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 pathes 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 ether 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 struct bpf_call_arg_meta {
238 struct bpf_map *map_ptr;
253 struct btf *btf_vmlinux;
255 static DEFINE_MUTEX(bpf_verifier_lock);
257 static const struct bpf_line_info *
258 find_linfo(const struct bpf_verifier_env *env, u32 insn_off)
260 const struct bpf_line_info *linfo;
261 const struct bpf_prog *prog;
265 nr_linfo = prog->aux->nr_linfo;
267 if (!nr_linfo || insn_off >= prog->len)
270 linfo = prog->aux->linfo;
271 for (i = 1; i < nr_linfo; i++)
272 if (insn_off < linfo[i].insn_off)
275 return &linfo[i - 1];
278 void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
283 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
285 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
286 "verifier log line truncated - local buffer too short\n");
288 n = min(log->len_total - log->len_used - 1, n);
291 if (log->level == BPF_LOG_KERNEL) {
292 pr_err("BPF:%s\n", log->kbuf);
295 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
301 static void bpf_vlog_reset(struct bpf_verifier_log *log, u32 new_pos)
305 if (!bpf_verifier_log_needed(log))
308 log->len_used = new_pos;
309 if (put_user(zero, log->ubuf + new_pos))
313 /* log_level controls verbosity level of eBPF verifier.
314 * bpf_verifier_log_write() is used to dump the verification trace to the log,
315 * so the user can figure out what's wrong with the program
317 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
318 const char *fmt, ...)
322 if (!bpf_verifier_log_needed(&env->log))
326 bpf_verifier_vlog(&env->log, fmt, args);
329 EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
331 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
333 struct bpf_verifier_env *env = private_data;
336 if (!bpf_verifier_log_needed(&env->log))
340 bpf_verifier_vlog(&env->log, fmt, args);
344 __printf(2, 3) void bpf_log(struct bpf_verifier_log *log,
345 const char *fmt, ...)
349 if (!bpf_verifier_log_needed(log))
353 bpf_verifier_vlog(log, fmt, args);
357 static const char *ltrim(const char *s)
365 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env,
367 const char *prefix_fmt, ...)
369 const struct bpf_line_info *linfo;
371 if (!bpf_verifier_log_needed(&env->log))
374 linfo = find_linfo(env, insn_off);
375 if (!linfo || linfo == env->prev_linfo)
381 va_start(args, prefix_fmt);
382 bpf_verifier_vlog(&env->log, prefix_fmt, args);
387 ltrim(btf_name_by_offset(env->prog->aux->btf,
390 env->prev_linfo = linfo;
393 static bool type_is_pkt_pointer(enum bpf_reg_type type)
395 return type == PTR_TO_PACKET ||
396 type == PTR_TO_PACKET_META;
399 static bool type_is_sk_pointer(enum bpf_reg_type type)
401 return type == PTR_TO_SOCKET ||
402 type == PTR_TO_SOCK_COMMON ||
403 type == PTR_TO_TCP_SOCK ||
404 type == PTR_TO_XDP_SOCK;
407 static bool reg_type_not_null(enum bpf_reg_type type)
409 return type == PTR_TO_SOCKET ||
410 type == PTR_TO_TCP_SOCK ||
411 type == PTR_TO_MAP_VALUE ||
412 type == PTR_TO_SOCK_COMMON;
415 static bool reg_type_may_be_null(enum bpf_reg_type type)
417 return type == PTR_TO_MAP_VALUE_OR_NULL ||
418 type == PTR_TO_SOCKET_OR_NULL ||
419 type == PTR_TO_SOCK_COMMON_OR_NULL ||
420 type == PTR_TO_TCP_SOCK_OR_NULL ||
421 type == PTR_TO_BTF_ID_OR_NULL ||
422 type == PTR_TO_MEM_OR_NULL ||
423 type == PTR_TO_RDONLY_BUF_OR_NULL ||
424 type == PTR_TO_RDWR_BUF_OR_NULL;
427 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
429 return reg->type == PTR_TO_MAP_VALUE &&
430 map_value_has_spin_lock(reg->map_ptr);
433 static bool reg_type_may_be_refcounted_or_null(enum bpf_reg_type type)
435 return type == PTR_TO_SOCKET ||
436 type == PTR_TO_SOCKET_OR_NULL ||
437 type == PTR_TO_TCP_SOCK ||
438 type == PTR_TO_TCP_SOCK_OR_NULL ||
439 type == PTR_TO_MEM ||
440 type == PTR_TO_MEM_OR_NULL;
443 static bool arg_type_may_be_refcounted(enum bpf_arg_type type)
445 return type == ARG_PTR_TO_SOCK_COMMON;
448 static bool arg_type_may_be_null(enum bpf_arg_type type)
450 return type == ARG_PTR_TO_MAP_VALUE_OR_NULL ||
451 type == ARG_PTR_TO_MEM_OR_NULL ||
452 type == ARG_PTR_TO_CTX_OR_NULL ||
453 type == ARG_PTR_TO_SOCKET_OR_NULL ||
454 type == ARG_PTR_TO_ALLOC_MEM_OR_NULL;
457 /* Determine whether the function releases some resources allocated by another
458 * function call. The first reference type argument will be assumed to be
459 * released by release_reference().
461 static bool is_release_function(enum bpf_func_id func_id)
463 return func_id == BPF_FUNC_sk_release ||
464 func_id == BPF_FUNC_ringbuf_submit ||
465 func_id == BPF_FUNC_ringbuf_discard;
468 static bool may_be_acquire_function(enum bpf_func_id func_id)
470 return func_id == BPF_FUNC_sk_lookup_tcp ||
471 func_id == BPF_FUNC_sk_lookup_udp ||
472 func_id == BPF_FUNC_skc_lookup_tcp ||
473 func_id == BPF_FUNC_map_lookup_elem ||
474 func_id == BPF_FUNC_ringbuf_reserve;
477 static bool is_acquire_function(enum bpf_func_id func_id,
478 const struct bpf_map *map)
480 enum bpf_map_type map_type = map ? map->map_type : BPF_MAP_TYPE_UNSPEC;
482 if (func_id == BPF_FUNC_sk_lookup_tcp ||
483 func_id == BPF_FUNC_sk_lookup_udp ||
484 func_id == BPF_FUNC_skc_lookup_tcp ||
485 func_id == BPF_FUNC_ringbuf_reserve)
488 if (func_id == BPF_FUNC_map_lookup_elem &&
489 (map_type == BPF_MAP_TYPE_SOCKMAP ||
490 map_type == BPF_MAP_TYPE_SOCKHASH))
496 static bool is_ptr_cast_function(enum bpf_func_id func_id)
498 return func_id == BPF_FUNC_tcp_sock ||
499 func_id == BPF_FUNC_sk_fullsock ||
500 func_id == BPF_FUNC_skc_to_tcp_sock ||
501 func_id == BPF_FUNC_skc_to_tcp6_sock ||
502 func_id == BPF_FUNC_skc_to_udp6_sock ||
503 func_id == BPF_FUNC_skc_to_tcp_timewait_sock ||
504 func_id == BPF_FUNC_skc_to_tcp_request_sock;
507 static bool is_cmpxchg_insn(const struct bpf_insn *insn)
509 return BPF_CLASS(insn->code) == BPF_STX &&
510 BPF_MODE(insn->code) == BPF_ATOMIC &&
511 insn->imm == BPF_CMPXCHG;
514 /* string representation of 'enum bpf_reg_type' */
515 static const char * const reg_type_str[] = {
517 [SCALAR_VALUE] = "inv",
518 [PTR_TO_CTX] = "ctx",
519 [CONST_PTR_TO_MAP] = "map_ptr",
520 [PTR_TO_MAP_VALUE] = "map_value",
521 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
522 [PTR_TO_STACK] = "fp",
523 [PTR_TO_PACKET] = "pkt",
524 [PTR_TO_PACKET_META] = "pkt_meta",
525 [PTR_TO_PACKET_END] = "pkt_end",
526 [PTR_TO_FLOW_KEYS] = "flow_keys",
527 [PTR_TO_SOCKET] = "sock",
528 [PTR_TO_SOCKET_OR_NULL] = "sock_or_null",
529 [PTR_TO_SOCK_COMMON] = "sock_common",
530 [PTR_TO_SOCK_COMMON_OR_NULL] = "sock_common_or_null",
531 [PTR_TO_TCP_SOCK] = "tcp_sock",
532 [PTR_TO_TCP_SOCK_OR_NULL] = "tcp_sock_or_null",
533 [PTR_TO_TP_BUFFER] = "tp_buffer",
534 [PTR_TO_XDP_SOCK] = "xdp_sock",
535 [PTR_TO_BTF_ID] = "ptr_",
536 [PTR_TO_BTF_ID_OR_NULL] = "ptr_or_null_",
537 [PTR_TO_PERCPU_BTF_ID] = "percpu_ptr_",
538 [PTR_TO_MEM] = "mem",
539 [PTR_TO_MEM_OR_NULL] = "mem_or_null",
540 [PTR_TO_RDONLY_BUF] = "rdonly_buf",
541 [PTR_TO_RDONLY_BUF_OR_NULL] = "rdonly_buf_or_null",
542 [PTR_TO_RDWR_BUF] = "rdwr_buf",
543 [PTR_TO_RDWR_BUF_OR_NULL] = "rdwr_buf_or_null",
546 static char slot_type_char[] = {
547 [STACK_INVALID] = '?',
553 static void print_liveness(struct bpf_verifier_env *env,
554 enum bpf_reg_liveness live)
556 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
558 if (live & REG_LIVE_READ)
560 if (live & REG_LIVE_WRITTEN)
562 if (live & REG_LIVE_DONE)
566 static struct bpf_func_state *func(struct bpf_verifier_env *env,
567 const struct bpf_reg_state *reg)
569 struct bpf_verifier_state *cur = env->cur_state;
571 return cur->frame[reg->frameno];
574 static const char *kernel_type_name(const struct btf* btf, u32 id)
576 return btf_name_by_offset(btf, btf_type_by_id(btf, id)->name_off);
579 static void print_verifier_state(struct bpf_verifier_env *env,
580 const struct bpf_func_state *state)
582 const struct bpf_reg_state *reg;
587 verbose(env, " frame%d:", state->frameno);
588 for (i = 0; i < MAX_BPF_REG; i++) {
589 reg = &state->regs[i];
593 verbose(env, " R%d", i);
594 print_liveness(env, reg->live);
595 verbose(env, "=%s", reg_type_str[t]);
596 if (t == SCALAR_VALUE && reg->precise)
598 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
599 tnum_is_const(reg->var_off)) {
600 /* reg->off should be 0 for SCALAR_VALUE */
601 verbose(env, "%lld", reg->var_off.value + reg->off);
603 if (t == PTR_TO_BTF_ID ||
604 t == PTR_TO_BTF_ID_OR_NULL ||
605 t == PTR_TO_PERCPU_BTF_ID)
606 verbose(env, "%s", kernel_type_name(reg->btf, reg->btf_id));
607 verbose(env, "(id=%d", reg->id);
608 if (reg_type_may_be_refcounted_or_null(t))
609 verbose(env, ",ref_obj_id=%d", reg->ref_obj_id);
610 if (t != SCALAR_VALUE)
611 verbose(env, ",off=%d", reg->off);
612 if (type_is_pkt_pointer(t))
613 verbose(env, ",r=%d", reg->range);
614 else if (t == CONST_PTR_TO_MAP ||
615 t == PTR_TO_MAP_VALUE ||
616 t == PTR_TO_MAP_VALUE_OR_NULL)
617 verbose(env, ",ks=%d,vs=%d",
618 reg->map_ptr->key_size,
619 reg->map_ptr->value_size);
620 if (tnum_is_const(reg->var_off)) {
621 /* Typically an immediate SCALAR_VALUE, but
622 * could be a pointer whose offset is too big
625 verbose(env, ",imm=%llx", reg->var_off.value);
627 if (reg->smin_value != reg->umin_value &&
628 reg->smin_value != S64_MIN)
629 verbose(env, ",smin_value=%lld",
630 (long long)reg->smin_value);
631 if (reg->smax_value != reg->umax_value &&
632 reg->smax_value != S64_MAX)
633 verbose(env, ",smax_value=%lld",
634 (long long)reg->smax_value);
635 if (reg->umin_value != 0)
636 verbose(env, ",umin_value=%llu",
637 (unsigned long long)reg->umin_value);
638 if (reg->umax_value != U64_MAX)
639 verbose(env, ",umax_value=%llu",
640 (unsigned long long)reg->umax_value);
641 if (!tnum_is_unknown(reg->var_off)) {
644 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
645 verbose(env, ",var_off=%s", tn_buf);
647 if (reg->s32_min_value != reg->smin_value &&
648 reg->s32_min_value != S32_MIN)
649 verbose(env, ",s32_min_value=%d",
650 (int)(reg->s32_min_value));
651 if (reg->s32_max_value != reg->smax_value &&
652 reg->s32_max_value != S32_MAX)
653 verbose(env, ",s32_max_value=%d",
654 (int)(reg->s32_max_value));
655 if (reg->u32_min_value != reg->umin_value &&
656 reg->u32_min_value != U32_MIN)
657 verbose(env, ",u32_min_value=%d",
658 (int)(reg->u32_min_value));
659 if (reg->u32_max_value != reg->umax_value &&
660 reg->u32_max_value != U32_MAX)
661 verbose(env, ",u32_max_value=%d",
662 (int)(reg->u32_max_value));
667 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
668 char types_buf[BPF_REG_SIZE + 1];
672 for (j = 0; j < BPF_REG_SIZE; j++) {
673 if (state->stack[i].slot_type[j] != STACK_INVALID)
675 types_buf[j] = slot_type_char[
676 state->stack[i].slot_type[j]];
678 types_buf[BPF_REG_SIZE] = 0;
681 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
682 print_liveness(env, state->stack[i].spilled_ptr.live);
683 if (state->stack[i].slot_type[0] == STACK_SPILL) {
684 reg = &state->stack[i].spilled_ptr;
686 verbose(env, "=%s", reg_type_str[t]);
687 if (t == SCALAR_VALUE && reg->precise)
689 if (t == SCALAR_VALUE && tnum_is_const(reg->var_off))
690 verbose(env, "%lld", reg->var_off.value + reg->off);
692 verbose(env, "=%s", types_buf);
695 if (state->acquired_refs && state->refs[0].id) {
696 verbose(env, " refs=%d", state->refs[0].id);
697 for (i = 1; i < state->acquired_refs; i++)
698 if (state->refs[i].id)
699 verbose(env, ",%d", state->refs[i].id);
704 #define COPY_STATE_FN(NAME, COUNT, FIELD, SIZE) \
705 static int copy_##NAME##_state(struct bpf_func_state *dst, \
706 const struct bpf_func_state *src) \
710 if (WARN_ON_ONCE(dst->COUNT < src->COUNT)) { \
711 /* internal bug, make state invalid to reject the program */ \
712 memset(dst, 0, sizeof(*dst)); \
715 memcpy(dst->FIELD, src->FIELD, \
716 sizeof(*src->FIELD) * (src->COUNT / SIZE)); \
719 /* copy_reference_state() */
720 COPY_STATE_FN(reference, acquired_refs, refs, 1)
721 /* copy_stack_state() */
722 COPY_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
725 #define REALLOC_STATE_FN(NAME, COUNT, FIELD, SIZE) \
726 static int realloc_##NAME##_state(struct bpf_func_state *state, int size, \
729 u32 old_size = state->COUNT; \
730 struct bpf_##NAME##_state *new_##FIELD; \
731 int slot = size / SIZE; \
733 if (size <= old_size || !size) { \
736 state->COUNT = slot * SIZE; \
737 if (!size && old_size) { \
738 kfree(state->FIELD); \
739 state->FIELD = NULL; \
743 new_##FIELD = kmalloc_array(slot, sizeof(struct bpf_##NAME##_state), \
749 memcpy(new_##FIELD, state->FIELD, \
750 sizeof(*new_##FIELD) * (old_size / SIZE)); \
751 memset(new_##FIELD + old_size / SIZE, 0, \
752 sizeof(*new_##FIELD) * (size - old_size) / SIZE); \
754 state->COUNT = slot * SIZE; \
755 kfree(state->FIELD); \
756 state->FIELD = new_##FIELD; \
759 /* realloc_reference_state() */
760 REALLOC_STATE_FN(reference, acquired_refs, refs, 1)
761 /* realloc_stack_state() */
762 REALLOC_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
763 #undef REALLOC_STATE_FN
765 /* do_check() starts with zero-sized stack in struct bpf_verifier_state to
766 * make it consume minimal amount of memory. check_stack_write() access from
767 * the program calls into realloc_func_state() to grow the stack size.
768 * Note there is a non-zero 'parent' pointer inside bpf_verifier_state
769 * which realloc_stack_state() copies over. It points to previous
770 * bpf_verifier_state which is never reallocated.
772 static int realloc_func_state(struct bpf_func_state *state, int stack_size,
773 int refs_size, bool copy_old)
775 int err = realloc_reference_state(state, refs_size, copy_old);
778 return realloc_stack_state(state, stack_size, copy_old);
781 /* Acquire a pointer id from the env and update the state->refs to include
782 * this new pointer reference.
783 * On success, returns a valid pointer id to associate with the register
784 * On failure, returns a negative errno.
786 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
788 struct bpf_func_state *state = cur_func(env);
789 int new_ofs = state->acquired_refs;
792 err = realloc_reference_state(state, state->acquired_refs + 1, true);
796 state->refs[new_ofs].id = id;
797 state->refs[new_ofs].insn_idx = insn_idx;
802 /* release function corresponding to acquire_reference_state(). Idempotent. */
803 static int release_reference_state(struct bpf_func_state *state, int ptr_id)
807 last_idx = state->acquired_refs - 1;
808 for (i = 0; i < state->acquired_refs; i++) {
809 if (state->refs[i].id == ptr_id) {
810 if (last_idx && i != last_idx)
811 memcpy(&state->refs[i], &state->refs[last_idx],
812 sizeof(*state->refs));
813 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
814 state->acquired_refs--;
821 static int transfer_reference_state(struct bpf_func_state *dst,
822 struct bpf_func_state *src)
824 int err = realloc_reference_state(dst, src->acquired_refs, false);
827 err = copy_reference_state(dst, src);
833 static void free_func_state(struct bpf_func_state *state)
842 static void clear_jmp_history(struct bpf_verifier_state *state)
844 kfree(state->jmp_history);
845 state->jmp_history = NULL;
846 state->jmp_history_cnt = 0;
849 static void free_verifier_state(struct bpf_verifier_state *state,
854 for (i = 0; i <= state->curframe; i++) {
855 free_func_state(state->frame[i]);
856 state->frame[i] = NULL;
858 clear_jmp_history(state);
863 /* copy verifier state from src to dst growing dst stack space
864 * when necessary to accommodate larger src stack
866 static int copy_func_state(struct bpf_func_state *dst,
867 const struct bpf_func_state *src)
871 err = realloc_func_state(dst, src->allocated_stack, src->acquired_refs,
875 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
876 err = copy_reference_state(dst, src);
879 return copy_stack_state(dst, src);
882 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
883 const struct bpf_verifier_state *src)
885 struct bpf_func_state *dst;
886 u32 jmp_sz = sizeof(struct bpf_idx_pair) * src->jmp_history_cnt;
889 if (dst_state->jmp_history_cnt < src->jmp_history_cnt) {
890 kfree(dst_state->jmp_history);
891 dst_state->jmp_history = kmalloc(jmp_sz, GFP_USER);
892 if (!dst_state->jmp_history)
895 memcpy(dst_state->jmp_history, src->jmp_history, jmp_sz);
896 dst_state->jmp_history_cnt = src->jmp_history_cnt;
898 /* if dst has more stack frames then src frame, free them */
899 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
900 free_func_state(dst_state->frame[i]);
901 dst_state->frame[i] = NULL;
903 dst_state->speculative = src->speculative;
904 dst_state->curframe = src->curframe;
905 dst_state->active_spin_lock = src->active_spin_lock;
906 dst_state->branches = src->branches;
907 dst_state->parent = src->parent;
908 dst_state->first_insn_idx = src->first_insn_idx;
909 dst_state->last_insn_idx = src->last_insn_idx;
910 for (i = 0; i <= src->curframe; i++) {
911 dst = dst_state->frame[i];
913 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
916 dst_state->frame[i] = dst;
918 err = copy_func_state(dst, src->frame[i]);
925 static void update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
928 u32 br = --st->branches;
930 /* WARN_ON(br > 1) technically makes sense here,
931 * but see comment in push_stack(), hence:
933 WARN_ONCE((int)br < 0,
934 "BUG update_branch_counts:branches_to_explore=%d\n",
942 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
943 int *insn_idx, bool pop_log)
945 struct bpf_verifier_state *cur = env->cur_state;
946 struct bpf_verifier_stack_elem *elem, *head = env->head;
949 if (env->head == NULL)
953 err = copy_verifier_state(cur, &head->st);
958 bpf_vlog_reset(&env->log, head->log_pos);
960 *insn_idx = head->insn_idx;
962 *prev_insn_idx = head->prev_insn_idx;
964 free_verifier_state(&head->st, false);
971 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
972 int insn_idx, int prev_insn_idx,
975 struct bpf_verifier_state *cur = env->cur_state;
976 struct bpf_verifier_stack_elem *elem;
979 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
983 elem->insn_idx = insn_idx;
984 elem->prev_insn_idx = prev_insn_idx;
985 elem->next = env->head;
986 elem->log_pos = env->log.len_used;
989 err = copy_verifier_state(&elem->st, cur);
992 elem->st.speculative |= speculative;
993 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
994 verbose(env, "The sequence of %d jumps is too complex.\n",
998 if (elem->st.parent) {
999 ++elem->st.parent->branches;
1000 /* WARN_ON(branches > 2) technically makes sense here,
1002 * 1. speculative states will bump 'branches' for non-branch
1004 * 2. is_state_visited() heuristics may decide not to create
1005 * a new state for a sequence of branches and all such current
1006 * and cloned states will be pointing to a single parent state
1007 * which might have large 'branches' count.
1012 free_verifier_state(env->cur_state, true);
1013 env->cur_state = NULL;
1014 /* pop all elements and return */
1015 while (!pop_stack(env, NULL, NULL, false));
1019 #define CALLER_SAVED_REGS 6
1020 static const int caller_saved[CALLER_SAVED_REGS] = {
1021 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
1024 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1025 struct bpf_reg_state *reg);
1027 /* This helper doesn't clear reg->id */
1028 static void ___mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1030 reg->var_off = tnum_const(imm);
1031 reg->smin_value = (s64)imm;
1032 reg->smax_value = (s64)imm;
1033 reg->umin_value = imm;
1034 reg->umax_value = imm;
1036 reg->s32_min_value = (s32)imm;
1037 reg->s32_max_value = (s32)imm;
1038 reg->u32_min_value = (u32)imm;
1039 reg->u32_max_value = (u32)imm;
1042 /* Mark the unknown part of a register (variable offset or scalar value) as
1043 * known to have the value @imm.
1045 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1047 /* Clear id, off, and union(map_ptr, range) */
1048 memset(((u8 *)reg) + sizeof(reg->type), 0,
1049 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
1050 ___mark_reg_known(reg, imm);
1053 static void __mark_reg32_known(struct bpf_reg_state *reg, u64 imm)
1055 reg->var_off = tnum_const_subreg(reg->var_off, imm);
1056 reg->s32_min_value = (s32)imm;
1057 reg->s32_max_value = (s32)imm;
1058 reg->u32_min_value = (u32)imm;
1059 reg->u32_max_value = (u32)imm;
1062 /* Mark the 'variable offset' part of a register as zero. This should be
1063 * used only on registers holding a pointer type.
1065 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
1067 __mark_reg_known(reg, 0);
1070 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
1072 __mark_reg_known(reg, 0);
1073 reg->type = SCALAR_VALUE;
1076 static void mark_reg_known_zero(struct bpf_verifier_env *env,
1077 struct bpf_reg_state *regs, u32 regno)
1079 if (WARN_ON(regno >= MAX_BPF_REG)) {
1080 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
1081 /* Something bad happened, let's kill all regs */
1082 for (regno = 0; regno < MAX_BPF_REG; regno++)
1083 __mark_reg_not_init(env, regs + regno);
1086 __mark_reg_known_zero(regs + regno);
1089 static void mark_ptr_not_null_reg(struct bpf_reg_state *reg)
1091 switch (reg->type) {
1092 case PTR_TO_MAP_VALUE_OR_NULL: {
1093 const struct bpf_map *map = reg->map_ptr;
1095 if (map->inner_map_meta) {
1096 reg->type = CONST_PTR_TO_MAP;
1097 reg->map_ptr = map->inner_map_meta;
1098 } else if (map->map_type == BPF_MAP_TYPE_XSKMAP) {
1099 reg->type = PTR_TO_XDP_SOCK;
1100 } else if (map->map_type == BPF_MAP_TYPE_SOCKMAP ||
1101 map->map_type == BPF_MAP_TYPE_SOCKHASH) {
1102 reg->type = PTR_TO_SOCKET;
1104 reg->type = PTR_TO_MAP_VALUE;
1108 case PTR_TO_SOCKET_OR_NULL:
1109 reg->type = PTR_TO_SOCKET;
1111 case PTR_TO_SOCK_COMMON_OR_NULL:
1112 reg->type = PTR_TO_SOCK_COMMON;
1114 case PTR_TO_TCP_SOCK_OR_NULL:
1115 reg->type = PTR_TO_TCP_SOCK;
1117 case PTR_TO_BTF_ID_OR_NULL:
1118 reg->type = PTR_TO_BTF_ID;
1120 case PTR_TO_MEM_OR_NULL:
1121 reg->type = PTR_TO_MEM;
1123 case PTR_TO_RDONLY_BUF_OR_NULL:
1124 reg->type = PTR_TO_RDONLY_BUF;
1126 case PTR_TO_RDWR_BUF_OR_NULL:
1127 reg->type = PTR_TO_RDWR_BUF;
1130 WARN_ONCE(1, "unknown nullable register type");
1134 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
1136 return type_is_pkt_pointer(reg->type);
1139 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
1141 return reg_is_pkt_pointer(reg) ||
1142 reg->type == PTR_TO_PACKET_END;
1145 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
1146 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
1147 enum bpf_reg_type which)
1149 /* The register can already have a range from prior markings.
1150 * This is fine as long as it hasn't been advanced from its
1153 return reg->type == which &&
1156 tnum_equals_const(reg->var_off, 0);
1159 /* Reset the min/max bounds of a register */
1160 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
1162 reg->smin_value = S64_MIN;
1163 reg->smax_value = S64_MAX;
1164 reg->umin_value = 0;
1165 reg->umax_value = U64_MAX;
1167 reg->s32_min_value = S32_MIN;
1168 reg->s32_max_value = S32_MAX;
1169 reg->u32_min_value = 0;
1170 reg->u32_max_value = U32_MAX;
1173 static void __mark_reg64_unbounded(struct bpf_reg_state *reg)
1175 reg->smin_value = S64_MIN;
1176 reg->smax_value = S64_MAX;
1177 reg->umin_value = 0;
1178 reg->umax_value = U64_MAX;
1181 static void __mark_reg32_unbounded(struct bpf_reg_state *reg)
1183 reg->s32_min_value = S32_MIN;
1184 reg->s32_max_value = S32_MAX;
1185 reg->u32_min_value = 0;
1186 reg->u32_max_value = U32_MAX;
1189 static void __update_reg32_bounds(struct bpf_reg_state *reg)
1191 struct tnum var32_off = tnum_subreg(reg->var_off);
1193 /* min signed is max(sign bit) | min(other bits) */
1194 reg->s32_min_value = max_t(s32, reg->s32_min_value,
1195 var32_off.value | (var32_off.mask & S32_MIN));
1196 /* max signed is min(sign bit) | max(other bits) */
1197 reg->s32_max_value = min_t(s32, reg->s32_max_value,
1198 var32_off.value | (var32_off.mask & S32_MAX));
1199 reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)var32_off.value);
1200 reg->u32_max_value = min(reg->u32_max_value,
1201 (u32)(var32_off.value | var32_off.mask));
1204 static void __update_reg64_bounds(struct bpf_reg_state *reg)
1206 /* min signed is max(sign bit) | min(other bits) */
1207 reg->smin_value = max_t(s64, reg->smin_value,
1208 reg->var_off.value | (reg->var_off.mask & S64_MIN));
1209 /* max signed is min(sign bit) | max(other bits) */
1210 reg->smax_value = min_t(s64, reg->smax_value,
1211 reg->var_off.value | (reg->var_off.mask & S64_MAX));
1212 reg->umin_value = max(reg->umin_value, reg->var_off.value);
1213 reg->umax_value = min(reg->umax_value,
1214 reg->var_off.value | reg->var_off.mask);
1217 static void __update_reg_bounds(struct bpf_reg_state *reg)
1219 __update_reg32_bounds(reg);
1220 __update_reg64_bounds(reg);
1223 /* Uses signed min/max values to inform unsigned, and vice-versa */
1224 static void __reg32_deduce_bounds(struct bpf_reg_state *reg)
1226 /* Learn sign from signed bounds.
1227 * If we cannot cross the sign boundary, then signed and unsigned bounds
1228 * are the same, so combine. This works even in the negative case, e.g.
1229 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1231 if (reg->s32_min_value >= 0 || reg->s32_max_value < 0) {
1232 reg->s32_min_value = reg->u32_min_value =
1233 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1234 reg->s32_max_value = reg->u32_max_value =
1235 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1238 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1239 * boundary, so we must be careful.
1241 if ((s32)reg->u32_max_value >= 0) {
1242 /* Positive. We can't learn anything from the smin, but smax
1243 * is positive, hence safe.
1245 reg->s32_min_value = reg->u32_min_value;
1246 reg->s32_max_value = reg->u32_max_value =
1247 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1248 } else if ((s32)reg->u32_min_value < 0) {
1249 /* Negative. We can't learn anything from the smax, but smin
1250 * is negative, hence safe.
1252 reg->s32_min_value = reg->u32_min_value =
1253 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1254 reg->s32_max_value = reg->u32_max_value;
1258 static void __reg64_deduce_bounds(struct bpf_reg_state *reg)
1260 /* Learn sign from signed bounds.
1261 * If we cannot cross the sign boundary, then signed and unsigned bounds
1262 * are the same, so combine. This works even in the negative case, e.g.
1263 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1265 if (reg->smin_value >= 0 || reg->smax_value < 0) {
1266 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1268 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1272 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1273 * boundary, so we must be careful.
1275 if ((s64)reg->umax_value >= 0) {
1276 /* Positive. We can't learn anything from the smin, but smax
1277 * is positive, hence safe.
1279 reg->smin_value = reg->umin_value;
1280 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1282 } else if ((s64)reg->umin_value < 0) {
1283 /* Negative. We can't learn anything from the smax, but smin
1284 * is negative, hence safe.
1286 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1288 reg->smax_value = reg->umax_value;
1292 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
1294 __reg32_deduce_bounds(reg);
1295 __reg64_deduce_bounds(reg);
1298 /* Attempts to improve var_off based on unsigned min/max information */
1299 static void __reg_bound_offset(struct bpf_reg_state *reg)
1301 struct tnum var64_off = tnum_intersect(reg->var_off,
1302 tnum_range(reg->umin_value,
1304 struct tnum var32_off = tnum_intersect(tnum_subreg(reg->var_off),
1305 tnum_range(reg->u32_min_value,
1306 reg->u32_max_value));
1308 reg->var_off = tnum_or(tnum_clear_subreg(var64_off), var32_off);
1311 static void __reg_assign_32_into_64(struct bpf_reg_state *reg)
1313 reg->umin_value = reg->u32_min_value;
1314 reg->umax_value = reg->u32_max_value;
1315 /* Attempt to pull 32-bit signed bounds into 64-bit bounds
1316 * but must be positive otherwise set to worse case bounds
1317 * and refine later from tnum.
1319 if (reg->s32_min_value >= 0 && reg->s32_max_value >= 0)
1320 reg->smax_value = reg->s32_max_value;
1322 reg->smax_value = U32_MAX;
1323 if (reg->s32_min_value >= 0)
1324 reg->smin_value = reg->s32_min_value;
1326 reg->smin_value = 0;
1329 static void __reg_combine_32_into_64(struct bpf_reg_state *reg)
1331 /* special case when 64-bit register has upper 32-bit register
1332 * zeroed. Typically happens after zext or <<32, >>32 sequence
1333 * allowing us to use 32-bit bounds directly,
1335 if (tnum_equals_const(tnum_clear_subreg(reg->var_off), 0)) {
1336 __reg_assign_32_into_64(reg);
1338 /* Otherwise the best we can do is push lower 32bit known and
1339 * unknown bits into register (var_off set from jmp logic)
1340 * then learn as much as possible from the 64-bit tnum
1341 * known and unknown bits. The previous smin/smax bounds are
1342 * invalid here because of jmp32 compare so mark them unknown
1343 * so they do not impact tnum bounds calculation.
1345 __mark_reg64_unbounded(reg);
1346 __update_reg_bounds(reg);
1349 /* Intersecting with the old var_off might have improved our bounds
1350 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1351 * then new var_off is (0; 0x7f...fc) which improves our umax.
1353 __reg_deduce_bounds(reg);
1354 __reg_bound_offset(reg);
1355 __update_reg_bounds(reg);
1358 static bool __reg64_bound_s32(s64 a)
1360 return a > S32_MIN && a < S32_MAX;
1363 static bool __reg64_bound_u32(u64 a)
1365 if (a > U32_MIN && a < U32_MAX)
1370 static void __reg_combine_64_into_32(struct bpf_reg_state *reg)
1372 __mark_reg32_unbounded(reg);
1374 if (__reg64_bound_s32(reg->smin_value) && __reg64_bound_s32(reg->smax_value)) {
1375 reg->s32_min_value = (s32)reg->smin_value;
1376 reg->s32_max_value = (s32)reg->smax_value;
1378 if (__reg64_bound_u32(reg->umin_value))
1379 reg->u32_min_value = (u32)reg->umin_value;
1380 if (__reg64_bound_u32(reg->umax_value))
1381 reg->u32_max_value = (u32)reg->umax_value;
1383 /* Intersecting with the old var_off might have improved our bounds
1384 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1385 * then new var_off is (0; 0x7f...fc) which improves our umax.
1387 __reg_deduce_bounds(reg);
1388 __reg_bound_offset(reg);
1389 __update_reg_bounds(reg);
1392 /* Mark a register as having a completely unknown (scalar) value. */
1393 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
1394 struct bpf_reg_state *reg)
1397 * Clear type, id, off, and union(map_ptr, range) and
1398 * padding between 'type' and union
1400 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
1401 reg->type = SCALAR_VALUE;
1402 reg->var_off = tnum_unknown;
1404 reg->precise = env->subprog_cnt > 1 || !env->bpf_capable;
1405 __mark_reg_unbounded(reg);
1408 static void mark_reg_unknown(struct bpf_verifier_env *env,
1409 struct bpf_reg_state *regs, u32 regno)
1411 if (WARN_ON(regno >= MAX_BPF_REG)) {
1412 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
1413 /* Something bad happened, let's kill all regs except FP */
1414 for (regno = 0; regno < BPF_REG_FP; regno++)
1415 __mark_reg_not_init(env, regs + regno);
1418 __mark_reg_unknown(env, regs + regno);
1421 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1422 struct bpf_reg_state *reg)
1424 __mark_reg_unknown(env, reg);
1425 reg->type = NOT_INIT;
1428 static void mark_reg_not_init(struct bpf_verifier_env *env,
1429 struct bpf_reg_state *regs, u32 regno)
1431 if (WARN_ON(regno >= MAX_BPF_REG)) {
1432 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
1433 /* Something bad happened, let's kill all regs except FP */
1434 for (regno = 0; regno < BPF_REG_FP; regno++)
1435 __mark_reg_not_init(env, regs + regno);
1438 __mark_reg_not_init(env, regs + regno);
1441 static void mark_btf_ld_reg(struct bpf_verifier_env *env,
1442 struct bpf_reg_state *regs, u32 regno,
1443 enum bpf_reg_type reg_type,
1444 struct btf *btf, u32 btf_id)
1446 if (reg_type == SCALAR_VALUE) {
1447 mark_reg_unknown(env, regs, regno);
1450 mark_reg_known_zero(env, regs, regno);
1451 regs[regno].type = PTR_TO_BTF_ID;
1452 regs[regno].btf = btf;
1453 regs[regno].btf_id = btf_id;
1456 #define DEF_NOT_SUBREG (0)
1457 static void init_reg_state(struct bpf_verifier_env *env,
1458 struct bpf_func_state *state)
1460 struct bpf_reg_state *regs = state->regs;
1463 for (i = 0; i < MAX_BPF_REG; i++) {
1464 mark_reg_not_init(env, regs, i);
1465 regs[i].live = REG_LIVE_NONE;
1466 regs[i].parent = NULL;
1467 regs[i].subreg_def = DEF_NOT_SUBREG;
1471 regs[BPF_REG_FP].type = PTR_TO_STACK;
1472 mark_reg_known_zero(env, regs, BPF_REG_FP);
1473 regs[BPF_REG_FP].frameno = state->frameno;
1476 #define BPF_MAIN_FUNC (-1)
1477 static void init_func_state(struct bpf_verifier_env *env,
1478 struct bpf_func_state *state,
1479 int callsite, int frameno, int subprogno)
1481 state->callsite = callsite;
1482 state->frameno = frameno;
1483 state->subprogno = subprogno;
1484 init_reg_state(env, state);
1488 SRC_OP, /* register is used as source operand */
1489 DST_OP, /* register is used as destination operand */
1490 DST_OP_NO_MARK /* same as above, check only, don't mark */
1493 static int cmp_subprogs(const void *a, const void *b)
1495 return ((struct bpf_subprog_info *)a)->start -
1496 ((struct bpf_subprog_info *)b)->start;
1499 static int find_subprog(struct bpf_verifier_env *env, int off)
1501 struct bpf_subprog_info *p;
1503 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
1504 sizeof(env->subprog_info[0]), cmp_subprogs);
1507 return p - env->subprog_info;
1511 static int add_subprog(struct bpf_verifier_env *env, int off)
1513 int insn_cnt = env->prog->len;
1516 if (off >= insn_cnt || off < 0) {
1517 verbose(env, "call to invalid destination\n");
1520 ret = find_subprog(env, off);
1523 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
1524 verbose(env, "too many subprograms\n");
1527 env->subprog_info[env->subprog_cnt++].start = off;
1528 sort(env->subprog_info, env->subprog_cnt,
1529 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
1533 static int check_subprogs(struct bpf_verifier_env *env)
1535 int i, ret, subprog_start, subprog_end, off, cur_subprog = 0;
1536 struct bpf_subprog_info *subprog = env->subprog_info;
1537 struct bpf_insn *insn = env->prog->insnsi;
1538 int insn_cnt = env->prog->len;
1540 /* Add entry function. */
1541 ret = add_subprog(env, 0);
1545 /* determine subprog starts. The end is one before the next starts */
1546 for (i = 0; i < insn_cnt; i++) {
1547 if (!bpf_pseudo_call(insn + i))
1549 if (!env->bpf_capable) {
1551 "function calls to other bpf functions are allowed for CAP_BPF and CAP_SYS_ADMIN\n");
1554 ret = add_subprog(env, i + insn[i].imm + 1);
1559 /* Add a fake 'exit' subprog which could simplify subprog iteration
1560 * logic. 'subprog_cnt' should not be increased.
1562 subprog[env->subprog_cnt].start = insn_cnt;
1564 if (env->log.level & BPF_LOG_LEVEL2)
1565 for (i = 0; i < env->subprog_cnt; i++)
1566 verbose(env, "func#%d @%d\n", i, subprog[i].start);
1568 /* now check that all jumps are within the same subprog */
1569 subprog_start = subprog[cur_subprog].start;
1570 subprog_end = subprog[cur_subprog + 1].start;
1571 for (i = 0; i < insn_cnt; i++) {
1572 u8 code = insn[i].code;
1574 if (code == (BPF_JMP | BPF_CALL) &&
1575 insn[i].imm == BPF_FUNC_tail_call &&
1576 insn[i].src_reg != BPF_PSEUDO_CALL)
1577 subprog[cur_subprog].has_tail_call = true;
1578 if (BPF_CLASS(code) == BPF_LD &&
1579 (BPF_MODE(code) == BPF_ABS || BPF_MODE(code) == BPF_IND))
1580 subprog[cur_subprog].has_ld_abs = true;
1581 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
1583 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
1585 off = i + insn[i].off + 1;
1586 if (off < subprog_start || off >= subprog_end) {
1587 verbose(env, "jump out of range from insn %d to %d\n", i, off);
1591 if (i == subprog_end - 1) {
1592 /* to avoid fall-through from one subprog into another
1593 * the last insn of the subprog should be either exit
1594 * or unconditional jump back
1596 if (code != (BPF_JMP | BPF_EXIT) &&
1597 code != (BPF_JMP | BPF_JA)) {
1598 verbose(env, "last insn is not an exit or jmp\n");
1601 subprog_start = subprog_end;
1603 if (cur_subprog < env->subprog_cnt)
1604 subprog_end = subprog[cur_subprog + 1].start;
1610 /* Parentage chain of this register (or stack slot) should take care of all
1611 * issues like callee-saved registers, stack slot allocation time, etc.
1613 static int mark_reg_read(struct bpf_verifier_env *env,
1614 const struct bpf_reg_state *state,
1615 struct bpf_reg_state *parent, u8 flag)
1617 bool writes = parent == state->parent; /* Observe write marks */
1621 /* if read wasn't screened by an earlier write ... */
1622 if (writes && state->live & REG_LIVE_WRITTEN)
1624 if (parent->live & REG_LIVE_DONE) {
1625 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
1626 reg_type_str[parent->type],
1627 parent->var_off.value, parent->off);
1630 /* The first condition is more likely to be true than the
1631 * second, checked it first.
1633 if ((parent->live & REG_LIVE_READ) == flag ||
1634 parent->live & REG_LIVE_READ64)
1635 /* The parentage chain never changes and
1636 * this parent was already marked as LIVE_READ.
1637 * There is no need to keep walking the chain again and
1638 * keep re-marking all parents as LIVE_READ.
1639 * This case happens when the same register is read
1640 * multiple times without writes into it in-between.
1641 * Also, if parent has the stronger REG_LIVE_READ64 set,
1642 * then no need to set the weak REG_LIVE_READ32.
1645 /* ... then we depend on parent's value */
1646 parent->live |= flag;
1647 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
1648 if (flag == REG_LIVE_READ64)
1649 parent->live &= ~REG_LIVE_READ32;
1651 parent = state->parent;
1656 if (env->longest_mark_read_walk < cnt)
1657 env->longest_mark_read_walk = cnt;
1661 /* This function is supposed to be used by the following 32-bit optimization
1662 * code only. It returns TRUE if the source or destination register operates
1663 * on 64-bit, otherwise return FALSE.
1665 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
1666 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
1671 class = BPF_CLASS(code);
1673 if (class == BPF_JMP) {
1674 /* BPF_EXIT for "main" will reach here. Return TRUE
1679 if (op == BPF_CALL) {
1680 /* BPF to BPF call will reach here because of marking
1681 * caller saved clobber with DST_OP_NO_MARK for which we
1682 * don't care the register def because they are anyway
1683 * marked as NOT_INIT already.
1685 if (insn->src_reg == BPF_PSEUDO_CALL)
1687 /* Helper call will reach here because of arg type
1688 * check, conservatively return TRUE.
