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>
22 #include <linux/error-injection.h>
23 #include <linux/bpf_lsm.h>
24 #include <linux/btf_ids.h>
28 static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
29 #define BPF_PROG_TYPE(_id, _name, prog_ctx_type, kern_ctx_type) \
30 [_id] = & _name ## _verifier_ops,
31 #define BPF_MAP_TYPE(_id, _ops)
32 #define BPF_LINK_TYPE(_id, _name)
33 #include <linux/bpf_types.h>
39 /* bpf_check() is a static code analyzer that walks eBPF program
40 * instruction by instruction and updates register/stack state.
41 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
43 * The first pass is depth-first-search to check that the program is a DAG.
44 * It rejects the following programs:
45 * - larger than BPF_MAXINSNS insns
46 * - if loop is present (detected via back-edge)
47 * - unreachable insns exist (shouldn't be a forest. program = one function)
48 * - out of bounds or malformed jumps
49 * The second pass is all possible path descent from the 1st insn.
50 * Since it's analyzing all paths through the program, the length of the
51 * analysis is limited to 64k insn, which may be hit even if total number of
52 * insn is less then 4K, but there are too many branches that change stack/regs.
53 * Number of 'branches to be analyzed' is limited to 1k
55 * On entry to each instruction, each register has a type, and the instruction
56 * changes the types of the registers depending on instruction semantics.
57 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
60 * All registers are 64-bit.
61 * R0 - return register
62 * R1-R5 argument passing registers
63 * R6-R9 callee saved registers
64 * R10 - frame pointer read-only
66 * At the start of BPF program the register R1 contains a pointer to bpf_context
67 * and has type PTR_TO_CTX.
69 * Verifier tracks arithmetic operations on pointers in case:
70 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
71 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
72 * 1st insn copies R10 (which has FRAME_PTR) type into R1
73 * and 2nd arithmetic instruction is pattern matched to recognize
74 * that it wants to construct a pointer to some element within stack.
75 * So after 2nd insn, the register R1 has type PTR_TO_STACK
76 * (and -20 constant is saved for further stack bounds checking).
77 * Meaning that this reg is a pointer to stack plus known immediate constant.
79 * Most of the time the registers have SCALAR_VALUE type, which
80 * means the register has some value, but it's not a valid pointer.
81 * (like pointer plus pointer becomes SCALAR_VALUE type)
83 * When verifier sees load or store instructions the type of base register
84 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK, PTR_TO_SOCKET. These are
85 * four pointer types recognized by check_mem_access() function.
87 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
88 * and the range of [ptr, ptr + map's value_size) is accessible.
90 * registers used to pass values to function calls are checked against
91 * function argument constraints.
93 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
94 * It means that the register type passed to this function must be
95 * PTR_TO_STACK and it will be used inside the function as
96 * 'pointer to map element key'
98 * For example the argument constraints for bpf_map_lookup_elem():
99 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
100 * .arg1_type = ARG_CONST_MAP_PTR,
101 * .arg2_type = ARG_PTR_TO_MAP_KEY,
103 * ret_type says that this function returns 'pointer to map elem value or null'
104 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
105 * 2nd argument should be a pointer to stack, which will be used inside
106 * the helper function as a pointer to map element key.
108 * On the kernel side the helper function looks like:
109 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
111 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
112 * void *key = (void *) (unsigned long) r2;
115 * here kernel can access 'key' and 'map' pointers safely, knowing that
116 * [key, key + map->key_size) bytes are valid and were initialized on
117 * the stack of eBPF program.
120 * Corresponding eBPF program may look like:
121 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
122 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
123 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
124 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
125 * here verifier looks at prototype of map_lookup_elem() and sees:
126 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
127 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
129 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
130 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
131 * and were initialized prior to this call.
132 * If it's ok, then verifier allows this BPF_CALL insn and looks at
133 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
134 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
135 * returns either pointer to map value or NULL.
137 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
138 * insn, the register holding that pointer in the true branch changes state to
139 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
140 * branch. See check_cond_jmp_op().
142 * After the call R0 is set to return type of the function and registers R1-R5
143 * are set to NOT_INIT to indicate that they are no longer readable.
145 * The following reference types represent a potential reference to a kernel
146 * resource which, after first being allocated, must be checked and freed by
148 * - PTR_TO_SOCKET_OR_NULL, PTR_TO_SOCKET
150 * When the verifier sees a helper call return a reference type, it allocates a
151 * pointer id for the reference and stores it in the current function state.
152 * Similar to the way that PTR_TO_MAP_VALUE_OR_NULL is converted into
153 * PTR_TO_MAP_VALUE, PTR_TO_SOCKET_OR_NULL becomes PTR_TO_SOCKET when the type
154 * passes through a NULL-check conditional. For the branch wherein the state is
155 * changed to CONST_IMM, the verifier releases the reference.
157 * For each helper function that allocates a reference, such as
158 * bpf_sk_lookup_tcp(), there is a corresponding release function, such as
159 * bpf_sk_release(). When a reference type passes into the release function,
160 * the verifier also releases the reference. If any unchecked or unreleased
161 * reference remains at the end of the program, the verifier rejects it.
164 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
165 struct bpf_verifier_stack_elem {
166 /* verifer state is 'st'
167 * before processing instruction 'insn_idx'
168 * and after processing instruction 'prev_insn_idx'
170 struct bpf_verifier_state st;
173 struct bpf_verifier_stack_elem *next;
174 /* length of verifier log at the time this state was pushed on stack */
178 #define BPF_COMPLEXITY_LIMIT_JMP_SEQ 8192
179 #define BPF_COMPLEXITY_LIMIT_STATES 64
181 #define BPF_MAP_KEY_POISON (1ULL << 63)
182 #define BPF_MAP_KEY_SEEN (1ULL << 62)
184 #define BPF_MAP_PTR_UNPRIV 1UL
185 #define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \
186 POISON_POINTER_DELTA))
187 #define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
189 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
191 return BPF_MAP_PTR(aux->map_ptr_state) == BPF_MAP_PTR_POISON;
194 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
196 return aux->map_ptr_state & BPF_MAP_PTR_UNPRIV;
199 static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
200 const struct bpf_map *map, bool unpriv)
202 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON & BPF_MAP_PTR_UNPRIV);
203 unpriv |= bpf_map_ptr_unpriv(aux);
204 aux->map_ptr_state = (unsigned long)map |
205 (unpriv ? BPF_MAP_PTR_UNPRIV : 0UL);
208 static bool bpf_map_key_poisoned(const struct bpf_insn_aux_data *aux)
210 return aux->map_key_state & BPF_MAP_KEY_POISON;
213 static bool bpf_map_key_unseen(const struct bpf_insn_aux_data *aux)
215 return !(aux->map_key_state & BPF_MAP_KEY_SEEN);
218 static u64 bpf_map_key_immediate(const struct bpf_insn_aux_data *aux)
220 return aux->map_key_state & ~(BPF_MAP_KEY_SEEN | BPF_MAP_KEY_POISON);
223 static void bpf_map_key_store(struct bpf_insn_aux_data *aux, u64 state)
225 bool poisoned = bpf_map_key_poisoned(aux);
227 aux->map_key_state = state | BPF_MAP_KEY_SEEN |
228 (poisoned ? BPF_MAP_KEY_POISON : 0ULL);
231 static bool bpf_pseudo_call(const struct bpf_insn *insn)
233 return insn->code == (BPF_JMP | BPF_CALL) &&
234 insn->src_reg == BPF_PSEUDO_CALL;
237 static bool bpf_pseudo_kfunc_call(const struct bpf_insn *insn)
239 return insn->code == (BPF_JMP | BPF_CALL) &&
240 insn->src_reg == BPF_PSEUDO_KFUNC_CALL;
243 static bool bpf_pseudo_func(const struct bpf_insn *insn)
245 return insn->code == (BPF_LD | BPF_IMM | BPF_DW) &&
246 insn->src_reg == BPF_PSEUDO_FUNC;
249 struct bpf_call_arg_meta {
250 struct bpf_map *map_ptr;
267 struct btf *btf_vmlinux;
269 static DEFINE_MUTEX(bpf_verifier_lock);
271 static const struct bpf_line_info *
272 find_linfo(const struct bpf_verifier_env *env, u32 insn_off)
274 const struct bpf_line_info *linfo;
275 const struct bpf_prog *prog;
279 nr_linfo = prog->aux->nr_linfo;
281 if (!nr_linfo || insn_off >= prog->len)
284 linfo = prog->aux->linfo;
285 for (i = 1; i < nr_linfo; i++)
286 if (insn_off < linfo[i].insn_off)
289 return &linfo[i - 1];
292 void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
297 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
299 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
300 "verifier log line truncated - local buffer too short\n");
302 n = min(log->len_total - log->len_used - 1, n);
305 if (log->level == BPF_LOG_KERNEL) {
306 pr_err("BPF:%s\n", log->kbuf);
309 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
315 static void bpf_vlog_reset(struct bpf_verifier_log *log, u32 new_pos)
319 if (!bpf_verifier_log_needed(log))
322 log->len_used = new_pos;
323 if (put_user(zero, log->ubuf + new_pos))
327 /* log_level controls verbosity level of eBPF verifier.
328 * bpf_verifier_log_write() is used to dump the verification trace to the log,
329 * so the user can figure out what's wrong with the program
331 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
332 const char *fmt, ...)
336 if (!bpf_verifier_log_needed(&env->log))
340 bpf_verifier_vlog(&env->log, fmt, args);
343 EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
345 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
347 struct bpf_verifier_env *env = private_data;
350 if (!bpf_verifier_log_needed(&env->log))
354 bpf_verifier_vlog(&env->log, fmt, args);
358 __printf(2, 3) void bpf_log(struct bpf_verifier_log *log,
359 const char *fmt, ...)
363 if (!bpf_verifier_log_needed(log))
367 bpf_verifier_vlog(log, fmt, args);
371 static const char *ltrim(const char *s)
379 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env,
381 const char *prefix_fmt, ...)
383 const struct bpf_line_info *linfo;
385 if (!bpf_verifier_log_needed(&env->log))
388 linfo = find_linfo(env, insn_off);
389 if (!linfo || linfo == env->prev_linfo)
395 va_start(args, prefix_fmt);
396 bpf_verifier_vlog(&env->log, prefix_fmt, args);
401 ltrim(btf_name_by_offset(env->prog->aux->btf,
404 env->prev_linfo = linfo;
407 static void verbose_invalid_scalar(struct bpf_verifier_env *env,
408 struct bpf_reg_state *reg,
409 struct tnum *range, const char *ctx,
410 const char *reg_name)
414 verbose(env, "At %s the register %s ", ctx, reg_name);
415 if (!tnum_is_unknown(reg->var_off)) {
416 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
417 verbose(env, "has value %s", tn_buf);
419 verbose(env, "has unknown scalar value");
421 tnum_strn(tn_buf, sizeof(tn_buf), *range);
422 verbose(env, " should have been in %s\n", tn_buf);
425 static bool type_is_pkt_pointer(enum bpf_reg_type type)
427 return type == PTR_TO_PACKET ||
428 type == PTR_TO_PACKET_META;
431 static bool type_is_sk_pointer(enum bpf_reg_type type)
433 return type == PTR_TO_SOCKET ||
434 type == PTR_TO_SOCK_COMMON ||
435 type == PTR_TO_TCP_SOCK ||
436 type == PTR_TO_XDP_SOCK;
439 static bool reg_type_not_null(enum bpf_reg_type type)
441 return type == PTR_TO_SOCKET ||
442 type == PTR_TO_TCP_SOCK ||
443 type == PTR_TO_MAP_VALUE ||
444 type == PTR_TO_MAP_KEY ||
445 type == PTR_TO_SOCK_COMMON;
448 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
450 return reg->type == PTR_TO_MAP_VALUE &&
451 map_value_has_spin_lock(reg->map_ptr);
454 static bool reg_type_may_be_refcounted_or_null(enum bpf_reg_type type)
456 return base_type(type) == PTR_TO_SOCKET ||
457 base_type(type) == PTR_TO_TCP_SOCK ||
458 base_type(type) == PTR_TO_MEM;
461 static bool type_is_rdonly_mem(u32 type)
463 return type & MEM_RDONLY;
466 static bool arg_type_may_be_refcounted(enum bpf_arg_type type)
468 return type == ARG_PTR_TO_SOCK_COMMON;
471 static bool type_may_be_null(u32 type)
473 return type & PTR_MAYBE_NULL;
476 /* Determine whether the function releases some resources allocated by another
477 * function call. The first reference type argument will be assumed to be
478 * released by release_reference().
480 static bool is_release_function(enum bpf_func_id func_id)
482 return func_id == BPF_FUNC_sk_release ||
483 func_id == BPF_FUNC_ringbuf_submit ||
484 func_id == BPF_FUNC_ringbuf_discard;
487 static bool may_be_acquire_function(enum bpf_func_id func_id)
489 return func_id == BPF_FUNC_sk_lookup_tcp ||
490 func_id == BPF_FUNC_sk_lookup_udp ||
491 func_id == BPF_FUNC_skc_lookup_tcp ||
492 func_id == BPF_FUNC_map_lookup_elem ||
493 func_id == BPF_FUNC_ringbuf_reserve;
496 static bool is_acquire_function(enum bpf_func_id func_id,
497 const struct bpf_map *map)
499 enum bpf_map_type map_type = map ? map->map_type : BPF_MAP_TYPE_UNSPEC;
501 if (func_id == BPF_FUNC_sk_lookup_tcp ||
502 func_id == BPF_FUNC_sk_lookup_udp ||
503 func_id == BPF_FUNC_skc_lookup_tcp ||
504 func_id == BPF_FUNC_ringbuf_reserve)
507 if (func_id == BPF_FUNC_map_lookup_elem &&
508 (map_type == BPF_MAP_TYPE_SOCKMAP ||
509 map_type == BPF_MAP_TYPE_SOCKHASH))
515 static bool is_ptr_cast_function(enum bpf_func_id func_id)
517 return func_id == BPF_FUNC_tcp_sock ||
518 func_id == BPF_FUNC_sk_fullsock ||
519 func_id == BPF_FUNC_skc_to_tcp_sock ||
520 func_id == BPF_FUNC_skc_to_tcp6_sock ||
521 func_id == BPF_FUNC_skc_to_udp6_sock ||
522 func_id == BPF_FUNC_skc_to_tcp_timewait_sock ||
523 func_id == BPF_FUNC_skc_to_tcp_request_sock;
526 static bool is_cmpxchg_insn(const struct bpf_insn *insn)
528 return BPF_CLASS(insn->code) == BPF_STX &&
529 BPF_MODE(insn->code) == BPF_ATOMIC &&
530 insn->imm == BPF_CMPXCHG;
533 /* string representation of 'enum bpf_reg_type'
535 * Note that reg_type_str() can not appear more than once in a single verbose()
538 static const char *reg_type_str(struct bpf_verifier_env *env,
539 enum bpf_reg_type type)
541 char postfix[16] = {0}, prefix[16] = {0};
542 static const char * const str[] = {
544 [SCALAR_VALUE] = "inv",
545 [PTR_TO_CTX] = "ctx",
546 [CONST_PTR_TO_MAP] = "map_ptr",
547 [PTR_TO_MAP_VALUE] = "map_value",
548 [PTR_TO_STACK] = "fp",
549 [PTR_TO_PACKET] = "pkt",
550 [PTR_TO_PACKET_META] = "pkt_meta",
551 [PTR_TO_PACKET_END] = "pkt_end",
552 [PTR_TO_FLOW_KEYS] = "flow_keys",
553 [PTR_TO_SOCKET] = "sock",
554 [PTR_TO_SOCK_COMMON] = "sock_common",
555 [PTR_TO_TCP_SOCK] = "tcp_sock",
556 [PTR_TO_TP_BUFFER] = "tp_buffer",
557 [PTR_TO_XDP_SOCK] = "xdp_sock",
558 [PTR_TO_BTF_ID] = "ptr_",
559 [PTR_TO_PERCPU_BTF_ID] = "percpu_ptr_",
560 [PTR_TO_MEM] = "mem",
561 [PTR_TO_BUF] = "buf",
562 [PTR_TO_FUNC] = "func",
563 [PTR_TO_MAP_KEY] = "map_key",
566 if (type & PTR_MAYBE_NULL) {
567 if (base_type(type) == PTR_TO_BTF_ID ||
568 base_type(type) == PTR_TO_PERCPU_BTF_ID)
569 strncpy(postfix, "or_null_", 16);
571 strncpy(postfix, "_or_null", 16);
574 if (type & MEM_RDONLY)
575 strncpy(prefix, "rdonly_", 16);
577 snprintf(env->type_str_buf, TYPE_STR_BUF_LEN, "%s%s%s",
578 prefix, str[base_type(type)], postfix);
579 return env->type_str_buf;
582 static char slot_type_char[] = {
583 [STACK_INVALID] = '?',
589 static void print_liveness(struct bpf_verifier_env *env,
590 enum bpf_reg_liveness live)
592 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
594 if (live & REG_LIVE_READ)
596 if (live & REG_LIVE_WRITTEN)
598 if (live & REG_LIVE_DONE)
602 static struct bpf_func_state *func(struct bpf_verifier_env *env,
603 const struct bpf_reg_state *reg)
605 struct bpf_verifier_state *cur = env->cur_state;
607 return cur->frame[reg->frameno];
610 static const char *kernel_type_name(const struct btf* btf, u32 id)
612 return btf_name_by_offset(btf, btf_type_by_id(btf, id)->name_off);
615 static void print_verifier_state(struct bpf_verifier_env *env,
616 const struct bpf_func_state *state)
618 const struct bpf_reg_state *reg;
623 verbose(env, " frame%d:", state->frameno);
624 for (i = 0; i < MAX_BPF_REG; i++) {
625 reg = &state->regs[i];
629 verbose(env, " R%d", i);
630 print_liveness(env, reg->live);
631 verbose(env, "=%s", reg_type_str(env, t));
632 if (t == SCALAR_VALUE && reg->precise)
634 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
635 tnum_is_const(reg->var_off)) {
636 /* reg->off should be 0 for SCALAR_VALUE */
637 verbose(env, "%lld", reg->var_off.value + reg->off);
639 if (base_type(t) == PTR_TO_BTF_ID ||
640 base_type(t) == PTR_TO_PERCPU_BTF_ID)
641 verbose(env, "%s", kernel_type_name(reg->btf, reg->btf_id));
642 verbose(env, "(id=%d", reg->id);
643 if (reg_type_may_be_refcounted_or_null(t))
644 verbose(env, ",ref_obj_id=%d", reg->ref_obj_id);
645 if (t != SCALAR_VALUE)
646 verbose(env, ",off=%d", reg->off);
647 if (type_is_pkt_pointer(t))
648 verbose(env, ",r=%d", reg->range);
649 else if (base_type(t) == CONST_PTR_TO_MAP ||
650 base_type(t) == PTR_TO_MAP_KEY ||
651 base_type(t) == PTR_TO_MAP_VALUE)
652 verbose(env, ",ks=%d,vs=%d",
653 reg->map_ptr->key_size,
654 reg->map_ptr->value_size);
655 if (tnum_is_const(reg->var_off)) {
656 /* Typically an immediate SCALAR_VALUE, but
657 * could be a pointer whose offset is too big
660 verbose(env, ",imm=%llx", reg->var_off.value);
662 if (reg->smin_value != reg->umin_value &&
663 reg->smin_value != S64_MIN)
664 verbose(env, ",smin_value=%lld",
665 (long long)reg->smin_value);
666 if (reg->smax_value != reg->umax_value &&
667 reg->smax_value != S64_MAX)
668 verbose(env, ",smax_value=%lld",
669 (long long)reg->smax_value);
670 if (reg->umin_value != 0)
671 verbose(env, ",umin_value=%llu",
672 (unsigned long long)reg->umin_value);
673 if (reg->umax_value != U64_MAX)
674 verbose(env, ",umax_value=%llu",
675 (unsigned long long)reg->umax_value);
676 if (!tnum_is_unknown(reg->var_off)) {
679 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
680 verbose(env, ",var_off=%s", tn_buf);
682 if (reg->s32_min_value != reg->smin_value &&
683 reg->s32_min_value != S32_MIN)
684 verbose(env, ",s32_min_value=%d",
685 (int)(reg->s32_min_value));
686 if (reg->s32_max_value != reg->smax_value &&
687 reg->s32_max_value != S32_MAX)
688 verbose(env, ",s32_max_value=%d",
689 (int)(reg->s32_max_value));
690 if (reg->u32_min_value != reg->umin_value &&
691 reg->u32_min_value != U32_MIN)
692 verbose(env, ",u32_min_value=%d",
693 (int)(reg->u32_min_value));
694 if (reg->u32_max_value != reg->umax_value &&
695 reg->u32_max_value != U32_MAX)
696 verbose(env, ",u32_max_value=%d",
697 (int)(reg->u32_max_value));
702 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
703 char types_buf[BPF_REG_SIZE + 1];
707 for (j = 0; j < BPF_REG_SIZE; j++) {
708 if (state->stack[i].slot_type[j] != STACK_INVALID)
710 types_buf[j] = slot_type_char[
711 state->stack[i].slot_type[j]];
713 types_buf[BPF_REG_SIZE] = 0;
716 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
717 print_liveness(env, state->stack[i].spilled_ptr.live);
718 if (state->stack[i].slot_type[0] == STACK_SPILL) {
719 reg = &state->stack[i].spilled_ptr;
721 verbose(env, "=%s", reg_type_str(env, t));
722 if (t == SCALAR_VALUE && reg->precise)
724 if (t == SCALAR_VALUE && tnum_is_const(reg->var_off))
725 verbose(env, "%lld", reg->var_off.value + reg->off);
727 verbose(env, "=%s", types_buf);
730 if (state->acquired_refs && state->refs[0].id) {
731 verbose(env, " refs=%d", state->refs[0].id);
732 for (i = 1; i < state->acquired_refs; i++)
733 if (state->refs[i].id)
734 verbose(env, ",%d", state->refs[i].id);
736 if (state->in_callback_fn)
738 if (state->in_async_callback_fn)
739 verbose(env, " async_cb");
743 /* copy array src of length n * size bytes to dst. dst is reallocated if it's too
744 * small to hold src. This is different from krealloc since we don't want to preserve
745 * the contents of dst.
747 * Leaves dst untouched if src is NULL or length is zero. Returns NULL if memory could
750 static void *copy_array(void *dst, const void *src, size_t n, size_t size, gfp_t flags)
754 if (ZERO_OR_NULL_PTR(src))
757 if (unlikely(check_mul_overflow(n, size, &bytes)))
760 if (ksize(dst) < bytes) {
762 dst = kmalloc_track_caller(bytes, flags);
767 memcpy(dst, src, bytes);
769 return dst ? dst : ZERO_SIZE_PTR;
772 /* resize an array from old_n items to new_n items. the array is reallocated if it's too
773 * small to hold new_n items. new items are zeroed out if the array grows.
775 * Contrary to krealloc_array, does not free arr if new_n is zero.
777 static void *realloc_array(void *arr, size_t old_n, size_t new_n, size_t size)
779 if (!new_n || old_n == new_n)
782 arr = krealloc_array(arr, new_n, size, GFP_KERNEL);
787 memset(arr + old_n * size, 0, (new_n - old_n) * size);
790 return arr ? arr : ZERO_SIZE_PTR;
793 static int copy_reference_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
795 dst->refs = copy_array(dst->refs, src->refs, src->acquired_refs,
796 sizeof(struct bpf_reference_state), GFP_KERNEL);
800 dst->acquired_refs = src->acquired_refs;
804 static int copy_stack_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
806 size_t n = src->allocated_stack / BPF_REG_SIZE;
808 dst->stack = copy_array(dst->stack, src->stack, n, sizeof(struct bpf_stack_state),
813 dst->allocated_stack = src->allocated_stack;
817 static int resize_reference_state(struct bpf_func_state *state, size_t n)
819 state->refs = realloc_array(state->refs, state->acquired_refs, n,
820 sizeof(struct bpf_reference_state));
824 state->acquired_refs = n;
828 static int grow_stack_state(struct bpf_func_state *state, int size)
830 size_t old_n = state->allocated_stack / BPF_REG_SIZE, n = size / BPF_REG_SIZE;
835 state->stack = realloc_array(state->stack, old_n, n, sizeof(struct bpf_stack_state));
839 state->allocated_stack = size;
843 /* Acquire a pointer id from the env and update the state->refs to include
844 * this new pointer reference.
845 * On success, returns a valid pointer id to associate with the register
846 * On failure, returns a negative errno.
848 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
850 struct bpf_func_state *state = cur_func(env);
851 int new_ofs = state->acquired_refs;
854 err = resize_reference_state(state, state->acquired_refs + 1);
858 state->refs[new_ofs].id = id;
859 state->refs[new_ofs].insn_idx = insn_idx;
864 /* release function corresponding to acquire_reference_state(). Idempotent. */
865 static int release_reference_state(struct bpf_func_state *state, int ptr_id)
869 last_idx = state->acquired_refs - 1;
870 for (i = 0; i < state->acquired_refs; i++) {
871 if (state->refs[i].id == ptr_id) {
872 if (last_idx && i != last_idx)
873 memcpy(&state->refs[i], &state->refs[last_idx],
874 sizeof(*state->refs));
875 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
876 state->acquired_refs--;
883 static void free_func_state(struct bpf_func_state *state)
892 static void clear_jmp_history(struct bpf_verifier_state *state)
894 kfree(state->jmp_history);
895 state->jmp_history = NULL;
896 state->jmp_history_cnt = 0;
899 static void free_verifier_state(struct bpf_verifier_state *state,
904 for (i = 0; i <= state->curframe; i++) {
905 free_func_state(state->frame[i]);
906 state->frame[i] = NULL;
908 clear_jmp_history(state);
913 /* copy verifier state from src to dst growing dst stack space
914 * when necessary to accommodate larger src stack
916 static int copy_func_state(struct bpf_func_state *dst,
917 const struct bpf_func_state *src)
921 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
922 err = copy_reference_state(dst, src);
925 return copy_stack_state(dst, src);
928 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
929 const struct bpf_verifier_state *src)
931 struct bpf_func_state *dst;
934 dst_state->jmp_history = copy_array(dst_state->jmp_history, src->jmp_history,
935 src->jmp_history_cnt, sizeof(struct bpf_idx_pair),
937 if (!dst_state->jmp_history)
939 dst_state->jmp_history_cnt = src->jmp_history_cnt;
941 /* if dst has more stack frames then src frame, free them */
942 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
943 free_func_state(dst_state->frame[i]);
944 dst_state->frame[i] = NULL;
946 dst_state->speculative = src->speculative;
947 dst_state->curframe = src->curframe;
948 dst_state->active_spin_lock = src->active_spin_lock;
949 dst_state->branches = src->branches;
950 dst_state->parent = src->parent;
951 dst_state->first_insn_idx = src->first_insn_idx;
952 dst_state->last_insn_idx = src->last_insn_idx;
953 for (i = 0; i <= src->curframe; i++) {
954 dst = dst_state->frame[i];
956 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
959 dst_state->frame[i] = dst;
961 err = copy_func_state(dst, src->frame[i]);
968 static void update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
971 u32 br = --st->branches;
973 /* WARN_ON(br > 1) technically makes sense here,
974 * but see comment in push_stack(), hence:
976 WARN_ONCE((int)br < 0,
977 "BUG update_branch_counts:branches_to_explore=%d\n",
985 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
986 int *insn_idx, bool pop_log)
988 struct bpf_verifier_state *cur = env->cur_state;
989 struct bpf_verifier_stack_elem *elem, *head = env->head;
992 if (env->head == NULL)
996 err = copy_verifier_state(cur, &head->st);
1001 bpf_vlog_reset(&env->log, head->log_pos);
1003 *insn_idx = head->insn_idx;
1005 *prev_insn_idx = head->prev_insn_idx;
1007 free_verifier_state(&head->st, false);
1014 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
1015 int insn_idx, int prev_insn_idx,
1018 struct bpf_verifier_state *cur = env->cur_state;
1019 struct bpf_verifier_stack_elem *elem;
1022 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
1026 elem->insn_idx = insn_idx;
1027 elem->prev_insn_idx = prev_insn_idx;
1028 elem->next = env->head;
1029 elem->log_pos = env->log.len_used;
1032 err = copy_verifier_state(&elem->st, cur);
1035 elem->st.speculative |= speculative;
1036 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
1037 verbose(env, "The sequence of %d jumps is too complex.\n",
1041 if (elem->st.parent) {
1042 ++elem->st.parent->branches;
1043 /* WARN_ON(branches > 2) technically makes sense here,
1045 * 1. speculative states will bump 'branches' for non-branch
1047 * 2. is_state_visited() heuristics may decide not to create
1048 * a new state for a sequence of branches and all such current
1049 * and cloned states will be pointing to a single parent state
1050 * which might have large 'branches' count.
1055 free_verifier_state(env->cur_state, true);
1056 env->cur_state = NULL;
1057 /* pop all elements and return */
1058 while (!pop_stack(env, NULL, NULL, false));
1062 #define CALLER_SAVED_REGS 6
1063 static const int caller_saved[CALLER_SAVED_REGS] = {
1064 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
1067 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1068 struct bpf_reg_state *reg);
1070 /* This helper doesn't clear reg->id */
1071 static void ___mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1073 reg->var_off = tnum_const(imm);
1074 reg->smin_value = (s64)imm;
1075 reg->smax_value = (s64)imm;
1076 reg->umin_value = imm;
1077 reg->umax_value = imm;
1079 reg->s32_min_value = (s32)imm;
1080 reg->s32_max_value = (s32)imm;
1081 reg->u32_min_value = (u32)imm;
1082 reg->u32_max_value = (u32)imm;
1085 /* Mark the unknown part of a register (variable offset or scalar value) as
1086 * known to have the value @imm.
1088 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1090 /* Clear id, off, and union(map_ptr, range) */
1091 memset(((u8 *)reg) + sizeof(reg->type), 0,
1092 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
1093 ___mark_reg_known(reg, imm);
1096 static void __mark_reg32_known(struct bpf_reg_state *reg, u64 imm)
1098 reg->var_off = tnum_const_subreg(reg->var_off, imm);
1099 reg->s32_min_value = (s32)imm;
1100 reg->s32_max_value = (s32)imm;
1101 reg->u32_min_value = (u32)imm;
1102 reg->u32_max_value = (u32)imm;
1105 /* Mark the 'variable offset' part of a register as zero. This should be
1106 * used only on registers holding a pointer type.
1108 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
1110 __mark_reg_known(reg, 0);
1113 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
1115 __mark_reg_known(reg, 0);
1116 reg->type = SCALAR_VALUE;
1119 static void mark_reg_known_zero(struct bpf_verifier_env *env,
1120 struct bpf_reg_state *regs, u32 regno)
1122 if (WARN_ON(regno >= MAX_BPF_REG)) {
1123 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
1124 /* Something bad happened, let's kill all regs */
1125 for (regno = 0; regno < MAX_BPF_REG; regno++)
1126 __mark_reg_not_init(env, regs + regno);
1129 __mark_reg_known_zero(regs + regno);
1132 static void mark_ptr_not_null_reg(struct bpf_reg_state *reg)
1134 if (base_type(reg->type) == PTR_TO_MAP_VALUE) {
1135 const struct bpf_map *map = reg->map_ptr;
1137 if (map->inner_map_meta) {
1138 reg->type = CONST_PTR_TO_MAP;
1139 reg->map_ptr = map->inner_map_meta;
1140 /* transfer reg's id which is unique for every map_lookup_elem
1141 * as UID of the inner map.
1143 if (map_value_has_timer(map->inner_map_meta))
1144 reg->map_uid = reg->id;
1145 } else if (map->map_type == BPF_MAP_TYPE_XSKMAP) {
1146 reg->type = PTR_TO_XDP_SOCK;
1147 } else if (map->map_type == BPF_MAP_TYPE_SOCKMAP ||
1148 map->map_type == BPF_MAP_TYPE_SOCKHASH) {
1149 reg->type = PTR_TO_SOCKET;
1151 reg->type = PTR_TO_MAP_VALUE;
1156 reg->type &= ~PTR_MAYBE_NULL;
1159 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
1161 return type_is_pkt_pointer(reg->type);
1164 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
1166 return reg_is_pkt_pointer(reg) ||
1167 reg->type == PTR_TO_PACKET_END;
1170 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
1171 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
1172 enum bpf_reg_type which)
1174 /* The register can already have a range from prior markings.
1175 * This is fine as long as it hasn't been advanced from its
1178 return reg->type == which &&
1181 tnum_equals_const(reg->var_off, 0);
1184 /* Reset the min/max bounds of a register */
1185 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
1187 reg->smin_value = S64_MIN;
1188 reg->smax_value = S64_MAX;
1189 reg->umin_value = 0;
1190 reg->umax_value = U64_MAX;
1192 reg->s32_min_value = S32_MIN;
1193 reg->s32_max_value = S32_MAX;
1194 reg->u32_min_value = 0;
1195 reg->u32_max_value = U32_MAX;
1198 static void __mark_reg64_unbounded(struct bpf_reg_state *reg)
1200 reg->smin_value = S64_MIN;
1201 reg->smax_value = S64_MAX;
1202 reg->umin_value = 0;
1203 reg->umax_value = U64_MAX;
1206 static void __mark_reg32_unbounded(struct bpf_reg_state *reg)
1208 reg->s32_min_value = S32_MIN;
1209 reg->s32_max_value = S32_MAX;
1210 reg->u32_min_value = 0;
1211 reg->u32_max_value = U32_MAX;
1214 static void __update_reg32_bounds(struct bpf_reg_state *reg)
1216 struct tnum var32_off = tnum_subreg(reg->var_off);
1218 /* min signed is max(sign bit) | min(other bits) */
1219 reg->s32_min_value = max_t(s32, reg->s32_min_value,
1220 var32_off.value | (var32_off.mask & S32_MIN));
1221 /* max signed is min(sign bit) | max(other bits) */
1222 reg->s32_max_value = min_t(s32, reg->s32_max_value,
1223 var32_off.value | (var32_off.mask & S32_MAX));
1224 reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)var32_off.value);
1225 reg->u32_max_value = min(reg->u32_max_value,
1226 (u32)(var32_off.value | var32_off.mask));
1229 static void __update_reg64_bounds(struct bpf_reg_state *reg)
1231 /* min signed is max(sign bit) | min(other bits) */
1232 reg->smin_value = max_t(s64, reg->smin_value,
1233 reg->var_off.value | (reg->var_off.mask & S64_MIN));
1234 /* max signed is min(sign bit) | max(other bits) */
1235 reg->smax_value = min_t(s64, reg->smax_value,
1236 reg->var_off.value | (reg->var_off.mask & S64_MAX));
1237 reg->umin_value = max(reg->umin_value, reg->var_off.value);
1238 reg->umax_value = min(reg->umax_value,
1239 reg->var_off.value | reg->var_off.mask);
1242 static void __update_reg_bounds(struct bpf_reg_state *reg)
1244 __update_reg32_bounds(reg);
1245 __update_reg64_bounds(reg);
1248 /* Uses signed min/max values to inform unsigned, and vice-versa */
1249 static void __reg32_deduce_bounds(struct bpf_reg_state *reg)
1251 /* Learn sign from signed bounds.
1252 * If we cannot cross the sign boundary, then signed and unsigned bounds
1253 * are the same, so combine. This works even in the negative case, e.g.
1254 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1256 if (reg->s32_min_value >= 0 || reg->s32_max_value < 0) {
1257 reg->s32_min_value = reg->u32_min_value =
1258 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1259 reg->s32_max_value = reg->u32_max_value =
1260 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1263 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1264 * boundary, so we must be careful.
1266 if ((s32)reg->u32_max_value >= 0) {
1267 /* Positive. We can't learn anything from the smin, but smax
1268 * is positive, hence safe.
1270 reg->s32_min_value = reg->u32_min_value;
1271 reg->s32_max_value = reg->u32_max_value =
1272 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1273 } else if ((s32)reg->u32_min_value < 0) {
1274 /* Negative. We can't learn anything from the smax, but smin
1275 * is negative, hence safe.
1277 reg->s32_min_value = reg->u32_min_value =
1278 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1279 reg->s32_max_value = reg->u32_max_value;
1283 static void __reg64_deduce_bounds(struct bpf_reg_state *reg)
1285 /* Learn sign from signed bounds.
1286 * If we cannot cross the sign boundary, then signed and unsigned bounds
1287 * are the same, so combine. This works even in the negative case, e.g.
1288 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1290 if (reg->smin_value >= 0 || reg->smax_value < 0) {
1291 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1293 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1297 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1298 * boundary, so we must be careful.
1300 if ((s64)reg->umax_value >= 0) {
1301 /* Positive. We can't learn anything from the smin, but smax
1302 * is positive, hence safe.
1304 reg->smin_value = reg->umin_value;
1305 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1307 } else if ((s64)reg->umin_value < 0) {
1308 /* Negative. We can't learn anything from the smax, but smin
1309 * is negative, hence safe.
1311 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1313 reg->smax_value = reg->umax_value;
1317 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
1319 __reg32_deduce_bounds(reg);
1320 __reg64_deduce_bounds(reg);
1323 /* Attempts to improve var_off based on unsigned min/max information */
1324 static void __reg_bound_offset(struct bpf_reg_state *reg)
1326 struct tnum var64_off = tnum_intersect(reg->var_off,
1327 tnum_range(reg->umin_value,
1329 struct tnum var32_off = tnum_intersect(tnum_subreg(reg->var_off),
1330 tnum_range(reg->u32_min_value,
1331 reg->u32_max_value));
1333 reg->var_off = tnum_or(tnum_clear_subreg(var64_off), var32_off);
1336 static bool __reg32_bound_s64(s32 a)
1338 return a >= 0 && a <= S32_MAX;
1341 static void __reg_assign_32_into_64(struct bpf_reg_state *reg)
1343 reg->umin_value = reg->u32_min_value;
1344 reg->umax_value = reg->u32_max_value;
1346 /* Attempt to pull 32-bit signed bounds into 64-bit bounds but must
1347 * be positive otherwise set to worse case bounds and refine later
1350 if (__reg32_bound_s64(reg->s32_min_value) &&
1351 __reg32_bound_s64(reg->s32_max_value)) {
1352 reg->smin_value = reg->s32_min_value;
1353 reg->smax_value = reg->s32_max_value;
1355 reg->smin_value = 0;
1356 reg->smax_value = U32_MAX;
1360 static void __reg_combine_32_into_64(struct bpf_reg_state *reg)
1362 /* special case when 64-bit register has upper 32-bit register
1363 * zeroed. Typically happens after zext or <<32, >>32 sequence
1364 * allowing us to use 32-bit bounds directly,
1366 if (tnum_equals_const(tnum_clear_subreg(reg->var_off), 0)) {
1367 __reg_assign_32_into_64(reg);
1369 /* Otherwise the best we can do is push lower 32bit known and
1370 * unknown bits into register (var_off set from jmp logic)
1371 * then learn as much as possible from the 64-bit tnum
1372 * known and unknown bits. The previous smin/smax bounds are
1373 * invalid here because of jmp32 compare so mark them unknown
1374 * so they do not impact tnum bounds calculation.
1376 __mark_reg64_unbounded(reg);
1377 __update_reg_bounds(reg);
1380 /* Intersecting with the old var_off might have improved our bounds
1381 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1382 * then new var_off is (0; 0x7f...fc) which improves our umax.
1384 __reg_deduce_bounds(reg);
1385 __reg_bound_offset(reg);
1386 __update_reg_bounds(reg);
1389 static bool __reg64_bound_s32(s64 a)
1391 return a >= S32_MIN && a <= S32_MAX;
1394 static bool __reg64_bound_u32(u64 a)
1396 return a >= U32_MIN && a <= U32_MAX;
1399 static void __reg_combine_64_into_32(struct bpf_reg_state *reg)
1401 __mark_reg32_unbounded(reg);
1403 if (__reg64_bound_s32(reg->smin_value) && __reg64_bound_s32(reg->smax_value)) {
1404 reg->s32_min_value = (s32)reg->smin_value;
1405 reg->s32_max_value = (s32)reg->smax_value;
1407 if (__reg64_bound_u32(reg->umin_value) && __reg64_bound_u32(reg->umax_value)) {
1408 reg->u32_min_value = (u32)reg->umin_value;
1409 reg->u32_max_value = (u32)reg->umax_value;
1412 /* Intersecting with the old var_off might have improved our bounds
1413 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1414 * then new var_off is (0; 0x7f...fc) which improves our umax.
1416 __reg_deduce_bounds(reg);
1417 __reg_bound_offset(reg);
1418 __update_reg_bounds(reg);
1421 /* Mark a register as having a completely unknown (scalar) value. */
1422 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
1423 struct bpf_reg_state *reg)
1426 * Clear type, id, off, and union(map_ptr, range) and
1427 * padding between 'type' and union
1429 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
1430 reg->type = SCALAR_VALUE;
1431 reg->var_off = tnum_unknown;
1433 reg->precise = env->subprog_cnt > 1 || !env->bpf_capable;
1434 __mark_reg_unbounded(reg);
1437 static void mark_reg_unknown(struct bpf_verifier_env *env,
1438 struct bpf_reg_state *regs, u32 regno)
1440 if (WARN_ON(regno >= MAX_BPF_REG)) {
1441 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
1442 /* Something bad happened, let's kill all regs except FP */
1443 for (regno = 0; regno < BPF_REG_FP; regno++)
1444 __mark_reg_not_init(env, regs + regno);
1447 __mark_reg_unknown(env, regs + regno);
1450 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1451 struct bpf_reg_state *reg)
1453 __mark_reg_unknown(env, reg);
1454 reg->type = NOT_INIT;
1457 static void mark_reg_not_init(struct bpf_verifier_env *env,
1458 struct bpf_reg_state *regs, u32 regno)
1460 if (WARN_ON(regno >= MAX_BPF_REG)) {
1461 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
1462 /* Something bad happened, let's kill all regs except FP */
1463 for (regno = 0; regno < BPF_REG_FP; regno++)
1464 __mark_reg_not_init(env, regs + regno);
1467 __mark_reg_not_init(env, regs + regno);
1470 static void mark_btf_ld_reg(struct bpf_verifier_env *env,
1471 struct bpf_reg_state *regs, u32 regno,
1472 enum bpf_reg_type reg_type,
1473 struct btf *btf, u32 btf_id)
1475 if (reg_type == SCALAR_VALUE) {
1476 mark_reg_unknown(env, regs, regno);
1479 mark_reg_known_zero(env, regs, regno);
1480 regs[regno].type = PTR_TO_BTF_ID;
1481 regs[regno].btf = btf;
1482 regs[regno].btf_id = btf_id;
1485 #define DEF_NOT_SUBREG (0)
1486 static void init_reg_state(struct bpf_verifier_env *env,
1487 struct bpf_func_state *state)
1489 struct bpf_reg_state *regs = state->regs;
1492 for (i = 0; i < MAX_BPF_REG; i++) {
1493 mark_reg_not_init(env, regs, i);
1494 regs[i].live = REG_LIVE_NONE;
1495 regs[i].parent = NULL;
1496 regs[i].subreg_def = DEF_NOT_SUBREG;
1500 regs[BPF_REG_FP].type = PTR_TO_STACK;
1501 mark_reg_known_zero(env, regs, BPF_REG_FP);
1502 regs[BPF_REG_FP].frameno = state->frameno;
1505 #define BPF_MAIN_FUNC (-1)
1506 static void init_func_state(struct bpf_verifier_env *env,
1507 struct bpf_func_state *state,
1508 int callsite, int frameno, int subprogno)
1510 state->callsite = callsite;
1511 state->frameno = frameno;
1512 state->subprogno = subprogno;
1513 init_reg_state(env, state);
1516 /* Similar to push_stack(), but for async callbacks */
1517 static struct bpf_verifier_state *push_async_cb(struct bpf_verifier_env *env,
1518 int insn_idx, int prev_insn_idx,
1521 struct bpf_verifier_stack_elem *elem;
1522 struct bpf_func_state *frame;
1524 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
1528 elem->insn_idx = insn_idx;
1529 elem->prev_insn_idx = prev_insn_idx;
1530 elem->next = env->head;
1531 elem->log_pos = env->log.len_used;
1534 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
1536 "The sequence of %d jumps is too complex for async cb.\n",
1540 /* Unlike push_stack() do not copy_verifier_state().
1541 * The caller state doesn't matter.
1542 * This is async callback. It starts in a fresh stack.
1543 * Initialize it similar to do_check_common().
1545 elem->st.branches = 1;
1546 frame = kzalloc(sizeof(*frame), GFP_KERNEL);
1549 init_func_state(env, frame,
1550 BPF_MAIN_FUNC /* callsite */,
1551 0 /* frameno within this callchain */,
1552 subprog /* subprog number within this prog */);
1553 elem->st.frame[0] = frame;
1556 free_verifier_state(env->cur_state, true);
1557 env->cur_state = NULL;
1558 /* pop all elements and return */
1559 while (!pop_stack(env, NULL, NULL, false));
1565 SRC_OP, /* register is used as source operand */
1566 DST_OP, /* register is used as destination operand */
1567 DST_OP_NO_MARK /* same as above, check only, don't mark */
1570 static int cmp_subprogs(const void *a, const void *b)
1572 return ((struct bpf_subprog_info *)a)->start -
1573 ((struct bpf_subprog_info *)b)->start;
1576 static int find_subprog(struct bpf_verifier_env *env, int off)
1578 struct bpf_subprog_info *p;
1580 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
1581 sizeof(env->subprog_info[0]), cmp_subprogs);
1584 return p - env->subprog_info;
1588 static int add_subprog(struct bpf_verifier_env *env, int off)
1590 int insn_cnt = env->prog->len;
1593 if (off >= insn_cnt || off < 0) {
1594 verbose(env, "call to invalid destination\n");
1597 ret = find_subprog(env, off);
1600 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
1601 verbose(env, "too many subprograms\n");
1604 /* determine subprog starts. The end is one before the next starts */
1605 env->subprog_info[env->subprog_cnt++].start = off;
1606 sort(env->subprog_info, env->subprog_cnt,
1607 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
1608 return env->subprog_cnt - 1;
1611 struct bpf_kfunc_desc {
1612 struct btf_func_model func_model;
1617 #define MAX_KFUNC_DESCS 256
1618 struct bpf_kfunc_desc_tab {
1619 struct bpf_kfunc_desc descs[MAX_KFUNC_DESCS];
1623 static int kfunc_desc_cmp_by_id(const void *a, const void *b)
1625 const struct bpf_kfunc_desc *d0 = a;
1626 const struct bpf_kfunc_desc *d1 = b;
1628 /* func_id is not greater than BTF_MAX_TYPE */
1629 return d0->func_id - d1->func_id;
1632 static const struct bpf_kfunc_desc *
1633 find_kfunc_desc(const struct bpf_prog *prog, u32 func_id)
1635 struct bpf_kfunc_desc desc = {
1638 struct bpf_kfunc_desc_tab *tab;
1640 tab = prog->aux->kfunc_tab;
1641 return bsearch(&desc, tab->descs, tab->nr_descs,
1642 sizeof(tab->descs[0]), kfunc_desc_cmp_by_id);
1645 static int add_kfunc_call(struct bpf_verifier_env *env, u32 func_id)
1647 const struct btf_type *func, *func_proto;
1648 struct bpf_kfunc_desc_tab *tab;
1649 struct bpf_prog_aux *prog_aux;
1650 struct bpf_kfunc_desc *desc;
1651 const char *func_name;
1655 prog_aux = env->prog->aux;
1656 tab = prog_aux->kfunc_tab;
1659 verbose(env, "calling kernel function is not supported without CONFIG_DEBUG_INFO_BTF\n");
1663 if (!env->prog->jit_requested) {
1664 verbose(env, "JIT is required for calling kernel function\n");
1668 if (!bpf_jit_supports_kfunc_call()) {
1669 verbose(env, "JIT does not support calling kernel function\n");
1673 if (!env->prog->gpl_compatible) {
1674 verbose(env, "cannot call kernel function from non-GPL compatible program\n");
1678 tab = kzalloc(sizeof(*tab), GFP_KERNEL);
1681 prog_aux->kfunc_tab = tab;
1684 if (find_kfunc_desc(env->prog, func_id))
1687 if (tab->nr_descs == MAX_KFUNC_DESCS) {
1688 verbose(env, "too many different kernel function calls\n");
1692 func = btf_type_by_id(btf_vmlinux, func_id);
1693 if (!func || !btf_type_is_func(func)) {
1694 verbose(env, "kernel btf_id %u is not a function\n",
1698 func_proto = btf_type_by_id(btf_vmlinux, func->type);
1699 if (!func_proto || !btf_type_is_func_proto(func_proto)) {
1700 verbose(env, "kernel function btf_id %u does not have a valid func_proto\n",
1705 func_name = btf_name_by_offset(btf_vmlinux, func->name_off);
1706 addr = kallsyms_lookup_name(func_name);
1708 verbose(env, "cannot find address for kernel function %s\n",
1713 desc = &tab->descs[tab->nr_descs++];
1714 desc->func_id = func_id;
1715 desc->imm = BPF_CAST_CALL(addr) - __bpf_call_base;
1716 err = btf_distill_func_proto(&env->log, btf_vmlinux,
1717 func_proto, func_name,
1720 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
1721 kfunc_desc_cmp_by_id, NULL);
1725 static int kfunc_desc_cmp_by_imm(const void *a, const void *b)
1727 const struct bpf_kfunc_desc *d0 = a;
1728 const struct bpf_kfunc_desc *d1 = b;
1730 if (d0->imm > d1->imm)
1732 else if (d0->imm < d1->imm)
1737 static void sort_kfunc_descs_by_imm(struct bpf_prog *prog)
1739 struct bpf_kfunc_desc_tab *tab;
1741 tab = prog->aux->kfunc_tab;
1745 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
1746 kfunc_desc_cmp_by_imm, NULL);
1749 bool bpf_prog_has_kfunc_call(const struct bpf_prog *prog)
1751 return !!prog->aux->kfunc_tab;
1754 const struct btf_func_model *
1755 bpf_jit_find_kfunc_model(const struct bpf_prog *prog,
1756 const struct bpf_insn *insn)
1758 const struct bpf_kfunc_desc desc = {
1761 const struct bpf_kfunc_desc *res;
1762 struct bpf_kfunc_desc_tab *tab;
1764 tab = prog->aux->kfunc_tab;
1765 res = bsearch(&desc, tab->descs, tab->nr_descs,
1766 sizeof(tab->descs[0]), kfunc_desc_cmp_by_imm);
1768 return res ? &res->func_model : NULL;
1771 static int add_subprog_and_kfunc(struct bpf_verifier_env *env)
1773 struct bpf_subprog_info *subprog = env->subprog_info;
1774 struct bpf_insn *insn = env->prog->insnsi;
1775 int i, ret, insn_cnt = env->prog->len;
1777 /* Add entry function. */
1778 ret = add_subprog(env, 0);
1782 for (i = 0; i < insn_cnt; i++, insn++) {
1783 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn) &&
1784 !bpf_pseudo_kfunc_call(insn))
1787 if (!env->bpf_capable) {
1788 verbose(env, "loading/calling other bpf or kernel functions are allowed for CAP_BPF and CAP_SYS_ADMIN\n");
1792 if (bpf_pseudo_func(insn)) {
1793 ret = add_subprog(env, i + insn->imm + 1);
1795 /* remember subprog */
1797 } else if (bpf_pseudo_call(insn)) {
1798 ret = add_subprog(env, i + insn->imm + 1);
1800 ret = add_kfunc_call(env, insn->imm);
1807 /* Add a fake 'exit' subprog which could simplify subprog iteration
1808 * logic. 'subprog_cnt' should not be increased.
