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 struct bpf_call_arg_meta {
232 struct bpf_map *map_ptr;
245 struct btf *btf_vmlinux;
247 static DEFINE_MUTEX(bpf_verifier_lock);
249 static const struct bpf_line_info *
250 find_linfo(const struct bpf_verifier_env *env, u32 insn_off)
252 const struct bpf_line_info *linfo;
253 const struct bpf_prog *prog;
257 nr_linfo = prog->aux->nr_linfo;
259 if (!nr_linfo || insn_off >= prog->len)
262 linfo = prog->aux->linfo;
263 for (i = 1; i < nr_linfo; i++)
264 if (insn_off < linfo[i].insn_off)
267 return &linfo[i - 1];
270 void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
275 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
277 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
278 "verifier log line truncated - local buffer too short\n");
280 n = min(log->len_total - log->len_used - 1, n);
283 if (log->level == BPF_LOG_KERNEL) {
284 pr_err("BPF:%s\n", log->kbuf);
287 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
293 static void bpf_vlog_reset(struct bpf_verifier_log *log, u32 new_pos)
297 if (!bpf_verifier_log_needed(log))
300 log->len_used = new_pos;
301 if (put_user(zero, log->ubuf + new_pos))
305 /* log_level controls verbosity level of eBPF verifier.
306 * bpf_verifier_log_write() is used to dump the verification trace to the log,
307 * so the user can figure out what's wrong with the program
309 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
310 const char *fmt, ...)
314 if (!bpf_verifier_log_needed(&env->log))
318 bpf_verifier_vlog(&env->log, fmt, args);
321 EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
323 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
325 struct bpf_verifier_env *env = private_data;
328 if (!bpf_verifier_log_needed(&env->log))
332 bpf_verifier_vlog(&env->log, fmt, args);
336 __printf(2, 3) void bpf_log(struct bpf_verifier_log *log,
337 const char *fmt, ...)
341 if (!bpf_verifier_log_needed(log))
345 bpf_verifier_vlog(log, fmt, args);
349 static const char *ltrim(const char *s)
357 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env,
359 const char *prefix_fmt, ...)
361 const struct bpf_line_info *linfo;
363 if (!bpf_verifier_log_needed(&env->log))
366 linfo = find_linfo(env, insn_off);
367 if (!linfo || linfo == env->prev_linfo)
373 va_start(args, prefix_fmt);
374 bpf_verifier_vlog(&env->log, prefix_fmt, args);
379 ltrim(btf_name_by_offset(env->prog->aux->btf,
382 env->prev_linfo = linfo;
385 static bool type_is_pkt_pointer(enum bpf_reg_type type)
387 return type == PTR_TO_PACKET ||
388 type == PTR_TO_PACKET_META;
391 static bool type_is_sk_pointer(enum bpf_reg_type type)
393 return type == PTR_TO_SOCKET ||
394 type == PTR_TO_SOCK_COMMON ||
395 type == PTR_TO_TCP_SOCK ||
396 type == PTR_TO_XDP_SOCK;
399 static bool reg_type_not_null(enum bpf_reg_type type)
401 return type == PTR_TO_SOCKET ||
402 type == PTR_TO_TCP_SOCK ||
403 type == PTR_TO_MAP_VALUE ||
404 type == PTR_TO_SOCK_COMMON;
407 static bool reg_type_may_be_null(enum bpf_reg_type type)
409 return type == PTR_TO_MAP_VALUE_OR_NULL ||
410 type == PTR_TO_SOCKET_OR_NULL ||
411 type == PTR_TO_SOCK_COMMON_OR_NULL ||
412 type == PTR_TO_TCP_SOCK_OR_NULL ||
413 type == PTR_TO_BTF_ID_OR_NULL ||
414 type == PTR_TO_MEM_OR_NULL ||
415 type == PTR_TO_RDONLY_BUF_OR_NULL ||
416 type == PTR_TO_RDWR_BUF_OR_NULL;
419 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
421 return reg->type == PTR_TO_MAP_VALUE &&
422 map_value_has_spin_lock(reg->map_ptr);
425 static bool reg_type_may_be_refcounted_or_null(enum bpf_reg_type type)
427 return type == PTR_TO_SOCKET ||
428 type == PTR_TO_SOCKET_OR_NULL ||
429 type == PTR_TO_TCP_SOCK ||
430 type == PTR_TO_TCP_SOCK_OR_NULL ||
431 type == PTR_TO_MEM ||
432 type == PTR_TO_MEM_OR_NULL;
435 static bool arg_type_may_be_refcounted(enum bpf_arg_type type)
437 return type == ARG_PTR_TO_SOCK_COMMON;
440 static bool arg_type_may_be_null(enum bpf_arg_type type)
442 return type == ARG_PTR_TO_MAP_VALUE_OR_NULL ||
443 type == ARG_PTR_TO_MEM_OR_NULL ||
444 type == ARG_PTR_TO_CTX_OR_NULL ||
445 type == ARG_PTR_TO_SOCKET_OR_NULL ||
446 type == ARG_PTR_TO_ALLOC_MEM_OR_NULL;
449 /* Determine whether the function releases some resources allocated by another
450 * function call. The first reference type argument will be assumed to be
451 * released by release_reference().