1697 if (class == BPF_ALU64 || class == BPF_JMP ||
1698 /* BPF_END always use BPF_ALU class. */
1699 (class == BPF_ALU && op == BPF_END && insn->imm == 64))
1702 if (class == BPF_ALU || class == BPF_JMP32)
1705 if (class == BPF_LDX) {
1707 return BPF_SIZE(code) == BPF_DW;
1708 /* LDX source must be ptr. */
1712 if (class == BPF_STX) {
1713 /* BPF_STX (including atomic variants) has multiple source
1714 * operands, one of which is a ptr. Check whether the caller is
1717 if (t == SRC_OP && reg->type != SCALAR_VALUE)
1719 return BPF_SIZE(code) == BPF_DW;
1722 if (class == BPF_LD) {
1723 u8 mode = BPF_MODE(code);
1726 if (mode == BPF_IMM)
1729 /* Both LD_IND and LD_ABS return 32-bit data. */
1733 /* Implicit ctx ptr. */
1734 if (regno == BPF_REG_6)
1737 /* Explicit source could be any width. */
1741 if (class == BPF_ST)
1742 /* The only source register for BPF_ST is a ptr. */
1745 /* Conservatively return true at default. */
1749 /* Return the regno defined by the insn, or -1. */
1750 static int insn_def_regno(const struct bpf_insn *insn)
1752 switch (BPF_CLASS(insn->code)) {
1758 if (BPF_MODE(insn->code) == BPF_ATOMIC &&
1759 (insn->imm & BPF_FETCH)) {
1760 if (insn->imm == BPF_CMPXCHG)
1763 return insn->src_reg;
1768 return insn->dst_reg;
1772 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
1773 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
1775 int dst_reg = insn_def_regno(insn);
1780 return !is_reg64(env, insn, dst_reg, NULL, DST_OP);
1783 static void mark_insn_zext(struct bpf_verifier_env *env,
1784 struct bpf_reg_state *reg)
1786 s32 def_idx = reg->subreg_def;
1788 if (def_idx == DEF_NOT_SUBREG)
1791 env->insn_aux_data[def_idx - 1].zext_dst = true;
1792 /* The dst will be zero extended, so won't be sub-register anymore. */
1793 reg->subreg_def = DEF_NOT_SUBREG;
1796 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
1797 enum reg_arg_type t)
1799 struct bpf_verifier_state *vstate = env->cur_state;
1800 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1801 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
1802 struct bpf_reg_state *reg, *regs = state->regs;
1805 if (regno >= MAX_BPF_REG) {
1806 verbose(env, "R%d is invalid\n", regno);
1811 rw64 = is_reg64(env, insn, regno, reg, t);
1813 /* check whether register used as source operand can be read */
1814 if (reg->type == NOT_INIT) {
1815 verbose(env, "R%d !read_ok\n", regno);
1818 /* We don't need to worry about FP liveness because it's read-only */
1819 if (regno == BPF_REG_FP)
1823 mark_insn_zext(env, reg);
1825 return mark_reg_read(env, reg, reg->parent,
1826 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
1828 /* check whether register used as dest operand can be written to */
1829 if (regno == BPF_REG_FP) {
1830 verbose(env, "frame pointer is read only\n");
1833 reg->live |= REG_LIVE_WRITTEN;
1834 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
1836 mark_reg_unknown(env, regs, regno);
1841 /* for any branch, call, exit record the history of jmps in the given state */
1842 static int push_jmp_history(struct bpf_verifier_env *env,
1843 struct bpf_verifier_state *cur)
1845 u32 cnt = cur->jmp_history_cnt;
1846 struct bpf_idx_pair *p;
1849 p = krealloc(cur->jmp_history, cnt * sizeof(*p), GFP_USER);
1852 p[cnt - 1].idx = env->insn_idx;
1853 p[cnt - 1].prev_idx = env->prev_insn_idx;
1854 cur->jmp_history = p;
1855 cur->jmp_history_cnt = cnt;
1859 /* Backtrack one insn at a time. If idx is not at the top of recorded
1860 * history then previous instruction came from straight line execution.
1862 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
1867 if (cnt && st->jmp_history[cnt - 1].idx == i) {
1868 i = st->jmp_history[cnt - 1].prev_idx;
1876 /* For given verifier state backtrack_insn() is called from the last insn to
1877 * the first insn. Its purpose is to compute a bitmask of registers and
1878 * stack slots that needs precision in the parent verifier state.
1880 static int backtrack_insn(struct bpf_verifier_env *env, int idx,
1881 u32 *reg_mask, u64 *stack_mask)
1883 const struct bpf_insn_cbs cbs = {
1884 .cb_print = verbose,
1885 .private_data = env,
1887 struct bpf_insn *insn = env->prog->insnsi + idx;
1888 u8 class = BPF_CLASS(insn->code);
1889 u8 opcode = BPF_OP(insn->code);
1890 u8 mode = BPF_MODE(insn->code);
1891 u32 dreg = 1u << insn->dst_reg;
1892 u32 sreg = 1u << insn->src_reg;
1895 if (insn->code == 0)
1897 if (env->log.level & BPF_LOG_LEVEL) {
1898 verbose(env, "regs=%x stack=%llx before ", *reg_mask, *stack_mask);
1899 verbose(env, "%d: ", idx);
1900 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
1903 if (class == BPF_ALU || class == BPF_ALU64) {
1904 if (!(*reg_mask & dreg))
1906 if (opcode == BPF_MOV) {
1907 if (BPF_SRC(insn->code) == BPF_X) {
1909 * dreg needs precision after this insn
1910 * sreg needs precision before this insn
1916 * dreg needs precision after this insn.
1917 * Corresponding register is already marked
1918 * as precise=true in this verifier state.
1919 * No further markings in parent are necessary
1924 if (BPF_SRC(insn->code) == BPF_X) {
1926 * both dreg and sreg need precision
1931 * dreg still needs precision before this insn
1934 } else if (class == BPF_LDX) {
1935 if (!(*reg_mask & dreg))
1939 /* scalars can only be spilled into stack w/o losing precision.
1940 * Load from any other memory can be zero extended.
1941 * The desire to keep that precision is already indicated
1942 * by 'precise' mark in corresponding register of this state.
1943 * No further tracking necessary.
1945 if (insn->src_reg != BPF_REG_FP)
1947 if (BPF_SIZE(insn->code) != BPF_DW)
1950 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
1951 * that [fp - off] slot contains scalar that needs to be
1952 * tracked with precision
1954 spi = (-insn->off - 1) / BPF_REG_SIZE;
1956 verbose(env, "BUG spi %d\n", spi);
1957 WARN_ONCE(1, "verifier backtracking bug");
1960 *stack_mask |= 1ull << spi;
1961 } else if (class == BPF_STX || class == BPF_ST) {
1962 if (*reg_mask & dreg)
1963 /* stx & st shouldn't be using _scalar_ dst_reg
1964 * to access memory. It means backtracking
1965 * encountered a case of pointer subtraction.
1968 /* scalars can only be spilled into stack */
1969 if (insn->dst_reg != BPF_REG_FP)
1971 if (BPF_SIZE(insn->code) != BPF_DW)
1973 spi = (-insn->off - 1) / BPF_REG_SIZE;
1975 verbose(env, "BUG spi %d\n", spi);
1976 WARN_ONCE(1, "verifier backtracking bug");
1979 if (!(*stack_mask & (1ull << spi)))
1981 *stack_mask &= ~(1ull << spi);
1982 if (class == BPF_STX)
1984 } else if (class == BPF_JMP || class == BPF_JMP32) {
1985 if (opcode == BPF_CALL) {
1986 if (insn->src_reg == BPF_PSEUDO_CALL)
1988 /* regular helper call sets R0 */
1990 if (*reg_mask & 0x3f) {
1991 /* if backtracing was looking for registers R1-R5
1992 * they should have been found already.
1994 verbose(env, "BUG regs %x\n", *reg_mask);
1995 WARN_ONCE(1, "verifier backtracking bug");
1998 } else if (opcode == BPF_EXIT) {
2001 } else if (class == BPF_LD) {
2002 if (!(*reg_mask & dreg))
2005 /* It's ld_imm64 or ld_abs or ld_ind.
2006 * For ld_imm64 no further tracking of precision
2007 * into parent is necessary
2009 if (mode == BPF_IND || mode == BPF_ABS)
2010 /* to be analyzed */
2016 /* the scalar precision tracking algorithm:
2017 * . at the start all registers have precise=false.
2018 * . scalar ranges are tracked as normal through alu and jmp insns.
2019 * . once precise value of the scalar register is used in:
2020 * . ptr + scalar alu
2021 * . if (scalar cond K|scalar)
2022 * . helper_call(.., scalar, ...) where ARG_CONST is expected
2023 * backtrack through the verifier states and mark all registers and
2024 * stack slots with spilled constants that these scalar regisers
2025 * should be precise.
2026 * . during state pruning two registers (or spilled stack slots)
2027 * are equivalent if both are not precise.
2029 * Note the verifier cannot simply walk register parentage chain,
2030 * since many different registers and stack slots could have been
2031 * used to compute single precise scalar.
2033 * The approach of starting with precise=true for all registers and then
2034 * backtrack to mark a register as not precise when the verifier detects
2035 * that program doesn't care about specific value (e.g., when helper
2036 * takes register as ARG_ANYTHING parameter) is not safe.
2038 * It's ok to walk single parentage chain of the verifier states.
2039 * It's possible that this backtracking will go all the way till 1st insn.
2040 * All other branches will be explored for needing precision later.
2042 * The backtracking needs to deal with cases like:
2043 * 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)
2046 * if r5 > 0x79f goto pc+7
2047 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
2050 * call bpf_perf_event_output#25
2051 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
2055 * call foo // uses callee's r6 inside to compute r0
2059 * to track above reg_mask/stack_mask needs to be independent for each frame.
2061 * Also if parent's curframe > frame where backtracking started,
2062 * the verifier need to mark registers in both frames, otherwise callees
2063 * may incorrectly prune callers. This is similar to
2064 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
2066 * For now backtracking falls back into conservative marking.
2068 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
2069 struct bpf_verifier_state *st)
2071 struct bpf_func_state *func;
2072 struct bpf_reg_state *reg;
2075 /* big hammer: mark all scalars precise in this path.
2076 * pop_stack may still get !precise scalars.
2078 for (; st; st = st->parent)
2079 for (i = 0; i <= st->curframe; i++) {
2080 func = st->frame[i];
2081 for (j = 0; j < BPF_REG_FP; j++) {
2082 reg = &func->regs[j];
2083 if (reg->type != SCALAR_VALUE)
2085 reg->precise = true;
2087 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
2088 if (func->stack[j].slot_type[0] != STACK_SPILL)
2090 reg = &func->stack[j].spilled_ptr;
2091 if (reg->type != SCALAR_VALUE)
2093 reg->precise = true;
2098 static int __mark_chain_precision(struct bpf_verifier_env *env, int regno,
2101 struct bpf_verifier_state *st = env->cur_state;
2102 int first_idx = st->first_insn_idx;
2103 int last_idx = env->insn_idx;
2104 struct bpf_func_state *func;
2105 struct bpf_reg_state *reg;
2106 u32 reg_mask = regno >= 0 ? 1u << regno : 0;
2107 u64 stack_mask = spi >= 0 ? 1ull << spi : 0;
2108 bool skip_first = true;
2109 bool new_marks = false;
2112 if (!env->bpf_capable)
2115 func = st->frame[st->curframe];
2117 reg = &func->regs[regno];
2118 if (reg->type != SCALAR_VALUE) {
2119 WARN_ONCE(1, "backtracing misuse");
2126 reg->precise = true;
2130 if (func->stack[spi].slot_type[0] != STACK_SPILL) {
2134 reg = &func->stack[spi].spilled_ptr;
2135 if (reg->type != SCALAR_VALUE) {
2143 reg->precise = true;
2149 if (!reg_mask && !stack_mask)
2152 DECLARE_BITMAP(mask, 64);
2153 u32 history = st->jmp_history_cnt;
2155 if (env->log.level & BPF_LOG_LEVEL)
2156 verbose(env, "last_idx %d first_idx %d\n", last_idx, first_idx);
2157 for (i = last_idx;;) {
2162 err = backtrack_insn(env, i, ®_mask, &stack_mask);
2164 if (err == -ENOTSUPP) {
2165 mark_all_scalars_precise(env, st);
2170 if (!reg_mask && !stack_mask)
2171 /* Found assignment(s) into tracked register in this state.
2172 * Since this state is already marked, just return.
2173 * Nothing to be tracked further in the parent state.
2178 i = get_prev_insn_idx(st, i, &history);
2179 if (i >= env->prog->len) {
2180 /* This can happen if backtracking reached insn 0
2181 * and there are still reg_mask or stack_mask
2183 * It means the backtracking missed the spot where
2184 * particular register was initialized with a constant.
2186 verbose(env, "BUG backtracking idx %d\n", i);
2187 WARN_ONCE(1, "verifier backtracking bug");
2196 func = st->frame[st->curframe];
2197 bitmap_from_u64(mask, reg_mask);
2198 for_each_set_bit(i, mask, 32) {
2199 reg = &func->regs[i];
2200 if (reg->type != SCALAR_VALUE) {
2201 reg_mask &= ~(1u << i);
2206 reg->precise = true;
2209 bitmap_from_u64(mask, stack_mask);
2210 for_each_set_bit(i, mask, 64) {
2211 if (i >= func->allocated_stack / BPF_REG_SIZE) {
2212 /* the sequence of instructions:
2214 * 3: (7b) *(u64 *)(r3 -8) = r0
2215 * 4: (79) r4 = *(u64 *)(r10 -8)
2216 * doesn't contain jmps. It's backtracked
2217 * as a single block.
2218 * During backtracking insn 3 is not recognized as
2219 * stack access, so at the end of backtracking
2220 * stack slot fp-8 is still marked in stack_mask.
2221 * However the parent state may not have accessed
2222 * fp-8 and it's "unallocated" stack space.
2223 * In such case fallback to conservative.
2225 mark_all_scalars_precise(env, st);
2229 if (func->stack[i].slot_type[0] != STACK_SPILL) {
2230 stack_mask &= ~(1ull << i);
2233 reg = &func->stack[i].spilled_ptr;
2234 if (reg->type != SCALAR_VALUE) {
2235 stack_mask &= ~(1ull << i);
2240 reg->precise = true;
2242 if (env->log.level & BPF_LOG_LEVEL) {
2243 print_verifier_state(env, func);
2244 verbose(env, "parent %s regs=%x stack=%llx marks\n",
2245 new_marks ? "didn't have" : "already had",
2246 reg_mask, stack_mask);
2249 if (!reg_mask && !stack_mask)
2254 last_idx = st->last_insn_idx;
2255 first_idx = st->first_insn_idx;
2260 static int mark_chain_precision(struct bpf_verifier_env *env, int regno)
2262 return __mark_chain_precision(env, regno, -1);
2265 static int mark_chain_precision_stack(struct bpf_verifier_env *env, int spi)
2267 return __mark_chain_precision(env, -1, spi);
2270 static bool is_spillable_regtype(enum bpf_reg_type type)
2273 case PTR_TO_MAP_VALUE:
2274 case PTR_TO_MAP_VALUE_OR_NULL:
2278 case PTR_TO_PACKET_META:
2279 case PTR_TO_PACKET_END:
2280 case PTR_TO_FLOW_KEYS:
2281 case CONST_PTR_TO_MAP:
2283 case PTR_TO_SOCKET_OR_NULL:
2284 case PTR_TO_SOCK_COMMON:
2285 case PTR_TO_SOCK_COMMON_OR_NULL:
2286 case PTR_TO_TCP_SOCK:
2287 case PTR_TO_TCP_SOCK_OR_NULL:
2288 case PTR_TO_XDP_SOCK:
2290 case PTR_TO_BTF_ID_OR_NULL:
2291 case PTR_TO_RDONLY_BUF:
2292 case PTR_TO_RDONLY_BUF_OR_NULL:
2293 case PTR_TO_RDWR_BUF:
2294 case PTR_TO_RDWR_BUF_OR_NULL:
2295 case PTR_TO_PERCPU_BTF_ID:
2297 case PTR_TO_MEM_OR_NULL:
2304 /* Does this register contain a constant zero? */
2305 static bool register_is_null(struct bpf_reg_state *reg)
2307 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
2310 static bool register_is_const(struct bpf_reg_state *reg)
2312 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
2315 static bool __is_scalar_unbounded(struct bpf_reg_state *reg)
2317 return tnum_is_unknown(reg->var_off) &&
2318 reg->smin_value == S64_MIN && reg->smax_value == S64_MAX &&
2319 reg->umin_value == 0 && reg->umax_value == U64_MAX &&
2320 reg->s32_min_value == S32_MIN && reg->s32_max_value == S32_MAX &&
2321 reg->u32_min_value == 0 && reg->u32_max_value == U32_MAX;
2324 static bool register_is_bounded(struct bpf_reg_state *reg)
2326 return reg->type == SCALAR_VALUE && !__is_scalar_unbounded(reg);
2329 static bool __is_pointer_value(bool allow_ptr_leaks,
2330 const struct bpf_reg_state *reg)
2332 if (allow_ptr_leaks)
2335 return reg->type != SCALAR_VALUE;
2338 static void save_register_state(struct bpf_func_state *state,
2339 int spi, struct bpf_reg_state *reg)
2343 state->stack[spi].spilled_ptr = *reg;
2344 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2346 for (i = 0; i < BPF_REG_SIZE; i++)
2347 state->stack[spi].slot_type[i] = STACK_SPILL;
2350 /* check_stack_{read,write}_fixed_off functions track spill/fill of registers,
2351 * stack boundary and alignment are checked in check_mem_access()
2353 static int check_stack_write_fixed_off(struct bpf_verifier_env *env,
2354 /* stack frame we're writing to */
2355 struct bpf_func_state *state,
2356 int off, int size, int value_regno,
2359 struct bpf_func_state *cur; /* state of the current function */
2360 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
2361 u32 dst_reg = env->prog->insnsi[insn_idx].dst_reg;
2362 struct bpf_reg_state *reg = NULL;
2364 err = realloc_func_state(state, round_up(slot + 1, BPF_REG_SIZE),
2365 state->acquired_refs, true);
2368 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
2369 * so it's aligned access and [off, off + size) are within stack limits
2371 if (!env->allow_ptr_leaks &&
2372 state->stack[spi].slot_type[0] == STACK_SPILL &&
2373 size != BPF_REG_SIZE) {
2374 verbose(env, "attempt to corrupt spilled pointer on stack\n");
2378 cur = env->cur_state->frame[env->cur_state->curframe];
2379 if (value_regno >= 0)
2380 reg = &cur->regs[value_regno];
2382 if (reg && size == BPF_REG_SIZE && register_is_bounded(reg) &&
2383 !register_is_null(reg) && env->bpf_capable) {
2384 if (dst_reg != BPF_REG_FP) {
2385 /* The backtracking logic can only recognize explicit
2386 * stack slot address like [fp - 8]. Other spill of
2387 * scalar via different register has to be conervative.
2388 * Backtrack from here and mark all registers as precise
2389 * that contributed into 'reg' being a constant.
2391 err = mark_chain_precision(env, value_regno);
2395 save_register_state(state, spi, reg);
2396 } else if (reg && is_spillable_regtype(reg->type)) {
2397 /* register containing pointer is being spilled into stack */
2398 if (size != BPF_REG_SIZE) {
2399 verbose_linfo(env, insn_idx, "; ");
2400 verbose(env, "invalid size of register spill\n");
2404 if (state != cur && reg->type == PTR_TO_STACK) {
2405 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
2409 if (!env->bypass_spec_v4) {
2410 bool sanitize = false;
2412 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
2413 register_is_const(&state->stack[spi].spilled_ptr))
2415 for (i = 0; i < BPF_REG_SIZE; i++)
2416 if (state->stack[spi].slot_type[i] == STACK_MISC) {
2421 int *poff = &env->insn_aux_data[insn_idx].sanitize_stack_off;
2422 int soff = (-spi - 1) * BPF_REG_SIZE;
2424 /* detected reuse of integer stack slot with a pointer
2425 * which means either llvm is reusing stack slot or
2426 * an attacker is trying to exploit CVE-2018-3639
2427 * (speculative store bypass)
2428 * Have to sanitize that slot with preemptive
2431 if (*poff && *poff != soff) {
2432 /* disallow programs where single insn stores
2433 * into two different stack slots, since verifier
2434 * cannot sanitize them
2437 "insn %d cannot access two stack slots fp%d and fp%d",
2438 insn_idx, *poff, soff);
2444 save_register_state(state, spi, reg);
2446 u8 type = STACK_MISC;
2448 /* regular write of data into stack destroys any spilled ptr */
2449 state->stack[spi].spilled_ptr.type = NOT_INIT;
2450 /* Mark slots as STACK_MISC if they belonged to spilled ptr. */
2451 if (state->stack[spi].slot_type[0] == STACK_SPILL)
2452 for (i = 0; i < BPF_REG_SIZE; i++)
2453 state->stack[spi].slot_type[i] = STACK_MISC;
2455 /* only mark the slot as written if all 8 bytes were written
2456 * otherwise read propagation may incorrectly stop too soon
2457 * when stack slots are partially written.
2458 * This heuristic means that read propagation will be
2459 * conservative, since it will add reg_live_read marks
2460 * to stack slots all the way to first state when programs
2461 * writes+reads less than 8 bytes
2463 if (size == BPF_REG_SIZE)
2464 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2466 /* when we zero initialize stack slots mark them as such */
2467 if (reg && register_is_null(reg)) {
2468 /* backtracking doesn't work for STACK_ZERO yet. */
2469 err = mark_chain_precision(env, value_regno);
2475 /* Mark slots affected by this stack write. */
2476 for (i = 0; i < size; i++)
2477 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
2483 /* Write the stack: 'stack[ptr_regno + off] = value_regno'. 'ptr_regno' is
2484 * known to contain a variable offset.
2485 * This function checks whether the write is permitted and conservatively
2486 * tracks the effects of the write, considering that each stack slot in the
2487 * dynamic range is potentially written to.
2489 * 'off' includes 'regno->off'.
2490 * 'value_regno' can be -1, meaning that an unknown value is being written to
2493 * Spilled pointers in range are not marked as written because we don't know
2494 * what's going to be actually written. This means that read propagation for
2495 * future reads cannot be terminated by this write.
2497 * For privileged programs, uninitialized stack slots are considered
2498 * initialized by this write (even though we don't know exactly what offsets
2499 * are going to be written to). The idea is that we don't want the verifier to
2500 * reject future reads that access slots written to through variable offsets.
2502 static int check_stack_write_var_off(struct bpf_verifier_env *env,
2503 /* func where register points to */
2504 struct bpf_func_state *state,
2505 int ptr_regno, int off, int size,
2506 int value_regno, int insn_idx)
2508 struct bpf_func_state *cur; /* state of the current function */
2509 int min_off, max_off;
2511 struct bpf_reg_state *ptr_reg = NULL, *value_reg = NULL;
2512 bool writing_zero = false;
2513 /* set if the fact that we're writing a zero is used to let any
2514 * stack slots remain STACK_ZERO
2516 bool zero_used = false;
2518 cur = env->cur_state->frame[env->cur_state->curframe];
2519 ptr_reg = &cur->regs[ptr_regno];
2520 min_off = ptr_reg->smin_value + off;
2521 max_off = ptr_reg->smax_value + off + size;
2522 if (value_regno >= 0)
2523 value_reg = &cur->regs[value_regno];
2524 if (value_reg && register_is_null(value_reg))
2525 writing_zero = true;
2527 err = realloc_func_state(state, round_up(-min_off, BPF_REG_SIZE),
2528 state->acquired_refs, true);
2533 /* Variable offset writes destroy any spilled pointers in range. */
2534 for (i = min_off; i < max_off; i++) {
2535 u8 new_type, *stype;
2539 spi = slot / BPF_REG_SIZE;
2540 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
2542 if (!env->allow_ptr_leaks
2543 && *stype != NOT_INIT
2544 && *stype != SCALAR_VALUE) {
2545 /* Reject the write if there's are spilled pointers in
2546 * range. If we didn't reject here, the ptr status
2547 * would be erased below (even though not all slots are
2548 * actually overwritten), possibly opening the door to
2551 verbose(env, "spilled ptr in range of var-offset stack write; insn %d, ptr off: %d",
2556 /* Erase all spilled pointers. */
2557 state->stack[spi].spilled_ptr.type = NOT_INIT;
2559 /* Update the slot type. */
2560 new_type = STACK_MISC;
2561 if (writing_zero && *stype == STACK_ZERO) {
2562 new_type = STACK_ZERO;
2565 /* If the slot is STACK_INVALID, we check whether it's OK to
2566 * pretend that it will be initialized by this write. The slot
2567 * might not actually be written to, and so if we mark it as
2568 * initialized future reads might leak uninitialized memory.
2569 * For privileged programs, we will accept such reads to slots
2570 * that may or may not be written because, if we're reject
2571 * them, the error would be too confusing.
2573 if (*stype == STACK_INVALID && !env->allow_uninit_stack) {
2574 verbose(env, "uninit stack in range of var-offset write prohibited for !root; insn %d, off: %d",
2581 /* backtracking doesn't work for STACK_ZERO yet. */
2582 err = mark_chain_precision(env, value_regno);
2589 /* When register 'dst_regno' is assigned some values from stack[min_off,
2590 * max_off), we set the register's type according to the types of the
2591 * respective stack slots. If all the stack values are known to be zeros, then
2592 * so is the destination reg. Otherwise, the register is considered to be
2593 * SCALAR. This function does not deal with register filling; the caller must
2594 * ensure that all spilled registers in the stack range have been marked as
2597 static void mark_reg_stack_read(struct bpf_verifier_env *env,
2598 /* func where src register points to */
2599 struct bpf_func_state *ptr_state,
2600 int min_off, int max_off, int dst_regno)
2602 struct bpf_verifier_state *vstate = env->cur_state;
2603 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2608 for (i = min_off; i < max_off; i++) {
2610 spi = slot / BPF_REG_SIZE;
2611 stype = ptr_state->stack[spi].slot_type;
2612 if (stype[slot % BPF_REG_SIZE] != STACK_ZERO)
2616 if (zeros == max_off - min_off) {
2617 /* any access_size read into register is zero extended,
2618 * so the whole register == const_zero
2620 __mark_reg_const_zero(&state->regs[dst_regno]);
2621 /* backtracking doesn't support STACK_ZERO yet,
2622 * so mark it precise here, so that later
2623 * backtracking can stop here.
2624 * Backtracking may not need this if this register
2625 * doesn't participate in pointer adjustment.
2626 * Forward propagation of precise flag is not
2627 * necessary either. This mark is only to stop
2628 * backtracking. Any register that contributed
2629 * to const 0 was marked precise before spill.
2631 state->regs[dst_regno].precise = true;
2633 /* have read misc data from the stack */
2634 mark_reg_unknown(env, state->regs, dst_regno);
2636 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
2639 /* Read the stack at 'off' and put the results into the register indicated by
2640 * 'dst_regno'. It handles reg filling if the addressed stack slot is a
2643 * 'dst_regno' can be -1, meaning that the read value is not going to a
2646 * The access is assumed to be within the current stack bounds.
2648 static int check_stack_read_fixed_off(struct bpf_verifier_env *env,
2649 /* func where src register points to */
2650 struct bpf_func_state *reg_state,
2651 int off, int size, int dst_regno)
2653 struct bpf_verifier_state *vstate = env->cur_state;
2654 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2655 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
2656 struct bpf_reg_state *reg;
2659 stype = reg_state->stack[spi].slot_type;
2660 reg = ®_state->stack[spi].spilled_ptr;
2662 if (stype[0] == STACK_SPILL) {
2663 if (size != BPF_REG_SIZE) {
2664 if (reg->type != SCALAR_VALUE) {
2665 verbose_linfo(env, env->insn_idx, "; ");
2666 verbose(env, "invalid size of register fill\n");
2669 if (dst_regno >= 0) {
2670 mark_reg_unknown(env, state->regs, dst_regno);
2671 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
2673 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2676 for (i = 1; i < BPF_REG_SIZE; i++) {
2677 if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) {
2678 verbose(env, "corrupted spill memory\n");
2683 if (dst_regno >= 0) {
2684 /* restore register state from stack */
2685 state->regs[dst_regno] = *reg;
2686 /* mark reg as written since spilled pointer state likely
2687 * has its liveness marks cleared by is_state_visited()
2688 * which resets stack/reg liveness for state transitions
2690 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
2691 } else if (__is_pointer_value(env->allow_ptr_leaks, reg)) {
2692 /* If dst_regno==-1, the caller is asking us whether
2693 * it is acceptable to use this value as a SCALAR_VALUE
2695 * We must not allow unprivileged callers to do that
2696 * with spilled pointers.
2698 verbose(env, "leaking pointer from stack off %d\n",
2702 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2706 for (i = 0; i < size; i++) {
2707 type = stype[(slot - i) % BPF_REG_SIZE];
2708 if (type == STACK_MISC)
2710 if (type == STACK_ZERO)
2712 verbose(env, "invalid read from stack off %d+%d size %d\n",
2716 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2718 mark_reg_stack_read(env, reg_state, off, off + size, dst_regno);
2723 enum stack_access_src {
2724 ACCESS_DIRECT = 1, /* the access is performed by an instruction */
2725 ACCESS_HELPER = 2, /* the access is performed by a helper */
2728 static int check_stack_range_initialized(struct bpf_verifier_env *env,
2729 int regno, int off, int access_size,
2730 bool zero_size_allowed,
2731 enum stack_access_src type,
2732 struct bpf_call_arg_meta *meta);
2734 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
2736 return cur_regs(env) + regno;
2739 /* Read the stack at 'ptr_regno + off' and put the result into the register
2741 * 'off' includes the pointer register's fixed offset(i.e. 'ptr_regno.off'),
2742 * but not its variable offset.
2743 * 'size' is assumed to be <= reg size and the access is assumed to be aligned.
2745 * As opposed to check_stack_read_fixed_off, this function doesn't deal with
2746 * filling registers (i.e. reads of spilled register cannot be detected when
2747 * the offset is not fixed). We conservatively mark 'dst_regno' as containing
2748 * SCALAR_VALUE. That's why we assert that the 'ptr_regno' has a variable
2749 * offset; for a fixed offset check_stack_read_fixed_off should be used
2752 static int check_stack_read_var_off(struct bpf_verifier_env *env,
2753 int ptr_regno, int off, int size, int dst_regno)
2755 /* The state of the source register. */
2756 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
2757 struct bpf_func_state *ptr_state = func(env, reg);
2759 int min_off, max_off;
2761 /* Note that we pass a NULL meta, so raw access will not be permitted.
2763 err = check_stack_range_initialized(env, ptr_regno, off, size,
2764 false, ACCESS_DIRECT, NULL);
2768 min_off = reg->smin_value + off;
2769 max_off = reg->smax_value + off;
2770 mark_reg_stack_read(env, ptr_state, min_off, max_off + size, dst_regno);
2774 /* check_stack_read dispatches to check_stack_read_fixed_off or
2775 * check_stack_read_var_off.
2777 * The caller must ensure that the offset falls within the allocated stack
2780 * 'dst_regno' is a register which will receive the value from the stack. It
2781 * can be -1, meaning that the read value is not going to a register.
2783 static int check_stack_read(struct bpf_verifier_env *env,
2784 int ptr_regno, int off, int size,
2787 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
2788 struct bpf_func_state *state = func(env, reg);
2790 /* Some accesses are only permitted with a static offset. */
2791 bool var_off = !tnum_is_const(reg->var_off);
2793 /* The offset is required to be static when reads don't go to a
2794 * register, in order to not leak pointers (see
2795 * check_stack_read_fixed_off).
2797 if (dst_regno < 0 && var_off) {
2800 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2801 verbose(env, "variable offset stack pointer cannot be passed into helper function; var_off=%s off=%d size=%d\n",
2805 /* Variable offset is prohibited for unprivileged mode for simplicity
2806 * since it requires corresponding support in Spectre masking for stack
2807 * ALU. See also retrieve_ptr_limit().
2809 if (!env->bypass_spec_v1 && var_off) {
2812 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2813 verbose(env, "R%d variable offset stack access prohibited for !root, var_off=%s\n",
2819 off += reg->var_off.value;
2820 err = check_stack_read_fixed_off(env, state, off, size,
2823 /* Variable offset stack reads need more conservative handling
2824 * than fixed offset ones. Note that dst_regno >= 0 on this
2827 err = check_stack_read_var_off(env, ptr_regno, off, size,
2834 /* check_stack_write dispatches to check_stack_write_fixed_off or
2835 * check_stack_write_var_off.
2837 * 'ptr_regno' is the register used as a pointer into the stack.
2838 * 'off' includes 'ptr_regno->off', but not its variable offset (if any).
2839 * 'value_regno' is the register whose value we're writing to the stack. It can
2840 * be -1, meaning that we're not writing from a register.
2842 * The caller must ensure that the offset falls within the maximum stack size.
2844 static int check_stack_write(struct bpf_verifier_env *env,
2845 int ptr_regno, int off, int size,
2846 int value_regno, int insn_idx)
2848 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
2849 struct bpf_func_state *state = func(env, reg);
2852 if (tnum_is_const(reg->var_off)) {
2853 off += reg->var_off.value;
2854 err = check_stack_write_fixed_off(env, state, off, size,
2855 value_regno, insn_idx);
2857 /* Variable offset stack reads need more conservative handling
2858 * than fixed offset ones.
2860 err = check_stack_write_var_off(env, state,
2861 ptr_regno, off, size,
2862 value_regno, insn_idx);
2867 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
2868 int off, int size, enum bpf_access_type type)
2870 struct bpf_reg_state *regs = cur_regs(env);
2871 struct bpf_map *map = regs[regno].map_ptr;
2872 u32 cap = bpf_map_flags_to_cap(map);
2874 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
2875 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
2876 map->value_size, off, size);
2880 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
2881 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
2882 map->value_size, off, size);
2889 /* check read/write into memory region (e.g., map value, ringbuf sample, etc) */
2890 static int __check_mem_access(struct bpf_verifier_env *env, int regno,
2891 int off, int size, u32 mem_size,
2892 bool zero_size_allowed)
2894 bool size_ok = size > 0 || (size == 0 && zero_size_allowed);
2895 struct bpf_reg_state *reg;
2897 if (off >= 0 && size_ok && (u64)off + size <= mem_size)
2900 reg = &cur_regs(env)[regno];
2901 switch (reg->type) {
2902 case PTR_TO_MAP_VALUE:
2903 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
2904 mem_size, off, size);
2907 case PTR_TO_PACKET_META:
2908 case PTR_TO_PACKET_END:
2909 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
2910 off, size, regno, reg->id, off, mem_size);
2914 verbose(env, "invalid access to memory, mem_size=%u off=%d size=%d\n",
2915 mem_size, off, size);
2921 /* check read/write into a memory region with possible variable offset */
2922 static int check_mem_region_access(struct bpf_verifier_env *env, u32 regno,
2923 int off, int size, u32 mem_size,
2924 bool zero_size_allowed)
2926 struct bpf_verifier_state *vstate = env->cur_state;
2927 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2928 struct bpf_reg_state *reg = &state->regs[regno];
2931 /* We may have adjusted the register pointing to memory region, so we
2932 * need to try adding each of min_value and max_value to off
2933 * to make sure our theoretical access will be safe.
2935 if (env->log.level & BPF_LOG_LEVEL)
2936 print_verifier_state(env, state);
2938 /* The minimum value is only important with signed
2939 * comparisons where we can't assume the floor of a
2940 * value is 0. If we are using signed variables for our
2941 * index'es we need to make sure that whatever we use
2942 * will have a set floor within our range.
2944 if (reg->smin_value < 0 &&
2945 (reg->smin_value == S64_MIN ||
2946 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
2947 reg->smin_value + off < 0)) {
2948 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2952 err = __check_mem_access(env, regno, reg->smin_value + off, size,
2953 mem_size, zero_size_allowed);
2955 verbose(env, "R%d min value is outside of the allowed memory range\n",
2960 /* If we haven't set a max value then we need to bail since we can't be
2961 * sure we won't do bad things.
2962 * If reg->umax_value + off could overflow, treat that as unbounded too.
2964 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
2965 verbose(env, "R%d unbounded memory access, make sure to bounds check any such access\n",
2969 err = __check_mem_access(env, regno, reg->umax_value + off, size,
2970 mem_size, zero_size_allowed);
2972 verbose(env, "R%d max value is outside of the allowed memory range\n",
2980 /* check read/write into a map element with possible variable offset */
2981 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
2982 int off, int size, bool zero_size_allowed)
2984 struct bpf_verifier_state *vstate = env->cur_state;
2985 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2986 struct bpf_reg_state *reg = &state->regs[regno];
2987 struct bpf_map *map = reg->map_ptr;
2990 err = check_mem_region_access(env, regno, off, size, map->value_size,
2995 if (map_value_has_spin_lock(map)) {
2996 u32 lock = map->spin_lock_off;
2998 /* if any part of struct bpf_spin_lock can be touched by
2999 * load/store reject this program.
3000 * To check that [x1, x2) overlaps with [y1, y2)
3001 * it is sufficient to check x1 < y2 && y1 < x2.
3003 if (reg->smin_value + off < lock + sizeof(struct bpf_spin_lock) &&
3004 lock < reg->umax_value + off + size) {
3005 verbose(env, "bpf_spin_lock cannot be accessed directly by load/store\n");
3012 #define MAX_PACKET_OFF 0xffff
3014 static enum bpf_prog_type resolve_prog_type(struct bpf_prog *prog)
3016 return prog->aux->dst_prog ? prog->aux->dst_prog->type : prog->type;
3019 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
3020 const struct bpf_call_arg_meta *meta,
3021 enum bpf_access_type t)
3023 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
3025 switch (prog_type) {
3026 /* Program types only with direct read access go here! */
3027 case BPF_PROG_TYPE_LWT_IN:
3028 case BPF_PROG_TYPE_LWT_OUT:
3029 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
3030 case BPF_PROG_TYPE_SK_REUSEPORT:
3031 case BPF_PROG_TYPE_FLOW_DISSECTOR:
3032 case BPF_PROG_TYPE_CGROUP_SKB:
3037 /* Program types with direct read + write access go here! */
3038 case BPF_PROG_TYPE_SCHED_CLS:
3039 case BPF_PROG_TYPE_SCHED_ACT:
3040 case BPF_PROG_TYPE_XDP:
3041 case BPF_PROG_TYPE_LWT_XMIT:
3042 case BPF_PROG_TYPE_SK_SKB:
3043 case BPF_PROG_TYPE_SK_MSG:
3045 return meta->pkt_access;
3047 env->seen_direct_write = true;
3050 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
3052 env->seen_direct_write = true;
3061 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
3062 int size, bool zero_size_allowed)
3064 struct bpf_reg_state *regs = cur_regs(env);
3065 struct bpf_reg_state *reg = ®s[regno];
3068 /* We may have added a variable offset to the packet pointer; but any
3069 * reg->range we have comes after that. We are only checking the fixed
3073 /* We don't allow negative numbers, because we aren't tracking enough
3074 * detail to prove they're safe.
3076 if (reg->smin_value < 0) {
3077 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3082 err = reg->range < 0 ? -EINVAL :
3083 __check_mem_access(env, regno, off, size, reg->range,
3086 verbose(env, "R%d offset is outside of the packet\n", regno);
3090 /* __check_mem_access has made sure "off + size - 1" is within u16.
3091 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
3092 * otherwise find_good_pkt_pointers would have refused to set range info
3093 * that __check_mem_access would have rejected this pkt access.
3094 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
3096 env->prog->aux->max_pkt_offset =
3097 max_t(u32, env->prog->aux->max_pkt_offset,
3098 off + reg->umax_value + size - 1);
3103 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
3104 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
3105 enum bpf_access_type t, enum bpf_reg_type *reg_type,
3106 struct btf **btf, u32 *btf_id)
3108 struct bpf_insn_access_aux info = {
3109 .reg_type = *reg_type,
3113 if (env->ops->is_valid_access &&
3114 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
3115 /* A non zero info.ctx_field_size indicates that this field is a
3116 * candidate for later verifier transformation to load the whole
3117 * field and then apply a mask when accessed with a narrower
3118 * access than actual ctx access size. A zero info.ctx_field_size
3119 * will only allow for whole field access and rejects any other
3120 * type of narrower access.
3122 *reg_type = info.reg_type;
3124 if (*reg_type == PTR_TO_BTF_ID || *reg_type == PTR_TO_BTF_ID_OR_NULL) {
3126 *btf_id = info.btf_id;
3128 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
3130 /* remember the offset of last byte accessed in ctx */
3131 if (env->prog->aux->max_ctx_offset < off + size)
3132 env->prog->aux->max_ctx_offset = off + size;
3136 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
3140 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
3143 if (size < 0 || off < 0 ||
3144 (u64)off + size > sizeof(struct bpf_flow_keys)) {
3145 verbose(env, "invalid access to flow keys off=%d size=%d\n",
3152 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
3153 u32 regno, int off, int size,
3154 enum bpf_access_type t)
3156 struct bpf_reg_state *regs = cur_regs(env);
3157 struct bpf_reg_state *reg = ®s[regno];
3158 struct bpf_insn_access_aux info = {};
3161 if (reg->smin_value < 0) {
3162 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3167 switch (reg->type) {
3168 case PTR_TO_SOCK_COMMON:
3169 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
3172 valid = bpf_sock_is_valid_access(off, size, t, &info);
3174 case PTR_TO_TCP_SOCK:
3175 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
3177 case PTR_TO_XDP_SOCK:
3178 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
3186 env->insn_aux_data[insn_idx].ctx_field_size =
3187 info.ctx_field_size;
3191 verbose(env, "R%d invalid %s access off=%d size=%d\n",
3192 regno, reg_type_str[reg->type], off, size);
3197 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
3199 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
3202 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
3204 const struct bpf_reg_state *reg = reg_state(env, regno);
3206 return reg->type == PTR_TO_CTX;
3209 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
3211 const struct bpf_reg_state *reg = reg_state(env, regno);
3213 return type_is_sk_pointer(reg->type);
3216 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
3218 const struct bpf_reg_state *reg = reg_state(env, regno);
3220 return type_is_pkt_pointer(reg->type);
3223 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
3225 const struct bpf_reg_state *reg = reg_state(env, regno);
3227 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
3228 return reg->type == PTR_TO_FLOW_KEYS;
3231 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
3232 const struct bpf_reg_state *reg,
3233 int off, int size, bool strict)
3235 struct tnum reg_off;
3238 /* Byte size accesses are always allowed. */
3239 if (!strict || size == 1)
3242 /* For platforms that do not have a Kconfig enabling
3243 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
3244 * NET_IP_ALIGN is universally set to '2'. And on platforms
3245 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
3246 * to this code only in strict mode where we want to emulate
3247 * the NET_IP_ALIGN==2 checking. Therefore use an
3248 * unconditional IP align value of '2'.
3252 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
3253 if (!tnum_is_aligned(reg_off, size)) {
3256 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3258 "misaligned packet access off %d+%s+%d+%d size %d\n",
3259 ip_align, tn_buf, reg->off, off, size);
3266 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
3267 const struct bpf_reg_state *reg,
3268 const char *pointer_desc,
3269 int off, int size, bool strict)
3271 struct tnum reg_off;
3273 /* Byte size accesses are always allowed. */
3274 if (!strict || size == 1)
3277 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
3278 if (!tnum_is_aligned(reg_off, size)) {
3281 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3282 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
3283 pointer_desc, tn_buf, reg->off, off, size);
3290 static int check_ptr_alignment(struct bpf_verifier_env *env,
3291 const struct bpf_reg_state *reg, int off,
3292 int size, bool strict_alignment_once)
3294 bool strict = env->strict_alignment || strict_alignment_once;
3295 const char *pointer_desc = "";
3297 switch (reg->type) {
3299 case PTR_TO_PACKET_META:
3300 /* Special case, because of NET_IP_ALIGN. Given metadata sits
3301 * right in front, treat it the very same way.