1810 subprog[env->subprog_cnt].start = insn_cnt;
1812 if (env->log.level & BPF_LOG_LEVEL2)
1813 for (i = 0; i < env->subprog_cnt; i++)
1814 verbose(env, "func#%d @%d\n", i, subprog[i].start);
1819 static int check_subprogs(struct bpf_verifier_env *env)
1821 int i, subprog_start, subprog_end, off, cur_subprog = 0;
1822 struct bpf_subprog_info *subprog = env->subprog_info;
1823 struct bpf_insn *insn = env->prog->insnsi;
1824 int insn_cnt = env->prog->len;
1826 /* now check that all jumps are within the same subprog */
1827 subprog_start = subprog[cur_subprog].start;
1828 subprog_end = subprog[cur_subprog + 1].start;
1829 for (i = 0; i < insn_cnt; i++) {
1830 u8 code = insn[i].code;
1832 if (code == (BPF_JMP | BPF_CALL) &&
1833 insn[i].imm == BPF_FUNC_tail_call &&
1834 insn[i].src_reg != BPF_PSEUDO_CALL)
1835 subprog[cur_subprog].has_tail_call = true;
1836 if (BPF_CLASS(code) == BPF_LD &&
1837 (BPF_MODE(code) == BPF_ABS || BPF_MODE(code) == BPF_IND))
1838 subprog[cur_subprog].has_ld_abs = true;
1839 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
1841 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
1843 off = i + insn[i].off + 1;
1844 if (off < subprog_start || off >= subprog_end) {
1845 verbose(env, "jump out of range from insn %d to %d\n", i, off);
1849 if (i == subprog_end - 1) {
1850 /* to avoid fall-through from one subprog into another
1851 * the last insn of the subprog should be either exit
1852 * or unconditional jump back
1854 if (code != (BPF_JMP | BPF_EXIT) &&
1855 code != (BPF_JMP | BPF_JA)) {
1856 verbose(env, "last insn is not an exit or jmp\n");
1859 subprog_start = subprog_end;
1861 if (cur_subprog < env->subprog_cnt)
1862 subprog_end = subprog[cur_subprog + 1].start;
1868 /* Parentage chain of this register (or stack slot) should take care of all
1869 * issues like callee-saved registers, stack slot allocation time, etc.
1871 static int mark_reg_read(struct bpf_verifier_env *env,
1872 const struct bpf_reg_state *state,
1873 struct bpf_reg_state *parent, u8 flag)
1875 bool writes = parent == state->parent; /* Observe write marks */
1879 /* if read wasn't screened by an earlier write ... */
1880 if (writes && state->live & REG_LIVE_WRITTEN)
1882 if (parent->live & REG_LIVE_DONE) {
1883 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
1884 reg_type_str(env, parent->type),
1885 parent->var_off.value, parent->off);
1888 /* The first condition is more likely to be true than the
1889 * second, checked it first.
1891 if ((parent->live & REG_LIVE_READ) == flag ||
1892 parent->live & REG_LIVE_READ64)
1893 /* The parentage chain never changes and
1894 * this parent was already marked as LIVE_READ.
1895 * There is no need to keep walking the chain again and
1896 * keep re-marking all parents as LIVE_READ.
1897 * This case happens when the same register is read
1898 * multiple times without writes into it in-between.
1899 * Also, if parent has the stronger REG_LIVE_READ64 set,
1900 * then no need to set the weak REG_LIVE_READ32.
1903 /* ... then we depend on parent's value */
1904 parent->live |= flag;
1905 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
1906 if (flag == REG_LIVE_READ64)
1907 parent->live &= ~REG_LIVE_READ32;
1909 parent = state->parent;
1914 if (env->longest_mark_read_walk < cnt)
1915 env->longest_mark_read_walk = cnt;
1919 /* This function is supposed to be used by the following 32-bit optimization
1920 * code only. It returns TRUE if the source or destination register operates
1921 * on 64-bit, otherwise return FALSE.
1923 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
1924 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
1929 class = BPF_CLASS(code);
1931 if (class == BPF_JMP) {
1932 /* BPF_EXIT for "main" will reach here. Return TRUE
1937 if (op == BPF_CALL) {
1938 /* BPF to BPF call will reach here because of marking
1939 * caller saved clobber with DST_OP_NO_MARK for which we
1940 * don't care the register def because they are anyway
1941 * marked as NOT_INIT already.
1943 if (insn->src_reg == BPF_PSEUDO_CALL)
1945 /* Helper call will reach here because of arg type
1946 * check, conservatively return TRUE.
1955 if (class == BPF_ALU64 || class == BPF_JMP ||
1956 /* BPF_END always use BPF_ALU class. */
1957 (class == BPF_ALU && op == BPF_END && insn->imm == 64))
1960 if (class == BPF_ALU || class == BPF_JMP32)
1963 if (class == BPF_LDX) {
1965 return BPF_SIZE(code) == BPF_DW;
1966 /* LDX source must be ptr. */
1970 if (class == BPF_STX) {
1971 /* BPF_STX (including atomic variants) has multiple source
1972 * operands, one of which is a ptr. Check whether the caller is
1975 if (t == SRC_OP && reg->type != SCALAR_VALUE)
1977 return BPF_SIZE(code) == BPF_DW;
1980 if (class == BPF_LD) {
1981 u8 mode = BPF_MODE(code);
1984 if (mode == BPF_IMM)
1987 /* Both LD_IND and LD_ABS return 32-bit data. */
1991 /* Implicit ctx ptr. */
1992 if (regno == BPF_REG_6)
1995 /* Explicit source could be any width. */
1999 if (class == BPF_ST)
2000 /* The only source register for BPF_ST is a ptr. */
2003 /* Conservatively return true at default. */
2007 /* Return the regno defined by the insn, or -1. */
2008 static int insn_def_regno(const struct bpf_insn *insn)
2010 switch (BPF_CLASS(insn->code)) {
2016 if (BPF_MODE(insn->code) == BPF_ATOMIC &&
2017 (insn->imm & BPF_FETCH)) {
2018 if (insn->imm == BPF_CMPXCHG)
2021 return insn->src_reg;
2026 return insn->dst_reg;
2030 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
2031 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
2033 int dst_reg = insn_def_regno(insn);
2038 return !is_reg64(env, insn, dst_reg, NULL, DST_OP);
2041 static void mark_insn_zext(struct bpf_verifier_env *env,
2042 struct bpf_reg_state *reg)
2044 s32 def_idx = reg->subreg_def;
2046 if (def_idx == DEF_NOT_SUBREG)
2049 env->insn_aux_data[def_idx - 1].zext_dst = true;
2050 /* The dst will be zero extended, so won't be sub-register anymore. */
2051 reg->subreg_def = DEF_NOT_SUBREG;
2054 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
2055 enum reg_arg_type t)
2057 struct bpf_verifier_state *vstate = env->cur_state;
2058 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2059 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
2060 struct bpf_reg_state *reg, *regs = state->regs;
2063 if (regno >= MAX_BPF_REG) {
2064 verbose(env, "R%d is invalid\n", regno);
2069 rw64 = is_reg64(env, insn, regno, reg, t);
2071 /* check whether register used as source operand can be read */
2072 if (reg->type == NOT_INIT) {
2073 verbose(env, "R%d !read_ok\n", regno);
2076 /* We don't need to worry about FP liveness because it's read-only */
2077 if (regno == BPF_REG_FP)
2081 mark_insn_zext(env, reg);
2083 return mark_reg_read(env, reg, reg->parent,
2084 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
2086 /* check whether register used as dest operand can be written to */
2087 if (regno == BPF_REG_FP) {
2088 verbose(env, "frame pointer is read only\n");
2091 reg->live |= REG_LIVE_WRITTEN;
2092 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
2094 mark_reg_unknown(env, regs, regno);
2099 /* for any branch, call, exit record the history of jmps in the given state */
2100 static int push_jmp_history(struct bpf_verifier_env *env,
2101 struct bpf_verifier_state *cur)
2103 u32 cnt = cur->jmp_history_cnt;
2104 struct bpf_idx_pair *p;
2107 p = krealloc(cur->jmp_history, cnt * sizeof(*p), GFP_USER);
2110 p[cnt - 1].idx = env->insn_idx;
2111 p[cnt - 1].prev_idx = env->prev_insn_idx;
2112 cur->jmp_history = p;
2113 cur->jmp_history_cnt = cnt;
2117 /* Backtrack one insn at a time. If idx is not at the top of recorded
2118 * history then previous instruction came from straight line execution.
2120 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
2125 if (cnt && st->jmp_history[cnt - 1].idx == i) {
2126 i = st->jmp_history[cnt - 1].prev_idx;
2134 static const char *disasm_kfunc_name(void *data, const struct bpf_insn *insn)
2136 const struct btf_type *func;
2138 if (insn->src_reg != BPF_PSEUDO_KFUNC_CALL)
2141 func = btf_type_by_id(btf_vmlinux, insn->imm);
2142 return btf_name_by_offset(btf_vmlinux, func->name_off);
2145 /* For given verifier state backtrack_insn() is called from the last insn to
2146 * the first insn. Its purpose is to compute a bitmask of registers and
2147 * stack slots that needs precision in the parent verifier state.
2149 static int backtrack_insn(struct bpf_verifier_env *env, int idx,
2150 u32 *reg_mask, u64 *stack_mask)
2152 const struct bpf_insn_cbs cbs = {
2153 .cb_call = disasm_kfunc_name,
2154 .cb_print = verbose,
2155 .private_data = env,
2157 struct bpf_insn *insn = env->prog->insnsi + idx;
2158 u8 class = BPF_CLASS(insn->code);
2159 u8 opcode = BPF_OP(insn->code);
2160 u8 mode = BPF_MODE(insn->code);
2161 u32 dreg = 1u << insn->dst_reg;
2162 u32 sreg = 1u << insn->src_reg;
2165 if (insn->code == 0)
2167 if (env->log.level & BPF_LOG_LEVEL) {
2168 verbose(env, "regs=%x stack=%llx before ", *reg_mask, *stack_mask);
2169 verbose(env, "%d: ", idx);
2170 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
2173 if (class == BPF_ALU || class == BPF_ALU64) {
2174 if (!(*reg_mask & dreg))
2176 if (opcode == BPF_MOV) {
2177 if (BPF_SRC(insn->code) == BPF_X) {
2179 * dreg needs precision after this insn
2180 * sreg needs precision before this insn
2186 * dreg needs precision after this insn.
2187 * Corresponding register is already marked
2188 * as precise=true in this verifier state.
2189 * No further markings in parent are necessary
2194 if (BPF_SRC(insn->code) == BPF_X) {
2196 * both dreg and sreg need precision
2201 * dreg still needs precision before this insn
2204 } else if (class == BPF_LDX) {
2205 if (!(*reg_mask & dreg))
2209 /* scalars can only be spilled into stack w/o losing precision.
2210 * Load from any other memory can be zero extended.
2211 * The desire to keep that precision is already indicated
2212 * by 'precise' mark in corresponding register of this state.
2213 * No further tracking necessary.
2215 if (insn->src_reg != BPF_REG_FP)
2217 if (BPF_SIZE(insn->code) != BPF_DW)
2220 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
2221 * that [fp - off] slot contains scalar that needs to be
2222 * tracked with precision
2224 spi = (-insn->off - 1) / BPF_REG_SIZE;
2226 verbose(env, "BUG spi %d\n", spi);
2227 WARN_ONCE(1, "verifier backtracking bug");
2230 *stack_mask |= 1ull << spi;
2231 } else if (class == BPF_STX || class == BPF_ST) {
2232 if (*reg_mask & dreg)
2233 /* stx & st shouldn't be using _scalar_ dst_reg
2234 * to access memory. It means backtracking
2235 * encountered a case of pointer subtraction.
2238 /* scalars can only be spilled into stack */
2239 if (insn->dst_reg != BPF_REG_FP)
2241 if (BPF_SIZE(insn->code) != BPF_DW)
2243 spi = (-insn->off - 1) / BPF_REG_SIZE;
2245 verbose(env, "BUG spi %d\n", spi);
2246 WARN_ONCE(1, "verifier backtracking bug");
2249 if (!(*stack_mask & (1ull << spi)))
2251 *stack_mask &= ~(1ull << spi);
2252 if (class == BPF_STX)
2254 } else if (class == BPF_JMP || class == BPF_JMP32) {
2255 if (opcode == BPF_CALL) {
2256 if (insn->src_reg == BPF_PSEUDO_CALL)
2258 /* regular helper call sets R0 */
2260 if (*reg_mask & 0x3f) {
2261 /* if backtracing was looking for registers R1-R5
2262 * they should have been found already.
2264 verbose(env, "BUG regs %x\n", *reg_mask);
2265 WARN_ONCE(1, "verifier backtracking bug");
2268 } else if (opcode == BPF_EXIT) {
2271 } else if (class == BPF_LD) {
2272 if (!(*reg_mask & dreg))
2275 /* It's ld_imm64 or ld_abs or ld_ind.
2276 * For ld_imm64 no further tracking of precision
2277 * into parent is necessary
2279 if (mode == BPF_IND || mode == BPF_ABS)
2280 /* to be analyzed */
2286 /* the scalar precision tracking algorithm:
2287 * . at the start all registers have precise=false.
2288 * . scalar ranges are tracked as normal through alu and jmp insns.
2289 * . once precise value of the scalar register is used in:
2290 * . ptr + scalar alu
2291 * . if (scalar cond K|scalar)
2292 * . helper_call(.., scalar, ...) where ARG_CONST is expected
2293 * backtrack through the verifier states and mark all registers and
2294 * stack slots with spilled constants that these scalar regisers
2295 * should be precise.
2296 * . during state pruning two registers (or spilled stack slots)
2297 * are equivalent if both are not precise.
2299 * Note the verifier cannot simply walk register parentage chain,
2300 * since many different registers and stack slots could have been
2301 * used to compute single precise scalar.
2303 * The approach of starting with precise=true for all registers and then
2304 * backtrack to mark a register as not precise when the verifier detects
2305 * that program doesn't care about specific value (e.g., when helper
2306 * takes register as ARG_ANYTHING parameter) is not safe.
2308 * It's ok to walk single parentage chain of the verifier states.
2309 * It's possible that this backtracking will go all the way till 1st insn.
2310 * All other branches will be explored for needing precision later.
2312 * The backtracking needs to deal with cases like:
2313 * 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)
2316 * if r5 > 0x79f goto pc+7
2317 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
2320 * call bpf_perf_event_output#25
2321 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
2325 * call foo // uses callee's r6 inside to compute r0
2329 * to track above reg_mask/stack_mask needs to be independent for each frame.
2331 * Also if parent's curframe > frame where backtracking started,
2332 * the verifier need to mark registers in both frames, otherwise callees
2333 * may incorrectly prune callers. This is similar to
2334 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
2336 * For now backtracking falls back into conservative marking.
2338 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
2339 struct bpf_verifier_state *st)
2341 struct bpf_func_state *func;
2342 struct bpf_reg_state *reg;
2345 /* big hammer: mark all scalars precise in this path.
2346 * pop_stack may still get !precise scalars.
2348 for (; st; st = st->parent)
2349 for (i = 0; i <= st->curframe; i++) {
2350 func = st->frame[i];
2351 for (j = 0; j < BPF_REG_FP; j++) {
2352 reg = &func->regs[j];
2353 if (reg->type != SCALAR_VALUE)
2355 reg->precise = true;
2357 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
2358 if (func->stack[j].slot_type[0] != STACK_SPILL)
2360 reg = &func->stack[j].spilled_ptr;
2361 if (reg->type != SCALAR_VALUE)
2363 reg->precise = true;
2368 static int __mark_chain_precision(struct bpf_verifier_env *env, int regno,
2371 struct bpf_verifier_state *st = env->cur_state;
2372 int first_idx = st->first_insn_idx;
2373 int last_idx = env->insn_idx;
2374 struct bpf_func_state *func;
2375 struct bpf_reg_state *reg;
2376 u32 reg_mask = regno >= 0 ? 1u << regno : 0;
2377 u64 stack_mask = spi >= 0 ? 1ull << spi : 0;
2378 bool skip_first = true;
2379 bool new_marks = false;
2382 if (!env->bpf_capable)
2385 func = st->frame[st->curframe];
2387 reg = &func->regs[regno];
2388 if (reg->type != SCALAR_VALUE) {
2389 WARN_ONCE(1, "backtracing misuse");
2396 reg->precise = true;
2400 if (func->stack[spi].slot_type[0] != STACK_SPILL) {
2404 reg = &func->stack[spi].spilled_ptr;
2405 if (reg->type != SCALAR_VALUE) {
2413 reg->precise = true;
2419 if (!reg_mask && !stack_mask)
2422 DECLARE_BITMAP(mask, 64);
2423 u32 history = st->jmp_history_cnt;
2425 if (env->log.level & BPF_LOG_LEVEL)
2426 verbose(env, "last_idx %d first_idx %d\n", last_idx, first_idx);
2427 for (i = last_idx;;) {
2432 err = backtrack_insn(env, i, ®_mask, &stack_mask);
2434 if (err == -ENOTSUPP) {
2435 mark_all_scalars_precise(env, st);
2440 if (!reg_mask && !stack_mask)
2441 /* Found assignment(s) into tracked register in this state.
2442 * Since this state is already marked, just return.
2443 * Nothing to be tracked further in the parent state.
2448 i = get_prev_insn_idx(st, i, &history);
2449 if (i >= env->prog->len) {
2450 /* This can happen if backtracking reached insn 0
2451 * and there are still reg_mask or stack_mask
2453 * It means the backtracking missed the spot where
2454 * particular register was initialized with a constant.
2456 verbose(env, "BUG backtracking idx %d\n", i);
2457 WARN_ONCE(1, "verifier backtracking bug");
2466 func = st->frame[st->curframe];
2467 bitmap_from_u64(mask, reg_mask);
2468 for_each_set_bit(i, mask, 32) {
2469 reg = &func->regs[i];
2470 if (reg->type != SCALAR_VALUE) {
2471 reg_mask &= ~(1u << i);
2476 reg->precise = true;
2479 bitmap_from_u64(mask, stack_mask);
2480 for_each_set_bit(i, mask, 64) {
2481 if (i >= func->allocated_stack / BPF_REG_SIZE) {
2482 /* the sequence of instructions:
2484 * 3: (7b) *(u64 *)(r3 -8) = r0
2485 * 4: (79) r4 = *(u64 *)(r10 -8)
2486 * doesn't contain jmps. It's backtracked
2487 * as a single block.
2488 * During backtracking insn 3 is not recognized as
2489 * stack access, so at the end of backtracking
2490 * stack slot fp-8 is still marked in stack_mask.
2491 * However the parent state may not have accessed
2492 * fp-8 and it's "unallocated" stack space.
2493 * In such case fallback to conservative.
2495 mark_all_scalars_precise(env, st);
2499 if (func->stack[i].slot_type[0] != STACK_SPILL) {
2500 stack_mask &= ~(1ull << i);
2503 reg = &func->stack[i].spilled_ptr;
2504 if (reg->type != SCALAR_VALUE) {
2505 stack_mask &= ~(1ull << i);
2510 reg->precise = true;
2512 if (env->log.level & BPF_LOG_LEVEL) {
2513 print_verifier_state(env, func);
2514 verbose(env, "parent %s regs=%x stack=%llx marks\n",
2515 new_marks ? "didn't have" : "already had",
2516 reg_mask, stack_mask);
2519 if (!reg_mask && !stack_mask)
2524 last_idx = st->last_insn_idx;
2525 first_idx = st->first_insn_idx;
2530 static int mark_chain_precision(struct bpf_verifier_env *env, int regno)
2532 return __mark_chain_precision(env, regno, -1);
2535 static int mark_chain_precision_stack(struct bpf_verifier_env *env, int spi)
2537 return __mark_chain_precision(env, -1, spi);
2540 static bool is_spillable_regtype(enum bpf_reg_type type)
2542 switch (base_type(type)) {
2543 case PTR_TO_MAP_VALUE:
2547 case PTR_TO_PACKET_META:
2548 case PTR_TO_PACKET_END:
2549 case PTR_TO_FLOW_KEYS:
2550 case CONST_PTR_TO_MAP:
2552 case PTR_TO_SOCK_COMMON:
2553 case PTR_TO_TCP_SOCK:
2554 case PTR_TO_XDP_SOCK:
2557 case PTR_TO_PERCPU_BTF_ID:
2560 case PTR_TO_MAP_KEY:
2567 /* Does this register contain a constant zero? */
2568 static bool register_is_null(struct bpf_reg_state *reg)
2570 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
2573 static bool register_is_const(struct bpf_reg_state *reg)
2575 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
2578 static bool __is_scalar_unbounded(struct bpf_reg_state *reg)
2580 return tnum_is_unknown(reg->var_off) &&
2581 reg->smin_value == S64_MIN && reg->smax_value == S64_MAX &&
2582 reg->umin_value == 0 && reg->umax_value == U64_MAX &&
2583 reg->s32_min_value == S32_MIN && reg->s32_max_value == S32_MAX &&
2584 reg->u32_min_value == 0 && reg->u32_max_value == U32_MAX;
2587 static bool register_is_bounded(struct bpf_reg_state *reg)
2589 return reg->type == SCALAR_VALUE && !__is_scalar_unbounded(reg);
2592 static bool __is_pointer_value(bool allow_ptr_leaks,
2593 const struct bpf_reg_state *reg)
2595 if (allow_ptr_leaks)
2598 return reg->type != SCALAR_VALUE;
2601 static void save_register_state(struct bpf_func_state *state,
2602 int spi, struct bpf_reg_state *reg)
2606 state->stack[spi].spilled_ptr = *reg;
2607 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2609 for (i = 0; i < BPF_REG_SIZE; i++)
2610 state->stack[spi].slot_type[i] = STACK_SPILL;
2613 /* check_stack_{read,write}_fixed_off functions track spill/fill of registers,
2614 * stack boundary and alignment are checked in check_mem_access()
2616 static int check_stack_write_fixed_off(struct bpf_verifier_env *env,
2617 /* stack frame we're writing to */
2618 struct bpf_func_state *state,
2619 int off, int size, int value_regno,
2622 struct bpf_func_state *cur; /* state of the current function */
2623 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
2624 u32 dst_reg = env->prog->insnsi[insn_idx].dst_reg;
2625 struct bpf_reg_state *reg = NULL;
2627 err = grow_stack_state(state, round_up(slot + 1, BPF_REG_SIZE));
2630 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
2631 * so it's aligned access and [off, off + size) are within stack limits
2633 if (!env->allow_ptr_leaks &&
2634 state->stack[spi].slot_type[0] == STACK_SPILL &&
2635 size != BPF_REG_SIZE) {
2636 verbose(env, "attempt to corrupt spilled pointer on stack\n");
2640 cur = env->cur_state->frame[env->cur_state->curframe];
2641 if (value_regno >= 0)
2642 reg = &cur->regs[value_regno];
2643 if (!env->bypass_spec_v4) {
2644 bool sanitize = reg && is_spillable_regtype(reg->type);
2646 for (i = 0; i < size; i++) {
2647 if (state->stack[spi].slot_type[i] == STACK_INVALID) {
2654 env->insn_aux_data[insn_idx].sanitize_stack_spill = true;
2657 if (reg && size == BPF_REG_SIZE && register_is_bounded(reg) &&
2658 !register_is_null(reg) && env->bpf_capable) {
2659 if (dst_reg != BPF_REG_FP) {
2660 /* The backtracking logic can only recognize explicit
2661 * stack slot address like [fp - 8]. Other spill of
2662 * scalar via different register has to be conservative.
2663 * Backtrack from here and mark all registers as precise
2664 * that contributed into 'reg' being a constant.
2666 err = mark_chain_precision(env, value_regno);
2670 save_register_state(state, spi, reg);
2671 } else if (reg && is_spillable_regtype(reg->type)) {
2672 /* register containing pointer is being spilled into stack */
2673 if (size != BPF_REG_SIZE) {
2674 verbose_linfo(env, insn_idx, "; ");
2675 verbose(env, "invalid size of register spill\n");
2678 if (state != cur && reg->type == PTR_TO_STACK) {
2679 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
2682 save_register_state(state, spi, reg);
2684 u8 type = STACK_MISC;
2686 /* regular write of data into stack destroys any spilled ptr */
2687 state->stack[spi].spilled_ptr.type = NOT_INIT;
2688 /* Mark slots as STACK_MISC if they belonged to spilled ptr. */
2689 if (state->stack[spi].slot_type[0] == STACK_SPILL)
2690 for (i = 0; i < BPF_REG_SIZE; i++)
2691 state->stack[spi].slot_type[i] = STACK_MISC;
2693 /* only mark the slot as written if all 8 bytes were written
2694 * otherwise read propagation may incorrectly stop too soon
2695 * when stack slots are partially written.
2696 * This heuristic means that read propagation will be
2697 * conservative, since it will add reg_live_read marks
2698 * to stack slots all the way to first state when programs
2699 * writes+reads less than 8 bytes
2701 if (size == BPF_REG_SIZE)
2702 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2704 /* when we zero initialize stack slots mark them as such */
2705 if (reg && register_is_null(reg)) {
2706 /* backtracking doesn't work for STACK_ZERO yet. */
2707 err = mark_chain_precision(env, value_regno);
2713 /* Mark slots affected by this stack write. */
2714 for (i = 0; i < size; i++)
2715 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
2721 /* Write the stack: 'stack[ptr_regno + off] = value_regno'. 'ptr_regno' is
2722 * known to contain a variable offset.
2723 * This function checks whether the write is permitted and conservatively
2724 * tracks the effects of the write, considering that each stack slot in the
2725 * dynamic range is potentially written to.
2727 * 'off' includes 'regno->off'.
2728 * 'value_regno' can be -1, meaning that an unknown value is being written to
2731 * Spilled pointers in range are not marked as written because we don't know
2732 * what's going to be actually written. This means that read propagation for
2733 * future reads cannot be terminated by this write.
2735 * For privileged programs, uninitialized stack slots are considered
2736 * initialized by this write (even though we don't know exactly what offsets
2737 * are going to be written to). The idea is that we don't want the verifier to
2738 * reject future reads that access slots written to through variable offsets.
2740 static int check_stack_write_var_off(struct bpf_verifier_env *env,
2741 /* func where register points to */
2742 struct bpf_func_state *state,
2743 int ptr_regno, int off, int size,
2744 int value_regno, int insn_idx)
2746 struct bpf_func_state *cur; /* state of the current function */
2747 int min_off, max_off;
2749 struct bpf_reg_state *ptr_reg = NULL, *value_reg = NULL;
2750 bool writing_zero = false;
2751 /* set if the fact that we're writing a zero is used to let any
2752 * stack slots remain STACK_ZERO
2754 bool zero_used = false;
2756 cur = env->cur_state->frame[env->cur_state->curframe];
2757 ptr_reg = &cur->regs[ptr_regno];
2758 min_off = ptr_reg->smin_value + off;
2759 max_off = ptr_reg->smax_value + off + size;
2760 if (value_regno >= 0)
2761 value_reg = &cur->regs[value_regno];
2762 if (value_reg && register_is_null(value_reg))
2763 writing_zero = true;
2765 err = grow_stack_state(state, round_up(-min_off, BPF_REG_SIZE));
2770 /* Variable offset writes destroy any spilled pointers in range. */
2771 for (i = min_off; i < max_off; i++) {
2772 u8 new_type, *stype;
2776 spi = slot / BPF_REG_SIZE;
2777 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
2779 if (!env->allow_ptr_leaks
2780 && *stype != NOT_INIT
2781 && *stype != SCALAR_VALUE) {
2782 /* Reject the write if there's are spilled pointers in
2783 * range. If we didn't reject here, the ptr status
2784 * would be erased below (even though not all slots are
2785 * actually overwritten), possibly opening the door to
2788 verbose(env, "spilled ptr in range of var-offset stack write; insn %d, ptr off: %d",
2793 /* Erase all spilled pointers. */
2794 state->stack[spi].spilled_ptr.type = NOT_INIT;
2796 /* Update the slot type. */
2797 new_type = STACK_MISC;
2798 if (writing_zero && *stype == STACK_ZERO) {
2799 new_type = STACK_ZERO;
2802 /* If the slot is STACK_INVALID, we check whether it's OK to
2803 * pretend that it will be initialized by this write. The slot
2804 * might not actually be written to, and so if we mark it as
2805 * initialized future reads might leak uninitialized memory.
2806 * For privileged programs, we will accept such reads to slots
2807 * that may or may not be written because, if we're reject
2808 * them, the error would be too confusing.
2810 if (*stype == STACK_INVALID && !env->allow_uninit_stack) {
2811 verbose(env, "uninit stack in range of var-offset write prohibited for !root; insn %d, off: %d",
2818 /* backtracking doesn't work for STACK_ZERO yet. */
2819 err = mark_chain_precision(env, value_regno);
2826 /* When register 'dst_regno' is assigned some values from stack[min_off,
2827 * max_off), we set the register's type according to the types of the
2828 * respective stack slots. If all the stack values are known to be zeros, then
2829 * so is the destination reg. Otherwise, the register is considered to be
2830 * SCALAR. This function does not deal with register filling; the caller must
2831 * ensure that all spilled registers in the stack range have been marked as
2834 static void mark_reg_stack_read(struct bpf_verifier_env *env,
2835 /* func where src register points to */
2836 struct bpf_func_state *ptr_state,
2837 int min_off, int max_off, int dst_regno)
2839 struct bpf_verifier_state *vstate = env->cur_state;
2840 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2845 for (i = min_off; i < max_off; i++) {
2847 spi = slot / BPF_REG_SIZE;
2848 stype = ptr_state->stack[spi].slot_type;
2849 if (stype[slot % BPF_REG_SIZE] != STACK_ZERO)
2853 if (zeros == max_off - min_off) {
2854 /* any access_size read into register is zero extended,
2855 * so the whole register == const_zero
2857 __mark_reg_const_zero(&state->regs[dst_regno]);
2858 /* backtracking doesn't support STACK_ZERO yet,
2859 * so mark it precise here, so that later
2860 * backtracking can stop here.
2861 * Backtracking may not need this if this register
2862 * doesn't participate in pointer adjustment.
2863 * Forward propagation of precise flag is not
2864 * necessary either. This mark is only to stop
2865 * backtracking. Any register that contributed
2866 * to const 0 was marked precise before spill.
2868 state->regs[dst_regno].precise = true;
2870 /* have read misc data from the stack */
2871 mark_reg_unknown(env, state->regs, dst_regno);
2873 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
2876 /* Read the stack at 'off' and put the results into the register indicated by
2877 * 'dst_regno'. It handles reg filling if the addressed stack slot is a
2880 * 'dst_regno' can be -1, meaning that the read value is not going to a
2883 * The access is assumed to be within the current stack bounds.
2885 static int check_stack_read_fixed_off(struct bpf_verifier_env *env,
2886 /* func where src register points to */
2887 struct bpf_func_state *reg_state,
2888 int off, int size, int dst_regno)
2890 struct bpf_verifier_state *vstate = env->cur_state;
2891 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2892 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
2893 struct bpf_reg_state *reg;
2896 stype = reg_state->stack[spi].slot_type;
2897 reg = ®_state->stack[spi].spilled_ptr;
2899 if (stype[0] == STACK_SPILL) {
2900 if (size != BPF_REG_SIZE) {
2901 if (reg->type != SCALAR_VALUE) {
2902 verbose_linfo(env, env->insn_idx, "; ");
2903 verbose(env, "invalid size of register fill\n");
2906 if (dst_regno >= 0) {
2907 mark_reg_unknown(env, state->regs, dst_regno);
2908 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
2910 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2913 for (i = 1; i < BPF_REG_SIZE; i++) {
2914 if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) {
2915 verbose(env, "corrupted spill memory\n");
2920 if (dst_regno >= 0) {
2921 /* restore register state from stack */
2922 state->regs[dst_regno] = *reg;
2923 /* mark reg as written since spilled pointer state likely
2924 * has its liveness marks cleared by is_state_visited()
2925 * which resets stack/reg liveness for state transitions
2927 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
2928 } else if (__is_pointer_value(env->allow_ptr_leaks, reg)) {
2929 /* If dst_regno==-1, the caller is asking us whether
2930 * it is acceptable to use this value as a SCALAR_VALUE
2932 * We must not allow unprivileged callers to do that
2933 * with spilled pointers.
2935 verbose(env, "leaking pointer from stack off %d\n",
2939 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2943 for (i = 0; i < size; i++) {
2944 type = stype[(slot - i) % BPF_REG_SIZE];
2945 if (type == STACK_MISC)
2947 if (type == STACK_ZERO)
2949 verbose(env, "invalid read from stack off %d+%d size %d\n",
2953 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2955 mark_reg_stack_read(env, reg_state, off, off + size, dst_regno);
2960 enum stack_access_src {
2961 ACCESS_DIRECT = 1, /* the access is performed by an instruction */
2962 ACCESS_HELPER = 2, /* the access is performed by a helper */
2965 static int check_stack_range_initialized(struct bpf_verifier_env *env,
2966 int regno, int off, int access_size,
2967 bool zero_size_allowed,
2968 enum stack_access_src type,
2969 struct bpf_call_arg_meta *meta);
2971 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
2973 return cur_regs(env) + regno;
2976 /* Read the stack at 'ptr_regno + off' and put the result into the register
2978 * 'off' includes the pointer register's fixed offset(i.e. 'ptr_regno.off'),
2979 * but not its variable offset.
2980 * 'size' is assumed to be <= reg size and the access is assumed to be aligned.
2982 * As opposed to check_stack_read_fixed_off, this function doesn't deal with
2983 * filling registers (i.e. reads of spilled register cannot be detected when
2984 * the offset is not fixed). We conservatively mark 'dst_regno' as containing
2985 * SCALAR_VALUE. That's why we assert that the 'ptr_regno' has a variable
2986 * offset; for a fixed offset check_stack_read_fixed_off should be used
2989 static int check_stack_read_var_off(struct bpf_verifier_env *env,
2990 int ptr_regno, int off, int size, int dst_regno)
2992 /* The state of the source register. */
2993 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
2994 struct bpf_func_state *ptr_state = func(env, reg);
2996 int min_off, max_off;
2998 /* Note that we pass a NULL meta, so raw access will not be permitted.
3000 err = check_stack_range_initialized(env, ptr_regno, off, size,
3001 false, ACCESS_DIRECT, NULL);
3005 min_off = reg->smin_value + off;
3006 max_off = reg->smax_value + off;
3007 mark_reg_stack_read(env, ptr_state, min_off, max_off + size, dst_regno);
3011 /* check_stack_read dispatches to check_stack_read_fixed_off or
3012 * check_stack_read_var_off.
3014 * The caller must ensure that the offset falls within the allocated stack
3017 * 'dst_regno' is a register which will receive the value from the stack. It
3018 * can be -1, meaning that the read value is not going to a register.
3020 static int check_stack_read(struct bpf_verifier_env *env,
3021 int ptr_regno, int off, int size,
3024 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
3025 struct bpf_func_state *state = func(env, reg);
3027 /* Some accesses are only permitted with a static offset. */
3028 bool var_off = !tnum_is_const(reg->var_off);
3030 /* The offset is required to be static when reads don't go to a
3031 * register, in order to not leak pointers (see
3032 * check_stack_read_fixed_off).
3034 if (dst_regno < 0 && var_off) {
3037 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3038 verbose(env, "variable offset stack pointer cannot be passed into helper function; var_off=%s off=%d size=%d\n",
3042 /* Variable offset is prohibited for unprivileged mode for simplicity
3043 * since it requires corresponding support in Spectre masking for stack
3044 * ALU. See also retrieve_ptr_limit().
3046 if (!env->bypass_spec_v1 && var_off) {
3049 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3050 verbose(env, "R%d variable offset stack access prohibited for !root, var_off=%s\n",
3056 off += reg->var_off.value;
3057 err = check_stack_read_fixed_off(env, state, off, size,
3060 /* Variable offset stack reads need more conservative handling
3061 * than fixed offset ones. Note that dst_regno >= 0 on this
3064 err = check_stack_read_var_off(env, ptr_regno, off, size,
3071 /* check_stack_write dispatches to check_stack_write_fixed_off or
3072 * check_stack_write_var_off.
3074 * 'ptr_regno' is the register used as a pointer into the stack.
3075 * 'off' includes 'ptr_regno->off', but not its variable offset (if any).
3076 * 'value_regno' is the register whose value we're writing to the stack. It can
3077 * be -1, meaning that we're not writing from a register.
3079 * The caller must ensure that the offset falls within the maximum stack size.
3081 static int check_stack_write(struct bpf_verifier_env *env,
3082 int ptr_regno, int off, int size,
3083 int value_regno, int insn_idx)
3085 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
3086 struct bpf_func_state *state = func(env, reg);
3089 if (tnum_is_const(reg->var_off)) {
3090 off += reg->var_off.value;
3091 err = check_stack_write_fixed_off(env, state, off, size,
3092 value_regno, insn_idx);
3094 /* Variable offset stack reads need more conservative handling
3095 * than fixed offset ones.
3097 err = check_stack_write_var_off(env, state,
3098 ptr_regno, off, size,
3099 value_regno, insn_idx);
3104 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
3105 int off, int size, enum bpf_access_type type)
3107 struct bpf_reg_state *regs = cur_regs(env);
3108 struct bpf_map *map = regs[regno].map_ptr;
3109 u32 cap = bpf_map_flags_to_cap(map);
3111 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
3112 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
3113 map->value_size, off, size);
3117 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
3118 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
3119 map->value_size, off, size);
3126 /* check read/write into memory region (e.g., map value, ringbuf sample, etc) */
3127 static int __check_mem_access(struct bpf_verifier_env *env, int regno,
3128 int off, int size, u32 mem_size,
3129 bool zero_size_allowed)
3131 bool size_ok = size > 0 || (size == 0 && zero_size_allowed);
3132 struct bpf_reg_state *reg;
3134 if (off >= 0 && size_ok && (u64)off + size <= mem_size)
3137 reg = &cur_regs(env)[regno];
3138 switch (reg->type) {
3139 case PTR_TO_MAP_KEY:
3140 verbose(env, "invalid access to map key, key_size=%d off=%d size=%d\n",
3141 mem_size, off, size);
3143 case PTR_TO_MAP_VALUE:
3144 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
3145 mem_size, off, size);
3148 case PTR_TO_PACKET_META:
3149 case PTR_TO_PACKET_END:
3150 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
3151 off, size, regno, reg->id, off, mem_size);
3155 verbose(env, "invalid access to memory, mem_size=%u off=%d size=%d\n",
3156 mem_size, off, size);
3162 /* check read/write into a memory region with possible variable offset */
3163 static int check_mem_region_access(struct bpf_verifier_env *env, u32 regno,
3164 int off, int size, u32 mem_size,
3165 bool zero_size_allowed)
3167 struct bpf_verifier_state *vstate = env->cur_state;
3168 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3169 struct bpf_reg_state *reg = &state->regs[regno];
3172 /* We may have adjusted the register pointing to memory region, so we
3173 * need to try adding each of min_value and max_value to off
3174 * to make sure our theoretical access will be safe.
3176 if (env->log.level & BPF_LOG_LEVEL)
3177 print_verifier_state(env, state);
3179 /* The minimum value is only important with signed
3180 * comparisons where we can't assume the floor of a
3181 * value is 0. If we are using signed variables for our
3182 * index'es we need to make sure that whatever we use
3183 * will have a set floor within our range.
3185 if (reg->smin_value < 0 &&
3186 (reg->smin_value == S64_MIN ||
3187 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
3188 reg->smin_value + off < 0)) {
3189 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3193 err = __check_mem_access(env, regno, reg->smin_value + off, size,
3194 mem_size, zero_size_allowed);
3196 verbose(env, "R%d min value is outside of the allowed memory range\n",
3201 /* If we haven't set a max value then we need to bail since we can't be
3202 * sure we won't do bad things.
3203 * If reg->umax_value + off could overflow, treat that as unbounded too.
3205 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
3206 verbose(env, "R%d unbounded memory access, make sure to bounds check any such access\n",
3210 err = __check_mem_access(env, regno, reg->umax_value + off, size,
3211 mem_size, zero_size_allowed);
3213 verbose(env, "R%d max value is outside of the allowed memory range\n",
3221 /* check read/write into a map element with possible variable offset */
3222 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
3223 int off, int size, bool zero_size_allowed)
3225 struct bpf_verifier_state *vstate = env->cur_state;
3226 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3227 struct bpf_reg_state *reg = &state->regs[regno];
3228 struct bpf_map *map = reg->map_ptr;
3231 err = check_mem_region_access(env, regno, off, size, map->value_size,
3236 if (map_value_has_spin_lock(map)) {
3237 u32 lock = map->spin_lock_off;
3239 /* if any part of struct bpf_spin_lock can be touched by
3240 * load/store reject this program.
3241 * To check that [x1, x2) overlaps with [y1, y2)
3242 * it is sufficient to check x1 < y2 && y1 < x2.
3244 if (reg->smin_value + off < lock + sizeof(struct bpf_spin_lock) &&
3245 lock < reg->umax_value + off + size) {
3246 verbose(env, "bpf_spin_lock cannot be accessed directly by load/store\n");
3250 if (map_value_has_timer(map)) {
3251 u32 t = map->timer_off;
3253 if (reg->smin_value + off < t + sizeof(struct bpf_timer) &&
3254 t < reg->umax_value + off + size) {
3255 verbose(env, "bpf_timer cannot be accessed directly by load/store\n");
3262 #define MAX_PACKET_OFF 0xffff
3264 static enum bpf_prog_type resolve_prog_type(struct bpf_prog *prog)
3266 return prog->aux->dst_prog ? prog->aux->dst_prog->type : prog->type;
3269 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
3270 const struct bpf_call_arg_meta *meta,
3271 enum bpf_access_type t)
3273 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
3275 switch (prog_type) {
3276 /* Program types only with direct read access go here! */
3277 case BPF_PROG_TYPE_LWT_IN:
3278 case BPF_PROG_TYPE_LWT_OUT:
3279 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
3280 case BPF_PROG_TYPE_SK_REUSEPORT:
3281 case BPF_PROG_TYPE_FLOW_DISSECTOR:
3282 case BPF_PROG_TYPE_CGROUP_SKB:
3287 /* Program types with direct read + write access go here! */
3288 case BPF_PROG_TYPE_SCHED_CLS:
3289 case BPF_PROG_TYPE_SCHED_ACT:
3290 case BPF_PROG_TYPE_XDP:
3291 case BPF_PROG_TYPE_LWT_XMIT:
3292 case BPF_PROG_TYPE_SK_SKB:
3293 case BPF_PROG_TYPE_SK_MSG:
3295 return meta->pkt_access;
3297 env->seen_direct_write = true;
3300 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
3302 env->seen_direct_write = true;
3311 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
3312 int size, bool zero_size_allowed)
3314 struct bpf_reg_state *regs = cur_regs(env);
3315 struct bpf_reg_state *reg = ®s[regno];
3318 /* We may have added a variable offset to the packet pointer; but any
3319 * reg->range we have comes after that. We are only checking the fixed
3323 /* We don't allow negative numbers, because we aren't tracking enough
3324 * detail to prove they're safe.
3326 if (reg->smin_value < 0) {
3327 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3332 err = reg->range < 0 ? -EINVAL :
3333 __check_mem_access(env, regno, off, size, reg->range,
3336 verbose(env, "R%d offset is outside of the packet\n", regno);
3340 /* __check_mem_access has made sure "off + size - 1" is within u16.
3341 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
3342 * otherwise find_good_pkt_pointers would have refused to set range info
3343 * that __check_mem_access would have rejected this pkt access.
3344 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
3346 env->prog->aux->max_pkt_offset =
3347 max_t(u32, env->prog->aux->max_pkt_offset,
3348 off + reg->umax_value + size - 1);
3353 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
3354 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
3355 enum bpf_access_type t, enum bpf_reg_type *reg_type,
3356 struct btf **btf, u32 *btf_id)
3358 struct bpf_insn_access_aux info = {
3359 .reg_type = *reg_type,
3363 if (env->ops->is_valid_access &&
3364 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
3365 /* A non zero info.ctx_field_size indicates that this field is a
3366 * candidate for later verifier transformation to load the whole
3367 * field and then apply a mask when accessed with a narrower
3368 * access than actual ctx access size. A zero info.ctx_field_size
3369 * will only allow for whole field access and rejects any other
3370 * type of narrower access.