453 static bool is_release_function(enum bpf_func_id func_id)
455 return func_id == BPF_FUNC_sk_release ||
456 func_id == BPF_FUNC_ringbuf_submit ||
457 func_id == BPF_FUNC_ringbuf_discard;
460 static bool may_be_acquire_function(enum bpf_func_id func_id)
462 return func_id == BPF_FUNC_sk_lookup_tcp ||
463 func_id == BPF_FUNC_sk_lookup_udp ||
464 func_id == BPF_FUNC_skc_lookup_tcp ||
465 func_id == BPF_FUNC_map_lookup_elem ||
466 func_id == BPF_FUNC_ringbuf_reserve;
469 static bool is_acquire_function(enum bpf_func_id func_id,
470 const struct bpf_map *map)
472 enum bpf_map_type map_type = map ? map->map_type : BPF_MAP_TYPE_UNSPEC;
474 if (func_id == BPF_FUNC_sk_lookup_tcp ||
475 func_id == BPF_FUNC_sk_lookup_udp ||
476 func_id == BPF_FUNC_skc_lookup_tcp ||
477 func_id == BPF_FUNC_ringbuf_reserve)
480 if (func_id == BPF_FUNC_map_lookup_elem &&
481 (map_type == BPF_MAP_TYPE_SOCKMAP ||
482 map_type == BPF_MAP_TYPE_SOCKHASH))
488 static bool is_ptr_cast_function(enum bpf_func_id func_id)
490 return func_id == BPF_FUNC_tcp_sock ||
491 func_id == BPF_FUNC_sk_fullsock ||
492 func_id == BPF_FUNC_skc_to_tcp_sock ||
493 func_id == BPF_FUNC_skc_to_tcp6_sock ||
494 func_id == BPF_FUNC_skc_to_udp6_sock ||
495 func_id == BPF_FUNC_skc_to_tcp_timewait_sock ||
496 func_id == BPF_FUNC_skc_to_tcp_request_sock;
499 /* string representation of 'enum bpf_reg_type' */
500 static const char * const reg_type_str[] = {
502 [SCALAR_VALUE] = "inv",
503 [PTR_TO_CTX] = "ctx",
504 [CONST_PTR_TO_MAP] = "map_ptr",
505 [PTR_TO_MAP_VALUE] = "map_value",
506 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
507 [PTR_TO_STACK] = "fp",
508 [PTR_TO_PACKET] = "pkt",
509 [PTR_TO_PACKET_META] = "pkt_meta",
510 [PTR_TO_PACKET_END] = "pkt_end",
511 [PTR_TO_FLOW_KEYS] = "flow_keys",
512 [PTR_TO_SOCKET] = "sock",
513 [PTR_TO_SOCKET_OR_NULL] = "sock_or_null",
514 [PTR_TO_SOCK_COMMON] = "sock_common",
515 [PTR_TO_SOCK_COMMON_OR_NULL] = "sock_common_or_null",
516 [PTR_TO_TCP_SOCK] = "tcp_sock",
517 [PTR_TO_TCP_SOCK_OR_NULL] = "tcp_sock_or_null",
518 [PTR_TO_TP_BUFFER] = "tp_buffer",
519 [PTR_TO_XDP_SOCK] = "xdp_sock",
520 [PTR_TO_BTF_ID] = "ptr_",
521 [PTR_TO_BTF_ID_OR_NULL] = "ptr_or_null_",
522 [PTR_TO_PERCPU_BTF_ID] = "percpu_ptr_",
523 [PTR_TO_MEM] = "mem",
524 [PTR_TO_MEM_OR_NULL] = "mem_or_null",
525 [PTR_TO_RDONLY_BUF] = "rdonly_buf",
526 [PTR_TO_RDONLY_BUF_OR_NULL] = "rdonly_buf_or_null",
527 [PTR_TO_RDWR_BUF] = "rdwr_buf",
528 [PTR_TO_RDWR_BUF_OR_NULL] = "rdwr_buf_or_null",
531 static char slot_type_char[] = {
532 [STACK_INVALID] = '?',
538 static void print_liveness(struct bpf_verifier_env *env,
539 enum bpf_reg_liveness live)
541 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
543 if (live & REG_LIVE_READ)
545 if (live & REG_LIVE_WRITTEN)
547 if (live & REG_LIVE_DONE)
551 static struct bpf_func_state *func(struct bpf_verifier_env *env,
552 const struct bpf_reg_state *reg)
554 struct bpf_verifier_state *cur = env->cur_state;
556 return cur->frame[reg->frameno];
559 const char *kernel_type_name(u32 id)
561 return btf_name_by_offset(btf_vmlinux,
562 btf_type_by_id(btf_vmlinux, id)->name_off);
565 static void print_verifier_state(struct bpf_verifier_env *env,
566 const struct bpf_func_state *state)
568 const struct bpf_reg_state *reg;
573 verbose(env, " frame%d:", state->frameno);
574 for (i = 0; i < MAX_BPF_REG; i++) {
575 reg = &state->regs[i];
579 verbose(env, " R%d", i);
580 print_liveness(env, reg->live);
581 verbose(env, "=%s", reg_type_str[t]);
582 if (t == SCALAR_VALUE && reg->precise)
584 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
585 tnum_is_const(reg->var_off)) {
586 /* reg->off should be 0 for SCALAR_VALUE */
587 verbose(env, "%lld", reg->var_off.value + reg->off);
589 if (t == PTR_TO_BTF_ID ||
590 t == PTR_TO_BTF_ID_OR_NULL ||
591 t == PTR_TO_PERCPU_BTF_ID)
592 verbose(env, "%s", kernel_type_name(reg->btf_id));
593 verbose(env, "(id=%d", reg->id);
594 if (reg_type_may_be_refcounted_or_null(t))
595 verbose(env, ",ref_obj_id=%d", reg->ref_obj_id);
596 if (t != SCALAR_VALUE)
597 verbose(env, ",off=%d", reg->off);
598 if (type_is_pkt_pointer(t))
599 verbose(env, ",r=%d", reg->range);
600 else if (t == CONST_PTR_TO_MAP ||
601 t == PTR_TO_MAP_VALUE ||
602 t == PTR_TO_MAP_VALUE_OR_NULL)
603 verbose(env, ",ks=%d,vs=%d",
604 reg->map_ptr->key_size,
605 reg->map_ptr->value_size);
606 if (tnum_is_const(reg->var_off)) {
607 /* Typically an immediate SCALAR_VALUE, but
608 * could be a pointer whose offset is too big
611 verbose(env, ",imm=%llx", reg->var_off.value);
613 if (reg->smin_value != reg->umin_value &&
614 reg->smin_value != S64_MIN)
615 verbose(env, ",smin_value=%lld",
616 (long long)reg->smin_value);
617 if (reg->smax_value != reg->umax_value &&
618 reg->smax_value != S64_MAX)
619 verbose(env, ",smax_value=%lld",
620 (long long)reg->smax_value);
621 if (reg->umin_value != 0)
622 verbose(env, ",umin_value=%llu",
623 (unsigned long long)reg->umin_value);
624 if (reg->umax_value != U64_MAX)
625 verbose(env, ",umax_value=%llu",
626 (unsigned long long)reg->umax_value);
627 if (!tnum_is_unknown(reg->var_off)) {
630 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
631 verbose(env, ",var_off=%s", tn_buf);
633 if (reg->s32_min_value != reg->smin_value &&
634 reg->s32_min_value != S32_MIN)
635 verbose(env, ",s32_min_value=%d",
636 (int)(reg->s32_min_value));
637 if (reg->s32_max_value != reg->smax_value &&
638 reg->s32_max_value != S32_MAX)
639 verbose(env, ",s32_max_value=%d",
640 (int)(reg->s32_max_value));
641 if (reg->u32_min_value != reg->umin_value &&
642 reg->u32_min_value != U32_MIN)
643 verbose(env, ",u32_min_value=%d",
644 (int)(reg->u32_min_value));
645 if (reg->u32_max_value != reg->umax_value &&
646 reg->u32_max_value != U32_MAX)
647 verbose(env, ",u32_max_value=%d",
648 (int)(reg->u32_max_value));
653 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
654 char types_buf[BPF_REG_SIZE + 1];
658 for (j = 0; j < BPF_REG_SIZE; j++) {
659 if (state->stack[i].slot_type[j] != STACK_INVALID)
661 types_buf[j] = slot_type_char[
662 state->stack[i].slot_type[j]];
664 types_buf[BPF_REG_SIZE] = 0;
667 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
668 print_liveness(env, state->stack[i].spilled_ptr.live);
669 if (state->stack[i].slot_type[0] == STACK_SPILL) {
670 reg = &state->stack[i].spilled_ptr;
672 verbose(env, "=%s", reg_type_str[t]);
673 if (t == SCALAR_VALUE && reg->precise)
675 if (t == SCALAR_VALUE && tnum_is_const(reg->var_off))
676 verbose(env, "%lld", reg->var_off.value + reg->off);
678 verbose(env, "=%s", types_buf);
681 if (state->acquired_refs && state->refs[0].id) {
682 verbose(env, " refs=%d", state->refs[0].id);
683 for (i = 1; i < state->acquired_refs; i++)
684 if (state->refs[i].id)
685 verbose(env, ",%d", state->refs[i].id);
690 #define COPY_STATE_FN(NAME, COUNT, FIELD, SIZE) \
691 static int copy_##NAME##_state(struct bpf_func_state *dst, \
692 const struct bpf_func_state *src) \
696 if (WARN_ON_ONCE(dst->COUNT < src->COUNT)) { \
697 /* internal bug, make state invalid to reject the program */ \
