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 struct bpf_call_arg_meta {
244 struct bpf_map *map_ptr;
261 struct btf *btf_vmlinux;
263 static DEFINE_MUTEX(bpf_verifier_lock);
265 static const struct bpf_line_info *
266 find_linfo(const struct bpf_verifier_env *env, u32 insn_off)
268 const struct bpf_line_info *linfo;
269 const struct bpf_prog *prog;
273 nr_linfo = prog->aux->nr_linfo;
275 if (!nr_linfo || insn_off >= prog->len)
278 linfo = prog->aux->linfo;
279 for (i = 1; i < nr_linfo; i++)
280 if (insn_off < linfo[i].insn_off)
283 return &linfo[i - 1];
286 void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
291 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
293 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
294 "verifier log line truncated - local buffer too short\n");
296 n = min(log->len_total - log->len_used - 1, n);
299 if (log->level == BPF_LOG_KERNEL) {
300 pr_err("BPF:%s\n", log->kbuf);
303 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
309 static void bpf_vlog_reset(struct bpf_verifier_log *log, u32 new_pos)
313 if (!bpf_verifier_log_needed(log))
316 log->len_used = new_pos;
317 if (put_user(zero, log->ubuf + new_pos))
321 /* log_level controls verbosity level of eBPF verifier.
322 * bpf_verifier_log_write() is used to dump the verification trace to the log,
323 * so the user can figure out what's wrong with the program
325 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
326 const char *fmt, ...)
330 if (!bpf_verifier_log_needed(&env->log))
334 bpf_verifier_vlog(&env->log, fmt, args);
337 EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
339 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
341 struct bpf_verifier_env *env = private_data;
344 if (!bpf_verifier_log_needed(&env->log))
348 bpf_verifier_vlog(&env->log, fmt, args);
352 __printf(2, 3) void bpf_log(struct bpf_verifier_log *log,
353 const char *fmt, ...)
357 if (!bpf_verifier_log_needed(log))
361 bpf_verifier_vlog(log, fmt, args);
365 static const char *ltrim(const char *s)
373 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env,
375 const char *prefix_fmt, ...)
377 const struct bpf_line_info *linfo;
379 if (!bpf_verifier_log_needed(&env->log))
382 linfo = find_linfo(env, insn_off);
383 if (!linfo || linfo == env->prev_linfo)
389 va_start(args, prefix_fmt);
390 bpf_verifier_vlog(&env->log, prefix_fmt, args);
395 ltrim(btf_name_by_offset(env->prog->aux->btf,
398 env->prev_linfo = linfo;
401 static void verbose_invalid_scalar(struct bpf_verifier_env *env,
402 struct bpf_reg_state *reg,
403 struct tnum *range, const char *ctx,
404 const char *reg_name)
408 verbose(env, "At %s the register %s ", ctx, reg_name);
409 if (!tnum_is_unknown(reg->var_off)) {
410 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
411 verbose(env, "has value %s", tn_buf);
413 verbose(env, "has unknown scalar value");
415 tnum_strn(tn_buf, sizeof(tn_buf), *range);
416 verbose(env, " should have been in %s\n", tn_buf);
419 static bool type_is_pkt_pointer(enum bpf_reg_type type)
421 return type == PTR_TO_PACKET ||
422 type == PTR_TO_PACKET_META;
425 static bool type_is_sk_pointer(enum bpf_reg_type type)
427 return type == PTR_TO_SOCKET ||
428 type == PTR_TO_SOCK_COMMON ||
429 type == PTR_TO_TCP_SOCK ||
430 type == PTR_TO_XDP_SOCK;
433 static bool reg_type_not_null(enum bpf_reg_type type)
435 return type == PTR_TO_SOCKET ||
436 type == PTR_TO_TCP_SOCK ||
437 type == PTR_TO_MAP_VALUE ||
438 type == PTR_TO_MAP_KEY ||
439 type == PTR_TO_SOCK_COMMON;
442 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
444 return reg->type == PTR_TO_MAP_VALUE &&
445 map_value_has_spin_lock(reg->map_ptr);
448 static bool reg_type_may_be_refcounted_or_null(enum bpf_reg_type type)
450 return base_type(type) == PTR_TO_SOCKET ||
451 base_type(type) == PTR_TO_TCP_SOCK ||
452 base_type(type) == PTR_TO_MEM;
455 static bool type_is_rdonly_mem(u32 type)
457 return type & MEM_RDONLY;
460 static bool arg_type_may_be_refcounted(enum bpf_arg_type type)
462 return type == ARG_PTR_TO_SOCK_COMMON;
465 static bool type_may_be_null(u32 type)
467 return type & PTR_MAYBE_NULL;
470 /* Determine whether the function releases some resources allocated by another
471 * function call. The first reference type argument will be assumed to be
472 * released by release_reference().
474 static bool is_release_function(enum bpf_func_id func_id)
476 return func_id == BPF_FUNC_sk_release ||
477 func_id == BPF_FUNC_ringbuf_submit ||
478 func_id == BPF_FUNC_ringbuf_discard;
481 static bool may_be_acquire_function(enum bpf_func_id func_id)
483 return func_id == BPF_FUNC_sk_lookup_tcp ||
484 func_id == BPF_FUNC_sk_lookup_udp ||
485 func_id == BPF_FUNC_skc_lookup_tcp ||
486 func_id == BPF_FUNC_map_lookup_elem ||
487 func_id == BPF_FUNC_ringbuf_reserve;
490 static bool is_acquire_function(enum bpf_func_id func_id,
491 const struct bpf_map *map)
493 enum bpf_map_type map_type = map ? map->map_type : BPF_MAP_TYPE_UNSPEC;
495 if (func_id == BPF_FUNC_sk_lookup_tcp ||
496 func_id == BPF_FUNC_sk_lookup_udp ||
497 func_id == BPF_FUNC_skc_lookup_tcp ||
498 func_id == BPF_FUNC_ringbuf_reserve)
501 if (func_id == BPF_FUNC_map_lookup_elem &&
502 (map_type == BPF_MAP_TYPE_SOCKMAP ||
503 map_type == BPF_MAP_TYPE_SOCKHASH))
509 static bool is_ptr_cast_function(enum bpf_func_id func_id)
511 return func_id == BPF_FUNC_tcp_sock ||
512 func_id == BPF_FUNC_sk_fullsock ||
513 func_id == BPF_FUNC_skc_to_tcp_sock ||
514 func_id == BPF_FUNC_skc_to_tcp6_sock ||
515 func_id == BPF_FUNC_skc_to_udp6_sock ||
516 func_id == BPF_FUNC_skc_to_tcp_timewait_sock ||
517 func_id == BPF_FUNC_skc_to_tcp_request_sock;
520 static bool is_cmpxchg_insn(const struct bpf_insn *insn)
522 return BPF_CLASS(insn->code) == BPF_STX &&
523 BPF_MODE(insn->code) == BPF_ATOMIC &&
524 insn->imm == BPF_CMPXCHG;
527 /* string representation of 'enum bpf_reg_type'
529 * Note that reg_type_str() can not appear more than once in a single verbose()
532 static const char *reg_type_str(struct bpf_verifier_env *env,
533 enum bpf_reg_type type)
535 char postfix[16] = {0}, prefix[16] = {0};
536 static const char * const str[] = {
538 [SCALAR_VALUE] = "inv",
539 [PTR_TO_CTX] = "ctx",
540 [CONST_PTR_TO_MAP] = "map_ptr",
541 [PTR_TO_MAP_VALUE] = "map_value",
542 [PTR_TO_STACK] = "fp",
543 [PTR_TO_PACKET] = "pkt",
544 [PTR_TO_PACKET_META] = "pkt_meta",
545 [PTR_TO_PACKET_END] = "pkt_end",
546 [PTR_TO_FLOW_KEYS] = "flow_keys",
547 [PTR_TO_SOCKET] = "sock",
548 [PTR_TO_SOCK_COMMON] = "sock_common",
549 [PTR_TO_TCP_SOCK] = "tcp_sock",
550 [PTR_TO_TP_BUFFER] = "tp_buffer",
551 [PTR_TO_XDP_SOCK] = "xdp_sock",
552 [PTR_TO_BTF_ID] = "ptr_",
553 [PTR_TO_PERCPU_BTF_ID] = "percpu_ptr_",
554 [PTR_TO_MEM] = "mem",
555 [PTR_TO_BUF] = "buf",
556 [PTR_TO_FUNC] = "func",
557 [PTR_TO_MAP_KEY] = "map_key",
560 if (type & PTR_MAYBE_NULL) {
561 if (base_type(type) == PTR_TO_BTF_ID ||
562 base_type(type) == PTR_TO_PERCPU_BTF_ID)
563 strncpy(postfix, "or_null_", 16);
565 strncpy(postfix, "_or_null", 16);
568 if (type & MEM_RDONLY)
569 strncpy(prefix, "rdonly_", 16);
571 snprintf(env->type_str_buf, TYPE_STR_BUF_LEN, "%s%s%s",
572 prefix, str[base_type(type)], postfix);
573 return env->type_str_buf;
576 static char slot_type_char[] = {
577 [STACK_INVALID] = '?',
583 static void print_liveness(struct bpf_verifier_env *env,
584 enum bpf_reg_liveness live)
586 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
588 if (live & REG_LIVE_READ)
590 if (live & REG_LIVE_WRITTEN)
592 if (live & REG_LIVE_DONE)
596 static struct bpf_func_state *func(struct bpf_verifier_env *env,
597 const struct bpf_reg_state *reg)
599 struct bpf_verifier_state *cur = env->cur_state;
601 return cur->frame[reg->frameno];
604 static const char *kernel_type_name(const struct btf* btf, u32 id)
606 return btf_name_by_offset(btf, btf_type_by_id(btf, id)->name_off);
609 static void print_verifier_state(struct bpf_verifier_env *env,
610 const struct bpf_func_state *state)
612 const struct bpf_reg_state *reg;
617 verbose(env, " frame%d:", state->frameno);
618 for (i = 0; i < MAX_BPF_REG; i++) {
619 reg = &state->regs[i];
623 verbose(env, " R%d", i);
624 print_liveness(env, reg->live);
625 verbose(env, "=%s", reg_type_str(env, t));
626 if (t == SCALAR_VALUE && reg->precise)
628 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
629 tnum_is_const(reg->var_off)) {
630 /* reg->off should be 0 for SCALAR_VALUE */
631 verbose(env, "%lld", reg->var_off.value + reg->off);
633 if (base_type(t) == PTR_TO_BTF_ID ||
634 base_type(t) == PTR_TO_PERCPU_BTF_ID)
635 verbose(env, "%s", kernel_type_name(reg->btf, reg->btf_id));
636 verbose(env, "(id=%d", reg->id);
637 if (reg_type_may_be_refcounted_or_null(t))
638 verbose(env, ",ref_obj_id=%d", reg->ref_obj_id);
639 if (t != SCALAR_VALUE)
640 verbose(env, ",off=%d", reg->off);
641 if (type_is_pkt_pointer(t))
642 verbose(env, ",r=%d", reg->range);
643 else if (base_type(t) == CONST_PTR_TO_MAP ||
644 base_type(t) == PTR_TO_MAP_KEY ||
645 base_type(t) == PTR_TO_MAP_VALUE)
646 verbose(env, ",ks=%d,vs=%d",
647 reg->map_ptr->key_size,
648 reg->map_ptr->value_size);
649 if (tnum_is_const(reg->var_off)) {
650 /* Typically an immediate SCALAR_VALUE, but
651 * could be a pointer whose offset is too big
654 verbose(env, ",imm=%llx", reg->var_off.value);
656 if (reg->smin_value != reg->umin_value &&
657 reg->smin_value != S64_MIN)
658 verbose(env, ",smin_value=%lld",
659 (long long)reg->smin_value);
660 if (reg->smax_value != reg->umax_value &&
661 reg->smax_value != S64_MAX)
662 verbose(env, ",smax_value=%lld",
663 (long long)reg->smax_value);
664 if (reg->umin_value != 0)
665 verbose(env, ",umin_value=%llu",
666 (unsigned long long)reg->umin_value);
667 if (reg->umax_value != U64_MAX)
668 verbose(env, ",umax_value=%llu",
669 (unsigned long long)reg->umax_value);
670 if (!tnum_is_unknown(reg->var_off)) {
673 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
674 verbose(env, ",var_off=%s", tn_buf);
676 if (reg->s32_min_value != reg->smin_value &&
677 reg->s32_min_value != S32_MIN)
678 verbose(env, ",s32_min_value=%d",
679 (int)(reg->s32_min_value));
680 if (reg->s32_max_value != reg->smax_value &&
681 reg->s32_max_value != S32_MAX)
682 verbose(env, ",s32_max_value=%d",
683 (int)(reg->s32_max_value));
684 if (reg->u32_min_value != reg->umin_value &&
685 reg->u32_min_value != U32_MIN)
686 verbose(env, ",u32_min_value=%d",
687 (int)(reg->u32_min_value));
688 if (reg->u32_max_value != reg->umax_value &&
689 reg->u32_max_value != U32_MAX)
690 verbose(env, ",u32_max_value=%d",
691 (int)(reg->u32_max_value));
696 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
697 char types_buf[BPF_REG_SIZE + 1];
701 for (j = 0; j < BPF_REG_SIZE; j++) {
702 if (state->stack[i].slot_type[j] != STACK_INVALID)
704 types_buf[j] = slot_type_char[
705 state->stack[i].slot_type[j]];
707 types_buf[BPF_REG_SIZE] = 0;
710 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
711 print_liveness(env, state->stack[i].spilled_ptr.live);
712 if (state->stack[i].slot_type[0] == STACK_SPILL) {
713 reg = &state->stack[i].spilled_ptr;
715 verbose(env, "=%s", reg_type_str(env, t));
716 if (t == SCALAR_VALUE && reg->precise)
718 if (t == SCALAR_VALUE && tnum_is_const(reg->var_off))
719 verbose(env, "%lld", reg->var_off.value + reg->off);
721 verbose(env, "=%s", types_buf);
724 if (state->acquired_refs && state->refs[0].id) {
725 verbose(env, " refs=%d", state->refs[0].id);
726 for (i = 1; i < state->acquired_refs; i++)
727 if (state->refs[i].id)
728 verbose(env, ",%d", state->refs[i].id);
730 if (state->in_callback_fn)
732 if (state->in_async_callback_fn)
733 verbose(env, " async_cb");
737 /* copy array src of length n * size bytes to dst. dst is reallocated if it's too
738 * small to hold src. This is different from krealloc since we don't want to preserve
739 * the contents of dst.
741 * Leaves dst untouched if src is NULL or length is zero. Returns NULL if memory could
744 static void *copy_array(void *dst, const void *src, size_t n, size_t size, gfp_t flags)
748 if (ZERO_OR_NULL_PTR(src))
751 if (unlikely(check_mul_overflow(n, size, &bytes)))
754 if (ksize(dst) < bytes) {
756 dst = kmalloc_track_caller(bytes, flags);
761 memcpy(dst, src, bytes);
763 return dst ? dst : ZERO_SIZE_PTR;
766 /* resize an array from old_n items to new_n items. the array is reallocated if it's too
767 * small to hold new_n items. new items are zeroed out if the array grows.
769 * Contrary to krealloc_array, does not free arr if new_n is zero.
771 static void *realloc_array(void *arr, size_t old_n, size_t new_n, size_t size)
775 if (!new_n || old_n == new_n)
778 new_arr = krealloc_array(arr, new_n, size, GFP_KERNEL);
786 memset(arr + old_n * size, 0, (new_n - old_n) * size);
789 return arr ? arr : ZERO_SIZE_PTR;
792 static int copy_reference_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
794 dst->refs = copy_array(dst->refs, src->refs, src->acquired_refs,
795 sizeof(struct bpf_reference_state), GFP_KERNEL);
799 dst->acquired_refs = src->acquired_refs;
803 static int copy_stack_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
805 size_t n = src->allocated_stack / BPF_REG_SIZE;
807 dst->stack = copy_array(dst->stack, src->stack, n, sizeof(struct bpf_stack_state),
812 dst->allocated_stack = src->allocated_stack;
816 static int resize_reference_state(struct bpf_func_state *state, size_t n)
818 state->refs = realloc_array(state->refs, state->acquired_refs, n,
819 sizeof(struct bpf_reference_state));
823 state->acquired_refs = n;
827 static int grow_stack_state(struct bpf_func_state *state, int size)
829 size_t old_n = state->allocated_stack / BPF_REG_SIZE, n = size / BPF_REG_SIZE;
834 state->stack = realloc_array(state->stack, old_n, n, sizeof(struct bpf_stack_state));
838 state->allocated_stack = size;
842 /* Acquire a pointer id from the env and update the state->refs to include
843 * this new pointer reference.
844 * On success, returns a valid pointer id to associate with the register
845 * On failure, returns a negative errno.
847 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
849 struct bpf_func_state *state = cur_func(env);
850 int new_ofs = state->acquired_refs;
853 err = resize_reference_state(state, state->acquired_refs + 1);
857 state->refs[new_ofs].id = id;
858 state->refs[new_ofs].insn_idx = insn_idx;
859 state->refs[new_ofs].callback_ref = state->in_callback_fn ? state->frameno : 0;
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 /* Cannot release caller references in callbacks */
873 if (state->in_callback_fn && state->refs[i].callback_ref != state->frameno)
875 if (last_idx && i != last_idx)
876 memcpy(&state->refs[i], &state->refs[last_idx],
877 sizeof(*state->refs));
878 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
879 state->acquired_refs--;
886 static void free_func_state(struct bpf_func_state *state)
895 static void clear_jmp_history(struct bpf_verifier_state *state)
897 kfree(state->jmp_history);
898 state->jmp_history = NULL;
899 state->jmp_history_cnt = 0;
902 static void free_verifier_state(struct bpf_verifier_state *state,
907 for (i = 0; i <= state->curframe; i++) {
908 free_func_state(state->frame[i]);
909 state->frame[i] = NULL;
911 clear_jmp_history(state);
916 /* copy verifier state from src to dst growing dst stack space
917 * when necessary to accommodate larger src stack
919 static int copy_func_state(struct bpf_func_state *dst,
920 const struct bpf_func_state *src)
924 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
925 err = copy_reference_state(dst, src);
928 return copy_stack_state(dst, src);
931 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
932 const struct bpf_verifier_state *src)
934 struct bpf_func_state *dst;
937 dst_state->jmp_history = copy_array(dst_state->jmp_history, src->jmp_history,
938 src->jmp_history_cnt, sizeof(struct bpf_idx_pair),
940 if (!dst_state->jmp_history)
942 dst_state->jmp_history_cnt = src->jmp_history_cnt;
944 /* if dst has more stack frames then src frame, free them */
945 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
946 free_func_state(dst_state->frame[i]);
947 dst_state->frame[i] = NULL;
949 dst_state->speculative = src->speculative;
950 dst_state->curframe = src->curframe;
951 dst_state->active_spin_lock = src->active_spin_lock;
952 dst_state->branches = src->branches;
953 dst_state->parent = src->parent;
954 dst_state->first_insn_idx = src->first_insn_idx;
955 dst_state->last_insn_idx = src->last_insn_idx;
956 for (i = 0; i <= src->curframe; i++) {
957 dst = dst_state->frame[i];
959 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
962 dst_state->frame[i] = dst;
964 err = copy_func_state(dst, src->frame[i]);
971 static void update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
974 u32 br = --st->branches;
976 /* WARN_ON(br > 1) technically makes sense here,
977 * but see comment in push_stack(), hence:
979 WARN_ONCE((int)br < 0,
980 "BUG update_branch_counts:branches_to_explore=%d\n",
988 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
989 int *insn_idx, bool pop_log)
991 struct bpf_verifier_state *cur = env->cur_state;
992 struct bpf_verifier_stack_elem *elem, *head = env->head;
995 if (env->head == NULL)
999 err = copy_verifier_state(cur, &head->st);
1004 bpf_vlog_reset(&env->log, head->log_pos);
1006 *insn_idx = head->insn_idx;
1008 *prev_insn_idx = head->prev_insn_idx;
1010 free_verifier_state(&head->st, false);
1017 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
1018 int insn_idx, int prev_insn_idx,
1021 struct bpf_verifier_state *cur = env->cur_state;
1022 struct bpf_verifier_stack_elem *elem;
1025 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
1029 elem->insn_idx = insn_idx;
1030 elem->prev_insn_idx = prev_insn_idx;
1031 elem->next = env->head;
1032 elem->log_pos = env->log.len_used;
1035 err = copy_verifier_state(&elem->st, cur);
1038 elem->st.speculative |= speculative;
1039 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
1040 verbose(env, "The sequence of %d jumps is too complex.\n",
1044 if (elem->st.parent) {
1045 ++elem->st.parent->branches;
1046 /* WARN_ON(branches > 2) technically makes sense here,
1048 * 1. speculative states will bump 'branches' for non-branch
1050 * 2. is_state_visited() heuristics may decide not to create
1051 * a new state for a sequence of branches and all such current
1052 * and cloned states will be pointing to a single parent state
1053 * which might have large 'branches' count.
1058 free_verifier_state(env->cur_state, true);
1059 env->cur_state = NULL;
1060 /* pop all elements and return */
1061 while (!pop_stack(env, NULL, NULL, false));
1065 #define CALLER_SAVED_REGS 6
1066 static const int caller_saved[CALLER_SAVED_REGS] = {
1067 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
1070 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1071 struct bpf_reg_state *reg);
1073 /* This helper doesn't clear reg->id */
1074 static void ___mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1076 reg->var_off = tnum_const(imm);
1077 reg->smin_value = (s64)imm;
1078 reg->smax_value = (s64)imm;
1079 reg->umin_value = imm;
1080 reg->umax_value = imm;
1082 reg->s32_min_value = (s32)imm;
1083 reg->s32_max_value = (s32)imm;
1084 reg->u32_min_value = (u32)imm;
1085 reg->u32_max_value = (u32)imm;
1088 /* Mark the unknown part of a register (variable offset or scalar value) as
1089 * known to have the value @imm.
1091 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1093 /* Clear id, off, and union(map_ptr, range) */
1094 memset(((u8 *)reg) + sizeof(reg->type), 0,
1095 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
1096 ___mark_reg_known(reg, imm);
1099 static void __mark_reg32_known(struct bpf_reg_state *reg, u64 imm)
1101 reg->var_off = tnum_const_subreg(reg->var_off, imm);
1102 reg->s32_min_value = (s32)imm;
1103 reg->s32_max_value = (s32)imm;
1104 reg->u32_min_value = (u32)imm;
1105 reg->u32_max_value = (u32)imm;
1108 /* Mark the 'variable offset' part of a register as zero. This should be
1109 * used only on registers holding a pointer type.
1111 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
1113 __mark_reg_known(reg, 0);
1116 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
1118 __mark_reg_known(reg, 0);
1119 reg->type = SCALAR_VALUE;
1122 static void mark_reg_known_zero(struct bpf_verifier_env *env,
1123 struct bpf_reg_state *regs, u32 regno)
1125 if (WARN_ON(regno >= MAX_BPF_REG)) {
1126 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
1127 /* Something bad happened, let's kill all regs */
1128 for (regno = 0; regno < MAX_BPF_REG; regno++)
1129 __mark_reg_not_init(env, regs + regno);
1132 __mark_reg_known_zero(regs + regno);
1135 static void mark_ptr_not_null_reg(struct bpf_reg_state *reg)
1137 if (base_type(reg->type) == PTR_TO_MAP_VALUE) {
1138 const struct bpf_map *map = reg->map_ptr;
1140 if (map->inner_map_meta) {
1141 reg->type = CONST_PTR_TO_MAP;
1142 reg->map_ptr = map->inner_map_meta;
1143 /* transfer reg's id which is unique for every map_lookup_elem
1144 * as UID of the inner map.
1146 if (map_value_has_timer(map->inner_map_meta))
1147 reg->map_uid = reg->id;
1148 } else if (map->map_type == BPF_MAP_TYPE_XSKMAP) {
1149 reg->type = PTR_TO_XDP_SOCK;
1150 } else if (map->map_type == BPF_MAP_TYPE_SOCKMAP ||
1151 map->map_type == BPF_MAP_TYPE_SOCKHASH) {
1152 reg->type = PTR_TO_SOCKET;
1154 reg->type = PTR_TO_MAP_VALUE;
1159 reg->type &= ~PTR_MAYBE_NULL;
1162 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
1164 return type_is_pkt_pointer(reg->type);
1167 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
1169 return reg_is_pkt_pointer(reg) ||
1170 reg->type == PTR_TO_PACKET_END;
1173 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
1174 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
1175 enum bpf_reg_type which)
1177 /* The register can already have a range from prior markings.
1178 * This is fine as long as it hasn't been advanced from its
1181 return reg->type == which &&
1184 tnum_equals_const(reg->var_off, 0);
1187 /* Reset the min/max bounds of a register */
1188 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
1190 reg->smin_value = S64_MIN;
1191 reg->smax_value = S64_MAX;
1192 reg->umin_value = 0;
1193 reg->umax_value = U64_MAX;
1195 reg->s32_min_value = S32_MIN;
1196 reg->s32_max_value = S32_MAX;
1197 reg->u32_min_value = 0;
1198 reg->u32_max_value = U32_MAX;
1201 static void __mark_reg64_unbounded(struct bpf_reg_state *reg)
1203 reg->smin_value = S64_MIN;
1204 reg->smax_value = S64_MAX;
1205 reg->umin_value = 0;
1206 reg->umax_value = U64_MAX;
1209 static void __mark_reg32_unbounded(struct bpf_reg_state *reg)
1211 reg->s32_min_value = S32_MIN;
1212 reg->s32_max_value = S32_MAX;
1213 reg->u32_min_value = 0;
1214 reg->u32_max_value = U32_MAX;
1217 static void __update_reg32_bounds(struct bpf_reg_state *reg)
1219 struct tnum var32_off = tnum_subreg(reg->var_off);
1221 /* min signed is max(sign bit) | min(other bits) */
1222 reg->s32_min_value = max_t(s32, reg->s32_min_value,
1223 var32_off.value | (var32_off.mask & S32_MIN));
1224 /* max signed is min(sign bit) | max(other bits) */
1225 reg->s32_max_value = min_t(s32, reg->s32_max_value,
1226 var32_off.value | (var32_off.mask & S32_MAX));
1227 reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)var32_off.value);
1228 reg->u32_max_value = min(reg->u32_max_value,
1229 (u32)(var32_off.value | var32_off.mask));
1232 static void __update_reg64_bounds(struct bpf_reg_state *reg)
1234 /* min signed is max(sign bit) | min(other bits) */
1235 reg->smin_value = max_t(s64, reg->smin_value,
1236 reg->var_off.value | (reg->var_off.mask & S64_MIN));
1237 /* max signed is min(sign bit) | max(other bits) */
1238 reg->smax_value = min_t(s64, reg->smax_value,
1239 reg->var_off.value | (reg->var_off.mask & S64_MAX));
1240 reg->umin_value = max(reg->umin_value, reg->var_off.value);
1241 reg->umax_value = min(reg->umax_value,
1242 reg->var_off.value | reg->var_off.mask);
1245 static void __update_reg_bounds(struct bpf_reg_state *reg)
1247 __update_reg32_bounds(reg);
1248 __update_reg64_bounds(reg);
1251 /* Uses signed min/max values to inform unsigned, and vice-versa */
1252 static void __reg32_deduce_bounds(struct bpf_reg_state *reg)
1254 /* Learn sign from signed bounds.
1255 * If we cannot cross the sign boundary, then signed and unsigned bounds
1256 * are the same, so combine. This works even in the negative case, e.g.
1257 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1259 if (reg->s32_min_value >= 0 || reg->s32_max_value < 0) {
1260 reg->s32_min_value = reg->u32_min_value =
1261 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1262 reg->s32_max_value = reg->u32_max_value =
1263 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1266 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1267 * boundary, so we must be careful.
1269 if ((s32)reg->u32_max_value >= 0) {
1270 /* Positive. We can't learn anything from the smin, but smax
1271 * is positive, hence safe.
1273 reg->s32_min_value = reg->u32_min_value;
1274 reg->s32_max_value = reg->u32_max_value =
1275 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1276 } else if ((s32)reg->u32_min_value < 0) {
1277 /* Negative. We can't learn anything from the smax, but smin
1278 * is negative, hence safe.
1280 reg->s32_min_value = reg->u32_min_value =
1281 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1282 reg->s32_max_value = reg->u32_max_value;
1286 static void __reg64_deduce_bounds(struct bpf_reg_state *reg)
1288 /* Learn sign from signed bounds.
1289 * If we cannot cross the sign boundary, then signed and unsigned bounds
1290 * are the same, so combine. This works even in the negative case, e.g.
1291 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1293 if (reg->smin_value >= 0 || reg->smax_value < 0) {
1294 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1296 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1300 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1301 * boundary, so we must be careful.
1303 if ((s64)reg->umax_value >= 0) {
1304 /* Positive. We can't learn anything from the smin, but smax
1305 * is positive, hence safe.
1307 reg->smin_value = reg->umin_value;
1308 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1310 } else if ((s64)reg->umin_value < 0) {
1311 /* Negative. We can't learn anything from the smax, but smin
1312 * is negative, hence safe.
1314 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1316 reg->smax_value = reg->umax_value;
1320 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
1322 __reg32_deduce_bounds(reg);
1323 __reg64_deduce_bounds(reg);
1326 /* Attempts to improve var_off based on unsigned min/max information */
1327 static void __reg_bound_offset(struct bpf_reg_state *reg)
1329 struct tnum var64_off = tnum_intersect(reg->var_off,
1330 tnum_range(reg->umin_value,
1332 struct tnum var32_off = tnum_intersect(tnum_subreg(reg->var_off),
1333 tnum_range(reg->u32_min_value,
1334 reg->u32_max_value));
1336 reg->var_off = tnum_or(tnum_clear_subreg(var64_off), var32_off);
1339 static void reg_bounds_sync(struct bpf_reg_state *reg)
1341 /* We might have learned new bounds from the var_off. */
1342 __update_reg_bounds(reg);
1343 /* We might have learned something about the sign bit. */
1344 __reg_deduce_bounds(reg);
1345 /* We might have learned some bits from the bounds. */
1346 __reg_bound_offset(reg);
1347 /* Intersecting with the old var_off might have improved our bounds
1348 * slightly, e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1349 * then new var_off is (0; 0x7f...fc) which improves our umax.
1351 __update_reg_bounds(reg);
1354 static bool __reg32_bound_s64(s32 a)
1356 return a >= 0 && a <= S32_MAX;
1359 static void __reg_assign_32_into_64(struct bpf_reg_state *reg)
1361 reg->umin_value = reg->u32_min_value;
1362 reg->umax_value = reg->u32_max_value;
1364 /* Attempt to pull 32-bit signed bounds into 64-bit bounds but must
1365 * be positive otherwise set to worse case bounds and refine later
1368 if (__reg32_bound_s64(reg->s32_min_value) &&
1369 __reg32_bound_s64(reg->s32_max_value)) {
1370 reg->smin_value = reg->s32_min_value;
1371 reg->smax_value = reg->s32_max_value;
1373 reg->smin_value = 0;
1374 reg->smax_value = U32_MAX;
1378 static void __reg_combine_32_into_64(struct bpf_reg_state *reg)
1380 /* special case when 64-bit register has upper 32-bit register
1381 * zeroed. Typically happens after zext or <<32, >>32 sequence
1382 * allowing us to use 32-bit bounds directly,
1384 if (tnum_equals_const(tnum_clear_subreg(reg->var_off), 0)) {
1385 __reg_assign_32_into_64(reg);
1387 /* Otherwise the best we can do is push lower 32bit known and
1388 * unknown bits into register (var_off set from jmp logic)
1389 * then learn as much as possible from the 64-bit tnum
1390 * known and unknown bits. The previous smin/smax bounds are
1391 * invalid here because of jmp32 compare so mark them unknown
1392 * so they do not impact tnum bounds calculation.
1394 __mark_reg64_unbounded(reg);
1396 reg_bounds_sync(reg);
1399 static bool __reg64_bound_s32(s64 a)
1401 return a >= S32_MIN && a <= S32_MAX;
1404 static bool __reg64_bound_u32(u64 a)
1406 return a >= U32_MIN && a <= U32_MAX;
1409 static void __reg_combine_64_into_32(struct bpf_reg_state *reg)
1411 __mark_reg32_unbounded(reg);
1412 if (__reg64_bound_s32(reg->smin_value) && __reg64_bound_s32(reg->smax_value)) {
1413 reg->s32_min_value = (s32)reg->smin_value;
1414 reg->s32_max_value = (s32)reg->smax_value;
1416 if (__reg64_bound_u32(reg->umin_value) && __reg64_bound_u32(reg->umax_value)) {
1417 reg->u32_min_value = (u32)reg->umin_value;
1418 reg->u32_max_value = (u32)reg->umax_value;
1420 reg_bounds_sync(reg);
1423 /* Mark a register as having a completely unknown (scalar) value. */
1424 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
1425 struct bpf_reg_state *reg)
1428 * Clear type, id, off, and union(map_ptr, range) and
1429 * padding between 'type' and union
1431 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
1432 reg->type = SCALAR_VALUE;
1433 reg->var_off = tnum_unknown;
1435 reg->precise = env->subprog_cnt > 1 || !env->bpf_capable;
1436 __mark_reg_unbounded(reg);
1439 static void mark_reg_unknown(struct bpf_verifier_env *env,
1440 struct bpf_reg_state *regs, u32 regno)
1442 if (WARN_ON(regno >= MAX_BPF_REG)) {
1443 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
1444 /* Something bad happened, let's kill all regs except FP */
1445 for (regno = 0; regno < BPF_REG_FP; regno++)
1446 __mark_reg_not_init(env, regs + regno);
1449 __mark_reg_unknown(env, regs + regno);
1452 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1453 struct bpf_reg_state *reg)
1455 __mark_reg_unknown(env, reg);
1456 reg->type = NOT_INIT;
1459 static void mark_reg_not_init(struct bpf_verifier_env *env,
1460 struct bpf_reg_state *regs, u32 regno)
1462 if (WARN_ON(regno >= MAX_BPF_REG)) {
1463 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
1464 /* Something bad happened, let's kill all regs except FP */
1465 for (regno = 0; regno < BPF_REG_FP; regno++)
1466 __mark_reg_not_init(env, regs + regno);
1469 __mark_reg_not_init(env, regs + regno);
1472 static void mark_btf_ld_reg(struct bpf_verifier_env *env,
1473 struct bpf_reg_state *regs, u32 regno,
1474 enum bpf_reg_type reg_type,
1475 struct btf *btf, u32 btf_id)
1477 if (reg_type == SCALAR_VALUE) {
1478 mark_reg_unknown(env, regs, regno);
1481 mark_reg_known_zero(env, regs, regno);
1482 regs[regno].type = PTR_TO_BTF_ID;
1483 regs[regno].btf = btf;
1484 regs[regno].btf_id = btf_id;
1487 #define DEF_NOT_SUBREG (0)
1488 static void init_reg_state(struct bpf_verifier_env *env,
1489 struct bpf_func_state *state)
1491 struct bpf_reg_state *regs = state->regs;
1494 for (i = 0; i < MAX_BPF_REG; i++) {
1495 mark_reg_not_init(env, regs, i);
1496 regs[i].live = REG_LIVE_NONE;
1497 regs[i].parent = NULL;
1498 regs[i].subreg_def = DEF_NOT_SUBREG;
1502 regs[BPF_REG_FP].type = PTR_TO_STACK;
1503 mark_reg_known_zero(env, regs, BPF_REG_FP);
1504 regs[BPF_REG_FP].frameno = state->frameno;
1507 #define BPF_MAIN_FUNC (-1)
1508 static void init_func_state(struct bpf_verifier_env *env,
1509 struct bpf_func_state *state,
1510 int callsite, int frameno, int subprogno)
1512 state->callsite = callsite;
1513 state->frameno = frameno;
1514 state->subprogno = subprogno;
1515 init_reg_state(env, state);
1518 /* Similar to push_stack(), but for async callbacks */
1519 static struct bpf_verifier_state *push_async_cb(struct bpf_verifier_env *env,
1520 int insn_idx, int prev_insn_idx,
1523 struct bpf_verifier_stack_elem *elem;
1524 struct bpf_func_state *frame;
1526 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
1530 elem->insn_idx = insn_idx;
1531 elem->prev_insn_idx = prev_insn_idx;
1532 elem->next = env->head;
1533 elem->log_pos = env->log.len_used;
1536 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
1538 "The sequence of %d jumps is too complex for async cb.\n",
1542 /* Unlike push_stack() do not copy_verifier_state().
1543 * The caller state doesn't matter.
1544 * This is async callback. It starts in a fresh stack.
1545 * Initialize it similar to do_check_common().
1547 elem->st.branches = 1;
1548 frame = kzalloc(sizeof(*frame), GFP_KERNEL);
1551 init_func_state(env, frame,
1552 BPF_MAIN_FUNC /* callsite */,
1553 0 /* frameno within this callchain */,
1554 subprog /* subprog number within this prog */);
1555 elem->st.frame[0] = frame;
1558 free_verifier_state(env->cur_state, true);
1559 env->cur_state = NULL;
1560 /* pop all elements and return */
1561 while (!pop_stack(env, NULL, NULL, false));
1567 SRC_OP, /* register is used as source operand */
1568 DST_OP, /* register is used as destination operand */
1569 DST_OP_NO_MARK /* same as above, check only, don't mark */
1572 static int cmp_subprogs(const void *a, const void *b)
1574 return ((struct bpf_subprog_info *)a)->start -
1575 ((struct bpf_subprog_info *)b)->start;
1578 static int find_subprog(struct bpf_verifier_env *env, int off)
1580 struct bpf_subprog_info *p;
1582 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
1583 sizeof(env->subprog_info[0]), cmp_subprogs);
1586 return p - env->subprog_info;
1590 static int add_subprog(struct bpf_verifier_env *env, int off)
1592 int insn_cnt = env->prog->len;
1595 if (off >= insn_cnt || off < 0) {
1596 verbose(env, "call to invalid destination\n");
1599 ret = find_subprog(env, off);
1602 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
1603 verbose(env, "too many subprograms\n");
1606 /* determine subprog starts. The end is one before the next starts */
1607 env->subprog_info[env->subprog_cnt++].start = off;
1608 sort(env->subprog_info, env->subprog_cnt,
1609 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
1610 return env->subprog_cnt - 1;
1613 struct bpf_kfunc_desc {
1614 struct btf_func_model func_model;
1619 #define MAX_KFUNC_DESCS 256
1620 struct bpf_kfunc_desc_tab {
1621 struct bpf_kfunc_desc descs[MAX_KFUNC_DESCS];
1625 static int kfunc_desc_cmp_by_id(const void *a, const void *b)
1627 const struct bpf_kfunc_desc *d0 = a;
1628 const struct bpf_kfunc_desc *d1 = b;
1630 /* func_id is not greater than BTF_MAX_TYPE */
1631 return d0->func_id - d1->func_id;
1634 static const struct bpf_kfunc_desc *
1635 find_kfunc_desc(const struct bpf_prog *prog, u32 func_id)
1637 struct bpf_kfunc_desc desc = {
1640 struct bpf_kfunc_desc_tab *tab;
1642 tab = prog->aux->kfunc_tab;
1643 return bsearch(&desc, tab->descs, tab->nr_descs,
1644 sizeof(tab->descs[0]), kfunc_desc_cmp_by_id);
1647 static int add_kfunc_call(struct bpf_verifier_env *env, u32 func_id)
1649 const struct btf_type *func, *func_proto;
1650 struct bpf_kfunc_desc_tab *tab;
1651 struct bpf_prog_aux *prog_aux;
1652 struct bpf_kfunc_desc *desc;
1653 const char *func_name;
1657 prog_aux = env->prog->aux;
1658 tab = prog_aux->kfunc_tab;
1661 verbose(env, "calling kernel function is not supported without CONFIG_DEBUG_INFO_BTF\n");
1665 if (!env->prog->jit_requested) {
1666 verbose(env, "JIT is required for calling kernel function\n");
1670 if (!bpf_jit_supports_kfunc_call()) {
1671 verbose(env, "JIT does not support calling kernel function\n");
1675 if (!env->prog->gpl_compatible) {
1676 verbose(env, "cannot call kernel function from non-GPL compatible program\n");
1680 tab = kzalloc(sizeof(*tab), GFP_KERNEL);
1683 prog_aux->kfunc_tab = tab;
1686 if (find_kfunc_desc(env->prog, func_id))
1689 if (tab->nr_descs == MAX_KFUNC_DESCS) {
1690 verbose(env, "too many different kernel function calls\n");
1694 func = btf_type_by_id(btf_vmlinux, func_id);
1695 if (!func || !btf_type_is_func(func)) {
1696 verbose(env, "kernel btf_id %u is not a function\n",
1700 func_proto = btf_type_by_id(btf_vmlinux, func->type);
1701 if (!func_proto || !btf_type_is_func_proto(func_proto)) {
1702 verbose(env, "kernel function btf_id %u does not have a valid func_proto\n",
1707 func_name = btf_name_by_offset(btf_vmlinux, func->name_off);
1708 addr = kallsyms_lookup_name(func_name);
1710 verbose(env, "cannot find address for kernel function %s\n",
1715 desc = &tab->descs[tab->nr_descs++];
1716 desc->func_id = func_id;
1717 desc->imm = BPF_CAST_CALL(addr) - __bpf_call_base;
1718 err = btf_distill_func_proto(&env->log, btf_vmlinux,
1719 func_proto, func_name,
1722 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
1723 kfunc_desc_cmp_by_id, NULL);
1727 static int kfunc_desc_cmp_by_imm(const void *a, const void *b)
1729 const struct bpf_kfunc_desc *d0 = a;
1730 const struct bpf_kfunc_desc *d1 = b;
1732 if (d0->imm > d1->imm)
1734 else if (d0->imm < d1->imm)
1739 static void sort_kfunc_descs_by_imm(struct bpf_prog *prog)
1741 struct bpf_kfunc_desc_tab *tab;
1743 tab = prog->aux->kfunc_tab;
1747 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
1748 kfunc_desc_cmp_by_imm, NULL);
1751 bool bpf_prog_has_kfunc_call(const struct bpf_prog *prog)
1753 return !!prog->aux->kfunc_tab;
1756 const struct btf_func_model *
1757 bpf_jit_find_kfunc_model(const struct bpf_prog *prog,
1758 const struct bpf_insn *insn)
1760 const struct bpf_kfunc_desc desc = {
1763 const struct bpf_kfunc_desc *res;
1764 struct bpf_kfunc_desc_tab *tab;
1766 tab = prog->aux->kfunc_tab;
1767 res = bsearch(&desc, tab->descs, tab->nr_descs,
1768 sizeof(tab->descs[0]), kfunc_desc_cmp_by_imm);
1770 return res ? &res->func_model : NULL;
1773 static int add_subprog_and_kfunc(struct bpf_verifier_env *env)
1775 struct bpf_subprog_info *subprog = env->subprog_info;
1776 struct bpf_insn *insn = env->prog->insnsi;
1777 int i, ret, insn_cnt = env->prog->len;
1779 /* Add entry function. */
1780 ret = add_subprog(env, 0);
1784 for (i = 0; i < insn_cnt; i++, insn++) {
1785 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn) &&
1786 !bpf_pseudo_kfunc_call(insn))
1789 if (!env->bpf_capable) {
1790 verbose(env, "loading/calling other bpf or kernel functions are allowed for CAP_BPF and CAP_SYS_ADMIN\n");
1794 if (bpf_pseudo_func(insn) || bpf_pseudo_call(insn))
1795 ret = add_subprog(env, i + insn->imm + 1);
1797 ret = add_kfunc_call(env, insn->imm);
1803 /* Add a fake 'exit' subprog which could simplify subprog iteration
1804 * logic. 'subprog_cnt' should not be increased.
1806 subprog[env->subprog_cnt].start = insn_cnt;
1808 if (env->log.level & BPF_LOG_LEVEL2)
1809 for (i = 0; i < env->subprog_cnt; i++)
1810 verbose(env, "func#%d @%d\n", i, subprog[i].start);
1815 static int check_subprogs(struct bpf_verifier_env *env)
1817 int i, subprog_start, subprog_end, off, cur_subprog = 0;
1818 struct bpf_subprog_info *subprog = env->subprog_info;
1819 struct bpf_insn *insn = env->prog->insnsi;
1820 int insn_cnt = env->prog->len;
1822 /* now check that all jumps are within the same subprog */
1823 subprog_start = subprog[cur_subprog].start;
1824 subprog_end = subprog[cur_subprog + 1].start;
1825 for (i = 0; i < insn_cnt; i++) {
1826 u8 code = insn[i].code;
1828 if (code == (BPF_JMP | BPF_CALL) &&
1829 insn[i].imm == BPF_FUNC_tail_call &&
1830 insn[i].src_reg != BPF_PSEUDO_CALL)
1831 subprog[cur_subprog].has_tail_call = true;
1832 if (BPF_CLASS(code) == BPF_LD &&
1833 (BPF_MODE(code) == BPF_ABS || BPF_MODE(code) == BPF_IND))
1834 subprog[cur_subprog].has_ld_abs = true;
1835 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
1837 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
1839 off = i + insn[i].off + 1;
1840 if (off < subprog_start || off >= subprog_end) {
1841 verbose(env, "jump out of range from insn %d to %d\n", i, off);
1845 if (i == subprog_end - 1) {
1846 /* to avoid fall-through from one subprog into another
1847 * the last insn of the subprog should be either exit
1848 * or unconditional jump back
1850 if (code != (BPF_JMP | BPF_EXIT) &&
1851 code != (BPF_JMP | BPF_JA)) {
1852 verbose(env, "last insn is not an exit or jmp\n");
1855 subprog_start = subprog_end;
1857 if (cur_subprog < env->subprog_cnt)
1858 subprog_end = subprog[cur_subprog + 1].start;
1864 /* Parentage chain of this register (or stack slot) should take care of all
1865 * issues like callee-saved registers, stack slot allocation time, etc.
1867 static int mark_reg_read(struct bpf_verifier_env *env,
1868 const struct bpf_reg_state *state,
1869 struct bpf_reg_state *parent, u8 flag)
1871 bool writes = parent == state->parent; /* Observe write marks */
1875 /* if read wasn't screened by an earlier write ... */
1876 if (writes && state->live & REG_LIVE_WRITTEN)
1878 if (parent->live & REG_LIVE_DONE) {
1879 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
1880 reg_type_str(env, parent->type),
1881 parent->var_off.value, parent->off);
1884 /* The first condition is more likely to be true than the
1885 * second, checked it first.
1887 if ((parent->live & REG_LIVE_READ) == flag ||
1888 parent->live & REG_LIVE_READ64)
1889 /* The parentage chain never changes and
1890 * this parent was already marked as LIVE_READ.
1891 * There is no need to keep walking the chain again and
1892 * keep re-marking all parents as LIVE_READ.
1893 * This case happens when the same register is read
1894 * multiple times without writes into it in-between.
1895 * Also, if parent has the stronger REG_LIVE_READ64 set,
1896 * then no need to set the weak REG_LIVE_READ32.