3372 *reg_type = info.reg_type;
3374 if (base_type(*reg_type) == PTR_TO_BTF_ID) {
3376 *btf_id = info.btf_id;
3378 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
3380 /* remember the offset of last byte accessed in ctx */
3381 if (env->prog->aux->max_ctx_offset < off + size)
3382 env->prog->aux->max_ctx_offset = off + size;
3386 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
3390 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
3393 if (size < 0 || off < 0 ||
3394 (u64)off + size > sizeof(struct bpf_flow_keys)) {
3395 verbose(env, "invalid access to flow keys off=%d size=%d\n",
3402 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
3403 u32 regno, int off, int size,
3404 enum bpf_access_type t)
3406 struct bpf_reg_state *regs = cur_regs(env);
3407 struct bpf_reg_state *reg = ®s[regno];
3408 struct bpf_insn_access_aux info = {};
3411 if (reg->smin_value < 0) {
3412 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3417 switch (reg->type) {
3418 case PTR_TO_SOCK_COMMON:
3419 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
3422 valid = bpf_sock_is_valid_access(off, size, t, &info);
3424 case PTR_TO_TCP_SOCK:
3425 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
3427 case PTR_TO_XDP_SOCK:
3428 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
3436 env->insn_aux_data[insn_idx].ctx_field_size =
3437 info.ctx_field_size;
3441 verbose(env, "R%d invalid %s access off=%d size=%d\n",
3442 regno, reg_type_str(env, reg->type), off, size);
3447 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
3449 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
3452 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
3454 const struct bpf_reg_state *reg = reg_state(env, regno);
3456 return reg->type == PTR_TO_CTX;
3459 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
3461 const struct bpf_reg_state *reg = reg_state(env, regno);
3463 return type_is_sk_pointer(reg->type);
3466 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
3468 const struct bpf_reg_state *reg = reg_state(env, regno);
3470 return type_is_pkt_pointer(reg->type);
3473 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
3475 const struct bpf_reg_state *reg = reg_state(env, regno);
3477 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
3478 return reg->type == PTR_TO_FLOW_KEYS;
3481 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
3482 const struct bpf_reg_state *reg,
3483 int off, int size, bool strict)
3485 struct tnum reg_off;
3488 /* Byte size accesses are always allowed. */
3489 if (!strict || size == 1)
3492 /* For platforms that do not have a Kconfig enabling
3493 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
3494 * NET_IP_ALIGN is universally set to '2'. And on platforms
3495 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
3496 * to this code only in strict mode where we want to emulate
3497 * the NET_IP_ALIGN==2 checking. Therefore use an
3498 * unconditional IP align value of '2'.
3502 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
3503 if (!tnum_is_aligned(reg_off, size)) {
3506 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3508 "misaligned packet access off %d+%s+%d+%d size %d\n",
3509 ip_align, tn_buf, reg->off, off, size);
3516 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
3517 const struct bpf_reg_state *reg,
3518 const char *pointer_desc,
3519 int off, int size, bool strict)
3521 struct tnum reg_off;
3523 /* Byte size accesses are always allowed. */
3524 if (!strict || size == 1)
3527 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
3528 if (!tnum_is_aligned(reg_off, size)) {
3531 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3532 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
3533 pointer_desc, tn_buf, reg->off, off, size);
3540 static int check_ptr_alignment(struct bpf_verifier_env *env,
3541 const struct bpf_reg_state *reg, int off,
3542 int size, bool strict_alignment_once)
3544 bool strict = env->strict_alignment || strict_alignment_once;
3545 const char *pointer_desc = "";
3547 switch (reg->type) {
3549 case PTR_TO_PACKET_META:
3550 /* Special case, because of NET_IP_ALIGN. Given metadata sits
3551 * right in front, treat it the very same way.
3553 return check_pkt_ptr_alignment(env, reg, off, size, strict);
3554 case PTR_TO_FLOW_KEYS:
3555 pointer_desc = "flow keys ";
3557 case PTR_TO_MAP_KEY:
3558 pointer_desc = "key ";
3560 case PTR_TO_MAP_VALUE:
3561 pointer_desc = "value ";
3564 pointer_desc = "context ";
3567 pointer_desc = "stack ";
3568 /* The stack spill tracking logic in check_stack_write_fixed_off()
3569 * and check_stack_read_fixed_off() relies on stack accesses being
3575 pointer_desc = "sock ";
3577 case PTR_TO_SOCK_COMMON:
3578 pointer_desc = "sock_common ";
3580 case PTR_TO_TCP_SOCK:
3581 pointer_desc = "tcp_sock ";
3583 case PTR_TO_XDP_SOCK:
3584 pointer_desc = "xdp_sock ";
3589 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
3593 static int update_stack_depth(struct bpf_verifier_env *env,
3594 const struct bpf_func_state *func,
3597 u16 stack = env->subprog_info[func->subprogno].stack_depth;
3602 /* update known max for given subprogram */
3603 env->subprog_info[func->subprogno].stack_depth = -off;
3607 /* starting from main bpf function walk all instructions of the function
3608 * and recursively walk all callees that given function can call.
3609 * Ignore jump and exit insns.
3610 * Since recursion is prevented by check_cfg() this algorithm
3611 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
3613 static int check_max_stack_depth(struct bpf_verifier_env *env)
3615 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
3616 struct bpf_subprog_info *subprog = env->subprog_info;
3617 struct bpf_insn *insn = env->prog->insnsi;
3618 bool tail_call_reachable = false;
3619 int ret_insn[MAX_CALL_FRAMES];
3620 int ret_prog[MAX_CALL_FRAMES];
3624 /* protect against potential stack overflow that might happen when
3625 * bpf2bpf calls get combined with tailcalls. Limit the caller's stack
3626 * depth for such case down to 256 so that the worst case scenario
3627 * would result in 8k stack size (32 which is tailcall limit * 256 =
3630 * To get the idea what might happen, see an example:
3631 * func1 -> sub rsp, 128
3632 * subfunc1 -> sub rsp, 256
3633 * tailcall1 -> add rsp, 256
3634 * func2 -> sub rsp, 192 (total stack size = 128 + 192 = 320)
3635 * subfunc2 -> sub rsp, 64
3636 * subfunc22 -> sub rsp, 128
3637 * tailcall2 -> add rsp, 128
3638 * func3 -> sub rsp, 32 (total stack size 128 + 192 + 64 + 32 = 416)
3640 * tailcall will unwind the current stack frame but it will not get rid
3641 * of caller's stack as shown on the example above.
3643 if (idx && subprog[idx].has_tail_call && depth >= 256) {
3645 "tail_calls are not allowed when call stack of previous frames is %d bytes. Too large\n",
3649 /* round up to 32-bytes, since this is granularity
3650 * of interpreter stack size
3652 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
3653 if (depth > MAX_BPF_STACK) {
3654 verbose(env, "combined stack size of %d calls is %d. Too large\n",
3659 subprog_end = subprog[idx + 1].start;
3660 for (; i < subprog_end; i++) {
3663 if (!bpf_pseudo_call(insn + i) && !bpf_pseudo_func(insn + i))
3665 /* remember insn and function to return to */
3666 ret_insn[frame] = i + 1;
3667 ret_prog[frame] = idx;
3669 /* find the callee */
3670 next_insn = i + insn[i].imm + 1;
3671 idx = find_subprog(env, next_insn);
3673 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3677 if (subprog[idx].is_async_cb) {
3678 if (subprog[idx].has_tail_call) {
3679 verbose(env, "verifier bug. subprog has tail_call and async cb\n");
3682 /* async callbacks don't increase bpf prog stack size */
3687 if (subprog[idx].has_tail_call)
3688 tail_call_reachable = true;
3691 if (frame >= MAX_CALL_FRAMES) {
3692 verbose(env, "the call stack of %d frames is too deep !\n",
3698 /* if tail call got detected across bpf2bpf calls then mark each of the
3699 * currently present subprog frames as tail call reachable subprogs;
3700 * this info will be utilized by JIT so that we will be preserving the
3701 * tail call counter throughout bpf2bpf calls combined with tailcalls
3703 if (tail_call_reachable)
3704 for (j = 0; j < frame; j++)
3705 subprog[ret_prog[j]].tail_call_reachable = true;
3706 if (subprog[0].tail_call_reachable)
3707 env->prog->aux->tail_call_reachable = true;
3709 /* end of for() loop means the last insn of the 'subprog'
3710 * was reached. Doesn't matter whether it was JA or EXIT
3714 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
3716 i = ret_insn[frame];
3717 idx = ret_prog[frame];
3721 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
3722 static int get_callee_stack_depth(struct bpf_verifier_env *env,
3723 const struct bpf_insn *insn, int idx)
3725 int start = idx + insn->imm + 1, subprog;
3727 subprog = find_subprog(env, start);
3729 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3733 return env->subprog_info[subprog].stack_depth;
3737 int check_ctx_reg(struct bpf_verifier_env *env,
3738 const struct bpf_reg_state *reg, int regno)
3740 /* Access to ctx or passing it to a helper is only allowed in
3741 * its original, unmodified form.
3745 verbose(env, "dereference of modified ctx ptr R%d off=%d disallowed\n",
3750 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3753 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3754 verbose(env, "variable ctx access var_off=%s disallowed\n", tn_buf);
3761 static int __check_buffer_access(struct bpf_verifier_env *env,
3762 const char *buf_info,
3763 const struct bpf_reg_state *reg,
3764 int regno, int off, int size)
3768 "R%d invalid %s buffer access: off=%d, size=%d\n",
3769 regno, buf_info, off, size);
3772 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3775 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3777 "R%d invalid variable buffer offset: off=%d, var_off=%s\n",
3778 regno, off, tn_buf);
3785 static int check_tp_buffer_access(struct bpf_verifier_env *env,
3786 const struct bpf_reg_state *reg,
3787 int regno, int off, int size)
3791 err = __check_buffer_access(env, "tracepoint", reg, regno, off, size);
3795 if (off + size > env->prog->aux->max_tp_access)
3796 env->prog->aux->max_tp_access = off + size;
3801 static int check_buffer_access(struct bpf_verifier_env *env,
3802 const struct bpf_reg_state *reg,
3803 int regno, int off, int size,
3804 bool zero_size_allowed,
3805 const char *buf_info,
3810 err = __check_buffer_access(env, buf_info, reg, regno, off, size);
3814 if (off + size > *max_access)
3815 *max_access = off + size;
3820 /* BPF architecture zero extends alu32 ops into 64-bit registesr */
3821 static void zext_32_to_64(struct bpf_reg_state *reg)
3823 reg->var_off = tnum_subreg(reg->var_off);
3824 __reg_assign_32_into_64(reg);
3827 /* truncate register to smaller size (in bytes)
3828 * must be called with size < BPF_REG_SIZE
3830 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
3834 /* clear high bits in bit representation */
3835 reg->var_off = tnum_cast(reg->var_off, size);
3837 /* fix arithmetic bounds */
3838 mask = ((u64)1 << (size * 8)) - 1;
3839 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
3840 reg->umin_value &= mask;
3841 reg->umax_value &= mask;
3843 reg->umin_value = 0;
3844 reg->umax_value = mask;
3846 reg->smin_value = reg->umin_value;
3847 reg->smax_value = reg->umax_value;
3849 /* If size is smaller than 32bit register the 32bit register
3850 * values are also truncated so we push 64-bit bounds into
3851 * 32-bit bounds. Above were truncated < 32-bits already.
3855 __reg_combine_64_into_32(reg);
3858 static bool bpf_map_is_rdonly(const struct bpf_map *map)
3860 /* A map is considered read-only if the following condition are true:
3862 * 1) BPF program side cannot change any of the map content. The
3863 * BPF_F_RDONLY_PROG flag is throughout the lifetime of a map
3864 * and was set at map creation time.
3865 * 2) The map value(s) have been initialized from user space by a
3866 * loader and then "frozen", such that no new map update/delete
3867 * operations from syscall side are possible for the rest of
3868 * the map's lifetime from that point onwards.
3869 * 3) Any parallel/pending map update/delete operations from syscall
3870 * side have been completed. Only after that point, it's safe to
3871 * assume that map value(s) are immutable.
3873 return (map->map_flags & BPF_F_RDONLY_PROG) &&
3874 READ_ONCE(map->frozen) &&
3875 !bpf_map_write_active(map);
3878 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val)
3884 err = map->ops->map_direct_value_addr(map, &addr, off);
3887 ptr = (void *)(long)addr + off;
3891 *val = (u64)*(u8 *)ptr;
3894 *val = (u64)*(u16 *)ptr;
3897 *val = (u64)*(u32 *)ptr;
3908 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
3909 struct bpf_reg_state *regs,
3910 int regno, int off, int size,
3911 enum bpf_access_type atype,
3914 struct bpf_reg_state *reg = regs + regno;
3915 const struct btf_type *t = btf_type_by_id(reg->btf, reg->btf_id);
3916 const char *tname = btf_name_by_offset(reg->btf, t->name_off);
3922 "R%d is ptr_%s invalid negative access: off=%d\n",
3926 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3929 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3931 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
3932 regno, tname, off, tn_buf);
3936 if (env->ops->btf_struct_access) {
3937 ret = env->ops->btf_struct_access(&env->log, reg->btf, t,
3938 off, size, atype, &btf_id);
3940 if (atype != BPF_READ) {
3941 verbose(env, "only read is supported\n");
3945 ret = btf_struct_access(&env->log, reg->btf, t, off, size,
3952 if (atype == BPF_READ && value_regno >= 0)
3953 mark_btf_ld_reg(env, regs, value_regno, ret, reg->btf, btf_id);
3958 static int check_ptr_to_map_access(struct bpf_verifier_env *env,
3959 struct bpf_reg_state *regs,
3960 int regno, int off, int size,
3961 enum bpf_access_type atype,
3964 struct bpf_reg_state *reg = regs + regno;
3965 struct bpf_map *map = reg->map_ptr;
3966 const struct btf_type *t;
3972 verbose(env, "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n");
3976 if (!map->ops->map_btf_id || !*map->ops->map_btf_id) {
3977 verbose(env, "map_ptr access not supported for map type %d\n",
3982 t = btf_type_by_id(btf_vmlinux, *map->ops->map_btf_id);
3983 tname = btf_name_by_offset(btf_vmlinux, t->name_off);
3985 if (!env->allow_ptr_to_map_access) {
3987 "%s access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
3993 verbose(env, "R%d is %s invalid negative access: off=%d\n",
3998 if (atype != BPF_READ) {
3999 verbose(env, "only read from %s is supported\n", tname);
4003 ret = btf_struct_access(&env->log, btf_vmlinux, t, off, size, atype, &btf_id);
4007 if (value_regno >= 0)
4008 mark_btf_ld_reg(env, regs, value_regno, ret, btf_vmlinux, btf_id);
4013 /* Check that the stack access at the given offset is within bounds. The
4014 * maximum valid offset is -1.
4016 * The minimum valid offset is -MAX_BPF_STACK for writes, and
4017 * -state->allocated_stack for reads.
4019 static int check_stack_slot_within_bounds(int off,
4020 struct bpf_func_state *state,
4021 enum bpf_access_type t)
4026 min_valid_off = -MAX_BPF_STACK;
4028 min_valid_off = -state->allocated_stack;
4030 if (off < min_valid_off || off > -1)
4035 /* Check that the stack access at 'regno + off' falls within the maximum stack
4038 * 'off' includes `regno->offset`, but not its dynamic part (if any).
4040 static int check_stack_access_within_bounds(
4041 struct bpf_verifier_env *env,
4042 int regno, int off, int access_size,
4043 enum stack_access_src src, enum bpf_access_type type)
4045 struct bpf_reg_state *regs = cur_regs(env);
4046 struct bpf_reg_state *reg = regs + regno;
4047 struct bpf_func_state *state = func(env, reg);
4048 int min_off, max_off;
4052 if (src == ACCESS_HELPER)
4053 /* We don't know if helpers are reading or writing (or both). */
4054 err_extra = " indirect access to";
4055 else if (type == BPF_READ)
4056 err_extra = " read from";
4058 err_extra = " write to";
4060 if (tnum_is_const(reg->var_off)) {
4061 min_off = reg->var_off.value + off;
4062 if (access_size > 0)
4063 max_off = min_off + access_size - 1;
4067 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
4068 reg->smin_value <= -BPF_MAX_VAR_OFF) {
4069 verbose(env, "invalid unbounded variable-offset%s stack R%d\n",
4073 min_off = reg->smin_value + off;
4074 if (access_size > 0)
4075 max_off = reg->smax_value + off + access_size - 1;
4080 err = check_stack_slot_within_bounds(min_off, state, type);
4082 err = check_stack_slot_within_bounds(max_off, state, type);
4085 if (tnum_is_const(reg->var_off)) {
4086 verbose(env, "invalid%s stack R%d off=%d size=%d\n",
4087 err_extra, regno, off, access_size);
4091 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4092 verbose(env, "invalid variable-offset%s stack R%d var_off=%s size=%d\n",
4093 err_extra, regno, tn_buf, access_size);
4099 /* check whether memory at (regno + off) is accessible for t = (read | write)
4100 * if t==write, value_regno is a register which value is stored into memory
4101 * if t==read, value_regno is a register which will receive the value from memory
4102 * if t==write && value_regno==-1, some unknown value is stored into memory
4103 * if t==read && value_regno==-1, don't care what we read from memory
4105 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
4106 int off, int bpf_size, enum bpf_access_type t,
4107 int value_regno, bool strict_alignment_once)
4109 struct bpf_reg_state *regs = cur_regs(env);
4110 struct bpf_reg_state *reg = regs + regno;
4111 struct bpf_func_state *state;
4114 size = bpf_size_to_bytes(bpf_size);
4118 /* alignment checks will add in reg->off themselves */
4119 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
4123 /* for access checks, reg->off is just part of off */
4126 if (reg->type == PTR_TO_MAP_KEY) {
4127 if (t == BPF_WRITE) {
4128 verbose(env, "write to change key R%d not allowed\n", regno);
4132 err = check_mem_region_access(env, regno, off, size,
4133 reg->map_ptr->key_size, false);
4136 if (value_regno >= 0)
4137 mark_reg_unknown(env, regs, value_regno);
4138 } else if (reg->type == PTR_TO_MAP_VALUE) {
4139 if (t == BPF_WRITE && value_regno >= 0 &&
4140 is_pointer_value(env, value_regno)) {
4141 verbose(env, "R%d leaks addr into map\n", value_regno);
4144 err = check_map_access_type(env, regno, off, size, t);
4147 err = check_map_access(env, regno, off, size, false);
4148 if (!err && t == BPF_READ && value_regno >= 0) {
4149 struct bpf_map *map = reg->map_ptr;
4151 /* if map is read-only, track its contents as scalars */
4152 if (tnum_is_const(reg->var_off) &&
4153 bpf_map_is_rdonly(map) &&
4154 map->ops->map_direct_value_addr) {
4155 int map_off = off + reg->var_off.value;
4158 err = bpf_map_direct_read(map, map_off, size,
4163 regs[value_regno].type = SCALAR_VALUE;
4164 __mark_reg_known(®s[value_regno], val);
4166 mark_reg_unknown(env, regs, value_regno);
4169 } else if (base_type(reg->type) == PTR_TO_MEM) {
4170 bool rdonly_mem = type_is_rdonly_mem(reg->type);
4172 if (type_may_be_null(reg->type)) {
4173 verbose(env, "R%d invalid mem access '%s'\n", regno,
4174 reg_type_str(env, reg->type));
4178 if (t == BPF_WRITE && rdonly_mem) {
4179 verbose(env, "R%d cannot write into %s\n",
4180 regno, reg_type_str(env, reg->type));
4184 if (t == BPF_WRITE && value_regno >= 0 &&
4185 is_pointer_value(env, value_regno)) {
4186 verbose(env, "R%d leaks addr into mem\n", value_regno);
4190 err = check_mem_region_access(env, regno, off, size,
4191 reg->mem_size, false);
4192 if (!err && value_regno >= 0 && (t == BPF_READ || rdonly_mem))
4193 mark_reg_unknown(env, regs, value_regno);
4194 } else if (reg->type == PTR_TO_CTX) {
4195 enum bpf_reg_type reg_type = SCALAR_VALUE;
4196 struct btf *btf = NULL;
4199 if (t == BPF_WRITE && value_regno >= 0 &&
4200 is_pointer_value(env, value_regno)) {
4201 verbose(env, "R%d leaks addr into ctx\n", value_regno);
4205 err = check_ctx_reg(env, reg, regno);
4209 err = check_ctx_access(env, insn_idx, off, size, t, ®_type, &btf, &btf_id);
4211 verbose_linfo(env, insn_idx, "; ");
4212 if (!err && t == BPF_READ && value_regno >= 0) {
4213 /* ctx access returns either a scalar, or a
4214 * PTR_TO_PACKET[_META,_END]. In the latter
4215 * case, we know the offset is zero.
4217 if (reg_type == SCALAR_VALUE) {
4218 mark_reg_unknown(env, regs, value_regno);
4220 mark_reg_known_zero(env, regs,
4222 if (type_may_be_null(reg_type))
4223 regs[value_regno].id = ++env->id_gen;
4224 /* A load of ctx field could have different
4225 * actual load size with the one encoded in the
4226 * insn. When the dst is PTR, it is for sure not
4229 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
4230 if (base_type(reg_type) == PTR_TO_BTF_ID) {
4231 regs[value_regno].btf = btf;
4232 regs[value_regno].btf_id = btf_id;
4235 regs[value_regno].type = reg_type;
4238 } else if (reg->type == PTR_TO_STACK) {
4239 /* Basic bounds checks. */
4240 err = check_stack_access_within_bounds(env, regno, off, size, ACCESS_DIRECT, t);
4244 state = func(env, reg);
4245 err = update_stack_depth(env, state, off);
4250 err = check_stack_read(env, regno, off, size,
4253 err = check_stack_write(env, regno, off, size,
4254 value_regno, insn_idx);
4255 } else if (reg_is_pkt_pointer(reg)) {
4256 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
4257 verbose(env, "cannot write into packet\n");
4260 if (t == BPF_WRITE && value_regno >= 0 &&
4261 is_pointer_value(env, value_regno)) {
4262 verbose(env, "R%d leaks addr into packet\n",
4266 err = check_packet_access(env, regno, off, size, false);
4267 if (!err && t == BPF_READ && value_regno >= 0)
4268 mark_reg_unknown(env, regs, value_regno);
4269 } else if (reg->type == PTR_TO_FLOW_KEYS) {
4270 if (t == BPF_WRITE && value_regno >= 0 &&
4271 is_pointer_value(env, value_regno)) {
4272 verbose(env, "R%d leaks addr into flow keys\n",
4277 err = check_flow_keys_access(env, off, size);
4278 if (!err && t == BPF_READ && value_regno >= 0)
4279 mark_reg_unknown(env, regs, value_regno);
4280 } else if (type_is_sk_pointer(reg->type)) {
4281 if (t == BPF_WRITE) {
4282 verbose(env, "R%d cannot write into %s\n",
4283 regno, reg_type_str(env, reg->type));
4286 err = check_sock_access(env, insn_idx, regno, off, size, t);
4287 if (!err && value_regno >= 0)
4288 mark_reg_unknown(env, regs, value_regno);
4289 } else if (reg->type == PTR_TO_TP_BUFFER) {
4290 err = check_tp_buffer_access(env, reg, regno, off, size);
4291 if (!err && t == BPF_READ && value_regno >= 0)
4292 mark_reg_unknown(env, regs, value_regno);
4293 } else if (reg->type == PTR_TO_BTF_ID) {
4294 err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
4296 } else if (reg->type == CONST_PTR_TO_MAP) {
4297 err = check_ptr_to_map_access(env, regs, regno, off, size, t,
4299 } else if (base_type(reg->type) == PTR_TO_BUF) {
4300 bool rdonly_mem = type_is_rdonly_mem(reg->type);
4301 const char *buf_info;
4305 if (t == BPF_WRITE) {
4306 verbose(env, "R%d cannot write into %s\n",
4307 regno, reg_type_str(env, reg->type));
4310 buf_info = "rdonly";
4311 max_access = &env->prog->aux->max_rdonly_access;
4314 max_access = &env->prog->aux->max_rdwr_access;
4317 err = check_buffer_access(env, reg, regno, off, size, false,
4318 buf_info, max_access);
4320 if (!err && value_regno >= 0 && (rdonly_mem || t == BPF_READ))
4321 mark_reg_unknown(env, regs, value_regno);
4323 verbose(env, "R%d invalid mem access '%s'\n", regno,
4324 reg_type_str(env, reg->type));
4328 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
4329 regs[value_regno].type == SCALAR_VALUE) {
4330 /* b/h/w load zero-extends, mark upper bits as known 0 */
4331 coerce_reg_to_size(®s[value_regno], size);
4336 static int check_atomic(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
4341 switch (insn->imm) {
4343 case BPF_ADD | BPF_FETCH:
4345 case BPF_AND | BPF_FETCH:
4347 case BPF_OR | BPF_FETCH:
4349 case BPF_XOR | BPF_FETCH:
4354 verbose(env, "BPF_ATOMIC uses invalid atomic opcode %02x\n", insn->imm);
4358 if (BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) {
4359 verbose(env, "invalid atomic operand size\n");
4363 /* check src1 operand */
4364 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4368 /* check src2 operand */
4369 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4373 if (insn->imm == BPF_CMPXCHG) {
4374 /* Check comparison of R0 with memory location */
4375 const u32 aux_reg = BPF_REG_0;
4377 err = check_reg_arg(env, aux_reg, SRC_OP);
4381 if (is_pointer_value(env, aux_reg)) {
4382 verbose(env, "R%d leaks addr into mem\n", aux_reg);
4387 if (is_pointer_value(env, insn->src_reg)) {
4388 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
4392 if (is_ctx_reg(env, insn->dst_reg) ||
4393 is_pkt_reg(env, insn->dst_reg) ||
4394 is_flow_key_reg(env, insn->dst_reg) ||
4395 is_sk_reg(env, insn->dst_reg)) {
4396 verbose(env, "BPF_ATOMIC stores into R%d %s is not allowed\n",
4398 reg_type_str(env, reg_state(env, insn->dst_reg)->type));
4402 if (insn->imm & BPF_FETCH) {
4403 if (insn->imm == BPF_CMPXCHG)
4404 load_reg = BPF_REG_0;
4406 load_reg = insn->src_reg;
4408 /* check and record load of old value */
4409 err = check_reg_arg(env, load_reg, DST_OP);
4413 /* This instruction accesses a memory location but doesn't
4414 * actually load it into a register.
4419 /* Check whether we can read the memory, with second call for fetch
4420 * case to simulate the register fill.
4422 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4423 BPF_SIZE(insn->code), BPF_READ, -1, true);
4424 if (!err && load_reg >= 0)
4425 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4426 BPF_SIZE(insn->code), BPF_READ, load_reg,
4431 /* Check whether we can write into the same memory. */
4432 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4433 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
4440 /* When register 'regno' is used to read the stack (either directly or through
4441 * a helper function) make sure that it's within stack boundary and, depending
4442 * on the access type, that all elements of the stack are initialized.
4444 * 'off' includes 'regno->off', but not its dynamic part (if any).
4446 * All registers that have been spilled on the stack in the slots within the
4447 * read offsets are marked as read.
4449 static int check_stack_range_initialized(
4450 struct bpf_verifier_env *env, int regno, int off,
4451 int access_size, bool zero_size_allowed,
4452 enum stack_access_src type, struct bpf_call_arg_meta *meta)
4454 struct bpf_reg_state *reg = reg_state(env, regno);
4455 struct bpf_func_state *state = func(env, reg);
4456 int err, min_off, max_off, i, j, slot, spi;
4457 char *err_extra = type == ACCESS_HELPER ? " indirect" : "";
4458 enum bpf_access_type bounds_check_type;
4459 /* Some accesses can write anything into the stack, others are
4462 bool clobber = false;
4464 if (access_size == 0 && !zero_size_allowed) {
4465 verbose(env, "invalid zero-sized read\n");
4469 if (type == ACCESS_HELPER) {
4470 /* The bounds checks for writes are more permissive than for
4471 * reads. However, if raw_mode is not set, we'll do extra
4474 bounds_check_type = BPF_WRITE;
4477 bounds_check_type = BPF_READ;
4479 err = check_stack_access_within_bounds(env, regno, off, access_size,
4480 type, bounds_check_type);
4485 if (tnum_is_const(reg->var_off)) {
4486 min_off = max_off = reg->var_off.value + off;
4488 /* Variable offset is prohibited for unprivileged mode for
4489 * simplicity since it requires corresponding support in
4490 * Spectre masking for stack ALU.
4491 * See also retrieve_ptr_limit().
4493 if (!env->bypass_spec_v1) {
4496 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4497 verbose(env, "R%d%s variable offset stack access prohibited for !root, var_off=%s\n",
4498 regno, err_extra, tn_buf);
4501 /* Only initialized buffer on stack is allowed to be accessed
4502 * with variable offset. With uninitialized buffer it's hard to
4503 * guarantee that whole memory is marked as initialized on
4504 * helper return since specific bounds are unknown what may
4505 * cause uninitialized stack leaking.
4507 if (meta && meta->raw_mode)
4510 min_off = reg->smin_value + off;
4511 max_off = reg->smax_value + off;
4514 if (meta && meta->raw_mode) {
4515 meta->access_size = access_size;
4516 meta->regno = regno;
4520 for (i = min_off; i < max_off + access_size; i++) {
4524 spi = slot / BPF_REG_SIZE;
4525 if (state->allocated_stack <= slot)
4527 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
4528 if (*stype == STACK_MISC)
4530 if (*stype == STACK_ZERO) {
4532 /* helper can write anything into the stack */
4533 *stype = STACK_MISC;
4538 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
4539 state->stack[spi].spilled_ptr.type == PTR_TO_BTF_ID)
4542 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
4543 (state->stack[spi].spilled_ptr.type == SCALAR_VALUE ||
4544 env->allow_ptr_leaks)) {
4546 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
4547 for (j = 0; j < BPF_REG_SIZE; j++)
4548 state->stack[spi].slot_type[j] = STACK_MISC;
4554 if (tnum_is_const(reg->var_off)) {
4555 verbose(env, "invalid%s read from stack R%d off %d+%d size %d\n",
4556 err_extra, regno, min_off, i - min_off, access_size);
4560 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4561 verbose(env, "invalid%s read from stack R%d var_off %s+%d size %d\n",
4562 err_extra, regno, tn_buf, i - min_off, access_size);
4566 /* reading any byte out of 8-byte 'spill_slot' will cause
4567 * the whole slot to be marked as 'read'
4569 mark_reg_read(env, &state->stack[spi].spilled_ptr,
4570 state->stack[spi].spilled_ptr.parent,
4573 return update_stack_depth(env, state, min_off);
4576 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
4577 int access_size, bool zero_size_allowed,
4578 struct bpf_call_arg_meta *meta)
4580 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4581 const char *buf_info;
4584 switch (base_type(reg->type)) {
4586 case PTR_TO_PACKET_META:
4587 return check_packet_access(env, regno, reg->off, access_size,
4589 case PTR_TO_MAP_KEY:
4590 if (meta && meta->raw_mode) {
4591 verbose(env, "R%d cannot write into %s\n", regno,
4592 reg_type_str(env, reg->type));
4595 return check_mem_region_access(env, regno, reg->off, access_size,
4596 reg->map_ptr->key_size, false);
4597 case PTR_TO_MAP_VALUE:
4598 if (check_map_access_type(env, regno, reg->off, access_size,
4599 meta && meta->raw_mode ? BPF_WRITE :
4602 return check_map_access(env, regno, reg->off, access_size,
4605 if (type_is_rdonly_mem(reg->type)) {
4606 if (meta && meta->raw_mode) {
4607 verbose(env, "R%d cannot write into %s\n", regno,
4608 reg_type_str(env, reg->type));
4612 return check_mem_region_access(env, regno, reg->off,
4613 access_size, reg->mem_size,
4616 if (type_is_rdonly_mem(reg->type)) {
4617 if (meta && meta->raw_mode) {
4618 verbose(env, "R%d cannot write into %s\n", regno,
4619 reg_type_str(env, reg->type));
4623 buf_info = "rdonly";
4624 max_access = &env->prog->aux->max_rdonly_access;
4627 max_access = &env->prog->aux->max_rdwr_access;
4629 return check_buffer_access(env, reg, regno, reg->off,
4630 access_size, zero_size_allowed,
4631 buf_info, max_access);
4633 return check_stack_range_initialized(
4635 regno, reg->off, access_size,
4636 zero_size_allowed, ACCESS_HELPER, meta);
4637 default: /* scalar_value or invalid ptr */
4638 /* Allow zero-byte read from NULL, regardless of pointer type */
4639 if (zero_size_allowed && access_size == 0 &&
4640 register_is_null(reg))
4643 verbose(env, "R%d type=%s ", regno,
4644 reg_type_str(env, reg->type));
4645 verbose(env, "expected=%s\n", reg_type_str(env, PTR_TO_STACK));
4650 int check_mem_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
4651 u32 regno, u32 mem_size)
4653 if (register_is_null(reg))
4656 if (type_may_be_null(reg->type)) {
4657 /* Assuming that the register contains a value check if the memory
4658 * access is safe. Temporarily save and restore the register's state as
4659 * the conversion shouldn't be visible to a caller.
4661 const struct bpf_reg_state saved_reg = *reg;
4664 mark_ptr_not_null_reg(reg);
4665 rv = check_helper_mem_access(env, regno, mem_size, true, NULL);
4670 return check_helper_mem_access(env, regno, mem_size, true, NULL);
4673 /* Implementation details:
4674 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL
4675 * Two bpf_map_lookups (even with the same key) will have different reg->id.
4676 * For traditional PTR_TO_MAP_VALUE the verifier clears reg->id after
4677 * value_or_null->value transition, since the verifier only cares about
4678 * the range of access to valid map value pointer and doesn't care about actual
4679 * address of the map element.
4680 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
4681 * reg->id > 0 after value_or_null->value transition. By doing so
4682 * two bpf_map_lookups will be considered two different pointers that
4683 * point to different bpf_spin_locks.
4684 * The verifier allows taking only one bpf_spin_lock at a time to avoid
4686 * Since only one bpf_spin_lock is allowed the checks are simpler than
4687 * reg_is_refcounted() logic. The verifier needs to remember only
4688 * one spin_lock instead of array of acquired_refs.
4689 * cur_state->active_spin_lock remembers which map value element got locked
4690 * and clears it after bpf_spin_unlock.
4692 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
4695 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4696 struct bpf_verifier_state *cur = env->cur_state;
4697 bool is_const = tnum_is_const(reg->var_off);
4698 struct bpf_map *map = reg->map_ptr;
4699 u64 val = reg->var_off.value;
4703 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
4709 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
4713 if (!map_value_has_spin_lock(map)) {
4714 if (map->spin_lock_off == -E2BIG)
4716 "map '%s' has more than one 'struct bpf_spin_lock'\n",
4718 else if (map->spin_lock_off == -ENOENT)
4720 "map '%s' doesn't have 'struct bpf_spin_lock'\n",
4724 "map '%s' is not a struct type or bpf_spin_lock is mangled\n",
4728 if (map->spin_lock_off != val + reg->off) {
4729 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock'\n",
4734 if (cur->active_spin_lock) {
4736 "Locking two bpf_spin_locks are not allowed\n");
4739 cur->active_spin_lock = reg->id;
4741 if (!cur->active_spin_lock) {
4742 verbose(env, "bpf_spin_unlock without taking a lock\n");
4745 if (cur->active_spin_lock != reg->id) {
4746 verbose(env, "bpf_spin_unlock of different lock\n");
4749 cur->active_spin_lock = 0;
4754 static int process_timer_func(struct bpf_verifier_env *env, int regno,
4755 struct bpf_call_arg_meta *meta)
4757 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4758 bool is_const = tnum_is_const(reg->var_off);
4759 struct bpf_map *map = reg->map_ptr;
4760 u64 val = reg->var_off.value;
4764 "R%d doesn't have constant offset. bpf_timer has to be at the constant offset\n",
4769 verbose(env, "map '%s' has to have BTF in order to use bpf_timer\n",
4773 if (!map_value_has_timer(map)) {
4774 if (map->timer_off == -E2BIG)
4776 "map '%s' has more than one 'struct bpf_timer'\n",
4778 else if (map->timer_off == -ENOENT)
4780 "map '%s' doesn't have 'struct bpf_timer'\n",
4784 "map '%s' is not a struct type or bpf_timer is mangled\n",
4788 if (map->timer_off != val + reg->off) {
4789 verbose(env, "off %lld doesn't point to 'struct bpf_timer' that is at %d\n",
4790 val + reg->off, map->timer_off);
4793 if (meta->map_ptr) {
4794 verbose(env, "verifier bug. Two map pointers in a timer helper\n");
4797 meta->map_uid = reg->map_uid;
4798 meta->map_ptr = map;
4802 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
4804 return base_type(type) == ARG_PTR_TO_MEM ||
4805 base_type(type) == ARG_PTR_TO_UNINIT_MEM;
4808 static bool arg_type_is_mem_size(enum bpf_arg_type type)
4810 return type == ARG_CONST_SIZE ||
4811 type == ARG_CONST_SIZE_OR_ZERO;
4814 static bool arg_type_is_alloc_size(enum bpf_arg_type type)
4816 return type == ARG_CONST_ALLOC_SIZE_OR_ZERO;
4819 static bool arg_type_is_int_ptr(enum bpf_arg_type type)
4821 return type == ARG_PTR_TO_INT ||
4822 type == ARG_PTR_TO_LONG;
4825 static int int_ptr_type_to_size(enum bpf_arg_type type)
4827 if (type == ARG_PTR_TO_INT)
4829 else if (type == ARG_PTR_TO_LONG)
4835 static int resolve_map_arg_type(struct bpf_verifier_env *env,
4836 const struct bpf_call_arg_meta *meta,
4837 enum bpf_arg_type *arg_type)
4839 if (!meta->map_ptr) {
4840 /* kernel subsystem misconfigured verifier */
4841 verbose(env, "invalid map_ptr to access map->type\n");
4845 switch (meta->map_ptr->map_type) {
4846 case BPF_MAP_TYPE_SOCKMAP:
4847 case BPF_MAP_TYPE_SOCKHASH:
4848 if (*arg_type == ARG_PTR_TO_MAP_VALUE) {
4849 *arg_type = ARG_PTR_TO_BTF_ID_SOCK_COMMON;
4851 verbose(env, "invalid arg_type for sockmap/sockhash\n");
4862 struct bpf_reg_types {
4863 const enum bpf_reg_type types[10];
4867 static const struct bpf_reg_types map_key_value_types = {
4877 static const struct bpf_reg_types sock_types = {
4887 static const struct bpf_reg_types btf_id_sock_common_types = {
4895 .btf_id = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
4899 static const struct bpf_reg_types mem_types = {
4911 static const struct bpf_reg_types int_ptr_types = {
4921 static const struct bpf_reg_types fullsock_types = { .types = { PTR_TO_SOCKET } };
4922 static const struct bpf_reg_types scalar_types = { .types = { SCALAR_VALUE } };
4923 static const struct bpf_reg_types context_types = { .types = { PTR_TO_CTX } };
4924 static const struct bpf_reg_types alloc_mem_types = { .types = { PTR_TO_MEM } };
4925 static const struct bpf_reg_types const_map_ptr_types = { .types = { CONST_PTR_TO_MAP } };
4926 static const struct bpf_reg_types btf_ptr_types = { .types = { PTR_TO_BTF_ID } };
4927 static const struct bpf_reg_types spin_lock_types = { .types = { PTR_TO_MAP_VALUE } };
4928 static const struct bpf_reg_types percpu_btf_ptr_types = { .types = { PTR_TO_PERCPU_BTF_ID } };
4929 static const struct bpf_reg_types func_ptr_types = { .types = { PTR_TO_FUNC } };
4930 static const struct bpf_reg_types stack_ptr_types = { .types = { PTR_TO_STACK } };
4931 static const struct bpf_reg_types const_str_ptr_types = { .types = { PTR_TO_MAP_VALUE } };
4932 static const struct bpf_reg_types timer_types = { .types = { PTR_TO_MAP_VALUE } };
4934 static const struct bpf_reg_types *compatible_reg_types[__BPF_ARG_TYPE_MAX] = {
4935 [ARG_PTR_TO_MAP_KEY] = &map_key_value_types,
4936 [ARG_PTR_TO_MAP_VALUE] = &map_key_value_types,
4937 [ARG_PTR_TO_UNINIT_MAP_VALUE] = &map_key_value_types,
4938 [ARG_CONST_SIZE] = &scalar_types,
4939 [ARG_CONST_SIZE_OR_ZERO] = &scalar_types,
4940 [ARG_CONST_ALLOC_SIZE_OR_ZERO] = &scalar_types,
4941 [ARG_CONST_MAP_PTR] = &const_map_ptr_types,
4942 [ARG_PTR_TO_CTX] = &context_types,
4943 [ARG_PTR_TO_SOCK_COMMON] = &sock_types,
4945 [ARG_PTR_TO_BTF_ID_SOCK_COMMON] = &btf_id_sock_common_types,
4947 [ARG_PTR_TO_SOCKET] = &fullsock_types,
4948 [ARG_PTR_TO_BTF_ID] = &btf_ptr_types,
4949 [ARG_PTR_TO_SPIN_LOCK] = &spin_lock_types,
4950 [ARG_PTR_TO_MEM] = &mem_types,
4951 [ARG_PTR_TO_UNINIT_MEM] = &mem_types,
4952 [ARG_PTR_TO_ALLOC_MEM] = &alloc_mem_types,
4953 [ARG_PTR_TO_INT] = &int_ptr_types,
4954 [ARG_PTR_TO_LONG] = &int_ptr_types,
4955 [ARG_PTR_TO_PERCPU_BTF_ID] = &percpu_btf_ptr_types,
4956 [ARG_PTR_TO_FUNC] = &func_ptr_types,
4957 [ARG_PTR_TO_STACK] = &stack_ptr_types,
4958 [ARG_PTR_TO_CONST_STR] = &const_str_ptr_types,
4959 [ARG_PTR_TO_TIMER] = &timer_types,
4962 static int check_reg_type(struct bpf_verifier_env *env, u32 regno,
4963 enum bpf_arg_type arg_type,
4964 const u32 *arg_btf_id)
4966 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4967 enum bpf_reg_type expected, type = reg->type;
4968 const struct bpf_reg_types *compatible;
4971 compatible = compatible_reg_types[base_type(arg_type)];
4973 verbose(env, "verifier internal error: unsupported arg type %d\n", arg_type);
4977 /* ARG_PTR_TO_MEM + RDONLY is compatible with PTR_TO_MEM and PTR_TO_MEM + RDONLY,
4978 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM and NOT with PTR_TO_MEM + RDONLY
4980 * Same for MAYBE_NULL:
4982 * ARG_PTR_TO_MEM + MAYBE_NULL is compatible with PTR_TO_MEM and PTR_TO_MEM + MAYBE_NULL,
4983 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM but NOT with PTR_TO_MEM + MAYBE_NULL
4985 * Therefore we fold these flags depending on the arg_type before comparison.
4987 if (arg_type & MEM_RDONLY)
4988 type &= ~MEM_RDONLY;
4989 if (arg_type & PTR_MAYBE_NULL)
4990 type &= ~PTR_MAYBE_NULL;
4992 for (i = 0; i < ARRAY_SIZE(compatible->types); i++) {
4993 expected = compatible->types[i];
4994 if (expected == NOT_INIT)
4997 if (type == expected)
5001 verbose(env, "R%d type=%s expected=", regno, reg_type_str(env, reg->type));
5002 for (j = 0; j + 1 < i; j++)
5003 verbose(env, "%s, ", reg_type_str(env, compatible->types[j]));
5004 verbose(env, "%s\n", reg_type_str(env, compatible->types[j]));
5008 if (reg->type == PTR_TO_BTF_ID) {
5010 if (!compatible->btf_id) {
5011 verbose(env, "verifier internal error: missing arg compatible BTF ID\n");
5014 arg_btf_id = compatible->btf_id;
5017 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
5018 btf_vmlinux, *arg_btf_id)) {
5019 verbose(env, "R%d is of type %s but %s is expected\n",
5020 regno, kernel_type_name(reg->btf, reg->btf_id),
5021 kernel_type_name(btf_vmlinux, *arg_btf_id));
5025 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
5026 verbose(env, "R%d is a pointer to in-kernel struct with non-zero offset\n",
5035 static int check_func_arg(struct bpf_verifier_env *env, u32 arg,
5036 struct bpf_call_arg_meta *meta,
5037 const struct bpf_func_proto *fn)
5039 u32 regno = BPF_REG_1 + arg;
5040 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
5041 enum bpf_arg_type arg_type = fn->arg_type[arg];
5042 enum bpf_reg_type type = reg->type;
5045 if (arg_type == ARG_DONTCARE)
5048 err = check_reg_arg(env, regno, SRC_OP);
5052 if (arg_type == ARG_ANYTHING) {
5053 if (is_pointer_value(env, regno)) {
5054 verbose(env, "R%d leaks addr into helper function\n",
5061 if (type_is_pkt_pointer(type) &&
5062 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
5063 verbose(env, "helper access to the packet is not allowed\n");
5067 if (base_type(arg_type) == ARG_PTR_TO_MAP_VALUE ||
5068 base_type(arg_type) == ARG_PTR_TO_UNINIT_MAP_VALUE) {
5069 err = resolve_map_arg_type(env, meta, &arg_type);
5074 if (register_is_null(reg) && type_may_be_null(arg_type))
5075 /* A NULL register has a SCALAR_VALUE type, so skip
5078 goto skip_type_check;
5080 err = check_reg_type(env, regno, arg_type, fn->arg_btf_id[arg]);
5084 if (type == PTR_TO_CTX) {
5085 err = check_ctx_reg(env, reg, regno);
5091 if (reg->ref_obj_id) {
5092 if (meta->ref_obj_id) {
5093 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
5094 regno, reg->ref_obj_id,
5098 meta->ref_obj_id = reg->ref_obj_id;
5101 if (arg_type == ARG_CONST_MAP_PTR) {
5102 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
5103 if (meta->map_ptr) {
5104 /* Use map_uid (which is unique id of inner map) to reject:
5105 * inner_map1 = bpf_map_lookup_elem(outer_map, key1)
5106 * inner_map2 = bpf_map_lookup_elem(outer_map, key2)
5107 * if (inner_map1 && inner_map2) {
5108 * timer = bpf_map_lookup_elem(inner_map1);
5110 * // mismatch would have been allowed
5111 * bpf_timer_init(timer, inner_map2);
5114 * Comparing map_ptr is enough to distinguish normal and outer maps.
5116 if (meta->map_ptr != reg->map_ptr ||
5117 meta->map_uid != reg->map_uid) {
5119 "timer pointer in R1 map_uid=%d doesn't match map pointer in R2 map_uid=%d\n",
5120 meta->map_uid, reg->map_uid);
5124 meta->map_ptr = reg->map_ptr;
5125 meta->map_uid = reg->map_uid;
5126 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
5127 /* bpf_map_xxx(..., map_ptr, ..., key) call:
5128 * check that [key, key + map->key_size) are within
5129 * stack limits and initialized
5131 if (!meta->map_ptr) {
5132 /* in function declaration map_ptr must come before
5133 * map_key, so that it's verified and known before
5134 * we have to check map_key here. Otherwise it means
5135 * that kernel subsystem misconfigured verifier
5137 verbose(env, "invalid map_ptr to access map->key\n");
5140 err = check_helper_mem_access(env, regno,
5141 meta->map_ptr->key_size, false,
5143 } else if (base_type(arg_type) == ARG_PTR_TO_MAP_VALUE ||
5144 base_type(arg_type) == ARG_PTR_TO_UNINIT_MAP_VALUE) {
5145 if (type_may_be_null(arg_type) && register_is_null(reg))
5148 /* bpf_map_xxx(..., map_ptr, ..., value) call:
5149 * check [value, value + map->value_size) validity
5151 if (!meta->map_ptr) {
5152 /* kernel subsystem misconfigured verifier */
5153 verbose(env, "invalid map_ptr to access map->value\n");
5156 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE);
5157 err = check_helper_mem_access(env, regno,
5158 meta->map_ptr->value_size, false,
5160 } else if (arg_type == ARG_PTR_TO_PERCPU_BTF_ID) {
5162 verbose(env, "Helper has invalid btf_id in R%d\n", regno);
5165 meta->ret_btf = reg->btf;
5166 meta->ret_btf_id = reg->btf_id;
5167 } else if (arg_type == ARG_PTR_TO_SPIN_LOCK) {
5168 if (meta->func_id == BPF_FUNC_spin_lock) {
5169 if (process_spin_lock(env, regno, true))
5171 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
5172 if (process_spin_lock(env, regno, false))
5175 verbose(env, "verifier internal error\n");
5178 } else if (arg_type == ARG_PTR_TO_TIMER) {
5179 if (process_timer_func(env, regno, meta))
5181 } else if (arg_type == ARG_PTR_TO_FUNC) {
5182 meta->subprogno = reg->subprogno;
5183 } else if (arg_type_is_mem_ptr(arg_type)) {
5184 /* The access to this pointer is only checked when we hit the
5185 * next is_mem_size argument below.
5187 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MEM);
5188 } else if (arg_type_is_mem_size(arg_type)) {
5189 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
5191 /* This is used to refine r0 return value bounds for helpers
5192 * that enforce this value as an upper bound on return values.
5193 * See do_refine_retval_range() for helpers that can refine
5194 * the return value. C type of helper is u32 so we pull register
5195 * bound from umax_value however, if negative verifier errors
5196 * out. Only upper bounds can be learned because retval is an
5197 * int type and negative retvals are allowed.
