0; i < old->allocated_stack; i++) {
5972 spi = i / BPF_REG_SIZE;
5974 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
5975 i += BPF_REG_SIZE - 1;
5976 /* explored state didn't use this */
5980 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
5983 /* explored stack has more populated slots than current stack
5984 * and these slots were used
5986 if (i >= cur->allocated_stack)
5989 /* if old state was safe with misc data in the stack
5990 * it will be safe with zero-initialized stack.
5991 * The opposite is not true
5993 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
5994 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
5996 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
5997 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
5998 /* Ex: old explored (safe) state has STACK_SPILL in
5999 * this stack slot, but current has has STACK_MISC ->
6000 * this verifier states are not equivalent,
6001 * return false to continue verification of this path
6004 if (i % BPF_REG_SIZE)
6006 if (old->stack[spi].slot_type[0] != STACK_SPILL)
6008 if (!regsafe(&old->stack[spi].spilled_ptr,
6009 &cur->stack[spi].spilled_ptr,
6011 /* when explored and current stack slot are both storing
6012 * spilled registers, check that stored pointers types
6013 * are the same as well.
6014 * Ex: explored safe path could have stored
6015 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
6016 * but current path has stored:
6017 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
6018 * such verifier states are not equivalent.
6019 * return false to continue verification of this path
6026 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
6028 if (old->acquired_refs != cur->acquired_refs)
6030 return !memcmp(old->refs, cur->refs,
6031 sizeof(*old->refs) * old->acquired_refs);
6034 /* compare two verifier states
6036 * all states stored in state_list are known to be valid, since
6037 * verifier reached 'bpf_exit' instruction through them
6039 * this function is called when verifier exploring different branches of
6040 * execution popped from the state stack. If it sees an old state that has
6041 * more strict register state and more strict stack state then this execution
6042 * branch doesn't need to be explored further, since verifier already
6043 * concluded that more strict state leads to valid finish.
6045 * Therefore two states are equivalent if register state is more conservative
6046 * and explored stack state is more conservative than the current one.
6049 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
6050 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
6052 * In other words if current stack state (one being explored) has more
6053 * valid slots than old one that already passed validation, it means
6054 * the verifier can stop exploring and conclude that current state is valid too
6056 * Similarly with registers. If explored state has register type as invalid
6057 * whereas register type in current state is meaningful, it means that
6058 * the current state will reach 'bpf_exit' instruction safely
6060 static bool func_states_equal(struct bpf_func_state *old,
6061 struct bpf_func_state *cur)
6063 struct idpair *idmap;
6067 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
6068 /* If we failed to allocate the idmap, just say it's not safe */
6072 for (i = 0; i < MAX_BPF_REG; i++) {
6073 if (!regsafe(&old->regs[i], &cur->regs[i], idmap))
6077 if (!stacksafe(old, cur, idmap))
6080 if (!refsafe(old, cur))
6088 static bool states_equal(struct bpf_verifier_env *env,
6089 struct bpf_verifier_state *old,
6090 struct bpf_verifier_state *cur)
6094 if (old->curframe != cur->curframe)
6097 /* Verification state from speculative execution simulation
6098 * must never prune a non-speculative execution one.
6100 if (old->speculative && !cur->speculative)
6103 if (old->active_spin_lock != cur->active_spin_lock)
6106 /* for states to be equal callsites have to be the same
6107 * and all frame states need to be equivalent
6109 for (i = 0; i <= old->curframe; i++) {
6110 if (old->frame[i]->callsite != cur->frame[i]->callsite)
6112 if (!func_states_equal(old->frame[i], cur->frame[i]))
6118 /* A write screens off any subsequent reads; but write marks come from the
6119 * straight-line code between a state and its parent. When we arrive at an
6120 * equivalent state (jump target or such) we didn't arrive by the straight-line
6121 * code, so read marks in the state must propagate to the parent regardless
6122 * of the state's write marks. That's what 'parent == state->parent' comparison
6123 * in mark_reg_read() is for.
6125 static int propagate_liveness(struct bpf_verifier_env *env,
6126 const struct bpf_verifier_state *vstate,
6127 struct bpf_verifier_state *vparent)
6129 int i, frame, err = 0;
6130 struct bpf_func_state *state, *parent;
6132 if (vparent->curframe != vstate->curframe) {
6133 WARN(1, "propagate_live: parent frame %d current frame %d\n",
6134 vparent->curframe, vstate->curframe);
6137 /* Propagate read liveness of registers... */
6138 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
6139 for (frame = 0; frame <= vstate->curframe; frame++) {
6140 /* We don't need to worry about FP liveness, it's read-only */
6141 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
6142 if (vparent->frame[frame]->regs[i].live & REG_LIVE_READ)
6144 if (vstate->frame[frame]->regs[i].live & REG_LIVE_READ) {
6145 err = mark_reg_read(env, &vstate->frame[frame]->regs[i],
6146 &vparent->frame[frame]->regs[i]);
6153 /* ... and stack slots */
6154 for (frame = 0; frame <= vstate->curframe; frame++) {
6155 state = vstate->frame[frame];
6156 parent = vparent->frame[frame];
6157 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
6158 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
6159 if (parent->stack[i].spilled_ptr.live & REG_LIVE_READ)
6161 if (state->stack[i].spilled_ptr.live & REG_LIVE_READ)
6162 mark_reg_read(env, &state->stack[i].spilled_ptr,
6163 &parent->stack[i].spilled_ptr);
6169 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
6171 struct bpf_verifier_state_list *new_sl;
6172 struct bpf_verifier_state_list *sl, **pprev;
6173 struct bpf_verifier_state *cur = env->cur_state, *new;
6174 int i, j, err, states_cnt = 0;
6176 pprev = &env->explored_states[insn_idx];
6180 /* this 'insn_idx' instruction wasn't marked, so we will not
6181 * be doing state search here
6185 clean_live_states(env, insn_idx, cur);
6187 while (sl != STATE_LIST_MARK) {
6188 if (states_equal(env, &sl->state, cur)) {
6190 /* reached equivalent register/stack state,
6192 * Registers read by the continuation are read by us.
6193 * If we have any write marks in env->cur_state, they
6194 * will prevent corresponding reads in the continuation
6195 * from reaching our parent (an explored_state). Our
6196 * own state will get the read marks recorded, but
6197 * they'll be immediately forgotten as we're pruning
6198 * this state and will pop a new one.
6200 err = propagate_liveness(env, &sl->state, cur);
6207 /* heuristic to determine whether this state is beneficial
6208 * to keep checking from state equivalence point of view.
6209 * Higher numbers increase max_states_per_insn and verification time,
6210 * but do not meaningfully decrease insn_processed.
