1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
3 * Copyright (c) 2016 Facebook
4 * Copyright (c) 2018 Covalent IO, Inc. http://covalent.io
6 #include <uapi/linux/btf.h>
7 #include <linux/kernel.h>
8 #include <linux/types.h>
9 #include <linux/slab.h>
10 #include <linux/bpf.h>
11 #include <linux/btf.h>
12 #include <linux/bpf_verifier.h>
13 #include <linux/filter.h>
14 #include <net/netlink.h>
15 #include <linux/file.h>
16 #include <linux/vmalloc.h>
17 #include <linux/stringify.h>
18 #include <linux/bsearch.h>
19 #include <linux/sort.h>
20 #include <linux/perf_event.h>
21 #include <linux/ctype.h>
25 static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
26 #define BPF_PROG_TYPE(_id, _name, prog_ctx_type, kern_ctx_type) \
27 [_id] = & _name ## _verifier_ops,
28 #define BPF_MAP_TYPE(_id, _ops)
29 #include <linux/bpf_types.h>
34 /* bpf_check() is a static code analyzer that walks eBPF program
35 * instruction by instruction and updates register/stack state.
36 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
38 * The first pass is depth-first-search to check that the program is a DAG.
39 * It rejects the following programs:
40 * - larger than BPF_MAXINSNS insns
41 * - if loop is present (detected via back-edge)
42 * - unreachable insns exist (shouldn't be a forest. program = one function)
43 * - out of bounds or malformed jumps
44 * The second pass is all possible path descent from the 1st insn.
45 * Since it's analyzing all pathes through the program, the length of the
46 * analysis is limited to 64k insn, which may be hit even if total number of
47 * insn is less then 4K, but there are too many branches that change stack/regs.
48 * Number of 'branches to be analyzed' is limited to 1k
50 * On entry to each instruction, each register has a type, and the instruction
51 * changes the types of the registers depending on instruction semantics.
52 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
55 * All registers are 64-bit.
56 * R0 - return register
57 * R1-R5 argument passing registers
58 * R6-R9 callee saved registers
59 * R10 - frame pointer read-only
61 * At the start of BPF program the register R1 contains a pointer to bpf_context
62 * and has type PTR_TO_CTX.
64 * Verifier tracks arithmetic operations on pointers in case:
65 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
66 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
67 * 1st insn copies R10 (which has FRAME_PTR) type into R1
68 * and 2nd arithmetic instruction is pattern matched to recognize
69 * that it wants to construct a pointer to some element within stack.
70 * So after 2nd insn, the register R1 has type PTR_TO_STACK
71 * (and -20 constant is saved for further stack bounds checking).
72 * Meaning that this reg is a pointer to stack plus known immediate constant.
74 * Most of the time the registers have SCALAR_VALUE type, which
75 * means the register has some value, but it's not a valid pointer.
76 * (like pointer plus pointer becomes SCALAR_VALUE type)
78 * When verifier sees load or store instructions the type of base register
79 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK, PTR_TO_SOCKET. These are
80 * four pointer types recognized by check_mem_access() function.
82 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
83 * and the range of [ptr, ptr + map's value_size) is accessible.
85 * registers used to pass values to function calls are checked against
86 * function argument constraints.
88 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
89 * It means that the register type passed to this function must be
90 * PTR_TO_STACK and it will be used inside the function as
91 * 'pointer to map element key'
93 * For example the argument constraints for bpf_map_lookup_elem():
94 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
95 * .arg1_type = ARG_CONST_MAP_PTR,
96 * .arg2_type = ARG_PTR_TO_MAP_KEY,
98 * ret_type says that this function returns 'pointer to map elem value or null'
99 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
100 * 2nd argument should be a pointer to stack, which will be used inside
101 * the helper function as a pointer to map element key.
103 * On the kernel side the helper function looks like:
104 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
106 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
107 * void *key = (void *) (unsigned long) r2;
110 * here kernel can access 'key' and 'map' pointers safely, knowing that
111 * [key, key + map->key_size) bytes are valid and were initialized on
112 * the stack of eBPF program.
115 * Corresponding eBPF program may look like:
116 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
117 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
118 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
119 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
120 * here verifier looks at prototype of map_lookup_elem() and sees:
121 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
122 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
124 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
125 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
126 * and were initialized prior to this call.
127 * If it's ok, then verifier allows this BPF_CALL insn and looks at
128 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
129 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
130 * returns ether pointer to map value or NULL.
132 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
133 * insn, the register holding that pointer in the true branch changes state to
134 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
135 * branch. See check_cond_jmp_op().
137 * After the call R0 is set to return type of the function and registers R1-R5
138 * are set to NOT_INIT to indicate that they are no longer readable.
140 * The following reference types represent a potential reference to a kernel
141 * resource which, after first being allocated, must be checked and freed by
143 * - PTR_TO_SOCKET_OR_NULL, PTR_TO_SOCKET
145 * When the verifier sees a helper call return a reference type, it allocates a
146 * pointer id for the reference and stores it in the current function state.
147 * Similar to the way that PTR_TO_MAP_VALUE_OR_NULL is converted into
148 * PTR_TO_MAP_VALUE, PTR_TO_SOCKET_OR_NULL becomes PTR_TO_SOCKET when the type
149 * passes through a NULL-check conditional. For the branch wherein the state is
150 * changed to CONST_IMM, the verifier releases the reference.
152 * For each helper function that allocates a reference, such as
153 * bpf_sk_lookup_tcp(), there is a corresponding release function, such as
154 * bpf_sk_release(). When a reference type passes into the release function,
155 * the verifier also releases the reference. If any unchecked or unreleased
156 * reference remains at the end of the program, the verifier rejects it.
159 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
160 struct bpf_verifier_stack_elem {
161 /* verifer state is 'st'
162 * before processing instruction 'insn_idx'
163 * and after processing instruction 'prev_insn_idx'
165 struct bpf_verifier_state st;
168 struct bpf_verifier_stack_elem *next;
171 #define BPF_COMPLEXITY_LIMIT_JMP_SEQ 8192
172 #define BPF_COMPLEXITY_LIMIT_STATES 64
174 #define BPF_MAP_KEY_POISON (1ULL << 63)
175 #define BPF_MAP_KEY_SEEN (1ULL << 62)
177 #define BPF_MAP_PTR_UNPRIV 1UL
178 #define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \
179 POISON_POINTER_DELTA))
180 #define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
182 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
184 return BPF_MAP_PTR(aux->map_ptr_state) == BPF_MAP_PTR_POISON;
187 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
189 return aux->map_ptr_state & BPF_MAP_PTR_UNPRIV;
192 static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
193 const struct bpf_map *map, bool unpriv)
195 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON & BPF_MAP_PTR_UNPRIV);
196 unpriv |= bpf_map_ptr_unpriv(aux);
197 aux->map_ptr_state = (unsigned long)map |
198 (unpriv ? BPF_MAP_PTR_UNPRIV : 0UL);
201 static bool bpf_map_key_poisoned(const struct bpf_insn_aux_data *aux)
203 return aux->map_key_state & BPF_MAP_KEY_POISON;
206 static bool bpf_map_key_unseen(const struct bpf_insn_aux_data *aux)
208 return !(aux->map_key_state & BPF_MAP_KEY_SEEN);
211 static u64 bpf_map_key_immediate(const struct bpf_insn_aux_data *aux)
213 return aux->map_key_state & ~(BPF_MAP_KEY_SEEN | BPF_MAP_KEY_POISON);
216 static void bpf_map_key_store(struct bpf_insn_aux_data *aux, u64 state)
218 bool poisoned = bpf_map_key_poisoned(aux);
220 aux->map_key_state = state | BPF_MAP_KEY_SEEN |
221 (poisoned ? BPF_MAP_KEY_POISON : 0ULL);
224 struct bpf_call_arg_meta {
225 struct bpf_map *map_ptr;
230 s64 msize_smax_value;
231 u64 msize_umax_value;
237 struct btf *btf_vmlinux;
239 static DEFINE_MUTEX(bpf_verifier_lock);
241 static const struct bpf_line_info *
242 find_linfo(const struct bpf_verifier_env *env, u32 insn_off)
244 const struct bpf_line_info *linfo;
245 const struct bpf_prog *prog;
249 nr_linfo = prog->aux->nr_linfo;
251 if (!nr_linfo || insn_off >= prog->len)
254 linfo = prog->aux->linfo;
255 for (i = 1; i < nr_linfo; i++)
256 if (insn_off < linfo[i].insn_off)
259 return &linfo[i - 1];
262 void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
267 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
269 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
270 "verifier log line truncated - local buffer too short\n");
272 n = min(log->len_total - log->len_used - 1, n);
275 if (log->level == BPF_LOG_KERNEL) {
276 pr_err("BPF:%s\n", log->kbuf);
279 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
285 /* log_level controls verbosity level of eBPF verifier.
286 * bpf_verifier_log_write() is used to dump the verification trace to the log,
287 * so the user can figure out what's wrong with the program
289 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
290 const char *fmt, ...)
294 if (!bpf_verifier_log_needed(&env->log))
298 bpf_verifier_vlog(&env->log, fmt, args);
301 EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
303 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
305 struct bpf_verifier_env *env = private_data;
308 if (!bpf_verifier_log_needed(&env->log))
312 bpf_verifier_vlog(&env->log, fmt, args);
316 __printf(2, 3) void bpf_log(struct bpf_verifier_log *log,
317 const char *fmt, ...)
321 if (!bpf_verifier_log_needed(log))
325 bpf_verifier_vlog(log, fmt, args);
329 static const char *ltrim(const char *s)
337 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env,
339 const char *prefix_fmt, ...)
341 const struct bpf_line_info *linfo;
343 if (!bpf_verifier_log_needed(&env->log))
346 linfo = find_linfo(env, insn_off);
347 if (!linfo || linfo == env->prev_linfo)
353 va_start(args, prefix_fmt);
354 bpf_verifier_vlog(&env->log, prefix_fmt, args);
359 ltrim(btf_name_by_offset(env->prog->aux->btf,
362 env->prev_linfo = linfo;
365 static bool type_is_pkt_pointer(enum bpf_reg_type type)
367 return type == PTR_TO_PACKET ||
368 type == PTR_TO_PACKET_META;
371 static bool type_is_sk_pointer(enum bpf_reg_type type)
373 return type == PTR_TO_SOCKET ||
374 type == PTR_TO_SOCK_COMMON ||
375 type == PTR_TO_TCP_SOCK ||
376 type == PTR_TO_XDP_SOCK;
379 static bool reg_type_may_be_null(enum bpf_reg_type type)
381 return type == PTR_TO_MAP_VALUE_OR_NULL ||
382 type == PTR_TO_SOCKET_OR_NULL ||
383 type == PTR_TO_SOCK_COMMON_OR_NULL ||
384 type == PTR_TO_TCP_SOCK_OR_NULL;
387 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
389 return reg->type == PTR_TO_MAP_VALUE &&
390 map_value_has_spin_lock(reg->map_ptr);
393 static bool reg_type_may_be_refcounted_or_null(enum bpf_reg_type type)
395 return type == PTR_TO_SOCKET ||
396 type == PTR_TO_SOCKET_OR_NULL ||
397 type == PTR_TO_TCP_SOCK ||
398 type == PTR_TO_TCP_SOCK_OR_NULL;
401 static bool arg_type_may_be_refcounted(enum bpf_arg_type type)
403 return type == ARG_PTR_TO_SOCK_COMMON;
406 /* Determine whether the function releases some resources allocated by another
407 * function call. The first reference type argument will be assumed to be
408 * released by release_reference().
410 static bool is_release_function(enum bpf_func_id func_id)
412 return func_id == BPF_FUNC_sk_release;
415 static bool is_acquire_function(enum bpf_func_id func_id)
417 return func_id == BPF_FUNC_sk_lookup_tcp ||
418 func_id == BPF_FUNC_sk_lookup_udp ||
419 func_id == BPF_FUNC_skc_lookup_tcp;
422 static bool is_ptr_cast_function(enum bpf_func_id func_id)
424 return func_id == BPF_FUNC_tcp_sock ||
425 func_id == BPF_FUNC_sk_fullsock;
428 /* string representation of 'enum bpf_reg_type' */
429 static const char * const reg_type_str[] = {
431 [SCALAR_VALUE] = "inv",
432 [PTR_TO_CTX] = "ctx",
433 [CONST_PTR_TO_MAP] = "map_ptr",
434 [PTR_TO_MAP_VALUE] = "map_value",
435 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
436 [PTR_TO_STACK] = "fp",
437 [PTR_TO_PACKET] = "pkt",
438 [PTR_TO_PACKET_META] = "pkt_meta",
439 [PTR_TO_PACKET_END] = "pkt_end",
440 [PTR_TO_FLOW_KEYS] = "flow_keys",
441 [PTR_TO_SOCKET] = "sock",
442 [PTR_TO_SOCKET_OR_NULL] = "sock_or_null",
443 [PTR_TO_SOCK_COMMON] = "sock_common",
444 [PTR_TO_SOCK_COMMON_OR_NULL] = "sock_common_or_null",
445 [PTR_TO_TCP_SOCK] = "tcp_sock",
446 [PTR_TO_TCP_SOCK_OR_NULL] = "tcp_sock_or_null",
447 [PTR_TO_TP_BUFFER] = "tp_buffer",
448 [PTR_TO_XDP_SOCK] = "xdp_sock",
449 [PTR_TO_BTF_ID] = "ptr_",
452 static char slot_type_char[] = {
453 [STACK_INVALID] = '?',
459 static void print_liveness(struct bpf_verifier_env *env,
460 enum bpf_reg_liveness live)
462 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
464 if (live & REG_LIVE_READ)
466 if (live & REG_LIVE_WRITTEN)
468 if (live & REG_LIVE_DONE)
472 static struct bpf_func_state *func(struct bpf_verifier_env *env,
473 const struct bpf_reg_state *reg)
475 struct bpf_verifier_state *cur = env->cur_state;
477 return cur->frame[reg->frameno];
480 const char *kernel_type_name(u32 id)
482 return btf_name_by_offset(btf_vmlinux,
483 btf_type_by_id(btf_vmlinux, id)->name_off);
486 static void print_verifier_state(struct bpf_verifier_env *env,
487 const struct bpf_func_state *state)
489 const struct bpf_reg_state *reg;
494 verbose(env, " frame%d:", state->frameno);
495 for (i = 0; i < MAX_BPF_REG; i++) {
496 reg = &state->regs[i];
500 verbose(env, " R%d", i);
501 print_liveness(env, reg->live);
502 verbose(env, "=%s", reg_type_str[t]);
503 if (t == SCALAR_VALUE && reg->precise)
505 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
506 tnum_is_const(reg->var_off)) {
507 /* reg->off should be 0 for SCALAR_VALUE */
508 verbose(env, "%lld", reg->var_off.value + reg->off);
510 if (t == PTR_TO_BTF_ID)
511 verbose(env, "%s", kernel_type_name(reg->btf_id));
512 verbose(env, "(id=%d", reg->id);
513 if (reg_type_may_be_refcounted_or_null(t))
514 verbose(env, ",ref_obj_id=%d", reg->ref_obj_id);
515 if (t != SCALAR_VALUE)
516 verbose(env, ",off=%d", reg->off);
517 if (type_is_pkt_pointer(t))
518 verbose(env, ",r=%d", reg->range);
519 else if (t == CONST_PTR_TO_MAP ||
520 t == PTR_TO_MAP_VALUE ||
521 t == PTR_TO_MAP_VALUE_OR_NULL)
522 verbose(env, ",ks=%d,vs=%d",
523 reg->map_ptr->key_size,
524 reg->map_ptr->value_size);
525 if (tnum_is_const(reg->var_off)) {
526 /* Typically an immediate SCALAR_VALUE, but
527 * could be a pointer whose offset is too big
530 verbose(env, ",imm=%llx", reg->var_off.value);
532 if (reg->smin_value != reg->umin_value &&
533 reg->smin_value != S64_MIN)
534 verbose(env, ",smin_value=%lld",
535 (long long)reg->smin_value);
536 if (reg->smax_value != reg->umax_value &&
537 reg->smax_value != S64_MAX)
538 verbose(env, ",smax_value=%lld",
539 (long long)reg->smax_value);
540 if (reg->umin_value != 0)
541 verbose(env, ",umin_value=%llu",
542 (unsigned long long)reg->umin_value);
543 if (reg->umax_value != U64_MAX)
544 verbose(env, ",umax_value=%llu",
545 (unsigned long long)reg->umax_value);
546 if (!tnum_is_unknown(reg->var_off)) {
549 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
550 verbose(env, ",var_off=%s", tn_buf);
556 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
557 char types_buf[BPF_REG_SIZE + 1];
561 for (j = 0; j < BPF_REG_SIZE; j++) {
562 if (state->stack[i].slot_type[j] != STACK_INVALID)
564 types_buf[j] = slot_type_char[
565 state->stack[i].slot_type[j]];
567 types_buf[BPF_REG_SIZE] = 0;
570 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
571 print_liveness(env, state->stack[i].spilled_ptr.live);
572 if (state->stack[i].slot_type[0] == STACK_SPILL) {
573 reg = &state->stack[i].spilled_ptr;
575 verbose(env, "=%s", reg_type_str[t]);
576 if (t == SCALAR_VALUE && reg->precise)
578 if (t == SCALAR_VALUE && tnum_is_const(reg->var_off))
579 verbose(env, "%lld", reg->var_off.value + reg->off);
581 verbose(env, "=%s", types_buf);
584 if (state->acquired_refs && state->refs[0].id) {
585 verbose(env, " refs=%d", state->refs[0].id);
586 for (i = 1; i < state->acquired_refs; i++)
587 if (state->refs[i].id)
588 verbose(env, ",%d", state->refs[i].id);
593 #define COPY_STATE_FN(NAME, COUNT, FIELD, SIZE) \
594 static int copy_##NAME##_state(struct bpf_func_state *dst, \
595 const struct bpf_func_state *src) \
599 if (WARN_ON_ONCE(dst->COUNT < src->COUNT)) { \
600 /* internal bug, make state invalid to reject the program */ \
601 memset(dst, 0, sizeof(*dst)); \
604 memcpy(dst->FIELD, src->FIELD, \
605 sizeof(*src->FIELD) * (src->COUNT / SIZE)); \
608 /* copy_reference_state() */
609 COPY_STATE_FN(reference, acquired_refs, refs, 1)
610 /* copy_stack_state() */
611 COPY_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
614 #define REALLOC_STATE_FN(NAME, COUNT, FIELD, SIZE) \
615 static int realloc_##NAME##_state(struct bpf_func_state *state, int size, \
618 u32 old_size = state->COUNT; \
619 struct bpf_##NAME##_state *new_##FIELD; \
620 int slot = size / SIZE; \
622 if (size <= old_size || !size) { \
625 state->COUNT = slot * SIZE; \
626 if (!size && old_size) { \
627 kfree(state->FIELD); \
628 state->FIELD = NULL; \
632 new_##FIELD = kmalloc_array(slot, sizeof(struct bpf_##NAME##_state), \
638 memcpy(new_##FIELD, state->FIELD, \
639 sizeof(*new_##FIELD) * (old_size / SIZE)); \
640 memset(new_##FIELD + old_size / SIZE, 0, \
641 sizeof(*new_##FIELD) * (size - old_size) / SIZE); \
643 state->COUNT = slot * SIZE; \
644 kfree(state->FIELD); \
645 state->FIELD = new_##FIELD; \
648 /* realloc_reference_state() */
649 REALLOC_STATE_FN(reference, acquired_refs, refs, 1)
650 /* realloc_stack_state() */
651 REALLOC_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
652 #undef REALLOC_STATE_FN
654 /* do_check() starts with zero-sized stack in struct bpf_verifier_state to
655 * make it consume minimal amount of memory. check_stack_write() access from
656 * the program calls into realloc_func_state() to grow the stack size.
657 * Note there is a non-zero 'parent' pointer inside bpf_verifier_state
658 * which realloc_stack_state() copies over. It points to previous
659 * bpf_verifier_state which is never reallocated.
661 static int realloc_func_state(struct bpf_func_state *state, int stack_size,
662 int refs_size, bool copy_old)
664 int err = realloc_reference_state(state, refs_size, copy_old);
667 return realloc_stack_state(state, stack_size, copy_old);
670 /* Acquire a pointer id from the env and update the state->refs to include
671 * this new pointer reference.
672 * On success, returns a valid pointer id to associate with the register
673 * On failure, returns a negative errno.
675 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
677 struct bpf_func_state *state = cur_func(env);
678 int new_ofs = state->acquired_refs;
681 err = realloc_reference_state(state, state->acquired_refs + 1, true);
685 state->refs[new_ofs].id = id;
686 state->refs[new_ofs].insn_idx = insn_idx;
691 /* release function corresponding to acquire_reference_state(). Idempotent. */
692 static int release_reference_state(struct bpf_func_state *state, int ptr_id)
696 last_idx = state->acquired_refs - 1;
697 for (i = 0; i < state->acquired_refs; i++) {
698 if (state->refs[i].id == ptr_id) {
699 if (last_idx && i != last_idx)
700 memcpy(&state->refs[i], &state->refs[last_idx],
701 sizeof(*state->refs));
702 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
703 state->acquired_refs--;
710 static int transfer_reference_state(struct bpf_func_state *dst,
711 struct bpf_func_state *src)
713 int err = realloc_reference_state(dst, src->acquired_refs, false);
716 err = copy_reference_state(dst, src);
722 static void free_func_state(struct bpf_func_state *state)
731 static void clear_jmp_history(struct bpf_verifier_state *state)
733 kfree(state->jmp_history);
734 state->jmp_history = NULL;
735 state->jmp_history_cnt = 0;
738 static void free_verifier_state(struct bpf_verifier_state *state,
743 for (i = 0; i <= state->curframe; i++) {
744 free_func_state(state->frame[i]);
745 state->frame[i] = NULL;
747 clear_jmp_history(state);
752 /* copy verifier state from src to dst growing dst stack space
753 * when necessary to accommodate larger src stack
755 static int copy_func_state(struct bpf_func_state *dst,
756 const struct bpf_func_state *src)
760 err = realloc_func_state(dst, src->allocated_stack, src->acquired_refs,
764 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
765 err = copy_reference_state(dst, src);
768 return copy_stack_state(dst, src);
771 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
772 const struct bpf_verifier_state *src)
774 struct bpf_func_state *dst;
775 u32 jmp_sz = sizeof(struct bpf_idx_pair) * src->jmp_history_cnt;
778 if (dst_state->jmp_history_cnt < src->jmp_history_cnt) {
779 kfree(dst_state->jmp_history);
780 dst_state->jmp_history = kmalloc(jmp_sz, GFP_USER);
781 if (!dst_state->jmp_history)
784 memcpy(dst_state->jmp_history, src->jmp_history, jmp_sz);
785 dst_state->jmp_history_cnt = src->jmp_history_cnt;
787 /* if dst has more stack frames then src frame, free them */
788 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
789 free_func_state(dst_state->frame[i]);
790 dst_state->frame[i] = NULL;
792 dst_state->speculative = src->speculative;
793 dst_state->curframe = src->curframe;
794 dst_state->active_spin_lock = src->active_spin_lock;
795 dst_state->branches = src->branches;
796 dst_state->parent = src->parent;
797 dst_state->first_insn_idx = src->first_insn_idx;
798 dst_state->last_insn_idx = src->last_insn_idx;
799 for (i = 0; i <= src->curframe; i++) {
800 dst = dst_state->frame[i];
802 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
805 dst_state->frame[i] = dst;
807 err = copy_func_state(dst, src->frame[i]);
814 static void update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
817 u32 br = --st->branches;
819 /* WARN_ON(br > 1) technically makes sense here,
820 * but see comment in push_stack(), hence:
822 WARN_ONCE((int)br < 0,
823 "BUG update_branch_counts:branches_to_explore=%d\n",
831 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
834 struct bpf_verifier_state *cur = env->cur_state;
835 struct bpf_verifier_stack_elem *elem, *head = env->head;
838 if (env->head == NULL)
842 err = copy_verifier_state(cur, &head->st);
847 *insn_idx = head->insn_idx;
849 *prev_insn_idx = head->prev_insn_idx;
851 free_verifier_state(&head->st, false);
858 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
859 int insn_idx, int prev_insn_idx,
862 struct bpf_verifier_state *cur = env->cur_state;
863 struct bpf_verifier_stack_elem *elem;
866 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
870 elem->insn_idx = insn_idx;
871 elem->prev_insn_idx = prev_insn_idx;
872 elem->next = env->head;
875 err = copy_verifier_state(&elem->st, cur);
878 elem->st.speculative |= speculative;
879 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
880 verbose(env, "The sequence of %d jumps is too complex.\n",
884 if (elem->st.parent) {
885 ++elem->st.parent->branches;
886 /* WARN_ON(branches > 2) technically makes sense here,
888 * 1. speculative states will bump 'branches' for non-branch
890 * 2. is_state_visited() heuristics may decide not to create
891 * a new state for a sequence of branches and all such current
892 * and cloned states will be pointing to a single parent state
893 * which might have large 'branches' count.
898 free_verifier_state(env->cur_state, true);
899 env->cur_state = NULL;
900 /* pop all elements and return */
901 while (!pop_stack(env, NULL, NULL));
905 #define CALLER_SAVED_REGS 6
906 static const int caller_saved[CALLER_SAVED_REGS] = {
907 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
910 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
911 struct bpf_reg_state *reg);
913 /* Mark the unknown part of a register (variable offset or scalar value) as
914 * known to have the value @imm.
916 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
918 /* Clear id, off, and union(map_ptr, range) */
919 memset(((u8 *)reg) + sizeof(reg->type), 0,
920 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
921 reg->var_off = tnum_const(imm);
922 reg->smin_value = (s64)imm;
923 reg->smax_value = (s64)imm;
924 reg->umin_value = imm;
925 reg->umax_value = imm;
928 /* Mark the 'variable offset' part of a register as zero. This should be
929 * used only on registers holding a pointer type.
931 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
933 __mark_reg_known(reg, 0);
936 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
938 __mark_reg_known(reg, 0);
939 reg->type = SCALAR_VALUE;
942 static void mark_reg_known_zero(struct bpf_verifier_env *env,
943 struct bpf_reg_state *regs, u32 regno)
945 if (WARN_ON(regno >= MAX_BPF_REG)) {
946 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
947 /* Something bad happened, let's kill all regs */
948 for (regno = 0; regno < MAX_BPF_REG; regno++)
949 __mark_reg_not_init(env, regs + regno);
952 __mark_reg_known_zero(regs + regno);
955 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
957 return type_is_pkt_pointer(reg->type);
960 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
962 return reg_is_pkt_pointer(reg) ||
963 reg->type == PTR_TO_PACKET_END;
966 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
967 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
968 enum bpf_reg_type which)
970 /* The register can already have a range from prior markings.
971 * This is fine as long as it hasn't been advanced from its
974 return reg->type == which &&
977 tnum_equals_const(reg->var_off, 0);
980 /* Attempts to improve min/max values based on var_off information */
981 static void __update_reg_bounds(struct bpf_reg_state *reg)
983 /* min signed is max(sign bit) | min(other bits) */
984 reg->smin_value = max_t(s64, reg->smin_value,
985 reg->var_off.value | (reg->var_off.mask & S64_MIN));
986 /* max signed is min(sign bit) | max(other bits) */
987 reg->smax_value = min_t(s64, reg->smax_value,
988 reg->var_off.value | (reg->var_off.mask & S64_MAX));
989 reg->umin_value = max(reg->umin_value, reg->var_off.value);
990 reg->umax_value = min(reg->umax_value,
991 reg->var_off.value | reg->var_off.mask);
994 /* Uses signed min/max values to inform unsigned, and vice-versa */
995 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
997 /* Learn sign from signed bounds.
998 * If we cannot cross the sign boundary, then signed and unsigned bounds
999 * are the same, so combine. This works even in the negative case, e.g.
1000 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1002 if (reg->smin_value >= 0 || reg->smax_value < 0) {
1003 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1005 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1009 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1010 * boundary, so we must be careful.
1012 if ((s64)reg->umax_value >= 0) {
1013 /* Positive. We can't learn anything from the smin, but smax
1014 * is positive, hence safe.
1016 reg->smin_value = reg->umin_value;
1017 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1019 } else if ((s64)reg->umin_value < 0) {
1020 /* Negative. We can't learn anything from the smax, but smin
1021 * is negative, hence safe.
1023 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1025 reg->smax_value = reg->umax_value;
1029 /* Attempts to improve var_off based on unsigned min/max information */
1030 static void __reg_bound_offset(struct bpf_reg_state *reg)
1032 reg->var_off = tnum_intersect(reg->var_off,
1033 tnum_range(reg->umin_value,
1037 static void __reg_bound_offset32(struct bpf_reg_state *reg)
1039 u64 mask = 0xffffFFFF;
1040 struct tnum range = tnum_range(reg->umin_value & mask,
1041 reg->umax_value & mask);
1042 struct tnum lo32 = tnum_cast(reg->var_off, 4);
1043 struct tnum hi32 = tnum_lshift(tnum_rshift(reg->var_off, 32), 32);
1045 reg->var_off = tnum_or(hi32, tnum_intersect(lo32, range));
1048 /* Reset the min/max bounds of a register */
1049 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
1051 reg->smin_value = S64_MIN;
1052 reg->smax_value = S64_MAX;
1053 reg->umin_value = 0;
1054 reg->umax_value = U64_MAX;
1057 /* Mark a register as having a completely unknown (scalar) value. */
1058 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
1059 struct bpf_reg_state *reg)
1062 * Clear type, id, off, and union(map_ptr, range) and
1063 * padding between 'type' and union
1065 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
1066 reg->type = SCALAR_VALUE;
1067 reg->var_off = tnum_unknown;
1069 reg->precise = env->subprog_cnt > 1 || !env->allow_ptr_leaks ?
1071 __mark_reg_unbounded(reg);
1074 static void mark_reg_unknown(struct bpf_verifier_env *env,
1075 struct bpf_reg_state *regs, u32 regno)
1077 if (WARN_ON(regno >= MAX_BPF_REG)) {
1078 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
1079 /* Something bad happened, let's kill all regs except FP */
1080 for (regno = 0; regno < BPF_REG_FP; regno++)
1081 __mark_reg_not_init(env, regs + regno);
1084 __mark_reg_unknown(env, regs + regno);
1087 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1088 struct bpf_reg_state *reg)
1090 __mark_reg_unknown(env, reg);
1091 reg->type = NOT_INIT;
1094 static void mark_reg_not_init(struct bpf_verifier_env *env,
1095 struct bpf_reg_state *regs, u32 regno)
1097 if (WARN_ON(regno >= MAX_BPF_REG)) {
1098 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
1099 /* Something bad happened, let's kill all regs except FP */
1100 for (regno = 0; regno < BPF_REG_FP; regno++)
1101 __mark_reg_not_init(env, regs + regno);
1104 __mark_reg_not_init(env, regs + regno);
1107 #define DEF_NOT_SUBREG (0)
1108 static void init_reg_state(struct bpf_verifier_env *env,
1109 struct bpf_func_state *state)
1111 struct bpf_reg_state *regs = state->regs;
1114 for (i = 0; i < MAX_BPF_REG; i++) {
1115 mark_reg_not_init(env, regs, i);
1116 regs[i].live = REG_LIVE_NONE;
1117 regs[i].parent = NULL;
1118 regs[i].subreg_def = DEF_NOT_SUBREG;
1122 regs[BPF_REG_FP].type = PTR_TO_STACK;
1123 mark_reg_known_zero(env, regs, BPF_REG_FP);
1124 regs[BPF_REG_FP].frameno = state->frameno;
1127 #define BPF_MAIN_FUNC (-1)
1128 static void init_func_state(struct bpf_verifier_env *env,
1129 struct bpf_func_state *state,
1130 int callsite, int frameno, int subprogno)
1132 state->callsite = callsite;
1133 state->frameno = frameno;
1134 state->subprogno = subprogno;
1135 init_reg_state(env, state);
1139 SRC_OP, /* register is used as source operand */
1140 DST_OP, /* register is used as destination operand */
1141 DST_OP_NO_MARK /* same as above, check only, don't mark */
1144 static int cmp_subprogs(const void *a, const void *b)
1146 return ((struct bpf_subprog_info *)a)->start -
1147 ((struct bpf_subprog_info *)b)->start;
1150 static int find_subprog(struct bpf_verifier_env *env, int off)
1152 struct bpf_subprog_info *p;
1154 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
1155 sizeof(env->subprog_info[0]), cmp_subprogs);
1158 return p - env->subprog_info;
1162 static int add_subprog(struct bpf_verifier_env *env, int off)
1164 int insn_cnt = env->prog->len;
1167 if (off >= insn_cnt || off < 0) {
1168 verbose(env, "call to invalid destination\n");
1171 ret = find_subprog(env, off);
1174 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
1175 verbose(env, "too many subprograms\n");
1178 env->subprog_info[env->subprog_cnt++].start = off;
1179 sort(env->subprog_info, env->subprog_cnt,
1180 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
1184 static int check_subprogs(struct bpf_verifier_env *env)
1186 int i, ret, subprog_start, subprog_end, off, cur_subprog = 0;
1187 struct bpf_subprog_info *subprog = env->subprog_info;
1188 struct bpf_insn *insn = env->prog->insnsi;
1189 int insn_cnt = env->prog->len;
1191 /* Add entry function. */
1192 ret = add_subprog(env, 0);
1196 /* determine subprog starts. The end is one before the next starts */
1197 for (i = 0; i < insn_cnt; i++) {
1198 if (insn[i].code != (BPF_JMP | BPF_CALL))
1200 if (insn[i].src_reg != BPF_PSEUDO_CALL)
1202 if (!env->allow_ptr_leaks) {
1203 verbose(env, "function calls to other bpf functions are allowed for root only\n");
1206 ret = add_subprog(env, i + insn[i].imm + 1);
1211 /* Add a fake 'exit' subprog which could simplify subprog iteration
1212 * logic. 'subprog_cnt' should not be increased.