5199 meta->msize_max_value = reg->umax_value;
5201 /* The register is SCALAR_VALUE; the access check
5202 * happens using its boundaries.
5204 if (!tnum_is_const(reg->var_off))
5205 /* For unprivileged variable accesses, disable raw
5206 * mode so that the program is required to
5207 * initialize all the memory that the helper could
5208 * just partially fill up.
5212 if (reg->smin_value < 0) {
5213 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
5218 if (reg->umin_value == 0) {
5219 err = check_helper_mem_access(env, regno - 1, 0,
5226 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
5227 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
5231 err = check_helper_mem_access(env, regno - 1,
5233 zero_size_allowed, meta);
5235 err = mark_chain_precision(env, regno);
5236 } else if (arg_type_is_alloc_size(arg_type)) {
5237 if (!tnum_is_const(reg->var_off)) {
5238 verbose(env, "R%d is not a known constant'\n",
5242 meta->mem_size = reg->var_off.value;
5243 } else if (arg_type_is_int_ptr(arg_type)) {
5244 int size = int_ptr_type_to_size(arg_type);
5246 err = check_helper_mem_access(env, regno, size, false, meta);
5249 err = check_ptr_alignment(env, reg, 0, size, true);
5250 } else if (arg_type == ARG_PTR_TO_CONST_STR) {
5251 struct bpf_map *map = reg->map_ptr;
5256 if (!bpf_map_is_rdonly(map)) {
5257 verbose(env, "R%d does not point to a readonly map'\n", regno);
5261 if (!tnum_is_const(reg->var_off)) {
5262 verbose(env, "R%d is not a constant address'\n", regno);
5266 if (!map->ops->map_direct_value_addr) {
5267 verbose(env, "no direct value access support for this map type\n");
5271 err = check_map_access(env, regno, reg->off,
5272 map->value_size - reg->off, false);
5276 map_off = reg->off + reg->var_off.value;
5277 err = map->ops->map_direct_value_addr(map, &map_addr, map_off);
5279 verbose(env, "direct value access on string failed\n");
5283 str_ptr = (char *)(long)(map_addr);
5284 if (!strnchr(str_ptr + map_off, map->value_size - map_off, 0)) {
5285 verbose(env, "string is not zero-terminated\n");
5293 static bool may_update_sockmap(struct bpf_verifier_env *env, int func_id)
5295 enum bpf_attach_type eatype = env->prog->expected_attach_type;
5296 enum bpf_prog_type type = resolve_prog_type(env->prog);
5298 if (func_id != BPF_FUNC_map_update_elem)
5301 /* It's not possible to get access to a locked struct sock in these
5302 * contexts, so updating is safe.
5305 case BPF_PROG_TYPE_TRACING:
5306 if (eatype == BPF_TRACE_ITER)
5309 case BPF_PROG_TYPE_SOCKET_FILTER:
5310 case BPF_PROG_TYPE_SCHED_CLS:
5311 case BPF_PROG_TYPE_SCHED_ACT:
5312 case BPF_PROG_TYPE_XDP:
5313 case BPF_PROG_TYPE_SK_REUSEPORT:
5314 case BPF_PROG_TYPE_FLOW_DISSECTOR:
5315 case BPF_PROG_TYPE_SK_LOOKUP:
5321 verbose(env, "cannot update sockmap in this context\n");
5325 static bool allow_tail_call_in_subprogs(struct bpf_verifier_env *env)
5327 return env->prog->jit_requested && IS_ENABLED(CONFIG_X86_64);
5330 static int check_map_func_compatibility(struct bpf_verifier_env *env,
5331 struct bpf_map *map, int func_id)
5336 /* We need a two way check, first is from map perspective ... */
5337 switch (map->map_type) {
5338 case BPF_MAP_TYPE_PROG_ARRAY:
5339 if (func_id != BPF_FUNC_tail_call)
5342 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
5343 if (func_id != BPF_FUNC_perf_event_read &&
5344 func_id != BPF_FUNC_perf_event_output &&
5345 func_id != BPF_FUNC_skb_output &&
5346 func_id != BPF_FUNC_perf_event_read_value &&
5347 func_id != BPF_FUNC_xdp_output)
5350 case BPF_MAP_TYPE_RINGBUF:
5351 if (func_id != BPF_FUNC_ringbuf_output &&
5352 func_id != BPF_FUNC_ringbuf_reserve &&
5353 func_id != BPF_FUNC_ringbuf_query)
5356 case BPF_MAP_TYPE_STACK_TRACE:
5357 if (func_id != BPF_FUNC_get_stackid)
5360 case BPF_MAP_TYPE_CGROUP_ARRAY:
5361 if (func_id != BPF_FUNC_skb_under_cgroup &&
5362 func_id != BPF_FUNC_current_task_under_cgroup)
5365 case BPF_MAP_TYPE_CGROUP_STORAGE:
5366 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
5367 if (func_id != BPF_FUNC_get_local_storage)
5370 case BPF_MAP_TYPE_DEVMAP:
5371 case BPF_MAP_TYPE_DEVMAP_HASH:
5372 if (func_id != BPF_FUNC_redirect_map &&
5373 func_id != BPF_FUNC_map_lookup_elem)
5376 /* Restrict bpf side of cpumap and xskmap, open when use-cases
5379 case BPF_MAP_TYPE_CPUMAP:
5380 if (func_id != BPF_FUNC_redirect_map)
5383 case BPF_MAP_TYPE_XSKMAP:
5384 if (func_id != BPF_FUNC_redirect_map &&
5385 func_id != BPF_FUNC_map_lookup_elem)
5388 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
5389 case BPF_MAP_TYPE_HASH_OF_MAPS:
5390 if (func_id != BPF_FUNC_map_lookup_elem)
5393 case BPF_MAP_TYPE_SOCKMAP:
5394 if (func_id != BPF_FUNC_sk_redirect_map &&
5395 func_id != BPF_FUNC_sock_map_update &&
5396 func_id != BPF_FUNC_map_delete_elem &&
5397 func_id != BPF_FUNC_msg_redirect_map &&
5398 func_id != BPF_FUNC_sk_select_reuseport &&
5399 func_id != BPF_FUNC_map_lookup_elem &&
5400 !may_update_sockmap(env, func_id))
5403 case BPF_MAP_TYPE_SOCKHASH:
5404 if (func_id != BPF_FUNC_sk_redirect_hash &&
5405 func_id != BPF_FUNC_sock_hash_update &&
5406 func_id != BPF_FUNC_map_delete_elem &&
5407 func_id != BPF_FUNC_msg_redirect_hash &&
5408 func_id != BPF_FUNC_sk_select_reuseport &&
5409 func_id != BPF_FUNC_map_lookup_elem &&
5410 !may_update_sockmap(env, func_id))
5413 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
5414 if (func_id != BPF_FUNC_sk_select_reuseport)
5417 case BPF_MAP_TYPE_QUEUE:
5418 case BPF_MAP_TYPE_STACK:
5419 if (func_id != BPF_FUNC_map_peek_elem &&
5420 func_id != BPF_FUNC_map_pop_elem &&
5421 func_id != BPF_FUNC_map_push_elem)
5424 case BPF_MAP_TYPE_SK_STORAGE:
5425 if (func_id != BPF_FUNC_sk_storage_get &&
5426 func_id != BPF_FUNC_sk_storage_delete)
5429 case BPF_MAP_TYPE_INODE_STORAGE:
5430 if (func_id != BPF_FUNC_inode_storage_get &&
5431 func_id != BPF_FUNC_inode_storage_delete)
5434 case BPF_MAP_TYPE_TASK_STORAGE:
5435 if (func_id != BPF_FUNC_task_storage_get &&
5436 func_id != BPF_FUNC_task_storage_delete)
5443 /* ... and second from the function itself. */
5445 case BPF_FUNC_tail_call:
5446 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
5448 if (env->subprog_cnt > 1 && !allow_tail_call_in_subprogs(env)) {
5449 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
5453 case BPF_FUNC_perf_event_read:
5454 case BPF_FUNC_perf_event_output:
5455 case BPF_FUNC_perf_event_read_value:
5456 case BPF_FUNC_skb_output:
5457 case BPF_FUNC_xdp_output:
5458 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
5461 case BPF_FUNC_ringbuf_output:
5462 case BPF_FUNC_ringbuf_reserve:
5463 case BPF_FUNC_ringbuf_query:
5464 if (map->map_type != BPF_MAP_TYPE_RINGBUF)
5467 case BPF_FUNC_get_stackid:
5468 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
5471 case BPF_FUNC_current_task_under_cgroup:
5472 case BPF_FUNC_skb_under_cgroup:
5473 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
5476 case BPF_FUNC_redirect_map:
5477 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
5478 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
5479 map->map_type != BPF_MAP_TYPE_CPUMAP &&
5480 map->map_type != BPF_MAP_TYPE_XSKMAP)
5483 case BPF_FUNC_sk_redirect_map:
5484 case BPF_FUNC_msg_redirect_map:
5485 case BPF_FUNC_sock_map_update:
5486 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
5489 case BPF_FUNC_sk_redirect_hash:
5490 case BPF_FUNC_msg_redirect_hash:
5491 case BPF_FUNC_sock_hash_update:
5492 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
5495 case BPF_FUNC_get_local_storage:
5496 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
5497 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
5500 case BPF_FUNC_sk_select_reuseport:
5501 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
5502 map->map_type != BPF_MAP_TYPE_SOCKMAP &&
5503 map->map_type != BPF_MAP_TYPE_SOCKHASH)
5506 case BPF_FUNC_map_peek_elem:
5507 case BPF_FUNC_map_pop_elem:
5508 case BPF_FUNC_map_push_elem:
5509 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
5510 map->map_type != BPF_MAP_TYPE_STACK)
5513 case BPF_FUNC_sk_storage_get:
5514 case BPF_FUNC_sk_storage_delete:
5515 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
5518 case BPF_FUNC_inode_storage_get:
5519 case BPF_FUNC_inode_storage_delete:
5520 if (map->map_type != BPF_MAP_TYPE_INODE_STORAGE)
5523 case BPF_FUNC_task_storage_get:
5524 case BPF_FUNC_task_storage_delete:
5525 if (map->map_type != BPF_MAP_TYPE_TASK_STORAGE)
5534 verbose(env, "cannot pass map_type %d into func %s#%d\n",
5535 map->map_type, func_id_name(func_id), func_id);
5539 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
5543 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
5545 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
5547 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
5549 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
5551 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
5554 /* We only support one arg being in raw mode at the moment,
5555 * which is sufficient for the helper functions we have
5561 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
5562 enum bpf_arg_type arg_next)
5564 return (arg_type_is_mem_ptr(arg_curr) &&
5565 !arg_type_is_mem_size(arg_next)) ||
5566 (!arg_type_is_mem_ptr(arg_curr) &&
5567 arg_type_is_mem_size(arg_next));
5570 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
5572 /* bpf_xxx(..., buf, len) call will access 'len'
5573 * bytes from memory 'buf'. Both arg types need
5574 * to be paired, so make sure there's no buggy
5575 * helper function specification.
5577 if (arg_type_is_mem_size(fn->arg1_type) ||
5578 arg_type_is_mem_ptr(fn->arg5_type) ||
5579 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
5580 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
5581 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
5582 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
5588 static bool check_refcount_ok(const struct bpf_func_proto *fn, int func_id)
5592 if (arg_type_may_be_refcounted(fn->arg1_type))
5594 if (arg_type_may_be_refcounted(fn->arg2_type))
5596 if (arg_type_may_be_refcounted(fn->arg3_type))
5598 if (arg_type_may_be_refcounted(fn->arg4_type))
5600 if (arg_type_may_be_refcounted(fn->arg5_type))
5603 /* A reference acquiring function cannot acquire
5604 * another refcounted ptr.
5606 if (may_be_acquire_function(func_id) && count)
5609 /* We only support one arg being unreferenced at the moment,
5610 * which is sufficient for the helper functions we have right now.
5615 static bool check_btf_id_ok(const struct bpf_func_proto *fn)
5619 for (i = 0; i < ARRAY_SIZE(fn->arg_type); i++) {
5620 if (fn->arg_type[i] == ARG_PTR_TO_BTF_ID && !fn->arg_btf_id[i])
5623 if (fn->arg_type[i] != ARG_PTR_TO_BTF_ID && fn->arg_btf_id[i])
5630 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
5632 return check_raw_mode_ok(fn) &&
5633 check_arg_pair_ok(fn) &&
5634 check_btf_id_ok(fn) &&
5635 check_refcount_ok(fn, func_id) ? 0 : -EINVAL;
5638 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
5639 * are now invalid, so turn them into unknown SCALAR_VALUE.
5641 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
5642 struct bpf_func_state *state)
5644 struct bpf_reg_state *regs = state->regs, *reg;
5647 for (i = 0; i < MAX_BPF_REG; i++)
5648 if (reg_is_pkt_pointer_any(®s[i]))
5649 mark_reg_unknown(env, regs, i);
5651 bpf_for_each_spilled_reg(i, state, reg) {
5654 if (reg_is_pkt_pointer_any(reg))
5655 __mark_reg_unknown(env, reg);
5659 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
5661 struct bpf_verifier_state *vstate = env->cur_state;
5664 for (i = 0; i <= vstate->curframe; i++)
5665 __clear_all_pkt_pointers(env, vstate->frame[i]);
5670 BEYOND_PKT_END = -2,
5673 static void mark_pkt_end(struct bpf_verifier_state *vstate, int regn, bool range_open)
5675 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5676 struct bpf_reg_state *reg = &state->regs[regn];
5678 if (reg->type != PTR_TO_PACKET)
5679 /* PTR_TO_PACKET_META is not supported yet */
5682 /* The 'reg' is pkt > pkt_end or pkt >= pkt_end.
5683 * How far beyond pkt_end it goes is unknown.
5684 * if (!range_open) it's the case of pkt >= pkt_end
5685 * if (range_open) it's the case of pkt > pkt_end
5686 * hence this pointer is at least 1 byte bigger than pkt_end
5689 reg->range = BEYOND_PKT_END;
5691 reg->range = AT_PKT_END;
5694 static void release_reg_references(struct bpf_verifier_env *env,
5695 struct bpf_func_state *state,
5698 struct bpf_reg_state *regs = state->regs, *reg;
5701 for (i = 0; i < MAX_BPF_REG; i++)
5702 if (regs[i].ref_obj_id == ref_obj_id)
5703 mark_reg_unknown(env, regs, i);
5705 bpf_for_each_spilled_reg(i, state, reg) {
5708 if (reg->ref_obj_id == ref_obj_id)
5709 __mark_reg_unknown(env, reg);
5713 /* The pointer with the specified id has released its reference to kernel
5714 * resources. Identify all copies of the same pointer and clear the reference.
5716 static int release_reference(struct bpf_verifier_env *env,
5719 struct bpf_verifier_state *vstate = env->cur_state;
5723 err = release_reference_state(cur_func(env), ref_obj_id);
5727 for (i = 0; i <= vstate->curframe; i++)
5728 release_reg_references(env, vstate->frame[i], ref_obj_id);
5733 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
5734 struct bpf_reg_state *regs)
5738 /* after the call registers r0 - r5 were scratched */
5739 for (i = 0; i < CALLER_SAVED_REGS; i++) {
5740 mark_reg_not_init(env, regs, caller_saved[i]);
5741 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
5745 typedef int (*set_callee_state_fn)(struct bpf_verifier_env *env,
5746 struct bpf_func_state *caller,
5747 struct bpf_func_state *callee,
5750 static int __check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
5751 int *insn_idx, int subprog,
5752 set_callee_state_fn set_callee_state_cb)
5754 struct bpf_verifier_state *state = env->cur_state;
5755 struct bpf_func_info_aux *func_info_aux;
5756 struct bpf_func_state *caller, *callee;
5758 bool is_global = false;
5760 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
5761 verbose(env, "the call stack of %d frames is too deep\n",
5762 state->curframe + 2);
5766 caller = state->frame[state->curframe];
5767 if (state->frame[state->curframe + 1]) {
5768 verbose(env, "verifier bug. Frame %d already allocated\n",
5769 state->curframe + 1);
5773 func_info_aux = env->prog->aux->func_info_aux;
5775 is_global = func_info_aux[subprog].linkage == BTF_FUNC_GLOBAL;
5776 err = btf_check_subprog_arg_match(env, subprog, caller->regs);
5781 verbose(env, "Caller passes invalid args into func#%d\n",
5785 if (env->log.level & BPF_LOG_LEVEL)
5787 "Func#%d is global and valid. Skipping.\n",
5789 clear_caller_saved_regs(env, caller->regs);
5791 /* All global functions return a 64-bit SCALAR_VALUE */
5792 mark_reg_unknown(env, caller->regs, BPF_REG_0);
5793 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
5795 /* continue with next insn after call */
5800 if (insn->code == (BPF_JMP | BPF_CALL) &&
5801 insn->src_reg == 0 &&
5802 insn->imm == BPF_FUNC_timer_set_callback) {
5803 struct bpf_verifier_state *async_cb;
5805 /* there is no real recursion here. timer callbacks are async */
5806 env->subprog_info[subprog].is_async_cb = true;
5807 async_cb = push_async_cb(env, env->subprog_info[subprog].start,
5808 *insn_idx, subprog);
5811 callee = async_cb->frame[0];
5812 callee->async_entry_cnt = caller->async_entry_cnt + 1;
5814 /* Convert bpf_timer_set_callback() args into timer callback args */
5815 err = set_callee_state_cb(env, caller, callee, *insn_idx);
5819 clear_caller_saved_regs(env, caller->regs);
5820 mark_reg_unknown(env, caller->regs, BPF_REG_0);
5821 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
5822 /* continue with next insn after call */
5826 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
5829 state->frame[state->curframe + 1] = callee;
5831 /* callee cannot access r0, r6 - r9 for reading and has to write
5832 * into its own stack before reading from it.
5833 * callee can read/write into caller's stack
5835 init_func_state(env, callee,
5836 /* remember the callsite, it will be used by bpf_exit */
5837 *insn_idx /* callsite */,
5838 state->curframe + 1 /* frameno within this callchain */,
5839 subprog /* subprog number within this prog */);
5841 /* Transfer references to the callee */
5842 err = copy_reference_state(callee, caller);
5846 err = set_callee_state_cb(env, caller, callee, *insn_idx);
5850 clear_caller_saved_regs(env, caller->regs);
5852 /* only increment it after check_reg_arg() finished */
5855 /* and go analyze first insn of the callee */
5856 *insn_idx = env->subprog_info[subprog].start - 1;
5858 if (env->log.level & BPF_LOG_LEVEL) {
5859 verbose(env, "caller:\n");
5860 print_verifier_state(env, caller);
5861 verbose(env, "callee:\n");
5862 print_verifier_state(env, callee);
5867 int map_set_for_each_callback_args(struct bpf_verifier_env *env,
5868 struct bpf_func_state *caller,
5869 struct bpf_func_state *callee)
5871 /* bpf_for_each_map_elem(struct bpf_map *map, void *callback_fn,
5872 * void *callback_ctx, u64 flags);
5873 * callback_fn(struct bpf_map *map, void *key, void *value,
5874 * void *callback_ctx);
5876 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
5878 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
5879 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
5880 callee->regs[BPF_REG_2].map_ptr = caller->regs[BPF_REG_1].map_ptr;
5882 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
5883 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
5884 callee->regs[BPF_REG_3].map_ptr = caller->regs[BPF_REG_1].map_ptr;
5886 /* pointer to stack or null */
5887 callee->regs[BPF_REG_4] = caller->regs[BPF_REG_3];
5890 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
5894 static int set_callee_state(struct bpf_verifier_env *env,
5895 struct bpf_func_state *caller,
5896 struct bpf_func_state *callee, int insn_idx)
5900 /* copy r1 - r5 args that callee can access. The copy includes parent
5901 * pointers, which connects us up to the liveness chain
5903 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
5904 callee->regs[i] = caller->regs[i];
5908 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
5911 int subprog, target_insn;
5913 target_insn = *insn_idx + insn->imm + 1;
5914 subprog = find_subprog(env, target_insn);
5916 verbose(env, "verifier bug. No program starts at insn %d\n",
5921 return __check_func_call(env, insn, insn_idx, subprog, set_callee_state);
5924 static int set_map_elem_callback_state(struct bpf_verifier_env *env,
5925 struct bpf_func_state *caller,
5926 struct bpf_func_state *callee,
5929 struct bpf_insn_aux_data *insn_aux = &env->insn_aux_data[insn_idx];
5930 struct bpf_map *map;
5933 if (bpf_map_ptr_poisoned(insn_aux)) {
5934 verbose(env, "tail_call abusing map_ptr\n");
5938 map = BPF_MAP_PTR(insn_aux->map_ptr_state);
5939 if (!map->ops->map_set_for_each_callback_args ||
5940 !map->ops->map_for_each_callback) {
5941 verbose(env, "callback function not allowed for map\n");
5945 err = map->ops->map_set_for_each_callback_args(env, caller, callee);
5949 callee->in_callback_fn = true;
5953 static int set_timer_callback_state(struct bpf_verifier_env *env,
5954 struct bpf_func_state *caller,
5955 struct bpf_func_state *callee,
5958 struct bpf_map *map_ptr = caller->regs[BPF_REG_1].map_ptr;
5960 /* bpf_timer_set_callback(struct bpf_timer *timer, void *callback_fn);
5961 * callback_fn(struct bpf_map *map, void *key, void *value);
5963 callee->regs[BPF_REG_1].type = CONST_PTR_TO_MAP;
5964 __mark_reg_known_zero(&callee->regs[BPF_REG_1]);
5965 callee->regs[BPF_REG_1].map_ptr = map_ptr;
5967 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
5968 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
5969 callee->regs[BPF_REG_2].map_ptr = map_ptr;
5971 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
5972 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
5973 callee->regs[BPF_REG_3].map_ptr = map_ptr;
5976 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
5977 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
5978 callee->in_async_callback_fn = true;
5982 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
5984 struct bpf_verifier_state *state = env->cur_state;
5985 struct bpf_func_state *caller, *callee;
5986 struct bpf_reg_state *r0;
5989 callee = state->frame[state->curframe];
5990 r0 = &callee->regs[BPF_REG_0];
5991 if (r0->type == PTR_TO_STACK) {
5992 /* technically it's ok to return caller's stack pointer
5993 * (or caller's caller's pointer) back to the caller,
5994 * since these pointers are valid. Only current stack
5995 * pointer will be invalid as soon as function exits,
5996 * but let's be conservative
5998 verbose(env, "cannot return stack pointer to the caller\n");
6003 caller = state->frame[state->curframe];
6004 if (callee->in_callback_fn) {
6005 /* enforce R0 return value range [0, 1]. */
6006 struct tnum range = tnum_range(0, 1);
6008 if (r0->type != SCALAR_VALUE) {
6009 verbose(env, "R0 not a scalar value\n");
6012 if (!tnum_in(range, r0->var_off)) {
6013 verbose_invalid_scalar(env, r0, &range, "callback return", "R0");
6017 /* return to the caller whatever r0 had in the callee */
6018 caller->regs[BPF_REG_0] = *r0;
6021 /* Transfer references to the caller */
6022 err = copy_reference_state(caller, callee);
6026 *insn_idx = callee->callsite + 1;
6027 if (env->log.level & BPF_LOG_LEVEL) {
6028 verbose(env, "returning from callee:\n");
6029 print_verifier_state(env, callee);
6030 verbose(env, "to caller at %d:\n", *insn_idx);
6031 print_verifier_state(env, caller);
6033 /* clear everything in the callee */
6034 free_func_state(callee);
6035 state->frame[state->curframe + 1] = NULL;
6039 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
6041 struct bpf_call_arg_meta *meta)
6043 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
6045 if (ret_type != RET_INTEGER ||
6046 (func_id != BPF_FUNC_get_stack &&
6047 func_id != BPF_FUNC_get_task_stack &&
6048 func_id != BPF_FUNC_probe_read_str &&
6049 func_id != BPF_FUNC_probe_read_kernel_str &&
6050 func_id != BPF_FUNC_probe_read_user_str))
6053 ret_reg->smax_value = meta->msize_max_value;
6054 ret_reg->s32_max_value = meta->msize_max_value;
6055 ret_reg->smin_value = -MAX_ERRNO;
6056 ret_reg->s32_min_value = -MAX_ERRNO;
6057 __reg_deduce_bounds(ret_reg);
6058 __reg_bound_offset(ret_reg);
6059 __update_reg_bounds(ret_reg);
6063 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
6064 int func_id, int insn_idx)
6066 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
6067 struct bpf_map *map = meta->map_ptr;
6069 if (func_id != BPF_FUNC_tail_call &&
6070 func_id != BPF_FUNC_map_lookup_elem &&
6071 func_id != BPF_FUNC_map_update_elem &&
6072 func_id != BPF_FUNC_map_delete_elem &&
6073 func_id != BPF_FUNC_map_push_elem &&
6074 func_id != BPF_FUNC_map_pop_elem &&
6075 func_id != BPF_FUNC_map_peek_elem &&
6076 func_id != BPF_FUNC_for_each_map_elem &&
6077 func_id != BPF_FUNC_redirect_map)
6081 verbose(env, "kernel subsystem misconfigured verifier\n");
6085 /* In case of read-only, some additional restrictions
6086 * need to be applied in order to prevent altering the
6087 * state of the map from program side.
6089 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
6090 (func_id == BPF_FUNC_map_delete_elem ||
6091 func_id == BPF_FUNC_map_update_elem ||
6092 func_id == BPF_FUNC_map_push_elem ||
6093 func_id == BPF_FUNC_map_pop_elem)) {
6094 verbose(env, "write into map forbidden\n");
6098 if (!BPF_MAP_PTR(aux->map_ptr_state))
6099 bpf_map_ptr_store(aux, meta->map_ptr,
6100 !meta->map_ptr->bypass_spec_v1);
6101 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
6102 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
6103 !meta->map_ptr->bypass_spec_v1);
6108 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
6109 int func_id, int insn_idx)
6111 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
6112 struct bpf_reg_state *regs = cur_regs(env), *reg;
6113 struct bpf_map *map = meta->map_ptr;
6118 if (func_id != BPF_FUNC_tail_call)
6120 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
6121 verbose(env, "kernel subsystem misconfigured verifier\n");
6125 range = tnum_range(0, map->max_entries - 1);
6126 reg = ®s[BPF_REG_3];
6128 if (!register_is_const(reg) || !tnum_in(range, reg->var_off)) {
6129 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
6133 err = mark_chain_precision(env, BPF_REG_3);
6137 val = reg->var_off.value;
6138 if (bpf_map_key_unseen(aux))
6139 bpf_map_key_store(aux, val);
6140 else if (!bpf_map_key_poisoned(aux) &&
6141 bpf_map_key_immediate(aux) != val)
6142 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
6146 static int check_reference_leak(struct bpf_verifier_env *env)
6148 struct bpf_func_state *state = cur_func(env);
6151 for (i = 0; i < state->acquired_refs; i++) {
6152 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
6153 state->refs[i].id, state->refs[i].insn_idx);
6155 return state->acquired_refs ? -EINVAL : 0;
6158 static int check_bpf_snprintf_call(struct bpf_verifier_env *env,
6159 struct bpf_reg_state *regs)
6161 struct bpf_reg_state *fmt_reg = ®s[BPF_REG_3];
6162 struct bpf_reg_state *data_len_reg = ®s[BPF_REG_5];
6163 struct bpf_map *fmt_map = fmt_reg->map_ptr;
6164 int err, fmt_map_off, num_args;
6168 /* data must be an array of u64 */
6169 if (data_len_reg->var_off.value % 8)
6171 num_args = data_len_reg->var_off.value / 8;
6173 /* fmt being ARG_PTR_TO_CONST_STR guarantees that var_off is const
6174 * and map_direct_value_addr is set.
6176 fmt_map_off = fmt_reg->off + fmt_reg->var_off.value;
6177 err = fmt_map->ops->map_direct_value_addr(fmt_map, &fmt_addr,
6180 verbose(env, "verifier bug\n");
6183 fmt = (char *)(long)fmt_addr + fmt_map_off;
6185 /* We are also guaranteed that fmt+fmt_map_off is NULL terminated, we
6186 * can focus on validating the format specifiers.
6188 err = bpf_bprintf_prepare(fmt, UINT_MAX, NULL, NULL, num_args);
6190 verbose(env, "Invalid format string\n");
6195 static int check_get_func_ip(struct bpf_verifier_env *env)
6197 enum bpf_attach_type eatype = env->prog->expected_attach_type;
6198 enum bpf_prog_type type = resolve_prog_type(env->prog);
6199 int func_id = BPF_FUNC_get_func_ip;
6201 if (type == BPF_PROG_TYPE_TRACING) {
6202 if (eatype != BPF_TRACE_FENTRY && eatype != BPF_TRACE_FEXIT &&
6203 eatype != BPF_MODIFY_RETURN) {
6204 verbose(env, "func %s#%d supported only for fentry/fexit/fmod_ret programs\n",
6205 func_id_name(func_id), func_id);
6209 } else if (type == BPF_PROG_TYPE_KPROBE) {
6213 verbose(env, "func %s#%d not supported for program type %d\n",
6214 func_id_name(func_id), func_id, type);
6218 static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
6221 const struct bpf_func_proto *fn = NULL;
6222 enum bpf_return_type ret_type;
6223 enum bpf_type_flag ret_flag;
6224 struct bpf_reg_state *regs;
6225 struct bpf_call_arg_meta meta;
6226 int insn_idx = *insn_idx_p;
6228 int i, err, func_id;
6230 /* find function prototype */
6231 func_id = insn->imm;
6232 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
6233 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
6238 if (env->ops->get_func_proto)
6239 fn = env->ops->get_func_proto(func_id, env->prog);
6241 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
6246 /* eBPF programs must be GPL compatible to use GPL-ed functions */
6247 if (!env->prog->gpl_compatible && fn->gpl_only) {
6248 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
6252 if (fn->allowed && !fn->allowed(env->prog)) {
6253 verbose(env, "helper call is not allowed in probe\n");
6257 /* With LD_ABS/IND some JITs save/restore skb from r1. */
6258 changes_data = bpf_helper_changes_pkt_data(fn->func);
6259 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
6260 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
6261 func_id_name(func_id), func_id);
6265 memset(&meta, 0, sizeof(meta));
6266 meta.pkt_access = fn->pkt_access;
6268 err = check_func_proto(fn, func_id);
6270 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
6271 func_id_name(func_id), func_id);
6275 meta.func_id = func_id;
6277 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++) {
6278 err = check_func_arg(env, i, &meta, fn);
6283 err = record_func_map(env, &meta, func_id, insn_idx);
6287 err = record_func_key(env, &meta, func_id, insn_idx);
6291 /* Mark slots with STACK_MISC in case of raw mode, stack offset
6292 * is inferred from register state.
6294 for (i = 0; i < meta.access_size; i++) {
6295 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
6296 BPF_WRITE, -1, false);
6301 if (func_id == BPF_FUNC_tail_call) {
6302 err = check_reference_leak(env);
6304 verbose(env, "tail_call would lead to reference leak\n");
6307 } else if (is_release_function(func_id)) {
6308 err = release_reference(env, meta.ref_obj_id);
6310 verbose(env, "func %s#%d reference has not been acquired before\n",
6311 func_id_name(func_id), func_id);
6316 regs = cur_regs(env);
6318 /* check that flags argument in get_local_storage(map, flags) is 0,
6319 * this is required because get_local_storage() can't return an error.
6321 if (func_id == BPF_FUNC_get_local_storage &&
6322 !register_is_null(®s[BPF_REG_2])) {
6323 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
6327 if (func_id == BPF_FUNC_for_each_map_elem) {
6328 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
6329 set_map_elem_callback_state);
6334 if (func_id == BPF_FUNC_timer_set_callback) {
6335 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
6336 set_timer_callback_state);
6341 if (func_id == BPF_FUNC_snprintf) {
6342 err = check_bpf_snprintf_call(env, regs);
6347 /* reset caller saved regs */
6348 for (i = 0; i < CALLER_SAVED_REGS; i++) {
6349 mark_reg_not_init(env, regs, caller_saved[i]);
6350 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
6353 /* helper call returns 64-bit value. */
6354 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
6356 /* update return register (already marked as written above) */
6357 ret_type = fn->ret_type;
6358 ret_flag = type_flag(fn->ret_type);
6359 if (ret_type == RET_INTEGER) {
6360 /* sets type to SCALAR_VALUE */
6361 mark_reg_unknown(env, regs, BPF_REG_0);
6362 } else if (ret_type == RET_VOID) {
6363 regs[BPF_REG_0].type = NOT_INIT;
6364 } else if (base_type(ret_type) == RET_PTR_TO_MAP_VALUE) {
6365 /* There is no offset yet applied, variable or fixed */
6366 mark_reg_known_zero(env, regs, BPF_REG_0);
6367 /* remember map_ptr, so that check_map_access()
6368 * can check 'value_size' boundary of memory access
6369 * to map element returned from bpf_map_lookup_elem()
6371 if (meta.map_ptr == NULL) {
6373 "kernel subsystem misconfigured verifier\n");
6376 regs[BPF_REG_0].map_ptr = meta.map_ptr;
6377 regs[BPF_REG_0].map_uid = meta.map_uid;
6378 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE | ret_flag;
6379 if (!type_may_be_null(ret_type) &&
6380 map_value_has_spin_lock(meta.map_ptr)) {
6381 regs[BPF_REG_0].id = ++env->id_gen;
6383 } else if (base_type(ret_type) == RET_PTR_TO_SOCKET) {
6384 mark_reg_known_zero(env, regs, BPF_REG_0);
6385 regs[BPF_REG_0].type = PTR_TO_SOCKET | ret_flag;
6386 } else if (base_type(ret_type) == RET_PTR_TO_SOCK_COMMON) {
6387 mark_reg_known_zero(env, regs, BPF_REG_0);
6388 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON | ret_flag;
6389 } else if (base_type(ret_type) == RET_PTR_TO_TCP_SOCK) {
6390 mark_reg_known_zero(env, regs, BPF_REG_0);
6391 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK | ret_flag;
6392 } else if (base_type(ret_type) == RET_PTR_TO_ALLOC_MEM) {
6393 mark_reg_known_zero(env, regs, BPF_REG_0);
6394 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
6395 regs[BPF_REG_0].mem_size = meta.mem_size;
6396 } else if (base_type(ret_type) == RET_PTR_TO_MEM_OR_BTF_ID) {
6397 const struct btf_type *t;
6399 mark_reg_known_zero(env, regs, BPF_REG_0);
6400 t = btf_type_skip_modifiers(meta.ret_btf, meta.ret_btf_id, NULL);
6401 if (!btf_type_is_struct(t)) {
6403 const struct btf_type *ret;
6406 /* resolve the type size of ksym. */
6407 ret = btf_resolve_size(meta.ret_btf, t, &tsize);
6409 tname = btf_name_by_offset(meta.ret_btf, t->name_off);
6410 verbose(env, "unable to resolve the size of type '%s': %ld\n",
6411 tname, PTR_ERR(ret));
6414 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
6415 regs[BPF_REG_0].mem_size = tsize;
6417 /* MEM_RDONLY may be carried from ret_flag, but it
6418 * doesn't apply on PTR_TO_BTF_ID. Fold it, otherwise
6419 * it will confuse the check of PTR_TO_BTF_ID in
6420 * check_mem_access().
6422 ret_flag &= ~MEM_RDONLY;
6424 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
6425 regs[BPF_REG_0].btf = meta.ret_btf;
6426 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
6428 } else if (base_type(ret_type) == RET_PTR_TO_BTF_ID) {
6431 mark_reg_known_zero(env, regs, BPF_REG_0);
6432 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
6433 ret_btf_id = *fn->ret_btf_id;
6434 if (ret_btf_id == 0) {
6435 verbose(env, "invalid return type %u of func %s#%d\n",
6436 base_type(ret_type), func_id_name(func_id),
6440 /* current BPF helper definitions are only coming from
6441 * built-in code with type IDs from vmlinux BTF
6443 regs[BPF_REG_0].btf = btf_vmlinux;
6444 regs[BPF_REG_0].btf_id = ret_btf_id;
6446 verbose(env, "unknown return type %u of func %s#%d\n",
6447 base_type(ret_type), func_id_name(func_id), func_id);
6451 if (type_may_be_null(regs[BPF_REG_0].type))
6452 regs[BPF_REG_0].id = ++env->id_gen;
6454 if (is_ptr_cast_function(func_id)) {
6455 /* For release_reference() */
6456 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
6457 } else if (is_acquire_function(func_id, meta.map_ptr)) {
6458 int id = acquire_reference_state(env, insn_idx);
6462 /* For mark_ptr_or_null_reg() */
6463 regs[BPF_REG_0].id = id;
6464 /* For release_reference() */
6465 regs[BPF_REG_0].ref_obj_id = id;
6468 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
6470 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
6474 if ((func_id == BPF_FUNC_get_stack ||
6475 func_id == BPF_FUNC_get_task_stack) &&
6476 !env->prog->has_callchain_buf) {
6477 const char *err_str;
6479 #ifdef CONFIG_PERF_EVENTS
6480 err = get_callchain_buffers(sysctl_perf_event_max_stack);
6481 err_str = "cannot get callchain buffer for func %s#%d\n";
6484 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
6487 verbose(env, err_str, func_id_name(func_id), func_id);
6491 env->prog->has_callchain_buf = true;
6494 if (func_id == BPF_FUNC_get_stackid || func_id == BPF_FUNC_get_stack)
6495 env->prog->call_get_stack = true;
6497 if (func_id == BPF_FUNC_get_func_ip) {
6498 if (check_get_func_ip(env))
6500 env->prog->call_get_func_ip = true;
6504 clear_all_pkt_pointers(env);
6508 /* mark_btf_func_reg_size() is used when the reg size is determined by
6509 * the BTF func_proto's return value size and argument.
6511 static void mark_btf_func_reg_size(struct bpf_verifier_env *env, u32 regno,
6514 struct bpf_reg_state *reg = &cur_regs(env)[regno];
6516 if (regno == BPF_REG_0) {
6517 /* Function return value */
6518 reg->live |= REG_LIVE_WRITTEN;
6519 reg->subreg_def = reg_size == sizeof(u64) ?
6520 DEF_NOT_SUBREG : env->insn_idx + 1;
6522 /* Function argument */
6523 if (reg_size == sizeof(u64)) {
6524 mark_insn_zext(env, reg);
6525 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
6527 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ32);
6532 static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn)
6534 const struct btf_type *t, *func, *func_proto, *ptr_type;
6535 struct bpf_reg_state *regs = cur_regs(env);
6536 const char *func_name, *ptr_type_name;
6537 u32 i, nargs, func_id, ptr_type_id;
6538 const struct btf_param *args;
6541 func_id = insn->imm;
6542 func = btf_type_by_id(btf_vmlinux, func_id);
6543 func_name = btf_name_by_offset(btf_vmlinux, func->name_off);
6544 func_proto = btf_type_by_id(btf_vmlinux, func->type);
6546 if (!env->ops->check_kfunc_call ||
6547 !env->ops->check_kfunc_call(func_id)) {
6548 verbose(env, "calling kernel function %s is not allowed\n",
6553 /* Check the arguments */
6554 err = btf_check_kfunc_arg_match(env, btf_vmlinux, func_id, regs);
6558 for (i = 0; i < CALLER_SAVED_REGS; i++)
6559 mark_reg_not_init(env, regs, caller_saved[i]);
6561 /* Check return type */
6562 t = btf_type_skip_modifiers(btf_vmlinux, func_proto->type, NULL);
6563 if (btf_type_is_scalar(t)) {
6564 mark_reg_unknown(env, regs, BPF_REG_0);
6565 mark_btf_func_reg_size(env, BPF_REG_0, t->size);
6566 } else if (btf_type_is_ptr(t)) {
6567 ptr_type = btf_type_skip_modifiers(btf_vmlinux, t->type,
6569 if (!btf_type_is_struct(ptr_type)) {
6570 ptr_type_name = btf_name_by_offset(btf_vmlinux,
6571 ptr_type->name_off);
6572 verbose(env, "kernel function %s returns pointer type %s %s is not supported\n",
6573 func_name, btf_type_str(ptr_type),
6577 mark_reg_known_zero(env, regs, BPF_REG_0);
6578 regs[BPF_REG_0].btf = btf_vmlinux;
6579 regs[BPF_REG_0].type = PTR_TO_BTF_ID;
6580 regs[BPF_REG_0].btf_id = ptr_type_id;
6581 mark_btf_func_reg_size(env, BPF_REG_0, sizeof(void *));
6582 } /* else { add_kfunc_call() ensures it is btf_type_is_void(t) } */
6584 nargs = btf_type_vlen(func_proto);
6585 args = (const struct btf_param *)(func_proto + 1);
6586 for (i = 0; i < nargs; i++) {
6589 t = btf_type_skip_modifiers(btf_vmlinux, args[i].type, NULL);
6590 if (btf_type_is_ptr(t))
6591 mark_btf_func_reg_size(env, regno, sizeof(void *));
6593 /* scalar. ensured by btf_check_kfunc_arg_match() */
6594 mark_btf_func_reg_size(env, regno, t->size);
6600 static bool signed_add_overflows(s64 a, s64 b)
6602 /* Do the add in u64, where overflow is well-defined */
6603 s64 res = (s64)((u64)a + (u64)b);
6610 static bool signed_add32_overflows(s32 a, s32 b)
6612 /* Do the add in u32, where overflow is well-defined */
6613 s32 res = (s32)((u32)a + (u32)b);
6620 static bool signed_sub_overflows(s64 a, s64 b)
6622 /* Do the sub in u64, where overflow is well-defined */
6623 s64 res = (s64)((u64)a - (u64)b);
6630 static bool signed_sub32_overflows(s32 a, s32 b)
6632 /* Do the sub in u32, where overflow is well-defined */
6633 s32 res = (s32)((u32)a - (u32)b);
6640 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
6641 const struct bpf_reg_state *reg,
6642 enum bpf_reg_type type)
6644 bool known = tnum_is_const(reg->var_off);
6645 s64 val = reg->var_off.value;
6646 s64 smin = reg->smin_value;
6648 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
6649 verbose(env, "math between %s pointer and %lld is not allowed\n",
6650 reg_type_str(env, type), val);
6654 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
6655 verbose(env, "%s pointer offset %d is not allowed\n",
6656 reg_type_str(env, type), reg->off);
6660 if (smin == S64_MIN) {
6661 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
6662 reg_type_str(env, type));
6666 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
6667 verbose(env, "value %lld makes %s pointer be out of bounds\n",
6668 smin, reg_type_str(env, type));
6675 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
6677 return &env->insn_aux_data[env->insn_idx];
6688 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
6689 u32 *alu_limit, bool mask_to_left)
6691 u32 max = 0, ptr_limit = 0;
6693 switch (ptr_reg->type) {
6695 /* Offset 0 is out-of-bounds, but acceptable start for the
6696 * left direction, see BPF_REG_FP. Also, unknown scalar
6697 * offset where we would need to deal with min/max bounds is
6698 * currently prohibited for unprivileged.
6700 max = MAX_BPF_STACK + mask_to_left;
6701 ptr_limit = -(ptr_reg->var_off.value + ptr_reg->off);
6703 case PTR_TO_MAP_VALUE:
6704 max = ptr_reg->map_ptr->value_size;
6705 ptr_limit = (mask_to_left ?
6706 ptr_reg->smin_value :
6707 ptr_reg->umax_value) + ptr_reg->off;
6713 if (ptr_limit >= max)
6714 return REASON_LIMIT;
6715 *alu_limit = ptr_limit;
6719 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
6720 const struct bpf_insn *insn)
6722 return env->bypass_spec_v1 || BPF_SRC(insn->code) == BPF_K;
6725 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
6726 u32 alu_state, u32 alu_limit)
6728 /* If we arrived here from different branches with different
6729 * state or limits to sanitize, then this won't work.
6731 if (aux->alu_state &&
6732 (aux->alu_state != alu_state ||
6733 aux->alu_limit != alu_limit))
6734 return REASON_PATHS;
6736 /* Corresponding fixup done in do_misc_fixups(). */
6737 aux->alu_state = alu_state;
6738 aux->alu_limit = alu_limit;
6742 static int sanitize_val_alu(struct bpf_verifier_env *env,
6743 struct bpf_insn *insn)
6745 struct bpf_insn_aux_data *aux = cur_aux(env);
6747 if (can_skip_alu_sanitation(env, insn))
6750 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
6753 static bool sanitize_needed(u8 opcode)
6755 return opcode == BPF_ADD || opcode == BPF_SUB;
6758 struct bpf_sanitize_info {
6759 struct bpf_insn_aux_data aux;
6763 static struct bpf_verifier_state *
6764 sanitize_speculative_path(struct bpf_verifier_env *env,
6765 const struct bpf_insn *insn,
6766 u32 next_idx, u32 curr_idx)
6768 struct bpf_verifier_state *branch;
6769 struct bpf_reg_state *regs;
6771 branch = push_stack(env, next_idx, curr_idx, true);
6772 if (branch && insn) {
6773 regs = branch->frame[branch->curframe]->regs;
6774 if (BPF_SRC(insn->code) == BPF_K) {
6775 mark_reg_unknown(env, regs, insn->dst_reg);
6776 } else if (BPF_SRC(insn->code) == BPF_X) {
6777 mark_reg_unknown(env, regs, insn->dst_reg);
6778 mark_reg_unknown(env, regs, insn->src_reg);
6784 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
6785 struct bpf_insn *insn,
6786 const struct bpf_reg_state *ptr_reg,
6787 const struct bpf_reg_state *off_reg,
6788 struct bpf_reg_state *dst_reg,
6789 struct bpf_sanitize_info *info,
6790 const bool commit_window)
6792 struct bpf_insn_aux_data *aux = commit_window ? cur_aux(env) : &info->aux;
6793 struct bpf_verifier_state *vstate = env->cur_state;
6794 bool off_is_imm = tnum_is_const(off_reg->var_off);
6795 bool off_is_neg = off_reg->smin_value < 0;
6796 bool ptr_is_dst_reg = ptr_reg == dst_reg;
6797 u8 opcode = BPF_OP(insn->code);
6798 u32 alu_state, alu_limit;
6799 struct bpf_reg_state tmp;
6803 if (can_skip_alu_sanitation(env, insn))
6806 /* We already marked aux for masking from non-speculative
6807 * paths, thus we got here in the first place. We only care
6808 * to explore bad access from here.
6810 if (vstate->speculative)
6813 if (!commit_window) {
6814 if (!tnum_is_const(off_reg->var_off) &&
6815 (off_reg->smin_value < 0) != (off_reg->smax_value < 0))
6816 return REASON_BOUNDS;
6818 info->mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
6819 (opcode == BPF_SUB && !off_is_neg);
6822 err = retrieve_ptr_limit(ptr_reg, &alu_limit, info->mask_to_left);
6826 if (commit_window) {
6827 /* In commit phase we narrow the masking window based on
6828 * the observed pointer move after the simulated operation.
6830 alu_state = info->aux.alu_state;
6831 alu_limit = abs(info->aux.alu_limit - alu_limit);
6833 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
6834 alu_state |= off_is_imm ? BPF_ALU_IMMEDIATE : 0;
6835 alu_state |= ptr_is_dst_reg ?
6836 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
6838 /* Limit pruning on unknown scalars to enable deep search for
6839 * potential masking differences from other program paths.
6842 env->explore_alu_limits = true;
6845 err = update_alu_sanitation_state(aux, alu_state, alu_limit);
6849 /* If we're in commit phase, we're done here given we already
6850 * pushed the truncated dst_reg into the speculative verification
6853 * Also, when register is a known constant, we rewrite register-based
6854 * operation to immediate-based, and thus do not need masking (and as
6855 * a consequence, do not need to simulate the zero-truncation either).
6857 if (commit_window || off_is_imm)
6860 /* Simulate and find potential out-of-bounds access under
6861 * speculative execution from truncation as a result of
6862 * masking when off was not within expected range. If off
6863 * sits in dst, then we temporarily need to move ptr there
6864 * to simulate dst (== 0) +/-= ptr. Needed, for example,
6865 * for cases where we use K-based arithmetic in one direction
6866 * and truncated reg-based in the other in order to explore
6869 if (!ptr_is_dst_reg) {
6871 *dst_reg = *ptr_reg;
6873 ret = sanitize_speculative_path(env, NULL, env->insn_idx + 1,
6875 if (!ptr_is_dst_reg && ret)
6877 return !ret ? REASON_STACK : 0;
6880 static void sanitize_mark_insn_seen(struct bpf_verifier_env *env)
6882 struct bpf_verifier_state *vstate = env->cur_state;
6884 /* If we simulate paths under speculation, we don't update the
6885 * insn as 'seen' such that when we verify unreachable paths in
6886 * the non-speculative domain, sanitize_dead_code() can still
6887 * rewrite/sanitize them.