6212 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
6213 /* the state is unlikely to be useful. Remove it to
6214 * speed up verification
6217 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
6218 free_verifier_state(&sl->state, false);
6222 /* cannot free this state, since parentage chain may
6223 * walk it later. Add it for free_list instead to
6224 * be freed at the end of verification
6226 sl->next = env->free_list;
6227 env->free_list = sl;
6236 if (env->max_states_per_insn < states_cnt)
6237 env->max_states_per_insn = states_cnt;
6239 if (!env->allow_ptr_leaks && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
6242 /* there were no equivalent states, remember current one.
6243 * technically the current state is not proven to be safe yet,
6244 * but it will either reach outer most bpf_exit (which means it's safe)
6245 * or it will be rejected. Since there are no loops, we won't be
6246 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
6247 * again on the way to bpf_exit
6249 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
6252 env->total_states++;
6255 /* add new state to the head of linked list */
6256 new = &new_sl->state;
6257 err = copy_verifier_state(new, cur);
6259 free_verifier_state(new, false);
6263 new_sl->next = env->explored_states[insn_idx];
6264 env->explored_states[insn_idx] = new_sl;
6265 /* connect new state to parentage chain. Current frame needs all
6266 * registers connected. Only r6 - r9 of the callers are alive (pushed
6267 * to the stack implicitly by JITs) so in callers' frames connect just
6268 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
6269 * the state of the call instruction (with WRITTEN set), and r0 comes
6270 * from callee with its full parentage chain, anyway.
6272 for (j = 0; j <= cur->curframe; j++)
6273 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
6274 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
6275 /* clear write marks in current state: the writes we did are not writes
6276 * our child did, so they don't screen off its reads from us.
6277 * (There are no read marks in current state, because reads always mark
6278 * their parent and current state never has children yet. Only
6279 * explored_states can get read marks.)
6281 for (i = 0; i < BPF_REG_FP; i++)
6282 cur->frame[cur->curframe]->regs[i].live = REG_LIVE_NONE;
6284 /* all stack frames are accessible from callee, clear them all */
6285 for (j = 0; j <= cur->curframe; j++) {
6286 struct bpf_func_state *frame = cur->frame[j];
6287 struct bpf_func_state *newframe = new->frame[j];
6289 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
6290 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
6291 frame->stack[i].spilled_ptr.parent =
6292 &newframe->stack[i].spilled_ptr;
6298 /* Return true if it's OK to have the same insn return a different type. */
6299 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
6304 case PTR_TO_SOCKET_OR_NULL:
6305 case PTR_TO_SOCK_COMMON:
6306 case PTR_TO_SOCK_COMMON_OR_NULL:
6307 case PTR_TO_TCP_SOCK:
6308 case PTR_TO_TCP_SOCK_OR_NULL:
6315 /* If an instruction was previously used with particular pointer types, then we
6316 * need to be careful to avoid cases such as the below, where it may be ok
6317 * for one branch accessing the pointer, but not ok for the other branch:
6322 * R1 = some_other_valid_ptr;
6325 * R2 = *(u32 *)(R1 + 0);
6327 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
6329 return src != prev && (!reg_type_mismatch_ok(src) ||
6330 !reg_type_mismatch_ok(prev));
6333 static int do_check(struct bpf_verifier_env *env)
6335 struct bpf_verifier_state *state;
6336 struct bpf_insn *insns = env->prog->insnsi;
6337 struct bpf_reg_state *regs;
6338 int insn_cnt = env->prog->len;
6339 bool do_print_state = false;
6341 env->prev_linfo = NULL;
6343 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
6346 state->curframe = 0;
6347 state->speculative = false;
6348 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
6349 if (!state->frame[0]) {
6353 env->cur_state = state;
6354 init_func_state(env, state->frame[0],
6355 BPF_MAIN_FUNC /* callsite */,
6357 0 /* subprogno, zero == main subprog */);
6360 struct bpf_insn *insn;
6364 if (env->insn_idx >= insn_cnt) {
6365 verbose(env, "invalid insn idx %d insn_cnt %d\n",
6366 env->insn_idx, insn_cnt);
6370 insn = &insns[env->insn_idx];
6371 class = BPF_CLASS(insn->code);
6373 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
6375 "BPF program is too large. Processed %d insn\n",
6376 env->insn_processed);
6380 err = is_state_visited(env, env->insn_idx);
6384 /* found equivalent state, can prune the search */
6385 if (env->log.level & BPF_LOG_LEVEL) {
6387 verbose(env, "\nfrom %d to %d%s: safe\n",
6388 env->prev_insn_idx, env->insn_idx,
6389 env->cur_state->speculative ?
6390 " (speculative execution)" : "");
6392 verbose(env, "%d: safe\n", env->insn_idx);
6394 goto process_bpf_exit;
6397 if (signal_pending(current))
6403 if (env->log.level & BPF_LOG_LEVEL2 ||
6404 (env->log.level & BPF_LOG_LEVEL && do_print_state)) {
6405 if (env->log.level & BPF_LOG_LEVEL2)
6406 verbose(env, "%d:", env->insn_idx);
6408 verbose(env, "\nfrom %d to %d%s:",
6409 env->prev_insn_idx, env->insn_idx,
6410 env->cur_state->speculative ?
6411 " (speculative execution)" : "");
6412 print_verifier_state(env, state->frame[state->curframe]);
6413 do_print_state = false;
6416 if (env->log.level & BPF_LOG_LEVEL) {
6417 const struct bpf_insn_cbs cbs = {
6418 .cb_print = verbose,
6419 .private_data = env,
6422 verbose_linfo(env, env->insn_idx, "; ");
6423 verbose(env, "%d: ", env->insn_idx);
6424 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
6427 if (bpf_prog_is_dev_bound(env->prog->aux)) {
6428 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
6429 env->prev_insn_idx);
6434 regs = cur_regs(env);
6435 env->insn_aux_data[env->insn_idx].seen = true;
6437 if (class == BPF_ALU || class == BPF_ALU64) {
6438 err = check_alu_op(env, insn);
6442 } else if (class == BPF_LDX) {
6443 enum bpf_reg_type *prev_src_type, src_reg_type;
6445 /* check for reserved fields is already done */
6447 /* check src operand */
6448 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6452 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
6456 src_reg_type = regs[insn->src_reg].type;
6458 /* check that memory (src_reg + off) is readable,
6459 * the state of dst_reg will be updated by this func
6461 err = check_mem_access(env, env->insn_idx, insn->src_reg,
6462 insn->off, BPF_SIZE(insn->code),
6463 BPF_READ, insn->dst_reg, false);
6467 prev_src_type = &env->insn_aux_data[env->insn_idx].ptr_type;
6469 if (*prev_src_type == NOT_INIT) {
6471 * dst_reg = *(u32 *)(src_reg + off)
6472 * save type to validate intersecting paths
6474 *prev_src_type = src_reg_type;
6476 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
6477 /* ABuser program is trying to use the same insn
6478 * dst_reg = *(u32*) (src_reg + off)
6479 * with different pointer types:
6480 * src_reg == ctx in one branch and
6481 * src_reg == stack|map in some other branch.