1214 subprog[env->subprog_cnt].start = insn_cnt;
1216 if (env->log.level & BPF_LOG_LEVEL2)
1217 for (i = 0; i < env->subprog_cnt; i++)
1218 verbose(env, "func#%d @%d\n", i, subprog[i].start);
1220 /* now check that all jumps are within the same subprog */
1221 subprog_start = subprog[cur_subprog].start;
1222 subprog_end = subprog[cur_subprog + 1].start;
1223 for (i = 0; i < insn_cnt; i++) {
1224 u8 code = insn[i].code;
1226 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
1228 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
1230 off = i + insn[i].off + 1;
1231 if (off < subprog_start || off >= subprog_end) {
1232 verbose(env, "jump out of range from insn %d to %d\n", i, off);
1236 if (i == subprog_end - 1) {
1237 /* to avoid fall-through from one subprog into another
1238 * the last insn of the subprog should be either exit
1239 * or unconditional jump back
1241 if (code != (BPF_JMP | BPF_EXIT) &&
1242 code != (BPF_JMP | BPF_JA)) {
1243 verbose(env, "last insn is not an exit or jmp\n");
1246 subprog_start = subprog_end;
1248 if (cur_subprog < env->subprog_cnt)
1249 subprog_end = subprog[cur_subprog + 1].start;
1255 /* Parentage chain of this register (or stack slot) should take care of all
1256 * issues like callee-saved registers, stack slot allocation time, etc.
1258 static int mark_reg_read(struct bpf_verifier_env *env,
1259 const struct bpf_reg_state *state,
1260 struct bpf_reg_state *parent, u8 flag)
1262 bool writes = parent == state->parent; /* Observe write marks */
1266 /* if read wasn't screened by an earlier write ... */
1267 if (writes && state->live & REG_LIVE_WRITTEN)
1269 if (parent->live & REG_LIVE_DONE) {
1270 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
1271 reg_type_str[parent->type],
1272 parent->var_off.value, parent->off);
1275 /* The first condition is more likely to be true than the
1276 * second, checked it first.
1278 if ((parent->live & REG_LIVE_READ) == flag ||
1279 parent->live & REG_LIVE_READ64)
1280 /* The parentage chain never changes and
1281 * this parent was already marked as LIVE_READ.
1282 * There is no need to keep walking the chain again and
1283 * keep re-marking all parents as LIVE_READ.
1284 * This case happens when the same register is read
1285 * multiple times without writes into it in-between.
1286 * Also, if parent has the stronger REG_LIVE_READ64 set,
1287 * then no need to set the weak REG_LIVE_READ32.
1290 /* ... then we depend on parent's value */
1291 parent->live |= flag;
1292 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
1293 if (flag == REG_LIVE_READ64)
1294 parent->live &= ~REG_LIVE_READ32;
1296 parent = state->parent;
1301 if (env->longest_mark_read_walk < cnt)
1302 env->longest_mark_read_walk = cnt;
1306 /* This function is supposed to be used by the following 32-bit optimization
1307 * code only. It returns TRUE if the source or destination register operates
1308 * on 64-bit, otherwise return FALSE.
1310 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
1311 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
1316 class = BPF_CLASS(code);
1318 if (class == BPF_JMP) {
1319 /* BPF_EXIT for "main" will reach here. Return TRUE
1324 if (op == BPF_CALL) {
1325 /* BPF to BPF call will reach here because of marking
1326 * caller saved clobber with DST_OP_NO_MARK for which we
1327 * don't care the register def because they are anyway
1328 * marked as NOT_INIT already.
1330 if (insn->src_reg == BPF_PSEUDO_CALL)
1332 /* Helper call will reach here because of arg type
1333 * check, conservatively return TRUE.
1342 if (class == BPF_ALU64 || class == BPF_JMP ||
1343 /* BPF_END always use BPF_ALU class. */
1344 (class == BPF_ALU && op == BPF_END && insn->imm == 64))
1347 if (class == BPF_ALU || class == BPF_JMP32)
1350 if (class == BPF_LDX) {
1352 return BPF_SIZE(code) == BPF_DW;
1353 /* LDX source must be ptr. */
1357 if (class == BPF_STX) {
1358 if (reg->type != SCALAR_VALUE)
1360 return BPF_SIZE(code) == BPF_DW;
1363 if (class == BPF_LD) {
1364 u8 mode = BPF_MODE(code);
1367 if (mode == BPF_IMM)
1370 /* Both LD_IND and LD_ABS return 32-bit data. */
1374 /* Implicit ctx ptr. */
1375 if (regno == BPF_REG_6)
1378 /* Explicit source could be any width. */
1382 if (class == BPF_ST)
1383 /* The only source register for BPF_ST is a ptr. */
1386 /* Conservatively return true at default. */
1390 /* Return TRUE if INSN doesn't have explicit value define. */
1391 static bool insn_no_def(struct bpf_insn *insn)
1393 u8 class = BPF_CLASS(insn->code);
1395 return (class == BPF_JMP || class == BPF_JMP32 ||
1396 class == BPF_STX || class == BPF_ST);
1399 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
1400 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
1402 if (insn_no_def(insn))
1405 return !is_reg64(env, insn, insn->dst_reg, NULL, DST_OP);
1408 static void mark_insn_zext(struct bpf_verifier_env *env,
1409 struct bpf_reg_state *reg)
1411 s32 def_idx = reg->subreg_def;
1413 if (def_idx == DEF_NOT_SUBREG)
1416 env->insn_aux_data[def_idx - 1].zext_dst = true;
1417 /* The dst will be zero extended, so won't be sub-register anymore. */
1418 reg->subreg_def = DEF_NOT_SUBREG;
1421 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
1422 enum reg_arg_type t)
1424 struct bpf_verifier_state *vstate = env->cur_state;
1425 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1426 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
1427 struct bpf_reg_state *reg, *regs = state->regs;
1430 if (regno >= MAX_BPF_REG) {
1431 verbose(env, "R%d is invalid\n", regno);
1436 rw64 = is_reg64(env, insn, regno, reg, t);
1438 /* check whether register used as source operand can be read */
1439 if (reg->type == NOT_INIT) {
1440 verbose(env, "R%d !read_ok\n", regno);
1443 /* We don't need to worry about FP liveness because it's read-only */
1444 if (regno == BPF_REG_FP)
1448 mark_insn_zext(env, reg);
1450 return mark_reg_read(env, reg, reg->parent,
1451 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
1453 /* check whether register used as dest operand can be written to */
1454 if (regno == BPF_REG_FP) {
1455 verbose(env, "frame pointer is read only\n");
1458 reg->live |= REG_LIVE_WRITTEN;
1459 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
1461 mark_reg_unknown(env, regs, regno);
1466 /* for any branch, call, exit record the history of jmps in the given state */
1467 static int push_jmp_history(struct bpf_verifier_env *env,
1468 struct bpf_verifier_state *cur)
1470 u32 cnt = cur->jmp_history_cnt;
1471 struct bpf_idx_pair *p;
1474 p = krealloc(cur->jmp_history, cnt * sizeof(*p), GFP_USER);
1477 p[cnt - 1].idx = env->insn_idx;
1478 p[cnt - 1].prev_idx = env->prev_insn_idx;
1479 cur->jmp_history = p;
1480 cur->jmp_history_cnt = cnt;
1484 /* Backtrack one insn at a time. If idx is not at the top of recorded
1485 * history then previous instruction came from straight line execution.
1487 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
1492 if (cnt && st->jmp_history[cnt - 1].idx == i) {
1493 i = st->jmp_history[cnt - 1].prev_idx;
1501 /* For given verifier state backtrack_insn() is called from the last insn to
1502 * the first insn. Its purpose is to compute a bitmask of registers and
1503 * stack slots that needs precision in the parent verifier state.
1505 static int backtrack_insn(struct bpf_verifier_env *env, int idx,
1506 u32 *reg_mask, u64 *stack_mask)
1508 const struct bpf_insn_cbs cbs = {
1509 .cb_print = verbose,
1510 .private_data = env,
1512 struct bpf_insn *insn = env->prog->insnsi + idx;
1513 u8 class = BPF_CLASS(insn->code);
1514 u8 opcode = BPF_OP(insn->code);
1515 u8 mode = BPF_MODE(insn->code);
1516 u32 dreg = 1u << insn->dst_reg;
1517 u32 sreg = 1u << insn->src_reg;
1520 if (insn->code == 0)
1522 if (env->log.level & BPF_LOG_LEVEL) {
1523 verbose(env, "regs=%x stack=%llx before ", *reg_mask, *stack_mask);
1524 verbose(env, "%d: ", idx);
1525 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
1528 if (class == BPF_ALU || class == BPF_ALU64) {
1529 if (!(*reg_mask & dreg))
1531 if (opcode == BPF_MOV) {
1532 if (BPF_SRC(insn->code) == BPF_X) {
1534 * dreg needs precision after this insn
1535 * sreg needs precision before this insn
1541 * dreg needs precision after this insn.
1542 * Corresponding register is already marked
1543 * as precise=true in this verifier state.
1544 * No further markings in parent are necessary
1549 if (BPF_SRC(insn->code) == BPF_X) {
1551 * both dreg and sreg need precision
1556 * dreg still needs precision before this insn
1559 } else if (class == BPF_LDX) {
1560 if (!(*reg_mask & dreg))
1564 /* scalars can only be spilled into stack w/o losing precision.
1565 * Load from any other memory can be zero extended.
1566 * The desire to keep that precision is already indicated
1567 * by 'precise' mark in corresponding register of this state.
1568 * No further tracking necessary.
1570 if (insn->src_reg != BPF_REG_FP)
1572 if (BPF_SIZE(insn->code) != BPF_DW)
1575 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
1576 * that [fp - off] slot contains scalar that needs to be
1577 * tracked with precision
1579 spi = (-insn->off - 1) / BPF_REG_SIZE;
1581 verbose(env, "BUG spi %d\n", spi);
1582 WARN_ONCE(1, "verifier backtracking bug");
1585 *stack_mask |= 1ull << spi;
1586 } else if (class == BPF_STX || class == BPF_ST) {
1587 if (*reg_mask & dreg)
1588 /* stx & st shouldn't be using _scalar_ dst_reg
1589 * to access memory. It means backtracking
1590 * encountered a case of pointer subtraction.
1593 /* scalars can only be spilled into stack */
1594 if (insn->dst_reg != BPF_REG_FP)
1596 if (BPF_SIZE(insn->code) != BPF_DW)
1598 spi = (-insn->off - 1) / BPF_REG_SIZE;
1600 verbose(env, "BUG spi %d\n", spi);
1601 WARN_ONCE(1, "verifier backtracking bug");
1604 if (!(*stack_mask & (1ull << spi)))
1606 *stack_mask &= ~(1ull << spi);
1607 if (class == BPF_STX)
1609 } else if (class == BPF_JMP || class == BPF_JMP32) {
1610 if (opcode == BPF_CALL) {
1611 if (insn->src_reg == BPF_PSEUDO_CALL)
1613 /* regular helper call sets R0 */
1615 if (*reg_mask & 0x3f) {
1616 /* if backtracing was looking for registers R1-R5
1617 * they should have been found already.
1619 verbose(env, "BUG regs %x\n", *reg_mask);
1620 WARN_ONCE(1, "verifier backtracking bug");
1623 } else if (opcode == BPF_EXIT) {
1626 } else if (class == BPF_LD) {
1627 if (!(*reg_mask & dreg))
1630 /* It's ld_imm64 or ld_abs or ld_ind.
1631 * For ld_imm64 no further tracking of precision
1632 * into parent is necessary
1634 if (mode == BPF_IND || mode == BPF_ABS)
1635 /* to be analyzed */
1641 /* the scalar precision tracking algorithm:
1642 * . at the start all registers have precise=false.
1643 * . scalar ranges are tracked as normal through alu and jmp insns.
1644 * . once precise value of the scalar register is used in:
1645 * . ptr + scalar alu
1646 * . if (scalar cond K|scalar)
1647 * . helper_call(.., scalar, ...) where ARG_CONST is expected
1648 * backtrack through the verifier states and mark all registers and
1649 * stack slots with spilled constants that these scalar regisers
1650 * should be precise.
1651 * . during state pruning two registers (or spilled stack slots)
1652 * are equivalent if both are not precise.
1654 * Note the verifier cannot simply walk register parentage chain,
1655 * since many different registers and stack slots could have been
1656 * used to compute single precise scalar.
1658 * The approach of starting with precise=true for all registers and then
1659 * backtrack to mark a register as not precise when the verifier detects
1660 * that program doesn't care about specific value (e.g., when helper
1661 * takes register as ARG_ANYTHING parameter) is not safe.
1663 * It's ok to walk single parentage chain of the verifier states.
1664 * It's possible that this backtracking will go all the way till 1st insn.
1665 * All other branches will be explored for needing precision later.
1667 * The backtracking needs to deal with cases like:
1668 * R8=map_value(id=0,off=0,ks=4,vs=1952,imm=0) R9_w=map_value(id=0,off=40,ks=4,vs=1952,imm=0)
1671 * if r5 > 0x79f goto pc+7
1672 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
1675 * call bpf_perf_event_output#25
1676 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
1680 * call foo // uses callee's r6 inside to compute r0
1684 * to track above reg_mask/stack_mask needs to be independent for each frame.
1686 * Also if parent's curframe > frame where backtracking started,
1687 * the verifier need to mark registers in both frames, otherwise callees
1688 * may incorrectly prune callers. This is similar to
1689 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
1691 * For now backtracking falls back into conservative marking.
1693 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
1694 struct bpf_verifier_state *st)
1696 struct bpf_func_state *func;
1697 struct bpf_reg_state *reg;
1700 /* big hammer: mark all scalars precise in this path.
1701 * pop_stack may still get !precise scalars.
1703 for (; st; st = st->parent)
1704 for (i = 0; i <= st->curframe; i++) {
1705 func = st->frame[i];
1706 for (j = 0; j < BPF_REG_FP; j++) {
1707 reg = &func->regs[j];
1708 if (reg->type != SCALAR_VALUE)
1710 reg->precise = true;
1712 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
1713 if (func->stack[j].slot_type[0] != STACK_SPILL)
1715 reg = &func->stack[j].spilled_ptr;
1716 if (reg->type != SCALAR_VALUE)
1718 reg->precise = true;
1723 static int __mark_chain_precision(struct bpf_verifier_env *env, int regno,
1726 struct bpf_verifier_state *st = env->cur_state;
1727 int first_idx = st->first_insn_idx;
1728 int last_idx = env->insn_idx;
1729 struct bpf_func_state *func;
1730 struct bpf_reg_state *reg;
1731 u32 reg_mask = regno >= 0 ? 1u << regno : 0;
1732 u64 stack_mask = spi >= 0 ? 1ull << spi : 0;
1733 bool skip_first = true;
1734 bool new_marks = false;
1737 if (!env->allow_ptr_leaks)
1738 /* backtracking is root only for now */
1741 func = st->frame[st->curframe];
1743 reg = &func->regs[regno];
1744 if (reg->type != SCALAR_VALUE) {
1745 WARN_ONCE(1, "backtracing misuse");
1752 reg->precise = true;
1756 if (func->stack[spi].slot_type[0] != STACK_SPILL) {
1760 reg = &func->stack[spi].spilled_ptr;
1761 if (reg->type != SCALAR_VALUE) {
1769 reg->precise = true;
1775 if (!reg_mask && !stack_mask)
1778 DECLARE_BITMAP(mask, 64);
1779 u32 history = st->jmp_history_cnt;
1781 if (env->log.level & BPF_LOG_LEVEL)
1782 verbose(env, "last_idx %d first_idx %d\n", last_idx, first_idx);
1783 for (i = last_idx;;) {
1788 err = backtrack_insn(env, i, ®_mask, &stack_mask);
1790 if (err == -ENOTSUPP) {
1791 mark_all_scalars_precise(env, st);
1796 if (!reg_mask && !stack_mask)
1797 /* Found assignment(s) into tracked register in this state.
1798 * Since this state is already marked, just return.
1799 * Nothing to be tracked further in the parent state.
1804 i = get_prev_insn_idx(st, i, &history);
1805 if (i >= env->prog->len) {
1806 /* This can happen if backtracking reached insn 0
1807 * and there are still reg_mask or stack_mask
1809 * It means the backtracking missed the spot where
1810 * particular register was initialized with a constant.
1812 verbose(env, "BUG backtracking idx %d\n", i);
1813 WARN_ONCE(1, "verifier backtracking bug");
1822 func = st->frame[st->curframe];
1823 bitmap_from_u64(mask, reg_mask);
1824 for_each_set_bit(i, mask, 32) {
1825 reg = &func->regs[i];
1826 if (reg->type != SCALAR_VALUE) {
1827 reg_mask &= ~(1u << i);
1832 reg->precise = true;
1835 bitmap_from_u64(mask, stack_mask);
1836 for_each_set_bit(i, mask, 64) {
1837 if (i >= func->allocated_stack / BPF_REG_SIZE) {
1838 /* the sequence of instructions:
1840 * 3: (7b) *(u64 *)(r3 -8) = r0
1841 * 4: (79) r4 = *(u64 *)(r10 -8)
1842 * doesn't contain jmps. It's backtracked
1843 * as a single block.
1844 * During backtracking insn 3 is not recognized as
1845 * stack access, so at the end of backtracking
1846 * stack slot fp-8 is still marked in stack_mask.
1847 * However the parent state may not have accessed
1848 * fp-8 and it's "unallocated" stack space.
1849 * In such case fallback to conservative.
1851 mark_all_scalars_precise(env, st);
1855 if (func->stack[i].slot_type[0] != STACK_SPILL) {
1856 stack_mask &= ~(1ull << i);
1859 reg = &func->stack[i].spilled_ptr;
1860 if (reg->type != SCALAR_VALUE) {
1861 stack_mask &= ~(1ull << i);
1866 reg->precise = true;
1868 if (env->log.level & BPF_LOG_LEVEL) {
1869 print_verifier_state(env, func);
1870 verbose(env, "parent %s regs=%x stack=%llx marks\n",
1871 new_marks ? "didn't have" : "already had",
1872 reg_mask, stack_mask);
1875 if (!reg_mask && !stack_mask)
1880 last_idx = st->last_insn_idx;
1881 first_idx = st->first_insn_idx;
1886 static int mark_chain_precision(struct bpf_verifier_env *env, int regno)
1888 return __mark_chain_precision(env, regno, -1);
1891 static int mark_chain_precision_stack(struct bpf_verifier_env *env, int spi)
1893 return __mark_chain_precision(env, -1, spi);
1896 static bool is_spillable_regtype(enum bpf_reg_type type)
1899 case PTR_TO_MAP_VALUE:
1900 case PTR_TO_MAP_VALUE_OR_NULL:
1904 case PTR_TO_PACKET_META:
1905 case PTR_TO_PACKET_END:
1906 case PTR_TO_FLOW_KEYS:
1907 case CONST_PTR_TO_MAP:
1909 case PTR_TO_SOCKET_OR_NULL:
1910 case PTR_TO_SOCK_COMMON:
1911 case PTR_TO_SOCK_COMMON_OR_NULL:
1912 case PTR_TO_TCP_SOCK:
1913 case PTR_TO_TCP_SOCK_OR_NULL:
1914 case PTR_TO_XDP_SOCK:
1922 /* Does this register contain a constant zero? */
1923 static bool register_is_null(struct bpf_reg_state *reg)
1925 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
1928 static bool register_is_const(struct bpf_reg_state *reg)
1930 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
1933 static void save_register_state(struct bpf_func_state *state,
1934 int spi, struct bpf_reg_state *reg)
1938 state->stack[spi].spilled_ptr = *reg;
1939 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1941 for (i = 0; i < BPF_REG_SIZE; i++)
1942 state->stack[spi].slot_type[i] = STACK_SPILL;
1945 /* check_stack_read/write functions track spill/fill of registers,
1946 * stack boundary and alignment are checked in check_mem_access()
1948 static int check_stack_write(struct bpf_verifier_env *env,
1949 struct bpf_func_state *state, /* func where register points to */
1950 int off, int size, int value_regno, int insn_idx)
1952 struct bpf_func_state *cur; /* state of the current function */
1953 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
1954 u32 dst_reg = env->prog->insnsi[insn_idx].dst_reg;
1955 struct bpf_reg_state *reg = NULL;
1957 err = realloc_func_state(state, round_up(slot + 1, BPF_REG_SIZE),
1958 state->acquired_refs, true);
1961 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
1962 * so it's aligned access and [off, off + size) are within stack limits
1964 if (!env->allow_ptr_leaks &&
1965 state->stack[spi].slot_type[0] == STACK_SPILL &&
1966 size != BPF_REG_SIZE) {
1967 verbose(env, "attempt to corrupt spilled pointer on stack\n");
1971 cur = env->cur_state->frame[env->cur_state->curframe];
1972 if (value_regno >= 0)
1973 reg = &cur->regs[value_regno];
1975 if (reg && size == BPF_REG_SIZE && register_is_const(reg) &&
1976 !register_is_null(reg) && env->allow_ptr_leaks) {
1977 if (dst_reg != BPF_REG_FP) {
1978 /* The backtracking logic can only recognize explicit
1979 * stack slot address like [fp - 8]. Other spill of
1980 * scalar via different register has to be conervative.
1981 * Backtrack from here and mark all registers as precise
1982 * that contributed into 'reg' being a constant.
1984 err = mark_chain_precision(env, value_regno);
1988 save_register_state(state, spi, reg);
1989 } else if (reg && is_spillable_regtype(reg->type)) {
1990 /* register containing pointer is being spilled into stack */
1991 if (size != BPF_REG_SIZE) {
1992 verbose_linfo(env, insn_idx, "; ");
1993 verbose(env, "invalid size of register spill\n");
1997 if (state != cur && reg->type == PTR_TO_STACK) {
1998 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
2002 if (!env->allow_ptr_leaks) {
2003 bool sanitize = false;
2005 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
2006 register_is_const(&state->stack[spi].spilled_ptr))
2008 for (i = 0; i < BPF_REG_SIZE; i++)
2009 if (state->stack[spi].slot_type[i] == STACK_MISC) {
2014 int *poff = &env->insn_aux_data[insn_idx].sanitize_stack_off;
2015 int soff = (-spi - 1) * BPF_REG_SIZE;
2017 /* detected reuse of integer stack slot with a pointer
2018 * which means either llvm is reusing stack slot or
2019 * an attacker is trying to exploit CVE-2018-3639
2020 * (speculative store bypass)
2021 * Have to sanitize that slot with preemptive
2024 if (*poff && *poff != soff) {
2025 /* disallow programs where single insn stores
2026 * into two different stack slots, since verifier
2027 * cannot sanitize them
2030 "insn %d cannot access two stack slots fp%d and fp%d",
2031 insn_idx, *poff, soff);
2037 save_register_state(state, spi, reg);
2039 u8 type = STACK_MISC;
2041 /* regular write of data into stack destroys any spilled ptr */
2042 state->stack[spi].spilled_ptr.type = NOT_INIT;
2043 /* Mark slots as STACK_MISC if they belonged to spilled ptr. */
2044 if (state->stack[spi].slot_type[0] == STACK_SPILL)
2045 for (i = 0; i < BPF_REG_SIZE; i++)
2046 state->stack[spi].slot_type[i] = STACK_MISC;
2048 /* only mark the slot as written if all 8 bytes were written
2049 * otherwise read propagation may incorrectly stop too soon
2050 * when stack slots are partially written.
2051 * This heuristic means that read propagation will be
2052 * conservative, since it will add reg_live_read marks
2053 * to stack slots all the way to first state when programs
2054 * writes+reads less than 8 bytes
2056 if (size == BPF_REG_SIZE)
2057 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2059 /* when we zero initialize stack slots mark them as such */
2060 if (reg && register_is_null(reg)) {
2061 /* backtracking doesn't work for STACK_ZERO yet. */
2062 err = mark_chain_precision(env, value_regno);
2068 /* Mark slots affected by this stack write. */
2069 for (i = 0; i < size; i++)
2070 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
2076 static int check_stack_read(struct bpf_verifier_env *env,
2077 struct bpf_func_state *reg_state /* func where register points to */,
2078 int off, int size, int value_regno)
2080 struct bpf_verifier_state *vstate = env->cur_state;
2081 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2082 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
2083 struct bpf_reg_state *reg;
2086 if (reg_state->allocated_stack <= slot) {
2087 verbose(env, "invalid read from stack off %d+0 size %d\n",
2091 stype = reg_state->stack[spi].slot_type;
2092 reg = ®_state->stack[spi].spilled_ptr;
2094 if (stype[0] == STACK_SPILL) {
2095 if (size != BPF_REG_SIZE) {
2096 if (reg->type != SCALAR_VALUE) {
2097 verbose_linfo(env, env->insn_idx, "; ");
2098 verbose(env, "invalid size of register fill\n");
2101 if (value_regno >= 0) {
2102 mark_reg_unknown(env, state->regs, value_regno);
2103 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
2105 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2108 for (i = 1; i < BPF_REG_SIZE; i++) {
2109 if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) {
2110 verbose(env, "corrupted spill memory\n");
2115 if (value_regno >= 0) {
2116 /* restore register state from stack */
2117 state->regs[value_regno] = *reg;
2118 /* mark reg as written since spilled pointer state likely
2119 * has its liveness marks cleared by is_state_visited()
2120 * which resets stack/reg liveness for state transitions
2122 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
2124 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2128 for (i = 0; i < size; i++) {
2129 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_MISC)
2131 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_ZERO) {
2135 verbose(env, "invalid read from stack off %d+%d size %d\n",
2139 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2140 if (value_regno >= 0) {
2141 if (zeros == size) {
2142 /* any size read into register is zero extended,
2143 * so the whole register == const_zero
2145 __mark_reg_const_zero(&state->regs[value_regno]);
2146 /* backtracking doesn't support STACK_ZERO yet,
2147 * so mark it precise here, so that later
2148 * backtracking can stop here.
2149 * Backtracking may not need this if this register
2150 * doesn't participate in pointer adjustment.
2151 * Forward propagation of precise flag is not
2152 * necessary either. This mark is only to stop
2153 * backtracking. Any register that contributed
2154 * to const 0 was marked precise before spill.
2156 state->regs[value_regno].precise = true;
2158 /* have read misc data from the stack */
2159 mark_reg_unknown(env, state->regs, value_regno);
2161 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
2167 static int check_stack_access(struct bpf_verifier_env *env,
2168 const struct bpf_reg_state *reg,
2171 /* Stack accesses must be at a fixed offset, so that we
2172 * can determine what type of data were returned. See
2173 * check_stack_read().
2175 if (!tnum_is_const(reg->var_off)) {
2178 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2179 verbose(env, "variable stack access var_off=%s off=%d size=%d\n",
2184 if (off >= 0 || off < -MAX_BPF_STACK) {
2185 verbose(env, "invalid stack off=%d size=%d\n", off, size);
2192 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
2193 int off, int size, enum bpf_access_type type)
2195 struct bpf_reg_state *regs = cur_regs(env);
2196 struct bpf_map *map = regs[regno].map_ptr;
2197 u32 cap = bpf_map_flags_to_cap(map);
2199 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
2200 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
2201 map->value_size, off, size);
2205 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
2206 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
2207 map->value_size, off, size);
2214 /* check read/write into map element returned by bpf_map_lookup_elem() */
2215 static int __check_map_access(struct bpf_verifier_env *env, u32 regno, int off,
2216 int size, bool zero_size_allowed)
2218 struct bpf_reg_state *regs = cur_regs(env);
2219 struct bpf_map *map = regs[regno].map_ptr;
2221 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
2222 off + size > map->value_size) {
2223 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
2224 map->value_size, off, size);
2230 /* check read/write into a map element with possible variable offset */
2231 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
2232 int off, int size, bool zero_size_allowed)
2234 struct bpf_verifier_state *vstate = env->cur_state;
2235 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2236 struct bpf_reg_state *reg = &state->regs[regno];
2239 /* We may have adjusted the register to this map value, so we
2240 * need to try adding each of min_value and max_value to off
2241 * to make sure our theoretical access will be safe.
2243 if (env->log.level & BPF_LOG_LEVEL)
2244 print_verifier_state(env, state);
2246 /* The minimum value is only important with signed
2247 * comparisons where we can't assume the floor of a
2248 * value is 0. If we are using signed variables for our
2249 * index'es we need to make sure that whatever we use
2250 * will have a set floor within our range.
2252 if (reg->smin_value < 0 &&
2253 (reg->smin_value == S64_MIN ||
2254 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
2255 reg->smin_value + off < 0)) {
2256 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2260 err = __check_map_access(env, regno, reg->smin_value + off, size,
2263 verbose(env, "R%d min value is outside of the array range\n",
2268 /* If we haven't set a max value then we need to bail since we can't be
2269 * sure we won't do bad things.
2270 * If reg->umax_value + off could overflow, treat that as unbounded too.
2272 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
2273 verbose(env, "R%d unbounded memory access, make sure to bounds check any array access into a map\n",
2277 err = __check_map_access(env, regno, reg->umax_value + off, size,
2280 verbose(env, "R%d max value is outside of the array range\n",
2283 if (map_value_has_spin_lock(reg->map_ptr)) {
2284 u32 lock = reg->map_ptr->spin_lock_off;
2286 /* if any part of struct bpf_spin_lock can be touched by
2287 * load/store reject this program.
2288 * To check that [x1, x2) overlaps with [y1, y2)
2289 * it is sufficient to check x1 < y2 && y1 < x2.
2291 if (reg->smin_value + off < lock + sizeof(struct bpf_spin_lock) &&
2292 lock < reg->umax_value + off + size) {
2293 verbose(env, "bpf_spin_lock cannot be accessed directly by load/store\n");
2300 #define MAX_PACKET_OFF 0xffff
2302 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
2303 const struct bpf_call_arg_meta *meta,
2304 enum bpf_access_type t)
2306 switch (env->prog->type) {
2307 /* Program types only with direct read access go here! */
2308 case BPF_PROG_TYPE_LWT_IN:
2309 case BPF_PROG_TYPE_LWT_OUT:
2310 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
2311 case BPF_PROG_TYPE_SK_REUSEPORT:
2312 case BPF_PROG_TYPE_FLOW_DISSECTOR:
2313 case BPF_PROG_TYPE_CGROUP_SKB:
2318 /* Program types with direct read + write access go here! */
2319 case BPF_PROG_TYPE_SCHED_CLS:
2320 case BPF_PROG_TYPE_SCHED_ACT:
2321 case BPF_PROG_TYPE_XDP:
2322 case BPF_PROG_TYPE_LWT_XMIT:
2323 case BPF_PROG_TYPE_SK_SKB:
2324 case BPF_PROG_TYPE_SK_MSG:
2326 return meta->pkt_access;
2328 env->seen_direct_write = true;
2331 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
2333 env->seen_direct_write = true;
2342 static int __check_packet_access(struct bpf_verifier_env *env, u32 regno,
2343 int off, int size, bool zero_size_allowed)
2345 struct bpf_reg_state *regs = cur_regs(env);
2346 struct bpf_reg_state *reg = ®s[regno];
2348 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
2349 (u64)off + size > reg->range) {
2350 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
2351 off, size, regno, reg->id, reg->off, reg->range);
2357 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
2358 int size, bool zero_size_allowed)
2360 struct bpf_reg_state *regs = cur_regs(env);
2361 struct bpf_reg_state *reg = ®s[regno];
2364 /* We may have added a variable offset to the packet pointer; but any
2365 * reg->range we have comes after that. We are only checking the fixed
2369 /* We don't allow negative numbers, because we aren't tracking enough
2370 * detail to prove they're safe.
2372 if (reg->smin_value < 0) {
2373 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2377 err = __check_packet_access(env, regno, off, size, zero_size_allowed);
2379 verbose(env, "R%d offset is outside of the packet\n", regno);
2383 /* __check_packet_access has made sure "off + size - 1" is within u16.
2384 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
2385 * otherwise find_good_pkt_pointers would have refused to set range info
2386 * that __check_packet_access would have rejected this pkt access.
2387 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
2389 env->prog->aux->max_pkt_offset =
2390 max_t(u32, env->prog->aux->max_pkt_offset,
2391 off + reg->umax_value + size - 1);
2396 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
2397 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
2398 enum bpf_access_type t, enum bpf_reg_type *reg_type,
2401 struct bpf_insn_access_aux info = {
2402 .reg_type = *reg_type,
2406 if (env->ops->is_valid_access &&
2407 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
2408 /* A non zero info.ctx_field_size indicates that this field is a
2409 * candidate for later verifier transformation to load the whole
2410 * field and then apply a mask when accessed with a narrower
2411 * access than actual ctx access size. A zero info.ctx_field_size
2412 * will only allow for whole field access and rejects any other
2413 * type of narrower access.