1899 /* ... then we depend on parent's value */
1900 parent->live |= flag;
1901 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
1902 if (flag == REG_LIVE_READ64)
1903 parent->live &= ~REG_LIVE_READ32;
1905 parent = state->parent;
1910 if (env->longest_mark_read_walk < cnt)
1911 env->longest_mark_read_walk = cnt;
1915 /* This function is supposed to be used by the following 32-bit optimization
1916 * code only. It returns TRUE if the source or destination register operates
1917 * on 64-bit, otherwise return FALSE.
1919 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
1920 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
1925 class = BPF_CLASS(code);
1927 if (class == BPF_JMP) {
1928 /* BPF_EXIT for "main" will reach here. Return TRUE
1933 if (op == BPF_CALL) {
1934 /* BPF to BPF call will reach here because of marking
1935 * caller saved clobber with DST_OP_NO_MARK for which we
1936 * don't care the register def because they are anyway
1937 * marked as NOT_INIT already.
1939 if (insn->src_reg == BPF_PSEUDO_CALL)
1941 /* Helper call will reach here because of arg type
1942 * check, conservatively return TRUE.
1951 if (class == BPF_ALU64 || class == BPF_JMP ||
1952 /* BPF_END always use BPF_ALU class. */
1953 (class == BPF_ALU && op == BPF_END && insn->imm == 64))
1956 if (class == BPF_ALU || class == BPF_JMP32)
1959 if (class == BPF_LDX) {
1961 return BPF_SIZE(code) == BPF_DW;
1962 /* LDX source must be ptr. */
1966 if (class == BPF_STX) {
1967 /* BPF_STX (including atomic variants) has multiple source
1968 * operands, one of which is a ptr. Check whether the caller is
1971 if (t == SRC_OP && reg->type != SCALAR_VALUE)
1973 return BPF_SIZE(code) == BPF_DW;
1976 if (class == BPF_LD) {
1977 u8 mode = BPF_MODE(code);
1980 if (mode == BPF_IMM)
1983 /* Both LD_IND and LD_ABS return 32-bit data. */
1987 /* Implicit ctx ptr. */
1988 if (regno == BPF_REG_6)
1991 /* Explicit source could be any width. */
1995 if (class == BPF_ST)
1996 /* The only source register for BPF_ST is a ptr. */
1999 /* Conservatively return true at default. */
2003 /* Return the regno defined by the insn, or -1. */
2004 static int insn_def_regno(const struct bpf_insn *insn)
2006 switch (BPF_CLASS(insn->code)) {
2012 if (BPF_MODE(insn->code) == BPF_ATOMIC &&
2013 (insn->imm & BPF_FETCH)) {
2014 if (insn->imm == BPF_CMPXCHG)
2017 return insn->src_reg;
2022 return insn->dst_reg;
2026 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
2027 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
2029 int dst_reg = insn_def_regno(insn);
2034 return !is_reg64(env, insn, dst_reg, NULL, DST_OP);
2037 static void mark_insn_zext(struct bpf_verifier_env *env,
2038 struct bpf_reg_state *reg)
2040 s32 def_idx = reg->subreg_def;
2042 if (def_idx == DEF_NOT_SUBREG)
2045 env->insn_aux_data[def_idx - 1].zext_dst = true;
2046 /* The dst will be zero extended, so won't be sub-register anymore. */
2047 reg->subreg_def = DEF_NOT_SUBREG;
2050 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
2051 enum reg_arg_type t)
2053 struct bpf_verifier_state *vstate = env->cur_state;
2054 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2055 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
2056 struct bpf_reg_state *reg, *regs = state->regs;
2059 if (regno >= MAX_BPF_REG) {
2060 verbose(env, "R%d is invalid\n", regno);
2065 rw64 = is_reg64(env, insn, regno, reg, t);
2067 /* check whether register used as source operand can be read */
2068 if (reg->type == NOT_INIT) {
2069 verbose(env, "R%d !read_ok\n", regno);
2072 /* We don't need to worry about FP liveness because it's read-only */
2073 if (regno == BPF_REG_FP)
2077 mark_insn_zext(env, reg);
2079 return mark_reg_read(env, reg, reg->parent,
2080 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
2082 /* check whether register used as dest operand can be written to */
2083 if (regno == BPF_REG_FP) {
2084 verbose(env, "frame pointer is read only\n");
2087 reg->live |= REG_LIVE_WRITTEN;
2088 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
2090 mark_reg_unknown(env, regs, regno);
2095 /* for any branch, call, exit record the history of jmps in the given state */
2096 static int push_jmp_history(struct bpf_verifier_env *env,
2097 struct bpf_verifier_state *cur)
2099 u32 cnt = cur->jmp_history_cnt;
2100 struct bpf_idx_pair *p;
2103 p = krealloc(cur->jmp_history, cnt * sizeof(*p), GFP_USER);
2106 p[cnt - 1].idx = env->insn_idx;
2107 p[cnt - 1].prev_idx = env->prev_insn_idx;
2108 cur->jmp_history = p;
2109 cur->jmp_history_cnt = cnt;
2113 /* Backtrack one insn at a time. If idx is not at the top of recorded
2114 * history then previous instruction came from straight line execution.
2116 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
2121 if (cnt && st->jmp_history[cnt - 1].idx == i) {
2122 i = st->jmp_history[cnt - 1].prev_idx;
2130 static const char *disasm_kfunc_name(void *data, const struct bpf_insn *insn)
2132 const struct btf_type *func;
2134 if (insn->src_reg != BPF_PSEUDO_KFUNC_CALL)
2137 func = btf_type_by_id(btf_vmlinux, insn->imm);
2138 return btf_name_by_offset(btf_vmlinux, func->name_off);
2141 /* For given verifier state backtrack_insn() is called from the last insn to
2142 * the first insn. Its purpose is to compute a bitmask of registers and
2143 * stack slots that needs precision in the parent verifier state.
2145 static int backtrack_insn(struct bpf_verifier_env *env, int idx,
2146 u32 *reg_mask, u64 *stack_mask)
2148 const struct bpf_insn_cbs cbs = {
2149 .cb_call = disasm_kfunc_name,
2150 .cb_print = verbose,
2151 .private_data = env,
2153 struct bpf_insn *insn = env->prog->insnsi + idx;
2154 u8 class = BPF_CLASS(insn->code);
2155 u8 opcode = BPF_OP(insn->code);
2156 u8 mode = BPF_MODE(insn->code);
2157 u32 dreg = 1u << insn->dst_reg;
2158 u32 sreg = 1u << insn->src_reg;
2161 if (insn->code == 0)
2163 if (env->log.level & BPF_LOG_LEVEL) {
2164 verbose(env, "regs=%x stack=%llx before ", *reg_mask, *stack_mask);
2165 verbose(env, "%d: ", idx);
2166 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
2169 if (class == BPF_ALU || class == BPF_ALU64) {
2170 if (!(*reg_mask & dreg))
2172 if (opcode == BPF_MOV) {
2173 if (BPF_SRC(insn->code) == BPF_X) {
2175 * dreg needs precision after this insn
2176 * sreg needs precision before this insn
2182 * dreg needs precision after this insn.
2183 * Corresponding register is already marked
2184 * as precise=true in this verifier state.
2185 * No further markings in parent are necessary
2190 if (BPF_SRC(insn->code) == BPF_X) {
2192 * both dreg and sreg need precision
2197 * dreg still needs precision before this insn
2200 } else if (class == BPF_LDX) {
2201 if (!(*reg_mask & dreg))
2205 /* scalars can only be spilled into stack w/o losing precision.
2206 * Load from any other memory can be zero extended.
2207 * The desire to keep that precision is already indicated
2208 * by 'precise' mark in corresponding register of this state.
2209 * No further tracking necessary.
2211 if (insn->src_reg != BPF_REG_FP)
2213 if (BPF_SIZE(insn->code) != BPF_DW)
2216 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
2217 * that [fp - off] slot contains scalar that needs to be
2218 * tracked with precision
2220 spi = (-insn->off - 1) / BPF_REG_SIZE;
2222 verbose(env, "BUG spi %d\n", spi);
2223 WARN_ONCE(1, "verifier backtracking bug");
2226 *stack_mask |= 1ull << spi;
2227 } else if (class == BPF_STX || class == BPF_ST) {
2228 if (*reg_mask & dreg)
2229 /* stx & st shouldn't be using _scalar_ dst_reg
2230 * to access memory. It means backtracking
2231 * encountered a case of pointer subtraction.
2234 /* scalars can only be spilled into stack */
2235 if (insn->dst_reg != BPF_REG_FP)
2237 if (BPF_SIZE(insn->code) != BPF_DW)
2239 spi = (-insn->off - 1) / BPF_REG_SIZE;
2241 verbose(env, "BUG spi %d\n", spi);
2242 WARN_ONCE(1, "verifier backtracking bug");
2245 if (!(*stack_mask & (1ull << spi)))
2247 *stack_mask &= ~(1ull << spi);
2248 if (class == BPF_STX)
2250 } else if (class == BPF_JMP || class == BPF_JMP32) {
2251 if (opcode == BPF_CALL) {
2252 if (insn->src_reg == BPF_PSEUDO_CALL)
2254 /* regular helper call sets R0 */
2256 if (*reg_mask & 0x3f) {
2257 /* if backtracing was looking for registers R1-R5
2258 * they should have been found already.
2260 verbose(env, "BUG regs %x\n", *reg_mask);
2261 WARN_ONCE(1, "verifier backtracking bug");
2264 } else if (opcode == BPF_EXIT) {
2267 } else if (class == BPF_LD) {
2268 if (!(*reg_mask & dreg))
2271 /* It's ld_imm64 or ld_abs or ld_ind.
2272 * For ld_imm64 no further tracking of precision
2273 * into parent is necessary
2275 if (mode == BPF_IND || mode == BPF_ABS)
2276 /* to be analyzed */
2282 /* the scalar precision tracking algorithm:
2283 * . at the start all registers have precise=false.
2284 * . scalar ranges are tracked as normal through alu and jmp insns.
2285 * . once precise value of the scalar register is used in:
2286 * . ptr + scalar alu
2287 * . if (scalar cond K|scalar)
2288 * . helper_call(.., scalar, ...) where ARG_CONST is expected
2289 * backtrack through the verifier states and mark all registers and
2290 * stack slots with spilled constants that these scalar regisers
2291 * should be precise.
2292 * . during state pruning two registers (or spilled stack slots)
2293 * are equivalent if both are not precise.
2295 * Note the verifier cannot simply walk register parentage chain,
2296 * since many different registers and stack slots could have been
2297 * used to compute single precise scalar.
2299 * The approach of starting with precise=true for all registers and then
2300 * backtrack to mark a register as not precise when the verifier detects
2301 * that program doesn't care about specific value (e.g., when helper
2302 * takes register as ARG_ANYTHING parameter) is not safe.
2304 * It's ok to walk single parentage chain of the verifier states.
2305 * It's possible that this backtracking will go all the way till 1st insn.
2306 * All other branches will be explored for needing precision later.
2308 * The backtracking needs to deal with cases like:
2309 * 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)
2312 * if r5 > 0x79f goto pc+7
2313 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
2316 * call bpf_perf_event_output#25
2317 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
2321 * call foo // uses callee's r6 inside to compute r0
2325 * to track above reg_mask/stack_mask needs to be independent for each frame.
2327 * Also if parent's curframe > frame where backtracking started,
2328 * the verifier need to mark registers in both frames, otherwise callees
2329 * may incorrectly prune callers. This is similar to
2330 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
2332 * For now backtracking falls back into conservative marking.
2334 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
2335 struct bpf_verifier_state *st)
2337 struct bpf_func_state *func;
2338 struct bpf_reg_state *reg;
2341 /* big hammer: mark all scalars precise in this path.
2342 * pop_stack may still get !precise scalars.
2344 for (; st; st = st->parent)
2345 for (i = 0; i <= st->curframe; i++) {
2346 func = st->frame[i];
2347 for (j = 0; j < BPF_REG_FP; j++) {
2348 reg = &func->regs[j];
2349 if (reg->type != SCALAR_VALUE)
2351 reg->precise = true;
2353 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
2354 if (func->stack[j].slot_type[0] != STACK_SPILL)
2356 reg = &func->stack[j].spilled_ptr;
2357 if (reg->type != SCALAR_VALUE)
2359 reg->precise = true;
2364 static int __mark_chain_precision(struct bpf_verifier_env *env, int regno,
2367 struct bpf_verifier_state *st = env->cur_state;
2368 int first_idx = st->first_insn_idx;
2369 int last_idx = env->insn_idx;
2370 struct bpf_func_state *func;
2371 struct bpf_reg_state *reg;
2372 u32 reg_mask = regno >= 0 ? 1u << regno : 0;
2373 u64 stack_mask = spi >= 0 ? 1ull << spi : 0;
2374 bool skip_first = true;
2375 bool new_marks = false;
2378 if (!env->bpf_capable)
2381 func = st->frame[st->curframe];
2383 reg = &func->regs[regno];
2384 if (reg->type != SCALAR_VALUE) {
2385 WARN_ONCE(1, "backtracing misuse");
2392 reg->precise = true;
2396 if (func->stack[spi].slot_type[0] != STACK_SPILL) {
2400 reg = &func->stack[spi].spilled_ptr;
2401 if (reg->type != SCALAR_VALUE) {
2409 reg->precise = true;
2415 if (!reg_mask && !stack_mask)
2418 DECLARE_BITMAP(mask, 64);
2419 u32 history = st->jmp_history_cnt;
2421 if (env->log.level & BPF_LOG_LEVEL)
2422 verbose(env, "last_idx %d first_idx %d\n", last_idx, first_idx);
2423 for (i = last_idx;;) {
2428 err = backtrack_insn(env, i, ®_mask, &stack_mask);
2430 if (err == -ENOTSUPP) {
2431 mark_all_scalars_precise(env, st);
2436 if (!reg_mask && !stack_mask)
2437 /* Found assignment(s) into tracked register in this state.
2438 * Since this state is already marked, just return.
2439 * Nothing to be tracked further in the parent state.
2444 i = get_prev_insn_idx(st, i, &history);
2445 if (i >= env->prog->len) {
2446 /* This can happen if backtracking reached insn 0
2447 * and there are still reg_mask or stack_mask
2449 * It means the backtracking missed the spot where
2450 * particular register was initialized with a constant.
2452 verbose(env, "BUG backtracking idx %d\n", i);
2453 WARN_ONCE(1, "verifier backtracking bug");
2462 func = st->frame[st->curframe];
2463 bitmap_from_u64(mask, reg_mask);
2464 for_each_set_bit(i, mask, 32) {
2465 reg = &func->regs[i];
2466 if (reg->type != SCALAR_VALUE) {
2467 reg_mask &= ~(1u << i);
2472 reg->precise = true;
2475 bitmap_from_u64(mask, stack_mask);
2476 for_each_set_bit(i, mask, 64) {
2477 if (i >= func->allocated_stack / BPF_REG_SIZE) {
2478 /* the sequence of instructions:
2480 * 3: (7b) *(u64 *)(r3 -8) = r0
2481 * 4: (79) r4 = *(u64 *)(r10 -8)
2482 * doesn't contain jmps. It's backtracked
2483 * as a single block.
2484 * During backtracking insn 3 is not recognized as
2485 * stack access, so at the end of backtracking
2486 * stack slot fp-8 is still marked in stack_mask.
2487 * However the parent state may not have accessed
2488 * fp-8 and it's "unallocated" stack space.
2489 * In such case fallback to conservative.
2491 mark_all_scalars_precise(env, st);
2495 if (func->stack[i].slot_type[0] != STACK_SPILL) {
2496 stack_mask &= ~(1ull << i);
2499 reg = &func->stack[i].spilled_ptr;
2500 if (reg->type != SCALAR_VALUE) {
2501 stack_mask &= ~(1ull << i);
2506 reg->precise = true;
2508 if (env->log.level & BPF_LOG_LEVEL) {
2509 print_verifier_state(env, func);
2510 verbose(env, "parent %s regs=%x stack=%llx marks\n",
2511 new_marks ? "didn't have" : "already had",
2512 reg_mask, stack_mask);
2515 if (!reg_mask && !stack_mask)
2520 last_idx = st->last_insn_idx;
2521 first_idx = st->first_insn_idx;
2526 static int mark_chain_precision(struct bpf_verifier_env *env, int regno)
2528 return __mark_chain_precision(env, regno, -1);
2531 static int mark_chain_precision_stack(struct bpf_verifier_env *env, int spi)
2533 return __mark_chain_precision(env, -1, spi);
2536 static bool is_spillable_regtype(enum bpf_reg_type type)
2538 switch (base_type(type)) {
2539 case PTR_TO_MAP_VALUE:
2543 case PTR_TO_PACKET_META:
2544 case PTR_TO_PACKET_END:
2545 case PTR_TO_FLOW_KEYS:
2546 case CONST_PTR_TO_MAP:
2548 case PTR_TO_SOCK_COMMON:
2549 case PTR_TO_TCP_SOCK:
2550 case PTR_TO_XDP_SOCK:
2553 case PTR_TO_PERCPU_BTF_ID:
2556 case PTR_TO_MAP_KEY:
2563 /* Does this register contain a constant zero? */
2564 static bool register_is_null(struct bpf_reg_state *reg)
2566 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
2569 static bool register_is_const(struct bpf_reg_state *reg)
2571 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
2574 static bool __is_scalar_unbounded(struct bpf_reg_state *reg)
2576 return tnum_is_unknown(reg->var_off) &&
2577 reg->smin_value == S64_MIN && reg->smax_value == S64_MAX &&
2578 reg->umin_value == 0 && reg->umax_value == U64_MAX &&
2579 reg->s32_min_value == S32_MIN && reg->s32_max_value == S32_MAX &&
2580 reg->u32_min_value == 0 && reg->u32_max_value == U32_MAX;
2583 static bool register_is_bounded(struct bpf_reg_state *reg)
2585 return reg->type == SCALAR_VALUE && !__is_scalar_unbounded(reg);
2588 static bool __is_pointer_value(bool allow_ptr_leaks,
2589 const struct bpf_reg_state *reg)
2591 if (allow_ptr_leaks)
2594 return reg->type != SCALAR_VALUE;
2597 static void save_register_state(struct bpf_func_state *state,
2598 int spi, struct bpf_reg_state *reg)
2602 state->stack[spi].spilled_ptr = *reg;
2603 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2605 for (i = 0; i < BPF_REG_SIZE; i++)
2606 state->stack[spi].slot_type[i] = STACK_SPILL;
2609 /* check_stack_{read,write}_fixed_off functions track spill/fill of registers,
2610 * stack boundary and alignment are checked in check_mem_access()
2612 static int check_stack_write_fixed_off(struct bpf_verifier_env *env,
2613 /* stack frame we're writing to */
2614 struct bpf_func_state *state,
2615 int off, int size, int value_regno,
2618 struct bpf_func_state *cur; /* state of the current function */
2619 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
2620 u32 dst_reg = env->prog->insnsi[insn_idx].dst_reg;
2621 struct bpf_reg_state *reg = NULL;
2623 err = grow_stack_state(state, round_up(slot + 1, BPF_REG_SIZE));
2626 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
2627 * so it's aligned access and [off, off + size) are within stack limits
2629 if (!env->allow_ptr_leaks &&
2630 state->stack[spi].slot_type[0] == STACK_SPILL &&
2631 size != BPF_REG_SIZE) {
2632 verbose(env, "attempt to corrupt spilled pointer on stack\n");
2636 cur = env->cur_state->frame[env->cur_state->curframe];
2637 if (value_regno >= 0)
2638 reg = &cur->regs[value_regno];
2639 if (!env->bypass_spec_v4) {
2640 bool sanitize = reg && is_spillable_regtype(reg->type);
2642 for (i = 0; i < size; i++) {
2643 if (state->stack[spi].slot_type[i] == STACK_INVALID) {
2650 env->insn_aux_data[insn_idx].sanitize_stack_spill = true;
2653 if (reg && size == BPF_REG_SIZE && register_is_bounded(reg) &&
2654 !register_is_null(reg) && env->bpf_capable) {
2655 if (dst_reg != BPF_REG_FP) {
2656 /* The backtracking logic can only recognize explicit
2657 * stack slot address like [fp - 8]. Other spill of
2658 * scalar via different register has to be conservative.
2659 * Backtrack from here and mark all registers as precise
2660 * that contributed into 'reg' being a constant.
2662 err = mark_chain_precision(env, value_regno);
2666 save_register_state(state, spi, reg);
2667 } else if (reg && is_spillable_regtype(reg->type)) {
2668 /* register containing pointer is being spilled into stack */
2669 if (size != BPF_REG_SIZE) {
2670 verbose_linfo(env, insn_idx, "; ");
2671 verbose(env, "invalid size of register spill\n");
2674 if (state != cur && reg->type == PTR_TO_STACK) {
2675 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
2678 save_register_state(state, spi, reg);
2680 u8 type = STACK_MISC;
2682 /* regular write of data into stack destroys any spilled ptr */
2683 state->stack[spi].spilled_ptr.type = NOT_INIT;
2684 /* Mark slots as STACK_MISC if they belonged to spilled ptr. */
2685 if (state->stack[spi].slot_type[0] == STACK_SPILL)
2686 for (i = 0; i < BPF_REG_SIZE; i++)
2687 state->stack[spi].slot_type[i] = STACK_MISC;
2689 /* only mark the slot as written if all 8 bytes were written
2690 * otherwise read propagation may incorrectly stop too soon
2691 * when stack slots are partially written.
2692 * This heuristic means that read propagation will be
2693 * conservative, since it will add reg_live_read marks
2694 * to stack slots all the way to first state when programs
2695 * writes+reads less than 8 bytes
2697 if (size == BPF_REG_SIZE)
2698 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2700 /* when we zero initialize stack slots mark them as such */
2701 if (reg && register_is_null(reg)) {
2702 /* backtracking doesn't work for STACK_ZERO yet. */
2703 err = mark_chain_precision(env, value_regno);
2709 /* Mark slots affected by this stack write. */
2710 for (i = 0; i < size; i++)
2711 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
2717 /* Write the stack: 'stack[ptr_regno + off] = value_regno'. 'ptr_regno' is
2718 * known to contain a variable offset.
2719 * This function checks whether the write is permitted and conservatively
2720 * tracks the effects of the write, considering that each stack slot in the
2721 * dynamic range is potentially written to.
2723 * 'off' includes 'regno->off'.
2724 * 'value_regno' can be -1, meaning that an unknown value is being written to
2727 * Spilled pointers in range are not marked as written because we don't know
2728 * what's going to be actually written. This means that read propagation for
2729 * future reads cannot be terminated by this write.
2731 * For privileged programs, uninitialized stack slots are considered
2732 * initialized by this write (even though we don't know exactly what offsets
2733 * are going to be written to). The idea is that we don't want the verifier to
2734 * reject future reads that access slots written to through variable offsets.
2736 static int check_stack_write_var_off(struct bpf_verifier_env *env,
2737 /* func where register points to */
2738 struct bpf_func_state *state,
2739 int ptr_regno, int off, int size,
2740 int value_regno, int insn_idx)
2742 struct bpf_func_state *cur; /* state of the current function */
2743 int min_off, max_off;
2745 struct bpf_reg_state *ptr_reg = NULL, *value_reg = NULL;
2746 bool writing_zero = false;
2747 /* set if the fact that we're writing a zero is used to let any
2748 * stack slots remain STACK_ZERO
2750 bool zero_used = false;
2752 cur = env->cur_state->frame[env->cur_state->curframe];
2753 ptr_reg = &cur->regs[ptr_regno];
2754 min_off = ptr_reg->smin_value + off;
2755 max_off = ptr_reg->smax_value + off + size;
2756 if (value_regno >= 0)
2757 value_reg = &cur->regs[value_regno];
2758 if (value_reg && register_is_null(value_reg))
2759 writing_zero = true;
2761 err = grow_stack_state(state, round_up(-min_off, BPF_REG_SIZE));
2766 /* Variable offset writes destroy any spilled pointers in range. */
2767 for (i = min_off; i < max_off; i++) {
2768 u8 new_type, *stype;
2772 spi = slot / BPF_REG_SIZE;
2773 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
2775 if (!env->allow_ptr_leaks
2776 && *stype != NOT_INIT
2777 && *stype != SCALAR_VALUE) {
2778 /* Reject the write if there's are spilled pointers in
2779 * range. If we didn't reject here, the ptr status
2780 * would be erased below (even though not all slots are
2781 * actually overwritten), possibly opening the door to
2784 verbose(env, "spilled ptr in range of var-offset stack write; insn %d, ptr off: %d",
2789 /* Erase all spilled pointers. */
2790 state->stack[spi].spilled_ptr.type = NOT_INIT;
2792 /* Update the slot type. */
2793 new_type = STACK_MISC;
2794 if (writing_zero && *stype == STACK_ZERO) {
2795 new_type = STACK_ZERO;
2798 /* If the slot is STACK_INVALID, we check whether it's OK to
2799 * pretend that it will be initialized by this write. The slot
2800 * might not actually be written to, and so if we mark it as
2801 * initialized future reads might leak uninitialized memory.
2802 * For privileged programs, we will accept such reads to slots
2803 * that may or may not be written because, if we're reject
2804 * them, the error would be too confusing.
2806 if (*stype == STACK_INVALID && !env->allow_uninit_stack) {
2807 verbose(env, "uninit stack in range of var-offset write prohibited for !root; insn %d, off: %d",
2814 /* backtracking doesn't work for STACK_ZERO yet. */
2815 err = mark_chain_precision(env, value_regno);
2822 /* When register 'dst_regno' is assigned some values from stack[min_off,
2823 * max_off), we set the register's type according to the types of the
2824 * respective stack slots. If all the stack values are known to be zeros, then
2825 * so is the destination reg. Otherwise, the register is considered to be
2826 * SCALAR. This function does not deal with register filling; the caller must
2827 * ensure that all spilled registers in the stack range have been marked as
2830 static void mark_reg_stack_read(struct bpf_verifier_env *env,
2831 /* func where src register points to */
2832 struct bpf_func_state *ptr_state,
2833 int min_off, int max_off, int dst_regno)
2835 struct bpf_verifier_state *vstate = env->cur_state;
2836 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2841 for (i = min_off; i < max_off; i++) {
2843 spi = slot / BPF_REG_SIZE;
2844 stype = ptr_state->stack[spi].slot_type;
2845 if (stype[slot % BPF_REG_SIZE] != STACK_ZERO)
2849 if (zeros == max_off - min_off) {
2850 /* any access_size read into register is zero extended,
2851 * so the whole register == const_zero
2853 __mark_reg_const_zero(&state->regs[dst_regno]);
2854 /* backtracking doesn't support STACK_ZERO yet,
2855 * so mark it precise here, so that later
2856 * backtracking can stop here.
2857 * Backtracking may not need this if this register
2858 * doesn't participate in pointer adjustment.
2859 * Forward propagation of precise flag is not
2860 * necessary either. This mark is only to stop
2861 * backtracking. Any register that contributed
2862 * to const 0 was marked precise before spill.
2864 state->regs[dst_regno].precise = true;
2866 /* have read misc data from the stack */
2867 mark_reg_unknown(env, state->regs, dst_regno);
2869 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
2872 /* Read the stack at 'off' and put the results into the register indicated by
2873 * 'dst_regno'. It handles reg filling if the addressed stack slot is a
2876 * 'dst_regno' can be -1, meaning that the read value is not going to a
2879 * The access is assumed to be within the current stack bounds.
2881 static int check_stack_read_fixed_off(struct bpf_verifier_env *env,
2882 /* func where src register points to */
2883 struct bpf_func_state *reg_state,
2884 int off, int size, int dst_regno)
2886 struct bpf_verifier_state *vstate = env->cur_state;
2887 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2888 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
2889 struct bpf_reg_state *reg;
2892 stype = reg_state->stack[spi].slot_type;
2893 reg = ®_state->stack[spi].spilled_ptr;
2895 if (stype[0] == STACK_SPILL) {
2896 if (size != BPF_REG_SIZE) {
2897 if (reg->type != SCALAR_VALUE) {
2898 verbose_linfo(env, env->insn_idx, "; ");
2899 verbose(env, "invalid size of register fill\n");
2902 if (dst_regno >= 0) {
2903 mark_reg_unknown(env, state->regs, dst_regno);
2904 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
2906 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2909 for (i = 1; i < BPF_REG_SIZE; i++) {
2910 if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) {
2911 verbose(env, "corrupted spill memory\n");
2916 if (dst_regno >= 0) {
2917 /* restore register state from stack */
2918 state->regs[dst_regno] = *reg;
2919 /* mark reg as written since spilled pointer state likely
2920 * has its liveness marks cleared by is_state_visited()
2921 * which resets stack/reg liveness for state transitions
2923 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
2924 } else if (__is_pointer_value(env->allow_ptr_leaks, reg)) {
2925 /* If dst_regno==-1, the caller is asking us whether
2926 * it is acceptable to use this value as a SCALAR_VALUE
2928 * We must not allow unprivileged callers to do that
2929 * with spilled pointers.
2931 verbose(env, "leaking pointer from stack off %d\n",
2935 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2939 for (i = 0; i < size; i++) {
2940 type = stype[(slot - i) % BPF_REG_SIZE];
2941 if (type == STACK_MISC)
2943 if (type == STACK_ZERO)
2945 verbose(env, "invalid read from stack off %d+%d size %d\n",
2949 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2951 mark_reg_stack_read(env, reg_state, off, off + size, dst_regno);
2956 enum stack_access_src {
2957 ACCESS_DIRECT = 1, /* the access is performed by an instruction */
2958 ACCESS_HELPER = 2, /* the access is performed by a helper */
2961 static int check_stack_range_initialized(struct bpf_verifier_env *env,
2962 int regno, int off, int access_size,
2963 bool zero_size_allowed,
2964 enum stack_access_src type,
2965 struct bpf_call_arg_meta *meta);
2967 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
2969 return cur_regs(env) + regno;
2972 /* Read the stack at 'ptr_regno + off' and put the result into the register
2974 * 'off' includes the pointer register's fixed offset(i.e. 'ptr_regno.off'),
2975 * but not its variable offset.
2976 * 'size' is assumed to be <= reg size and the access is assumed to be aligned.
2978 * As opposed to check_stack_read_fixed_off, this function doesn't deal with
2979 * filling registers (i.e. reads of spilled register cannot be detected when
2980 * the offset is not fixed). We conservatively mark 'dst_regno' as containing
2981 * SCALAR_VALUE. That's why we assert that the 'ptr_regno' has a variable
2982 * offset; for a fixed offset check_stack_read_fixed_off should be used
2985 static int check_stack_read_var_off(struct bpf_verifier_env *env,
2986 int ptr_regno, int off, int size, int dst_regno)
2988 /* The state of the source register. */
2989 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
2990 struct bpf_func_state *ptr_state = func(env, reg);
2992 int min_off, max_off;
2994 /* Note that we pass a NULL meta, so raw access will not be permitted.
2996 err = check_stack_range_initialized(env, ptr_regno, off, size,
2997 false, ACCESS_DIRECT, NULL);
3001 min_off = reg->smin_value + off;
3002 max_off = reg->smax_value + off;
3003 mark_reg_stack_read(env, ptr_state, min_off, max_off + size, dst_regno);
3007 /* check_stack_read dispatches to check_stack_read_fixed_off or
3008 * check_stack_read_var_off.
3010 * The caller must ensure that the offset falls within the allocated stack
3013 * 'dst_regno' is a register which will receive the value from the stack. It
3014 * can be -1, meaning that the read value is not going to a register.
3016 static int check_stack_read(struct bpf_verifier_env *env,
3017 int ptr_regno, int off, int size,
3020 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
3021 struct bpf_func_state *state = func(env, reg);
3023 /* Some accesses are only permitted with a static offset. */
3024 bool var_off = !tnum_is_const(reg->var_off);
3026 /* The offset is required to be static when reads don't go to a
3027 * register, in order to not leak pointers (see
3028 * check_stack_read_fixed_off).
3030 if (dst_regno < 0 && var_off) {
3033 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3034 verbose(env, "variable offset stack pointer cannot be passed into helper function; var_off=%s off=%d size=%d\n",
3038 /* Variable offset is prohibited for unprivileged mode for simplicity
3039 * since it requires corresponding support in Spectre masking for stack
3040 * ALU. See also retrieve_ptr_limit().
3042 if (!env->bypass_spec_v1 && var_off) {
3045 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3046 verbose(env, "R%d variable offset stack access prohibited for !root, var_off=%s\n",
3052 off += reg->var_off.value;
3053 err = check_stack_read_fixed_off(env, state, off, size,
3056 /* Variable offset stack reads need more conservative handling
3057 * than fixed offset ones. Note that dst_regno >= 0 on this
3060 err = check_stack_read_var_off(env, ptr_regno, off, size,
3067 /* check_stack_write dispatches to check_stack_write_fixed_off or
3068 * check_stack_write_var_off.
3070 * 'ptr_regno' is the register used as a pointer into the stack.
3071 * 'off' includes 'ptr_regno->off', but not its variable offset (if any).
3072 * 'value_regno' is the register whose value we're writing to the stack. It can
3073 * be -1, meaning that we're not writing from a register.
3075 * The caller must ensure that the offset falls within the maximum stack size.
3077 static int check_stack_write(struct bpf_verifier_env *env,
3078 int ptr_regno, int off, int size,
3079 int value_regno, int insn_idx)
3081 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
3082 struct bpf_func_state *state = func(env, reg);
3085 if (tnum_is_const(reg->var_off)) {
3086 off += reg->var_off.value;
3087 err = check_stack_write_fixed_off(env, state, off, size,
3088 value_regno, insn_idx);
3090 /* Variable offset stack reads need more conservative handling
3091 * than fixed offset ones.
3093 err = check_stack_write_var_off(env, state,
3094 ptr_regno, off, size,
3095 value_regno, insn_idx);
3100 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
3101 int off, int size, enum bpf_access_type type)
3103 struct bpf_reg_state *regs = cur_regs(env);
3104 struct bpf_map *map = regs[regno].map_ptr;
3105 u32 cap = bpf_map_flags_to_cap(map);
3107 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
3108 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
3109 map->value_size, off, size);
3113 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
3114 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
3115 map->value_size, off, size);
3122 /* check read/write into memory region (e.g., map value, ringbuf sample, etc) */
3123 static int __check_mem_access(struct bpf_verifier_env *env, int regno,
3124 int off, int size, u32 mem_size,
3125 bool zero_size_allowed)
3127 bool size_ok = size > 0 || (size == 0 && zero_size_allowed);
3128 struct bpf_reg_state *reg;
3130 if (off >= 0 && size_ok && (u64)off + size <= mem_size)
3133 reg = &cur_regs(env)[regno];
3134 switch (reg->type) {
3135 case PTR_TO_MAP_KEY:
3136 verbose(env, "invalid access to map key, key_size=%d off=%d size=%d\n",
3137 mem_size, off, size);
3139 case PTR_TO_MAP_VALUE:
3140 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
3141 mem_size, off, size);
3144 case PTR_TO_PACKET_META:
3145 case PTR_TO_PACKET_END:
3146 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
3147 off, size, regno, reg->id, off, mem_size);
3151 verbose(env, "invalid access to memory, mem_size=%u off=%d size=%d\n",
3152 mem_size, off, size);
3158 /* check read/write into a memory region with possible variable offset */
3159 static int check_mem_region_access(struct bpf_verifier_env *env, u32 regno,
3160 int off, int size, u32 mem_size,
3161 bool zero_size_allowed)
3163 struct bpf_verifier_state *vstate = env->cur_state;
3164 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3165 struct bpf_reg_state *reg = &state->regs[regno];
3168 /* We may have adjusted the register pointing to memory region, so we
3169 * need to try adding each of min_value and max_value to off
3170 * to make sure our theoretical access will be safe.
3172 if (env->log.level & BPF_LOG_LEVEL)
3173 print_verifier_state(env, state);
3175 /* The minimum value is only important with signed
3176 * comparisons where we can't assume the floor of a
3177 * value is 0. If we are using signed variables for our
3178 * index'es we need to make sure that whatever we use
3179 * will have a set floor within our range.
3181 if (reg->smin_value < 0 &&
3182 (reg->smin_value == S64_MIN ||
3183 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
3184 reg->smin_value + off < 0)) {
3185 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3189 err = __check_mem_access(env, regno, reg->smin_value + off, size,
3190 mem_size, zero_size_allowed);
3192 verbose(env, "R%d min value is outside of the allowed memory range\n",
3197 /* If we haven't set a max value then we need to bail since we can't be
3198 * sure we won't do bad things.
3199 * If reg->umax_value + off could overflow, treat that as unbounded too.
3201 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
3202 verbose(env, "R%d unbounded memory access, make sure to bounds check any such access\n",
3206 err = __check_mem_access(env, regno, reg->umax_value + off, size,
3207 mem_size, zero_size_allowed);
3209 verbose(env, "R%d max value is outside of the allowed memory range\n",
3217 /* check read/write into a map element with possible variable offset */
3218 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
3219 int off, int size, bool zero_size_allowed)
3221 struct bpf_verifier_state *vstate = env->cur_state;
3222 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3223 struct bpf_reg_state *reg = &state->regs[regno];
3224 struct bpf_map *map = reg->map_ptr;
3227 err = check_mem_region_access(env, regno, off, size, map->value_size,
3232 if (map_value_has_spin_lock(map)) {
3233 u32 lock = map->spin_lock_off;
3235 /* if any part of struct bpf_spin_lock can be touched by
3236 * load/store reject this program.
3237 * To check that [x1, x2) overlaps with [y1, y2)
3238 * it is sufficient to check x1 < y2 && y1 < x2.
3240 if (reg->smin_value + off < lock + sizeof(struct bpf_spin_lock) &&
3241 lock < reg->umax_value + off + size) {
3242 verbose(env, "bpf_spin_lock cannot be accessed directly by load/store\n");
3246 if (map_value_has_timer(map)) {
3247 u32 t = map->timer_off;
3249 if (reg->smin_value + off < t + sizeof(struct bpf_timer) &&
3250 t < reg->umax_value + off + size) {
3251 verbose(env, "bpf_timer cannot be accessed directly by load/store\n");
3258 #define MAX_PACKET_OFF 0xffff
3260 static enum bpf_prog_type resolve_prog_type(struct bpf_prog *prog)
3262 return prog->aux->dst_prog ? prog->aux->dst_prog->type : prog->type;
3265 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
3266 const struct bpf_call_arg_meta *meta,
3267 enum bpf_access_type t)
3269 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
3271 switch (prog_type) {
3272 /* Program types only with direct read access go here! */
3273 case BPF_PROG_TYPE_LWT_IN:
3274 case BPF_PROG_TYPE_LWT_OUT:
3275 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
3276 case BPF_PROG_TYPE_SK_REUSEPORT:
3277 case BPF_PROG_TYPE_FLOW_DISSECTOR:
3278 case BPF_PROG_TYPE_CGROUP_SKB:
3283 /* Program types with direct read + write access go here! */
3284 case BPF_PROG_TYPE_SCHED_CLS:
3285 case BPF_PROG_TYPE_SCHED_ACT:
3286 case BPF_PROG_TYPE_XDP:
3287 case BPF_PROG_TYPE_LWT_XMIT:
3288 case BPF_PROG_TYPE_SK_SKB:
3289 case BPF_PROG_TYPE_SK_MSG:
3291 return meta->pkt_access;
3293 env->seen_direct_write = true;
3296 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
3298 env->seen_direct_write = true;
3307 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
3308 int size, bool zero_size_allowed)
3310 struct bpf_reg_state *regs = cur_regs(env);
3311 struct bpf_reg_state *reg = ®s[regno];
3314 /* We may have added a variable offset to the packet pointer; but any
3315 * reg->range we have comes after that. We are only checking the fixed
3319 /* We don't allow negative numbers, because we aren't tracking enough
3320 * detail to prove they're safe.
3322 if (reg->smin_value < 0) {
3323 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3328 err = reg->range < 0 ? -EINVAL :
3329 __check_mem_access(env, regno, off, size, reg->range,
3332 verbose(env, "R%d offset is outside of the packet\n", regno);
3336 /* __check_mem_access has made sure "off + size - 1" is within u16.
3337 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
3338 * otherwise find_good_pkt_pointers would have refused to set range info
3339 * that __check_mem_access would have rejected this pkt access.
3340 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
3342 env->prog->aux->max_pkt_offset =
3343 max_t(u32, env->prog->aux->max_pkt_offset,
3344 off + reg->umax_value + size - 1);
3349 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
3350 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
3351 enum bpf_access_type t, enum bpf_reg_type *reg_type,
3352 struct btf **btf, u32 *btf_id)
3354 struct bpf_insn_access_aux info = {
3355 .reg_type = *reg_type,
3359 if (env->ops->is_valid_access &&
3360 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
3361 /* A non zero info.ctx_field_size indicates that this field is a
3362 * candidate for later verifier transformation to load the whole
3363 * field and then apply a mask when accessed with a narrower
3364 * access than actual ctx access size. A zero info.ctx_field_size
3365 * will only allow for whole field access and rejects any other
3366 * type of narrower access.
3368 *reg_type = info.reg_type;
3370 if (base_type(*reg_type) == PTR_TO_BTF_ID) {
3372 *btf_id = info.btf_id;
3374 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
3376 /* remember the offset of last byte accessed in ctx */
3377 if (env->prog->aux->max_ctx_offset < off + size)
3378 env->prog->aux->max_ctx_offset = off + size;
3382 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
3386 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
3389 if (size < 0 || off < 0 ||
3390 (u64)off + size > sizeof(struct bpf_flow_keys)) {
3391 verbose(env, "invalid access to flow keys off=%d size=%d\n",
3398 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
3399 u32 regno, int off, int size,
3400 enum bpf_access_type t)
3402 struct bpf_reg_state *regs = cur_regs(env);
3403 struct bpf_reg_state *reg = ®s[regno];
3404 struct bpf_insn_access_aux info = {};
3407 if (reg->smin_value < 0) {
3408 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3413 switch (reg->type) {
3414 case PTR_TO_SOCK_COMMON:
3415 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
3418 valid = bpf_sock_is_valid_access(off, size, t, &info);
3420 case PTR_TO_TCP_SOCK:
3421 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
3423 case PTR_TO_XDP_SOCK:
3424 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
3432 env->insn_aux_data[insn_idx].ctx_field_size =
3433 info.ctx_field_size;
3437 verbose(env, "R%d invalid %s access off=%d size=%d\n",
3438 regno, reg_type_str(env, reg->type), off, size);
3443 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
3445 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
3448 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
3450 const struct bpf_reg_state *reg = reg_state(env, regno);
3452 return reg->type == PTR_TO_CTX;
3455 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
3457 const struct bpf_reg_state *reg = reg_state(env, regno);
3459 return type_is_sk_pointer(reg->type);
3462 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
3464 const struct bpf_reg_state *reg = reg_state(env, regno);
3466 return type_is_pkt_pointer(reg->type);
3469 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
3471 const struct bpf_reg_state *reg = reg_state(env, regno);
3473 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
3474 return reg->type == PTR_TO_FLOW_KEYS;
3477 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
3478 const struct bpf_reg_state *reg,
3479 int off, int size, bool strict)
3481 struct tnum reg_off;
3484 /* Byte size accesses are always allowed. */
3485 if (!strict || size == 1)
3488 /* For platforms that do not have a Kconfig enabling
3489 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
3490 * NET_IP_ALIGN is universally set to '2'. And on platforms
3491 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
3492 * to this code only in strict mode where we want to emulate
3493 * the NET_IP_ALIGN==2 checking. Therefore use an
3494 * unconditional IP align value of '2'.
3498 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
3499 if (!tnum_is_aligned(reg_off, size)) {
3502 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3504 "misaligned packet access off %d+%s+%d+%d size %d\n",
3505 ip_align, tn_buf, reg->off, off, size);
3512 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
3513 const struct bpf_reg_state *reg,
3514 const char *pointer_desc,
3515 int off, int size, bool strict)
3517 struct tnum reg_off;
3519 /* Byte size accesses are always allowed. */
3520 if (!strict || size == 1)
3523 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
3524 if (!tnum_is_aligned(reg_off, size)) {
3527 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3528 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
3529 pointer_desc, tn_buf, reg->off, off, size);
3536 static int check_ptr_alignment(struct bpf_verifier_env *env,
3537 const struct bpf_reg_state *reg, int off,
3538 int size, bool strict_alignment_once)
3540 bool strict = env->strict_alignment || strict_alignment_once;
3541 const char *pointer_desc = "";
3543 switch (reg->type) {
3545 case PTR_TO_PACKET_META:
3546 /* Special case, because of NET_IP_ALIGN. Given metadata sits
3547 * right in front, treat it the very same way.
3549 return check_pkt_ptr_alignment(env, reg, off, size, strict);
3550 case PTR_TO_FLOW_KEYS:
3551 pointer_desc = "flow keys ";
3553 case PTR_TO_MAP_KEY:
3554 pointer_desc = "key ";
3556 case PTR_TO_MAP_VALUE:
3557 pointer_desc = "value ";
3560 pointer_desc = "context ";
3563 pointer_desc = "stack ";
3564 /* The stack spill tracking logic in check_stack_write_fixed_off()
3565 * and check_stack_read_fixed_off() relies on stack accesses being
3571 pointer_desc = "sock ";
3573 case PTR_TO_SOCK_COMMON:
3574 pointer_desc = "sock_common ";
3576 case PTR_TO_TCP_SOCK:
3577 pointer_desc = "tcp_sock ";
3579 case PTR_TO_XDP_SOCK:
3580 pointer_desc = "xdp_sock ";
3585 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
3589 static int update_stack_depth(struct bpf_verifier_env *env,
3590 const struct bpf_func_state *func,
3593 u16 stack = env->subprog_info[func->subprogno].stack_depth;
3598 /* update known max for given subprogram */
3599 env->subprog_info[func->subprogno].stack_depth = -off;
3603 /* starting from main bpf function walk all instructions of the function
3604 * and recursively walk all callees that given function can call.
3605 * Ignore jump and exit insns.
3606 * Since recursion is prevented by check_cfg() this algorithm
3607 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
3609 static int check_max_stack_depth(struct bpf_verifier_env *env)
3611 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
3612 struct bpf_subprog_info *subprog = env->subprog_info;
3613 struct bpf_insn *insn = env->prog->insnsi;
3614 bool tail_call_reachable = false;
3615 int ret_insn[MAX_CALL_FRAMES];
3616 int ret_prog[MAX_CALL_FRAMES];
3620 /* protect against potential stack overflow that might happen when
3621 * bpf2bpf calls get combined with tailcalls. Limit the caller's stack
3622 * depth for such case down to 256 so that the worst case scenario
3623 * would result in 8k stack size (32 which is tailcall limit * 256 =
3626 * To get the idea what might happen, see an example:
3627 * func1 -> sub rsp, 128
3628 * subfunc1 -> sub rsp, 256
3629 * tailcall1 -> add rsp, 256
3630 * func2 -> sub rsp, 192 (total stack size = 128 + 192 = 320)
3631 * subfunc2 -> sub rsp, 64
3632 * subfunc22 -> sub rsp, 128
3633 * tailcall2 -> add rsp, 128
3634 * func3 -> sub rsp, 32 (total stack size 128 + 192 + 64 + 32 = 416)
3636 * tailcall will unwind the current stack frame but it will not get rid
3637 * of caller's stack as shown on the example above.