3303 return check_pkt_ptr_alignment(env, reg, off, size, strict);
3304 case PTR_TO_FLOW_KEYS:
3305 pointer_desc = "flow keys ";
3307 case PTR_TO_MAP_VALUE:
3308 pointer_desc = "value ";
3311 pointer_desc = "context ";
3314 pointer_desc = "stack ";
3315 /* The stack spill tracking logic in check_stack_write_fixed_off()
3316 * and check_stack_read_fixed_off() relies on stack accesses being
3322 pointer_desc = "sock ";
3324 case PTR_TO_SOCK_COMMON:
3325 pointer_desc = "sock_common ";
3327 case PTR_TO_TCP_SOCK:
3328 pointer_desc = "tcp_sock ";
3330 case PTR_TO_XDP_SOCK:
3331 pointer_desc = "xdp_sock ";
3336 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
3340 static int update_stack_depth(struct bpf_verifier_env *env,
3341 const struct bpf_func_state *func,
3344 u16 stack = env->subprog_info[func->subprogno].stack_depth;
3349 /* update known max for given subprogram */
3350 env->subprog_info[func->subprogno].stack_depth = -off;
3354 /* starting from main bpf function walk all instructions of the function
3355 * and recursively walk all callees that given function can call.
3356 * Ignore jump and exit insns.
3357 * Since recursion is prevented by check_cfg() this algorithm
3358 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
3360 static int check_max_stack_depth(struct bpf_verifier_env *env)
3362 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
3363 struct bpf_subprog_info *subprog = env->subprog_info;
3364 struct bpf_insn *insn = env->prog->insnsi;
3365 bool tail_call_reachable = false;
3366 int ret_insn[MAX_CALL_FRAMES];
3367 int ret_prog[MAX_CALL_FRAMES];
3371 /* protect against potential stack overflow that might happen when
3372 * bpf2bpf calls get combined with tailcalls. Limit the caller's stack
3373 * depth for such case down to 256 so that the worst case scenario
3374 * would result in 8k stack size (32 which is tailcall limit * 256 =
3377 * To get the idea what might happen, see an example:
3378 * func1 -> sub rsp, 128
3379 * subfunc1 -> sub rsp, 256
3380 * tailcall1 -> add rsp, 256
3381 * func2 -> sub rsp, 192 (total stack size = 128 + 192 = 320)
3382 * subfunc2 -> sub rsp, 64
3383 * subfunc22 -> sub rsp, 128
3384 * tailcall2 -> add rsp, 128
3385 * func3 -> sub rsp, 32 (total stack size 128 + 192 + 64 + 32 = 416)
3387 * tailcall will unwind the current stack frame but it will not get rid
3388 * of caller's stack as shown on the example above.
3390 if (idx && subprog[idx].has_tail_call && depth >= 256) {
3392 "tail_calls are not allowed when call stack of previous frames is %d bytes. Too large\n",
3396 /* round up to 32-bytes, since this is granularity
3397 * of interpreter stack size
3399 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
3400 if (depth > MAX_BPF_STACK) {
3401 verbose(env, "combined stack size of %d calls is %d. Too large\n",
3406 subprog_end = subprog[idx + 1].start;
3407 for (; i < subprog_end; i++) {
3408 if (!bpf_pseudo_call(insn + i))
3410 /* remember insn and function to return to */
3411 ret_insn[frame] = i + 1;
3412 ret_prog[frame] = idx;
3414 /* find the callee */
3415 i = i + insn[i].imm + 1;
3416 idx = find_subprog(env, i);
3418 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3423 if (subprog[idx].has_tail_call)
3424 tail_call_reachable = true;
3427 if (frame >= MAX_CALL_FRAMES) {
3428 verbose(env, "the call stack of %d frames is too deep !\n",
3434 /* if tail call got detected across bpf2bpf calls then mark each of the
3435 * currently present subprog frames as tail call reachable subprogs;
3436 * this info will be utilized by JIT so that we will be preserving the
3437 * tail call counter throughout bpf2bpf calls combined with tailcalls
3439 if (tail_call_reachable)
3440 for (j = 0; j < frame; j++)
3441 subprog[ret_prog[j]].tail_call_reachable = true;
3443 /* end of for() loop means the last insn of the 'subprog'
3444 * was reached. Doesn't matter whether it was JA or EXIT
3448 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
3450 i = ret_insn[frame];
3451 idx = ret_prog[frame];
3455 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
3456 static int get_callee_stack_depth(struct bpf_verifier_env *env,
3457 const struct bpf_insn *insn, int idx)
3459 int start = idx + insn->imm + 1, subprog;
3461 subprog = find_subprog(env, start);
3463 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3467 return env->subprog_info[subprog].stack_depth;
3471 int check_ctx_reg(struct bpf_verifier_env *env,
3472 const struct bpf_reg_state *reg, int regno)
3474 /* Access to ctx or passing it to a helper is only allowed in
3475 * its original, unmodified form.
3479 verbose(env, "dereference of modified ctx ptr R%d off=%d disallowed\n",
3484 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3487 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3488 verbose(env, "variable ctx access var_off=%s disallowed\n", tn_buf);
3495 static int __check_buffer_access(struct bpf_verifier_env *env,
3496 const char *buf_info,
3497 const struct bpf_reg_state *reg,
3498 int regno, int off, int size)
3502 "R%d invalid %s buffer access: off=%d, size=%d\n",
3503 regno, buf_info, off, size);
3506 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3509 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3511 "R%d invalid variable buffer offset: off=%d, var_off=%s\n",
3512 regno, off, tn_buf);
3519 static int check_tp_buffer_access(struct bpf_verifier_env *env,
3520 const struct bpf_reg_state *reg,
3521 int regno, int off, int size)
3525 err = __check_buffer_access(env, "tracepoint", reg, regno, off, size);
3529 if (off + size > env->prog->aux->max_tp_access)
3530 env->prog->aux->max_tp_access = off + size;
3535 static int check_buffer_access(struct bpf_verifier_env *env,
3536 const struct bpf_reg_state *reg,
3537 int regno, int off, int size,
3538 bool zero_size_allowed,
3539 const char *buf_info,
3544 err = __check_buffer_access(env, buf_info, reg, regno, off, size);
3548 if (off + size > *max_access)
3549 *max_access = off + size;
3554 /* BPF architecture zero extends alu32 ops into 64-bit registesr */
3555 static void zext_32_to_64(struct bpf_reg_state *reg)
3557 reg->var_off = tnum_subreg(reg->var_off);
3558 __reg_assign_32_into_64(reg);
3561 /* truncate register to smaller size (in bytes)
3562 * must be called with size < BPF_REG_SIZE
3564 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
3568 /* clear high bits in bit representation */
3569 reg->var_off = tnum_cast(reg->var_off, size);
3571 /* fix arithmetic bounds */
3572 mask = ((u64)1 << (size * 8)) - 1;
3573 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
3574 reg->umin_value &= mask;
3575 reg->umax_value &= mask;
3577 reg->umin_value = 0;
3578 reg->umax_value = mask;
3580 reg->smin_value = reg->umin_value;
3581 reg->smax_value = reg->umax_value;
3583 /* If size is smaller than 32bit register the 32bit register
3584 * values are also truncated so we push 64-bit bounds into
3585 * 32-bit bounds. Above were truncated < 32-bits already.
3589 __reg_combine_64_into_32(reg);
3592 static bool bpf_map_is_rdonly(const struct bpf_map *map)
3594 return (map->map_flags & BPF_F_RDONLY_PROG) && map->frozen;
3597 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val)
3603 err = map->ops->map_direct_value_addr(map, &addr, off);
3606 ptr = (void *)(long)addr + off;
3610 *val = (u64)*(u8 *)ptr;
3613 *val = (u64)*(u16 *)ptr;
3616 *val = (u64)*(u32 *)ptr;
3627 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
3628 struct bpf_reg_state *regs,
3629 int regno, int off, int size,
3630 enum bpf_access_type atype,
3633 struct bpf_reg_state *reg = regs + regno;
3634 const struct btf_type *t = btf_type_by_id(reg->btf, reg->btf_id);
3635 const char *tname = btf_name_by_offset(reg->btf, t->name_off);
3641 "R%d is ptr_%s invalid negative access: off=%d\n",
3645 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3648 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3650 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
3651 regno, tname, off, tn_buf);
3655 if (env->ops->btf_struct_access) {
3656 ret = env->ops->btf_struct_access(&env->log, reg->btf, t,
3657 off, size, atype, &btf_id);
3659 if (atype != BPF_READ) {
3660 verbose(env, "only read is supported\n");
3664 ret = btf_struct_access(&env->log, reg->btf, t, off, size,
3671 if (atype == BPF_READ && value_regno >= 0)
3672 mark_btf_ld_reg(env, regs, value_regno, ret, reg->btf, btf_id);
3677 static int check_ptr_to_map_access(struct bpf_verifier_env *env,
3678 struct bpf_reg_state *regs,
3679 int regno, int off, int size,
3680 enum bpf_access_type atype,
3683 struct bpf_reg_state *reg = regs + regno;
3684 struct bpf_map *map = reg->map_ptr;
3685 const struct btf_type *t;
3691 verbose(env, "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n");
3695 if (!map->ops->map_btf_id || !*map->ops->map_btf_id) {
3696 verbose(env, "map_ptr access not supported for map type %d\n",
3701 t = btf_type_by_id(btf_vmlinux, *map->ops->map_btf_id);
3702 tname = btf_name_by_offset(btf_vmlinux, t->name_off);
3704 if (!env->allow_ptr_to_map_access) {
3706 "%s access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
3712 verbose(env, "R%d is %s invalid negative access: off=%d\n",
3717 if (atype != BPF_READ) {
3718 verbose(env, "only read from %s is supported\n", tname);
3722 ret = btf_struct_access(&env->log, btf_vmlinux, t, off, size, atype, &btf_id);
3726 if (value_regno >= 0)
3727 mark_btf_ld_reg(env, regs, value_regno, ret, btf_vmlinux, btf_id);
3732 /* Check that the stack access at the given offset is within bounds. The
3733 * maximum valid offset is -1.
3735 * The minimum valid offset is -MAX_BPF_STACK for writes, and
3736 * -state->allocated_stack for reads.
3738 static int check_stack_slot_within_bounds(int off,
3739 struct bpf_func_state *state,
3740 enum bpf_access_type t)
3745 min_valid_off = -MAX_BPF_STACK;
3747 min_valid_off = -state->allocated_stack;
3749 if (off < min_valid_off || off > -1)
3754 /* Check that the stack access at 'regno + off' falls within the maximum stack
3757 * 'off' includes `regno->offset`, but not its dynamic part (if any).
3759 static int check_stack_access_within_bounds(
3760 struct bpf_verifier_env *env,
3761 int regno, int off, int access_size,
3762 enum stack_access_src src, enum bpf_access_type type)
3764 struct bpf_reg_state *regs = cur_regs(env);
3765 struct bpf_reg_state *reg = regs + regno;
3766 struct bpf_func_state *state = func(env, reg);
3767 int min_off, max_off;
3771 if (src == ACCESS_HELPER)
3772 /* We don't know if helpers are reading or writing (or both). */
3773 err_extra = " indirect access to";
3774 else if (type == BPF_READ)
3775 err_extra = " read from";
3777 err_extra = " write to";
3779 if (tnum_is_const(reg->var_off)) {
3780 min_off = reg->var_off.value + off;
3781 if (access_size > 0)
3782 max_off = min_off + access_size - 1;
3786 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
3787 reg->smin_value <= -BPF_MAX_VAR_OFF) {
3788 verbose(env, "invalid unbounded variable-offset%s stack R%d\n",
3792 min_off = reg->smin_value + off;
3793 if (access_size > 0)
3794 max_off = reg->smax_value + off + access_size - 1;
3799 err = check_stack_slot_within_bounds(min_off, state, type);
3801 err = check_stack_slot_within_bounds(max_off, state, type);
3804 if (tnum_is_const(reg->var_off)) {
3805 verbose(env, "invalid%s stack R%d off=%d size=%d\n",
3806 err_extra, regno, off, access_size);
3810 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3811 verbose(env, "invalid variable-offset%s stack R%d var_off=%s size=%d\n",
3812 err_extra, regno, tn_buf, access_size);
3818 /* check whether memory at (regno + off) is accessible for t = (read | write)
3819 * if t==write, value_regno is a register which value is stored into memory
3820 * if t==read, value_regno is a register which will receive the value from memory
3821 * if t==write && value_regno==-1, some unknown value is stored into memory
3822 * if t==read && value_regno==-1, don't care what we read from memory
3824 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
3825 int off, int bpf_size, enum bpf_access_type t,
3826 int value_regno, bool strict_alignment_once)
3828 struct bpf_reg_state *regs = cur_regs(env);
3829 struct bpf_reg_state *reg = regs + regno;
3830 struct bpf_func_state *state;
3833 size = bpf_size_to_bytes(bpf_size);
3837 /* alignment checks will add in reg->off themselves */
3838 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
3842 /* for access checks, reg->off is just part of off */
3845 if (reg->type == PTR_TO_MAP_VALUE) {
3846 if (t == BPF_WRITE && value_regno >= 0 &&
3847 is_pointer_value(env, value_regno)) {
3848 verbose(env, "R%d leaks addr into map\n", value_regno);
3851 err = check_map_access_type(env, regno, off, size, t);
3854 err = check_map_access(env, regno, off, size, false);
3855 if (!err && t == BPF_READ && value_regno >= 0) {
3856 struct bpf_map *map = reg->map_ptr;
3858 /* if map is read-only, track its contents as scalars */
3859 if (tnum_is_const(reg->var_off) &&
3860 bpf_map_is_rdonly(map) &&
3861 map->ops->map_direct_value_addr) {
3862 int map_off = off + reg->var_off.value;
3865 err = bpf_map_direct_read(map, map_off, size,
3870 regs[value_regno].type = SCALAR_VALUE;
3871 __mark_reg_known(®s[value_regno], val);
3873 mark_reg_unknown(env, regs, value_regno);
3876 } else if (reg->type == PTR_TO_MEM) {
3877 if (t == BPF_WRITE && value_regno >= 0 &&
3878 is_pointer_value(env, value_regno)) {
3879 verbose(env, "R%d leaks addr into mem\n", value_regno);
3882 err = check_mem_region_access(env, regno, off, size,
3883 reg->mem_size, false);
3884 if (!err && t == BPF_READ && value_regno >= 0)
3885 mark_reg_unknown(env, regs, value_regno);
3886 } else if (reg->type == PTR_TO_CTX) {
3887 enum bpf_reg_type reg_type = SCALAR_VALUE;
3888 struct btf *btf = NULL;
3891 if (t == BPF_WRITE && value_regno >= 0 &&
3892 is_pointer_value(env, value_regno)) {
3893 verbose(env, "R%d leaks addr into ctx\n", value_regno);
3897 err = check_ctx_reg(env, reg, regno);
3901 err = check_ctx_access(env, insn_idx, off, size, t, ®_type, &btf, &btf_id);
3903 verbose_linfo(env, insn_idx, "; ");
3904 if (!err && t == BPF_READ && value_regno >= 0) {
3905 /* ctx access returns either a scalar, or a
3906 * PTR_TO_PACKET[_META,_END]. In the latter
3907 * case, we know the offset is zero.
3909 if (reg_type == SCALAR_VALUE) {
3910 mark_reg_unknown(env, regs, value_regno);
3912 mark_reg_known_zero(env, regs,
3914 if (reg_type_may_be_null(reg_type))
3915 regs[value_regno].id = ++env->id_gen;
3916 /* A load of ctx field could have different
3917 * actual load size with the one encoded in the
3918 * insn. When the dst is PTR, it is for sure not
3921 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
3922 if (reg_type == PTR_TO_BTF_ID ||
3923 reg_type == PTR_TO_BTF_ID_OR_NULL) {
3924 regs[value_regno].btf = btf;
3925 regs[value_regno].btf_id = btf_id;
3928 regs[value_regno].type = reg_type;
3931 } else if (reg->type == PTR_TO_STACK) {
3932 /* Basic bounds checks. */
3933 err = check_stack_access_within_bounds(env, regno, off, size, ACCESS_DIRECT, t);
3937 state = func(env, reg);
3938 err = update_stack_depth(env, state, off);
3943 err = check_stack_read(env, regno, off, size,
3946 err = check_stack_write(env, regno, off, size,
3947 value_regno, insn_idx);
3948 } else if (reg_is_pkt_pointer(reg)) {
3949 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
3950 verbose(env, "cannot write into packet\n");
3953 if (t == BPF_WRITE && value_regno >= 0 &&
3954 is_pointer_value(env, value_regno)) {
3955 verbose(env, "R%d leaks addr into packet\n",
3959 err = check_packet_access(env, regno, off, size, false);
3960 if (!err && t == BPF_READ && value_regno >= 0)
3961 mark_reg_unknown(env, regs, value_regno);
3962 } else if (reg->type == PTR_TO_FLOW_KEYS) {
3963 if (t == BPF_WRITE && value_regno >= 0 &&
3964 is_pointer_value(env, value_regno)) {
3965 verbose(env, "R%d leaks addr into flow keys\n",
3970 err = check_flow_keys_access(env, off, size);
3971 if (!err && t == BPF_READ && value_regno >= 0)
3972 mark_reg_unknown(env, regs, value_regno);
3973 } else if (type_is_sk_pointer(reg->type)) {
3974 if (t == BPF_WRITE) {
3975 verbose(env, "R%d cannot write into %s\n",
3976 regno, reg_type_str[reg->type]);
3979 err = check_sock_access(env, insn_idx, regno, off, size, t);
3980 if (!err && value_regno >= 0)
3981 mark_reg_unknown(env, regs, value_regno);
3982 } else if (reg->type == PTR_TO_TP_BUFFER) {
3983 err = check_tp_buffer_access(env, reg, regno, off, size);
3984 if (!err && t == BPF_READ && value_regno >= 0)
3985 mark_reg_unknown(env, regs, value_regno);
3986 } else if (reg->type == PTR_TO_BTF_ID) {
3987 err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
3989 } else if (reg->type == CONST_PTR_TO_MAP) {
3990 err = check_ptr_to_map_access(env, regs, regno, off, size, t,
3992 } else if (reg->type == PTR_TO_RDONLY_BUF) {
3993 if (t == BPF_WRITE) {
3994 verbose(env, "R%d cannot write into %s\n",
3995 regno, reg_type_str[reg->type]);
3998 err = check_buffer_access(env, reg, regno, off, size, false,
4000 &env->prog->aux->max_rdonly_access);
4001 if (!err && value_regno >= 0)
4002 mark_reg_unknown(env, regs, value_regno);
4003 } else if (reg->type == PTR_TO_RDWR_BUF) {
4004 err = check_buffer_access(env, reg, regno, off, size, false,
4006 &env->prog->aux->max_rdwr_access);
4007 if (!err && t == BPF_READ && value_regno >= 0)
4008 mark_reg_unknown(env, regs, value_regno);
4010 verbose(env, "R%d invalid mem access '%s'\n", regno,
4011 reg_type_str[reg->type]);
4015 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
4016 regs[value_regno].type == SCALAR_VALUE) {
4017 /* b/h/w load zero-extends, mark upper bits as known 0 */
4018 coerce_reg_to_size(®s[value_regno], size);
4023 static int check_atomic(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
4028 switch (insn->imm) {
4030 case BPF_ADD | BPF_FETCH:
4032 case BPF_AND | BPF_FETCH:
4034 case BPF_OR | BPF_FETCH:
4036 case BPF_XOR | BPF_FETCH:
4041 verbose(env, "BPF_ATOMIC uses invalid atomic opcode %02x\n", insn->imm);
4045 if (BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) {
4046 verbose(env, "invalid atomic operand size\n");
4050 /* check src1 operand */
4051 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4055 /* check src2 operand */
4056 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4060 if (insn->imm == BPF_CMPXCHG) {
4061 /* Check comparison of R0 with memory location */
4062 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
4067 if (is_pointer_value(env, insn->src_reg)) {
4068 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
4072 if (is_ctx_reg(env, insn->dst_reg) ||
4073 is_pkt_reg(env, insn->dst_reg) ||
4074 is_flow_key_reg(env, insn->dst_reg) ||
4075 is_sk_reg(env, insn->dst_reg)) {
4076 verbose(env, "BPF_ATOMIC stores into R%d %s is not allowed\n",
4078 reg_type_str[reg_state(env, insn->dst_reg)->type]);
4082 if (insn->imm & BPF_FETCH) {
4083 if (insn->imm == BPF_CMPXCHG)
4084 load_reg = BPF_REG_0;
4086 load_reg = insn->src_reg;
4088 /* check and record load of old value */
4089 err = check_reg_arg(env, load_reg, DST_OP);
4093 /* This instruction accesses a memory location but doesn't
4094 * actually load it into a register.
4099 /* check whether we can read the memory */
4100 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4101 BPF_SIZE(insn->code), BPF_READ, load_reg, true);
4105 /* check whether we can write into the same memory */
4106 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4107 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
4114 /* When register 'regno' is used to read the stack (either directly or through
4115 * a helper function) make sure that it's within stack boundary and, depending
4116 * on the access type, that all elements of the stack are initialized.
4118 * 'off' includes 'regno->off', but not its dynamic part (if any).
4120 * All registers that have been spilled on the stack in the slots within the
4121 * read offsets are marked as read.
4123 static int check_stack_range_initialized(
4124 struct bpf_verifier_env *env, int regno, int off,
4125 int access_size, bool zero_size_allowed,
4126 enum stack_access_src type, struct bpf_call_arg_meta *meta)
4128 struct bpf_reg_state *reg = reg_state(env, regno);
4129 struct bpf_func_state *state = func(env, reg);
4130 int err, min_off, max_off, i, j, slot, spi;
4131 char *err_extra = type == ACCESS_HELPER ? " indirect" : "";
4132 enum bpf_access_type bounds_check_type;
4133 /* Some accesses can write anything into the stack, others are
4136 bool clobber = false;
4138 if (access_size == 0 && !zero_size_allowed) {
4139 verbose(env, "invalid zero-sized read\n");
4143 if (type == ACCESS_HELPER) {
4144 /* The bounds checks for writes are more permissive than for
4145 * reads. However, if raw_mode is not set, we'll do extra
4148 bounds_check_type = BPF_WRITE;
4151 bounds_check_type = BPF_READ;
4153 err = check_stack_access_within_bounds(env, regno, off, access_size,
4154 type, bounds_check_type);
4159 if (tnum_is_const(reg->var_off)) {
4160 min_off = max_off = reg->var_off.value + off;
4162 /* Variable offset is prohibited for unprivileged mode for
4163 * simplicity since it requires corresponding support in
4164 * Spectre masking for stack ALU.
4165 * See also retrieve_ptr_limit().
4167 if (!env->bypass_spec_v1) {
4170 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4171 verbose(env, "R%d%s variable offset stack access prohibited for !root, var_off=%s\n",
4172 regno, err_extra, tn_buf);
4175 /* Only initialized buffer on stack is allowed to be accessed
4176 * with variable offset. With uninitialized buffer it's hard to
4177 * guarantee that whole memory is marked as initialized on
4178 * helper return since specific bounds are unknown what may
4179 * cause uninitialized stack leaking.
4181 if (meta && meta->raw_mode)
4184 min_off = reg->smin_value + off;
4185 max_off = reg->smax_value + off;
4188 if (meta && meta->raw_mode) {
4189 meta->access_size = access_size;
4190 meta->regno = regno;
4194 for (i = min_off; i < max_off + access_size; i++) {
4198 spi = slot / BPF_REG_SIZE;
4199 if (state->allocated_stack <= slot)
4201 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
4202 if (*stype == STACK_MISC)
4204 if (*stype == STACK_ZERO) {
4206 /* helper can write anything into the stack */
4207 *stype = STACK_MISC;
4212 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
4213 state->stack[spi].spilled_ptr.type == PTR_TO_BTF_ID)
4216 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
4217 (state->stack[spi].spilled_ptr.type == SCALAR_VALUE ||
4218 env->allow_ptr_leaks)) {
4220 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
4221 for (j = 0; j < BPF_REG_SIZE; j++)
4222 state->stack[spi].slot_type[j] = STACK_MISC;
4228 if (tnum_is_const(reg->var_off)) {
4229 verbose(env, "invalid%s read from stack R%d off %d+%d size %d\n",
4230 err_extra, regno, min_off, i - min_off, access_size);
4234 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4235 verbose(env, "invalid%s read from stack R%d var_off %s+%d size %d\n",
4236 err_extra, regno, tn_buf, i - min_off, access_size);
4240 /* reading any byte out of 8-byte 'spill_slot' will cause
4241 * the whole slot to be marked as 'read'
4243 mark_reg_read(env, &state->stack[spi].spilled_ptr,
4244 state->stack[spi].spilled_ptr.parent,
4247 return update_stack_depth(env, state, min_off);
4250 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
4251 int access_size, bool zero_size_allowed,
4252 struct bpf_call_arg_meta *meta)
4254 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4256 switch (reg->type) {
4258 case PTR_TO_PACKET_META:
4259 return check_packet_access(env, regno, reg->off, access_size,
4261 case PTR_TO_MAP_VALUE:
4262 if (check_map_access_type(env, regno, reg->off, access_size,
4263 meta && meta->raw_mode ? BPF_WRITE :
4266 return check_map_access(env, regno, reg->off, access_size,
4269 return check_mem_region_access(env, regno, reg->off,
4270 access_size, reg->mem_size,
4272 case PTR_TO_RDONLY_BUF:
4273 if (meta && meta->raw_mode)
4275 return check_buffer_access(env, reg, regno, reg->off,
4276 access_size, zero_size_allowed,
4278 &env->prog->aux->max_rdonly_access);
4279 case PTR_TO_RDWR_BUF:
4280 return check_buffer_access(env, reg, regno, reg->off,
4281 access_size, zero_size_allowed,
4283 &env->prog->aux->max_rdwr_access);
4285 return check_stack_range_initialized(
4287 regno, reg->off, access_size,
4288 zero_size_allowed, ACCESS_HELPER, meta);
4289 default: /* scalar_value or invalid ptr */
4290 /* Allow zero-byte read from NULL, regardless of pointer type */
4291 if (zero_size_allowed && access_size == 0 &&
4292 register_is_null(reg))
4295 verbose(env, "R%d type=%s expected=%s\n", regno,
4296 reg_type_str[reg->type],
4297 reg_type_str[PTR_TO_STACK]);
4302 int check_mem_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
4303 u32 regno, u32 mem_size)
4305 if (register_is_null(reg))
4308 if (reg_type_may_be_null(reg->type)) {
4309 /* Assuming that the register contains a value check if the memory
4310 * access is safe. Temporarily save and restore the register's state as
4311 * the conversion shouldn't be visible to a caller.
4313 const struct bpf_reg_state saved_reg = *reg;
4316 mark_ptr_not_null_reg(reg);
4317 rv = check_helper_mem_access(env, regno, mem_size, true, NULL);
4322 return check_helper_mem_access(env, regno, mem_size, true, NULL);
4325 /* Implementation details:
4326 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL
4327 * Two bpf_map_lookups (even with the same key) will have different reg->id.
4328 * For traditional PTR_TO_MAP_VALUE the verifier clears reg->id after
4329 * value_or_null->value transition, since the verifier only cares about
4330 * the range of access to valid map value pointer and doesn't care about actual
4331 * address of the map element.
4332 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
4333 * reg->id > 0 after value_or_null->value transition. By doing so
4334 * two bpf_map_lookups will be considered two different pointers that
4335 * point to different bpf_spin_locks.
4336 * The verifier allows taking only one bpf_spin_lock at a time to avoid
4338 * Since only one bpf_spin_lock is allowed the checks are simpler than
4339 * reg_is_refcounted() logic. The verifier needs to remember only
4340 * one spin_lock instead of array of acquired_refs.
4341 * cur_state->active_spin_lock remembers which map value element got locked
4342 * and clears it after bpf_spin_unlock.
4344 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
4347 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4348 struct bpf_verifier_state *cur = env->cur_state;
4349 bool is_const = tnum_is_const(reg->var_off);
4350 struct bpf_map *map = reg->map_ptr;
4351 u64 val = reg->var_off.value;
4355 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
4361 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
4365 if (!map_value_has_spin_lock(map)) {
4366 if (map->spin_lock_off == -E2BIG)
4368 "map '%s' has more than one 'struct bpf_spin_lock'\n",
4370 else if (map->spin_lock_off == -ENOENT)
4372 "map '%s' doesn't have 'struct bpf_spin_lock'\n",
4376 "map '%s' is not a struct type or bpf_spin_lock is mangled\n",
4380 if (map->spin_lock_off != val + reg->off) {
4381 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock'\n",
4386 if (cur->active_spin_lock) {
4388 "Locking two bpf_spin_locks are not allowed\n");
4391 cur->active_spin_lock = reg->id;
4393 if (!cur->active_spin_lock) {
4394 verbose(env, "bpf_spin_unlock without taking a lock\n");
4397 if (cur->active_spin_lock != reg->id) {
4398 verbose(env, "bpf_spin_unlock of different lock\n");
4401 cur->active_spin_lock = 0;
4406 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
4408 return type == ARG_PTR_TO_MEM ||
4409 type == ARG_PTR_TO_MEM_OR_NULL ||
4410 type == ARG_PTR_TO_UNINIT_MEM;
4413 static bool arg_type_is_mem_size(enum bpf_arg_type type)
4415 return type == ARG_CONST_SIZE ||
4416 type == ARG_CONST_SIZE_OR_ZERO;
4419 static bool arg_type_is_alloc_size(enum bpf_arg_type type)
4421 return type == ARG_CONST_ALLOC_SIZE_OR_ZERO;
4424 static bool arg_type_is_int_ptr(enum bpf_arg_type type)
4426 return type == ARG_PTR_TO_INT ||
4427 type == ARG_PTR_TO_LONG;
4430 static int int_ptr_type_to_size(enum bpf_arg_type type)
4432 if (type == ARG_PTR_TO_INT)
4434 else if (type == ARG_PTR_TO_LONG)
4440 static int resolve_map_arg_type(struct bpf_verifier_env *env,
4441 const struct bpf_call_arg_meta *meta,
4442 enum bpf_arg_type *arg_type)
4444 if (!meta->map_ptr) {
4445 /* kernel subsystem misconfigured verifier */
4446 verbose(env, "invalid map_ptr to access map->type\n");
4450 switch (meta->map_ptr->map_type) {
4451 case BPF_MAP_TYPE_SOCKMAP:
4452 case BPF_MAP_TYPE_SOCKHASH:
4453 if (*arg_type == ARG_PTR_TO_MAP_VALUE) {
4454 *arg_type = ARG_PTR_TO_BTF_ID_SOCK_COMMON;
4456 verbose(env, "invalid arg_type for sockmap/sockhash\n");
4467 struct bpf_reg_types {
4468 const enum bpf_reg_type types[10];
4472 static const struct bpf_reg_types map_key_value_types = {
4481 static const struct bpf_reg_types sock_types = {
4491 static const struct bpf_reg_types btf_id_sock_common_types = {
4499 .btf_id = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
4503 static const struct bpf_reg_types mem_types = {
4515 static const struct bpf_reg_types int_ptr_types = {
4524 static const struct bpf_reg_types fullsock_types = { .types = { PTR_TO_SOCKET } };
4525 static const struct bpf_reg_types scalar_types = { .types = { SCALAR_VALUE } };
4526 static const struct bpf_reg_types context_types = { .types = { PTR_TO_CTX } };
4527 static const struct bpf_reg_types alloc_mem_types = { .types = { PTR_TO_MEM } };
4528 static const struct bpf_reg_types const_map_ptr_types = { .types = { CONST_PTR_TO_MAP } };
4529 static const struct bpf_reg_types btf_ptr_types = { .types = { PTR_TO_BTF_ID } };
4530 static const struct bpf_reg_types spin_lock_types = { .types = { PTR_TO_MAP_VALUE } };
4531 static const struct bpf_reg_types percpu_btf_ptr_types = { .types = { PTR_TO_PERCPU_BTF_ID } };
4533 static const struct bpf_reg_types *compatible_reg_types[__BPF_ARG_TYPE_MAX] = {
4534 [ARG_PTR_TO_MAP_KEY] = &map_key_value_types,
4535 [ARG_PTR_TO_MAP_VALUE] = &map_key_value_types,
4536 [ARG_PTR_TO_UNINIT_MAP_VALUE] = &map_key_value_types,
4537 [ARG_PTR_TO_MAP_VALUE_OR_NULL] = &map_key_value_types,
4538 [ARG_CONST_SIZE] = &scalar_types,
4539 [ARG_CONST_SIZE_OR_ZERO] = &scalar_types,
4540 [ARG_CONST_ALLOC_SIZE_OR_ZERO] = &scalar_types,
4541 [ARG_CONST_MAP_PTR] = &const_map_ptr_types,
4542 [ARG_PTR_TO_CTX] = &context_types,
4543 [ARG_PTR_TO_CTX_OR_NULL] = &context_types,
4544 [ARG_PTR_TO_SOCK_COMMON] = &sock_types,
4546 [ARG_PTR_TO_BTF_ID_SOCK_COMMON] = &btf_id_sock_common_types,
4548 [ARG_PTR_TO_SOCKET] = &fullsock_types,
4549 [ARG_PTR_TO_SOCKET_OR_NULL] = &fullsock_types,
4550 [ARG_PTR_TO_BTF_ID] = &btf_ptr_types,
4551 [ARG_PTR_TO_SPIN_LOCK] = &spin_lock_types,
4552 [ARG_PTR_TO_MEM] = &mem_types,
4553 [ARG_PTR_TO_MEM_OR_NULL] = &mem_types,
4554 [ARG_PTR_TO_UNINIT_MEM] = &mem_types,
4555 [ARG_PTR_TO_ALLOC_MEM] = &alloc_mem_types,
4556 [ARG_PTR_TO_ALLOC_MEM_OR_NULL] = &alloc_mem_types,
4557 [ARG_PTR_TO_INT] = &int_ptr_types,
4558 [ARG_PTR_TO_LONG] = &int_ptr_types,
4559 [ARG_PTR_TO_PERCPU_BTF_ID] = &percpu_btf_ptr_types,
4562 static int check_reg_type(struct bpf_verifier_env *env, u32 regno,
4563 enum bpf_arg_type arg_type,
4564 const u32 *arg_btf_id)
4566 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4567 enum bpf_reg_type expected, type = reg->type;
4568 const struct bpf_reg_types *compatible;
4571 compatible = compatible_reg_types[arg_type];
4573 verbose(env, "verifier internal error: unsupported arg type %d\n", arg_type);
4577 for (i = 0; i < ARRAY_SIZE(compatible->types); i++) {
4578 expected = compatible->types[i];
4579 if (expected == NOT_INIT)
4582 if (type == expected)
4586 verbose(env, "R%d type=%s expected=", regno, reg_type_str[type]);
4587 for (j = 0; j + 1 < i; j++)
4588 verbose(env, "%s, ", reg_type_str[compatible->types[j]]);
4589 verbose(env, "%s\n", reg_type_str[compatible->types[j]]);
4593 if (type == PTR_TO_BTF_ID) {
4595 if (!compatible->btf_id) {
4596 verbose(env, "verifier internal error: missing arg compatible BTF ID\n");
4599 arg_btf_id = compatible->btf_id;
4602 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
4603 btf_vmlinux, *arg_btf_id)) {
4604 verbose(env, "R%d is of type %s but %s is expected\n",
4605 regno, kernel_type_name(reg->btf, reg->btf_id),
4606 kernel_type_name(btf_vmlinux, *arg_btf_id));
4610 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
4611 verbose(env, "R%d is a pointer to in-kernel struct with non-zero offset\n",
4620 static int check_func_arg(struct bpf_verifier_env *env, u32 arg,
4621 struct bpf_call_arg_meta *meta,
4622 const struct bpf_func_proto *fn)
4624 u32 regno = BPF_REG_1 + arg;
4625 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4626 enum bpf_arg_type arg_type = fn->arg_type[arg];
4627 enum bpf_reg_type type = reg->type;
4630 if (arg_type == ARG_DONTCARE)
4633 err = check_reg_arg(env, regno, SRC_OP);
4637 if (arg_type == ARG_ANYTHING) {
4638 if (is_pointer_value(env, regno)) {
4639 verbose(env, "R%d leaks addr into helper function\n",
4646 if (type_is_pkt_pointer(type) &&
4647 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
4648 verbose(env, "helper access to the packet is not allowed\n");
4652 if (arg_type == ARG_PTR_TO_MAP_VALUE ||
4653 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE ||
4654 arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL) {
4655 err = resolve_map_arg_type(env, meta, &arg_type);
4660 if (register_is_null(reg) && arg_type_may_be_null(arg_type))
4661 /* A NULL register has a SCALAR_VALUE type, so skip
4664 goto skip_type_check;
4666 err = check_reg_type(env, regno, arg_type, fn->arg_btf_id[arg]);
4670 if (type == PTR_TO_CTX) {
4671 err = check_ctx_reg(env, reg, regno);
4677 if (reg->ref_obj_id) {
4678 if (meta->ref_obj_id) {
4679 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
4680 regno, reg->ref_obj_id,
4684 meta->ref_obj_id = reg->ref_obj_id;
4687 if (arg_type == ARG_CONST_MAP_PTR) {
4688 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
4689 meta->map_ptr = reg->map_ptr;
4690 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
4691 /* bpf_map_xxx(..., map_ptr, ..., key) call:
4692 * check that [key, key + map->key_size) are within
4693 * stack limits and initialized
4695 if (!meta->map_ptr) {
4696 /* in function declaration map_ptr must come before
4697 * map_key, so that it's verified and known before
4698 * we have to check map_key here. Otherwise it means
4699 * that kernel subsystem misconfigured verifier
4701 verbose(env, "invalid map_ptr to access map->key\n");
4704 err = check_helper_mem_access(env, regno,
4705 meta->map_ptr->key_size, false,
4707 } else if (arg_type == ARG_PTR_TO_MAP_VALUE ||
4708 (arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL &&
4709 !register_is_null(reg)) ||
4710 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
4711 /* bpf_map_xxx(..., map_ptr, ..., value) call:
4712 * check [value, value + map->value_size) validity
4714 if (!meta->map_ptr) {
4715 /* kernel subsystem misconfigured verifier */
4716 verbose(env, "invalid map_ptr to access map->value\n");
4719 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE);
4720 err = check_helper_mem_access(env, regno,
4721 meta->map_ptr->value_size, false,
4723 } else if (arg_type == ARG_PTR_TO_PERCPU_BTF_ID) {
4725 verbose(env, "Helper has invalid btf_id in R%d\n", regno);
4728 meta->ret_btf = reg->btf;
4729 meta->ret_btf_id = reg->btf_id;
4730 } else if (arg_type == ARG_PTR_TO_SPIN_LOCK) {
4731 if (meta->func_id == BPF_FUNC_spin_lock) {
4732 if (process_spin_lock(env, regno, true))
4734 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
4735 if (process_spin_lock(env, regno, false))
4738 verbose(env, "verifier internal error\n");
4741 } else if (arg_type_is_mem_ptr(arg_type)) {
4742 /* The access to this pointer is only checked when we hit the
4743 * next is_mem_size argument below.
4745 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MEM);
4746 } else if (arg_type_is_mem_size(arg_type)) {
4747 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
4749 /* This is used to refine r0 return value bounds for helpers
4750 * that enforce this value as an upper bound on return values.
4751 * See do_refine_retval_range() for helpers that can refine
4752 * the return value. C type of helper is u32 so we pull register
4753 * bound from umax_value however, if negative verifier errors
4754 * out. Only upper bounds can be learned because retval is an
4755 * int type and negative retvals are allowed.
4757 meta->msize_max_value = reg->umax_value;
4759 /* The register is SCALAR_VALUE; the access check
4760 * happens using its boundaries.
4762 if (!tnum_is_const(reg->var_off))
4763 /* For unprivileged variable accesses, disable raw
4764 * mode so that the program is required to
4765 * initialize all the memory that the helper could
4766 * just partially fill up.
4770 if (reg->smin_value < 0) {
4771 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
4776 if (reg->umin_value == 0) {
4777 err = check_helper_mem_access(env, regno - 1, 0,
4784 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
4785 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
4789 err = check_helper_mem_access(env, regno - 1,
4791 zero_size_allowed, meta);
4793 err = mark_chain_precision(env, regno);
4794 } else if (arg_type_is_alloc_size(arg_type)) {
4795 if (!tnum_is_const(reg->var_off)) {
4796 verbose(env, "R%d is not a known constant'\n",
4800 meta->mem_size = reg->var_off.value;
4801 } else if (arg_type_is_int_ptr(arg_type)) {
4802 int size = int_ptr_type_to_size(arg_type);
4804 err = check_helper_mem_access(env, regno, size, false, meta);
4807 err = check_ptr_alignment(env, reg, 0, size, true);
4813 static bool may_update_sockmap(struct bpf_verifier_env *env, int func_id)
4815 enum bpf_attach_type eatype = env->prog->expected_attach_type;
4816 enum bpf_prog_type type = resolve_prog_type(env->prog);
4818 if (func_id != BPF_FUNC_map_update_elem)
4821 /* It's not possible to get access to a locked struct sock in these
4822 * contexts, so updating is safe.