2415 *reg_type = info.reg_type;
2417 if (*reg_type == PTR_TO_BTF_ID)
2418 *btf_id = info.btf_id;
2420 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
2421 /* remember the offset of last byte accessed in ctx */
2422 if (env->prog->aux->max_ctx_offset < off + size)
2423 env->prog->aux->max_ctx_offset = off + size;
2427 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
2431 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
2434 if (size < 0 || off < 0 ||
2435 (u64)off + size > sizeof(struct bpf_flow_keys)) {
2436 verbose(env, "invalid access to flow keys off=%d size=%d\n",
2443 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
2444 u32 regno, int off, int size,
2445 enum bpf_access_type t)
2447 struct bpf_reg_state *regs = cur_regs(env);
2448 struct bpf_reg_state *reg = ®s[regno];
2449 struct bpf_insn_access_aux info = {};
2452 if (reg->smin_value < 0) {
2453 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2458 switch (reg->type) {
2459 case PTR_TO_SOCK_COMMON:
2460 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
2463 valid = bpf_sock_is_valid_access(off, size, t, &info);
2465 case PTR_TO_TCP_SOCK:
2466 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
2468 case PTR_TO_XDP_SOCK:
2469 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
2477 env->insn_aux_data[insn_idx].ctx_field_size =
2478 info.ctx_field_size;
2482 verbose(env, "R%d invalid %s access off=%d size=%d\n",
2483 regno, reg_type_str[reg->type], off, size);
2488 static bool __is_pointer_value(bool allow_ptr_leaks,
2489 const struct bpf_reg_state *reg)
2491 if (allow_ptr_leaks)
2494 return reg->type != SCALAR_VALUE;
2497 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
2499 return cur_regs(env) + regno;
2502 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
2504 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
2507 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
2509 const struct bpf_reg_state *reg = reg_state(env, regno);
2511 return reg->type == PTR_TO_CTX;
2514 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
2516 const struct bpf_reg_state *reg = reg_state(env, regno);
2518 return type_is_sk_pointer(reg->type);
2521 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
2523 const struct bpf_reg_state *reg = reg_state(env, regno);
2525 return type_is_pkt_pointer(reg->type);
2528 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
2530 const struct bpf_reg_state *reg = reg_state(env, regno);
2532 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
2533 return reg->type == PTR_TO_FLOW_KEYS;
2536 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
2537 const struct bpf_reg_state *reg,
2538 int off, int size, bool strict)
2540 struct tnum reg_off;
2543 /* Byte size accesses are always allowed. */
2544 if (!strict || size == 1)
2547 /* For platforms that do not have a Kconfig enabling
2548 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
2549 * NET_IP_ALIGN is universally set to '2'. And on platforms
2550 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
2551 * to this code only in strict mode where we want to emulate
2552 * the NET_IP_ALIGN==2 checking. Therefore use an
2553 * unconditional IP align value of '2'.
2557 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
2558 if (!tnum_is_aligned(reg_off, size)) {
2561 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2563 "misaligned packet access off %d+%s+%d+%d size %d\n",
2564 ip_align, tn_buf, reg->off, off, size);
2571 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
2572 const struct bpf_reg_state *reg,
2573 const char *pointer_desc,
2574 int off, int size, bool strict)
2576 struct tnum reg_off;
2578 /* Byte size accesses are always allowed. */
2579 if (!strict || size == 1)
2582 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
2583 if (!tnum_is_aligned(reg_off, size)) {
2586 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2587 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
2588 pointer_desc, tn_buf, reg->off, off, size);
2595 static int check_ptr_alignment(struct bpf_verifier_env *env,
2596 const struct bpf_reg_state *reg, int off,
2597 int size, bool strict_alignment_once)
2599 bool strict = env->strict_alignment || strict_alignment_once;
2600 const char *pointer_desc = "";
2602 switch (reg->type) {
2604 case PTR_TO_PACKET_META:
2605 /* Special case, because of NET_IP_ALIGN. Given metadata sits
2606 * right in front, treat it the very same way.
2608 return check_pkt_ptr_alignment(env, reg, off, size, strict);
2609 case PTR_TO_FLOW_KEYS:
2610 pointer_desc = "flow keys ";
2612 case PTR_TO_MAP_VALUE:
2613 pointer_desc = "value ";
2616 pointer_desc = "context ";
2619 pointer_desc = "stack ";
2620 /* The stack spill tracking logic in check_stack_write()
2621 * and check_stack_read() relies on stack accesses being
2627 pointer_desc = "sock ";
2629 case PTR_TO_SOCK_COMMON:
2630 pointer_desc = "sock_common ";
2632 case PTR_TO_TCP_SOCK:
2633 pointer_desc = "tcp_sock ";
2635 case PTR_TO_XDP_SOCK:
2636 pointer_desc = "xdp_sock ";
2641 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
2645 static int update_stack_depth(struct bpf_verifier_env *env,
2646 const struct bpf_func_state *func,
2649 u16 stack = env->subprog_info[func->subprogno].stack_depth;
2654 /* update known max for given subprogram */
2655 env->subprog_info[func->subprogno].stack_depth = -off;
2659 /* starting from main bpf function walk all instructions of the function
2660 * and recursively walk all callees that given function can call.
2661 * Ignore jump and exit insns.
2662 * Since recursion is prevented by check_cfg() this algorithm
2663 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
2665 static int check_max_stack_depth(struct bpf_verifier_env *env)
2667 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
2668 struct bpf_subprog_info *subprog = env->subprog_info;
2669 struct bpf_insn *insn = env->prog->insnsi;
2670 int ret_insn[MAX_CALL_FRAMES];
2671 int ret_prog[MAX_CALL_FRAMES];
2674 /* round up to 32-bytes, since this is granularity
2675 * of interpreter stack size
2677 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
2678 if (depth > MAX_BPF_STACK) {
2679 verbose(env, "combined stack size of %d calls is %d. Too large\n",
2684 subprog_end = subprog[idx + 1].start;
2685 for (; i < subprog_end; i++) {
2686 if (insn[i].code != (BPF_JMP | BPF_CALL))
2688 if (insn[i].src_reg != BPF_PSEUDO_CALL)
2690 /* remember insn and function to return to */
2691 ret_insn[frame] = i + 1;
2692 ret_prog[frame] = idx;
2694 /* find the callee */
2695 i = i + insn[i].imm + 1;
2696 idx = find_subprog(env, i);
2698 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
2703 if (frame >= MAX_CALL_FRAMES) {
2704 verbose(env, "the call stack of %d frames is too deep !\n",
2710 /* end of for() loop means the last insn of the 'subprog'
2711 * was reached. Doesn't matter whether it was JA or EXIT
2715 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
2717 i = ret_insn[frame];
2718 idx = ret_prog[frame];
2722 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
2723 static int get_callee_stack_depth(struct bpf_verifier_env *env,
2724 const struct bpf_insn *insn, int idx)
2726 int start = idx + insn->imm + 1, subprog;
2728 subprog = find_subprog(env, start);
2730 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
2734 return env->subprog_info[subprog].stack_depth;
2738 int check_ctx_reg(struct bpf_verifier_env *env,
2739 const struct bpf_reg_state *reg, int regno)
2741 /* Access to ctx or passing it to a helper is only allowed in
2742 * its original, unmodified form.
2746 verbose(env, "dereference of modified ctx ptr R%d off=%d disallowed\n",
2751 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
2754 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2755 verbose(env, "variable ctx access var_off=%s disallowed\n", tn_buf);
2762 static int check_tp_buffer_access(struct bpf_verifier_env *env,
2763 const struct bpf_reg_state *reg,
2764 int regno, int off, int size)
2768 "R%d invalid tracepoint buffer access: off=%d, size=%d",
2772 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
2775 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2777 "R%d invalid variable buffer offset: off=%d, var_off=%s",
2778 regno, off, tn_buf);
2781 if (off + size > env->prog->aux->max_tp_access)
2782 env->prog->aux->max_tp_access = off + size;
2788 /* truncate register to smaller size (in bytes)
2789 * must be called with size < BPF_REG_SIZE
2791 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
2795 /* clear high bits in bit representation */
2796 reg->var_off = tnum_cast(reg->var_off, size);
2798 /* fix arithmetic bounds */
2799 mask = ((u64)1 << (size * 8)) - 1;
2800 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
2801 reg->umin_value &= mask;
2802 reg->umax_value &= mask;
2804 reg->umin_value = 0;
2805 reg->umax_value = mask;
2807 reg->smin_value = reg->umin_value;
2808 reg->smax_value = reg->umax_value;
2811 static bool bpf_map_is_rdonly(const struct bpf_map *map)
2813 return (map->map_flags & BPF_F_RDONLY_PROG) && map->frozen;
2816 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val)
2822 err = map->ops->map_direct_value_addr(map, &addr, off);
2825 ptr = (void *)(long)addr + off;
2829 *val = (u64)*(u8 *)ptr;
2832 *val = (u64)*(u16 *)ptr;
2835 *val = (u64)*(u32 *)ptr;
2846 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
2847 struct bpf_reg_state *regs,
2848 int regno, int off, int size,
2849 enum bpf_access_type atype,
2852 struct bpf_reg_state *reg = regs + regno;
2853 const struct btf_type *t = btf_type_by_id(btf_vmlinux, reg->btf_id);
2854 const char *tname = btf_name_by_offset(btf_vmlinux, t->name_off);
2860 "R%d is ptr_%s invalid negative access: off=%d\n",
2864 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
2867 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2869 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
2870 regno, tname, off, tn_buf);
2874 if (env->ops->btf_struct_access) {
2875 ret = env->ops->btf_struct_access(&env->log, t, off, size,
2878 if (atype != BPF_READ) {
2879 verbose(env, "only read is supported\n");
2883 ret = btf_struct_access(&env->log, t, off, size, atype,
2890 if (atype == BPF_READ) {
2891 if (ret == SCALAR_VALUE) {
2892 mark_reg_unknown(env, regs, value_regno);
2895 mark_reg_known_zero(env, regs, value_regno);
2896 regs[value_regno].type = PTR_TO_BTF_ID;
2897 regs[value_regno].btf_id = btf_id;
2903 /* check whether memory at (regno + off) is accessible for t = (read | write)
2904 * if t==write, value_regno is a register which value is stored into memory
2905 * if t==read, value_regno is a register which will receive the value from memory
2906 * if t==write && value_regno==-1, some unknown value is stored into memory
2907 * if t==read && value_regno==-1, don't care what we read from memory
2909 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
2910 int off, int bpf_size, enum bpf_access_type t,
2911 int value_regno, bool strict_alignment_once)
2913 struct bpf_reg_state *regs = cur_regs(env);
2914 struct bpf_reg_state *reg = regs + regno;
2915 struct bpf_func_state *state;
2918 size = bpf_size_to_bytes(bpf_size);
2922 /* alignment checks will add in reg->off themselves */
2923 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
2927 /* for access checks, reg->off is just part of off */
2930 if (reg->type == PTR_TO_MAP_VALUE) {
2931 if (t == BPF_WRITE && value_regno >= 0 &&
2932 is_pointer_value(env, value_regno)) {
2933 verbose(env, "R%d leaks addr into map\n", value_regno);
2936 err = check_map_access_type(env, regno, off, size, t);
2939 err = check_map_access(env, regno, off, size, false);
2940 if (!err && t == BPF_READ && value_regno >= 0) {
2941 struct bpf_map *map = reg->map_ptr;
2943 /* if map is read-only, track its contents as scalars */
2944 if (tnum_is_const(reg->var_off) &&
2945 bpf_map_is_rdonly(map) &&
2946 map->ops->map_direct_value_addr) {
2947 int map_off = off + reg->var_off.value;
2950 err = bpf_map_direct_read(map, map_off, size,
2955 regs[value_regno].type = SCALAR_VALUE;
2956 __mark_reg_known(®s[value_regno], val);
2958 mark_reg_unknown(env, regs, value_regno);
2961 } else if (reg->type == PTR_TO_CTX) {
2962 enum bpf_reg_type reg_type = SCALAR_VALUE;
2965 if (t == BPF_WRITE && value_regno >= 0 &&
2966 is_pointer_value(env, value_regno)) {
2967 verbose(env, "R%d leaks addr into ctx\n", value_regno);
2971 err = check_ctx_reg(env, reg, regno);
2975 err = check_ctx_access(env, insn_idx, off, size, t, ®_type, &btf_id);
2977 verbose_linfo(env, insn_idx, "; ");
2978 if (!err && t == BPF_READ && value_regno >= 0) {
2979 /* ctx access returns either a scalar, or a
2980 * PTR_TO_PACKET[_META,_END]. In the latter
2981 * case, we know the offset is zero.
2983 if (reg_type == SCALAR_VALUE) {
2984 mark_reg_unknown(env, regs, value_regno);
2986 mark_reg_known_zero(env, regs,
2988 if (reg_type_may_be_null(reg_type))
2989 regs[value_regno].id = ++env->id_gen;
2990 /* A load of ctx field could have different
2991 * actual load size with the one encoded in the
2992 * insn. When the dst is PTR, it is for sure not
2995 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
2996 if (reg_type == PTR_TO_BTF_ID)
2997 regs[value_regno].btf_id = btf_id;
2999 regs[value_regno].type = reg_type;
3002 } else if (reg->type == PTR_TO_STACK) {
3003 off += reg->var_off.value;
3004 err = check_stack_access(env, reg, off, size);
3008 state = func(env, reg);
3009 err = update_stack_depth(env, state, off);
3014 err = check_stack_write(env, state, off, size,
3015 value_regno, insn_idx);
3017 err = check_stack_read(env, state, off, size,
3019 } else if (reg_is_pkt_pointer(reg)) {
3020 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
3021 verbose(env, "cannot write into packet\n");
3024 if (t == BPF_WRITE && value_regno >= 0 &&
3025 is_pointer_value(env, value_regno)) {
3026 verbose(env, "R%d leaks addr into packet\n",
3030 err = check_packet_access(env, regno, off, size, false);
3031 if (!err && t == BPF_READ && value_regno >= 0)
3032 mark_reg_unknown(env, regs, value_regno);
3033 } else if (reg->type == PTR_TO_FLOW_KEYS) {
3034 if (t == BPF_WRITE && value_regno >= 0 &&
3035 is_pointer_value(env, value_regno)) {
3036 verbose(env, "R%d leaks addr into flow keys\n",
3041 err = check_flow_keys_access(env, off, size);
3042 if (!err && t == BPF_READ && value_regno >= 0)
3043 mark_reg_unknown(env, regs, value_regno);
3044 } else if (type_is_sk_pointer(reg->type)) {
3045 if (t == BPF_WRITE) {
3046 verbose(env, "R%d cannot write into %s\n",
3047 regno, reg_type_str[reg->type]);
3050 err = check_sock_access(env, insn_idx, regno, off, size, t);
3051 if (!err && value_regno >= 0)
3052 mark_reg_unknown(env, regs, value_regno);
3053 } else if (reg->type == PTR_TO_TP_BUFFER) {
3054 err = check_tp_buffer_access(env, reg, regno, off, size);
3055 if (!err && t == BPF_READ && value_regno >= 0)
3056 mark_reg_unknown(env, regs, value_regno);
3057 } else if (reg->type == PTR_TO_BTF_ID) {
3058 err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
3061 verbose(env, "R%d invalid mem access '%s'\n", regno,
3062 reg_type_str[reg->type]);
3066 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
3067 regs[value_regno].type == SCALAR_VALUE) {
3068 /* b/h/w load zero-extends, mark upper bits as known 0 */
3069 coerce_reg_to_size(®s[value_regno], size);
3074 static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
3078 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
3080 verbose(env, "BPF_XADD uses reserved fields\n");
3084 /* check src1 operand */
3085 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3089 /* check src2 operand */
3090 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3094 if (is_pointer_value(env, insn->src_reg)) {
3095 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
3099 if (is_ctx_reg(env, insn->dst_reg) ||
3100 is_pkt_reg(env, insn->dst_reg) ||
3101 is_flow_key_reg(env, insn->dst_reg) ||
3102 is_sk_reg(env, insn->dst_reg)) {
3103 verbose(env, "BPF_XADD stores into R%d %s is not allowed\n",
3105 reg_type_str[reg_state(env, insn->dst_reg)->type]);
3109 /* check whether atomic_add can read the memory */
3110 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3111 BPF_SIZE(insn->code), BPF_READ, -1, true);
3115 /* check whether atomic_add can write into the same memory */
3116 return check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3117 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
3120 static int __check_stack_boundary(struct bpf_verifier_env *env, u32 regno,
3121 int off, int access_size,
3122 bool zero_size_allowed)
3124 struct bpf_reg_state *reg = reg_state(env, regno);
3126 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
3127 access_size < 0 || (access_size == 0 && !zero_size_allowed)) {
3128 if (tnum_is_const(reg->var_off)) {
3129 verbose(env, "invalid stack type R%d off=%d access_size=%d\n",
3130 regno, off, access_size);
3134 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3135 verbose(env, "invalid stack type R%d var_off=%s access_size=%d\n",
3136 regno, tn_buf, access_size);
3143 /* when register 'regno' is passed into function that will read 'access_size'
3144 * bytes from that pointer, make sure that it's within stack boundary
3145 * and all elements of stack are initialized.
3146 * Unlike most pointer bounds-checking functions, this one doesn't take an
3147 * 'off' argument, so it has to add in reg->off itself.
3149 static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
3150 int access_size, bool zero_size_allowed,
3151 struct bpf_call_arg_meta *meta)
3153 struct bpf_reg_state *reg = reg_state(env, regno);
3154 struct bpf_func_state *state = func(env, reg);
3155 int err, min_off, max_off, i, j, slot, spi;
3157 if (reg->type != PTR_TO_STACK) {
3158 /* Allow zero-byte read from NULL, regardless of pointer type */
3159 if (zero_size_allowed && access_size == 0 &&
3160 register_is_null(reg))
3163 verbose(env, "R%d type=%s expected=%s\n", regno,
3164 reg_type_str[reg->type],
3165 reg_type_str[PTR_TO_STACK]);
3169 if (tnum_is_const(reg->var_off)) {
3170 min_off = max_off = reg->var_off.value + reg->off;
3171 err = __check_stack_boundary(env, regno, min_off, access_size,
3176 /* Variable offset is prohibited for unprivileged mode for
3177 * simplicity since it requires corresponding support in
3178 * Spectre masking for stack ALU.
3179 * See also retrieve_ptr_limit().
3181 if (!env->allow_ptr_leaks) {
3184 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3185 verbose(env, "R%d indirect variable offset stack access prohibited for !root, var_off=%s\n",
3189 /* Only initialized buffer on stack is allowed to be accessed
3190 * with variable offset. With uninitialized buffer it's hard to
3191 * guarantee that whole memory is marked as initialized on
3192 * helper return since specific bounds are unknown what may
3193 * cause uninitialized stack leaking.
3195 if (meta && meta->raw_mode)
3198 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
3199 reg->smax_value <= -BPF_MAX_VAR_OFF) {
3200 verbose(env, "R%d unbounded indirect variable offset stack access\n",
3204 min_off = reg->smin_value + reg->off;
3205 max_off = reg->smax_value + reg->off;
3206 err = __check_stack_boundary(env, regno, min_off, access_size,
3209 verbose(env, "R%d min value is outside of stack bound\n",
3213 err = __check_stack_boundary(env, regno, max_off, access_size,
3216 verbose(env, "R%d max value is outside of stack bound\n",
3222 if (meta && meta->raw_mode) {
3223 meta->access_size = access_size;
3224 meta->regno = regno;
3228 for (i = min_off; i < max_off + access_size; i++) {
3232 spi = slot / BPF_REG_SIZE;
3233 if (state->allocated_stack <= slot)
3235 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
3236 if (*stype == STACK_MISC)
3238 if (*stype == STACK_ZERO) {
3239 /* helper can write anything into the stack */
3240 *stype = STACK_MISC;
3243 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
3244 state->stack[spi].spilled_ptr.type == SCALAR_VALUE) {
3245 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
3246 for (j = 0; j < BPF_REG_SIZE; j++)
3247 state->stack[spi].slot_type[j] = STACK_MISC;
3252 if (tnum_is_const(reg->var_off)) {
3253 verbose(env, "invalid indirect read from stack off %d+%d size %d\n",
3254 min_off, i - min_off, access_size);
3258 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3259 verbose(env, "invalid indirect read from stack var_off %s+%d size %d\n",
3260 tn_buf, i - min_off, access_size);
3264 /* reading any byte out of 8-byte 'spill_slot' will cause
3265 * the whole slot to be marked as 'read'
3267 mark_reg_read(env, &state->stack[spi].spilled_ptr,
3268 state->stack[spi].spilled_ptr.parent,
3271 return update_stack_depth(env, state, min_off);
3274 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
3275 int access_size, bool zero_size_allowed,
3276 struct bpf_call_arg_meta *meta)
3278 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
3280 switch (reg->type) {
3282 case PTR_TO_PACKET_META:
3283 return check_packet_access(env, regno, reg->off, access_size,
3285 case PTR_TO_MAP_VALUE:
3286 if (check_map_access_type(env, regno, reg->off, access_size,
3287 meta && meta->raw_mode ? BPF_WRITE :
3290 return check_map_access(env, regno, reg->off, access_size,
3292 default: /* scalar_value|ptr_to_stack or invalid ptr */
3293 return check_stack_boundary(env, regno, access_size,
3294 zero_size_allowed, meta);
3298 /* Implementation details:
3299 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL
3300 * Two bpf_map_lookups (even with the same key) will have different reg->id.
3301 * For traditional PTR_TO_MAP_VALUE the verifier clears reg->id after
3302 * value_or_null->value transition, since the verifier only cares about
3303 * the range of access to valid map value pointer and doesn't care about actual
3304 * address of the map element.
3305 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
3306 * reg->id > 0 after value_or_null->value transition. By doing so
3307 * two bpf_map_lookups will be considered two different pointers that
3308 * point to different bpf_spin_locks.
3309 * The verifier allows taking only one bpf_spin_lock at a time to avoid
3311 * Since only one bpf_spin_lock is allowed the checks are simpler than
3312 * reg_is_refcounted() logic. The verifier needs to remember only
3313 * one spin_lock instead of array of acquired_refs.
3314 * cur_state->active_spin_lock remembers which map value element got locked
3315 * and clears it after bpf_spin_unlock.
3317 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
3320 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
3321 struct bpf_verifier_state *cur = env->cur_state;
3322 bool is_const = tnum_is_const(reg->var_off);
3323 struct bpf_map *map = reg->map_ptr;
3324 u64 val = reg->var_off.value;
3326 if (reg->type != PTR_TO_MAP_VALUE) {
3327 verbose(env, "R%d is not a pointer to map_value\n", regno);
3332 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
3338 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
3342 if (!map_value_has_spin_lock(map)) {
3343 if (map->spin_lock_off == -E2BIG)
3345 "map '%s' has more than one 'struct bpf_spin_lock'\n",
3347 else if (map->spin_lock_off == -ENOENT)
3349 "map '%s' doesn't have 'struct bpf_spin_lock'\n",
3353 "map '%s' is not a struct type or bpf_spin_lock is mangled\n",
3357 if (map->spin_lock_off != val + reg->off) {
3358 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock'\n",
3363 if (cur->active_spin_lock) {
3365 "Locking two bpf_spin_locks are not allowed\n");
3368 cur->active_spin_lock = reg->id;
3370 if (!cur->active_spin_lock) {
3371 verbose(env, "bpf_spin_unlock without taking a lock\n");
3374 if (cur->active_spin_lock != reg->id) {
3375 verbose(env, "bpf_spin_unlock of different lock\n");
3378 cur->active_spin_lock = 0;
3383 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
3385 return type == ARG_PTR_TO_MEM ||
3386 type == ARG_PTR_TO_MEM_OR_NULL ||
3387 type == ARG_PTR_TO_UNINIT_MEM;
3390 static bool arg_type_is_mem_size(enum bpf_arg_type type)
3392 return type == ARG_CONST_SIZE ||
3393 type == ARG_CONST_SIZE_OR_ZERO;
3396 static bool arg_type_is_int_ptr(enum bpf_arg_type type)
3398 return type == ARG_PTR_TO_INT ||
3399 type == ARG_PTR_TO_LONG;
3402 static int int_ptr_type_to_size(enum bpf_arg_type type)
3404 if (type == ARG_PTR_TO_INT)
3406 else if (type == ARG_PTR_TO_LONG)
3412 static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
3413 enum bpf_arg_type arg_type,
3414 struct bpf_call_arg_meta *meta)
3416 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
3417 enum bpf_reg_type expected_type, type = reg->type;
3420 if (arg_type == ARG_DONTCARE)
3423 err = check_reg_arg(env, regno, SRC_OP);
3427 if (arg_type == ARG_ANYTHING) {
3428 if (is_pointer_value(env, regno)) {
3429 verbose(env, "R%d leaks addr into helper function\n",
3436 if (type_is_pkt_pointer(type) &&
3437 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
3438 verbose(env, "helper access to the packet is not allowed\n");
3442 if (arg_type == ARG_PTR_TO_MAP_KEY ||
3443 arg_type == ARG_PTR_TO_MAP_VALUE ||
3444 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE ||
3445 arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL) {
3446 expected_type = PTR_TO_STACK;
3447 if (register_is_null(reg) &&
3448 arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL)
3449 /* final test in check_stack_boundary() */;
3450 else if (!type_is_pkt_pointer(type) &&
3451 type != PTR_TO_MAP_VALUE &&
3452 type != expected_type)
3454 } else if (arg_type == ARG_CONST_SIZE ||
3455 arg_type == ARG_CONST_SIZE_OR_ZERO) {
3456 expected_type = SCALAR_VALUE;
3457 if (type != expected_type)
3459 } else if (arg_type == ARG_CONST_MAP_PTR) {
3460 expected_type = CONST_PTR_TO_MAP;
3461 if (type != expected_type)
3463 } else if (arg_type == ARG_PTR_TO_CTX) {
3464 expected_type = PTR_TO_CTX;
3465 if (type != expected_type)
3467 err = check_ctx_reg(env, reg, regno);
3470 } else if (arg_type == ARG_PTR_TO_SOCK_COMMON) {
3471 expected_type = PTR_TO_SOCK_COMMON;
3472 /* Any sk pointer can be ARG_PTR_TO_SOCK_COMMON */
3473 if (!type_is_sk_pointer(type))
3475 if (reg->ref_obj_id) {
3476 if (meta->ref_obj_id) {
3477 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
3478 regno, reg->ref_obj_id,
3482 meta->ref_obj_id = reg->ref_obj_id;
3484 } else if (arg_type == ARG_PTR_TO_SOCKET) {
3485 expected_type = PTR_TO_SOCKET;
3486 if (type != expected_type)
3488 } else if (arg_type == ARG_PTR_TO_BTF_ID) {
3489 expected_type = PTR_TO_BTF_ID;
3490 if (type != expected_type)
3492 if (reg->btf_id != meta->btf_id) {
3493 verbose(env, "Helper has type %s got %s in R%d\n",
3494 kernel_type_name(meta->btf_id),
3495 kernel_type_name(reg->btf_id), regno);
3499 if (!tnum_is_const(reg->var_off) || reg->var_off.value || reg->off) {
3500 verbose(env, "R%d is a pointer to in-kernel struct with non-zero offset\n",
3504 } else if (arg_type == ARG_PTR_TO_SPIN_LOCK) {
3505 if (meta->func_id == BPF_FUNC_spin_lock) {
3506 if (process_spin_lock(env, regno, true))
3508 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
3509 if (process_spin_lock(env, regno, false))
3512 verbose(env, "verifier internal error\n");
3515 } else if (arg_type_is_mem_ptr(arg_type)) {
3516 expected_type = PTR_TO_STACK;
3517 /* One exception here. In case function allows for NULL to be
3518 * passed in as argument, it's a SCALAR_VALUE type. Final test
3519 * happens during stack boundary checking.
3521 if (register_is_null(reg) &&
3522 arg_type == ARG_PTR_TO_MEM_OR_NULL)
3523 /* final test in check_stack_boundary() */;
3524 else if (!type_is_pkt_pointer(type) &&
3525 type != PTR_TO_MAP_VALUE &&
3526 type != expected_type)
3528 meta->raw_mode = arg_type == ARG_PTR_TO_UNINIT_MEM;
3529 } else if (arg_type_is_int_ptr(arg_type)) {
3530 expected_type = PTR_TO_STACK;
3531 if (!type_is_pkt_pointer(type) &&
3532 type != PTR_TO_MAP_VALUE &&
3533 type != expected_type)
3536 verbose(env, "unsupported arg_type %d\n", arg_type);
3540 if (arg_type == ARG_CONST_MAP_PTR) {
3541 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
3542 meta->map_ptr = reg->map_ptr;
3543 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
3544 /* bpf_map_xxx(..., map_ptr, ..., key) call:
3545 * check that [key, key + map->key_size) are within
3546 * stack limits and initialized
3548 if (!meta->map_ptr) {
3549 /* in function declaration map_ptr must come before
3550 * map_key, so that it's verified and known before
3551 * we have to check map_key here. Otherwise it means
3552 * that kernel subsystem misconfigured verifier
3554 verbose(env, "invalid map_ptr to access map->key\n");
3557 err = check_helper_mem_access(env, regno,
3558 meta->map_ptr->key_size, false,
3560 } else if (arg_type == ARG_PTR_TO_MAP_VALUE ||
3561 (arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL &&
3562 !register_is_null(reg)) ||
3563 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
3564 /* bpf_map_xxx(..., map_ptr, ..., value) call:
3565 * check [value, value + map->value_size) validity
3567 if (!meta->map_ptr) {
3568 /* kernel subsystem misconfigured verifier */
3569 verbose(env, "invalid map_ptr to access map->value\n");
3572 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE);
3573 err = check_helper_mem_access(env, regno,
3574 meta->map_ptr->value_size, false,
3576 } else if (arg_type_is_mem_size(arg_type)) {
3577 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
3579 /* remember the mem_size which may be used later
3580 * to refine return values.
3582 meta->msize_smax_value = reg->smax_value;
3583 meta->msize_umax_value = reg->umax_value;
3585 /* The register is SCALAR_VALUE; the access check
3586 * happens using its boundaries.
3588 if (!tnum_is_const(reg->var_off))
3589 /* For unprivileged variable accesses, disable raw
3590 * mode so that the program is required to
3591 * initialize all the memory that the helper could
3592 * just partially fill up.
3596 if (reg->smin_value < 0) {
3597 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
3602 if (reg->umin_value == 0) {
3603 err = check_helper_mem_access(env, regno - 1, 0,
3610 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
3611 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
3615 err = check_helper_mem_access(env, regno - 1,
3617 zero_size_allowed, meta);
3619 err = mark_chain_precision(env, regno);
3620 } else if (arg_type_is_int_ptr(arg_type)) {
3621 int size = int_ptr_type_to_size(arg_type);
3623 err = check_helper_mem_access(env, regno, size, false, meta);
3626 err = check_ptr_alignment(env, reg, 0, size, true);
3631 verbose(env, "R%d type=%s expected=%s\n", regno,
3632 reg_type_str[type], reg_type_str[expected_type]);
3636 static int check_map_func_compatibility(struct bpf_verifier_env *env,
3637 struct bpf_map *map, int func_id)
3642 /* We need a two way check, first is from map perspective ... */
3643 switch (map->map_type) {
3644 case BPF_MAP_TYPE_PROG_ARRAY:
3645 if (func_id != BPF_FUNC_tail_call)
3648 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
3649 if (func_id != BPF_FUNC_perf_event_read &&
3650 func_id != BPF_FUNC_perf_event_output &&
3651 func_id != BPF_FUNC_skb_output &&
3652 func_id != BPF_FUNC_perf_event_read_value)
3655 case BPF_MAP_TYPE_STACK_TRACE:
3656 if (func_id != BPF_FUNC_get_stackid)
3659 case BPF_MAP_TYPE_CGROUP_ARRAY:
3660 if (func_id != BPF_FUNC_skb_under_cgroup &&
3661 func_id != BPF_FUNC_current_task_under_cgroup)
3664 case BPF_MAP_TYPE_CGROUP_STORAGE:
3665 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
3666 if (func_id != BPF_FUNC_get_local_storage)
3669 case BPF_MAP_TYPE_DEVMAP:
3670 case BPF_MAP_TYPE_DEVMAP_HASH:
3671 if (func_id != BPF_FUNC_redirect_map &&
3672 func_id != BPF_FUNC_map_lookup_elem)
3675 /* Restrict bpf side of cpumap and xskmap, open when use-cases
3678 case BPF_MAP_TYPE_CPUMAP:
3679 if (func_id != BPF_FUNC_redirect_map)
3682 case BPF_MAP_TYPE_XSKMAP:
3683 if (func_id != BPF_FUNC_redirect_map &&
3684 func_id != BPF_FUNC_map_lookup_elem)
3687 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
3688 case BPF_MAP_TYPE_HASH_OF_MAPS:
3689 if (func_id != BPF_FUNC_map_lookup_elem)
3692 case BPF_MAP_TYPE_SOCKMAP:
3693 if (func_id != BPF_FUNC_sk_redirect_map &&
3694 func_id != BPF_FUNC_sock_map_update &&
3695 func_id != BPF_FUNC_map_delete_elem &&
3696 func_id != BPF_FUNC_msg_redirect_map &&
3697 func_id != BPF_FUNC_sk_select_reuseport)
3700 case BPF_MAP_TYPE_SOCKHASH:
3701 if (func_id != BPF_FUNC_sk_redirect_hash &&
3702 func_id != BPF_FUNC_sock_hash_update &&
3703 func_id != BPF_FUNC_map_delete_elem &&
3704 func_id != BPF_FUNC_msg_redirect_hash &&
3705 func_id != BPF_FUNC_sk_select_reuseport)
3708 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
3709 if (func_id != BPF_FUNC_sk_select_reuseport)
3712 case BPF_MAP_TYPE_QUEUE:
3713 case BPF_MAP_TYPE_STACK:
3714 if (func_id != BPF_FUNC_map_peek_elem &&
3715 func_id != BPF_FUNC_map_pop_elem &&
3716 func_id != BPF_FUNC_map_push_elem)
3719 case BPF_MAP_TYPE_SK_STORAGE:
3720 if (func_id != BPF_FUNC_sk_storage_get &&
3721 func_id != BPF_FUNC_sk_storage_delete)
3728 /* ... and second from the function itself. */
3730 case BPF_FUNC_tail_call:
3731 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
3733 if (env->subprog_cnt > 1) {
3734 verbose(env, "tail_calls are not allowed in programs with bpf-to-bpf calls\n");
3738 case BPF_FUNC_perf_event_read:
3739 case BPF_FUNC_perf_event_output:
3740 case BPF_FUNC_perf_event_read_value:
3741 case BPF_FUNC_skb_output:
3742 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
3745 case BPF_FUNC_get_stackid:
3746 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
3749 case BPF_FUNC_current_task_under_cgroup:
3750 case BPF_FUNC_skb_under_cgroup:
3751 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
3754 case BPF_FUNC_redirect_map:
3755 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
3756 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
3757 map->map_type != BPF_MAP_TYPE_CPUMAP &&
3758 map->map_type != BPF_MAP_TYPE_XSKMAP)
3761 case BPF_FUNC_sk_redirect_map:
3762 case BPF_FUNC_msg_redirect_map:
3763 case BPF_FUNC_sock_map_update:
3764 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
3767 case BPF_FUNC_sk_redirect_hash:
3768 case BPF_FUNC_msg_redirect_hash:
3769 case BPF_FUNC_sock_hash_update:
3770 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
3773 case BPF_FUNC_get_local_storage:
3774 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
3775 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
3778 case BPF_FUNC_sk_select_reuseport:
3779 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
3780 map->map_type != BPF_MAP_TYPE_SOCKMAP &&
3781 map->map_type != BPF_MAP_TYPE_SOCKHASH)
3784 case BPF_FUNC_map_peek_elem:
3785 case BPF_FUNC_map_pop_elem:
3786 case BPF_FUNC_map_push_elem:
3787 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
3788 map->map_type != BPF_MAP_TYPE_STACK)
3791 case BPF_FUNC_sk_storage_get:
3792 case BPF_FUNC_sk_storage_delete:
3793 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
3802 verbose(env, "cannot pass map_type %d into func %s#%d\n",
3803 map->map_type, func_id_name(func_id), func_id);
3807 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
3811 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
3813 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
3815 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
3817 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
3819 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
3822 /* We only support one arg being in raw mode at the moment,
3823 * which is sufficient for the helper functions we have
3829 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
3830 enum bpf_arg_type arg_next)
3832 return (arg_type_is_mem_ptr(arg_curr) &&
3833 !arg_type_is_mem_size(arg_next)) ||
3834 (!arg_type_is_mem_ptr(arg_curr) &&
3835 arg_type_is_mem_size(arg_next));
3838 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
3840 /* bpf_xxx(..., buf, len) call will access 'len'
3841 * bytes from memory 'buf'. Both arg types need
3842 * to be paired, so make sure there's no buggy
3843 * helper function specification.