3639 if (idx && subprog[idx].has_tail_call && depth >= 256) {
3641 "tail_calls are not allowed when call stack of previous frames is %d bytes. Too large\n",
3645 /* round up to 32-bytes, since this is granularity
3646 * of interpreter stack size
3648 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
3649 if (depth > MAX_BPF_STACK) {
3650 verbose(env, "combined stack size of %d calls is %d. Too large\n",
3655 subprog_end = subprog[idx + 1].start;
3656 for (; i < subprog_end; i++) {
3659 if (!bpf_pseudo_call(insn + i) && !bpf_pseudo_func(insn + i))
3661 /* remember insn and function to return to */
3662 ret_insn[frame] = i + 1;
3663 ret_prog[frame] = idx;
3665 /* find the callee */
3666 next_insn = i + insn[i].imm + 1;
3667 idx = find_subprog(env, next_insn);
3669 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3673 if (subprog[idx].is_async_cb) {
3674 if (subprog[idx].has_tail_call) {
3675 verbose(env, "verifier bug. subprog has tail_call and async cb\n");
3678 /* async callbacks don't increase bpf prog stack size */
3683 if (subprog[idx].has_tail_call)
3684 tail_call_reachable = true;
3687 if (frame >= MAX_CALL_FRAMES) {
3688 verbose(env, "the call stack of %d frames is too deep !\n",
3694 /* if tail call got detected across bpf2bpf calls then mark each of the
3695 * currently present subprog frames as tail call reachable subprogs;
3696 * this info will be utilized by JIT so that we will be preserving the
3697 * tail call counter throughout bpf2bpf calls combined with tailcalls
3699 if (tail_call_reachable)
3700 for (j = 0; j < frame; j++)
3701 subprog[ret_prog[j]].tail_call_reachable = true;
3702 if (subprog[0].tail_call_reachable)
3703 env->prog->aux->tail_call_reachable = true;
3705 /* end of for() loop means the last insn of the 'subprog'
3706 * was reached. Doesn't matter whether it was JA or EXIT
3710 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
3712 i = ret_insn[frame];
3713 idx = ret_prog[frame];
3717 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
3718 static int get_callee_stack_depth(struct bpf_verifier_env *env,
3719 const struct bpf_insn *insn, int idx)
3721 int start = idx + insn->imm + 1, subprog;
3723 subprog = find_subprog(env, start);
3725 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3729 return env->subprog_info[subprog].stack_depth;
3733 int check_ctx_reg(struct bpf_verifier_env *env,
3734 const struct bpf_reg_state *reg, int regno)
3736 /* Access to ctx or passing it to a helper is only allowed in
3737 * its original, unmodified form.
3741 verbose(env, "dereference of modified ctx ptr R%d off=%d disallowed\n",
3746 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3749 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3750 verbose(env, "variable ctx access var_off=%s disallowed\n", tn_buf);
3757 static int __check_buffer_access(struct bpf_verifier_env *env,
3758 const char *buf_info,
3759 const struct bpf_reg_state *reg,
3760 int regno, int off, int size)
3764 "R%d invalid %s buffer access: off=%d, size=%d\n",
3765 regno, buf_info, off, size);
3768 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3771 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3773 "R%d invalid variable buffer offset: off=%d, var_off=%s\n",
3774 regno, off, tn_buf);
3781 static int check_tp_buffer_access(struct bpf_verifier_env *env,
3782 const struct bpf_reg_state *reg,
3783 int regno, int off, int size)
3787 err = __check_buffer_access(env, "tracepoint", reg, regno, off, size);
3791 if (off + size > env->prog->aux->max_tp_access)
3792 env->prog->aux->max_tp_access = off + size;
3797 static int check_buffer_access(struct bpf_verifier_env *env,
3798 const struct bpf_reg_state *reg,
3799 int regno, int off, int size,
3800 bool zero_size_allowed,
3801 const char *buf_info,
3806 err = __check_buffer_access(env, buf_info, reg, regno, off, size);
3810 if (off + size > *max_access)
3811 *max_access = off + size;
3816 /* BPF architecture zero extends alu32 ops into 64-bit registesr */
3817 static void zext_32_to_64(struct bpf_reg_state *reg)
3819 reg->var_off = tnum_subreg(reg->var_off);
3820 __reg_assign_32_into_64(reg);
3823 /* truncate register to smaller size (in bytes)
3824 * must be called with size < BPF_REG_SIZE
3826 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
3830 /* clear high bits in bit representation */
3831 reg->var_off = tnum_cast(reg->var_off, size);
3833 /* fix arithmetic bounds */
3834 mask = ((u64)1 << (size * 8)) - 1;
3835 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
3836 reg->umin_value &= mask;
3837 reg->umax_value &= mask;
3839 reg->umin_value = 0;
3840 reg->umax_value = mask;
3842 reg->smin_value = reg->umin_value;
3843 reg->smax_value = reg->umax_value;
3845 /* If size is smaller than 32bit register the 32bit register
3846 * values are also truncated so we push 64-bit bounds into
3847 * 32-bit bounds. Above were truncated < 32-bits already.
3851 __reg_combine_64_into_32(reg);
3854 static bool bpf_map_is_rdonly(const struct bpf_map *map)
3856 /* A map is considered read-only if the following condition are true:
3858 * 1) BPF program side cannot change any of the map content. The
3859 * BPF_F_RDONLY_PROG flag is throughout the lifetime of a map
3860 * and was set at map creation time.
3861 * 2) The map value(s) have been initialized from user space by a
3862 * loader and then "frozen", such that no new map update/delete
3863 * operations from syscall side are possible for the rest of
3864 * the map's lifetime from that point onwards.
3865 * 3) Any parallel/pending map update/delete operations from syscall
3866 * side have been completed. Only after that point, it's safe to
3867 * assume that map value(s) are immutable.
3869 return (map->map_flags & BPF_F_RDONLY_PROG) &&
3870 READ_ONCE(map->frozen) &&
3871 !bpf_map_write_active(map);
3874 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val)
3880 err = map->ops->map_direct_value_addr(map, &addr, off);
3883 ptr = (void *)(long)addr + off;
3887 *val = (u64)*(u8 *)ptr;
3890 *val = (u64)*(u16 *)ptr;
3893 *val = (u64)*(u32 *)ptr;
3904 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
3905 struct bpf_reg_state *regs,
3906 int regno, int off, int size,
3907 enum bpf_access_type atype,
3910 struct bpf_reg_state *reg = regs + regno;
3911 const struct btf_type *t = btf_type_by_id(reg->btf, reg->btf_id);
3912 const char *tname = btf_name_by_offset(reg->btf, t->name_off);
3918 "R%d is ptr_%s invalid negative access: off=%d\n",
3922 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3925 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3927 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
3928 regno, tname, off, tn_buf);
3932 if (env->ops->btf_struct_access) {
3933 ret = env->ops->btf_struct_access(&env->log, reg->btf, t,
3934 off, size, atype, &btf_id);
3936 if (atype != BPF_READ) {
3937 verbose(env, "only read is supported\n");
3941 ret = btf_struct_access(&env->log, reg->btf, t, off, size,
3948 if (atype == BPF_READ && value_regno >= 0)
3949 mark_btf_ld_reg(env, regs, value_regno, ret, reg->btf, btf_id);
3954 static int check_ptr_to_map_access(struct bpf_verifier_env *env,
3955 struct bpf_reg_state *regs,
3956 int regno, int off, int size,
3957 enum bpf_access_type atype,
3960 struct bpf_reg_state *reg = regs + regno;
3961 struct bpf_map *map = reg->map_ptr;
3962 const struct btf_type *t;
3968 verbose(env, "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n");
3972 if (!map->ops->map_btf_id || !*map->ops->map_btf_id) {
3973 verbose(env, "map_ptr access not supported for map type %d\n",
3978 t = btf_type_by_id(btf_vmlinux, *map->ops->map_btf_id);
3979 tname = btf_name_by_offset(btf_vmlinux, t->name_off);
3981 if (!env->allow_ptr_to_map_access) {
3983 "%s access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
3989 verbose(env, "R%d is %s invalid negative access: off=%d\n",
3994 if (atype != BPF_READ) {
3995 verbose(env, "only read from %s is supported\n", tname);
3999 ret = btf_struct_access(&env->log, btf_vmlinux, t, off, size, atype, &btf_id);
4003 if (value_regno >= 0)
4004 mark_btf_ld_reg(env, regs, value_regno, ret, btf_vmlinux, btf_id);
4009 /* Check that the stack access at the given offset is within bounds. The
4010 * maximum valid offset is -1.
4012 * The minimum valid offset is -MAX_BPF_STACK for writes, and
4013 * -state->allocated_stack for reads.
4015 static int check_stack_slot_within_bounds(int off,
4016 struct bpf_func_state *state,
4017 enum bpf_access_type t)
4022 min_valid_off = -MAX_BPF_STACK;
4024 min_valid_off = -state->allocated_stack;
4026 if (off < min_valid_off || off > -1)
4031 /* Check that the stack access at 'regno + off' falls within the maximum stack
4034 * 'off' includes `regno->offset`, but not its dynamic part (if any).
4036 static int check_stack_access_within_bounds(
4037 struct bpf_verifier_env *env,
4038 int regno, int off, int access_size,
4039 enum stack_access_src src, enum bpf_access_type type)
4041 struct bpf_reg_state *regs = cur_regs(env);
4042 struct bpf_reg_state *reg = regs + regno;
4043 struct bpf_func_state *state = func(env, reg);
4044 int min_off, max_off;
4048 if (src == ACCESS_HELPER)
4049 /* We don't know if helpers are reading or writing (or both). */
4050 err_extra = " indirect access to";
4051 else if (type == BPF_READ)
4052 err_extra = " read from";
4054 err_extra = " write to";
4056 if (tnum_is_const(reg->var_off)) {
4057 min_off = reg->var_off.value + off;
4058 if (access_size > 0)
4059 max_off = min_off + access_size - 1;
4063 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
4064 reg->smin_value <= -BPF_MAX_VAR_OFF) {
4065 verbose(env, "invalid unbounded variable-offset%s stack R%d\n",
4069 min_off = reg->smin_value + off;
4070 if (access_size > 0)
4071 max_off = reg->smax_value + off + access_size - 1;
4076 err = check_stack_slot_within_bounds(min_off, state, type);
4078 err = check_stack_slot_within_bounds(max_off, state, type);
4081 if (tnum_is_const(reg->var_off)) {
4082 verbose(env, "invalid%s stack R%d off=%d size=%d\n",
4083 err_extra, regno, off, access_size);
4087 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4088 verbose(env, "invalid variable-offset%s stack R%d var_off=%s size=%d\n",
4089 err_extra, regno, tn_buf, access_size);
4095 /* check whether memory at (regno + off) is accessible for t = (read | write)
4096 * if t==write, value_regno is a register which value is stored into memory
4097 * if t==read, value_regno is a register which will receive the value from memory
4098 * if t==write && value_regno==-1, some unknown value is stored into memory
4099 * if t==read && value_regno==-1, don't care what we read from memory
4101 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
4102 int off, int bpf_size, enum bpf_access_type t,
4103 int value_regno, bool strict_alignment_once)
4105 struct bpf_reg_state *regs = cur_regs(env);
4106 struct bpf_reg_state *reg = regs + regno;
4107 struct bpf_func_state *state;
4110 size = bpf_size_to_bytes(bpf_size);
4114 /* alignment checks will add in reg->off themselves */
4115 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
4119 /* for access checks, reg->off is just part of off */
4122 if (reg->type == PTR_TO_MAP_KEY) {
4123 if (t == BPF_WRITE) {
4124 verbose(env, "write to change key R%d not allowed\n", regno);
4128 err = check_mem_region_access(env, regno, off, size,
4129 reg->map_ptr->key_size, false);
4132 if (value_regno >= 0)
4133 mark_reg_unknown(env, regs, value_regno);
4134 } else if (reg->type == PTR_TO_MAP_VALUE) {
4135 if (t == BPF_WRITE && value_regno >= 0 &&
4136 is_pointer_value(env, value_regno)) {
4137 verbose(env, "R%d leaks addr into map\n", value_regno);
4140 err = check_map_access_type(env, regno, off, size, t);
4143 err = check_map_access(env, regno, off, size, false);
4144 if (!err && t == BPF_READ && value_regno >= 0) {
4145 struct bpf_map *map = reg->map_ptr;
4147 /* if map is read-only, track its contents as scalars */
4148 if (tnum_is_const(reg->var_off) &&
4149 bpf_map_is_rdonly(map) &&
4150 map->ops->map_direct_value_addr) {
4151 int map_off = off + reg->var_off.value;
4154 err = bpf_map_direct_read(map, map_off, size,
4159 regs[value_regno].type = SCALAR_VALUE;
4160 __mark_reg_known(®s[value_regno], val);
4162 mark_reg_unknown(env, regs, value_regno);
4165 } else if (base_type(reg->type) == PTR_TO_MEM) {
4166 bool rdonly_mem = type_is_rdonly_mem(reg->type);
4168 if (type_may_be_null(reg->type)) {
4169 verbose(env, "R%d invalid mem access '%s'\n", regno,
4170 reg_type_str(env, reg->type));
4174 if (t == BPF_WRITE && rdonly_mem) {
4175 verbose(env, "R%d cannot write into %s\n",
4176 regno, reg_type_str(env, reg->type));
4180 if (t == BPF_WRITE && value_regno >= 0 &&
4181 is_pointer_value(env, value_regno)) {
4182 verbose(env, "R%d leaks addr into mem\n", value_regno);
4186 err = check_mem_region_access(env, regno, off, size,
4187 reg->mem_size, false);
4188 if (!err && value_regno >= 0 && (t == BPF_READ || rdonly_mem))
4189 mark_reg_unknown(env, regs, value_regno);
4190 } else if (reg->type == PTR_TO_CTX) {
4191 enum bpf_reg_type reg_type = SCALAR_VALUE;
4192 struct btf *btf = NULL;
4195 if (t == BPF_WRITE && value_regno >= 0 &&
4196 is_pointer_value(env, value_regno)) {
4197 verbose(env, "R%d leaks addr into ctx\n", value_regno);
4201 err = check_ctx_reg(env, reg, regno);
4205 err = check_ctx_access(env, insn_idx, off, size, t, ®_type, &btf, &btf_id);
4207 verbose_linfo(env, insn_idx, "; ");
4208 if (!err && t == BPF_READ && value_regno >= 0) {
4209 /* ctx access returns either a scalar, or a
4210 * PTR_TO_PACKET[_META,_END]. In the latter
4211 * case, we know the offset is zero.
4213 if (reg_type == SCALAR_VALUE) {
4214 mark_reg_unknown(env, regs, value_regno);
4216 mark_reg_known_zero(env, regs,
4218 if (type_may_be_null(reg_type))
4219 regs[value_regno].id = ++env->id_gen;
4220 /* A load of ctx field could have different
4221 * actual load size with the one encoded in the
4222 * insn. When the dst is PTR, it is for sure not
4225 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
4226 if (base_type(reg_type) == PTR_TO_BTF_ID) {
4227 regs[value_regno].btf = btf;
4228 regs[value_regno].btf_id = btf_id;
4231 regs[value_regno].type = reg_type;
4234 } else if (reg->type == PTR_TO_STACK) {
4235 /* Basic bounds checks. */
4236 err = check_stack_access_within_bounds(env, regno, off, size, ACCESS_DIRECT, t);
4240 state = func(env, reg);
4241 err = update_stack_depth(env, state, off);
4246 err = check_stack_read(env, regno, off, size,
4249 err = check_stack_write(env, regno, off, size,
4250 value_regno, insn_idx);
4251 } else if (reg_is_pkt_pointer(reg)) {
4252 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
4253 verbose(env, "cannot write into packet\n");
4256 if (t == BPF_WRITE && value_regno >= 0 &&
4257 is_pointer_value(env, value_regno)) {
4258 verbose(env, "R%d leaks addr into packet\n",
4262 err = check_packet_access(env, regno, off, size, false);
4263 if (!err && t == BPF_READ && value_regno >= 0)
4264 mark_reg_unknown(env, regs, value_regno);
4265 } else if (reg->type == PTR_TO_FLOW_KEYS) {
4266 if (t == BPF_WRITE && value_regno >= 0 &&
4267 is_pointer_value(env, value_regno)) {
4268 verbose(env, "R%d leaks addr into flow keys\n",
4273 err = check_flow_keys_access(env, off, size);
4274 if (!err && t == BPF_READ && value_regno >= 0)
4275 mark_reg_unknown(env, regs, value_regno);
4276 } else if (type_is_sk_pointer(reg->type)) {
4277 if (t == BPF_WRITE) {
4278 verbose(env, "R%d cannot write into %s\n",
4279 regno, reg_type_str(env, reg->type));
4282 err = check_sock_access(env, insn_idx, regno, off, size, t);
4283 if (!err && value_regno >= 0)
4284 mark_reg_unknown(env, regs, value_regno);
4285 } else if (reg->type == PTR_TO_TP_BUFFER) {
4286 err = check_tp_buffer_access(env, reg, regno, off, size);
4287 if (!err && t == BPF_READ && value_regno >= 0)
4288 mark_reg_unknown(env, regs, value_regno);
4289 } else if (reg->type == PTR_TO_BTF_ID) {
4290 err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
4292 } else if (reg->type == CONST_PTR_TO_MAP) {
4293 err = check_ptr_to_map_access(env, regs, regno, off, size, t,
4295 } else if (base_type(reg->type) == PTR_TO_BUF) {
4296 bool rdonly_mem = type_is_rdonly_mem(reg->type);
4297 const char *buf_info;
4301 if (t == BPF_WRITE) {
4302 verbose(env, "R%d cannot write into %s\n",
4303 regno, reg_type_str(env, reg->type));
4306 buf_info = "rdonly";
4307 max_access = &env->prog->aux->max_rdonly_access;
4310 max_access = &env->prog->aux->max_rdwr_access;
4313 err = check_buffer_access(env, reg, regno, off, size, false,
4314 buf_info, max_access);
4316 if (!err && value_regno >= 0 && (rdonly_mem || t == BPF_READ))
4317 mark_reg_unknown(env, regs, value_regno);
4319 verbose(env, "R%d invalid mem access '%s'\n", regno,
4320 reg_type_str(env, reg->type));
4324 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
4325 regs[value_regno].type == SCALAR_VALUE) {
4326 /* b/h/w load zero-extends, mark upper bits as known 0 */
4327 coerce_reg_to_size(®s[value_regno], size);
4332 static int check_atomic(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
4337 switch (insn->imm) {
4339 case BPF_ADD | BPF_FETCH:
4341 case BPF_AND | BPF_FETCH:
4343 case BPF_OR | BPF_FETCH:
4345 case BPF_XOR | BPF_FETCH:
4350 verbose(env, "BPF_ATOMIC uses invalid atomic opcode %02x\n", insn->imm);
4354 if (BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) {
4355 verbose(env, "invalid atomic operand size\n");
4359 /* check src1 operand */
4360 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4364 /* check src2 operand */
4365 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4369 if (insn->imm == BPF_CMPXCHG) {
4370 /* Check comparison of R0 with memory location */
4371 const u32 aux_reg = BPF_REG_0;
4373 err = check_reg_arg(env, aux_reg, SRC_OP);
4377 if (is_pointer_value(env, aux_reg)) {
4378 verbose(env, "R%d leaks addr into mem\n", aux_reg);
4383 if (is_pointer_value(env, insn->src_reg)) {
4384 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
4388 if (is_ctx_reg(env, insn->dst_reg) ||
4389 is_pkt_reg(env, insn->dst_reg) ||
4390 is_flow_key_reg(env, insn->dst_reg) ||
4391 is_sk_reg(env, insn->dst_reg)) {
4392 verbose(env, "BPF_ATOMIC stores into R%d %s is not allowed\n",
4394 reg_type_str(env, reg_state(env, insn->dst_reg)->type));
4398 if (insn->imm & BPF_FETCH) {
4399 if (insn->imm == BPF_CMPXCHG)
4400 load_reg = BPF_REG_0;
4402 load_reg = insn->src_reg;
4404 /* check and record load of old value */
4405 err = check_reg_arg(env, load_reg, DST_OP);
4409 /* This instruction accesses a memory location but doesn't
4410 * actually load it into a register.
4415 /* Check whether we can read the memory, with second call for fetch
4416 * case to simulate the register fill.
4418 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4419 BPF_SIZE(insn->code), BPF_READ, -1, true);
4420 if (!err && load_reg >= 0)
4421 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4422 BPF_SIZE(insn->code), BPF_READ, load_reg,
4427 /* Check whether we can write into the same memory. */
4428 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4429 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
4436 /* When register 'regno' is used to read the stack (either directly or through
4437 * a helper function) make sure that it's within stack boundary and, depending
4438 * on the access type, that all elements of the stack are initialized.
4440 * 'off' includes 'regno->off', but not its dynamic part (if any).
4442 * All registers that have been spilled on the stack in the slots within the
4443 * read offsets are marked as read.
4445 static int check_stack_range_initialized(
4446 struct bpf_verifier_env *env, int regno, int off,
4447 int access_size, bool zero_size_allowed,
4448 enum stack_access_src type, struct bpf_call_arg_meta *meta)
4450 struct bpf_reg_state *reg = reg_state(env, regno);
4451 struct bpf_func_state *state = func(env, reg);
4452 int err, min_off, max_off, i, j, slot, spi;
4453 char *err_extra = type == ACCESS_HELPER ? " indirect" : "";
4454 enum bpf_access_type bounds_check_type;
4455 /* Some accesses can write anything into the stack, others are
4458 bool clobber = false;
4460 if (access_size == 0 && !zero_size_allowed) {
4461 verbose(env, "invalid zero-sized read\n");
4465 if (type == ACCESS_HELPER) {
4466 /* The bounds checks for writes are more permissive than for
4467 * reads. However, if raw_mode is not set, we'll do extra
4470 bounds_check_type = BPF_WRITE;
4473 bounds_check_type = BPF_READ;
4475 err = check_stack_access_within_bounds(env, regno, off, access_size,
4476 type, bounds_check_type);
4481 if (tnum_is_const(reg->var_off)) {
4482 min_off = max_off = reg->var_off.value + off;
4484 /* Variable offset is prohibited for unprivileged mode for
4485 * simplicity since it requires corresponding support in
4486 * Spectre masking for stack ALU.
4487 * See also retrieve_ptr_limit().
4489 if (!env->bypass_spec_v1) {
4492 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4493 verbose(env, "R%d%s variable offset stack access prohibited for !root, var_off=%s\n",
4494 regno, err_extra, tn_buf);
4497 /* Only initialized buffer on stack is allowed to be accessed
4498 * with variable offset. With uninitialized buffer it's hard to
4499 * guarantee that whole memory is marked as initialized on
4500 * helper return since specific bounds are unknown what may
4501 * cause uninitialized stack leaking.
4503 if (meta && meta->raw_mode)
4506 min_off = reg->smin_value + off;
4507 max_off = reg->smax_value + off;
4510 if (meta && meta->raw_mode) {
4511 meta->access_size = access_size;
4512 meta->regno = regno;
4516 for (i = min_off; i < max_off + access_size; i++) {
4520 spi = slot / BPF_REG_SIZE;
4521 if (state->allocated_stack <= slot)
4523 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
4524 if (*stype == STACK_MISC)
4526 if (*stype == STACK_ZERO) {
4528 /* helper can write anything into the stack */
4529 *stype = STACK_MISC;
4534 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
4535 state->stack[spi].spilled_ptr.type == PTR_TO_BTF_ID)
4538 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
4539 (state->stack[spi].spilled_ptr.type == SCALAR_VALUE ||
4540 env->allow_ptr_leaks)) {
4542 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
4543 for (j = 0; j < BPF_REG_SIZE; j++)
4544 state->stack[spi].slot_type[j] = STACK_MISC;
4550 if (tnum_is_const(reg->var_off)) {
4551 verbose(env, "invalid%s read from stack R%d off %d+%d size %d\n",
4552 err_extra, regno, min_off, i - min_off, access_size);
4556 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4557 verbose(env, "invalid%s read from stack R%d var_off %s+%d size %d\n",
4558 err_extra, regno, tn_buf, i - min_off, access_size);
4562 /* reading any byte out of 8-byte 'spill_slot' will cause
4563 * the whole slot to be marked as 'read'
4565 mark_reg_read(env, &state->stack[spi].spilled_ptr,
4566 state->stack[spi].spilled_ptr.parent,
4569 return update_stack_depth(env, state, min_off);
4572 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
4573 int access_size, bool zero_size_allowed,
4574 struct bpf_call_arg_meta *meta)
4576 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4577 const char *buf_info;
4580 switch (base_type(reg->type)) {
4582 case PTR_TO_PACKET_META:
4583 return check_packet_access(env, regno, reg->off, access_size,
4585 case PTR_TO_MAP_KEY:
4586 if (meta && meta->raw_mode) {
4587 verbose(env, "R%d cannot write into %s\n", regno,
4588 reg_type_str(env, reg->type));
4591 return check_mem_region_access(env, regno, reg->off, access_size,
4592 reg->map_ptr->key_size, false);
4593 case PTR_TO_MAP_VALUE:
4594 if (check_map_access_type(env, regno, reg->off, access_size,
4595 meta && meta->raw_mode ? BPF_WRITE :
4598 return check_map_access(env, regno, reg->off, access_size,
4601 if (type_is_rdonly_mem(reg->type)) {
4602 if (meta && meta->raw_mode) {
4603 verbose(env, "R%d cannot write into %s\n", regno,
4604 reg_type_str(env, reg->type));
4608 return check_mem_region_access(env, regno, reg->off,
4609 access_size, reg->mem_size,
4612 if (type_is_rdonly_mem(reg->type)) {
4613 if (meta && meta->raw_mode) {
4614 verbose(env, "R%d cannot write into %s\n", regno,
4615 reg_type_str(env, reg->type));
4619 buf_info = "rdonly";
4620 max_access = &env->prog->aux->max_rdonly_access;
4623 max_access = &env->prog->aux->max_rdwr_access;
4625 return check_buffer_access(env, reg, regno, reg->off,
4626 access_size, zero_size_allowed,
4627 buf_info, max_access);
4629 return check_stack_range_initialized(
4631 regno, reg->off, access_size,
4632 zero_size_allowed, ACCESS_HELPER, meta);
4633 default: /* scalar_value or invalid ptr */
4634 /* Allow zero-byte read from NULL, regardless of pointer type */
4635 if (zero_size_allowed && access_size == 0 &&
4636 register_is_null(reg))
4639 verbose(env, "R%d type=%s ", regno,
4640 reg_type_str(env, reg->type));
4641 verbose(env, "expected=%s\n", reg_type_str(env, PTR_TO_STACK));
4646 int check_mem_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
4647 u32 regno, u32 mem_size)
4649 if (register_is_null(reg))
4652 if (type_may_be_null(reg->type)) {
4653 /* Assuming that the register contains a value check if the memory
4654 * access is safe. Temporarily save and restore the register's state as
4655 * the conversion shouldn't be visible to a caller.
4657 const struct bpf_reg_state saved_reg = *reg;
4660 mark_ptr_not_null_reg(reg);
4661 rv = check_helper_mem_access(env, regno, mem_size, true, NULL);
4666 return check_helper_mem_access(env, regno, mem_size, true, NULL);
4669 /* Implementation details:
4670 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL
4671 * Two bpf_map_lookups (even with the same key) will have different reg->id.
4672 * For traditional PTR_TO_MAP_VALUE the verifier clears reg->id after
4673 * value_or_null->value transition, since the verifier only cares about
4674 * the range of access to valid map value pointer and doesn't care about actual
4675 * address of the map element.
4676 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
4677 * reg->id > 0 after value_or_null->value transition. By doing so
4678 * two bpf_map_lookups will be considered two different pointers that
4679 * point to different bpf_spin_locks.
4680 * The verifier allows taking only one bpf_spin_lock at a time to avoid
4682 * Since only one bpf_spin_lock is allowed the checks are simpler than
4683 * reg_is_refcounted() logic. The verifier needs to remember only
4684 * one spin_lock instead of array of acquired_refs.
4685 * cur_state->active_spin_lock remembers which map value element got locked
4686 * and clears it after bpf_spin_unlock.
4688 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
4691 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4692 struct bpf_verifier_state *cur = env->cur_state;
4693 bool is_const = tnum_is_const(reg->var_off);
4694 struct bpf_map *map = reg->map_ptr;
4695 u64 val = reg->var_off.value;
4699 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
4705 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
4709 if (!map_value_has_spin_lock(map)) {
4710 if (map->spin_lock_off == -E2BIG)
4712 "map '%s' has more than one 'struct bpf_spin_lock'\n",
4714 else if (map->spin_lock_off == -ENOENT)
4716 "map '%s' doesn't have 'struct bpf_spin_lock'\n",
4720 "map '%s' is not a struct type or bpf_spin_lock is mangled\n",
4724 if (map->spin_lock_off != val + reg->off) {
4725 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock'\n",
4730 if (cur->active_spin_lock) {
4732 "Locking two bpf_spin_locks are not allowed\n");
4735 cur->active_spin_lock = reg->id;
4737 if (!cur->active_spin_lock) {
4738 verbose(env, "bpf_spin_unlock without taking a lock\n");
4741 if (cur->active_spin_lock != reg->id) {
4742 verbose(env, "bpf_spin_unlock of different lock\n");
4745 cur->active_spin_lock = 0;
4750 static int process_timer_func(struct bpf_verifier_env *env, int regno,
4751 struct bpf_call_arg_meta *meta)
4753 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4754 bool is_const = tnum_is_const(reg->var_off);
4755 struct bpf_map *map = reg->map_ptr;
4756 u64 val = reg->var_off.value;
4760 "R%d doesn't have constant offset. bpf_timer has to be at the constant offset\n",
4765 verbose(env, "map '%s' has to have BTF in order to use bpf_timer\n",
4769 if (!map_value_has_timer(map)) {
4770 if (map->timer_off == -E2BIG)
4772 "map '%s' has more than one 'struct bpf_timer'\n",
4774 else if (map->timer_off == -ENOENT)
4776 "map '%s' doesn't have 'struct bpf_timer'\n",
4780 "map '%s' is not a struct type or bpf_timer is mangled\n",
4784 if (map->timer_off != val + reg->off) {
4785 verbose(env, "off %lld doesn't point to 'struct bpf_timer' that is at %d\n",
4786 val + reg->off, map->timer_off);
4789 if (meta->map_ptr) {
4790 verbose(env, "verifier bug. Two map pointers in a timer helper\n");
4793 meta->map_uid = reg->map_uid;
4794 meta->map_ptr = map;
4798 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
4800 return base_type(type) == ARG_PTR_TO_MEM ||
4801 base_type(type) == ARG_PTR_TO_UNINIT_MEM;
4804 static bool arg_type_is_mem_size(enum bpf_arg_type type)
4806 return type == ARG_CONST_SIZE ||
4807 type == ARG_CONST_SIZE_OR_ZERO;
4810 static bool arg_type_is_alloc_size(enum bpf_arg_type type)
4812 return type == ARG_CONST_ALLOC_SIZE_OR_ZERO;
4815 static bool arg_type_is_int_ptr(enum bpf_arg_type type)
4817 return type == ARG_PTR_TO_INT ||
4818 type == ARG_PTR_TO_LONG;
4821 static int int_ptr_type_to_size(enum bpf_arg_type type)
4823 if (type == ARG_PTR_TO_INT)
4825 else if (type == ARG_PTR_TO_LONG)
4831 static int resolve_map_arg_type(struct bpf_verifier_env *env,
4832 const struct bpf_call_arg_meta *meta,
4833 enum bpf_arg_type *arg_type)
4835 if (!meta->map_ptr) {
4836 /* kernel subsystem misconfigured verifier */
4837 verbose(env, "invalid map_ptr to access map->type\n");
4841 switch (meta->map_ptr->map_type) {
4842 case BPF_MAP_TYPE_SOCKMAP:
4843 case BPF_MAP_TYPE_SOCKHASH:
4844 if (*arg_type == ARG_PTR_TO_MAP_VALUE) {
4845 *arg_type = ARG_PTR_TO_BTF_ID_SOCK_COMMON;
4847 verbose(env, "invalid arg_type for sockmap/sockhash\n");
4858 struct bpf_reg_types {
4859 const enum bpf_reg_type types[10];
4863 static const struct bpf_reg_types map_key_value_types = {
4873 static const struct bpf_reg_types sock_types = {
4883 static const struct bpf_reg_types btf_id_sock_common_types = {
4891 .btf_id = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
4895 static const struct bpf_reg_types mem_types = {
4907 static const struct bpf_reg_types int_ptr_types = {
4917 static const struct bpf_reg_types fullsock_types = { .types = { PTR_TO_SOCKET } };
4918 static const struct bpf_reg_types scalar_types = { .types = { SCALAR_VALUE } };
4919 static const struct bpf_reg_types context_types = { .types = { PTR_TO_CTX } };
4920 static const struct bpf_reg_types alloc_mem_types = { .types = { PTR_TO_MEM } };
4921 static const struct bpf_reg_types const_map_ptr_types = { .types = { CONST_PTR_TO_MAP } };
4922 static const struct bpf_reg_types btf_ptr_types = { .types = { PTR_TO_BTF_ID } };
4923 static const struct bpf_reg_types spin_lock_types = { .types = { PTR_TO_MAP_VALUE } };
4924 static const struct bpf_reg_types percpu_btf_ptr_types = { .types = { PTR_TO_PERCPU_BTF_ID } };
4925 static const struct bpf_reg_types func_ptr_types = { .types = { PTR_TO_FUNC } };
4926 static const struct bpf_reg_types stack_ptr_types = { .types = { PTR_TO_STACK } };
4927 static const struct bpf_reg_types const_str_ptr_types = { .types = { PTR_TO_MAP_VALUE } };
4928 static const struct bpf_reg_types timer_types = { .types = { PTR_TO_MAP_VALUE } };
4930 static const struct bpf_reg_types *compatible_reg_types[__BPF_ARG_TYPE_MAX] = {
4931 [ARG_PTR_TO_MAP_KEY] = &map_key_value_types,
4932 [ARG_PTR_TO_MAP_VALUE] = &map_key_value_types,
4933 [ARG_PTR_TO_UNINIT_MAP_VALUE] = &map_key_value_types,
4934 [ARG_CONST_SIZE] = &scalar_types,
4935 [ARG_CONST_SIZE_OR_ZERO] = &scalar_types,
4936 [ARG_CONST_ALLOC_SIZE_OR_ZERO] = &scalar_types,
4937 [ARG_CONST_MAP_PTR] = &const_map_ptr_types,
4938 [ARG_PTR_TO_CTX] = &context_types,
4939 [ARG_PTR_TO_SOCK_COMMON] = &sock_types,
4941 [ARG_PTR_TO_BTF_ID_SOCK_COMMON] = &btf_id_sock_common_types,
4943 [ARG_PTR_TO_SOCKET] = &fullsock_types,
4944 [ARG_PTR_TO_BTF_ID] = &btf_ptr_types,
4945 [ARG_PTR_TO_SPIN_LOCK] = &spin_lock_types,
4946 [ARG_PTR_TO_MEM] = &mem_types,
4947 [ARG_PTR_TO_UNINIT_MEM] = &mem_types,
4948 [ARG_PTR_TO_ALLOC_MEM] = &alloc_mem_types,
4949 [ARG_PTR_TO_INT] = &int_ptr_types,
4950 [ARG_PTR_TO_LONG] = &int_ptr_types,
4951 [ARG_PTR_TO_PERCPU_BTF_ID] = &percpu_btf_ptr_types,
4952 [ARG_PTR_TO_FUNC] = &func_ptr_types,
4953 [ARG_PTR_TO_STACK] = &stack_ptr_types,
4954 [ARG_PTR_TO_CONST_STR] = &const_str_ptr_types,
4955 [ARG_PTR_TO_TIMER] = &timer_types,
4958 static int check_reg_type(struct bpf_verifier_env *env, u32 regno,
4959 enum bpf_arg_type arg_type,
4960 const u32 *arg_btf_id)
4962 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4963 enum bpf_reg_type expected, type = reg->type;
4964 const struct bpf_reg_types *compatible;
4967 compatible = compatible_reg_types[base_type(arg_type)];
4969 verbose(env, "verifier internal error: unsupported arg type %d\n", arg_type);
4973 /* ARG_PTR_TO_MEM + RDONLY is compatible with PTR_TO_MEM and PTR_TO_MEM + RDONLY,
4974 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM and NOT with PTR_TO_MEM + RDONLY
4976 * Same for MAYBE_NULL:
4978 * ARG_PTR_TO_MEM + MAYBE_NULL is compatible with PTR_TO_MEM and PTR_TO_MEM + MAYBE_NULL,
4979 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM but NOT with PTR_TO_MEM + MAYBE_NULL
4981 * Therefore we fold these flags depending on the arg_type before comparison.
4983 if (arg_type & MEM_RDONLY)
4984 type &= ~MEM_RDONLY;
4985 if (arg_type & PTR_MAYBE_NULL)
4986 type &= ~PTR_MAYBE_NULL;
4988 for (i = 0; i < ARRAY_SIZE(compatible->types); i++) {
4989 expected = compatible->types[i];
4990 if (expected == NOT_INIT)
4993 if (type == expected)
4997 verbose(env, "R%d type=%s expected=", regno, reg_type_str(env, reg->type));
4998 for (j = 0; j + 1 < i; j++)
4999 verbose(env, "%s, ", reg_type_str(env, compatible->types[j]));
5000 verbose(env, "%s\n", reg_type_str(env, compatible->types[j]));
5004 if (reg->type == PTR_TO_BTF_ID) {
5006 if (!compatible->btf_id) {
5007 verbose(env, "verifier internal error: missing arg compatible BTF ID\n");
5010 arg_btf_id = compatible->btf_id;
5013 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
5014 btf_vmlinux, *arg_btf_id)) {
5015 verbose(env, "R%d is of type %s but %s is expected\n",
5016 regno, kernel_type_name(reg->btf, reg->btf_id),
5017 kernel_type_name(btf_vmlinux, *arg_btf_id));
5021 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
5022 verbose(env, "R%d is a pointer to in-kernel struct with non-zero offset\n",
5031 static int check_func_arg(struct bpf_verifier_env *env, u32 arg,
5032 struct bpf_call_arg_meta *meta,
5033 const struct bpf_func_proto *fn)
5035 u32 regno = BPF_REG_1 + arg;
5036 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
5037 enum bpf_arg_type arg_type = fn->arg_type[arg];
5038 enum bpf_reg_type type = reg->type;
5041 if (arg_type == ARG_DONTCARE)
5044 err = check_reg_arg(env, regno, SRC_OP);
5048 if (arg_type == ARG_ANYTHING) {
5049 if (is_pointer_value(env, regno)) {
5050 verbose(env, "R%d leaks addr into helper function\n",
5057 if (type_is_pkt_pointer(type) &&
5058 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
5059 verbose(env, "helper access to the packet is not allowed\n");
5063 if (base_type(arg_type) == ARG_PTR_TO_MAP_VALUE ||
5064 base_type(arg_type) == ARG_PTR_TO_UNINIT_MAP_VALUE) {
5065 err = resolve_map_arg_type(env, meta, &arg_type);
5070 if (register_is_null(reg) && type_may_be_null(arg_type))
5071 /* A NULL register has a SCALAR_VALUE type, so skip
5074 goto skip_type_check;
5076 err = check_reg_type(env, regno, arg_type, fn->arg_btf_id[arg]);
5080 if (type == PTR_TO_CTX) {
5081 err = check_ctx_reg(env, reg, regno);
5087 if (reg->ref_obj_id) {
5088 if (meta->ref_obj_id) {
5089 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
5090 regno, reg->ref_obj_id,
5094 meta->ref_obj_id = reg->ref_obj_id;
5097 if (arg_type == ARG_CONST_MAP_PTR) {
5098 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
5099 if (meta->map_ptr) {
5100 /* Use map_uid (which is unique id of inner map) to reject:
5101 * inner_map1 = bpf_map_lookup_elem(outer_map, key1)
5102 * inner_map2 = bpf_map_lookup_elem(outer_map, key2)
5103 * if (inner_map1 && inner_map2) {
5104 * timer = bpf_map_lookup_elem(inner_map1);
5106 * // mismatch would have been allowed
5107 * bpf_timer_init(timer, inner_map2);
5110 * Comparing map_ptr is enough to distinguish normal and outer maps.
5112 if (meta->map_ptr != reg->map_ptr ||
5113 meta->map_uid != reg->map_uid) {
5115 "timer pointer in R1 map_uid=%d doesn't match map pointer in R2 map_uid=%d\n",
5116 meta->map_uid, reg->map_uid);
5120 meta->map_ptr = reg->map_ptr;
5121 meta->map_uid = reg->map_uid;
5122 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
5123 /* bpf_map_xxx(..., map_ptr, ..., key) call:
5124 * check that [key, key + map->key_size) are within
5125 * stack limits and initialized
5127 if (!meta->map_ptr) {
5128 /* in function declaration map_ptr must come before
5129 * map_key, so that it's verified and known before
5130 * we have to check map_key here. Otherwise it means
5131 * that kernel subsystem misconfigured verifier
5133 verbose(env, "invalid map_ptr to access map->key\n");
5136 err = check_helper_mem_access(env, regno,
5137 meta->map_ptr->key_size, false,
5139 } else if (base_type(arg_type) == ARG_PTR_TO_MAP_VALUE ||
5140 base_type(arg_type) == ARG_PTR_TO_UNINIT_MAP_VALUE) {
5141 if (type_may_be_null(arg_type) && register_is_null(reg))
5144 /* bpf_map_xxx(..., map_ptr, ..., value) call:
5145 * check [value, value + map->value_size) validity
5147 if (!meta->map_ptr) {
5148 /* kernel subsystem misconfigured verifier */
5149 verbose(env, "invalid map_ptr to access map->value\n");
5152 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE);
5153 err = check_helper_mem_access(env, regno,
5154 meta->map_ptr->value_size, false,
5156 } else if (arg_type == ARG_PTR_TO_PERCPU_BTF_ID) {
5158 verbose(env, "Helper has invalid btf_id in R%d\n", regno);
5161 meta->ret_btf = reg->btf;
5162 meta->ret_btf_id = reg->btf_id;
5163 } else if (arg_type == ARG_PTR_TO_SPIN_LOCK) {
5164 if (meta->func_id == BPF_FUNC_spin_lock) {
5165 if (process_spin_lock(env, regno, true))
5167 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
5168 if (process_spin_lock(env, regno, false))
5171 verbose(env, "verifier internal error\n");
5174 } else if (arg_type == ARG_PTR_TO_TIMER) {
5175 if (process_timer_func(env, regno, meta))
5177 } else if (arg_type == ARG_PTR_TO_FUNC) {
5178 meta->subprogno = reg->subprogno;
5179 } else if (arg_type_is_mem_ptr(arg_type)) {
5180 /* The access to this pointer is only checked when we hit the
5181 * next is_mem_size argument below.
5183 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MEM);
5184 } else if (arg_type_is_mem_size(arg_type)) {
5185 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
5187 /* This is used to refine r0 return value bounds for helpers
5188 * that enforce this value as an upper bound on return values.
5189 * See do_refine_retval_range() for helpers that can refine
5190 * the return value. C type of helper is u32 so we pull register
5191 * bound from umax_value however, if negative verifier errors
5192 * out. Only upper bounds can be learned because retval is an
5193 * int type and negative retvals are allowed.
5195 meta->msize_max_value = reg->umax_value;
5197 /* The register is SCALAR_VALUE; the access check
5198 * happens using its boundaries.
5200 if (!tnum_is_const(reg->var_off))
5201 /* For unprivileged variable accesses, disable raw
5202 * mode so that the program is required to
5203 * initialize all the memory that the helper could
5204 * just partially fill up.
5208 if (reg->smin_value < 0) {
5209 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
5214 if (reg->umin_value == 0) {
5215 err = check_helper_mem_access(env, regno - 1, 0,
5222 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
5223 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
5227 err = check_helper_mem_access(env, regno - 1,
5229 zero_size_allowed, meta);
5231 err = mark_chain_precision(env, regno);
5232 } else if (arg_type_is_alloc_size(arg_type)) {
5233 if (!tnum_is_const(reg->var_off)) {
5234 verbose(env, "R%d is not a known constant'\n",
5238 meta->mem_size = reg->var_off.value;
5239 } else if (arg_type_is_int_ptr(arg_type)) {
5240 int size = int_ptr_type_to_size(arg_type);
5242 err = check_helper_mem_access(env, regno, size, false, meta);
5245 err = check_ptr_alignment(env, reg, 0, size, true);
5246 } else if (arg_type == ARG_PTR_TO_CONST_STR) {
5247 struct bpf_map *map = reg->map_ptr;
5252 if (!bpf_map_is_rdonly(map)) {
5253 verbose(env, "R%d does not point to a readonly map'\n", regno);
5257 if (!tnum_is_const(reg->var_off)) {
5258 verbose(env, "R%d is not a constant address'\n", regno);
5262 if (!map->ops->map_direct_value_addr) {
5263 verbose(env, "no direct value access support for this map type\n");
5267 err = check_map_access(env, regno, reg->off,
5268 map->value_size - reg->off, false);
5272 map_off = reg->off + reg->var_off.value;
5273 err = map->ops->map_direct_value_addr(map, &map_addr, map_off);
5275 verbose(env, "direct value access on string failed\n");
5279 str_ptr = (char *)(long)(map_addr);
5280 if (!strnchr(str_ptr + map_off, map->value_size - map_off, 0)) {
5281 verbose(env, "string is not zero-terminated\n");
5289 static bool may_update_sockmap(struct bpf_verifier_env *env, int func_id)
5291 enum bpf_attach_type eatype = env->prog->expected_attach_type;
5292 enum bpf_prog_type type = resolve_prog_type(env->prog);
5294 if (func_id != BPF_FUNC_map_update_elem)
5297 /* It's not possible to get access to a locked struct sock in these
5298 * contexts, so updating is safe.