6484 verbose(env, "same insn cannot be used with different pointers\n");
6488 } else if (class == BPF_STX) {
6489 enum bpf_reg_type *prev_dst_type, dst_reg_type;
6491 if (BPF_MODE(insn->code) == BPF_XADD) {
6492 err = check_xadd(env, env->insn_idx, insn);
6499 /* check src1 operand */
6500 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6503 /* check src2 operand */
6504 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
6508 dst_reg_type = regs[insn->dst_reg].type;
6510 /* check that memory (dst_reg + off) is writeable */
6511 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
6512 insn->off, BPF_SIZE(insn->code),
6513 BPF_WRITE, insn->src_reg, false);
6517 prev_dst_type = &env->insn_aux_data[env->insn_idx].ptr_type;
6519 if (*prev_dst_type == NOT_INIT) {
6520 *prev_dst_type = dst_reg_type;
6521 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
6522 verbose(env, "same insn cannot be used with different pointers\n");
6526 } else if (class == BPF_ST) {
6527 if (BPF_MODE(insn->code) != BPF_MEM ||
6528 insn->src_reg != BPF_REG_0) {
6529 verbose(env, "BPF_ST uses reserved fields\n");
6532 /* check src operand */
6533 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
6537 if (is_ctx_reg(env, insn->dst_reg)) {
6538 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
6540 reg_type_str[reg_state(env, insn->dst_reg)->type]);
6544 /* check that memory (dst_reg + off) is writeable */
6545 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
6546 insn->off, BPF_SIZE(insn->code),
6547 BPF_WRITE, -1, false);
6551 } else if (class == BPF_JMP || class == BPF_JMP32) {
6552 u8 opcode = BPF_OP(insn->code);
6554 if (opcode == BPF_CALL) {
6555 if (BPF_SRC(insn->code) != BPF_K ||
6557 (insn->src_reg != BPF_REG_0 &&
6558 insn->src_reg != BPF_PSEUDO_CALL) ||
6559 insn->dst_reg != BPF_REG_0 ||
6560 class == BPF_JMP32) {
6561 verbose(env, "BPF_CALL uses reserved fields\n");
6565 if (env->cur_state->active_spin_lock &&
6566 (insn->src_reg == BPF_PSEUDO_CALL ||
6567 insn->imm != BPF_FUNC_spin_unlock)) {
6568 verbose(env, "function calls are not allowed while holding a lock\n");
6571 if (insn->src_reg == BPF_PSEUDO_CALL)
6572 err = check_func_call(env, insn, &env->insn_idx);
6574 err = check_helper_call(env, insn->imm, env->insn_idx);
6578 } else if (opcode == BPF_JA) {
6579 if (BPF_SRC(insn->code) != BPF_K ||
6581 insn->src_reg != BPF_REG_0 ||
6582 insn->dst_reg != BPF_REG_0 ||
6583 class == BPF_JMP32) {
6584 verbose(env, "BPF_JA uses reserved fields\n");
6588 env->insn_idx += insn->off + 1;
6591 } else if (opcode == BPF_EXIT) {
6592 if (BPF_SRC(insn->code) != BPF_K ||
6594 insn->src_reg != BPF_REG_0 ||
6595 insn->dst_reg != BPF_REG_0 ||
6596 class == BPF_JMP32) {
6597 verbose(env, "BPF_EXIT uses reserved fields\n");
6601 if (env->cur_state->active_spin_lock) {
6602 verbose(env, "bpf_spin_unlock is missing\n");
6606 if (state->curframe) {
6607 /* exit from nested function */
6608 env->prev_insn_idx = env->insn_idx;
6609 err = prepare_func_exit(env, &env->insn_idx);
6612 do_print_state = true;
6616 err = check_reference_leak(env);
6620 /* eBPF calling convetion is such that R0 is used
6621 * to return the value from eBPF program.
6622 * Make sure that it's readable at this time
6623 * of bpf_exit, which means that program wrote
6624 * something into it earlier
6626 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
6630 if (is_pointer_value(env, BPF_REG_0)) {
6631 verbose(env, "R0 leaks addr as return value\n");
6635 err = check_return_code(env);
6639 err = pop_stack(env, &env->prev_insn_idx,
6646 do_print_state = true;
6650 err = check_cond_jmp_op(env, insn, &env->insn_idx);
6654 } else if (class == BPF_LD) {
6655 u8 mode = BPF_MODE(insn->code);
6657 if (mode == BPF_ABS || mode == BPF_IND) {
6658 err = check_ld_abs(env, insn);
6662 } else if (mode == BPF_IMM) {
6663 err = check_ld_imm(env, insn);
6668 env->insn_aux_data[env->insn_idx].seen = true;
6670 verbose(env, "invalid BPF_LD mode\n");
6674 verbose(env, "unknown insn class %d\n", class);
6681 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
6685 static int check_map_prealloc(struct bpf_map *map)
6687 return (map->map_type != BPF_MAP_TYPE_HASH &&
6688 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
6689 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
6690 !(map->map_flags & BPF_F_NO_PREALLOC);
6693 static bool is_tracing_prog_type(enum bpf_prog_type type)
6696 case BPF_PROG_TYPE_KPROBE:
6697 case BPF_PROG_TYPE_TRACEPOINT:
6698 case BPF_PROG_TYPE_PERF_EVENT:
6699 case BPF_PROG_TYPE_RAW_TRACEPOINT:
6706 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
6707 struct bpf_map *map,
6708 struct bpf_prog *prog)
6711 /* Make sure that BPF_PROG_TYPE_PERF_EVENT programs only use
6712 * preallocated hash maps, since doing memory allocation
6713 * in overflow_handler can crash depending on where nmi got
6716 if (prog->type == BPF_PROG_TYPE_PERF_EVENT) {
6717 if (!check_map_prealloc(map)) {
6718 verbose(env, "perf_event programs can only use preallocated hash map\n");
6721 if (map->inner_map_meta &&
6722 !check_map_prealloc(map->inner_map_meta)) {
6723 verbose(env, "perf_event programs can only use preallocated inner hash map\n");
6728 if ((is_tracing_prog_type(prog->type) ||
6729 prog->type == BPF_PROG_TYPE_SOCKET_FILTER) &&
6730 map_value_has_spin_lock(map)) {
6731 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
6735 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
6736 !bpf_offload_prog_map_match(prog, map)) {
6737 verbose(env, "offload device mismatch between prog and map\n");
6744 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
6746 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
6747 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
6750 /* look for pseudo eBPF instructions that access map FDs and
6751 * replace them with actual map pointers
6753 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env *env)
6755 struct bpf_insn *insn = env->prog->insnsi;
6756 int insn_cnt = env->prog->len;
6759 err = bpf_prog_calc_tag(env->prog);
6763 for (i = 0; i < insn_cnt; i++, insn++) {
6764 if (BPF_CLASS(insn->code) == BPF_LDX &&
6765 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
6766 verbose(env, "BPF_LDX uses reserved fields\n");