698 memset(dst, 0, sizeof(*dst)); \
701 memcpy(dst->FIELD, src->FIELD, \
702 sizeof(*src->FIELD) * (src->COUNT / SIZE)); \
705 /* copy_reference_state() */
706 COPY_STATE_FN(reference, acquired_refs, refs, 1)
707 /* copy_stack_state() */
708 COPY_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
711 #define REALLOC_STATE_FN(NAME, COUNT, FIELD, SIZE) \
712 static int realloc_##NAME##_state(struct bpf_func_state *state, int size, \
715 u32 old_size = state->COUNT; \
716 struct bpf_##NAME##_state *new_##FIELD; \
717 int slot = size / SIZE; \
719 if (size <= old_size || !size) { \
722 state->COUNT = slot * SIZE; \
723 if (!size && old_size) { \
724 kfree(state->FIELD); \
725 state->FIELD = NULL; \
729 new_##FIELD = kmalloc_array(slot, sizeof(struct bpf_##NAME##_state), \
735 memcpy(new_##FIELD, state->FIELD, \
736 sizeof(*new_##FIELD) * (old_size / SIZE)); \
737 memset(new_##FIELD + old_size / SIZE, 0, \
738 sizeof(*new_##FIELD) * (size - old_size) / SIZE); \
740 state->COUNT = slot * SIZE; \
741 kfree(state->FIELD); \
742 state->FIELD = new_##FIELD; \
745 /* realloc_reference_state() */
746 REALLOC_STATE_FN(reference, acquired_refs, refs, 1)
747 /* realloc_stack_state() */
748 REALLOC_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
749 #undef REALLOC_STATE_FN
751 /* do_check() starts with zero-sized stack in struct bpf_verifier_state to
752 * make it consume minimal amount of memory. check_stack_write() access from
753 * the program calls into realloc_func_state() to grow the stack size.
754 * Note there is a non-zero 'parent' pointer inside bpf_verifier_state
755 * which realloc_stack_state() copies over. It points to previous
756 * bpf_verifier_state which is never reallocated.
758 static int realloc_func_state(struct bpf_func_state *state, int stack_size,
759 int refs_size, bool copy_old)
761 int err = realloc_reference_state(state, refs_size, copy_old);
764 return realloc_stack_state(state, stack_size, copy_old);
767 /* Acquire a pointer id from the env and update the state->refs to include
768 * this new pointer reference.
769 * On success, returns a valid pointer id to associate with the register
770 * On failure, returns a negative errno.
772 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
774 struct bpf_func_state *state = cur_func(env);
775 int new_ofs = state->acquired_refs;
778 err = realloc_reference_state(state, state->acquired_refs + 1, true);
782 state->refs[new_ofs].id = id;
783 state->refs[new_ofs].insn_idx = insn_idx;
788 /* release function corresponding to acquire_reference_state(). Idempotent. */
789 static int release_reference_state(struct bpf_func_state *state, int ptr_id)
793 last_idx = state->acquired_refs - 1;
794 for (i = 0; i < state->acquired_refs; i++) {
795 if (state->refs[i].id == ptr_id) {
796 if (last_idx && i != last_idx)
797 memcpy(&state->refs[i], &state->refs[last_idx],
798 sizeof(*state->refs));
799 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
800 state->acquired_refs--;
807 static int transfer_reference_state(struct bpf_func_state *dst,
808 struct bpf_func_state *src)
810 int err = realloc_reference_state(dst, src->acquired_refs, false);
813 err = copy_reference_state(dst, src);
819 static void free_func_state(struct bpf_func_state *state)
828 static void clear_jmp_history(struct bpf_verifier_state *state)
830 kfree(state->jmp_history);
831 state->jmp_history = NULL;
832 state->jmp_history_cnt = 0;
835 static void free_verifier_state(struct bpf_verifier_state *state,
840 for (i = 0; i <= state->curframe; i++) {
841 free_func_state(state->frame[i]);
842 state->frame[i] = NULL;
844 clear_jmp_history(state);
849 /* copy verifier state from src to dst growing dst stack space
850 * when necessary to accommodate larger src stack
852 static int copy_func_state(struct bpf_func_state *dst,
853 const struct bpf_func_state *src)
857 err = realloc_func_state(dst, src->allocated_stack, src->acquired_refs,
861 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
862 err = copy_reference_state(dst, src);
865 return copy_stack_state(dst, src);
868 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
869 const struct bpf_verifier_state *src)
871 struct bpf_func_state *dst;
872 u32 jmp_sz = sizeof(struct bpf_idx_pair) * src->jmp_history_cnt;
875 if (dst_state->jmp_history_cnt < src->jmp_history_cnt) {
876 kfree(dst_state->jmp_history);
877 dst_state->jmp_history = kmalloc(jmp_sz, GFP_USER);
878 if (!dst_state->jmp_history)
881 memcpy(dst_state->jmp_history, src->jmp_history, jmp_sz);
882 dst_state->jmp_history_cnt = src->jmp_history_cnt;
884 /* if dst has more stack frames then src frame, free them */
885 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
886 free_func_state(dst_state->frame[i]);
887 dst_state->frame[i] = NULL;
889 dst_state->speculative = src->speculative;
890 dst_state->curframe = src->curframe;
891 dst_state->active_spin_lock = src->active_spin_lock;
892 dst_state->branches = src->branches;
893 dst_state->parent = src->parent;
894 dst_state->first_insn_idx = src->first_insn_idx;
895 dst_state->last_insn_idx = src->last_insn_idx;
896 for (i = 0; i <= src->curframe; i++) {
897 dst = dst_state->frame[i];
899 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
902 dst_state->frame[i] = dst;
904 err = copy_func_state(dst, src->frame[i]);
911 static void update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
914 u32 br = --st->branches;
916 /* WARN_ON(br > 1) technically makes sense here,
917 * but see comment in push_stack(), hence:
919 WARN_ONCE((int)br < 0,
920 "BUG update_branch_counts:branches_to_explore=%d\n",
928 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
929 int *insn_idx, bool pop_log)
931 struct bpf_verifier_state *cur = env->cur_state;
932 struct bpf_verifier_stack_elem *elem, *head = env->head;
935 if (env->head == NULL)
939 err = copy_verifier_state(cur, &head->st);
944 bpf_vlog_reset(&env->log, head->log_pos);
946 *insn_idx = head->insn_idx;
948 *prev_insn_idx = head->prev_insn_idx;
950 free_verifier_state(&head->st, false);
957 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
958 int insn_idx, int prev_insn_idx,
961 struct bpf_verifier_state *cur = env->cur_state;
962 struct bpf_verifier_stack_elem *elem;
965 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
969 elem->insn_idx = insn_idx;
970 elem->prev_insn_idx = prev_insn_idx;
971 elem->next = env->head;
972 elem->log_pos = env->log.len_used;
975 err = copy_verifier_state(&elem->st, cur);
978 elem->st.speculative |= speculative;
979 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
980 verbose(env, "The sequence of %d jumps is too complex.\n",
984 if (elem->st.parent) {
985 ++elem->st.parent->branches;
986 /* WARN_ON(branches > 2) technically makes sense here,
988 * 1. speculative states will bump 'branches' for non-branch
990 * 2. is_state_visited() heuristics may decide not to create
991 * a new state for a sequence of branches and all such current
992 * and cloned states will be pointing to a single parent state
993 * which might have large 'branches' count.