5301 case BPF_PROG_TYPE_TRACING:
5302 if (eatype == BPF_TRACE_ITER)
5305 case BPF_PROG_TYPE_SOCKET_FILTER:
5306 case BPF_PROG_TYPE_SCHED_CLS:
5307 case BPF_PROG_TYPE_SCHED_ACT:
5308 case BPF_PROG_TYPE_XDP:
5309 case BPF_PROG_TYPE_SK_REUSEPORT:
5310 case BPF_PROG_TYPE_FLOW_DISSECTOR:
5311 case BPF_PROG_TYPE_SK_LOOKUP:
5317 verbose(env, "cannot update sockmap in this context\n");
5321 static bool allow_tail_call_in_subprogs(struct bpf_verifier_env *env)
5323 return env->prog->jit_requested && IS_ENABLED(CONFIG_X86_64);
5326 static int check_map_func_compatibility(struct bpf_verifier_env *env,
5327 struct bpf_map *map, int func_id)
5332 /* We need a two way check, first is from map perspective ... */
5333 switch (map->map_type) {
5334 case BPF_MAP_TYPE_PROG_ARRAY:
5335 if (func_id != BPF_FUNC_tail_call)
5338 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
5339 if (func_id != BPF_FUNC_perf_event_read &&
5340 func_id != BPF_FUNC_perf_event_output &&
5341 func_id != BPF_FUNC_skb_output &&
5342 func_id != BPF_FUNC_perf_event_read_value &&
5343 func_id != BPF_FUNC_xdp_output)
5346 case BPF_MAP_TYPE_RINGBUF:
5347 if (func_id != BPF_FUNC_ringbuf_output &&
5348 func_id != BPF_FUNC_ringbuf_reserve &&
5349 func_id != BPF_FUNC_ringbuf_query)
5352 case BPF_MAP_TYPE_STACK_TRACE:
5353 if (func_id != BPF_FUNC_get_stackid)
5356 case BPF_MAP_TYPE_CGROUP_ARRAY:
5357 if (func_id != BPF_FUNC_skb_under_cgroup &&
5358 func_id != BPF_FUNC_current_task_under_cgroup)
5361 case BPF_MAP_TYPE_CGROUP_STORAGE:
5362 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
5363 if (func_id != BPF_FUNC_get_local_storage)
5366 case BPF_MAP_TYPE_DEVMAP:
5367 case BPF_MAP_TYPE_DEVMAP_HASH:
5368 if (func_id != BPF_FUNC_redirect_map &&
5369 func_id != BPF_FUNC_map_lookup_elem)
5372 /* Restrict bpf side of cpumap and xskmap, open when use-cases
5375 case BPF_MAP_TYPE_CPUMAP:
5376 if (func_id != BPF_FUNC_redirect_map)
5379 case BPF_MAP_TYPE_XSKMAP:
5380 if (func_id != BPF_FUNC_redirect_map &&
5381 func_id != BPF_FUNC_map_lookup_elem)
5384 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
5385 case BPF_MAP_TYPE_HASH_OF_MAPS:
5386 if (func_id != BPF_FUNC_map_lookup_elem)
5389 case BPF_MAP_TYPE_SOCKMAP:
5390 if (func_id != BPF_FUNC_sk_redirect_map &&
5391 func_id != BPF_FUNC_sock_map_update &&
5392 func_id != BPF_FUNC_map_delete_elem &&
5393 func_id != BPF_FUNC_msg_redirect_map &&
5394 func_id != BPF_FUNC_sk_select_reuseport &&
5395 func_id != BPF_FUNC_map_lookup_elem &&
5396 !may_update_sockmap(env, func_id))
5399 case BPF_MAP_TYPE_SOCKHASH:
5400 if (func_id != BPF_FUNC_sk_redirect_hash &&
5401 func_id != BPF_FUNC_sock_hash_update &&
5402 func_id != BPF_FUNC_map_delete_elem &&
5403 func_id != BPF_FUNC_msg_redirect_hash &&
5404 func_id != BPF_FUNC_sk_select_reuseport &&
5405 func_id != BPF_FUNC_map_lookup_elem &&
5406 !may_update_sockmap(env, func_id))
5409 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
5410 if (func_id != BPF_FUNC_sk_select_reuseport)
5413 case BPF_MAP_TYPE_QUEUE:
5414 case BPF_MAP_TYPE_STACK:
5415 if (func_id != BPF_FUNC_map_peek_elem &&
5416 func_id != BPF_FUNC_map_pop_elem &&
5417 func_id != BPF_FUNC_map_push_elem)
5420 case BPF_MAP_TYPE_SK_STORAGE:
5421 if (func_id != BPF_FUNC_sk_storage_get &&
5422 func_id != BPF_FUNC_sk_storage_delete)
5425 case BPF_MAP_TYPE_INODE_STORAGE:
5426 if (func_id != BPF_FUNC_inode_storage_get &&
5427 func_id != BPF_FUNC_inode_storage_delete)
5430 case BPF_MAP_TYPE_TASK_STORAGE:
5431 if (func_id != BPF_FUNC_task_storage_get &&
5432 func_id != BPF_FUNC_task_storage_delete)
5439 /* ... and second from the function itself. */
5441 case BPF_FUNC_tail_call:
5442 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
5444 if (env->subprog_cnt > 1 && !allow_tail_call_in_subprogs(env)) {
5445 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
5449 case BPF_FUNC_perf_event_read:
5450 case BPF_FUNC_perf_event_output:
5451 case BPF_FUNC_perf_event_read_value:
5452 case BPF_FUNC_skb_output:
5453 case BPF_FUNC_xdp_output:
5454 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
5457 case BPF_FUNC_ringbuf_output:
5458 case BPF_FUNC_ringbuf_reserve:
5459 case BPF_FUNC_ringbuf_query:
5460 if (map->map_type != BPF_MAP_TYPE_RINGBUF)
5463 case BPF_FUNC_get_stackid:
5464 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
5467 case BPF_FUNC_current_task_under_cgroup:
5468 case BPF_FUNC_skb_under_cgroup:
5469 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
5472 case BPF_FUNC_redirect_map:
5473 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
5474 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
5475 map->map_type != BPF_MAP_TYPE_CPUMAP &&
5476 map->map_type != BPF_MAP_TYPE_XSKMAP)
5479 case BPF_FUNC_sk_redirect_map:
5480 case BPF_FUNC_msg_redirect_map:
5481 case BPF_FUNC_sock_map_update:
5482 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
5485 case BPF_FUNC_sk_redirect_hash:
5486 case BPF_FUNC_msg_redirect_hash:
5487 case BPF_FUNC_sock_hash_update:
5488 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
5491 case BPF_FUNC_get_local_storage:
5492 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
5493 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
5496 case BPF_FUNC_sk_select_reuseport:
5497 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
5498 map->map_type != BPF_MAP_TYPE_SOCKMAP &&
5499 map->map_type != BPF_MAP_TYPE_SOCKHASH)
5502 case BPF_FUNC_map_peek_elem:
5503 case BPF_FUNC_map_pop_elem:
5504 case BPF_FUNC_map_push_elem:
5505 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
5506 map->map_type != BPF_MAP_TYPE_STACK)
5509 case BPF_FUNC_sk_storage_get:
5510 case BPF_FUNC_sk_storage_delete:
5511 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
5514 case BPF_FUNC_inode_storage_get:
5515 case BPF_FUNC_inode_storage_delete:
5516 if (map->map_type != BPF_MAP_TYPE_INODE_STORAGE)
5519 case BPF_FUNC_task_storage_get:
5520 case BPF_FUNC_task_storage_delete:
5521 if (map->map_type != BPF_MAP_TYPE_TASK_STORAGE)
5530 verbose(env, "cannot pass map_type %d into func %s#%d\n",
5531 map->map_type, func_id_name(func_id), func_id);
5535 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
5539 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
5541 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
5543 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
5545 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
5547 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
5550 /* We only support one arg being in raw mode at the moment,
5551 * which is sufficient for the helper functions we have
5557 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
5558 enum bpf_arg_type arg_next)
5560 return (arg_type_is_mem_ptr(arg_curr) &&
5561 !arg_type_is_mem_size(arg_next)) ||
5562 (!arg_type_is_mem_ptr(arg_curr) &&
5563 arg_type_is_mem_size(arg_next));
5566 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
5568 /* bpf_xxx(..., buf, len) call will access 'len'
5569 * bytes from memory 'buf'. Both arg types need
5570 * to be paired, so make sure there's no buggy
5571 * helper function specification.
5573 if (arg_type_is_mem_size(fn->arg1_type) ||
5574 arg_type_is_mem_ptr(fn->arg5_type) ||
5575 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
5576 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
5577 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
5578 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
5584 static bool check_refcount_ok(const struct bpf_func_proto *fn, int func_id)
5588 if (arg_type_may_be_refcounted(fn->arg1_type))
5590 if (arg_type_may_be_refcounted(fn->arg2_type))
5592 if (arg_type_may_be_refcounted(fn->arg3_type))
5594 if (arg_type_may_be_refcounted(fn->arg4_type))
5596 if (arg_type_may_be_refcounted(fn->arg5_type))
5599 /* A reference acquiring function cannot acquire
5600 * another refcounted ptr.
5602 if (may_be_acquire_function(func_id) && count)
5605 /* We only support one arg being unreferenced at the moment,
5606 * which is sufficient for the helper functions we have right now.
5611 static bool check_btf_id_ok(const struct bpf_func_proto *fn)
5615 for (i = 0; i < ARRAY_SIZE(fn->arg_type); i++) {
5616 if (fn->arg_type[i] == ARG_PTR_TO_BTF_ID && !fn->arg_btf_id[i])
5619 if (fn->arg_type[i] != ARG_PTR_TO_BTF_ID && fn->arg_btf_id[i])
5626 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
5628 return check_raw_mode_ok(fn) &&
5629 check_arg_pair_ok(fn) &&
5630 check_btf_id_ok(fn) &&
5631 check_refcount_ok(fn, func_id) ? 0 : -EINVAL;
5634 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
5635 * are now invalid, so turn them into unknown SCALAR_VALUE.
5637 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
5639 struct bpf_func_state *state;
5640 struct bpf_reg_state *reg;
5642 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
5643 if (reg_is_pkt_pointer_any(reg))
5644 __mark_reg_unknown(env, reg);
5650 BEYOND_PKT_END = -2,
5653 static void mark_pkt_end(struct bpf_verifier_state *vstate, int regn, bool range_open)
5655 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5656 struct bpf_reg_state *reg = &state->regs[regn];
5658 if (reg->type != PTR_TO_PACKET)
5659 /* PTR_TO_PACKET_META is not supported yet */
5662 /* The 'reg' is pkt > pkt_end or pkt >= pkt_end.
5663 * How far beyond pkt_end it goes is unknown.
5664 * if (!range_open) it's the case of pkt >= pkt_end
5665 * if (range_open) it's the case of pkt > pkt_end
5666 * hence this pointer is at least 1 byte bigger than pkt_end
5669 reg->range = BEYOND_PKT_END;
5671 reg->range = AT_PKT_END;
5674 /* The pointer with the specified id has released its reference to kernel
5675 * resources. Identify all copies of the same pointer and clear the reference.
5677 static int release_reference(struct bpf_verifier_env *env,
5680 struct bpf_func_state *state;
5681 struct bpf_reg_state *reg;
5684 err = release_reference_state(cur_func(env), ref_obj_id);
5688 bpf_for_each_reg_in_vstate(env->cur_state, state, reg, ({
5689 if (reg->ref_obj_id == ref_obj_id) {
5690 if (!env->allow_ptr_leaks)
5691 __mark_reg_not_init(env, reg);
5693 __mark_reg_unknown(env, reg);
5700 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
5701 struct bpf_reg_state *regs)
5705 /* after the call registers r0 - r5 were scratched */
5706 for (i = 0; i < CALLER_SAVED_REGS; i++) {
5707 mark_reg_not_init(env, regs, caller_saved[i]);
5708 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
5712 typedef int (*set_callee_state_fn)(struct bpf_verifier_env *env,
5713 struct bpf_func_state *caller,
5714 struct bpf_func_state *callee,
5717 static int __check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
5718 int *insn_idx, int subprog,
5719 set_callee_state_fn set_callee_state_cb)
5721 struct bpf_verifier_state *state = env->cur_state;
5722 struct bpf_func_info_aux *func_info_aux;
5723 struct bpf_func_state *caller, *callee;
5725 bool is_global = false;
5727 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
5728 verbose(env, "the call stack of %d frames is too deep\n",
5729 state->curframe + 2);
5733 caller = state->frame[state->curframe];
5734 if (state->frame[state->curframe + 1]) {
5735 verbose(env, "verifier bug. Frame %d already allocated\n",
5736 state->curframe + 1);
5740 func_info_aux = env->prog->aux->func_info_aux;
5742 is_global = func_info_aux[subprog].linkage == BTF_FUNC_GLOBAL;
5743 err = btf_check_subprog_arg_match(env, subprog, caller->regs);
5748 verbose(env, "Caller passes invalid args into func#%d\n",
5752 if (env->log.level & BPF_LOG_LEVEL)
5754 "Func#%d is global and valid. Skipping.\n",
5756 clear_caller_saved_regs(env, caller->regs);
5758 /* All global functions return a 64-bit SCALAR_VALUE */
5759 mark_reg_unknown(env, caller->regs, BPF_REG_0);
5760 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
5762 /* continue with next insn after call */
5767 if (insn->code == (BPF_JMP | BPF_CALL) &&
5768 insn->src_reg == 0 &&
5769 insn->imm == BPF_FUNC_timer_set_callback) {
5770 struct bpf_verifier_state *async_cb;
5772 /* there is no real recursion here. timer callbacks are async */
5773 env->subprog_info[subprog].is_async_cb = true;
5774 async_cb = push_async_cb(env, env->subprog_info[subprog].start,
5775 *insn_idx, subprog);
5778 callee = async_cb->frame[0];
5779 callee->async_entry_cnt = caller->async_entry_cnt + 1;
5781 /* Convert bpf_timer_set_callback() args into timer callback args */
5782 err = set_callee_state_cb(env, caller, callee, *insn_idx);
5786 clear_caller_saved_regs(env, caller->regs);
5787 mark_reg_unknown(env, caller->regs, BPF_REG_0);
5788 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
5789 /* continue with next insn after call */
5793 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
5796 state->frame[state->curframe + 1] = callee;
5798 /* callee cannot access r0, r6 - r9 for reading and has to write
5799 * into its own stack before reading from it.
5800 * callee can read/write into caller's stack
5802 init_func_state(env, callee,
5803 /* remember the callsite, it will be used by bpf_exit */
5804 *insn_idx /* callsite */,
5805 state->curframe + 1 /* frameno within this callchain */,
5806 subprog /* subprog number within this prog */);
5808 /* Transfer references to the callee */
5809 err = copy_reference_state(callee, caller);
5813 err = set_callee_state_cb(env, caller, callee, *insn_idx);
5817 clear_caller_saved_regs(env, caller->regs);
5819 /* only increment it after check_reg_arg() finished */
5822 /* and go analyze first insn of the callee */
5823 *insn_idx = env->subprog_info[subprog].start - 1;
5825 if (env->log.level & BPF_LOG_LEVEL) {
5826 verbose(env, "caller:\n");
5827 print_verifier_state(env, caller);
5828 verbose(env, "callee:\n");
5829 print_verifier_state(env, callee);
5834 int map_set_for_each_callback_args(struct bpf_verifier_env *env,
5835 struct bpf_func_state *caller,
5836 struct bpf_func_state *callee)
5838 /* bpf_for_each_map_elem(struct bpf_map *map, void *callback_fn,
5839 * void *callback_ctx, u64 flags);
5840 * callback_fn(struct bpf_map *map, void *key, void *value,
5841 * void *callback_ctx);
5843 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
5845 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
5846 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
5847 callee->regs[BPF_REG_2].map_ptr = caller->regs[BPF_REG_1].map_ptr;
5849 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
5850 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
5851 callee->regs[BPF_REG_3].map_ptr = caller->regs[BPF_REG_1].map_ptr;
5853 /* pointer to stack or null */
5854 callee->regs[BPF_REG_4] = caller->regs[BPF_REG_3];
5857 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
5861 static int set_callee_state(struct bpf_verifier_env *env,
5862 struct bpf_func_state *caller,
5863 struct bpf_func_state *callee, int insn_idx)
5867 /* copy r1 - r5 args that callee can access. The copy includes parent
5868 * pointers, which connects us up to the liveness chain
5870 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
5871 callee->regs[i] = caller->regs[i];
5875 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
5878 int subprog, target_insn;
5880 target_insn = *insn_idx + insn->imm + 1;
5881 subprog = find_subprog(env, target_insn);
5883 verbose(env, "verifier bug. No program starts at insn %d\n",
5888 return __check_func_call(env, insn, insn_idx, subprog, set_callee_state);
5891 static int set_map_elem_callback_state(struct bpf_verifier_env *env,
5892 struct bpf_func_state *caller,
5893 struct bpf_func_state *callee,
5896 struct bpf_insn_aux_data *insn_aux = &env->insn_aux_data[insn_idx];
5897 struct bpf_map *map;
5900 if (bpf_map_ptr_poisoned(insn_aux)) {
5901 verbose(env, "tail_call abusing map_ptr\n");
5905 map = BPF_MAP_PTR(insn_aux->map_ptr_state);
5906 if (!map->ops->map_set_for_each_callback_args ||
5907 !map->ops->map_for_each_callback) {
5908 verbose(env, "callback function not allowed for map\n");
5912 err = map->ops->map_set_for_each_callback_args(env, caller, callee);
5916 callee->in_callback_fn = true;
5920 static int set_timer_callback_state(struct bpf_verifier_env *env,
5921 struct bpf_func_state *caller,
5922 struct bpf_func_state *callee,
5925 struct bpf_map *map_ptr = caller->regs[BPF_REG_1].map_ptr;
5927 /* bpf_timer_set_callback(struct bpf_timer *timer, void *callback_fn);
5928 * callback_fn(struct bpf_map *map, void *key, void *value);
5930 callee->regs[BPF_REG_1].type = CONST_PTR_TO_MAP;
5931 __mark_reg_known_zero(&callee->regs[BPF_REG_1]);
5932 callee->regs[BPF_REG_1].map_ptr = map_ptr;
5934 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
5935 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
5936 callee->regs[BPF_REG_2].map_ptr = map_ptr;
5938 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
5939 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
5940 callee->regs[BPF_REG_3].map_ptr = map_ptr;
5943 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
5944 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
5945 callee->in_async_callback_fn = true;
5949 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
5951 struct bpf_verifier_state *state = env->cur_state;
5952 struct bpf_func_state *caller, *callee;
5953 struct bpf_reg_state *r0;
5956 callee = state->frame[state->curframe];
5957 r0 = &callee->regs[BPF_REG_0];
5958 if (r0->type == PTR_TO_STACK) {
5959 /* technically it's ok to return caller's stack pointer
5960 * (or caller's caller's pointer) back to the caller,
5961 * since these pointers are valid. Only current stack
5962 * pointer will be invalid as soon as function exits,
5963 * but let's be conservative
5965 verbose(env, "cannot return stack pointer to the caller\n");
5970 caller = state->frame[state->curframe];
5971 if (callee->in_callback_fn) {
5972 /* enforce R0 return value range [0, 1]. */
5973 struct tnum range = tnum_range(0, 1);
5975 if (r0->type != SCALAR_VALUE) {
5976 verbose(env, "R0 not a scalar value\n");
5979 if (!tnum_in(range, r0->var_off)) {
5980 verbose_invalid_scalar(env, r0, &range, "callback return", "R0");
5984 /* return to the caller whatever r0 had in the callee */
5985 caller->regs[BPF_REG_0] = *r0;
5988 /* callback_fn frame should have released its own additions to parent's
5989 * reference state at this point, or check_reference_leak would
5990 * complain, hence it must be the same as the caller. There is no need
5993 if (!callee->in_callback_fn) {
5994 /* Transfer references to the caller */
5995 err = copy_reference_state(caller, callee);
6000 *insn_idx = callee->callsite + 1;
6001 if (env->log.level & BPF_LOG_LEVEL) {
6002 verbose(env, "returning from callee:\n");
6003 print_verifier_state(env, callee);
6004 verbose(env, "to caller at %d:\n", *insn_idx);
6005 print_verifier_state(env, caller);
6007 /* clear everything in the callee */
6008 free_func_state(callee);
6009 state->frame[state->curframe + 1] = NULL;
6013 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
6015 struct bpf_call_arg_meta *meta)
6017 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
6019 if (ret_type != RET_INTEGER ||
6020 (func_id != BPF_FUNC_get_stack &&
6021 func_id != BPF_FUNC_get_task_stack &&
6022 func_id != BPF_FUNC_probe_read_str &&
6023 func_id != BPF_FUNC_probe_read_kernel_str &&
6024 func_id != BPF_FUNC_probe_read_user_str))
6027 ret_reg->smax_value = meta->msize_max_value;
6028 ret_reg->s32_max_value = meta->msize_max_value;
6029 ret_reg->smin_value = -MAX_ERRNO;
6030 ret_reg->s32_min_value = -MAX_ERRNO;
6031 reg_bounds_sync(ret_reg);
6035 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
6036 int func_id, int insn_idx)
6038 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
6039 struct bpf_map *map = meta->map_ptr;
6041 if (func_id != BPF_FUNC_tail_call &&
6042 func_id != BPF_FUNC_map_lookup_elem &&
6043 func_id != BPF_FUNC_map_update_elem &&
6044 func_id != BPF_FUNC_map_delete_elem &&
6045 func_id != BPF_FUNC_map_push_elem &&
6046 func_id != BPF_FUNC_map_pop_elem &&
6047 func_id != BPF_FUNC_map_peek_elem &&
6048 func_id != BPF_FUNC_for_each_map_elem &&
6049 func_id != BPF_FUNC_redirect_map)
6053 verbose(env, "kernel subsystem misconfigured verifier\n");
6057 /* In case of read-only, some additional restrictions
6058 * need to be applied in order to prevent altering the
6059 * state of the map from program side.
6061 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
6062 (func_id == BPF_FUNC_map_delete_elem ||
6063 func_id == BPF_FUNC_map_update_elem ||
6064 func_id == BPF_FUNC_map_push_elem ||
6065 func_id == BPF_FUNC_map_pop_elem)) {
6066 verbose(env, "write into map forbidden\n");
6070 if (!BPF_MAP_PTR(aux->map_ptr_state))
6071 bpf_map_ptr_store(aux, meta->map_ptr,
6072 !meta->map_ptr->bypass_spec_v1);
6073 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
6074 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
6075 !meta->map_ptr->bypass_spec_v1);
6080 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
6081 int func_id, int insn_idx)
6083 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
6084 struct bpf_reg_state *regs = cur_regs(env), *reg;
6085 struct bpf_map *map = meta->map_ptr;
6089 if (func_id != BPF_FUNC_tail_call)
6091 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
6092 verbose(env, "kernel subsystem misconfigured verifier\n");
6096 reg = ®s[BPF_REG_3];
6097 val = reg->var_off.value;
6098 max = map->max_entries;
6100 if (!(register_is_const(reg) && val < max)) {
6101 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
6105 err = mark_chain_precision(env, BPF_REG_3);
6108 if (bpf_map_key_unseen(aux))
6109 bpf_map_key_store(aux, val);
6110 else if (!bpf_map_key_poisoned(aux) &&
6111 bpf_map_key_immediate(aux) != val)
6112 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
6116 static int check_reference_leak(struct bpf_verifier_env *env)
6118 struct bpf_func_state *state = cur_func(env);
6119 bool refs_lingering = false;
6122 if (state->frameno && !state->in_callback_fn)
6125 for (i = 0; i < state->acquired_refs; i++) {
6126 if (state->in_callback_fn && state->refs[i].callback_ref != state->frameno)
6128 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
6129 state->refs[i].id, state->refs[i].insn_idx);
6130 refs_lingering = true;
6132 return refs_lingering ? -EINVAL : 0;
6135 static int check_bpf_snprintf_call(struct bpf_verifier_env *env,
6136 struct bpf_reg_state *regs)
6138 struct bpf_reg_state *fmt_reg = ®s[BPF_REG_3];
6139 struct bpf_reg_state *data_len_reg = ®s[BPF_REG_5];
6140 struct bpf_map *fmt_map = fmt_reg->map_ptr;
6141 int err, fmt_map_off, num_args;
6145 /* data must be an array of u64 */
6146 if (data_len_reg->var_off.value % 8)
6148 num_args = data_len_reg->var_off.value / 8;
6150 /* fmt being ARG_PTR_TO_CONST_STR guarantees that var_off is const
6151 * and map_direct_value_addr is set.
6153 fmt_map_off = fmt_reg->off + fmt_reg->var_off.value;
6154 err = fmt_map->ops->map_direct_value_addr(fmt_map, &fmt_addr,
6157 verbose(env, "verifier bug\n");
6160 fmt = (char *)(long)fmt_addr + fmt_map_off;
6162 /* We are also guaranteed that fmt+fmt_map_off is NULL terminated, we
6163 * can focus on validating the format specifiers.
6165 err = bpf_bprintf_prepare(fmt, UINT_MAX, NULL, NULL, num_args);
6167 verbose(env, "Invalid format string\n");
6172 static int check_get_func_ip(struct bpf_verifier_env *env)
6174 enum bpf_attach_type eatype = env->prog->expected_attach_type;
6175 enum bpf_prog_type type = resolve_prog_type(env->prog);
6176 int func_id = BPF_FUNC_get_func_ip;
6178 if (type == BPF_PROG_TYPE_TRACING) {
6179 if (eatype != BPF_TRACE_FENTRY && eatype != BPF_TRACE_FEXIT &&
6180 eatype != BPF_MODIFY_RETURN) {
6181 verbose(env, "func %s#%d supported only for fentry/fexit/fmod_ret programs\n",
6182 func_id_name(func_id), func_id);
6186 } else if (type == BPF_PROG_TYPE_KPROBE) {
6190 verbose(env, "func %s#%d not supported for program type %d\n",
6191 func_id_name(func_id), func_id, type);
6195 static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
6198 const struct bpf_func_proto *fn = NULL;
6199 enum bpf_return_type ret_type;
6200 enum bpf_type_flag ret_flag;
6201 struct bpf_reg_state *regs;
6202 struct bpf_call_arg_meta meta;
6203 int insn_idx = *insn_idx_p;
6205 int i, err, func_id;
6207 /* find function prototype */
6208 func_id = insn->imm;
6209 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
6210 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
6215 if (env->ops->get_func_proto)
6216 fn = env->ops->get_func_proto(func_id, env->prog);
6218 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
6223 /* eBPF programs must be GPL compatible to use GPL-ed functions */
6224 if (!env->prog->gpl_compatible && fn->gpl_only) {
6225 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
6229 if (fn->allowed && !fn->allowed(env->prog)) {
6230 verbose(env, "helper call is not allowed in probe\n");
6234 /* With LD_ABS/IND some JITs save/restore skb from r1. */
6235 changes_data = bpf_helper_changes_pkt_data(fn->func);
6236 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
6237 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
6238 func_id_name(func_id), func_id);
6242 memset(&meta, 0, sizeof(meta));
6243 meta.pkt_access = fn->pkt_access;
6245 err = check_func_proto(fn, func_id);
6247 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
6248 func_id_name(func_id), func_id);
6252 meta.func_id = func_id;
6254 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++) {
6255 err = check_func_arg(env, i, &meta, fn);
6260 err = record_func_map(env, &meta, func_id, insn_idx);
6264 err = record_func_key(env, &meta, func_id, insn_idx);
6268 /* Mark slots with STACK_MISC in case of raw mode, stack offset
6269 * is inferred from register state.
6271 for (i = 0; i < meta.access_size; i++) {
6272 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
6273 BPF_WRITE, -1, false);
6278 if (func_id == BPF_FUNC_tail_call) {
6279 err = check_reference_leak(env);
6281 verbose(env, "tail_call would lead to reference leak\n");
6284 } else if (is_release_function(func_id)) {
6285 err = release_reference(env, meta.ref_obj_id);
6287 verbose(env, "func %s#%d reference has not been acquired before\n",
6288 func_id_name(func_id), func_id);
6293 regs = cur_regs(env);
6295 /* check that flags argument in get_local_storage(map, flags) is 0,
6296 * this is required because get_local_storage() can't return an error.
6298 if (func_id == BPF_FUNC_get_local_storage &&
6299 !register_is_null(®s[BPF_REG_2])) {
6300 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
6304 if (func_id == BPF_FUNC_for_each_map_elem) {
6305 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
6306 set_map_elem_callback_state);
6311 if (func_id == BPF_FUNC_timer_set_callback) {
6312 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
6313 set_timer_callback_state);
6318 if (func_id == BPF_FUNC_snprintf) {
6319 err = check_bpf_snprintf_call(env, regs);
6324 /* reset caller saved regs */
6325 for (i = 0; i < CALLER_SAVED_REGS; i++) {
6326 mark_reg_not_init(env, regs, caller_saved[i]);
6327 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
6330 /* helper call returns 64-bit value. */
6331 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
6333 /* update return register (already marked as written above) */
6334 ret_type = fn->ret_type;
6335 ret_flag = type_flag(fn->ret_type);
6336 if (ret_type == RET_INTEGER) {
6337 /* sets type to SCALAR_VALUE */
6338 mark_reg_unknown(env, regs, BPF_REG_0);
6339 } else if (ret_type == RET_VOID) {
6340 regs[BPF_REG_0].type = NOT_INIT;
6341 } else if (base_type(ret_type) == RET_PTR_TO_MAP_VALUE) {
6342 /* There is no offset yet applied, variable or fixed */
6343 mark_reg_known_zero(env, regs, BPF_REG_0);
6344 /* remember map_ptr, so that check_map_access()
6345 * can check 'value_size' boundary of memory access
6346 * to map element returned from bpf_map_lookup_elem()
6348 if (meta.map_ptr == NULL) {
6350 "kernel subsystem misconfigured verifier\n");
6353 regs[BPF_REG_0].map_ptr = meta.map_ptr;
6354 regs[BPF_REG_0].map_uid = meta.map_uid;
6355 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE | ret_flag;
6356 if (!type_may_be_null(ret_type) &&
6357 map_value_has_spin_lock(meta.map_ptr)) {
6358 regs[BPF_REG_0].id = ++env->id_gen;
6360 } else if (base_type(ret_type) == RET_PTR_TO_SOCKET) {
6361 mark_reg_known_zero(env, regs, BPF_REG_0);
6362 regs[BPF_REG_0].type = PTR_TO_SOCKET | ret_flag;
6363 } else if (base_type(ret_type) == RET_PTR_TO_SOCK_COMMON) {
6364 mark_reg_known_zero(env, regs, BPF_REG_0);
6365 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON | ret_flag;
6366 } else if (base_type(ret_type) == RET_PTR_TO_TCP_SOCK) {
6367 mark_reg_known_zero(env, regs, BPF_REG_0);
6368 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK | ret_flag;
6369 } else if (base_type(ret_type) == RET_PTR_TO_ALLOC_MEM) {
6370 mark_reg_known_zero(env, regs, BPF_REG_0);
6371 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
6372 regs[BPF_REG_0].mem_size = meta.mem_size;
6373 } else if (base_type(ret_type) == RET_PTR_TO_MEM_OR_BTF_ID) {
6374 const struct btf_type *t;
6376 mark_reg_known_zero(env, regs, BPF_REG_0);
6377 t = btf_type_skip_modifiers(meta.ret_btf, meta.ret_btf_id, NULL);
6378 if (!btf_type_is_struct(t)) {
6380 const struct btf_type *ret;
6383 /* resolve the type size of ksym. */
6384 ret = btf_resolve_size(meta.ret_btf, t, &tsize);
6386 tname = btf_name_by_offset(meta.ret_btf, t->name_off);
6387 verbose(env, "unable to resolve the size of type '%s': %ld\n",
6388 tname, PTR_ERR(ret));
6391 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
6392 regs[BPF_REG_0].mem_size = tsize;
6394 /* MEM_RDONLY may be carried from ret_flag, but it
6395 * doesn't apply on PTR_TO_BTF_ID. Fold it, otherwise
6396 * it will confuse the check of PTR_TO_BTF_ID in
6397 * check_mem_access().
6399 ret_flag &= ~MEM_RDONLY;
6401 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
6402 regs[BPF_REG_0].btf = meta.ret_btf;
6403 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
6405 } else if (base_type(ret_type) == RET_PTR_TO_BTF_ID) {
6408 mark_reg_known_zero(env, regs, BPF_REG_0);
6409 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
6410 ret_btf_id = *fn->ret_btf_id;
6411 if (ret_btf_id == 0) {
6412 verbose(env, "invalid return type %u of func %s#%d\n",
6413 base_type(ret_type), func_id_name(func_id),
6417 /* current BPF helper definitions are only coming from
6418 * built-in code with type IDs from vmlinux BTF
6420 regs[BPF_REG_0].btf = btf_vmlinux;
6421 regs[BPF_REG_0].btf_id = ret_btf_id;
6423 verbose(env, "unknown return type %u of func %s#%d\n",
6424 base_type(ret_type), func_id_name(func_id), func_id);
6428 if (type_may_be_null(regs[BPF_REG_0].type))
6429 regs[BPF_REG_0].id = ++env->id_gen;
6431 if (is_ptr_cast_function(func_id)) {
6432 /* For release_reference() */
6433 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
6434 } else if (is_acquire_function(func_id, meta.map_ptr)) {
6435 int id = acquire_reference_state(env, insn_idx);
6439 /* For mark_ptr_or_null_reg() */
6440 regs[BPF_REG_0].id = id;
6441 /* For release_reference() */
6442 regs[BPF_REG_0].ref_obj_id = id;
6445 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
6447 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
6451 if ((func_id == BPF_FUNC_get_stack ||
6452 func_id == BPF_FUNC_get_task_stack) &&
6453 !env->prog->has_callchain_buf) {
6454 const char *err_str;
6456 #ifdef CONFIG_PERF_EVENTS
6457 err = get_callchain_buffers(sysctl_perf_event_max_stack);
6458 err_str = "cannot get callchain buffer for func %s#%d\n";
6461 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
6464 verbose(env, err_str, func_id_name(func_id), func_id);
6468 env->prog->has_callchain_buf = true;
6471 if (func_id == BPF_FUNC_get_stackid || func_id == BPF_FUNC_get_stack)
6472 env->prog->call_get_stack = true;
6474 if (func_id == BPF_FUNC_get_func_ip) {
6475 if (check_get_func_ip(env))
6477 env->prog->call_get_func_ip = true;
6481 clear_all_pkt_pointers(env);
6485 /* mark_btf_func_reg_size() is used when the reg size is determined by
6486 * the BTF func_proto's return value size and argument.
6488 static void mark_btf_func_reg_size(struct bpf_verifier_env *env, u32 regno,
6491 struct bpf_reg_state *reg = &cur_regs(env)[regno];
6493 if (regno == BPF_REG_0) {
6494 /* Function return value */
6495 reg->live |= REG_LIVE_WRITTEN;
6496 reg->subreg_def = reg_size == sizeof(u64) ?
6497 DEF_NOT_SUBREG : env->insn_idx + 1;
6499 /* Function argument */
6500 if (reg_size == sizeof(u64)) {
6501 mark_insn_zext(env, reg);
6502 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
6504 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ32);
6509 static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn)
6511 const struct btf_type *t, *func, *func_proto, *ptr_type;
6512 struct bpf_reg_state *regs = cur_regs(env);
6513 const char *func_name, *ptr_type_name;
6514 u32 i, nargs, func_id, ptr_type_id;
6515 const struct btf_param *args;
6518 func_id = insn->imm;
6519 func = btf_type_by_id(btf_vmlinux, func_id);
6520 func_name = btf_name_by_offset(btf_vmlinux, func->name_off);
6521 func_proto = btf_type_by_id(btf_vmlinux, func->type);
6523 if (!env->ops->check_kfunc_call ||
6524 !env->ops->check_kfunc_call(func_id)) {
6525 verbose(env, "calling kernel function %s is not allowed\n",
6530 /* Check the arguments */
6531 err = btf_check_kfunc_arg_match(env, btf_vmlinux, func_id, regs);
6535 for (i = 0; i < CALLER_SAVED_REGS; i++)
6536 mark_reg_not_init(env, regs, caller_saved[i]);
6538 /* Check return type */
6539 t = btf_type_skip_modifiers(btf_vmlinux, func_proto->type, NULL);
6540 if (btf_type_is_scalar(t)) {
6541 mark_reg_unknown(env, regs, BPF_REG_0);
6542 mark_btf_func_reg_size(env, BPF_REG_0, t->size);
6543 } else if (btf_type_is_ptr(t)) {
6544 ptr_type = btf_type_skip_modifiers(btf_vmlinux, t->type,
6546 if (!btf_type_is_struct(ptr_type)) {
6547 ptr_type_name = btf_name_by_offset(btf_vmlinux,
6548 ptr_type->name_off);
6549 verbose(env, "kernel function %s returns pointer type %s %s is not supported\n",
6550 func_name, btf_type_str(ptr_type),
6554 mark_reg_known_zero(env, regs, BPF_REG_0);
6555 regs[BPF_REG_0].btf = btf_vmlinux;
6556 regs[BPF_REG_0].type = PTR_TO_BTF_ID;
6557 regs[BPF_REG_0].btf_id = ptr_type_id;
6558 mark_btf_func_reg_size(env, BPF_REG_0, sizeof(void *));
6559 } /* else { add_kfunc_call() ensures it is btf_type_is_void(t) } */
6561 nargs = btf_type_vlen(func_proto);
6562 args = (const struct btf_param *)(func_proto + 1);
6563 for (i = 0; i < nargs; i++) {
6566 t = btf_type_skip_modifiers(btf_vmlinux, args[i].type, NULL);
6567 if (btf_type_is_ptr(t))
6568 mark_btf_func_reg_size(env, regno, sizeof(void *));
6570 /* scalar. ensured by btf_check_kfunc_arg_match() */
6571 mark_btf_func_reg_size(env, regno, t->size);
6577 static bool signed_add_overflows(s64 a, s64 b)
6579 /* Do the add in u64, where overflow is well-defined */
6580 s64 res = (s64)((u64)a + (u64)b);
6587 static bool signed_add32_overflows(s32 a, s32 b)
6589 /* Do the add in u32, where overflow is well-defined */
6590 s32 res = (s32)((u32)a + (u32)b);
6597 static bool signed_sub_overflows(s64 a, s64 b)
6599 /* Do the sub in u64, where overflow is well-defined */
6600 s64 res = (s64)((u64)a - (u64)b);
6607 static bool signed_sub32_overflows(s32 a, s32 b)
6609 /* Do the sub in u32, where overflow is well-defined */
6610 s32 res = (s32)((u32)a - (u32)b);
6617 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
6618 const struct bpf_reg_state *reg,
6619 enum bpf_reg_type type)
6621 bool known = tnum_is_const(reg->var_off);
6622 s64 val = reg->var_off.value;
6623 s64 smin = reg->smin_value;
6625 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
6626 verbose(env, "math between %s pointer and %lld is not allowed\n",
6627 reg_type_str(env, type), val);
6631 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
6632 verbose(env, "%s pointer offset %d is not allowed\n",
6633 reg_type_str(env, type), reg->off);
6637 if (smin == S64_MIN) {
6638 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
6639 reg_type_str(env, type));
6643 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
6644 verbose(env, "value %lld makes %s pointer be out of bounds\n",
6645 smin, reg_type_str(env, type));
6652 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
6654 return &env->insn_aux_data[env->insn_idx];
6665 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
6666 u32 *alu_limit, bool mask_to_left)
6668 u32 max = 0, ptr_limit = 0;
6670 switch (ptr_reg->type) {
6672 /* Offset 0 is out-of-bounds, but acceptable start for the
6673 * left direction, see BPF_REG_FP. Also, unknown scalar
6674 * offset where we would need to deal with min/max bounds is
6675 * currently prohibited for unprivileged.
6677 max = MAX_BPF_STACK + mask_to_left;
6678 ptr_limit = -(ptr_reg->var_off.value + ptr_reg->off);
6680 case PTR_TO_MAP_VALUE:
6681 max = ptr_reg->map_ptr->value_size;
6682 ptr_limit = (mask_to_left ?
6683 ptr_reg->smin_value :
6684 ptr_reg->umax_value) + ptr_reg->off;
6690 if (ptr_limit >= max)
6691 return REASON_LIMIT;
6692 *alu_limit = ptr_limit;
6696 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
6697 const struct bpf_insn *insn)
6699 return env->bypass_spec_v1 || BPF_SRC(insn->code) == BPF_K;
6702 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
6703 u32 alu_state, u32 alu_limit)
6705 /* If we arrived here from different branches with different
6706 * state or limits to sanitize, then this won't work.
6708 if (aux->alu_state &&
6709 (aux->alu_state != alu_state ||
6710 aux->alu_limit != alu_limit))
6711 return REASON_PATHS;
6713 /* Corresponding fixup done in do_misc_fixups(). */
6714 aux->alu_state = alu_state;
6715 aux->alu_limit = alu_limit;
6719 static int sanitize_val_alu(struct bpf_verifier_env *env,
6720 struct bpf_insn *insn)
6722 struct bpf_insn_aux_data *aux = cur_aux(env);
6724 if (can_skip_alu_sanitation(env, insn))
6727 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
6730 static bool sanitize_needed(u8 opcode)
6732 return opcode == BPF_ADD || opcode == BPF_SUB;
6735 struct bpf_sanitize_info {
6736 struct bpf_insn_aux_data aux;
6740 static struct bpf_verifier_state *
6741 sanitize_speculative_path(struct bpf_verifier_env *env,
6742 const struct bpf_insn *insn,
6743 u32 next_idx, u32 curr_idx)
6745 struct bpf_verifier_state *branch;
6746 struct bpf_reg_state *regs;
6748 branch = push_stack(env, next_idx, curr_idx, true);
6749 if (branch && insn) {
6750 regs = branch->frame[branch->curframe]->regs;
6751 if (BPF_SRC(insn->code) == BPF_K) {
6752 mark_reg_unknown(env, regs, insn->dst_reg);
6753 } else if (BPF_SRC(insn->code) == BPF_X) {
6754 mark_reg_unknown(env, regs, insn->dst_reg);
6755 mark_reg_unknown(env, regs, insn->src_reg);
6761 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
6762 struct bpf_insn *insn,
6763 const struct bpf_reg_state *ptr_reg,
6764 const struct bpf_reg_state *off_reg,
6765 struct bpf_reg_state *dst_reg,
6766 struct bpf_sanitize_info *info,
6767 const bool commit_window)
6769 struct bpf_insn_aux_data *aux = commit_window ? cur_aux(env) : &info->aux;
6770 struct bpf_verifier_state *vstate = env->cur_state;
6771 bool off_is_imm = tnum_is_const(off_reg->var_off);
6772 bool off_is_neg = off_reg->smin_value < 0;
6773 bool ptr_is_dst_reg = ptr_reg == dst_reg;
6774 u8 opcode = BPF_OP(insn->code);
6775 u32 alu_state, alu_limit;
6776 struct bpf_reg_state tmp;
6780 if (can_skip_alu_sanitation(env, insn))
6783 /* We already marked aux for masking from non-speculative
6784 * paths, thus we got here in the first place. We only care
6785 * to explore bad access from here.
6787 if (vstate->speculative)
6790 if (!commit_window) {
6791 if (!tnum_is_const(off_reg->var_off) &&
6792 (off_reg->smin_value < 0) != (off_reg->smax_value < 0))
6793 return REASON_BOUNDS;
6795 info->mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
6796 (opcode == BPF_SUB && !off_is_neg);
6799 err = retrieve_ptr_limit(ptr_reg, &alu_limit, info->mask_to_left);
6803 if (commit_window) {
6804 /* In commit phase we narrow the masking window based on
6805 * the observed pointer move after the simulated operation.
6807 alu_state = info->aux.alu_state;
6808 alu_limit = abs(info->aux.alu_limit - alu_limit);
6810 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
6811 alu_state |= off_is_imm ? BPF_ALU_IMMEDIATE : 0;
6812 alu_state |= ptr_is_dst_reg ?
6813 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
6815 /* Limit pruning on unknown scalars to enable deep search for
6816 * potential masking differences from other program paths.
6819 env->explore_alu_limits = true;
6822 err = update_alu_sanitation_state(aux, alu_state, alu_limit);
6826 /* If we're in commit phase, we're done here given we already
6827 * pushed the truncated dst_reg into the speculative verification
6830 * Also, when register is a known constant, we rewrite register-based
6831 * operation to immediate-based, and thus do not need masking (and as
6832 * a consequence, do not need to simulate the zero-truncation either).
6834 if (commit_window || off_is_imm)
6837 /* Simulate and find potential out-of-bounds access under
6838 * speculative execution from truncation as a result of
6839 * masking when off was not within expected range. If off
6840 * sits in dst, then we temporarily need to move ptr there
6841 * to simulate dst (== 0) +/-= ptr. Needed, for example,
6842 * for cases where we use K-based arithmetic in one direction
6843 * and truncated reg-based in the other in order to explore
6846 if (!ptr_is_dst_reg) {
6848 *dst_reg = *ptr_reg;
6850 ret = sanitize_speculative_path(env, NULL, env->insn_idx + 1,
6852 if (!ptr_is_dst_reg && ret)
6854 return !ret ? REASON_STACK : 0;
6857 static void sanitize_mark_insn_seen(struct bpf_verifier_env *env)
6859 struct bpf_verifier_state *vstate = env->cur_state;
6861 /* If we simulate paths under speculation, we don't update the
6862 * insn as 'seen' such that when we verify unreachable paths in
6863 * the non-speculative domain, sanitize_dead_code() can still
6864 * rewrite/sanitize them.
6866 if (!vstate->speculative)
6867 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
6870 static int sanitize_err(struct bpf_verifier_env *env,
6871 const struct bpf_insn *insn, int reason,
6872 const struct bpf_reg_state *off_reg,
6873 const struct bpf_reg_state *dst_reg)
6875 static const char *err = "pointer arithmetic with it prohibited for !root";
6876 const char *op = BPF_OP(insn->code) == BPF_ADD ? "add" : "sub";
6877 u32 dst = insn->dst_reg, src = insn->src_reg;
6881 verbose(env, "R%d has unknown scalar with mixed signed bounds, %s\n",
6882 off_reg == dst_reg ? dst : src, err);
6885 verbose(env, "R%d has pointer with unsupported alu operation, %s\n",
6886 off_reg == dst_reg ? src : dst, err);
6889 verbose(env, "R%d tried to %s from different maps, paths or scalars, %s\n",
6893 verbose(env, "R%d tried to %s beyond pointer bounds, %s\n",
6897 verbose(env, "R%d could not be pushed for speculative verification, %s\n",
6901 verbose(env, "verifier internal error: unknown reason (%d)\n",
6909 /* check that stack access falls within stack limits and that 'reg' doesn't
6910 * have a variable offset.
6912 * Variable offset is prohibited for unprivileged mode for simplicity since it
6913 * requires corresponding support in Spectre masking for stack ALU. See also
6914 * retrieve_ptr_limit().
6917 * 'off' includes 'reg->off'.
6919 static int check_stack_access_for_ptr_arithmetic(
6920 struct bpf_verifier_env *env,
6922 const struct bpf_reg_state *reg,
6925 if (!tnum_is_const(reg->var_off)) {
6928 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6929 verbose(env, "R%d variable stack access prohibited for !root, var_off=%s off=%d\n",
6930 regno, tn_buf, off);
6934 if (off >= 0 || off < -MAX_BPF_STACK) {
6935 verbose(env, "R%d stack pointer arithmetic goes out of range, "
6936 "prohibited for !root; off=%d\n", regno, off);
6943 static int sanitize_check_bounds(struct bpf_verifier_env *env,
6944 const struct bpf_insn *insn,
6945 const struct bpf_reg_state *dst_reg)
6947 u32 dst = insn->dst_reg;
6949 /* For unprivileged we require that resulting offset must be in bounds
6950 * in order to be able to sanitize access later on.
6952 if (env->bypass_spec_v1)
6955 switch (dst_reg->type) {
6957 if (check_stack_access_for_ptr_arithmetic(env, dst, dst_reg,
6958 dst_reg->off + dst_reg->var_off.value))
6961 case PTR_TO_MAP_VALUE:
6962 if (check_map_access(env, dst, dst_reg->off, 1, false)) {
6963 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
6964 "prohibited for !root\n", dst);
6975 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
6976 * Caller should also handle BPF_MOV case separately.