6770 if (BPF_CLASS(insn->code) == BPF_STX &&
6771 ((BPF_MODE(insn->code) != BPF_MEM &&
6772 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
6773 verbose(env, "BPF_STX uses reserved fields\n");
6777 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
6778 struct bpf_map *map;
6781 if (i == insn_cnt - 1 || insn[1].code != 0 ||
6782 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
6784 verbose(env, "invalid bpf_ld_imm64 insn\n");
6788 if (insn->src_reg == 0)
6789 /* valid generic load 64-bit imm */
6792 if (insn[0].src_reg != BPF_PSEUDO_MAP_FD ||
6794 verbose(env, "unrecognized bpf_ld_imm64 insn\n");
6798 f = fdget(insn[0].imm);
6799 map = __bpf_map_get(f);
6801 verbose(env, "fd %d is not pointing to valid bpf_map\n",
6803 return PTR_ERR(map);
6806 err = check_map_prog_compatibility(env, map, env->prog);
6812 /* store map pointer inside BPF_LD_IMM64 instruction */
6813 insn[0].imm = (u32) (unsigned long) map;
6814 insn[1].imm = ((u64) (unsigned long) map) >> 32;
6816 /* check whether we recorded this map already */
6817 for (j = 0; j < env->used_map_cnt; j++)
6818 if (env->used_maps[j] == map) {
6823 if (env->used_map_cnt >= MAX_USED_MAPS) {
6828 /* hold the map. If the program is rejected by verifier,
6829 * the map will be released by release_maps() or it
6830 * will be used by the valid program until it's unloaded
6831 * and all maps are released in free_used_maps()
6833 map = bpf_map_inc(map, false);
6836 return PTR_ERR(map);
6838 env->used_maps[env->used_map_cnt++] = map;
6840 if (bpf_map_is_cgroup_storage(map) &&
6841 bpf_cgroup_storage_assign(env->prog, map)) {
6842 verbose(env, "only one cgroup storage of each type is allowed\n");
6854 /* Basic sanity check before we invest more work here. */
6855 if (!bpf_opcode_in_insntable(insn->code)) {
6856 verbose(env, "unknown opcode %02x\n", insn->code);
6861 /* now all pseudo BPF_LD_IMM64 instructions load valid
6862 * 'struct bpf_map *' into a register instead of user map_fd.
6863 * These pointers will be used later by verifier to validate map access.
6868 /* drop refcnt of maps used by the rejected program */
6869 static void release_maps(struct bpf_verifier_env *env)
6871 enum bpf_cgroup_storage_type stype;
6874 for_each_cgroup_storage_type(stype) {
6875 if (!env->prog->aux->cgroup_storage[stype])
6877 bpf_cgroup_storage_release(env->prog,
6878 env->prog->aux->cgroup_storage[stype]);
6881 for (i = 0; i < env->used_map_cnt; i++)
6882 bpf_map_put(env->used_maps[i]);
6885 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
6886 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
6888 struct bpf_insn *insn = env->prog->insnsi;
6889 int insn_cnt = env->prog->len;
6892 for (i = 0; i < insn_cnt; i++, insn++)
6893 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
6897 /* single env->prog->insni[off] instruction was replaced with the range
6898 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
6899 * [0, off) and [off, end) to new locations, so the patched range stays zero
6901 static int adjust_insn_aux_data(struct bpf_verifier_env *env, u32 prog_len,
6904 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
6909 new_data = vzalloc(array_size(prog_len,
6910 sizeof(struct bpf_insn_aux_data)));
6913 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
6914 memcpy(new_data + off + cnt - 1, old_data + off,
6915 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
6916 for (i = off; i < off + cnt - 1; i++)
6917 new_data[i].seen = true;
6918 env->insn_aux_data = new_data;
6923 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
6929 /* NOTE: fake 'exit' subprog should be updated as well. */
6930 for (i = 0; i <= env->subprog_cnt; i++) {
6931 if (env->subprog_info[i].start <= off)
6933 env->subprog_info[i].start += len - 1;
6937 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
6938 const struct bpf_insn *patch, u32 len)
6940 struct bpf_prog *new_prog;
6942 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
6943 if (IS_ERR(new_prog)) {
6944 if (PTR_ERR(new_prog) == -ERANGE)
6946 "insn %d cannot be patched due to 16-bit range\n",
6947 env->insn_aux_data[off].orig_idx);
6950 if (adjust_insn_aux_data(env, new_prog->len, off, len))
6952 adjust_subprog_starts(env, off, len);
6956 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
6961 /* find first prog starting at or after off (first to remove) */
6962 for (i = 0; i < env->subprog_cnt; i++)
6963 if (env->subprog_info[i].start >= off)
6965 /* find first prog starting at or after off + cnt (first to stay) */
6966 for (j = i; j < env->subprog_cnt; j++)
6967 if (env->subprog_info[j].start >= off + cnt)
6969 /* if j doesn't start exactly at off + cnt, we are just removing
6970 * the front of previous prog
6972 if (env->subprog_info[j].start != off + cnt)
6976 struct bpf_prog_aux *aux = env->prog->aux;
6979 /* move fake 'exit' subprog as well */
6980 move = env->subprog_cnt + 1 - j;
6982 memmove(env->subprog_info + i,
6983 env->subprog_info + j,
6984 sizeof(*env->subprog_info) * move);
6985 env->subprog_cnt -= j - i;
6987 /* remove func_info */
6988 if (aux->func_info) {
6989 move = aux->func_info_cnt - j;
6991 memmove(aux->func_info + i,
6993 sizeof(*aux->func_info) * move);
6994 aux->func_info_cnt -= j - i;
6995 /* func_info->insn_off is set after all code rewrites,
6996 * in adjust_btf_func() - no need to adjust
7000 /* convert i from "first prog to remove" to "first to adjust" */
7001 if (env->subprog_info[i].start == off)
7005 /* update fake 'exit' subprog as well */
7006 for (; i <= env->subprog_cnt; i++)
7007 env->subprog_info[i].start -= cnt;
7012 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
7015 struct bpf_prog *prog = env->prog;
7016 u32 i, l_off, l_cnt, nr_linfo;
7017 struct bpf_line_info *linfo;
7019 nr_linfo = prog->aux->nr_linfo;
7023 linfo = prog->aux->linfo;
7025 /* find first line info to remove, count lines to be removed */
7026 for (i = 0; i < nr_linfo; i++)
7027 if (linfo[i].insn_off >= off)
7032 for (; i < nr_linfo; i++)
7033 if (linfo[i].insn_off < off + cnt)
7038 /* First live insn doesn't match first live linfo, it needs to "inherit"
7039 * last removed linfo. prog is already modified, so prog->len == off
7040 * means no live instructions after (tail of the program was removed).