4825 case BPF_PROG_TYPE_TRACING:
4826 if (eatype == BPF_TRACE_ITER)
4829 case BPF_PROG_TYPE_SOCKET_FILTER:
4830 case BPF_PROG_TYPE_SCHED_CLS:
4831 case BPF_PROG_TYPE_SCHED_ACT:
4832 case BPF_PROG_TYPE_XDP:
4833 case BPF_PROG_TYPE_SK_REUSEPORT:
4834 case BPF_PROG_TYPE_FLOW_DISSECTOR:
4835 case BPF_PROG_TYPE_SK_LOOKUP:
4841 verbose(env, "cannot update sockmap in this context\n");
4845 static bool allow_tail_call_in_subprogs(struct bpf_verifier_env *env)
4847 return env->prog->jit_requested && IS_ENABLED(CONFIG_X86_64);
4850 static int check_map_func_compatibility(struct bpf_verifier_env *env,
4851 struct bpf_map *map, int func_id)
4856 /* We need a two way check, first is from map perspective ... */
4857 switch (map->map_type) {
4858 case BPF_MAP_TYPE_PROG_ARRAY:
4859 if (func_id != BPF_FUNC_tail_call)
4862 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
4863 if (func_id != BPF_FUNC_perf_event_read &&
4864 func_id != BPF_FUNC_perf_event_output &&
4865 func_id != BPF_FUNC_skb_output &&
4866 func_id != BPF_FUNC_perf_event_read_value &&
4867 func_id != BPF_FUNC_xdp_output)
4870 case BPF_MAP_TYPE_RINGBUF:
4871 if (func_id != BPF_FUNC_ringbuf_output &&
4872 func_id != BPF_FUNC_ringbuf_reserve &&
4873 func_id != BPF_FUNC_ringbuf_submit &&
4874 func_id != BPF_FUNC_ringbuf_discard &&
4875 func_id != BPF_FUNC_ringbuf_query)
4878 case BPF_MAP_TYPE_STACK_TRACE:
4879 if (func_id != BPF_FUNC_get_stackid)
4882 case BPF_MAP_TYPE_CGROUP_ARRAY:
4883 if (func_id != BPF_FUNC_skb_under_cgroup &&
4884 func_id != BPF_FUNC_current_task_under_cgroup)
4887 case BPF_MAP_TYPE_CGROUP_STORAGE:
4888 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
4889 if (func_id != BPF_FUNC_get_local_storage)
4892 case BPF_MAP_TYPE_DEVMAP:
4893 case BPF_MAP_TYPE_DEVMAP_HASH:
4894 if (func_id != BPF_FUNC_redirect_map &&
4895 func_id != BPF_FUNC_map_lookup_elem)
4898 /* Restrict bpf side of cpumap and xskmap, open when use-cases
4901 case BPF_MAP_TYPE_CPUMAP:
4902 if (func_id != BPF_FUNC_redirect_map)
4905 case BPF_MAP_TYPE_XSKMAP:
4906 if (func_id != BPF_FUNC_redirect_map &&
4907 func_id != BPF_FUNC_map_lookup_elem)
4910 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
4911 case BPF_MAP_TYPE_HASH_OF_MAPS:
4912 if (func_id != BPF_FUNC_map_lookup_elem)
4915 case BPF_MAP_TYPE_SOCKMAP:
4916 if (func_id != BPF_FUNC_sk_redirect_map &&
4917 func_id != BPF_FUNC_sock_map_update &&
4918 func_id != BPF_FUNC_map_delete_elem &&
4919 func_id != BPF_FUNC_msg_redirect_map &&
4920 func_id != BPF_FUNC_sk_select_reuseport &&
4921 func_id != BPF_FUNC_map_lookup_elem &&
4922 !may_update_sockmap(env, func_id))
4925 case BPF_MAP_TYPE_SOCKHASH:
4926 if (func_id != BPF_FUNC_sk_redirect_hash &&
4927 func_id != BPF_FUNC_sock_hash_update &&
4928 func_id != BPF_FUNC_map_delete_elem &&
4929 func_id != BPF_FUNC_msg_redirect_hash &&
4930 func_id != BPF_FUNC_sk_select_reuseport &&
4931 func_id != BPF_FUNC_map_lookup_elem &&
4932 !may_update_sockmap(env, func_id))
4935 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
4936 if (func_id != BPF_FUNC_sk_select_reuseport)
4939 case BPF_MAP_TYPE_QUEUE:
4940 case BPF_MAP_TYPE_STACK:
4941 if (func_id != BPF_FUNC_map_peek_elem &&
4942 func_id != BPF_FUNC_map_pop_elem &&
4943 func_id != BPF_FUNC_map_push_elem)
4946 case BPF_MAP_TYPE_SK_STORAGE:
4947 if (func_id != BPF_FUNC_sk_storage_get &&
4948 func_id != BPF_FUNC_sk_storage_delete)
4951 case BPF_MAP_TYPE_INODE_STORAGE:
4952 if (func_id != BPF_FUNC_inode_storage_get &&
4953 func_id != BPF_FUNC_inode_storage_delete)
4956 case BPF_MAP_TYPE_TASK_STORAGE:
4957 if (func_id != BPF_FUNC_task_storage_get &&
4958 func_id != BPF_FUNC_task_storage_delete)
4965 /* ... and second from the function itself. */
4967 case BPF_FUNC_tail_call:
4968 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
4970 if (env->subprog_cnt > 1 && !allow_tail_call_in_subprogs(env)) {
4971 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
4975 case BPF_FUNC_perf_event_read:
4976 case BPF_FUNC_perf_event_output:
4977 case BPF_FUNC_perf_event_read_value:
4978 case BPF_FUNC_skb_output:
4979 case BPF_FUNC_xdp_output:
4980 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
4983 case BPF_FUNC_get_stackid:
4984 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
4987 case BPF_FUNC_current_task_under_cgroup:
4988 case BPF_FUNC_skb_under_cgroup:
4989 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
4992 case BPF_FUNC_redirect_map:
4993 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
4994 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
4995 map->map_type != BPF_MAP_TYPE_CPUMAP &&
4996 map->map_type != BPF_MAP_TYPE_XSKMAP)
4999 case BPF_FUNC_sk_redirect_map:
5000 case BPF_FUNC_msg_redirect_map:
5001 case BPF_FUNC_sock_map_update:
5002 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
5005 case BPF_FUNC_sk_redirect_hash:
5006 case BPF_FUNC_msg_redirect_hash:
5007 case BPF_FUNC_sock_hash_update:
5008 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
5011 case BPF_FUNC_get_local_storage:
5012 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
5013 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
5016 case BPF_FUNC_sk_select_reuseport:
5017 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
5018 map->map_type != BPF_MAP_TYPE_SOCKMAP &&
5019 map->map_type != BPF_MAP_TYPE_SOCKHASH)
5022 case BPF_FUNC_map_peek_elem:
5023 case BPF_FUNC_map_pop_elem:
5024 case BPF_FUNC_map_push_elem:
5025 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
5026 map->map_type != BPF_MAP_TYPE_STACK)
5029 case BPF_FUNC_sk_storage_get:
5030 case BPF_FUNC_sk_storage_delete:
5031 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
5034 case BPF_FUNC_inode_storage_get:
5035 case BPF_FUNC_inode_storage_delete:
5036 if (map->map_type != BPF_MAP_TYPE_INODE_STORAGE)
5039 case BPF_FUNC_task_storage_get:
5040 case BPF_FUNC_task_storage_delete:
5041 if (map->map_type != BPF_MAP_TYPE_TASK_STORAGE)
5050 verbose(env, "cannot pass map_type %d into func %s#%d\n",
5051 map->map_type, func_id_name(func_id), func_id);
5055 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
5059 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
5061 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
5063 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
5065 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
5067 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
5070 /* We only support one arg being in raw mode at the moment,
5071 * which is sufficient for the helper functions we have
5077 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
5078 enum bpf_arg_type arg_next)
5080 return (arg_type_is_mem_ptr(arg_curr) &&
5081 !arg_type_is_mem_size(arg_next)) ||
5082 (!arg_type_is_mem_ptr(arg_curr) &&
5083 arg_type_is_mem_size(arg_next));
5086 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
5088 /* bpf_xxx(..., buf, len) call will access 'len'
5089 * bytes from memory 'buf'. Both arg types need
5090 * to be paired, so make sure there's no buggy
5091 * helper function specification.
5093 if (arg_type_is_mem_size(fn->arg1_type) ||
5094 arg_type_is_mem_ptr(fn->arg5_type) ||
5095 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
5096 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
5097 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
5098 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
5104 static bool check_refcount_ok(const struct bpf_func_proto *fn, int func_id)
5108 if (arg_type_may_be_refcounted(fn->arg1_type))
5110 if (arg_type_may_be_refcounted(fn->arg2_type))
5112 if (arg_type_may_be_refcounted(fn->arg3_type))
5114 if (arg_type_may_be_refcounted(fn->arg4_type))
5116 if (arg_type_may_be_refcounted(fn->arg5_type))
5119 /* A reference acquiring function cannot acquire
5120 * another refcounted ptr.
5122 if (may_be_acquire_function(func_id) && count)
5125 /* We only support one arg being unreferenced at the moment,
5126 * which is sufficient for the helper functions we have right now.
5131 static bool check_btf_id_ok(const struct bpf_func_proto *fn)
5135 for (i = 0; i < ARRAY_SIZE(fn->arg_type); i++) {
5136 if (fn->arg_type[i] == ARG_PTR_TO_BTF_ID && !fn->arg_btf_id[i])
5139 if (fn->arg_type[i] != ARG_PTR_TO_BTF_ID && fn->arg_btf_id[i])
5146 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
5148 return check_raw_mode_ok(fn) &&
5149 check_arg_pair_ok(fn) &&
5150 check_btf_id_ok(fn) &&
5151 check_refcount_ok(fn, func_id) ? 0 : -EINVAL;
5154 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
5155 * are now invalid, so turn them into unknown SCALAR_VALUE.
5157 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
5158 struct bpf_func_state *state)
5160 struct bpf_reg_state *regs = state->regs, *reg;
5163 for (i = 0; i < MAX_BPF_REG; i++)
5164 if (reg_is_pkt_pointer_any(®s[i]))
5165 mark_reg_unknown(env, regs, i);
5167 bpf_for_each_spilled_reg(i, state, reg) {
5170 if (reg_is_pkt_pointer_any(reg))
5171 __mark_reg_unknown(env, reg);
5175 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
5177 struct bpf_verifier_state *vstate = env->cur_state;
5180 for (i = 0; i <= vstate->curframe; i++)
5181 __clear_all_pkt_pointers(env, vstate->frame[i]);
5186 BEYOND_PKT_END = -2,
5189 static void mark_pkt_end(struct bpf_verifier_state *vstate, int regn, bool range_open)
5191 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5192 struct bpf_reg_state *reg = &state->regs[regn];
5194 if (reg->type != PTR_TO_PACKET)
5195 /* PTR_TO_PACKET_META is not supported yet */
5198 /* The 'reg' is pkt > pkt_end or pkt >= pkt_end.
5199 * How far beyond pkt_end it goes is unknown.
5200 * if (!range_open) it's the case of pkt >= pkt_end
5201 * if (range_open) it's the case of pkt > pkt_end
5202 * hence this pointer is at least 1 byte bigger than pkt_end
5205 reg->range = BEYOND_PKT_END;
5207 reg->range = AT_PKT_END;
5210 static void release_reg_references(struct bpf_verifier_env *env,
5211 struct bpf_func_state *state,
5214 struct bpf_reg_state *regs = state->regs, *reg;
5217 for (i = 0; i < MAX_BPF_REG; i++)
5218 if (regs[i].ref_obj_id == ref_obj_id)
5219 mark_reg_unknown(env, regs, i);
5221 bpf_for_each_spilled_reg(i, state, reg) {
5224 if (reg->ref_obj_id == ref_obj_id)
5225 __mark_reg_unknown(env, reg);
5229 /* The pointer with the specified id has released its reference to kernel
5230 * resources. Identify all copies of the same pointer and clear the reference.
5232 static int release_reference(struct bpf_verifier_env *env,
5235 struct bpf_verifier_state *vstate = env->cur_state;
5239 err = release_reference_state(cur_func(env), ref_obj_id);
5243 for (i = 0; i <= vstate->curframe; i++)
5244 release_reg_references(env, vstate->frame[i], ref_obj_id);
5249 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
5250 struct bpf_reg_state *regs)
5254 /* after the call registers r0 - r5 were scratched */
5255 for (i = 0; i < CALLER_SAVED_REGS; i++) {
5256 mark_reg_not_init(env, regs, caller_saved[i]);
5257 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
5261 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
5264 struct bpf_verifier_state *state = env->cur_state;
5265 struct bpf_func_info_aux *func_info_aux;
5266 struct bpf_func_state *caller, *callee;
5267 int i, err, subprog, target_insn;
5268 bool is_global = false;
5270 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
5271 verbose(env, "the call stack of %d frames is too deep\n",
5272 state->curframe + 2);
5276 target_insn = *insn_idx + insn->imm;
5277 subprog = find_subprog(env, target_insn + 1);
5279 verbose(env, "verifier bug. No program starts at insn %d\n",
5284 caller = state->frame[state->curframe];
5285 if (state->frame[state->curframe + 1]) {
5286 verbose(env, "verifier bug. Frame %d already allocated\n",
5287 state->curframe + 1);
5291 func_info_aux = env->prog->aux->func_info_aux;
5293 is_global = func_info_aux[subprog].linkage == BTF_FUNC_GLOBAL;
5294 err = btf_check_func_arg_match(env, subprog, caller->regs);
5299 verbose(env, "Caller passes invalid args into func#%d\n",
5303 if (env->log.level & BPF_LOG_LEVEL)
5305 "Func#%d is global and valid. Skipping.\n",
5307 clear_caller_saved_regs(env, caller->regs);
5309 /* All global functions return a 64-bit SCALAR_VALUE */
5310 mark_reg_unknown(env, caller->regs, BPF_REG_0);
5311 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
5313 /* continue with next insn after call */
5318 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
5321 state->frame[state->curframe + 1] = callee;
5323 /* callee cannot access r0, r6 - r9 for reading and has to write
5324 * into its own stack before reading from it.
5325 * callee can read/write into caller's stack
5327 init_func_state(env, callee,
5328 /* remember the callsite, it will be used by bpf_exit */
5329 *insn_idx /* callsite */,
5330 state->curframe + 1 /* frameno within this callchain */,
5331 subprog /* subprog number within this prog */);
5333 /* Transfer references to the callee */
5334 err = transfer_reference_state(callee, caller);
5338 /* copy r1 - r5 args that callee can access. The copy includes parent
5339 * pointers, which connects us up to the liveness chain
5341 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
5342 callee->regs[i] = caller->regs[i];
5344 clear_caller_saved_regs(env, caller->regs);
5346 /* only increment it after check_reg_arg() finished */
5349 /* and go analyze first insn of the callee */
5350 *insn_idx = target_insn;
5352 if (env->log.level & BPF_LOG_LEVEL) {
5353 verbose(env, "caller:\n");
5354 print_verifier_state(env, caller);
5355 verbose(env, "callee:\n");
5356 print_verifier_state(env, callee);
5361 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
5363 struct bpf_verifier_state *state = env->cur_state;
5364 struct bpf_func_state *caller, *callee;
5365 struct bpf_reg_state *r0;
5368 callee = state->frame[state->curframe];
5369 r0 = &callee->regs[BPF_REG_0];
5370 if (r0->type == PTR_TO_STACK) {
5371 /* technically it's ok to return caller's stack pointer
5372 * (or caller's caller's pointer) back to the caller,
5373 * since these pointers are valid. Only current stack
5374 * pointer will be invalid as soon as function exits,
5375 * but let's be conservative
5377 verbose(env, "cannot return stack pointer to the caller\n");
5382 caller = state->frame[state->curframe];
5383 /* return to the caller whatever r0 had in the callee */
5384 caller->regs[BPF_REG_0] = *r0;
5386 /* Transfer references to the caller */
5387 err = transfer_reference_state(caller, callee);
5391 *insn_idx = callee->callsite + 1;
5392 if (env->log.level & BPF_LOG_LEVEL) {
5393 verbose(env, "returning from callee:\n");
5394 print_verifier_state(env, callee);
5395 verbose(env, "to caller at %d:\n", *insn_idx);
5396 print_verifier_state(env, caller);
5398 /* clear everything in the callee */
5399 free_func_state(callee);
5400 state->frame[state->curframe + 1] = NULL;
5404 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
5406 struct bpf_call_arg_meta *meta)
5408 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
5410 if (ret_type != RET_INTEGER ||
5411 (func_id != BPF_FUNC_get_stack &&
5412 func_id != BPF_FUNC_probe_read_str &&
5413 func_id != BPF_FUNC_probe_read_kernel_str &&
5414 func_id != BPF_FUNC_probe_read_user_str))
5417 ret_reg->smax_value = meta->msize_max_value;
5418 ret_reg->s32_max_value = meta->msize_max_value;
5419 ret_reg->smin_value = -MAX_ERRNO;
5420 ret_reg->s32_min_value = -MAX_ERRNO;
5421 __reg_deduce_bounds(ret_reg);
5422 __reg_bound_offset(ret_reg);
5423 __update_reg_bounds(ret_reg);
5427 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
5428 int func_id, int insn_idx)
5430 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
5431 struct bpf_map *map = meta->map_ptr;
5433 if (func_id != BPF_FUNC_tail_call &&
5434 func_id != BPF_FUNC_map_lookup_elem &&
5435 func_id != BPF_FUNC_map_update_elem &&
5436 func_id != BPF_FUNC_map_delete_elem &&
5437 func_id != BPF_FUNC_map_push_elem &&
5438 func_id != BPF_FUNC_map_pop_elem &&
5439 func_id != BPF_FUNC_map_peek_elem)
5443 verbose(env, "kernel subsystem misconfigured verifier\n");
5447 /* In case of read-only, some additional restrictions
5448 * need to be applied in order to prevent altering the
5449 * state of the map from program side.
5451 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
5452 (func_id == BPF_FUNC_map_delete_elem ||
5453 func_id == BPF_FUNC_map_update_elem ||
5454 func_id == BPF_FUNC_map_push_elem ||
5455 func_id == BPF_FUNC_map_pop_elem)) {
5456 verbose(env, "write into map forbidden\n");
5460 if (!BPF_MAP_PTR(aux->map_ptr_state))
5461 bpf_map_ptr_store(aux, meta->map_ptr,
5462 !meta->map_ptr->bypass_spec_v1);
5463 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
5464 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
5465 !meta->map_ptr->bypass_spec_v1);
5470 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
5471 int func_id, int insn_idx)
5473 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
5474 struct bpf_reg_state *regs = cur_regs(env), *reg;
5475 struct bpf_map *map = meta->map_ptr;
5480 if (func_id != BPF_FUNC_tail_call)
5482 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
5483 verbose(env, "kernel subsystem misconfigured verifier\n");
5487 range = tnum_range(0, map->max_entries - 1);
5488 reg = ®s[BPF_REG_3];
5490 if (!register_is_const(reg) || !tnum_in(range, reg->var_off)) {
5491 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
5495 err = mark_chain_precision(env, BPF_REG_3);
5499 val = reg->var_off.value;
5500 if (bpf_map_key_unseen(aux))
5501 bpf_map_key_store(aux, val);
5502 else if (!bpf_map_key_poisoned(aux) &&
5503 bpf_map_key_immediate(aux) != val)
5504 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
5508 static int check_reference_leak(struct bpf_verifier_env *env)
5510 struct bpf_func_state *state = cur_func(env);
5513 for (i = 0; i < state->acquired_refs; i++) {
5514 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
5515 state->refs[i].id, state->refs[i].insn_idx);
5517 return state->acquired_refs ? -EINVAL : 0;
5520 static int check_helper_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
5522 const struct bpf_func_proto *fn = NULL;
5523 struct bpf_reg_state *regs;
5524 struct bpf_call_arg_meta meta;
5528 /* find function prototype */
5529 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
5530 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
5535 if (env->ops->get_func_proto)
5536 fn = env->ops->get_func_proto(func_id, env->prog);
5538 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
5543 /* eBPF programs must be GPL compatible to use GPL-ed functions */
5544 if (!env->prog->gpl_compatible && fn->gpl_only) {
5545 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
5549 if (fn->allowed && !fn->allowed(env->prog)) {
5550 verbose(env, "helper call is not allowed in probe\n");
5554 /* With LD_ABS/IND some JITs save/restore skb from r1. */
5555 changes_data = bpf_helper_changes_pkt_data(fn->func);
5556 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
5557 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
5558 func_id_name(func_id), func_id);
5562 memset(&meta, 0, sizeof(meta));
5563 meta.pkt_access = fn->pkt_access;
5565 err = check_func_proto(fn, func_id);
5567 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
5568 func_id_name(func_id), func_id);
5572 meta.func_id = func_id;
5574 for (i = 0; i < 5; i++) {
5575 err = check_func_arg(env, i, &meta, fn);
5580 err = record_func_map(env, &meta, func_id, insn_idx);
5584 err = record_func_key(env, &meta, func_id, insn_idx);
5588 /* Mark slots with STACK_MISC in case of raw mode, stack offset
5589 * is inferred from register state.
5591 for (i = 0; i < meta.access_size; i++) {
5592 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
5593 BPF_WRITE, -1, false);
5598 if (func_id == BPF_FUNC_tail_call) {
5599 err = check_reference_leak(env);
5601 verbose(env, "tail_call would lead to reference leak\n");
5604 } else if (is_release_function(func_id)) {
5605 err = release_reference(env, meta.ref_obj_id);
5607 verbose(env, "func %s#%d reference has not been acquired before\n",
5608 func_id_name(func_id), func_id);
5613 regs = cur_regs(env);
5615 /* check that flags argument in get_local_storage(map, flags) is 0,
5616 * this is required because get_local_storage() can't return an error.
5618 if (func_id == BPF_FUNC_get_local_storage &&
5619 !register_is_null(®s[BPF_REG_2])) {
5620 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
5624 /* reset caller saved regs */
5625 for (i = 0; i < CALLER_SAVED_REGS; i++) {
5626 mark_reg_not_init(env, regs, caller_saved[i]);
5627 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
5630 /* helper call returns 64-bit value. */
5631 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
5633 /* update return register (already marked as written above) */
5634 if (fn->ret_type == RET_INTEGER) {
5635 /* sets type to SCALAR_VALUE */
5636 mark_reg_unknown(env, regs, BPF_REG_0);
5637 } else if (fn->ret_type == RET_VOID) {
5638 regs[BPF_REG_0].type = NOT_INIT;
5639 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL ||
5640 fn->ret_type == RET_PTR_TO_MAP_VALUE) {
5641 /* There is no offset yet applied, variable or fixed */
5642 mark_reg_known_zero(env, regs, BPF_REG_0);
5643 /* remember map_ptr, so that check_map_access()
5644 * can check 'value_size' boundary of memory access
5645 * to map element returned from bpf_map_lookup_elem()
5647 if (meta.map_ptr == NULL) {
5649 "kernel subsystem misconfigured verifier\n");
5652 regs[BPF_REG_0].map_ptr = meta.map_ptr;
5653 if (fn->ret_type == RET_PTR_TO_MAP_VALUE) {
5654 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE;
5655 if (map_value_has_spin_lock(meta.map_ptr))
5656 regs[BPF_REG_0].id = ++env->id_gen;
5658 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
5660 } else if (fn->ret_type == RET_PTR_TO_SOCKET_OR_NULL) {
5661 mark_reg_known_zero(env, regs, BPF_REG_0);
5662 regs[BPF_REG_0].type = PTR_TO_SOCKET_OR_NULL;
5663 } else if (fn->ret_type == RET_PTR_TO_SOCK_COMMON_OR_NULL) {
5664 mark_reg_known_zero(env, regs, BPF_REG_0);
5665 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON_OR_NULL;
5666 } else if (fn->ret_type == RET_PTR_TO_TCP_SOCK_OR_NULL) {
5667 mark_reg_known_zero(env, regs, BPF_REG_0);
5668 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK_OR_NULL;
5669 } else if (fn->ret_type == RET_PTR_TO_ALLOC_MEM_OR_NULL) {
5670 mark_reg_known_zero(env, regs, BPF_REG_0);
5671 regs[BPF_REG_0].type = PTR_TO_MEM_OR_NULL;
5672 regs[BPF_REG_0].mem_size = meta.mem_size;
5673 } else if (fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID_OR_NULL ||
5674 fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID) {
5675 const struct btf_type *t;
5677 mark_reg_known_zero(env, regs, BPF_REG_0);
5678 t = btf_type_skip_modifiers(meta.ret_btf, meta.ret_btf_id, NULL);
5679 if (!btf_type_is_struct(t)) {
5681 const struct btf_type *ret;
5684 /* resolve the type size of ksym. */
5685 ret = btf_resolve_size(meta.ret_btf, t, &tsize);
5687 tname = btf_name_by_offset(meta.ret_btf, t->name_off);
5688 verbose(env, "unable to resolve the size of type '%s': %ld\n",
5689 tname, PTR_ERR(ret));
5692 regs[BPF_REG_0].type =
5693 fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID ?
5694 PTR_TO_MEM : PTR_TO_MEM_OR_NULL;
5695 regs[BPF_REG_0].mem_size = tsize;
5697 regs[BPF_REG_0].type =
5698 fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID ?
5699 PTR_TO_BTF_ID : PTR_TO_BTF_ID_OR_NULL;
5700 regs[BPF_REG_0].btf = meta.ret_btf;
5701 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
5703 } else if (fn->ret_type == RET_PTR_TO_BTF_ID_OR_NULL ||
5704 fn->ret_type == RET_PTR_TO_BTF_ID) {
5707 mark_reg_known_zero(env, regs, BPF_REG_0);
5708 regs[BPF_REG_0].type = fn->ret_type == RET_PTR_TO_BTF_ID ?
5710 PTR_TO_BTF_ID_OR_NULL;
5711 ret_btf_id = *fn->ret_btf_id;
5712 if (ret_btf_id == 0) {
5713 verbose(env, "invalid return type %d of func %s#%d\n",
5714 fn->ret_type, func_id_name(func_id), func_id);
5717 /* current BPF helper definitions are only coming from
5718 * built-in code with type IDs from vmlinux BTF
5720 regs[BPF_REG_0].btf = btf_vmlinux;
5721 regs[BPF_REG_0].btf_id = ret_btf_id;
5723 verbose(env, "unknown return type %d of func %s#%d\n",
5724 fn->ret_type, func_id_name(func_id), func_id);
5728 if (reg_type_may_be_null(regs[BPF_REG_0].type))
5729 regs[BPF_REG_0].id = ++env->id_gen;
5731 if (is_ptr_cast_function(func_id)) {
5732 /* For release_reference() */
5733 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
5734 } else if (is_acquire_function(func_id, meta.map_ptr)) {
5735 int id = acquire_reference_state(env, insn_idx);
5739 /* For mark_ptr_or_null_reg() */
5740 regs[BPF_REG_0].id = id;
5741 /* For release_reference() */
5742 regs[BPF_REG_0].ref_obj_id = id;
5745 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
5747 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
5751 if ((func_id == BPF_FUNC_get_stack ||
5752 func_id == BPF_FUNC_get_task_stack) &&
5753 !env->prog->has_callchain_buf) {
5754 const char *err_str;
5756 #ifdef CONFIG_PERF_EVENTS
5757 err = get_callchain_buffers(sysctl_perf_event_max_stack);
5758 err_str = "cannot get callchain buffer for func %s#%d\n";
5761 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
5764 verbose(env, err_str, func_id_name(func_id), func_id);
5768 env->prog->has_callchain_buf = true;
5771 if (func_id == BPF_FUNC_get_stackid || func_id == BPF_FUNC_get_stack)
5772 env->prog->call_get_stack = true;
5775 clear_all_pkt_pointers(env);
5779 static bool signed_add_overflows(s64 a, s64 b)
5781 /* Do the add in u64, where overflow is well-defined */
5782 s64 res = (s64)((u64)a + (u64)b);
5789 static bool signed_add32_overflows(s32 a, s32 b)
5791 /* Do the add in u32, where overflow is well-defined */
5792 s32 res = (s32)((u32)a + (u32)b);
5799 static bool signed_sub_overflows(s64 a, s64 b)
5801 /* Do the sub in u64, where overflow is well-defined */
5802 s64 res = (s64)((u64)a - (u64)b);
5809 static bool signed_sub32_overflows(s32 a, s32 b)
5811 /* Do the sub in u32, where overflow is well-defined */
5812 s32 res = (s32)((u32)a - (u32)b);
5819 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
5820 const struct bpf_reg_state *reg,
5821 enum bpf_reg_type type)
5823 bool known = tnum_is_const(reg->var_off);
5824 s64 val = reg->var_off.value;
5825 s64 smin = reg->smin_value;
5827 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
5828 verbose(env, "math between %s pointer and %lld is not allowed\n",
5829 reg_type_str[type], val);
5833 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
5834 verbose(env, "%s pointer offset %d is not allowed\n",
5835 reg_type_str[type], reg->off);
5839 if (smin == S64_MIN) {
5840 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
5841 reg_type_str[type]);
5845 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
5846 verbose(env, "value %lld makes %s pointer be out of bounds\n",
5847 smin, reg_type_str[type]);
5854 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
5856 return &env->insn_aux_data[env->insn_idx];
5867 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
5868 const struct bpf_reg_state *off_reg,
5869 u32 *alu_limit, u8 opcode)
5871 bool off_is_neg = off_reg->smin_value < 0;
5872 bool mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
5873 (opcode == BPF_SUB && !off_is_neg);
5874 u32 max = 0, ptr_limit = 0;
5876 if (!tnum_is_const(off_reg->var_off) &&
5877 (off_reg->smin_value < 0) != (off_reg->smax_value < 0))
5878 return REASON_BOUNDS;
5880 switch (ptr_reg->type) {
5882 /* Offset 0 is out-of-bounds, but acceptable start for the
5883 * left direction, see BPF_REG_FP. Also, unknown scalar
5884 * offset where we would need to deal with min/max bounds is
5885 * currently prohibited for unprivileged.
5887 max = MAX_BPF_STACK + mask_to_left;
5888 ptr_limit = -(ptr_reg->var_off.value + ptr_reg->off);
5890 case PTR_TO_MAP_VALUE:
5891 max = ptr_reg->map_ptr->value_size;
5892 ptr_limit = (mask_to_left ?
5893 ptr_reg->smin_value :
5894 ptr_reg->umax_value) + ptr_reg->off;
5900 if (ptr_limit >= max)
5901 return REASON_LIMIT;
5902 *alu_limit = ptr_limit;
5906 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
5907 const struct bpf_insn *insn)
5909 return env->bypass_spec_v1 || BPF_SRC(insn->code) == BPF_K;
5912 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
5913 u32 alu_state, u32 alu_limit)
5915 /* If we arrived here from different branches with different
5916 * state or limits to sanitize, then this won't work.
5918 if (aux->alu_state &&
5919 (aux->alu_state != alu_state ||
5920 aux->alu_limit != alu_limit))
5921 return REASON_PATHS;
5923 /* Corresponding fixup done in fixup_bpf_calls(). */
5924 aux->alu_state = alu_state;
5925 aux->alu_limit = alu_limit;
5929 static int sanitize_val_alu(struct bpf_verifier_env *env,
5930 struct bpf_insn *insn)
5932 struct bpf_insn_aux_data *aux = cur_aux(env);
5934 if (can_skip_alu_sanitation(env, insn))
5937 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
5940 static bool sanitize_needed(u8 opcode)
5942 return opcode == BPF_ADD || opcode == BPF_SUB;
5945 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
5946 struct bpf_insn *insn,
5947 const struct bpf_reg_state *ptr_reg,
5948 const struct bpf_reg_state *off_reg,
5949 struct bpf_reg_state *dst_reg,
5950 struct bpf_insn_aux_data *tmp_aux,
5951 const bool commit_window)
5953 struct bpf_insn_aux_data *aux = commit_window ? cur_aux(env) : tmp_aux;
5954 struct bpf_verifier_state *vstate = env->cur_state;
5955 bool off_is_neg = off_reg->smin_value < 0;
5956 bool ptr_is_dst_reg = ptr_reg == dst_reg;
5957 u8 opcode = BPF_OP(insn->code);
5958 u32 alu_state, alu_limit;
5959 struct bpf_reg_state tmp;
5963 if (can_skip_alu_sanitation(env, insn))
5966 /* We already marked aux for masking from non-speculative
5967 * paths, thus we got here in the first place. We only care
5968 * to explore bad access from here.
5970 if (vstate->speculative)
5973 err = retrieve_ptr_limit(ptr_reg, off_reg, &alu_limit, opcode);
5977 if (commit_window) {
5978 /* In commit phase we narrow the masking window based on
5979 * the observed pointer move after the simulated operation.
5981 alu_state = tmp_aux->alu_state;
5982 alu_limit = abs(tmp_aux->alu_limit - alu_limit);
5984 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
5985 alu_state |= ptr_is_dst_reg ?
5986 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
5989 err = update_alu_sanitation_state(aux, alu_state, alu_limit);
5993 /* If we're in commit phase, we're done here given we already
5994 * pushed the truncated dst_reg into the speculative verification
6000 /* Simulate and find potential out-of-bounds access under
6001 * speculative execution from truncation as a result of
6002 * masking when off was not within expected range. If off
6003 * sits in dst, then we temporarily need to move ptr there
6004 * to simulate dst (== 0) +/-= ptr. Needed, for example,
6005 * for cases where we use K-based arithmetic in one direction
6006 * and truncated reg-based in the other in order to explore
6009 if (!ptr_is_dst_reg) {
6011 *dst_reg = *ptr_reg;
6013 ret = push_stack(env, env->insn_idx + 1, env->insn_idx, true);
6014 if (!ptr_is_dst_reg && ret)
6016 return !ret ? REASON_STACK : 0;
6019 static int sanitize_err(struct bpf_verifier_env *env,
6020 const struct bpf_insn *insn, int reason,
6021 const struct bpf_reg_state *off_reg,
6022 const struct bpf_reg_state *dst_reg)
6024 static const char *err = "pointer arithmetic with it prohibited for !root";
6025 const char *op = BPF_OP(insn->code) == BPF_ADD ? "add" : "sub";
6026 u32 dst = insn->dst_reg, src = insn->src_reg;
6030 verbose(env, "R%d has unknown scalar with mixed signed bounds, %s\n",
6031 off_reg == dst_reg ? dst : src, err);
6034 verbose(env, "R%d has pointer with unsupported alu operation, %s\n",
6035 off_reg == dst_reg ? src : dst, err);
6038 verbose(env, "R%d tried to %s from different maps, paths or scalars, %s\n",
6042 verbose(env, "R%d tried to %s beyond pointer bounds, %s\n",
6046 verbose(env, "R%d could not be pushed for speculative verification, %s\n",
6050 verbose(env, "verifier internal error: unknown reason (%d)\n",
6058 /* check that stack access falls within stack limits and that 'reg' doesn't
6059 * have a variable offset.
6061 * Variable offset is prohibited for unprivileged mode for simplicity since it
6062 * requires corresponding support in Spectre masking for stack ALU. See also
6063 * retrieve_ptr_limit().
6066 * 'off' includes 'reg->off'.
6068 static int check_stack_access_for_ptr_arithmetic(
6069 struct bpf_verifier_env *env,
6071 const struct bpf_reg_state *reg,
6074 if (!tnum_is_const(reg->var_off)) {
6077 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6078 verbose(env, "R%d variable stack access prohibited for !root, var_off=%s off=%d\n",
6079 regno, tn_buf, off);
6083 if (off >= 0 || off < -MAX_BPF_STACK) {
6084 verbose(env, "R%d stack pointer arithmetic goes out of range, "
6085 "prohibited for !root; off=%d\n", regno, off);
6092 static int sanitize_check_bounds(struct bpf_verifier_env *env,
6093 const struct bpf_insn *insn,
6094 const struct bpf_reg_state *dst_reg)
6096 u32 dst = insn->dst_reg;
6098 /* For unprivileged we require that resulting offset must be in bounds
6099 * in order to be able to sanitize access later on.
6101 if (env->bypass_spec_v1)
6104 switch (dst_reg->type) {
6106 if (check_stack_access_for_ptr_arithmetic(env, dst, dst_reg,
6107 dst_reg->off + dst_reg->var_off.value))
6110 case PTR_TO_MAP_VALUE:
6111 if (check_map_access(env, dst, dst_reg->off, 1, false)) {
6112 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
6113 "prohibited for !root\n", dst);
6124 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
6125 * Caller should also handle BPF_MOV case separately.
6126 * If we return -EACCES, caller may want to try again treating pointer as a
6127 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
6129 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
6130 struct bpf_insn *insn,
6131 const struct bpf_reg_state *ptr_reg,
6132 const struct bpf_reg_state *off_reg)
6134 struct bpf_verifier_state *vstate = env->cur_state;
6135 struct bpf_func_state *state = vstate->frame[vstate->curframe];
6136 struct bpf_reg_state *regs = state->regs, *dst_reg;
6137 bool known = tnum_is_const(off_reg->var_off);
6138 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
6139 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
6140 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
6141 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
6142 struct bpf_insn_aux_data tmp_aux = {};
6143 u8 opcode = BPF_OP(insn->code);
6144 u32 dst = insn->dst_reg;
6147 dst_reg = ®s[dst];
6149 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
6150 smin_val > smax_val || umin_val > umax_val) {
6151 /* Taint dst register if offset had invalid bounds derived from
6152 * e.g. dead branches.
6154 __mark_reg_unknown(env, dst_reg);
6158 if (BPF_CLASS(insn->code) != BPF_ALU64) {
6159 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
6160 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
6161 __mark_reg_unknown(env, dst_reg);
6166 "R%d 32-bit pointer arithmetic prohibited\n",
6171 switch (ptr_reg->type) {
6172 case PTR_TO_MAP_VALUE_OR_NULL:
6173 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
6174 dst, reg_type_str[ptr_reg->type]);
6176 case CONST_PTR_TO_MAP:
6177 /* smin_val represents the known value */
6178 if (known && smin_val == 0 && opcode == BPF_ADD)
6181 case PTR_TO_PACKET_END:
6183 case PTR_TO_SOCKET_OR_NULL:
6184 case PTR_TO_SOCK_COMMON:
6185 case PTR_TO_SOCK_COMMON_OR_NULL:
6186 case PTR_TO_TCP_SOCK:
6187 case PTR_TO_TCP_SOCK_OR_NULL:
6188 case PTR_TO_XDP_SOCK:
6189 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
6190 dst, reg_type_str[ptr_reg->type]);
6196 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
6197 * The id may be overwritten later if we create a new variable offset.
6199 dst_reg->type = ptr_reg->type;
6200 dst_reg->id = ptr_reg->id;
6202 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
6203 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
6206 /* pointer types do not carry 32-bit bounds at the moment. */
6207 __mark_reg32_unbounded(dst_reg);
6209 if (sanitize_needed(opcode)) {
6210 ret = sanitize_ptr_alu(env, insn, ptr_reg, off_reg, dst_reg,
6213 return sanitize_err(env, insn, ret, off_reg, dst_reg);
6218 /* We can take a fixed offset as long as it doesn't overflow
6219 * the s32 'off' field
6221 if (known && (ptr_reg->off + smin_val ==
6222 (s64)(s32)(ptr_reg->off + smin_val))) {
6223 /* pointer += K. Accumulate it into fixed offset */
6224 dst_reg->smin_value = smin_ptr;
6225 dst_reg->smax_value = smax_ptr;
6226 dst_reg->umin_value = umin_ptr;
6227 dst_reg->umax_value = umax_ptr;
6228 dst_reg->var_off = ptr_reg->var_off;
6229 dst_reg->off = ptr_reg->off + smin_val;
6230 dst_reg->raw = ptr_reg->raw;
6233 /* A new variable offset is created. Note that off_reg->off
6234 * == 0, since it's a scalar.
6235 * dst_reg gets the pointer type and since some positive
6236 * integer value was added to the pointer, give it a new 'id'
6237 * if it's a PTR_TO_PACKET.
6238 * this creates a new 'base' pointer, off_reg (variable) gets
6239 * added into the variable offset, and we copy the fixed offset
6242 if (signed_add_overflows(smin_ptr, smin_val) ||
6243 signed_add_overflows(smax_ptr, smax_val)) {
6244 dst_reg->smin_value = S64_MIN;
6245 dst_reg->smax_value = S64_MAX;
6247 dst_reg->smin_value = smin_ptr + smin_val;
6248 dst_reg->smax_value = smax_ptr + smax_val;
6250 if (umin_ptr + umin_val < umin_ptr ||
6251 umax_ptr + umax_val < umax_ptr) {
6252 dst_reg->umin_value = 0;
6253 dst_reg->umax_value = U64_MAX;
6255 dst_reg->umin_value = umin_ptr + umin_val;
6256 dst_reg->umax_value = umax_ptr + umax_val;
6258 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
6259 dst_reg->off = ptr_reg->off;
6260 dst_reg->raw = ptr_reg->raw;
6261 if (reg_is_pkt_pointer(ptr_reg)) {
6262 dst_reg->id = ++env->id_gen;
6263 /* something was added to pkt_ptr, set range to zero */
6264 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
6268 if (dst_reg == off_reg) {
6269 /* scalar -= pointer. Creates an unknown scalar */
6270 verbose(env, "R%d tried to subtract pointer from scalar\n",
6274 /* We don't allow subtraction from FP, because (according to
6275 * test_verifier.c test "invalid fp arithmetic", JITs might not
6276 * be able to deal with it.
6278 if (ptr_reg->type == PTR_TO_STACK) {
6279 verbose(env, "R%d subtraction from stack pointer prohibited\n",
6283 if (known && (ptr_reg->off - smin_val ==
6284 (s64)(s32)(ptr_reg->off - smin_val))) {
6285 /* pointer -= K. Subtract it from fixed offset */
6286 dst_reg->smin_value = smin_ptr;
6287 dst_reg->smax_value = smax_ptr;
6288 dst_reg->umin_value = umin_ptr;
6289 dst_reg->umax_value = umax_ptr;
6290 dst_reg->var_off = ptr_reg->var_off;
6291 dst_reg->id = ptr_reg->id;
6292 dst_reg->off = ptr_reg->off - smin_val;
6293 dst_reg->raw = ptr_reg->raw;
6296 /* A new variable offset is created. If the subtrahend is known
6297 * nonnegative, then any reg->range we had before is still good.