6977 * If we return -EACCES, caller may want to try again treating pointer as a
6978 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
6980 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
6981 struct bpf_insn *insn,
6982 const struct bpf_reg_state *ptr_reg,
6983 const struct bpf_reg_state *off_reg)
6985 struct bpf_verifier_state *vstate = env->cur_state;
6986 struct bpf_func_state *state = vstate->frame[vstate->curframe];
6987 struct bpf_reg_state *regs = state->regs, *dst_reg;
6988 bool known = tnum_is_const(off_reg->var_off);
6989 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
6990 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
6991 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
6992 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
6993 struct bpf_sanitize_info info = {};
6994 u8 opcode = BPF_OP(insn->code);
6995 u32 dst = insn->dst_reg;
6998 dst_reg = ®s[dst];
7000 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
7001 smin_val > smax_val || umin_val > umax_val) {
7002 /* Taint dst register if offset had invalid bounds derived from
7003 * e.g. dead branches.
7005 __mark_reg_unknown(env, dst_reg);
7009 if (BPF_CLASS(insn->code) != BPF_ALU64) {
7010 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
7011 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
7012 __mark_reg_unknown(env, dst_reg);
7017 "R%d 32-bit pointer arithmetic prohibited\n",
7022 if (ptr_reg->type & PTR_MAYBE_NULL) {
7023 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
7024 dst, reg_type_str(env, ptr_reg->type));
7028 switch (base_type(ptr_reg->type)) {
7029 case CONST_PTR_TO_MAP:
7030 /* smin_val represents the known value */
7031 if (known && smin_val == 0 && opcode == BPF_ADD)
7034 case PTR_TO_PACKET_END:
7036 case PTR_TO_SOCK_COMMON:
7037 case PTR_TO_TCP_SOCK:
7038 case PTR_TO_XDP_SOCK:
7040 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
7041 dst, reg_type_str(env, ptr_reg->type));
7044 if (type_may_be_null(ptr_reg->type))
7049 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
7050 * The id may be overwritten later if we create a new variable offset.
7052 dst_reg->type = ptr_reg->type;
7053 dst_reg->id = ptr_reg->id;
7055 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
7056 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
7059 /* pointer types do not carry 32-bit bounds at the moment. */
7060 __mark_reg32_unbounded(dst_reg);
7062 if (sanitize_needed(opcode)) {
7063 ret = sanitize_ptr_alu(env, insn, ptr_reg, off_reg, dst_reg,
7066 return sanitize_err(env, insn, ret, off_reg, dst_reg);
7071 /* We can take a fixed offset as long as it doesn't overflow
7072 * the s32 'off' field
7074 if (known && (ptr_reg->off + smin_val ==
7075 (s64)(s32)(ptr_reg->off + smin_val))) {
7076 /* pointer += K. Accumulate it into fixed offset */
7077 dst_reg->smin_value = smin_ptr;
7078 dst_reg->smax_value = smax_ptr;
7079 dst_reg->umin_value = umin_ptr;
7080 dst_reg->umax_value = umax_ptr;
7081 dst_reg->var_off = ptr_reg->var_off;
7082 dst_reg->off = ptr_reg->off + smin_val;
7083 dst_reg->raw = ptr_reg->raw;
7086 /* A new variable offset is created. Note that off_reg->off
7087 * == 0, since it's a scalar.
7088 * dst_reg gets the pointer type and since some positive
7089 * integer value was added to the pointer, give it a new 'id'
7090 * if it's a PTR_TO_PACKET.
7091 * this creates a new 'base' pointer, off_reg (variable) gets
7092 * added into the variable offset, and we copy the fixed offset
7095 if (signed_add_overflows(smin_ptr, smin_val) ||
7096 signed_add_overflows(smax_ptr, smax_val)) {
7097 dst_reg->smin_value = S64_MIN;
7098 dst_reg->smax_value = S64_MAX;
7100 dst_reg->smin_value = smin_ptr + smin_val;
7101 dst_reg->smax_value = smax_ptr + smax_val;
7103 if (umin_ptr + umin_val < umin_ptr ||
7104 umax_ptr + umax_val < umax_ptr) {
7105 dst_reg->umin_value = 0;
7106 dst_reg->umax_value = U64_MAX;
7108 dst_reg->umin_value = umin_ptr + umin_val;
7109 dst_reg->umax_value = umax_ptr + umax_val;
7111 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
7112 dst_reg->off = ptr_reg->off;
7113 dst_reg->raw = ptr_reg->raw;
7114 if (reg_is_pkt_pointer(ptr_reg)) {
7115 dst_reg->id = ++env->id_gen;
7116 /* something was added to pkt_ptr, set range to zero */
7117 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
7121 if (dst_reg == off_reg) {
7122 /* scalar -= pointer. Creates an unknown scalar */
7123 verbose(env, "R%d tried to subtract pointer from scalar\n",
7127 /* We don't allow subtraction from FP, because (according to
7128 * test_verifier.c test "invalid fp arithmetic", JITs might not
7129 * be able to deal with it.
7131 if (ptr_reg->type == PTR_TO_STACK) {
7132 verbose(env, "R%d subtraction from stack pointer prohibited\n",
7136 if (known && (ptr_reg->off - smin_val ==
7137 (s64)(s32)(ptr_reg->off - smin_val))) {
7138 /* pointer -= K. Subtract it from fixed offset */
7139 dst_reg->smin_value = smin_ptr;
7140 dst_reg->smax_value = smax_ptr;
7141 dst_reg->umin_value = umin_ptr;
7142 dst_reg->umax_value = umax_ptr;
7143 dst_reg->var_off = ptr_reg->var_off;
7144 dst_reg->id = ptr_reg->id;
7145 dst_reg->off = ptr_reg->off - smin_val;
7146 dst_reg->raw = ptr_reg->raw;
7149 /* A new variable offset is created. If the subtrahend is known
7150 * nonnegative, then any reg->range we had before is still good.
7152 if (signed_sub_overflows(smin_ptr, smax_val) ||
7153 signed_sub_overflows(smax_ptr, smin_val)) {
7154 /* Overflow possible, we know nothing */
7155 dst_reg->smin_value = S64_MIN;
7156 dst_reg->smax_value = S64_MAX;
7158 dst_reg->smin_value = smin_ptr - smax_val;
7159 dst_reg->smax_value = smax_ptr - smin_val;
7161 if (umin_ptr < umax_val) {
7162 /* Overflow possible, we know nothing */
7163 dst_reg->umin_value = 0;
7164 dst_reg->umax_value = U64_MAX;
7166 /* Cannot overflow (as long as bounds are consistent) */
7167 dst_reg->umin_value = umin_ptr - umax_val;
7168 dst_reg->umax_value = umax_ptr - umin_val;
7170 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
7171 dst_reg->off = ptr_reg->off;
7172 dst_reg->raw = ptr_reg->raw;
7173 if (reg_is_pkt_pointer(ptr_reg)) {
7174 dst_reg->id = ++env->id_gen;
7175 /* something was added to pkt_ptr, set range to zero */
7177 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
7183 /* bitwise ops on pointers are troublesome, prohibit. */
7184 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
7185 dst, bpf_alu_string[opcode >> 4]);
7188 /* other operators (e.g. MUL,LSH) produce non-pointer results */
7189 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
7190 dst, bpf_alu_string[opcode >> 4]);
7194 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
7196 reg_bounds_sync(dst_reg);
7197 if (sanitize_check_bounds(env, insn, dst_reg) < 0)
7199 if (sanitize_needed(opcode)) {
7200 ret = sanitize_ptr_alu(env, insn, dst_reg, off_reg, dst_reg,
7203 return sanitize_err(env, insn, ret, off_reg, dst_reg);
7209 static void scalar32_min_max_add(struct bpf_reg_state *dst_reg,
7210 struct bpf_reg_state *src_reg)
7212 s32 smin_val = src_reg->s32_min_value;
7213 s32 smax_val = src_reg->s32_max_value;
7214 u32 umin_val = src_reg->u32_min_value;
7215 u32 umax_val = src_reg->u32_max_value;
7217 if (signed_add32_overflows(dst_reg->s32_min_value, smin_val) ||
7218 signed_add32_overflows(dst_reg->s32_max_value, smax_val)) {
7219 dst_reg->s32_min_value = S32_MIN;
7220 dst_reg->s32_max_value = S32_MAX;
7222 dst_reg->s32_min_value += smin_val;
7223 dst_reg->s32_max_value += smax_val;
7225 if (dst_reg->u32_min_value + umin_val < umin_val ||
7226 dst_reg->u32_max_value + umax_val < umax_val) {
7227 dst_reg->u32_min_value = 0;
7228 dst_reg->u32_max_value = U32_MAX;
7230 dst_reg->u32_min_value += umin_val;
7231 dst_reg->u32_max_value += umax_val;
7235 static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
7236 struct bpf_reg_state *src_reg)
7238 s64 smin_val = src_reg->smin_value;
7239 s64 smax_val = src_reg->smax_value;
7240 u64 umin_val = src_reg->umin_value;
7241 u64 umax_val = src_reg->umax_value;
7243 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
7244 signed_add_overflows(dst_reg->smax_value, smax_val)) {
7245 dst_reg->smin_value = S64_MIN;
7246 dst_reg->smax_value = S64_MAX;
7248 dst_reg->smin_value += smin_val;
7249 dst_reg->smax_value += smax_val;
7251 if (dst_reg->umin_value + umin_val < umin_val ||
7252 dst_reg->umax_value + umax_val < umax_val) {
7253 dst_reg->umin_value = 0;
7254 dst_reg->umax_value = U64_MAX;
7256 dst_reg->umin_value += umin_val;
7257 dst_reg->umax_value += umax_val;
7261 static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg,
7262 struct bpf_reg_state *src_reg)
7264 s32 smin_val = src_reg->s32_min_value;
7265 s32 smax_val = src_reg->s32_max_value;
7266 u32 umin_val = src_reg->u32_min_value;
7267 u32 umax_val = src_reg->u32_max_value;
7269 if (signed_sub32_overflows(dst_reg->s32_min_value, smax_val) ||
7270 signed_sub32_overflows(dst_reg->s32_max_value, smin_val)) {
7271 /* Overflow possible, we know nothing */
7272 dst_reg->s32_min_value = S32_MIN;
7273 dst_reg->s32_max_value = S32_MAX;
7275 dst_reg->s32_min_value -= smax_val;
7276 dst_reg->s32_max_value -= smin_val;
7278 if (dst_reg->u32_min_value < umax_val) {
7279 /* Overflow possible, we know nothing */
7280 dst_reg->u32_min_value = 0;
7281 dst_reg->u32_max_value = U32_MAX;
7283 /* Cannot overflow (as long as bounds are consistent) */
7284 dst_reg->u32_min_value -= umax_val;
7285 dst_reg->u32_max_value -= umin_val;
7289 static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
7290 struct bpf_reg_state *src_reg)
7292 s64 smin_val = src_reg->smin_value;
7293 s64 smax_val = src_reg->smax_value;
7294 u64 umin_val = src_reg->umin_value;
7295 u64 umax_val = src_reg->umax_value;
7297 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
7298 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
7299 /* Overflow possible, we know nothing */
7300 dst_reg->smin_value = S64_MIN;
7301 dst_reg->smax_value = S64_MAX;
7303 dst_reg->smin_value -= smax_val;
7304 dst_reg->smax_value -= smin_val;
7306 if (dst_reg->umin_value < umax_val) {
7307 /* Overflow possible, we know nothing */
7308 dst_reg->umin_value = 0;
7309 dst_reg->umax_value = U64_MAX;
7311 /* Cannot overflow (as long as bounds are consistent) */
7312 dst_reg->umin_value -= umax_val;
7313 dst_reg->umax_value -= umin_val;
7317 static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg,
7318 struct bpf_reg_state *src_reg)
7320 s32 smin_val = src_reg->s32_min_value;
7321 u32 umin_val = src_reg->u32_min_value;
7322 u32 umax_val = src_reg->u32_max_value;
7324 if (smin_val < 0 || dst_reg->s32_min_value < 0) {
7325 /* Ain't nobody got time to multiply that sign */
7326 __mark_reg32_unbounded(dst_reg);
7329 /* Both values are positive, so we can work with unsigned and
7330 * copy the result to signed (unless it exceeds S32_MAX).
7332 if (umax_val > U16_MAX || dst_reg->u32_max_value > U16_MAX) {
7333 /* Potential overflow, we know nothing */
7334 __mark_reg32_unbounded(dst_reg);
7337 dst_reg->u32_min_value *= umin_val;
7338 dst_reg->u32_max_value *= umax_val;
7339 if (dst_reg->u32_max_value > S32_MAX) {
7340 /* Overflow possible, we know nothing */
7341 dst_reg->s32_min_value = S32_MIN;
7342 dst_reg->s32_max_value = S32_MAX;
7344 dst_reg->s32_min_value = dst_reg->u32_min_value;
7345 dst_reg->s32_max_value = dst_reg->u32_max_value;
7349 static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
7350 struct bpf_reg_state *src_reg)
7352 s64 smin_val = src_reg->smin_value;
7353 u64 umin_val = src_reg->umin_value;
7354 u64 umax_val = src_reg->umax_value;
7356 if (smin_val < 0 || dst_reg->smin_value < 0) {
7357 /* Ain't nobody got time to multiply that sign */
7358 __mark_reg64_unbounded(dst_reg);
7361 /* Both values are positive, so we can work with unsigned and
7362 * copy the result to signed (unless it exceeds S64_MAX).
7364 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
7365 /* Potential overflow, we know nothing */
7366 __mark_reg64_unbounded(dst_reg);
7369 dst_reg->umin_value *= umin_val;
7370 dst_reg->umax_value *= umax_val;
7371 if (dst_reg->umax_value > S64_MAX) {
7372 /* Overflow possible, we know nothing */
7373 dst_reg->smin_value = S64_MIN;
7374 dst_reg->smax_value = S64_MAX;
7376 dst_reg->smin_value = dst_reg->umin_value;
7377 dst_reg->smax_value = dst_reg->umax_value;
7381 static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
7382 struct bpf_reg_state *src_reg)
7384 bool src_known = tnum_subreg_is_const(src_reg->var_off);
7385 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
7386 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
7387 s32 smin_val = src_reg->s32_min_value;
7388 u32 umax_val = src_reg->u32_max_value;
7390 if (src_known && dst_known) {
7391 __mark_reg32_known(dst_reg, var32_off.value);
7395 /* We get our minimum from the var_off, since that's inherently
7396 * bitwise. Our maximum is the minimum of the operands' maxima.
7398 dst_reg->u32_min_value = var32_off.value;
7399 dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val);
7400 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
7401 /* Lose signed bounds when ANDing negative numbers,
7402 * ain't nobody got time for that.
7404 dst_reg->s32_min_value = S32_MIN;
7405 dst_reg->s32_max_value = S32_MAX;
7407 /* ANDing two positives gives a positive, so safe to
7408 * cast result into s64.
7410 dst_reg->s32_min_value = dst_reg->u32_min_value;
7411 dst_reg->s32_max_value = dst_reg->u32_max_value;
7415 static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
7416 struct bpf_reg_state *src_reg)
7418 bool src_known = tnum_is_const(src_reg->var_off);
7419 bool dst_known = tnum_is_const(dst_reg->var_off);
7420 s64 smin_val = src_reg->smin_value;
7421 u64 umax_val = src_reg->umax_value;
7423 if (src_known && dst_known) {
7424 __mark_reg_known(dst_reg, dst_reg->var_off.value);
7428 /* We get our minimum from the var_off, since that's inherently
7429 * bitwise. Our maximum is the minimum of the operands' maxima.
7431 dst_reg->umin_value = dst_reg->var_off.value;
7432 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
7433 if (dst_reg->smin_value < 0 || smin_val < 0) {
7434 /* Lose signed bounds when ANDing negative numbers,
7435 * ain't nobody got time for that.
7437 dst_reg->smin_value = S64_MIN;
7438 dst_reg->smax_value = S64_MAX;
7440 /* ANDing two positives gives a positive, so safe to
7441 * cast result into s64.
7443 dst_reg->smin_value = dst_reg->umin_value;
7444 dst_reg->smax_value = dst_reg->umax_value;
7446 /* We may learn something more from the var_off */
7447 __update_reg_bounds(dst_reg);
7450 static void scalar32_min_max_or(struct bpf_reg_state *dst_reg,
7451 struct bpf_reg_state *src_reg)
7453 bool src_known = tnum_subreg_is_const(src_reg->var_off);
7454 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
7455 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
7456 s32 smin_val = src_reg->s32_min_value;
7457 u32 umin_val = src_reg->u32_min_value;
7459 if (src_known && dst_known) {
7460 __mark_reg32_known(dst_reg, var32_off.value);
7464 /* We get our maximum from the var_off, and our minimum is the
7465 * maximum of the operands' minima
7467 dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val);
7468 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
7469 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
7470 /* Lose signed bounds when ORing negative numbers,
7471 * ain't nobody got time for that.
7473 dst_reg->s32_min_value = S32_MIN;
7474 dst_reg->s32_max_value = S32_MAX;
7476 /* ORing two positives gives a positive, so safe to
7477 * cast result into s64.
7479 dst_reg->s32_min_value = dst_reg->u32_min_value;
7480 dst_reg->s32_max_value = dst_reg->u32_max_value;
7484 static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
7485 struct bpf_reg_state *src_reg)
7487 bool src_known = tnum_is_const(src_reg->var_off);
7488 bool dst_known = tnum_is_const(dst_reg->var_off);
7489 s64 smin_val = src_reg->smin_value;
7490 u64 umin_val = src_reg->umin_value;
7492 if (src_known && dst_known) {
7493 __mark_reg_known(dst_reg, dst_reg->var_off.value);
7497 /* We get our maximum from the var_off, and our minimum is the
7498 * maximum of the operands' minima
7500 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
7501 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
7502 if (dst_reg->smin_value < 0 || smin_val < 0) {
7503 /* Lose signed bounds when ORing negative numbers,
7504 * ain't nobody got time for that.
7506 dst_reg->smin_value = S64_MIN;
7507 dst_reg->smax_value = S64_MAX;
7509 /* ORing two positives gives a positive, so safe to
7510 * cast result into s64.
7512 dst_reg->smin_value = dst_reg->umin_value;
7513 dst_reg->smax_value = dst_reg->umax_value;
7515 /* We may learn something more from the var_off */
7516 __update_reg_bounds(dst_reg);
7519 static void scalar32_min_max_xor(struct bpf_reg_state *dst_reg,
7520 struct bpf_reg_state *src_reg)
7522 bool src_known = tnum_subreg_is_const(src_reg->var_off);
7523 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
7524 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
7525 s32 smin_val = src_reg->s32_min_value;
7527 if (src_known && dst_known) {
7528 __mark_reg32_known(dst_reg, var32_off.value);
7532 /* We get both minimum and maximum from the var32_off. */
7533 dst_reg->u32_min_value = var32_off.value;
7534 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
7536 if (dst_reg->s32_min_value >= 0 && smin_val >= 0) {
7537 /* XORing two positive sign numbers gives a positive,
7538 * so safe to cast u32 result into s32.
7540 dst_reg->s32_min_value = dst_reg->u32_min_value;
7541 dst_reg->s32_max_value = dst_reg->u32_max_value;
7543 dst_reg->s32_min_value = S32_MIN;
7544 dst_reg->s32_max_value = S32_MAX;
7548 static void scalar_min_max_xor(struct bpf_reg_state *dst_reg,
7549 struct bpf_reg_state *src_reg)
7551 bool src_known = tnum_is_const(src_reg->var_off);
7552 bool dst_known = tnum_is_const(dst_reg->var_off);
7553 s64 smin_val = src_reg->smin_value;
7555 if (src_known && dst_known) {
7556 /* dst_reg->var_off.value has been updated earlier */
7557 __mark_reg_known(dst_reg, dst_reg->var_off.value);
7561 /* We get both minimum and maximum from the var_off. */
7562 dst_reg->umin_value = dst_reg->var_off.value;
7563 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
7565 if (dst_reg->smin_value >= 0 && smin_val >= 0) {
7566 /* XORing two positive sign numbers gives a positive,
7567 * so safe to cast u64 result into s64.
7569 dst_reg->smin_value = dst_reg->umin_value;
7570 dst_reg->smax_value = dst_reg->umax_value;
7572 dst_reg->smin_value = S64_MIN;
7573 dst_reg->smax_value = S64_MAX;
7576 __update_reg_bounds(dst_reg);
7579 static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
7580 u64 umin_val, u64 umax_val)
7582 /* We lose all sign bit information (except what we can pick
7585 dst_reg->s32_min_value = S32_MIN;
7586 dst_reg->s32_max_value = S32_MAX;
7587 /* If we might shift our top bit out, then we know nothing */
7588 if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) {
7589 dst_reg->u32_min_value = 0;
7590 dst_reg->u32_max_value = U32_MAX;
7592 dst_reg->u32_min_value <<= umin_val;
7593 dst_reg->u32_max_value <<= umax_val;
7597 static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
7598 struct bpf_reg_state *src_reg)
7600 u32 umax_val = src_reg->u32_max_value;
7601 u32 umin_val = src_reg->u32_min_value;
7602 /* u32 alu operation will zext upper bits */
7603 struct tnum subreg = tnum_subreg(dst_reg->var_off);
7605 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
7606 dst_reg->var_off = tnum_subreg(tnum_lshift(subreg, umin_val));
7607 /* Not required but being careful mark reg64 bounds as unknown so
7608 * that we are forced to pick them up from tnum and zext later and
7609 * if some path skips this step we are still safe.
7611 __mark_reg64_unbounded(dst_reg);
7612 __update_reg32_bounds(dst_reg);
7615 static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg,
7616 u64 umin_val, u64 umax_val)
7618 /* Special case <<32 because it is a common compiler pattern to sign
7619 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
7620 * positive we know this shift will also be positive so we can track
7621 * bounds correctly. Otherwise we lose all sign bit information except
7622 * what we can pick up from var_off. Perhaps we can generalize this
7623 * later to shifts of any length.
7625 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0)
7626 dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32;
7628 dst_reg->smax_value = S64_MAX;
7630 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0)
7631 dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32;
7633 dst_reg->smin_value = S64_MIN;
7635 /* If we might shift our top bit out, then we know nothing */
7636 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
7637 dst_reg->umin_value = 0;
7638 dst_reg->umax_value = U64_MAX;
7640 dst_reg->umin_value <<= umin_val;
7641 dst_reg->umax_value <<= umax_val;
7645 static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
7646 struct bpf_reg_state *src_reg)
7648 u64 umax_val = src_reg->umax_value;
7649 u64 umin_val = src_reg->umin_value;
7651 /* scalar64 calc uses 32bit unshifted bounds so must be called first */
7652 __scalar64_min_max_lsh(dst_reg, umin_val, umax_val);
7653 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
7655 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
7656 /* We may learn something more from the var_off */
7657 __update_reg_bounds(dst_reg);
7660 static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg,
7661 struct bpf_reg_state *src_reg)
7663 struct tnum subreg = tnum_subreg(dst_reg->var_off);
7664 u32 umax_val = src_reg->u32_max_value;
7665 u32 umin_val = src_reg->u32_min_value;
7667 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
7668 * be negative, then either:
7669 * 1) src_reg might be zero, so the sign bit of the result is
7670 * unknown, so we lose our signed bounds
7671 * 2) it's known negative, thus the unsigned bounds capture the
7673 * 3) the signed bounds cross zero, so they tell us nothing
7675 * If the value in dst_reg is known nonnegative, then again the
7676 * unsigned bounds capture the signed bounds.
7677 * Thus, in all cases it suffices to blow away our signed bounds
7678 * and rely on inferring new ones from the unsigned bounds and
7679 * var_off of the result.
7681 dst_reg->s32_min_value = S32_MIN;
7682 dst_reg->s32_max_value = S32_MAX;
7684 dst_reg->var_off = tnum_rshift(subreg, umin_val);
7685 dst_reg->u32_min_value >>= umax_val;
7686 dst_reg->u32_max_value >>= umin_val;
7688 __mark_reg64_unbounded(dst_reg);
7689 __update_reg32_bounds(dst_reg);
7692 static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
7693 struct bpf_reg_state *src_reg)
7695 u64 umax_val = src_reg->umax_value;
7696 u64 umin_val = src_reg->umin_value;
7698 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
7699 * be negative, then either:
7700 * 1) src_reg might be zero, so the sign bit of the result is
7701 * unknown, so we lose our signed bounds
7702 * 2) it's known negative, thus the unsigned bounds capture the
7704 * 3) the signed bounds cross zero, so they tell us nothing
7706 * If the value in dst_reg is known nonnegative, then again the
7707 * unsigned bounds capture the signed bounds.
7708 * Thus, in all cases it suffices to blow away our signed bounds
7709 * and rely on inferring new ones from the unsigned bounds and
7710 * var_off of the result.
7712 dst_reg->smin_value = S64_MIN;
7713 dst_reg->smax_value = S64_MAX;
7714 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
7715 dst_reg->umin_value >>= umax_val;
7716 dst_reg->umax_value >>= umin_val;
7718 /* Its not easy to operate on alu32 bounds here because it depends
7719 * on bits being shifted in. Take easy way out and mark unbounded
7720 * so we can recalculate later from tnum.
7722 __mark_reg32_unbounded(dst_reg);
7723 __update_reg_bounds(dst_reg);
7726 static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg,
7727 struct bpf_reg_state *src_reg)
7729 u64 umin_val = src_reg->u32_min_value;
7731 /* Upon reaching here, src_known is true and
7732 * umax_val is equal to umin_val.
7734 dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val);
7735 dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val);
7737 dst_reg->var_off = tnum_arshift(tnum_subreg(dst_reg->var_off), umin_val, 32);
7739 /* blow away the dst_reg umin_value/umax_value and rely on
7740 * dst_reg var_off to refine the result.
7742 dst_reg->u32_min_value = 0;
7743 dst_reg->u32_max_value = U32_MAX;
7745 __mark_reg64_unbounded(dst_reg);
7746 __update_reg32_bounds(dst_reg);
7749 static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
7750 struct bpf_reg_state *src_reg)
7752 u64 umin_val = src_reg->umin_value;
7754 /* Upon reaching here, src_known is true and umax_val is equal
7757 dst_reg->smin_value >>= umin_val;
7758 dst_reg->smax_value >>= umin_val;
7760 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val, 64);
7762 /* blow away the dst_reg umin_value/umax_value and rely on
7763 * dst_reg var_off to refine the result.
7765 dst_reg->umin_value = 0;
7766 dst_reg->umax_value = U64_MAX;
7768 /* Its not easy to operate on alu32 bounds here because it depends
7769 * on bits being shifted in from upper 32-bits. Take easy way out
7770 * and mark unbounded so we can recalculate later from tnum.
7772 __mark_reg32_unbounded(dst_reg);
7773 __update_reg_bounds(dst_reg);
7776 /* WARNING: This function does calculations on 64-bit values, but the actual
7777 * execution may occur on 32-bit values. Therefore, things like bitshifts
7778 * need extra checks in the 32-bit case.
7780 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
7781 struct bpf_insn *insn,
7782 struct bpf_reg_state *dst_reg,
7783 struct bpf_reg_state src_reg)
7785 struct bpf_reg_state *regs = cur_regs(env);
7786 u8 opcode = BPF_OP(insn->code);
7788 s64 smin_val, smax_val;
7789 u64 umin_val, umax_val;
7790 s32 s32_min_val, s32_max_val;
7791 u32 u32_min_val, u32_max_val;
7792 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
7793 bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
7796 smin_val = src_reg.smin_value;
7797 smax_val = src_reg.smax_value;
7798 umin_val = src_reg.umin_value;
7799 umax_val = src_reg.umax_value;
7801 s32_min_val = src_reg.s32_min_value;
7802 s32_max_val = src_reg.s32_max_value;
7803 u32_min_val = src_reg.u32_min_value;
7804 u32_max_val = src_reg.u32_max_value;
7807 src_known = tnum_subreg_is_const(src_reg.var_off);
7809 (s32_min_val != s32_max_val || u32_min_val != u32_max_val)) ||
7810 s32_min_val > s32_max_val || u32_min_val > u32_max_val) {
7811 /* Taint dst register if offset had invalid bounds
7812 * derived from e.g. dead branches.
7814 __mark_reg_unknown(env, dst_reg);
7818 src_known = tnum_is_const(src_reg.var_off);
7820 (smin_val != smax_val || umin_val != umax_val)) ||
7821 smin_val > smax_val || umin_val > umax_val) {
7822 /* Taint dst register if offset had invalid bounds
7823 * derived from e.g. dead branches.
7825 __mark_reg_unknown(env, dst_reg);
7831 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
7832 __mark_reg_unknown(env, dst_reg);
7836 if (sanitize_needed(opcode)) {
7837 ret = sanitize_val_alu(env, insn);
7839 return sanitize_err(env, insn, ret, NULL, NULL);
7842 /* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
7843 * There are two classes of instructions: The first class we track both
7844 * alu32 and alu64 sign/unsigned bounds independently this provides the
7845 * greatest amount of precision when alu operations are mixed with jmp32
7846 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
7847 * and BPF_OR. This is possible because these ops have fairly easy to
7848 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
7849 * See alu32 verifier tests for examples. The second class of
7850 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
7851 * with regards to tracking sign/unsigned bounds because the bits may
7852 * cross subreg boundaries in the alu64 case. When this happens we mark
7853 * the reg unbounded in the subreg bound space and use the resulting
7854 * tnum to calculate an approximation of the sign/unsigned bounds.
7858 scalar32_min_max_add(dst_reg, &src_reg);
7859 scalar_min_max_add(dst_reg, &src_reg);
7860 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
7863 scalar32_min_max_sub(dst_reg, &src_reg);
7864 scalar_min_max_sub(dst_reg, &src_reg);
7865 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
7868 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
7869 scalar32_min_max_mul(dst_reg, &src_reg);
7870 scalar_min_max_mul(dst_reg, &src_reg);
7873 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
7874 scalar32_min_max_and(dst_reg, &src_reg);
7875 scalar_min_max_and(dst_reg, &src_reg);
7878 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
7879 scalar32_min_max_or(dst_reg, &src_reg);
7880 scalar_min_max_or(dst_reg, &src_reg);
7883 dst_reg->var_off = tnum_xor(dst_reg->var_off, src_reg.var_off);
7884 scalar32_min_max_xor(dst_reg, &src_reg);
7885 scalar_min_max_xor(dst_reg, &src_reg);
7888 if (umax_val >= insn_bitness) {
7889 /* Shifts greater than 31 or 63 are undefined.
7890 * This includes shifts by a negative number.
7892 mark_reg_unknown(env, regs, insn->dst_reg);
7896 scalar32_min_max_lsh(dst_reg, &src_reg);
7898 scalar_min_max_lsh(dst_reg, &src_reg);
7901 if (umax_val >= insn_bitness) {
7902 /* Shifts greater than 31 or 63 are undefined.
7903 * This includes shifts by a negative number.
7905 mark_reg_unknown(env, regs, insn->dst_reg);
7909 scalar32_min_max_rsh(dst_reg, &src_reg);
7911 scalar_min_max_rsh(dst_reg, &src_reg);
7914 if (umax_val >= insn_bitness) {
7915 /* Shifts greater than 31 or 63 are undefined.
7916 * This includes shifts by a negative number.
7918 mark_reg_unknown(env, regs, insn->dst_reg);
7922 scalar32_min_max_arsh(dst_reg, &src_reg);
7924 scalar_min_max_arsh(dst_reg, &src_reg);
7927 mark_reg_unknown(env, regs, insn->dst_reg);
7931 /* ALU32 ops are zero extended into 64bit register */
7933 zext_32_to_64(dst_reg);
7934 reg_bounds_sync(dst_reg);
7938 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
7941 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
7942 struct bpf_insn *insn)
7944 struct bpf_verifier_state *vstate = env->cur_state;
7945 struct bpf_func_state *state = vstate->frame[vstate->curframe];
7946 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
7947 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
7948 u8 opcode = BPF_OP(insn->code);
7951 dst_reg = ®s[insn->dst_reg];
7953 if (dst_reg->type != SCALAR_VALUE)
7956 /* Make sure ID is cleared otherwise dst_reg min/max could be
7957 * incorrectly propagated into other registers by find_equal_scalars()
7960 if (BPF_SRC(insn->code) == BPF_X) {
7961 src_reg = ®s[insn->src_reg];
7962 if (src_reg->type != SCALAR_VALUE) {
7963 if (dst_reg->type != SCALAR_VALUE) {
7964 /* Combining two pointers by any ALU op yields
7965 * an arbitrary scalar. Disallow all math except
7966 * pointer subtraction
7968 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
7969 mark_reg_unknown(env, regs, insn->dst_reg);
7972 verbose(env, "R%d pointer %s pointer prohibited\n",
7974 bpf_alu_string[opcode >> 4]);
7977 /* scalar += pointer
7978 * This is legal, but we have to reverse our
7979 * src/dest handling in computing the range
7981 err = mark_chain_precision(env, insn->dst_reg);
7984 return adjust_ptr_min_max_vals(env, insn,
7987 } else if (ptr_reg) {
7988 /* pointer += scalar */
7989 err = mark_chain_precision(env, insn->src_reg);
7992 return adjust_ptr_min_max_vals(env, insn,
7996 /* Pretend the src is a reg with a known value, since we only
7997 * need to be able to read from this state.
7999 off_reg.type = SCALAR_VALUE;
8000 __mark_reg_known(&off_reg, insn->imm);
8002 if (ptr_reg) /* pointer += K */
8003 return adjust_ptr_min_max_vals(env, insn,
8007 /* Got here implies adding two SCALAR_VALUEs */
8008 if (WARN_ON_ONCE(ptr_reg)) {
8009 print_verifier_state(env, state);
8010 verbose(env, "verifier internal error: unexpected ptr_reg\n");
8013 if (WARN_ON(!src_reg)) {
8014 print_verifier_state(env, state);
8015 verbose(env, "verifier internal error: no src_reg\n");
8018 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
8021 /* check validity of 32-bit and 64-bit arithmetic operations */
8022 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
8024 struct bpf_reg_state *regs = cur_regs(env);
8025 u8 opcode = BPF_OP(insn->code);
8028 if (opcode == BPF_END || opcode == BPF_NEG) {
8029 if (opcode == BPF_NEG) {
8030 if (BPF_SRC(insn->code) != 0 ||
8031 insn->src_reg != BPF_REG_0 ||
8032 insn->off != 0 || insn->imm != 0) {
8033 verbose(env, "BPF_NEG uses reserved fields\n");
8037 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
8038 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
8039 BPF_CLASS(insn->code) == BPF_ALU64) {
8040 verbose(env, "BPF_END uses reserved fields\n");
8045 /* check src operand */
8046 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
8050 if (is_pointer_value(env, insn->dst_reg)) {
8051 verbose(env, "R%d pointer arithmetic prohibited\n",
8056 /* check dest operand */
8057 err = check_reg_arg(env, insn->dst_reg, DST_OP);
8061 } else if (opcode == BPF_MOV) {
8063 if (BPF_SRC(insn->code) == BPF_X) {
8064 if (insn->imm != 0 || insn->off != 0) {
8065 verbose(env, "BPF_MOV uses reserved fields\n");
8069 /* check src operand */
8070 err = check_reg_arg(env, insn->src_reg, SRC_OP);
8074 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
8075 verbose(env, "BPF_MOV uses reserved fields\n");
8080 /* check dest operand, mark as required later */
8081 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
8085 if (BPF_SRC(insn->code) == BPF_X) {
8086 struct bpf_reg_state *src_reg = regs + insn->src_reg;
8087 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
8089 if (BPF_CLASS(insn->code) == BPF_ALU64) {
8091 * copy register state to dest reg
8093 if (src_reg->type == SCALAR_VALUE && !src_reg->id)
8094 /* Assign src and dst registers the same ID
8095 * that will be used by find_equal_scalars()
8096 * to propagate min/max range.
8098 src_reg->id = ++env->id_gen;
8099 *dst_reg = *src_reg;
8100 dst_reg->live |= REG_LIVE_WRITTEN;
8101 dst_reg->subreg_def = DEF_NOT_SUBREG;
8104 if (is_pointer_value(env, insn->src_reg)) {
8106 "R%d partial copy of pointer\n",
8109 } else if (src_reg->type == SCALAR_VALUE) {
8110 *dst_reg = *src_reg;
8111 /* Make sure ID is cleared otherwise
8112 * dst_reg min/max could be incorrectly
8113 * propagated into src_reg by find_equal_scalars()
8116 dst_reg->live |= REG_LIVE_WRITTEN;
8117 dst_reg->subreg_def = env->insn_idx + 1;
8119 mark_reg_unknown(env, regs,
8122 zext_32_to_64(dst_reg);
8123 reg_bounds_sync(dst_reg);
8127 * remember the value we stored into this reg
8129 /* clear any state __mark_reg_known doesn't set */
8130 mark_reg_unknown(env, regs, insn->dst_reg);
8131 regs[insn->dst_reg].type = SCALAR_VALUE;
8132 if (BPF_CLASS(insn->code) == BPF_ALU64) {
8133 __mark_reg_known(regs + insn->dst_reg,
8136 __mark_reg_known(regs + insn->dst_reg,
8141 } else if (opcode > BPF_END) {
8142 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
8145 } else { /* all other ALU ops: and, sub, xor, add, ... */
8147 if (BPF_SRC(insn->code) == BPF_X) {
8148 if (insn->imm != 0 || insn->off != 0) {
8149 verbose(env, "BPF_ALU uses reserved fields\n");
8152 /* check src1 operand */
8153 err = check_reg_arg(env, insn->src_reg, SRC_OP);
8157 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
8158 verbose(env, "BPF_ALU uses reserved fields\n");
8163 /* check src2 operand */
8164 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
8168 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
8169 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
8170 verbose(env, "div by zero\n");
8174 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
8175 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
8176 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
8178 if (insn->imm < 0 || insn->imm >= size) {
8179 verbose(env, "invalid shift %d\n", insn->imm);
8184 /* check dest operand */
8185 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
8189 return adjust_reg_min_max_vals(env, insn);
8195 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
8196 struct bpf_reg_state *dst_reg,
8197 enum bpf_reg_type type,
8198 bool range_right_open)
8200 struct bpf_func_state *state;
8201 struct bpf_reg_state *reg;
8204 if (dst_reg->off < 0 ||
8205 (dst_reg->off == 0 && range_right_open))
8206 /* This doesn't give us any range */
8209 if (dst_reg->umax_value > MAX_PACKET_OFF ||
8210 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
8211 /* Risk of overflow. For instance, ptr + (1<<63) may be less
8212 * than pkt_end, but that's because it's also less than pkt.
8216 new_range = dst_reg->off;
8217 if (range_right_open)
8220 /* Examples for register markings:
8222 * pkt_data in dst register:
8226 * if (r2 > pkt_end) goto <handle exception>
8231 * if (r2 < pkt_end) goto <access okay>
8232 * <handle exception>
8235 * r2 == dst_reg, pkt_end == src_reg
8236 * r2=pkt(id=n,off=8,r=0)
8237 * r3=pkt(id=n,off=0,r=0)
8239 * pkt_data in src register:
8243 * if (pkt_end >= r2) goto <access okay>
8244 * <handle exception>
8248 * if (pkt_end <= r2) goto <handle exception>
8252 * pkt_end == dst_reg, r2 == src_reg
8253 * r2=pkt(id=n,off=8,r=0)
8254 * r3=pkt(id=n,off=0,r=0)
8256 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
8257 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
8258 * and [r3, r3 + 8-1) respectively is safe to access depending on
8262 /* If our ids match, then we must have the same max_value. And we
8263 * don't care about the other reg's fixed offset, since if it's too big
8264 * the range won't allow anything.
8265 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
8267 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
8268 if (reg->type == type && reg->id == dst_reg->id)
8269 /* keep the maximum range already checked */
8270 reg->range = max(reg->range, new_range);
8274 static int is_branch32_taken(struct bpf_reg_state *reg, u32 val, u8 opcode)
8276 struct tnum subreg = tnum_subreg(reg->var_off);
8277 s32 sval = (s32)val;
8281 if (tnum_is_const(subreg))
8282 return !!tnum_equals_const(subreg, val);
8285 if (tnum_is_const(subreg))
8286 return !tnum_equals_const(subreg, val);
8289 if ((~subreg.mask & subreg.value) & val)
8291 if (!((subreg.mask | subreg.value) & val))
8295 if (reg->u32_min_value > val)
8297 else if (reg->u32_max_value <= val)
8301 if (reg->s32_min_value > sval)
8303 else if (reg->s32_max_value <= sval)
8307 if (reg->u32_max_value < val)
8309 else if (reg->u32_min_value >= val)
8313 if (reg->s32_max_value < sval)
8315 else if (reg->s32_min_value >= sval)
8319 if (reg->u32_min_value >= val)
8321 else if (reg->u32_max_value < val)
8325 if (reg->s32_min_value >= sval)
8327 else if (reg->s32_max_value < sval)
8331 if (reg->u32_max_value <= val)
8333 else if (reg->u32_min_value > val)
8337 if (reg->s32_max_value <= sval)
8339 else if (reg->s32_min_value > sval)
8348 static int is_branch64_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
8350 s64 sval = (s64)val;
8354 if (tnum_is_const(reg->var_off))
8355 return !!tnum_equals_const(reg->var_off, val);
8358 if (tnum_is_const(reg->var_off))
8359 return !tnum_equals_const(reg->var_off, val);
8362 if ((~reg->var_off.mask & reg->var_off.value) & val)
8364 if (!((reg->var_off.mask | reg->var_off.value) & val))
8368 if (reg->umin_value > val)
8370 else if (reg->umax_value <= val)
8374 if (reg->smin_value > sval)
8376 else if (reg->smax_value <= sval)
8380 if (reg->umax_value < val)
8382 else if (reg->umin_value >= val)
8386 if (reg->smax_value < sval)
8388 else if (reg->smin_value >= sval)
8392 if (reg->umin_value >= val)
8394 else if (reg->umax_value < val)
8398 if (reg->smin_value >= sval)
8400 else if (reg->smax_value < sval)
8404 if (reg->umax_value <= val)
8406 else if (reg->umin_value > val)
8410 if (reg->smax_value <= sval)
8412 else if (reg->smin_value > sval)
8420 /* compute branch direction of the expression "if (reg opcode val) goto target;"
8422 * 1 - branch will be taken and "goto target" will be executed
8423 * 0 - branch will not be taken and fall-through to next insn
8424 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value
8427 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
8430 if (__is_pointer_value(false, reg)) {
8431 if (!reg_type_not_null(reg->type))
8434 /* If pointer is valid tests against zero will fail so we can
8435 * use this to direct branch taken.
8451 return is_branch32_taken(reg, val, opcode);
8452 return is_branch64_taken(reg, val, opcode);
8455 static int flip_opcode(u32 opcode)
8457 /* How can we transform "a <op> b" into "b <op> a"? */
8458 static const u8 opcode_flip[16] = {
8459 /* these stay the same */
8460 [BPF_JEQ >> 4] = BPF_JEQ,
8461 [BPF_JNE >> 4] = BPF_JNE,
8462 [BPF_JSET >> 4] = BPF_JSET,
8463 /* these swap "lesser" and "greater" (L and G in the opcodes) */
8464 [BPF_JGE >> 4] = BPF_JLE,
8465 [BPF_JGT >> 4] = BPF_JLT,
8466 [BPF_JLE >> 4] = BPF_JGE,
8467 [BPF_JLT >> 4] = BPF_JGT,
8468 [BPF_JSGE >> 4] = BPF_JSLE,
8469 [BPF_JSGT >> 4] = BPF_JSLT,
8470 [BPF_JSLE >> 4] = BPF_JSGE,
8471 [BPF_JSLT >> 4] = BPF_JSGT
8473 return opcode_flip[opcode >> 4];
8476 static int is_pkt_ptr_branch_taken(struct bpf_reg_state *dst_reg,
8477 struct bpf_reg_state *src_reg,
8480 struct bpf_reg_state *pkt;
8482 if (src_reg->type == PTR_TO_PACKET_END) {
8484 } else if (dst_reg->type == PTR_TO_PACKET_END) {
8486 opcode = flip_opcode(opcode);
8491 if (pkt->range >= 0)
8496 /* pkt <= pkt_end */
8500 if (pkt->range == BEYOND_PKT_END)
8501 /* pkt has at last one extra byte beyond pkt_end */
8502 return opcode == BPF_JGT;
8508 /* pkt >= pkt_end */
8509 if (pkt->range == BEYOND_PKT_END || pkt->range == AT_PKT_END)
8510 return opcode == BPF_JGE;
8516 /* Adjusts the register min/max values in the case that the dst_reg is the
8517 * variable register that we are working on, and src_reg is a constant or we're
8518 * simply doing a BPF_K check.
8519 * In JEQ/JNE cases we also adjust the var_off values.
8521 static void reg_set_min_max(struct bpf_reg_state *true_reg,
8522 struct bpf_reg_state *false_reg,
8524 u8 opcode, bool is_jmp32)
8526 struct tnum false_32off = tnum_subreg(false_reg->var_off);
8527 struct tnum false_64off = false_reg->var_off;
8528 struct tnum true_32off = tnum_subreg(true_reg->var_off);
8529 struct tnum true_64off = true_reg->var_off;
8530 s64 sval = (s64)val;
8531 s32 sval32 = (s32)val32;
8533 /* If the dst_reg is a pointer, we can't learn anything about its
8534 * variable offset from the compare (unless src_reg were a pointer into
8535 * the same object, but we don't bother with that.
8536 * Since false_reg and true_reg have the same type by construction, we
8537 * only need to check one of them for pointerness.
8539 if (__is_pointer_value(false, false_reg))
8543 /* JEQ/JNE comparison doesn't change the register equivalence.
8546 * if (r1 == 42) goto label;
8548 * label: // here both r1 and r2 are known to be 42.
8550 * Hence when marking register as known preserve it's ID.