7042 if (prog->len != off && l_cnt &&
7043 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
7045 linfo[--i].insn_off = off + cnt;
7048 /* remove the line info which refer to the removed instructions */
7050 memmove(linfo + l_off, linfo + i,
7051 sizeof(*linfo) * (nr_linfo - i));
7053 prog->aux->nr_linfo -= l_cnt;
7054 nr_linfo = prog->aux->nr_linfo;
7057 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
7058 for (i = l_off; i < nr_linfo; i++)
7059 linfo[i].insn_off -= cnt;
7061 /* fix up all subprogs (incl. 'exit') which start >= off */
7062 for (i = 0; i <= env->subprog_cnt; i++)
7063 if (env->subprog_info[i].linfo_idx > l_off) {
7064 /* program may have started in the removed region but
7065 * may not be fully removed
7067 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
7068 env->subprog_info[i].linfo_idx -= l_cnt;
7070 env->subprog_info[i].linfo_idx = l_off;
7076 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
7078 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
7079 unsigned int orig_prog_len = env->prog->len;
7082 if (bpf_prog_is_dev_bound(env->prog->aux))
7083 bpf_prog_offload_remove_insns(env, off, cnt);
7085 err = bpf_remove_insns(env->prog, off, cnt);
7089 err = adjust_subprog_starts_after_remove(env, off, cnt);
7093 err = bpf_adj_linfo_after_remove(env, off, cnt);
7097 memmove(aux_data + off, aux_data + off + cnt,
7098 sizeof(*aux_data) * (orig_prog_len - off - cnt));
7103 /* The verifier does more data flow analysis than llvm and will not
7104 * explore branches that are dead at run time. Malicious programs can
7105 * have dead code too. Therefore replace all dead at-run-time code
7108 * Just nops are not optimal, e.g. if they would sit at the end of the
7109 * program and through another bug we would manage to jump there, then
7110 * we'd execute beyond program memory otherwise. Returning exception
7111 * code also wouldn't work since we can have subprogs where the dead
7112 * code could be located.
7114 static void sanitize_dead_code(struct bpf_verifier_env *env)
7116 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
7117 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
7118 struct bpf_insn *insn = env->prog->insnsi;
7119 const int insn_cnt = env->prog->len;
7122 for (i = 0; i < insn_cnt; i++) {
7123 if (aux_data[i].seen)
7125 memcpy(insn + i, &trap, sizeof(trap));
7129 static bool insn_is_cond_jump(u8 code)
7133 if (BPF_CLASS(code) == BPF_JMP32)
7136 if (BPF_CLASS(code) != BPF_JMP)
7140 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
7143 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
7145 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
7146 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
7147 struct bpf_insn *insn = env->prog->insnsi;
7148 const int insn_cnt = env->prog->len;
7151 for (i = 0; i < insn_cnt; i++, insn++) {
7152 if (!insn_is_cond_jump(insn->code))
7155 if (!aux_data[i + 1].seen)
7157 else if (!aux_data[i + 1 + insn->off].seen)
7162 if (bpf_prog_is_dev_bound(env->prog->aux))
7163 bpf_prog_offload_replace_insn(env, i, &ja);
7165 memcpy(insn, &ja, sizeof(ja));
7169 static int opt_remove_dead_code(struct bpf_verifier_env *env)
7171 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
7172 int insn_cnt = env->prog->len;
7175 for (i = 0; i < insn_cnt; i++) {
7179 while (i + j < insn_cnt && !aux_data[i + j].seen)
7184 err = verifier_remove_insns(env, i, j);
7187 insn_cnt = env->prog->len;
7193 static int opt_remove_nops(struct bpf_verifier_env *env)
7195 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
7196 struct bpf_insn *insn = env->prog->insnsi;
7197 int insn_cnt = env->prog->len;
7200 for (i = 0; i < insn_cnt; i++) {
7201 if (memcmp(&insn[i], &ja, sizeof(ja)))
7204 err = verifier_remove_insns(env, i, 1);
7214 /* convert load instructions that access fields of a context type into a
7215 * sequence of instructions that access fields of the underlying structure:
7216 * struct __sk_buff -> struct sk_buff
7217 * struct bpf_sock_ops -> struct sock
7219 static int convert_ctx_accesses(struct bpf_verifier_env *env)
7221 const struct bpf_verifier_ops *ops = env->ops;
7222 int i, cnt, size, ctx_field_size, delta = 0;
7223 const int insn_cnt = env->prog->len;
7224 struct bpf_insn insn_buf[16], *insn;
7225 u32 target_size, size_default, off;
7226 struct bpf_prog *new_prog;
7227 enum bpf_access_type type;
7228 bool is_narrower_load;
7230 if (ops->gen_prologue || env->seen_direct_write) {
7231 if (!ops->gen_prologue) {
7232 verbose(env, "bpf verifier is misconfigured\n");
7235 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
7237 if (cnt >= ARRAY_SIZE(insn_buf)) {
7238 verbose(env, "bpf verifier is misconfigured\n");
7241 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
7245 env->prog = new_prog;
7250 if (bpf_prog_is_dev_bound(env->prog->aux))
7253 insn = env->prog->insnsi + delta;
7255 for (i = 0; i < insn_cnt; i++, insn++) {
7256 bpf_convert_ctx_access_t convert_ctx_access;
7258 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
7259 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
7260 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
7261 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
7263 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
7264 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
7265 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
7266 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
7271 if (type == BPF_WRITE &&
7272 env->insn_aux_data[i + delta].sanitize_stack_off) {
7273 struct bpf_insn patch[] = {
7274 /* Sanitize suspicious stack slot with zero.
7275 * There are no memory dependencies for this store,
7276 * since it's only using frame pointer and immediate
7279 BPF_ST_MEM(BPF_DW, BPF_REG_FP,
7280 env->insn_aux_data[i + delta].sanitize_stack_off,
7282 /* the original STX instruction will immediately
7283 * overwrite the same stack slot with appropriate value
7288 cnt = ARRAY_SIZE(patch);
7289 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
7294 env->prog = new_prog;
7295 insn = new_prog->insnsi + i + delta;
7299 switch (env->insn_aux_data[i + delta].ptr_type) {
7301 if (!ops->convert_ctx_access)
7303 convert_ctx_access = ops->convert_ctx_access;
7306 case PTR_TO_SOCK_COMMON:
7307 convert_ctx_access = bpf_sock_convert_ctx_access;
7309 case PTR_TO_TCP_SOCK:
7310 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
7316 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
7317 size = BPF_LDST_BYTES(insn);
7319 /* If the read access is a narrower load of the field,
7320 * convert to a 4/8-byte load, to minimum program type specific
7321 * convert_ctx_access changes. If conversion is successful,
7322 * we will apply proper mask to the result.