6299 if (signed_sub_overflows(smin_ptr, smax_val) ||
6300 signed_sub_overflows(smax_ptr, smin_val)) {
6301 /* Overflow possible, we know nothing */
6302 dst_reg->smin_value = S64_MIN;
6303 dst_reg->smax_value = S64_MAX;
6305 dst_reg->smin_value = smin_ptr - smax_val;
6306 dst_reg->smax_value = smax_ptr - smin_val;
6308 if (umin_ptr < umax_val) {
6309 /* Overflow possible, we know nothing */
6310 dst_reg->umin_value = 0;
6311 dst_reg->umax_value = U64_MAX;
6313 /* Cannot overflow (as long as bounds are consistent) */
6314 dst_reg->umin_value = umin_ptr - umax_val;
6315 dst_reg->umax_value = umax_ptr - umin_val;
6317 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
6318 dst_reg->off = ptr_reg->off;
6319 dst_reg->raw = ptr_reg->raw;
6320 if (reg_is_pkt_pointer(ptr_reg)) {
6321 dst_reg->id = ++env->id_gen;
6322 /* something was added to pkt_ptr, set range to zero */
6324 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
6330 /* bitwise ops on pointers are troublesome, prohibit. */
6331 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
6332 dst, bpf_alu_string[opcode >> 4]);
6335 /* other operators (e.g. MUL,LSH) produce non-pointer results */
6336 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
6337 dst, bpf_alu_string[opcode >> 4]);
6341 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
6344 __update_reg_bounds(dst_reg);
6345 __reg_deduce_bounds(dst_reg);
6346 __reg_bound_offset(dst_reg);
6348 if (sanitize_check_bounds(env, insn, dst_reg) < 0)
6350 if (sanitize_needed(opcode)) {
6351 ret = sanitize_ptr_alu(env, insn, dst_reg, off_reg, dst_reg,
6354 return sanitize_err(env, insn, ret, off_reg, dst_reg);
6360 static void scalar32_min_max_add(struct bpf_reg_state *dst_reg,
6361 struct bpf_reg_state *src_reg)
6363 s32 smin_val = src_reg->s32_min_value;
6364 s32 smax_val = src_reg->s32_max_value;
6365 u32 umin_val = src_reg->u32_min_value;
6366 u32 umax_val = src_reg->u32_max_value;
6368 if (signed_add32_overflows(dst_reg->s32_min_value, smin_val) ||
6369 signed_add32_overflows(dst_reg->s32_max_value, smax_val)) {
6370 dst_reg->s32_min_value = S32_MIN;
6371 dst_reg->s32_max_value = S32_MAX;
6373 dst_reg->s32_min_value += smin_val;
6374 dst_reg->s32_max_value += smax_val;
6376 if (dst_reg->u32_min_value + umin_val < umin_val ||
6377 dst_reg->u32_max_value + umax_val < umax_val) {
6378 dst_reg->u32_min_value = 0;
6379 dst_reg->u32_max_value = U32_MAX;
6381 dst_reg->u32_min_value += umin_val;
6382 dst_reg->u32_max_value += umax_val;
6386 static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
6387 struct bpf_reg_state *src_reg)
6389 s64 smin_val = src_reg->smin_value;
6390 s64 smax_val = src_reg->smax_value;
6391 u64 umin_val = src_reg->umin_value;
6392 u64 umax_val = src_reg->umax_value;
6394 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
6395 signed_add_overflows(dst_reg->smax_value, smax_val)) {
6396 dst_reg->smin_value = S64_MIN;
6397 dst_reg->smax_value = S64_MAX;
6399 dst_reg->smin_value += smin_val;
6400 dst_reg->smax_value += smax_val;
6402 if (dst_reg->umin_value + umin_val < umin_val ||
6403 dst_reg->umax_value + umax_val < umax_val) {
6404 dst_reg->umin_value = 0;
6405 dst_reg->umax_value = U64_MAX;
6407 dst_reg->umin_value += umin_val;
6408 dst_reg->umax_value += umax_val;
6412 static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg,
6413 struct bpf_reg_state *src_reg)
6415 s32 smin_val = src_reg->s32_min_value;
6416 s32 smax_val = src_reg->s32_max_value;
6417 u32 umin_val = src_reg->u32_min_value;
6418 u32 umax_val = src_reg->u32_max_value;
6420 if (signed_sub32_overflows(dst_reg->s32_min_value, smax_val) ||
6421 signed_sub32_overflows(dst_reg->s32_max_value, smin_val)) {
6422 /* Overflow possible, we know nothing */
6423 dst_reg->s32_min_value = S32_MIN;
6424 dst_reg->s32_max_value = S32_MAX;
6426 dst_reg->s32_min_value -= smax_val;
6427 dst_reg->s32_max_value -= smin_val;
6429 if (dst_reg->u32_min_value < umax_val) {
6430 /* Overflow possible, we know nothing */
6431 dst_reg->u32_min_value = 0;
6432 dst_reg->u32_max_value = U32_MAX;
6434 /* Cannot overflow (as long as bounds are consistent) */
6435 dst_reg->u32_min_value -= umax_val;
6436 dst_reg->u32_max_value -= umin_val;
6440 static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
6441 struct bpf_reg_state *src_reg)
6443 s64 smin_val = src_reg->smin_value;
6444 s64 smax_val = src_reg->smax_value;
6445 u64 umin_val = src_reg->umin_value;
6446 u64 umax_val = src_reg->umax_value;
6448 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
6449 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
6450 /* Overflow possible, we know nothing */
6451 dst_reg->smin_value = S64_MIN;
6452 dst_reg->smax_value = S64_MAX;
6454 dst_reg->smin_value -= smax_val;
6455 dst_reg->smax_value -= smin_val;
6457 if (dst_reg->umin_value < umax_val) {
6458 /* Overflow possible, we know nothing */
6459 dst_reg->umin_value = 0;
6460 dst_reg->umax_value = U64_MAX;
6462 /* Cannot overflow (as long as bounds are consistent) */
6463 dst_reg->umin_value -= umax_val;
6464 dst_reg->umax_value -= umin_val;
6468 static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg,
6469 struct bpf_reg_state *src_reg)
6471 s32 smin_val = src_reg->s32_min_value;
6472 u32 umin_val = src_reg->u32_min_value;
6473 u32 umax_val = src_reg->u32_max_value;
6475 if (smin_val < 0 || dst_reg->s32_min_value < 0) {
6476 /* Ain't nobody got time to multiply that sign */
6477 __mark_reg32_unbounded(dst_reg);
6480 /* Both values are positive, so we can work with unsigned and
6481 * copy the result to signed (unless it exceeds S32_MAX).
6483 if (umax_val > U16_MAX || dst_reg->u32_max_value > U16_MAX) {
6484 /* Potential overflow, we know nothing */
6485 __mark_reg32_unbounded(dst_reg);
6488 dst_reg->u32_min_value *= umin_val;
6489 dst_reg->u32_max_value *= umax_val;
6490 if (dst_reg->u32_max_value > S32_MAX) {
6491 /* Overflow possible, we know nothing */
6492 dst_reg->s32_min_value = S32_MIN;
6493 dst_reg->s32_max_value = S32_MAX;
6495 dst_reg->s32_min_value = dst_reg->u32_min_value;
6496 dst_reg->s32_max_value = dst_reg->u32_max_value;
6500 static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
6501 struct bpf_reg_state *src_reg)
6503 s64 smin_val = src_reg->smin_value;
6504 u64 umin_val = src_reg->umin_value;
6505 u64 umax_val = src_reg->umax_value;
6507 if (smin_val < 0 || dst_reg->smin_value < 0) {
6508 /* Ain't nobody got time to multiply that sign */
6509 __mark_reg64_unbounded(dst_reg);
6512 /* Both values are positive, so we can work with unsigned and
6513 * copy the result to signed (unless it exceeds S64_MAX).
6515 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
6516 /* Potential overflow, we know nothing */
6517 __mark_reg64_unbounded(dst_reg);
6520 dst_reg->umin_value *= umin_val;
6521 dst_reg->umax_value *= umax_val;
6522 if (dst_reg->umax_value > S64_MAX) {
6523 /* Overflow possible, we know nothing */
6524 dst_reg->smin_value = S64_MIN;
6525 dst_reg->smax_value = S64_MAX;
6527 dst_reg->smin_value = dst_reg->umin_value;
6528 dst_reg->smax_value = dst_reg->umax_value;
6532 static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
6533 struct bpf_reg_state *src_reg)
6535 bool src_known = tnum_subreg_is_const(src_reg->var_off);
6536 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
6537 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
6538 s32 smin_val = src_reg->s32_min_value;
6539 u32 umax_val = src_reg->u32_max_value;
6541 /* Assuming scalar64_min_max_and will be called so its safe
6542 * to skip updating register for known 32-bit case.
6544 if (src_known && dst_known)
6547 /* We get our minimum from the var_off, since that's inherently
6548 * bitwise. Our maximum is the minimum of the operands' maxima.
6550 dst_reg->u32_min_value = var32_off.value;
6551 dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val);
6552 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
6553 /* Lose signed bounds when ANDing negative numbers,
6554 * ain't nobody got time for that.
6556 dst_reg->s32_min_value = S32_MIN;
6557 dst_reg->s32_max_value = S32_MAX;
6559 /* ANDing two positives gives a positive, so safe to
6560 * cast result into s64.
6562 dst_reg->s32_min_value = dst_reg->u32_min_value;
6563 dst_reg->s32_max_value = dst_reg->u32_max_value;
6568 static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
6569 struct bpf_reg_state *src_reg)
6571 bool src_known = tnum_is_const(src_reg->var_off);
6572 bool dst_known = tnum_is_const(dst_reg->var_off);
6573 s64 smin_val = src_reg->smin_value;
6574 u64 umax_val = src_reg->umax_value;
6576 if (src_known && dst_known) {
6577 __mark_reg_known(dst_reg, dst_reg->var_off.value);
6581 /* We get our minimum from the var_off, since that's inherently
6582 * bitwise. Our maximum is the minimum of the operands' maxima.
6584 dst_reg->umin_value = dst_reg->var_off.value;
6585 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
6586 if (dst_reg->smin_value < 0 || smin_val < 0) {
6587 /* Lose signed bounds when ANDing negative numbers,
6588 * ain't nobody got time for that.
6590 dst_reg->smin_value = S64_MIN;
6591 dst_reg->smax_value = S64_MAX;
6593 /* ANDing two positives gives a positive, so safe to
6594 * cast result into s64.
6596 dst_reg->smin_value = dst_reg->umin_value;
6597 dst_reg->smax_value = dst_reg->umax_value;
6599 /* We may learn something more from the var_off */
6600 __update_reg_bounds(dst_reg);
6603 static void scalar32_min_max_or(struct bpf_reg_state *dst_reg,
6604 struct bpf_reg_state *src_reg)
6606 bool src_known = tnum_subreg_is_const(src_reg->var_off);
6607 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
6608 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
6609 s32 smin_val = src_reg->s32_min_value;
6610 u32 umin_val = src_reg->u32_min_value;
6612 /* Assuming scalar64_min_max_or will be called so it is safe
6613 * to skip updating register for known case.
6615 if (src_known && dst_known)
6618 /* We get our maximum from the var_off, and our minimum is the
6619 * maximum of the operands' minima
6621 dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val);
6622 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
6623 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
6624 /* Lose signed bounds when ORing negative numbers,
6625 * ain't nobody got time for that.
6627 dst_reg->s32_min_value = S32_MIN;
6628 dst_reg->s32_max_value = S32_MAX;
6630 /* ORing two positives gives a positive, so safe to
6631 * cast result into s64.
6633 dst_reg->s32_min_value = dst_reg->u32_min_value;
6634 dst_reg->s32_max_value = dst_reg->u32_max_value;
6638 static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
6639 struct bpf_reg_state *src_reg)
6641 bool src_known = tnum_is_const(src_reg->var_off);
6642 bool dst_known = tnum_is_const(dst_reg->var_off);
6643 s64 smin_val = src_reg->smin_value;
6644 u64 umin_val = src_reg->umin_value;
6646 if (src_known && dst_known) {
6647 __mark_reg_known(dst_reg, dst_reg->var_off.value);
6651 /* We get our maximum from the var_off, and our minimum is the
6652 * maximum of the operands' minima
6654 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
6655 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
6656 if (dst_reg->smin_value < 0 || smin_val < 0) {
6657 /* Lose signed bounds when ORing negative numbers,
6658 * ain't nobody got time for that.
6660 dst_reg->smin_value = S64_MIN;
6661 dst_reg->smax_value = S64_MAX;
6663 /* ORing two positives gives a positive, so safe to
6664 * cast result into s64.
6666 dst_reg->smin_value = dst_reg->umin_value;
6667 dst_reg->smax_value = dst_reg->umax_value;
6669 /* We may learn something more from the var_off */
6670 __update_reg_bounds(dst_reg);
6673 static void scalar32_min_max_xor(struct bpf_reg_state *dst_reg,
6674 struct bpf_reg_state *src_reg)
6676 bool src_known = tnum_subreg_is_const(src_reg->var_off);
6677 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
6678 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
6679 s32 smin_val = src_reg->s32_min_value;
6681 /* Assuming scalar64_min_max_xor will be called so it is safe
6682 * to skip updating register for known case.
6684 if (src_known && dst_known)
6687 /* We get both minimum and maximum from the var32_off. */
6688 dst_reg->u32_min_value = var32_off.value;
6689 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
6691 if (dst_reg->s32_min_value >= 0 && smin_val >= 0) {
6692 /* XORing two positive sign numbers gives a positive,
6693 * so safe to cast u32 result into s32.
6695 dst_reg->s32_min_value = dst_reg->u32_min_value;
6696 dst_reg->s32_max_value = dst_reg->u32_max_value;
6698 dst_reg->s32_min_value = S32_MIN;
6699 dst_reg->s32_max_value = S32_MAX;
6703 static void scalar_min_max_xor(struct bpf_reg_state *dst_reg,
6704 struct bpf_reg_state *src_reg)
6706 bool src_known = tnum_is_const(src_reg->var_off);
6707 bool dst_known = tnum_is_const(dst_reg->var_off);
6708 s64 smin_val = src_reg->smin_value;
6710 if (src_known && dst_known) {
6711 /* dst_reg->var_off.value has been updated earlier */
6712 __mark_reg_known(dst_reg, dst_reg->var_off.value);
6716 /* We get both minimum and maximum from the var_off. */
6717 dst_reg->umin_value = dst_reg->var_off.value;
6718 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
6720 if (dst_reg->smin_value >= 0 && smin_val >= 0) {
6721 /* XORing two positive sign numbers gives a positive,
6722 * so safe to cast u64 result into s64.
6724 dst_reg->smin_value = dst_reg->umin_value;
6725 dst_reg->smax_value = dst_reg->umax_value;
6727 dst_reg->smin_value = S64_MIN;
6728 dst_reg->smax_value = S64_MAX;
6731 __update_reg_bounds(dst_reg);
6734 static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
6735 u64 umin_val, u64 umax_val)
6737 /* We lose all sign bit information (except what we can pick
6740 dst_reg->s32_min_value = S32_MIN;
6741 dst_reg->s32_max_value = S32_MAX;
6742 /* If we might shift our top bit out, then we know nothing */
6743 if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) {
6744 dst_reg->u32_min_value = 0;
6745 dst_reg->u32_max_value = U32_MAX;
6747 dst_reg->u32_min_value <<= umin_val;
6748 dst_reg->u32_max_value <<= umax_val;
6752 static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
6753 struct bpf_reg_state *src_reg)
6755 u32 umax_val = src_reg->u32_max_value;
6756 u32 umin_val = src_reg->u32_min_value;
6757 /* u32 alu operation will zext upper bits */
6758 struct tnum subreg = tnum_subreg(dst_reg->var_off);
6760 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
6761 dst_reg->var_off = tnum_subreg(tnum_lshift(subreg, umin_val));
6762 /* Not required but being careful mark reg64 bounds as unknown so
6763 * that we are forced to pick them up from tnum and zext later and
6764 * if some path skips this step we are still safe.
6766 __mark_reg64_unbounded(dst_reg);
6767 __update_reg32_bounds(dst_reg);
6770 static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg,
6771 u64 umin_val, u64 umax_val)
6773 /* Special case <<32 because it is a common compiler pattern to sign
6774 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
6775 * positive we know this shift will also be positive so we can track
6776 * bounds correctly. Otherwise we lose all sign bit information except
6777 * what we can pick up from var_off. Perhaps we can generalize this
6778 * later to shifts of any length.
6780 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0)
6781 dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32;
6783 dst_reg->smax_value = S64_MAX;
6785 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0)
6786 dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32;
6788 dst_reg->smin_value = S64_MIN;
6790 /* If we might shift our top bit out, then we know nothing */
6791 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
6792 dst_reg->umin_value = 0;
6793 dst_reg->umax_value = U64_MAX;
6795 dst_reg->umin_value <<= umin_val;
6796 dst_reg->umax_value <<= umax_val;
6800 static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
6801 struct bpf_reg_state *src_reg)
6803 u64 umax_val = src_reg->umax_value;
6804 u64 umin_val = src_reg->umin_value;
6806 /* scalar64 calc uses 32bit unshifted bounds so must be called first */
6807 __scalar64_min_max_lsh(dst_reg, umin_val, umax_val);
6808 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
6810 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
6811 /* We may learn something more from the var_off */
6812 __update_reg_bounds(dst_reg);
6815 static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg,
6816 struct bpf_reg_state *src_reg)
6818 struct tnum subreg = tnum_subreg(dst_reg->var_off);
6819 u32 umax_val = src_reg->u32_max_value;
6820 u32 umin_val = src_reg->u32_min_value;
6822 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
6823 * be negative, then either:
6824 * 1) src_reg might be zero, so the sign bit of the result is
6825 * unknown, so we lose our signed bounds
6826 * 2) it's known negative, thus the unsigned bounds capture the
6828 * 3) the signed bounds cross zero, so they tell us nothing
6830 * If the value in dst_reg is known nonnegative, then again the
6831 * unsigned bounds capture the signed bounds.
6832 * Thus, in all cases it suffices to blow away our signed bounds
6833 * and rely on inferring new ones from the unsigned bounds and
6834 * var_off of the result.
6836 dst_reg->s32_min_value = S32_MIN;
6837 dst_reg->s32_max_value = S32_MAX;
6839 dst_reg->var_off = tnum_rshift(subreg, umin_val);
6840 dst_reg->u32_min_value >>= umax_val;
6841 dst_reg->u32_max_value >>= umin_val;
6843 __mark_reg64_unbounded(dst_reg);
6844 __update_reg32_bounds(dst_reg);
6847 static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
6848 struct bpf_reg_state *src_reg)
6850 u64 umax_val = src_reg->umax_value;
6851 u64 umin_val = src_reg->umin_value;
6853 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
6854 * be negative, then either:
6855 * 1) src_reg might be zero, so the sign bit of the result is
6856 * unknown, so we lose our signed bounds
6857 * 2) it's known negative, thus the unsigned bounds capture the
6859 * 3) the signed bounds cross zero, so they tell us nothing
6861 * If the value in dst_reg is known nonnegative, then again the
6862 * unsigned bounds capture the signed bounds.
6863 * Thus, in all cases it suffices to blow away our signed bounds
6864 * and rely on inferring new ones from the unsigned bounds and
6865 * var_off of the result.
6867 dst_reg->smin_value = S64_MIN;
6868 dst_reg->smax_value = S64_MAX;
6869 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
6870 dst_reg->umin_value >>= umax_val;
6871 dst_reg->umax_value >>= umin_val;
6873 /* Its not easy to operate on alu32 bounds here because it depends
6874 * on bits being shifted in. Take easy way out and mark unbounded
6875 * so we can recalculate later from tnum.
6877 __mark_reg32_unbounded(dst_reg);
6878 __update_reg_bounds(dst_reg);
6881 static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg,
6882 struct bpf_reg_state *src_reg)
6884 u64 umin_val = src_reg->u32_min_value;
6886 /* Upon reaching here, src_known is true and
6887 * umax_val is equal to umin_val.
6889 dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val);
6890 dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val);
6892 dst_reg->var_off = tnum_arshift(tnum_subreg(dst_reg->var_off), umin_val, 32);
6894 /* blow away the dst_reg umin_value/umax_value and rely on
6895 * dst_reg var_off to refine the result.
6897 dst_reg->u32_min_value = 0;
6898 dst_reg->u32_max_value = U32_MAX;
6900 __mark_reg64_unbounded(dst_reg);
6901 __update_reg32_bounds(dst_reg);
6904 static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
6905 struct bpf_reg_state *src_reg)
6907 u64 umin_val = src_reg->umin_value;
6909 /* Upon reaching here, src_known is true and umax_val is equal
6912 dst_reg->smin_value >>= umin_val;
6913 dst_reg->smax_value >>= umin_val;
6915 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val, 64);
6917 /* blow away the dst_reg umin_value/umax_value and rely on
6918 * dst_reg var_off to refine the result.
6920 dst_reg->umin_value = 0;
6921 dst_reg->umax_value = U64_MAX;
6923 /* Its not easy to operate on alu32 bounds here because it depends
6924 * on bits being shifted in from upper 32-bits. Take easy way out
6925 * and mark unbounded so we can recalculate later from tnum.
6927 __mark_reg32_unbounded(dst_reg);
6928 __update_reg_bounds(dst_reg);
6931 /* WARNING: This function does calculations on 64-bit values, but the actual
6932 * execution may occur on 32-bit values. Therefore, things like bitshifts
6933 * need extra checks in the 32-bit case.
6935 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
6936 struct bpf_insn *insn,
6937 struct bpf_reg_state *dst_reg,
6938 struct bpf_reg_state src_reg)
6940 struct bpf_reg_state *regs = cur_regs(env);
6941 u8 opcode = BPF_OP(insn->code);
6943 s64 smin_val, smax_val;
6944 u64 umin_val, umax_val;
6945 s32 s32_min_val, s32_max_val;
6946 u32 u32_min_val, u32_max_val;
6947 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
6948 bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
6951 smin_val = src_reg.smin_value;
6952 smax_val = src_reg.smax_value;
6953 umin_val = src_reg.umin_value;
6954 umax_val = src_reg.umax_value;
6956 s32_min_val = src_reg.s32_min_value;
6957 s32_max_val = src_reg.s32_max_value;
6958 u32_min_val = src_reg.u32_min_value;
6959 u32_max_val = src_reg.u32_max_value;
6962 src_known = tnum_subreg_is_const(src_reg.var_off);
6964 (s32_min_val != s32_max_val || u32_min_val != u32_max_val)) ||
6965 s32_min_val > s32_max_val || u32_min_val > u32_max_val) {
6966 /* Taint dst register if offset had invalid bounds
6967 * derived from e.g. dead branches.
6969 __mark_reg_unknown(env, dst_reg);
6973 src_known = tnum_is_const(src_reg.var_off);
6975 (smin_val != smax_val || umin_val != umax_val)) ||
6976 smin_val > smax_val || umin_val > umax_val) {
6977 /* Taint dst register if offset had invalid bounds
6978 * derived from e.g. dead branches.
6980 __mark_reg_unknown(env, dst_reg);
6986 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
6987 __mark_reg_unknown(env, dst_reg);
6991 if (sanitize_needed(opcode)) {
6992 ret = sanitize_val_alu(env, insn);
6994 return sanitize_err(env, insn, ret, NULL, NULL);
6997 /* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
6998 * There are two classes of instructions: The first class we track both
6999 * alu32 and alu64 sign/unsigned bounds independently this provides the
7000 * greatest amount of precision when alu operations are mixed with jmp32
7001 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
7002 * and BPF_OR. This is possible because these ops have fairly easy to
7003 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
7004 * See alu32 verifier tests for examples. The second class of
7005 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
7006 * with regards to tracking sign/unsigned bounds because the bits may
7007 * cross subreg boundaries in the alu64 case. When this happens we mark
7008 * the reg unbounded in the subreg bound space and use the resulting
7009 * tnum to calculate an approximation of the sign/unsigned bounds.
7013 scalar32_min_max_add(dst_reg, &src_reg);
7014 scalar_min_max_add(dst_reg, &src_reg);
7015 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
7018 scalar32_min_max_sub(dst_reg, &src_reg);
7019 scalar_min_max_sub(dst_reg, &src_reg);
7020 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
7023 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
7024 scalar32_min_max_mul(dst_reg, &src_reg);
7025 scalar_min_max_mul(dst_reg, &src_reg);
7028 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
7029 scalar32_min_max_and(dst_reg, &src_reg);
7030 scalar_min_max_and(dst_reg, &src_reg);
7033 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
7034 scalar32_min_max_or(dst_reg, &src_reg);
7035 scalar_min_max_or(dst_reg, &src_reg);
7038 dst_reg->var_off = tnum_xor(dst_reg->var_off, src_reg.var_off);
7039 scalar32_min_max_xor(dst_reg, &src_reg);
7040 scalar_min_max_xor(dst_reg, &src_reg);
7043 if (umax_val >= insn_bitness) {
7044 /* Shifts greater than 31 or 63 are undefined.
7045 * This includes shifts by a negative number.
7047 mark_reg_unknown(env, regs, insn->dst_reg);
7051 scalar32_min_max_lsh(dst_reg, &src_reg);
7053 scalar_min_max_lsh(dst_reg, &src_reg);
7056 if (umax_val >= insn_bitness) {
7057 /* Shifts greater than 31 or 63 are undefined.
7058 * This includes shifts by a negative number.
7060 mark_reg_unknown(env, regs, insn->dst_reg);
7064 scalar32_min_max_rsh(dst_reg, &src_reg);
7066 scalar_min_max_rsh(dst_reg, &src_reg);
7069 if (umax_val >= insn_bitness) {
7070 /* Shifts greater than 31 or 63 are undefined.
7071 * This includes shifts by a negative number.
7073 mark_reg_unknown(env, regs, insn->dst_reg);
7077 scalar32_min_max_arsh(dst_reg, &src_reg);
7079 scalar_min_max_arsh(dst_reg, &src_reg);
7082 mark_reg_unknown(env, regs, insn->dst_reg);
7086 /* ALU32 ops are zero extended into 64bit register */
7088 zext_32_to_64(dst_reg);
7090 __update_reg_bounds(dst_reg);
7091 __reg_deduce_bounds(dst_reg);
7092 __reg_bound_offset(dst_reg);
7096 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
7099 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
7100 struct bpf_insn *insn)
7102 struct bpf_verifier_state *vstate = env->cur_state;
7103 struct bpf_func_state *state = vstate->frame[vstate->curframe];
7104 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
7105 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
7106 u8 opcode = BPF_OP(insn->code);
7109 dst_reg = ®s[insn->dst_reg];
7111 if (dst_reg->type != SCALAR_VALUE)
7114 /* Make sure ID is cleared otherwise dst_reg min/max could be
7115 * incorrectly propagated into other registers by find_equal_scalars()
7118 if (BPF_SRC(insn->code) == BPF_X) {
7119 src_reg = ®s[insn->src_reg];
7120 if (src_reg->type != SCALAR_VALUE) {
7121 if (dst_reg->type != SCALAR_VALUE) {
7122 /* Combining two pointers by any ALU op yields
7123 * an arbitrary scalar. Disallow all math except
7124 * pointer subtraction
7126 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
7127 mark_reg_unknown(env, regs, insn->dst_reg);
7130 verbose(env, "R%d pointer %s pointer prohibited\n",
7132 bpf_alu_string[opcode >> 4]);
7135 /* scalar += pointer
7136 * This is legal, but we have to reverse our
7137 * src/dest handling in computing the range
7139 err = mark_chain_precision(env, insn->dst_reg);
7142 return adjust_ptr_min_max_vals(env, insn,
7145 } else if (ptr_reg) {
7146 /* pointer += scalar */
7147 err = mark_chain_precision(env, insn->src_reg);
7150 return adjust_ptr_min_max_vals(env, insn,
7154 /* Pretend the src is a reg with a known value, since we only
7155 * need to be able to read from this state.
7157 off_reg.type = SCALAR_VALUE;
7158 __mark_reg_known(&off_reg, insn->imm);
7160 if (ptr_reg) /* pointer += K */
7161 return adjust_ptr_min_max_vals(env, insn,
7165 /* Got here implies adding two SCALAR_VALUEs */
7166 if (WARN_ON_ONCE(ptr_reg)) {
7167 print_verifier_state(env, state);
7168 verbose(env, "verifier internal error: unexpected ptr_reg\n");
7171 if (WARN_ON(!src_reg)) {
7172 print_verifier_state(env, state);
7173 verbose(env, "verifier internal error: no src_reg\n");
7176 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
7179 /* check validity of 32-bit and 64-bit arithmetic operations */
7180 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
7182 struct bpf_reg_state *regs = cur_regs(env);
7183 u8 opcode = BPF_OP(insn->code);
7186 if (opcode == BPF_END || opcode == BPF_NEG) {
7187 if (opcode == BPF_NEG) {
7188 if (BPF_SRC(insn->code) != 0 ||
7189 insn->src_reg != BPF_REG_0 ||
7190 insn->off != 0 || insn->imm != 0) {
7191 verbose(env, "BPF_NEG uses reserved fields\n");
7195 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
7196 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
7197 BPF_CLASS(insn->code) == BPF_ALU64) {
7198 verbose(env, "BPF_END uses reserved fields\n");
7203 /* check src operand */
7204 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
7208 if (is_pointer_value(env, insn->dst_reg)) {
7209 verbose(env, "R%d pointer arithmetic prohibited\n",
7214 /* check dest operand */
7215 err = check_reg_arg(env, insn->dst_reg, DST_OP);
7219 } else if (opcode == BPF_MOV) {
7221 if (BPF_SRC(insn->code) == BPF_X) {
7222 if (insn->imm != 0 || insn->off != 0) {
7223 verbose(env, "BPF_MOV uses reserved fields\n");
7227 /* check src operand */
7228 err = check_reg_arg(env, insn->src_reg, SRC_OP);
7232 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
7233 verbose(env, "BPF_MOV uses reserved fields\n");
7238 /* check dest operand, mark as required later */
7239 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
7243 if (BPF_SRC(insn->code) == BPF_X) {
7244 struct bpf_reg_state *src_reg = regs + insn->src_reg;
7245 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
7247 if (BPF_CLASS(insn->code) == BPF_ALU64) {
7249 * copy register state to dest reg
7251 if (src_reg->type == SCALAR_VALUE && !src_reg->id)
7252 /* Assign src and dst registers the same ID
7253 * that will be used by find_equal_scalars()
7254 * to propagate min/max range.
7256 src_reg->id = ++env->id_gen;
7257 *dst_reg = *src_reg;
7258 dst_reg->live |= REG_LIVE_WRITTEN;
7259 dst_reg->subreg_def = DEF_NOT_SUBREG;
7262 if (is_pointer_value(env, insn->src_reg)) {
7264 "R%d partial copy of pointer\n",
7267 } else if (src_reg->type == SCALAR_VALUE) {
7268 *dst_reg = *src_reg;
7269 /* Make sure ID is cleared otherwise
7270 * dst_reg min/max could be incorrectly
7271 * propagated into src_reg by find_equal_scalars()
7274 dst_reg->live |= REG_LIVE_WRITTEN;
7275 dst_reg->subreg_def = env->insn_idx + 1;
7277 mark_reg_unknown(env, regs,
7280 zext_32_to_64(dst_reg);
7284 * remember the value we stored into this reg
7286 /* clear any state __mark_reg_known doesn't set */
7287 mark_reg_unknown(env, regs, insn->dst_reg);
7288 regs[insn->dst_reg].type = SCALAR_VALUE;
7289 if (BPF_CLASS(insn->code) == BPF_ALU64) {
7290 __mark_reg_known(regs + insn->dst_reg,
7293 __mark_reg_known(regs + insn->dst_reg,
7298 } else if (opcode > BPF_END) {
7299 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
7302 } else { /* all other ALU ops: and, sub, xor, add, ... */
7304 if (BPF_SRC(insn->code) == BPF_X) {
7305 if (insn->imm != 0 || insn->off != 0) {
7306 verbose(env, "BPF_ALU uses reserved fields\n");
7309 /* check src1 operand */
7310 err = check_reg_arg(env, insn->src_reg, SRC_OP);
7314 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
7315 verbose(env, "BPF_ALU uses reserved fields\n");
7320 /* check src2 operand */
7321 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
7325 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
7326 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
7327 verbose(env, "div by zero\n");
7331 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
7332 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
7333 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
7335 if (insn->imm < 0 || insn->imm >= size) {
7336 verbose(env, "invalid shift %d\n", insn->imm);
7341 /* check dest operand */
7342 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
7346 return adjust_reg_min_max_vals(env, insn);
7352 static void __find_good_pkt_pointers(struct bpf_func_state *state,
7353 struct bpf_reg_state *dst_reg,
7354 enum bpf_reg_type type, int new_range)
7356 struct bpf_reg_state *reg;
7359 for (i = 0; i < MAX_BPF_REG; i++) {
7360 reg = &state->regs[i];
7361 if (reg->type == type && reg->id == dst_reg->id)
7362 /* keep the maximum range already checked */
7363 reg->range = max(reg->range, new_range);
7366 bpf_for_each_spilled_reg(i, state, reg) {
7369 if (reg->type == type && reg->id == dst_reg->id)
7370 reg->range = max(reg->range, new_range);
7374 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
7375 struct bpf_reg_state *dst_reg,
7376 enum bpf_reg_type type,
7377 bool range_right_open)
7381 if (dst_reg->off < 0 ||
7382 (dst_reg->off == 0 && range_right_open))
7383 /* This doesn't give us any range */
7386 if (dst_reg->umax_value > MAX_PACKET_OFF ||
7387 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
7388 /* Risk of overflow. For instance, ptr + (1<<63) may be less
7389 * than pkt_end, but that's because it's also less than pkt.
7393 new_range = dst_reg->off;
7394 if (range_right_open)
7397 /* Examples for register markings:
7399 * pkt_data in dst register:
7403 * if (r2 > pkt_end) goto <handle exception>
7408 * if (r2 < pkt_end) goto <access okay>
7409 * <handle exception>
7412 * r2 == dst_reg, pkt_end == src_reg
7413 * r2=pkt(id=n,off=8,r=0)
7414 * r3=pkt(id=n,off=0,r=0)
7416 * pkt_data in src register:
7420 * if (pkt_end >= r2) goto <access okay>
7421 * <handle exception>
7425 * if (pkt_end <= r2) goto <handle exception>
7429 * pkt_end == dst_reg, r2 == src_reg
7430 * r2=pkt(id=n,off=8,r=0)
7431 * r3=pkt(id=n,off=0,r=0)
7433 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
7434 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
7435 * and [r3, r3 + 8-1) respectively is safe to access depending on
7439 /* If our ids match, then we must have the same max_value. And we
7440 * don't care about the other reg's fixed offset, since if it's too big
7441 * the range won't allow anything.
7442 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
7444 for (i = 0; i <= vstate->curframe; i++)
7445 __find_good_pkt_pointers(vstate->frame[i], dst_reg, type,
7449 static int is_branch32_taken(struct bpf_reg_state *reg, u32 val, u8 opcode)
7451 struct tnum subreg = tnum_subreg(reg->var_off);
7452 s32 sval = (s32)val;
7456 if (tnum_is_const(subreg))
7457 return !!tnum_equals_const(subreg, val);
7460 if (tnum_is_const(subreg))
7461 return !tnum_equals_const(subreg, val);
7464 if ((~subreg.mask & subreg.value) & val)
7466 if (!((subreg.mask | subreg.value) & val))
7470 if (reg->u32_min_value > val)
7472 else if (reg->u32_max_value <= val)
7476 if (reg->s32_min_value > sval)
7478 else if (reg->s32_max_value <= sval)
7482 if (reg->u32_max_value < val)
7484 else if (reg->u32_min_value >= val)
7488 if (reg->s32_max_value < sval)
7490 else if (reg->s32_min_value >= sval)
7494 if (reg->u32_min_value >= val)
7496 else if (reg->u32_max_value < val)
7500 if (reg->s32_min_value >= sval)
7502 else if (reg->s32_max_value < sval)
7506 if (reg->u32_max_value <= val)
7508 else if (reg->u32_min_value > val)
7512 if (reg->s32_max_value <= sval)
7514 else if (reg->s32_min_value > sval)
7523 static int is_branch64_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
7525 s64 sval = (s64)val;
7529 if (tnum_is_const(reg->var_off))
7530 return !!tnum_equals_const(reg->var_off, val);
7533 if (tnum_is_const(reg->var_off))
7534 return !tnum_equals_const(reg->var_off, val);
7537 if ((~reg->var_off.mask & reg->var_off.value) & val)
7539 if (!((reg->var_off.mask | reg->var_off.value) & val))
7543 if (reg->umin_value > val)
7545 else if (reg->umax_value <= val)
7549 if (reg->smin_value > sval)
7551 else if (reg->smax_value <= sval)
7555 if (reg->umax_value < val)
7557 else if (reg->umin_value >= val)
7561 if (reg->smax_value < sval)
7563 else if (reg->smin_value >= sval)
7567 if (reg->umin_value >= val)
7569 else if (reg->umax_value < val)
7573 if (reg->smin_value >= sval)
7575 else if (reg->smax_value < sval)
7579 if (reg->umax_value <= val)
7581 else if (reg->umin_value > val)
7585 if (reg->smax_value <= sval)
7587 else if (reg->smin_value > sval)
7595 /* compute branch direction of the expression "if (reg opcode val) goto target;"
7597 * 1 - branch will be taken and "goto target" will be executed
7598 * 0 - branch will not be taken and fall-through to next insn
7599 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value
7602 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
7605 if (__is_pointer_value(false, reg)) {
7606 if (!reg_type_not_null(reg->type))
7609 /* If pointer is valid tests against zero will fail so we can
7610 * use this to direct branch taken.
7626 return is_branch32_taken(reg, val, opcode);
7627 return is_branch64_taken(reg, val, opcode);
7630 static int flip_opcode(u32 opcode)
7632 /* How can we transform "a <op> b" into "b <op> a"? */
7633 static const u8 opcode_flip[16] = {
7634 /* these stay the same */
7635 [BPF_JEQ >> 4] = BPF_JEQ,
7636 [BPF_JNE >> 4] = BPF_JNE,
7637 [BPF_JSET >> 4] = BPF_JSET,
7638 /* these swap "lesser" and "greater" (L and G in the opcodes) */
7639 [BPF_JGE >> 4] = BPF_JLE,
7640 [BPF_JGT >> 4] = BPF_JLT,
7641 [BPF_JLE >> 4] = BPF_JGE,
7642 [BPF_JLT >> 4] = BPF_JGT,
7643 [BPF_JSGE >> 4] = BPF_JSLE,
7644 [BPF_JSGT >> 4] = BPF_JSLT,
7645 [BPF_JSLE >> 4] = BPF_JSGE,
7646 [BPF_JSLT >> 4] = BPF_JSGT
7648 return opcode_flip[opcode >> 4];
7651 static int is_pkt_ptr_branch_taken(struct bpf_reg_state *dst_reg,
7652 struct bpf_reg_state *src_reg,
7655 struct bpf_reg_state *pkt;
7657 if (src_reg->type == PTR_TO_PACKET_END) {
7659 } else if (dst_reg->type == PTR_TO_PACKET_END) {
7661 opcode = flip_opcode(opcode);
7666 if (pkt->range >= 0)
7671 /* pkt <= pkt_end */
7675 if (pkt->range == BEYOND_PKT_END)
7676 /* pkt has at last one extra byte beyond pkt_end */
7677 return opcode == BPF_JGT;
7683 /* pkt >= pkt_end */
7684 if (pkt->range == BEYOND_PKT_END || pkt->range == AT_PKT_END)
7685 return opcode == BPF_JGE;
7691 /* Adjusts the register min/max values in the case that the dst_reg is the
7692 * variable register that we are working on, and src_reg is a constant or we're
7693 * simply doing a BPF_K check.
7694 * In JEQ/JNE cases we also adjust the var_off values.
7696 static void reg_set_min_max(struct bpf_reg_state *true_reg,
7697 struct bpf_reg_state *false_reg,
7699 u8 opcode, bool is_jmp32)
7701 struct tnum false_32off = tnum_subreg(false_reg->var_off);
7702 struct tnum false_64off = false_reg->var_off;
7703 struct tnum true_32off = tnum_subreg(true_reg->var_off);
7704 struct tnum true_64off = true_reg->var_off;
7705 s64 sval = (s64)val;
7706 s32 sval32 = (s32)val32;
7708 /* If the dst_reg is a pointer, we can't learn anything about its
7709 * variable offset from the compare (unless src_reg were a pointer into
7710 * the same object, but we don't bother with that.
7711 * Since false_reg and true_reg have the same type by construction, we
7712 * only need to check one of them for pointerness.
7714 if (__is_pointer_value(false, false_reg))
7721 struct bpf_reg_state *reg =
7722 opcode == BPF_JEQ ? true_reg : false_reg;
7724 /* JEQ/JNE comparison doesn't change the register equivalence.
7726 * if (r1 == 42) goto label;
7728 * label: // here both r1 and r2 are known to be 42.
7730 * Hence when marking register as known preserve it's ID.
7733 __mark_reg32_known(reg, val32);
7735 ___mark_reg_known(reg, val);
7740 false_32off = tnum_and(false_32off, tnum_const(~val32));
7741 if (is_power_of_2(val32))
7742 true_32off = tnum_or(true_32off,
7745 false_64off = tnum_and(false_64off, tnum_const(~val));
7746 if (is_power_of_2(val))
7747 true_64off = tnum_or(true_64off,
7755 u32 false_umax = opcode == BPF_JGT ? val32 : val32 - 1;
7756 u32 true_umin = opcode == BPF_JGT ? val32 + 1 : val32;
7758 false_reg->u32_max_value = min(false_reg->u32_max_value,
7760 true_reg->u32_min_value = max(true_reg->u32_min_value,
7763 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
7764 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
7766 false_reg->umax_value = min(false_reg->umax_value, false_umax);
7767 true_reg->umin_value = max(true_reg->umin_value, true_umin);
7775 s32 false_smax = opcode == BPF_JSGT ? sval32 : sval32 - 1;
7776 s32 true_smin = opcode == BPF_JSGT ? sval32 + 1 : sval32;
7778 false_reg->s32_max_value = min(false_reg->s32_max_value, false_smax);
7779 true_reg->s32_min_value = max(true_reg->s32_min_value, true_smin);
7781 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
7782 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
7784 false_reg->smax_value = min(false_reg->smax_value, false_smax);
7785 true_reg->smin_value = max(true_reg->smin_value, true_smin);
7793 u32 false_umin = opcode == BPF_JLT ? val32 : val32 + 1;
7794 u32 true_umax = opcode == BPF_JLT ? val32 - 1 : val32;
7796 false_reg->u32_min_value = max(false_reg->u32_min_value,
7798 true_reg->u32_max_value = min(true_reg->u32_max_value,
7801 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
7802 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
7804 false_reg->umin_value = max(false_reg->umin_value, false_umin);
7805 true_reg->umax_value = min(true_reg->umax_value, true_umax);
7813 s32 false_smin = opcode == BPF_JSLT ? sval32 : sval32 + 1;
7814 s32 true_smax = opcode == BPF_JSLT ? sval32 - 1 : sval32;
7816 false_reg->s32_min_value = max(false_reg->s32_min_value, false_smin);
7817 true_reg->s32_max_value = min(true_reg->s32_max_value, true_smax);
7819 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
7820 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
7822 false_reg->smin_value = max(false_reg->smin_value, false_smin);
7823 true_reg->smax_value = min(true_reg->smax_value, true_smax);
7832 false_reg->var_off = tnum_or(tnum_clear_subreg(false_64off),
7833 tnum_subreg(false_32off));
7834 true_reg->var_off = tnum_or(tnum_clear_subreg(true_64off),
7835 tnum_subreg(true_32off));
7836 __reg_combine_32_into_64(false_reg);
7837 __reg_combine_32_into_64(true_reg);
7839 false_reg->var_off = false_64off;
7840 true_reg->var_off = true_64off;
7841 __reg_combine_64_into_32(false_reg);
7842 __reg_combine_64_into_32(true_reg);
7846 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
7849 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
7850 struct bpf_reg_state *false_reg,
7852 u8 opcode, bool is_jmp32)
7854 opcode = flip_opcode(opcode);
7855 /* This uses zero as "not present in table"; luckily the zero opcode,
7856 * BPF_JA, can't get here.
7859 reg_set_min_max(true_reg, false_reg, val, val32, opcode, is_jmp32);
7862 /* Regs are known to be equal, so intersect their min/max/var_off */
7863 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
7864 struct bpf_reg_state *dst_reg)
7866 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
7867 dst_reg->umin_value);
7868 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
7869 dst_reg->umax_value);
7870 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
7871 dst_reg->smin_value);
7872 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
7873 dst_reg->smax_value);
7874 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
7876 /* We might have learned new bounds from the var_off. */
7877 __update_reg_bounds(src_reg);
7878 __update_reg_bounds(dst_reg);
7879 /* We might have learned something about the sign bit. */
7880 __reg_deduce_bounds(src_reg);
7881 __reg_deduce_bounds(dst_reg);
7882 /* We might have learned some bits from the bounds. */
7883 __reg_bound_offset(src_reg);
7884 __reg_bound_offset(dst_reg);
7885 /* Intersecting with the old var_off might have improved our bounds
7886 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
7887 * then new var_off is (0; 0x7f...fc) which improves our umax.
7889 __update_reg_bounds(src_reg);
7890 __update_reg_bounds(dst_reg);
7893 static void reg_combine_min_max(struct bpf_reg_state *true_src,
7894 struct bpf_reg_state *true_dst,
7895 struct bpf_reg_state *false_src,
7896 struct bpf_reg_state *false_dst,
7901 __reg_combine_min_max(true_src, true_dst);
7904 __reg_combine_min_max(false_src, false_dst);
7909 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
7910 struct bpf_reg_state *reg, u32 id,
7913 if (reg_type_may_be_null(reg->type) && reg->id == id &&
7914 !WARN_ON_ONCE(!reg->id)) {
7915 /* Old offset (both fixed and variable parts) should
7916 * have been known-zero, because we don't allow pointer
7917 * arithmetic on pointers that might be NULL.
7919 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
7920 !tnum_equals_const(reg->var_off, 0) ||
7922 __mark_reg_known_zero(reg);
7926 reg->type = SCALAR_VALUE;
7927 /* We don't need id and ref_obj_id from this point
7928 * onwards anymore, thus we should better reset it,
7929 * so that state pruning has chances to take effect.