3845 if (arg_type_is_mem_size(fn->arg1_type) ||
3846 arg_type_is_mem_ptr(fn->arg5_type) ||
3847 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
3848 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
3849 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
3850 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
3856 static bool check_refcount_ok(const struct bpf_func_proto *fn, int func_id)
3860 if (arg_type_may_be_refcounted(fn->arg1_type))
3862 if (arg_type_may_be_refcounted(fn->arg2_type))
3864 if (arg_type_may_be_refcounted(fn->arg3_type))
3866 if (arg_type_may_be_refcounted(fn->arg4_type))
3868 if (arg_type_may_be_refcounted(fn->arg5_type))
3871 /* A reference acquiring function cannot acquire
3872 * another refcounted ptr.
3874 if (is_acquire_function(func_id) && count)
3877 /* We only support one arg being unreferenced at the moment,
3878 * which is sufficient for the helper functions we have right now.
3883 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
3885 return check_raw_mode_ok(fn) &&
3886 check_arg_pair_ok(fn) &&
3887 check_refcount_ok(fn, func_id) ? 0 : -EINVAL;
3890 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
3891 * are now invalid, so turn them into unknown SCALAR_VALUE.
3893 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
3894 struct bpf_func_state *state)
3896 struct bpf_reg_state *regs = state->regs, *reg;
3899 for (i = 0; i < MAX_BPF_REG; i++)
3900 if (reg_is_pkt_pointer_any(®s[i]))
3901 mark_reg_unknown(env, regs, i);
3903 bpf_for_each_spilled_reg(i, state, reg) {
3906 if (reg_is_pkt_pointer_any(reg))
3907 __mark_reg_unknown(env, reg);
3911 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
3913 struct bpf_verifier_state *vstate = env->cur_state;
3916 for (i = 0; i <= vstate->curframe; i++)
3917 __clear_all_pkt_pointers(env, vstate->frame[i]);
3920 static void release_reg_references(struct bpf_verifier_env *env,
3921 struct bpf_func_state *state,
3924 struct bpf_reg_state *regs = state->regs, *reg;
3927 for (i = 0; i < MAX_BPF_REG; i++)
3928 if (regs[i].ref_obj_id == ref_obj_id)
3929 mark_reg_unknown(env, regs, i);
3931 bpf_for_each_spilled_reg(i, state, reg) {
3934 if (reg->ref_obj_id == ref_obj_id)
3935 __mark_reg_unknown(env, reg);
3939 /* The pointer with the specified id has released its reference to kernel
3940 * resources. Identify all copies of the same pointer and clear the reference.
3942 static int release_reference(struct bpf_verifier_env *env,
3945 struct bpf_verifier_state *vstate = env->cur_state;
3949 err = release_reference_state(cur_func(env), ref_obj_id);
3953 for (i = 0; i <= vstate->curframe; i++)
3954 release_reg_references(env, vstate->frame[i], ref_obj_id);
3959 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
3960 struct bpf_reg_state *regs)
3964 /* after the call registers r0 - r5 were scratched */
3965 for (i = 0; i < CALLER_SAVED_REGS; i++) {
3966 mark_reg_not_init(env, regs, caller_saved[i]);
3967 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
3971 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
3974 struct bpf_verifier_state *state = env->cur_state;
3975 struct bpf_func_info_aux *func_info_aux;
3976 struct bpf_func_state *caller, *callee;
3977 int i, err, subprog, target_insn;
3978 bool is_global = false;
3980 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
3981 verbose(env, "the call stack of %d frames is too deep\n",
3982 state->curframe + 2);
3986 target_insn = *insn_idx + insn->imm;
3987 subprog = find_subprog(env, target_insn + 1);
3989 verbose(env, "verifier bug. No program starts at insn %d\n",
3994 caller = state->frame[state->curframe];
3995 if (state->frame[state->curframe + 1]) {
3996 verbose(env, "verifier bug. Frame %d already allocated\n",
3997 state->curframe + 1);
4001 func_info_aux = env->prog->aux->func_info_aux;
4003 is_global = func_info_aux[subprog].linkage == BTF_FUNC_GLOBAL;
4004 err = btf_check_func_arg_match(env, subprog, caller->regs);
4009 verbose(env, "Caller passes invalid args into func#%d\n",
4013 if (env->log.level & BPF_LOG_LEVEL)
4015 "Func#%d is global and valid. Skipping.\n",
4017 clear_caller_saved_regs(env, caller->regs);
4019 /* All global functions return SCALAR_VALUE */
4020 mark_reg_unknown(env, caller->regs, BPF_REG_0);
4022 /* continue with next insn after call */
4027 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
4030 state->frame[state->curframe + 1] = callee;
4032 /* callee cannot access r0, r6 - r9 for reading and has to write
4033 * into its own stack before reading from it.
4034 * callee can read/write into caller's stack
4036 init_func_state(env, callee,
4037 /* remember the callsite, it will be used by bpf_exit */
4038 *insn_idx /* callsite */,
4039 state->curframe + 1 /* frameno within this callchain */,
4040 subprog /* subprog number within this prog */);
4042 /* Transfer references to the callee */
4043 err = transfer_reference_state(callee, caller);
4047 /* copy r1 - r5 args that callee can access. The copy includes parent
4048 * pointers, which connects us up to the liveness chain
4050 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
4051 callee->regs[i] = caller->regs[i];
4053 clear_caller_saved_regs(env, caller->regs);
4055 /* only increment it after check_reg_arg() finished */
4058 /* and go analyze first insn of the callee */
4059 *insn_idx = target_insn;
4061 if (env->log.level & BPF_LOG_LEVEL) {
4062 verbose(env, "caller:\n");
4063 print_verifier_state(env, caller);
4064 verbose(env, "callee:\n");
4065 print_verifier_state(env, callee);
4070 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
4072 struct bpf_verifier_state *state = env->cur_state;
4073 struct bpf_func_state *caller, *callee;
4074 struct bpf_reg_state *r0;
4077 callee = state->frame[state->curframe];
4078 r0 = &callee->regs[BPF_REG_0];
4079 if (r0->type == PTR_TO_STACK) {
4080 /* technically it's ok to return caller's stack pointer
4081 * (or caller's caller's pointer) back to the caller,
4082 * since these pointers are valid. Only current stack
4083 * pointer will be invalid as soon as function exits,
4084 * but let's be conservative
4086 verbose(env, "cannot return stack pointer to the caller\n");
4091 caller = state->frame[state->curframe];
4092 /* return to the caller whatever r0 had in the callee */
4093 caller->regs[BPF_REG_0] = *r0;
4095 /* Transfer references to the caller */
4096 err = transfer_reference_state(caller, callee);
4100 *insn_idx = callee->callsite + 1;
4101 if (env->log.level & BPF_LOG_LEVEL) {
4102 verbose(env, "returning from callee:\n");
4103 print_verifier_state(env, callee);
4104 verbose(env, "to caller at %d:\n", *insn_idx);
4105 print_verifier_state(env, caller);
4107 /* clear everything in the callee */
4108 free_func_state(callee);
4109 state->frame[state->curframe + 1] = NULL;
4113 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
4115 struct bpf_call_arg_meta *meta)
4117 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
4119 if (ret_type != RET_INTEGER ||
4120 (func_id != BPF_FUNC_get_stack &&
4121 func_id != BPF_FUNC_probe_read_str))
4124 ret_reg->smax_value = meta->msize_smax_value;
4125 ret_reg->umax_value = meta->msize_umax_value;
4126 __reg_deduce_bounds(ret_reg);
4127 __reg_bound_offset(ret_reg);
4131 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
4132 int func_id, int insn_idx)
4134 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
4135 struct bpf_map *map = meta->map_ptr;
4137 if (func_id != BPF_FUNC_tail_call &&
4138 func_id != BPF_FUNC_map_lookup_elem &&
4139 func_id != BPF_FUNC_map_update_elem &&
4140 func_id != BPF_FUNC_map_delete_elem &&
4141 func_id != BPF_FUNC_map_push_elem &&
4142 func_id != BPF_FUNC_map_pop_elem &&
4143 func_id != BPF_FUNC_map_peek_elem)
4147 verbose(env, "kernel subsystem misconfigured verifier\n");
4151 /* In case of read-only, some additional restrictions
4152 * need to be applied in order to prevent altering the
4153 * state of the map from program side.
4155 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
4156 (func_id == BPF_FUNC_map_delete_elem ||
4157 func_id == BPF_FUNC_map_update_elem ||
4158 func_id == BPF_FUNC_map_push_elem ||
4159 func_id == BPF_FUNC_map_pop_elem)) {
4160 verbose(env, "write into map forbidden\n");
4164 if (!BPF_MAP_PTR(aux->map_ptr_state))
4165 bpf_map_ptr_store(aux, meta->map_ptr,
4166 meta->map_ptr->unpriv_array);
4167 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
4168 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
4169 meta->map_ptr->unpriv_array);
4174 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
4175 int func_id, int insn_idx)
4177 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
4178 struct bpf_reg_state *regs = cur_regs(env), *reg;
4179 struct bpf_map *map = meta->map_ptr;
4184 if (func_id != BPF_FUNC_tail_call)
4186 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
4187 verbose(env, "kernel subsystem misconfigured verifier\n");
4191 range = tnum_range(0, map->max_entries - 1);
4192 reg = ®s[BPF_REG_3];
4194 if (!register_is_const(reg) || !tnum_in(range, reg->var_off)) {
4195 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
4199 err = mark_chain_precision(env, BPF_REG_3);
4203 val = reg->var_off.value;
4204 if (bpf_map_key_unseen(aux))
4205 bpf_map_key_store(aux, val);
4206 else if (!bpf_map_key_poisoned(aux) &&
4207 bpf_map_key_immediate(aux) != val)
4208 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
4212 static int check_reference_leak(struct bpf_verifier_env *env)
4214 struct bpf_func_state *state = cur_func(env);
4217 for (i = 0; i < state->acquired_refs; i++) {
4218 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
4219 state->refs[i].id, state->refs[i].insn_idx);
4221 return state->acquired_refs ? -EINVAL : 0;
4224 static int check_helper_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
4226 const struct bpf_func_proto *fn = NULL;
4227 struct bpf_reg_state *regs;
4228 struct bpf_call_arg_meta meta;
4232 /* find function prototype */
4233 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
4234 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
4239 if (env->ops->get_func_proto)
4240 fn = env->ops->get_func_proto(func_id, env->prog);
4242 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
4247 /* eBPF programs must be GPL compatible to use GPL-ed functions */
4248 if (!env->prog->gpl_compatible && fn->gpl_only) {
4249 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
4253 /* With LD_ABS/IND some JITs save/restore skb from r1. */
4254 changes_data = bpf_helper_changes_pkt_data(fn->func);
4255 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
4256 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
4257 func_id_name(func_id), func_id);
4261 memset(&meta, 0, sizeof(meta));
4262 meta.pkt_access = fn->pkt_access;
4264 err = check_func_proto(fn, func_id);
4266 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
4267 func_id_name(func_id), func_id);
4271 meta.func_id = func_id;
4273 for (i = 0; i < 5; i++) {
4274 err = btf_resolve_helper_id(&env->log, fn, i);
4277 err = check_func_arg(env, BPF_REG_1 + i, fn->arg_type[i], &meta);
4282 err = record_func_map(env, &meta, func_id, insn_idx);
4286 err = record_func_key(env, &meta, func_id, insn_idx);
4290 /* Mark slots with STACK_MISC in case of raw mode, stack offset
4291 * is inferred from register state.
4293 for (i = 0; i < meta.access_size; i++) {
4294 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
4295 BPF_WRITE, -1, false);
4300 if (func_id == BPF_FUNC_tail_call) {
4301 err = check_reference_leak(env);
4303 verbose(env, "tail_call would lead to reference leak\n");
4306 } else if (is_release_function(func_id)) {
4307 err = release_reference(env, meta.ref_obj_id);
4309 verbose(env, "func %s#%d reference has not been acquired before\n",
4310 func_id_name(func_id), func_id);
4315 regs = cur_regs(env);
4317 /* check that flags argument in get_local_storage(map, flags) is 0,
4318 * this is required because get_local_storage() can't return an error.
4320 if (func_id == BPF_FUNC_get_local_storage &&
4321 !register_is_null(®s[BPF_REG_2])) {
4322 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
4326 /* reset caller saved regs */
4327 for (i = 0; i < CALLER_SAVED_REGS; i++) {
4328 mark_reg_not_init(env, regs, caller_saved[i]);
4329 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
4332 /* helper call returns 64-bit value. */
4333 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
4335 /* update return register (already marked as written above) */
4336 if (fn->ret_type == RET_INTEGER) {
4337 /* sets type to SCALAR_VALUE */
4338 mark_reg_unknown(env, regs, BPF_REG_0);
4339 } else if (fn->ret_type == RET_VOID) {
4340 regs[BPF_REG_0].type = NOT_INIT;
4341 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL ||
4342 fn->ret_type == RET_PTR_TO_MAP_VALUE) {
4343 /* There is no offset yet applied, variable or fixed */
4344 mark_reg_known_zero(env, regs, BPF_REG_0);
4345 /* remember map_ptr, so that check_map_access()
4346 * can check 'value_size' boundary of memory access
4347 * to map element returned from bpf_map_lookup_elem()
4349 if (meta.map_ptr == NULL) {
4351 "kernel subsystem misconfigured verifier\n");
4354 regs[BPF_REG_0].map_ptr = meta.map_ptr;
4355 if (fn->ret_type == RET_PTR_TO_MAP_VALUE) {
4356 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE;
4357 if (map_value_has_spin_lock(meta.map_ptr))
4358 regs[BPF_REG_0].id = ++env->id_gen;
4360 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
4361 regs[BPF_REG_0].id = ++env->id_gen;
4363 } else if (fn->ret_type == RET_PTR_TO_SOCKET_OR_NULL) {
4364 mark_reg_known_zero(env, regs, BPF_REG_0);
4365 regs[BPF_REG_0].type = PTR_TO_SOCKET_OR_NULL;
4366 regs[BPF_REG_0].id = ++env->id_gen;
4367 } else if (fn->ret_type == RET_PTR_TO_SOCK_COMMON_OR_NULL) {
4368 mark_reg_known_zero(env, regs, BPF_REG_0);
4369 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON_OR_NULL;
4370 regs[BPF_REG_0].id = ++env->id_gen;
4371 } else if (fn->ret_type == RET_PTR_TO_TCP_SOCK_OR_NULL) {
4372 mark_reg_known_zero(env, regs, BPF_REG_0);
4373 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK_OR_NULL;
4374 regs[BPF_REG_0].id = ++env->id_gen;
4376 verbose(env, "unknown return type %d of func %s#%d\n",
4377 fn->ret_type, func_id_name(func_id), func_id);
4381 if (is_ptr_cast_function(func_id)) {
4382 /* For release_reference() */
4383 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
4384 } else if (is_acquire_function(func_id)) {
4385 int id = acquire_reference_state(env, insn_idx);
4389 /* For mark_ptr_or_null_reg() */
4390 regs[BPF_REG_0].id = id;
4391 /* For release_reference() */
4392 regs[BPF_REG_0].ref_obj_id = id;
4395 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
4397 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
4401 if (func_id == BPF_FUNC_get_stack && !env->prog->has_callchain_buf) {
4402 const char *err_str;
4404 #ifdef CONFIG_PERF_EVENTS
4405 err = get_callchain_buffers(sysctl_perf_event_max_stack);
4406 err_str = "cannot get callchain buffer for func %s#%d\n";
4409 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
4412 verbose(env, err_str, func_id_name(func_id), func_id);
4416 env->prog->has_callchain_buf = true;
4420 clear_all_pkt_pointers(env);
4424 static bool signed_add_overflows(s64 a, s64 b)
4426 /* Do the add in u64, where overflow is well-defined */
4427 s64 res = (s64)((u64)a + (u64)b);
4434 static bool signed_sub_overflows(s64 a, s64 b)
4436 /* Do the sub in u64, where overflow is well-defined */
4437 s64 res = (s64)((u64)a - (u64)b);
4444 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
4445 const struct bpf_reg_state *reg,
4446 enum bpf_reg_type type)
4448 bool known = tnum_is_const(reg->var_off);
4449 s64 val = reg->var_off.value;
4450 s64 smin = reg->smin_value;
4452 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
4453 verbose(env, "math between %s pointer and %lld is not allowed\n",
4454 reg_type_str[type], val);
4458 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
4459 verbose(env, "%s pointer offset %d is not allowed\n",
4460 reg_type_str[type], reg->off);
4464 if (smin == S64_MIN) {
4465 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
4466 reg_type_str[type]);
4470 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
4471 verbose(env, "value %lld makes %s pointer be out of bounds\n",
4472 smin, reg_type_str[type]);
4479 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
4481 return &env->insn_aux_data[env->insn_idx];
4484 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
4485 u32 *ptr_limit, u8 opcode, bool off_is_neg)
4487 bool mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
4488 (opcode == BPF_SUB && !off_is_neg);
4491 switch (ptr_reg->type) {
4493 /* Indirect variable offset stack access is prohibited in
4494 * unprivileged mode so it's not handled here.
4496 off = ptr_reg->off + ptr_reg->var_off.value;
4498 *ptr_limit = MAX_BPF_STACK + off;
4502 case PTR_TO_MAP_VALUE:
4504 *ptr_limit = ptr_reg->umax_value + ptr_reg->off;
4506 off = ptr_reg->smin_value + ptr_reg->off;
4507 *ptr_limit = ptr_reg->map_ptr->value_size - off;
4515 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
4516 const struct bpf_insn *insn)
4518 return env->allow_ptr_leaks || BPF_SRC(insn->code) == BPF_K;
4521 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
4522 u32 alu_state, u32 alu_limit)
4524 /* If we arrived here from different branches with different
4525 * state or limits to sanitize, then this won't work.
4527 if (aux->alu_state &&
4528 (aux->alu_state != alu_state ||
4529 aux->alu_limit != alu_limit))
4532 /* Corresponding fixup done in fixup_bpf_calls(). */
4533 aux->alu_state = alu_state;
4534 aux->alu_limit = alu_limit;
4538 static int sanitize_val_alu(struct bpf_verifier_env *env,
4539 struct bpf_insn *insn)
4541 struct bpf_insn_aux_data *aux = cur_aux(env);
4543 if (can_skip_alu_sanitation(env, insn))
4546 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
4549 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
4550 struct bpf_insn *insn,
4551 const struct bpf_reg_state *ptr_reg,
4552 struct bpf_reg_state *dst_reg,
4555 struct bpf_verifier_state *vstate = env->cur_state;
4556 struct bpf_insn_aux_data *aux = cur_aux(env);
4557 bool ptr_is_dst_reg = ptr_reg == dst_reg;
4558 u8 opcode = BPF_OP(insn->code);
4559 u32 alu_state, alu_limit;
4560 struct bpf_reg_state tmp;
4563 if (can_skip_alu_sanitation(env, insn))
4566 /* We already marked aux for masking from non-speculative
4567 * paths, thus we got here in the first place. We only care
4568 * to explore bad access from here.
4570 if (vstate->speculative)
4573 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
4574 alu_state |= ptr_is_dst_reg ?
4575 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
4577 if (retrieve_ptr_limit(ptr_reg, &alu_limit, opcode, off_is_neg))
4579 if (update_alu_sanitation_state(aux, alu_state, alu_limit))
4582 /* Simulate and find potential out-of-bounds access under
4583 * speculative execution from truncation as a result of
4584 * masking when off was not within expected range. If off
4585 * sits in dst, then we temporarily need to move ptr there
4586 * to simulate dst (== 0) +/-= ptr. Needed, for example,
4587 * for cases where we use K-based arithmetic in one direction
4588 * and truncated reg-based in the other in order to explore
4591 if (!ptr_is_dst_reg) {
4593 *dst_reg = *ptr_reg;
4595 ret = push_stack(env, env->insn_idx + 1, env->insn_idx, true);
4596 if (!ptr_is_dst_reg && ret)
4598 return !ret ? -EFAULT : 0;
4601 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
4602 * Caller should also handle BPF_MOV case separately.
4603 * If we return -EACCES, caller may want to try again treating pointer as a
4604 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
4606 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
4607 struct bpf_insn *insn,
4608 const struct bpf_reg_state *ptr_reg,
4609 const struct bpf_reg_state *off_reg)
4611 struct bpf_verifier_state *vstate = env->cur_state;
4612 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4613 struct bpf_reg_state *regs = state->regs, *dst_reg;
4614 bool known = tnum_is_const(off_reg->var_off);
4615 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
4616 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
4617 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
4618 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
4619 u32 dst = insn->dst_reg, src = insn->src_reg;
4620 u8 opcode = BPF_OP(insn->code);
4623 dst_reg = ®s[dst];
4625 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
4626 smin_val > smax_val || umin_val > umax_val) {
4627 /* Taint dst register if offset had invalid bounds derived from
4628 * e.g. dead branches.
4630 __mark_reg_unknown(env, dst_reg);
4634 if (BPF_CLASS(insn->code) != BPF_ALU64) {
4635 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
4637 "R%d 32-bit pointer arithmetic prohibited\n",
4642 switch (ptr_reg->type) {
4643 case PTR_TO_MAP_VALUE_OR_NULL:
4644 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
4645 dst, reg_type_str[ptr_reg->type]);
4647 case CONST_PTR_TO_MAP:
4648 case PTR_TO_PACKET_END:
4650 case PTR_TO_SOCKET_OR_NULL:
4651 case PTR_TO_SOCK_COMMON:
4652 case PTR_TO_SOCK_COMMON_OR_NULL:
4653 case PTR_TO_TCP_SOCK:
4654 case PTR_TO_TCP_SOCK_OR_NULL:
4655 case PTR_TO_XDP_SOCK:
4656 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
4657 dst, reg_type_str[ptr_reg->type]);
4659 case PTR_TO_MAP_VALUE:
4660 if (!env->allow_ptr_leaks && !known && (smin_val < 0) != (smax_val < 0)) {
4661 verbose(env, "R%d has unknown scalar with mixed signed bounds, pointer arithmetic with it prohibited for !root\n",
4662 off_reg == dst_reg ? dst : src);
4670 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
4671 * The id may be overwritten later if we create a new variable offset.
4673 dst_reg->type = ptr_reg->type;
4674 dst_reg->id = ptr_reg->id;
4676 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
4677 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
4682 ret = sanitize_ptr_alu(env, insn, ptr_reg, dst_reg, smin_val < 0);
4684 verbose(env, "R%d tried to add from different maps or paths\n", dst);
4687 /* We can take a fixed offset as long as it doesn't overflow
4688 * the s32 'off' field
4690 if (known && (ptr_reg->off + smin_val ==
4691 (s64)(s32)(ptr_reg->off + smin_val))) {
4692 /* pointer += K. Accumulate it into fixed offset */
4693 dst_reg->smin_value = smin_ptr;
4694 dst_reg->smax_value = smax_ptr;
4695 dst_reg->umin_value = umin_ptr;
4696 dst_reg->umax_value = umax_ptr;
4697 dst_reg->var_off = ptr_reg->var_off;
4698 dst_reg->off = ptr_reg->off + smin_val;
4699 dst_reg->raw = ptr_reg->raw;
4702 /* A new variable offset is created. Note that off_reg->off
4703 * == 0, since it's a scalar.
4704 * dst_reg gets the pointer type and since some positive
4705 * integer value was added to the pointer, give it a new 'id'
4706 * if it's a PTR_TO_PACKET.
4707 * this creates a new 'base' pointer, off_reg (variable) gets
4708 * added into the variable offset, and we copy the fixed offset
4711 if (signed_add_overflows(smin_ptr, smin_val) ||
4712 signed_add_overflows(smax_ptr, smax_val)) {
4713 dst_reg->smin_value = S64_MIN;
4714 dst_reg->smax_value = S64_MAX;
4716 dst_reg->smin_value = smin_ptr + smin_val;
4717 dst_reg->smax_value = smax_ptr + smax_val;
4719 if (umin_ptr + umin_val < umin_ptr ||
4720 umax_ptr + umax_val < umax_ptr) {
4721 dst_reg->umin_value = 0;
4722 dst_reg->umax_value = U64_MAX;
4724 dst_reg->umin_value = umin_ptr + umin_val;
4725 dst_reg->umax_value = umax_ptr + umax_val;
4727 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
4728 dst_reg->off = ptr_reg->off;
4729 dst_reg->raw = ptr_reg->raw;
4730 if (reg_is_pkt_pointer(ptr_reg)) {
4731 dst_reg->id = ++env->id_gen;
4732 /* something was added to pkt_ptr, set range to zero */
4737 ret = sanitize_ptr_alu(env, insn, ptr_reg, dst_reg, smin_val < 0);
4739 verbose(env, "R%d tried to sub from different maps or paths\n", dst);
4742 if (dst_reg == off_reg) {
4743 /* scalar -= pointer. Creates an unknown scalar */
4744 verbose(env, "R%d tried to subtract pointer from scalar\n",
4748 /* We don't allow subtraction from FP, because (according to
4749 * test_verifier.c test "invalid fp arithmetic", JITs might not
4750 * be able to deal with it.
4752 if (ptr_reg->type == PTR_TO_STACK) {
4753 verbose(env, "R%d subtraction from stack pointer prohibited\n",
4757 if (known && (ptr_reg->off - smin_val ==
4758 (s64)(s32)(ptr_reg->off - smin_val))) {
4759 /* pointer -= K. Subtract it from fixed offset */
4760 dst_reg->smin_value = smin_ptr;
4761 dst_reg->smax_value = smax_ptr;
4762 dst_reg->umin_value = umin_ptr;
4763 dst_reg->umax_value = umax_ptr;
4764 dst_reg->var_off = ptr_reg->var_off;
4765 dst_reg->id = ptr_reg->id;
4766 dst_reg->off = ptr_reg->off - smin_val;
4767 dst_reg->raw = ptr_reg->raw;
4770 /* A new variable offset is created. If the subtrahend is known
4771 * nonnegative, then any reg->range we had before is still good.
4773 if (signed_sub_overflows(smin_ptr, smax_val) ||
4774 signed_sub_overflows(smax_ptr, smin_val)) {
4775 /* Overflow possible, we know nothing */
4776 dst_reg->smin_value = S64_MIN;
4777 dst_reg->smax_value = S64_MAX;
4779 dst_reg->smin_value = smin_ptr - smax_val;
4780 dst_reg->smax_value = smax_ptr - smin_val;
4782 if (umin_ptr < umax_val) {
4783 /* Overflow possible, we know nothing */
4784 dst_reg->umin_value = 0;
4785 dst_reg->umax_value = U64_MAX;
4787 /* Cannot overflow (as long as bounds are consistent) */
4788 dst_reg->umin_value = umin_ptr - umax_val;
4789 dst_reg->umax_value = umax_ptr - umin_val;
4791 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
4792 dst_reg->off = ptr_reg->off;
4793 dst_reg->raw = ptr_reg->raw;
4794 if (reg_is_pkt_pointer(ptr_reg)) {
4795 dst_reg->id = ++env->id_gen;
4796 /* something was added to pkt_ptr, set range to zero */
4804 /* bitwise ops on pointers are troublesome, prohibit. */
4805 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
4806 dst, bpf_alu_string[opcode >> 4]);
4809 /* other operators (e.g. MUL,LSH) produce non-pointer results */
4810 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
4811 dst, bpf_alu_string[opcode >> 4]);
4815 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
4818 __update_reg_bounds(dst_reg);
4819 __reg_deduce_bounds(dst_reg);
4820 __reg_bound_offset(dst_reg);
4822 /* For unprivileged we require that resulting offset must be in bounds
4823 * in order to be able to sanitize access later on.
4825 if (!env->allow_ptr_leaks) {
4826 if (dst_reg->type == PTR_TO_MAP_VALUE &&
4827 check_map_access(env, dst, dst_reg->off, 1, false)) {
4828 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
4829 "prohibited for !root\n", dst);
4831 } else if (dst_reg->type == PTR_TO_STACK &&
4832 check_stack_access(env, dst_reg, dst_reg->off +
4833 dst_reg->var_off.value, 1)) {
4834 verbose(env, "R%d stack pointer arithmetic goes out of range, "
4835 "prohibited for !root\n", dst);
4843 /* WARNING: This function does calculations on 64-bit values, but the actual
4844 * execution may occur on 32-bit values. Therefore, things like bitshifts
4845 * need extra checks in the 32-bit case.
4847 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
4848 struct bpf_insn *insn,
4849 struct bpf_reg_state *dst_reg,
4850 struct bpf_reg_state src_reg)
4852 struct bpf_reg_state *regs = cur_regs(env);
4853 u8 opcode = BPF_OP(insn->code);
4854 bool src_known, dst_known;
4855 s64 smin_val, smax_val;
4856 u64 umin_val, umax_val;
4857 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
4858 u32 dst = insn->dst_reg;
4861 if (insn_bitness == 32) {
4862 /* Relevant for 32-bit RSH: Information can propagate towards
4863 * LSB, so it isn't sufficient to only truncate the output to
4866 coerce_reg_to_size(dst_reg, 4);
4867 coerce_reg_to_size(&src_reg, 4);
4870 smin_val = src_reg.smin_value;
4871 smax_val = src_reg.smax_value;
4872 umin_val = src_reg.umin_value;
4873 umax_val = src_reg.umax_value;
4874 src_known = tnum_is_const(src_reg.var_off);
4875 dst_known = tnum_is_const(dst_reg->var_off);
4877 if ((src_known && (smin_val != smax_val || umin_val != umax_val)) ||
4878 smin_val > smax_val || umin_val > umax_val) {
4879 /* Taint dst register if offset had invalid bounds derived from
4880 * e.g. dead branches.
4882 __mark_reg_unknown(env, dst_reg);
4887 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
4888 __mark_reg_unknown(env, dst_reg);
4894 ret = sanitize_val_alu(env, insn);
4896 verbose(env, "R%d tried to add from different pointers or scalars\n", dst);
4899 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
4900 signed_add_overflows(dst_reg->smax_value, smax_val)) {
4901 dst_reg->smin_value = S64_MIN;
4902 dst_reg->smax_value = S64_MAX;
4904 dst_reg->smin_value += smin_val;
4905 dst_reg->smax_value += smax_val;
4907 if (dst_reg->umin_value + umin_val < umin_val ||
4908 dst_reg->umax_value + umax_val < umax_val) {
4909 dst_reg->umin_value = 0;
4910 dst_reg->umax_value = U64_MAX;
4912 dst_reg->umin_value += umin_val;
4913 dst_reg->umax_value += umax_val;
4915 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
4918 ret = sanitize_val_alu(env, insn);
4920 verbose(env, "R%d tried to sub from different pointers or scalars\n", dst);
4923 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
4924 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
4925 /* Overflow possible, we know nothing */
4926 dst_reg->smin_value = S64_MIN;
4927 dst_reg->smax_value = S64_MAX;
4929 dst_reg->smin_value -= smax_val;
4930 dst_reg->smax_value -= smin_val;
4932 if (dst_reg->umin_value < umax_val) {
4933 /* Overflow possible, we know nothing */
4934 dst_reg->umin_value = 0;
4935 dst_reg->umax_value = U64_MAX;
4937 /* Cannot overflow (as long as bounds are consistent) */
4938 dst_reg->umin_value -= umax_val;
4939 dst_reg->umax_value -= umin_val;
4941 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
4944 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
4945 if (smin_val < 0 || dst_reg->smin_value < 0) {
4946 /* Ain't nobody got time to multiply that sign */
4947 __mark_reg_unbounded(dst_reg);
4948 __update_reg_bounds(dst_reg);
4951 /* Both values are positive, so we can work with unsigned and
4952 * copy the result to signed (unless it exceeds S64_MAX).