998 free_verifier_state(env->cur_state, true);
999 env->cur_state = NULL;
1000 /* pop all elements and return */
1001 while (!pop_stack(env, NULL, NULL, false));
1005 #define CALLER_SAVED_REGS 6
1006 static const int caller_saved[CALLER_SAVED_REGS] = {
1007 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
1010 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1011 struct bpf_reg_state *reg);
1013 /* This helper doesn't clear reg->id */
1014 static void ___mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1016 reg->var_off = tnum_const(imm);
1017 reg->smin_value = (s64)imm;
1018 reg->smax_value = (s64)imm;
1019 reg->umin_value = imm;
1020 reg->umax_value = imm;
1022 reg->s32_min_value = (s32)imm;
1023 reg->s32_max_value = (s32)imm;
1024 reg->u32_min_value = (u32)imm;
1025 reg->u32_max_value = (u32)imm;
1028 /* Mark the unknown part of a register (variable offset or scalar value) as
1029 * known to have the value @imm.
1031 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1033 /* Clear id, off, and union(map_ptr, range) */
1034 memset(((u8 *)reg) + sizeof(reg->type), 0,
1035 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
1036 ___mark_reg_known(reg, imm);
1039 static void __mark_reg32_known(struct bpf_reg_state *reg, u64 imm)
1041 reg->var_off = tnum_const_subreg(reg->var_off, imm);
1042 reg->s32_min_value = (s32)imm;
1043 reg->s32_max_value = (s32)imm;
1044 reg->u32_min_value = (u32)imm;
1045 reg->u32_max_value = (u32)imm;
1048 /* Mark the 'variable offset' part of a register as zero. This should be
1049 * used only on registers holding a pointer type.
1051 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
1053 __mark_reg_known(reg, 0);
1056 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
1058 __mark_reg_known(reg, 0);
1059 reg->type = SCALAR_VALUE;
1062 static void mark_reg_known_zero(struct bpf_verifier_env *env,
1063 struct bpf_reg_state *regs, u32 regno)
1065 if (WARN_ON(regno >= MAX_BPF_REG)) {
1066 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
1067 /* Something bad happened, let's kill all regs */
1068 for (regno = 0; regno < MAX_BPF_REG; regno++)
1069 __mark_reg_not_init(env, regs + regno);
1072 __mark_reg_known_zero(regs + regno);
1075 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
1077 return type_is_pkt_pointer(reg->type);
1080 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
1082 return reg_is_pkt_pointer(reg) ||
1083 reg->type == PTR_TO_PACKET_END;
1086 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
1087 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
1088 enum bpf_reg_type which)
1090 /* The register can already have a range from prior markings.
1091 * This is fine as long as it hasn't been advanced from its
1094 return reg->type == which &&
1097 tnum_equals_const(reg->var_off, 0);
1100 /* Reset the min/max bounds of a register */
1101 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
1103 reg->smin_value = S64_MIN;
1104 reg->smax_value = S64_MAX;
1105 reg->umin_value = 0;
1106 reg->umax_value = U64_MAX;
1108 reg->s32_min_value = S32_MIN;
1109 reg->s32_max_value = S32_MAX;
1110 reg->u32_min_value = 0;
1111 reg->u32_max_value = U32_MAX;
1114 static void __mark_reg64_unbounded(struct bpf_reg_state *reg)
1116 reg->smin_value = S64_MIN;
1117 reg->smax_value = S64_MAX;
1118 reg->umin_value = 0;
1119 reg->umax_value = U64_MAX;
1122 static void __mark_reg32_unbounded(struct bpf_reg_state *reg)
1124 reg->s32_min_value = S32_MIN;
1125 reg->s32_max_value = S32_MAX;
1126 reg->u32_min_value = 0;
1127 reg->u32_max_value = U32_MAX;
1130 static void __update_reg32_bounds(struct bpf_reg_state *reg)
1132 struct tnum var32_off = tnum_subreg(reg->var_off);
1134 /* min signed is max(sign bit) | min(other bits) */
1135 reg->s32_min_value = max_t(s32, reg->s32_min_value,
1136 var32_off.value | (var32_off.mask & S32_MIN));
1137 /* max signed is min(sign bit) | max(other bits) */
1138 reg->s32_max_value = min_t(s32, reg->s32_max_value,
1139 var32_off.value | (var32_off.mask & S32_MAX));
1140 reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)var32_off.value);
1141 reg->u32_max_value = min(reg->u32_max_value,
1142 (u32)(var32_off.value | var32_off.mask));
1145 static void __update_reg64_bounds(struct bpf_reg_state *reg)
1147 /* min signed is max(sign bit) | min(other bits) */
1148 reg->smin_value = max_t(s64, reg->smin_value,
1149 reg->var_off.value | (reg->var_off.mask & S64_MIN));
1150 /* max signed is min(sign bit) | max(other bits) */
1151 reg->smax_value = min_t(s64, reg->smax_value,
1152 reg->var_off.value | (reg->var_off.mask & S64_MAX));
1153 reg->umin_value = max(reg->umin_value, reg->var_off.value);
1154 reg->umax_value = min(reg->umax_value,
1155 reg->var_off.value | reg->var_off.mask);
1158 static void __update_reg_bounds(struct bpf_reg_state *reg)
1160 __update_reg32_bounds(reg);
1161 __update_reg64_bounds(reg);
1164 /* Uses signed min/max values to inform unsigned, and vice-versa */
1165 static void __reg32_deduce_bounds(struct bpf_reg_state *reg)
1167 /* Learn sign from signed bounds.
1168 * If we cannot cross the sign boundary, then signed and unsigned bounds
1169 * are the same, so combine. This works even in the negative case, e.g.
1170 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1172 if (reg->s32_min_value >= 0 || reg->s32_max_value < 0) {
1173 reg->s32_min_value = reg->u32_min_value =
1174 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1175 reg->s32_max_value = reg->u32_max_value =
1176 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1179 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1180 * boundary, so we must be careful.
1182 if ((s32)reg->u32_max_value >= 0) {
1183 /* Positive. We can't learn anything from the smin, but smax
1184 * is positive, hence safe.
1186 reg->s32_min_value = reg->u32_min_value;
1187 reg->s32_max_value = reg->u32_max_value =
1188 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1189 } else if ((s32)reg->u32_min_value < 0) {
1190 /* Negative. We can't learn anything from the smax, but smin
1191 * is negative, hence safe.