7932 reg->ref_obj_id = 0;
7937 mark_ptr_not_null_reg(reg);
7939 if (!reg_may_point_to_spin_lock(reg)) {
7940 /* For not-NULL ptr, reg->ref_obj_id will be reset
7941 * in release_reg_references().
7943 * reg->id is still used by spin_lock ptr. Other
7944 * than spin_lock ptr type, reg->id can be reset.
7951 static void __mark_ptr_or_null_regs(struct bpf_func_state *state, u32 id,
7954 struct bpf_reg_state *reg;
7957 for (i = 0; i < MAX_BPF_REG; i++)
7958 mark_ptr_or_null_reg(state, &state->regs[i], id, is_null);
7960 bpf_for_each_spilled_reg(i, state, reg) {
7963 mark_ptr_or_null_reg(state, reg, id, is_null);
7967 /* The logic is similar to find_good_pkt_pointers(), both could eventually
7968 * be folded together at some point.
7970 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
7973 struct bpf_func_state *state = vstate->frame[vstate->curframe];
7974 struct bpf_reg_state *regs = state->regs;
7975 u32 ref_obj_id = regs[regno].ref_obj_id;
7976 u32 id = regs[regno].id;
7979 if (ref_obj_id && ref_obj_id == id && is_null)
7980 /* regs[regno] is in the " == NULL" branch.
7981 * No one could have freed the reference state before
7982 * doing the NULL check.
7984 WARN_ON_ONCE(release_reference_state(state, id));
7986 for (i = 0; i <= vstate->curframe; i++)
7987 __mark_ptr_or_null_regs(vstate->frame[i], id, is_null);
7990 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
7991 struct bpf_reg_state *dst_reg,
7992 struct bpf_reg_state *src_reg,
7993 struct bpf_verifier_state *this_branch,
7994 struct bpf_verifier_state *other_branch)
7996 if (BPF_SRC(insn->code) != BPF_X)
7999 /* Pointers are always 64-bit. */
8000 if (BPF_CLASS(insn->code) == BPF_JMP32)
8003 switch (BPF_OP(insn->code)) {
8005 if ((dst_reg->type == PTR_TO_PACKET &&
8006 src_reg->type == PTR_TO_PACKET_END) ||
8007 (dst_reg->type == PTR_TO_PACKET_META &&
8008 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
8009 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
8010 find_good_pkt_pointers(this_branch, dst_reg,
8011 dst_reg->type, false);
8012 mark_pkt_end(other_branch, insn->dst_reg, true);
8013 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
8014 src_reg->type == PTR_TO_PACKET) ||
8015 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
8016 src_reg->type == PTR_TO_PACKET_META)) {
8017 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
8018 find_good_pkt_pointers(other_branch, src_reg,
8019 src_reg->type, true);
8020 mark_pkt_end(this_branch, insn->src_reg, false);
8026 if ((dst_reg->type == PTR_TO_PACKET &&
8027 src_reg->type == PTR_TO_PACKET_END) ||
8028 (dst_reg->type == PTR_TO_PACKET_META &&
8029 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
8030 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
8031 find_good_pkt_pointers(other_branch, dst_reg,
8032 dst_reg->type, true);
8033 mark_pkt_end(this_branch, insn->dst_reg, false);
8034 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
8035 src_reg->type == PTR_TO_PACKET) ||
8036 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
8037 src_reg->type == PTR_TO_PACKET_META)) {
8038 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
8039 find_good_pkt_pointers(this_branch, src_reg,
8040 src_reg->type, false);
8041 mark_pkt_end(other_branch, insn->src_reg, true);
8047 if ((dst_reg->type == PTR_TO_PACKET &&
8048 src_reg->type == PTR_TO_PACKET_END) ||
8049 (dst_reg->type == PTR_TO_PACKET_META &&
8050 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
8051 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
8052 find_good_pkt_pointers(this_branch, dst_reg,
8053 dst_reg->type, true);
8054 mark_pkt_end(other_branch, insn->dst_reg, false);
8055 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
8056 src_reg->type == PTR_TO_PACKET) ||
8057 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
8058 src_reg->type == PTR_TO_PACKET_META)) {
8059 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
8060 find_good_pkt_pointers(other_branch, src_reg,
8061 src_reg->type, false);
8062 mark_pkt_end(this_branch, insn->src_reg, true);
8068 if ((dst_reg->type == PTR_TO_PACKET &&
8069 src_reg->type == PTR_TO_PACKET_END) ||
8070 (dst_reg->type == PTR_TO_PACKET_META &&
8071 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
8072 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
8073 find_good_pkt_pointers(other_branch, dst_reg,
8074 dst_reg->type, false);
8075 mark_pkt_end(this_branch, insn->dst_reg, true);
8076 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
8077 src_reg->type == PTR_TO_PACKET) ||
8078 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
8079 src_reg->type == PTR_TO_PACKET_META)) {
8080 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
8081 find_good_pkt_pointers(this_branch, src_reg,
8082 src_reg->type, true);
8083 mark_pkt_end(other_branch, insn->src_reg, false);
8095 static void find_equal_scalars(struct bpf_verifier_state *vstate,
8096 struct bpf_reg_state *known_reg)
8098 struct bpf_func_state *state;
8099 struct bpf_reg_state *reg;
8102 for (i = 0; i <= vstate->curframe; i++) {
8103 state = vstate->frame[i];
8104 for (j = 0; j < MAX_BPF_REG; j++) {
8105 reg = &state->regs[j];
8106 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
8110 bpf_for_each_spilled_reg(j, state, reg) {
8113 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
8119 static int check_cond_jmp_op(struct bpf_verifier_env *env,
8120 struct bpf_insn *insn, int *insn_idx)
8122 struct bpf_verifier_state *this_branch = env->cur_state;
8123 struct bpf_verifier_state *other_branch;
8124 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
8125 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
8126 u8 opcode = BPF_OP(insn->code);
8131 /* Only conditional jumps are expected to reach here. */
8132 if (opcode == BPF_JA || opcode > BPF_JSLE) {
8133 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
8137 if (BPF_SRC(insn->code) == BPF_X) {
8138 if (insn->imm != 0) {
8139 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
8143 /* check src1 operand */
8144 err = check_reg_arg(env, insn->src_reg, SRC_OP);
8148 if (is_pointer_value(env, insn->src_reg)) {
8149 verbose(env, "R%d pointer comparison prohibited\n",
8153 src_reg = ®s[insn->src_reg];
8155 if (insn->src_reg != BPF_REG_0) {
8156 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
8161 /* check src2 operand */
8162 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
8166 dst_reg = ®s[insn->dst_reg];
8167 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
8169 if (BPF_SRC(insn->code) == BPF_K) {
8170 pred = is_branch_taken(dst_reg, insn->imm, opcode, is_jmp32);
8171 } else if (src_reg->type == SCALAR_VALUE &&
8172 is_jmp32 && tnum_is_const(tnum_subreg(src_reg->var_off))) {
8173 pred = is_branch_taken(dst_reg,
8174 tnum_subreg(src_reg->var_off).value,
8177 } else if (src_reg->type == SCALAR_VALUE &&
8178 !is_jmp32 && tnum_is_const(src_reg->var_off)) {
8179 pred = is_branch_taken(dst_reg,
8180 src_reg->var_off.value,
8183 } else if (reg_is_pkt_pointer_any(dst_reg) &&
8184 reg_is_pkt_pointer_any(src_reg) &&
8186 pred = is_pkt_ptr_branch_taken(dst_reg, src_reg, opcode);
8190 /* If we get here with a dst_reg pointer type it is because
8191 * above is_branch_taken() special cased the 0 comparison.
8193 if (!__is_pointer_value(false, dst_reg))
8194 err = mark_chain_precision(env, insn->dst_reg);
8195 if (BPF_SRC(insn->code) == BPF_X && !err &&
8196 !__is_pointer_value(false, src_reg))
8197 err = mark_chain_precision(env, insn->src_reg);
8202 /* only follow the goto, ignore fall-through */
8203 *insn_idx += insn->off;
8205 } else if (pred == 0) {
8206 /* only follow fall-through branch, since
8207 * that's where the program will go
8212 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
8216 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
8218 /* detect if we are comparing against a constant value so we can adjust
8219 * our min/max values for our dst register.
8220 * this is only legit if both are scalars (or pointers to the same
8221 * object, I suppose, but we don't support that right now), because
8222 * otherwise the different base pointers mean the offsets aren't
8225 if (BPF_SRC(insn->code) == BPF_X) {
8226 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
8228 if (dst_reg->type == SCALAR_VALUE &&
8229 src_reg->type == SCALAR_VALUE) {
8230 if (tnum_is_const(src_reg->var_off) ||
8232 tnum_is_const(tnum_subreg(src_reg->var_off))))
8233 reg_set_min_max(&other_branch_regs[insn->dst_reg],
8235 src_reg->var_off.value,
8236 tnum_subreg(src_reg->var_off).value,
8238 else if (tnum_is_const(dst_reg->var_off) ||
8240 tnum_is_const(tnum_subreg(dst_reg->var_off))))
8241 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
8243 dst_reg->var_off.value,
8244 tnum_subreg(dst_reg->var_off).value,
8246 else if (!is_jmp32 &&
8247 (opcode == BPF_JEQ || opcode == BPF_JNE))
8248 /* Comparing for equality, we can combine knowledge */
8249 reg_combine_min_max(&other_branch_regs[insn->src_reg],
8250 &other_branch_regs[insn->dst_reg],
8251 src_reg, dst_reg, opcode);
8253 !WARN_ON_ONCE(src_reg->id != other_branch_regs[insn->src_reg].id)) {
8254 find_equal_scalars(this_branch, src_reg);
8255 find_equal_scalars(other_branch, &other_branch_regs[insn->src_reg]);
8259 } else if (dst_reg->type == SCALAR_VALUE) {
8260 reg_set_min_max(&other_branch_regs[insn->dst_reg],
8261 dst_reg, insn->imm, (u32)insn->imm,
8265 if (dst_reg->type == SCALAR_VALUE && dst_reg->id &&
8266 !WARN_ON_ONCE(dst_reg->id != other_branch_regs[insn->dst_reg].id)) {
8267 find_equal_scalars(this_branch, dst_reg);
8268 find_equal_scalars(other_branch, &other_branch_regs[insn->dst_reg]);
8271 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
8272 * NOTE: these optimizations below are related with pointer comparison
8273 * which will never be JMP32.
8275 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
8276 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
8277 reg_type_may_be_null(dst_reg->type)) {
8278 /* Mark all identical registers in each branch as either
8279 * safe or unknown depending R == 0 or R != 0 conditional.
8281 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
8283 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
8285 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
8286 this_branch, other_branch) &&
8287 is_pointer_value(env, insn->dst_reg)) {
8288 verbose(env, "R%d pointer comparison prohibited\n",
8292 if (env->log.level & BPF_LOG_LEVEL)
8293 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
8297 /* verify BPF_LD_IMM64 instruction */
8298 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
8300 struct bpf_insn_aux_data *aux = cur_aux(env);
8301 struct bpf_reg_state *regs = cur_regs(env);
8302 struct bpf_reg_state *dst_reg;
8303 struct bpf_map *map;
8306 if (BPF_SIZE(insn->code) != BPF_DW) {
8307 verbose(env, "invalid BPF_LD_IMM insn\n");
8310 if (insn->off != 0) {
8311 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
8315 err = check_reg_arg(env, insn->dst_reg, DST_OP);
8319 dst_reg = ®s[insn->dst_reg];
8320 if (insn->src_reg == 0) {
8321 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
8323 dst_reg->type = SCALAR_VALUE;
8324 __mark_reg_known(®s[insn->dst_reg], imm);
8328 if (insn->src_reg == BPF_PSEUDO_BTF_ID) {
8329 mark_reg_known_zero(env, regs, insn->dst_reg);
8331 dst_reg->type = aux->btf_var.reg_type;
8332 switch (dst_reg->type) {
8334 dst_reg->mem_size = aux->btf_var.mem_size;
8337 case PTR_TO_PERCPU_BTF_ID:
8338 dst_reg->btf = aux->btf_var.btf;
8339 dst_reg->btf_id = aux->btf_var.btf_id;
8342 verbose(env, "bpf verifier is misconfigured\n");
8348 map = env->used_maps[aux->map_index];
8349 mark_reg_known_zero(env, regs, insn->dst_reg);
8350 dst_reg->map_ptr = map;
8352 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE) {
8353 dst_reg->type = PTR_TO_MAP_VALUE;
8354 dst_reg->off = aux->map_off;
8355 if (map_value_has_spin_lock(map))
8356 dst_reg->id = ++env->id_gen;
8357 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD) {
8358 dst_reg->type = CONST_PTR_TO_MAP;
8360 verbose(env, "bpf verifier is misconfigured\n");
8367 static bool may_access_skb(enum bpf_prog_type type)
8370 case BPF_PROG_TYPE_SOCKET_FILTER:
8371 case BPF_PROG_TYPE_SCHED_CLS:
8372 case BPF_PROG_TYPE_SCHED_ACT:
8379 /* verify safety of LD_ABS|LD_IND instructions:
8380 * - they can only appear in the programs where ctx == skb
8381 * - since they are wrappers of function calls, they scratch R1-R5 registers,
8382 * preserve R6-R9, and store return value into R0
8385 * ctx == skb == R6 == CTX
8388 * SRC == any register
8389 * IMM == 32-bit immediate
8392 * R0 - 8/16/32-bit skb data converted to cpu endianness
8394 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
8396 struct bpf_reg_state *regs = cur_regs(env);
8397 static const int ctx_reg = BPF_REG_6;
8398 u8 mode = BPF_MODE(insn->code);
8401 if (!may_access_skb(resolve_prog_type(env->prog))) {
8402 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
8406 if (!env->ops->gen_ld_abs) {
8407 verbose(env, "bpf verifier is misconfigured\n");
8411 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
8412 BPF_SIZE(insn->code) == BPF_DW ||
8413 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
8414 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
8418 /* check whether implicit source operand (register R6) is readable */
8419 err = check_reg_arg(env, ctx_reg, SRC_OP);
8423 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
8424 * gen_ld_abs() may terminate the program at runtime, leading to
8427 err = check_reference_leak(env);
8429 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
8433 if (env->cur_state->active_spin_lock) {
8434 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
8438 if (regs[ctx_reg].type != PTR_TO_CTX) {
8440 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
8444 if (mode == BPF_IND) {
8445 /* check explicit source operand */
8446 err = check_reg_arg(env, insn->src_reg, SRC_OP);
8451 err = check_ctx_reg(env, ®s[ctx_reg], ctx_reg);
8455 /* reset caller saved regs to unreadable */
8456 for (i = 0; i < CALLER_SAVED_REGS; i++) {
8457 mark_reg_not_init(env, regs, caller_saved[i]);
8458 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
8461 /* mark destination R0 register as readable, since it contains
8462 * the value fetched from the packet.
8463 * Already marked as written above.
8465 mark_reg_unknown(env, regs, BPF_REG_0);
8466 /* ld_abs load up to 32-bit skb data. */
8467 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
8471 static int check_return_code(struct bpf_verifier_env *env)
8473 struct tnum enforce_attach_type_range = tnum_unknown;
8474 const struct bpf_prog *prog = env->prog;
8475 struct bpf_reg_state *reg;
8476 struct tnum range = tnum_range(0, 1);
8477 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
8479 const bool is_subprog = env->cur_state->frame[0]->subprogno;
8481 /* LSM and struct_ops func-ptr's return type could be "void" */
8483 (prog_type == BPF_PROG_TYPE_STRUCT_OPS ||
8484 prog_type == BPF_PROG_TYPE_LSM) &&
8485 !prog->aux->attach_func_proto->type)
8488 /* eBPF calling convetion is such that R0 is used
8489 * to return the value from eBPF program.
8490 * Make sure that it's readable at this time
8491 * of bpf_exit, which means that program wrote
8492 * something into it earlier
8494 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
8498 if (is_pointer_value(env, BPF_REG_0)) {
8499 verbose(env, "R0 leaks addr as return value\n");
8503 reg = cur_regs(env) + BPF_REG_0;
8505 if (reg->type != SCALAR_VALUE) {
8506 verbose(env, "At subprogram exit the register R0 is not a scalar value (%s)\n",
8507 reg_type_str[reg->type]);
8513 switch (prog_type) {
8514 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
8515 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
8516 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG ||
8517 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETPEERNAME ||
8518 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETPEERNAME ||
8519 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETSOCKNAME ||
8520 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETSOCKNAME)
8521 range = tnum_range(1, 1);
8522 if (env->prog->expected_attach_type == BPF_CGROUP_INET4_BIND ||
8523 env->prog->expected_attach_type == BPF_CGROUP_INET6_BIND)
8524 range = tnum_range(0, 3);
8526 case BPF_PROG_TYPE_CGROUP_SKB:
8527 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
8528 range = tnum_range(0, 3);
8529 enforce_attach_type_range = tnum_range(2, 3);
8532 case BPF_PROG_TYPE_CGROUP_SOCK:
8533 case BPF_PROG_TYPE_SOCK_OPS:
8534 case BPF_PROG_TYPE_CGROUP_DEVICE:
8535 case BPF_PROG_TYPE_CGROUP_SYSCTL:
8536 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
8538 case BPF_PROG_TYPE_RAW_TRACEPOINT:
8539 if (!env->prog->aux->attach_btf_id)
8541 range = tnum_const(0);
8543 case BPF_PROG_TYPE_TRACING:
8544 switch (env->prog->expected_attach_type) {
8545 case BPF_TRACE_FENTRY:
8546 case BPF_TRACE_FEXIT:
8547 range = tnum_const(0);
8549 case BPF_TRACE_RAW_TP:
8550 case BPF_MODIFY_RETURN:
8552 case BPF_TRACE_ITER:
8558 case BPF_PROG_TYPE_SK_LOOKUP:
8559 range = tnum_range(SK_DROP, SK_PASS);
8561 case BPF_PROG_TYPE_EXT:
8562 /* freplace program can return anything as its return value
8563 * depends on the to-be-replaced kernel func or bpf program.
8569 if (reg->type != SCALAR_VALUE) {
8570 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
8571 reg_type_str[reg->type]);
8575 if (!tnum_in(range, reg->var_off)) {
8578 verbose(env, "At program exit the register R0 ");
8579 if (!tnum_is_unknown(reg->var_off)) {
8580 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
8581 verbose(env, "has value %s", tn_buf);
8583 verbose(env, "has unknown scalar value");
8585 tnum_strn(tn_buf, sizeof(tn_buf), range);
8586 verbose(env, " should have been in %s\n", tn_buf);
8590 if (!tnum_is_unknown(enforce_attach_type_range) &&
8591 tnum_in(enforce_attach_type_range, reg->var_off))
8592 env->prog->enforce_expected_attach_type = 1;
8596 /* non-recursive DFS pseudo code
8597 * 1 procedure DFS-iterative(G,v):
8598 * 2 label v as discovered
8599 * 3 let S be a stack
8601 * 5 while S is not empty
8603 * 7 if t is what we're looking for:
8605 * 9 for all edges e in G.adjacentEdges(t) do
8606 * 10 if edge e is already labelled
8607 * 11 continue with the next edge
8608 * 12 w <- G.adjacentVertex(t,e)
8609 * 13 if vertex w is not discovered and not explored
8610 * 14 label e as tree-edge
8611 * 15 label w as discovered
8614 * 18 else if vertex w is discovered
8615 * 19 label e as back-edge
8617 * 21 // vertex w is explored
8618 * 22 label e as forward- or cross-edge
8619 * 23 label t as explored
8624 * 0x11 - discovered and fall-through edge labelled
8625 * 0x12 - discovered and fall-through and branch edges labelled
8636 static u32 state_htab_size(struct bpf_verifier_env *env)
8638 return env->prog->len;
8641 static struct bpf_verifier_state_list **explored_state(
8642 struct bpf_verifier_env *env,
8645 struct bpf_verifier_state *cur = env->cur_state;
8646 struct bpf_func_state *state = cur->frame[cur->curframe];
8648 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
8651 static void init_explored_state(struct bpf_verifier_env *env, int idx)
8653 env->insn_aux_data[idx].prune_point = true;
8661 /* t, w, e - match pseudo-code above:
8662 * t - index of current instruction
8663 * w - next instruction
8666 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env,
8669 int *insn_stack = env->cfg.insn_stack;
8670 int *insn_state = env->cfg.insn_state;
8672 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
8673 return DONE_EXPLORING;
8675 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
8676 return DONE_EXPLORING;
8678 if (w < 0 || w >= env->prog->len) {
8679 verbose_linfo(env, t, "%d: ", t);
8680 verbose(env, "jump out of range from insn %d to %d\n", t, w);
8685 /* mark branch target for state pruning */
8686 init_explored_state(env, w);
8688 if (insn_state[w] == 0) {
8690 insn_state[t] = DISCOVERED | e;
8691 insn_state[w] = DISCOVERED;
8692 if (env->cfg.cur_stack >= env->prog->len)
8694 insn_stack[env->cfg.cur_stack++] = w;
8695 return KEEP_EXPLORING;
8696 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
8697 if (loop_ok && env->bpf_capable)
8698 return DONE_EXPLORING;
8699 verbose_linfo(env, t, "%d: ", t);
8700 verbose_linfo(env, w, "%d: ", w);
8701 verbose(env, "back-edge from insn %d to %d\n", t, w);
8703 } else if (insn_state[w] == EXPLORED) {
8704 /* forward- or cross-edge */
8705 insn_state[t] = DISCOVERED | e;
8707 verbose(env, "insn state internal bug\n");
8710 return DONE_EXPLORING;
8713 /* Visits the instruction at index t and returns one of the following:
8714 * < 0 - an error occurred
8715 * DONE_EXPLORING - the instruction was fully explored
8716 * KEEP_EXPLORING - there is still work to be done before it is fully explored
8718 static int visit_insn(int t, int insn_cnt, struct bpf_verifier_env *env)
8720 struct bpf_insn *insns = env->prog->insnsi;
8723 /* All non-branch instructions have a single fall-through edge. */
8724 if (BPF_CLASS(insns[t].code) != BPF_JMP &&
8725 BPF_CLASS(insns[t].code) != BPF_JMP32)
8726 return push_insn(t, t + 1, FALLTHROUGH, env, false);
8728 switch (BPF_OP(insns[t].code)) {
8730 return DONE_EXPLORING;
8733 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
8737 if (t + 1 < insn_cnt)
8738 init_explored_state(env, t + 1);
8739 if (insns[t].src_reg == BPF_PSEUDO_CALL) {
8740 init_explored_state(env, t);
8741 ret = push_insn(t, t + insns[t].imm + 1, BRANCH,
8747 if (BPF_SRC(insns[t].code) != BPF_K)
8750 /* unconditional jump with single edge */
8751 ret = push_insn(t, t + insns[t].off + 1, FALLTHROUGH, env,
8756 /* unconditional jmp is not a good pruning point,
8757 * but it's marked, since backtracking needs
8758 * to record jmp history in is_state_visited().
8760 init_explored_state(env, t + insns[t].off + 1);
8761 /* tell verifier to check for equivalent states
8762 * after every call and jump
8764 if (t + 1 < insn_cnt)
8765 init_explored_state(env, t + 1);
8770 /* conditional jump with two edges */
8771 init_explored_state(env, t);
8772 ret = push_insn(t, t + 1, FALLTHROUGH, env, true);
8776 return push_insn(t, t + insns[t].off + 1, BRANCH, env, true);
8780 /* non-recursive depth-first-search to detect loops in BPF program
8781 * loop == back-edge in directed graph
8783 static int check_cfg(struct bpf_verifier_env *env)
8785 int insn_cnt = env->prog->len;
8786 int *insn_stack, *insn_state;
8790 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
8794 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
8800 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
8801 insn_stack[0] = 0; /* 0 is the first instruction */
8802 env->cfg.cur_stack = 1;
8804 while (env->cfg.cur_stack > 0) {
8805 int t = insn_stack[env->cfg.cur_stack - 1];
8807 ret = visit_insn(t, insn_cnt, env);
8809 case DONE_EXPLORING:
8810 insn_state[t] = EXPLORED;
8811 env->cfg.cur_stack--;
8813 case KEEP_EXPLORING:
8817 verbose(env, "visit_insn internal bug\n");
8824 if (env->cfg.cur_stack < 0) {
8825 verbose(env, "pop stack internal bug\n");
8830 for (i = 0; i < insn_cnt; i++) {
8831 if (insn_state[i] != EXPLORED) {
8832 verbose(env, "unreachable insn %d\n", i);
8837 ret = 0; /* cfg looks good */
8842 env->cfg.insn_state = env->cfg.insn_stack = NULL;
8846 static int check_abnormal_return(struct bpf_verifier_env *env)
8850 for (i = 1; i < env->subprog_cnt; i++) {
8851 if (env->subprog_info[i].has_ld_abs) {
8852 verbose(env, "LD_ABS is not allowed in subprogs without BTF\n");
8855 if (env->subprog_info[i].has_tail_call) {
8856 verbose(env, "tail_call is not allowed in subprogs without BTF\n");
8863 /* The minimum supported BTF func info size */
8864 #define MIN_BPF_FUNCINFO_SIZE 8
8865 #define MAX_FUNCINFO_REC_SIZE 252
8867 static int check_btf_func(struct bpf_verifier_env *env,
8868 const union bpf_attr *attr,
8869 union bpf_attr __user *uattr)
8871 const struct btf_type *type, *func_proto, *ret_type;
8872 u32 i, nfuncs, urec_size, min_size;
8873 u32 krec_size = sizeof(struct bpf_func_info);
8874 struct bpf_func_info *krecord;
8875 struct bpf_func_info_aux *info_aux = NULL;
8876 struct bpf_prog *prog;
8877 const struct btf *btf;
8878 void __user *urecord;
8879 u32 prev_offset = 0;
8883 nfuncs = attr->func_info_cnt;
8885 if (check_abnormal_return(env))
8890 if (nfuncs != env->subprog_cnt) {
8891 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
8895 urec_size = attr->func_info_rec_size;
8896 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
8897 urec_size > MAX_FUNCINFO_REC_SIZE ||
8898 urec_size % sizeof(u32)) {
8899 verbose(env, "invalid func info rec size %u\n", urec_size);
8904 btf = prog->aux->btf;
8906 urecord = u64_to_user_ptr(attr->func_info);
8907 min_size = min_t(u32, krec_size, urec_size);
8909 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
8912 info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN);
8916 for (i = 0; i < nfuncs; i++) {
8917 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
8919 if (ret == -E2BIG) {
8920 verbose(env, "nonzero tailing record in func info");
8921 /* set the size kernel expects so loader can zero
8922 * out the rest of the record.
8924 if (put_user(min_size, &uattr->func_info_rec_size))
8930 if (copy_from_user(&krecord[i], urecord, min_size)) {
8935 /* check insn_off */
8938 if (krecord[i].insn_off) {
8940 "nonzero insn_off %u for the first func info record",
8941 krecord[i].insn_off);
8944 } else if (krecord[i].insn_off <= prev_offset) {
8946 "same or smaller insn offset (%u) than previous func info record (%u)",
8947 krecord[i].insn_off, prev_offset);
8951 if (env->subprog_info[i].start != krecord[i].insn_off) {
8952 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
8957 type = btf_type_by_id(btf, krecord[i].type_id);
8958 if (!type || !btf_type_is_func(type)) {
8959 verbose(env, "invalid type id %d in func info",
8960 krecord[i].type_id);
8963 info_aux[i].linkage = BTF_INFO_VLEN(type->info);
8965 func_proto = btf_type_by_id(btf, type->type);
8966 if (unlikely(!func_proto || !btf_type_is_func_proto(func_proto)))
8967 /* btf_func_check() already verified it during BTF load */
8969 ret_type = btf_type_skip_modifiers(btf, func_proto->type, NULL);
8971 btf_type_is_small_int(ret_type) || btf_type_is_enum(ret_type);
8972 if (i && !scalar_return && env->subprog_info[i].has_ld_abs) {
8973 verbose(env, "LD_ABS is only allowed in functions that return 'int'.\n");
8976 if (i && !scalar_return && env->subprog_info[i].has_tail_call) {
8977 verbose(env, "tail_call is only allowed in functions that return 'int'.\n");
8981 prev_offset = krecord[i].insn_off;
8982 urecord += urec_size;
8985 prog->aux->func_info = krecord;
8986 prog->aux->func_info_cnt = nfuncs;
8987 prog->aux->func_info_aux = info_aux;
8996 static void adjust_btf_func(struct bpf_verifier_env *env)
8998 struct bpf_prog_aux *aux = env->prog->aux;
9001 if (!aux->func_info)
9004 for (i = 0; i < env->subprog_cnt; i++)
9005 aux->func_info[i].insn_off = env->subprog_info[i].start;
9008 #define MIN_BPF_LINEINFO_SIZE (offsetof(struct bpf_line_info, line_col) + \
9009 sizeof(((struct bpf_line_info *)(0))->line_col))
9010 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
9012 static int check_btf_line(struct bpf_verifier_env *env,
9013 const union bpf_attr *attr,
9014 union bpf_attr __user *uattr)
9016 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
9017 struct bpf_subprog_info *sub;
9018 struct bpf_line_info *linfo;
9019 struct bpf_prog *prog;
9020 const struct btf *btf;
9021 void __user *ulinfo;
9024 nr_linfo = attr->line_info_cnt;
9028 rec_size = attr->line_info_rec_size;
9029 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
9030 rec_size > MAX_LINEINFO_REC_SIZE ||
9031 rec_size & (sizeof(u32) - 1))
9034 /* Need to zero it in case the userspace may
9035 * pass in a smaller bpf_line_info object.
9037 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
9038 GFP_KERNEL | __GFP_NOWARN);
9043 btf = prog->aux->btf;
9046 sub = env->subprog_info;
9047 ulinfo = u64_to_user_ptr(attr->line_info);
9048 expected_size = sizeof(struct bpf_line_info);
9049 ncopy = min_t(u32, expected_size, rec_size);
9050 for (i = 0; i < nr_linfo; i++) {
9051 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
9053 if (err == -E2BIG) {
9054 verbose(env, "nonzero tailing record in line_info");
9055 if (put_user(expected_size,
9056 &uattr->line_info_rec_size))
9062 if (copy_from_user(&linfo[i], ulinfo, ncopy)) {
9068 * Check insn_off to ensure
9069 * 1) strictly increasing AND
9070 * 2) bounded by prog->len
9072 * The linfo[0].insn_off == 0 check logically falls into
9073 * the later "missing bpf_line_info for func..." case
9074 * because the first linfo[0].insn_off must be the
9075 * first sub also and the first sub must have
9076 * subprog_info[0].start == 0.
9078 if ((i && linfo[i].insn_off <= prev_offset) ||
9079 linfo[i].insn_off >= prog->len) {
9080 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
9081 i, linfo[i].insn_off, prev_offset,
9087 if (!prog->insnsi[linfo[i].insn_off].code) {
9089 "Invalid insn code at line_info[%u].insn_off\n",
9095 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
9096 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
9097 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
9102 if (s != env->subprog_cnt) {
9103 if (linfo[i].insn_off == sub[s].start) {
9104 sub[s].linfo_idx = i;
9106 } else if (sub[s].start < linfo[i].insn_off) {
9107 verbose(env, "missing bpf_line_info for func#%u\n", s);
9113 prev_offset = linfo[i].insn_off;
9117 if (s != env->subprog_cnt) {
9118 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
9119 env->subprog_cnt - s, s);
9124 prog->aux->linfo = linfo;
9125 prog->aux->nr_linfo = nr_linfo;
9134 static int check_btf_info(struct bpf_verifier_env *env,
9135 const union bpf_attr *attr,
9136 union bpf_attr __user *uattr)
9141 if (!attr->func_info_cnt && !attr->line_info_cnt) {
9142 if (check_abnormal_return(env))
9147 btf = btf_get_by_fd(attr->prog_btf_fd);
9149 return PTR_ERR(btf);
9150 if (btf_is_kernel(btf)) {
9154 env->prog->aux->btf = btf;
9156 err = check_btf_func(env, attr, uattr);
9160 err = check_btf_line(env, attr, uattr);
9167 /* check %cur's range satisfies %old's */
9168 static bool range_within(struct bpf_reg_state *old,
9169 struct bpf_reg_state *cur)
9171 return old->umin_value <= cur->umin_value &&
9172 old->umax_value >= cur->umax_value &&
9173 old->smin_value <= cur->smin_value &&
9174 old->smax_value >= cur->smax_value &&
9175 old->u32_min_value <= cur->u32_min_value &&
9176 old->u32_max_value >= cur->u32_max_value &&
9177 old->s32_min_value <= cur->s32_min_value &&
9178 old->s32_max_value >= cur->s32_max_value;
9181 /* Maximum number of register states that can exist at once */
9182 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
9188 /* If in the old state two registers had the same id, then they need to have
9189 * the same id in the new state as well. But that id could be different from
9190 * the old state, so we need to track the mapping from old to new ids.
9191 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
9192 * regs with old id 5 must also have new id 9 for the new state to be safe. But
9193 * regs with a different old id could still have new id 9, we don't care about
9195 * So we look through our idmap to see if this old id has been seen before. If
9196 * so, we require the new id to match; otherwise, we add the id pair to the map.
9198 static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
9202 for (i = 0; i < ID_MAP_SIZE; i++) {
9203 if (!idmap[i].old) {
9204 /* Reached an empty slot; haven't seen this id before */
9205 idmap[i].old = old_id;
9206 idmap[i].cur = cur_id;
9209 if (idmap[i].old == old_id)
9210 return idmap[i].cur == cur_id;
9212 /* We ran out of idmap slots, which should be impossible */
9217 static void clean_func_state(struct bpf_verifier_env *env,
9218 struct bpf_func_state *st)
9220 enum bpf_reg_liveness live;
9223 for (i = 0; i < BPF_REG_FP; i++) {
9224 live = st->regs[i].live;
9225 /* liveness must not touch this register anymore */
9226 st->regs[i].live |= REG_LIVE_DONE;
9227 if (!(live & REG_LIVE_READ))
9228 /* since the register is unused, clear its state
9229 * to make further comparison simpler
9231 __mark_reg_not_init(env, &st->regs[i]);
9234 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
9235 live = st->stack[i].spilled_ptr.live;
9236 /* liveness must not touch this stack slot anymore */
9237 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
9238 if (!(live & REG_LIVE_READ)) {
9239 __mark_reg_not_init(env, &st->stack[i].spilled_ptr);
9240 for (j = 0; j < BPF_REG_SIZE; j++)
9241 st->stack[i].slot_type[j] = STACK_INVALID;
9246 static void clean_verifier_state(struct bpf_verifier_env *env,
9247 struct bpf_verifier_state *st)
9251 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
9252 /* all regs in this state in all frames were already marked */
9255 for (i = 0; i <= st->curframe; i++)
9256 clean_func_state(env, st->frame[i]);
9259 /* the parentage chains form a tree.
9260 * the verifier states are added to state lists at given insn and
9261 * pushed into state stack for future exploration.
9262 * when the verifier reaches bpf_exit insn some of the verifer states
9263 * stored in the state lists have their final liveness state already,
9264 * but a lot of states will get revised from liveness point of view when
9265 * the verifier explores other branches.
9268 * 2: if r1 == 100 goto pc+1
9271 * when the verifier reaches exit insn the register r0 in the state list of
9272 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
9273 * of insn 2 and goes exploring further. At the insn 4 it will walk the
9274 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
9276 * Since the verifier pushes the branch states as it sees them while exploring
9277 * the program the condition of walking the branch instruction for the second
9278 * time means that all states below this branch were already explored and
9279 * their final liveness markes are already propagated.
9280 * Hence when the verifier completes the search of state list in is_state_visited()
9281 * we can call this clean_live_states() function to mark all liveness states
9282 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
9284 * This function also clears the registers and stack for states that !READ
9285 * to simplify state merging.
9287 * Important note here that walking the same branch instruction in the callee
9288 * doesn't meant that the states are DONE. The verifier has to compare
9291 static void clean_live_states(struct bpf_verifier_env *env, int insn,
9292 struct bpf_verifier_state *cur)
9294 struct bpf_verifier_state_list *sl;
9297 sl = *explored_state(env, insn);
9299 if (sl->state.branches)
9301 if (sl->state.insn_idx != insn ||
9302 sl->state.curframe != cur->curframe)
9304 for (i = 0; i <= cur->curframe; i++)
9305 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
9307 clean_verifier_state(env, &sl->state);
9313 /* Returns true if (rold safe implies rcur safe) */
9314 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
9315 struct idpair *idmap)
9319 if (!(rold->live & REG_LIVE_READ))
9320 /* explored state didn't use this */
9323 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
9325 if (rold->type == PTR_TO_STACK)
9326 /* two stack pointers are equal only if they're pointing to
9327 * the same stack frame, since fp-8 in foo != fp-8 in bar
9329 return equal && rold->frameno == rcur->frameno;
9334 if (rold->type == NOT_INIT)
9335 /* explored state can't have used this */
9337 if (rcur->type == NOT_INIT)
9339 switch (rold->type) {
9341 if (rcur->type == SCALAR_VALUE) {
9342 if (!rold->precise && !rcur->precise)
9344 /* new val must satisfy old val knowledge */
9345 return range_within(rold, rcur) &&
9346 tnum_in(rold->var_off, rcur->var_off);
9348 /* We're trying to use a pointer in place of a scalar.
9349 * Even if the scalar was unbounded, this could lead to
9350 * pointer leaks because scalars are allowed to leak
9351 * while pointers are not. We could make this safe in
9352 * special cases if root is calling us, but it's
9353 * probably not worth the hassle.
9357 case PTR_TO_MAP_VALUE:
9358 /* If the new min/max/var_off satisfy the old ones and
9359 * everything else matches, we are OK.
9360 * 'id' is not compared, since it's only used for maps with
9361 * bpf_spin_lock inside map element and in such cases if
9362 * the rest of the prog is valid for one map element then
9363 * it's valid for all map elements regardless of the key
9364 * used in bpf_map_lookup()
9366 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
9367 range_within(rold, rcur) &&
9368 tnum_in(rold->var_off, rcur->var_off);
9369 case PTR_TO_MAP_VALUE_OR_NULL:
9370 /* a PTR_TO_MAP_VALUE could be safe to use as a
9371 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
9372 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
9373 * checked, doing so could have affected others with the same
9374 * id, and we can't check for that because we lost the id when
9375 * we converted to a PTR_TO_MAP_VALUE.
9377 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
9379 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
9381 /* Check our ids match any regs they're supposed to */
9382 return check_ids(rold->id, rcur->id, idmap);
9383 case PTR_TO_PACKET_META:
9385 if (rcur->type != rold->type)
9387 /* We must have at least as much range as the old ptr
9388 * did, so that any accesses which were safe before are
9389 * still safe. This is true even if old range < old off,
9390 * since someone could have accessed through (ptr - k), or
9391 * even done ptr -= k in a register, to get a safe access.
9393 if (rold->range > rcur->range)
9395 /* If the offsets don't match, we can't trust our alignment;
9396 * nor can we be sure that we won't fall out of range.
9398 if (rold->off != rcur->off)
9400 /* id relations must be preserved */
9401 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
9403 /* new val must satisfy old val knowledge */
9404 return range_within(rold, rcur) &&
9405 tnum_in(rold->var_off, rcur->var_off);
9407 case CONST_PTR_TO_MAP:
9408 case PTR_TO_PACKET_END:
9409 case PTR_TO_FLOW_KEYS:
9411 case PTR_TO_SOCKET_OR_NULL:
9412 case PTR_TO_SOCK_COMMON:
9413 case PTR_TO_SOCK_COMMON_OR_NULL:
9414 case PTR_TO_TCP_SOCK:
9415 case PTR_TO_TCP_SOCK_OR_NULL:
9416 case PTR_TO_XDP_SOCK:
9417 /* Only valid matches are exact, which memcmp() above
9418 * would have accepted
9421 /* Don't know what's going on, just say it's not safe */
9425 /* Shouldn't get here; if we do, say it's not safe */
9430 static bool stacksafe(struct bpf_func_state *old,
9431 struct bpf_func_state *cur,
9432 struct idpair *idmap)
9436 /* walk slots of the explored stack and ignore any additional
9437 * slots in the current stack, since explored(safe) state
9440 for (i = 0; i < old->allocated_stack; i++) {
9441 spi = i / BPF_REG_SIZE;
9443 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
9444 i += BPF_REG_SIZE - 1;
9445 /* explored state didn't use this */
9449 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
9452 /* explored stack has more populated slots than current stack
9453 * and these slots were used
9455 if (i >= cur->allocated_stack)
9458 /* if old state was safe with misc data in the stack
9459 * it will be safe with zero-initialized stack.
9460 * The opposite is not true
9462 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
9463 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
9465 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
9466 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
9467 /* Ex: old explored (safe) state has STACK_SPILL in
9468 * this stack slot, but current has STACK_MISC ->
9469 * this verifier states are not equivalent,
9470 * return false to continue verification of this path
9473 if (i % BPF_REG_SIZE)
9475 if (old->stack[spi].slot_type[0] != STACK_SPILL)
9477 if (!regsafe(&old->stack[spi].spilled_ptr,
9478 &cur->stack[spi].spilled_ptr,
9480 /* when explored and current stack slot are both storing
9481 * spilled registers, check that stored pointers types
9482 * are the same as well.
9483 * Ex: explored safe path could have stored
9484 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
9485 * but current path has stored:
9486 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
9487 * such verifier states are not equivalent.
9488 * return false to continue verification of this path
9495 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
9497 if (old->acquired_refs != cur->acquired_refs)
9499 return !memcmp(old->refs, cur->refs,
9500 sizeof(*old->refs) * old->acquired_refs);
9503 /* compare two verifier states
9505 * all states stored in state_list are known to be valid, since
9506 * verifier reached 'bpf_exit' instruction through them
9508 * this function is called when verifier exploring different branches of
9509 * execution popped from the state stack. If it sees an old state that has
9510 * more strict register state and more strict stack state then this execution
9511 * branch doesn't need to be explored further, since verifier already
9512 * concluded that more strict state leads to valid finish.
9514 * Therefore two states are equivalent if register state is more conservative
9515 * and explored stack state is more conservative than the current one.
9518 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
9519 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
9521 * In other words if current stack state (one being explored) has more
9522 * valid slots than old one that already passed validation, it means
9523 * the verifier can stop exploring and conclude that current state is valid too
9525 * Similarly with registers. If explored state has register type as invalid
9526 * whereas register type in current state is meaningful, it means that
9527 * the current state will reach 'bpf_exit' instruction safely
9529 static bool func_states_equal(struct bpf_func_state *old,
9530 struct bpf_func_state *cur)
9532 struct idpair *idmap;
9536 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
9537 /* If we failed to allocate the idmap, just say it's not safe */
9541 for (i = 0; i < MAX_BPF_REG; i++) {
9542 if (!regsafe(&old->regs[i], &cur->regs[i], idmap))
9546 if (!stacksafe(old, cur, idmap))
9549 if (!refsafe(old, cur))
9557 static bool states_equal(struct bpf_verifier_env *env,
9558 struct bpf_verifier_state *old,
9559 struct bpf_verifier_state *cur)
9563 if (old->curframe != cur->curframe)
9566 /* Verification state from speculative execution simulation
9567 * must never prune a non-speculative execution one.