4954 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
4955 /* Potential overflow, we know nothing */
4956 __mark_reg_unbounded(dst_reg);
4957 /* (except what we can learn from the var_off) */
4958 __update_reg_bounds(dst_reg);
4961 dst_reg->umin_value *= umin_val;
4962 dst_reg->umax_value *= umax_val;
4963 if (dst_reg->umax_value > S64_MAX) {
4964 /* Overflow possible, we know nothing */
4965 dst_reg->smin_value = S64_MIN;
4966 dst_reg->smax_value = S64_MAX;
4968 dst_reg->smin_value = dst_reg->umin_value;
4969 dst_reg->smax_value = dst_reg->umax_value;
4973 if (src_known && dst_known) {
4974 __mark_reg_known(dst_reg, dst_reg->var_off.value &
4975 src_reg.var_off.value);
4978 /* We get our minimum from the var_off, since that's inherently
4979 * bitwise. Our maximum is the minimum of the operands' maxima.
4981 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
4982 dst_reg->umin_value = dst_reg->var_off.value;
4983 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
4984 if (dst_reg->smin_value < 0 || smin_val < 0) {
4985 /* Lose signed bounds when ANDing negative numbers,
4986 * ain't nobody got time for that.
4988 dst_reg->smin_value = S64_MIN;
4989 dst_reg->smax_value = S64_MAX;
4991 /* ANDing two positives gives a positive, so safe to
4992 * cast result into s64.
4994 dst_reg->smin_value = dst_reg->umin_value;
4995 dst_reg->smax_value = dst_reg->umax_value;
4997 /* We may learn something more from the var_off */
4998 __update_reg_bounds(dst_reg);
5001 if (src_known && dst_known) {
5002 __mark_reg_known(dst_reg, dst_reg->var_off.value |
5003 src_reg.var_off.value);
5006 /* We get our maximum from the var_off, and our minimum is the
5007 * maximum of the operands' minima
5009 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
5010 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
5011 dst_reg->umax_value = dst_reg->var_off.value |
5012 dst_reg->var_off.mask;
5013 if (dst_reg->smin_value < 0 || smin_val < 0) {
5014 /* Lose signed bounds when ORing negative numbers,
5015 * ain't nobody got time for that.
5017 dst_reg->smin_value = S64_MIN;
5018 dst_reg->smax_value = S64_MAX;
5020 /* ORing two positives gives a positive, so safe to
5021 * cast result into s64.
5023 dst_reg->smin_value = dst_reg->umin_value;
5024 dst_reg->smax_value = dst_reg->umax_value;
5026 /* We may learn something more from the var_off */
5027 __update_reg_bounds(dst_reg);
5030 if (umax_val >= insn_bitness) {
5031 /* Shifts greater than 31 or 63 are undefined.
5032 * This includes shifts by a negative number.
5034 mark_reg_unknown(env, regs, insn->dst_reg);
5037 /* We lose all sign bit information (except what we can pick
5040 dst_reg->smin_value = S64_MIN;
5041 dst_reg->smax_value = S64_MAX;
5042 /* If we might shift our top bit out, then we know nothing */
5043 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
5044 dst_reg->umin_value = 0;
5045 dst_reg->umax_value = U64_MAX;
5047 dst_reg->umin_value <<= umin_val;
5048 dst_reg->umax_value <<= umax_val;
5050 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
5051 /* We may learn something more from the var_off */
5052 __update_reg_bounds(dst_reg);
5055 if (umax_val >= insn_bitness) {
5056 /* Shifts greater than 31 or 63 are undefined.
5057 * This includes shifts by a negative number.
5059 mark_reg_unknown(env, regs, insn->dst_reg);
5062 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
5063 * be negative, then either:
5064 * 1) src_reg might be zero, so the sign bit of the result is
5065 * unknown, so we lose our signed bounds
5066 * 2) it's known negative, thus the unsigned bounds capture the
5068 * 3) the signed bounds cross zero, so they tell us nothing
5070 * If the value in dst_reg is known nonnegative, then again the
5071 * unsigned bounts capture the signed bounds.
5072 * Thus, in all cases it suffices to blow away our signed bounds
5073 * and rely on inferring new ones from the unsigned bounds and
5074 * var_off of the result.
5076 dst_reg->smin_value = S64_MIN;
5077 dst_reg->smax_value = S64_MAX;
5078 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
5079 dst_reg->umin_value >>= umax_val;
5080 dst_reg->umax_value >>= umin_val;
5081 /* We may learn something more from the var_off */
5082 __update_reg_bounds(dst_reg);
5085 if (umax_val >= insn_bitness) {
5086 /* Shifts greater than 31 or 63 are undefined.
5087 * This includes shifts by a negative number.
5089 mark_reg_unknown(env, regs, insn->dst_reg);
5093 /* Upon reaching here, src_known is true and
5094 * umax_val is equal to umin_val.
5096 if (insn_bitness == 32) {
5097 dst_reg->smin_value = (u32)(((s32)dst_reg->smin_value) >> umin_val);
5098 dst_reg->smax_value = (u32)(((s32)dst_reg->smax_value) >> umin_val);
5100 dst_reg->smin_value >>= umin_val;
5101 dst_reg->smax_value >>= umin_val;
5104 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val,
5107 /* blow away the dst_reg umin_value/umax_value and rely on
5108 * dst_reg var_off to refine the result.
5110 dst_reg->umin_value = 0;
5111 dst_reg->umax_value = U64_MAX;
5112 __update_reg_bounds(dst_reg);
5115 mark_reg_unknown(env, regs, insn->dst_reg);
5119 if (BPF_CLASS(insn->code) != BPF_ALU64) {
5120 /* 32-bit ALU ops are (32,32)->32 */
5121 coerce_reg_to_size(dst_reg, 4);
5124 __reg_deduce_bounds(dst_reg);
5125 __reg_bound_offset(dst_reg);
5129 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
5132 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
5133 struct bpf_insn *insn)
5135 struct bpf_verifier_state *vstate = env->cur_state;
5136 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5137 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
5138 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
5139 u8 opcode = BPF_OP(insn->code);
5142 dst_reg = ®s[insn->dst_reg];
5144 if (dst_reg->type != SCALAR_VALUE)
5146 if (BPF_SRC(insn->code) == BPF_X) {
5147 src_reg = ®s[insn->src_reg];
5148 if (src_reg->type != SCALAR_VALUE) {
5149 if (dst_reg->type != SCALAR_VALUE) {
5150 /* Combining two pointers by any ALU op yields
5151 * an arbitrary scalar. Disallow all math except
5152 * pointer subtraction
5154 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
5155 mark_reg_unknown(env, regs, insn->dst_reg);
5158 verbose(env, "R%d pointer %s pointer prohibited\n",
5160 bpf_alu_string[opcode >> 4]);
5163 /* scalar += pointer
5164 * This is legal, but we have to reverse our
5165 * src/dest handling in computing the range
5167 err = mark_chain_precision(env, insn->dst_reg);
5170 return adjust_ptr_min_max_vals(env, insn,
5173 } else if (ptr_reg) {
5174 /* pointer += scalar */
5175 err = mark_chain_precision(env, insn->src_reg);
5178 return adjust_ptr_min_max_vals(env, insn,
5182 /* Pretend the src is a reg with a known value, since we only
5183 * need to be able to read from this state.
5185 off_reg.type = SCALAR_VALUE;
5186 __mark_reg_known(&off_reg, insn->imm);
5188 if (ptr_reg) /* pointer += K */
5189 return adjust_ptr_min_max_vals(env, insn,
5193 /* Got here implies adding two SCALAR_VALUEs */
5194 if (WARN_ON_ONCE(ptr_reg)) {
5195 print_verifier_state(env, state);
5196 verbose(env, "verifier internal error: unexpected ptr_reg\n");
5199 if (WARN_ON(!src_reg)) {
5200 print_verifier_state(env, state);
5201 verbose(env, "verifier internal error: no src_reg\n");
5204 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
5207 /* check validity of 32-bit and 64-bit arithmetic operations */
5208 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
5210 struct bpf_reg_state *regs = cur_regs(env);
5211 u8 opcode = BPF_OP(insn->code);
5214 if (opcode == BPF_END || opcode == BPF_NEG) {
5215 if (opcode == BPF_NEG) {
5216 if (BPF_SRC(insn->code) != 0 ||
5217 insn->src_reg != BPF_REG_0 ||
5218 insn->off != 0 || insn->imm != 0) {
5219 verbose(env, "BPF_NEG uses reserved fields\n");
5223 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
5224 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
5225 BPF_CLASS(insn->code) == BPF_ALU64) {
5226 verbose(env, "BPF_END uses reserved fields\n");
5231 /* check src operand */
5232 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
5236 if (is_pointer_value(env, insn->dst_reg)) {
5237 verbose(env, "R%d pointer arithmetic prohibited\n",
5242 /* check dest operand */
5243 err = check_reg_arg(env, insn->dst_reg, DST_OP);
5247 } else if (opcode == BPF_MOV) {
5249 if (BPF_SRC(insn->code) == BPF_X) {
5250 if (insn->imm != 0 || insn->off != 0) {
5251 verbose(env, "BPF_MOV uses reserved fields\n");
5255 /* check src operand */
5256 err = check_reg_arg(env, insn->src_reg, SRC_OP);
5260 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
5261 verbose(env, "BPF_MOV uses reserved fields\n");
5266 /* check dest operand, mark as required later */
5267 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
5271 if (BPF_SRC(insn->code) == BPF_X) {
5272 struct bpf_reg_state *src_reg = regs + insn->src_reg;
5273 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
5275 if (BPF_CLASS(insn->code) == BPF_ALU64) {
5277 * copy register state to dest reg
5279 *dst_reg = *src_reg;
5280 dst_reg->live |= REG_LIVE_WRITTEN;
5281 dst_reg->subreg_def = DEF_NOT_SUBREG;
5284 if (is_pointer_value(env, insn->src_reg)) {
5286 "R%d partial copy of pointer\n",
5289 } else if (src_reg->type == SCALAR_VALUE) {
5290 *dst_reg = *src_reg;
5291 dst_reg->live |= REG_LIVE_WRITTEN;
5292 dst_reg->subreg_def = env->insn_idx + 1;
5294 mark_reg_unknown(env, regs,
5297 coerce_reg_to_size(dst_reg, 4);
5301 * remember the value we stored into this reg
5303 /* clear any state __mark_reg_known doesn't set */
5304 mark_reg_unknown(env, regs, insn->dst_reg);
5305 regs[insn->dst_reg].type = SCALAR_VALUE;
5306 if (BPF_CLASS(insn->code) == BPF_ALU64) {
5307 __mark_reg_known(regs + insn->dst_reg,
5310 __mark_reg_known(regs + insn->dst_reg,
5315 } else if (opcode > BPF_END) {
5316 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
5319 } else { /* all other ALU ops: and, sub, xor, add, ... */
5321 if (BPF_SRC(insn->code) == BPF_X) {
5322 if (insn->imm != 0 || insn->off != 0) {
5323 verbose(env, "BPF_ALU uses reserved fields\n");
5326 /* check src1 operand */
5327 err = check_reg_arg(env, insn->src_reg, SRC_OP);
5331 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
5332 verbose(env, "BPF_ALU uses reserved fields\n");
5337 /* check src2 operand */
5338 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
5342 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
5343 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
5344 verbose(env, "div by zero\n");
5348 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
5349 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
5350 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
5352 if (insn->imm < 0 || insn->imm >= size) {
5353 verbose(env, "invalid shift %d\n", insn->imm);
5358 /* check dest operand */
5359 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
5363 return adjust_reg_min_max_vals(env, insn);
5369 static void __find_good_pkt_pointers(struct bpf_func_state *state,
5370 struct bpf_reg_state *dst_reg,
5371 enum bpf_reg_type type, u16 new_range)
5373 struct bpf_reg_state *reg;
5376 for (i = 0; i < MAX_BPF_REG; i++) {
5377 reg = &state->regs[i];
5378 if (reg->type == type && reg->id == dst_reg->id)
5379 /* keep the maximum range already checked */
5380 reg->range = max(reg->range, new_range);
5383 bpf_for_each_spilled_reg(i, state, reg) {
5386 if (reg->type == type && reg->id == dst_reg->id)
5387 reg->range = max(reg->range, new_range);
5391 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
5392 struct bpf_reg_state *dst_reg,
5393 enum bpf_reg_type type,
5394 bool range_right_open)
5399 if (dst_reg->off < 0 ||
5400 (dst_reg->off == 0 && range_right_open))
5401 /* This doesn't give us any range */
5404 if (dst_reg->umax_value > MAX_PACKET_OFF ||
5405 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
5406 /* Risk of overflow. For instance, ptr + (1<<63) may be less
5407 * than pkt_end, but that's because it's also less than pkt.
5411 new_range = dst_reg->off;
5412 if (range_right_open)
5415 /* Examples for register markings:
5417 * pkt_data in dst register:
5421 * if (r2 > pkt_end) goto <handle exception>
5426 * if (r2 < pkt_end) goto <access okay>
5427 * <handle exception>
5430 * r2 == dst_reg, pkt_end == src_reg
5431 * r2=pkt(id=n,off=8,r=0)
5432 * r3=pkt(id=n,off=0,r=0)
5434 * pkt_data in src register:
5438 * if (pkt_end >= r2) goto <access okay>
5439 * <handle exception>
5443 * if (pkt_end <= r2) goto <handle exception>
5447 * pkt_end == dst_reg, r2 == src_reg
5448 * r2=pkt(id=n,off=8,r=0)
5449 * r3=pkt(id=n,off=0,r=0)
5451 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
5452 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
5453 * and [r3, r3 + 8-1) respectively is safe to access depending on
5457 /* If our ids match, then we must have the same max_value. And we
5458 * don't care about the other reg's fixed offset, since if it's too big
5459 * the range won't allow anything.
5460 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
5462 for (i = 0; i <= vstate->curframe; i++)
5463 __find_good_pkt_pointers(vstate->frame[i], dst_reg, type,
5467 /* compute branch direction of the expression "if (reg opcode val) goto target;"
5469 * 1 - branch will be taken and "goto target" will be executed
5470 * 0 - branch will not be taken and fall-through to next insn
5471 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value range [0,10]
5473 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
5476 struct bpf_reg_state reg_lo;
5479 if (__is_pointer_value(false, reg))
5485 /* For JMP32, only low 32 bits are compared, coerce_reg_to_size
5486 * could truncate high bits and update umin/umax according to
5487 * information of low bits.
5489 coerce_reg_to_size(reg, 4);
5490 /* smin/smax need special handling. For example, after coerce,
5491 * if smin_value is 0x00000000ffffffffLL, the value is -1 when
5492 * used as operand to JMP32. It is a negative number from s32's
5493 * point of view, while it is a positive number when seen as
5494 * s64. The smin/smax are kept as s64, therefore, when used with
5495 * JMP32, they need to be transformed into s32, then sign
5496 * extended back to s64.
5498 * Also, smin/smax were copied from umin/umax. If umin/umax has
5499 * different sign bit, then min/max relationship doesn't
5500 * maintain after casting into s32, for this case, set smin/smax
5503 if ((reg->umax_value ^ reg->umin_value) &
5505 reg->smin_value = S32_MIN;
5506 reg->smax_value = S32_MAX;
5508 reg->smin_value = (s64)(s32)reg->smin_value;
5509 reg->smax_value = (s64)(s32)reg->smax_value;
5512 sval = (s64)(s32)val;
5519 if (tnum_is_const(reg->var_off))
5520 return !!tnum_equals_const(reg->var_off, val);
5523 if (tnum_is_const(reg->var_off))
5524 return !tnum_equals_const(reg->var_off, val);
5527 if ((~reg->var_off.mask & reg->var_off.value) & val)
5529 if (!((reg->var_off.mask | reg->var_off.value) & val))
5533 if (reg->umin_value > val)
5535 else if (reg->umax_value <= val)
5539 if (reg->smin_value > sval)
5541 else if (reg->smax_value < sval)
5545 if (reg->umax_value < val)
5547 else if (reg->umin_value >= val)
5551 if (reg->smax_value < sval)
5553 else if (reg->smin_value >= sval)
5557 if (reg->umin_value >= val)
5559 else if (reg->umax_value < val)
5563 if (reg->smin_value >= sval)
5565 else if (reg->smax_value < sval)
5569 if (reg->umax_value <= val)
5571 else if (reg->umin_value > val)
5575 if (reg->smax_value <= sval)
5577 else if (reg->smin_value > sval)
5585 /* Generate min value of the high 32-bit from TNUM info. */
5586 static u64 gen_hi_min(struct tnum var)
5588 return var.value & ~0xffffffffULL;
5591 /* Generate max value of the high 32-bit from TNUM info. */
5592 static u64 gen_hi_max(struct tnum var)
5594 return (var.value | var.mask) & ~0xffffffffULL;
5597 /* Return true if VAL is compared with a s64 sign extended from s32, and they
5598 * are with the same signedness.
5600 static bool cmp_val_with_extended_s64(s64 sval, struct bpf_reg_state *reg)
5602 return ((s32)sval >= 0 &&
5603 reg->smin_value >= 0 && reg->smax_value <= S32_MAX) ||
5605 reg->smax_value <= 0 && reg->smin_value >= S32_MIN);
5608 /* Adjusts the register min/max values in the case that the dst_reg is the
5609 * variable register that we are working on, and src_reg is a constant or we're
5610 * simply doing a BPF_K check.
5611 * In JEQ/JNE cases we also adjust the var_off values.
5613 static void reg_set_min_max(struct bpf_reg_state *true_reg,
5614 struct bpf_reg_state *false_reg, u64 val,
5615 u8 opcode, bool is_jmp32)
5619 /* If the dst_reg is a pointer, we can't learn anything about its
5620 * variable offset from the compare (unless src_reg were a pointer into
5621 * the same object, but we don't bother with that.
5622 * Since false_reg and true_reg have the same type by construction, we
5623 * only need to check one of them for pointerness.
5625 if (__is_pointer_value(false, false_reg))
5628 val = is_jmp32 ? (u32)val : val;
5629 sval = is_jmp32 ? (s64)(s32)val : (s64)val;
5635 struct bpf_reg_state *reg =
5636 opcode == BPF_JEQ ? true_reg : false_reg;
5638 /* For BPF_JEQ, if this is false we know nothing Jon Snow, but
5639 * if it is true we know the value for sure. Likewise for
5643 u64 old_v = reg->var_off.value;
5644 u64 hi_mask = ~0xffffffffULL;
5646 reg->var_off.value = (old_v & hi_mask) | val;
5647 reg->var_off.mask &= hi_mask;
5649 __mark_reg_known(reg, val);
5654 false_reg->var_off = tnum_and(false_reg->var_off,
5656 if (is_power_of_2(val))
5657 true_reg->var_off = tnum_or(true_reg->var_off,
5663 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
5664 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
5667 false_umax += gen_hi_max(false_reg->var_off);
5668 true_umin += gen_hi_min(true_reg->var_off);
5670 false_reg->umax_value = min(false_reg->umax_value, false_umax);
5671 true_reg->umin_value = max(true_reg->umin_value, true_umin);
5677 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
5678 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
5680 /* If the full s64 was not sign-extended from s32 then don't
5681 * deduct further info.
5683 if (is_jmp32 && !cmp_val_with_extended_s64(sval, false_reg))
5685 false_reg->smax_value = min(false_reg->smax_value, false_smax);
5686 true_reg->smin_value = max(true_reg->smin_value, true_smin);
5692 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
5693 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
5696 false_umin += gen_hi_min(false_reg->var_off);
5697 true_umax += gen_hi_max(true_reg->var_off);
5699 false_reg->umin_value = max(false_reg->umin_value, false_umin);
5700 true_reg->umax_value = min(true_reg->umax_value, true_umax);
5706 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
5707 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
5709 if (is_jmp32 && !cmp_val_with_extended_s64(sval, false_reg))
5711 false_reg->smin_value = max(false_reg->smin_value, false_smin);
5712 true_reg->smax_value = min(true_reg->smax_value, true_smax);
5719 __reg_deduce_bounds(false_reg);
5720 __reg_deduce_bounds(true_reg);
5721 /* We might have learned some bits from the bounds. */
5722 __reg_bound_offset(false_reg);
5723 __reg_bound_offset(true_reg);
5725 __reg_bound_offset32(false_reg);
5726 __reg_bound_offset32(true_reg);
5728 /* Intersecting with the old var_off might have improved our bounds
5729 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
5730 * then new var_off is (0; 0x7f...fc) which improves our umax.
5732 __update_reg_bounds(false_reg);
5733 __update_reg_bounds(true_reg);
5736 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
5739 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
5740 struct bpf_reg_state *false_reg, u64 val,
5741 u8 opcode, bool is_jmp32)
5745 if (__is_pointer_value(false, false_reg))
5748 val = is_jmp32 ? (u32)val : val;
5749 sval = is_jmp32 ? (s64)(s32)val : (s64)val;
5755 struct bpf_reg_state *reg =
5756 opcode == BPF_JEQ ? true_reg : false_reg;
5759 u64 old_v = reg->var_off.value;
5760 u64 hi_mask = ~0xffffffffULL;
5762 reg->var_off.value = (old_v & hi_mask) | val;
5763 reg->var_off.mask &= hi_mask;
5765 __mark_reg_known(reg, val);
5770 false_reg->var_off = tnum_and(false_reg->var_off,
5772 if (is_power_of_2(val))
5773 true_reg->var_off = tnum_or(true_reg->var_off,
5779 u64 false_umin = opcode == BPF_JGT ? val : val + 1;
5780 u64 true_umax = opcode == BPF_JGT ? val - 1 : val;
5783 false_umin += gen_hi_min(false_reg->var_off);
5784 true_umax += gen_hi_max(true_reg->var_off);
5786 false_reg->umin_value = max(false_reg->umin_value, false_umin);
5787 true_reg->umax_value = min(true_reg->umax_value, true_umax);
5793 s64 false_smin = opcode == BPF_JSGT ? sval : sval + 1;
5794 s64 true_smax = opcode == BPF_JSGT ? sval - 1 : sval;
5796 if (is_jmp32 && !cmp_val_with_extended_s64(sval, false_reg))
5798 false_reg->smin_value = max(false_reg->smin_value, false_smin);
5799 true_reg->smax_value = min(true_reg->smax_value, true_smax);
5805 u64 false_umax = opcode == BPF_JLT ? val : val - 1;
5806 u64 true_umin = opcode == BPF_JLT ? val + 1 : val;
5809 false_umax += gen_hi_max(false_reg->var_off);
5810 true_umin += gen_hi_min(true_reg->var_off);
5812 false_reg->umax_value = min(false_reg->umax_value, false_umax);
5813 true_reg->umin_value = max(true_reg->umin_value, true_umin);
5819 s64 false_smax = opcode == BPF_JSLT ? sval : sval - 1;
5820 s64 true_smin = opcode == BPF_JSLT ? sval + 1 : sval;
5822 if (is_jmp32 && !cmp_val_with_extended_s64(sval, false_reg))
5824 false_reg->smax_value = min(false_reg->smax_value, false_smax);
5825 true_reg->smin_value = max(true_reg->smin_value, true_smin);
5832 __reg_deduce_bounds(false_reg);
5833 __reg_deduce_bounds(true_reg);
5834 /* We might have learned some bits from the bounds. */
5835 __reg_bound_offset(false_reg);
5836 __reg_bound_offset(true_reg);
5838 __reg_bound_offset32(false_reg);
5839 __reg_bound_offset32(true_reg);
5841 /* Intersecting with the old var_off might have improved our bounds
5842 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
5843 * then new var_off is (0; 0x7f...fc) which improves our umax.
5845 __update_reg_bounds(false_reg);
5846 __update_reg_bounds(true_reg);
5849 /* Regs are known to be equal, so intersect their min/max/var_off */
5850 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
5851 struct bpf_reg_state *dst_reg)
5853 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
5854 dst_reg->umin_value);
5855 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
5856 dst_reg->umax_value);
5857 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
5858 dst_reg->smin_value);
5859 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
5860 dst_reg->smax_value);
5861 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
5863 /* We might have learned new bounds from the var_off. */
5864 __update_reg_bounds(src_reg);
5865 __update_reg_bounds(dst_reg);
5866 /* We might have learned something about the sign bit. */
5867 __reg_deduce_bounds(src_reg);
5868 __reg_deduce_bounds(dst_reg);
5869 /* We might have learned some bits from the bounds. */
5870 __reg_bound_offset(src_reg);
5871 __reg_bound_offset(dst_reg);
5872 /* Intersecting with the old var_off might have improved our bounds
5873 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
5874 * then new var_off is (0; 0x7f...fc) which improves our umax.
5876 __update_reg_bounds(src_reg);
5877 __update_reg_bounds(dst_reg);
5880 static void reg_combine_min_max(struct bpf_reg_state *true_src,
5881 struct bpf_reg_state *true_dst,
5882 struct bpf_reg_state *false_src,
5883 struct bpf_reg_state *false_dst,
5888 __reg_combine_min_max(true_src, true_dst);
5891 __reg_combine_min_max(false_src, false_dst);
5896 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
5897 struct bpf_reg_state *reg, u32 id,
5900 if (reg_type_may_be_null(reg->type) && reg->id == id) {
5901 /* Old offset (both fixed and variable parts) should
5902 * have been known-zero, because we don't allow pointer
5903 * arithmetic on pointers that might be NULL.
5905 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
5906 !tnum_equals_const(reg->var_off, 0) ||
5908 __mark_reg_known_zero(reg);
5912 reg->type = SCALAR_VALUE;
5913 } else if (reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
5914 if (reg->map_ptr->inner_map_meta) {
5915 reg->type = CONST_PTR_TO_MAP;
5916 reg->map_ptr = reg->map_ptr->inner_map_meta;
5917 } else if (reg->map_ptr->map_type ==
5918 BPF_MAP_TYPE_XSKMAP) {
5919 reg->type = PTR_TO_XDP_SOCK;
5921 reg->type = PTR_TO_MAP_VALUE;
5923 } else if (reg->type == PTR_TO_SOCKET_OR_NULL) {
5924 reg->type = PTR_TO_SOCKET;
5925 } else if (reg->type == PTR_TO_SOCK_COMMON_OR_NULL) {
5926 reg->type = PTR_TO_SOCK_COMMON;
5927 } else if (reg->type == PTR_TO_TCP_SOCK_OR_NULL) {
5928 reg->type = PTR_TO_TCP_SOCK;
5931 /* We don't need id and ref_obj_id from this point
5932 * onwards anymore, thus we should better reset it,
5933 * so that state pruning has chances to take effect.
5936 reg->ref_obj_id = 0;
5937 } else if (!reg_may_point_to_spin_lock(reg)) {
5938 /* For not-NULL ptr, reg->ref_obj_id will be reset
5939 * in release_reg_references().
5941 * reg->id is still used by spin_lock ptr. Other
5942 * than spin_lock ptr type, reg->id can be reset.
5949 static void __mark_ptr_or_null_regs(struct bpf_func_state *state, u32 id,
5952 struct bpf_reg_state *reg;
5955 for (i = 0; i < MAX_BPF_REG; i++)
5956 mark_ptr_or_null_reg(state, &state->regs[i], id, is_null);
5958 bpf_for_each_spilled_reg(i, state, reg) {
5961 mark_ptr_or_null_reg(state, reg, id, is_null);
5965 /* The logic is similar to find_good_pkt_pointers(), both could eventually
5966 * be folded together at some point.
5968 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
5971 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5972 struct bpf_reg_state *regs = state->regs;
5973 u32 ref_obj_id = regs[regno].ref_obj_id;
5974 u32 id = regs[regno].id;
5977 if (ref_obj_id && ref_obj_id == id && is_null)
5978 /* regs[regno] is in the " == NULL" branch.
5979 * No one could have freed the reference state before
5980 * doing the NULL check.
5982 WARN_ON_ONCE(release_reference_state(state, id));
5984 for (i = 0; i <= vstate->curframe; i++)
5985 __mark_ptr_or_null_regs(vstate->frame[i], id, is_null);
5988 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
5989 struct bpf_reg_state *dst_reg,
5990 struct bpf_reg_state *src_reg,
5991 struct bpf_verifier_state *this_branch,
5992 struct bpf_verifier_state *other_branch)
5994 if (BPF_SRC(insn->code) != BPF_X)
5997 /* Pointers are always 64-bit. */
5998 if (BPF_CLASS(insn->code) == BPF_JMP32)
6001 switch (BPF_OP(insn->code)) {
6003 if ((dst_reg->type == PTR_TO_PACKET &&
6004 src_reg->type == PTR_TO_PACKET_END) ||
6005 (dst_reg->type == PTR_TO_PACKET_META &&
6006 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
6007 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
6008 find_good_pkt_pointers(this_branch, dst_reg,
6009 dst_reg->type, false);
6010 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
6011 src_reg->type == PTR_TO_PACKET) ||
6012 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
6013 src_reg->type == PTR_TO_PACKET_META)) {
6014 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
6015 find_good_pkt_pointers(other_branch, src_reg,
6016 src_reg->type, true);
6022 if ((dst_reg->type == PTR_TO_PACKET &&
6023 src_reg->type == PTR_TO_PACKET_END) ||
6024 (dst_reg->type == PTR_TO_PACKET_META &&
6025 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
6026 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
6027 find_good_pkt_pointers(other_branch, dst_reg,
6028 dst_reg->type, true);
6029 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
6030 src_reg->type == PTR_TO_PACKET) ||
6031 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
6032 src_reg->type == PTR_TO_PACKET_META)) {
6033 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
6034 find_good_pkt_pointers(this_branch, src_reg,
6035 src_reg->type, false);
6041 if ((dst_reg->type == PTR_TO_PACKET &&
6042 src_reg->type == PTR_TO_PACKET_END) ||
6043 (dst_reg->type == PTR_TO_PACKET_META &&
6044 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
6045 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
6046 find_good_pkt_pointers(this_branch, dst_reg,
6047 dst_reg->type, true);
6048 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
6049 src_reg->type == PTR_TO_PACKET) ||
6050 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
6051 src_reg->type == PTR_TO_PACKET_META)) {
6052 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
6053 find_good_pkt_pointers(other_branch, src_reg,
6054 src_reg->type, false);
6060 if ((dst_reg->type == PTR_TO_PACKET &&
6061 src_reg->type == PTR_TO_PACKET_END) ||
6062 (dst_reg->type == PTR_TO_PACKET_META &&
6063 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
6064 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
6065 find_good_pkt_pointers(other_branch, dst_reg,
6066 dst_reg->type, false);
6067 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
6068 src_reg->type == PTR_TO_PACKET) ||
6069 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
6070 src_reg->type == PTR_TO_PACKET_META)) {
6071 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
6072 find_good_pkt_pointers(this_branch, src_reg,
6073 src_reg->type, true);
6085 static int check_cond_jmp_op(struct bpf_verifier_env *env,
6086 struct bpf_insn *insn, int *insn_idx)
6088 struct bpf_verifier_state *this_branch = env->cur_state;
6089 struct bpf_verifier_state *other_branch;
6090 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
6091 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
6092 u8 opcode = BPF_OP(insn->code);
6097 /* Only conditional jumps are expected to reach here. */
6098 if (opcode == BPF_JA || opcode > BPF_JSLE) {
6099 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
6103 if (BPF_SRC(insn->code) == BPF_X) {
6104 if (insn->imm != 0) {
6105 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
6109 /* check src1 operand */
6110 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6114 if (is_pointer_value(env, insn->src_reg)) {
6115 verbose(env, "R%d pointer comparison prohibited\n",
6119 src_reg = ®s[insn->src_reg];
6121 if (insn->src_reg != BPF_REG_0) {
6122 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
6127 /* check src2 operand */
6128 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
6132 dst_reg = ®s[insn->dst_reg];
6133 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
6135 if (BPF_SRC(insn->code) == BPF_K)
6136 pred = is_branch_taken(dst_reg, insn->imm,
6138 else if (src_reg->type == SCALAR_VALUE &&
6139 tnum_is_const(src_reg->var_off))
6140 pred = is_branch_taken(dst_reg, src_reg->var_off.value,
6143 err = mark_chain_precision(env, insn->dst_reg);
6144 if (BPF_SRC(insn->code) == BPF_X && !err)
6145 err = mark_chain_precision(env, insn->src_reg);
6150 /* only follow the goto, ignore fall-through */
6151 *insn_idx += insn->off;
6153 } else if (pred == 0) {
6154 /* only follow fall-through branch, since
6155 * that's where the program will go
6160 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
6164 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
6166 /* detect if we are comparing against a constant value so we can adjust
6167 * our min/max values for our dst register.
6168 * this is only legit if both are scalars (or pointers to the same
6169 * object, I suppose, but we don't support that right now), because
6170 * otherwise the different base pointers mean the offsets aren't
6173 if (BPF_SRC(insn->code) == BPF_X) {
6174 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
6175 struct bpf_reg_state lo_reg0 = *dst_reg;
6176 struct bpf_reg_state lo_reg1 = *src_reg;
6177 struct bpf_reg_state *src_lo, *dst_lo;
6181 coerce_reg_to_size(dst_lo, 4);
6182 coerce_reg_to_size(src_lo, 4);
6184 if (dst_reg->type == SCALAR_VALUE &&
6185 src_reg->type == SCALAR_VALUE) {
6186 if (tnum_is_const(src_reg->var_off) ||
6187 (is_jmp32 && tnum_is_const(src_lo->var_off)))
6188 reg_set_min_max(&other_branch_regs[insn->dst_reg],
6191 ? src_lo->var_off.value
6192 : src_reg->var_off.value,
6194 else if (tnum_is_const(dst_reg->var_off) ||
6195 (is_jmp32 && tnum_is_const(dst_lo->var_off)))
6196 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
6199 ? dst_lo->var_off.value
6200 : dst_reg->var_off.value,
6202 else if (!is_jmp32 &&
6203 (opcode == BPF_JEQ || opcode == BPF_JNE))
6204 /* Comparing for equality, we can combine knowledge */
6205 reg_combine_min_max(&other_branch_regs[insn->src_reg],
6206 &other_branch_regs[insn->dst_reg],
6207 src_reg, dst_reg, opcode);
6209 } else if (dst_reg->type == SCALAR_VALUE) {
6210 reg_set_min_max(&other_branch_regs[insn->dst_reg],
6211 dst_reg, insn->imm, opcode, is_jmp32);
6214 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
6215 * NOTE: these optimizations below are related with pointer comparison
6216 * which will never be JMP32.