6889 if (!vstate->speculative)
6890 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
6893 static int sanitize_err(struct bpf_verifier_env *env,
6894 const struct bpf_insn *insn, int reason,
6895 const struct bpf_reg_state *off_reg,
6896 const struct bpf_reg_state *dst_reg)
6898 static const char *err = "pointer arithmetic with it prohibited for !root";
6899 const char *op = BPF_OP(insn->code) == BPF_ADD ? "add" : "sub";
6900 u32 dst = insn->dst_reg, src = insn->src_reg;
6904 verbose(env, "R%d has unknown scalar with mixed signed bounds, %s\n",
6905 off_reg == dst_reg ? dst : src, err);
6908 verbose(env, "R%d has pointer with unsupported alu operation, %s\n",
6909 off_reg == dst_reg ? src : dst, err);
6912 verbose(env, "R%d tried to %s from different maps, paths or scalars, %s\n",
6916 verbose(env, "R%d tried to %s beyond pointer bounds, %s\n",
6920 verbose(env, "R%d could not be pushed for speculative verification, %s\n",
6924 verbose(env, "verifier internal error: unknown reason (%d)\n",
6932 /* check that stack access falls within stack limits and that 'reg' doesn't
6933 * have a variable offset.
6935 * Variable offset is prohibited for unprivileged mode for simplicity since it
6936 * requires corresponding support in Spectre masking for stack ALU. See also
6937 * retrieve_ptr_limit().
6940 * 'off' includes 'reg->off'.
6942 static int check_stack_access_for_ptr_arithmetic(
6943 struct bpf_verifier_env *env,
6945 const struct bpf_reg_state *reg,
6948 if (!tnum_is_const(reg->var_off)) {
6951 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6952 verbose(env, "R%d variable stack access prohibited for !root, var_off=%s off=%d\n",
6953 regno, tn_buf, off);
6957 if (off >= 0 || off < -MAX_BPF_STACK) {
6958 verbose(env, "R%d stack pointer arithmetic goes out of range, "
6959 "prohibited for !root; off=%d\n", regno, off);
6966 static int sanitize_check_bounds(struct bpf_verifier_env *env,
6967 const struct bpf_insn *insn,
6968 const struct bpf_reg_state *dst_reg)
6970 u32 dst = insn->dst_reg;
6972 /* For unprivileged we require that resulting offset must be in bounds
6973 * in order to be able to sanitize access later on.
6975 if (env->bypass_spec_v1)
6978 switch (dst_reg->type) {
6980 if (check_stack_access_for_ptr_arithmetic(env, dst, dst_reg,
6981 dst_reg->off + dst_reg->var_off.value))
6984 case PTR_TO_MAP_VALUE:
6985 if (check_map_access(env, dst, dst_reg->off, 1, false)) {
6986 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
6987 "prohibited for !root\n", dst);
6998 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
6999 * Caller should also handle BPF_MOV case separately.
7000 * If we return -EACCES, caller may want to try again treating pointer as a
7001 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
7003 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
7004 struct bpf_insn *insn,
7005 const struct bpf_reg_state *ptr_reg,
7006 const struct bpf_reg_state *off_reg)
7008 struct bpf_verifier_state *vstate = env->cur_state;
7009 struct bpf_func_state *state = vstate->frame[vstate->curframe];
7010 struct bpf_reg_state *regs = state->regs, *dst_reg;
7011 bool known = tnum_is_const(off_reg->var_off);
7012 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
7013 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
7014 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
7015 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
7016 struct bpf_sanitize_info info = {};
7017 u8 opcode = BPF_OP(insn->code);
7018 u32 dst = insn->dst_reg;
7021 dst_reg = ®s[dst];
7023 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
7024 smin_val > smax_val || umin_val > umax_val) {
7025 /* Taint dst register if offset had invalid bounds derived from
7026 * e.g. dead branches.
7028 __mark_reg_unknown(env, dst_reg);
7032 if (BPF_CLASS(insn->code) != BPF_ALU64) {
7033 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
7034 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
7035 __mark_reg_unknown(env, dst_reg);
7040 "R%d 32-bit pointer arithmetic prohibited\n",
7045 if (ptr_reg->type & PTR_MAYBE_NULL) {
7046 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
7047 dst, reg_type_str(env, ptr_reg->type));
7051 switch (base_type(ptr_reg->type)) {
7052 case CONST_PTR_TO_MAP:
7053 /* smin_val represents the known value */
7054 if (known && smin_val == 0 && opcode == BPF_ADD)
7057 case PTR_TO_PACKET_END:
7059 case PTR_TO_SOCK_COMMON:
7060 case PTR_TO_TCP_SOCK:
7061 case PTR_TO_XDP_SOCK:
7063 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
7064 dst, reg_type_str(env, ptr_reg->type));
7067 if (type_may_be_null(ptr_reg->type))
7072 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
7073 * The id may be overwritten later if we create a new variable offset.
7075 dst_reg->type = ptr_reg->type;
7076 dst_reg->id = ptr_reg->id;
7078 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
7079 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
7082 /* pointer types do not carry 32-bit bounds at the moment. */
7083 __mark_reg32_unbounded(dst_reg);
7085 if (sanitize_needed(opcode)) {
7086 ret = sanitize_ptr_alu(env, insn, ptr_reg, off_reg, dst_reg,
7089 return sanitize_err(env, insn, ret, off_reg, dst_reg);
7094 /* We can take a fixed offset as long as it doesn't overflow
7095 * the s32 'off' field
7097 if (known && (ptr_reg->off + smin_val ==
7098 (s64)(s32)(ptr_reg->off + smin_val))) {
7099 /* pointer += K. Accumulate it into fixed offset */
7100 dst_reg->smin_value = smin_ptr;
7101 dst_reg->smax_value = smax_ptr;
7102 dst_reg->umin_value = umin_ptr;
7103 dst_reg->umax_value = umax_ptr;
7104 dst_reg->var_off = ptr_reg->var_off;
7105 dst_reg->off = ptr_reg->off + smin_val;
7106 dst_reg->raw = ptr_reg->raw;
7109 /* A new variable offset is created. Note that off_reg->off
7110 * == 0, since it's a scalar.
7111 * dst_reg gets the pointer type and since some positive
7112 * integer value was added to the pointer, give it a new 'id'
7113 * if it's a PTR_TO_PACKET.
7114 * this creates a new 'base' pointer, off_reg (variable) gets
7115 * added into the variable offset, and we copy the fixed offset
7118 if (signed_add_overflows(smin_ptr, smin_val) ||
7119 signed_add_overflows(smax_ptr, smax_val)) {
7120 dst_reg->smin_value = S64_MIN;
7121 dst_reg->smax_value = S64_MAX;
7123 dst_reg->smin_value = smin_ptr + smin_val;
7124 dst_reg->smax_value = smax_ptr + smax_val;
7126 if (umin_ptr + umin_val < umin_ptr ||
7127 umax_ptr + umax_val < umax_ptr) {
7128 dst_reg->umin_value = 0;
7129 dst_reg->umax_value = U64_MAX;
7131 dst_reg->umin_value = umin_ptr + umin_val;
7132 dst_reg->umax_value = umax_ptr + umax_val;
7134 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
7135 dst_reg->off = ptr_reg->off;
7136 dst_reg->raw = ptr_reg->raw;
7137 if (reg_is_pkt_pointer(ptr_reg)) {
7138 dst_reg->id = ++env->id_gen;
7139 /* something was added to pkt_ptr, set range to zero */
7140 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
7144 if (dst_reg == off_reg) {
7145 /* scalar -= pointer. Creates an unknown scalar */
7146 verbose(env, "R%d tried to subtract pointer from scalar\n",
7150 /* We don't allow subtraction from FP, because (according to
7151 * test_verifier.c test "invalid fp arithmetic", JITs might not
7152 * be able to deal with it.
7154 if (ptr_reg->type == PTR_TO_STACK) {
7155 verbose(env, "R%d subtraction from stack pointer prohibited\n",
7159 if (known && (ptr_reg->off - smin_val ==
7160 (s64)(s32)(ptr_reg->off - smin_val))) {
7161 /* pointer -= K. Subtract it from fixed offset */
7162 dst_reg->smin_value = smin_ptr;
7163 dst_reg->smax_value = smax_ptr;
7164 dst_reg->umin_value = umin_ptr;
7165 dst_reg->umax_value = umax_ptr;
7166 dst_reg->var_off = ptr_reg->var_off;
7167 dst_reg->id = ptr_reg->id;
7168 dst_reg->off = ptr_reg->off - smin_val;
7169 dst_reg->raw = ptr_reg->raw;
7172 /* A new variable offset is created. If the subtrahend is known
7173 * nonnegative, then any reg->range we had before is still good.
7175 if (signed_sub_overflows(smin_ptr, smax_val) ||
7176 signed_sub_overflows(smax_ptr, smin_val)) {
7177 /* Overflow possible, we know nothing */
7178 dst_reg->smin_value = S64_MIN;
7179 dst_reg->smax_value = S64_MAX;
7181 dst_reg->smin_value = smin_ptr - smax_val;
7182 dst_reg->smax_value = smax_ptr - smin_val;
7184 if (umin_ptr < umax_val) {
7185 /* Overflow possible, we know nothing */
7186 dst_reg->umin_value = 0;
7187 dst_reg->umax_value = U64_MAX;
7189 /* Cannot overflow (as long as bounds are consistent) */
7190 dst_reg->umin_value = umin_ptr - umax_val;
7191 dst_reg->umax_value = umax_ptr - umin_val;
7193 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
7194 dst_reg->off = ptr_reg->off;
7195 dst_reg->raw = ptr_reg->raw;
7196 if (reg_is_pkt_pointer(ptr_reg)) {
7197 dst_reg->id = ++env->id_gen;
7198 /* something was added to pkt_ptr, set range to zero */
7200 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
7206 /* bitwise ops on pointers are troublesome, prohibit. */
7207 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
7208 dst, bpf_alu_string[opcode >> 4]);
7211 /* other operators (e.g. MUL,LSH) produce non-pointer results */
7212 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
7213 dst, bpf_alu_string[opcode >> 4]);
7217 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
7220 __update_reg_bounds(dst_reg);
7221 __reg_deduce_bounds(dst_reg);
7222 __reg_bound_offset(dst_reg);
7224 if (sanitize_check_bounds(env, insn, dst_reg) < 0)
7226 if (sanitize_needed(opcode)) {
7227 ret = sanitize_ptr_alu(env, insn, dst_reg, off_reg, dst_reg,
7230 return sanitize_err(env, insn, ret, off_reg, dst_reg);
7236 static void scalar32_min_max_add(struct bpf_reg_state *dst_reg,
7237 struct bpf_reg_state *src_reg)
7239 s32 smin_val = src_reg->s32_min_value;
7240 s32 smax_val = src_reg->s32_max_value;
7241 u32 umin_val = src_reg->u32_min_value;
7242 u32 umax_val = src_reg->u32_max_value;
7244 if (signed_add32_overflows(dst_reg->s32_min_value, smin_val) ||
7245 signed_add32_overflows(dst_reg->s32_max_value, smax_val)) {
7246 dst_reg->s32_min_value = S32_MIN;
7247 dst_reg->s32_max_value = S32_MAX;
7249 dst_reg->s32_min_value += smin_val;
7250 dst_reg->s32_max_value += smax_val;
7252 if (dst_reg->u32_min_value + umin_val < umin_val ||
7253 dst_reg->u32_max_value + umax_val < umax_val) {
7254 dst_reg->u32_min_value = 0;
7255 dst_reg->u32_max_value = U32_MAX;
7257 dst_reg->u32_min_value += umin_val;
7258 dst_reg->u32_max_value += umax_val;
7262 static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
7263 struct bpf_reg_state *src_reg)
7265 s64 smin_val = src_reg->smin_value;
7266 s64 smax_val = src_reg->smax_value;
7267 u64 umin_val = src_reg->umin_value;
7268 u64 umax_val = src_reg->umax_value;
7270 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
7271 signed_add_overflows(dst_reg->smax_value, smax_val)) {
7272 dst_reg->smin_value = S64_MIN;
7273 dst_reg->smax_value = S64_MAX;
7275 dst_reg->smin_value += smin_val;
7276 dst_reg->smax_value += smax_val;
7278 if (dst_reg->umin_value + umin_val < umin_val ||
7279 dst_reg->umax_value + umax_val < umax_val) {
7280 dst_reg->umin_value = 0;
7281 dst_reg->umax_value = U64_MAX;
7283 dst_reg->umin_value += umin_val;
7284 dst_reg->umax_value += umax_val;
7288 static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg,
7289 struct bpf_reg_state *src_reg)
7291 s32 smin_val = src_reg->s32_min_value;
7292 s32 smax_val = src_reg->s32_max_value;
7293 u32 umin_val = src_reg->u32_min_value;
7294 u32 umax_val = src_reg->u32_max_value;
7296 if (signed_sub32_overflows(dst_reg->s32_min_value, smax_val) ||
7297 signed_sub32_overflows(dst_reg->s32_max_value, smin_val)) {
7298 /* Overflow possible, we know nothing */
7299 dst_reg->s32_min_value = S32_MIN;
7300 dst_reg->s32_max_value = S32_MAX;
7302 dst_reg->s32_min_value -= smax_val;
7303 dst_reg->s32_max_value -= smin_val;
7305 if (dst_reg->u32_min_value < umax_val) {
7306 /* Overflow possible, we know nothing */
7307 dst_reg->u32_min_value = 0;
7308 dst_reg->u32_max_value = U32_MAX;
7310 /* Cannot overflow (as long as bounds are consistent) */
7311 dst_reg->u32_min_value -= umax_val;
7312 dst_reg->u32_max_value -= umin_val;
7316 static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
7317 struct bpf_reg_state *src_reg)
7319 s64 smin_val = src_reg->smin_value;
7320 s64 smax_val = src_reg->smax_value;
7321 u64 umin_val = src_reg->umin_value;
7322 u64 umax_val = src_reg->umax_value;
7324 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
7325 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
7326 /* Overflow possible, we know nothing */
7327 dst_reg->smin_value = S64_MIN;
7328 dst_reg->smax_value = S64_MAX;
7330 dst_reg->smin_value -= smax_val;
7331 dst_reg->smax_value -= smin_val;
7333 if (dst_reg->umin_value < umax_val) {
7334 /* Overflow possible, we know nothing */
7335 dst_reg->umin_value = 0;
7336 dst_reg->umax_value = U64_MAX;
7338 /* Cannot overflow (as long as bounds are consistent) */
7339 dst_reg->umin_value -= umax_val;
7340 dst_reg->umax_value -= umin_val;
7344 static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg,
7345 struct bpf_reg_state *src_reg)
7347 s32 smin_val = src_reg->s32_min_value;
7348 u32 umin_val = src_reg->u32_min_value;
7349 u32 umax_val = src_reg->u32_max_value;
7351 if (smin_val < 0 || dst_reg->s32_min_value < 0) {
7352 /* Ain't nobody got time to multiply that sign */
7353 __mark_reg32_unbounded(dst_reg);
7356 /* Both values are positive, so we can work with unsigned and
7357 * copy the result to signed (unless it exceeds S32_MAX).
7359 if (umax_val > U16_MAX || dst_reg->u32_max_value > U16_MAX) {
7360 /* Potential overflow, we know nothing */
7361 __mark_reg32_unbounded(dst_reg);
7364 dst_reg->u32_min_value *= umin_val;
7365 dst_reg->u32_max_value *= umax_val;
7366 if (dst_reg->u32_max_value > S32_MAX) {
7367 /* Overflow possible, we know nothing */
7368 dst_reg->s32_min_value = S32_MIN;
7369 dst_reg->s32_max_value = S32_MAX;
7371 dst_reg->s32_min_value = dst_reg->u32_min_value;
7372 dst_reg->s32_max_value = dst_reg->u32_max_value;
7376 static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
7377 struct bpf_reg_state *src_reg)
7379 s64 smin_val = src_reg->smin_value;
7380 u64 umin_val = src_reg->umin_value;
7381 u64 umax_val = src_reg->umax_value;
7383 if (smin_val < 0 || dst_reg->smin_value < 0) {
7384 /* Ain't nobody got time to multiply that sign */
7385 __mark_reg64_unbounded(dst_reg);
7388 /* Both values are positive, so we can work with unsigned and
7389 * copy the result to signed (unless it exceeds S64_MAX).
7391 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
7392 /* Potential overflow, we know nothing */
7393 __mark_reg64_unbounded(dst_reg);
7396 dst_reg->umin_value *= umin_val;
7397 dst_reg->umax_value *= umax_val;
7398 if (dst_reg->umax_value > S64_MAX) {
7399 /* Overflow possible, we know nothing */
7400 dst_reg->smin_value = S64_MIN;
7401 dst_reg->smax_value = S64_MAX;
7403 dst_reg->smin_value = dst_reg->umin_value;
7404 dst_reg->smax_value = dst_reg->umax_value;
7408 static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
7409 struct bpf_reg_state *src_reg)
7411 bool src_known = tnum_subreg_is_const(src_reg->var_off);
7412 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
7413 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
7414 s32 smin_val = src_reg->s32_min_value;
7415 u32 umax_val = src_reg->u32_max_value;
7417 if (src_known && dst_known) {
7418 __mark_reg32_known(dst_reg, var32_off.value);
7422 /* We get our minimum from the var_off, since that's inherently
7423 * bitwise. Our maximum is the minimum of the operands' maxima.
7425 dst_reg->u32_min_value = var32_off.value;
7426 dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val);
7427 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
7428 /* Lose signed bounds when ANDing negative numbers,
7429 * ain't nobody got time for that.
7431 dst_reg->s32_min_value = S32_MIN;
7432 dst_reg->s32_max_value = S32_MAX;
7434 /* ANDing two positives gives a positive, so safe to
7435 * cast result into s64.
7437 dst_reg->s32_min_value = dst_reg->u32_min_value;
7438 dst_reg->s32_max_value = dst_reg->u32_max_value;
7442 static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
7443 struct bpf_reg_state *src_reg)
7445 bool src_known = tnum_is_const(src_reg->var_off);
7446 bool dst_known = tnum_is_const(dst_reg->var_off);
7447 s64 smin_val = src_reg->smin_value;
7448 u64 umax_val = src_reg->umax_value;
7450 if (src_known && dst_known) {
7451 __mark_reg_known(dst_reg, dst_reg->var_off.value);
7455 /* We get our minimum from the var_off, since that's inherently
7456 * bitwise. Our maximum is the minimum of the operands' maxima.
7458 dst_reg->umin_value = dst_reg->var_off.value;
7459 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
7460 if (dst_reg->smin_value < 0 || smin_val < 0) {
7461 /* Lose signed bounds when ANDing negative numbers,
7462 * ain't nobody got time for that.
7464 dst_reg->smin_value = S64_MIN;
7465 dst_reg->smax_value = S64_MAX;
7467 /* ANDing two positives gives a positive, so safe to
7468 * cast result into s64.
7470 dst_reg->smin_value = dst_reg->umin_value;
7471 dst_reg->smax_value = dst_reg->umax_value;
7473 /* We may learn something more from the var_off */
7474 __update_reg_bounds(dst_reg);
7477 static void scalar32_min_max_or(struct bpf_reg_state *dst_reg,
7478 struct bpf_reg_state *src_reg)
7480 bool src_known = tnum_subreg_is_const(src_reg->var_off);
7481 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
7482 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
7483 s32 smin_val = src_reg->s32_min_value;
7484 u32 umin_val = src_reg->u32_min_value;
7486 if (src_known && dst_known) {
7487 __mark_reg32_known(dst_reg, var32_off.value);
7491 /* We get our maximum from the var_off, and our minimum is the
7492 * maximum of the operands' minima
7494 dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val);
7495 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
7496 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
7497 /* Lose signed bounds when ORing negative numbers,
7498 * ain't nobody got time for that.
7500 dst_reg->s32_min_value = S32_MIN;
7501 dst_reg->s32_max_value = S32_MAX;
7503 /* ORing two positives gives a positive, so safe to
7504 * cast result into s64.
7506 dst_reg->s32_min_value = dst_reg->u32_min_value;
7507 dst_reg->s32_max_value = dst_reg->u32_max_value;
7511 static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
7512 struct bpf_reg_state *src_reg)
7514 bool src_known = tnum_is_const(src_reg->var_off);
7515 bool dst_known = tnum_is_const(dst_reg->var_off);
7516 s64 smin_val = src_reg->smin_value;
7517 u64 umin_val = src_reg->umin_value;
7519 if (src_known && dst_known) {
7520 __mark_reg_known(dst_reg, dst_reg->var_off.value);
7524 /* We get our maximum from the var_off, and our minimum is the
7525 * maximum of the operands' minima
7527 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
7528 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
7529 if (dst_reg->smin_value < 0 || smin_val < 0) {
7530 /* Lose signed bounds when ORing negative numbers,
7531 * ain't nobody got time for that.
7533 dst_reg->smin_value = S64_MIN;
7534 dst_reg->smax_value = S64_MAX;
7536 /* ORing two positives gives a positive, so safe to
7537 * cast result into s64.
7539 dst_reg->smin_value = dst_reg->umin_value;
7540 dst_reg->smax_value = dst_reg->umax_value;
7542 /* We may learn something more from the var_off */
7543 __update_reg_bounds(dst_reg);
7546 static void scalar32_min_max_xor(struct bpf_reg_state *dst_reg,
7547 struct bpf_reg_state *src_reg)
7549 bool src_known = tnum_subreg_is_const(src_reg->var_off);
7550 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
7551 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
7552 s32 smin_val = src_reg->s32_min_value;
7554 if (src_known && dst_known) {
7555 __mark_reg32_known(dst_reg, var32_off.value);
7559 /* We get both minimum and maximum from the var32_off. */
7560 dst_reg->u32_min_value = var32_off.value;
7561 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
7563 if (dst_reg->s32_min_value >= 0 && smin_val >= 0) {
7564 /* XORing two positive sign numbers gives a positive,
7565 * so safe to cast u32 result into s32.
7567 dst_reg->s32_min_value = dst_reg->u32_min_value;
7568 dst_reg->s32_max_value = dst_reg->u32_max_value;
7570 dst_reg->s32_min_value = S32_MIN;
7571 dst_reg->s32_max_value = S32_MAX;
7575 static void scalar_min_max_xor(struct bpf_reg_state *dst_reg,
7576 struct bpf_reg_state *src_reg)
7578 bool src_known = tnum_is_const(src_reg->var_off);
7579 bool dst_known = tnum_is_const(dst_reg->var_off);
7580 s64 smin_val = src_reg->smin_value;
7582 if (src_known && dst_known) {
7583 /* dst_reg->var_off.value has been updated earlier */
7584 __mark_reg_known(dst_reg, dst_reg->var_off.value);
7588 /* We get both minimum and maximum from the var_off. */
7589 dst_reg->umin_value = dst_reg->var_off.value;
7590 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
7592 if (dst_reg->smin_value >= 0 && smin_val >= 0) {
7593 /* XORing two positive sign numbers gives a positive,
7594 * so safe to cast u64 result into s64.
7596 dst_reg->smin_value = dst_reg->umin_value;
7597 dst_reg->smax_value = dst_reg->umax_value;
7599 dst_reg->smin_value = S64_MIN;
7600 dst_reg->smax_value = S64_MAX;
7603 __update_reg_bounds(dst_reg);
7606 static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
7607 u64 umin_val, u64 umax_val)
7609 /* We lose all sign bit information (except what we can pick
7612 dst_reg->s32_min_value = S32_MIN;
7613 dst_reg->s32_max_value = S32_MAX;
7614 /* If we might shift our top bit out, then we know nothing */
7615 if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) {
7616 dst_reg->u32_min_value = 0;
7617 dst_reg->u32_max_value = U32_MAX;
7619 dst_reg->u32_min_value <<= umin_val;
7620 dst_reg->u32_max_value <<= umax_val;
7624 static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
7625 struct bpf_reg_state *src_reg)
7627 u32 umax_val = src_reg->u32_max_value;
7628 u32 umin_val = src_reg->u32_min_value;
7629 /* u32 alu operation will zext upper bits */
7630 struct tnum subreg = tnum_subreg(dst_reg->var_off);
7632 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
7633 dst_reg->var_off = tnum_subreg(tnum_lshift(subreg, umin_val));
7634 /* Not required but being careful mark reg64 bounds as unknown so
7635 * that we are forced to pick them up from tnum and zext later and
7636 * if some path skips this step we are still safe.
7638 __mark_reg64_unbounded(dst_reg);
7639 __update_reg32_bounds(dst_reg);
7642 static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg,
7643 u64 umin_val, u64 umax_val)
7645 /* Special case <<32 because it is a common compiler pattern to sign
7646 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
7647 * positive we know this shift will also be positive so we can track
7648 * bounds correctly. Otherwise we lose all sign bit information except
7649 * what we can pick up from var_off. Perhaps we can generalize this
7650 * later to shifts of any length.
7652 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0)
7653 dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32;
7655 dst_reg->smax_value = S64_MAX;
7657 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0)
7658 dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32;
7660 dst_reg->smin_value = S64_MIN;
7662 /* If we might shift our top bit out, then we know nothing */
7663 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
7664 dst_reg->umin_value = 0;
7665 dst_reg->umax_value = U64_MAX;
7667 dst_reg->umin_value <<= umin_val;
7668 dst_reg->umax_value <<= umax_val;
7672 static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
7673 struct bpf_reg_state *src_reg)
7675 u64 umax_val = src_reg->umax_value;
7676 u64 umin_val = src_reg->umin_value;
7678 /* scalar64 calc uses 32bit unshifted bounds so must be called first */
7679 __scalar64_min_max_lsh(dst_reg, umin_val, umax_val);
7680 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
7682 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
7683 /* We may learn something more from the var_off */
7684 __update_reg_bounds(dst_reg);
7687 static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg,
7688 struct bpf_reg_state *src_reg)
7690 struct tnum subreg = tnum_subreg(dst_reg->var_off);
7691 u32 umax_val = src_reg->u32_max_value;
7692 u32 umin_val = src_reg->u32_min_value;
7694 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
7695 * be negative, then either:
7696 * 1) src_reg might be zero, so the sign bit of the result is
7697 * unknown, so we lose our signed bounds
7698 * 2) it's known negative, thus the unsigned bounds capture the
7700 * 3) the signed bounds cross zero, so they tell us nothing
7702 * If the value in dst_reg is known nonnegative, then again the
7703 * unsigned bounds capture the signed bounds.
7704 * Thus, in all cases it suffices to blow away our signed bounds
7705 * and rely on inferring new ones from the unsigned bounds and
7706 * var_off of the result.
7708 dst_reg->s32_min_value = S32_MIN;
7709 dst_reg->s32_max_value = S32_MAX;
7711 dst_reg->var_off = tnum_rshift(subreg, umin_val);
7712 dst_reg->u32_min_value >>= umax_val;
7713 dst_reg->u32_max_value >>= umin_val;
7715 __mark_reg64_unbounded(dst_reg);
7716 __update_reg32_bounds(dst_reg);
7719 static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
7720 struct bpf_reg_state *src_reg)
7722 u64 umax_val = src_reg->umax_value;
7723 u64 umin_val = src_reg->umin_value;
7725 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
7726 * be negative, then either:
7727 * 1) src_reg might be zero, so the sign bit of the result is
7728 * unknown, so we lose our signed bounds
7729 * 2) it's known negative, thus the unsigned bounds capture the
7731 * 3) the signed bounds cross zero, so they tell us nothing
7733 * If the value in dst_reg is known nonnegative, then again the
7734 * unsigned bounds capture the signed bounds.
7735 * Thus, in all cases it suffices to blow away our signed bounds
7736 * and rely on inferring new ones from the unsigned bounds and
7737 * var_off of the result.
7739 dst_reg->smin_value = S64_MIN;
7740 dst_reg->smax_value = S64_MAX;
7741 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
7742 dst_reg->umin_value >>= umax_val;
7743 dst_reg->umax_value >>= umin_val;
7745 /* Its not easy to operate on alu32 bounds here because it depends
7746 * on bits being shifted in. Take easy way out and mark unbounded
7747 * so we can recalculate later from tnum.
7749 __mark_reg32_unbounded(dst_reg);
7750 __update_reg_bounds(dst_reg);
7753 static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg,
7754 struct bpf_reg_state *src_reg)
7756 u64 umin_val = src_reg->u32_min_value;
7758 /* Upon reaching here, src_known is true and
7759 * umax_val is equal to umin_val.
7761 dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val);
7762 dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val);
7764 dst_reg->var_off = tnum_arshift(tnum_subreg(dst_reg->var_off), umin_val, 32);
7766 /* blow away the dst_reg umin_value/umax_value and rely on
7767 * dst_reg var_off to refine the result.
7769 dst_reg->u32_min_value = 0;
7770 dst_reg->u32_max_value = U32_MAX;
7772 __mark_reg64_unbounded(dst_reg);
7773 __update_reg32_bounds(dst_reg);
7776 static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
7777 struct bpf_reg_state *src_reg)
7779 u64 umin_val = src_reg->umin_value;
7781 /* Upon reaching here, src_known is true and umax_val is equal
7784 dst_reg->smin_value >>= umin_val;
7785 dst_reg->smax_value >>= umin_val;
7787 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val, 64);
7789 /* blow away the dst_reg umin_value/umax_value and rely on
7790 * dst_reg var_off to refine the result.
7792 dst_reg->umin_value = 0;
7793 dst_reg->umax_value = U64_MAX;
7795 /* Its not easy to operate on alu32 bounds here because it depends
7796 * on bits being shifted in from upper 32-bits. Take easy way out
7797 * and mark unbounded so we can recalculate later from tnum.
7799 __mark_reg32_unbounded(dst_reg);
7800 __update_reg_bounds(dst_reg);
7803 /* WARNING: This function does calculations on 64-bit values, but the actual
7804 * execution may occur on 32-bit values. Therefore, things like bitshifts
7805 * need extra checks in the 32-bit case.
7807 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
7808 struct bpf_insn *insn,
7809 struct bpf_reg_state *dst_reg,
7810 struct bpf_reg_state src_reg)
7812 struct bpf_reg_state *regs = cur_regs(env);
7813 u8 opcode = BPF_OP(insn->code);
7815 s64 smin_val, smax_val;
7816 u64 umin_val, umax_val;
7817 s32 s32_min_val, s32_max_val;
7818 u32 u32_min_val, u32_max_val;
7819 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
7820 bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
7823 smin_val = src_reg.smin_value;
7824 smax_val = src_reg.smax_value;
7825 umin_val = src_reg.umin_value;
7826 umax_val = src_reg.umax_value;
7828 s32_min_val = src_reg.s32_min_value;
7829 s32_max_val = src_reg.s32_max_value;
7830 u32_min_val = src_reg.u32_min_value;
7831 u32_max_val = src_reg.u32_max_value;
7834 src_known = tnum_subreg_is_const(src_reg.var_off);
7836 (s32_min_val != s32_max_val || u32_min_val != u32_max_val)) ||
7837 s32_min_val > s32_max_val || u32_min_val > u32_max_val) {
7838 /* Taint dst register if offset had invalid bounds
7839 * derived from e.g. dead branches.
7841 __mark_reg_unknown(env, dst_reg);
7845 src_known = tnum_is_const(src_reg.var_off);
7847 (smin_val != smax_val || umin_val != umax_val)) ||
7848 smin_val > smax_val || umin_val > umax_val) {
7849 /* Taint dst register if offset had invalid bounds
7850 * derived from e.g. dead branches.
7852 __mark_reg_unknown(env, dst_reg);
7858 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
7859 __mark_reg_unknown(env, dst_reg);
7863 if (sanitize_needed(opcode)) {
7864 ret = sanitize_val_alu(env, insn);
7866 return sanitize_err(env, insn, ret, NULL, NULL);
7869 /* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
7870 * There are two classes of instructions: The first class we track both
7871 * alu32 and alu64 sign/unsigned bounds independently this provides the
7872 * greatest amount of precision when alu operations are mixed with jmp32
7873 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
7874 * and BPF_OR. This is possible because these ops have fairly easy to
7875 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
7876 * See alu32 verifier tests for examples. The second class of
7877 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
7878 * with regards to tracking sign/unsigned bounds because the bits may
7879 * cross subreg boundaries in the alu64 case. When this happens we mark
7880 * the reg unbounded in the subreg bound space and use the resulting
7881 * tnum to calculate an approximation of the sign/unsigned bounds.
7885 scalar32_min_max_add(dst_reg, &src_reg);
7886 scalar_min_max_add(dst_reg, &src_reg);
7887 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
7890 scalar32_min_max_sub(dst_reg, &src_reg);
7891 scalar_min_max_sub(dst_reg, &src_reg);
7892 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
7895 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
7896 scalar32_min_max_mul(dst_reg, &src_reg);
7897 scalar_min_max_mul(dst_reg, &src_reg);
7900 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
7901 scalar32_min_max_and(dst_reg, &src_reg);
7902 scalar_min_max_and(dst_reg, &src_reg);
7905 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
7906 scalar32_min_max_or(dst_reg, &src_reg);
7907 scalar_min_max_or(dst_reg, &src_reg);
7910 dst_reg->var_off = tnum_xor(dst_reg->var_off, src_reg.var_off);
7911 scalar32_min_max_xor(dst_reg, &src_reg);
7912 scalar_min_max_xor(dst_reg, &src_reg);
7915 if (umax_val >= insn_bitness) {
7916 /* Shifts greater than 31 or 63 are undefined.
7917 * This includes shifts by a negative number.
7919 mark_reg_unknown(env, regs, insn->dst_reg);
7923 scalar32_min_max_lsh(dst_reg, &src_reg);
7925 scalar_min_max_lsh(dst_reg, &src_reg);
7928 if (umax_val >= insn_bitness) {
7929 /* Shifts greater than 31 or 63 are undefined.
7930 * This includes shifts by a negative number.
7932 mark_reg_unknown(env, regs, insn->dst_reg);
7936 scalar32_min_max_rsh(dst_reg, &src_reg);
7938 scalar_min_max_rsh(dst_reg, &src_reg);
7941 if (umax_val >= insn_bitness) {
7942 /* Shifts greater than 31 or 63 are undefined.
7943 * This includes shifts by a negative number.
7945 mark_reg_unknown(env, regs, insn->dst_reg);
7949 scalar32_min_max_arsh(dst_reg, &src_reg);
7951 scalar_min_max_arsh(dst_reg, &src_reg);
7954 mark_reg_unknown(env, regs, insn->dst_reg);
7958 /* ALU32 ops are zero extended into 64bit register */
7960 zext_32_to_64(dst_reg);
7962 __update_reg_bounds(dst_reg);
7963 __reg_deduce_bounds(dst_reg);
7964 __reg_bound_offset(dst_reg);
7968 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
7971 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
7972 struct bpf_insn *insn)
7974 struct bpf_verifier_state *vstate = env->cur_state;
7975 struct bpf_func_state *state = vstate->frame[vstate->curframe];
7976 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
7977 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
7978 u8 opcode = BPF_OP(insn->code);
7981 dst_reg = ®s[insn->dst_reg];
7983 if (dst_reg->type != SCALAR_VALUE)
7986 /* Make sure ID is cleared otherwise dst_reg min/max could be
7987 * incorrectly propagated into other registers by find_equal_scalars()
7990 if (BPF_SRC(insn->code) == BPF_X) {
7991 src_reg = ®s[insn->src_reg];
7992 if (src_reg->type != SCALAR_VALUE) {
7993 if (dst_reg->type != SCALAR_VALUE) {
7994 /* Combining two pointers by any ALU op yields
7995 * an arbitrary scalar. Disallow all math except
7996 * pointer subtraction
7998 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
7999 mark_reg_unknown(env, regs, insn->dst_reg);
8002 verbose(env, "R%d pointer %s pointer prohibited\n",
8004 bpf_alu_string[opcode >> 4]);
8007 /* scalar += pointer
8008 * This is legal, but we have to reverse our
8009 * src/dest handling in computing the range
8011 err = mark_chain_precision(env, insn->dst_reg);
8014 return adjust_ptr_min_max_vals(env, insn,
8017 } else if (ptr_reg) {
8018 /* pointer += scalar */
8019 err = mark_chain_precision(env, insn->src_reg);
8022 return adjust_ptr_min_max_vals(env, insn,
8026 /* Pretend the src is a reg with a known value, since we only
8027 * need to be able to read from this state.
8029 off_reg.type = SCALAR_VALUE;
8030 __mark_reg_known(&off_reg, insn->imm);
8032 if (ptr_reg) /* pointer += K */
8033 return adjust_ptr_min_max_vals(env, insn,
8037 /* Got here implies adding two SCALAR_VALUEs */
8038 if (WARN_ON_ONCE(ptr_reg)) {
8039 print_verifier_state(env, state);
8040 verbose(env, "verifier internal error: unexpected ptr_reg\n");
8043 if (WARN_ON(!src_reg)) {
8044 print_verifier_state(env, state);
8045 verbose(env, "verifier internal error: no src_reg\n");
8048 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
8051 /* check validity of 32-bit and 64-bit arithmetic operations */
8052 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
8054 struct bpf_reg_state *regs = cur_regs(env);
8055 u8 opcode = BPF_OP(insn->code);
8058 if (opcode == BPF_END || opcode == BPF_NEG) {
8059 if (opcode == BPF_NEG) {
8060 if (BPF_SRC(insn->code) != 0 ||
8061 insn->src_reg != BPF_REG_0 ||
8062 insn->off != 0 || insn->imm != 0) {
8063 verbose(env, "BPF_NEG uses reserved fields\n");
8067 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
8068 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
8069 BPF_CLASS(insn->code) == BPF_ALU64) {
8070 verbose(env, "BPF_END uses reserved fields\n");
8075 /* check src operand */
8076 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
8080 if (is_pointer_value(env, insn->dst_reg)) {
8081 verbose(env, "R%d pointer arithmetic prohibited\n",
8086 /* check dest operand */
8087 err = check_reg_arg(env, insn->dst_reg, DST_OP);
8091 } else if (opcode == BPF_MOV) {
8093 if (BPF_SRC(insn->code) == BPF_X) {
8094 if (insn->imm != 0 || insn->off != 0) {
8095 verbose(env, "BPF_MOV uses reserved fields\n");
8099 /* check src operand */
8100 err = check_reg_arg(env, insn->src_reg, SRC_OP);
8104 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
8105 verbose(env, "BPF_MOV uses reserved fields\n");
8110 /* check dest operand, mark as required later */
8111 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
8115 if (BPF_SRC(insn->code) == BPF_X) {
8116 struct bpf_reg_state *src_reg = regs + insn->src_reg;
8117 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
8119 if (BPF_CLASS(insn->code) == BPF_ALU64) {
8121 * copy register state to dest reg
8123 if (src_reg->type == SCALAR_VALUE && !src_reg->id)
8124 /* Assign src and dst registers the same ID
8125 * that will be used by find_equal_scalars()
8126 * to propagate min/max range.
8128 src_reg->id = ++env->id_gen;
8129 *dst_reg = *src_reg;
8130 dst_reg->live |= REG_LIVE_WRITTEN;
8131 dst_reg->subreg_def = DEF_NOT_SUBREG;
8134 if (is_pointer_value(env, insn->src_reg)) {
8136 "R%d partial copy of pointer\n",
8139 } else if (src_reg->type == SCALAR_VALUE) {
8140 *dst_reg = *src_reg;
8141 /* Make sure ID is cleared otherwise
8142 * dst_reg min/max could be incorrectly
8143 * propagated into src_reg by find_equal_scalars()
8146 dst_reg->live |= REG_LIVE_WRITTEN;
8147 dst_reg->subreg_def = env->insn_idx + 1;
8149 mark_reg_unknown(env, regs,
8152 zext_32_to_64(dst_reg);
8154 __update_reg_bounds(dst_reg);
8155 __reg_deduce_bounds(dst_reg);
8156 __reg_bound_offset(dst_reg);
8160 * remember the value we stored into this reg
8162 /* clear any state __mark_reg_known doesn't set */
8163 mark_reg_unknown(env, regs, insn->dst_reg);
8164 regs[insn->dst_reg].type = SCALAR_VALUE;
8165 if (BPF_CLASS(insn->code) == BPF_ALU64) {
8166 __mark_reg_known(regs + insn->dst_reg,
8169 __mark_reg_known(regs + insn->dst_reg,
8174 } else if (opcode > BPF_END) {
8175 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
8178 } else { /* all other ALU ops: and, sub, xor, add, ... */
8180 if (BPF_SRC(insn->code) == BPF_X) {
8181 if (insn->imm != 0 || insn->off != 0) {
8182 verbose(env, "BPF_ALU uses reserved fields\n");
8185 /* check src1 operand */
8186 err = check_reg_arg(env, insn->src_reg, SRC_OP);
8190 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
8191 verbose(env, "BPF_ALU uses reserved fields\n");
8196 /* check src2 operand */
8197 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
8201 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
8202 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
8203 verbose(env, "div by zero\n");
8207 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
8208 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
8209 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
8211 if (insn->imm < 0 || insn->imm >= size) {
8212 verbose(env, "invalid shift %d\n", insn->imm);
8217 /* check dest operand */
8218 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
8222 return adjust_reg_min_max_vals(env, insn);
8228 static void __find_good_pkt_pointers(struct bpf_func_state *state,
8229 struct bpf_reg_state *dst_reg,
8230 enum bpf_reg_type type, int new_range)
8232 struct bpf_reg_state *reg;
8235 for (i = 0; i < MAX_BPF_REG; i++) {
8236 reg = &state->regs[i];
8237 if (reg->type == type && reg->id == dst_reg->id)
8238 /* keep the maximum range already checked */
8239 reg->range = max(reg->range, new_range);
8242 bpf_for_each_spilled_reg(i, state, reg) {
8245 if (reg->type == type && reg->id == dst_reg->id)
8246 reg->range = max(reg->range, new_range);
8250 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
8251 struct bpf_reg_state *dst_reg,
8252 enum bpf_reg_type type,
8253 bool range_right_open)
8257 if (dst_reg->off < 0 ||
8258 (dst_reg->off == 0 && range_right_open))
8259 /* This doesn't give us any range */
8262 if (dst_reg->umax_value > MAX_PACKET_OFF ||
8263 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
8264 /* Risk of overflow. For instance, ptr + (1<<63) may be less
8265 * than pkt_end, but that's because it's also less than pkt.
8269 new_range = dst_reg->off;
8270 if (range_right_open)
8273 /* Examples for register markings:
8275 * pkt_data in dst register:
8279 * if (r2 > pkt_end) goto <handle exception>
8284 * if (r2 < pkt_end) goto <access okay>
8285 * <handle exception>
8288 * r2 == dst_reg, pkt_end == src_reg
8289 * r2=pkt(id=n,off=8,r=0)
8290 * r3=pkt(id=n,off=0,r=0)
8292 * pkt_data in src register:
8296 * if (pkt_end >= r2) goto <access okay>
8297 * <handle exception>
8301 * if (pkt_end <= r2) goto <handle exception>
8305 * pkt_end == dst_reg, r2 == src_reg
8306 * r2=pkt(id=n,off=8,r=0)
8307 * r3=pkt(id=n,off=0,r=0)
8309 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
8310 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
8311 * and [r3, r3 + 8-1) respectively is safe to access depending on
8315 /* If our ids match, then we must have the same max_value. And we
8316 * don't care about the other reg's fixed offset, since if it's too big
8317 * the range won't allow anything.
8318 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
8320 for (i = 0; i <= vstate->curframe; i++)
8321 __find_good_pkt_pointers(vstate->frame[i], dst_reg, type,
8325 static int is_branch32_taken(struct bpf_reg_state *reg, u32 val, u8 opcode)
8327 struct tnum subreg = tnum_subreg(reg->var_off);
8328 s32 sval = (s32)val;
8332 if (tnum_is_const(subreg))
8333 return !!tnum_equals_const(subreg, val);
8336 if (tnum_is_const(subreg))
8337 return !tnum_equals_const(subreg, val);
8340 if ((~subreg.mask & subreg.value) & val)
8342 if (!((subreg.mask | subreg.value) & val))
8346 if (reg->u32_min_value > val)
8348 else if (reg->u32_max_value <= val)
8352 if (reg->s32_min_value > sval)
8354 else if (reg->s32_max_value <= sval)
8358 if (reg->u32_max_value < val)
8360 else if (reg->u32_min_value >= val)
8364 if (reg->s32_max_value < sval)
8366 else if (reg->s32_min_value >= sval)
8370 if (reg->u32_min_value >= val)
8372 else if (reg->u32_max_value < val)
8376 if (reg->s32_min_value >= sval)
8378 else if (reg->s32_max_value < sval)
8382 if (reg->u32_max_value <= val)
8384 else if (reg->u32_min_value > val)
8388 if (reg->s32_max_value <= sval)
8390 else if (reg->s32_min_value > sval)
8399 static int is_branch64_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
8401 s64 sval = (s64)val;
8405 if (tnum_is_const(reg->var_off))
8406 return !!tnum_equals_const(reg->var_off, val);
8409 if (tnum_is_const(reg->var_off))
8410 return !tnum_equals_const(reg->var_off, val);
8413 if ((~reg->var_off.mask & reg->var_off.value) & val)
8415 if (!((reg->var_off.mask | reg->var_off.value) & val))
8419 if (reg->umin_value > val)
8421 else if (reg->umax_value <= val)
8425 if (reg->smin_value > sval)
8427 else if (reg->smax_value <= sval)
8431 if (reg->umax_value < val)
8433 else if (reg->umin_value >= val)
8437 if (reg->smax_value < sval)
8439 else if (reg->smin_value >= sval)
8443 if (reg->umin_value >= val)
8445 else if (reg->umax_value < val)
8449 if (reg->smin_value >= sval)
8451 else if (reg->smax_value < sval)
8455 if (reg->umax_value <= val)
8457 else if (reg->umin_value > val)
8461 if (reg->smax_value <= sval)
8463 else if (reg->smin_value > sval)
8471 /* compute branch direction of the expression "if (reg opcode val) goto target;"
8473 * 1 - branch will be taken and "goto target" will be executed
8474 * 0 - branch will not be taken and fall-through to next insn
8475 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value
8478 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
8481 if (__is_pointer_value(false, reg)) {
8482 if (!reg_type_not_null(reg->type))
8485 /* If pointer is valid tests against zero will fail so we can
8486 * use this to direct branch taken.
8502 return is_branch32_taken(reg, val, opcode);
8503 return is_branch64_taken(reg, val, opcode);
8506 static int flip_opcode(u32 opcode)
8508 /* How can we transform "a <op> b" into "b <op> a"? */
8509 static const u8 opcode_flip[16] = {
8510 /* these stay the same */
8511 [BPF_JEQ >> 4] = BPF_JEQ,
8512 [BPF_JNE >> 4] = BPF_JNE,
8513 [BPF_JSET >> 4] = BPF_JSET,
8514 /* these swap "lesser" and "greater" (L and G in the opcodes) */
8515 [BPF_JGE >> 4] = BPF_JLE,
8516 [BPF_JGT >> 4] = BPF_JLT,
8517 [BPF_JLE >> 4] = BPF_JGE,
8518 [BPF_JLT >> 4] = BPF_JGT,
8519 [BPF_JSGE >> 4] = BPF_JSLE,
8520 [BPF_JSGT >> 4] = BPF_JSLT,
8521 [BPF_JSLE >> 4] = BPF_JSGE,
8522 [BPF_JSLT >> 4] = BPF_JSGT
8524 return opcode_flip[opcode >> 4];
8527 static int is_pkt_ptr_branch_taken(struct bpf_reg_state *dst_reg,
8528 struct bpf_reg_state *src_reg,
8531 struct bpf_reg_state *pkt;
8533 if (src_reg->type == PTR_TO_PACKET_END) {
8535 } else if (dst_reg->type == PTR_TO_PACKET_END) {
8537 opcode = flip_opcode(opcode);
8542 if (pkt->range >= 0)
8547 /* pkt <= pkt_end */
8551 if (pkt->range == BEYOND_PKT_END)
8552 /* pkt has at last one extra byte beyond pkt_end */
8553 return opcode == BPF_JGT;
8559 /* pkt >= pkt_end */
8560 if (pkt->range == BEYOND_PKT_END || pkt->range == AT_PKT_END)
8561 return opcode == BPF_JGE;
8567 /* Adjusts the register min/max values in the case that the dst_reg is the
8568 * variable register that we are working on, and src_reg is a constant or we're
8569 * simply doing a BPF_K check.