7324 is_narrower_load = size < ctx_field_size;
7325 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
7327 if (is_narrower_load) {
7330 if (type == BPF_WRITE) {
7331 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
7336 if (ctx_field_size == 4)
7338 else if (ctx_field_size == 8)
7341 insn->off = off & ~(size_default - 1);
7342 insn->code = BPF_LDX | BPF_MEM | size_code;
7346 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
7348 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
7349 (ctx_field_size && !target_size)) {
7350 verbose(env, "bpf verifier is misconfigured\n");
7354 if (is_narrower_load && size < target_size) {
7355 u8 shift = (off & (size_default - 1)) * 8;
7357 if (ctx_field_size <= 4) {
7359 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
7362 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
7363 (1 << size * 8) - 1);
7366 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
7369 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
7370 (1 << size * 8) - 1);
7374 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
7380 /* keep walking new program and skip insns we just inserted */
7381 env->prog = new_prog;
7382 insn = new_prog->insnsi + i + delta;
7388 static int jit_subprogs(struct bpf_verifier_env *env)
7390 struct bpf_prog *prog = env->prog, **func, *tmp;
7391 int i, j, subprog_start, subprog_end = 0, len, subprog;
7392 struct bpf_insn *insn;
7396 if (env->subprog_cnt <= 1)
7399 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
7400 if (insn->code != (BPF_JMP | BPF_CALL) ||
7401 insn->src_reg != BPF_PSEUDO_CALL)
7403 /* Upon error here we cannot fall back to interpreter but
7404 * need a hard reject of the program. Thus -EFAULT is
7405 * propagated in any case.
7407 subprog = find_subprog(env, i + insn->imm + 1);
7409 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
7413 /* temporarily remember subprog id inside insn instead of
7414 * aux_data, since next loop will split up all insns into funcs
7416 insn->off = subprog;
7417 /* remember original imm in case JIT fails and fallback
7418 * to interpreter will be needed
7420 env->insn_aux_data[i].call_imm = insn->imm;
7421 /* point imm to __bpf_call_base+1 from JITs point of view */
7425 err = bpf_prog_alloc_jited_linfo(prog);
7430 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
7434 for (i = 0; i < env->subprog_cnt; i++) {
7435 subprog_start = subprog_end;
7436 subprog_end = env->subprog_info[i + 1].start;
7438 len = subprog_end - subprog_start;
7439 /* BPF_PROG_RUN doesn't call subprogs directly,
7440 * hence main prog stats include the runtime of subprogs.
7441 * subprogs don't have IDs and not reachable via prog_get_next_id
7442 * func[i]->aux->stats will never be accessed and stays NULL
7444 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
7447 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
7448 len * sizeof(struct bpf_insn));
7449 func[i]->type = prog->type;
7451 if (bpf_prog_calc_tag(func[i]))
7453 func[i]->is_func = 1;
7454 func[i]->aux->func_idx = i;
7455 /* the btf and func_info will be freed only at prog->aux */
7456 func[i]->aux->btf = prog->aux->btf;
7457 func[i]->aux->func_info = prog->aux->func_info;
7459 /* Use bpf_prog_F_tag to indicate functions in stack traces.
7460 * Long term would need debug info to populate names
7462 func[i]->aux->name[0] = 'F';
7463 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
7464 func[i]->jit_requested = 1;
7465 func[i]->aux->linfo = prog->aux->linfo;
7466 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
7467 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
7468 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
7469 func[i] = bpf_int_jit_compile(func[i]);
7470 if (!func[i]->jited) {
7476 /* at this point all bpf functions were successfully JITed
7477 * now populate all bpf_calls with correct addresses and
7478 * run last pass of JIT
7480 for (i = 0; i < env->subprog_cnt; i++) {
7481 insn = func[i]->insnsi;
7482 for (j = 0; j < func[i]->len; j++, insn++) {
7483 if (insn->code != (BPF_JMP | BPF_CALL) ||
7484 insn->src_reg != BPF_PSEUDO_CALL)
7486 subprog = insn->off;
7487 insn->imm = (u64 (*)(u64, u64, u64, u64, u64))
7488 func[subprog]->bpf_func -
7492 /* we use the aux data to keep a list of the start addresses
7493 * of the JITed images for each function in the program
7495 * for some architectures, such as powerpc64, the imm field
7496 * might not be large enough to hold the offset of the start
7497 * address of the callee's JITed image from __bpf_call_base
7499 * in such cases, we can lookup the start address of a callee
7500 * by using its subprog id, available from the off field of
7501 * the call instruction, as an index for this list
7503 func[i]->aux->func = func;
7504 func[i]->aux->func_cnt = env->subprog_cnt;
7506 for (i = 0; i < env->subprog_cnt; i++) {
7507 old_bpf_func = func[i]->bpf_func;
7508 tmp = bpf_int_jit_compile(func[i]);
7509 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
7510 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
7517 /* finally lock prog and jit images for all functions and
7520 for (i = 0; i < env->subprog_cnt; i++) {
7521 bpf_prog_lock_ro(func[i]);
7522 bpf_prog_kallsyms_add(func[i]);
7525 /* Last step: make now unused interpreter insns from main
7526 * prog consistent for later dump requests, so they can
7527 * later look the same as if they were interpreted only.