9569 if (old->speculative && !cur->speculative)
9572 if (old->active_spin_lock != cur->active_spin_lock)
9575 /* for states to be equal callsites have to be the same
9576 * and all frame states need to be equivalent
9578 for (i = 0; i <= old->curframe; i++) {
9579 if (old->frame[i]->callsite != cur->frame[i]->callsite)
9581 if (!func_states_equal(old->frame[i], cur->frame[i]))
9587 /* Return 0 if no propagation happened. Return negative error code if error
9588 * happened. Otherwise, return the propagated bit.
9590 static int propagate_liveness_reg(struct bpf_verifier_env *env,
9591 struct bpf_reg_state *reg,
9592 struct bpf_reg_state *parent_reg)
9594 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
9595 u8 flag = reg->live & REG_LIVE_READ;
9598 /* When comes here, read flags of PARENT_REG or REG could be any of
9599 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
9600 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
9602 if (parent_flag == REG_LIVE_READ64 ||
9603 /* Or if there is no read flag from REG. */
9605 /* Or if the read flag from REG is the same as PARENT_REG. */
9606 parent_flag == flag)
9609 err = mark_reg_read(env, reg, parent_reg, flag);
9616 /* A write screens off any subsequent reads; but write marks come from the
9617 * straight-line code between a state and its parent. When we arrive at an
9618 * equivalent state (jump target or such) we didn't arrive by the straight-line
9619 * code, so read marks in the state must propagate to the parent regardless
9620 * of the state's write marks. That's what 'parent == state->parent' comparison
9621 * in mark_reg_read() is for.
9623 static int propagate_liveness(struct bpf_verifier_env *env,
9624 const struct bpf_verifier_state *vstate,
9625 struct bpf_verifier_state *vparent)
9627 struct bpf_reg_state *state_reg, *parent_reg;
9628 struct bpf_func_state *state, *parent;
9629 int i, frame, err = 0;
9631 if (vparent->curframe != vstate->curframe) {
9632 WARN(1, "propagate_live: parent frame %d current frame %d\n",
9633 vparent->curframe, vstate->curframe);
9636 /* Propagate read liveness of registers... */
9637 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
9638 for (frame = 0; frame <= vstate->curframe; frame++) {
9639 parent = vparent->frame[frame];
9640 state = vstate->frame[frame];
9641 parent_reg = parent->regs;
9642 state_reg = state->regs;
9643 /* We don't need to worry about FP liveness, it's read-only */
9644 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
9645 err = propagate_liveness_reg(env, &state_reg[i],
9649 if (err == REG_LIVE_READ64)
9650 mark_insn_zext(env, &parent_reg[i]);
9653 /* Propagate stack slots. */
9654 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
9655 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
9656 parent_reg = &parent->stack[i].spilled_ptr;
9657 state_reg = &state->stack[i].spilled_ptr;
9658 err = propagate_liveness_reg(env, state_reg,
9667 /* find precise scalars in the previous equivalent state and
9668 * propagate them into the current state
9670 static int propagate_precision(struct bpf_verifier_env *env,
9671 const struct bpf_verifier_state *old)
9673 struct bpf_reg_state *state_reg;
9674 struct bpf_func_state *state;
9677 state = old->frame[old->curframe];
9678 state_reg = state->regs;
9679 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
9680 if (state_reg->type != SCALAR_VALUE ||
9681 !state_reg->precise)
9683 if (env->log.level & BPF_LOG_LEVEL2)
9684 verbose(env, "propagating r%d\n", i);
9685 err = mark_chain_precision(env, i);
9690 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
9691 if (state->stack[i].slot_type[0] != STACK_SPILL)
9693 state_reg = &state->stack[i].spilled_ptr;
9694 if (state_reg->type != SCALAR_VALUE ||
9695 !state_reg->precise)
9697 if (env->log.level & BPF_LOG_LEVEL2)
9698 verbose(env, "propagating fp%d\n",
9699 (-i - 1) * BPF_REG_SIZE);
9700 err = mark_chain_precision_stack(env, i);
9707 static bool states_maybe_looping(struct bpf_verifier_state *old,
9708 struct bpf_verifier_state *cur)
9710 struct bpf_func_state *fold, *fcur;
9711 int i, fr = cur->curframe;
9713 if (old->curframe != fr)
9716 fold = old->frame[fr];
9717 fcur = cur->frame[fr];
9718 for (i = 0; i < MAX_BPF_REG; i++)
9719 if (memcmp(&fold->regs[i], &fcur->regs[i],
9720 offsetof(struct bpf_reg_state, parent)))
9726 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
9728 struct bpf_verifier_state_list *new_sl;
9729 struct bpf_verifier_state_list *sl, **pprev;
9730 struct bpf_verifier_state *cur = env->cur_state, *new;
9731 int i, j, err, states_cnt = 0;
9732 bool add_new_state = env->test_state_freq ? true : false;
9734 cur->last_insn_idx = env->prev_insn_idx;
9735 if (!env->insn_aux_data[insn_idx].prune_point)
9736 /* this 'insn_idx' instruction wasn't marked, so we will not
9737 * be doing state search here
9741 /* bpf progs typically have pruning point every 4 instructions
9742 * http://vger.kernel.org/bpfconf2019.html#session-1
9743 * Do not add new state for future pruning if the verifier hasn't seen
9744 * at least 2 jumps and at least 8 instructions.
9745 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
9746 * In tests that amounts to up to 50% reduction into total verifier
9747 * memory consumption and 20% verifier time speedup.
9749 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
9750 env->insn_processed - env->prev_insn_processed >= 8)
9751 add_new_state = true;
9753 pprev = explored_state(env, insn_idx);
9756 clean_live_states(env, insn_idx, cur);
9760 if (sl->state.insn_idx != insn_idx)
9762 if (sl->state.branches) {
9763 if (states_maybe_looping(&sl->state, cur) &&
9764 states_equal(env, &sl->state, cur)) {
9765 verbose_linfo(env, insn_idx, "; ");
9766 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
9769 /* if the verifier is processing a loop, avoid adding new state
9770 * too often, since different loop iterations have distinct
9771 * states and may not help future pruning.
9772 * This threshold shouldn't be too low to make sure that
9773 * a loop with large bound will be rejected quickly.
9774 * The most abusive loop will be:
9776 * if r1 < 1000000 goto pc-2
9777 * 1M insn_procssed limit / 100 == 10k peak states.
9778 * This threshold shouldn't be too high either, since states
9779 * at the end of the loop are likely to be useful in pruning.
9781 if (env->jmps_processed - env->prev_jmps_processed < 20 &&
9782 env->insn_processed - env->prev_insn_processed < 100)
9783 add_new_state = false;
9786 if (states_equal(env, &sl->state, cur)) {
9788 /* reached equivalent register/stack state,
9790 * Registers read by the continuation are read by us.
9791 * If we have any write marks in env->cur_state, they
9792 * will prevent corresponding reads in the continuation
9793 * from reaching our parent (an explored_state). Our
9794 * own state will get the read marks recorded, but
9795 * they'll be immediately forgotten as we're pruning
9796 * this state and will pop a new one.
9798 err = propagate_liveness(env, &sl->state, cur);
9800 /* if previous state reached the exit with precision and
9801 * current state is equivalent to it (except precsion marks)
9802 * the precision needs to be propagated back in
9803 * the current state.
9805 err = err ? : push_jmp_history(env, cur);
9806 err = err ? : propagate_precision(env, &sl->state);
9812 /* when new state is not going to be added do not increase miss count.
9813 * Otherwise several loop iterations will remove the state
9814 * recorded earlier. The goal of these heuristics is to have
9815 * states from some iterations of the loop (some in the beginning
9816 * and some at the end) to help pruning.
9820 /* heuristic to determine whether this state is beneficial
9821 * to keep checking from state equivalence point of view.
9822 * Higher numbers increase max_states_per_insn and verification time,
9823 * but do not meaningfully decrease insn_processed.
9825 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
9826 /* the state is unlikely to be useful. Remove it to
9827 * speed up verification
9830 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
9831 u32 br = sl->state.branches;
9834 "BUG live_done but branches_to_explore %d\n",
9836 free_verifier_state(&sl->state, false);
9840 /* cannot free this state, since parentage chain may
9841 * walk it later. Add it for free_list instead to
9842 * be freed at the end of verification
9844 sl->next = env->free_list;
9845 env->free_list = sl;
9855 if (env->max_states_per_insn < states_cnt)
9856 env->max_states_per_insn = states_cnt;
9858 if (!env->bpf_capable && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
9859 return push_jmp_history(env, cur);
9862 return push_jmp_history(env, cur);
9864 /* There were no equivalent states, remember the current one.
9865 * Technically the current state is not proven to be safe yet,
9866 * but it will either reach outer most bpf_exit (which means it's safe)
9867 * or it will be rejected. When there are no loops the verifier won't be
9868 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
9869 * again on the way to bpf_exit.
9870 * When looping the sl->state.branches will be > 0 and this state
9871 * will not be considered for equivalence until branches == 0.
9873 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
9876 env->total_states++;
9878 env->prev_jmps_processed = env->jmps_processed;
9879 env->prev_insn_processed = env->insn_processed;
9881 /* add new state to the head of linked list */
9882 new = &new_sl->state;
9883 err = copy_verifier_state(new, cur);
9885 free_verifier_state(new, false);
9889 new->insn_idx = insn_idx;
9890 WARN_ONCE(new->branches != 1,
9891 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
9894 cur->first_insn_idx = insn_idx;
9895 clear_jmp_history(cur);
9896 new_sl->next = *explored_state(env, insn_idx);
9897 *explored_state(env, insn_idx) = new_sl;
9898 /* connect new state to parentage chain. Current frame needs all
9899 * registers connected. Only r6 - r9 of the callers are alive (pushed
9900 * to the stack implicitly by JITs) so in callers' frames connect just
9901 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
9902 * the state of the call instruction (with WRITTEN set), and r0 comes
9903 * from callee with its full parentage chain, anyway.
9905 /* clear write marks in current state: the writes we did are not writes
9906 * our child did, so they don't screen off its reads from us.
9907 * (There are no read marks in current state, because reads always mark
9908 * their parent and current state never has children yet. Only
9909 * explored_states can get read marks.)
9911 for (j = 0; j <= cur->curframe; j++) {
9912 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
9913 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
9914 for (i = 0; i < BPF_REG_FP; i++)
9915 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
9918 /* all stack frames are accessible from callee, clear them all */
9919 for (j = 0; j <= cur->curframe; j++) {
9920 struct bpf_func_state *frame = cur->frame[j];
9921 struct bpf_func_state *newframe = new->frame[j];
9923 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
9924 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
9925 frame->stack[i].spilled_ptr.parent =
9926 &newframe->stack[i].spilled_ptr;
9932 /* Return true if it's OK to have the same insn return a different type. */
9933 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
9938 case PTR_TO_SOCKET_OR_NULL:
9939 case PTR_TO_SOCK_COMMON:
9940 case PTR_TO_SOCK_COMMON_OR_NULL:
9941 case PTR_TO_TCP_SOCK:
9942 case PTR_TO_TCP_SOCK_OR_NULL:
9943 case PTR_TO_XDP_SOCK:
9945 case PTR_TO_BTF_ID_OR_NULL:
9952 /* If an instruction was previously used with particular pointer types, then we
9953 * need to be careful to avoid cases such as the below, where it may be ok
9954 * for one branch accessing the pointer, but not ok for the other branch:
9959 * R1 = some_other_valid_ptr;
9962 * R2 = *(u32 *)(R1 + 0);
9964 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
9966 return src != prev && (!reg_type_mismatch_ok(src) ||
9967 !reg_type_mismatch_ok(prev));
9970 static int do_check(struct bpf_verifier_env *env)
9972 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
9973 struct bpf_verifier_state *state = env->cur_state;
9974 struct bpf_insn *insns = env->prog->insnsi;
9975 struct bpf_reg_state *regs;
9976 int insn_cnt = env->prog->len;
9977 bool do_print_state = false;
9978 int prev_insn_idx = -1;
9981 struct bpf_insn *insn;
9985 env->prev_insn_idx = prev_insn_idx;
9986 if (env->insn_idx >= insn_cnt) {
9987 verbose(env, "invalid insn idx %d insn_cnt %d\n",
9988 env->insn_idx, insn_cnt);
9992 insn = &insns[env->insn_idx];
9993 class = BPF_CLASS(insn->code);
9995 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
9997 "BPF program is too large. Processed %d insn\n",
9998 env->insn_processed);
10002 err = is_state_visited(env, env->insn_idx);
10006 /* found equivalent state, can prune the search */
10007 if (env->log.level & BPF_LOG_LEVEL) {
10008 if (do_print_state)
10009 verbose(env, "\nfrom %d to %d%s: safe\n",
10010 env->prev_insn_idx, env->insn_idx,
10011 env->cur_state->speculative ?
10012 " (speculative execution)" : "");
10014 verbose(env, "%d: safe\n", env->insn_idx);
10016 goto process_bpf_exit;
10019 if (signal_pending(current))
10022 if (need_resched())
10025 if (env->log.level & BPF_LOG_LEVEL2 ||
10026 (env->log.level & BPF_LOG_LEVEL && do_print_state)) {
10027 if (env->log.level & BPF_LOG_LEVEL2)
10028 verbose(env, "%d:", env->insn_idx);
10030 verbose(env, "\nfrom %d to %d%s:",
10031 env->prev_insn_idx, env->insn_idx,
10032 env->cur_state->speculative ?
10033 " (speculative execution)" : "");
10034 print_verifier_state(env, state->frame[state->curframe]);
10035 do_print_state = false;
10038 if (env->log.level & BPF_LOG_LEVEL) {
10039 const struct bpf_insn_cbs cbs = {
10040 .cb_print = verbose,
10041 .private_data = env,
10044 verbose_linfo(env, env->insn_idx, "; ");
10045 verbose(env, "%d: ", env->insn_idx);
10046 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
10049 if (bpf_prog_is_dev_bound(env->prog->aux)) {
10050 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
10051 env->prev_insn_idx);
10056 regs = cur_regs(env);
10057 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
10058 prev_insn_idx = env->insn_idx;
10060 if (class == BPF_ALU || class == BPF_ALU64) {
10061 err = check_alu_op(env, insn);
10065 } else if (class == BPF_LDX) {
10066 enum bpf_reg_type *prev_src_type, src_reg_type;
10068 /* check for reserved fields is already done */
10070 /* check src operand */
10071 err = check_reg_arg(env, insn->src_reg, SRC_OP);
10075 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
10079 src_reg_type = regs[insn->src_reg].type;
10081 /* check that memory (src_reg + off) is readable,
10082 * the state of dst_reg will be updated by this func
10084 err = check_mem_access(env, env->insn_idx, insn->src_reg,
10085 insn->off, BPF_SIZE(insn->code),
10086 BPF_READ, insn->dst_reg, false);
10090 prev_src_type = &env->insn_aux_data[env->insn_idx].ptr_type;
10092 if (*prev_src_type == NOT_INIT) {
10093 /* saw a valid insn
10094 * dst_reg = *(u32 *)(src_reg + off)
10095 * save type to validate intersecting paths
10097 *prev_src_type = src_reg_type;
10099 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
10100 /* ABuser program is trying to use the same insn
10101 * dst_reg = *(u32*) (src_reg + off)
10102 * with different pointer types:
10103 * src_reg == ctx in one branch and
10104 * src_reg == stack|map in some other branch.
10107 verbose(env, "same insn cannot be used with different pointers\n");
10111 } else if (class == BPF_STX) {
10112 enum bpf_reg_type *prev_dst_type, dst_reg_type;
10114 if (BPF_MODE(insn->code) == BPF_ATOMIC) {
10115 err = check_atomic(env, env->insn_idx, insn);
10122 if (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0) {
10123 verbose(env, "BPF_STX uses reserved fields\n");
10127 /* check src1 operand */
10128 err = check_reg_arg(env, insn->src_reg, SRC_OP);
10131 /* check src2 operand */
10132 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
10136 dst_reg_type = regs[insn->dst_reg].type;
10138 /* check that memory (dst_reg + off) is writeable */
10139 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
10140 insn->off, BPF_SIZE(insn->code),
10141 BPF_WRITE, insn->src_reg, false);
10145 prev_dst_type = &env->insn_aux_data[env->insn_idx].ptr_type;
10147 if (*prev_dst_type == NOT_INIT) {
10148 *prev_dst_type = dst_reg_type;
10149 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
10150 verbose(env, "same insn cannot be used with different pointers\n");
10154 } else if (class == BPF_ST) {
10155 if (BPF_MODE(insn->code) != BPF_MEM ||
10156 insn->src_reg != BPF_REG_0) {
10157 verbose(env, "BPF_ST uses reserved fields\n");
10160 /* check src operand */
10161 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
10165 if (is_ctx_reg(env, insn->dst_reg)) {
10166 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
10168 reg_type_str[reg_state(env, insn->dst_reg)->type]);
10172 /* check that memory (dst_reg + off) is writeable */
10173 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
10174 insn->off, BPF_SIZE(insn->code),
10175 BPF_WRITE, -1, false);
10179 } else if (class == BPF_JMP || class == BPF_JMP32) {
10180 u8 opcode = BPF_OP(insn->code);
10182 env->jmps_processed++;
10183 if (opcode == BPF_CALL) {
10184 if (BPF_SRC(insn->code) != BPF_K ||
10186 (insn->src_reg != BPF_REG_0 &&
10187 insn->src_reg != BPF_PSEUDO_CALL) ||
10188 insn->dst_reg != BPF_REG_0 ||
10189 class == BPF_JMP32) {
10190 verbose(env, "BPF_CALL uses reserved fields\n");
10194 if (env->cur_state->active_spin_lock &&
10195 (insn->src_reg == BPF_PSEUDO_CALL ||
10196 insn->imm != BPF_FUNC_spin_unlock)) {
10197 verbose(env, "function calls are not allowed while holding a lock\n");
10200 if (insn->src_reg == BPF_PSEUDO_CALL)
10201 err = check_func_call(env, insn, &env->insn_idx);
10203 err = check_helper_call(env, insn->imm, env->insn_idx);
10207 } else if (opcode == BPF_JA) {
10208 if (BPF_SRC(insn->code) != BPF_K ||
10210 insn->src_reg != BPF_REG_0 ||
10211 insn->dst_reg != BPF_REG_0 ||
10212 class == BPF_JMP32) {
10213 verbose(env, "BPF_JA uses reserved fields\n");
10217 env->insn_idx += insn->off + 1;
10220 } else if (opcode == BPF_EXIT) {
10221 if (BPF_SRC(insn->code) != BPF_K ||
10223 insn->src_reg != BPF_REG_0 ||
10224 insn->dst_reg != BPF_REG_0 ||
10225 class == BPF_JMP32) {
10226 verbose(env, "BPF_EXIT uses reserved fields\n");
10230 if (env->cur_state->active_spin_lock) {
10231 verbose(env, "bpf_spin_unlock is missing\n");
10235 if (state->curframe) {
10236 /* exit from nested function */
10237 err = prepare_func_exit(env, &env->insn_idx);
10240 do_print_state = true;
10244 err = check_reference_leak(env);
10248 err = check_return_code(env);
10252 update_branch_counts(env, env->cur_state);
10253 err = pop_stack(env, &prev_insn_idx,
10254 &env->insn_idx, pop_log);
10256 if (err != -ENOENT)
10260 do_print_state = true;
10264 err = check_cond_jmp_op(env, insn, &env->insn_idx);
10268 } else if (class == BPF_LD) {
10269 u8 mode = BPF_MODE(insn->code);
10271 if (mode == BPF_ABS || mode == BPF_IND) {
10272 err = check_ld_abs(env, insn);
10276 } else if (mode == BPF_IMM) {
10277 err = check_ld_imm(env, insn);
10282 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
10284 verbose(env, "invalid BPF_LD mode\n");
10288 verbose(env, "unknown insn class %d\n", class);
10298 static int find_btf_percpu_datasec(struct btf *btf)
10300 const struct btf_type *t;
10305 * Both vmlinux and module each have their own ".data..percpu"
10306 * DATASECs in BTF. So for module's case, we need to skip vmlinux BTF
10307 * types to look at only module's own BTF types.
10309 n = btf_nr_types(btf);
10310 if (btf_is_module(btf))
10311 i = btf_nr_types(btf_vmlinux);
10315 for(; i < n; i++) {
10316 t = btf_type_by_id(btf, i);
10317 if (BTF_INFO_KIND(t->info) != BTF_KIND_DATASEC)
10320 tname = btf_name_by_offset(btf, t->name_off);
10321 if (!strcmp(tname, ".data..percpu"))
10328 /* replace pseudo btf_id with kernel symbol address */
10329 static int check_pseudo_btf_id(struct bpf_verifier_env *env,
10330 struct bpf_insn *insn,
10331 struct bpf_insn_aux_data *aux)
10333 const struct btf_var_secinfo *vsi;
10334 const struct btf_type *datasec;
10335 struct btf_mod_pair *btf_mod;
10336 const struct btf_type *t;
10337 const char *sym_name;
10338 bool percpu = false;
10339 u32 type, id = insn->imm;
10343 int i, btf_fd, err;
10345 btf_fd = insn[1].imm;
10347 btf = btf_get_by_fd(btf_fd);
10349 verbose(env, "invalid module BTF object FD specified.\n");
10353 if (!btf_vmlinux) {
10354 verbose(env, "kernel is missing BTF, make sure CONFIG_DEBUG_INFO_BTF=y is specified in Kconfig.\n");
10361 t = btf_type_by_id(btf, id);
10363 verbose(env, "ldimm64 insn specifies invalid btf_id %d.\n", id);
10368 if (!btf_type_is_var(t)) {
10369 verbose(env, "pseudo btf_id %d in ldimm64 isn't KIND_VAR.\n", id);
10374 sym_name = btf_name_by_offset(btf, t->name_off);
10375 addr = kallsyms_lookup_name(sym_name);
10377 verbose(env, "ldimm64 failed to find the address for kernel symbol '%s'.\n",
10383 datasec_id = find_btf_percpu_datasec(btf);
10384 if (datasec_id > 0) {
10385 datasec = btf_type_by_id(btf, datasec_id);
10386 for_each_vsi(i, datasec, vsi) {
10387 if (vsi->type == id) {
10394 insn[0].imm = (u32)addr;
10395 insn[1].imm = addr >> 32;
10398 t = btf_type_skip_modifiers(btf, type, NULL);
10400 aux->btf_var.reg_type = PTR_TO_PERCPU_BTF_ID;
10401 aux->btf_var.btf = btf;
10402 aux->btf_var.btf_id = type;
10403 } else if (!btf_type_is_struct(t)) {
10404 const struct btf_type *ret;
10408 /* resolve the type size of ksym. */
10409 ret = btf_resolve_size(btf, t, &tsize);
10411 tname = btf_name_by_offset(btf, t->name_off);
10412 verbose(env, "ldimm64 unable to resolve the size of type '%s': %ld\n",
10413 tname, PTR_ERR(ret));
10417 aux->btf_var.reg_type = PTR_TO_MEM;
10418 aux->btf_var.mem_size = tsize;
10420 aux->btf_var.reg_type = PTR_TO_BTF_ID;
10421 aux->btf_var.btf = btf;
10422 aux->btf_var.btf_id = type;
10425 /* check whether we recorded this BTF (and maybe module) already */
10426 for (i = 0; i < env->used_btf_cnt; i++) {
10427 if (env->used_btfs[i].btf == btf) {
10433 if (env->used_btf_cnt >= MAX_USED_BTFS) {
10438 btf_mod = &env->used_btfs[env->used_btf_cnt];
10439 btf_mod->btf = btf;
10440 btf_mod->module = NULL;
10442 /* if we reference variables from kernel module, bump its refcount */
10443 if (btf_is_module(btf)) {
10444 btf_mod->module = btf_try_get_module(btf);
10445 if (!btf_mod->module) {
10451 env->used_btf_cnt++;
10459 static int check_map_prealloc(struct bpf_map *map)
10461 return (map->map_type != BPF_MAP_TYPE_HASH &&
10462 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
10463 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
10464 !(map->map_flags & BPF_F_NO_PREALLOC);
10467 static bool is_tracing_prog_type(enum bpf_prog_type type)
10470 case BPF_PROG_TYPE_KPROBE:
10471 case BPF_PROG_TYPE_TRACEPOINT:
10472 case BPF_PROG_TYPE_PERF_EVENT:
10473 case BPF_PROG_TYPE_RAW_TRACEPOINT:
10480 static bool is_preallocated_map(struct bpf_map *map)
10482 if (!check_map_prealloc(map))
10484 if (map->inner_map_meta && !check_map_prealloc(map->inner_map_meta))
10489 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
10490 struct bpf_map *map,
10491 struct bpf_prog *prog)
10494 enum bpf_prog_type prog_type = resolve_prog_type(prog);
10496 * Validate that trace type programs use preallocated hash maps.
10498 * For programs attached to PERF events this is mandatory as the
10499 * perf NMI can hit any arbitrary code sequence.
10501 * All other trace types using preallocated hash maps are unsafe as
10502 * well because tracepoint or kprobes can be inside locked regions
10503 * of the memory allocator or at a place where a recursion into the
10504 * memory allocator would see inconsistent state.
10506 * On RT enabled kernels run-time allocation of all trace type
10507 * programs is strictly prohibited due to lock type constraints. On
10508 * !RT kernels it is allowed for backwards compatibility reasons for
10509 * now, but warnings are emitted so developers are made aware of
10510 * the unsafety and can fix their programs before this is enforced.
10512 if (is_tracing_prog_type(prog_type) && !is_preallocated_map(map)) {
10513 if (prog_type == BPF_PROG_TYPE_PERF_EVENT) {
10514 verbose(env, "perf_event programs can only use preallocated hash map\n");
10517 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
10518 verbose(env, "trace type programs can only use preallocated hash map\n");
10521 WARN_ONCE(1, "trace type BPF program uses run-time allocation\n");
10522 verbose(env, "trace type programs with run-time allocated hash maps are unsafe. Switch to preallocated hash maps.\n");
10525 if (map_value_has_spin_lock(map)) {
10526 if (prog_type == BPF_PROG_TYPE_SOCKET_FILTER) {
10527 verbose(env, "socket filter progs cannot use bpf_spin_lock yet\n");
10531 if (is_tracing_prog_type(prog_type)) {
10532 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
10536 if (prog->aux->sleepable) {
10537 verbose(env, "sleepable progs cannot use bpf_spin_lock yet\n");
10542 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
10543 !bpf_offload_prog_map_match(prog, map)) {
10544 verbose(env, "offload device mismatch between prog and map\n");
10548 if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
10549 verbose(env, "bpf_struct_ops map cannot be used in prog\n");
10553 if (prog->aux->sleepable)
10554 switch (map->map_type) {
10555 case BPF_MAP_TYPE_HASH:
10556 case BPF_MAP_TYPE_LRU_HASH:
10557 case BPF_MAP_TYPE_ARRAY:
10558 case BPF_MAP_TYPE_PERCPU_HASH:
10559 case BPF_MAP_TYPE_PERCPU_ARRAY:
10560 case BPF_MAP_TYPE_LRU_PERCPU_HASH:
10561 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
10562 case BPF_MAP_TYPE_HASH_OF_MAPS:
10563 if (!is_preallocated_map(map)) {
10565 "Sleepable programs can only use preallocated maps\n");
10569 case BPF_MAP_TYPE_RINGBUF:
10573 "Sleepable programs can only use array, hash, and ringbuf maps\n");
10580 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
10582 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
10583 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
10586 /* find and rewrite pseudo imm in ld_imm64 instructions:
10588 * 1. if it accesses map FD, replace it with actual map pointer.
10589 * 2. if it accesses btf_id of a VAR, replace it with pointer to the var.
10591 * NOTE: btf_vmlinux is required for converting pseudo btf_id.
10593 static int resolve_pseudo_ldimm64(struct bpf_verifier_env *env)
10595 struct bpf_insn *insn = env->prog->insnsi;
10596 int insn_cnt = env->prog->len;
10599 err = bpf_prog_calc_tag(env->prog);
10603 for (i = 0; i < insn_cnt; i++, insn++) {
10604 if (BPF_CLASS(insn->code) == BPF_LDX &&
10605 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
10606 verbose(env, "BPF_LDX uses reserved fields\n");
10610 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
10611 struct bpf_insn_aux_data *aux;
10612 struct bpf_map *map;
10616 if (i == insn_cnt - 1 || insn[1].code != 0 ||
10617 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
10618 insn[1].off != 0) {
10619 verbose(env, "invalid bpf_ld_imm64 insn\n");
10623 if (insn[0].src_reg == 0)
10624 /* valid generic load 64-bit imm */
10627 if (insn[0].src_reg == BPF_PSEUDO_BTF_ID) {
10628 aux = &env->insn_aux_data[i];
10629 err = check_pseudo_btf_id(env, insn, aux);
10635 /* In final convert_pseudo_ld_imm64() step, this is
10636 * converted into regular 64-bit imm load insn.
10638 if ((insn[0].src_reg != BPF_PSEUDO_MAP_FD &&
10639 insn[0].src_reg != BPF_PSEUDO_MAP_VALUE) ||
10640 (insn[0].src_reg == BPF_PSEUDO_MAP_FD &&
10641 insn[1].imm != 0)) {
10643 "unrecognized bpf_ld_imm64 insn\n");
10647 f = fdget(insn[0].imm);
10648 map = __bpf_map_get(f);
10650 verbose(env, "fd %d is not pointing to valid bpf_map\n",
10652 return PTR_ERR(map);
10655 err = check_map_prog_compatibility(env, map, env->prog);
10661 aux = &env->insn_aux_data[i];
10662 if (insn->src_reg == BPF_PSEUDO_MAP_FD) {
10663 addr = (unsigned long)map;
10665 u32 off = insn[1].imm;
10667 if (off >= BPF_MAX_VAR_OFF) {
10668 verbose(env, "direct value offset of %u is not allowed\n", off);
10673 if (!map->ops->map_direct_value_addr) {
10674 verbose(env, "no direct value access support for this map type\n");
10679 err = map->ops->map_direct_value_addr(map, &addr, off);
10681 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
10682 map->value_size, off);
10687 aux->map_off = off;
10691 insn[0].imm = (u32)addr;
10692 insn[1].imm = addr >> 32;
10694 /* check whether we recorded this map already */
10695 for (j = 0; j < env->used_map_cnt; j++) {
10696 if (env->used_maps[j] == map) {
10697 aux->map_index = j;
10703 if (env->used_map_cnt >= MAX_USED_MAPS) {
10708 /* hold the map. If the program is rejected by verifier,
10709 * the map will be released by release_maps() or it
10710 * will be used by the valid program until it's unloaded
10711 * and all maps are released in free_used_maps()
10715 aux->map_index = env->used_map_cnt;
10716 env->used_maps[env->used_map_cnt++] = map;
10718 if (bpf_map_is_cgroup_storage(map) &&
10719 bpf_cgroup_storage_assign(env->prog->aux, map)) {
10720 verbose(env, "only one cgroup storage of each type is allowed\n");
10732 /* Basic sanity check before we invest more work here. */
10733 if (!bpf_opcode_in_insntable(insn->code)) {
10734 verbose(env, "unknown opcode %02x\n", insn->code);
10739 /* now all pseudo BPF_LD_IMM64 instructions load valid
10740 * 'struct bpf_map *' into a register instead of user map_fd.
10741 * These pointers will be used later by verifier to validate map access.
10746 /* drop refcnt of maps used by the rejected program */
10747 static void release_maps(struct bpf_verifier_env *env)
10749 __bpf_free_used_maps(env->prog->aux, env->used_maps,
10750 env->used_map_cnt);
10753 /* drop refcnt of maps used by the rejected program */
10754 static void release_btfs(struct bpf_verifier_env *env)
10756 __bpf_free_used_btfs(env->prog->aux, env->used_btfs,
10757 env->used_btf_cnt);
10760 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
10761 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
10763 struct bpf_insn *insn = env->prog->insnsi;
10764 int insn_cnt = env->prog->len;
10767 for (i = 0; i < insn_cnt; i++, insn++)
10768 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
10772 /* single env->prog->insni[off] instruction was replaced with the range
10773 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
10774 * [0, off) and [off, end) to new locations, so the patched range stays zero
10776 static int adjust_insn_aux_data(struct bpf_verifier_env *env,
10777 struct bpf_prog *new_prog, u32 off, u32 cnt)
10779 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
10780 struct bpf_insn *insn = new_prog->insnsi;
10784 /* aux info at OFF always needs adjustment, no matter fast path
10785 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
10786 * original insn at old prog.
10788 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
10792 prog_len = new_prog->len;
10793 new_data = vzalloc(array_size(prog_len,
10794 sizeof(struct bpf_insn_aux_data)));
10797 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
10798 memcpy(new_data + off + cnt - 1, old_data + off,
10799 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
10800 for (i = off; i < off + cnt - 1; i++) {
10801 new_data[i].seen = env->pass_cnt;
10802 new_data[i].zext_dst = insn_has_def32(env, insn + i);
10804 env->insn_aux_data = new_data;
10809 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
10815 /* NOTE: fake 'exit' subprog should be updated as well. */
10816 for (i = 0; i <= env->subprog_cnt; i++) {
10817 if (env->subprog_info[i].start <= off)
10819 env->subprog_info[i].start += len - 1;
10823 static void adjust_poke_descs(struct bpf_prog *prog, u32 len)
10825 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
10826 int i, sz = prog->aux->size_poke_tab;
10827 struct bpf_jit_poke_descriptor *desc;
10829 for (i = 0; i < sz; i++) {
10831 desc->insn_idx += len - 1;
10835 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
10836 const struct bpf_insn *patch, u32 len)
10838 struct bpf_prog *new_prog;
10840 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
10841 if (IS_ERR(new_prog)) {
10842 if (PTR_ERR(new_prog) == -ERANGE)
10844 "insn %d cannot be patched due to 16-bit range\n",
10845 env->insn_aux_data[off].orig_idx);
10848 if (adjust_insn_aux_data(env, new_prog, off, len))
10850 adjust_subprog_starts(env, off, len);
10851 adjust_poke_descs(new_prog, len);
10855 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
10860 /* find first prog starting at or after off (first to remove) */
10861 for (i = 0; i < env->subprog_cnt; i++)
10862 if (env->subprog_info[i].start >= off)
10864 /* find first prog starting at or after off + cnt (first to stay) */
10865 for (j = i; j < env->subprog_cnt; j++)
10866 if (env->subprog_info[j].start >= off + cnt)
10868 /* if j doesn't start exactly at off + cnt, we are just removing
10869 * the front of previous prog
10871 if (env->subprog_info[j].start != off + cnt)
10875 struct bpf_prog_aux *aux = env->prog->aux;
10878 /* move fake 'exit' subprog as well */
10879 move = env->subprog_cnt + 1 - j;
10881 memmove(env->subprog_info + i,
10882 env->subprog_info + j,
10883 sizeof(*env->subprog_info) * move);
10884 env->subprog_cnt -= j - i;
10886 /* remove func_info */
10887 if (aux->func_info) {
10888 move = aux->func_info_cnt - j;
10890 memmove(aux->func_info + i,
10891 aux->func_info + j,
10892 sizeof(*aux->func_info) * move);
10893 aux->func_info_cnt -= j - i;
10894 /* func_info->insn_off is set after all code rewrites,
10895 * in adjust_btf_func() - no need to adjust
10899 /* convert i from "first prog to remove" to "first to adjust" */
10900 if (env->subprog_info[i].start == off)
10904 /* update fake 'exit' subprog as well */
10905 for (; i <= env->subprog_cnt; i++)
10906 env->subprog_info[i].start -= cnt;
10911 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
10914 struct bpf_prog *prog = env->prog;
10915 u32 i, l_off, l_cnt, nr_linfo;
10916 struct bpf_line_info *linfo;
10918 nr_linfo = prog->aux->nr_linfo;
10922 linfo = prog->aux->linfo;
10924 /* find first line info to remove, count lines to be removed */
10925 for (i = 0; i < nr_linfo; i++)
10926 if (linfo[i].insn_off >= off)
10931 for (; i < nr_linfo; i++)
10932 if (linfo[i].insn_off < off + cnt)
10937 /* First live insn doesn't match first live linfo, it needs to "inherit"
10938 * last removed linfo. prog is already modified, so prog->len == off
10939 * means no live instructions after (tail of the program was removed).
10941 if (prog->len != off && l_cnt &&
10942 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
10944 linfo[--i].insn_off = off + cnt;
10947 /* remove the line info which refer to the removed instructions */
10949 memmove(linfo + l_off, linfo + i,
10950 sizeof(*linfo) * (nr_linfo - i));
10952 prog->aux->nr_linfo -= l_cnt;
10953 nr_linfo = prog->aux->nr_linfo;
10956 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
10957 for (i = l_off; i < nr_linfo; i++)
10958 linfo[i].insn_off -= cnt;
10960 /* fix up all subprogs (incl. 'exit') which start >= off */
10961 for (i = 0; i <= env->subprog_cnt; i++)
10962 if (env->subprog_info[i].linfo_idx > l_off) {
10963 /* program may have started in the removed region but
10964 * may not be fully removed
10966 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
10967 env->subprog_info[i].linfo_idx -= l_cnt;
10969 env->subprog_info[i].linfo_idx = l_off;
10975 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
10977 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
10978 unsigned int orig_prog_len = env->prog->len;
10981 if (bpf_prog_is_dev_bound(env->prog->aux))
10982 bpf_prog_offload_remove_insns(env, off, cnt);
10984 err = bpf_remove_insns(env->prog, off, cnt);
10988 err = adjust_subprog_starts_after_remove(env, off, cnt);
10992 err = bpf_adj_linfo_after_remove(env, off, cnt);
10996 memmove(aux_data + off, aux_data + off + cnt,
10997 sizeof(*aux_data) * (orig_prog_len - off - cnt));
11002 /* The verifier does more data flow analysis than llvm and will not
11003 * explore branches that are dead at run time. Malicious programs can
11004 * have dead code too. Therefore replace all dead at-run-time code
11007 * Just nops are not optimal, e.g. if they would sit at the end of the
11008 * program and through another bug we would manage to jump there, then
11009 * we'd execute beyond program memory otherwise. Returning exception
11010 * code also wouldn't work since we can have subprogs where the dead
11011 * code could be located.
11013 static void sanitize_dead_code(struct bpf_verifier_env *env)
11015 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
11016 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
11017 struct bpf_insn *insn = env->prog->insnsi;
11018 const int insn_cnt = env->prog->len;
11021 for (i = 0; i < insn_cnt; i++) {
11022 if (aux_data[i].seen)
11024 memcpy(insn + i, &trap, sizeof(trap));
11028 static bool insn_is_cond_jump(u8 code)
11032 if (BPF_CLASS(code) == BPF_JMP32)
11035 if (BPF_CLASS(code) != BPF_JMP)
11039 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
11042 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
11044 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
11045 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
11046 struct bpf_insn *insn = env->prog->insnsi;
11047 const int insn_cnt = env->prog->len;
11050 for (i = 0; i < insn_cnt; i++, insn++) {
11051 if (!insn_is_cond_jump(insn->code))
11054 if (!aux_data[i + 1].seen)
11055 ja.off = insn->off;
11056 else if (!aux_data[i + 1 + insn->off].seen)
11061 if (bpf_prog_is_dev_bound(env->prog->aux))
11062 bpf_prog_offload_replace_insn(env, i, &ja);
11064 memcpy(insn, &ja, sizeof(ja));
11068 static int opt_remove_dead_code(struct bpf_verifier_env *env)
11070 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
11071 int insn_cnt = env->prog->len;
11074 for (i = 0; i < insn_cnt; i++) {
11078 while (i + j < insn_cnt && !aux_data[i + j].seen)
11083 err = verifier_remove_insns(env, i, j);
11086 insn_cnt = env->prog->len;
11092 static int opt_remove_nops(struct bpf_verifier_env *env)
11094 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
11095 struct bpf_insn *insn = env->prog->insnsi;
11096 int insn_cnt = env->prog->len;
11099 for (i = 0; i < insn_cnt; i++) {
11100 if (memcmp(&insn[i], &ja, sizeof(ja)))
11103 err = verifier_remove_insns(env, i, 1);
11113 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
11114 const union bpf_attr *attr)
11116 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
11117 struct bpf_insn_aux_data *aux = env->insn_aux_data;
11118 int i, patch_len, delta = 0, len = env->prog->len;
11119 struct bpf_insn *insns = env->prog->insnsi;
11120 struct bpf_prog *new_prog;
11123 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
11124 zext_patch[1] = BPF_ZEXT_REG(0);
11125 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
11126 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
11127 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
11128 for (i = 0; i < len; i++) {
11129 int adj_idx = i + delta;
11130 struct bpf_insn insn;
11133 insn = insns[adj_idx];
11134 load_reg = insn_def_regno(&insn);
11135 if (!aux[adj_idx].zext_dst) {
11143 class = BPF_CLASS(code);
11144 if (load_reg == -1)
11147 /* NOTE: arg "reg" (the fourth one) is only used for
11148 * BPF_STX + SRC_OP, so it is safe to pass NULL
11151 if (is_reg64(env, &insn, load_reg, NULL, DST_OP)) {
11152 if (class == BPF_LD &&
11153 BPF_MODE(code) == BPF_IMM)
11158 /* ctx load could be transformed into wider load. */
11159 if (class == BPF_LDX &&
11160 aux[adj_idx].ptr_type == PTR_TO_CTX)
11163 imm_rnd = get_random_int();
11164 rnd_hi32_patch[0] = insn;
11165 rnd_hi32_patch[1].imm = imm_rnd;
11166 rnd_hi32_patch[3].dst_reg = load_reg;
11167 patch = rnd_hi32_patch;
11169 goto apply_patch_buffer;
11172 /* Add in an zero-extend instruction if a) the JIT has requested
11173 * it or b) it's a CMPXCHG.
11175 * The latter is because: BPF_CMPXCHG always loads a value into
11176 * R0, therefore always zero-extends. However some archs'
11177 * equivalent instruction only does this load when the
11178 * comparison is successful. This detail of CMPXCHG is
11179 * orthogonal to the general zero-extension behaviour of the
11180 * CPU, so it's treated independently of bpf_jit_needs_zext.
11182 if (!bpf_jit_needs_zext() && !is_cmpxchg_insn(&insn))
11185 if (WARN_ON(load_reg == -1)) {
11186 verbose(env, "verifier bug. zext_dst is set, but no reg is defined\n");
11190 zext_patch[0] = insn;
11191 zext_patch[1].dst_reg = load_reg;
11192 zext_patch[1].src_reg = load_reg;
11193 patch = zext_patch;
11195 apply_patch_buffer:
11196 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
11199 env->prog = new_prog;
11200 insns = new_prog->insnsi;
11201 aux = env->insn_aux_data;
11202 delta += patch_len - 1;
11208 /* convert load instructions that access fields of a context type into a
11209 * sequence of instructions that access fields of the underlying structure:
11210 * struct __sk_buff -> struct sk_buff
11211 * struct bpf_sock_ops -> struct sock
11213 static int convert_ctx_accesses(struct bpf_verifier_env *env)
11215 const struct bpf_verifier_ops *ops = env->ops;
11216 int i, cnt, size, ctx_field_size, delta = 0;
11217 const int insn_cnt = env->prog->len;
11218 struct bpf_insn insn_buf[16], *insn;
11219 u32 target_size, size_default, off;
11220 struct bpf_prog *new_prog;
11221 enum bpf_access_type type;
11222 bool is_narrower_load;
11224 if (ops->gen_prologue || env->seen_direct_write) {
11225 if (!ops->gen_prologue) {
11226 verbose(env, "bpf verifier is misconfigured\n");
11229 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
11231 if (cnt >= ARRAY_SIZE(insn_buf)) {
11232 verbose(env, "bpf verifier is misconfigured\n");
11235 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
11239 env->prog = new_prog;
11244 if (bpf_prog_is_dev_bound(env->prog->aux))
11247 insn = env->prog->insnsi + delta;
11249 for (i = 0; i < insn_cnt; i++, insn++) {
11250 bpf_convert_ctx_access_t convert_ctx_access;
11252 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
11253 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
11254 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
11255 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
11257 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
11258 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
11259 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
11260 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
11265 if (type == BPF_WRITE &&
11266 env->insn_aux_data[i + delta].sanitize_stack_off) {
11267 struct bpf_insn patch[] = {
11268 /* Sanitize suspicious stack slot with zero.