1193 reg->s32_min_value = reg->u32_min_value =
1194 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1195 reg->s32_max_value = reg->u32_max_value;
1199 static void __reg64_deduce_bounds(struct bpf_reg_state *reg)
1201 /* Learn sign from signed bounds.
1202 * If we cannot cross the sign boundary, then signed and unsigned bounds
1203 * are the same, so combine. This works even in the negative case, e.g.
1204 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1206 if (reg->smin_value >= 0 || reg->smax_value < 0) {
1207 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1209 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1213 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1214 * boundary, so we must be careful.
1216 if ((s64)reg->umax_value >= 0) {
1217 /* Positive. We can't learn anything from the smin, but smax
1218 * is positive, hence safe.
1220 reg->smin_value = reg->umin_value;
1221 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1223 } else if ((s64)reg->umin_value < 0) {
1224 /* Negative. We can't learn anything from the smax, but smin
1225 * is negative, hence safe.
1227 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1229 reg->smax_value = reg->umax_value;
1233 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
1235 __reg32_deduce_bounds(reg);
1236 __reg64_deduce_bounds(reg);
1239 /* Attempts to improve var_off based on unsigned min/max information */
1240 static void __reg_bound_offset(struct bpf_reg_state *reg)
1242 struct tnum var64_off = tnum_intersect(reg->var_off,
1243 tnum_range(reg->umin_value,
1245 struct tnum var32_off = tnum_intersect(tnum_subreg(reg->var_off),
1246 tnum_range(reg->u32_min_value,
1247 reg->u32_max_value));
1249 reg->var_off = tnum_or(tnum_clear_subreg(var64_off), var32_off);
1252 static void __reg_assign_32_into_64(struct bpf_reg_state *reg)
1254 reg->umin_value = reg->u32_min_value;
1255 reg->umax_value = reg->u32_max_value;
1256 /* Attempt to pull 32-bit signed bounds into 64-bit bounds
1257 * but must be positive otherwise set to worse case bounds
1258 * and refine later from tnum.
1260 if (reg->s32_min_value >= 0 && reg->s32_max_value >= 0)
1261 reg->smax_value = reg->s32_max_value;
1263 reg->smax_value = U32_MAX;
1264 if (reg->s32_min_value >= 0)
1265 reg->smin_value = reg->s32_min_value;
1267 reg->smin_value = 0;
1270 static void __reg_combine_32_into_64(struct bpf_reg_state *reg)
1272 /* special case when 64-bit register has upper 32-bit register
1273 * zeroed. Typically happens after zext or <<32, >>32 sequence
1274 * allowing us to use 32-bit bounds directly,
1276 if (tnum_equals_const(tnum_clear_subreg(reg->var_off), 0)) {
1277 __reg_assign_32_into_64(reg);
1279 /* Otherwise the best we can do is push lower 32bit known and
1280 * unknown bits into register (var_off set from jmp logic)
1281 * then learn as much as possible from the 64-bit tnum
1282 * known and unknown bits. The previous smin/smax bounds are
1283 * invalid here because of jmp32 compare so mark them unknown
1284 * so they do not impact tnum bounds calculation.
1286 __mark_reg64_unbounded(reg);
1287 __update_reg_bounds(reg);
1290 /* Intersecting with the old var_off might have improved our bounds
1291 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1292 * then new var_off is (0; 0x7f...fc) which improves our umax.
1294 __reg_deduce_bounds(reg);
1295 __reg_bound_offset(reg);
1296 __update_reg_bounds(reg);
1299 static bool __reg64_bound_s32(s64 a)
1301 return a > S32_MIN && a < S32_MAX;
1304 static bool __reg64_bound_u32(u64 a)
1306 if (a > U32_MIN && a < U32_MAX)
1311 static void __reg_combine_64_into_32(struct bpf_reg_state *reg)
1313 __mark_reg32_unbounded(reg);
1315 if (__reg64_bound_s32(reg->smin_value) && __reg64_bound_s32(reg->smax_value)) {
1316 reg->s32_min_value = (s32)reg->smin_value;
1317 reg->s32_max_value = (s32)reg->smax_value;
1319 if (__reg64_bound_u32(reg->umin_value))
1320 reg->u32_min_value = (u32)reg->umin_value;
1321 if (__reg64_bound_u32(reg->umax_value))
1322 reg->u32_max_value = (u32)reg->umax_value;
1324 /* Intersecting with the old var_off might have improved our bounds
1325 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1326 * then new var_off is (0; 0x7f...fc) which improves our umax.
1328 __reg_deduce_bounds(reg);
1329 __reg_bound_offset(reg);
1330 __update_reg_bounds(reg);
1333 /* Mark a register as having a completely unknown (scalar) value. */
1334 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
1335 struct bpf_reg_state *reg)
1338 * Clear type, id, off, and union(map_ptr, range) and
1339 * padding between 'type' and union
1341 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
1342 reg->type = SCALAR_VALUE;
1343 reg->var_off = tnum_unknown;
1345 reg->precise = env->subprog_cnt > 1 || !env->bpf_capable;
1346 __mark_reg_unbounded(reg);
1349 static void mark_reg_unknown(struct bpf_verifier_env *env,
1350 struct bpf_reg_state *regs, u32 regno)
1352 if (WARN_ON(regno >= MAX_BPF_REG)) {
1353 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
1354 /* Something bad happened, let's kill all regs except FP */
1355 for (regno = 0; regno < BPF_REG_FP; regno++)
1356 __mark_reg_not_init(env, regs + regno);
1359 __mark_reg_unknown(env, regs + regno);
1362 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1363 struct bpf_reg_state *reg)
1365 __mark_reg_unknown(env, reg);
1366 reg->type = NOT_INIT;
1369 static void mark_reg_not_init(struct bpf_verifier_env *env,
1370 struct bpf_reg_state *regs, u32 regno)
1372 if (WARN_ON(regno >= MAX_BPF_REG)) {
1373 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
1374 /* Something bad happened, let's kill all regs except FP */
1375 for (regno = 0; regno < BPF_REG_FP; regno++)
1376 __mark_reg_not_init(env, regs + regno);
1379 __mark_reg_not_init(env, regs + regno);
1382 static void mark_btf_ld_reg(struct bpf_verifier_env *env,
1383 struct bpf_reg_state *regs, u32 regno,
1384 enum bpf_reg_type reg_type, u32 btf_id)
1386 if (reg_type == SCALAR_VALUE) {
1387 mark_reg_unknown(env, regs, regno);
1390 mark_reg_known_zero(env, regs, regno);
1391 regs[regno].type = PTR_TO_BTF_ID;
1392 regs[regno].btf_id = btf_id;
1395 #define DEF_NOT_SUBREG (0)
1396 static void init_reg_state(struct bpf_verifier_env *env,
1397 struct bpf_func_state *state)
1399 struct bpf_reg_state *regs = state->regs;
1402 for (i = 0; i < MAX_BPF_REG; i++) {