7529 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
7530 if (insn->code != (BPF_JMP | BPF_CALL) ||
7531 insn->src_reg != BPF_PSEUDO_CALL)
7533 insn->off = env->insn_aux_data[i].call_imm;
7534 subprog = find_subprog(env, i + insn->off + 1);
7535 insn->imm = subprog;
7539 prog->bpf_func = func[0]->bpf_func;
7540 prog->aux->func = func;
7541 prog->aux->func_cnt = env->subprog_cnt;
7542 bpf_prog_free_unused_jited_linfo(prog);
7545 for (i = 0; i < env->subprog_cnt; i++)
7547 bpf_jit_free(func[i]);
7550 /* cleanup main prog to be interpreted */
7551 prog->jit_requested = 0;
7552 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
7553 if (insn->code != (BPF_JMP | BPF_CALL) ||
7554 insn->src_reg != BPF_PSEUDO_CALL)
7557 insn->imm = env->insn_aux_data[i].call_imm;
7559 bpf_prog_free_jited_linfo(prog);
7563 static int fixup_call_args(struct bpf_verifier_env *env)
7565 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
7566 struct bpf_prog *prog = env->prog;
7567 struct bpf_insn *insn = prog->insnsi;
7572 if (env->prog->jit_requested &&
7573 !bpf_prog_is_dev_bound(env->prog->aux)) {
7574 err = jit_subprogs(env);
7580 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
7581 for (i = 0; i < prog->len; i++, insn++) {
7582 if (insn->code != (BPF_JMP | BPF_CALL) ||
7583 insn->src_reg != BPF_PSEUDO_CALL)
7585 depth = get_callee_stack_depth(env, insn, i);
7588 bpf_patch_call_args(insn, depth);
7595 /* fixup insn->imm field of bpf_call instructions
7596 * and inline eligible helpers as explicit sequence of BPF instructions
7598 * this function is called after eBPF program passed verification
7600 static int fixup_bpf_calls(struct bpf_verifier_env *env)
7602 struct bpf_prog *prog = env->prog;
7603 struct bpf_insn *insn = prog->insnsi;
7604 const struct bpf_func_proto *fn;
7605 const int insn_cnt = prog->len;
7606 const struct bpf_map_ops *ops;
7607 struct bpf_insn_aux_data *aux;
7608 struct bpf_insn insn_buf[16];
7609 struct bpf_prog *new_prog;
7610 struct bpf_map *map_ptr;
7611 int i, cnt, delta = 0;
7613 for (i = 0; i < insn_cnt; i++, insn++) {
7614 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
7615 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
7616 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
7617 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
7618 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
7619 struct bpf_insn mask_and_div[] = {
7620 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
7622 BPF_JMP_IMM(BPF_JNE, insn->src_reg, 0, 2),
7623 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
7624 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
7627 struct bpf_insn mask_and_mod[] = {
7628 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
7629 /* Rx mod 0 -> Rx */
7630 BPF_JMP_IMM(BPF_JEQ, insn->src_reg, 0, 1),
7633 struct bpf_insn *patchlet;
7635 if (insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
7636 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
7637 patchlet = mask_and_div + (is64 ? 1 : 0);
7638 cnt = ARRAY_SIZE(mask_and_div) - (is64 ? 1 : 0);
7640 patchlet = mask_and_mod + (is64 ? 1 : 0);
7641 cnt = ARRAY_SIZE(mask_and_mod) - (is64 ? 1 : 0);
7644 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
7649 env->prog = prog = new_prog;
7650 insn = new_prog->insnsi + i + delta;
7654 if (BPF_CLASS(insn->code) == BPF_LD &&
7655 (BPF_MODE(insn->code) == BPF_ABS ||
7656 BPF_MODE(insn->code) == BPF_IND)) {
7657 cnt = env->ops->gen_ld_abs(insn, insn_buf);
7658 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
7659 verbose(env, "bpf verifier is misconfigured\n");
7663 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
7668 env->prog = prog = new_prog;
7669 insn = new_prog->insnsi + i + delta;
7673 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
7674 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
7675 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
7676 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
7677 struct bpf_insn insn_buf[16];
7678 struct bpf_insn *patch = &insn_buf[0];
7682 aux = &env->insn_aux_data[i + delta];
7683 if (!aux->alu_state ||
7684 aux->alu_state == BPF_ALU_NON_POINTER)
7687 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
7688 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
7689 BPF_ALU_SANITIZE_SRC;
7691 off_reg = issrc ? insn->src_reg : insn->dst_reg;
7693 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
7694 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit - 1);
7695 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
7696 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
7697 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
7698 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
7700 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX,
7702 insn->src_reg = BPF_REG_AX;
7704 *patch++ = BPF_ALU64_REG(BPF_AND, off_reg,
7708 insn->code = insn->code == code_add ?
7709 code_sub : code_add;
7712 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
7713 cnt = patch - insn_buf;
7715 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
7720 env->prog = prog = new_prog;
7721 insn = new_prog->insnsi + i + delta;
7725 if (insn->code != (BPF_JMP | BPF_CALL))
7727 if (insn->src_reg == BPF_PSEUDO_CALL)
7730 if (insn->imm == BPF_FUNC_get_route_realm)
7731 prog->dst_needed = 1;
7732 if (insn->imm == BPF_FUNC_get_prandom_u32)
7733 bpf_user_rnd_init_once();
7734 if (insn->imm == BPF_FUNC_override_return)
7735 prog->kprobe_override = 1;
7736 if (insn->imm == BPF_FUNC_tail_call) {
7737 /* If we tail call into other programs, we
7738 * cannot make any assumptions since they can
7739 * be replaced dynamically during runtime in
7740 * the program array.
7742 prog->cb_access = 1;
7743 env->prog->aux->stack_depth = MAX_BPF_STACK;
7744 env->prog->aux->max_pkt_offset = MAX_PACKET_OFF;
7746 /* mark bpf_tail_call as different opcode to avoid
7747 * conditional branch in the interpeter for every normal
7748 * call and to prevent accidental JITing by JIT compiler
7749 * that doesn't support bpf_tail_call yet
7752 insn->code = BPF_JMP | BPF_TAIL_CALL;
7754 aux = &env->insn_aux_data[i + delta];
7755 if (!bpf_map_ptr_unpriv(aux))
7758 /* instead of changing every JIT dealing with tail_call
7759 * emit two extra insns:
7760 * if (index >= max_entries) goto out;
7761 * index &= array->index_mask;
7762 * to avoid out-of-bounds cpu speculation
7764 if (bpf_map_ptr_poisoned(aux)) {
7765 verbose(env, "tail_call abusing map_ptr\n");
7769 map_ptr = BPF_MAP_PTR(aux->map_state);
7770 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
7771 map_ptr->max_entries, 2);
7772 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
7773 container_of(map_ptr,
7776 insn_buf[2] = *insn;
7778 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
7783 env->prog = prog = new_prog;
7784 insn = new_prog->insnsi + i + delta;
7788 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
7789 * and other inlining handlers are currently limited to 64 bit
7792 if (prog->jit_requested && BITS_PER_LONG == 64 &&
7793 (insn->imm == BPF_FUNC_map_lookup_elem ||
7794 insn->imm == BPF_FUNC_map_update_elem ||
7795 insn->imm == BPF_FUNC_map_delete_elem ||
7796 insn->imm == BPF_FUNC_map_push_elem ||