6218 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
6219 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
6220 reg_type_may_be_null(dst_reg->type)) {
6221 /* Mark all identical registers in each branch as either
6222 * safe or unknown depending R == 0 or R != 0 conditional.
6224 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
6226 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
6228 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
6229 this_branch, other_branch) &&
6230 is_pointer_value(env, insn->dst_reg)) {
6231 verbose(env, "R%d pointer comparison prohibited\n",
6235 if (env->log.level & BPF_LOG_LEVEL)
6236 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
6240 /* verify BPF_LD_IMM64 instruction */
6241 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
6243 struct bpf_insn_aux_data *aux = cur_aux(env);
6244 struct bpf_reg_state *regs = cur_regs(env);
6245 struct bpf_map *map;
6248 if (BPF_SIZE(insn->code) != BPF_DW) {
6249 verbose(env, "invalid BPF_LD_IMM insn\n");
6252 if (insn->off != 0) {
6253 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
6257 err = check_reg_arg(env, insn->dst_reg, DST_OP);
6261 if (insn->src_reg == 0) {
6262 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
6264 regs[insn->dst_reg].type = SCALAR_VALUE;
6265 __mark_reg_known(®s[insn->dst_reg], imm);
6269 map = env->used_maps[aux->map_index];
6270 mark_reg_known_zero(env, regs, insn->dst_reg);
6271 regs[insn->dst_reg].map_ptr = map;
6273 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE) {
6274 regs[insn->dst_reg].type = PTR_TO_MAP_VALUE;
6275 regs[insn->dst_reg].off = aux->map_off;
6276 if (map_value_has_spin_lock(map))
6277 regs[insn->dst_reg].id = ++env->id_gen;
6278 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD) {
6279 regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
6281 verbose(env, "bpf verifier is misconfigured\n");
6288 static bool may_access_skb(enum bpf_prog_type type)
6291 case BPF_PROG_TYPE_SOCKET_FILTER:
6292 case BPF_PROG_TYPE_SCHED_CLS:
6293 case BPF_PROG_TYPE_SCHED_ACT:
6300 /* verify safety of LD_ABS|LD_IND instructions:
6301 * - they can only appear in the programs where ctx == skb
6302 * - since they are wrappers of function calls, they scratch R1-R5 registers,
6303 * preserve R6-R9, and store return value into R0
6306 * ctx == skb == R6 == CTX
6309 * SRC == any register
6310 * IMM == 32-bit immediate
6313 * R0 - 8/16/32-bit skb data converted to cpu endianness
6315 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
6317 struct bpf_reg_state *regs = cur_regs(env);
6318 static const int ctx_reg = BPF_REG_6;
6319 u8 mode = BPF_MODE(insn->code);
6322 if (!may_access_skb(env->prog->type)) {
6323 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
6327 if (!env->ops->gen_ld_abs) {
6328 verbose(env, "bpf verifier is misconfigured\n");
6332 if (env->subprog_cnt > 1) {
6333 /* when program has LD_ABS insn JITs and interpreter assume
6334 * that r1 == ctx == skb which is not the case for callees
6335 * that can have arbitrary arguments. It's problematic
6336 * for main prog as well since JITs would need to analyze
6337 * all functions in order to make proper register save/restore
6338 * decisions in the main prog. Hence disallow LD_ABS with calls
6340 verbose(env, "BPF_LD_[ABS|IND] instructions cannot be mixed with bpf-to-bpf calls\n");
6344 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
6345 BPF_SIZE(insn->code) == BPF_DW ||
6346 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
6347 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
6351 /* check whether implicit source operand (register R6) is readable */
6352 err = check_reg_arg(env, ctx_reg, SRC_OP);
6356 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
6357 * gen_ld_abs() may terminate the program at runtime, leading to
6360 err = check_reference_leak(env);
6362 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
6366 if (env->cur_state->active_spin_lock) {
6367 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
6371 if (regs[ctx_reg].type != PTR_TO_CTX) {
6373 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
6377 if (mode == BPF_IND) {
6378 /* check explicit source operand */
6379 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6384 err = check_ctx_reg(env, ®s[ctx_reg], ctx_reg);
6388 /* reset caller saved regs to unreadable */
6389 for (i = 0; i < CALLER_SAVED_REGS; i++) {
6390 mark_reg_not_init(env, regs, caller_saved[i]);
6391 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
6394 /* mark destination R0 register as readable, since it contains
6395 * the value fetched from the packet.
6396 * Already marked as written above.
6398 mark_reg_unknown(env, regs, BPF_REG_0);
6399 /* ld_abs load up to 32-bit skb data. */
6400 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
6404 static int check_return_code(struct bpf_verifier_env *env)
6406 struct tnum enforce_attach_type_range = tnum_unknown;
6407 const struct bpf_prog *prog = env->prog;
6408 struct bpf_reg_state *reg;
6409 struct tnum range = tnum_range(0, 1);
6412 /* The struct_ops func-ptr's return type could be "void" */
6413 if (env->prog->type == BPF_PROG_TYPE_STRUCT_OPS &&
6414 !prog->aux->attach_func_proto->type)
6417 /* eBPF calling convetion is such that R0 is used
6418 * to return the value from eBPF program.
6419 * Make sure that it's readable at this time
6420 * of bpf_exit, which means that program wrote
6421 * something into it earlier
6423 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
6427 if (is_pointer_value(env, BPF_REG_0)) {
6428 verbose(env, "R0 leaks addr as return value\n");
6432 switch (env->prog->type) {
6433 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
6434 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
6435 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG)
6436 range = tnum_range(1, 1);
6438 case BPF_PROG_TYPE_CGROUP_SKB:
6439 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
6440 range = tnum_range(0, 3);
6441 enforce_attach_type_range = tnum_range(2, 3);
6444 case BPF_PROG_TYPE_CGROUP_SOCK:
6445 case BPF_PROG_TYPE_SOCK_OPS:
6446 case BPF_PROG_TYPE_CGROUP_DEVICE:
6447 case BPF_PROG_TYPE_CGROUP_SYSCTL:
6448 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
6450 case BPF_PROG_TYPE_RAW_TRACEPOINT:
6451 if (!env->prog->aux->attach_btf_id)
6453 range = tnum_const(0);
6459 reg = cur_regs(env) + BPF_REG_0;
6460 if (reg->type != SCALAR_VALUE) {
6461 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
6462 reg_type_str[reg->type]);
6466 if (!tnum_in(range, reg->var_off)) {
6469 verbose(env, "At program exit the register R0 ");
6470 if (!tnum_is_unknown(reg->var_off)) {
6471 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6472 verbose(env, "has value %s", tn_buf);
6474 verbose(env, "has unknown scalar value");
6476 tnum_strn(tn_buf, sizeof(tn_buf), range);
6477 verbose(env, " should have been in %s\n", tn_buf);
6481 if (!tnum_is_unknown(enforce_attach_type_range) &&
6482 tnum_in(enforce_attach_type_range, reg->var_off))
6483 env->prog->enforce_expected_attach_type = 1;
6487 /* non-recursive DFS pseudo code
6488 * 1 procedure DFS-iterative(G,v):
6489 * 2 label v as discovered
6490 * 3 let S be a stack
6492 * 5 while S is not empty
6494 * 7 if t is what we're looking for:
6496 * 9 for all edges e in G.adjacentEdges(t) do
6497 * 10 if edge e is already labelled
6498 * 11 continue with the next edge
6499 * 12 w <- G.adjacentVertex(t,e)
6500 * 13 if vertex w is not discovered and not explored
6501 * 14 label e as tree-edge
6502 * 15 label w as discovered
6505 * 18 else if vertex w is discovered
6506 * 19 label e as back-edge
6508 * 21 // vertex w is explored
6509 * 22 label e as forward- or cross-edge
6510 * 23 label t as explored
6515 * 0x11 - discovered and fall-through edge labelled
6516 * 0x12 - discovered and fall-through and branch edges labelled
6527 static u32 state_htab_size(struct bpf_verifier_env *env)
6529 return env->prog->len;
6532 static struct bpf_verifier_state_list **explored_state(
6533 struct bpf_verifier_env *env,
6536 struct bpf_verifier_state *cur = env->cur_state;
6537 struct bpf_func_state *state = cur->frame[cur->curframe];
6539 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
6542 static void init_explored_state(struct bpf_verifier_env *env, int idx)
6544 env->insn_aux_data[idx].prune_point = true;
6547 /* t, w, e - match pseudo-code above:
6548 * t - index of current instruction
6549 * w - next instruction
6552 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env,
6555 int *insn_stack = env->cfg.insn_stack;
6556 int *insn_state = env->cfg.insn_state;
6558 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
6561 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
6564 if (w < 0 || w >= env->prog->len) {
6565 verbose_linfo(env, t, "%d: ", t);
6566 verbose(env, "jump out of range from insn %d to %d\n", t, w);
6571 /* mark branch target for state pruning */
6572 init_explored_state(env, w);
6574 if (insn_state[w] == 0) {
6576 insn_state[t] = DISCOVERED | e;
6577 insn_state[w] = DISCOVERED;
6578 if (env->cfg.cur_stack >= env->prog->len)
6580 insn_stack[env->cfg.cur_stack++] = w;
6582 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
6583 if (loop_ok && env->allow_ptr_leaks)
6585 verbose_linfo(env, t, "%d: ", t);
6586 verbose_linfo(env, w, "%d: ", w);
6587 verbose(env, "back-edge from insn %d to %d\n", t, w);
6589 } else if (insn_state[w] == EXPLORED) {
6590 /* forward- or cross-edge */
6591 insn_state[t] = DISCOVERED | e;
6593 verbose(env, "insn state internal bug\n");
6599 /* non-recursive depth-first-search to detect loops in BPF program
6600 * loop == back-edge in directed graph
6602 static int check_cfg(struct bpf_verifier_env *env)
6604 struct bpf_insn *insns = env->prog->insnsi;
6605 int insn_cnt = env->prog->len;
6606 int *insn_stack, *insn_state;
6610 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
6614 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
6620 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
6621 insn_stack[0] = 0; /* 0 is the first instruction */
6622 env->cfg.cur_stack = 1;
6625 if (env->cfg.cur_stack == 0)
6627 t = insn_stack[env->cfg.cur_stack - 1];
6629 if (BPF_CLASS(insns[t].code) == BPF_JMP ||
6630 BPF_CLASS(insns[t].code) == BPF_JMP32) {
6631 u8 opcode = BPF_OP(insns[t].code);
6633 if (opcode == BPF_EXIT) {
6635 } else if (opcode == BPF_CALL) {
6636 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
6641 if (t + 1 < insn_cnt)
6642 init_explored_state(env, t + 1);
6643 if (insns[t].src_reg == BPF_PSEUDO_CALL) {
6644 init_explored_state(env, t);
6645 ret = push_insn(t, t + insns[t].imm + 1, BRANCH,
6652 } else if (opcode == BPF_JA) {
6653 if (BPF_SRC(insns[t].code) != BPF_K) {
6657 /* unconditional jump with single edge */
6658 ret = push_insn(t, t + insns[t].off + 1,
6659 FALLTHROUGH, env, true);
6664 /* unconditional jmp is not a good pruning point,
6665 * but it's marked, since backtracking needs
6666 * to record jmp history in is_state_visited().
6668 init_explored_state(env, t + insns[t].off + 1);
6669 /* tell verifier to check for equivalent states
6670 * after every call and jump
6672 if (t + 1 < insn_cnt)
6673 init_explored_state(env, t + 1);
6675 /* conditional jump with two edges */
6676 init_explored_state(env, t);
6677 ret = push_insn(t, t + 1, FALLTHROUGH, env, true);
6683 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env, true);
6690 /* all other non-branch instructions with single
6693 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
6701 insn_state[t] = EXPLORED;
6702 if (env->cfg.cur_stack-- <= 0) {
6703 verbose(env, "pop stack internal bug\n");
6710 for (i = 0; i < insn_cnt; i++) {
6711 if (insn_state[i] != EXPLORED) {
6712 verbose(env, "unreachable insn %d\n", i);
6717 ret = 0; /* cfg looks good */
6722 env->cfg.insn_state = env->cfg.insn_stack = NULL;
6726 /* The minimum supported BTF func info size */
6727 #define MIN_BPF_FUNCINFO_SIZE 8
6728 #define MAX_FUNCINFO_REC_SIZE 252
6730 static int check_btf_func(struct bpf_verifier_env *env,
6731 const union bpf_attr *attr,
6732 union bpf_attr __user *uattr)
6734 u32 i, nfuncs, urec_size, min_size;
6735 u32 krec_size = sizeof(struct bpf_func_info);
6736 struct bpf_func_info *krecord;
6737 struct bpf_func_info_aux *info_aux = NULL;
6738 const struct btf_type *type;
6739 struct bpf_prog *prog;
6740 const struct btf *btf;
6741 void __user *urecord;
6742 u32 prev_offset = 0;
6745 nfuncs = attr->func_info_cnt;
6749 if (nfuncs != env->subprog_cnt) {
6750 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
6754 urec_size = attr->func_info_rec_size;
6755 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
6756 urec_size > MAX_FUNCINFO_REC_SIZE ||
6757 urec_size % sizeof(u32)) {
6758 verbose(env, "invalid func info rec size %u\n", urec_size);
6763 btf = prog->aux->btf;
6765 urecord = u64_to_user_ptr(attr->func_info);
6766 min_size = min_t(u32, krec_size, urec_size);
6768 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
6771 info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN);
6775 for (i = 0; i < nfuncs; i++) {
6776 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
6778 if (ret == -E2BIG) {
6779 verbose(env, "nonzero tailing record in func info");
6780 /* set the size kernel expects so loader can zero
6781 * out the rest of the record.
6783 if (put_user(min_size, &uattr->func_info_rec_size))
6789 if (copy_from_user(&krecord[i], urecord, min_size)) {
6794 /* check insn_off */
6796 if (krecord[i].insn_off) {
6798 "nonzero insn_off %u for the first func info record",
6799 krecord[i].insn_off);
6803 } else if (krecord[i].insn_off <= prev_offset) {
6805 "same or smaller insn offset (%u) than previous func info record (%u)",
6806 krecord[i].insn_off, prev_offset);
6811 if (env->subprog_info[i].start != krecord[i].insn_off) {
6812 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
6818 type = btf_type_by_id(btf, krecord[i].type_id);
6819 if (!type || !btf_type_is_func(type)) {
6820 verbose(env, "invalid type id %d in func info",
6821 krecord[i].type_id);
6825 info_aux[i].linkage = BTF_INFO_VLEN(type->info);
6826 prev_offset = krecord[i].insn_off;
6827 urecord += urec_size;
6830 prog->aux->func_info = krecord;
6831 prog->aux->func_info_cnt = nfuncs;
6832 prog->aux->func_info_aux = info_aux;
6841 static void adjust_btf_func(struct bpf_verifier_env *env)
6843 struct bpf_prog_aux *aux = env->prog->aux;
6846 if (!aux->func_info)
6849 for (i = 0; i < env->subprog_cnt; i++)
6850 aux->func_info[i].insn_off = env->subprog_info[i].start;
6853 #define MIN_BPF_LINEINFO_SIZE (offsetof(struct bpf_line_info, line_col) + \
6854 sizeof(((struct bpf_line_info *)(0))->line_col))
6855 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
6857 static int check_btf_line(struct bpf_verifier_env *env,
6858 const union bpf_attr *attr,
6859 union bpf_attr __user *uattr)
6861 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
6862 struct bpf_subprog_info *sub;
6863 struct bpf_line_info *linfo;
6864 struct bpf_prog *prog;
6865 const struct btf *btf;
6866 void __user *ulinfo;
6869 nr_linfo = attr->line_info_cnt;
6873 rec_size = attr->line_info_rec_size;
6874 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
6875 rec_size > MAX_LINEINFO_REC_SIZE ||
6876 rec_size & (sizeof(u32) - 1))
6879 /* Need to zero it in case the userspace may
6880 * pass in a smaller bpf_line_info object.
6882 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
6883 GFP_KERNEL | __GFP_NOWARN);
6888 btf = prog->aux->btf;
6891 sub = env->subprog_info;
6892 ulinfo = u64_to_user_ptr(attr->line_info);
6893 expected_size = sizeof(struct bpf_line_info);
6894 ncopy = min_t(u32, expected_size, rec_size);
6895 for (i = 0; i < nr_linfo; i++) {
6896 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
6898 if (err == -E2BIG) {
6899 verbose(env, "nonzero tailing record in line_info");
6900 if (put_user(expected_size,
6901 &uattr->line_info_rec_size))
6907 if (copy_from_user(&linfo[i], ulinfo, ncopy)) {
6913 * Check insn_off to ensure
6914 * 1) strictly increasing AND
6915 * 2) bounded by prog->len
6917 * The linfo[0].insn_off == 0 check logically falls into
6918 * the later "missing bpf_line_info for func..." case
6919 * because the first linfo[0].insn_off must be the
6920 * first sub also and the first sub must have
6921 * subprog_info[0].start == 0.
6923 if ((i && linfo[i].insn_off <= prev_offset) ||
6924 linfo[i].insn_off >= prog->len) {
6925 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
6926 i, linfo[i].insn_off, prev_offset,
6932 if (!prog->insnsi[linfo[i].insn_off].code) {
6934 "Invalid insn code at line_info[%u].insn_off\n",
6940 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
6941 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
6942 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
6947 if (s != env->subprog_cnt) {
6948 if (linfo[i].insn_off == sub[s].start) {
6949 sub[s].linfo_idx = i;
6951 } else if (sub[s].start < linfo[i].insn_off) {
6952 verbose(env, "missing bpf_line_info for func#%u\n", s);
6958 prev_offset = linfo[i].insn_off;
6962 if (s != env->subprog_cnt) {
6963 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
6964 env->subprog_cnt - s, s);
6969 prog->aux->linfo = linfo;
6970 prog->aux->nr_linfo = nr_linfo;
6979 static int check_btf_info(struct bpf_verifier_env *env,
6980 const union bpf_attr *attr,
6981 union bpf_attr __user *uattr)
6986 if (!attr->func_info_cnt && !attr->line_info_cnt)
6989 btf = btf_get_by_fd(attr->prog_btf_fd);
6991 return PTR_ERR(btf);
6992 env->prog->aux->btf = btf;
6994 err = check_btf_func(env, attr, uattr);
6998 err = check_btf_line(env, attr, uattr);
7005 /* check %cur's range satisfies %old's */
7006 static bool range_within(struct bpf_reg_state *old,
7007 struct bpf_reg_state *cur)
7009 return old->umin_value <= cur->umin_value &&
7010 old->umax_value >= cur->umax_value &&
7011 old->smin_value <= cur->smin_value &&
7012 old->smax_value >= cur->smax_value;
7015 /* Maximum number of register states that can exist at once */
7016 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
7022 /* If in the old state two registers had the same id, then they need to have
7023 * the same id in the new state as well. But that id could be different from
7024 * the old state, so we need to track the mapping from old to new ids.
7025 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
7026 * regs with old id 5 must also have new id 9 for the new state to be safe. But
7027 * regs with a different old id could still have new id 9, we don't care about
7029 * So we look through our idmap to see if this old id has been seen before. If
7030 * so, we require the new id to match; otherwise, we add the id pair to the map.
7032 static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
7036 for (i = 0; i < ID_MAP_SIZE; i++) {
7037 if (!idmap[i].old) {
7038 /* Reached an empty slot; haven't seen this id before */
7039 idmap[i].old = old_id;
7040 idmap[i].cur = cur_id;
7043 if (idmap[i].old == old_id)
7044 return idmap[i].cur == cur_id;
7046 /* We ran out of idmap slots, which should be impossible */
7051 static void clean_func_state(struct bpf_verifier_env *env,
7052 struct bpf_func_state *st)
7054 enum bpf_reg_liveness live;
7057 for (i = 0; i < BPF_REG_FP; i++) {
7058 live = st->regs[i].live;
7059 /* liveness must not touch this register anymore */
7060 st->regs[i].live |= REG_LIVE_DONE;
7061 if (!(live & REG_LIVE_READ))
7062 /* since the register is unused, clear its state
7063 * to make further comparison simpler
7065 __mark_reg_not_init(env, &st->regs[i]);
7068 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
7069 live = st->stack[i].spilled_ptr.live;
7070 /* liveness must not touch this stack slot anymore */
7071 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
7072 if (!(live & REG_LIVE_READ)) {
7073 __mark_reg_not_init(env, &st->stack[i].spilled_ptr);
7074 for (j = 0; j < BPF_REG_SIZE; j++)
7075 st->stack[i].slot_type[j] = STACK_INVALID;
7080 static void clean_verifier_state(struct bpf_verifier_env *env,
7081 struct bpf_verifier_state *st)
7085 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
7086 /* all regs in this state in all frames were already marked */
7089 for (i = 0; i <= st->curframe; i++)
7090 clean_func_state(env, st->frame[i]);
7093 /* the parentage chains form a tree.
7094 * the verifier states are added to state lists at given insn and
7095 * pushed into state stack for future exploration.
7096 * when the verifier reaches bpf_exit insn some of the verifer states
7097 * stored in the state lists have their final liveness state already,
7098 * but a lot of states will get revised from liveness point of view when
7099 * the verifier explores other branches.
7102 * 2: if r1 == 100 goto pc+1
7105 * when the verifier reaches exit insn the register r0 in the state list of
7106 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
7107 * of insn 2 and goes exploring further. At the insn 4 it will walk the
7108 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
7110 * Since the verifier pushes the branch states as it sees them while exploring
7111 * the program the condition of walking the branch instruction for the second
7112 * time means that all states below this branch were already explored and
7113 * their final liveness markes are already propagated.
7114 * Hence when the verifier completes the search of state list in is_state_visited()
7115 * we can call this clean_live_states() function to mark all liveness states
7116 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
7118 * This function also clears the registers and stack for states that !READ
7119 * to simplify state merging.
7121 * Important note here that walking the same branch instruction in the callee
7122 * doesn't meant that the states are DONE. The verifier has to compare
7125 static void clean_live_states(struct bpf_verifier_env *env, int insn,
7126 struct bpf_verifier_state *cur)
7128 struct bpf_verifier_state_list *sl;
7131 sl = *explored_state(env, insn);
7133 if (sl->state.branches)
7135 if (sl->state.insn_idx != insn ||
7136 sl->state.curframe != cur->curframe)
7138 for (i = 0; i <= cur->curframe; i++)
7139 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
7141 clean_verifier_state(env, &sl->state);
7147 /* Returns true if (rold safe implies rcur safe) */
7148 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
7149 struct idpair *idmap)
7153 if (!(rold->live & REG_LIVE_READ))
7154 /* explored state didn't use this */
7157 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
7159 if (rold->type == PTR_TO_STACK)
7160 /* two stack pointers are equal only if they're pointing to
7161 * the same stack frame, since fp-8 in foo != fp-8 in bar
7163 return equal && rold->frameno == rcur->frameno;
7168 if (rold->type == NOT_INIT)
7169 /* explored state can't have used this */
7171 if (rcur->type == NOT_INIT)
7173 switch (rold->type) {
7175 if (rcur->type == SCALAR_VALUE) {
7176 if (!rold->precise && !rcur->precise)
7178 /* new val must satisfy old val knowledge */
7179 return range_within(rold, rcur) &&
7180 tnum_in(rold->var_off, rcur->var_off);
7182 /* We're trying to use a pointer in place of a scalar.
7183 * Even if the scalar was unbounded, this could lead to
7184 * pointer leaks because scalars are allowed to leak
7185 * while pointers are not. We could make this safe in
7186 * special cases if root is calling us, but it's
7187 * probably not worth the hassle.
7191 case PTR_TO_MAP_VALUE:
7192 /* If the new min/max/var_off satisfy the old ones and
7193 * everything else matches, we are OK.
7194 * 'id' is not compared, since it's only used for maps with
7195 * bpf_spin_lock inside map element and in such cases if
7196 * the rest of the prog is valid for one map element then
7197 * it's valid for all map elements regardless of the key
7198 * used in bpf_map_lookup()
7200 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
7201 range_within(rold, rcur) &&
7202 tnum_in(rold->var_off, rcur->var_off);
7203 case PTR_TO_MAP_VALUE_OR_NULL:
7204 /* a PTR_TO_MAP_VALUE could be safe to use as a
7205 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
7206 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
7207 * checked, doing so could have affected others with the same
7208 * id, and we can't check for that because we lost the id when
7209 * we converted to a PTR_TO_MAP_VALUE.
7211 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
7213 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
7215 /* Check our ids match any regs they're supposed to */
7216 return check_ids(rold->id, rcur->id, idmap);
7217 case PTR_TO_PACKET_META:
7219 if (rcur->type != rold->type)
7221 /* We must have at least as much range as the old ptr
7222 * did, so that any accesses which were safe before are
7223 * still safe. This is true even if old range < old off,
7224 * since someone could have accessed through (ptr - k), or
7225 * even done ptr -= k in a register, to get a safe access.
7227 if (rold->range > rcur->range)
7229 /* If the offsets don't match, we can't trust our alignment;
7230 * nor can we be sure that we won't fall out of range.
7232 if (rold->off != rcur->off)
7234 /* id relations must be preserved */
7235 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
7237 /* new val must satisfy old val knowledge */
7238 return range_within(rold, rcur) &&
7239 tnum_in(rold->var_off, rcur->var_off);
7241 case CONST_PTR_TO_MAP:
7242 case PTR_TO_PACKET_END:
7243 case PTR_TO_FLOW_KEYS:
7245 case PTR_TO_SOCKET_OR_NULL:
7246 case PTR_TO_SOCK_COMMON:
7247 case PTR_TO_SOCK_COMMON_OR_NULL:
7248 case PTR_TO_TCP_SOCK:
7249 case PTR_TO_TCP_SOCK_OR_NULL:
7250 case PTR_TO_XDP_SOCK:
7251 /* Only valid matches are exact, which memcmp() above
7252 * would have accepted
7255 /* Don't know what's going on, just say it's not safe */
7259 /* Shouldn't get here; if we do, say it's not safe */
7264 static bool stacksafe(struct bpf_func_state *old,
7265 struct bpf_func_state *cur,
7266 struct idpair *idmap)
7270 /* walk slots of the explored stack and ignore any additional
7271 * slots in the current stack, since explored(safe) state
7274 for (i = 0; i < old->allocated_stack; i++) {
7275 spi = i / BPF_REG_SIZE;
7277 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
7278 i += BPF_REG_SIZE - 1;
7279 /* explored state didn't use this */
7283 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
7286 /* explored stack has more populated slots than current stack
7287 * and these slots were used
7289 if (i >= cur->allocated_stack)
7292 /* if old state was safe with misc data in the stack
7293 * it will be safe with zero-initialized stack.
7294 * The opposite is not true
7296 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
7297 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
7299 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
7300 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
7301 /* Ex: old explored (safe) state has STACK_SPILL in
7302 * this stack slot, but current has has STACK_MISC ->
7303 * this verifier states are not equivalent,
7304 * return false to continue verification of this path
7307 if (i % BPF_REG_SIZE)
7309 if (old->stack[spi].slot_type[0] != STACK_SPILL)
7311 if (!regsafe(&old->stack[spi].spilled_ptr,
7312 &cur->stack[spi].spilled_ptr,
7314 /* when explored and current stack slot are both storing
7315 * spilled registers, check that stored pointers types
7316 * are the same as well.
7317 * Ex: explored safe path could have stored
7318 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
7319 * but current path has stored:
7320 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
7321 * such verifier states are not equivalent.
7322 * return false to continue verification of this path
7329 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
7331 if (old->acquired_refs != cur->acquired_refs)
7333 return !memcmp(old->refs, cur->refs,
7334 sizeof(*old->refs) * old->acquired_refs);
7337 /* compare two verifier states
7339 * all states stored in state_list are known to be valid, since
7340 * verifier reached 'bpf_exit' instruction through them
7342 * this function is called when verifier exploring different branches of
7343 * execution popped from the state stack. If it sees an old state that has
7344 * more strict register state and more strict stack state then this execution
7345 * branch doesn't need to be explored further, since verifier already
7346 * concluded that more strict state leads to valid finish.
7348 * Therefore two states are equivalent if register state is more conservative
7349 * and explored stack state is more conservative than the current one.
7352 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
7353 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
7355 * In other words if current stack state (one being explored) has more
7356 * valid slots than old one that already passed validation, it means
7357 * the verifier can stop exploring and conclude that current state is valid too
7359 * Similarly with registers. If explored state has register type as invalid
7360 * whereas register type in current state is meaningful, it means that
7361 * the current state will reach 'bpf_exit' instruction safely
7363 static bool func_states_equal(struct bpf_func_state *old,
7364 struct bpf_func_state *cur)
7366 struct idpair *idmap;
7370 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
7371 /* If we failed to allocate the idmap, just say it's not safe */
7375 for (i = 0; i < MAX_BPF_REG; i++) {
7376 if (!regsafe(&old->regs[i], &cur->regs[i], idmap))
7380 if (!stacksafe(old, cur, idmap))
7383 if (!refsafe(old, cur))
7391 static bool states_equal(struct bpf_verifier_env *env,
7392 struct bpf_verifier_state *old,
7393 struct bpf_verifier_state *cur)
7397 if (old->curframe != cur->curframe)
7400 /* Verification state from speculative execution simulation
7401 * must never prune a non-speculative execution one.
7403 if (old->speculative && !cur->speculative)
7406 if (old->active_spin_lock != cur->active_spin_lock)
7409 /* for states to be equal callsites have to be the same
7410 * and all frame states need to be equivalent
7412 for (i = 0; i <= old->curframe; i++) {
7413 if (old->frame[i]->callsite != cur->frame[i]->callsite)
7415 if (!func_states_equal(old->frame[i], cur->frame[i]))
7421 /* Return 0 if no propagation happened. Return negative error code if error
7422 * happened. Otherwise, return the propagated bit.
7424 static int propagate_liveness_reg(struct bpf_verifier_env *env,
7425 struct bpf_reg_state *reg,
7426 struct bpf_reg_state *parent_reg)
7428 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
7429 u8 flag = reg->live & REG_LIVE_READ;
7432 /* When comes here, read flags of PARENT_REG or REG could be any of
7433 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
7434 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
7436 if (parent_flag == REG_LIVE_READ64 ||
7437 /* Or if there is no read flag from REG. */
7439 /* Or if the read flag from REG is the same as PARENT_REG. */
7440 parent_flag == flag)
7443 err = mark_reg_read(env, reg, parent_reg, flag);
7450 /* A write screens off any subsequent reads; but write marks come from the
7451 * straight-line code between a state and its parent. When we arrive at an
7452 * equivalent state (jump target or such) we didn't arrive by the straight-line
7453 * code, so read marks in the state must propagate to the parent regardless
7454 * of the state's write marks. That's what 'parent == state->parent' comparison
7455 * in mark_reg_read() is for.
7457 static int propagate_liveness(struct bpf_verifier_env *env,
7458 const struct bpf_verifier_state *vstate,
7459 struct bpf_verifier_state *vparent)
7461 struct bpf_reg_state *state_reg, *parent_reg;
7462 struct bpf_func_state *state, *parent;
7463 int i, frame, err = 0;
7465 if (vparent->curframe != vstate->curframe) {
7466 WARN(1, "propagate_live: parent frame %d current frame %d\n",
7467 vparent->curframe, vstate->curframe);
7470 /* Propagate read liveness of registers... */
7471 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
7472 for (frame = 0; frame <= vstate->curframe; frame++) {
7473 parent = vparent->frame[frame];
7474 state = vstate->frame[frame];
7475 parent_reg = parent->regs;
7476 state_reg = state->regs;
7477 /* We don't need to worry about FP liveness, it's read-only */
7478 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
7479 err = propagate_liveness_reg(env, &state_reg[i],
7483 if (err == REG_LIVE_READ64)
7484 mark_insn_zext(env, &parent_reg[i]);
7487 /* Propagate stack slots. */
7488 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
7489 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
7490 parent_reg = &parent->stack[i].spilled_ptr;
7491 state_reg = &state->stack[i].spilled_ptr;
7492 err = propagate_liveness_reg(env, state_reg,
7501 /* find precise scalars in the previous equivalent state and
7502 * propagate them into the current state
7504 static int propagate_precision(struct bpf_verifier_env *env,
7505 const struct bpf_verifier_state *old)
7507 struct bpf_reg_state *state_reg;
7508 struct bpf_func_state *state;
7511 state = old->frame[old->curframe];
7512 state_reg = state->regs;
7513 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
7514 if (state_reg->type != SCALAR_VALUE ||
7515 !state_reg->precise)
7517 if (env->log.level & BPF_LOG_LEVEL2)
7518 verbose(env, "propagating r%d\n", i);
7519 err = mark_chain_precision(env, i);
7524 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
7525 if (state->stack[i].slot_type[0] != STACK_SPILL)
7527 state_reg = &state->stack[i].spilled_ptr;
7528 if (state_reg->type != SCALAR_VALUE ||
7529 !state_reg->precise)
7531 if (env->log.level & BPF_LOG_LEVEL2)
7532 verbose(env, "propagating fp%d\n",
7533 (-i - 1) * BPF_REG_SIZE);
7534 err = mark_chain_precision_stack(env, i);
7541 static bool states_maybe_looping(struct bpf_verifier_state *old,
7542 struct bpf_verifier_state *cur)
7544 struct bpf_func_state *fold, *fcur;
7545 int i, fr = cur->curframe;
7547 if (old->curframe != fr)
7550 fold = old->frame[fr];
7551 fcur = cur->frame[fr];
7552 for (i = 0; i < MAX_BPF_REG; i++)
7553 if (memcmp(&fold->regs[i], &fcur->regs[i],
7554 offsetof(struct bpf_reg_state, parent)))
7560 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
7562 struct bpf_verifier_state_list *new_sl;
7563 struct bpf_verifier_state_list *sl, **pprev;
7564 struct bpf_verifier_state *cur = env->cur_state, *new;
7565 int i, j, err, states_cnt = 0;
7566 bool add_new_state = env->test_state_freq ? true : false;
7568 cur->last_insn_idx = env->prev_insn_idx;
7569 if (!env->insn_aux_data[insn_idx].prune_point)
7570 /* this 'insn_idx' instruction wasn't marked, so we will not
7571 * be doing state search here
7575 /* bpf progs typically have pruning point every 4 instructions
7576 * http://vger.kernel.org/bpfconf2019.html#session-1
7577 * Do not add new state for future pruning if the verifier hasn't seen
7578 * at least 2 jumps and at least 8 instructions.