1403 mark_reg_not_init(env, regs, i);
1404 regs[i].live = REG_LIVE_NONE;
1405 regs[i].parent = NULL;
1406 regs[i].subreg_def = DEF_NOT_SUBREG;
1410 regs[BPF_REG_FP].type = PTR_TO_STACK;
1411 mark_reg_known_zero(env, regs, BPF_REG_FP);
1412 regs[BPF_REG_FP].frameno = state->frameno;
1415 #define BPF_MAIN_FUNC (-1)
1416 static void init_func_state(struct bpf_verifier_env *env,
1417 struct bpf_func_state *state,
1418 int callsite, int frameno, int subprogno)
1420 state->callsite = callsite;
1421 state->frameno = frameno;
1422 state->subprogno = subprogno;
1423 init_reg_state(env, state);
1427 SRC_OP, /* register is used as source operand */
1428 DST_OP, /* register is used as destination operand */
1429 DST_OP_NO_MARK /* same as above, check only, don't mark */
1432 static int cmp_subprogs(const void *a, const void *b)
1434 return ((struct bpf_subprog_info *)a)->start -
1435 ((struct bpf_subprog_info *)b)->start;
1438 static int find_subprog(struct bpf_verifier_env *env, int off)
1440 struct bpf_subprog_info *p;
1442 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
1443 sizeof(env->subprog_info[0]), cmp_subprogs);
1446 return p - env->subprog_info;
1450 static int add_subprog(struct bpf_verifier_env *env, int off)
1452 int insn_cnt = env->prog->len;
1455 if (off >= insn_cnt || off < 0) {
1456 verbose(env, "call to invalid destination\n");
1459 ret = find_subprog(env, off);
1462 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
1463 verbose(env, "too many subprograms\n");
1466 env->subprog_info[env->subprog_cnt++].start = off;
1467 sort(env->subprog_info, env->subprog_cnt,
1468 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
1472 static int check_subprogs(struct bpf_verifier_env *env)
1474 int i, ret, subprog_start, subprog_end, off, cur_subprog = 0;
1475 struct bpf_subprog_info *subprog = env->subprog_info;
1476 struct bpf_insn *insn = env->prog->insnsi;
1477 int insn_cnt = env->prog->len;
1479 /* Add entry function. */
1480 ret = add_subprog(env, 0);
1484 /* determine subprog starts. The end is one before the next starts */
1485 for (i = 0; i < insn_cnt; i++) {
1486 if (insn[i].code != (BPF_JMP | BPF_CALL))
1488 if (insn[i].src_reg != BPF_PSEUDO_CALL)
1490 if (!env->bpf_capable) {
1492 "function calls to other bpf functions are allowed for CAP_BPF and CAP_SYS_ADMIN\n");
1495 ret = add_subprog(env, i + insn[i].imm + 1);
1500 /* Add a fake 'exit' subprog which could simplify subprog iteration
1501 * logic. 'subprog_cnt' should not be increased.
1503 subprog[env->subprog_cnt].start = insn_cnt;
1505 if (env->log.level & BPF_LOG_LEVEL2)
1506 for (i = 0; i < env->subprog_cnt; i++)
1507 verbose(env, "func#%d @%d\n", i, subprog[i].start);
1509 /* now check that all jumps are within the same subprog */
1510 subprog_start = subprog[cur_subprog].start;
1511 subprog_end = subprog[cur_subprog + 1].start;
1512 for (i = 0; i < insn_cnt; i++) {
1513 u8 code = insn[i].code;
1515 if (code == (BPF_JMP | BPF_CALL) &&
1516 insn[i].imm == BPF_FUNC_tail_call &&
1517 insn[i].src_reg != BPF_PSEUDO_CALL)
1518 subprog[cur_subprog].has_tail_call = true;
1519 if (BPF_CLASS(code) == BPF_LD &&
1520 (BPF_MODE(code) == BPF_ABS || BPF_MODE(code) == BPF_IND))
1521 subprog[cur_subprog].has_ld_abs = true;
1522 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
1524 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
1526 off = i + insn[i].off + 1;
1527 if (off < subprog_start || off >= subprog_end) {
1528 verbose(env, "jump out of range from insn %d to %d\n", i, off);
1532 if (i == subprog_end - 1) {
1533 /* to avoid fall-through from one subprog into another
1534 * the last insn of the subprog should be either exit
1535 * or unconditional jump back
1537 if (code != (BPF_JMP | BPF_EXIT) &&
1538 code != (BPF_JMP | BPF_JA)) {
1539 verbose(env, "last insn is not an exit or jmp\n");
1542 subprog_start = subprog_end;
1544 if (cur_subprog < env->subprog_cnt)
1545 subprog_end = subprog[cur_subprog + 1].start;
1551 /* Parentage chain of this register (or stack slot) should take care of all
1552 * issues like callee-saved registers, stack slot allocation time, etc.
1554 static int mark_reg_read(struct bpf_verifier_env *env,
1555 const struct bpf_reg_state *state,
1556 struct bpf_reg_state *parent, u8 flag)
1558 bool writes = parent == state->parent; /* Observe write marks */
1562 /* if read wasn't screened by an earlier write ... */
1563 if (writes && state->live & REG_LIVE_WRITTEN)
1565 if (parent->live & REG_LIVE_DONE) {
1566 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
1567 reg_type_str[parent->type],
1568 parent->var_off.value, parent->off);
1571 /* The first condition is more likely to be true than the
1572 * second, checked it first.
1574 if ((parent->live & REG_LIVE_READ) == flag ||
1575 parent->live & REG_LIVE_READ64)
1576 /* The parentage chain never changes and
1577 * this parent was already marked as LIVE_READ.
1578 * There is no need to keep walking the chain again and
1579 * keep re-marking all parents as LIVE_READ.
1580 * This case happens when the same register is read
1581 * multiple times without writes into it in-between.
1582 * Also, if parent has the stronger REG_LIVE_READ64 set,
1583 * then no need to set the weak REG_LIVE_READ32.
1586 /* ... then we depend on parent's value */
1587 parent->live |= flag;
1588 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
1589 if (flag == REG_LIVE_READ64)
1590 parent->live &= ~REG_LIVE_READ32;
1592 parent = state->parent;
1597 if (env->longest_mark_read_walk < cnt)
1598 env->longest_mark_read_walk = cnt;
1602 /* This function is supposed to be used by the following 32-bit optimization
1603 * code only. It returns TRUE if the source or destination register operates
1604 * on 64-bit, otherwise return FALSE.
1606 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
1607 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
1612 class = BPF_CLASS(code);
1614 if (class == BPF_JMP) {
1615 /* BPF_EXIT for "main" will reach here. Return TRUE
1620 if (op == BPF_CALL) {
1621 /* BPF to BPF call will reach here because of marking
1622 * caller saved clobber with DST_OP_NO_MARK for which we
1623 * don't care the register def because they are anyway
1624 * marked as NOT_INIT already.
1626 if (insn->src_reg == BPF_PSEUDO_CALL)
1628 /* Helper call will reach here because of arg type
1629 * check, conservatively return TRUE.