7797 insn->imm == BPF_FUNC_map_pop_elem ||
7798 insn->imm == BPF_FUNC_map_peek_elem)) {
7799 aux = &env->insn_aux_data[i + delta];
7800 if (bpf_map_ptr_poisoned(aux))
7801 goto patch_call_imm;
7803 map_ptr = BPF_MAP_PTR(aux->map_state);
7805 if (insn->imm == BPF_FUNC_map_lookup_elem &&
7806 ops->map_gen_lookup) {
7807 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
7808 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
7809 verbose(env, "bpf verifier is misconfigured\n");
7813 new_prog = bpf_patch_insn_data(env, i + delta,
7819 env->prog = prog = new_prog;
7820 insn = new_prog->insnsi + i + delta;
7824 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
7825 (void *(*)(struct bpf_map *map, void *key))NULL));
7826 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
7827 (int (*)(struct bpf_map *map, void *key))NULL));
7828 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
7829 (int (*)(struct bpf_map *map, void *key, void *value,
7831 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
7832 (int (*)(struct bpf_map *map, void *value,
7834 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
7835 (int (*)(struct bpf_map *map, void *value))NULL));
7836 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
7837 (int (*)(struct bpf_map *map, void *value))NULL));
7839 switch (insn->imm) {
7840 case BPF_FUNC_map_lookup_elem:
7841 insn->imm = BPF_CAST_CALL(ops->map_lookup_elem) -
7844 case BPF_FUNC_map_update_elem:
7845 insn->imm = BPF_CAST_CALL(ops->map_update_elem) -
7848 case BPF_FUNC_map_delete_elem:
7849 insn->imm = BPF_CAST_CALL(ops->map_delete_elem) -
7852 case BPF_FUNC_map_push_elem:
7853 insn->imm = BPF_CAST_CALL(ops->map_push_elem) -
7856 case BPF_FUNC_map_pop_elem:
7857 insn->imm = BPF_CAST_CALL(ops->map_pop_elem) -
7860 case BPF_FUNC_map_peek_elem:
7861 insn->imm = BPF_CAST_CALL(ops->map_peek_elem) -
7866 goto patch_call_imm;
7870 fn = env->ops->get_func_proto(insn->imm, env->prog);
7871 /* all functions that have prototype and verifier allowed
7872 * programs to call them, must be real in-kernel functions
7876 "kernel subsystem misconfigured func %s#%d\n",
7877 func_id_name(insn->imm), insn->imm);
7880 insn->imm = fn->func - __bpf_call_base;
7886 static void free_states(struct bpf_verifier_env *env)
7888 struct bpf_verifier_state_list *sl, *sln;
7891 sl = env->free_list;
7894 free_verifier_state(&sl->state, false);
7899 if (!env->explored_states)
7902 for (i = 0; i < env->prog->len; i++) {
7903 sl = env->explored_states[i];
7906 while (sl != STATE_LIST_MARK) {
7908 free_verifier_state(&sl->state, false);
7914 kvfree(env->explored_states);
7917 static void print_verification_stats(struct bpf_verifier_env *env)
7921 if (env->log.level & BPF_LOG_STATS) {
7922 verbose(env, "verification time %lld usec\n",
7923 div_u64(env->verification_time, 1000));
7924 verbose(env, "stack depth ");
7925 for (i = 0; i < env->subprog_cnt; i++) {
7926 u32 depth = env->subprog_info[i].stack_depth;
7928 verbose(env, "%d", depth);
7929 if (i + 1 < env->subprog_cnt)
7934 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
7935 "total_states %d peak_states %d mark_read %d\n",
7936 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
7937 env->max_states_per_insn, env->total_states,
7938 env->peak_states, env->longest_mark_read_walk);
7941 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr,
7942 union bpf_attr __user *uattr)
7944 u64 start_time = ktime_get_ns();
7945 struct bpf_verifier_env *env;
7946 struct bpf_verifier_log *log;
7947 int i, len, ret = -EINVAL;
7950 /* no program is valid */
7951 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
7954 /* 'struct bpf_verifier_env' can be global, but since it's not small,
7955 * allocate/free it every time bpf_check() is called
7957 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
7963 env->insn_aux_data =
7964 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
7966 if (!env->insn_aux_data)
7968 for (i = 0; i < len; i++)
7969 env->insn_aux_data[i].orig_idx = i;
7971 env->ops = bpf_verifier_ops[env->prog->type];
7973 /* grab the mutex to protect few globals used by verifier */
7974 mutex_lock(&bpf_verifier_lock);
7976 if (attr->log_level || attr->log_buf || attr->log_size) {
7977 /* user requested verbose verifier output
7978 * and supplied buffer to store the verification trace
7980 log->level = attr->log_level;
7981 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
7982 log->len_total = attr->log_size;
7985 /* log attributes have to be sane */
7986 if (log->len_total < 128 || log->len_total > UINT_MAX >> 2 ||
7987 !log->level || !log->ubuf || log->level & ~BPF_LOG_MASK)
7991 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
7992 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
7993 env->strict_alignment = true;
7994 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
7995 env->strict_alignment = false;
7997 is_priv = capable(CAP_SYS_ADMIN);
7998 env->allow_ptr_leaks = is_priv;
8000 ret = replace_map_fd_with_map_ptr(env);
8002 goto skip_full_check;
8004 if (bpf_prog_is_dev_bound(env->prog->aux)) {
8005 ret = bpf_prog_offload_verifier_prep(env->prog);
8007 goto skip_full_check;
8010 env->explored_states = kvcalloc(env->prog->len,
8011 sizeof(struct bpf_verifier_state_list *),
8014 if (!env->explored_states)
8015 goto skip_full_check;
8017 ret = check_subprogs(env);
8019 goto skip_full_check;
8021 ret = check_btf_info(env, attr, uattr);
8023 goto skip_full_check;
8025 ret = check_cfg(env);
8027 goto skip_full_check;
8029 ret = do_check(env);
8030 if (env->cur_state) {
8031 free_verifier_state(env->cur_state, true);
8032 env->cur_state = NULL;
8035 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
8036 ret = bpf_prog_offload_finalize(env);
8039 while (!pop_stack(env, NULL, NULL));
8043 ret = check_max_stack_depth(env);
8045 /* instruction rewrites happen after this point */
8048 opt_hard_wire_dead_code_branches(env);
8050 ret = opt_remove_dead_code(env);
8052 ret = opt_remove_nops(env);
8055 sanitize_dead_code(env);
8059 /* program is valid, convert *(u32*)(ctx + off) accesses */
8060 ret = convert_ctx_accesses(env);
8063 ret = fixup_bpf_calls(env);
8066 ret = fixup_call_args(env);
8068 env->verification_time = ktime_get_ns() - start_time;
8069 print_verification_stats(env);
8071 if (log->level && bpf_verifier_log_full(log))
8073 if (log->level && !log->ubuf) {
8075 goto err_release_maps;
8078 if (ret == 0 && env->used_map_cnt) {
8079 /* if program passed verifier, update used_maps in bpf_prog_info */
8080 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
8081 sizeof(env->used_maps[0]),
8084 if (!env->prog->aux->used_maps) {
8086 goto err_release_maps;
8089 memcpy(env->prog->aux->used_maps, env->used_maps,
8090 sizeof(env->used_maps[0]) * env->used_map_cnt);
8091 env->prog->aux->used_map_cnt = env->used_map_cnt;
8093 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
8094 * bpf_ld_imm64 instructions
8096 convert_pseudo_ld_imm64(env);
8100 adjust_btf_func(env);
8103 if (!env->prog->aux->used_maps)
8104 /* if we didn't copy map pointers into bpf_prog_info, release
8105 * them now. Otherwise free_used_maps() will release them.
8110 mutex_unlock(&bpf_verifier_lock);
8111 vfree(env->insn_aux_data);