8570 * In JEQ/JNE cases we also adjust the var_off values.
8572 static void reg_set_min_max(struct bpf_reg_state *true_reg,
8573 struct bpf_reg_state *false_reg,
8575 u8 opcode, bool is_jmp32)
8577 struct tnum false_32off = tnum_subreg(false_reg->var_off);
8578 struct tnum false_64off = false_reg->var_off;
8579 struct tnum true_32off = tnum_subreg(true_reg->var_off);
8580 struct tnum true_64off = true_reg->var_off;
8581 s64 sval = (s64)val;
8582 s32 sval32 = (s32)val32;
8584 /* If the dst_reg is a pointer, we can't learn anything about its
8585 * variable offset from the compare (unless src_reg were a pointer into
8586 * the same object, but we don't bother with that.
8587 * Since false_reg and true_reg have the same type by construction, we
8588 * only need to check one of them for pointerness.
8590 if (__is_pointer_value(false, false_reg))
8597 struct bpf_reg_state *reg =
8598 opcode == BPF_JEQ ? true_reg : false_reg;
8600 /* JEQ/JNE comparison doesn't change the register equivalence.
8602 * if (r1 == 42) goto label;
8604 * label: // here both r1 and r2 are known to be 42.
8606 * Hence when marking register as known preserve it's ID.
8609 __mark_reg32_known(reg, val32);
8611 ___mark_reg_known(reg, val);
8616 false_32off = tnum_and(false_32off, tnum_const(~val32));
8617 if (is_power_of_2(val32))
8618 true_32off = tnum_or(true_32off,
8621 false_64off = tnum_and(false_64off, tnum_const(~val));
8622 if (is_power_of_2(val))
8623 true_64off = tnum_or(true_64off,
8631 u32 false_umax = opcode == BPF_JGT ? val32 : val32 - 1;
8632 u32 true_umin = opcode == BPF_JGT ? val32 + 1 : val32;
8634 false_reg->u32_max_value = min(false_reg->u32_max_value,
8636 true_reg->u32_min_value = max(true_reg->u32_min_value,
8639 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
8640 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
8642 false_reg->umax_value = min(false_reg->umax_value, false_umax);
8643 true_reg->umin_value = max(true_reg->umin_value, true_umin);
8651 s32 false_smax = opcode == BPF_JSGT ? sval32 : sval32 - 1;
8652 s32 true_smin = opcode == BPF_JSGT ? sval32 + 1 : sval32;
8654 false_reg->s32_max_value = min(false_reg->s32_max_value, false_smax);
8655 true_reg->s32_min_value = max(true_reg->s32_min_value, true_smin);
8657 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
8658 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
8660 false_reg->smax_value = min(false_reg->smax_value, false_smax);
8661 true_reg->smin_value = max(true_reg->smin_value, true_smin);
8669 u32 false_umin = opcode == BPF_JLT ? val32 : val32 + 1;
8670 u32 true_umax = opcode == BPF_JLT ? val32 - 1 : val32;
8672 false_reg->u32_min_value = max(false_reg->u32_min_value,
8674 true_reg->u32_max_value = min(true_reg->u32_max_value,
8677 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
8678 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
8680 false_reg->umin_value = max(false_reg->umin_value, false_umin);
8681 true_reg->umax_value = min(true_reg->umax_value, true_umax);
8689 s32 false_smin = opcode == BPF_JSLT ? sval32 : sval32 + 1;
8690 s32 true_smax = opcode == BPF_JSLT ? sval32 - 1 : sval32;
8692 false_reg->s32_min_value = max(false_reg->s32_min_value, false_smin);
8693 true_reg->s32_max_value = min(true_reg->s32_max_value, true_smax);
8695 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
8696 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
8698 false_reg->smin_value = max(false_reg->smin_value, false_smin);
8699 true_reg->smax_value = min(true_reg->smax_value, true_smax);
8708 false_reg->var_off = tnum_or(tnum_clear_subreg(false_64off),
8709 tnum_subreg(false_32off));
8710 true_reg->var_off = tnum_or(tnum_clear_subreg(true_64off),
8711 tnum_subreg(true_32off));
8712 __reg_combine_32_into_64(false_reg);
8713 __reg_combine_32_into_64(true_reg);
8715 false_reg->var_off = false_64off;
8716 true_reg->var_off = true_64off;
8717 __reg_combine_64_into_32(false_reg);
8718 __reg_combine_64_into_32(true_reg);
8722 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
8725 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
8726 struct bpf_reg_state *false_reg,
8728 u8 opcode, bool is_jmp32)
8730 opcode = flip_opcode(opcode);
8731 /* This uses zero as "not present in table"; luckily the zero opcode,
8732 * BPF_JA, can't get here.
8735 reg_set_min_max(true_reg, false_reg, val, val32, opcode, is_jmp32);
8738 /* Regs are known to be equal, so intersect their min/max/var_off */
8739 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
8740 struct bpf_reg_state *dst_reg)
8742 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
8743 dst_reg->umin_value);
8744 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
8745 dst_reg->umax_value);
8746 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
8747 dst_reg->smin_value);
8748 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
8749 dst_reg->smax_value);
8750 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
8752 /* We might have learned new bounds from the var_off. */
8753 __update_reg_bounds(src_reg);
8754 __update_reg_bounds(dst_reg);
8755 /* We might have learned something about the sign bit. */
8756 __reg_deduce_bounds(src_reg);
8757 __reg_deduce_bounds(dst_reg);
8758 /* We might have learned some bits from the bounds. */
8759 __reg_bound_offset(src_reg);
8760 __reg_bound_offset(dst_reg);
8761 /* Intersecting with the old var_off might have improved our bounds
8762 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
8763 * then new var_off is (0; 0x7f...fc) which improves our umax.
8765 __update_reg_bounds(src_reg);
8766 __update_reg_bounds(dst_reg);
8769 static void reg_combine_min_max(struct bpf_reg_state *true_src,
8770 struct bpf_reg_state *true_dst,
8771 struct bpf_reg_state *false_src,
8772 struct bpf_reg_state *false_dst,
8777 __reg_combine_min_max(true_src, true_dst);
8780 __reg_combine_min_max(false_src, false_dst);
8785 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
8786 struct bpf_reg_state *reg, u32 id,
8789 if (type_may_be_null(reg->type) && reg->id == id &&
8790 !WARN_ON_ONCE(!reg->id)) {
8791 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
8792 !tnum_equals_const(reg->var_off, 0) ||
8794 /* Old offset (both fixed and variable parts) should
8795 * have been known-zero, because we don't allow pointer
8796 * arithmetic on pointers that might be NULL. If we
8797 * see this happening, don't convert the register.
8802 reg->type = SCALAR_VALUE;
8803 /* We don't need id and ref_obj_id from this point
8804 * onwards anymore, thus we should better reset it,
8805 * so that state pruning has chances to take effect.
8808 reg->ref_obj_id = 0;
8813 mark_ptr_not_null_reg(reg);
8815 if (!reg_may_point_to_spin_lock(reg)) {
8816 /* For not-NULL ptr, reg->ref_obj_id will be reset
8817 * in release_reg_references().
8819 * reg->id is still used by spin_lock ptr. Other
8820 * than spin_lock ptr type, reg->id can be reset.
8827 static void __mark_ptr_or_null_regs(struct bpf_func_state *state, u32 id,
8830 struct bpf_reg_state *reg;
8833 for (i = 0; i < MAX_BPF_REG; i++)
8834 mark_ptr_or_null_reg(state, &state->regs[i], id, is_null);
8836 bpf_for_each_spilled_reg(i, state, reg) {
8839 mark_ptr_or_null_reg(state, reg, id, is_null);
8843 /* The logic is similar to find_good_pkt_pointers(), both could eventually
8844 * be folded together at some point.
8846 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
8849 struct bpf_func_state *state = vstate->frame[vstate->curframe];
8850 struct bpf_reg_state *regs = state->regs;
8851 u32 ref_obj_id = regs[regno].ref_obj_id;
8852 u32 id = regs[regno].id;
8855 if (ref_obj_id && ref_obj_id == id && is_null)
8856 /* regs[regno] is in the " == NULL" branch.
8857 * No one could have freed the reference state before
8858 * doing the NULL check.
8860 WARN_ON_ONCE(release_reference_state(state, id));
8862 for (i = 0; i <= vstate->curframe; i++)
8863 __mark_ptr_or_null_regs(vstate->frame[i], id, is_null);
8866 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
8867 struct bpf_reg_state *dst_reg,
8868 struct bpf_reg_state *src_reg,
8869 struct bpf_verifier_state *this_branch,
8870 struct bpf_verifier_state *other_branch)
8872 if (BPF_SRC(insn->code) != BPF_X)
8875 /* Pointers are always 64-bit. */
8876 if (BPF_CLASS(insn->code) == BPF_JMP32)
8879 switch (BPF_OP(insn->code)) {
8881 if ((dst_reg->type == PTR_TO_PACKET &&
8882 src_reg->type == PTR_TO_PACKET_END) ||
8883 (dst_reg->type == PTR_TO_PACKET_META &&
8884 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
8885 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
8886 find_good_pkt_pointers(this_branch, dst_reg,
8887 dst_reg->type, false);
8888 mark_pkt_end(other_branch, insn->dst_reg, true);
8889 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
8890 src_reg->type == PTR_TO_PACKET) ||
8891 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
8892 src_reg->type == PTR_TO_PACKET_META)) {
8893 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
8894 find_good_pkt_pointers(other_branch, src_reg,
8895 src_reg->type, true);
8896 mark_pkt_end(this_branch, insn->src_reg, false);
8902 if ((dst_reg->type == PTR_TO_PACKET &&
8903 src_reg->type == PTR_TO_PACKET_END) ||
8904 (dst_reg->type == PTR_TO_PACKET_META &&
8905 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
8906 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
8907 find_good_pkt_pointers(other_branch, dst_reg,
8908 dst_reg->type, true);
8909 mark_pkt_end(this_branch, insn->dst_reg, false);
8910 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
8911 src_reg->type == PTR_TO_PACKET) ||
8912 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
8913 src_reg->type == PTR_TO_PACKET_META)) {
8914 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
8915 find_good_pkt_pointers(this_branch, src_reg,
8916 src_reg->type, false);
8917 mark_pkt_end(other_branch, insn->src_reg, true);
8923 if ((dst_reg->type == PTR_TO_PACKET &&
8924 src_reg->type == PTR_TO_PACKET_END) ||
8925 (dst_reg->type == PTR_TO_PACKET_META &&
8926 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
8927 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
8928 find_good_pkt_pointers(this_branch, dst_reg,
8929 dst_reg->type, true);
8930 mark_pkt_end(other_branch, insn->dst_reg, false);
8931 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
8932 src_reg->type == PTR_TO_PACKET) ||
8933 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
8934 src_reg->type == PTR_TO_PACKET_META)) {
8935 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
8936 find_good_pkt_pointers(other_branch, src_reg,
8937 src_reg->type, false);
8938 mark_pkt_end(this_branch, insn->src_reg, true);
8944 if ((dst_reg->type == PTR_TO_PACKET &&
8945 src_reg->type == PTR_TO_PACKET_END) ||
8946 (dst_reg->type == PTR_TO_PACKET_META &&
8947 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
8948 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
8949 find_good_pkt_pointers(other_branch, dst_reg,
8950 dst_reg->type, false);
8951 mark_pkt_end(this_branch, insn->dst_reg, true);
8952 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
8953 src_reg->type == PTR_TO_PACKET) ||
8954 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
8955 src_reg->type == PTR_TO_PACKET_META)) {
8956 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
8957 find_good_pkt_pointers(this_branch, src_reg,
8958 src_reg->type, true);
8959 mark_pkt_end(other_branch, insn->src_reg, false);
8971 static void find_equal_scalars(struct bpf_verifier_state *vstate,
8972 struct bpf_reg_state *known_reg)
8974 struct bpf_func_state *state;
8975 struct bpf_reg_state *reg;
8978 for (i = 0; i <= vstate->curframe; i++) {
8979 state = vstate->frame[i];
8980 for (j = 0; j < MAX_BPF_REG; j++) {
8981 reg = &state->regs[j];
8982 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
8986 bpf_for_each_spilled_reg(j, state, reg) {
8989 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
8995 static int check_cond_jmp_op(struct bpf_verifier_env *env,
8996 struct bpf_insn *insn, int *insn_idx)
8998 struct bpf_verifier_state *this_branch = env->cur_state;
8999 struct bpf_verifier_state *other_branch;
9000 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
9001 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
9002 u8 opcode = BPF_OP(insn->code);
9007 /* Only conditional jumps are expected to reach here. */
9008 if (opcode == BPF_JA || opcode > BPF_JSLE) {
9009 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
9013 if (BPF_SRC(insn->code) == BPF_X) {
9014 if (insn->imm != 0) {
9015 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
9019 /* check src1 operand */
9020 err = check_reg_arg(env, insn->src_reg, SRC_OP);
9024 if (is_pointer_value(env, insn->src_reg)) {
9025 verbose(env, "R%d pointer comparison prohibited\n",
9029 src_reg = ®s[insn->src_reg];
9031 if (insn->src_reg != BPF_REG_0) {
9032 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
9037 /* check src2 operand */
9038 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
9042 dst_reg = ®s[insn->dst_reg];
9043 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
9045 if (BPF_SRC(insn->code) == BPF_K) {
9046 pred = is_branch_taken(dst_reg, insn->imm, opcode, is_jmp32);
9047 } else if (src_reg->type == SCALAR_VALUE &&
9048 is_jmp32 && tnum_is_const(tnum_subreg(src_reg->var_off))) {
9049 pred = is_branch_taken(dst_reg,
9050 tnum_subreg(src_reg->var_off).value,
9053 } else if (src_reg->type == SCALAR_VALUE &&
9054 !is_jmp32 && tnum_is_const(src_reg->var_off)) {
9055 pred = is_branch_taken(dst_reg,
9056 src_reg->var_off.value,
9059 } else if (reg_is_pkt_pointer_any(dst_reg) &&
9060 reg_is_pkt_pointer_any(src_reg) &&
9062 pred = is_pkt_ptr_branch_taken(dst_reg, src_reg, opcode);
9066 /* If we get here with a dst_reg pointer type it is because
9067 * above is_branch_taken() special cased the 0 comparison.
9069 if (!__is_pointer_value(false, dst_reg))
9070 err = mark_chain_precision(env, insn->dst_reg);
9071 if (BPF_SRC(insn->code) == BPF_X && !err &&
9072 !__is_pointer_value(false, src_reg))
9073 err = mark_chain_precision(env, insn->src_reg);
9079 /* Only follow the goto, ignore fall-through. If needed, push
9080 * the fall-through branch for simulation under speculative
9083 if (!env->bypass_spec_v1 &&
9084 !sanitize_speculative_path(env, insn, *insn_idx + 1,
9087 *insn_idx += insn->off;
9089 } else if (pred == 0) {
9090 /* Only follow the fall-through branch, since that's where the
9091 * program will go. If needed, push the goto branch for
9092 * simulation under speculative execution.
9094 if (!env->bypass_spec_v1 &&
9095 !sanitize_speculative_path(env, insn,
9096 *insn_idx + insn->off + 1,
9102 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
9106 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
9108 /* detect if we are comparing against a constant value so we can adjust
9109 * our min/max values for our dst register.
9110 * this is only legit if both are scalars (or pointers to the same
9111 * object, I suppose, but we don't support that right now), because
9112 * otherwise the different base pointers mean the offsets aren't
9115 if (BPF_SRC(insn->code) == BPF_X) {
9116 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
9118 if (dst_reg->type == SCALAR_VALUE &&
9119 src_reg->type == SCALAR_VALUE) {
9120 if (tnum_is_const(src_reg->var_off) ||
9122 tnum_is_const(tnum_subreg(src_reg->var_off))))
9123 reg_set_min_max(&other_branch_regs[insn->dst_reg],
9125 src_reg->var_off.value,
9126 tnum_subreg(src_reg->var_off).value,
9128 else if (tnum_is_const(dst_reg->var_off) ||
9130 tnum_is_const(tnum_subreg(dst_reg->var_off))))
9131 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
9133 dst_reg->var_off.value,
9134 tnum_subreg(dst_reg->var_off).value,
9136 else if (!is_jmp32 &&
9137 (opcode == BPF_JEQ || opcode == BPF_JNE))
9138 /* Comparing for equality, we can combine knowledge */
9139 reg_combine_min_max(&other_branch_regs[insn->src_reg],
9140 &other_branch_regs[insn->dst_reg],
9141 src_reg, dst_reg, opcode);
9143 !WARN_ON_ONCE(src_reg->id != other_branch_regs[insn->src_reg].id)) {
9144 find_equal_scalars(this_branch, src_reg);
9145 find_equal_scalars(other_branch, &other_branch_regs[insn->src_reg]);
9149 } else if (dst_reg->type == SCALAR_VALUE) {
9150 reg_set_min_max(&other_branch_regs[insn->dst_reg],
9151 dst_reg, insn->imm, (u32)insn->imm,
9155 if (dst_reg->type == SCALAR_VALUE && dst_reg->id &&
9156 !WARN_ON_ONCE(dst_reg->id != other_branch_regs[insn->dst_reg].id)) {
9157 find_equal_scalars(this_branch, dst_reg);
9158 find_equal_scalars(other_branch, &other_branch_regs[insn->dst_reg]);
9161 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
9162 * NOTE: these optimizations below are related with pointer comparison
9163 * which will never be JMP32.
9165 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
9166 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
9167 type_may_be_null(dst_reg->type)) {
9168 /* Mark all identical registers in each branch as either
9169 * safe or unknown depending R == 0 or R != 0 conditional.
9171 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
9173 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
9175 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
9176 this_branch, other_branch) &&
9177 is_pointer_value(env, insn->dst_reg)) {
9178 verbose(env, "R%d pointer comparison prohibited\n",
9182 if (env->log.level & BPF_LOG_LEVEL)
9183 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
9187 /* verify BPF_LD_IMM64 instruction */
9188 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
9190 struct bpf_insn_aux_data *aux = cur_aux(env);
9191 struct bpf_reg_state *regs = cur_regs(env);
9192 struct bpf_reg_state *dst_reg;
9193 struct bpf_map *map;
9196 if (BPF_SIZE(insn->code) != BPF_DW) {
9197 verbose(env, "invalid BPF_LD_IMM insn\n");
9200 if (insn->off != 0) {
9201 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
9205 err = check_reg_arg(env, insn->dst_reg, DST_OP);
9209 dst_reg = ®s[insn->dst_reg];
9210 if (insn->src_reg == 0) {
9211 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
9213 dst_reg->type = SCALAR_VALUE;
9214 __mark_reg_known(®s[insn->dst_reg], imm);
9218 /* All special src_reg cases are listed below. From this point onwards
9219 * we either succeed and assign a corresponding dst_reg->type after
9220 * zeroing the offset, or fail and reject the program.
9222 mark_reg_known_zero(env, regs, insn->dst_reg);
9224 if (insn->src_reg == BPF_PSEUDO_BTF_ID) {
9225 dst_reg->type = aux->btf_var.reg_type;
9226 switch (base_type(dst_reg->type)) {
9228 dst_reg->mem_size = aux->btf_var.mem_size;
9231 case PTR_TO_PERCPU_BTF_ID:
9232 dst_reg->btf = aux->btf_var.btf;
9233 dst_reg->btf_id = aux->btf_var.btf_id;
9236 verbose(env, "bpf verifier is misconfigured\n");
9242 if (insn->src_reg == BPF_PSEUDO_FUNC) {
9243 struct bpf_prog_aux *aux = env->prog->aux;
9244 u32 subprogno = insn[1].imm;
9246 if (!aux->func_info) {
9247 verbose(env, "missing btf func_info\n");
9250 if (aux->func_info_aux[subprogno].linkage != BTF_FUNC_STATIC) {
9251 verbose(env, "callback function not static\n");
9255 dst_reg->type = PTR_TO_FUNC;
9256 dst_reg->subprogno = subprogno;
9260 map = env->used_maps[aux->map_index];
9261 dst_reg->map_ptr = map;
9263 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE ||
9264 insn->src_reg == BPF_PSEUDO_MAP_IDX_VALUE) {
9265 dst_reg->type = PTR_TO_MAP_VALUE;
9266 dst_reg->off = aux->map_off;
9267 if (map_value_has_spin_lock(map))
9268 dst_reg->id = ++env->id_gen;
9269 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD ||
9270 insn->src_reg == BPF_PSEUDO_MAP_IDX) {
9271 dst_reg->type = CONST_PTR_TO_MAP;
9273 verbose(env, "bpf verifier is misconfigured\n");
9280 static bool may_access_skb(enum bpf_prog_type type)
9283 case BPF_PROG_TYPE_SOCKET_FILTER:
9284 case BPF_PROG_TYPE_SCHED_CLS:
9285 case BPF_PROG_TYPE_SCHED_ACT:
9292 /* verify safety of LD_ABS|LD_IND instructions:
9293 * - they can only appear in the programs where ctx == skb
9294 * - since they are wrappers of function calls, they scratch R1-R5 registers,
9295 * preserve R6-R9, and store return value into R0
9298 * ctx == skb == R6 == CTX
9301 * SRC == any register
9302 * IMM == 32-bit immediate
9305 * R0 - 8/16/32-bit skb data converted to cpu endianness
9307 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
9309 struct bpf_reg_state *regs = cur_regs(env);
9310 static const int ctx_reg = BPF_REG_6;
9311 u8 mode = BPF_MODE(insn->code);
9314 if (!may_access_skb(resolve_prog_type(env->prog))) {
9315 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
9319 if (!env->ops->gen_ld_abs) {
9320 verbose(env, "bpf verifier is misconfigured\n");
9324 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
9325 BPF_SIZE(insn->code) == BPF_DW ||
9326 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
9327 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
9331 /* check whether implicit source operand (register R6) is readable */
9332 err = check_reg_arg(env, ctx_reg, SRC_OP);
9336 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
9337 * gen_ld_abs() may terminate the program at runtime, leading to
9340 err = check_reference_leak(env);
9342 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
9346 if (env->cur_state->active_spin_lock) {
9347 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
9351 if (regs[ctx_reg].type != PTR_TO_CTX) {
9353 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
9357 if (mode == BPF_IND) {
9358 /* check explicit source operand */
9359 err = check_reg_arg(env, insn->src_reg, SRC_OP);
9364 err = check_ctx_reg(env, ®s[ctx_reg], ctx_reg);
9368 /* reset caller saved regs to unreadable */
9369 for (i = 0; i < CALLER_SAVED_REGS; i++) {
9370 mark_reg_not_init(env, regs, caller_saved[i]);
9371 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
9374 /* mark destination R0 register as readable, since it contains
9375 * the value fetched from the packet.
9376 * Already marked as written above.
9378 mark_reg_unknown(env, regs, BPF_REG_0);
9379 /* ld_abs load up to 32-bit skb data. */
9380 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
9384 static int check_return_code(struct bpf_verifier_env *env)
9386 struct tnum enforce_attach_type_range = tnum_unknown;
9387 const struct bpf_prog *prog = env->prog;
9388 struct bpf_reg_state *reg;
9389 struct tnum range = tnum_range(0, 1);
9390 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
9392 struct bpf_func_state *frame = env->cur_state->frame[0];
9393 const bool is_subprog = frame->subprogno;
9395 /* LSM and struct_ops func-ptr's return type could be "void" */
9397 (prog_type == BPF_PROG_TYPE_STRUCT_OPS ||
9398 prog_type == BPF_PROG_TYPE_LSM) &&
9399 !prog->aux->attach_func_proto->type)
9402 /* eBPF calling convention is such that R0 is used
9403 * to return the value from eBPF program.
9404 * Make sure that it's readable at this time
9405 * of bpf_exit, which means that program wrote
9406 * something into it earlier
9408 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
9412 if (is_pointer_value(env, BPF_REG_0)) {
9413 verbose(env, "R0 leaks addr as return value\n");
9417 reg = cur_regs(env) + BPF_REG_0;
9419 if (frame->in_async_callback_fn) {
9420 /* enforce return zero from async callbacks like timer */
9421 if (reg->type != SCALAR_VALUE) {
9422 verbose(env, "In async callback the register R0 is not a known value (%s)\n",
9423 reg_type_str(env, reg->type));
9427 if (!tnum_in(tnum_const(0), reg->var_off)) {
9428 verbose_invalid_scalar(env, reg, &range, "async callback", "R0");
9435 if (reg->type != SCALAR_VALUE) {
9436 verbose(env, "At subprogram exit the register R0 is not a scalar value (%s)\n",
9437 reg_type_str(env, reg->type));
9443 switch (prog_type) {
9444 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
9445 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
9446 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG ||
9447 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETPEERNAME ||
9448 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETPEERNAME ||
9449 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETSOCKNAME ||
9450 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETSOCKNAME)
9451 range = tnum_range(1, 1);
9452 if (env->prog->expected_attach_type == BPF_CGROUP_INET4_BIND ||
9453 env->prog->expected_attach_type == BPF_CGROUP_INET6_BIND)
9454 range = tnum_range(0, 3);
9456 case BPF_PROG_TYPE_CGROUP_SKB:
9457 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
9458 range = tnum_range(0, 3);
9459 enforce_attach_type_range = tnum_range(2, 3);
9462 case BPF_PROG_TYPE_CGROUP_SOCK:
9463 case BPF_PROG_TYPE_SOCK_OPS:
9464 case BPF_PROG_TYPE_CGROUP_DEVICE:
9465 case BPF_PROG_TYPE_CGROUP_SYSCTL:
9466 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
9468 case BPF_PROG_TYPE_RAW_TRACEPOINT:
9469 if (!env->prog->aux->attach_btf_id)
9471 range = tnum_const(0);
9473 case BPF_PROG_TYPE_TRACING:
9474 switch (env->prog->expected_attach_type) {
9475 case BPF_TRACE_FENTRY:
9476 case BPF_TRACE_FEXIT:
9477 range = tnum_const(0);
9479 case BPF_TRACE_RAW_TP:
9480 case BPF_MODIFY_RETURN:
9482 case BPF_TRACE_ITER:
9488 case BPF_PROG_TYPE_SK_LOOKUP:
9489 range = tnum_range(SK_DROP, SK_PASS);
9491 case BPF_PROG_TYPE_EXT:
9492 /* freplace program can return anything as its return value
9493 * depends on the to-be-replaced kernel func or bpf program.
9499 if (reg->type != SCALAR_VALUE) {
9500 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
9501 reg_type_str(env, reg->type));
9505 if (!tnum_in(range, reg->var_off)) {
9506 verbose_invalid_scalar(env, reg, &range, "program exit", "R0");
9510 if (!tnum_is_unknown(enforce_attach_type_range) &&
9511 tnum_in(enforce_attach_type_range, reg->var_off))
9512 env->prog->enforce_expected_attach_type = 1;
9516 /* non-recursive DFS pseudo code
9517 * 1 procedure DFS-iterative(G,v):
9518 * 2 label v as discovered
9519 * 3 let S be a stack
9521 * 5 while S is not empty
9523 * 7 if t is what we're looking for:
9525 * 9 for all edges e in G.adjacentEdges(t) do
9526 * 10 if edge e is already labelled
9527 * 11 continue with the next edge
9528 * 12 w <- G.adjacentVertex(t,e)
9529 * 13 if vertex w is not discovered and not explored
9530 * 14 label e as tree-edge
9531 * 15 label w as discovered
9534 * 18 else if vertex w is discovered
9535 * 19 label e as back-edge
9537 * 21 // vertex w is explored
9538 * 22 label e as forward- or cross-edge
9539 * 23 label t as explored
9544 * 0x11 - discovered and fall-through edge labelled
9545 * 0x12 - discovered and fall-through and branch edges labelled
9556 static u32 state_htab_size(struct bpf_verifier_env *env)
9558 return env->prog->len;
9561 static struct bpf_verifier_state_list **explored_state(
9562 struct bpf_verifier_env *env,
9565 struct bpf_verifier_state *cur = env->cur_state;
9566 struct bpf_func_state *state = cur->frame[cur->curframe];
9568 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
9571 static void init_explored_state(struct bpf_verifier_env *env, int idx)
9573 env->insn_aux_data[idx].prune_point = true;
9581 /* t, w, e - match pseudo-code above:
9582 * t - index of current instruction
9583 * w - next instruction
9586 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env,
9589 int *insn_stack = env->cfg.insn_stack;
9590 int *insn_state = env->cfg.insn_state;
9592 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
9593 return DONE_EXPLORING;
9595 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
9596 return DONE_EXPLORING;
9598 if (w < 0 || w >= env->prog->len) {
9599 verbose_linfo(env, t, "%d: ", t);
9600 verbose(env, "jump out of range from insn %d to %d\n", t, w);
9605 /* mark branch target for state pruning */
9606 init_explored_state(env, w);
9608 if (insn_state[w] == 0) {
9610 insn_state[t] = DISCOVERED | e;
9611 insn_state[w] = DISCOVERED;
9612 if (env->cfg.cur_stack >= env->prog->len)
9614 insn_stack[env->cfg.cur_stack++] = w;
9615 return KEEP_EXPLORING;
9616 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
9617 if (loop_ok && env->bpf_capable)
9618 return DONE_EXPLORING;
9619 verbose_linfo(env, t, "%d: ", t);
9620 verbose_linfo(env, w, "%d: ", w);
9621 verbose(env, "back-edge from insn %d to %d\n", t, w);
9623 } else if (insn_state[w] == EXPLORED) {
9624 /* forward- or cross-edge */
9625 insn_state[t] = DISCOVERED | e;
9627 verbose(env, "insn state internal bug\n");
9630 return DONE_EXPLORING;
9633 static int visit_func_call_insn(int t, int insn_cnt,
9634 struct bpf_insn *insns,
9635 struct bpf_verifier_env *env,
9640 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
9644 if (t + 1 < insn_cnt)
9645 init_explored_state(env, t + 1);
9647 init_explored_state(env, t);
9648 ret = push_insn(t, t + insns[t].imm + 1, BRANCH, env,
9649 /* It's ok to allow recursion from CFG point of
9650 * view. __check_func_call() will do the actual
9653 bpf_pseudo_func(insns + t));
9658 /* Visits the instruction at index t and returns one of the following:
9659 * < 0 - an error occurred
9660 * DONE_EXPLORING - the instruction was fully explored
9661 * KEEP_EXPLORING - there is still work to be done before it is fully explored
9663 static int visit_insn(int t, int insn_cnt, struct bpf_verifier_env *env)
9665 struct bpf_insn *insns = env->prog->insnsi;
9668 if (bpf_pseudo_func(insns + t))
9669 return visit_func_call_insn(t, insn_cnt, insns, env, true);
9671 /* All non-branch instructions have a single fall-through edge. */
9672 if (BPF_CLASS(insns[t].code) != BPF_JMP &&
9673 BPF_CLASS(insns[t].code) != BPF_JMP32)
9674 return push_insn(t, t + 1, FALLTHROUGH, env, false);
9676 switch (BPF_OP(insns[t].code)) {
9678 return DONE_EXPLORING;
9681 if (insns[t].imm == BPF_FUNC_timer_set_callback)
9682 /* Mark this call insn to trigger is_state_visited() check
9683 * before call itself is processed by __check_func_call().
9684 * Otherwise new async state will be pushed for further
9687 init_explored_state(env, t);
9688 return visit_func_call_insn(t, insn_cnt, insns, env,
9689 insns[t].src_reg == BPF_PSEUDO_CALL);
9692 if (BPF_SRC(insns[t].code) != BPF_K)
9695 /* unconditional jump with single edge */
9696 ret = push_insn(t, t + insns[t].off + 1, FALLTHROUGH, env,
9701 /* unconditional jmp is not a good pruning point,
9702 * but it's marked, since backtracking needs
9703 * to record jmp history in is_state_visited().
9705 init_explored_state(env, t + insns[t].off + 1);
9706 /* tell verifier to check for equivalent states
9707 * after every call and jump
9709 if (t + 1 < insn_cnt)
9710 init_explored_state(env, t + 1);
9715 /* conditional jump with two edges */
9716 init_explored_state(env, t);
9717 ret = push_insn(t, t + 1, FALLTHROUGH, env, true);
9721 return push_insn(t, t + insns[t].off + 1, BRANCH, env, true);
9725 /* non-recursive depth-first-search to detect loops in BPF program
9726 * loop == back-edge in directed graph
9728 static int check_cfg(struct bpf_verifier_env *env)
9730 int insn_cnt = env->prog->len;
9731 int *insn_stack, *insn_state;
9735 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
9739 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
9745 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
9746 insn_stack[0] = 0; /* 0 is the first instruction */
9747 env->cfg.cur_stack = 1;
9749 while (env->cfg.cur_stack > 0) {
9750 int t = insn_stack[env->cfg.cur_stack - 1];
9752 ret = visit_insn(t, insn_cnt, env);
9754 case DONE_EXPLORING:
9755 insn_state[t] = EXPLORED;
9756 env->cfg.cur_stack--;
9758 case KEEP_EXPLORING:
9762 verbose(env, "visit_insn internal bug\n");
9769 if (env->cfg.cur_stack < 0) {
9770 verbose(env, "pop stack internal bug\n");
9775 for (i = 0; i < insn_cnt; i++) {
9776 if (insn_state[i] != EXPLORED) {
9777 verbose(env, "unreachable insn %d\n", i);
9782 ret = 0; /* cfg looks good */
9787 env->cfg.insn_state = env->cfg.insn_stack = NULL;
9791 static int check_abnormal_return(struct bpf_verifier_env *env)
9795 for (i = 1; i < env->subprog_cnt; i++) {
9796 if (env->subprog_info[i].has_ld_abs) {
9797 verbose(env, "LD_ABS is not allowed in subprogs without BTF\n");
9800 if (env->subprog_info[i].has_tail_call) {
9801 verbose(env, "tail_call is not allowed in subprogs without BTF\n");
9808 /* The minimum supported BTF func info size */
9809 #define MIN_BPF_FUNCINFO_SIZE 8
9810 #define MAX_FUNCINFO_REC_SIZE 252
9812 static int check_btf_func(struct bpf_verifier_env *env,
9813 const union bpf_attr *attr,
9816 const struct btf_type *type, *func_proto, *ret_type;
9817 u32 i, nfuncs, urec_size, min_size;
9818 u32 krec_size = sizeof(struct bpf_func_info);
9819 struct bpf_func_info *krecord;
9820 struct bpf_func_info_aux *info_aux = NULL;
9821 struct bpf_prog *prog;
9822 const struct btf *btf;
9824 u32 prev_offset = 0;
9828 nfuncs = attr->func_info_cnt;
9830 if (check_abnormal_return(env))
9835 if (nfuncs != env->subprog_cnt) {
9836 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
9840 urec_size = attr->func_info_rec_size;
9841 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
9842 urec_size > MAX_FUNCINFO_REC_SIZE ||
9843 urec_size % sizeof(u32)) {
9844 verbose(env, "invalid func info rec size %u\n", urec_size);
9849 btf = prog->aux->btf;
9851 urecord = make_bpfptr(attr->func_info, uattr.is_kernel);
9852 min_size = min_t(u32, krec_size, urec_size);
9854 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
9857 info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN);
9861 for (i = 0; i < nfuncs; i++) {
9862 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
9864 if (ret == -E2BIG) {
9865 verbose(env, "nonzero tailing record in func info");
9866 /* set the size kernel expects so loader can zero
9867 * out the rest of the record.
9869 if (copy_to_bpfptr_offset(uattr,
9870 offsetof(union bpf_attr, func_info_rec_size),
9871 &min_size, sizeof(min_size)))
9877 if (copy_from_bpfptr(&krecord[i], urecord, min_size)) {
9882 /* check insn_off */
9885 if (krecord[i].insn_off) {
9887 "nonzero insn_off %u for the first func info record",
9888 krecord[i].insn_off);
9891 } else if (krecord[i].insn_off <= prev_offset) {
9893 "same or smaller insn offset (%u) than previous func info record (%u)",
9894 krecord[i].insn_off, prev_offset);
9898 if (env->subprog_info[i].start != krecord[i].insn_off) {
9899 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
9904 type = btf_type_by_id(btf, krecord[i].type_id);
9905 if (!type || !btf_type_is_func(type)) {
9906 verbose(env, "invalid type id %d in func info",
9907 krecord[i].type_id);
9910 info_aux[i].linkage = BTF_INFO_VLEN(type->info);
9912 func_proto = btf_type_by_id(btf, type->type);
9913 if (unlikely(!func_proto || !btf_type_is_func_proto(func_proto)))
9914 /* btf_func_check() already verified it during BTF load */
9916 ret_type = btf_type_skip_modifiers(btf, func_proto->type, NULL);
9918 btf_type_is_small_int(ret_type) || btf_type_is_enum(ret_type);
9919 if (i && !scalar_return && env->subprog_info[i].has_ld_abs) {
9920 verbose(env, "LD_ABS is only allowed in functions that return 'int'.\n");
9923 if (i && !scalar_return && env->subprog_info[i].has_tail_call) {
9924 verbose(env, "tail_call is only allowed in functions that return 'int'.\n");
9928 prev_offset = krecord[i].insn_off;
9929 bpfptr_add(&urecord, urec_size);
9932 prog->aux->func_info = krecord;
9933 prog->aux->func_info_cnt = nfuncs;
9934 prog->aux->func_info_aux = info_aux;
9943 static void adjust_btf_func(struct bpf_verifier_env *env)
9945 struct bpf_prog_aux *aux = env->prog->aux;
9948 if (!aux->func_info)
9951 for (i = 0; i < env->subprog_cnt; i++)
9952 aux->func_info[i].insn_off = env->subprog_info[i].start;
9955 #define MIN_BPF_LINEINFO_SIZE (offsetof(struct bpf_line_info, line_col) + \
9956 sizeof(((struct bpf_line_info *)(0))->line_col))
9957 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
9959 static int check_btf_line(struct bpf_verifier_env *env,
9960 const union bpf_attr *attr,
9963 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
9964 struct bpf_subprog_info *sub;
9965 struct bpf_line_info *linfo;
9966 struct bpf_prog *prog;
9967 const struct btf *btf;
9971 nr_linfo = attr->line_info_cnt;
9974 if (nr_linfo > INT_MAX / sizeof(struct bpf_line_info))
9977 rec_size = attr->line_info_rec_size;
9978 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
9979 rec_size > MAX_LINEINFO_REC_SIZE ||
9980 rec_size & (sizeof(u32) - 1))
9983 /* Need to zero it in case the userspace may
9984 * pass in a smaller bpf_line_info object.
9986 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
9987 GFP_KERNEL | __GFP_NOWARN);
9992 btf = prog->aux->btf;
9995 sub = env->subprog_info;
9996 ulinfo = make_bpfptr(attr->line_info, uattr.is_kernel);
9997 expected_size = sizeof(struct bpf_line_info);
9998 ncopy = min_t(u32, expected_size, rec_size);
9999 for (i = 0; i < nr_linfo; i++) {
10000 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
10002 if (err == -E2BIG) {
10003 verbose(env, "nonzero tailing record in line_info");
10004 if (copy_to_bpfptr_offset(uattr,
10005 offsetof(union bpf_attr, line_info_rec_size),
10006 &expected_size, sizeof(expected_size)))
10012 if (copy_from_bpfptr(&linfo[i], ulinfo, ncopy)) {
10018 * Check insn_off to ensure
10019 * 1) strictly increasing AND
10020 * 2) bounded by prog->len
10022 * The linfo[0].insn_off == 0 check logically falls into
10023 * the later "missing bpf_line_info for func..." case
10024 * because the first linfo[0].insn_off must be the
10025 * first sub also and the first sub must have
10026 * subprog_info[0].start == 0.
10028 if ((i && linfo[i].insn_off <= prev_offset) ||
10029 linfo[i].insn_off >= prog->len) {
10030 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
10031 i, linfo[i].insn_off, prev_offset,
10037 if (!prog->insnsi[linfo[i].insn_off].code) {
10039 "Invalid insn code at line_info[%u].insn_off\n",
10045 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
10046 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
10047 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
10052 if (s != env->subprog_cnt) {
10053 if (linfo[i].insn_off == sub[s].start) {
10054 sub[s].linfo_idx = i;
10056 } else if (sub[s].start < linfo[i].insn_off) {
10057 verbose(env, "missing bpf_line_info for func#%u\n", s);
10063 prev_offset = linfo[i].insn_off;
10064 bpfptr_add(&ulinfo, rec_size);
10067 if (s != env->subprog_cnt) {
10068 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
10069 env->subprog_cnt - s, s);
10074 prog->aux->linfo = linfo;
10075 prog->aux->nr_linfo = nr_linfo;
10084 static int check_btf_info(struct bpf_verifier_env *env,
10085 const union bpf_attr *attr,
10091 if (!attr->func_info_cnt && !attr->line_info_cnt) {
10092 if (check_abnormal_return(env))
10097 btf = btf_get_by_fd(attr->prog_btf_fd);
10099 return PTR_ERR(btf);
10100 if (btf_is_kernel(btf)) {
10104 env->prog->aux->btf = btf;
10106 err = check_btf_func(env, attr, uattr);
10110 err = check_btf_line(env, attr, uattr);
10117 /* check %cur's range satisfies %old's */
10118 static bool range_within(struct bpf_reg_state *old,
10119 struct bpf_reg_state *cur)
10121 return old->umin_value <= cur->umin_value &&
10122 old->umax_value >= cur->umax_value &&
10123 old->smin_value <= cur->smin_value &&
10124 old->smax_value >= cur->smax_value &&
10125 old->u32_min_value <= cur->u32_min_value &&
10126 old->u32_max_value >= cur->u32_max_value &&
10127 old->s32_min_value <= cur->s32_min_value &&
10128 old->s32_max_value >= cur->s32_max_value;
10131 /* If in the old state two registers had the same id, then they need to have
10132 * the same id in the new state as well. But that id could be different from
10133 * the old state, so we need to track the mapping from old to new ids.
10134 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
10135 * regs with old id 5 must also have new id 9 for the new state to be safe. But
10136 * regs with a different old id could still have new id 9, we don't care about
10138 * So we look through our idmap to see if this old id has been seen before. If
10139 * so, we require the new id to match; otherwise, we add the id pair to the map.
10141 static bool check_ids(u32 old_id, u32 cur_id, struct bpf_id_pair *idmap)
10145 for (i = 0; i < BPF_ID_MAP_SIZE; i++) {
10146 if (!idmap[i].old) {
10147 /* Reached an empty slot; haven't seen this id before */
10148 idmap[i].old = old_id;
10149 idmap[i].cur = cur_id;
10152 if (idmap[i].old == old_id)
10153 return idmap[i].cur == cur_id;
10155 /* We ran out of idmap slots, which should be impossible */
10160 static void clean_func_state(struct bpf_verifier_env *env,
10161 struct bpf_func_state *st)
10163 enum bpf_reg_liveness live;
10166 for (i = 0; i < BPF_REG_FP; i++) {
10167 live = st->regs[i].live;
10168 /* liveness must not touch this register anymore */
10169 st->regs[i].live |= REG_LIVE_DONE;
10170 if (!(live & REG_LIVE_READ))
10171 /* since the register is unused, clear its state
10172 * to make further comparison simpler
10174 __mark_reg_not_init(env, &st->regs[i]);
10177 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
10178 live = st->stack[i].spilled_ptr.live;
10179 /* liveness must not touch this stack slot anymore */
10180 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
10181 if (!(live & REG_LIVE_READ)) {
10182 __mark_reg_not_init(env, &st->stack[i].spilled_ptr);
10183 for (j = 0; j < BPF_REG_SIZE; j++)
10184 st->stack[i].slot_type[j] = STACK_INVALID;
10189 static void clean_verifier_state(struct bpf_verifier_env *env,
10190 struct bpf_verifier_state *st)
10194 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
10195 /* all regs in this state in all frames were already marked */
10198 for (i = 0; i <= st->curframe; i++)
10199 clean_func_state(env, st->frame[i]);
10202 /* the parentage chains form a tree.
10203 * the verifier states are added to state lists at given insn and
10204 * pushed into state stack for future exploration.
10205 * when the verifier reaches bpf_exit insn some of the verifer states
10206 * stored in the state lists have their final liveness state already,
10207 * but a lot of states will get revised from liveness point of view when
10208 * the verifier explores other branches.
10211 * 2: if r1 == 100 goto pc+1
10214 * when the verifier reaches exit insn the register r0 in the state list of
10215 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
10216 * of insn 2 and goes exploring further. At the insn 4 it will walk the
10217 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
10219 * Since the verifier pushes the branch states as it sees them while exploring
10220 * the program the condition of walking the branch instruction for the second
10221 * time means that all states below this branch were already explored and
10222 * their final liveness marks are already propagated.
10223 * Hence when the verifier completes the search of state list in is_state_visited()
10224 * we can call this clean_live_states() function to mark all liveness states
10225 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
10226 * will not be used.
10227 * This function also clears the registers and stack for states that !READ
10228 * to simplify state merging.
10230 * Important note here that walking the same branch instruction in the callee
10231 * doesn't meant that the states are DONE. The verifier has to compare
10234 static void clean_live_states(struct bpf_verifier_env *env, int insn,
10235 struct bpf_verifier_state *cur)
10237 struct bpf_verifier_state_list *sl;
10240 sl = *explored_state(env, insn);
10242 if (sl->state.branches)
10244 if (sl->state.insn_idx != insn ||
10245 sl->state.curframe != cur->curframe)
10247 for (i = 0; i <= cur->curframe; i++)
10248 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
10250 clean_verifier_state(env, &sl->state);
10256 /* Returns true if (rold safe implies rcur safe) */
10257 static bool regsafe(struct bpf_verifier_env *env, struct bpf_reg_state *rold,
10258 struct bpf_reg_state *rcur, struct bpf_id_pair *idmap)
10262 if (!(rold->live & REG_LIVE_READ))
10263 /* explored state didn't use this */
10266 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
10268 if (rold->type == PTR_TO_STACK)
10269 /* two stack pointers are equal only if they're pointing to
10270 * the same stack frame, since fp-8 in foo != fp-8 in bar
10272 return equal && rold->frameno == rcur->frameno;
10277 if (rold->type == NOT_INIT)
10278 /* explored state can't have used this */
10280 if (rcur->type == NOT_INIT)
10282 switch (base_type(rold->type)) {
10284 if (env->explore_alu_limits)
10286 if (rcur->type == SCALAR_VALUE) {
10287 if (!rold->precise && !rcur->precise)
10289 /* new val must satisfy old val knowledge */
10290 return range_within(rold, rcur) &&
10291 tnum_in(rold->var_off, rcur->var_off);
10293 /* We're trying to use a pointer in place of a scalar.
10294 * Even if the scalar was unbounded, this could lead to
10295 * pointer leaks because scalars are allowed to leak
10296 * while pointers are not. We could make this safe in
10297 * special cases if root is calling us, but it's
10298 * probably not worth the hassle.
10302 case PTR_TO_MAP_KEY:
10303 case PTR_TO_MAP_VALUE:
10304 /* a PTR_TO_MAP_VALUE could be safe to use as a
10305 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
10306 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
10307 * checked, doing so could have affected others with the same
10308 * id, and we can't check for that because we lost the id when
10309 * we converted to a PTR_TO_MAP_VALUE.
10311 if (type_may_be_null(rold->type)) {
10312 if (!type_may_be_null(rcur->type))
10314 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
10316 /* Check our ids match any regs they're supposed to */
10317 return check_ids(rold->id, rcur->id, idmap);
10320 /* If the new min/max/var_off satisfy the old ones and
10321 * everything else matches, we are OK.
10322 * 'id' is not compared, since it's only used for maps with
10323 * bpf_spin_lock inside map element and in such cases if
10324 * the rest of the prog is valid for one map element then
10325 * it's valid for all map elements regardless of the key
10326 * used in bpf_map_lookup()
10328 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
10329 range_within(rold, rcur) &&
10330 tnum_in(rold->var_off, rcur->var_off);
10331 case PTR_TO_PACKET_META:
10332 case PTR_TO_PACKET:
10333 if (rcur->type != rold->type)
10335 /* We must have at least as much range as the old ptr
10336 * did, so that any accesses which were safe before are
10337 * still safe. This is true even if old range < old off,
10338 * since someone could have accessed through (ptr - k), or
10339 * even done ptr -= k in a register, to get a safe access.
10341 if (rold->range > rcur->range)
10343 /* If the offsets don't match, we can't trust our alignment;
10344 * nor can we be sure that we won't fall out of range.