7579 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
7580 * In tests that amounts to up to 50% reduction into total verifier
7581 * memory consumption and 20% verifier time speedup.
7583 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
7584 env->insn_processed - env->prev_insn_processed >= 8)
7585 add_new_state = true;
7587 pprev = explored_state(env, insn_idx);
7590 clean_live_states(env, insn_idx, cur);
7594 if (sl->state.insn_idx != insn_idx)
7596 if (sl->state.branches) {
7597 if (states_maybe_looping(&sl->state, cur) &&
7598 states_equal(env, &sl->state, cur)) {
7599 verbose_linfo(env, insn_idx, "; ");
7600 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
7603 /* if the verifier is processing a loop, avoid adding new state
7604 * too often, since different loop iterations have distinct
7605 * states and may not help future pruning.
7606 * This threshold shouldn't be too low to make sure that
7607 * a loop with large bound will be rejected quickly.
7608 * The most abusive loop will be:
7610 * if r1 < 1000000 goto pc-2
7611 * 1M insn_procssed limit / 100 == 10k peak states.
7612 * This threshold shouldn't be too high either, since states
7613 * at the end of the loop are likely to be useful in pruning.
7615 if (env->jmps_processed - env->prev_jmps_processed < 20 &&
7616 env->insn_processed - env->prev_insn_processed < 100)
7617 add_new_state = false;
7620 if (states_equal(env, &sl->state, cur)) {
7622 /* reached equivalent register/stack state,
7624 * Registers read by the continuation are read by us.
7625 * If we have any write marks in env->cur_state, they
7626 * will prevent corresponding reads in the continuation
7627 * from reaching our parent (an explored_state). Our
7628 * own state will get the read marks recorded, but
7629 * they'll be immediately forgotten as we're pruning
7630 * this state and will pop a new one.
7632 err = propagate_liveness(env, &sl->state, cur);
7634 /* if previous state reached the exit with precision and
7635 * current state is equivalent to it (except precsion marks)
7636 * the precision needs to be propagated back in
7637 * the current state.
7639 err = err ? : push_jmp_history(env, cur);
7640 err = err ? : propagate_precision(env, &sl->state);
7646 /* when new state is not going to be added do not increase miss count.
7647 * Otherwise several loop iterations will remove the state
7648 * recorded earlier. The goal of these heuristics is to have
7649 * states from some iterations of the loop (some in the beginning
7650 * and some at the end) to help pruning.
7654 /* heuristic to determine whether this state is beneficial
7655 * to keep checking from state equivalence point of view.
7656 * Higher numbers increase max_states_per_insn and verification time,
7657 * but do not meaningfully decrease insn_processed.
7659 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
7660 /* the state is unlikely to be useful. Remove it to
7661 * speed up verification
7664 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
7665 u32 br = sl->state.branches;
7668 "BUG live_done but branches_to_explore %d\n",
7670 free_verifier_state(&sl->state, false);
7674 /* cannot free this state, since parentage chain may
7675 * walk it later. Add it for free_list instead to
7676 * be freed at the end of verification
7678 sl->next = env->free_list;
7679 env->free_list = sl;
7689 if (env->max_states_per_insn < states_cnt)
7690 env->max_states_per_insn = states_cnt;
7692 if (!env->allow_ptr_leaks && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
7693 return push_jmp_history(env, cur);
7696 return push_jmp_history(env, cur);
7698 /* There were no equivalent states, remember the current one.
7699 * Technically the current state is not proven to be safe yet,
7700 * but it will either reach outer most bpf_exit (which means it's safe)
7701 * or it will be rejected. When there are no loops the verifier won't be
7702 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
7703 * again on the way to bpf_exit.
7704 * When looping the sl->state.branches will be > 0 and this state
7705 * will not be considered for equivalence until branches == 0.
7707 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
7710 env->total_states++;
7712 env->prev_jmps_processed = env->jmps_processed;
7713 env->prev_insn_processed = env->insn_processed;
7715 /* add new state to the head of linked list */
7716 new = &new_sl->state;
7717 err = copy_verifier_state(new, cur);
7719 free_verifier_state(new, false);
7723 new->insn_idx = insn_idx;
7724 WARN_ONCE(new->branches != 1,
7725 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
7728 cur->first_insn_idx = insn_idx;
7729 clear_jmp_history(cur);
7730 new_sl->next = *explored_state(env, insn_idx);
7731 *explored_state(env, insn_idx) = new_sl;
7732 /* connect new state to parentage chain. Current frame needs all
7733 * registers connected. Only r6 - r9 of the callers are alive (pushed
7734 * to the stack implicitly by JITs) so in callers' frames connect just
7735 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
7736 * the state of the call instruction (with WRITTEN set), and r0 comes
7737 * from callee with its full parentage chain, anyway.
7739 /* clear write marks in current state: the writes we did are not writes
7740 * our child did, so they don't screen off its reads from us.
7741 * (There are no read marks in current state, because reads always mark
7742 * their parent and current state never has children yet. Only
7743 * explored_states can get read marks.)
7745 for (j = 0; j <= cur->curframe; j++) {
7746 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
7747 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
7748 for (i = 0; i < BPF_REG_FP; i++)
7749 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
7752 /* all stack frames are accessible from callee, clear them all */
7753 for (j = 0; j <= cur->curframe; j++) {
7754 struct bpf_func_state *frame = cur->frame[j];
7755 struct bpf_func_state *newframe = new->frame[j];
7757 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
7758 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
7759 frame->stack[i].spilled_ptr.parent =
7760 &newframe->stack[i].spilled_ptr;
7766 /* Return true if it's OK to have the same insn return a different type. */
7767 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
7772 case PTR_TO_SOCKET_OR_NULL:
7773 case PTR_TO_SOCK_COMMON:
7774 case PTR_TO_SOCK_COMMON_OR_NULL:
7775 case PTR_TO_TCP_SOCK:
7776 case PTR_TO_TCP_SOCK_OR_NULL:
7777 case PTR_TO_XDP_SOCK:
7785 /* If an instruction was previously used with particular pointer types, then we
7786 * need to be careful to avoid cases such as the below, where it may be ok
7787 * for one branch accessing the pointer, but not ok for the other branch:
7792 * R1 = some_other_valid_ptr;
7795 * R2 = *(u32 *)(R1 + 0);
7797 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
7799 return src != prev && (!reg_type_mismatch_ok(src) ||
7800 !reg_type_mismatch_ok(prev));
7803 static int do_check(struct bpf_verifier_env *env)
7805 struct bpf_verifier_state *state = env->cur_state;
7806 struct bpf_insn *insns = env->prog->insnsi;
7807 struct bpf_reg_state *regs;
7808 int insn_cnt = env->prog->len;
7809 bool do_print_state = false;
7810 int prev_insn_idx = -1;
7813 struct bpf_insn *insn;
7817 env->prev_insn_idx = prev_insn_idx;
7818 if (env->insn_idx >= insn_cnt) {
7819 verbose(env, "invalid insn idx %d insn_cnt %d\n",
7820 env->insn_idx, insn_cnt);
7824 insn = &insns[env->insn_idx];
7825 class = BPF_CLASS(insn->code);
7827 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
7829 "BPF program is too large. Processed %d insn\n",
7830 env->insn_processed);
7834 err = is_state_visited(env, env->insn_idx);
7838 /* found equivalent state, can prune the search */
7839 if (env->log.level & BPF_LOG_LEVEL) {
7841 verbose(env, "\nfrom %d to %d%s: safe\n",
7842 env->prev_insn_idx, env->insn_idx,
7843 env->cur_state->speculative ?
7844 " (speculative execution)" : "");
7846 verbose(env, "%d: safe\n", env->insn_idx);
7848 goto process_bpf_exit;
7851 if (signal_pending(current))
7857 if (env->log.level & BPF_LOG_LEVEL2 ||
7858 (env->log.level & BPF_LOG_LEVEL && do_print_state)) {
7859 if (env->log.level & BPF_LOG_LEVEL2)
7860 verbose(env, "%d:", env->insn_idx);
7862 verbose(env, "\nfrom %d to %d%s:",
7863 env->prev_insn_idx, env->insn_idx,
7864 env->cur_state->speculative ?
7865 " (speculative execution)" : "");
7866 print_verifier_state(env, state->frame[state->curframe]);
7867 do_print_state = false;
7870 if (env->log.level & BPF_LOG_LEVEL) {
7871 const struct bpf_insn_cbs cbs = {
7872 .cb_print = verbose,
7873 .private_data = env,
7876 verbose_linfo(env, env->insn_idx, "; ");
7877 verbose(env, "%d: ", env->insn_idx);
7878 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
7881 if (bpf_prog_is_dev_bound(env->prog->aux)) {
7882 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
7883 env->prev_insn_idx);
7888 regs = cur_regs(env);
7889 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
7890 prev_insn_idx = env->insn_idx;
7892 if (class == BPF_ALU || class == BPF_ALU64) {
7893 err = check_alu_op(env, insn);
7897 } else if (class == BPF_LDX) {
7898 enum bpf_reg_type *prev_src_type, src_reg_type;
7900 /* check for reserved fields is already done */
7902 /* check src operand */
7903 err = check_reg_arg(env, insn->src_reg, SRC_OP);
7907 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
7911 src_reg_type = regs[insn->src_reg].type;
7913 /* check that memory (src_reg + off) is readable,
7914 * the state of dst_reg will be updated by this func
7916 err = check_mem_access(env, env->insn_idx, insn->src_reg,
7917 insn->off, BPF_SIZE(insn->code),
7918 BPF_READ, insn->dst_reg, false);
7922 prev_src_type = &env->insn_aux_data[env->insn_idx].ptr_type;
7924 if (*prev_src_type == NOT_INIT) {
7926 * dst_reg = *(u32 *)(src_reg + off)
7927 * save type to validate intersecting paths
7929 *prev_src_type = src_reg_type;
7931 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
7932 /* ABuser program is trying to use the same insn
7933 * dst_reg = *(u32*) (src_reg + off)
7934 * with different pointer types:
7935 * src_reg == ctx in one branch and
7936 * src_reg == stack|map in some other branch.
7939 verbose(env, "same insn cannot be used with different pointers\n");
7943 } else if (class == BPF_STX) {
7944 enum bpf_reg_type *prev_dst_type, dst_reg_type;
7946 if (BPF_MODE(insn->code) == BPF_XADD) {
7947 err = check_xadd(env, env->insn_idx, insn);
7954 /* check src1 operand */
7955 err = check_reg_arg(env, insn->src_reg, SRC_OP);
7958 /* check src2 operand */
7959 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
7963 dst_reg_type = regs[insn->dst_reg].type;
7965 /* check that memory (dst_reg + off) is writeable */
7966 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
7967 insn->off, BPF_SIZE(insn->code),
7968 BPF_WRITE, insn->src_reg, false);
7972 prev_dst_type = &env->insn_aux_data[env->insn_idx].ptr_type;
7974 if (*prev_dst_type == NOT_INIT) {
7975 *prev_dst_type = dst_reg_type;
7976 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
7977 verbose(env, "same insn cannot be used with different pointers\n");
7981 } else if (class == BPF_ST) {
7982 if (BPF_MODE(insn->code) != BPF_MEM ||
7983 insn->src_reg != BPF_REG_0) {
7984 verbose(env, "BPF_ST uses reserved fields\n");
7987 /* check src operand */
7988 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
7992 if (is_ctx_reg(env, insn->dst_reg)) {
7993 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
7995 reg_type_str[reg_state(env, insn->dst_reg)->type]);
7999 /* check that memory (dst_reg + off) is writeable */
8000 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
8001 insn->off, BPF_SIZE(insn->code),
8002 BPF_WRITE, -1, false);
8006 } else if (class == BPF_JMP || class == BPF_JMP32) {
8007 u8 opcode = BPF_OP(insn->code);
8009 env->jmps_processed++;
8010 if (opcode == BPF_CALL) {
8011 if (BPF_SRC(insn->code) != BPF_K ||
8013 (insn->src_reg != BPF_REG_0 &&
8014 insn->src_reg != BPF_PSEUDO_CALL) ||
8015 insn->dst_reg != BPF_REG_0 ||
8016 class == BPF_JMP32) {
8017 verbose(env, "BPF_CALL uses reserved fields\n");
8021 if (env->cur_state->active_spin_lock &&
8022 (insn->src_reg == BPF_PSEUDO_CALL ||
8023 insn->imm != BPF_FUNC_spin_unlock)) {
8024 verbose(env, "function calls are not allowed while holding a lock\n");
8027 if (insn->src_reg == BPF_PSEUDO_CALL)
8028 err = check_func_call(env, insn, &env->insn_idx);
8030 err = check_helper_call(env, insn->imm, env->insn_idx);
8034 } else if (opcode == BPF_JA) {
8035 if (BPF_SRC(insn->code) != BPF_K ||
8037 insn->src_reg != BPF_REG_0 ||
8038 insn->dst_reg != BPF_REG_0 ||
8039 class == BPF_JMP32) {
8040 verbose(env, "BPF_JA uses reserved fields\n");
8044 env->insn_idx += insn->off + 1;
8047 } else if (opcode == BPF_EXIT) {
8048 if (BPF_SRC(insn->code) != BPF_K ||
8050 insn->src_reg != BPF_REG_0 ||
8051 insn->dst_reg != BPF_REG_0 ||
8052 class == BPF_JMP32) {
8053 verbose(env, "BPF_EXIT uses reserved fields\n");
8057 if (env->cur_state->active_spin_lock) {
8058 verbose(env, "bpf_spin_unlock is missing\n");
8062 if (state->curframe) {
8063 /* exit from nested function */
8064 err = prepare_func_exit(env, &env->insn_idx);
8067 do_print_state = true;
8071 err = check_reference_leak(env);
8075 err = check_return_code(env);
8079 update_branch_counts(env, env->cur_state);
8080 err = pop_stack(env, &prev_insn_idx,
8087 do_print_state = true;
8091 err = check_cond_jmp_op(env, insn, &env->insn_idx);
8095 } else if (class == BPF_LD) {
8096 u8 mode = BPF_MODE(insn->code);
8098 if (mode == BPF_ABS || mode == BPF_IND) {
8099 err = check_ld_abs(env, insn);
8103 } else if (mode == BPF_IMM) {
8104 err = check_ld_imm(env, insn);
8109 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
8111 verbose(env, "invalid BPF_LD mode\n");
8115 verbose(env, "unknown insn class %d\n", class);
8125 static int check_map_prealloc(struct bpf_map *map)
8127 return (map->map_type != BPF_MAP_TYPE_HASH &&
8128 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
8129 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
8130 !(map->map_flags & BPF_F_NO_PREALLOC);
8133 static bool is_tracing_prog_type(enum bpf_prog_type type)
8136 case BPF_PROG_TYPE_KPROBE:
8137 case BPF_PROG_TYPE_TRACEPOINT:
8138 case BPF_PROG_TYPE_PERF_EVENT:
8139 case BPF_PROG_TYPE_RAW_TRACEPOINT:
8146 static bool is_preallocated_map(struct bpf_map *map)
8148 if (!check_map_prealloc(map))
8150 if (map->inner_map_meta && !check_map_prealloc(map->inner_map_meta))
8155 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
8156 struct bpf_map *map,
8157 struct bpf_prog *prog)
8161 * Validate that trace type programs use preallocated hash maps.
8163 * For programs attached to PERF events this is mandatory as the
8164 * perf NMI can hit any arbitrary code sequence.
8166 * All other trace types using preallocated hash maps are unsafe as
8167 * well because tracepoint or kprobes can be inside locked regions
8168 * of the memory allocator or at a place where a recursion into the
8169 * memory allocator would see inconsistent state.
8171 * On RT enabled kernels run-time allocation of all trace type
8172 * programs is strictly prohibited due to lock type constraints. On
8173 * !RT kernels it is allowed for backwards compatibility reasons for
8174 * now, but warnings are emitted so developers are made aware of
8175 * the unsafety and can fix their programs before this is enforced.
8177 if (is_tracing_prog_type(prog->type) && !is_preallocated_map(map)) {
8178 if (prog->type == BPF_PROG_TYPE_PERF_EVENT) {
8179 verbose(env, "perf_event programs can only use preallocated hash map\n");
8182 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
8183 verbose(env, "trace type programs can only use preallocated hash map\n");
8186 WARN_ONCE(1, "trace type BPF program uses run-time allocation\n");
8187 verbose(env, "trace type programs with run-time allocated hash maps are unsafe. Switch to preallocated hash maps.\n");
8190 if ((is_tracing_prog_type(prog->type) ||
8191 prog->type == BPF_PROG_TYPE_SOCKET_FILTER) &&
8192 map_value_has_spin_lock(map)) {
8193 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
8197 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
8198 !bpf_offload_prog_map_match(prog, map)) {
8199 verbose(env, "offload device mismatch between prog and map\n");
8203 if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
8204 verbose(env, "bpf_struct_ops map cannot be used in prog\n");
8211 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
8213 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
8214 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
8217 /* look for pseudo eBPF instructions that access map FDs and
8218 * replace them with actual map pointers
8220 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env *env)
8222 struct bpf_insn *insn = env->prog->insnsi;
8223 int insn_cnt = env->prog->len;
8226 err = bpf_prog_calc_tag(env->prog);
8230 for (i = 0; i < insn_cnt; i++, insn++) {
8231 if (BPF_CLASS(insn->code) == BPF_LDX &&
8232 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
8233 verbose(env, "BPF_LDX uses reserved fields\n");
8237 if (BPF_CLASS(insn->code) == BPF_STX &&
8238 ((BPF_MODE(insn->code) != BPF_MEM &&
8239 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
8240 verbose(env, "BPF_STX uses reserved fields\n");
8244 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
8245 struct bpf_insn_aux_data *aux;
8246 struct bpf_map *map;
8250 if (i == insn_cnt - 1 || insn[1].code != 0 ||
8251 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
8253 verbose(env, "invalid bpf_ld_imm64 insn\n");
8257 if (insn[0].src_reg == 0)
8258 /* valid generic load 64-bit imm */
8261 /* In final convert_pseudo_ld_imm64() step, this is
8262 * converted into regular 64-bit imm load insn.
8264 if ((insn[0].src_reg != BPF_PSEUDO_MAP_FD &&
8265 insn[0].src_reg != BPF_PSEUDO_MAP_VALUE) ||
8266 (insn[0].src_reg == BPF_PSEUDO_MAP_FD &&
8267 insn[1].imm != 0)) {
8269 "unrecognized bpf_ld_imm64 insn\n");
8273 f = fdget(insn[0].imm);
8274 map = __bpf_map_get(f);
8276 verbose(env, "fd %d is not pointing to valid bpf_map\n",
8278 return PTR_ERR(map);
8281 err = check_map_prog_compatibility(env, map, env->prog);
8287 aux = &env->insn_aux_data[i];
8288 if (insn->src_reg == BPF_PSEUDO_MAP_FD) {
8289 addr = (unsigned long)map;
8291 u32 off = insn[1].imm;
8293 if (off >= BPF_MAX_VAR_OFF) {
8294 verbose(env, "direct value offset of %u is not allowed\n", off);
8299 if (!map->ops->map_direct_value_addr) {
8300 verbose(env, "no direct value access support for this map type\n");
8305 err = map->ops->map_direct_value_addr(map, &addr, off);
8307 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
8308 map->value_size, off);
8317 insn[0].imm = (u32)addr;
8318 insn[1].imm = addr >> 32;
8320 /* check whether we recorded this map already */
8321 for (j = 0; j < env->used_map_cnt; j++) {
8322 if (env->used_maps[j] == map) {
8329 if (env->used_map_cnt >= MAX_USED_MAPS) {
8334 /* hold the map. If the program is rejected by verifier,
8335 * the map will be released by release_maps() or it
8336 * will be used by the valid program until it's unloaded
8337 * and all maps are released in free_used_maps()
8341 aux->map_index = env->used_map_cnt;
8342 env->used_maps[env->used_map_cnt++] = map;
8344 if (bpf_map_is_cgroup_storage(map) &&
8345 bpf_cgroup_storage_assign(env->prog->aux, map)) {
8346 verbose(env, "only one cgroup storage of each type is allowed\n");
8358 /* Basic sanity check before we invest more work here. */
8359 if (!bpf_opcode_in_insntable(insn->code)) {
8360 verbose(env, "unknown opcode %02x\n", insn->code);
8365 /* now all pseudo BPF_LD_IMM64 instructions load valid
8366 * 'struct bpf_map *' into a register instead of user map_fd.
8367 * These pointers will be used later by verifier to validate map access.
8372 /* drop refcnt of maps used by the rejected program */
8373 static void release_maps(struct bpf_verifier_env *env)
8375 __bpf_free_used_maps(env->prog->aux, env->used_maps,
8379 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
8380 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
8382 struct bpf_insn *insn = env->prog->insnsi;
8383 int insn_cnt = env->prog->len;
8386 for (i = 0; i < insn_cnt; i++, insn++)
8387 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
8391 /* single env->prog->insni[off] instruction was replaced with the range
8392 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
8393 * [0, off) and [off, end) to new locations, so the patched range stays zero
8395 static int adjust_insn_aux_data(struct bpf_verifier_env *env,
8396 struct bpf_prog *new_prog, u32 off, u32 cnt)
8398 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
8399 struct bpf_insn *insn = new_prog->insnsi;
8403 /* aux info at OFF always needs adjustment, no matter fast path
8404 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
8405 * original insn at old prog.
8407 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
8411 prog_len = new_prog->len;
8412 new_data = vzalloc(array_size(prog_len,
8413 sizeof(struct bpf_insn_aux_data)));
8416 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
8417 memcpy(new_data + off + cnt - 1, old_data + off,
8418 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
8419 for (i = off; i < off + cnt - 1; i++) {
8420 new_data[i].seen = env->pass_cnt;
8421 new_data[i].zext_dst = insn_has_def32(env, insn + i);
8423 env->insn_aux_data = new_data;
8428 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
8434 /* NOTE: fake 'exit' subprog should be updated as well. */
8435 for (i = 0; i <= env->subprog_cnt; i++) {
8436 if (env->subprog_info[i].start <= off)
8438 env->subprog_info[i].start += len - 1;
8442 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
8443 const struct bpf_insn *patch, u32 len)
8445 struct bpf_prog *new_prog;
8447 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
8448 if (IS_ERR(new_prog)) {
8449 if (PTR_ERR(new_prog) == -ERANGE)
8451 "insn %d cannot be patched due to 16-bit range\n",
8452 env->insn_aux_data[off].orig_idx);
8455 if (adjust_insn_aux_data(env, new_prog, off, len))
8457 adjust_subprog_starts(env, off, len);
8461 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
8466 /* find first prog starting at or after off (first to remove) */
8467 for (i = 0; i < env->subprog_cnt; i++)
8468 if (env->subprog_info[i].start >= off)
8470 /* find first prog starting at or after off + cnt (first to stay) */
8471 for (j = i; j < env->subprog_cnt; j++)
8472 if (env->subprog_info[j].start >= off + cnt)
8474 /* if j doesn't start exactly at off + cnt, we are just removing
8475 * the front of previous prog
8477 if (env->subprog_info[j].start != off + cnt)
8481 struct bpf_prog_aux *aux = env->prog->aux;
8484 /* move fake 'exit' subprog as well */
8485 move = env->subprog_cnt + 1 - j;
8487 memmove(env->subprog_info + i,
8488 env->subprog_info + j,
8489 sizeof(*env->subprog_info) * move);
8490 env->subprog_cnt -= j - i;
8492 /* remove func_info */
8493 if (aux->func_info) {
8494 move = aux->func_info_cnt - j;
8496 memmove(aux->func_info + i,
8498 sizeof(*aux->func_info) * move);
8499 aux->func_info_cnt -= j - i;
8500 /* func_info->insn_off is set after all code rewrites,
8501 * in adjust_btf_func() - no need to adjust
8505 /* convert i from "first prog to remove" to "first to adjust" */
8506 if (env->subprog_info[i].start == off)
8510 /* update fake 'exit' subprog as well */
8511 for (; i <= env->subprog_cnt; i++)
8512 env->subprog_info[i].start -= cnt;
8517 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
8520 struct bpf_prog *prog = env->prog;
8521 u32 i, l_off, l_cnt, nr_linfo;
8522 struct bpf_line_info *linfo;
8524 nr_linfo = prog->aux->nr_linfo;
8528 linfo = prog->aux->linfo;
8530 /* find first line info to remove, count lines to be removed */
8531 for (i = 0; i < nr_linfo; i++)
8532 if (linfo[i].insn_off >= off)
8537 for (; i < nr_linfo; i++)
8538 if (linfo[i].insn_off < off + cnt)
8543 /* First live insn doesn't match first live linfo, it needs to "inherit"
8544 * last removed linfo. prog is already modified, so prog->len == off
8545 * means no live instructions after (tail of the program was removed).
8547 if (prog->len != off && l_cnt &&
8548 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
8550 linfo[--i].insn_off = off + cnt;
8553 /* remove the line info which refer to the removed instructions */
8555 memmove(linfo + l_off, linfo + i,
8556 sizeof(*linfo) * (nr_linfo - i));
8558 prog->aux->nr_linfo -= l_cnt;
8559 nr_linfo = prog->aux->nr_linfo;
8562 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
8563 for (i = l_off; i < nr_linfo; i++)
8564 linfo[i].insn_off -= cnt;
8566 /* fix up all subprogs (incl. 'exit') which start >= off */
8567 for (i = 0; i <= env->subprog_cnt; i++)
8568 if (env->subprog_info[i].linfo_idx > l_off) {
8569 /* program may have started in the removed region but
8570 * may not be fully removed
8572 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
8573 env->subprog_info[i].linfo_idx -= l_cnt;
8575 env->subprog_info[i].linfo_idx = l_off;
8581 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
8583 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
8584 unsigned int orig_prog_len = env->prog->len;
8587 if (bpf_prog_is_dev_bound(env->prog->aux))
8588 bpf_prog_offload_remove_insns(env, off, cnt);
8590 err = bpf_remove_insns(env->prog, off, cnt);
8594 err = adjust_subprog_starts_after_remove(env, off, cnt);
8598 err = bpf_adj_linfo_after_remove(env, off, cnt);
8602 memmove(aux_data + off, aux_data + off + cnt,
8603 sizeof(*aux_data) * (orig_prog_len - off - cnt));
8608 /* The verifier does more data flow analysis than llvm and will not
8609 * explore branches that are dead at run time. Malicious programs can
8610 * have dead code too. Therefore replace all dead at-run-time code
8613 * Just nops are not optimal, e.g. if they would sit at the end of the
8614 * program and through another bug we would manage to jump there, then
8615 * we'd execute beyond program memory otherwise. Returning exception
8616 * code also wouldn't work since we can have subprogs where the dead
8617 * code could be located.
8619 static void sanitize_dead_code(struct bpf_verifier_env *env)
8621 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
8622 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
8623 struct bpf_insn *insn = env->prog->insnsi;
8624 const int insn_cnt = env->prog->len;
8627 for (i = 0; i < insn_cnt; i++) {
8628 if (aux_data[i].seen)
8630 memcpy(insn + i, &trap, sizeof(trap));
8634 static bool insn_is_cond_jump(u8 code)
8638 if (BPF_CLASS(code) == BPF_JMP32)
8641 if (BPF_CLASS(code) != BPF_JMP)
8645 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
8648 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
8650 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
8651 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
8652 struct bpf_insn *insn = env->prog->insnsi;
8653 const int insn_cnt = env->prog->len;
8656 for (i = 0; i < insn_cnt; i++, insn++) {
8657 if (!insn_is_cond_jump(insn->code))
8660 if (!aux_data[i + 1].seen)
8662 else if (!aux_data[i + 1 + insn->off].seen)
8667 if (bpf_prog_is_dev_bound(env->prog->aux))
8668 bpf_prog_offload_replace_insn(env, i, &ja);
8670 memcpy(insn, &ja, sizeof(ja));
8674 static int opt_remove_dead_code(struct bpf_verifier_env *env)
8676 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
8677 int insn_cnt = env->prog->len;
8680 for (i = 0; i < insn_cnt; i++) {
8684 while (i + j < insn_cnt && !aux_data[i + j].seen)
8689 err = verifier_remove_insns(env, i, j);
8692 insn_cnt = env->prog->len;
8698 static int opt_remove_nops(struct bpf_verifier_env *env)
8700 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
8701 struct bpf_insn *insn = env->prog->insnsi;
8702 int insn_cnt = env->prog->len;
8705 for (i = 0; i < insn_cnt; i++) {
8706 if (memcmp(&insn[i], &ja, sizeof(ja)))
8709 err = verifier_remove_insns(env, i, 1);
8719 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
8720 const union bpf_attr *attr)
8722 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
8723 struct bpf_insn_aux_data *aux = env->insn_aux_data;
8724 int i, patch_len, delta = 0, len = env->prog->len;
8725 struct bpf_insn *insns = env->prog->insnsi;
8726 struct bpf_prog *new_prog;
8729 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
8730 zext_patch[1] = BPF_ZEXT_REG(0);
8731 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
8732 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
8733 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
8734 for (i = 0; i < len; i++) {
8735 int adj_idx = i + delta;
8736 struct bpf_insn insn;
8738 insn = insns[adj_idx];
8739 if (!aux[adj_idx].zext_dst) {
8747 class = BPF_CLASS(code);
8748 if (insn_no_def(&insn))
8751 /* NOTE: arg "reg" (the fourth one) is only used for
8752 * BPF_STX which has been ruled out in above
8753 * check, it is safe to pass NULL here.
8755 if (is_reg64(env, &insn, insn.dst_reg, NULL, DST_OP)) {
8756 if (class == BPF_LD &&
8757 BPF_MODE(code) == BPF_IMM)
8762 /* ctx load could be transformed into wider load. */
8763 if (class == BPF_LDX &&
8764 aux[adj_idx].ptr_type == PTR_TO_CTX)
8767 imm_rnd = get_random_int();
8768 rnd_hi32_patch[0] = insn;
8769 rnd_hi32_patch[1].imm = imm_rnd;
8770 rnd_hi32_patch[3].dst_reg = insn.dst_reg;
8771 patch = rnd_hi32_patch;
8773 goto apply_patch_buffer;
8776 if (!bpf_jit_needs_zext())
8779 zext_patch[0] = insn;
8780 zext_patch[1].dst_reg = insn.dst_reg;
8781 zext_patch[1].src_reg = insn.dst_reg;
8785 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
8788 env->prog = new_prog;
8789 insns = new_prog->insnsi;
8790 aux = env->insn_aux_data;
8791 delta += patch_len - 1;
8797 /* convert load instructions that access fields of a context type into a
8798 * sequence of instructions that access fields of the underlying structure:
8799 * struct __sk_buff -> struct sk_buff
8800 * struct bpf_sock_ops -> struct sock
8802 static int convert_ctx_accesses(struct bpf_verifier_env *env)
8804 const struct bpf_verifier_ops *ops = env->ops;
8805 int i, cnt, size, ctx_field_size, delta = 0;
8806 const int insn_cnt = env->prog->len;
8807 struct bpf_insn insn_buf[16], *insn;
8808 u32 target_size, size_default, off;
8809 struct bpf_prog *new_prog;
8810 enum bpf_access_type type;
8811 bool is_narrower_load;
8813 if (ops->gen_prologue || env->seen_direct_write) {
8814 if (!ops->gen_prologue) {
8815 verbose(env, "bpf verifier is misconfigured\n");
8818 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
8820 if (cnt >= ARRAY_SIZE(insn_buf)) {
8821 verbose(env, "bpf verifier is misconfigured\n");
8824 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
8828 env->prog = new_prog;
8833 if (bpf_prog_is_dev_bound(env->prog->aux))
8836 insn = env->prog->insnsi + delta;
8838 for (i = 0; i < insn_cnt; i++, insn++) {
8839 bpf_convert_ctx_access_t convert_ctx_access;
8841 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
8842 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
8843 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
8844 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
8846 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
8847 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
8848 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
8849 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
8854 if (type == BPF_WRITE &&
8855 env->insn_aux_data[i + delta].sanitize_stack_off) {
8856 struct bpf_insn patch[] = {
8857 /* Sanitize suspicious stack slot with zero.