11269 * There are no memory dependencies for this store,
11270 * since it's only using frame pointer and immediate
11273 BPF_ST_MEM(BPF_DW, BPF_REG_FP,
11274 env->insn_aux_data[i + delta].sanitize_stack_off,
11276 /* the original STX instruction will immediately
11277 * overwrite the same stack slot with appropriate value
11282 cnt = ARRAY_SIZE(patch);
11283 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
11288 env->prog = new_prog;
11289 insn = new_prog->insnsi + i + delta;
11293 switch (env->insn_aux_data[i + delta].ptr_type) {
11295 if (!ops->convert_ctx_access)
11297 convert_ctx_access = ops->convert_ctx_access;
11299 case PTR_TO_SOCKET:
11300 case PTR_TO_SOCK_COMMON:
11301 convert_ctx_access = bpf_sock_convert_ctx_access;
11303 case PTR_TO_TCP_SOCK:
11304 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
11306 case PTR_TO_XDP_SOCK:
11307 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
11309 case PTR_TO_BTF_ID:
11310 if (type == BPF_READ) {
11311 insn->code = BPF_LDX | BPF_PROBE_MEM |
11312 BPF_SIZE((insn)->code);
11313 env->prog->aux->num_exentries++;
11314 } else if (resolve_prog_type(env->prog) != BPF_PROG_TYPE_STRUCT_OPS) {
11315 verbose(env, "Writes through BTF pointers are not allowed\n");
11323 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
11324 size = BPF_LDST_BYTES(insn);
11326 /* If the read access is a narrower load of the field,
11327 * convert to a 4/8-byte load, to minimum program type specific
11328 * convert_ctx_access changes. If conversion is successful,
11329 * we will apply proper mask to the result.
11331 is_narrower_load = size < ctx_field_size;
11332 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
11334 if (is_narrower_load) {
11337 if (type == BPF_WRITE) {
11338 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
11343 if (ctx_field_size == 4)
11345 else if (ctx_field_size == 8)
11346 size_code = BPF_DW;
11348 insn->off = off & ~(size_default - 1);
11349 insn->code = BPF_LDX | BPF_MEM | size_code;
11353 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
11355 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
11356 (ctx_field_size && !target_size)) {
11357 verbose(env, "bpf verifier is misconfigured\n");
11361 if (is_narrower_load && size < target_size) {
11362 u8 shift = bpf_ctx_narrow_access_offset(
11363 off, size, size_default) * 8;
11364 if (ctx_field_size <= 4) {
11366 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
11369 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
11370 (1 << size * 8) - 1);
11373 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
11376 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
11377 (1ULL << size * 8) - 1);
11381 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
11387 /* keep walking new program and skip insns we just inserted */
11388 env->prog = new_prog;
11389 insn = new_prog->insnsi + i + delta;
11395 static int jit_subprogs(struct bpf_verifier_env *env)
11397 struct bpf_prog *prog = env->prog, **func, *tmp;
11398 int i, j, subprog_start, subprog_end = 0, len, subprog;
11399 struct bpf_map *map_ptr;
11400 struct bpf_insn *insn;
11401 void *old_bpf_func;
11402 int err, num_exentries;
11404 if (env->subprog_cnt <= 1)
11407 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
11408 if (!bpf_pseudo_call(insn))
11410 /* Upon error here we cannot fall back to interpreter but
11411 * need a hard reject of the program. Thus -EFAULT is
11412 * propagated in any case.
11414 subprog = find_subprog(env, i + insn->imm + 1);
11416 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
11417 i + insn->imm + 1);
11420 /* temporarily remember subprog id inside insn instead of
11421 * aux_data, since next loop will split up all insns into funcs
11423 insn->off = subprog;
11424 /* remember original imm in case JIT fails and fallback
11425 * to interpreter will be needed
11427 env->insn_aux_data[i].call_imm = insn->imm;
11428 /* point imm to __bpf_call_base+1 from JITs point of view */
11432 err = bpf_prog_alloc_jited_linfo(prog);
11434 goto out_undo_insn;
11437 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
11439 goto out_undo_insn;
11441 for (i = 0; i < env->subprog_cnt; i++) {
11442 subprog_start = subprog_end;
11443 subprog_end = env->subprog_info[i + 1].start;
11445 len = subprog_end - subprog_start;
11446 /* BPF_PROG_RUN doesn't call subprogs directly,
11447 * hence main prog stats include the runtime of subprogs.
11448 * subprogs don't have IDs and not reachable via prog_get_next_id
11449 * func[i]->stats will never be accessed and stays NULL
11451 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
11454 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
11455 len * sizeof(struct bpf_insn));
11456 func[i]->type = prog->type;
11457 func[i]->len = len;
11458 if (bpf_prog_calc_tag(func[i]))
11460 func[i]->is_func = 1;
11461 func[i]->aux->func_idx = i;
11462 /* the btf and func_info will be freed only at prog->aux */
11463 func[i]->aux->btf = prog->aux->btf;
11464 func[i]->aux->func_info = prog->aux->func_info;
11466 for (j = 0; j < prog->aux->size_poke_tab; j++) {
11467 u32 insn_idx = prog->aux->poke_tab[j].insn_idx;
11470 if (!(insn_idx >= subprog_start &&
11471 insn_idx <= subprog_end))
11474 ret = bpf_jit_add_poke_descriptor(func[i],
11475 &prog->aux->poke_tab[j]);
11477 verbose(env, "adding tail call poke descriptor failed\n");
11481 func[i]->insnsi[insn_idx - subprog_start].imm = ret + 1;
11483 map_ptr = func[i]->aux->poke_tab[ret].tail_call.map;
11484 ret = map_ptr->ops->map_poke_track(map_ptr, func[i]->aux);
11486 verbose(env, "tracking tail call prog failed\n");
11491 /* Use bpf_prog_F_tag to indicate functions in stack traces.
11492 * Long term would need debug info to populate names
11494 func[i]->aux->name[0] = 'F';
11495 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
11496 func[i]->jit_requested = 1;
11497 func[i]->aux->linfo = prog->aux->linfo;
11498 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
11499 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
11500 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
11502 insn = func[i]->insnsi;
11503 for (j = 0; j < func[i]->len; j++, insn++) {
11504 if (BPF_CLASS(insn->code) == BPF_LDX &&
11505 BPF_MODE(insn->code) == BPF_PROBE_MEM)
11508 func[i]->aux->num_exentries = num_exentries;
11509 func[i]->aux->tail_call_reachable = env->subprog_info[i].tail_call_reachable;
11510 func[i] = bpf_int_jit_compile(func[i]);
11511 if (!func[i]->jited) {
11518 /* Untrack main program's aux structs so that during map_poke_run()
11519 * we will not stumble upon the unfilled poke descriptors; each
11520 * of the main program's poke descs got distributed across subprogs
11521 * and got tracked onto map, so we are sure that none of them will
11522 * be missed after the operation below
11524 for (i = 0; i < prog->aux->size_poke_tab; i++) {
11525 map_ptr = prog->aux->poke_tab[i].tail_call.map;
11527 map_ptr->ops->map_poke_untrack(map_ptr, prog->aux);
11530 /* at this point all bpf functions were successfully JITed
11531 * now populate all bpf_calls with correct addresses and
11532 * run last pass of JIT
11534 for (i = 0; i < env->subprog_cnt; i++) {
11535 insn = func[i]->insnsi;
11536 for (j = 0; j < func[i]->len; j++, insn++) {
11537 if (!bpf_pseudo_call(insn))
11539 subprog = insn->off;
11540 insn->imm = BPF_CAST_CALL(func[subprog]->bpf_func) -
11544 /* we use the aux data to keep a list of the start addresses
11545 * of the JITed images for each function in the program
11547 * for some architectures, such as powerpc64, the imm field
11548 * might not be large enough to hold the offset of the start
11549 * address of the callee's JITed image from __bpf_call_base
11551 * in such cases, we can lookup the start address of a callee
11552 * by using its subprog id, available from the off field of
11553 * the call instruction, as an index for this list
11555 func[i]->aux->func = func;
11556 func[i]->aux->func_cnt = env->subprog_cnt;
11558 for (i = 0; i < env->subprog_cnt; i++) {
11559 old_bpf_func = func[i]->bpf_func;
11560 tmp = bpf_int_jit_compile(func[i]);
11561 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
11562 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
11569 /* finally lock prog and jit images for all functions and
11570 * populate kallsysm
11572 for (i = 0; i < env->subprog_cnt; i++) {
11573 bpf_prog_lock_ro(func[i]);
11574 bpf_prog_kallsyms_add(func[i]);
11577 /* Last step: make now unused interpreter insns from main
11578 * prog consistent for later dump requests, so they can
11579 * later look the same as if they were interpreted only.
11581 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
11582 if (!bpf_pseudo_call(insn))
11584 insn->off = env->insn_aux_data[i].call_imm;
11585 subprog = find_subprog(env, i + insn->off + 1);
11586 insn->imm = subprog;
11590 prog->bpf_func = func[0]->bpf_func;
11591 prog->aux->func = func;
11592 prog->aux->func_cnt = env->subprog_cnt;
11593 bpf_prog_free_unused_jited_linfo(prog);
11596 for (i = 0; i < env->subprog_cnt; i++) {
11600 for (j = 0; j < func[i]->aux->size_poke_tab; j++) {
11601 map_ptr = func[i]->aux->poke_tab[j].tail_call.map;
11602 map_ptr->ops->map_poke_untrack(map_ptr, func[i]->aux);
11604 bpf_jit_free(func[i]);
11608 /* cleanup main prog to be interpreted */
11609 prog->jit_requested = 0;
11610 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
11611 if (!bpf_pseudo_call(insn))
11614 insn->imm = env->insn_aux_data[i].call_imm;
11616 bpf_prog_free_jited_linfo(prog);
11620 static int fixup_call_args(struct bpf_verifier_env *env)
11622 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
11623 struct bpf_prog *prog = env->prog;
11624 struct bpf_insn *insn = prog->insnsi;
11629 if (env->prog->jit_requested &&
11630 !bpf_prog_is_dev_bound(env->prog->aux)) {
11631 err = jit_subprogs(env);
11634 if (err == -EFAULT)
11637 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
11638 if (env->subprog_cnt > 1 && env->prog->aux->tail_call_reachable) {
11639 /* When JIT fails the progs with bpf2bpf calls and tail_calls
11640 * have to be rejected, since interpreter doesn't support them yet.
11642 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
11645 for (i = 0; i < prog->len; i++, insn++) {
11646 if (!bpf_pseudo_call(insn))
11648 depth = get_callee_stack_depth(env, insn, i);
11651 bpf_patch_call_args(insn, depth);
11658 /* fixup insn->imm field of bpf_call instructions
11659 * and inline eligible helpers as explicit sequence of BPF instructions
11661 * this function is called after eBPF program passed verification
11663 static int fixup_bpf_calls(struct bpf_verifier_env *env)
11665 struct bpf_prog *prog = env->prog;
11666 bool expect_blinding = bpf_jit_blinding_enabled(prog);
11667 struct bpf_insn *insn = prog->insnsi;
11668 const struct bpf_func_proto *fn;
11669 const int insn_cnt = prog->len;
11670 const struct bpf_map_ops *ops;
11671 struct bpf_insn_aux_data *aux;
11672 struct bpf_insn insn_buf[16];
11673 struct bpf_prog *new_prog;
11674 struct bpf_map *map_ptr;
11675 int i, ret, cnt, delta = 0;
11677 for (i = 0; i < insn_cnt; i++, insn++) {
11678 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
11679 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
11680 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
11681 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
11682 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
11683 bool isdiv = BPF_OP(insn->code) == BPF_DIV;
11684 struct bpf_insn *patchlet;
11685 struct bpf_insn chk_and_div[] = {
11686 /* [R,W]x div 0 -> 0 */
11687 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
11688 BPF_JNE | BPF_K, insn->src_reg,
11690 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
11691 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
11694 struct bpf_insn chk_and_mod[] = {
11695 /* [R,W]x mod 0 -> [R,W]x */
11696 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
11697 BPF_JEQ | BPF_K, insn->src_reg,
11698 0, 1 + (is64 ? 0 : 1), 0),
11700 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
11701 BPF_MOV32_REG(insn->dst_reg, insn->dst_reg),
11704 patchlet = isdiv ? chk_and_div : chk_and_mod;
11705 cnt = isdiv ? ARRAY_SIZE(chk_and_div) :
11706 ARRAY_SIZE(chk_and_mod) - (is64 ? 2 : 0);
11708 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
11713 env->prog = prog = new_prog;
11714 insn = new_prog->insnsi + i + delta;
11718 if (BPF_CLASS(insn->code) == BPF_LD &&
11719 (BPF_MODE(insn->code) == BPF_ABS ||
11720 BPF_MODE(insn->code) == BPF_IND)) {
11721 cnt = env->ops->gen_ld_abs(insn, insn_buf);
11722 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
11723 verbose(env, "bpf verifier is misconfigured\n");
11727 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
11732 env->prog = prog = new_prog;
11733 insn = new_prog->insnsi + i + delta;
11737 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
11738 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
11739 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
11740 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
11741 struct bpf_insn insn_buf[16];
11742 struct bpf_insn *patch = &insn_buf[0];
11746 aux = &env->insn_aux_data[i + delta];
11747 if (!aux->alu_state ||
11748 aux->alu_state == BPF_ALU_NON_POINTER)
11751 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
11752 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
11753 BPF_ALU_SANITIZE_SRC;
11755 off_reg = issrc ? insn->src_reg : insn->dst_reg;
11757 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
11758 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
11759 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
11760 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
11761 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
11762 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
11764 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX,
11766 insn->src_reg = BPF_REG_AX;
11768 *patch++ = BPF_ALU64_REG(BPF_AND, off_reg,
11772 insn->code = insn->code == code_add ?
11773 code_sub : code_add;
11775 if (issrc && isneg)
11776 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
11777 cnt = patch - insn_buf;
11779 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
11784 env->prog = prog = new_prog;
11785 insn = new_prog->insnsi + i + delta;
11789 if (insn->code != (BPF_JMP | BPF_CALL))
11791 if (insn->src_reg == BPF_PSEUDO_CALL)
11794 if (insn->imm == BPF_FUNC_get_route_realm)
11795 prog->dst_needed = 1;
11796 if (insn->imm == BPF_FUNC_get_prandom_u32)
11797 bpf_user_rnd_init_once();
11798 if (insn->imm == BPF_FUNC_override_return)
11799 prog->kprobe_override = 1;
11800 if (insn->imm == BPF_FUNC_tail_call) {
11801 /* If we tail call into other programs, we
11802 * cannot make any assumptions since they can
11803 * be replaced dynamically during runtime in
11804 * the program array.
11806 prog->cb_access = 1;
11807 if (!allow_tail_call_in_subprogs(env))
11808 prog->aux->stack_depth = MAX_BPF_STACK;
11809 prog->aux->max_pkt_offset = MAX_PACKET_OFF;
11811 /* mark bpf_tail_call as different opcode to avoid
11812 * conditional branch in the interpeter for every normal
11813 * call and to prevent accidental JITing by JIT compiler
11814 * that doesn't support bpf_tail_call yet
11817 insn->code = BPF_JMP | BPF_TAIL_CALL;
11819 aux = &env->insn_aux_data[i + delta];
11820 if (env->bpf_capable && !expect_blinding &&
11821 prog->jit_requested &&
11822 !bpf_map_key_poisoned(aux) &&
11823 !bpf_map_ptr_poisoned(aux) &&
11824 !bpf_map_ptr_unpriv(aux)) {
11825 struct bpf_jit_poke_descriptor desc = {
11826 .reason = BPF_POKE_REASON_TAIL_CALL,
11827 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
11828 .tail_call.key = bpf_map_key_immediate(aux),
11829 .insn_idx = i + delta,
11832 ret = bpf_jit_add_poke_descriptor(prog, &desc);
11834 verbose(env, "adding tail call poke descriptor failed\n");
11838 insn->imm = ret + 1;
11842 if (!bpf_map_ptr_unpriv(aux))
11845 /* instead of changing every JIT dealing with tail_call
11846 * emit two extra insns:
11847 * if (index >= max_entries) goto out;
11848 * index &= array->index_mask;
11849 * to avoid out-of-bounds cpu speculation
11851 if (bpf_map_ptr_poisoned(aux)) {
11852 verbose(env, "tail_call abusing map_ptr\n");
11856 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
11857 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
11858 map_ptr->max_entries, 2);
11859 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
11860 container_of(map_ptr,
11863 insn_buf[2] = *insn;
11865 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
11870 env->prog = prog = new_prog;
11871 insn = new_prog->insnsi + i + delta;
11875 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
11876 * and other inlining handlers are currently limited to 64 bit
11879 if (prog->jit_requested && BITS_PER_LONG == 64 &&
11880 (insn->imm == BPF_FUNC_map_lookup_elem ||
11881 insn->imm == BPF_FUNC_map_update_elem ||
11882 insn->imm == BPF_FUNC_map_delete_elem ||
11883 insn->imm == BPF_FUNC_map_push_elem ||
11884 insn->imm == BPF_FUNC_map_pop_elem ||
11885 insn->imm == BPF_FUNC_map_peek_elem)) {
11886 aux = &env->insn_aux_data[i + delta];
11887 if (bpf_map_ptr_poisoned(aux))
11888 goto patch_call_imm;
11890 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
11891 ops = map_ptr->ops;
11892 if (insn->imm == BPF_FUNC_map_lookup_elem &&
11893 ops->map_gen_lookup) {
11894 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
11895 if (cnt == -EOPNOTSUPP)
11896 goto patch_map_ops_generic;
11897 if (cnt <= 0 || cnt >= ARRAY_SIZE(insn_buf)) {
11898 verbose(env, "bpf verifier is misconfigured\n");
11902 new_prog = bpf_patch_insn_data(env, i + delta,
11908 env->prog = prog = new_prog;
11909 insn = new_prog->insnsi + i + delta;
11913 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
11914 (void *(*)(struct bpf_map *map, void *key))NULL));
11915 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
11916 (int (*)(struct bpf_map *map, void *key))NULL));
11917 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
11918 (int (*)(struct bpf_map *map, void *key, void *value,
11920 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
11921 (int (*)(struct bpf_map *map, void *value,
11923 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
11924 (int (*)(struct bpf_map *map, void *value))NULL));
11925 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
11926 (int (*)(struct bpf_map *map, void *value))NULL));
11927 patch_map_ops_generic:
11928 switch (insn->imm) {
11929 case BPF_FUNC_map_lookup_elem:
11930 insn->imm = BPF_CAST_CALL(ops->map_lookup_elem) -
11933 case BPF_FUNC_map_update_elem:
11934 insn->imm = BPF_CAST_CALL(ops->map_update_elem) -
11937 case BPF_FUNC_map_delete_elem:
11938 insn->imm = BPF_CAST_CALL(ops->map_delete_elem) -
11941 case BPF_FUNC_map_push_elem:
11942 insn->imm = BPF_CAST_CALL(ops->map_push_elem) -
11945 case BPF_FUNC_map_pop_elem:
11946 insn->imm = BPF_CAST_CALL(ops->map_pop_elem) -
11949 case BPF_FUNC_map_peek_elem:
11950 insn->imm = BPF_CAST_CALL(ops->map_peek_elem) -
11955 goto patch_call_imm;
11958 if (prog->jit_requested && BITS_PER_LONG == 64 &&
11959 insn->imm == BPF_FUNC_jiffies64) {
11960 struct bpf_insn ld_jiffies_addr[2] = {
11961 BPF_LD_IMM64(BPF_REG_0,
11962 (unsigned long)&jiffies),
11965 insn_buf[0] = ld_jiffies_addr[0];
11966 insn_buf[1] = ld_jiffies_addr[1];
11967 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
11971 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
11977 env->prog = prog = new_prog;
11978 insn = new_prog->insnsi + i + delta;
11983 fn = env->ops->get_func_proto(insn->imm, env->prog);
11984 /* all functions that have prototype and verifier allowed
11985 * programs to call them, must be real in-kernel functions
11989 "kernel subsystem misconfigured func %s#%d\n",
11990 func_id_name(insn->imm), insn->imm);
11993 insn->imm = fn->func - __bpf_call_base;
11996 /* Since poke tab is now finalized, publish aux to tracker. */
11997 for (i = 0; i < prog->aux->size_poke_tab; i++) {
11998 map_ptr = prog->aux->poke_tab[i].tail_call.map;
11999 if (!map_ptr->ops->map_poke_track ||
12000 !map_ptr->ops->map_poke_untrack ||
12001 !map_ptr->ops->map_poke_run) {
12002 verbose(env, "bpf verifier is misconfigured\n");
12006 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
12008 verbose(env, "tracking tail call prog failed\n");
12016 static void free_states(struct bpf_verifier_env *env)
12018 struct bpf_verifier_state_list *sl, *sln;
12021 sl = env->free_list;
12024 free_verifier_state(&sl->state, false);
12028 env->free_list = NULL;
12030 if (!env->explored_states)
12033 for (i = 0; i < state_htab_size(env); i++) {
12034 sl = env->explored_states[i];
12038 free_verifier_state(&sl->state, false);
12042 env->explored_states[i] = NULL;
12046 /* The verifier is using insn_aux_data[] to store temporary data during
12047 * verification and to store information for passes that run after the
12048 * verification like dead code sanitization. do_check_common() for subprogram N
12049 * may analyze many other subprograms. sanitize_insn_aux_data() clears all
12050 * temporary data after do_check_common() finds that subprogram N cannot be
12051 * verified independently. pass_cnt counts the number of times
12052 * do_check_common() was run and insn->aux->seen tells the pass number
12053 * insn_aux_data was touched. These variables are compared to clear temporary
12054 * data from failed pass. For testing and experiments do_check_common() can be
12055 * run multiple times even when prior attempt to verify is unsuccessful.
12057 static void sanitize_insn_aux_data(struct bpf_verifier_env *env)
12059 struct bpf_insn *insn = env->prog->insnsi;
12060 struct bpf_insn_aux_data *aux;
12063 for (i = 0; i < env->prog->len; i++) {
12064 class = BPF_CLASS(insn[i].code);
12065 if (class != BPF_LDX && class != BPF_STX)
12067 aux = &env->insn_aux_data[i];
12068 if (aux->seen != env->pass_cnt)
12070 memset(aux, 0, offsetof(typeof(*aux), orig_idx));
12074 static int do_check_common(struct bpf_verifier_env *env, int subprog)
12076 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
12077 struct bpf_verifier_state *state;
12078 struct bpf_reg_state *regs;
12081 env->prev_linfo = NULL;
12084 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
12087 state->curframe = 0;
12088 state->speculative = false;
12089 state->branches = 1;
12090 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
12091 if (!state->frame[0]) {
12095 env->cur_state = state;
12096 init_func_state(env, state->frame[0],
12097 BPF_MAIN_FUNC /* callsite */,
12101 regs = state->frame[state->curframe]->regs;
12102 if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
12103 ret = btf_prepare_func_args(env, subprog, regs);
12106 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
12107 if (regs[i].type == PTR_TO_CTX)
12108 mark_reg_known_zero(env, regs, i);
12109 else if (regs[i].type == SCALAR_VALUE)
12110 mark_reg_unknown(env, regs, i);
12111 else if (regs[i].type == PTR_TO_MEM_OR_NULL) {
12112 const u32 mem_size = regs[i].mem_size;
12114 mark_reg_known_zero(env, regs, i);
12115 regs[i].mem_size = mem_size;
12116 regs[i].id = ++env->id_gen;
12120 /* 1st arg to a function */
12121 regs[BPF_REG_1].type = PTR_TO_CTX;
12122 mark_reg_known_zero(env, regs, BPF_REG_1);
12123 ret = btf_check_func_arg_match(env, subprog, regs);
12124 if (ret == -EFAULT)
12125 /* unlikely verifier bug. abort.
12126 * ret == 0 and ret < 0 are sadly acceptable for
12127 * main() function due to backward compatibility.
12128 * Like socket filter program may be written as:
12129 * int bpf_prog(struct pt_regs *ctx)
12130 * and never dereference that ctx in the program.
12131 * 'struct pt_regs' is a type mismatch for socket
12132 * filter that should be using 'struct __sk_buff'.
12137 ret = do_check(env);
12139 /* check for NULL is necessary, since cur_state can be freed inside
12140 * do_check() under memory pressure.
12142 if (env->cur_state) {
12143 free_verifier_state(env->cur_state, true);
12144 env->cur_state = NULL;
12146 while (!pop_stack(env, NULL, NULL, false));
12147 if (!ret && pop_log)
12148 bpf_vlog_reset(&env->log, 0);
12151 /* clean aux data in case subprog was rejected */
12152 sanitize_insn_aux_data(env);
12156 /* Verify all global functions in a BPF program one by one based on their BTF.
12157 * All global functions must pass verification. Otherwise the whole program is rejected.
12168 * foo() will be verified first for R1=any_scalar_value. During verification it
12169 * will be assumed that bar() already verified successfully and call to bar()
12170 * from foo() will be checked for type match only. Later bar() will be verified
12171 * independently to check that it's safe for R1=any_scalar_value.
12173 static int do_check_subprogs(struct bpf_verifier_env *env)
12175 struct bpf_prog_aux *aux = env->prog->aux;
12178 if (!aux->func_info)
12181 for (i = 1; i < env->subprog_cnt; i++) {
12182 if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL)
12184 env->insn_idx = env->subprog_info[i].start;
12185 WARN_ON_ONCE(env->insn_idx == 0);
12186 ret = do_check_common(env, i);
12189 } else if (env->log.level & BPF_LOG_LEVEL) {
12191 "Func#%d is safe for any args that match its prototype\n",
12198 static int do_check_main(struct bpf_verifier_env *env)
12203 ret = do_check_common(env, 0);
12205 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
12210 static void print_verification_stats(struct bpf_verifier_env *env)
12214 if (env->log.level & BPF_LOG_STATS) {
12215 verbose(env, "verification time %lld usec\n",
12216 div_u64(env->verification_time, 1000));
12217 verbose(env, "stack depth ");
12218 for (i = 0; i < env->subprog_cnt; i++) {
12219 u32 depth = env->subprog_info[i].stack_depth;
12221 verbose(env, "%d", depth);
12222 if (i + 1 < env->subprog_cnt)
12225 verbose(env, "\n");
12227 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
12228 "total_states %d peak_states %d mark_read %d\n",
12229 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
12230 env->max_states_per_insn, env->total_states,
12231 env->peak_states, env->longest_mark_read_walk);
12234 static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
12236 const struct btf_type *t, *func_proto;
12237 const struct bpf_struct_ops *st_ops;
12238 const struct btf_member *member;
12239 struct bpf_prog *prog = env->prog;
12240 u32 btf_id, member_idx;
12243 if (!prog->gpl_compatible) {
12244 verbose(env, "struct ops programs must have a GPL compatible license\n");
12248 btf_id = prog->aux->attach_btf_id;
12249 st_ops = bpf_struct_ops_find(btf_id);
12251 verbose(env, "attach_btf_id %u is not a supported struct\n",
12257 member_idx = prog->expected_attach_type;
12258 if (member_idx >= btf_type_vlen(t)) {
12259 verbose(env, "attach to invalid member idx %u of struct %s\n",
12260 member_idx, st_ops->name);
12264 member = &btf_type_member(t)[member_idx];
12265 mname = btf_name_by_offset(btf_vmlinux, member->name_off);
12266 func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type,
12269 verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
12270 mname, member_idx, st_ops->name);
12274 if (st_ops->check_member) {
12275 int err = st_ops->check_member(t, member);
12278 verbose(env, "attach to unsupported member %s of struct %s\n",
12279 mname, st_ops->name);
12284 prog->aux->attach_func_proto = func_proto;
12285 prog->aux->attach_func_name = mname;
12286 env->ops = st_ops->verifier_ops;
12290 #define SECURITY_PREFIX "security_"
12292 static int check_attach_modify_return(unsigned long addr, const char *func_name)
12294 if (within_error_injection_list(addr) ||
12295 !strncmp(SECURITY_PREFIX, func_name, sizeof(SECURITY_PREFIX) - 1))
12301 /* list of non-sleepable functions that are otherwise on
12302 * ALLOW_ERROR_INJECTION list
12304 BTF_SET_START(btf_non_sleepable_error_inject)
12305 /* Three functions below can be called from sleepable and non-sleepable context.
12306 * Assume non-sleepable from bpf safety point of view.
12308 BTF_ID(func, __add_to_page_cache_locked)
12309 BTF_ID(func, should_fail_alloc_page)
12310 BTF_ID(func, should_failslab)
12311 BTF_SET_END(btf_non_sleepable_error_inject)
12313 static int check_non_sleepable_error_inject(u32 btf_id)
12315 return btf_id_set_contains(&btf_non_sleepable_error_inject, btf_id);
12318 int bpf_check_attach_target(struct bpf_verifier_log *log,
12319 const struct bpf_prog *prog,
12320 const struct bpf_prog *tgt_prog,
12322 struct bpf_attach_target_info *tgt_info)
12324 bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
12325 const char prefix[] = "btf_trace_";
12326 int ret = 0, subprog = -1, i;
12327 const struct btf_type *t;
12328 bool conservative = true;
12334 bpf_log(log, "Tracing programs must provide btf_id\n");
12337 btf = tgt_prog ? tgt_prog->aux->btf : prog->aux->attach_btf;
12340 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
12343 t = btf_type_by_id(btf, btf_id);
12345 bpf_log(log, "attach_btf_id %u is invalid\n", btf_id);
12348 tname = btf_name_by_offset(btf, t->name_off);
12350 bpf_log(log, "attach_btf_id %u doesn't have a name\n", btf_id);
12354 struct bpf_prog_aux *aux = tgt_prog->aux;
12356 for (i = 0; i < aux->func_info_cnt; i++)
12357 if (aux->func_info[i].type_id == btf_id) {
12361 if (subprog == -1) {
12362 bpf_log(log, "Subprog %s doesn't exist\n", tname);
12365 conservative = aux->func_info_aux[subprog].unreliable;
12366 if (prog_extension) {
12367 if (conservative) {
12369 "Cannot replace static functions\n");
12372 if (!prog->jit_requested) {
12374 "Extension programs should be JITed\n");
12378 if (!tgt_prog->jited) {
12379 bpf_log(log, "Can attach to only JITed progs\n");
12382 if (tgt_prog->type == prog->type) {
12383 /* Cannot fentry/fexit another fentry/fexit program.
12384 * Cannot attach program extension to another extension.
12385 * It's ok to attach fentry/fexit to extension program.
12387 bpf_log(log, "Cannot recursively attach\n");
12390 if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
12392 (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
12393 tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
12394 /* Program extensions can extend all program types
12395 * except fentry/fexit. The reason is the following.
12396 * The fentry/fexit programs are used for performance
12397 * analysis, stats and can be attached to any program
12398 * type except themselves. When extension program is
12399 * replacing XDP function it is necessary to allow
12400 * performance analysis of all functions. Both original
12401 * XDP program and its program extension. Hence
12402 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
12403 * allowed. If extending of fentry/fexit was allowed it
12404 * would be possible to create long call chain
12405 * fentry->extension->fentry->extension beyond
12406 * reasonable stack size. Hence extending fentry is not
12409 bpf_log(log, "Cannot extend fentry/fexit\n");
12413 if (prog_extension) {
12414 bpf_log(log, "Cannot replace kernel functions\n");
12419 switch (prog->expected_attach_type) {
12420 case BPF_TRACE_RAW_TP:
12423 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
12426 if (!btf_type_is_typedef(t)) {
12427 bpf_log(log, "attach_btf_id %u is not a typedef\n",
12431 if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
12432 bpf_log(log, "attach_btf_id %u points to wrong type name %s\n",
12436 tname += sizeof(prefix) - 1;
12437 t = btf_type_by_id(btf, t->type);
12438 if (!btf_type_is_ptr(t))
12439 /* should never happen in valid vmlinux build */
12441 t = btf_type_by_id(btf, t->type);
12442 if (!btf_type_is_func_proto(t))
12443 /* should never happen in valid vmlinux build */
12447 case BPF_TRACE_ITER:
12448 if (!btf_type_is_func(t)) {
12449 bpf_log(log, "attach_btf_id %u is not a function\n",
12453 t = btf_type_by_id(btf, t->type);
12454 if (!btf_type_is_func_proto(t))
12456 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
12461 if (!prog_extension)
12464 case BPF_MODIFY_RETURN:
12466 case BPF_TRACE_FENTRY:
12467 case BPF_TRACE_FEXIT:
12468 if (!btf_type_is_func(t)) {
12469 bpf_log(log, "attach_btf_id %u is not a function\n",
12473 if (prog_extension &&
12474 btf_check_type_match(log, prog, btf, t))
12476 t = btf_type_by_id(btf, t->type);
12477 if (!btf_type_is_func_proto(t))
12480 if ((prog->aux->saved_dst_prog_type || prog->aux->saved_dst_attach_type) &&
12481 (!tgt_prog || prog->aux->saved_dst_prog_type != tgt_prog->type ||
12482 prog->aux->saved_dst_attach_type != tgt_prog->expected_attach_type))
12485 if (tgt_prog && conservative)
12488 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
12494 addr = (long) tgt_prog->bpf_func;
12496 addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
12498 addr = kallsyms_lookup_name(tname);
12501 "The address of function %s cannot be found\n",
12507 if (prog->aux->sleepable) {
12509 switch (prog->type) {
12510 case BPF_PROG_TYPE_TRACING:
12511 /* fentry/fexit/fmod_ret progs can be sleepable only if they are
12512 * attached to ALLOW_ERROR_INJECTION and are not in denylist.
12514 if (!check_non_sleepable_error_inject(btf_id) &&
12515 within_error_injection_list(addr))
12518 case BPF_PROG_TYPE_LSM:
12519 /* LSM progs check that they are attached to bpf_lsm_*() funcs.
12520 * Only some of them are sleepable.
12522 if (bpf_lsm_is_sleepable_hook(btf_id))
12529 bpf_log(log, "%s is not sleepable\n", tname);
12532 } else if (prog->expected_attach_type == BPF_MODIFY_RETURN) {
12534 bpf_log(log, "can't modify return codes of BPF programs\n");
12537 ret = check_attach_modify_return(addr, tname);
12539 bpf_log(log, "%s() is not modifiable\n", tname);
12546 tgt_info->tgt_addr = addr;
12547 tgt_info->tgt_name = tname;
12548 tgt_info->tgt_type = t;
12552 static int check_attach_btf_id(struct bpf_verifier_env *env)
12554 struct bpf_prog *prog = env->prog;
12555 struct bpf_prog *tgt_prog = prog->aux->dst_prog;
12556 struct bpf_attach_target_info tgt_info = {};
12557 u32 btf_id = prog->aux->attach_btf_id;
12558 struct bpf_trampoline *tr;
12562 if (prog->aux->sleepable && prog->type != BPF_PROG_TYPE_TRACING &&
12563 prog->type != BPF_PROG_TYPE_LSM) {
12564 verbose(env, "Only fentry/fexit/fmod_ret and lsm programs can be sleepable\n");
12568 if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
12569 return check_struct_ops_btf_id(env);
12571 if (prog->type != BPF_PROG_TYPE_TRACING &&
12572 prog->type != BPF_PROG_TYPE_LSM &&
12573 prog->type != BPF_PROG_TYPE_EXT)
12576 ret = bpf_check_attach_target(&env->log, prog, tgt_prog, btf_id, &tgt_info);
12580 if (tgt_prog && prog->type == BPF_PROG_TYPE_EXT) {
12581 /* to make freplace equivalent to their targets, they need to
12582 * inherit env->ops and expected_attach_type for the rest of the
12585 env->ops = bpf_verifier_ops[tgt_prog->type];
12586 prog->expected_attach_type = tgt_prog->expected_attach_type;
12589 /* store info about the attachment target that will be used later */
12590 prog->aux->attach_func_proto = tgt_info.tgt_type;
12591 prog->aux->attach_func_name = tgt_info.tgt_name;
12594 prog->aux->saved_dst_prog_type = tgt_prog->type;
12595 prog->aux->saved_dst_attach_type = tgt_prog->expected_attach_type;
12598 if (prog->expected_attach_type == BPF_TRACE_RAW_TP) {
12599 prog->aux->attach_btf_trace = true;
12601 } else if (prog->expected_attach_type == BPF_TRACE_ITER) {
12602 if (!bpf_iter_prog_supported(prog))
12607 if (prog->type == BPF_PROG_TYPE_LSM) {
12608 ret = bpf_lsm_verify_prog(&env->log, prog);
12613 key = bpf_trampoline_compute_key(tgt_prog, prog->aux->attach_btf, btf_id);
12614 tr = bpf_trampoline_get(key, &tgt_info);
12618 prog->aux->dst_trampoline = tr;
12622 struct btf *bpf_get_btf_vmlinux(void)
12624 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
12625 mutex_lock(&bpf_verifier_lock);
12627 btf_vmlinux = btf_parse_vmlinux();
12628 mutex_unlock(&bpf_verifier_lock);
12630 return btf_vmlinux;
12633 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr,
12634 union bpf_attr __user *uattr)
12636 u64 start_time = ktime_get_ns();
12637 struct bpf_verifier_env *env;
12638 struct bpf_verifier_log *log;
12639 int i, len, ret = -EINVAL;
12642 /* no program is valid */
12643 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
12646 /* 'struct bpf_verifier_env' can be global, but since it's not small,
12647 * allocate/free it every time bpf_check() is called
12649 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
12654 len = (*prog)->len;
12655 env->insn_aux_data =
12656 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
12658 if (!env->insn_aux_data)
12660 for (i = 0; i < len; i++)
12661 env->insn_aux_data[i].orig_idx = i;
12663 env->ops = bpf_verifier_ops[env->prog->type];
12664 is_priv = bpf_capable();
12666 bpf_get_btf_vmlinux();
12668 /* grab the mutex to protect few globals used by verifier */
12670 mutex_lock(&bpf_verifier_lock);
12672 if (attr->log_level || attr->log_buf || attr->log_size) {
12673 /* user requested verbose verifier output
12674 * and supplied buffer to store the verification trace
12676 log->level = attr->log_level;
12677 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
12678 log->len_total = attr->log_size;
12681 /* log attributes have to be sane */
12682 if (log->len_total < 128 || log->len_total > UINT_MAX >> 2 ||
12683 !log->level || !log->ubuf || log->level & ~BPF_LOG_MASK)
12687 if (IS_ERR(btf_vmlinux)) {
12688 /* Either gcc or pahole or kernel are broken. */
12689 verbose(env, "in-kernel BTF is malformed\n");
12690 ret = PTR_ERR(btf_vmlinux);
12691 goto skip_full_check;
12694 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
12695 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
12696 env->strict_alignment = true;
12697 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
12698 env->strict_alignment = false;
12700 env->allow_ptr_leaks = bpf_allow_ptr_leaks();
12701 env->allow_uninit_stack = bpf_allow_uninit_stack();
12702 env->allow_ptr_to_map_access = bpf_allow_ptr_to_map_access();
12703 env->bypass_spec_v1 = bpf_bypass_spec_v1();
12704 env->bypass_spec_v4 = bpf_bypass_spec_v4();
12705 env->bpf_capable = bpf_capable();
12708 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
12710 if (bpf_prog_is_dev_bound(env->prog->aux)) {
12711 ret = bpf_prog_offload_verifier_prep(env->prog);
12713 goto skip_full_check;
12716 env->explored_states = kvcalloc(state_htab_size(env),
12717 sizeof(struct bpf_verifier_state_list *),
12720 if (!env->explored_states)
12721 goto skip_full_check;
12723 ret = check_subprogs(env);
12725 goto skip_full_check;
12727 ret = check_btf_info(env, attr, uattr);
12729 goto skip_full_check;
12731 ret = check_attach_btf_id(env);
12733 goto skip_full_check;
12735 ret = resolve_pseudo_ldimm64(env);
12737 goto skip_full_check;
12739 ret = check_cfg(env);
12741 goto skip_full_check;
12743 ret = do_check_subprogs(env);
12744 ret = ret ?: do_check_main(env);
12746 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
12747 ret = bpf_prog_offload_finalize(env);
12750 kvfree(env->explored_states);
12753 ret = check_max_stack_depth(env);
12755 /* instruction rewrites happen after this point */
12758 opt_hard_wire_dead_code_branches(env);
12760 ret = opt_remove_dead_code(env);
12762 ret = opt_remove_nops(env);
12765 sanitize_dead_code(env);
12769 /* program is valid, convert *(u32*)(ctx + off) accesses */
12770 ret = convert_ctx_accesses(env);
12773 ret = fixup_bpf_calls(env);
12775 /* do 32-bit optimization after insn patching has done so those patched
12776 * insns could be handled correctly.
12778 if (ret == 0 && !bpf_prog_is_dev_bound(env->prog->aux)) {
12779 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
12780 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
12785 ret = fixup_call_args(env);
12787 env->verification_time = ktime_get_ns() - start_time;
12788 print_verification_stats(env);
12790 if (log->level && bpf_verifier_log_full(log))
12792 if (log->level && !log->ubuf) {
12794 goto err_release_maps;
12798 goto err_release_maps;
12800 if (env->used_map_cnt) {
12801 /* if program passed verifier, update used_maps in bpf_prog_info */
12802 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
12803 sizeof(env->used_maps[0]),
12806 if (!env->prog->aux->used_maps) {
12808 goto err_release_maps;
12811 memcpy(env->prog->aux->used_maps, env->used_maps,
12812 sizeof(env->used_maps[0]) * env->used_map_cnt);
12813 env->prog->aux->used_map_cnt = env->used_map_cnt;
12815 if (env->used_btf_cnt) {
12816 /* if program passed verifier, update used_btfs in bpf_prog_aux */
12817 env->prog->aux->used_btfs = kmalloc_array(env->used_btf_cnt,
12818 sizeof(env->used_btfs[0]),
12820 if (!env->prog->aux->used_btfs) {
12822 goto err_release_maps;
12825 memcpy(env->prog->aux->used_btfs, env->used_btfs,
12826 sizeof(env->used_btfs[0]) * env->used_btf_cnt);
12827 env->prog->aux->used_btf_cnt = env->used_btf_cnt;
12829 if (env->used_map_cnt || env->used_btf_cnt) {
12830 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
12831 * bpf_ld_imm64 instructions
12833 convert_pseudo_ld_imm64(env);
12836 adjust_btf_func(env);
12839 if (!env->prog->aux->used_maps)
12840 /* if we didn't copy map pointers into bpf_prog_info, release
12841 * them now. Otherwise free_used_maps() will release them.
12844 if (!env->prog->aux->used_btfs)
12847 /* extension progs temporarily inherit the attach_type of their targets
12848 for verification purposes, so set it back to zero before returning
12850 if (env->prog->type == BPF_PROG_TYPE_EXT)
12851 env->prog->expected_attach_type = 0;
12856 mutex_unlock(&bpf_verifier_lock);
12857 vfree(env->insn_aux_data);