8554 __mark_reg32_known(true_reg, val32);
8555 true_32off = tnum_subreg(true_reg->var_off);
8557 ___mark_reg_known(true_reg, val);
8558 true_64off = true_reg->var_off;
8563 __mark_reg32_known(false_reg, val32);
8564 false_32off = tnum_subreg(false_reg->var_off);
8566 ___mark_reg_known(false_reg, val);
8567 false_64off = false_reg->var_off;
8572 false_32off = tnum_and(false_32off, tnum_const(~val32));
8573 if (is_power_of_2(val32))
8574 true_32off = tnum_or(true_32off,
8577 false_64off = tnum_and(false_64off, tnum_const(~val));
8578 if (is_power_of_2(val))
8579 true_64off = tnum_or(true_64off,
8587 u32 false_umax = opcode == BPF_JGT ? val32 : val32 - 1;
8588 u32 true_umin = opcode == BPF_JGT ? val32 + 1 : val32;
8590 false_reg->u32_max_value = min(false_reg->u32_max_value,
8592 true_reg->u32_min_value = max(true_reg->u32_min_value,
8595 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
8596 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
8598 false_reg->umax_value = min(false_reg->umax_value, false_umax);
8599 true_reg->umin_value = max(true_reg->umin_value, true_umin);
8607 s32 false_smax = opcode == BPF_JSGT ? sval32 : sval32 - 1;
8608 s32 true_smin = opcode == BPF_JSGT ? sval32 + 1 : sval32;
8610 false_reg->s32_max_value = min(false_reg->s32_max_value, false_smax);
8611 true_reg->s32_min_value = max(true_reg->s32_min_value, true_smin);
8613 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
8614 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
8616 false_reg->smax_value = min(false_reg->smax_value, false_smax);
8617 true_reg->smin_value = max(true_reg->smin_value, true_smin);
8625 u32 false_umin = opcode == BPF_JLT ? val32 : val32 + 1;
8626 u32 true_umax = opcode == BPF_JLT ? val32 - 1 : val32;
8628 false_reg->u32_min_value = max(false_reg->u32_min_value,
8630 true_reg->u32_max_value = min(true_reg->u32_max_value,
8633 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
8634 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
8636 false_reg->umin_value = max(false_reg->umin_value, false_umin);
8637 true_reg->umax_value = min(true_reg->umax_value, true_umax);
8645 s32 false_smin = opcode == BPF_JSLT ? sval32 : sval32 + 1;
8646 s32 true_smax = opcode == BPF_JSLT ? sval32 - 1 : sval32;
8648 false_reg->s32_min_value = max(false_reg->s32_min_value, false_smin);
8649 true_reg->s32_max_value = min(true_reg->s32_max_value, true_smax);
8651 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
8652 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
8654 false_reg->smin_value = max(false_reg->smin_value, false_smin);
8655 true_reg->smax_value = min(true_reg->smax_value, true_smax);
8664 false_reg->var_off = tnum_or(tnum_clear_subreg(false_64off),
8665 tnum_subreg(false_32off));
8666 true_reg->var_off = tnum_or(tnum_clear_subreg(true_64off),
8667 tnum_subreg(true_32off));
8668 __reg_combine_32_into_64(false_reg);
8669 __reg_combine_32_into_64(true_reg);
8671 false_reg->var_off = false_64off;
8672 true_reg->var_off = true_64off;
8673 __reg_combine_64_into_32(false_reg);
8674 __reg_combine_64_into_32(true_reg);
8678 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
8681 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
8682 struct bpf_reg_state *false_reg,
8684 u8 opcode, bool is_jmp32)
8686 opcode = flip_opcode(opcode);
8687 /* This uses zero as "not present in table"; luckily the zero opcode,
8688 * BPF_JA, can't get here.
8691 reg_set_min_max(true_reg, false_reg, val, val32, opcode, is_jmp32);
8694 /* Regs are known to be equal, so intersect their min/max/var_off */
8695 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
8696 struct bpf_reg_state *dst_reg)
8698 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
8699 dst_reg->umin_value);
8700 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
8701 dst_reg->umax_value);
8702 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
8703 dst_reg->smin_value);
8704 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
8705 dst_reg->smax_value);
8706 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
8708 reg_bounds_sync(src_reg);
8709 reg_bounds_sync(dst_reg);
8712 static void reg_combine_min_max(struct bpf_reg_state *true_src,
8713 struct bpf_reg_state *true_dst,
8714 struct bpf_reg_state *false_src,
8715 struct bpf_reg_state *false_dst,
8720 __reg_combine_min_max(true_src, true_dst);
8723 __reg_combine_min_max(false_src, false_dst);
8728 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
8729 struct bpf_reg_state *reg, u32 id,
8732 if (type_may_be_null(reg->type) && reg->id == id &&
8733 !WARN_ON_ONCE(!reg->id)) {
8734 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
8735 !tnum_equals_const(reg->var_off, 0) ||
8737 /* Old offset (both fixed and variable parts) should
8738 * have been known-zero, because we don't allow pointer
8739 * arithmetic on pointers that might be NULL. If we
8740 * see this happening, don't convert the register.
8745 reg->type = SCALAR_VALUE;
8746 /* We don't need id and ref_obj_id from this point
8747 * onwards anymore, thus we should better reset it,
8748 * so that state pruning has chances to take effect.
8751 reg->ref_obj_id = 0;
8756 mark_ptr_not_null_reg(reg);
8758 if (!reg_may_point_to_spin_lock(reg)) {
8759 /* For not-NULL ptr, reg->ref_obj_id will be reset
8760 * in release_reference().
8762 * reg->id is still used by spin_lock ptr. Other
8763 * than spin_lock ptr type, reg->id can be reset.
8770 /* The logic is similar to find_good_pkt_pointers(), both could eventually
8771 * be folded together at some point.
8773 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
8776 struct bpf_func_state *state = vstate->frame[vstate->curframe];
8777 struct bpf_reg_state *regs = state->regs, *reg;
8778 u32 ref_obj_id = regs[regno].ref_obj_id;
8779 u32 id = regs[regno].id;
8781 if (ref_obj_id && ref_obj_id == id && is_null)
8782 /* regs[regno] is in the " == NULL" branch.
8783 * No one could have freed the reference state before
8784 * doing the NULL check.
8786 WARN_ON_ONCE(release_reference_state(state, id));
8788 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
8789 mark_ptr_or_null_reg(state, reg, id, is_null);
8793 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
8794 struct bpf_reg_state *dst_reg,
8795 struct bpf_reg_state *src_reg,
8796 struct bpf_verifier_state *this_branch,
8797 struct bpf_verifier_state *other_branch)
8799 if (BPF_SRC(insn->code) != BPF_X)
8802 /* Pointers are always 64-bit. */
8803 if (BPF_CLASS(insn->code) == BPF_JMP32)
8806 switch (BPF_OP(insn->code)) {
8808 if ((dst_reg->type == PTR_TO_PACKET &&
8809 src_reg->type == PTR_TO_PACKET_END) ||
8810 (dst_reg->type == PTR_TO_PACKET_META &&
8811 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
8812 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
8813 find_good_pkt_pointers(this_branch, dst_reg,
8814 dst_reg->type, false);
8815 mark_pkt_end(other_branch, insn->dst_reg, true);
8816 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
8817 src_reg->type == PTR_TO_PACKET) ||
8818 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
8819 src_reg->type == PTR_TO_PACKET_META)) {
8820 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
8821 find_good_pkt_pointers(other_branch, src_reg,
8822 src_reg->type, true);
8823 mark_pkt_end(this_branch, insn->src_reg, false);
8829 if ((dst_reg->type == PTR_TO_PACKET &&
8830 src_reg->type == PTR_TO_PACKET_END) ||
8831 (dst_reg->type == PTR_TO_PACKET_META &&
8832 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
8833 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
8834 find_good_pkt_pointers(other_branch, dst_reg,
8835 dst_reg->type, true);
8836 mark_pkt_end(this_branch, insn->dst_reg, false);
8837 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
8838 src_reg->type == PTR_TO_PACKET) ||
8839 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
8840 src_reg->type == PTR_TO_PACKET_META)) {
8841 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
8842 find_good_pkt_pointers(this_branch, src_reg,
8843 src_reg->type, false);
8844 mark_pkt_end(other_branch, insn->src_reg, true);
8850 if ((dst_reg->type == PTR_TO_PACKET &&
8851 src_reg->type == PTR_TO_PACKET_END) ||
8852 (dst_reg->type == PTR_TO_PACKET_META &&
8853 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
8854 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
8855 find_good_pkt_pointers(this_branch, dst_reg,
8856 dst_reg->type, true);
8857 mark_pkt_end(other_branch, insn->dst_reg, false);
8858 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
8859 src_reg->type == PTR_TO_PACKET) ||
8860 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
8861 src_reg->type == PTR_TO_PACKET_META)) {
8862 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
8863 find_good_pkt_pointers(other_branch, src_reg,
8864 src_reg->type, false);
8865 mark_pkt_end(this_branch, insn->src_reg, true);
8871 if ((dst_reg->type == PTR_TO_PACKET &&
8872 src_reg->type == PTR_TO_PACKET_END) ||
8873 (dst_reg->type == PTR_TO_PACKET_META &&
8874 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
8875 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
8876 find_good_pkt_pointers(other_branch, dst_reg,
8877 dst_reg->type, false);
8878 mark_pkt_end(this_branch, insn->dst_reg, true);
8879 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
8880 src_reg->type == PTR_TO_PACKET) ||
8881 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
8882 src_reg->type == PTR_TO_PACKET_META)) {
8883 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
8884 find_good_pkt_pointers(this_branch, src_reg,
8885 src_reg->type, true);
8886 mark_pkt_end(other_branch, insn->src_reg, false);
8898 static void find_equal_scalars(struct bpf_verifier_state *vstate,
8899 struct bpf_reg_state *known_reg)
8901 struct bpf_func_state *state;
8902 struct bpf_reg_state *reg;
8904 bpf_for_each_reg_in_vstate(vstate, state, reg, ({
8905 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
8910 static int check_cond_jmp_op(struct bpf_verifier_env *env,
8911 struct bpf_insn *insn, int *insn_idx)
8913 struct bpf_verifier_state *this_branch = env->cur_state;
8914 struct bpf_verifier_state *other_branch;
8915 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
8916 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
8917 u8 opcode = BPF_OP(insn->code);
8922 /* Only conditional jumps are expected to reach here. */
8923 if (opcode == BPF_JA || opcode > BPF_JSLE) {
8924 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
8928 if (BPF_SRC(insn->code) == BPF_X) {
8929 if (insn->imm != 0) {
8930 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
8934 /* check src1 operand */
8935 err = check_reg_arg(env, insn->src_reg, SRC_OP);
8939 if (is_pointer_value(env, insn->src_reg)) {
8940 verbose(env, "R%d pointer comparison prohibited\n",
8944 src_reg = ®s[insn->src_reg];
8946 if (insn->src_reg != BPF_REG_0) {
8947 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
8952 /* check src2 operand */
8953 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
8957 dst_reg = ®s[insn->dst_reg];
8958 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
8960 if (BPF_SRC(insn->code) == BPF_K) {
8961 pred = is_branch_taken(dst_reg, insn->imm, opcode, is_jmp32);
8962 } else if (src_reg->type == SCALAR_VALUE &&
8963 is_jmp32 && tnum_is_const(tnum_subreg(src_reg->var_off))) {
8964 pred = is_branch_taken(dst_reg,
8965 tnum_subreg(src_reg->var_off).value,
8968 } else if (src_reg->type == SCALAR_VALUE &&
8969 !is_jmp32 && tnum_is_const(src_reg->var_off)) {
8970 pred = is_branch_taken(dst_reg,
8971 src_reg->var_off.value,
8974 } else if (reg_is_pkt_pointer_any(dst_reg) &&
8975 reg_is_pkt_pointer_any(src_reg) &&
8977 pred = is_pkt_ptr_branch_taken(dst_reg, src_reg, opcode);
8981 /* If we get here with a dst_reg pointer type it is because
8982 * above is_branch_taken() special cased the 0 comparison.
8984 if (!__is_pointer_value(false, dst_reg))
8985 err = mark_chain_precision(env, insn->dst_reg);
8986 if (BPF_SRC(insn->code) == BPF_X && !err &&
8987 !__is_pointer_value(false, src_reg))
8988 err = mark_chain_precision(env, insn->src_reg);
8994 /* Only follow the goto, ignore fall-through. If needed, push
8995 * the fall-through branch for simulation under speculative
8998 if (!env->bypass_spec_v1 &&
8999 !sanitize_speculative_path(env, insn, *insn_idx + 1,
9002 *insn_idx += insn->off;
9004 } else if (pred == 0) {
9005 /* Only follow the fall-through branch, since that's where the
9006 * program will go. If needed, push the goto branch for
9007 * simulation under speculative execution.
9009 if (!env->bypass_spec_v1 &&
9010 !sanitize_speculative_path(env, insn,
9011 *insn_idx + insn->off + 1,
9017 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
9021 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
9023 /* detect if we are comparing against a constant value so we can adjust
9024 * our min/max values for our dst register.
9025 * this is only legit if both are scalars (or pointers to the same
9026 * object, I suppose, but we don't support that right now), because
9027 * otherwise the different base pointers mean the offsets aren't
9030 if (BPF_SRC(insn->code) == BPF_X) {
9031 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
9033 if (dst_reg->type == SCALAR_VALUE &&
9034 src_reg->type == SCALAR_VALUE) {
9035 if (tnum_is_const(src_reg->var_off) ||
9037 tnum_is_const(tnum_subreg(src_reg->var_off))))
9038 reg_set_min_max(&other_branch_regs[insn->dst_reg],
9040 src_reg->var_off.value,
9041 tnum_subreg(src_reg->var_off).value,
9043 else if (tnum_is_const(dst_reg->var_off) ||
9045 tnum_is_const(tnum_subreg(dst_reg->var_off))))
9046 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
9048 dst_reg->var_off.value,
9049 tnum_subreg(dst_reg->var_off).value,
9051 else if (!is_jmp32 &&
9052 (opcode == BPF_JEQ || opcode == BPF_JNE))
9053 /* Comparing for equality, we can combine knowledge */
9054 reg_combine_min_max(&other_branch_regs[insn->src_reg],
9055 &other_branch_regs[insn->dst_reg],
9056 src_reg, dst_reg, opcode);
9058 !WARN_ON_ONCE(src_reg->id != other_branch_regs[insn->src_reg].id)) {
9059 find_equal_scalars(this_branch, src_reg);
9060 find_equal_scalars(other_branch, &other_branch_regs[insn->src_reg]);
9064 } else if (dst_reg->type == SCALAR_VALUE) {
9065 reg_set_min_max(&other_branch_regs[insn->dst_reg],
9066 dst_reg, insn->imm, (u32)insn->imm,
9070 if (dst_reg->type == SCALAR_VALUE && dst_reg->id &&
9071 !WARN_ON_ONCE(dst_reg->id != other_branch_regs[insn->dst_reg].id)) {
9072 find_equal_scalars(this_branch, dst_reg);
9073 find_equal_scalars(other_branch, &other_branch_regs[insn->dst_reg]);
9076 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
9077 * NOTE: these optimizations below are related with pointer comparison
9078 * which will never be JMP32.
9080 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
9081 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
9082 type_may_be_null(dst_reg->type)) {
9083 /* Mark all identical registers in each branch as either
9084 * safe or unknown depending R == 0 or R != 0 conditional.
9086 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
9088 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
9090 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
9091 this_branch, other_branch) &&
9092 is_pointer_value(env, insn->dst_reg)) {
9093 verbose(env, "R%d pointer comparison prohibited\n",
9097 if (env->log.level & BPF_LOG_LEVEL)
9098 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
9102 /* verify BPF_LD_IMM64 instruction */
9103 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
9105 struct bpf_insn_aux_data *aux = cur_aux(env);
9106 struct bpf_reg_state *regs = cur_regs(env);
9107 struct bpf_reg_state *dst_reg;
9108 struct bpf_map *map;
9111 if (BPF_SIZE(insn->code) != BPF_DW) {
9112 verbose(env, "invalid BPF_LD_IMM insn\n");
9115 if (insn->off != 0) {
9116 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
9120 err = check_reg_arg(env, insn->dst_reg, DST_OP);
9124 dst_reg = ®s[insn->dst_reg];
9125 if (insn->src_reg == 0) {
9126 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
9128 dst_reg->type = SCALAR_VALUE;
9129 __mark_reg_known(®s[insn->dst_reg], imm);
9133 /* All special src_reg cases are listed below. From this point onwards
9134 * we either succeed and assign a corresponding dst_reg->type after
9135 * zeroing the offset, or fail and reject the program.
9137 mark_reg_known_zero(env, regs, insn->dst_reg);
9139 if (insn->src_reg == BPF_PSEUDO_BTF_ID) {
9140 dst_reg->type = aux->btf_var.reg_type;
9141 switch (base_type(dst_reg->type)) {
9143 dst_reg->mem_size = aux->btf_var.mem_size;
9146 case PTR_TO_PERCPU_BTF_ID:
9147 dst_reg->btf = aux->btf_var.btf;
9148 dst_reg->btf_id = aux->btf_var.btf_id;
9151 verbose(env, "bpf verifier is misconfigured\n");
9157 if (insn->src_reg == BPF_PSEUDO_FUNC) {
9158 struct bpf_prog_aux *aux = env->prog->aux;
9159 u32 subprogno = find_subprog(env,
9160 env->insn_idx + insn->imm + 1);
9162 if (!aux->func_info) {
9163 verbose(env, "missing btf func_info\n");
9166 if (aux->func_info_aux[subprogno].linkage != BTF_FUNC_STATIC) {
9167 verbose(env, "callback function not static\n");
9171 dst_reg->type = PTR_TO_FUNC;
9172 dst_reg->subprogno = subprogno;
9176 map = env->used_maps[aux->map_index];
9177 dst_reg->map_ptr = map;
9179 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE ||
9180 insn->src_reg == BPF_PSEUDO_MAP_IDX_VALUE) {
9181 dst_reg->type = PTR_TO_MAP_VALUE;
9182 dst_reg->off = aux->map_off;
9183 if (map_value_has_spin_lock(map))
9184 dst_reg->id = ++env->id_gen;
9185 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD ||
9186 insn->src_reg == BPF_PSEUDO_MAP_IDX) {
9187 dst_reg->type = CONST_PTR_TO_MAP;
9189 verbose(env, "bpf verifier is misconfigured\n");
9196 static bool may_access_skb(enum bpf_prog_type type)
9199 case BPF_PROG_TYPE_SOCKET_FILTER:
9200 case BPF_PROG_TYPE_SCHED_CLS:
9201 case BPF_PROG_TYPE_SCHED_ACT:
9208 /* verify safety of LD_ABS|LD_IND instructions:
9209 * - they can only appear in the programs where ctx == skb
9210 * - since they are wrappers of function calls, they scratch R1-R5 registers,
9211 * preserve R6-R9, and store return value into R0
9214 * ctx == skb == R6 == CTX
9217 * SRC == any register
9218 * IMM == 32-bit immediate
9221 * R0 - 8/16/32-bit skb data converted to cpu endianness
9223 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
9225 struct bpf_reg_state *regs = cur_regs(env);
9226 static const int ctx_reg = BPF_REG_6;
9227 u8 mode = BPF_MODE(insn->code);
9230 if (!may_access_skb(resolve_prog_type(env->prog))) {
9231 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
9235 if (!env->ops->gen_ld_abs) {
9236 verbose(env, "bpf verifier is misconfigured\n");
9240 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
9241 BPF_SIZE(insn->code) == BPF_DW ||
9242 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
9243 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
9247 /* check whether implicit source operand (register R6) is readable */
9248 err = check_reg_arg(env, ctx_reg, SRC_OP);
9252 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
9253 * gen_ld_abs() may terminate the program at runtime, leading to
9256 err = check_reference_leak(env);
9258 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
9262 if (env->cur_state->active_spin_lock) {
9263 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
9267 if (regs[ctx_reg].type != PTR_TO_CTX) {
9269 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
9273 if (mode == BPF_IND) {
9274 /* check explicit source operand */
9275 err = check_reg_arg(env, insn->src_reg, SRC_OP);
9280 err = check_ctx_reg(env, ®s[ctx_reg], ctx_reg);
9284 /* reset caller saved regs to unreadable */
9285 for (i = 0; i < CALLER_SAVED_REGS; i++) {
9286 mark_reg_not_init(env, regs, caller_saved[i]);
9287 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
9290 /* mark destination R0 register as readable, since it contains
9291 * the value fetched from the packet.
9292 * Already marked as written above.
9294 mark_reg_unknown(env, regs, BPF_REG_0);
9295 /* ld_abs load up to 32-bit skb data. */
9296 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
9300 static int check_return_code(struct bpf_verifier_env *env)
9302 struct tnum enforce_attach_type_range = tnum_unknown;
9303 const struct bpf_prog *prog = env->prog;
9304 struct bpf_reg_state *reg;
9305 struct tnum range = tnum_range(0, 1);
9306 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
9308 struct bpf_func_state *frame = env->cur_state->frame[0];
9309 const bool is_subprog = frame->subprogno;
9311 /* LSM and struct_ops func-ptr's return type could be "void" */
9313 (prog_type == BPF_PROG_TYPE_STRUCT_OPS ||
9314 prog_type == BPF_PROG_TYPE_LSM) &&
9315 !prog->aux->attach_func_proto->type)
9318 /* eBPF calling convention is such that R0 is used
9319 * to return the value from eBPF program.
9320 * Make sure that it's readable at this time
9321 * of bpf_exit, which means that program wrote
9322 * something into it earlier
9324 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
9328 if (is_pointer_value(env, BPF_REG_0)) {
9329 verbose(env, "R0 leaks addr as return value\n");
9333 reg = cur_regs(env) + BPF_REG_0;
9335 if (frame->in_async_callback_fn) {
9336 /* enforce return zero from async callbacks like timer */
9337 if (reg->type != SCALAR_VALUE) {
9338 verbose(env, "In async callback the register R0 is not a known value (%s)\n",
9339 reg_type_str(env, reg->type));
9343 if (!tnum_in(tnum_const(0), reg->var_off)) {
9344 verbose_invalid_scalar(env, reg, &range, "async callback", "R0");
9351 if (reg->type != SCALAR_VALUE) {
9352 verbose(env, "At subprogram exit the register R0 is not a scalar value (%s)\n",
9353 reg_type_str(env, reg->type));
9359 switch (prog_type) {
9360 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
9361 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
9362 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG ||
9363 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETPEERNAME ||
9364 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETPEERNAME ||
9365 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETSOCKNAME ||
9366 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETSOCKNAME)
9367 range = tnum_range(1, 1);
9368 if (env->prog->expected_attach_type == BPF_CGROUP_INET4_BIND ||
9369 env->prog->expected_attach_type == BPF_CGROUP_INET6_BIND)
9370 range = tnum_range(0, 3);
9372 case BPF_PROG_TYPE_CGROUP_SKB:
9373 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
9374 range = tnum_range(0, 3);
9375 enforce_attach_type_range = tnum_range(2, 3);
9378 case BPF_PROG_TYPE_CGROUP_SOCK:
9379 case BPF_PROG_TYPE_SOCK_OPS:
9380 case BPF_PROG_TYPE_CGROUP_DEVICE:
9381 case BPF_PROG_TYPE_CGROUP_SYSCTL:
9382 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
9384 case BPF_PROG_TYPE_RAW_TRACEPOINT:
9385 if (!env->prog->aux->attach_btf_id)
9387 range = tnum_const(0);
9389 case BPF_PROG_TYPE_TRACING:
9390 switch (env->prog->expected_attach_type) {
9391 case BPF_TRACE_FENTRY:
9392 case BPF_TRACE_FEXIT:
9393 range = tnum_const(0);
9395 case BPF_TRACE_RAW_TP:
9396 case BPF_MODIFY_RETURN:
9398 case BPF_TRACE_ITER:
9404 case BPF_PROG_TYPE_SK_LOOKUP:
9405 range = tnum_range(SK_DROP, SK_PASS);
9407 case BPF_PROG_TYPE_EXT:
9408 /* freplace program can return anything as its return value
9409 * depends on the to-be-replaced kernel func or bpf program.
9415 if (reg->type != SCALAR_VALUE) {
9416 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
9417 reg_type_str(env, reg->type));
9421 if (!tnum_in(range, reg->var_off)) {
9422 verbose_invalid_scalar(env, reg, &range, "program exit", "R0");
9426 if (!tnum_is_unknown(enforce_attach_type_range) &&
9427 tnum_in(enforce_attach_type_range, reg->var_off))
9428 env->prog->enforce_expected_attach_type = 1;
9432 /* non-recursive DFS pseudo code
9433 * 1 procedure DFS-iterative(G,v):
9434 * 2 label v as discovered
9435 * 3 let S be a stack
9437 * 5 while S is not empty
9439 * 7 if t is what we're looking for:
9441 * 9 for all edges e in G.adjacentEdges(t) do
9442 * 10 if edge e is already labelled
9443 * 11 continue with the next edge
9444 * 12 w <- G.adjacentVertex(t,e)
9445 * 13 if vertex w is not discovered and not explored
9446 * 14 label e as tree-edge
9447 * 15 label w as discovered
9450 * 18 else if vertex w is discovered
9451 * 19 label e as back-edge
9453 * 21 // vertex w is explored
9454 * 22 label e as forward- or cross-edge
9455 * 23 label t as explored
9460 * 0x11 - discovered and fall-through edge labelled
9461 * 0x12 - discovered and fall-through and branch edges labelled
9472 static u32 state_htab_size(struct bpf_verifier_env *env)
9474 return env->prog->len;
9477 static struct bpf_verifier_state_list **explored_state(
9478 struct bpf_verifier_env *env,
9481 struct bpf_verifier_state *cur = env->cur_state;
9482 struct bpf_func_state *state = cur->frame[cur->curframe];
9484 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
9487 static void init_explored_state(struct bpf_verifier_env *env, int idx)
9489 env->insn_aux_data[idx].prune_point = true;
9497 /* t, w, e - match pseudo-code above:
9498 * t - index of current instruction
9499 * w - next instruction
9502 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env,
9505 int *insn_stack = env->cfg.insn_stack;
9506 int *insn_state = env->cfg.insn_state;
9508 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
9509 return DONE_EXPLORING;
9511 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
9512 return DONE_EXPLORING;
9514 if (w < 0 || w >= env->prog->len) {
9515 verbose_linfo(env, t, "%d: ", t);
9516 verbose(env, "jump out of range from insn %d to %d\n", t, w);
9521 /* mark branch target for state pruning */
9522 init_explored_state(env, w);
9524 if (insn_state[w] == 0) {
9526 insn_state[t] = DISCOVERED | e;
9527 insn_state[w] = DISCOVERED;
9528 if (env->cfg.cur_stack >= env->prog->len)
9530 insn_stack[env->cfg.cur_stack++] = w;
9531 return KEEP_EXPLORING;
9532 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
9533 if (loop_ok && env->bpf_capable)
9534 return DONE_EXPLORING;
9535 verbose_linfo(env, t, "%d: ", t);
9536 verbose_linfo(env, w, "%d: ", w);
9537 verbose(env, "back-edge from insn %d to %d\n", t, w);
9539 } else if (insn_state[w] == EXPLORED) {
9540 /* forward- or cross-edge */
9541 insn_state[t] = DISCOVERED | e;
9543 verbose(env, "insn state internal bug\n");
9546 return DONE_EXPLORING;
9549 static int visit_func_call_insn(int t, int insn_cnt,
9550 struct bpf_insn *insns,
9551 struct bpf_verifier_env *env,
9556 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
9560 if (t + 1 < insn_cnt)
9561 init_explored_state(env, t + 1);
9563 init_explored_state(env, t);
9564 ret = push_insn(t, t + insns[t].imm + 1, BRANCH, env,
9565 /* It's ok to allow recursion from CFG point of
9566 * view. __check_func_call() will do the actual
9569 bpf_pseudo_func(insns + t));
9574 /* Visits the instruction at index t and returns one of the following:
9575 * < 0 - an error occurred
9576 * DONE_EXPLORING - the instruction was fully explored
9577 * KEEP_EXPLORING - there is still work to be done before it is fully explored
9579 static int visit_insn(int t, int insn_cnt, struct bpf_verifier_env *env)
9581 struct bpf_insn *insns = env->prog->insnsi;
9584 if (bpf_pseudo_func(insns + t))
9585 return visit_func_call_insn(t, insn_cnt, insns, env, true);
9587 /* All non-branch instructions have a single fall-through edge. */
9588 if (BPF_CLASS(insns[t].code) != BPF_JMP &&
9589 BPF_CLASS(insns[t].code) != BPF_JMP32)
9590 return push_insn(t, t + 1, FALLTHROUGH, env, false);
9592 switch (BPF_OP(insns[t].code)) {
9594 return DONE_EXPLORING;
9597 if (insns[t].imm == BPF_FUNC_timer_set_callback)
9598 /* Mark this call insn to trigger is_state_visited() check
9599 * before call itself is processed by __check_func_call().
9600 * Otherwise new async state will be pushed for further
9603 init_explored_state(env, t);
9604 return visit_func_call_insn(t, insn_cnt, insns, env,
9605 insns[t].src_reg == BPF_PSEUDO_CALL);
9608 if (BPF_SRC(insns[t].code) != BPF_K)
9611 /* unconditional jump with single edge */
9612 ret = push_insn(t, t + insns[t].off + 1, FALLTHROUGH, env,
9617 /* unconditional jmp is not a good pruning point,
9618 * but it's marked, since backtracking needs
9619 * to record jmp history in is_state_visited().
9621 init_explored_state(env, t + insns[t].off + 1);
9622 /* tell verifier to check for equivalent states
9623 * after every call and jump
9625 if (t + 1 < insn_cnt)
9626 init_explored_state(env, t + 1);
9631 /* conditional jump with two edges */
9632 init_explored_state(env, t);
9633 ret = push_insn(t, t + 1, FALLTHROUGH, env, true);
9637 return push_insn(t, t + insns[t].off + 1, BRANCH, env, true);
9641 /* non-recursive depth-first-search to detect loops in BPF program
9642 * loop == back-edge in directed graph
9644 static int check_cfg(struct bpf_verifier_env *env)
9646 int insn_cnt = env->prog->len;
9647 int *insn_stack, *insn_state;
9651 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
9655 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
9661 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
9662 insn_stack[0] = 0; /* 0 is the first instruction */
9663 env->cfg.cur_stack = 1;
9665 while (env->cfg.cur_stack > 0) {
9666 int t = insn_stack[env->cfg.cur_stack - 1];
9668 ret = visit_insn(t, insn_cnt, env);
9670 case DONE_EXPLORING:
9671 insn_state[t] = EXPLORED;
9672 env->cfg.cur_stack--;
9674 case KEEP_EXPLORING:
9678 verbose(env, "visit_insn internal bug\n");
9685 if (env->cfg.cur_stack < 0) {
9686 verbose(env, "pop stack internal bug\n");
9691 for (i = 0; i < insn_cnt; i++) {
9692 if (insn_state[i] != EXPLORED) {
9693 verbose(env, "unreachable insn %d\n", i);
9698 ret = 0; /* cfg looks good */
9703 env->cfg.insn_state = env->cfg.insn_stack = NULL;
9707 static int check_abnormal_return(struct bpf_verifier_env *env)
9711 for (i = 1; i < env->subprog_cnt; i++) {
9712 if (env->subprog_info[i].has_ld_abs) {
9713 verbose(env, "LD_ABS is not allowed in subprogs without BTF\n");
9716 if (env->subprog_info[i].has_tail_call) {
9717 verbose(env, "tail_call is not allowed in subprogs without BTF\n");
9724 /* The minimum supported BTF func info size */
9725 #define MIN_BPF_FUNCINFO_SIZE 8
9726 #define MAX_FUNCINFO_REC_SIZE 252
9728 static int check_btf_func(struct bpf_verifier_env *env,
9729 const union bpf_attr *attr,
9732 const struct btf_type *type, *func_proto, *ret_type;
9733 u32 i, nfuncs, urec_size, min_size;
9734 u32 krec_size = sizeof(struct bpf_func_info);
9735 struct bpf_func_info *krecord;
9736 struct bpf_func_info_aux *info_aux = NULL;
9737 struct bpf_prog *prog;
9738 const struct btf *btf;
9740 u32 prev_offset = 0;
9744 nfuncs = attr->func_info_cnt;
9746 if (check_abnormal_return(env))
9751 if (nfuncs != env->subprog_cnt) {
9752 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
9756 urec_size = attr->func_info_rec_size;
9757 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
9758 urec_size > MAX_FUNCINFO_REC_SIZE ||
9759 urec_size % sizeof(u32)) {
9760 verbose(env, "invalid func info rec size %u\n", urec_size);
9765 btf = prog->aux->btf;
9767 urecord = make_bpfptr(attr->func_info, uattr.is_kernel);
9768 min_size = min_t(u32, krec_size, urec_size);
9770 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
9773 info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN);
9777 for (i = 0; i < nfuncs; i++) {
9778 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
9780 if (ret == -E2BIG) {
9781 verbose(env, "nonzero tailing record in func info");
9782 /* set the size kernel expects so loader can zero
9783 * out the rest of the record.
9785 if (copy_to_bpfptr_offset(uattr,
9786 offsetof(union bpf_attr, func_info_rec_size),
9787 &min_size, sizeof(min_size)))
9793 if (copy_from_bpfptr(&krecord[i], urecord, min_size)) {
9798 /* check insn_off */
9801 if (krecord[i].insn_off) {
9803 "nonzero insn_off %u for the first func info record",
9804 krecord[i].insn_off);
9807 } else if (krecord[i].insn_off <= prev_offset) {
9809 "same or smaller insn offset (%u) than previous func info record (%u)",
9810 krecord[i].insn_off, prev_offset);
9814 if (env->subprog_info[i].start != krecord[i].insn_off) {
9815 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
9820 type = btf_type_by_id(btf, krecord[i].type_id);
9821 if (!type || !btf_type_is_func(type)) {
9822 verbose(env, "invalid type id %d in func info",
9823 krecord[i].type_id);
9826 info_aux[i].linkage = BTF_INFO_VLEN(type->info);
9828 func_proto = btf_type_by_id(btf, type->type);
9829 if (unlikely(!func_proto || !btf_type_is_func_proto(func_proto)))
9830 /* btf_func_check() already verified it during BTF load */
9832 ret_type = btf_type_skip_modifiers(btf, func_proto->type, NULL);
9834 btf_type_is_small_int(ret_type) || btf_type_is_enum(ret_type);
9835 if (i && !scalar_return && env->subprog_info[i].has_ld_abs) {
9836 verbose(env, "LD_ABS is only allowed in functions that return 'int'.\n");
9839 if (i && !scalar_return && env->subprog_info[i].has_tail_call) {
9840 verbose(env, "tail_call is only allowed in functions that return 'int'.\n");
9844 prev_offset = krecord[i].insn_off;
9845 bpfptr_add(&urecord, urec_size);
9848 prog->aux->func_info = krecord;
9849 prog->aux->func_info_cnt = nfuncs;
9850 prog->aux->func_info_aux = info_aux;
9859 static void adjust_btf_func(struct bpf_verifier_env *env)
9861 struct bpf_prog_aux *aux = env->prog->aux;
9864 if (!aux->func_info)
9867 for (i = 0; i < env->subprog_cnt; i++)
9868 aux->func_info[i].insn_off = env->subprog_info[i].start;
9871 #define MIN_BPF_LINEINFO_SIZE (offsetof(struct bpf_line_info, line_col) + \
9872 sizeof(((struct bpf_line_info *)(0))->line_col))
9873 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
9875 static int check_btf_line(struct bpf_verifier_env *env,
9876 const union bpf_attr *attr,
9879 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
9880 struct bpf_subprog_info *sub;
9881 struct bpf_line_info *linfo;
9882 struct bpf_prog *prog;
9883 const struct btf *btf;
9887 nr_linfo = attr->line_info_cnt;
9890 if (nr_linfo > INT_MAX / sizeof(struct bpf_line_info))
9893 rec_size = attr->line_info_rec_size;
9894 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
9895 rec_size > MAX_LINEINFO_REC_SIZE ||
9896 rec_size & (sizeof(u32) - 1))
9899 /* Need to zero it in case the userspace may
9900 * pass in a smaller bpf_line_info object.
9902 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
9903 GFP_KERNEL | __GFP_NOWARN);
9908 btf = prog->aux->btf;
9911 sub = env->subprog_info;
9912 ulinfo = make_bpfptr(attr->line_info, uattr.is_kernel);
9913 expected_size = sizeof(struct bpf_line_info);
9914 ncopy = min_t(u32, expected_size, rec_size);
9915 for (i = 0; i < nr_linfo; i++) {
9916 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
9918 if (err == -E2BIG) {
9919 verbose(env, "nonzero tailing record in line_info");
9920 if (copy_to_bpfptr_offset(uattr,
9921 offsetof(union bpf_attr, line_info_rec_size),
9922 &expected_size, sizeof(expected_size)))
9928 if (copy_from_bpfptr(&linfo[i], ulinfo, ncopy)) {
9934 * Check insn_off to ensure
9935 * 1) strictly increasing AND
9936 * 2) bounded by prog->len
9938 * The linfo[0].insn_off == 0 check logically falls into
9939 * the later "missing bpf_line_info for func..." case
9940 * because the first linfo[0].insn_off must be the
9941 * first sub also and the first sub must have
9942 * subprog_info[0].start == 0.
9944 if ((i && linfo[i].insn_off <= prev_offset) ||
9945 linfo[i].insn_off >= prog->len) {
9946 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
9947 i, linfo[i].insn_off, prev_offset,
9953 if (!prog->insnsi[linfo[i].insn_off].code) {
9955 "Invalid insn code at line_info[%u].insn_off\n",
9961 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
9962 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
9963 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
9968 if (s != env->subprog_cnt) {
9969 if (linfo[i].insn_off == sub[s].start) {
9970 sub[s].linfo_idx = i;
9972 } else if (sub[s].start < linfo[i].insn_off) {
9973 verbose(env, "missing bpf_line_info for func#%u\n", s);
9979 prev_offset = linfo[i].insn_off;
9980 bpfptr_add(&ulinfo, rec_size);
9983 if (s != env->subprog_cnt) {
9984 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
9985 env->subprog_cnt - s, s);
9990 prog->aux->linfo = linfo;
9991 prog->aux->nr_linfo = nr_linfo;
10000 static int check_btf_info(struct bpf_verifier_env *env,
10001 const union bpf_attr *attr,
10007 if (!attr->func_info_cnt && !attr->line_info_cnt) {
10008 if (check_abnormal_return(env))
10013 btf = btf_get_by_fd(attr->prog_btf_fd);
10015 return PTR_ERR(btf);
10016 if (btf_is_kernel(btf)) {
10020 env->prog->aux->btf = btf;
10022 err = check_btf_func(env, attr, uattr);
10026 err = check_btf_line(env, attr, uattr);
10033 /* check %cur's range satisfies %old's */
10034 static bool range_within(struct bpf_reg_state *old,
10035 struct bpf_reg_state *cur)
10037 return old->umin_value <= cur->umin_value &&
10038 old->umax_value >= cur->umax_value &&
10039 old->smin_value <= cur->smin_value &&
10040 old->smax_value >= cur->smax_value &&
10041 old->u32_min_value <= cur->u32_min_value &&
10042 old->u32_max_value >= cur->u32_max_value &&
10043 old->s32_min_value <= cur->s32_min_value &&
10044 old->s32_max_value >= cur->s32_max_value;
10047 /* If in the old state two registers had the same id, then they need to have
10048 * the same id in the new state as well. But that id could be different from
10049 * the old state, so we need to track the mapping from old to new ids.
10050 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
10051 * regs with old id 5 must also have new id 9 for the new state to be safe. But
10052 * regs with a different old id could still have new id 9, we don't care about
10054 * So we look through our idmap to see if this old id has been seen before. If
10055 * so, we require the new id to match; otherwise, we add the id pair to the map.
10057 static bool check_ids(u32 old_id, u32 cur_id, struct bpf_id_pair *idmap)
10061 for (i = 0; i < BPF_ID_MAP_SIZE; i++) {
10062 if (!idmap[i].old) {
10063 /* Reached an empty slot; haven't seen this id before */
10064 idmap[i].old = old_id;
10065 idmap[i].cur = cur_id;
10068 if (idmap[i].old == old_id)
10069 return idmap[i].cur == cur_id;
10071 /* We ran out of idmap slots, which should be impossible */
10076 static void clean_func_state(struct bpf_verifier_env *env,
10077 struct bpf_func_state *st)
10079 enum bpf_reg_liveness live;
10082 for (i = 0; i < BPF_REG_FP; i++) {
10083 live = st->regs[i].live;
10084 /* liveness must not touch this register anymore */
10085 st->regs[i].live |= REG_LIVE_DONE;
10086 if (!(live & REG_LIVE_READ))
10087 /* since the register is unused, clear its state
10088 * to make further comparison simpler
10090 __mark_reg_not_init(env, &st->regs[i]);
10093 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
10094 live = st->stack[i].spilled_ptr.live;
10095 /* liveness must not touch this stack slot anymore */
10096 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
10097 if (!(live & REG_LIVE_READ)) {
10098 __mark_reg_not_init(env, &st->stack[i].spilled_ptr);
10099 for (j = 0; j < BPF_REG_SIZE; j++)
10100 st->stack[i].slot_type[j] = STACK_INVALID;
10105 static void clean_verifier_state(struct bpf_verifier_env *env,
10106 struct bpf_verifier_state *st)
10110 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
10111 /* all regs in this state in all frames were already marked */
10114 for (i = 0; i <= st->curframe; i++)
10115 clean_func_state(env, st->frame[i]);
10118 /* the parentage chains form a tree.
10119 * the verifier states are added to state lists at given insn and
10120 * pushed into state stack for future exploration.
10121 * when the verifier reaches bpf_exit insn some of the verifer states
10122 * stored in the state lists have their final liveness state already,
10123 * but a lot of states will get revised from liveness point of view when
10124 * the verifier explores other branches.
10127 * 2: if r1 == 100 goto pc+1
10130 * when the verifier reaches exit insn the register r0 in the state list of
10131 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
10132 * of insn 2 and goes exploring further. At the insn 4 it will walk the
10133 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
10135 * Since the verifier pushes the branch states as it sees them while exploring
10136 * the program the condition of walking the branch instruction for the second
10137 * time means that all states below this branch were already explored and
10138 * their final liveness marks are already propagated.
10139 * Hence when the verifier completes the search of state list in is_state_visited()
10140 * we can call this clean_live_states() function to mark all liveness states
10141 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
10142 * will not be used.
10143 * This function also clears the registers and stack for states that !READ
10144 * to simplify state merging.
10146 * Important note here that walking the same branch instruction in the callee
10147 * doesn't meant that the states are DONE. The verifier has to compare
10150 static void clean_live_states(struct bpf_verifier_env *env, int insn,
10151 struct bpf_verifier_state *cur)
10153 struct bpf_verifier_state_list *sl;
10156 sl = *explored_state(env, insn);
10158 if (sl->state.branches)
10160 if (sl->state.insn_idx != insn ||
10161 sl->state.curframe != cur->curframe)
10163 for (i = 0; i <= cur->curframe; i++)
10164 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
10166 clean_verifier_state(env, &sl->state);
10172 /* Returns true if (rold safe implies rcur safe) */
10173 static bool regsafe(struct bpf_verifier_env *env, struct bpf_reg_state *rold,
10174 struct bpf_reg_state *rcur, struct bpf_id_pair *idmap)
10178 if (!(rold->live & REG_LIVE_READ))
10179 /* explored state didn't use this */
10182 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
10184 if (rold->type == PTR_TO_STACK)
10185 /* two stack pointers are equal only if they're pointing to
10186 * the same stack frame, since fp-8 in foo != fp-8 in bar
10188 return equal && rold->frameno == rcur->frameno;
10193 if (rold->type == NOT_INIT)
10194 /* explored state can't have used this */
10196 if (rcur->type == NOT_INIT)
10198 switch (base_type(rold->type)) {
10200 if (env->explore_alu_limits)
10202 if (rcur->type == SCALAR_VALUE) {
10203 if (!rold->precise && !rcur->precise)
10205 /* new val must satisfy old val knowledge */
10206 return range_within(rold, rcur) &&
10207 tnum_in(rold->var_off, rcur->var_off);
10209 /* We're trying to use a pointer in place of a scalar.
10210 * Even if the scalar was unbounded, this could lead to
10211 * pointer leaks because scalars are allowed to leak
10212 * while pointers are not. We could make this safe in
10213 * special cases if root is calling us, but it's
10214 * probably not worth the hassle.
10218 case PTR_TO_MAP_KEY:
10219 case PTR_TO_MAP_VALUE:
10220 /* a PTR_TO_MAP_VALUE could be safe to use as a
10221 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
10222 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
10223 * checked, doing so could have affected others with the same
10224 * id, and we can't check for that because we lost the id when
10225 * we converted to a PTR_TO_MAP_VALUE.
10227 if (type_may_be_null(rold->type)) {
10228 if (!type_may_be_null(rcur->type))
10230 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
10232 /* Check our ids match any regs they're supposed to */
10233 return check_ids(rold->id, rcur->id, idmap);
10236 /* If the new min/max/var_off satisfy the old ones and
10237 * everything else matches, we are OK.
10238 * 'id' is not compared, since it's only used for maps with
10239 * bpf_spin_lock inside map element and in such cases if
10240 * the rest of the prog is valid for one map element then
10241 * it's valid for all map elements regardless of the key
10242 * used in bpf_map_lookup()
10244 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
10245 range_within(rold, rcur) &&
10246 tnum_in(rold->var_off, rcur->var_off);
10247 case PTR_TO_PACKET_META:
10248 case PTR_TO_PACKET:
10249 if (rcur->type != rold->type)
10251 /* We must have at least as much range as the old ptr
10252 * did, so that any accesses which were safe before are
10253 * still safe. This is true even if old range < old off,
10254 * since someone could have accessed through (ptr - k), or
10255 * even done ptr -= k in a register, to get a safe access.
10257 if (rold->range > rcur->range)
10259 /* If the offsets don't match, we can't trust our alignment;
10260 * nor can we be sure that we won't fall out of range.
10262 if (rold->off != rcur->off)
10264 /* id relations must be preserved */
10265 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
10267 /* new val must satisfy old val knowledge */
10268 return range_within(rold, rcur) &&
10269 tnum_in(rold->var_off, rcur->var_off);
10271 case CONST_PTR_TO_MAP:
10272 case PTR_TO_PACKET_END:
10273 case PTR_TO_FLOW_KEYS:
10274 case PTR_TO_SOCKET:
10275 case PTR_TO_SOCK_COMMON:
10276 case PTR_TO_TCP_SOCK:
10277 case PTR_TO_XDP_SOCK:
10278 /* Only valid matches are exact, which memcmp() above
10279 * would have accepted
10282 /* Don't know what's going on, just say it's not safe */
10286 /* Shouldn't get here; if we do, say it's not safe */
10291 static bool stacksafe(struct bpf_verifier_env *env, struct bpf_func_state *old,
10292 struct bpf_func_state *cur, struct bpf_id_pair *idmap)
10296 /* walk slots of the explored stack and ignore any additional
10297 * slots in the current stack, since explored(safe) state
10300 for (i = 0; i < old->allocated_stack; i++) {
10301 spi = i / BPF_REG_SIZE;
10303 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
10304 i += BPF_REG_SIZE - 1;
10305 /* explored state didn't use this */
10309 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
10312 /* explored stack has more populated slots than current stack
10313 * and these slots were used
10315 if (i >= cur->allocated_stack)
10318 /* if old state was safe with misc data in the stack
10319 * it will be safe with zero-initialized stack.
10320 * The opposite is not true
10322 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
10323 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
10325 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
10326 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
10327 /* Ex: old explored (safe) state has STACK_SPILL in
10328 * this stack slot, but current has STACK_MISC ->
10329 * this verifier states are not equivalent,
10330 * return false to continue verification of this path
10333 if (i % BPF_REG_SIZE)
10335 if (old->stack[spi].slot_type[0] != STACK_SPILL)
10337 if (!regsafe(env, &old->stack[spi].spilled_ptr,
10338 &cur->stack[spi].spilled_ptr, idmap))
10339 /* when explored and current stack slot are both storing
10340 * spilled registers, check that stored pointers types
10341 * are the same as well.
10342 * Ex: explored safe path could have stored
10343 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
10344 * but current path has stored:
10345 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
10346 * such verifier states are not equivalent.
10347 * return false to continue verification of this path
10354 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
10356 if (old->acquired_refs != cur->acquired_refs)
10358 return !memcmp(old->refs, cur->refs,
10359 sizeof(*old->refs) * old->acquired_refs);
10362 /* compare two verifier states
10364 * all states stored in state_list are known to be valid, since
10365 * verifier reached 'bpf_exit' instruction through them
10367 * this function is called when verifier exploring different branches of
10368 * execution popped from the state stack. If it sees an old state that has
10369 * more strict register state and more strict stack state then this execution
10370 * branch doesn't need to be explored further, since verifier already
10371 * concluded that more strict state leads to valid finish.