8858 * There are no memory dependencies for this store,
8859 * since it's only using frame pointer and immediate
8862 BPF_ST_MEM(BPF_DW, BPF_REG_FP,
8863 env->insn_aux_data[i + delta].sanitize_stack_off,
8865 /* the original STX instruction will immediately
8866 * overwrite the same stack slot with appropriate value
8871 cnt = ARRAY_SIZE(patch);
8872 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
8877 env->prog = new_prog;
8878 insn = new_prog->insnsi + i + delta;
8882 switch (env->insn_aux_data[i + delta].ptr_type) {
8884 if (!ops->convert_ctx_access)
8886 convert_ctx_access = ops->convert_ctx_access;
8889 case PTR_TO_SOCK_COMMON:
8890 convert_ctx_access = bpf_sock_convert_ctx_access;
8892 case PTR_TO_TCP_SOCK:
8893 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
8895 case PTR_TO_XDP_SOCK:
8896 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
8899 if (type == BPF_READ) {
8900 insn->code = BPF_LDX | BPF_PROBE_MEM |
8901 BPF_SIZE((insn)->code);
8902 env->prog->aux->num_exentries++;
8903 } else if (env->prog->type != BPF_PROG_TYPE_STRUCT_OPS) {
8904 verbose(env, "Writes through BTF pointers are not allowed\n");
8912 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
8913 size = BPF_LDST_BYTES(insn);
8915 /* If the read access is a narrower load of the field,
8916 * convert to a 4/8-byte load, to minimum program type specific
8917 * convert_ctx_access changes. If conversion is successful,
8918 * we will apply proper mask to the result.
8920 is_narrower_load = size < ctx_field_size;
8921 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
8923 if (is_narrower_load) {
8926 if (type == BPF_WRITE) {
8927 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
8932 if (ctx_field_size == 4)
8934 else if (ctx_field_size == 8)
8937 insn->off = off & ~(size_default - 1);
8938 insn->code = BPF_LDX | BPF_MEM | size_code;
8942 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
8944 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
8945 (ctx_field_size && !target_size)) {
8946 verbose(env, "bpf verifier is misconfigured\n");
8950 if (is_narrower_load && size < target_size) {
8951 u8 shift = bpf_ctx_narrow_access_offset(
8952 off, size, size_default) * 8;
8953 if (ctx_field_size <= 4) {
8955 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
8958 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
8959 (1 << size * 8) - 1);
8962 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
8965 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
8966 (1ULL << size * 8) - 1);
8970 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
8976 /* keep walking new program and skip insns we just inserted */
8977 env->prog = new_prog;
8978 insn = new_prog->insnsi + i + delta;
8984 static int jit_subprogs(struct bpf_verifier_env *env)
8986 struct bpf_prog *prog = env->prog, **func, *tmp;
8987 int i, j, subprog_start, subprog_end = 0, len, subprog;
8988 struct bpf_insn *insn;
8992 if (env->subprog_cnt <= 1)
8995 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
8996 if (insn->code != (BPF_JMP | BPF_CALL) ||
8997 insn->src_reg != BPF_PSEUDO_CALL)
8999 /* Upon error here we cannot fall back to interpreter but
9000 * need a hard reject of the program. Thus -EFAULT is
9001 * propagated in any case.
9003 subprog = find_subprog(env, i + insn->imm + 1);
9005 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
9009 /* temporarily remember subprog id inside insn instead of
9010 * aux_data, since next loop will split up all insns into funcs
9012 insn->off = subprog;
9013 /* remember original imm in case JIT fails and fallback
9014 * to interpreter will be needed
9016 env->insn_aux_data[i].call_imm = insn->imm;
9017 /* point imm to __bpf_call_base+1 from JITs point of view */
9021 err = bpf_prog_alloc_jited_linfo(prog);
9026 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
9030 for (i = 0; i < env->subprog_cnt; i++) {
9031 subprog_start = subprog_end;
9032 subprog_end = env->subprog_info[i + 1].start;
9034 len = subprog_end - subprog_start;
9035 /* BPF_PROG_RUN doesn't call subprogs directly,
9036 * hence main prog stats include the runtime of subprogs.
9037 * subprogs don't have IDs and not reachable via prog_get_next_id
9038 * func[i]->aux->stats will never be accessed and stays NULL
9040 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
9043 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
9044 len * sizeof(struct bpf_insn));
9045 func[i]->type = prog->type;
9047 if (bpf_prog_calc_tag(func[i]))
9049 func[i]->is_func = 1;
9050 func[i]->aux->func_idx = i;
9051 /* the btf and func_info will be freed only at prog->aux */
9052 func[i]->aux->btf = prog->aux->btf;
9053 func[i]->aux->func_info = prog->aux->func_info;
9055 /* Use bpf_prog_F_tag to indicate functions in stack traces.
9056 * Long term would need debug info to populate names
9058 func[i]->aux->name[0] = 'F';
9059 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
9060 func[i]->jit_requested = 1;
9061 func[i]->aux->linfo = prog->aux->linfo;
9062 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
9063 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
9064 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
9065 func[i] = bpf_int_jit_compile(func[i]);
9066 if (!func[i]->jited) {
9072 /* at this point all bpf functions were successfully JITed
9073 * now populate all bpf_calls with correct addresses and
9074 * run last pass of JIT
9076 for (i = 0; i < env->subprog_cnt; i++) {
9077 insn = func[i]->insnsi;
9078 for (j = 0; j < func[i]->len; j++, insn++) {
9079 if (insn->code != (BPF_JMP | BPF_CALL) ||
9080 insn->src_reg != BPF_PSEUDO_CALL)
9082 subprog = insn->off;
9083 insn->imm = BPF_CAST_CALL(func[subprog]->bpf_func) -
9087 /* we use the aux data to keep a list of the start addresses
9088 * of the JITed images for each function in the program
9090 * for some architectures, such as powerpc64, the imm field
9091 * might not be large enough to hold the offset of the start
9092 * address of the callee's JITed image from __bpf_call_base
9094 * in such cases, we can lookup the start address of a callee
9095 * by using its subprog id, available from the off field of
9096 * the call instruction, as an index for this list
9098 func[i]->aux->func = func;
9099 func[i]->aux->func_cnt = env->subprog_cnt;
9101 for (i = 0; i < env->subprog_cnt; i++) {
9102 old_bpf_func = func[i]->bpf_func;
9103 tmp = bpf_int_jit_compile(func[i]);
9104 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
9105 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
9112 /* finally lock prog and jit images for all functions and
9115 for (i = 0; i < env->subprog_cnt; i++) {
9116 bpf_prog_lock_ro(func[i]);
9117 bpf_prog_kallsyms_add(func[i]);
9120 /* Last step: make now unused interpreter insns from main
9121 * prog consistent for later dump requests, so they can
9122 * later look the same as if they were interpreted only.
9124 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
9125 if (insn->code != (BPF_JMP | BPF_CALL) ||
9126 insn->src_reg != BPF_PSEUDO_CALL)
9128 insn->off = env->insn_aux_data[i].call_imm;
9129 subprog = find_subprog(env, i + insn->off + 1);
9130 insn->imm = subprog;
9134 prog->bpf_func = func[0]->bpf_func;
9135 prog->aux->func = func;
9136 prog->aux->func_cnt = env->subprog_cnt;
9137 bpf_prog_free_unused_jited_linfo(prog);
9140 for (i = 0; i < env->subprog_cnt; i++)
9142 bpf_jit_free(func[i]);
9145 /* cleanup main prog to be interpreted */
9146 prog->jit_requested = 0;
9147 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
9148 if (insn->code != (BPF_JMP | BPF_CALL) ||
9149 insn->src_reg != BPF_PSEUDO_CALL)
9152 insn->imm = env->insn_aux_data[i].call_imm;
9154 bpf_prog_free_jited_linfo(prog);
9158 static int fixup_call_args(struct bpf_verifier_env *env)
9160 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
9161 struct bpf_prog *prog = env->prog;
9162 struct bpf_insn *insn = prog->insnsi;
9167 if (env->prog->jit_requested &&
9168 !bpf_prog_is_dev_bound(env->prog->aux)) {
9169 err = jit_subprogs(env);
9175 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
9176 for (i = 0; i < prog->len; i++, insn++) {
9177 if (insn->code != (BPF_JMP | BPF_CALL) ||
9178 insn->src_reg != BPF_PSEUDO_CALL)
9180 depth = get_callee_stack_depth(env, insn, i);
9183 bpf_patch_call_args(insn, depth);
9190 /* fixup insn->imm field of bpf_call instructions
9191 * and inline eligible helpers as explicit sequence of BPF instructions
9193 * this function is called after eBPF program passed verification
9195 static int fixup_bpf_calls(struct bpf_verifier_env *env)
9197 struct bpf_prog *prog = env->prog;
9198 bool expect_blinding = bpf_jit_blinding_enabled(prog);
9199 struct bpf_insn *insn = prog->insnsi;
9200 const struct bpf_func_proto *fn;
9201 const int insn_cnt = prog->len;
9202 const struct bpf_map_ops *ops;
9203 struct bpf_insn_aux_data *aux;
9204 struct bpf_insn insn_buf[16];
9205 struct bpf_prog *new_prog;
9206 struct bpf_map *map_ptr;
9207 int i, ret, cnt, delta = 0;
9209 for (i = 0; i < insn_cnt; i++, insn++) {
9210 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
9211 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
9212 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
9213 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
9214 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
9215 struct bpf_insn mask_and_div[] = {
9216 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
9218 BPF_JMP_IMM(BPF_JNE, insn->src_reg, 0, 2),
9219 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
9220 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
9223 struct bpf_insn mask_and_mod[] = {
9224 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
9225 /* Rx mod 0 -> Rx */
9226 BPF_JMP_IMM(BPF_JEQ, insn->src_reg, 0, 1),
9229 struct bpf_insn *patchlet;
9231 if (insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
9232 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
9233 patchlet = mask_and_div + (is64 ? 1 : 0);
9234 cnt = ARRAY_SIZE(mask_and_div) - (is64 ? 1 : 0);
9236 patchlet = mask_and_mod + (is64 ? 1 : 0);
9237 cnt = ARRAY_SIZE(mask_and_mod) - (is64 ? 1 : 0);
9240 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
9245 env->prog = prog = new_prog;
9246 insn = new_prog->insnsi + i + delta;
9250 if (BPF_CLASS(insn->code) == BPF_LD &&
9251 (BPF_MODE(insn->code) == BPF_ABS ||
9252 BPF_MODE(insn->code) == BPF_IND)) {
9253 cnt = env->ops->gen_ld_abs(insn, insn_buf);
9254 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
9255 verbose(env, "bpf verifier is misconfigured\n");
9259 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
9264 env->prog = prog = new_prog;
9265 insn = new_prog->insnsi + i + delta;
9269 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
9270 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
9271 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
9272 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
9273 struct bpf_insn insn_buf[16];
9274 struct bpf_insn *patch = &insn_buf[0];
9278 aux = &env->insn_aux_data[i + delta];
9279 if (!aux->alu_state ||
9280 aux->alu_state == BPF_ALU_NON_POINTER)
9283 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
9284 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
9285 BPF_ALU_SANITIZE_SRC;
9287 off_reg = issrc ? insn->src_reg : insn->dst_reg;
9289 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
9290 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit - 1);
9291 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
9292 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
9293 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
9294 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
9296 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX,
9298 insn->src_reg = BPF_REG_AX;
9300 *patch++ = BPF_ALU64_REG(BPF_AND, off_reg,
9304 insn->code = insn->code == code_add ?
9305 code_sub : code_add;
9308 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
9309 cnt = patch - insn_buf;
9311 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
9316 env->prog = prog = new_prog;
9317 insn = new_prog->insnsi + i + delta;
9321 if (insn->code != (BPF_JMP | BPF_CALL))
9323 if (insn->src_reg == BPF_PSEUDO_CALL)
9326 if (insn->imm == BPF_FUNC_get_route_realm)
9327 prog->dst_needed = 1;
9328 if (insn->imm == BPF_FUNC_get_prandom_u32)
9329 bpf_user_rnd_init_once();
9330 if (insn->imm == BPF_FUNC_override_return)
9331 prog->kprobe_override = 1;
9332 if (insn->imm == BPF_FUNC_tail_call) {
9333 /* If we tail call into other programs, we
9334 * cannot make any assumptions since they can
9335 * be replaced dynamically during runtime in
9336 * the program array.
9338 prog->cb_access = 1;
9339 env->prog->aux->stack_depth = MAX_BPF_STACK;
9340 env->prog->aux->max_pkt_offset = MAX_PACKET_OFF;
9342 /* mark bpf_tail_call as different opcode to avoid
9343 * conditional branch in the interpeter for every normal
9344 * call and to prevent accidental JITing by JIT compiler
9345 * that doesn't support bpf_tail_call yet
9348 insn->code = BPF_JMP | BPF_TAIL_CALL;
9350 aux = &env->insn_aux_data[i + delta];
9351 if (env->allow_ptr_leaks && !expect_blinding &&
9352 prog->jit_requested &&
9353 !bpf_map_key_poisoned(aux) &&
9354 !bpf_map_ptr_poisoned(aux) &&
9355 !bpf_map_ptr_unpriv(aux)) {
9356 struct bpf_jit_poke_descriptor desc = {
9357 .reason = BPF_POKE_REASON_TAIL_CALL,
9358 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
9359 .tail_call.key = bpf_map_key_immediate(aux),
9362 ret = bpf_jit_add_poke_descriptor(prog, &desc);
9364 verbose(env, "adding tail call poke descriptor failed\n");
9368 insn->imm = ret + 1;
9372 if (!bpf_map_ptr_unpriv(aux))
9375 /* instead of changing every JIT dealing with tail_call
9376 * emit two extra insns:
9377 * if (index >= max_entries) goto out;
9378 * index &= array->index_mask;
9379 * to avoid out-of-bounds cpu speculation
9381 if (bpf_map_ptr_poisoned(aux)) {
9382 verbose(env, "tail_call abusing map_ptr\n");
9386 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
9387 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
9388 map_ptr->max_entries, 2);
9389 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
9390 container_of(map_ptr,
9393 insn_buf[2] = *insn;
9395 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
9400 env->prog = prog = new_prog;
9401 insn = new_prog->insnsi + i + delta;
9405 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
9406 * and other inlining handlers are currently limited to 64 bit
9409 if (prog->jit_requested && BITS_PER_LONG == 64 &&
9410 (insn->imm == BPF_FUNC_map_lookup_elem ||
9411 insn->imm == BPF_FUNC_map_update_elem ||
9412 insn->imm == BPF_FUNC_map_delete_elem ||
9413 insn->imm == BPF_FUNC_map_push_elem ||
9414 insn->imm == BPF_FUNC_map_pop_elem ||
9415 insn->imm == BPF_FUNC_map_peek_elem)) {
9416 aux = &env->insn_aux_data[i + delta];
9417 if (bpf_map_ptr_poisoned(aux))
9418 goto patch_call_imm;
9420 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
9422 if (insn->imm == BPF_FUNC_map_lookup_elem &&
9423 ops->map_gen_lookup) {
9424 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
9425 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
9426 verbose(env, "bpf verifier is misconfigured\n");
9430 new_prog = bpf_patch_insn_data(env, i + delta,
9436 env->prog = prog = new_prog;
9437 insn = new_prog->insnsi + i + delta;
9441 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
9442 (void *(*)(struct bpf_map *map, void *key))NULL));
9443 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
9444 (int (*)(struct bpf_map *map, void *key))NULL));
9445 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
9446 (int (*)(struct bpf_map *map, void *key, void *value,
9448 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
9449 (int (*)(struct bpf_map *map, void *value,
9451 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
9452 (int (*)(struct bpf_map *map, void *value))NULL));
9453 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
9454 (int (*)(struct bpf_map *map, void *value))NULL));
9456 switch (insn->imm) {
9457 case BPF_FUNC_map_lookup_elem:
9458 insn->imm = BPF_CAST_CALL(ops->map_lookup_elem) -
9461 case BPF_FUNC_map_update_elem:
9462 insn->imm = BPF_CAST_CALL(ops->map_update_elem) -
9465 case BPF_FUNC_map_delete_elem:
9466 insn->imm = BPF_CAST_CALL(ops->map_delete_elem) -
9469 case BPF_FUNC_map_push_elem:
9470 insn->imm = BPF_CAST_CALL(ops->map_push_elem) -
9473 case BPF_FUNC_map_pop_elem:
9474 insn->imm = BPF_CAST_CALL(ops->map_pop_elem) -
9477 case BPF_FUNC_map_peek_elem:
9478 insn->imm = BPF_CAST_CALL(ops->map_peek_elem) -
9483 goto patch_call_imm;
9486 if (prog->jit_requested && BITS_PER_LONG == 64 &&
9487 insn->imm == BPF_FUNC_jiffies64) {
9488 struct bpf_insn ld_jiffies_addr[2] = {
9489 BPF_LD_IMM64(BPF_REG_0,
9490 (unsigned long)&jiffies),
9493 insn_buf[0] = ld_jiffies_addr[0];
9494 insn_buf[1] = ld_jiffies_addr[1];
9495 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
9499 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
9505 env->prog = prog = new_prog;
9506 insn = new_prog->insnsi + i + delta;
9511 fn = env->ops->get_func_proto(insn->imm, env->prog);
9512 /* all functions that have prototype and verifier allowed
9513 * programs to call them, must be real in-kernel functions
9517 "kernel subsystem misconfigured func %s#%d\n",
9518 func_id_name(insn->imm), insn->imm);
9521 insn->imm = fn->func - __bpf_call_base;
9524 /* Since poke tab is now finalized, publish aux to tracker. */
9525 for (i = 0; i < prog->aux->size_poke_tab; i++) {
9526 map_ptr = prog->aux->poke_tab[i].tail_call.map;
9527 if (!map_ptr->ops->map_poke_track ||
9528 !map_ptr->ops->map_poke_untrack ||
9529 !map_ptr->ops->map_poke_run) {
9530 verbose(env, "bpf verifier is misconfigured\n");
9534 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
9536 verbose(env, "tracking tail call prog failed\n");
9544 static void free_states(struct bpf_verifier_env *env)
9546 struct bpf_verifier_state_list *sl, *sln;
9549 sl = env->free_list;
9552 free_verifier_state(&sl->state, false);
9556 env->free_list = NULL;
9558 if (!env->explored_states)
9561 for (i = 0; i < state_htab_size(env); i++) {
9562 sl = env->explored_states[i];
9566 free_verifier_state(&sl->state, false);
9570 env->explored_states[i] = NULL;
9574 /* The verifier is using insn_aux_data[] to store temporary data during
9575 * verification and to store information for passes that run after the
9576 * verification like dead code sanitization. do_check_common() for subprogram N
9577 * may analyze many other subprograms. sanitize_insn_aux_data() clears all
9578 * temporary data after do_check_common() finds that subprogram N cannot be
9579 * verified independently. pass_cnt counts the number of times
9580 * do_check_common() was run and insn->aux->seen tells the pass number
9581 * insn_aux_data was touched. These variables are compared to clear temporary
9582 * data from failed pass. For testing and experiments do_check_common() can be
9583 * run multiple times even when prior attempt to verify is unsuccessful.
9585 static void sanitize_insn_aux_data(struct bpf_verifier_env *env)
9587 struct bpf_insn *insn = env->prog->insnsi;
9588 struct bpf_insn_aux_data *aux;
9591 for (i = 0; i < env->prog->len; i++) {
9592 class = BPF_CLASS(insn[i].code);
9593 if (class != BPF_LDX && class != BPF_STX)
9595 aux = &env->insn_aux_data[i];
9596 if (aux->seen != env->pass_cnt)
9598 memset(aux, 0, offsetof(typeof(*aux), orig_idx));
9602 static int do_check_common(struct bpf_verifier_env *env, int subprog)
9604 struct bpf_verifier_state *state;
9605 struct bpf_reg_state *regs;
9608 env->prev_linfo = NULL;
9611 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
9614 state->curframe = 0;
9615 state->speculative = false;
9616 state->branches = 1;
9617 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
9618 if (!state->frame[0]) {
9622 env->cur_state = state;
9623 init_func_state(env, state->frame[0],
9624 BPF_MAIN_FUNC /* callsite */,
9628 regs = state->frame[state->curframe]->regs;
9629 if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
9630 ret = btf_prepare_func_args(env, subprog, regs);
9633 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
9634 if (regs[i].type == PTR_TO_CTX)
9635 mark_reg_known_zero(env, regs, i);
9636 else if (regs[i].type == SCALAR_VALUE)
9637 mark_reg_unknown(env, regs, i);
9640 /* 1st arg to a function */
9641 regs[BPF_REG_1].type = PTR_TO_CTX;
9642 mark_reg_known_zero(env, regs, BPF_REG_1);
9643 ret = btf_check_func_arg_match(env, subprog, regs);
9645 /* unlikely verifier bug. abort.
9646 * ret == 0 and ret < 0 are sadly acceptable for
9647 * main() function due to backward compatibility.
9648 * Like socket filter program may be written as:
9649 * int bpf_prog(struct pt_regs *ctx)
9650 * and never dereference that ctx in the program.
9651 * 'struct pt_regs' is a type mismatch for socket
9652 * filter that should be using 'struct __sk_buff'.
9657 ret = do_check(env);
9659 /* check for NULL is necessary, since cur_state can be freed inside
9660 * do_check() under memory pressure.
9662 if (env->cur_state) {
9663 free_verifier_state(env->cur_state, true);
9664 env->cur_state = NULL;
9666 while (!pop_stack(env, NULL, NULL));
9669 /* clean aux data in case subprog was rejected */
9670 sanitize_insn_aux_data(env);
9674 /* Verify all global functions in a BPF program one by one based on their BTF.
9675 * All global functions must pass verification. Otherwise the whole program is rejected.
9686 * foo() will be verified first for R1=any_scalar_value. During verification it
9687 * will be assumed that bar() already verified successfully and call to bar()
9688 * from foo() will be checked for type match only. Later bar() will be verified
9689 * independently to check that it's safe for R1=any_scalar_value.
9691 static int do_check_subprogs(struct bpf_verifier_env *env)
9693 struct bpf_prog_aux *aux = env->prog->aux;
9696 if (!aux->func_info)
9699 for (i = 1; i < env->subprog_cnt; i++) {
9700 if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL)
9702 env->insn_idx = env->subprog_info[i].start;
9703 WARN_ON_ONCE(env->insn_idx == 0);
9704 ret = do_check_common(env, i);
9707 } else if (env->log.level & BPF_LOG_LEVEL) {
9709 "Func#%d is safe for any args that match its prototype\n",
9716 static int do_check_main(struct bpf_verifier_env *env)
9721 ret = do_check_common(env, 0);
9723 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
9728 static void print_verification_stats(struct bpf_verifier_env *env)
9732 if (env->log.level & BPF_LOG_STATS) {
9733 verbose(env, "verification time %lld usec\n",
9734 div_u64(env->verification_time, 1000));
9735 verbose(env, "stack depth ");
9736 for (i = 0; i < env->subprog_cnt; i++) {
9737 u32 depth = env->subprog_info[i].stack_depth;
9739 verbose(env, "%d", depth);
9740 if (i + 1 < env->subprog_cnt)
9745 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
9746 "total_states %d peak_states %d mark_read %d\n",
9747 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
9748 env->max_states_per_insn, env->total_states,
9749 env->peak_states, env->longest_mark_read_walk);
9752 static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
9754 const struct btf_type *t, *func_proto;
9755 const struct bpf_struct_ops *st_ops;
9756 const struct btf_member *member;
9757 struct bpf_prog *prog = env->prog;
9758 u32 btf_id, member_idx;
9761 btf_id = prog->aux->attach_btf_id;
9762 st_ops = bpf_struct_ops_find(btf_id);
9764 verbose(env, "attach_btf_id %u is not a supported struct\n",
9770 member_idx = prog->expected_attach_type;
9771 if (member_idx >= btf_type_vlen(t)) {
9772 verbose(env, "attach to invalid member idx %u of struct %s\n",
9773 member_idx, st_ops->name);
9777 member = &btf_type_member(t)[member_idx];
9778 mname = btf_name_by_offset(btf_vmlinux, member->name_off);
9779 func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type,
9782 verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
9783 mname, member_idx, st_ops->name);
9787 if (st_ops->check_member) {
9788 int err = st_ops->check_member(t, member);
9791 verbose(env, "attach to unsupported member %s of struct %s\n",
9792 mname, st_ops->name);
9797 prog->aux->attach_func_proto = func_proto;
9798 prog->aux->attach_func_name = mname;
9799 env->ops = st_ops->verifier_ops;
9804 static int check_attach_btf_id(struct bpf_verifier_env *env)
9806 struct bpf_prog *prog = env->prog;
9807 bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
9808 struct bpf_prog *tgt_prog = prog->aux->linked_prog;
9809 u32 btf_id = prog->aux->attach_btf_id;
9810 const char prefix[] = "btf_trace_";
9811 int ret = 0, subprog = -1, i;
9812 struct bpf_trampoline *tr;
9813 const struct btf_type *t;
9814 bool conservative = true;
9820 if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
9821 return check_struct_ops_btf_id(env);
9823 if (prog->type != BPF_PROG_TYPE_TRACING && !prog_extension)
9827 verbose(env, "Tracing programs must provide btf_id\n");
9830 btf = bpf_prog_get_target_btf(prog);
9833 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
9836 t = btf_type_by_id(btf, btf_id);
9838 verbose(env, "attach_btf_id %u is invalid\n", btf_id);
9841 tname = btf_name_by_offset(btf, t->name_off);
9843 verbose(env, "attach_btf_id %u doesn't have a name\n", btf_id);
9847 struct bpf_prog_aux *aux = tgt_prog->aux;
9849 for (i = 0; i < aux->func_info_cnt; i++)
9850 if (aux->func_info[i].type_id == btf_id) {
9854 if (subprog == -1) {
9855 verbose(env, "Subprog %s doesn't exist\n", tname);
9858 conservative = aux->func_info_aux[subprog].unreliable;
9859 if (prog_extension) {
9862 "Cannot replace static functions\n");
9865 if (!prog->jit_requested) {
9867 "Extension programs should be JITed\n");
9870 env->ops = bpf_verifier_ops[tgt_prog->type];
9872 if (!tgt_prog->jited) {
9873 verbose(env, "Can attach to only JITed progs\n");
9876 if (tgt_prog->type == prog->type) {
9877 /* Cannot fentry/fexit another fentry/fexit program.
9878 * Cannot attach program extension to another extension.
9879 * It's ok to attach fentry/fexit to extension program.
9881 verbose(env, "Cannot recursively attach\n");
9884 if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
9886 (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
9887 tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
9888 /* Program extensions can extend all program types
9889 * except fentry/fexit. The reason is the following.
9890 * The fentry/fexit programs are used for performance
9891 * analysis, stats and can be attached to any program
9892 * type except themselves. When extension program is
9893 * replacing XDP function it is necessary to allow
9894 * performance analysis of all functions. Both original
9895 * XDP program and its program extension. Hence
9896 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
9897 * allowed. If extending of fentry/fexit was allowed it
9898 * would be possible to create long call chain
9899 * fentry->extension->fentry->extension beyond
9900 * reasonable stack size. Hence extending fentry is not
9903 verbose(env, "Cannot extend fentry/fexit\n");
9906 key = ((u64)aux->id) << 32 | btf_id;
9908 if (prog_extension) {
9909 verbose(env, "Cannot replace kernel functions\n");
9915 switch (prog->expected_attach_type) {
9916 case BPF_TRACE_RAW_TP:
9919 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
9922 if (!btf_type_is_typedef(t)) {
9923 verbose(env, "attach_btf_id %u is not a typedef\n",
9927 if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
9928 verbose(env, "attach_btf_id %u points to wrong type name %s\n",
9932 tname += sizeof(prefix) - 1;
9933 t = btf_type_by_id(btf, t->type);
9934 if (!btf_type_is_ptr(t))
9935 /* should never happen in valid vmlinux build */
9937 t = btf_type_by_id(btf, t->type);
9938 if (!btf_type_is_func_proto(t))
9939 /* should never happen in valid vmlinux build */
9942 /* remember two read only pointers that are valid for
9943 * the life time of the kernel
9945 prog->aux->attach_func_name = tname;
9946 prog->aux->attach_func_proto = t;
9947 prog->aux->attach_btf_trace = true;
9950 if (!prog_extension)
9953 case BPF_TRACE_FENTRY:
9954 case BPF_TRACE_FEXIT:
9955 if (!btf_type_is_func(t)) {
9956 verbose(env, "attach_btf_id %u is not a function\n",
9960 if (prog_extension &&
9961 btf_check_type_match(env, prog, btf, t))
9963 t = btf_type_by_id(btf, t->type);
9964 if (!btf_type_is_func_proto(t))
9966 tr = bpf_trampoline_lookup(key);
9969 prog->aux->attach_func_name = tname;
9970 /* t is either vmlinux type or another program's type */
9971 prog->aux->attach_func_proto = t;
9972 mutex_lock(&tr->mutex);
9973 if (tr->func.addr) {
9974 prog->aux->trampoline = tr;
9977 if (tgt_prog && conservative) {
9978 prog->aux->attach_func_proto = NULL;
9981 ret = btf_distill_func_proto(&env->log, btf, t,
9982 tname, &tr->func.model);
9987 addr = (long) tgt_prog->bpf_func;
9989 addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
9991 addr = kallsyms_lookup_name(tname);
9994 "The address of function %s cannot be found\n",
10000 tr->func.addr = (void *)addr;
10001 prog->aux->trampoline = tr;
10003 mutex_unlock(&tr->mutex);
10005 bpf_trampoline_put(tr);
10010 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr,
10011 union bpf_attr __user *uattr)
10013 u64 start_time = ktime_get_ns();
10014 struct bpf_verifier_env *env;
10015 struct bpf_verifier_log *log;
10016 int i, len, ret = -EINVAL;
10019 /* no program is valid */
10020 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
10023 /* 'struct bpf_verifier_env' can be global, but since it's not small,
10024 * allocate/free it every time bpf_check() is called
10026 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
10031 len = (*prog)->len;
10032 env->insn_aux_data =
10033 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
10035 if (!env->insn_aux_data)
10037 for (i = 0; i < len; i++)
10038 env->insn_aux_data[i].orig_idx = i;
10040 env->ops = bpf_verifier_ops[env->prog->type];
10041 is_priv = capable(CAP_SYS_ADMIN);
10043 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
10044 mutex_lock(&bpf_verifier_lock);
10046 btf_vmlinux = btf_parse_vmlinux();
10047 mutex_unlock(&bpf_verifier_lock);
10050 /* grab the mutex to protect few globals used by verifier */
10052 mutex_lock(&bpf_verifier_lock);
10054 if (attr->log_level || attr->log_buf || attr->log_size) {
10055 /* user requested verbose verifier output
10056 * and supplied buffer to store the verification trace
10058 log->level = attr->log_level;
10059 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
10060 log->len_total = attr->log_size;
10063 /* log attributes have to be sane */
10064 if (log->len_total < 128 || log->len_total > UINT_MAX >> 2 ||
10065 !log->level || !log->ubuf || log->level & ~BPF_LOG_MASK)
10069 if (IS_ERR(btf_vmlinux)) {
10070 /* Either gcc or pahole or kernel are broken. */
10071 verbose(env, "in-kernel BTF is malformed\n");
10072 ret = PTR_ERR(btf_vmlinux);
10073 goto skip_full_check;
10076 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
10077 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
10078 env->strict_alignment = true;
10079 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
10080 env->strict_alignment = false;
10082 env->allow_ptr_leaks = is_priv;
10085 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
10087 ret = replace_map_fd_with_map_ptr(env);
10089 goto skip_full_check;
10091 if (bpf_prog_is_dev_bound(env->prog->aux)) {
10092 ret = bpf_prog_offload_verifier_prep(env->prog);
10094 goto skip_full_check;
10097 env->explored_states = kvcalloc(state_htab_size(env),
10098 sizeof(struct bpf_verifier_state_list *),
10101 if (!env->explored_states)
10102 goto skip_full_check;
10104 ret = check_subprogs(env);
10106 goto skip_full_check;
10108 ret = check_btf_info(env, attr, uattr);
10110 goto skip_full_check;
10112 ret = check_attach_btf_id(env);
10114 goto skip_full_check;
10116 ret = check_cfg(env);
10118 goto skip_full_check;
10120 ret = do_check_subprogs(env);
10121 ret = ret ?: do_check_main(env);
10123 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
10124 ret = bpf_prog_offload_finalize(env);
10127 kvfree(env->explored_states);
10130 ret = check_max_stack_depth(env);
10132 /* instruction rewrites happen after this point */
10135 opt_hard_wire_dead_code_branches(env);
10137 ret = opt_remove_dead_code(env);
10139 ret = opt_remove_nops(env);
10142 sanitize_dead_code(env);
10146 /* program is valid, convert *(u32*)(ctx + off) accesses */
10147 ret = convert_ctx_accesses(env);
10150 ret = fixup_bpf_calls(env);
10152 /* do 32-bit optimization after insn patching has done so those patched
10153 * insns could be handled correctly.
10155 if (ret == 0 && !bpf_prog_is_dev_bound(env->prog->aux)) {
10156 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
10157 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
10162 ret = fixup_call_args(env);
10164 env->verification_time = ktime_get_ns() - start_time;
10165 print_verification_stats(env);
10167 if (log->level && bpf_verifier_log_full(log))
10169 if (log->level && !log->ubuf) {
10171 goto err_release_maps;
10174 if (ret == 0 && env->used_map_cnt) {
10175 /* if program passed verifier, update used_maps in bpf_prog_info */
10176 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
10177 sizeof(env->used_maps[0]),
10180 if (!env->prog->aux->used_maps) {
10182 goto err_release_maps;
10185 memcpy(env->prog->aux->used_maps, env->used_maps,
10186 sizeof(env->used_maps[0]) * env->used_map_cnt);
10187 env->prog->aux->used_map_cnt = env->used_map_cnt;
10189 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
10190 * bpf_ld_imm64 instructions
10192 convert_pseudo_ld_imm64(env);
10196 adjust_btf_func(env);
10199 if (!env->prog->aux->used_maps)
10200 /* if we didn't copy map pointers into bpf_prog_info, release
10201 * them now. Otherwise free_used_maps() will release them.
10207 mutex_unlock(&bpf_verifier_lock);
10208 vfree(env->insn_aux_data);