1638 if (class == BPF_ALU64 || class == BPF_JMP ||
1639 /* BPF_END always use BPF_ALU class. */
1640 (class == BPF_ALU && op == BPF_END && insn->imm == 64))
1643 if (class == BPF_ALU || class == BPF_JMP32)
1646 if (class == BPF_LDX) {
1648 return BPF_SIZE(code) == BPF_DW;
1649 /* LDX source must be ptr. */
1653 if (class == BPF_STX) {
1654 if (reg->type != SCALAR_VALUE)
1656 return BPF_SIZE(code) == BPF_DW;
1659 if (class == BPF_LD) {
1660 u8 mode = BPF_MODE(code);
1663 if (mode == BPF_IMM)
1666 /* Both LD_IND and LD_ABS return 32-bit data. */
1670 /* Implicit ctx ptr. */
1671 if (regno == BPF_REG_6)
1674 /* Explicit source could be any width. */
1678 if (class == BPF_ST)
1679 /* The only source register for BPF_ST is a ptr. */
1682 /* Conservatively return true at default. */
1686 /* Return TRUE if INSN doesn't have explicit value define. */
1687 static bool insn_no_def(struct bpf_insn *insn)
1689 u8 class = BPF_CLASS(insn->code);
1691 return (class == BPF_JMP || class == BPF_JMP32 ||
1692 class == BPF_STX || class == BPF_ST);
1695 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
1696 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
1698 if (insn_no_def(insn))
1701 return !is_reg64(env, insn, insn->dst_reg, NULL, DST_OP);
1704 static void mark_insn_zext(struct bpf_verifier_env *env,
1705 struct bpf_reg_state *reg)
1707 s32 def_idx = reg->subreg_def;
1709 if (def_idx == DEF_NOT_SUBREG)
1712 env->insn_aux_data[def_idx - 1].zext_dst = true;
1713 /* The dst will be zero extended, so won't be sub-register anymore. */
1714 reg->subreg_def = DEF_NOT_SUBREG;
1717 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
1718 enum reg_arg_type t)
1720 struct bpf_verifier_state *vstate = env->cur_state;
1721 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1722 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
1723 struct bpf_reg_state *reg, *regs = state->regs;
1726 if (regno >= MAX_BPF_REG) {
1727 verbose(env, "R%d is invalid\n", regno);
1732 rw64 = is_reg64(env, insn, regno, reg, t);
1734 /* check whether register used as source operand can be read */
1735 if (reg->type == NOT_INIT) {
1736 verbose(env, "R%d !read_ok\n", regno);
1739 /* We don't need to worry about FP liveness because it's read-only */
1740 if (regno == BPF_REG_FP)
1744 mark_insn_zext(env, reg);
1746 return mark_reg_read(env, reg, reg->parent,
1747 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
1749 /* check whether register used as dest operand can be written to */
1750 if (regno == BPF_REG_FP) {
1751 verbose(env, "frame pointer is read only\n");
1754 reg->live |= REG_LIVE_WRITTEN;
1755 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
1757 mark_reg_unknown(env, regs, regno);
1762 /* for any branch, call, exit record the history of jmps in the given state */
1763 static int push_jmp_history(struct bpf_verifier_env *env,
1764 struct bpf_verifier_state *cur)
1766 u32 cnt = cur->jmp_history_cnt;
1767 struct bpf_idx_pair *p;
1770 p = krealloc(cur->jmp_history, cnt * sizeof(*p), GFP_USER);
1773 p[cnt - 1].idx = env->insn_idx;
1774 p[cnt - 1].prev_idx = env->prev_insn_idx;
1775 cur->jmp_history = p;
1776 cur->jmp_history_cnt = cnt;
1780 /* Backtrack one insn at a time. If idx is not at the top of recorded
1781 * history then previous instruction came from straight line execution.
1783 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
1788 if (cnt && st->jmp_history[cnt - 1].idx == i) {
1789 i = st->jmp_history[cnt - 1].prev_idx;
1797 /* For given verifier state backtrack_insn() is called from the last insn to
1798 * the first insn. Its purpose is to compute a bitmask of registers and
1799 * stack slots that needs precision in the parent verifier state.
1801 static int backtrack_insn(struct bpf_verifier_env *env, int idx,
1802 u32 *reg_mask, u64 *stack_mask)
1804 const struct bpf_insn_cbs cbs = {
1805 .cb_print = verbose,
1806 .private_data = env,
1808 struct bpf_insn *insn = env->prog->insnsi + idx;
1809 u8 class = BPF_CLASS(insn->code);
1810 u8 opcode = BPF_OP(insn->code);
1811 u8 mode = BPF_MODE(insn->code);
1812 u32 dreg = 1u << insn->dst_reg;
1813 u32 sreg = 1u << insn->src_reg;
1816 if (insn->code == 0)
1818 if (env->log.level & BPF_LOG_LEVEL) {
1819 verbose(env, "regs=%x stack=%llx before ", *reg_mask, *stack_mask);
1820 verbose(env, "%d: ", idx);
1821 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
1824 if (class == BPF_ALU || class == BPF_ALU64) {
1825 if (!(*reg_mask & dreg))
1827 if (opcode == BPF_MOV) {
1828 if (BPF_SRC(insn->code) == BPF_X) {
1830 * dreg needs precision after this insn
1831 * sreg needs precision before this insn
1837 * dreg needs precision after this insn.
1838 * Corresponding register is already marked
1839 * as precise=true in this verifier state.
1840 * No further markings in parent are necessary
1845 if (BPF_SRC(insn->code) == BPF_X) {
1847 * both dreg and sreg need precision
1852 * dreg still needs precision before this insn
1855 } else if (class == BPF_LDX) {
1856 if (!(*reg_mask & dreg))
1860 /* scalars can only be spilled into stack w/o losing precision.
1861 * Load from any other memory can be zero extended.
1862 * The desire to keep that precision is already indicated
1863 * by 'precise' mark in corresponding register of this state.
1864 * No further tracking necessary.
1866 if (insn->src_reg != BPF_REG_FP)
1868 if (BPF_SIZE(insn->code) != BPF_DW)
1871 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
1872 * that [fp - off] slot contains scalar that needs to be
1873 * tracked with precision
1875 spi = (-insn->off - 1) / BPF_REG_SIZE;
1877 verbose(env, "BUG spi %d\n", spi);
1878 WARN_ONCE(1, "verifier backtracking bug");
1881 *stack_mask |= 1ull << spi;
1882 } else if (class == BPF_STX || class == BPF_ST) {
1883 if (*reg_mask & dreg)
1884 /* stx & st shouldn't be using _scalar_ dst_reg
1885 * to access memory. It means backtracking
1886 * encountered a case of pointer subtraction.
1889 /* scalars can only be spilled into stack */
1890 if (insn->dst_reg != BPF_REG_FP)
1892 if (BPF_SIZE(insn->code) != BPF_DW)
1894 spi = (-insn->off - 1) / BPF_REG_SIZE;
1896 verbose(env, "BUG spi %d\n", spi);
1897 WARN_ONCE(1, "verifier backtracking bug");
1900 if (!(*stack_mask & (1ull << spi)))
1902 *stack_mask &= ~(1ull << spi);
1903 if (class == BPF_STX)
1905 } else if (class == BPF_JMP || class == BPF_JMP32) {
1906 if (opcode == BPF_CALL) {
1907 if (insn->src_reg == BPF_PSEUDO_CALL)
1909 /* regular helper call sets R0 */
1911 if (*reg_mask & 0x3f) {
1912 /* if backtracing was looking for registers R1-R5
1913 * they should have been found already.