10373 * Therefore two states are equivalent if register state is more conservative
10374 * and explored stack state is more conservative than the current one.
10377 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
10378 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
10380 * In other words if current stack state (one being explored) has more
10381 * valid slots than old one that already passed validation, it means
10382 * the verifier can stop exploring and conclude that current state is valid too
10384 * Similarly with registers. If explored state has register type as invalid
10385 * whereas register type in current state is meaningful, it means that
10386 * the current state will reach 'bpf_exit' instruction safely
10388 static bool func_states_equal(struct bpf_verifier_env *env, struct bpf_func_state *old,
10389 struct bpf_func_state *cur)
10393 memset(env->idmap_scratch, 0, sizeof(env->idmap_scratch));
10394 for (i = 0; i < MAX_BPF_REG; i++)
10395 if (!regsafe(env, &old->regs[i], &cur->regs[i],
10396 env->idmap_scratch))
10399 if (!stacksafe(env, old, cur, env->idmap_scratch))
10402 if (!refsafe(old, cur))
10408 static bool states_equal(struct bpf_verifier_env *env,
10409 struct bpf_verifier_state *old,
10410 struct bpf_verifier_state *cur)
10414 if (old->curframe != cur->curframe)
10417 /* Verification state from speculative execution simulation
10418 * must never prune a non-speculative execution one.
10420 if (old->speculative && !cur->speculative)
10423 if (old->active_spin_lock != cur->active_spin_lock)
10426 /* for states to be equal callsites have to be the same
10427 * and all frame states need to be equivalent
10429 for (i = 0; i <= old->curframe; i++) {
10430 if (old->frame[i]->callsite != cur->frame[i]->callsite)
10432 if (!func_states_equal(env, old->frame[i], cur->frame[i]))
10438 /* Return 0 if no propagation happened. Return negative error code if error
10439 * happened. Otherwise, return the propagated bit.
10441 static int propagate_liveness_reg(struct bpf_verifier_env *env,
10442 struct bpf_reg_state *reg,
10443 struct bpf_reg_state *parent_reg)
10445 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
10446 u8 flag = reg->live & REG_LIVE_READ;
10449 /* When comes here, read flags of PARENT_REG or REG could be any of
10450 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
10451 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
10453 if (parent_flag == REG_LIVE_READ64 ||
10454 /* Or if there is no read flag from REG. */
10456 /* Or if the read flag from REG is the same as PARENT_REG. */
10457 parent_flag == flag)
10460 err = mark_reg_read(env, reg, parent_reg, flag);
10467 /* A write screens off any subsequent reads; but write marks come from the
10468 * straight-line code between a state and its parent. When we arrive at an
10469 * equivalent state (jump target or such) we didn't arrive by the straight-line
10470 * code, so read marks in the state must propagate to the parent regardless
10471 * of the state's write marks. That's what 'parent == state->parent' comparison
10472 * in mark_reg_read() is for.
10474 static int propagate_liveness(struct bpf_verifier_env *env,
10475 const struct bpf_verifier_state *vstate,
10476 struct bpf_verifier_state *vparent)
10478 struct bpf_reg_state *state_reg, *parent_reg;
10479 struct bpf_func_state *state, *parent;
10480 int i, frame, err = 0;
10482 if (vparent->curframe != vstate->curframe) {
10483 WARN(1, "propagate_live: parent frame %d current frame %d\n",
10484 vparent->curframe, vstate->curframe);
10487 /* Propagate read liveness of registers... */
10488 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
10489 for (frame = 0; frame <= vstate->curframe; frame++) {
10490 parent = vparent->frame[frame];
10491 state = vstate->frame[frame];
10492 parent_reg = parent->regs;
10493 state_reg = state->regs;
10494 /* We don't need to worry about FP liveness, it's read-only */
10495 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
10496 err = propagate_liveness_reg(env, &state_reg[i],
10500 if (err == REG_LIVE_READ64)
10501 mark_insn_zext(env, &parent_reg[i]);
10504 /* Propagate stack slots. */
10505 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
10506 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
10507 parent_reg = &parent->stack[i].spilled_ptr;
10508 state_reg = &state->stack[i].spilled_ptr;
10509 err = propagate_liveness_reg(env, state_reg,
10518 /* find precise scalars in the previous equivalent state and
10519 * propagate them into the current state
10521 static int propagate_precision(struct bpf_verifier_env *env,
10522 const struct bpf_verifier_state *old)
10524 struct bpf_reg_state *state_reg;
10525 struct bpf_func_state *state;
10528 state = old->frame[old->curframe];
10529 state_reg = state->regs;
10530 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
10531 if (state_reg->type != SCALAR_VALUE ||
10532 !state_reg->precise)
10534 if (env->log.level & BPF_LOG_LEVEL2)
10535 verbose(env, "propagating r%d\n", i);
10536 err = mark_chain_precision(env, i);
10541 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
10542 if (state->stack[i].slot_type[0] != STACK_SPILL)
10544 state_reg = &state->stack[i].spilled_ptr;
10545 if (state_reg->type != SCALAR_VALUE ||
10546 !state_reg->precise)
10548 if (env->log.level & BPF_LOG_LEVEL2)
10549 verbose(env, "propagating fp%d\n",
10550 (-i - 1) * BPF_REG_SIZE);
10551 err = mark_chain_precision_stack(env, i);
10558 static bool states_maybe_looping(struct bpf_verifier_state *old,
10559 struct bpf_verifier_state *cur)
10561 struct bpf_func_state *fold, *fcur;
10562 int i, fr = cur->curframe;
10564 if (old->curframe != fr)
10567 fold = old->frame[fr];
10568 fcur = cur->frame[fr];
10569 for (i = 0; i < MAX_BPF_REG; i++)
10570 if (memcmp(&fold->regs[i], &fcur->regs[i],
10571 offsetof(struct bpf_reg_state, parent)))
10577 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
10579 struct bpf_verifier_state_list *new_sl;
10580 struct bpf_verifier_state_list *sl, **pprev;
10581 struct bpf_verifier_state *cur = env->cur_state, *new;
10582 int i, j, err, states_cnt = 0;
10583 bool add_new_state = env->test_state_freq ? true : false;
10585 cur->last_insn_idx = env->prev_insn_idx;
10586 if (!env->insn_aux_data[insn_idx].prune_point)
10587 /* this 'insn_idx' instruction wasn't marked, so we will not
10588 * be doing state search here
10592 /* bpf progs typically have pruning point every 4 instructions
10593 * http://vger.kernel.org/bpfconf2019.html#session-1
10594 * Do not add new state for future pruning if the verifier hasn't seen
10595 * at least 2 jumps and at least 8 instructions.
10596 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
10597 * In tests that amounts to up to 50% reduction into total verifier
10598 * memory consumption and 20% verifier time speedup.
10600 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
10601 env->insn_processed - env->prev_insn_processed >= 8)
10602 add_new_state = true;
10604 pprev = explored_state(env, insn_idx);
10607 clean_live_states(env, insn_idx, cur);
10611 if (sl->state.insn_idx != insn_idx)
10614 if (sl->state.branches) {
10615 struct bpf_func_state *frame = sl->state.frame[sl->state.curframe];
10617 if (frame->in_async_callback_fn &&
10618 frame->async_entry_cnt != cur->frame[cur->curframe]->async_entry_cnt) {
10619 /* Different async_entry_cnt means that the verifier is
10620 * processing another entry into async callback.
10621 * Seeing the same state is not an indication of infinite
10622 * loop or infinite recursion.
10623 * But finding the same state doesn't mean that it's safe
10624 * to stop processing the current state. The previous state
10625 * hasn't yet reached bpf_exit, since state.branches > 0.
10626 * Checking in_async_callback_fn alone is not enough either.
10627 * Since the verifier still needs to catch infinite loops
10628 * inside async callbacks.
10630 } else if (states_maybe_looping(&sl->state, cur) &&
10631 states_equal(env, &sl->state, cur)) {
10632 verbose_linfo(env, insn_idx, "; ");
10633 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
10636 /* if the verifier is processing a loop, avoid adding new state
10637 * too often, since different loop iterations have distinct
10638 * states and may not help future pruning.
10639 * This threshold shouldn't be too low to make sure that
10640 * a loop with large bound will be rejected quickly.
10641 * The most abusive loop will be:
10643 * if r1 < 1000000 goto pc-2
10644 * 1M insn_procssed limit / 100 == 10k peak states.
10645 * This threshold shouldn't be too high either, since states
10646 * at the end of the loop are likely to be useful in pruning.
10648 if (env->jmps_processed - env->prev_jmps_processed < 20 &&
10649 env->insn_processed - env->prev_insn_processed < 100)
10650 add_new_state = false;
10653 if (states_equal(env, &sl->state, cur)) {
10655 /* reached equivalent register/stack state,
10656 * prune the search.
10657 * Registers read by the continuation are read by us.
10658 * If we have any write marks in env->cur_state, they
10659 * will prevent corresponding reads in the continuation
10660 * from reaching our parent (an explored_state). Our
10661 * own state will get the read marks recorded, but
10662 * they'll be immediately forgotten as we're pruning
10663 * this state and will pop a new one.
10665 err = propagate_liveness(env, &sl->state, cur);
10667 /* if previous state reached the exit with precision and
10668 * current state is equivalent to it (except precsion marks)
10669 * the precision needs to be propagated back in
10670 * the current state.
10672 err = err ? : push_jmp_history(env, cur);
10673 err = err ? : propagate_precision(env, &sl->state);
10679 /* when new state is not going to be added do not increase miss count.
10680 * Otherwise several loop iterations will remove the state
10681 * recorded earlier. The goal of these heuristics is to have
10682 * states from some iterations of the loop (some in the beginning
10683 * and some at the end) to help pruning.
10687 /* heuristic to determine whether this state is beneficial
10688 * to keep checking from state equivalence point of view.
10689 * Higher numbers increase max_states_per_insn and verification time,
10690 * but do not meaningfully decrease insn_processed.
10692 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
10693 /* the state is unlikely to be useful. Remove it to
10694 * speed up verification
10697 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
10698 u32 br = sl->state.branches;
10701 "BUG live_done but branches_to_explore %d\n",
10703 free_verifier_state(&sl->state, false);
10705 env->peak_states--;
10707 /* cannot free this state, since parentage chain may
10708 * walk it later. Add it for free_list instead to
10709 * be freed at the end of verification
10711 sl->next = env->free_list;
10712 env->free_list = sl;
10722 if (env->max_states_per_insn < states_cnt)
10723 env->max_states_per_insn = states_cnt;
10725 if (!env->bpf_capable && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
10726 return push_jmp_history(env, cur);
10728 if (!add_new_state)
10729 return push_jmp_history(env, cur);
10731 /* There were no equivalent states, remember the current one.
10732 * Technically the current state is not proven to be safe yet,
10733 * but it will either reach outer most bpf_exit (which means it's safe)
10734 * or it will be rejected. When there are no loops the verifier won't be
10735 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
10736 * again on the way to bpf_exit.
10737 * When looping the sl->state.branches will be > 0 and this state
10738 * will not be considered for equivalence until branches == 0.
10740 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
10743 env->total_states++;
10744 env->peak_states++;
10745 env->prev_jmps_processed = env->jmps_processed;
10746 env->prev_insn_processed = env->insn_processed;
10748 /* add new state to the head of linked list */
10749 new = &new_sl->state;
10750 err = copy_verifier_state(new, cur);
10752 free_verifier_state(new, false);
10756 new->insn_idx = insn_idx;
10757 WARN_ONCE(new->branches != 1,
10758 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
10761 cur->first_insn_idx = insn_idx;
10762 clear_jmp_history(cur);
10763 new_sl->next = *explored_state(env, insn_idx);
10764 *explored_state(env, insn_idx) = new_sl;
10765 /* connect new state to parentage chain. Current frame needs all
10766 * registers connected. Only r6 - r9 of the callers are alive (pushed
10767 * to the stack implicitly by JITs) so in callers' frames connect just
10768 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
10769 * the state of the call instruction (with WRITTEN set), and r0 comes
10770 * from callee with its full parentage chain, anyway.
10772 /* clear write marks in current state: the writes we did are not writes
10773 * our child did, so they don't screen off its reads from us.
10774 * (There are no read marks in current state, because reads always mark
10775 * their parent and current state never has children yet. Only
10776 * explored_states can get read marks.)
10778 for (j = 0; j <= cur->curframe; j++) {
10779 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
10780 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
10781 for (i = 0; i < BPF_REG_FP; i++)
10782 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
10785 /* all stack frames are accessible from callee, clear them all */
10786 for (j = 0; j <= cur->curframe; j++) {
10787 struct bpf_func_state *frame = cur->frame[j];
10788 struct bpf_func_state *newframe = new->frame[j];
10790 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
10791 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
10792 frame->stack[i].spilled_ptr.parent =
10793 &newframe->stack[i].spilled_ptr;
10799 /* Return true if it's OK to have the same insn return a different type. */
10800 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
10802 switch (base_type(type)) {
10804 case PTR_TO_SOCKET:
10805 case PTR_TO_SOCK_COMMON:
10806 case PTR_TO_TCP_SOCK:
10807 case PTR_TO_XDP_SOCK:
10808 case PTR_TO_BTF_ID:
10815 /* If an instruction was previously used with particular pointer types, then we
10816 * need to be careful to avoid cases such as the below, where it may be ok
10817 * for one branch accessing the pointer, but not ok for the other branch:
10822 * R1 = some_other_valid_ptr;
10825 * R2 = *(u32 *)(R1 + 0);
10827 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
10829 return src != prev && (!reg_type_mismatch_ok(src) ||
10830 !reg_type_mismatch_ok(prev));
10833 static int do_check(struct bpf_verifier_env *env)
10835 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
10836 struct bpf_verifier_state *state = env->cur_state;
10837 struct bpf_insn *insns = env->prog->insnsi;
10838 struct bpf_reg_state *regs;
10839 int insn_cnt = env->prog->len;
10840 bool do_print_state = false;
10841 int prev_insn_idx = -1;
10844 struct bpf_insn *insn;
10848 env->prev_insn_idx = prev_insn_idx;
10849 if (env->insn_idx >= insn_cnt) {
10850 verbose(env, "invalid insn idx %d insn_cnt %d\n",
10851 env->insn_idx, insn_cnt);
10855 insn = &insns[env->insn_idx];
10856 class = BPF_CLASS(insn->code);
10858 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
10860 "BPF program is too large. Processed %d insn\n",
10861 env->insn_processed);
10865 err = is_state_visited(env, env->insn_idx);
10869 /* found equivalent state, can prune the search */
10870 if (env->log.level & BPF_LOG_LEVEL) {
10871 if (do_print_state)
10872 verbose(env, "\nfrom %d to %d%s: safe\n",
10873 env->prev_insn_idx, env->insn_idx,
10874 env->cur_state->speculative ?
10875 " (speculative execution)" : "");
10877 verbose(env, "%d: safe\n", env->insn_idx);
10879 goto process_bpf_exit;
10882 if (signal_pending(current))
10885 if (need_resched())
10888 if (env->log.level & BPF_LOG_LEVEL2 ||
10889 (env->log.level & BPF_LOG_LEVEL && do_print_state)) {
10890 if (env->log.level & BPF_LOG_LEVEL2)
10891 verbose(env, "%d:", env->insn_idx);
10893 verbose(env, "\nfrom %d to %d%s:",
10894 env->prev_insn_idx, env->insn_idx,
10895 env->cur_state->speculative ?
10896 " (speculative execution)" : "");
10897 print_verifier_state(env, state->frame[state->curframe]);
10898 do_print_state = false;
10901 if (env->log.level & BPF_LOG_LEVEL) {
10902 const struct bpf_insn_cbs cbs = {
10903 .cb_call = disasm_kfunc_name,
10904 .cb_print = verbose,
10905 .private_data = env,
10908 verbose_linfo(env, env->insn_idx, "; ");
10909 verbose(env, "%d: ", env->insn_idx);
10910 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
10913 if (bpf_prog_is_dev_bound(env->prog->aux)) {
10914 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
10915 env->prev_insn_idx);
10920 regs = cur_regs(env);
10921 sanitize_mark_insn_seen(env);
10922 prev_insn_idx = env->insn_idx;
10924 if (class == BPF_ALU || class == BPF_ALU64) {
10925 err = check_alu_op(env, insn);
10929 } else if (class == BPF_LDX) {
10930 enum bpf_reg_type *prev_src_type, src_reg_type;
10932 /* check for reserved fields is already done */
10934 /* check src operand */
10935 err = check_reg_arg(env, insn->src_reg, SRC_OP);
10939 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
10943 src_reg_type = regs[insn->src_reg].type;
10945 /* check that memory (src_reg + off) is readable,
10946 * the state of dst_reg will be updated by this func
10948 err = check_mem_access(env, env->insn_idx, insn->src_reg,
10949 insn->off, BPF_SIZE(insn->code),
10950 BPF_READ, insn->dst_reg, false);
10954 prev_src_type = &env->insn_aux_data[env->insn_idx].ptr_type;
10956 if (*prev_src_type == NOT_INIT) {
10957 /* saw a valid insn
10958 * dst_reg = *(u32 *)(src_reg + off)
10959 * save type to validate intersecting paths
10961 *prev_src_type = src_reg_type;
10963 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
10964 /* ABuser program is trying to use the same insn
10965 * dst_reg = *(u32*) (src_reg + off)
10966 * with different pointer types:
10967 * src_reg == ctx in one branch and
10968 * src_reg == stack|map in some other branch.
10971 verbose(env, "same insn cannot be used with different pointers\n");
10975 } else if (class == BPF_STX) {
10976 enum bpf_reg_type *prev_dst_type, dst_reg_type;
10978 if (BPF_MODE(insn->code) == BPF_ATOMIC) {
10979 err = check_atomic(env, env->insn_idx, insn);
10986 if (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0) {
10987 verbose(env, "BPF_STX uses reserved fields\n");
10991 /* check src1 operand */
10992 err = check_reg_arg(env, insn->src_reg, SRC_OP);
10995 /* check src2 operand */
10996 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
11000 dst_reg_type = regs[insn->dst_reg].type;
11002 /* check that memory (dst_reg + off) is writeable */
11003 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
11004 insn->off, BPF_SIZE(insn->code),
11005 BPF_WRITE, insn->src_reg, false);
11009 prev_dst_type = &env->insn_aux_data[env->insn_idx].ptr_type;
11011 if (*prev_dst_type == NOT_INIT) {
11012 *prev_dst_type = dst_reg_type;
11013 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
11014 verbose(env, "same insn cannot be used with different pointers\n");
11018 } else if (class == BPF_ST) {
11019 if (BPF_MODE(insn->code) != BPF_MEM ||
11020 insn->src_reg != BPF_REG_0) {
11021 verbose(env, "BPF_ST uses reserved fields\n");
11024 /* check src operand */
11025 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
11029 if (is_ctx_reg(env, insn->dst_reg)) {
11030 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
11032 reg_type_str(env, reg_state(env, insn->dst_reg)->type));
11036 /* check that memory (dst_reg + off) is writeable */
11037 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
11038 insn->off, BPF_SIZE(insn->code),
11039 BPF_WRITE, -1, false);
11043 } else if (class == BPF_JMP || class == BPF_JMP32) {
11044 u8 opcode = BPF_OP(insn->code);
11046 env->jmps_processed++;
11047 if (opcode == BPF_CALL) {
11048 if (BPF_SRC(insn->code) != BPF_K ||
11050 (insn->src_reg != BPF_REG_0 &&
11051 insn->src_reg != BPF_PSEUDO_CALL &&
11052 insn->src_reg != BPF_PSEUDO_KFUNC_CALL) ||
11053 insn->dst_reg != BPF_REG_0 ||
11054 class == BPF_JMP32) {
11055 verbose(env, "BPF_CALL uses reserved fields\n");
11059 if (env->cur_state->active_spin_lock &&
11060 (insn->src_reg == BPF_PSEUDO_CALL ||
11061 insn->imm != BPF_FUNC_spin_unlock)) {
11062 verbose(env, "function calls are not allowed while holding a lock\n");
11065 if (insn->src_reg == BPF_PSEUDO_CALL)
11066 err = check_func_call(env, insn, &env->insn_idx);
11067 else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL)
11068 err = check_kfunc_call(env, insn);
11070 err = check_helper_call(env, insn, &env->insn_idx);
11073 } else if (opcode == BPF_JA) {
11074 if (BPF_SRC(insn->code) != BPF_K ||
11076 insn->src_reg != BPF_REG_0 ||
11077 insn->dst_reg != BPF_REG_0 ||
11078 class == BPF_JMP32) {
11079 verbose(env, "BPF_JA uses reserved fields\n");
11083 env->insn_idx += insn->off + 1;
11086 } else if (opcode == BPF_EXIT) {
11087 if (BPF_SRC(insn->code) != BPF_K ||
11089 insn->src_reg != BPF_REG_0 ||
11090 insn->dst_reg != BPF_REG_0 ||
11091 class == BPF_JMP32) {
11092 verbose(env, "BPF_EXIT uses reserved fields\n");
11096 if (env->cur_state->active_spin_lock) {
11097 verbose(env, "bpf_spin_unlock is missing\n");
11101 /* We must do check_reference_leak here before
11102 * prepare_func_exit to handle the case when
11103 * state->curframe > 0, it may be a callback
11104 * function, for which reference_state must
11105 * match caller reference state when it exits.
11107 err = check_reference_leak(env);
11111 if (state->curframe) {
11112 /* exit from nested function */
11113 err = prepare_func_exit(env, &env->insn_idx);
11116 do_print_state = true;
11120 err = check_return_code(env);
11124 update_branch_counts(env, env->cur_state);
11125 err = pop_stack(env, &prev_insn_idx,
11126 &env->insn_idx, pop_log);
11128 if (err != -ENOENT)
11132 do_print_state = true;
11136 err = check_cond_jmp_op(env, insn, &env->insn_idx);
11140 } else if (class == BPF_LD) {
11141 u8 mode = BPF_MODE(insn->code);
11143 if (mode == BPF_ABS || mode == BPF_IND) {
11144 err = check_ld_abs(env, insn);
11148 } else if (mode == BPF_IMM) {
11149 err = check_ld_imm(env, insn);
11154 sanitize_mark_insn_seen(env);
11156 verbose(env, "invalid BPF_LD mode\n");
11160 verbose(env, "unknown insn class %d\n", class);
11170 static int find_btf_percpu_datasec(struct btf *btf)
11172 const struct btf_type *t;
11177 * Both vmlinux and module each have their own ".data..percpu"
11178 * DATASECs in BTF. So for module's case, we need to skip vmlinux BTF
11179 * types to look at only module's own BTF types.
11181 n = btf_nr_types(btf);
11182 if (btf_is_module(btf))
11183 i = btf_nr_types(btf_vmlinux);
11187 for(; i < n; i++) {
11188 t = btf_type_by_id(btf, i);
11189 if (BTF_INFO_KIND(t->info) != BTF_KIND_DATASEC)
11192 tname = btf_name_by_offset(btf, t->name_off);
11193 if (!strcmp(tname, ".data..percpu"))
11200 /* replace pseudo btf_id with kernel symbol address */
11201 static int check_pseudo_btf_id(struct bpf_verifier_env *env,
11202 struct bpf_insn *insn,
11203 struct bpf_insn_aux_data *aux)
11205 const struct btf_var_secinfo *vsi;
11206 const struct btf_type *datasec;
11207 struct btf_mod_pair *btf_mod;
11208 const struct btf_type *t;
11209 const char *sym_name;
11210 bool percpu = false;
11211 u32 type, id = insn->imm;
11215 int i, btf_fd, err;
11217 btf_fd = insn[1].imm;
11219 btf = btf_get_by_fd(btf_fd);
11221 verbose(env, "invalid module BTF object FD specified.\n");
11225 if (!btf_vmlinux) {
11226 verbose(env, "kernel is missing BTF, make sure CONFIG_DEBUG_INFO_BTF=y is specified in Kconfig.\n");
11233 t = btf_type_by_id(btf, id);
11235 verbose(env, "ldimm64 insn specifies invalid btf_id %d.\n", id);
11240 if (!btf_type_is_var(t)) {
11241 verbose(env, "pseudo btf_id %d in ldimm64 isn't KIND_VAR.\n", id);
11246 sym_name = btf_name_by_offset(btf, t->name_off);
11247 addr = kallsyms_lookup_name(sym_name);
11249 verbose(env, "ldimm64 failed to find the address for kernel symbol '%s'.\n",
11255 datasec_id = find_btf_percpu_datasec(btf);
11256 if (datasec_id > 0) {
11257 datasec = btf_type_by_id(btf, datasec_id);
11258 for_each_vsi(i, datasec, vsi) {
11259 if (vsi->type == id) {
11266 insn[0].imm = (u32)addr;
11267 insn[1].imm = addr >> 32;
11270 t = btf_type_skip_modifiers(btf, type, NULL);
11272 aux->btf_var.reg_type = PTR_TO_PERCPU_BTF_ID;
11273 aux->btf_var.btf = btf;
11274 aux->btf_var.btf_id = type;
11275 } else if (!btf_type_is_struct(t)) {
11276 const struct btf_type *ret;
11280 /* resolve the type size of ksym. */
11281 ret = btf_resolve_size(btf, t, &tsize);
11283 tname = btf_name_by_offset(btf, t->name_off);
11284 verbose(env, "ldimm64 unable to resolve the size of type '%s': %ld\n",
11285 tname, PTR_ERR(ret));
11289 aux->btf_var.reg_type = PTR_TO_MEM | MEM_RDONLY;
11290 aux->btf_var.mem_size = tsize;
11292 aux->btf_var.reg_type = PTR_TO_BTF_ID;
11293 aux->btf_var.btf = btf;
11294 aux->btf_var.btf_id = type;
11297 /* check whether we recorded this BTF (and maybe module) already */
11298 for (i = 0; i < env->used_btf_cnt; i++) {
11299 if (env->used_btfs[i].btf == btf) {
11305 if (env->used_btf_cnt >= MAX_USED_BTFS) {
11310 btf_mod = &env->used_btfs[env->used_btf_cnt];
11311 btf_mod->btf = btf;
11312 btf_mod->module = NULL;
11314 /* if we reference variables from kernel module, bump its refcount */
11315 if (btf_is_module(btf)) {
11316 btf_mod->module = btf_try_get_module(btf);
11317 if (!btf_mod->module) {
11323 env->used_btf_cnt++;
11331 static int check_map_prealloc(struct bpf_map *map)
11333 return (map->map_type != BPF_MAP_TYPE_HASH &&
11334 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
11335 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
11336 !(map->map_flags & BPF_F_NO_PREALLOC);
11339 static bool is_tracing_prog_type(enum bpf_prog_type type)
11342 case BPF_PROG_TYPE_KPROBE:
11343 case BPF_PROG_TYPE_TRACEPOINT:
11344 case BPF_PROG_TYPE_PERF_EVENT:
11345 case BPF_PROG_TYPE_RAW_TRACEPOINT:
11352 static bool is_preallocated_map(struct bpf_map *map)
11354 if (!check_map_prealloc(map))
11356 if (map->inner_map_meta && !check_map_prealloc(map->inner_map_meta))
11361 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
11362 struct bpf_map *map,
11363 struct bpf_prog *prog)
11366 enum bpf_prog_type prog_type = resolve_prog_type(prog);
11368 * Validate that trace type programs use preallocated hash maps.
11370 * For programs attached to PERF events this is mandatory as the
11371 * perf NMI can hit any arbitrary code sequence.
11373 * All other trace types using preallocated hash maps are unsafe as
11374 * well because tracepoint or kprobes can be inside locked regions
11375 * of the memory allocator or at a place where a recursion into the
11376 * memory allocator would see inconsistent state.
11378 * On RT enabled kernels run-time allocation of all trace type
11379 * programs is strictly prohibited due to lock type constraints. On
11380 * !RT kernels it is allowed for backwards compatibility reasons for
11381 * now, but warnings are emitted so developers are made aware of
11382 * the unsafety and can fix their programs before this is enforced.
11384 if (is_tracing_prog_type(prog_type) && !is_preallocated_map(map)) {
11385 if (prog_type == BPF_PROG_TYPE_PERF_EVENT) {
11386 verbose(env, "perf_event programs can only use preallocated hash map\n");
11389 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
11390 verbose(env, "trace type programs can only use preallocated hash map\n");
11393 WARN_ONCE(1, "trace type BPF program uses run-time allocation\n");
11394 verbose(env, "trace type programs with run-time allocated hash maps are unsafe. Switch to preallocated hash maps.\n");
11397 if (map_value_has_spin_lock(map)) {
11398 if (prog_type == BPF_PROG_TYPE_SOCKET_FILTER) {
11399 verbose(env, "socket filter progs cannot use bpf_spin_lock yet\n");
11403 if (is_tracing_prog_type(prog_type)) {
11404 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
11408 if (prog->aux->sleepable) {
11409 verbose(env, "sleepable progs cannot use bpf_spin_lock yet\n");
11414 if (map_value_has_timer(map)) {
11415 if (is_tracing_prog_type(prog_type)) {
11416 verbose(env, "tracing progs cannot use bpf_timer yet\n");
11421 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
11422 !bpf_offload_prog_map_match(prog, map)) {
11423 verbose(env, "offload device mismatch between prog and map\n");
11427 if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
11428 verbose(env, "bpf_struct_ops map cannot be used in prog\n");
11432 if (prog->aux->sleepable)
11433 switch (map->map_type) {
11434 case BPF_MAP_TYPE_HASH:
11435 case BPF_MAP_TYPE_LRU_HASH:
11436 case BPF_MAP_TYPE_ARRAY:
11437 case BPF_MAP_TYPE_PERCPU_HASH:
11438 case BPF_MAP_TYPE_PERCPU_ARRAY:
11439 case BPF_MAP_TYPE_LRU_PERCPU_HASH:
11440 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
11441 case BPF_MAP_TYPE_HASH_OF_MAPS:
11442 if (!is_preallocated_map(map)) {
11444 "Sleepable programs can only use preallocated maps\n");
11448 case BPF_MAP_TYPE_RINGBUF:
11452 "Sleepable programs can only use array, hash, and ringbuf maps\n");
11459 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
11461 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
11462 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
11465 /* find and rewrite pseudo imm in ld_imm64 instructions:
11467 * 1. if it accesses map FD, replace it with actual map pointer.
11468 * 2. if it accesses btf_id of a VAR, replace it with pointer to the var.
11470 * NOTE: btf_vmlinux is required for converting pseudo btf_id.
11472 static int resolve_pseudo_ldimm64(struct bpf_verifier_env *env)
11474 struct bpf_insn *insn = env->prog->insnsi;
11475 int insn_cnt = env->prog->len;
11478 err = bpf_prog_calc_tag(env->prog);
11482 for (i = 0; i < insn_cnt; i++, insn++) {
11483 if (BPF_CLASS(insn->code) == BPF_LDX &&
11484 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
11485 verbose(env, "BPF_LDX uses reserved fields\n");
11489 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
11490 struct bpf_insn_aux_data *aux;
11491 struct bpf_map *map;
11496 if (i == insn_cnt - 1 || insn[1].code != 0 ||
11497 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
11498 insn[1].off != 0) {
11499 verbose(env, "invalid bpf_ld_imm64 insn\n");
11503 if (insn[0].src_reg == 0)
11504 /* valid generic load 64-bit imm */
11507 if (insn[0].src_reg == BPF_PSEUDO_BTF_ID) {
11508 aux = &env->insn_aux_data[i];
11509 err = check_pseudo_btf_id(env, insn, aux);
11515 if (insn[0].src_reg == BPF_PSEUDO_FUNC) {
11516 aux = &env->insn_aux_data[i];
11517 aux->ptr_type = PTR_TO_FUNC;
11521 /* In final convert_pseudo_ld_imm64() step, this is
11522 * converted into regular 64-bit imm load insn.
11524 switch (insn[0].src_reg) {
11525 case BPF_PSEUDO_MAP_VALUE:
11526 case BPF_PSEUDO_MAP_IDX_VALUE:
11528 case BPF_PSEUDO_MAP_FD:
11529 case BPF_PSEUDO_MAP_IDX:
11530 if (insn[1].imm == 0)
11534 verbose(env, "unrecognized bpf_ld_imm64 insn\n");
11538 switch (insn[0].src_reg) {
11539 case BPF_PSEUDO_MAP_IDX_VALUE:
11540 case BPF_PSEUDO_MAP_IDX:
11541 if (bpfptr_is_null(env->fd_array)) {
11542 verbose(env, "fd_idx without fd_array is invalid\n");
11545 if (copy_from_bpfptr_offset(&fd, env->fd_array,
11546 insn[0].imm * sizeof(fd),
11556 map = __bpf_map_get(f);
11558 verbose(env, "fd %d is not pointing to valid bpf_map\n",
11560 return PTR_ERR(map);
11563 err = check_map_prog_compatibility(env, map, env->prog);
11569 aux = &env->insn_aux_data[i];
11570 if (insn[0].src_reg == BPF_PSEUDO_MAP_FD ||
11571 insn[0].src_reg == BPF_PSEUDO_MAP_IDX) {
11572 addr = (unsigned long)map;
11574 u32 off = insn[1].imm;
11576 if (off >= BPF_MAX_VAR_OFF) {
11577 verbose(env, "direct value offset of %u is not allowed\n", off);
11582 if (!map->ops->map_direct_value_addr) {
11583 verbose(env, "no direct value access support for this map type\n");
11588 err = map->ops->map_direct_value_addr(map, &addr, off);
11590 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
11591 map->value_size, off);
11596 aux->map_off = off;
11600 insn[0].imm = (u32)addr;
11601 insn[1].imm = addr >> 32;
11603 /* check whether we recorded this map already */
11604 for (j = 0; j < env->used_map_cnt; j++) {
11605 if (env->used_maps[j] == map) {
11606 aux->map_index = j;
11612 if (env->used_map_cnt >= MAX_USED_MAPS) {
11617 /* hold the map. If the program is rejected by verifier,
11618 * the map will be released by release_maps() or it
11619 * will be used by the valid program until it's unloaded
11620 * and all maps are released in free_used_maps()
11624 aux->map_index = env->used_map_cnt;
11625 env->used_maps[env->used_map_cnt++] = map;
11627 if (bpf_map_is_cgroup_storage(map) &&
11628 bpf_cgroup_storage_assign(env->prog->aux, map)) {
11629 verbose(env, "only one cgroup storage of each type is allowed\n");
11641 /* Basic sanity check before we invest more work here. */
11642 if (!bpf_opcode_in_insntable(insn->code)) {
11643 verbose(env, "unknown opcode %02x\n", insn->code);
11648 /* now all pseudo BPF_LD_IMM64 instructions load valid
11649 * 'struct bpf_map *' into a register instead of user map_fd.
11650 * These pointers will be used later by verifier to validate map access.
11655 /* drop refcnt of maps used by the rejected program */
11656 static void release_maps(struct bpf_verifier_env *env)
11658 __bpf_free_used_maps(env->prog->aux, env->used_maps,
11659 env->used_map_cnt);
11662 /* drop refcnt of maps used by the rejected program */
11663 static void release_btfs(struct bpf_verifier_env *env)
11665 __bpf_free_used_btfs(env->prog->aux, env->used_btfs,
11666 env->used_btf_cnt);
11669 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
11670 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
11672 struct bpf_insn *insn = env->prog->insnsi;
11673 int insn_cnt = env->prog->len;
11676 for (i = 0; i < insn_cnt; i++, insn++) {
11677 if (insn->code != (BPF_LD | BPF_IMM | BPF_DW))
11679 if (insn->src_reg == BPF_PSEUDO_FUNC)
11685 /* single env->prog->insni[off] instruction was replaced with the range
11686 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
11687 * [0, off) and [off, end) to new locations, so the patched range stays zero
11689 static void adjust_insn_aux_data(struct bpf_verifier_env *env,
11690 struct bpf_insn_aux_data *new_data,
11691 struct bpf_prog *new_prog, u32 off, u32 cnt)
11693 struct bpf_insn_aux_data *old_data = env->insn_aux_data;
11694 struct bpf_insn *insn = new_prog->insnsi;
11695 u32 old_seen = old_data[off].seen;
11699 /* aux info at OFF always needs adjustment, no matter fast path
11700 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
11701 * original insn at old prog.
11703 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
11707 prog_len = new_prog->len;
11709 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
11710 memcpy(new_data + off + cnt - 1, old_data + off,
11711 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
11712 for (i = off; i < off + cnt - 1; i++) {
11713 /* Expand insni[off]'s seen count to the patched range. */
11714 new_data[i].seen = old_seen;
11715 new_data[i].zext_dst = insn_has_def32(env, insn + i);
11717 env->insn_aux_data = new_data;
11721 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
11727 /* NOTE: fake 'exit' subprog should be updated as well. */
11728 for (i = 0; i <= env->subprog_cnt; i++) {
11729 if (env->subprog_info[i].start <= off)
11731 env->subprog_info[i].start += len - 1;
11735 static void adjust_poke_descs(struct bpf_prog *prog, u32 off, u32 len)
11737 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
11738 int i, sz = prog->aux->size_poke_tab;
11739 struct bpf_jit_poke_descriptor *desc;
11741 for (i = 0; i < sz; i++) {
11743 if (desc->insn_idx <= off)
11745 desc->insn_idx += len - 1;
11749 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
11750 const struct bpf_insn *patch, u32 len)
11752 struct bpf_prog *new_prog;
11753 struct bpf_insn_aux_data *new_data = NULL;
11756 new_data = vzalloc(array_size(env->prog->len + len - 1,
11757 sizeof(struct bpf_insn_aux_data)));
11762 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
11763 if (IS_ERR(new_prog)) {
11764 if (PTR_ERR(new_prog) == -ERANGE)
11766 "insn %d cannot be patched due to 16-bit range\n",
11767 env->insn_aux_data[off].orig_idx);
11771 adjust_insn_aux_data(env, new_data, new_prog, off, len);
11772 adjust_subprog_starts(env, off, len);
11773 adjust_poke_descs(new_prog, off, len);
11777 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
11782 /* find first prog starting at or after off (first to remove) */
11783 for (i = 0; i < env->subprog_cnt; i++)
11784 if (env->subprog_info[i].start >= off)
11786 /* find first prog starting at or after off + cnt (first to stay) */
11787 for (j = i; j < env->subprog_cnt; j++)
11788 if (env->subprog_info[j].start >= off + cnt)
11790 /* if j doesn't start exactly at off + cnt, we are just removing
11791 * the front of previous prog
11793 if (env->subprog_info[j].start != off + cnt)
11797 struct bpf_prog_aux *aux = env->prog->aux;
11800 /* move fake 'exit' subprog as well */
11801 move = env->subprog_cnt + 1 - j;
11803 memmove(env->subprog_info + i,
11804 env->subprog_info + j,
11805 sizeof(*env->subprog_info) * move);
11806 env->subprog_cnt -= j - i;
11808 /* remove func_info */
11809 if (aux->func_info) {
11810 move = aux->func_info_cnt - j;
11812 memmove(aux->func_info + i,
11813 aux->func_info + j,
11814 sizeof(*aux->func_info) * move);
11815 aux->func_info_cnt -= j - i;
11816 /* func_info->insn_off is set after all code rewrites,
11817 * in adjust_btf_func() - no need to adjust
11821 /* convert i from "first prog to remove" to "first to adjust" */
11822 if (env->subprog_info[i].start == off)
11826 /* update fake 'exit' subprog as well */
11827 for (; i <= env->subprog_cnt; i++)
11828 env->subprog_info[i].start -= cnt;
11833 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
11836 struct bpf_prog *prog = env->prog;
11837 u32 i, l_off, l_cnt, nr_linfo;
11838 struct bpf_line_info *linfo;
11840 nr_linfo = prog->aux->nr_linfo;
11844 linfo = prog->aux->linfo;
11846 /* find first line info to remove, count lines to be removed */
11847 for (i = 0; i < nr_linfo; i++)
11848 if (linfo[i].insn_off >= off)
11853 for (; i < nr_linfo; i++)
11854 if (linfo[i].insn_off < off + cnt)
11859 /* First live insn doesn't match first live linfo, it needs to "inherit"
11860 * last removed linfo. prog is already modified, so prog->len == off
11861 * means no live instructions after (tail of the program was removed).
11863 if (prog->len != off && l_cnt &&
11864 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
11866 linfo[--i].insn_off = off + cnt;
11869 /* remove the line info which refer to the removed instructions */
11871 memmove(linfo + l_off, linfo + i,
11872 sizeof(*linfo) * (nr_linfo - i));
11874 prog->aux->nr_linfo -= l_cnt;
11875 nr_linfo = prog->aux->nr_linfo;
11878 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
11879 for (i = l_off; i < nr_linfo; i++)
11880 linfo[i].insn_off -= cnt;
11882 /* fix up all subprogs (incl. 'exit') which start >= off */
11883 for (i = 0; i <= env->subprog_cnt; i++)
11884 if (env->subprog_info[i].linfo_idx > l_off) {
11885 /* program may have started in the removed region but
11886 * may not be fully removed
11888 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
11889 env->subprog_info[i].linfo_idx -= l_cnt;
11891 env->subprog_info[i].linfo_idx = l_off;
11897 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
11899 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
11900 unsigned int orig_prog_len = env->prog->len;
11903 if (bpf_prog_is_dev_bound(env->prog->aux))
11904 bpf_prog_offload_remove_insns(env, off, cnt);
11906 err = bpf_remove_insns(env->prog, off, cnt);
11910 err = adjust_subprog_starts_after_remove(env, off, cnt);
11914 err = bpf_adj_linfo_after_remove(env, off, cnt);
11918 memmove(aux_data + off, aux_data + off + cnt,
11919 sizeof(*aux_data) * (orig_prog_len - off - cnt));
11924 /* The verifier does more data flow analysis than llvm and will not
11925 * explore branches that are dead at run time. Malicious programs can
11926 * have dead code too. Therefore replace all dead at-run-time code
11929 * Just nops are not optimal, e.g. if they would sit at the end of the
11930 * program and through another bug we would manage to jump there, then
11931 * we'd execute beyond program memory otherwise. Returning exception
11932 * code also wouldn't work since we can have subprogs where the dead
11933 * code could be located.
11935 static void sanitize_dead_code(struct bpf_verifier_env *env)
11937 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
11938 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
11939 struct bpf_insn *insn = env->prog->insnsi;
11940 const int insn_cnt = env->prog->len;
11943 for (i = 0; i < insn_cnt; i++) {
11944 if (aux_data[i].seen)
11946 memcpy(insn + i, &trap, sizeof(trap));
11947 aux_data[i].zext_dst = false;
11951 static bool insn_is_cond_jump(u8 code)
11955 if (BPF_CLASS(code) == BPF_JMP32)
11958 if (BPF_CLASS(code) != BPF_JMP)
11962 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
11965 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
11967 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
11968 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
11969 struct bpf_insn *insn = env->prog->insnsi;
11970 const int insn_cnt = env->prog->len;
11973 for (i = 0; i < insn_cnt; i++, insn++) {
11974 if (!insn_is_cond_jump(insn->code))
11977 if (!aux_data[i + 1].seen)
11978 ja.off = insn->off;
11979 else if (!aux_data[i + 1 + insn->off].seen)
11984 if (bpf_prog_is_dev_bound(env->prog->aux))
11985 bpf_prog_offload_replace_insn(env, i, &ja);
11987 memcpy(insn, &ja, sizeof(ja));
11991 static int opt_remove_dead_code(struct bpf_verifier_env *env)
11993 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
11994 int insn_cnt = env->prog->len;
11997 for (i = 0; i < insn_cnt; i++) {
12001 while (i + j < insn_cnt && !aux_data[i + j].seen)
12006 err = verifier_remove_insns(env, i, j);
12009 insn_cnt = env->prog->len;
12015 static int opt_remove_nops(struct bpf_verifier_env *env)
12017 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
12018 struct bpf_insn *insn = env->prog->insnsi;
12019 int insn_cnt = env->prog->len;
12022 for (i = 0; i < insn_cnt; i++) {
12023 if (memcmp(&insn[i], &ja, sizeof(ja)))
12026 err = verifier_remove_insns(env, i, 1);
12036 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
12037 const union bpf_attr *attr)
12039 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
12040 struct bpf_insn_aux_data *aux = env->insn_aux_data;
12041 int i, patch_len, delta = 0, len = env->prog->len;
12042 struct bpf_insn *insns = env->prog->insnsi;
12043 struct bpf_prog *new_prog;
12046 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
12047 zext_patch[1] = BPF_ZEXT_REG(0);
12048 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
12049 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
12050 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
12051 for (i = 0; i < len; i++) {
12052 int adj_idx = i + delta;
12053 struct bpf_insn insn;
12056 insn = insns[adj_idx];
12057 load_reg = insn_def_regno(&insn);
12058 if (!aux[adj_idx].zext_dst) {
12066 class = BPF_CLASS(code);
12067 if (load_reg == -1)
12070 /* NOTE: arg "reg" (the fourth one) is only used for
12071 * BPF_STX + SRC_OP, so it is safe to pass NULL
12074 if (is_reg64(env, &insn, load_reg, NULL, DST_OP)) {
12075 if (class == BPF_LD &&
12076 BPF_MODE(code) == BPF_IMM)
12081 /* ctx load could be transformed into wider load. */
12082 if (class == BPF_LDX &&
12083 aux[adj_idx].ptr_type == PTR_TO_CTX)
12086 imm_rnd = get_random_int();
12087 rnd_hi32_patch[0] = insn;
12088 rnd_hi32_patch[1].imm = imm_rnd;
12089 rnd_hi32_patch[3].dst_reg = load_reg;
12090 patch = rnd_hi32_patch;
12092 goto apply_patch_buffer;
12095 /* Add in an zero-extend instruction if a) the JIT has requested
12096 * it or b) it's a CMPXCHG.
12098 * The latter is because: BPF_CMPXCHG always loads a value into
12099 * R0, therefore always zero-extends. However some archs'
12100 * equivalent instruction only does this load when the
12101 * comparison is successful. This detail of CMPXCHG is
12102 * orthogonal to the general zero-extension behaviour of the
12103 * CPU, so it's treated independently of bpf_jit_needs_zext.