10346 if (rold->off != rcur->off)
10348 /* id relations must be preserved */
10349 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
10351 /* new val must satisfy old val knowledge */
10352 return range_within(rold, rcur) &&
10353 tnum_in(rold->var_off, rcur->var_off);
10355 case CONST_PTR_TO_MAP:
10356 case PTR_TO_PACKET_END:
10357 case PTR_TO_FLOW_KEYS:
10358 case PTR_TO_SOCKET:
10359 case PTR_TO_SOCK_COMMON:
10360 case PTR_TO_TCP_SOCK:
10361 case PTR_TO_XDP_SOCK:
10362 /* Only valid matches are exact, which memcmp() above
10363 * would have accepted
10366 /* Don't know what's going on, just say it's not safe */
10370 /* Shouldn't get here; if we do, say it's not safe */
10375 static bool stacksafe(struct bpf_verifier_env *env, struct bpf_func_state *old,
10376 struct bpf_func_state *cur, struct bpf_id_pair *idmap)
10380 /* walk slots of the explored stack and ignore any additional
10381 * slots in the current stack, since explored(safe) state
10384 for (i = 0; i < old->allocated_stack; i++) {
10385 spi = i / BPF_REG_SIZE;
10387 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
10388 i += BPF_REG_SIZE - 1;
10389 /* explored state didn't use this */
10393 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
10396 /* explored stack has more populated slots than current stack
10397 * and these slots were used
10399 if (i >= cur->allocated_stack)
10402 /* if old state was safe with misc data in the stack
10403 * it will be safe with zero-initialized stack.
10404 * The opposite is not true
10406 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
10407 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
10409 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
10410 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
10411 /* Ex: old explored (safe) state has STACK_SPILL in
10412 * this stack slot, but current has STACK_MISC ->
10413 * this verifier states are not equivalent,
10414 * return false to continue verification of this path
10417 if (i % BPF_REG_SIZE)
10419 if (old->stack[spi].slot_type[0] != STACK_SPILL)
10421 if (!regsafe(env, &old->stack[spi].spilled_ptr,
10422 &cur->stack[spi].spilled_ptr, idmap))
10423 /* when explored and current stack slot are both storing
10424 * spilled registers, check that stored pointers types
10425 * are the same as well.
10426 * Ex: explored safe path could have stored
10427 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
10428 * but current path has stored:
10429 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
10430 * such verifier states are not equivalent.
10431 * return false to continue verification of this path
10438 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
10440 if (old->acquired_refs != cur->acquired_refs)
10442 return !memcmp(old->refs, cur->refs,
10443 sizeof(*old->refs) * old->acquired_refs);
10446 /* compare two verifier states
10448 * all states stored in state_list are known to be valid, since
10449 * verifier reached 'bpf_exit' instruction through them
10451 * this function is called when verifier exploring different branches of
10452 * execution popped from the state stack. If it sees an old state that has
10453 * more strict register state and more strict stack state then this execution
10454 * branch doesn't need to be explored further, since verifier already
10455 * concluded that more strict state leads to valid finish.
10457 * Therefore two states are equivalent if register state is more conservative
10458 * and explored stack state is more conservative than the current one.
10461 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
10462 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
10464 * In other words if current stack state (one being explored) has more
10465 * valid slots than old one that already passed validation, it means
10466 * the verifier can stop exploring and conclude that current state is valid too
10468 * Similarly with registers. If explored state has register type as invalid
10469 * whereas register type in current state is meaningful, it means that
10470 * the current state will reach 'bpf_exit' instruction safely
10472 static bool func_states_equal(struct bpf_verifier_env *env, struct bpf_func_state *old,
10473 struct bpf_func_state *cur)
10477 memset(env->idmap_scratch, 0, sizeof(env->idmap_scratch));
10478 for (i = 0; i < MAX_BPF_REG; i++)
10479 if (!regsafe(env, &old->regs[i], &cur->regs[i],
10480 env->idmap_scratch))
10483 if (!stacksafe(env, old, cur, env->idmap_scratch))
10486 if (!refsafe(old, cur))
10492 static bool states_equal(struct bpf_verifier_env *env,
10493 struct bpf_verifier_state *old,
10494 struct bpf_verifier_state *cur)
10498 if (old->curframe != cur->curframe)
10501 /* Verification state from speculative execution simulation
10502 * must never prune a non-speculative execution one.
10504 if (old->speculative && !cur->speculative)
10507 if (old->active_spin_lock != cur->active_spin_lock)
10510 /* for states to be equal callsites have to be the same
10511 * and all frame states need to be equivalent
10513 for (i = 0; i <= old->curframe; i++) {
10514 if (old->frame[i]->callsite != cur->frame[i]->callsite)
10516 if (!func_states_equal(env, old->frame[i], cur->frame[i]))
10522 /* Return 0 if no propagation happened. Return negative error code if error
10523 * happened. Otherwise, return the propagated bit.
10525 static int propagate_liveness_reg(struct bpf_verifier_env *env,
10526 struct bpf_reg_state *reg,
10527 struct bpf_reg_state *parent_reg)
10529 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
10530 u8 flag = reg->live & REG_LIVE_READ;
10533 /* When comes here, read flags of PARENT_REG or REG could be any of
10534 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
10535 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
10537 if (parent_flag == REG_LIVE_READ64 ||
10538 /* Or if there is no read flag from REG. */
10540 /* Or if the read flag from REG is the same as PARENT_REG. */
10541 parent_flag == flag)
10544 err = mark_reg_read(env, reg, parent_reg, flag);
10551 /* A write screens off any subsequent reads; but write marks come from the
10552 * straight-line code between a state and its parent. When we arrive at an
10553 * equivalent state (jump target or such) we didn't arrive by the straight-line
10554 * code, so read marks in the state must propagate to the parent regardless
10555 * of the state's write marks. That's what 'parent == state->parent' comparison
10556 * in mark_reg_read() is for.
10558 static int propagate_liveness(struct bpf_verifier_env *env,
10559 const struct bpf_verifier_state *vstate,
10560 struct bpf_verifier_state *vparent)
10562 struct bpf_reg_state *state_reg, *parent_reg;
10563 struct bpf_func_state *state, *parent;
10564 int i, frame, err = 0;
10566 if (vparent->curframe != vstate->curframe) {
10567 WARN(1, "propagate_live: parent frame %d current frame %d\n",
10568 vparent->curframe, vstate->curframe);
10571 /* Propagate read liveness of registers... */
10572 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
10573 for (frame = 0; frame <= vstate->curframe; frame++) {
10574 parent = vparent->frame[frame];
10575 state = vstate->frame[frame];
10576 parent_reg = parent->regs;
10577 state_reg = state->regs;
10578 /* We don't need to worry about FP liveness, it's read-only */
10579 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
10580 err = propagate_liveness_reg(env, &state_reg[i],
10584 if (err == REG_LIVE_READ64)
10585 mark_insn_zext(env, &parent_reg[i]);
10588 /* Propagate stack slots. */
10589 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
10590 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
10591 parent_reg = &parent->stack[i].spilled_ptr;
10592 state_reg = &state->stack[i].spilled_ptr;
10593 err = propagate_liveness_reg(env, state_reg,
10602 /* find precise scalars in the previous equivalent state and
10603 * propagate them into the current state
10605 static int propagate_precision(struct bpf_verifier_env *env,
10606 const struct bpf_verifier_state *old)
10608 struct bpf_reg_state *state_reg;
10609 struct bpf_func_state *state;
10612 state = old->frame[old->curframe];
10613 state_reg = state->regs;
10614 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
10615 if (state_reg->type != SCALAR_VALUE ||
10616 !state_reg->precise)
10618 if (env->log.level & BPF_LOG_LEVEL2)
10619 verbose(env, "propagating r%d\n", i);
10620 err = mark_chain_precision(env, i);
10625 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
10626 if (state->stack[i].slot_type[0] != STACK_SPILL)
10628 state_reg = &state->stack[i].spilled_ptr;
10629 if (state_reg->type != SCALAR_VALUE ||
10630 !state_reg->precise)
10632 if (env->log.level & BPF_LOG_LEVEL2)
10633 verbose(env, "propagating fp%d\n",
10634 (-i - 1) * BPF_REG_SIZE);
10635 err = mark_chain_precision_stack(env, i);
10642 static bool states_maybe_looping(struct bpf_verifier_state *old,
10643 struct bpf_verifier_state *cur)
10645 struct bpf_func_state *fold, *fcur;
10646 int i, fr = cur->curframe;
10648 if (old->curframe != fr)
10651 fold = old->frame[fr];
10652 fcur = cur->frame[fr];
10653 for (i = 0; i < MAX_BPF_REG; i++)
10654 if (memcmp(&fold->regs[i], &fcur->regs[i],
10655 offsetof(struct bpf_reg_state, parent)))
10661 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
10663 struct bpf_verifier_state_list *new_sl;
10664 struct bpf_verifier_state_list *sl, **pprev;
10665 struct bpf_verifier_state *cur = env->cur_state, *new;
10666 int i, j, err, states_cnt = 0;
10667 bool add_new_state = env->test_state_freq ? true : false;
10669 cur->last_insn_idx = env->prev_insn_idx;
10670 if (!env->insn_aux_data[insn_idx].prune_point)
10671 /* this 'insn_idx' instruction wasn't marked, so we will not
10672 * be doing state search here
10676 /* bpf progs typically have pruning point every 4 instructions
10677 * http://vger.kernel.org/bpfconf2019.html#session-1
10678 * Do not add new state for future pruning if the verifier hasn't seen
10679 * at least 2 jumps and at least 8 instructions.
10680 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
10681 * In tests that amounts to up to 50% reduction into total verifier
10682 * memory consumption and 20% verifier time speedup.
10684 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
10685 env->insn_processed - env->prev_insn_processed >= 8)
10686 add_new_state = true;
10688 pprev = explored_state(env, insn_idx);
10691 clean_live_states(env, insn_idx, cur);
10695 if (sl->state.insn_idx != insn_idx)
10698 if (sl->state.branches) {
10699 struct bpf_func_state *frame = sl->state.frame[sl->state.curframe];
10701 if (frame->in_async_callback_fn &&
10702 frame->async_entry_cnt != cur->frame[cur->curframe]->async_entry_cnt) {
10703 /* Different async_entry_cnt means that the verifier is
10704 * processing another entry into async callback.
10705 * Seeing the same state is not an indication of infinite
10706 * loop or infinite recursion.
10707 * But finding the same state doesn't mean that it's safe
10708 * to stop processing the current state. The previous state
10709 * hasn't yet reached bpf_exit, since state.branches > 0.
10710 * Checking in_async_callback_fn alone is not enough either.
10711 * Since the verifier still needs to catch infinite loops
10712 * inside async callbacks.
10714 } else if (states_maybe_looping(&sl->state, cur) &&
10715 states_equal(env, &sl->state, cur)) {
10716 verbose_linfo(env, insn_idx, "; ");
10717 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
10720 /* if the verifier is processing a loop, avoid adding new state
10721 * too often, since different loop iterations have distinct
10722 * states and may not help future pruning.
10723 * This threshold shouldn't be too low to make sure that
10724 * a loop with large bound will be rejected quickly.
10725 * The most abusive loop will be:
10727 * if r1 < 1000000 goto pc-2
10728 * 1M insn_procssed limit / 100 == 10k peak states.
10729 * This threshold shouldn't be too high either, since states
10730 * at the end of the loop are likely to be useful in pruning.
10732 if (env->jmps_processed - env->prev_jmps_processed < 20 &&
10733 env->insn_processed - env->prev_insn_processed < 100)
10734 add_new_state = false;
10737 if (states_equal(env, &sl->state, cur)) {
10739 /* reached equivalent register/stack state,
10740 * prune the search.
10741 * Registers read by the continuation are read by us.
10742 * If we have any write marks in env->cur_state, they
10743 * will prevent corresponding reads in the continuation
10744 * from reaching our parent (an explored_state). Our
10745 * own state will get the read marks recorded, but
10746 * they'll be immediately forgotten as we're pruning
10747 * this state and will pop a new one.
10749 err = propagate_liveness(env, &sl->state, cur);
10751 /* if previous state reached the exit with precision and
10752 * current state is equivalent to it (except precsion marks)
10753 * the precision needs to be propagated back in
10754 * the current state.
10756 err = err ? : push_jmp_history(env, cur);
10757 err = err ? : propagate_precision(env, &sl->state);
10763 /* when new state is not going to be added do not increase miss count.
10764 * Otherwise several loop iterations will remove the state
10765 * recorded earlier. The goal of these heuristics is to have
10766 * states from some iterations of the loop (some in the beginning
10767 * and some at the end) to help pruning.
10771 /* heuristic to determine whether this state is beneficial
10772 * to keep checking from state equivalence point of view.
10773 * Higher numbers increase max_states_per_insn and verification time,
10774 * but do not meaningfully decrease insn_processed.
10776 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
10777 /* the state is unlikely to be useful. Remove it to
10778 * speed up verification
10781 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
10782 u32 br = sl->state.branches;
10785 "BUG live_done but branches_to_explore %d\n",
10787 free_verifier_state(&sl->state, false);
10789 env->peak_states--;
10791 /* cannot free this state, since parentage chain may
10792 * walk it later. Add it for free_list instead to
10793 * be freed at the end of verification
10795 sl->next = env->free_list;
10796 env->free_list = sl;
10806 if (env->max_states_per_insn < states_cnt)
10807 env->max_states_per_insn = states_cnt;
10809 if (!env->bpf_capable && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
10810 return push_jmp_history(env, cur);
10812 if (!add_new_state)
10813 return push_jmp_history(env, cur);
10815 /* There were no equivalent states, remember the current one.
10816 * Technically the current state is not proven to be safe yet,
10817 * but it will either reach outer most bpf_exit (which means it's safe)
10818 * or it will be rejected. When there are no loops the verifier won't be
10819 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
10820 * again on the way to bpf_exit.
10821 * When looping the sl->state.branches will be > 0 and this state
10822 * will not be considered for equivalence until branches == 0.
10824 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
10827 env->total_states++;
10828 env->peak_states++;
10829 env->prev_jmps_processed = env->jmps_processed;
10830 env->prev_insn_processed = env->insn_processed;
10832 /* add new state to the head of linked list */
10833 new = &new_sl->state;
10834 err = copy_verifier_state(new, cur);
10836 free_verifier_state(new, false);
10840 new->insn_idx = insn_idx;
10841 WARN_ONCE(new->branches != 1,
10842 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
10845 cur->first_insn_idx = insn_idx;
10846 clear_jmp_history(cur);
10847 new_sl->next = *explored_state(env, insn_idx);
10848 *explored_state(env, insn_idx) = new_sl;
10849 /* connect new state to parentage chain. Current frame needs all
10850 * registers connected. Only r6 - r9 of the callers are alive (pushed
10851 * to the stack implicitly by JITs) so in callers' frames connect just
10852 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
10853 * the state of the call instruction (with WRITTEN set), and r0 comes
10854 * from callee with its full parentage chain, anyway.
10856 /* clear write marks in current state: the writes we did are not writes
10857 * our child did, so they don't screen off its reads from us.
10858 * (There are no read marks in current state, because reads always mark
10859 * their parent and current state never has children yet. Only
10860 * explored_states can get read marks.)
10862 for (j = 0; j <= cur->curframe; j++) {
10863 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
10864 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
10865 for (i = 0; i < BPF_REG_FP; i++)
10866 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
10869 /* all stack frames are accessible from callee, clear them all */
10870 for (j = 0; j <= cur->curframe; j++) {
10871 struct bpf_func_state *frame = cur->frame[j];
10872 struct bpf_func_state *newframe = new->frame[j];
10874 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
10875 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
10876 frame->stack[i].spilled_ptr.parent =
10877 &newframe->stack[i].spilled_ptr;
10883 /* Return true if it's OK to have the same insn return a different type. */
10884 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
10886 switch (base_type(type)) {
10888 case PTR_TO_SOCKET:
10889 case PTR_TO_SOCK_COMMON:
10890 case PTR_TO_TCP_SOCK:
10891 case PTR_TO_XDP_SOCK:
10892 case PTR_TO_BTF_ID:
10899 /* If an instruction was previously used with particular pointer types, then we
10900 * need to be careful to avoid cases such as the below, where it may be ok
10901 * for one branch accessing the pointer, but not ok for the other branch:
10906 * R1 = some_other_valid_ptr;
10909 * R2 = *(u32 *)(R1 + 0);
10911 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
10913 return src != prev && (!reg_type_mismatch_ok(src) ||
10914 !reg_type_mismatch_ok(prev));
10917 static int do_check(struct bpf_verifier_env *env)
10919 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
10920 struct bpf_verifier_state *state = env->cur_state;
10921 struct bpf_insn *insns = env->prog->insnsi;
10922 struct bpf_reg_state *regs;
10923 int insn_cnt = env->prog->len;
10924 bool do_print_state = false;
10925 int prev_insn_idx = -1;
10928 struct bpf_insn *insn;
10932 env->prev_insn_idx = prev_insn_idx;
10933 if (env->insn_idx >= insn_cnt) {
10934 verbose(env, "invalid insn idx %d insn_cnt %d\n",
10935 env->insn_idx, insn_cnt);
10939 insn = &insns[env->insn_idx];
10940 class = BPF_CLASS(insn->code);
10942 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
10944 "BPF program is too large. Processed %d insn\n",
10945 env->insn_processed);
10949 err = is_state_visited(env, env->insn_idx);
10953 /* found equivalent state, can prune the search */
10954 if (env->log.level & BPF_LOG_LEVEL) {
10955 if (do_print_state)
10956 verbose(env, "\nfrom %d to %d%s: safe\n",
10957 env->prev_insn_idx, env->insn_idx,
10958 env->cur_state->speculative ?
10959 " (speculative execution)" : "");
10961 verbose(env, "%d: safe\n", env->insn_idx);
10963 goto process_bpf_exit;
10966 if (signal_pending(current))
10969 if (need_resched())
10972 if (env->log.level & BPF_LOG_LEVEL2 ||
10973 (env->log.level & BPF_LOG_LEVEL && do_print_state)) {
10974 if (env->log.level & BPF_LOG_LEVEL2)
10975 verbose(env, "%d:", env->insn_idx);
10977 verbose(env, "\nfrom %d to %d%s:",
10978 env->prev_insn_idx, env->insn_idx,
10979 env->cur_state->speculative ?
10980 " (speculative execution)" : "");
10981 print_verifier_state(env, state->frame[state->curframe]);
10982 do_print_state = false;
10985 if (env->log.level & BPF_LOG_LEVEL) {
10986 const struct bpf_insn_cbs cbs = {
10987 .cb_call = disasm_kfunc_name,
10988 .cb_print = verbose,
10989 .private_data = env,
10992 verbose_linfo(env, env->insn_idx, "; ");
10993 verbose(env, "%d: ", env->insn_idx);
10994 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
10997 if (bpf_prog_is_dev_bound(env->prog->aux)) {
10998 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
10999 env->prev_insn_idx);
11004 regs = cur_regs(env);
11005 sanitize_mark_insn_seen(env);
11006 prev_insn_idx = env->insn_idx;
11008 if (class == BPF_ALU || class == BPF_ALU64) {
11009 err = check_alu_op(env, insn);
11013 } else if (class == BPF_LDX) {
11014 enum bpf_reg_type *prev_src_type, src_reg_type;
11016 /* check for reserved fields is already done */
11018 /* check src operand */
11019 err = check_reg_arg(env, insn->src_reg, SRC_OP);
11023 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
11027 src_reg_type = regs[insn->src_reg].type;
11029 /* check that memory (src_reg + off) is readable,
11030 * the state of dst_reg will be updated by this func
11032 err = check_mem_access(env, env->insn_idx, insn->src_reg,
11033 insn->off, BPF_SIZE(insn->code),
11034 BPF_READ, insn->dst_reg, false);
11038 prev_src_type = &env->insn_aux_data[env->insn_idx].ptr_type;
11040 if (*prev_src_type == NOT_INIT) {
11041 /* saw a valid insn
11042 * dst_reg = *(u32 *)(src_reg + off)
11043 * save type to validate intersecting paths
11045 *prev_src_type = src_reg_type;
11047 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
11048 /* ABuser program is trying to use the same insn
11049 * dst_reg = *(u32*) (src_reg + off)
11050 * with different pointer types:
11051 * src_reg == ctx in one branch and
11052 * src_reg == stack|map in some other branch.
11055 verbose(env, "same insn cannot be used with different pointers\n");
11059 } else if (class == BPF_STX) {
11060 enum bpf_reg_type *prev_dst_type, dst_reg_type;
11062 if (BPF_MODE(insn->code) == BPF_ATOMIC) {
11063 err = check_atomic(env, env->insn_idx, insn);
11070 if (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0) {
11071 verbose(env, "BPF_STX uses reserved fields\n");
11075 /* check src1 operand */
11076 err = check_reg_arg(env, insn->src_reg, SRC_OP);
11079 /* check src2 operand */
11080 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
11084 dst_reg_type = regs[insn->dst_reg].type;
11086 /* check that memory (dst_reg + off) is writeable */
11087 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
11088 insn->off, BPF_SIZE(insn->code),
11089 BPF_WRITE, insn->src_reg, false);
11093 prev_dst_type = &env->insn_aux_data[env->insn_idx].ptr_type;
11095 if (*prev_dst_type == NOT_INIT) {
11096 *prev_dst_type = dst_reg_type;
11097 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
11098 verbose(env, "same insn cannot be used with different pointers\n");
11102 } else if (class == BPF_ST) {
11103 if (BPF_MODE(insn->code) != BPF_MEM ||
11104 insn->src_reg != BPF_REG_0) {
11105 verbose(env, "BPF_ST uses reserved fields\n");
11108 /* check src operand */
11109 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
11113 if (is_ctx_reg(env, insn->dst_reg)) {
11114 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
11116 reg_type_str(env, reg_state(env, insn->dst_reg)->type));
11120 /* check that memory (dst_reg + off) is writeable */
11121 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
11122 insn->off, BPF_SIZE(insn->code),
11123 BPF_WRITE, -1, false);
11127 } else if (class == BPF_JMP || class == BPF_JMP32) {
11128 u8 opcode = BPF_OP(insn->code);
11130 env->jmps_processed++;
11131 if (opcode == BPF_CALL) {
11132 if (BPF_SRC(insn->code) != BPF_K ||
11134 (insn->src_reg != BPF_REG_0 &&
11135 insn->src_reg != BPF_PSEUDO_CALL &&
11136 insn->src_reg != BPF_PSEUDO_KFUNC_CALL) ||
11137 insn->dst_reg != BPF_REG_0 ||
11138 class == BPF_JMP32) {
11139 verbose(env, "BPF_CALL uses reserved fields\n");
11143 if (env->cur_state->active_spin_lock &&
11144 (insn->src_reg == BPF_PSEUDO_CALL ||
11145 insn->imm != BPF_FUNC_spin_unlock)) {
11146 verbose(env, "function calls are not allowed while holding a lock\n");
11149 if (insn->src_reg == BPF_PSEUDO_CALL)
11150 err = check_func_call(env, insn, &env->insn_idx);
11151 else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL)
11152 err = check_kfunc_call(env, insn);
11154 err = check_helper_call(env, insn, &env->insn_idx);
11157 } else if (opcode == BPF_JA) {
11158 if (BPF_SRC(insn->code) != BPF_K ||
11160 insn->src_reg != BPF_REG_0 ||
11161 insn->dst_reg != BPF_REG_0 ||
11162 class == BPF_JMP32) {
11163 verbose(env, "BPF_JA uses reserved fields\n");
11167 env->insn_idx += insn->off + 1;
11170 } else if (opcode == BPF_EXIT) {
11171 if (BPF_SRC(insn->code) != BPF_K ||
11173 insn->src_reg != BPF_REG_0 ||
11174 insn->dst_reg != BPF_REG_0 ||
11175 class == BPF_JMP32) {
11176 verbose(env, "BPF_EXIT uses reserved fields\n");
11180 if (env->cur_state->active_spin_lock) {
11181 verbose(env, "bpf_spin_unlock is missing\n");
11185 if (state->curframe) {
11186 /* exit from nested function */
11187 err = prepare_func_exit(env, &env->insn_idx);
11190 do_print_state = true;
11194 err = check_reference_leak(env);
11198 err = check_return_code(env);
11202 update_branch_counts(env, env->cur_state);
11203 err = pop_stack(env, &prev_insn_idx,
11204 &env->insn_idx, pop_log);
11206 if (err != -ENOENT)
11210 do_print_state = true;
11214 err = check_cond_jmp_op(env, insn, &env->insn_idx);
11218 } else if (class == BPF_LD) {
11219 u8 mode = BPF_MODE(insn->code);
11221 if (mode == BPF_ABS || mode == BPF_IND) {
11222 err = check_ld_abs(env, insn);
11226 } else if (mode == BPF_IMM) {
11227 err = check_ld_imm(env, insn);
11232 sanitize_mark_insn_seen(env);
11234 verbose(env, "invalid BPF_LD mode\n");
11238 verbose(env, "unknown insn class %d\n", class);
11248 static int find_btf_percpu_datasec(struct btf *btf)
11250 const struct btf_type *t;
11255 * Both vmlinux and module each have their own ".data..percpu"
11256 * DATASECs in BTF. So for module's case, we need to skip vmlinux BTF
11257 * types to look at only module's own BTF types.
11259 n = btf_nr_types(btf);
11260 if (btf_is_module(btf))
11261 i = btf_nr_types(btf_vmlinux);
11265 for(; i < n; i++) {
11266 t = btf_type_by_id(btf, i);
11267 if (BTF_INFO_KIND(t->info) != BTF_KIND_DATASEC)
11270 tname = btf_name_by_offset(btf, t->name_off);
11271 if (!strcmp(tname, ".data..percpu"))
11278 /* replace pseudo btf_id with kernel symbol address */
11279 static int check_pseudo_btf_id(struct bpf_verifier_env *env,
11280 struct bpf_insn *insn,
11281 struct bpf_insn_aux_data *aux)
11283 const struct btf_var_secinfo *vsi;
11284 const struct btf_type *datasec;
11285 struct btf_mod_pair *btf_mod;
11286 const struct btf_type *t;
11287 const char *sym_name;
11288 bool percpu = false;
11289 u32 type, id = insn->imm;
11293 int i, btf_fd, err;
11295 btf_fd = insn[1].imm;
11297 btf = btf_get_by_fd(btf_fd);
11299 verbose(env, "invalid module BTF object FD specified.\n");
11303 if (!btf_vmlinux) {
11304 verbose(env, "kernel is missing BTF, make sure CONFIG_DEBUG_INFO_BTF=y is specified in Kconfig.\n");
11311 t = btf_type_by_id(btf, id);
11313 verbose(env, "ldimm64 insn specifies invalid btf_id %d.\n", id);
11318 if (!btf_type_is_var(t)) {
11319 verbose(env, "pseudo btf_id %d in ldimm64 isn't KIND_VAR.\n", id);
11324 sym_name = btf_name_by_offset(btf, t->name_off);
11325 addr = kallsyms_lookup_name(sym_name);
11327 verbose(env, "ldimm64 failed to find the address for kernel symbol '%s'.\n",
11333 datasec_id = find_btf_percpu_datasec(btf);
11334 if (datasec_id > 0) {
11335 datasec = btf_type_by_id(btf, datasec_id);
11336 for_each_vsi(i, datasec, vsi) {
11337 if (vsi->type == id) {
11344 insn[0].imm = (u32)addr;
11345 insn[1].imm = addr >> 32;
11348 t = btf_type_skip_modifiers(btf, type, NULL);
11350 aux->btf_var.reg_type = PTR_TO_PERCPU_BTF_ID;
11351 aux->btf_var.btf = btf;
11352 aux->btf_var.btf_id = type;
11353 } else if (!btf_type_is_struct(t)) {
11354 const struct btf_type *ret;
11358 /* resolve the type size of ksym. */
11359 ret = btf_resolve_size(btf, t, &tsize);
11361 tname = btf_name_by_offset(btf, t->name_off);
11362 verbose(env, "ldimm64 unable to resolve the size of type '%s': %ld\n",
11363 tname, PTR_ERR(ret));
11367 aux->btf_var.reg_type = PTR_TO_MEM | MEM_RDONLY;
11368 aux->btf_var.mem_size = tsize;
11370 aux->btf_var.reg_type = PTR_TO_BTF_ID;
11371 aux->btf_var.btf = btf;
11372 aux->btf_var.btf_id = type;
11375 /* check whether we recorded this BTF (and maybe module) already */
11376 for (i = 0; i < env->used_btf_cnt; i++) {
11377 if (env->used_btfs[i].btf == btf) {
11383 if (env->used_btf_cnt >= MAX_USED_BTFS) {
11388 btf_mod = &env->used_btfs[env->used_btf_cnt];
11389 btf_mod->btf = btf;
11390 btf_mod->module = NULL;
11392 /* if we reference variables from kernel module, bump its refcount */
11393 if (btf_is_module(btf)) {
11394 btf_mod->module = btf_try_get_module(btf);
11395 if (!btf_mod->module) {
11401 env->used_btf_cnt++;
11409 static int check_map_prealloc(struct bpf_map *map)
11411 return (map->map_type != BPF_MAP_TYPE_HASH &&
11412 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
11413 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
11414 !(map->map_flags & BPF_F_NO_PREALLOC);
11417 static bool is_tracing_prog_type(enum bpf_prog_type type)
11420 case BPF_PROG_TYPE_KPROBE:
11421 case BPF_PROG_TYPE_TRACEPOINT:
11422 case BPF_PROG_TYPE_PERF_EVENT:
11423 case BPF_PROG_TYPE_RAW_TRACEPOINT:
11430 static bool is_preallocated_map(struct bpf_map *map)
11432 if (!check_map_prealloc(map))
11434 if (map->inner_map_meta && !check_map_prealloc(map->inner_map_meta))
11439 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
11440 struct bpf_map *map,
11441 struct bpf_prog *prog)
11444 enum bpf_prog_type prog_type = resolve_prog_type(prog);
11446 * Validate that trace type programs use preallocated hash maps.
11448 * For programs attached to PERF events this is mandatory as the
11449 * perf NMI can hit any arbitrary code sequence.
11451 * All other trace types using preallocated hash maps are unsafe as
11452 * well because tracepoint or kprobes can be inside locked regions
11453 * of the memory allocator or at a place where a recursion into the
11454 * memory allocator would see inconsistent state.
11456 * On RT enabled kernels run-time allocation of all trace type
11457 * programs is strictly prohibited due to lock type constraints. On
11458 * !RT kernels it is allowed for backwards compatibility reasons for
11459 * now, but warnings are emitted so developers are made aware of
11460 * the unsafety and can fix their programs before this is enforced.
11462 if (is_tracing_prog_type(prog_type) && !is_preallocated_map(map)) {
11463 if (prog_type == BPF_PROG_TYPE_PERF_EVENT) {
11464 verbose(env, "perf_event programs can only use preallocated hash map\n");
11467 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
11468 verbose(env, "trace type programs can only use preallocated hash map\n");
11471 WARN_ONCE(1, "trace type BPF program uses run-time allocation\n");
11472 verbose(env, "trace type programs with run-time allocated hash maps are unsafe. Switch to preallocated hash maps.\n");
11475 if (map_value_has_spin_lock(map)) {
11476 if (prog_type == BPF_PROG_TYPE_SOCKET_FILTER) {
11477 verbose(env, "socket filter progs cannot use bpf_spin_lock yet\n");
11481 if (is_tracing_prog_type(prog_type)) {
11482 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
11486 if (prog->aux->sleepable) {
11487 verbose(env, "sleepable progs cannot use bpf_spin_lock yet\n");
11492 if (map_value_has_timer(map)) {
11493 if (is_tracing_prog_type(prog_type)) {
11494 verbose(env, "tracing progs cannot use bpf_timer yet\n");
11499 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
11500 !bpf_offload_prog_map_match(prog, map)) {
11501 verbose(env, "offload device mismatch between prog and map\n");
11505 if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
11506 verbose(env, "bpf_struct_ops map cannot be used in prog\n");
11510 if (prog->aux->sleepable)
11511 switch (map->map_type) {
11512 case BPF_MAP_TYPE_HASH:
11513 case BPF_MAP_TYPE_LRU_HASH:
11514 case BPF_MAP_TYPE_ARRAY:
11515 case BPF_MAP_TYPE_PERCPU_HASH:
11516 case BPF_MAP_TYPE_PERCPU_ARRAY:
11517 case BPF_MAP_TYPE_LRU_PERCPU_HASH:
11518 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
11519 case BPF_MAP_TYPE_HASH_OF_MAPS:
11520 if (!is_preallocated_map(map)) {
11522 "Sleepable programs can only use preallocated maps\n");
11526 case BPF_MAP_TYPE_RINGBUF:
11530 "Sleepable programs can only use array, hash, and ringbuf maps\n");
11537 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
11539 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
11540 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
11543 /* find and rewrite pseudo imm in ld_imm64 instructions:
11545 * 1. if it accesses map FD, replace it with actual map pointer.
11546 * 2. if it accesses btf_id of a VAR, replace it with pointer to the var.
11548 * NOTE: btf_vmlinux is required for converting pseudo btf_id.
11550 static int resolve_pseudo_ldimm64(struct bpf_verifier_env *env)
11552 struct bpf_insn *insn = env->prog->insnsi;
11553 int insn_cnt = env->prog->len;
11556 err = bpf_prog_calc_tag(env->prog);
11560 for (i = 0; i < insn_cnt; i++, insn++) {
11561 if (BPF_CLASS(insn->code) == BPF_LDX &&
11562 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
11563 verbose(env, "BPF_LDX uses reserved fields\n");
11567 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
11568 struct bpf_insn_aux_data *aux;
11569 struct bpf_map *map;
11574 if (i == insn_cnt - 1 || insn[1].code != 0 ||
11575 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
11576 insn[1].off != 0) {
11577 verbose(env, "invalid bpf_ld_imm64 insn\n");
11581 if (insn[0].src_reg == 0)
11582 /* valid generic load 64-bit imm */
11585 if (insn[0].src_reg == BPF_PSEUDO_BTF_ID) {
11586 aux = &env->insn_aux_data[i];
11587 err = check_pseudo_btf_id(env, insn, aux);
11593 if (insn[0].src_reg == BPF_PSEUDO_FUNC) {
11594 aux = &env->insn_aux_data[i];
11595 aux->ptr_type = PTR_TO_FUNC;
11599 /* In final convert_pseudo_ld_imm64() step, this is
11600 * converted into regular 64-bit imm load insn.
11602 switch (insn[0].src_reg) {
11603 case BPF_PSEUDO_MAP_VALUE:
11604 case BPF_PSEUDO_MAP_IDX_VALUE:
11606 case BPF_PSEUDO_MAP_FD:
11607 case BPF_PSEUDO_MAP_IDX:
11608 if (insn[1].imm == 0)
11612 verbose(env, "unrecognized bpf_ld_imm64 insn\n");
11616 switch (insn[0].src_reg) {
11617 case BPF_PSEUDO_MAP_IDX_VALUE:
11618 case BPF_PSEUDO_MAP_IDX:
11619 if (bpfptr_is_null(env->fd_array)) {
11620 verbose(env, "fd_idx without fd_array is invalid\n");
11623 if (copy_from_bpfptr_offset(&fd, env->fd_array,
11624 insn[0].imm * sizeof(fd),
11634 map = __bpf_map_get(f);
11636 verbose(env, "fd %d is not pointing to valid bpf_map\n",
11638 return PTR_ERR(map);
11641 err = check_map_prog_compatibility(env, map, env->prog);
11647 aux = &env->insn_aux_data[i];
11648 if (insn[0].src_reg == BPF_PSEUDO_MAP_FD ||
11649 insn[0].src_reg == BPF_PSEUDO_MAP_IDX) {
11650 addr = (unsigned long)map;
11652 u32 off = insn[1].imm;
11654 if (off >= BPF_MAX_VAR_OFF) {
11655 verbose(env, "direct value offset of %u is not allowed\n", off);
11660 if (!map->ops->map_direct_value_addr) {
11661 verbose(env, "no direct value access support for this map type\n");
11666 err = map->ops->map_direct_value_addr(map, &addr, off);
11668 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
11669 map->value_size, off);
11674 aux->map_off = off;
11678 insn[0].imm = (u32)addr;
11679 insn[1].imm = addr >> 32;
11681 /* check whether we recorded this map already */
11682 for (j = 0; j < env->used_map_cnt; j++) {
11683 if (env->used_maps[j] == map) {
11684 aux->map_index = j;
11690 if (env->used_map_cnt >= MAX_USED_MAPS) {
11695 /* hold the map. If the program is rejected by verifier,
11696 * the map will be released by release_maps() or it
11697 * will be used by the valid program until it's unloaded
11698 * and all maps are released in free_used_maps()
11702 aux->map_index = env->used_map_cnt;
11703 env->used_maps[env->used_map_cnt++] = map;
11705 if (bpf_map_is_cgroup_storage(map) &&
11706 bpf_cgroup_storage_assign(env->prog->aux, map)) {
11707 verbose(env, "only one cgroup storage of each type is allowed\n");
11719 /* Basic sanity check before we invest more work here. */
11720 if (!bpf_opcode_in_insntable(insn->code)) {
11721 verbose(env, "unknown opcode %02x\n", insn->code);
11726 /* now all pseudo BPF_LD_IMM64 instructions load valid
11727 * 'struct bpf_map *' into a register instead of user map_fd.
11728 * These pointers will be used later by verifier to validate map access.
11733 /* drop refcnt of maps used by the rejected program */
11734 static void release_maps(struct bpf_verifier_env *env)
11736 __bpf_free_used_maps(env->prog->aux, env->used_maps,
11737 env->used_map_cnt);
11740 /* drop refcnt of maps used by the rejected program */
11741 static void release_btfs(struct bpf_verifier_env *env)
11743 __bpf_free_used_btfs(env->prog->aux, env->used_btfs,
11744 env->used_btf_cnt);
11747 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
11748 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
11750 struct bpf_insn *insn = env->prog->insnsi;
11751 int insn_cnt = env->prog->len;
11754 for (i = 0; i < insn_cnt; i++, insn++) {
11755 if (insn->code != (BPF_LD | BPF_IMM | BPF_DW))
11757 if (insn->src_reg == BPF_PSEUDO_FUNC)
11763 /* single env->prog->insni[off] instruction was replaced with the range
11764 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
11765 * [0, off) and [off, end) to new locations, so the patched range stays zero
11767 static void adjust_insn_aux_data(struct bpf_verifier_env *env,
11768 struct bpf_insn_aux_data *new_data,
11769 struct bpf_prog *new_prog, u32 off, u32 cnt)
11771 struct bpf_insn_aux_data *old_data = env->insn_aux_data;
11772 struct bpf_insn *insn = new_prog->insnsi;
11773 u32 old_seen = old_data[off].seen;
11777 /* aux info at OFF always needs adjustment, no matter fast path
11778 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
11779 * original insn at old prog.
11781 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
11785 prog_len = new_prog->len;
11787 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
11788 memcpy(new_data + off + cnt - 1, old_data + off,
11789 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
11790 for (i = off; i < off + cnt - 1; i++) {
11791 /* Expand insni[off]'s seen count to the patched range. */
11792 new_data[i].seen = old_seen;
11793 new_data[i].zext_dst = insn_has_def32(env, insn + i);
11795 env->insn_aux_data = new_data;
11799 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
11805 /* NOTE: fake 'exit' subprog should be updated as well. */
11806 for (i = 0; i <= env->subprog_cnt; i++) {
11807 if (env->subprog_info[i].start <= off)
11809 env->subprog_info[i].start += len - 1;
11813 static void adjust_poke_descs(struct bpf_prog *prog, u32 off, u32 len)
11815 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
11816 int i, sz = prog->aux->size_poke_tab;
11817 struct bpf_jit_poke_descriptor *desc;
11819 for (i = 0; i < sz; i++) {
11821 if (desc->insn_idx <= off)
11823 desc->insn_idx += len - 1;
11827 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
11828 const struct bpf_insn *patch, u32 len)
11830 struct bpf_prog *new_prog;
11831 struct bpf_insn_aux_data *new_data = NULL;
11834 new_data = vzalloc(array_size(env->prog->len + len - 1,
11835 sizeof(struct bpf_insn_aux_data)));
11840 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
11841 if (IS_ERR(new_prog)) {
11842 if (PTR_ERR(new_prog) == -ERANGE)
11844 "insn %d cannot be patched due to 16-bit range\n",
11845 env->insn_aux_data[off].orig_idx);
11849 adjust_insn_aux_data(env, new_data, new_prog, off, len);
11850 adjust_subprog_starts(env, off, len);
11851 adjust_poke_descs(new_prog, off, len);
11855 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
11860 /* find first prog starting at or after off (first to remove) */
11861 for (i = 0; i < env->subprog_cnt; i++)
11862 if (env->subprog_info[i].start >= off)
11864 /* find first prog starting at or after off + cnt (first to stay) */
11865 for (j = i; j < env->subprog_cnt; j++)
11866 if (env->subprog_info[j].start >= off + cnt)
11868 /* if j doesn't start exactly at off + cnt, we are just removing
11869 * the front of previous prog
11871 if (env->subprog_info[j].start != off + cnt)
11875 struct bpf_prog_aux *aux = env->prog->aux;
11878 /* move fake 'exit' subprog as well */
11879 move = env->subprog_cnt + 1 - j;
11881 memmove(env->subprog_info + i,
11882 env->subprog_info + j,
11883 sizeof(*env->subprog_info) * move);
11884 env->subprog_cnt -= j - i;
11886 /* remove func_info */
11887 if (aux->func_info) {
11888 move = aux->func_info_cnt - j;
11890 memmove(aux->func_info + i,
11891 aux->func_info + j,
11892 sizeof(*aux->func_info) * move);
11893 aux->func_info_cnt -= j - i;
11894 /* func_info->insn_off is set after all code rewrites,
11895 * in adjust_btf_func() - no need to adjust
11899 /* convert i from "first prog to remove" to "first to adjust" */
11900 if (env->subprog_info[i].start == off)
11904 /* update fake 'exit' subprog as well */
11905 for (; i <= env->subprog_cnt; i++)
11906 env->subprog_info[i].start -= cnt;
11911 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
11914 struct bpf_prog *prog = env->prog;
11915 u32 i, l_off, l_cnt, nr_linfo;
11916 struct bpf_line_info *linfo;
11918 nr_linfo = prog->aux->nr_linfo;
11922 linfo = prog->aux->linfo;
11924 /* find first line info to remove, count lines to be removed */
11925 for (i = 0; i < nr_linfo; i++)
11926 if (linfo[i].insn_off >= off)
11931 for (; i < nr_linfo; i++)
11932 if (linfo[i].insn_off < off + cnt)
11937 /* First live insn doesn't match first live linfo, it needs to "inherit"
11938 * last removed linfo. prog is already modified, so prog->len == off
11939 * means no live instructions after (tail of the program was removed).
11941 if (prog->len != off && l_cnt &&
11942 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
11944 linfo[--i].insn_off = off + cnt;
11947 /* remove the line info which refer to the removed instructions */
11949 memmove(linfo + l_off, linfo + i,
11950 sizeof(*linfo) * (nr_linfo - i));
11952 prog->aux->nr_linfo -= l_cnt;
11953 nr_linfo = prog->aux->nr_linfo;
11956 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
11957 for (i = l_off; i < nr_linfo; i++)
11958 linfo[i].insn_off -= cnt;
11960 /* fix up all subprogs (incl. 'exit') which start >= off */
11961 for (i = 0; i <= env->subprog_cnt; i++)
11962 if (env->subprog_info[i].linfo_idx > l_off) {
11963 /* program may have started in the removed region but
11964 * may not be fully removed
11966 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
11967 env->subprog_info[i].linfo_idx -= l_cnt;
11969 env->subprog_info[i].linfo_idx = l_off;
11975 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
11977 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
11978 unsigned int orig_prog_len = env->prog->len;
11981 if (bpf_prog_is_dev_bound(env->prog->aux))
11982 bpf_prog_offload_remove_insns(env, off, cnt);
11984 err = bpf_remove_insns(env->prog, off, cnt);
11988 err = adjust_subprog_starts_after_remove(env, off, cnt);
11992 err = bpf_adj_linfo_after_remove(env, off, cnt);
11996 memmove(aux_data + off, aux_data + off + cnt,
11997 sizeof(*aux_data) * (orig_prog_len - off - cnt));
12002 /* The verifier does more data flow analysis than llvm and will not
12003 * explore branches that are dead at run time. Malicious programs can
12004 * have dead code too. Therefore replace all dead at-run-time code
12007 * Just nops are not optimal, e.g. if they would sit at the end of the
12008 * program and through another bug we would manage to jump there, then
12009 * we'd execute beyond program memory otherwise. Returning exception
12010 * code also wouldn't work since we can have subprogs where the dead
12011 * code could be located.
12013 static void sanitize_dead_code(struct bpf_verifier_env *env)
12015 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
12016 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
12017 struct bpf_insn *insn = env->prog->insnsi;
12018 const int insn_cnt = env->prog->len;
12021 for (i = 0; i < insn_cnt; i++) {
12022 if (aux_data[i].seen)
12024 memcpy(insn + i, &trap, sizeof(trap));
12025 aux_data[i].zext_dst = false;
12029 static bool insn_is_cond_jump(u8 code)
12033 if (BPF_CLASS(code) == BPF_JMP32)
12036 if (BPF_CLASS(code) != BPF_JMP)
12040 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
12043 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
12045 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
12046 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
12047 struct bpf_insn *insn = env->prog->insnsi;
12048 const int insn_cnt = env->prog->len;
12051 for (i = 0; i < insn_cnt; i++, insn++) {
12052 if (!insn_is_cond_jump(insn->code))
12055 if (!aux_data[i + 1].seen)
12056 ja.off = insn->off;
12057 else if (!aux_data[i + 1 + insn->off].seen)
12062 if (bpf_prog_is_dev_bound(env->prog->aux))
12063 bpf_prog_offload_replace_insn(env, i, &ja);
12065 memcpy(insn, &ja, sizeof(ja));
12069 static int opt_remove_dead_code(struct bpf_verifier_env *env)
12071 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
12072 int insn_cnt = env->prog->len;
12075 for (i = 0; i < insn_cnt; i++) {
12079 while (i + j < insn_cnt && !aux_data[i + j].seen)
12084 err = verifier_remove_insns(env, i, j);
12087 insn_cnt = env->prog->len;
12093 static int opt_remove_nops(struct bpf_verifier_env *env)
12095 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
12096 struct bpf_insn *insn = env->prog->insnsi;
12097 int insn_cnt = env->prog->len;
12100 for (i = 0; i < insn_cnt; i++) {
12101 if (memcmp(&insn[i], &ja, sizeof(ja)))
12104 err = verifier_remove_insns(env, i, 1);
12114 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
12115 const union bpf_attr *attr)
12117 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
12118 struct bpf_insn_aux_data *aux = env->insn_aux_data;
12119 int i, patch_len, delta = 0, len = env->prog->len;
12120 struct bpf_insn *insns = env->prog->insnsi;
12121 struct bpf_prog *new_prog;
12124 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
12125 zext_patch[1] = BPF_ZEXT_REG(0);
12126 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
12127 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
12128 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
12129 for (i = 0; i < len; i++) {
12130 int adj_idx = i + delta;
12131 struct bpf_insn insn;
12134 insn = insns[adj_idx];
12135 load_reg = insn_def_regno(&insn);
12136 if (!aux[adj_idx].zext_dst) {
12144 class = BPF_CLASS(code);
12145 if (load_reg == -1)
12148 /* NOTE: arg "reg" (the fourth one) is only used for
12149 * BPF_STX + SRC_OP, so it is safe to pass NULL
12152 if (is_reg64(env, &insn, load_reg, NULL, DST_OP)) {
12153 if (class == BPF_LD &&
12154 BPF_MODE(code) == BPF_IMM)
12159 /* ctx load could be transformed into wider load. */
12160 if (class == BPF_LDX &&
12161 aux[adj_idx].ptr_type == PTR_TO_CTX)
12164 imm_rnd = get_random_int();
12165 rnd_hi32_patch[0] = insn;
12166 rnd_hi32_patch[1].imm = imm_rnd;
12167 rnd_hi32_patch[3].dst_reg = load_reg;
12168 patch = rnd_hi32_patch;
12170 goto apply_patch_buffer;
12173 /* Add in an zero-extend instruction if a) the JIT has requested
12174 * it or b) it's a CMPXCHG.
12176 * The latter is because: BPF_CMPXCHG always loads a value into
12177 * R0, therefore always zero-extends. However some archs'
12178 * equivalent instruction only does this load when the
12179 * comparison is successful. This detail of CMPXCHG is
12180 * orthogonal to the general zero-extension behaviour of the
12181 * CPU, so it's treated independently of bpf_jit_needs_zext.