12105 if (!bpf_jit_needs_zext() && !is_cmpxchg_insn(&insn))
12108 if (WARN_ON(load_reg == -1)) {
12109 verbose(env, "verifier bug. zext_dst is set, but no reg is defined\n");
12113 zext_patch[0] = insn;
12114 zext_patch[1].dst_reg = load_reg;
12115 zext_patch[1].src_reg = load_reg;
12116 patch = zext_patch;
12118 apply_patch_buffer:
12119 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
12122 env->prog = new_prog;
12123 insns = new_prog->insnsi;
12124 aux = env->insn_aux_data;
12125 delta += patch_len - 1;
12131 /* convert load instructions that access fields of a context type into a
12132 * sequence of instructions that access fields of the underlying structure:
12133 * struct __sk_buff -> struct sk_buff
12134 * struct bpf_sock_ops -> struct sock
12136 static int convert_ctx_accesses(struct bpf_verifier_env *env)
12138 const struct bpf_verifier_ops *ops = env->ops;
12139 int i, cnt, size, ctx_field_size, delta = 0;
12140 const int insn_cnt = env->prog->len;
12141 struct bpf_insn insn_buf[16], *insn;
12142 u32 target_size, size_default, off;
12143 struct bpf_prog *new_prog;
12144 enum bpf_access_type type;
12145 bool is_narrower_load;
12147 if (ops->gen_prologue || env->seen_direct_write) {
12148 if (!ops->gen_prologue) {
12149 verbose(env, "bpf verifier is misconfigured\n");
12152 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
12154 if (cnt >= ARRAY_SIZE(insn_buf)) {
12155 verbose(env, "bpf verifier is misconfigured\n");
12158 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
12162 env->prog = new_prog;
12167 if (bpf_prog_is_dev_bound(env->prog->aux))
12170 insn = env->prog->insnsi + delta;
12172 for (i = 0; i < insn_cnt; i++, insn++) {
12173 bpf_convert_ctx_access_t convert_ctx_access;
12176 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
12177 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
12178 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
12179 insn->code == (BPF_LDX | BPF_MEM | BPF_DW)) {
12182 } else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
12183 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
12184 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
12185 insn->code == (BPF_STX | BPF_MEM | BPF_DW) ||
12186 insn->code == (BPF_ST | BPF_MEM | BPF_B) ||
12187 insn->code == (BPF_ST | BPF_MEM | BPF_H) ||
12188 insn->code == (BPF_ST | BPF_MEM | BPF_W) ||
12189 insn->code == (BPF_ST | BPF_MEM | BPF_DW)) {
12191 ctx_access = BPF_CLASS(insn->code) == BPF_STX;
12196 if (type == BPF_WRITE &&
12197 env->insn_aux_data[i + delta].sanitize_stack_spill) {
12198 struct bpf_insn patch[] = {
12203 cnt = ARRAY_SIZE(patch);
12204 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
12209 env->prog = new_prog;
12210 insn = new_prog->insnsi + i + delta;
12217 switch (env->insn_aux_data[i + delta].ptr_type) {
12219 if (!ops->convert_ctx_access)
12221 convert_ctx_access = ops->convert_ctx_access;
12223 case PTR_TO_SOCKET:
12224 case PTR_TO_SOCK_COMMON:
12225 convert_ctx_access = bpf_sock_convert_ctx_access;
12227 case PTR_TO_TCP_SOCK:
12228 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
12230 case PTR_TO_XDP_SOCK:
12231 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
12233 case PTR_TO_BTF_ID:
12234 if (type == BPF_READ) {
12235 insn->code = BPF_LDX | BPF_PROBE_MEM |
12236 BPF_SIZE((insn)->code);
12237 env->prog->aux->num_exentries++;
12238 } else if (resolve_prog_type(env->prog) != BPF_PROG_TYPE_STRUCT_OPS) {
12239 verbose(env, "Writes through BTF pointers are not allowed\n");
12247 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
12248 size = BPF_LDST_BYTES(insn);
12250 /* If the read access is a narrower load of the field,
12251 * convert to a 4/8-byte load, to minimum program type specific
12252 * convert_ctx_access changes. If conversion is successful,
12253 * we will apply proper mask to the result.
12255 is_narrower_load = size < ctx_field_size;
12256 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
12258 if (is_narrower_load) {
12261 if (type == BPF_WRITE) {
12262 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
12267 if (ctx_field_size == 4)
12269 else if (ctx_field_size == 8)
12270 size_code = BPF_DW;
12272 insn->off = off & ~(size_default - 1);
12273 insn->code = BPF_LDX | BPF_MEM | size_code;
12277 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
12279 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
12280 (ctx_field_size && !target_size)) {
12281 verbose(env, "bpf verifier is misconfigured\n");
12285 if (is_narrower_load && size < target_size) {
12286 u8 shift = bpf_ctx_narrow_access_offset(
12287 off, size, size_default) * 8;
12288 if (shift && cnt + 1 >= ARRAY_SIZE(insn_buf)) {
12289 verbose(env, "bpf verifier narrow ctx load misconfigured\n");
12292 if (ctx_field_size <= 4) {
12294 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
12297 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
12298 (1 << size * 8) - 1);
12301 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
12304 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
12305 (1ULL << size * 8) - 1);
12309 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12315 /* keep walking new program and skip insns we just inserted */
12316 env->prog = new_prog;
12317 insn = new_prog->insnsi + i + delta;
12323 static int jit_subprogs(struct bpf_verifier_env *env)
12325 struct bpf_prog *prog = env->prog, **func, *tmp;
12326 int i, j, subprog_start, subprog_end = 0, len, subprog;
12327 struct bpf_map *map_ptr;
12328 struct bpf_insn *insn;
12329 void *old_bpf_func;
12330 int err, num_exentries;
12332 if (env->subprog_cnt <= 1)
12335 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
12336 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn))
12339 /* Upon error here we cannot fall back to interpreter but
12340 * need a hard reject of the program. Thus -EFAULT is
12341 * propagated in any case.
12343 subprog = find_subprog(env, i + insn->imm + 1);
12345 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
12346 i + insn->imm + 1);
12349 /* temporarily remember subprog id inside insn instead of
12350 * aux_data, since next loop will split up all insns into funcs
12352 insn->off = subprog;
12353 /* remember original imm in case JIT fails and fallback
12354 * to interpreter will be needed
12356 env->insn_aux_data[i].call_imm = insn->imm;
12357 /* point imm to __bpf_call_base+1 from JITs point of view */
12359 if (bpf_pseudo_func(insn))
12360 /* jit (e.g. x86_64) may emit fewer instructions
12361 * if it learns a u32 imm is the same as a u64 imm.
12362 * Force a non zero here.
12367 err = bpf_prog_alloc_jited_linfo(prog);
12369 goto out_undo_insn;
12372 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
12374 goto out_undo_insn;
12376 for (i = 0; i < env->subprog_cnt; i++) {
12377 subprog_start = subprog_end;
12378 subprog_end = env->subprog_info[i + 1].start;
12380 len = subprog_end - subprog_start;
12381 /* bpf_prog_run() doesn't call subprogs directly,
12382 * hence main prog stats include the runtime of subprogs.
12383 * subprogs don't have IDs and not reachable via prog_get_next_id
12384 * func[i]->stats will never be accessed and stays NULL
12386 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
12389 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
12390 len * sizeof(struct bpf_insn));
12391 func[i]->type = prog->type;
12392 func[i]->len = len;
12393 if (bpf_prog_calc_tag(func[i]))
12395 func[i]->is_func = 1;
12396 func[i]->aux->func_idx = i;
12397 /* Below members will be freed only at prog->aux */
12398 func[i]->aux->btf = prog->aux->btf;
12399 func[i]->aux->func_info = prog->aux->func_info;
12400 func[i]->aux->func_info_cnt = prog->aux->func_info_cnt;
12401 func[i]->aux->poke_tab = prog->aux->poke_tab;
12402 func[i]->aux->size_poke_tab = prog->aux->size_poke_tab;
12404 for (j = 0; j < prog->aux->size_poke_tab; j++) {
12405 struct bpf_jit_poke_descriptor *poke;
12407 poke = &prog->aux->poke_tab[j];
12408 if (poke->insn_idx < subprog_end &&
12409 poke->insn_idx >= subprog_start)
12410 poke->aux = func[i]->aux;
12413 func[i]->aux->name[0] = 'F';
12414 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
12415 func[i]->jit_requested = 1;
12416 func[i]->aux->kfunc_tab = prog->aux->kfunc_tab;
12417 func[i]->aux->linfo = prog->aux->linfo;
12418 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
12419 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
12420 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
12422 insn = func[i]->insnsi;
12423 for (j = 0; j < func[i]->len; j++, insn++) {
12424 if (BPF_CLASS(insn->code) == BPF_LDX &&
12425 BPF_MODE(insn->code) == BPF_PROBE_MEM)
12428 func[i]->aux->num_exentries = num_exentries;
12429 func[i]->aux->tail_call_reachable = env->subprog_info[i].tail_call_reachable;
12430 func[i] = bpf_int_jit_compile(func[i]);
12431 if (!func[i]->jited) {
12438 /* at this point all bpf functions were successfully JITed
12439 * now populate all bpf_calls with correct addresses and
12440 * run last pass of JIT
12442 for (i = 0; i < env->subprog_cnt; i++) {
12443 insn = func[i]->insnsi;
12444 for (j = 0; j < func[i]->len; j++, insn++) {
12445 if (bpf_pseudo_func(insn)) {
12446 subprog = insn->off;
12447 insn[0].imm = (u32)(long)func[subprog]->bpf_func;
12448 insn[1].imm = ((u64)(long)func[subprog]->bpf_func) >> 32;
12451 if (!bpf_pseudo_call(insn))
12453 subprog = insn->off;
12454 insn->imm = BPF_CAST_CALL(func[subprog]->bpf_func) -
12458 /* we use the aux data to keep a list of the start addresses
12459 * of the JITed images for each function in the program
12461 * for some architectures, such as powerpc64, the imm field
12462 * might not be large enough to hold the offset of the start
12463 * address of the callee's JITed image from __bpf_call_base
12465 * in such cases, we can lookup the start address of a callee
12466 * by using its subprog id, available from the off field of
12467 * the call instruction, as an index for this list
12469 func[i]->aux->func = func;
12470 func[i]->aux->func_cnt = env->subprog_cnt;
12472 for (i = 0; i < env->subprog_cnt; i++) {
12473 old_bpf_func = func[i]->bpf_func;
12474 tmp = bpf_int_jit_compile(func[i]);
12475 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
12476 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
12483 /* finally lock prog and jit images for all functions and
12484 * populate kallsysm
12486 for (i = 0; i < env->subprog_cnt; i++) {
12487 bpf_prog_lock_ro(func[i]);
12488 bpf_prog_kallsyms_add(func[i]);
12491 /* Last step: make now unused interpreter insns from main
12492 * prog consistent for later dump requests, so they can
12493 * later look the same as if they were interpreted only.
12495 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
12496 if (bpf_pseudo_func(insn)) {
12497 insn[0].imm = env->insn_aux_data[i].call_imm;
12498 insn[1].imm = insn->off;
12502 if (!bpf_pseudo_call(insn))
12504 insn->off = env->insn_aux_data[i].call_imm;
12505 subprog = find_subprog(env, i + insn->off + 1);
12506 insn->imm = subprog;
12510 prog->bpf_func = func[0]->bpf_func;
12511 prog->aux->func = func;
12512 prog->aux->func_cnt = env->subprog_cnt;
12513 bpf_prog_jit_attempt_done(prog);
12516 /* We failed JIT'ing, so at this point we need to unregister poke
12517 * descriptors from subprogs, so that kernel is not attempting to
12518 * patch it anymore as we're freeing the subprog JIT memory.
12520 for (i = 0; i < prog->aux->size_poke_tab; i++) {
12521 map_ptr = prog->aux->poke_tab[i].tail_call.map;
12522 map_ptr->ops->map_poke_untrack(map_ptr, prog->aux);
12524 /* At this point we're guaranteed that poke descriptors are not
12525 * live anymore. We can just unlink its descriptor table as it's
12526 * released with the main prog.
12528 for (i = 0; i < env->subprog_cnt; i++) {
12531 func[i]->aux->poke_tab = NULL;
12532 bpf_jit_free(func[i]);
12536 /* cleanup main prog to be interpreted */
12537 prog->jit_requested = 0;
12538 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
12539 if (!bpf_pseudo_call(insn))
12542 insn->imm = env->insn_aux_data[i].call_imm;
12544 bpf_prog_jit_attempt_done(prog);
12548 static int fixup_call_args(struct bpf_verifier_env *env)
12550 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
12551 struct bpf_prog *prog = env->prog;
12552 struct bpf_insn *insn = prog->insnsi;
12553 bool has_kfunc_call = bpf_prog_has_kfunc_call(prog);
12558 if (env->prog->jit_requested &&
12559 !bpf_prog_is_dev_bound(env->prog->aux)) {
12560 err = jit_subprogs(env);
12563 if (err == -EFAULT)
12566 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
12567 if (has_kfunc_call) {
12568 verbose(env, "calling kernel functions are not allowed in non-JITed programs\n");
12571 if (env->subprog_cnt > 1 && env->prog->aux->tail_call_reachable) {
12572 /* When JIT fails the progs with bpf2bpf calls and tail_calls
12573 * have to be rejected, since interpreter doesn't support them yet.
12575 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
12578 for (i = 0; i < prog->len; i++, insn++) {
12579 if (bpf_pseudo_func(insn)) {
12580 /* When JIT fails the progs with callback calls
12581 * have to be rejected, since interpreter doesn't support them yet.
12583 verbose(env, "callbacks are not allowed in non-JITed programs\n");
12587 if (!bpf_pseudo_call(insn))
12589 depth = get_callee_stack_depth(env, insn, i);
12592 bpf_patch_call_args(insn, depth);
12599 static int fixup_kfunc_call(struct bpf_verifier_env *env,
12600 struct bpf_insn *insn)
12602 const struct bpf_kfunc_desc *desc;
12604 /* insn->imm has the btf func_id. Replace it with
12605 * an address (relative to __bpf_base_call).
12607 desc = find_kfunc_desc(env->prog, insn->imm);
12609 verbose(env, "verifier internal error: kernel function descriptor not found for func_id %u\n",
12614 insn->imm = desc->imm;
12619 /* Do various post-verification rewrites in a single program pass.
12620 * These rewrites simplify JIT and interpreter implementations.
12622 static int do_misc_fixups(struct bpf_verifier_env *env)
12624 struct bpf_prog *prog = env->prog;
12625 bool expect_blinding = bpf_jit_blinding_enabled(prog);
12626 enum bpf_prog_type prog_type = resolve_prog_type(prog);
12627 struct bpf_insn *insn = prog->insnsi;
12628 const struct bpf_func_proto *fn;
12629 const int insn_cnt = prog->len;
12630 const struct bpf_map_ops *ops;
12631 struct bpf_insn_aux_data *aux;
12632 struct bpf_insn insn_buf[16];
12633 struct bpf_prog *new_prog;
12634 struct bpf_map *map_ptr;
12635 int i, ret, cnt, delta = 0;
12637 for (i = 0; i < insn_cnt; i++, insn++) {
12638 /* Make divide-by-zero exceptions impossible. */
12639 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
12640 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
12641 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
12642 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
12643 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
12644 bool isdiv = BPF_OP(insn->code) == BPF_DIV;
12645 struct bpf_insn *patchlet;
12646 struct bpf_insn chk_and_div[] = {
12647 /* [R,W]x div 0 -> 0 */
12648 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
12649 BPF_JNE | BPF_K, insn->src_reg,
12651 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
12652 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
12655 struct bpf_insn chk_and_mod[] = {
12656 /* [R,W]x mod 0 -> [R,W]x */
12657 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
12658 BPF_JEQ | BPF_K, insn->src_reg,
12659 0, 1 + (is64 ? 0 : 1), 0),
12661 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
12662 BPF_MOV32_REG(insn->dst_reg, insn->dst_reg),
12665 patchlet = isdiv ? chk_and_div : chk_and_mod;
12666 cnt = isdiv ? ARRAY_SIZE(chk_and_div) :
12667 ARRAY_SIZE(chk_and_mod) - (is64 ? 2 : 0);
12669 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
12674 env->prog = prog = new_prog;
12675 insn = new_prog->insnsi + i + delta;
12679 /* Implement LD_ABS and LD_IND with a rewrite, if supported by the program type. */
12680 if (BPF_CLASS(insn->code) == BPF_LD &&
12681 (BPF_MODE(insn->code) == BPF_ABS ||
12682 BPF_MODE(insn->code) == BPF_IND)) {
12683 cnt = env->ops->gen_ld_abs(insn, insn_buf);
12684 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
12685 verbose(env, "bpf verifier is misconfigured\n");
12689 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12694 env->prog = prog = new_prog;
12695 insn = new_prog->insnsi + i + delta;
12699 /* Rewrite pointer arithmetic to mitigate speculation attacks. */
12700 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
12701 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
12702 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
12703 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
12704 struct bpf_insn *patch = &insn_buf[0];
12705 bool issrc, isneg, isimm;
12708 aux = &env->insn_aux_data[i + delta];
12709 if (!aux->alu_state ||
12710 aux->alu_state == BPF_ALU_NON_POINTER)
12713 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
12714 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
12715 BPF_ALU_SANITIZE_SRC;
12716 isimm = aux->alu_state & BPF_ALU_IMMEDIATE;
12718 off_reg = issrc ? insn->src_reg : insn->dst_reg;
12720 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
12723 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
12724 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
12725 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
12726 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
12727 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
12728 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
12729 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX, off_reg);
12732 *patch++ = BPF_MOV64_REG(insn->dst_reg, insn->src_reg);
12733 insn->src_reg = BPF_REG_AX;
12735 insn->code = insn->code == code_add ?
12736 code_sub : code_add;
12738 if (issrc && isneg && !isimm)
12739 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
12740 cnt = patch - insn_buf;
12742 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12747 env->prog = prog = new_prog;
12748 insn = new_prog->insnsi + i + delta;
12752 if (insn->code != (BPF_JMP | BPF_CALL))
12754 if (insn->src_reg == BPF_PSEUDO_CALL)
12756 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
12757 ret = fixup_kfunc_call(env, insn);
12763 if (insn->imm == BPF_FUNC_get_route_realm)
12764 prog->dst_needed = 1;
12765 if (insn->imm == BPF_FUNC_get_prandom_u32)
12766 bpf_user_rnd_init_once();
12767 if (insn->imm == BPF_FUNC_override_return)
12768 prog->kprobe_override = 1;
12769 if (insn->imm == BPF_FUNC_tail_call) {
12770 /* If we tail call into other programs, we
12771 * cannot make any assumptions since they can
12772 * be replaced dynamically during runtime in
12773 * the program array.
12775 prog->cb_access = 1;
12776 if (!allow_tail_call_in_subprogs(env))
12777 prog->aux->stack_depth = MAX_BPF_STACK;
12778 prog->aux->max_pkt_offset = MAX_PACKET_OFF;
12780 /* mark bpf_tail_call as different opcode to avoid
12781 * conditional branch in the interpreter for every normal
12782 * call and to prevent accidental JITing by JIT compiler
12783 * that doesn't support bpf_tail_call yet
12786 insn->code = BPF_JMP | BPF_TAIL_CALL;
12788 aux = &env->insn_aux_data[i + delta];
12789 if (env->bpf_capable && !expect_blinding &&
12790 prog->jit_requested &&
12791 !bpf_map_key_poisoned(aux) &&
12792 !bpf_map_ptr_poisoned(aux) &&
12793 !bpf_map_ptr_unpriv(aux)) {
12794 struct bpf_jit_poke_descriptor desc = {
12795 .reason = BPF_POKE_REASON_TAIL_CALL,
12796 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
12797 .tail_call.key = bpf_map_key_immediate(aux),
12798 .insn_idx = i + delta,
12801 ret = bpf_jit_add_poke_descriptor(prog, &desc);
12803 verbose(env, "adding tail call poke descriptor failed\n");
12807 insn->imm = ret + 1;
12811 if (!bpf_map_ptr_unpriv(aux))
12814 /* instead of changing every JIT dealing with tail_call
12815 * emit two extra insns:
12816 * if (index >= max_entries) goto out;
12817 * index &= array->index_mask;
12818 * to avoid out-of-bounds cpu speculation
12820 if (bpf_map_ptr_poisoned(aux)) {
12821 verbose(env, "tail_call abusing map_ptr\n");
12825 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
12826 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
12827 map_ptr->max_entries, 2);
12828 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
12829 container_of(map_ptr,
12832 insn_buf[2] = *insn;
12834 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12839 env->prog = prog = new_prog;
12840 insn = new_prog->insnsi + i + delta;
12844 if (insn->imm == BPF_FUNC_timer_set_callback) {
12845 /* The verifier will process callback_fn as many times as necessary
12846 * with different maps and the register states prepared by
12847 * set_timer_callback_state will be accurate.
12849 * The following use case is valid:
12850 * map1 is shared by prog1, prog2, prog3.
12851 * prog1 calls bpf_timer_init for some map1 elements
12852 * prog2 calls bpf_timer_set_callback for some map1 elements.
12853 * Those that were not bpf_timer_init-ed will return -EINVAL.
12854 * prog3 calls bpf_timer_start for some map1 elements.
12855 * Those that were not both bpf_timer_init-ed and
12856 * bpf_timer_set_callback-ed will return -EINVAL.
12858 struct bpf_insn ld_addrs[2] = {
12859 BPF_LD_IMM64(BPF_REG_3, (long)prog->aux),
12862 insn_buf[0] = ld_addrs[0];
12863 insn_buf[1] = ld_addrs[1];
12864 insn_buf[2] = *insn;
12867 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12872 env->prog = prog = new_prog;
12873 insn = new_prog->insnsi + i + delta;
12874 goto patch_call_imm;
12877 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
12878 * and other inlining handlers are currently limited to 64 bit
12881 if (prog->jit_requested && BITS_PER_LONG == 64 &&
12882 (insn->imm == BPF_FUNC_map_lookup_elem ||
12883 insn->imm == BPF_FUNC_map_update_elem ||
12884 insn->imm == BPF_FUNC_map_delete_elem ||
12885 insn->imm == BPF_FUNC_map_push_elem ||
12886 insn->imm == BPF_FUNC_map_pop_elem ||
12887 insn->imm == BPF_FUNC_map_peek_elem ||
12888 insn->imm == BPF_FUNC_redirect_map)) {
12889 aux = &env->insn_aux_data[i + delta];
12890 if (bpf_map_ptr_poisoned(aux))
12891 goto patch_call_imm;
12893 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
12894 ops = map_ptr->ops;
12895 if (insn->imm == BPF_FUNC_map_lookup_elem &&
12896 ops->map_gen_lookup) {
12897 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
12898 if (cnt == -EOPNOTSUPP)
12899 goto patch_map_ops_generic;
12900 if (cnt <= 0 || cnt >= ARRAY_SIZE(insn_buf)) {
12901 verbose(env, "bpf verifier is misconfigured\n");
12905 new_prog = bpf_patch_insn_data(env, i + delta,
12911 env->prog = prog = new_prog;
12912 insn = new_prog->insnsi + i + delta;
12916 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
12917 (void *(*)(struct bpf_map *map, void *key))NULL));
12918 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
12919 (int (*)(struct bpf_map *map, void *key))NULL));
12920 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
12921 (int (*)(struct bpf_map *map, void *key, void *value,
12923 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
12924 (int (*)(struct bpf_map *map, void *value,
12926 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
12927 (int (*)(struct bpf_map *map, void *value))NULL));
12928 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
12929 (int (*)(struct bpf_map *map, void *value))NULL));
12930 BUILD_BUG_ON(!__same_type(ops->map_redirect,
12931 (int (*)(struct bpf_map *map, u32 ifindex, u64 flags))NULL));
12933 patch_map_ops_generic:
12934 switch (insn->imm) {
12935 case BPF_FUNC_map_lookup_elem:
12936 insn->imm = BPF_CAST_CALL(ops->map_lookup_elem) -
12939 case BPF_FUNC_map_update_elem:
12940 insn->imm = BPF_CAST_CALL(ops->map_update_elem) -
12943 case BPF_FUNC_map_delete_elem:
12944 insn->imm = BPF_CAST_CALL(ops->map_delete_elem) -
12947 case BPF_FUNC_map_push_elem:
12948 insn->imm = BPF_CAST_CALL(ops->map_push_elem) -
12951 case BPF_FUNC_map_pop_elem:
12952 insn->imm = BPF_CAST_CALL(ops->map_pop_elem) -
12955 case BPF_FUNC_map_peek_elem:
12956 insn->imm = BPF_CAST_CALL(ops->map_peek_elem) -
12959 case BPF_FUNC_redirect_map:
12960 insn->imm = BPF_CAST_CALL(ops->map_redirect) -
12965 goto patch_call_imm;
12968 /* Implement bpf_jiffies64 inline. */
12969 if (prog->jit_requested && BITS_PER_LONG == 64 &&
12970 insn->imm == BPF_FUNC_jiffies64) {
12971 struct bpf_insn ld_jiffies_addr[2] = {
12972 BPF_LD_IMM64(BPF_REG_0,
12973 (unsigned long)&jiffies),
12976 insn_buf[0] = ld_jiffies_addr[0];
12977 insn_buf[1] = ld_jiffies_addr[1];
12978 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
12982 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
12988 env->prog = prog = new_prog;
12989 insn = new_prog->insnsi + i + delta;
12993 /* Implement bpf_get_func_ip inline. */
12994 if (prog_type == BPF_PROG_TYPE_TRACING &&
12995 insn->imm == BPF_FUNC_get_func_ip) {
12996 /* Load IP address from ctx - 8 */
12997 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
12999 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
13003 env->prog = prog = new_prog;
13004 insn = new_prog->insnsi + i + delta;
13009 fn = env->ops->get_func_proto(insn->imm, env->prog);
13010 /* all functions that have prototype and verifier allowed
13011 * programs to call them, must be real in-kernel functions
13015 "kernel subsystem misconfigured func %s#%d\n",
13016 func_id_name(insn->imm), insn->imm);
13019 insn->imm = fn->func - __bpf_call_base;
13022 /* Since poke tab is now finalized, publish aux to tracker. */
13023 for (i = 0; i < prog->aux->size_poke_tab; i++) {
13024 map_ptr = prog->aux->poke_tab[i].tail_call.map;
13025 if (!map_ptr->ops->map_poke_track ||
13026 !map_ptr->ops->map_poke_untrack ||
13027 !map_ptr->ops->map_poke_run) {
13028 verbose(env, "bpf verifier is misconfigured\n");
13032 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
13034 verbose(env, "tracking tail call prog failed\n");
13039 sort_kfunc_descs_by_imm(env->prog);
13044 static void free_states(struct bpf_verifier_env *env)
13046 struct bpf_verifier_state_list *sl, *sln;
13049 sl = env->free_list;
13052 free_verifier_state(&sl->state, false);
13056 env->free_list = NULL;
13058 if (!env->explored_states)
13061 for (i = 0; i < state_htab_size(env); i++) {
13062 sl = env->explored_states[i];
13066 free_verifier_state(&sl->state, false);
13070 env->explored_states[i] = NULL;
13074 static int do_check_common(struct bpf_verifier_env *env, int subprog)
13076 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
13077 struct bpf_verifier_state *state;
13078 struct bpf_reg_state *regs;
13081 env->prev_linfo = NULL;
13084 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
13087 state->curframe = 0;
13088 state->speculative = false;
13089 state->branches = 1;
13090 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
13091 if (!state->frame[0]) {
13095 env->cur_state = state;
13096 init_func_state(env, state->frame[0],
13097 BPF_MAIN_FUNC /* callsite */,
13101 regs = state->frame[state->curframe]->regs;
13102 if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
13103 ret = btf_prepare_func_args(env, subprog, regs);
13106 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
13107 if (regs[i].type == PTR_TO_CTX)
13108 mark_reg_known_zero(env, regs, i);
13109 else if (regs[i].type == SCALAR_VALUE)
13110 mark_reg_unknown(env, regs, i);
13111 else if (base_type(regs[i].type) == PTR_TO_MEM) {
13112 const u32 mem_size = regs[i].mem_size;
13114 mark_reg_known_zero(env, regs, i);
13115 regs[i].mem_size = mem_size;
13116 regs[i].id = ++env->id_gen;
13120 /* 1st arg to a function */
13121 regs[BPF_REG_1].type = PTR_TO_CTX;
13122 mark_reg_known_zero(env, regs, BPF_REG_1);
13123 ret = btf_check_subprog_arg_match(env, subprog, regs);
13124 if (ret == -EFAULT)
13125 /* unlikely verifier bug. abort.
13126 * ret == 0 and ret < 0 are sadly acceptable for
13127 * main() function due to backward compatibility.
13128 * Like socket filter program may be written as:
13129 * int bpf_prog(struct pt_regs *ctx)
13130 * and never dereference that ctx in the program.
13131 * 'struct pt_regs' is a type mismatch for socket
13132 * filter that should be using 'struct __sk_buff'.
13137 ret = do_check(env);
13139 /* check for NULL is necessary, since cur_state can be freed inside
13140 * do_check() under memory pressure.
13142 if (env->cur_state) {
13143 free_verifier_state(env->cur_state, true);
13144 env->cur_state = NULL;
13146 while (!pop_stack(env, NULL, NULL, false));
13147 if (!ret && pop_log)
13148 bpf_vlog_reset(&env->log, 0);
13153 /* Verify all global functions in a BPF program one by one based on their BTF.
13154 * All global functions must pass verification. Otherwise the whole program is rejected.
13165 * foo() will be verified first for R1=any_scalar_value. During verification it
13166 * will be assumed that bar() already verified successfully and call to bar()
13167 * from foo() will be checked for type match only. Later bar() will be verified
13168 * independently to check that it's safe for R1=any_scalar_value.
13170 static int do_check_subprogs(struct bpf_verifier_env *env)
13172 struct bpf_prog_aux *aux = env->prog->aux;
13175 if (!aux->func_info)
13178 for (i = 1; i < env->subprog_cnt; i++) {
13179 if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL)
13181 env->insn_idx = env->subprog_info[i].start;
13182 WARN_ON_ONCE(env->insn_idx == 0);
13183 ret = do_check_common(env, i);
13186 } else if (env->log.level & BPF_LOG_LEVEL) {
13188 "Func#%d is safe for any args that match its prototype\n",
13195 static int do_check_main(struct bpf_verifier_env *env)
13200 ret = do_check_common(env, 0);
13202 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
13207 static void print_verification_stats(struct bpf_verifier_env *env)
13211 if (env->log.level & BPF_LOG_STATS) {
13212 verbose(env, "verification time %lld usec\n",
13213 div_u64(env->verification_time, 1000));
13214 verbose(env, "stack depth ");
13215 for (i = 0; i < env->subprog_cnt; i++) {
13216 u32 depth = env->subprog_info[i].stack_depth;
13218 verbose(env, "%d", depth);
13219 if (i + 1 < env->subprog_cnt)
13222 verbose(env, "\n");
13224 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
13225 "total_states %d peak_states %d mark_read %d\n",
13226 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
13227 env->max_states_per_insn, env->total_states,
13228 env->peak_states, env->longest_mark_read_walk);
13231 static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
13233 const struct btf_type *t, *func_proto;
13234 const struct bpf_struct_ops *st_ops;
13235 const struct btf_member *member;
13236 struct bpf_prog *prog = env->prog;
13237 u32 btf_id, member_idx;
13240 if (!prog->gpl_compatible) {
13241 verbose(env, "struct ops programs must have a GPL compatible license\n");
13245 btf_id = prog->aux->attach_btf_id;
13246 st_ops = bpf_struct_ops_find(btf_id);
13248 verbose(env, "attach_btf_id %u is not a supported struct\n",
13254 member_idx = prog->expected_attach_type;
13255 if (member_idx >= btf_type_vlen(t)) {
13256 verbose(env, "attach to invalid member idx %u of struct %s\n",
13257 member_idx, st_ops->name);
13261 member = &btf_type_member(t)[member_idx];
13262 mname = btf_name_by_offset(btf_vmlinux, member->name_off);
13263 func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type,
13266 verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
13267 mname, member_idx, st_ops->name);
13271 if (st_ops->check_member) {
13272 int err = st_ops->check_member(t, member);
13275 verbose(env, "attach to unsupported member %s of struct %s\n",
13276 mname, st_ops->name);
13281 prog->aux->attach_func_proto = func_proto;
13282 prog->aux->attach_func_name = mname;
13283 env->ops = st_ops->verifier_ops;
13287 #define SECURITY_PREFIX "security_"
13289 static int check_attach_modify_return(unsigned long addr, const char *func_name)
13291 if (within_error_injection_list(addr) ||
13292 !strncmp(SECURITY_PREFIX, func_name, sizeof(SECURITY_PREFIX) - 1))
13298 /* list of non-sleepable functions that are otherwise on
13299 * ALLOW_ERROR_INJECTION list
13301 BTF_SET_START(btf_non_sleepable_error_inject)
13302 /* Three functions below can be called from sleepable and non-sleepable context.
13303 * Assume non-sleepable from bpf safety point of view.
13305 BTF_ID(func, __add_to_page_cache_locked)
13306 BTF_ID(func, should_fail_alloc_page)
13307 BTF_ID(func, should_failslab)
13308 BTF_SET_END(btf_non_sleepable_error_inject)
13310 static int check_non_sleepable_error_inject(u32 btf_id)
13312 return btf_id_set_contains(&btf_non_sleepable_error_inject, btf_id);
13315 int bpf_check_attach_target(struct bpf_verifier_log *log,
13316 const struct bpf_prog *prog,
13317 const struct bpf_prog *tgt_prog,
13319 struct bpf_attach_target_info *tgt_info)
13321 bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
13322 const char prefix[] = "btf_trace_";
13323 int ret = 0, subprog = -1, i;
13324 const struct btf_type *t;
13325 bool conservative = true;
13331 bpf_log(log, "Tracing programs must provide btf_id\n");
13334 btf = tgt_prog ? tgt_prog->aux->btf : prog->aux->attach_btf;
13337 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
13340 t = btf_type_by_id(btf, btf_id);
13342 bpf_log(log, "attach_btf_id %u is invalid\n", btf_id);
13345 tname = btf_name_by_offset(btf, t->name_off);
13347 bpf_log(log, "attach_btf_id %u doesn't have a name\n", btf_id);
13351 struct bpf_prog_aux *aux = tgt_prog->aux;
13353 for (i = 0; i < aux->func_info_cnt; i++)
13354 if (aux->func_info[i].type_id == btf_id) {
13358 if (subprog == -1) {
13359 bpf_log(log, "Subprog %s doesn't exist\n", tname);
13362 conservative = aux->func_info_aux[subprog].unreliable;
13363 if (prog_extension) {
13364 if (conservative) {
13366 "Cannot replace static functions\n");
13369 if (!prog->jit_requested) {
13371 "Extension programs should be JITed\n");
13375 if (!tgt_prog->jited) {
13376 bpf_log(log, "Can attach to only JITed progs\n");
13379 if (tgt_prog->type == prog->type) {
13380 /* Cannot fentry/fexit another fentry/fexit program.
13381 * Cannot attach program extension to another extension.
13382 * It's ok to attach fentry/fexit to extension program.
13384 bpf_log(log, "Cannot recursively attach\n");
13387 if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
13389 (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
13390 tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
13391 /* Program extensions can extend all program types
13392 * except fentry/fexit. The reason is the following.
13393 * The fentry/fexit programs are used for performance
13394 * analysis, stats and can be attached to any program
13395 * type except themselves. When extension program is
13396 * replacing XDP function it is necessary to allow
13397 * performance analysis of all functions. Both original
13398 * XDP program and its program extension. Hence
13399 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
13400 * allowed. If extending of fentry/fexit was allowed it
13401 * would be possible to create long call chain
13402 * fentry->extension->fentry->extension beyond
13403 * reasonable stack size. Hence extending fentry is not
13406 bpf_log(log, "Cannot extend fentry/fexit\n");
13410 if (prog_extension) {
13411 bpf_log(log, "Cannot replace kernel functions\n");
13416 switch (prog->expected_attach_type) {
13417 case BPF_TRACE_RAW_TP:
13420 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
13423 if (!btf_type_is_typedef(t)) {
13424 bpf_log(log, "attach_btf_id %u is not a typedef\n",
13428 if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
13429 bpf_log(log, "attach_btf_id %u points to wrong type name %s\n",
13433 tname += sizeof(prefix) - 1;
13434 t = btf_type_by_id(btf, t->type);
13435 if (!btf_type_is_ptr(t))
13436 /* should never happen in valid vmlinux build */
13438 t = btf_type_by_id(btf, t->type);
13439 if (!btf_type_is_func_proto(t))
13440 /* should never happen in valid vmlinux build */
13444 case BPF_TRACE_ITER:
13445 if (!btf_type_is_func(t)) {
13446 bpf_log(log, "attach_btf_id %u is not a function\n",
13450 t = btf_type_by_id(btf, t->type);
13451 if (!btf_type_is_func_proto(t))
13453 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
13458 if (!prog_extension)
13461 case BPF_MODIFY_RETURN:
13463 case BPF_TRACE_FENTRY:
13464 case BPF_TRACE_FEXIT:
13465 if (!btf_type_is_func(t)) {
13466 bpf_log(log, "attach_btf_id %u is not a function\n",
13470 if (prog_extension &&
13471 btf_check_type_match(log, prog, btf, t))
13473 t = btf_type_by_id(btf, t->type);
13474 if (!btf_type_is_func_proto(t))
13477 if ((prog->aux->saved_dst_prog_type || prog->aux->saved_dst_attach_type) &&
13478 (!tgt_prog || prog->aux->saved_dst_prog_type != tgt_prog->type ||
13479 prog->aux->saved_dst_attach_type != tgt_prog->expected_attach_type))
13482 if (tgt_prog && conservative)
13485 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
13491 addr = (long) tgt_prog->bpf_func;
13493 addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
13495 addr = kallsyms_lookup_name(tname);
13498 "The address of function %s cannot be found\n",
13504 if (prog->aux->sleepable) {
13506 switch (prog->type) {
13507 case BPF_PROG_TYPE_TRACING:
13508 /* fentry/fexit/fmod_ret progs can be sleepable only if they are
13509 * attached to ALLOW_ERROR_INJECTION and are not in denylist.
13511 if (!check_non_sleepable_error_inject(btf_id) &&
13512 within_error_injection_list(addr))
13515 case BPF_PROG_TYPE_LSM:
13516 /* LSM progs check that they are attached to bpf_lsm_*() funcs.
13517 * Only some of them are sleepable.
13519 if (bpf_lsm_is_sleepable_hook(btf_id))
13526 bpf_log(log, "%s is not sleepable\n", tname);
13529 } else if (prog->expected_attach_type == BPF_MODIFY_RETURN) {
13531 bpf_log(log, "can't modify return codes of BPF programs\n");
13534 ret = check_attach_modify_return(addr, tname);
13536 bpf_log(log, "%s() is not modifiable\n", tname);
13543 tgt_info->tgt_addr = addr;
13544 tgt_info->tgt_name = tname;
13545 tgt_info->tgt_type = t;
13549 BTF_SET_START(btf_id_deny)
13552 BTF_ID(func, migrate_disable)
13553 BTF_ID(func, migrate_enable)
13555 #if !defined CONFIG_PREEMPT_RCU && !defined CONFIG_TINY_RCU
13556 BTF_ID(func, rcu_read_unlock_strict)
13558 BTF_SET_END(btf_id_deny)
13560 static int check_attach_btf_id(struct bpf_verifier_env *env)
13562 struct bpf_prog *prog = env->prog;
13563 struct bpf_prog *tgt_prog = prog->aux->dst_prog;
13564 struct bpf_attach_target_info tgt_info = {};
13565 u32 btf_id = prog->aux->attach_btf_id;
13566 struct bpf_trampoline *tr;
13570 if (prog->type == BPF_PROG_TYPE_SYSCALL) {
13571 if (prog->aux->sleepable)
13572 /* attach_btf_id checked to be zero already */
13574 verbose(env, "Syscall programs can only be sleepable\n");
13578 if (prog->aux->sleepable && prog->type != BPF_PROG_TYPE_TRACING &&
13579 prog->type != BPF_PROG_TYPE_LSM) {
13580 verbose(env, "Only fentry/fexit/fmod_ret and lsm programs can be sleepable\n");
13584 if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
13585 return check_struct_ops_btf_id(env);
13587 if (prog->type != BPF_PROG_TYPE_TRACING &&
13588 prog->type != BPF_PROG_TYPE_LSM &&
13589 prog->type != BPF_PROG_TYPE_EXT)
13592 ret = bpf_check_attach_target(&env->log, prog, tgt_prog, btf_id, &tgt_info);
13596 if (tgt_prog && prog->type == BPF_PROG_TYPE_EXT) {
13597 /* to make freplace equivalent to their targets, they need to
13598 * inherit env->ops and expected_attach_type for the rest of the
13601 env->ops = bpf_verifier_ops[tgt_prog->type];
13602 prog->expected_attach_type = tgt_prog->expected_attach_type;
13605 /* store info about the attachment target that will be used later */
13606 prog->aux->attach_func_proto = tgt_info.tgt_type;
13607 prog->aux->attach_func_name = tgt_info.tgt_name;
13610 prog->aux->saved_dst_prog_type = tgt_prog->type;
13611 prog->aux->saved_dst_attach_type = tgt_prog->expected_attach_type;
13614 if (prog->expected_attach_type == BPF_TRACE_RAW_TP) {
13615 prog->aux->attach_btf_trace = true;
13617 } else if (prog->expected_attach_type == BPF_TRACE_ITER) {
13618 if (!bpf_iter_prog_supported(prog))
13623 if (prog->type == BPF_PROG_TYPE_LSM) {
13624 ret = bpf_lsm_verify_prog(&env->log, prog);
13627 } else if (prog->type == BPF_PROG_TYPE_TRACING &&
13628 btf_id_set_contains(&btf_id_deny, btf_id)) {
13632 key = bpf_trampoline_compute_key(tgt_prog, prog->aux->attach_btf, btf_id);
13633 tr = bpf_trampoline_get(key, &tgt_info);
13637 prog->aux->dst_trampoline = tr;
13641 struct btf *bpf_get_btf_vmlinux(void)
13643 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
13644 mutex_lock(&bpf_verifier_lock);
13646 btf_vmlinux = btf_parse_vmlinux();
13647 mutex_unlock(&bpf_verifier_lock);
13649 return btf_vmlinux;
13652 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr, bpfptr_t uattr)
13654 u64 start_time = ktime_get_ns();
13655 struct bpf_verifier_env *env;
13656 struct bpf_verifier_log *log;
13657 int i, len, ret = -EINVAL;
13660 /* no program is valid */
13661 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
13664 /* 'struct bpf_verifier_env' can be global, but since it's not small,
13665 * allocate/free it every time bpf_check() is called
13667 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
13672 len = (*prog)->len;
13673 env->insn_aux_data =
13674 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
13676 if (!env->insn_aux_data)
13678 for (i = 0; i < len; i++)
13679 env->insn_aux_data[i].orig_idx = i;
13681 env->ops = bpf_verifier_ops[env->prog->type];
13682 env->fd_array = make_bpfptr(attr->fd_array, uattr.is_kernel);
13683 is_priv = bpf_capable();
13685 bpf_get_btf_vmlinux();
13687 /* grab the mutex to protect few globals used by verifier */
13689 mutex_lock(&bpf_verifier_lock);
13691 if (attr->log_level || attr->log_buf || attr->log_size) {
13692 /* user requested verbose verifier output
13693 * and supplied buffer to store the verification trace
13695 log->level = attr->log_level;
13696 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
13697 log->len_total = attr->log_size;
13699 /* log attributes have to be sane */
13700 if (!bpf_verifier_log_attr_valid(log)) {
13706 if (IS_ERR(btf_vmlinux)) {
13707 /* Either gcc or pahole or kernel are broken. */
13708 verbose(env, "in-kernel BTF is malformed\n");
13709 ret = PTR_ERR(btf_vmlinux);
13710 goto skip_full_check;
13713 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
13714 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
13715 env->strict_alignment = true;
13716 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
13717 env->strict_alignment = false;
13719 env->allow_ptr_leaks = bpf_allow_ptr_leaks();
13720 env->allow_uninit_stack = bpf_allow_uninit_stack();
13721 env->allow_ptr_to_map_access = bpf_allow_ptr_to_map_access();
13722 env->bypass_spec_v1 = bpf_bypass_spec_v1();
13723 env->bypass_spec_v4 = bpf_bypass_spec_v4();
13724 env->bpf_capable = bpf_capable();
13727 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
13729 env->explored_states = kvcalloc(state_htab_size(env),
13730 sizeof(struct bpf_verifier_state_list *),
13733 if (!env->explored_states)
13734 goto skip_full_check;
13736 ret = add_subprog_and_kfunc(env);
13738 goto skip_full_check;
13740 ret = check_subprogs(env);
13742 goto skip_full_check;
13744 ret = check_btf_info(env, attr, uattr);
13746 goto skip_full_check;
13748 ret = check_attach_btf_id(env);
13750 goto skip_full_check;
13752 ret = resolve_pseudo_ldimm64(env);
13754 goto skip_full_check;
13756 if (bpf_prog_is_dev_bound(env->prog->aux)) {
13757 ret = bpf_prog_offload_verifier_prep(env->prog);
13759 goto skip_full_check;
13762 ret = check_cfg(env);
13764 goto skip_full_check;
13766 ret = do_check_subprogs(env);
13767 ret = ret ?: do_check_main(env);
13769 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
13770 ret = bpf_prog_offload_finalize(env);
13773 kvfree(env->explored_states);
13776 ret = check_max_stack_depth(env);
13778 /* instruction rewrites happen after this point */
13781 opt_hard_wire_dead_code_branches(env);
13783 ret = opt_remove_dead_code(env);
13785 ret = opt_remove_nops(env);
13788 sanitize_dead_code(env);
13792 /* program is valid, convert *(u32*)(ctx + off) accesses */
13793 ret = convert_ctx_accesses(env);
13796 ret = do_misc_fixups(env);
13798 /* do 32-bit optimization after insn patching has done so those patched
13799 * insns could be handled correctly.
13801 if (ret == 0 && !bpf_prog_is_dev_bound(env->prog->aux)) {
13802 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
13803 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
13808 ret = fixup_call_args(env);
13810 env->verification_time = ktime_get_ns() - start_time;
13811 print_verification_stats(env);
13813 if (log->level && bpf_verifier_log_full(log))
13815 if (log->level && !log->ubuf) {
13817 goto err_release_maps;
13821 goto err_release_maps;
13823 if (env->used_map_cnt) {
13824 /* if program passed verifier, update used_maps in bpf_prog_info */
13825 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
13826 sizeof(env->used_maps[0]),
13829 if (!env->prog->aux->used_maps) {
13831 goto err_release_maps;
13834 memcpy(env->prog->aux->used_maps, env->used_maps,
13835 sizeof(env->used_maps[0]) * env->used_map_cnt);
13836 env->prog->aux->used_map_cnt = env->used_map_cnt;
13838 if (env->used_btf_cnt) {
13839 /* if program passed verifier, update used_btfs in bpf_prog_aux */
13840 env->prog->aux->used_btfs = kmalloc_array(env->used_btf_cnt,
13841 sizeof(env->used_btfs[0]),
13843 if (!env->prog->aux->used_btfs) {
13845 goto err_release_maps;
13848 memcpy(env->prog->aux->used_btfs, env->used_btfs,
13849 sizeof(env->used_btfs[0]) * env->used_btf_cnt);
13850 env->prog->aux->used_btf_cnt = env->used_btf_cnt;
13852 if (env->used_map_cnt || env->used_btf_cnt) {
13853 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
13854 * bpf_ld_imm64 instructions
13856 convert_pseudo_ld_imm64(env);
13859 adjust_btf_func(env);
13862 if (!env->prog->aux->used_maps)
13863 /* if we didn't copy map pointers into bpf_prog_info, release
13864 * them now. Otherwise free_used_maps() will release them.
13867 if (!env->prog->aux->used_btfs)
13870 /* extension progs temporarily inherit the attach_type of their targets
13871 for verification purposes, so set it back to zero before returning
13873 if (env->prog->type == BPF_PROG_TYPE_EXT)
13874 env->prog->expected_attach_type = 0;
13879 mutex_unlock(&bpf_verifier_lock);
13880 vfree(env->insn_aux_data);