12183 if (!bpf_jit_needs_zext() && !is_cmpxchg_insn(&insn))
12186 if (WARN_ON(load_reg == -1)) {
12187 verbose(env, "verifier bug. zext_dst is set, but no reg is defined\n");
12191 zext_patch[0] = insn;
12192 zext_patch[1].dst_reg = load_reg;
12193 zext_patch[1].src_reg = load_reg;
12194 patch = zext_patch;
12196 apply_patch_buffer:
12197 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
12200 env->prog = new_prog;
12201 insns = new_prog->insnsi;
12202 aux = env->insn_aux_data;
12203 delta += patch_len - 1;
12209 /* convert load instructions that access fields of a context type into a
12210 * sequence of instructions that access fields of the underlying structure:
12211 * struct __sk_buff -> struct sk_buff
12212 * struct bpf_sock_ops -> struct sock
12214 static int convert_ctx_accesses(struct bpf_verifier_env *env)
12216 const struct bpf_verifier_ops *ops = env->ops;
12217 int i, cnt, size, ctx_field_size, delta = 0;
12218 const int insn_cnt = env->prog->len;
12219 struct bpf_insn insn_buf[16], *insn;
12220 u32 target_size, size_default, off;
12221 struct bpf_prog *new_prog;
12222 enum bpf_access_type type;
12223 bool is_narrower_load;
12225 if (ops->gen_prologue || env->seen_direct_write) {
12226 if (!ops->gen_prologue) {
12227 verbose(env, "bpf verifier is misconfigured\n");
12230 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
12232 if (cnt >= ARRAY_SIZE(insn_buf)) {
12233 verbose(env, "bpf verifier is misconfigured\n");
12236 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
12240 env->prog = new_prog;
12245 if (bpf_prog_is_dev_bound(env->prog->aux))
12248 insn = env->prog->insnsi + delta;
12250 for (i = 0; i < insn_cnt; i++, insn++) {
12251 bpf_convert_ctx_access_t convert_ctx_access;
12254 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
12255 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
12256 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
12257 insn->code == (BPF_LDX | BPF_MEM | BPF_DW)) {
12260 } else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
12261 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
12262 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
12263 insn->code == (BPF_STX | BPF_MEM | BPF_DW) ||
12264 insn->code == (BPF_ST | BPF_MEM | BPF_B) ||
12265 insn->code == (BPF_ST | BPF_MEM | BPF_H) ||
12266 insn->code == (BPF_ST | BPF_MEM | BPF_W) ||
12267 insn->code == (BPF_ST | BPF_MEM | BPF_DW)) {
12269 ctx_access = BPF_CLASS(insn->code) == BPF_STX;
12274 if (type == BPF_WRITE &&
12275 env->insn_aux_data[i + delta].sanitize_stack_spill) {
12276 struct bpf_insn patch[] = {
12281 cnt = ARRAY_SIZE(patch);
12282 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
12287 env->prog = new_prog;
12288 insn = new_prog->insnsi + i + delta;
12295 switch (env->insn_aux_data[i + delta].ptr_type) {
12297 if (!ops->convert_ctx_access)
12299 convert_ctx_access = ops->convert_ctx_access;
12301 case PTR_TO_SOCKET:
12302 case PTR_TO_SOCK_COMMON:
12303 convert_ctx_access = bpf_sock_convert_ctx_access;
12305 case PTR_TO_TCP_SOCK:
12306 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
12308 case PTR_TO_XDP_SOCK:
12309 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
12311 case PTR_TO_BTF_ID:
12312 if (type == BPF_READ) {
12313 insn->code = BPF_LDX | BPF_PROBE_MEM |
12314 BPF_SIZE((insn)->code);
12315 env->prog->aux->num_exentries++;
12316 } else if (resolve_prog_type(env->prog) != BPF_PROG_TYPE_STRUCT_OPS) {
12317 verbose(env, "Writes through BTF pointers are not allowed\n");
12325 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
12326 size = BPF_LDST_BYTES(insn);
12328 /* If the read access is a narrower load of the field,
12329 * convert to a 4/8-byte load, to minimum program type specific
12330 * convert_ctx_access changes. If conversion is successful,
12331 * we will apply proper mask to the result.
12333 is_narrower_load = size < ctx_field_size;
12334 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
12336 if (is_narrower_load) {
12339 if (type == BPF_WRITE) {
12340 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
12345 if (ctx_field_size == 4)
12347 else if (ctx_field_size == 8)
12348 size_code = BPF_DW;
12350 insn->off = off & ~(size_default - 1);
12351 insn->code = BPF_LDX | BPF_MEM | size_code;
12355 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
12357 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
12358 (ctx_field_size && !target_size)) {
12359 verbose(env, "bpf verifier is misconfigured\n");
12363 if (is_narrower_load && size < target_size) {
12364 u8 shift = bpf_ctx_narrow_access_offset(
12365 off, size, size_default) * 8;
12366 if (shift && cnt + 1 >= ARRAY_SIZE(insn_buf)) {
12367 verbose(env, "bpf verifier narrow ctx load misconfigured\n");
12370 if (ctx_field_size <= 4) {
12372 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
12375 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
12376 (1 << size * 8) - 1);
12379 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
12382 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
12383 (1ULL << size * 8) - 1);
12387 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12393 /* keep walking new program and skip insns we just inserted */
12394 env->prog = new_prog;
12395 insn = new_prog->insnsi + i + delta;
12401 static int jit_subprogs(struct bpf_verifier_env *env)
12403 struct bpf_prog *prog = env->prog, **func, *tmp;
12404 int i, j, subprog_start, subprog_end = 0, len, subprog;
12405 struct bpf_map *map_ptr;
12406 struct bpf_insn *insn;
12407 void *old_bpf_func;
12408 int err, num_exentries;
12410 if (env->subprog_cnt <= 1)
12413 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
12414 if (bpf_pseudo_func(insn)) {
12415 env->insn_aux_data[i].call_imm = insn->imm;
12416 /* subprog is encoded in insn[1].imm */
12420 if (!bpf_pseudo_call(insn))
12422 /* Upon error here we cannot fall back to interpreter but
12423 * need a hard reject of the program. Thus -EFAULT is
12424 * propagated in any case.
12426 subprog = find_subprog(env, i + insn->imm + 1);
12428 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
12429 i + insn->imm + 1);
12432 /* temporarily remember subprog id inside insn instead of
12433 * aux_data, since next loop will split up all insns into funcs
12435 insn->off = subprog;
12436 /* remember original imm in case JIT fails and fallback
12437 * to interpreter will be needed
12439 env->insn_aux_data[i].call_imm = insn->imm;
12440 /* point imm to __bpf_call_base+1 from JITs point of view */
12444 err = bpf_prog_alloc_jited_linfo(prog);
12446 goto out_undo_insn;
12449 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
12451 goto out_undo_insn;
12453 for (i = 0; i < env->subprog_cnt; i++) {
12454 subprog_start = subprog_end;
12455 subprog_end = env->subprog_info[i + 1].start;
12457 len = subprog_end - subprog_start;
12458 /* bpf_prog_run() doesn't call subprogs directly,
12459 * hence main prog stats include the runtime of subprogs.
12460 * subprogs don't have IDs and not reachable via prog_get_next_id
12461 * func[i]->stats will never be accessed and stays NULL
12463 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
12466 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
12467 len * sizeof(struct bpf_insn));
12468 func[i]->type = prog->type;
12469 func[i]->len = len;
12470 if (bpf_prog_calc_tag(func[i]))
12472 func[i]->is_func = 1;
12473 func[i]->aux->func_idx = i;
12474 /* Below members will be freed only at prog->aux */
12475 func[i]->aux->btf = prog->aux->btf;
12476 func[i]->aux->func_info = prog->aux->func_info;
12477 func[i]->aux->poke_tab = prog->aux->poke_tab;
12478 func[i]->aux->size_poke_tab = prog->aux->size_poke_tab;
12480 for (j = 0; j < prog->aux->size_poke_tab; j++) {
12481 struct bpf_jit_poke_descriptor *poke;
12483 poke = &prog->aux->poke_tab[j];
12484 if (poke->insn_idx < subprog_end &&
12485 poke->insn_idx >= subprog_start)
12486 poke->aux = func[i]->aux;
12489 /* Use bpf_prog_F_tag to indicate functions in stack traces.
12490 * Long term would need debug info to populate names
12492 func[i]->aux->name[0] = 'F';
12493 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
12494 func[i]->jit_requested = 1;
12495 func[i]->aux->kfunc_tab = prog->aux->kfunc_tab;
12496 func[i]->aux->linfo = prog->aux->linfo;
12497 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
12498 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
12499 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
12501 insn = func[i]->insnsi;
12502 for (j = 0; j < func[i]->len; j++, insn++) {
12503 if (BPF_CLASS(insn->code) == BPF_LDX &&
12504 BPF_MODE(insn->code) == BPF_PROBE_MEM)
12507 func[i]->aux->num_exentries = num_exentries;
12508 func[i]->aux->tail_call_reachable = env->subprog_info[i].tail_call_reachable;
12509 func[i] = bpf_int_jit_compile(func[i]);
12510 if (!func[i]->jited) {
12517 /* at this point all bpf functions were successfully JITed
12518 * now populate all bpf_calls with correct addresses and
12519 * run last pass of JIT
12521 for (i = 0; i < env->subprog_cnt; i++) {
12522 insn = func[i]->insnsi;
12523 for (j = 0; j < func[i]->len; j++, insn++) {
12524 if (bpf_pseudo_func(insn)) {
12525 subprog = insn[1].imm;
12526 insn[0].imm = (u32)(long)func[subprog]->bpf_func;
12527 insn[1].imm = ((u64)(long)func[subprog]->bpf_func) >> 32;
12530 if (!bpf_pseudo_call(insn))
12532 subprog = insn->off;
12533 insn->imm = BPF_CAST_CALL(func[subprog]->bpf_func) -
12537 /* we use the aux data to keep a list of the start addresses
12538 * of the JITed images for each function in the program
12540 * for some architectures, such as powerpc64, the imm field
12541 * might not be large enough to hold the offset of the start
12542 * address of the callee's JITed image from __bpf_call_base
12544 * in such cases, we can lookup the start address of a callee
12545 * by using its subprog id, available from the off field of
12546 * the call instruction, as an index for this list
12548 func[i]->aux->func = func;
12549 func[i]->aux->func_cnt = env->subprog_cnt;
12551 for (i = 0; i < env->subprog_cnt; i++) {
12552 old_bpf_func = func[i]->bpf_func;
12553 tmp = bpf_int_jit_compile(func[i]);
12554 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
12555 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
12562 /* finally lock prog and jit images for all functions and
12563 * populate kallsysm
12565 for (i = 0; i < env->subprog_cnt; i++) {
12566 bpf_prog_lock_ro(func[i]);
12567 bpf_prog_kallsyms_add(func[i]);
12570 /* Last step: make now unused interpreter insns from main
12571 * prog consistent for later dump requests, so they can
12572 * later look the same as if they were interpreted only.
12574 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
12575 if (bpf_pseudo_func(insn)) {
12576 insn[0].imm = env->insn_aux_data[i].call_imm;
12577 insn[1].imm = find_subprog(env, i + insn[0].imm + 1);
12580 if (!bpf_pseudo_call(insn))
12582 insn->off = env->insn_aux_data[i].call_imm;
12583 subprog = find_subprog(env, i + insn->off + 1);
12584 insn->imm = subprog;
12588 prog->bpf_func = func[0]->bpf_func;
12589 prog->aux->func = func;
12590 prog->aux->func_cnt = env->subprog_cnt;
12591 bpf_prog_jit_attempt_done(prog);
12594 /* We failed JIT'ing, so at this point we need to unregister poke
12595 * descriptors from subprogs, so that kernel is not attempting to
12596 * patch it anymore as we're freeing the subprog JIT memory.
12598 for (i = 0; i < prog->aux->size_poke_tab; i++) {
12599 map_ptr = prog->aux->poke_tab[i].tail_call.map;
12600 map_ptr->ops->map_poke_untrack(map_ptr, prog->aux);
12602 /* At this point we're guaranteed that poke descriptors are not
12603 * live anymore. We can just unlink its descriptor table as it's
12604 * released with the main prog.
12606 for (i = 0; i < env->subprog_cnt; i++) {
12609 func[i]->aux->poke_tab = NULL;
12610 bpf_jit_free(func[i]);
12614 /* cleanup main prog to be interpreted */
12615 prog->jit_requested = 0;
12616 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
12617 if (!bpf_pseudo_call(insn))
12620 insn->imm = env->insn_aux_data[i].call_imm;
12622 bpf_prog_jit_attempt_done(prog);
12626 static int fixup_call_args(struct bpf_verifier_env *env)
12628 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
12629 struct bpf_prog *prog = env->prog;
12630 struct bpf_insn *insn = prog->insnsi;
12631 bool has_kfunc_call = bpf_prog_has_kfunc_call(prog);
12636 if (env->prog->jit_requested &&
12637 !bpf_prog_is_dev_bound(env->prog->aux)) {
12638 err = jit_subprogs(env);
12641 if (err == -EFAULT)
12644 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
12645 if (has_kfunc_call) {
12646 verbose(env, "calling kernel functions are not allowed in non-JITed programs\n");
12649 if (env->subprog_cnt > 1 && env->prog->aux->tail_call_reachable) {
12650 /* When JIT fails the progs with bpf2bpf calls and tail_calls
12651 * have to be rejected, since interpreter doesn't support them yet.
12653 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
12656 for (i = 0; i < prog->len; i++, insn++) {
12657 if (bpf_pseudo_func(insn)) {
12658 /* When JIT fails the progs with callback calls
12659 * have to be rejected, since interpreter doesn't support them yet.
12661 verbose(env, "callbacks are not allowed in non-JITed programs\n");
12665 if (!bpf_pseudo_call(insn))
12667 depth = get_callee_stack_depth(env, insn, i);
12670 bpf_patch_call_args(insn, depth);
12677 static int fixup_kfunc_call(struct bpf_verifier_env *env,
12678 struct bpf_insn *insn)
12680 const struct bpf_kfunc_desc *desc;
12682 /* insn->imm has the btf func_id. Replace it with
12683 * an address (relative to __bpf_base_call).
12685 desc = find_kfunc_desc(env->prog, insn->imm);
12687 verbose(env, "verifier internal error: kernel function descriptor not found for func_id %u\n",
12692 insn->imm = desc->imm;
12697 /* Do various post-verification rewrites in a single program pass.
12698 * These rewrites simplify JIT and interpreter implementations.
12700 static int do_misc_fixups(struct bpf_verifier_env *env)
12702 struct bpf_prog *prog = env->prog;
12703 bool expect_blinding = bpf_jit_blinding_enabled(prog);
12704 enum bpf_prog_type prog_type = resolve_prog_type(prog);
12705 struct bpf_insn *insn = prog->insnsi;
12706 const struct bpf_func_proto *fn;
12707 const int insn_cnt = prog->len;
12708 const struct bpf_map_ops *ops;
12709 struct bpf_insn_aux_data *aux;
12710 struct bpf_insn insn_buf[16];
12711 struct bpf_prog *new_prog;
12712 struct bpf_map *map_ptr;
12713 int i, ret, cnt, delta = 0;
12715 for (i = 0; i < insn_cnt; i++, insn++) {
12716 /* Make divide-by-zero exceptions impossible. */
12717 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
12718 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
12719 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
12720 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
12721 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
12722 bool isdiv = BPF_OP(insn->code) == BPF_DIV;
12723 struct bpf_insn *patchlet;
12724 struct bpf_insn chk_and_div[] = {
12725 /* [R,W]x div 0 -> 0 */
12726 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
12727 BPF_JNE | BPF_K, insn->src_reg,
12729 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
12730 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
12733 struct bpf_insn chk_and_mod[] = {
12734 /* [R,W]x mod 0 -> [R,W]x */
12735 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
12736 BPF_JEQ | BPF_K, insn->src_reg,
12737 0, 1 + (is64 ? 0 : 1), 0),
12739 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
12740 BPF_MOV32_REG(insn->dst_reg, insn->dst_reg),
12743 patchlet = isdiv ? chk_and_div : chk_and_mod;
12744 cnt = isdiv ? ARRAY_SIZE(chk_and_div) :
12745 ARRAY_SIZE(chk_and_mod) - (is64 ? 2 : 0);
12747 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
12752 env->prog = prog = new_prog;
12753 insn = new_prog->insnsi + i + delta;
12757 /* Implement LD_ABS and LD_IND with a rewrite, if supported by the program type. */
12758 if (BPF_CLASS(insn->code) == BPF_LD &&
12759 (BPF_MODE(insn->code) == BPF_ABS ||
12760 BPF_MODE(insn->code) == BPF_IND)) {
12761 cnt = env->ops->gen_ld_abs(insn, insn_buf);
12762 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
12763 verbose(env, "bpf verifier is misconfigured\n");
12767 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12772 env->prog = prog = new_prog;
12773 insn = new_prog->insnsi + i + delta;
12777 /* Rewrite pointer arithmetic to mitigate speculation attacks. */
12778 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
12779 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
12780 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
12781 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
12782 struct bpf_insn *patch = &insn_buf[0];
12783 bool issrc, isneg, isimm;
12786 aux = &env->insn_aux_data[i + delta];
12787 if (!aux->alu_state ||
12788 aux->alu_state == BPF_ALU_NON_POINTER)
12791 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
12792 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
12793 BPF_ALU_SANITIZE_SRC;
12794 isimm = aux->alu_state & BPF_ALU_IMMEDIATE;
12796 off_reg = issrc ? insn->src_reg : insn->dst_reg;
12798 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
12801 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
12802 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
12803 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
12804 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
12805 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
12806 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
12807 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX, off_reg);
12810 *patch++ = BPF_MOV64_REG(insn->dst_reg, insn->src_reg);
12811 insn->src_reg = BPF_REG_AX;
12813 insn->code = insn->code == code_add ?
12814 code_sub : code_add;
12816 if (issrc && isneg && !isimm)
12817 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
12818 cnt = patch - insn_buf;
12820 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12825 env->prog = prog = new_prog;
12826 insn = new_prog->insnsi + i + delta;
12830 if (insn->code != (BPF_JMP | BPF_CALL))
12832 if (insn->src_reg == BPF_PSEUDO_CALL)
12834 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
12835 ret = fixup_kfunc_call(env, insn);
12841 if (insn->imm == BPF_FUNC_get_route_realm)
12842 prog->dst_needed = 1;
12843 if (insn->imm == BPF_FUNC_get_prandom_u32)
12844 bpf_user_rnd_init_once();
12845 if (insn->imm == BPF_FUNC_override_return)
12846 prog->kprobe_override = 1;
12847 if (insn->imm == BPF_FUNC_tail_call) {
12848 /* If we tail call into other programs, we
12849 * cannot make any assumptions since they can
12850 * be replaced dynamically during runtime in
12851 * the program array.
12853 prog->cb_access = 1;
12854 if (!allow_tail_call_in_subprogs(env))
12855 prog->aux->stack_depth = MAX_BPF_STACK;
12856 prog->aux->max_pkt_offset = MAX_PACKET_OFF;
12858 /* mark bpf_tail_call as different opcode to avoid
12859 * conditional branch in the interpreter for every normal
12860 * call and to prevent accidental JITing by JIT compiler
12861 * that doesn't support bpf_tail_call yet
12864 insn->code = BPF_JMP | BPF_TAIL_CALL;
12866 aux = &env->insn_aux_data[i + delta];
12867 if (env->bpf_capable && !expect_blinding &&
12868 prog->jit_requested &&
12869 !bpf_map_key_poisoned(aux) &&
12870 !bpf_map_ptr_poisoned(aux) &&
12871 !bpf_map_ptr_unpriv(aux)) {
12872 struct bpf_jit_poke_descriptor desc = {
12873 .reason = BPF_POKE_REASON_TAIL_CALL,
12874 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
12875 .tail_call.key = bpf_map_key_immediate(aux),
12876 .insn_idx = i + delta,
12879 ret = bpf_jit_add_poke_descriptor(prog, &desc);
12881 verbose(env, "adding tail call poke descriptor failed\n");
12885 insn->imm = ret + 1;
12889 if (!bpf_map_ptr_unpriv(aux))
12892 /* instead of changing every JIT dealing with tail_call
12893 * emit two extra insns:
12894 * if (index >= max_entries) goto out;
12895 * index &= array->index_mask;
12896 * to avoid out-of-bounds cpu speculation
12898 if (bpf_map_ptr_poisoned(aux)) {
12899 verbose(env, "tail_call abusing map_ptr\n");
12903 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
12904 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
12905 map_ptr->max_entries, 2);
12906 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
12907 container_of(map_ptr,
12910 insn_buf[2] = *insn;
12912 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12917 env->prog = prog = new_prog;
12918 insn = new_prog->insnsi + i + delta;
12922 if (insn->imm == BPF_FUNC_timer_set_callback) {
12923 /* The verifier will process callback_fn as many times as necessary
12924 * with different maps and the register states prepared by
12925 * set_timer_callback_state will be accurate.
12927 * The following use case is valid:
12928 * map1 is shared by prog1, prog2, prog3.
12929 * prog1 calls bpf_timer_init for some map1 elements
12930 * prog2 calls bpf_timer_set_callback for some map1 elements.
12931 * Those that were not bpf_timer_init-ed will return -EINVAL.
12932 * prog3 calls bpf_timer_start for some map1 elements.
12933 * Those that were not both bpf_timer_init-ed and
12934 * bpf_timer_set_callback-ed will return -EINVAL.
12936 struct bpf_insn ld_addrs[2] = {
12937 BPF_LD_IMM64(BPF_REG_3, (long)prog->aux),
12940 insn_buf[0] = ld_addrs[0];
12941 insn_buf[1] = ld_addrs[1];
12942 insn_buf[2] = *insn;
12945 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12950 env->prog = prog = new_prog;
12951 insn = new_prog->insnsi + i + delta;
12952 goto patch_call_imm;
12955 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
12956 * and other inlining handlers are currently limited to 64 bit
12959 if (prog->jit_requested && BITS_PER_LONG == 64 &&
12960 (insn->imm == BPF_FUNC_map_lookup_elem ||
12961 insn->imm == BPF_FUNC_map_update_elem ||
12962 insn->imm == BPF_FUNC_map_delete_elem ||
12963 insn->imm == BPF_FUNC_map_push_elem ||
12964 insn->imm == BPF_FUNC_map_pop_elem ||
12965 insn->imm == BPF_FUNC_map_peek_elem ||
12966 insn->imm == BPF_FUNC_redirect_map)) {
12967 aux = &env->insn_aux_data[i + delta];
12968 if (bpf_map_ptr_poisoned(aux))
12969 goto patch_call_imm;
12971 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
12972 ops = map_ptr->ops;
12973 if (insn->imm == BPF_FUNC_map_lookup_elem &&
12974 ops->map_gen_lookup) {
12975 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
12976 if (cnt == -EOPNOTSUPP)
12977 goto patch_map_ops_generic;
12978 if (cnt <= 0 || cnt >= ARRAY_SIZE(insn_buf)) {
12979 verbose(env, "bpf verifier is misconfigured\n");
12983 new_prog = bpf_patch_insn_data(env, i + delta,
12989 env->prog = prog = new_prog;
12990 insn = new_prog->insnsi + i + delta;
12994 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
12995 (void *(*)(struct bpf_map *map, void *key))NULL));
12996 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
12997 (int (*)(struct bpf_map *map, void *key))NULL));
12998 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
12999 (int (*)(struct bpf_map *map, void *key, void *value,
13001 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
13002 (int (*)(struct bpf_map *map, void *value,
13004 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
13005 (int (*)(struct bpf_map *map, void *value))NULL));
13006 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
13007 (int (*)(struct bpf_map *map, void *value))NULL));
13008 BUILD_BUG_ON(!__same_type(ops->map_redirect,
13009 (int (*)(struct bpf_map *map, u32 ifindex, u64 flags))NULL));
13011 patch_map_ops_generic:
13012 switch (insn->imm) {
13013 case BPF_FUNC_map_lookup_elem:
13014 insn->imm = BPF_CAST_CALL(ops->map_lookup_elem) -
13017 case BPF_FUNC_map_update_elem:
13018 insn->imm = BPF_CAST_CALL(ops->map_update_elem) -
13021 case BPF_FUNC_map_delete_elem:
13022 insn->imm = BPF_CAST_CALL(ops->map_delete_elem) -
13025 case BPF_FUNC_map_push_elem:
13026 insn->imm = BPF_CAST_CALL(ops->map_push_elem) -
13029 case BPF_FUNC_map_pop_elem:
13030 insn->imm = BPF_CAST_CALL(ops->map_pop_elem) -
13033 case BPF_FUNC_map_peek_elem:
13034 insn->imm = BPF_CAST_CALL(ops->map_peek_elem) -
13037 case BPF_FUNC_redirect_map:
13038 insn->imm = BPF_CAST_CALL(ops->map_redirect) -
13043 goto patch_call_imm;
13046 /* Implement bpf_jiffies64 inline. */
13047 if (prog->jit_requested && BITS_PER_LONG == 64 &&
13048 insn->imm == BPF_FUNC_jiffies64) {
13049 struct bpf_insn ld_jiffies_addr[2] = {
13050 BPF_LD_IMM64(BPF_REG_0,
13051 (unsigned long)&jiffies),
13054 insn_buf[0] = ld_jiffies_addr[0];
13055 insn_buf[1] = ld_jiffies_addr[1];
13056 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
13060 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
13066 env->prog = prog = new_prog;
13067 insn = new_prog->insnsi + i + delta;
13071 /* Implement bpf_get_func_ip inline. */
13072 if (prog_type == BPF_PROG_TYPE_TRACING &&
13073 insn->imm == BPF_FUNC_get_func_ip) {
13074 /* Load IP address from ctx - 8 */
13075 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
13077 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
13081 env->prog = prog = new_prog;
13082 insn = new_prog->insnsi + i + delta;
13087 fn = env->ops->get_func_proto(insn->imm, env->prog);
13088 /* all functions that have prototype and verifier allowed
13089 * programs to call them, must be real in-kernel functions
13093 "kernel subsystem misconfigured func %s#%d\n",
13094 func_id_name(insn->imm), insn->imm);
13097 insn->imm = fn->func - __bpf_call_base;
13100 /* Since poke tab is now finalized, publish aux to tracker. */
13101 for (i = 0; i < prog->aux->size_poke_tab; i++) {
13102 map_ptr = prog->aux->poke_tab[i].tail_call.map;
13103 if (!map_ptr->ops->map_poke_track ||
13104 !map_ptr->ops->map_poke_untrack ||
13105 !map_ptr->ops->map_poke_run) {
13106 verbose(env, "bpf verifier is misconfigured\n");
13110 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
13112 verbose(env, "tracking tail call prog failed\n");
13117 sort_kfunc_descs_by_imm(env->prog);
13122 static void free_states(struct bpf_verifier_env *env)
13124 struct bpf_verifier_state_list *sl, *sln;
13127 sl = env->free_list;
13130 free_verifier_state(&sl->state, false);
13134 env->free_list = NULL;
13136 if (!env->explored_states)
13139 for (i = 0; i < state_htab_size(env); i++) {
13140 sl = env->explored_states[i];
13144 free_verifier_state(&sl->state, false);
13148 env->explored_states[i] = NULL;
13152 static int do_check_common(struct bpf_verifier_env *env, int subprog)
13154 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
13155 struct bpf_verifier_state *state;
13156 struct bpf_reg_state *regs;
13159 env->prev_linfo = NULL;
13162 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
13165 state->curframe = 0;
13166 state->speculative = false;
13167 state->branches = 1;
13168 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
13169 if (!state->frame[0]) {
13173 env->cur_state = state;
13174 init_func_state(env, state->frame[0],
13175 BPF_MAIN_FUNC /* callsite */,
13179 regs = state->frame[state->curframe]->regs;
13180 if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
13181 ret = btf_prepare_func_args(env, subprog, regs);
13184 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
13185 if (regs[i].type == PTR_TO_CTX)
13186 mark_reg_known_zero(env, regs, i);
13187 else if (regs[i].type == SCALAR_VALUE)
13188 mark_reg_unknown(env, regs, i);
13189 else if (base_type(regs[i].type) == PTR_TO_MEM) {
13190 const u32 mem_size = regs[i].mem_size;
13192 mark_reg_known_zero(env, regs, i);
13193 regs[i].mem_size = mem_size;
13194 regs[i].id = ++env->id_gen;
13198 /* 1st arg to a function */
13199 regs[BPF_REG_1].type = PTR_TO_CTX;
13200 mark_reg_known_zero(env, regs, BPF_REG_1);
13201 ret = btf_check_subprog_arg_match(env, subprog, regs);
13202 if (ret == -EFAULT)
13203 /* unlikely verifier bug. abort.
13204 * ret == 0 and ret < 0 are sadly acceptable for
13205 * main() function due to backward compatibility.
13206 * Like socket filter program may be written as:
13207 * int bpf_prog(struct pt_regs *ctx)
13208 * and never dereference that ctx in the program.
13209 * 'struct pt_regs' is a type mismatch for socket
13210 * filter that should be using 'struct __sk_buff'.
13215 ret = do_check(env);
13217 /* check for NULL is necessary, since cur_state can be freed inside
13218 * do_check() under memory pressure.
13220 if (env->cur_state) {
13221 free_verifier_state(env->cur_state, true);
13222 env->cur_state = NULL;
13224 while (!pop_stack(env, NULL, NULL, false));
13225 if (!ret && pop_log)
13226 bpf_vlog_reset(&env->log, 0);
13231 /* Verify all global functions in a BPF program one by one based on their BTF.
13232 * All global functions must pass verification. Otherwise the whole program is rejected.
13243 * foo() will be verified first for R1=any_scalar_value. During verification it
13244 * will be assumed that bar() already verified successfully and call to bar()
13245 * from foo() will be checked for type match only. Later bar() will be verified
13246 * independently to check that it's safe for R1=any_scalar_value.
13248 static int do_check_subprogs(struct bpf_verifier_env *env)
13250 struct bpf_prog_aux *aux = env->prog->aux;
13253 if (!aux->func_info)
13256 for (i = 1; i < env->subprog_cnt; i++) {
13257 if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL)
13259 env->insn_idx = env->subprog_info[i].start;
13260 WARN_ON_ONCE(env->insn_idx == 0);
13261 ret = do_check_common(env, i);
13264 } else if (env->log.level & BPF_LOG_LEVEL) {
13266 "Func#%d is safe for any args that match its prototype\n",
13273 static int do_check_main(struct bpf_verifier_env *env)
13278 ret = do_check_common(env, 0);
13280 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
13285 static void print_verification_stats(struct bpf_verifier_env *env)
13289 if (env->log.level & BPF_LOG_STATS) {
13290 verbose(env, "verification time %lld usec\n",
13291 div_u64(env->verification_time, 1000));
13292 verbose(env, "stack depth ");
13293 for (i = 0; i < env->subprog_cnt; i++) {
13294 u32 depth = env->subprog_info[i].stack_depth;
13296 verbose(env, "%d", depth);
13297 if (i + 1 < env->subprog_cnt)
13300 verbose(env, "\n");
13302 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
13303 "total_states %d peak_states %d mark_read %d\n",
13304 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
13305 env->max_states_per_insn, env->total_states,
13306 env->peak_states, env->longest_mark_read_walk);
13309 static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
13311 const struct btf_type *t, *func_proto;
13312 const struct bpf_struct_ops *st_ops;
13313 const struct btf_member *member;
13314 struct bpf_prog *prog = env->prog;
13315 u32 btf_id, member_idx;
13318 if (!prog->gpl_compatible) {
13319 verbose(env, "struct ops programs must have a GPL compatible license\n");
13323 btf_id = prog->aux->attach_btf_id;
13324 st_ops = bpf_struct_ops_find(btf_id);
13326 verbose(env, "attach_btf_id %u is not a supported struct\n",
13332 member_idx = prog->expected_attach_type;
13333 if (member_idx >= btf_type_vlen(t)) {
13334 verbose(env, "attach to invalid member idx %u of struct %s\n",
13335 member_idx, st_ops->name);
13339 member = &btf_type_member(t)[member_idx];
13340 mname = btf_name_by_offset(btf_vmlinux, member->name_off);
13341 func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type,
13344 verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
13345 mname, member_idx, st_ops->name);
13349 if (st_ops->check_member) {
13350 int err = st_ops->check_member(t, member);
13353 verbose(env, "attach to unsupported member %s of struct %s\n",
13354 mname, st_ops->name);
13359 prog->aux->attach_func_proto = func_proto;
13360 prog->aux->attach_func_name = mname;
13361 env->ops = st_ops->verifier_ops;
13365 #define SECURITY_PREFIX "security_"
13367 static int check_attach_modify_return(unsigned long addr, const char *func_name)
13369 if (within_error_injection_list(addr) ||
13370 !strncmp(SECURITY_PREFIX, func_name, sizeof(SECURITY_PREFIX) - 1))
13376 /* list of non-sleepable functions that are otherwise on
13377 * ALLOW_ERROR_INJECTION list
13379 BTF_SET_START(btf_non_sleepable_error_inject)
13380 /* Three functions below can be called from sleepable and non-sleepable context.
13381 * Assume non-sleepable from bpf safety point of view.
13383 BTF_ID(func, __add_to_page_cache_locked)
13384 BTF_ID(func, should_fail_alloc_page)
13385 BTF_ID(func, should_failslab)
13386 BTF_SET_END(btf_non_sleepable_error_inject)
13388 static int check_non_sleepable_error_inject(u32 btf_id)
13390 return btf_id_set_contains(&btf_non_sleepable_error_inject, btf_id);
13393 int bpf_check_attach_target(struct bpf_verifier_log *log,
13394 const struct bpf_prog *prog,
13395 const struct bpf_prog *tgt_prog,
13397 struct bpf_attach_target_info *tgt_info)
13399 bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
13400 const char prefix[] = "btf_trace_";
13401 int ret = 0, subprog = -1, i;
13402 const struct btf_type *t;
13403 bool conservative = true;
13409 bpf_log(log, "Tracing programs must provide btf_id\n");
13412 btf = tgt_prog ? tgt_prog->aux->btf : prog->aux->attach_btf;
13415 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
13418 t = btf_type_by_id(btf, btf_id);
13420 bpf_log(log, "attach_btf_id %u is invalid\n", btf_id);
13423 tname = btf_name_by_offset(btf, t->name_off);
13425 bpf_log(log, "attach_btf_id %u doesn't have a name\n", btf_id);
13429 struct bpf_prog_aux *aux = tgt_prog->aux;
13431 for (i = 0; i < aux->func_info_cnt; i++)
13432 if (aux->func_info[i].type_id == btf_id) {
13436 if (subprog == -1) {
13437 bpf_log(log, "Subprog %s doesn't exist\n", tname);
13440 conservative = aux->func_info_aux[subprog].unreliable;
13441 if (prog_extension) {
13442 if (conservative) {
13444 "Cannot replace static functions\n");
13447 if (!prog->jit_requested) {
13449 "Extension programs should be JITed\n");
13453 if (!tgt_prog->jited) {
13454 bpf_log(log, "Can attach to only JITed progs\n");
13457 if (tgt_prog->type == prog->type) {
13458 /* Cannot fentry/fexit another fentry/fexit program.
13459 * Cannot attach program extension to another extension.
13460 * It's ok to attach fentry/fexit to extension program.
13462 bpf_log(log, "Cannot recursively attach\n");
13465 if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
13467 (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
13468 tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
13469 /* Program extensions can extend all program types
13470 * except fentry/fexit. The reason is the following.
13471 * The fentry/fexit programs are used for performance
13472 * analysis, stats and can be attached to any program
13473 * type except themselves. When extension program is
13474 * replacing XDP function it is necessary to allow
13475 * performance analysis of all functions. Both original
13476 * XDP program and its program extension. Hence
13477 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
13478 * allowed. If extending of fentry/fexit was allowed it
13479 * would be possible to create long call chain
13480 * fentry->extension->fentry->extension beyond
13481 * reasonable stack size. Hence extending fentry is not
13484 bpf_log(log, "Cannot extend fentry/fexit\n");
13488 if (prog_extension) {
13489 bpf_log(log, "Cannot replace kernel functions\n");
13494 switch (prog->expected_attach_type) {
13495 case BPF_TRACE_RAW_TP:
13498 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
13501 if (!btf_type_is_typedef(t)) {
13502 bpf_log(log, "attach_btf_id %u is not a typedef\n",
13506 if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
13507 bpf_log(log, "attach_btf_id %u points to wrong type name %s\n",
13511 tname += sizeof(prefix) - 1;
13512 t = btf_type_by_id(btf, t->type);
13513 if (!btf_type_is_ptr(t))
13514 /* should never happen in valid vmlinux build */
13516 t = btf_type_by_id(btf, t->type);
13517 if (!btf_type_is_func_proto(t))
13518 /* should never happen in valid vmlinux build */
13522 case BPF_TRACE_ITER:
13523 if (!btf_type_is_func(t)) {
13524 bpf_log(log, "attach_btf_id %u is not a function\n",
13528 t = btf_type_by_id(btf, t->type);
13529 if (!btf_type_is_func_proto(t))
13531 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
13536 if (!prog_extension)
13539 case BPF_MODIFY_RETURN:
13541 case BPF_TRACE_FENTRY:
13542 case BPF_TRACE_FEXIT:
13543 if (!btf_type_is_func(t)) {
13544 bpf_log(log, "attach_btf_id %u is not a function\n",
13548 if (prog_extension &&
13549 btf_check_type_match(log, prog, btf, t))
13551 t = btf_type_by_id(btf, t->type);
13552 if (!btf_type_is_func_proto(t))
13555 if ((prog->aux->saved_dst_prog_type || prog->aux->saved_dst_attach_type) &&
13556 (!tgt_prog || prog->aux->saved_dst_prog_type != tgt_prog->type ||
13557 prog->aux->saved_dst_attach_type != tgt_prog->expected_attach_type))
13560 if (tgt_prog && conservative)
13563 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
13569 addr = (long) tgt_prog->bpf_func;
13571 addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
13573 addr = kallsyms_lookup_name(tname);
13576 "The address of function %s cannot be found\n",
13582 if (prog->aux->sleepable) {
13584 switch (prog->type) {
13585 case BPF_PROG_TYPE_TRACING:
13586 /* fentry/fexit/fmod_ret progs can be sleepable only if they are
13587 * attached to ALLOW_ERROR_INJECTION and are not in denylist.
13589 if (!check_non_sleepable_error_inject(btf_id) &&
13590 within_error_injection_list(addr))
13593 case BPF_PROG_TYPE_LSM:
13594 /* LSM progs check that they are attached to bpf_lsm_*() funcs.
13595 * Only some of them are sleepable.
13597 if (bpf_lsm_is_sleepable_hook(btf_id))
13604 bpf_log(log, "%s is not sleepable\n", tname);
13607 } else if (prog->expected_attach_type == BPF_MODIFY_RETURN) {
13609 bpf_log(log, "can't modify return codes of BPF programs\n");
13612 ret = check_attach_modify_return(addr, tname);
13614 bpf_log(log, "%s() is not modifiable\n", tname);
13621 tgt_info->tgt_addr = addr;
13622 tgt_info->tgt_name = tname;
13623 tgt_info->tgt_type = t;
13627 BTF_SET_START(btf_id_deny)
13630 BTF_ID(func, migrate_disable)
13631 BTF_ID(func, migrate_enable)
13633 #if !defined CONFIG_PREEMPT_RCU && !defined CONFIG_TINY_RCU
13634 BTF_ID(func, rcu_read_unlock_strict)
13636 BTF_SET_END(btf_id_deny)
13638 static int check_attach_btf_id(struct bpf_verifier_env *env)
13640 struct bpf_prog *prog = env->prog;
13641 struct bpf_prog *tgt_prog = prog->aux->dst_prog;
13642 struct bpf_attach_target_info tgt_info = {};
13643 u32 btf_id = prog->aux->attach_btf_id;
13644 struct bpf_trampoline *tr;
13648 if (prog->type == BPF_PROG_TYPE_SYSCALL) {
13649 if (prog->aux->sleepable)
13650 /* attach_btf_id checked to be zero already */
13652 verbose(env, "Syscall programs can only be sleepable\n");
13656 if (prog->aux->sleepable && prog->type != BPF_PROG_TYPE_TRACING &&
13657 prog->type != BPF_PROG_TYPE_LSM) {
13658 verbose(env, "Only fentry/fexit/fmod_ret and lsm programs can be sleepable\n");
13662 if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
13663 return check_struct_ops_btf_id(env);
13665 if (prog->type != BPF_PROG_TYPE_TRACING &&
13666 prog->type != BPF_PROG_TYPE_LSM &&
13667 prog->type != BPF_PROG_TYPE_EXT)
13670 ret = bpf_check_attach_target(&env->log, prog, tgt_prog, btf_id, &tgt_info);
13674 if (tgt_prog && prog->type == BPF_PROG_TYPE_EXT) {
13675 /* to make freplace equivalent to their targets, they need to
13676 * inherit env->ops and expected_attach_type for the rest of the
13679 env->ops = bpf_verifier_ops[tgt_prog->type];
13680 prog->expected_attach_type = tgt_prog->expected_attach_type;
13683 /* store info about the attachment target that will be used later */
13684 prog->aux->attach_func_proto = tgt_info.tgt_type;
13685 prog->aux->attach_func_name = tgt_info.tgt_name;
13688 prog->aux->saved_dst_prog_type = tgt_prog->type;
13689 prog->aux->saved_dst_attach_type = tgt_prog->expected_attach_type;
13692 if (prog->expected_attach_type == BPF_TRACE_RAW_TP) {
13693 prog->aux->attach_btf_trace = true;
13695 } else if (prog->expected_attach_type == BPF_TRACE_ITER) {
13696 if (!bpf_iter_prog_supported(prog))
13701 if (prog->type == BPF_PROG_TYPE_LSM) {
13702 ret = bpf_lsm_verify_prog(&env->log, prog);
13705 } else if (prog->type == BPF_PROG_TYPE_TRACING &&
13706 btf_id_set_contains(&btf_id_deny, btf_id)) {
13710 key = bpf_trampoline_compute_key(tgt_prog, prog->aux->attach_btf, btf_id);
13711 tr = bpf_trampoline_get(key, &tgt_info);
13715 prog->aux->dst_trampoline = tr;
13719 struct btf *bpf_get_btf_vmlinux(void)
13721 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
13722 mutex_lock(&bpf_verifier_lock);
13724 btf_vmlinux = btf_parse_vmlinux();
13725 mutex_unlock(&bpf_verifier_lock);
13727 return btf_vmlinux;
13730 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr, bpfptr_t uattr)
13732 u64 start_time = ktime_get_ns();
13733 struct bpf_verifier_env *env;
13734 struct bpf_verifier_log *log;
13735 int i, len, ret = -EINVAL;
13738 /* no program is valid */
13739 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
13742 /* 'struct bpf_verifier_env' can be global, but since it's not small,
13743 * allocate/free it every time bpf_check() is called
13745 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
13750 len = (*prog)->len;
13751 env->insn_aux_data =
13752 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
13754 if (!env->insn_aux_data)
13756 for (i = 0; i < len; i++)
13757 env->insn_aux_data[i].orig_idx = i;
13759 env->ops = bpf_verifier_ops[env->prog->type];
13760 env->fd_array = make_bpfptr(attr->fd_array, uattr.is_kernel);
13761 is_priv = bpf_capable();
13763 bpf_get_btf_vmlinux();
13765 /* grab the mutex to protect few globals used by verifier */
13767 mutex_lock(&bpf_verifier_lock);
13769 if (attr->log_level || attr->log_buf || attr->log_size) {
13770 /* user requested verbose verifier output
13771 * and supplied buffer to store the verification trace
13773 log->level = attr->log_level;
13774 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
13775 log->len_total = attr->log_size;
13777 /* log attributes have to be sane */
13778 if (!bpf_verifier_log_attr_valid(log)) {
13784 if (IS_ERR(btf_vmlinux)) {
13785 /* Either gcc or pahole or kernel are broken. */
13786 verbose(env, "in-kernel BTF is malformed\n");
13787 ret = PTR_ERR(btf_vmlinux);
13788 goto skip_full_check;
13791 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
13792 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
13793 env->strict_alignment = true;
13794 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
13795 env->strict_alignment = false;
13797 env->allow_ptr_leaks = bpf_allow_ptr_leaks();
13798 env->allow_uninit_stack = bpf_allow_uninit_stack();
13799 env->allow_ptr_to_map_access = bpf_allow_ptr_to_map_access();
13800 env->bypass_spec_v1 = bpf_bypass_spec_v1();
13801 env->bypass_spec_v4 = bpf_bypass_spec_v4();
13802 env->bpf_capable = bpf_capable();
13805 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
13807 env->explored_states = kvcalloc(state_htab_size(env),
13808 sizeof(struct bpf_verifier_state_list *),
13811 if (!env->explored_states)
13812 goto skip_full_check;
13814 ret = add_subprog_and_kfunc(env);
13816 goto skip_full_check;
13818 ret = check_subprogs(env);
13820 goto skip_full_check;
13822 ret = check_btf_info(env, attr, uattr);
13824 goto skip_full_check;
13826 ret = check_attach_btf_id(env);
13828 goto skip_full_check;
13830 ret = resolve_pseudo_ldimm64(env);
13832 goto skip_full_check;
13834 if (bpf_prog_is_dev_bound(env->prog->aux)) {
13835 ret = bpf_prog_offload_verifier_prep(env->prog);
13837 goto skip_full_check;
13840 ret = check_cfg(env);
13842 goto skip_full_check;
13844 ret = do_check_subprogs(env);
13845 ret = ret ?: do_check_main(env);
13847 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
13848 ret = bpf_prog_offload_finalize(env);
13851 kvfree(env->explored_states);
13854 ret = check_max_stack_depth(env);
13856 /* instruction rewrites happen after this point */
13859 opt_hard_wire_dead_code_branches(env);
13861 ret = opt_remove_dead_code(env);
13863 ret = opt_remove_nops(env);
13866 sanitize_dead_code(env);
13870 /* program is valid, convert *(u32*)(ctx + off) accesses */
13871 ret = convert_ctx_accesses(env);
13874 ret = do_misc_fixups(env);
13876 /* do 32-bit optimization after insn patching has done so those patched
13877 * insns could be handled correctly.
13879 if (ret == 0 && !bpf_prog_is_dev_bound(env->prog->aux)) {
13880 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
13881 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
13886 ret = fixup_call_args(env);
13888 env->verification_time = ktime_get_ns() - start_time;
13889 print_verification_stats(env);
13891 if (log->level && bpf_verifier_log_full(log))
13893 if (log->level && !log->ubuf) {
13895 goto err_release_maps;
13899 goto err_release_maps;
13901 if (env->used_map_cnt) {
13902 /* if program passed verifier, update used_maps in bpf_prog_info */
13903 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
13904 sizeof(env->used_maps[0]),
13907 if (!env->prog->aux->used_maps) {
13909 goto err_release_maps;
13912 memcpy(env->prog->aux->used_maps, env->used_maps,
13913 sizeof(env->used_maps[0]) * env->used_map_cnt);
13914 env->prog->aux->used_map_cnt = env->used_map_cnt;
13916 if (env->used_btf_cnt) {
13917 /* if program passed verifier, update used_btfs in bpf_prog_aux */
13918 env->prog->aux->used_btfs = kmalloc_array(env->used_btf_cnt,
13919 sizeof(env->used_btfs[0]),
13921 if (!env->prog->aux->used_btfs) {
13923 goto err_release_maps;
13926 memcpy(env->prog->aux->used_btfs, env->used_btfs,
13927 sizeof(env->used_btfs[0]) * env->used_btf_cnt);
13928 env->prog->aux->used_btf_cnt = env->used_btf_cnt;
13930 if (env->used_map_cnt || env->used_btf_cnt) {
13931 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
13932 * bpf_ld_imm64 instructions
13934 convert_pseudo_ld_imm64(env);
13937 adjust_btf_func(env);
13940 if (!env->prog->aux->used_maps)
13941 /* if we didn't copy map pointers into bpf_prog_info, release
13942 * them now. Otherwise free_used_maps() will release them.
13945 if (!env->prog->aux->used_btfs)
13948 /* extension progs temporarily inherit the attach_type of their targets
13949 for verification purposes, so set it back to zero before returning
13951 if (env->prog->type == BPF_PROG_TYPE_EXT)
13952 env->prog->expected_attach_type = 0;
13957 mutex_unlock(&bpf_verifier_lock);
13958 vfree(env->insn_aux_data);