1915 verbose(env, "BUG regs %x\n", *reg_mask);
1916 WARN_ONCE(1, "verifier backtracking bug");
1919 } else if (opcode == BPF_EXIT) {
1922 } else if (class == BPF_LD) {
1923 if (!(*reg_mask & dreg))
1926 /* It's ld_imm64 or ld_abs or ld_ind.
1927 * For ld_imm64 no further tracking of precision
1928 * into parent is necessary
1930 if (mode == BPF_IND || mode == BPF_ABS)
1931 /* to be analyzed */
1937 /* the scalar precision tracking algorithm:
1938 * . at the start all registers have precise=false.
1939 * . scalar ranges are tracked as normal through alu and jmp insns.
1940 * . once precise value of the scalar register is used in:
1941 * . ptr + scalar alu
1942 * . if (scalar cond K|scalar)
1943 * . helper_call(.., scalar, ...) where ARG_CONST is expected
1944 * backtrack through the verifier states and mark all registers and
1945 * stack slots with spilled constants that these scalar regisers
1946 * should be precise.
1947 * . during state pruning two registers (or spilled stack slots)
1948 * are equivalent if both are not precise.
1950 * Note the verifier cannot simply walk register parentage chain,
1951 * since many different registers and stack slots could have been
1952 * used to compute single precise scalar.
1954 * The approach of starting with precise=true for all registers and then
1955 * backtrack to mark a register as not precise when the verifier detects
1956 * that program doesn't care about specific value (e.g., when helper
1957 * takes register as ARG_ANYTHING parameter) is not safe.
1959 * It's ok to walk single parentage chain of the verifier states.
1960 * It's possible that this backtracking will go all the way till 1st insn.
1961 * All other branches will be explored for needing precision later.
1963 * The backtracking needs to deal with cases like:
1964 * 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)
1967 * if r5 > 0x79f goto pc+7
1968 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
1971 * call bpf_perf_event_output#25
1972 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
1976 * call foo // uses callee's r6 inside to compute r0
1980 * to track above reg_mask/stack_mask needs to be independent for each frame.
1982 * Also if parent's curframe > frame where backtracking started,
1983 * the verifier need to mark registers in both frames, otherwise callees
1984 * may incorrectly prune callers. This is similar to
1985 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
1987 * For now backtracking falls back into conservative marking.
1989 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
1990 struct bpf_verifier_state *st)
1992 struct bpf_func_state *func;
1993 struct bpf_reg_state *reg;
1996 /* big hammer: mark all scalars precise in this path.
1997 * pop_stack may still get !precise scalars.
1999 for (; st; st = st->parent)
2000 for (i = 0; i <= st->curframe; i++) {
2001 func = st->frame[i];
2002 for (j = 0; j < BPF_REG_FP; j++) {
2003 reg = &func->regs[j];
2004 if (reg->type != SCALAR_VALUE)
2006 reg->precise = true;
2008 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
2009 if (func->stack[j].slot_type[0] != STACK_SPILL)
2011 reg = &func->stack[j].spilled_ptr;
2012 if (reg->type != SCALAR_VALUE)
2014 reg->precise = true;
2019 static int __mark_chain_precision(struct bpf_verifier_env *env, int regno,
2022 struct bpf_verifier_state *st = env->cur_state;
2023 int first_idx = st->first_insn_idx;
2024 int last_idx = env->insn_idx;
2025 struct bpf_func_state *func;
2026 struct bpf_reg_state *reg;
2027 u32 reg_mask = regno >= 0 ? 1u << regno : 0;
2028 u64 stack_mask = spi >= 0 ? 1ull << spi : 0;
2029 bool skip_first = true;
2030 bool new_marks = false;
2033 if (!env->bpf_capable)
2036 func = st->frame[st->curframe];
2038 reg = &func->regs[regno];
2039 if (reg->type != SCALAR_VALUE) {
2040 WARN_ONCE(1, "backtracing misuse");
2047 reg->precise = true;
2051 if (func->stack[spi].slot_type[0] != STACK_SPILL) {
2055 reg = &func->stack[spi].spilled_ptr;
2056 if (reg->type != SCALAR_VALUE) {
2064 reg->precise = true;
2070 if (!reg_mask && !stack_mask)
2073 DECLARE_BITMAP(mask, 64);
2074 u32 history = st->jmp_history_cnt;
2076 if (env->log.level & BPF_LOG_LEVEL)
2077 verbose(env, "last_idx %d first_idx %d\n", last_idx, first_idx);
2078 for (i = last_idx;;) {
2083 err = backtrack_insn(env, i, ®_mask, &stack_mask);
2085 if (err == -ENOTSUPP) {
2086 mark_all_scalars_precise(env, st);
2091 if (!reg_mask && !stack_mask)
2092 /* Found assignment(s) into tracked register in this state.
2093 * Since this state is already marked, just return.
2094 * Nothing to be tracked further in the parent state.
2099 i = get_prev_insn_idx(st, i, &history);
2100 if (i >= env->prog->len) {
2101 /* This can happen if backtracking reached insn 0
2102 * and there are still reg_mask or stack_mask
2104 * It means the backtracking missed the spot where
2105 * particular register was initialized with a constant.
2107 verbose(env, "BUG backtracking idx %d\n", i);
2108 WARN_ONCE(1, "verifier backtracking bug");
2117 func = st->frame[st->curframe];
2118 bitmap_from_u64(mask, reg_mask);
2119 for_each_set_bit(i, mask, 32) {
2120 reg = &func->regs[i];
2121 if (reg->type != SCALAR_VALUE) {
2122 reg_mask &= ~(1u << i);
2127 reg->precise = true;
2130 bitmap_from_u64(mask, stack_mask);
2131 for_each_set_bit(i, mask, 64) {
2132 if (i >= func->allocated_stack / BPF_REG_SIZE) {
2133 /* the sequence of instructions:
2135 * 3: (7b) *(u64 *)(r3 -8) = r0
2136 * 4: (79) r4 = *(u64 *)(r10 -8)
2137 * doesn't contain jmps. It's backtracked
2138 * as a single block.
2139 * During backtracking insn 3 is not recognized as
2140 * stack access, so at the end of backtracking
2141 * stack slot fp-8 is still marked in stack_mask.
2142 * However the parent state may not have accessed
2143 * fp-8 and it's "unallocated" stack space.
2144 * In such case fallback to conservative.
2146 mark_all_scalars_precise(env, st);
2150 if (func->stack[i].slot_type[0] != STACK_SPILL) {
2151 stack_mask &= ~(1ull << i);
2154 reg = &func->stack[i].spilled_ptr;
2155 if (reg->type != SCALAR_VALUE) {
2156 stack_mask &= ~(1ull << i);
2161 reg->precise = true;
2163 if (env->log.level & BPF_LOG_LEVEL) {
2164 print_verifier_state(env, func);
2165 verbose(env, "parent %s regs=%x stack=%llx marks\n",
2166 new_marks ? "didn't have" : "already had",
2167 reg_mask, stack_mask);
2170 if (!reg_mask && !stack_mask)
2175 last_idx = st->last_insn_idx;
2176 first_idx = st->first_insn_idx;
2181 static int mark_chain_precision(struct bpf_verifier_env *env, int regno)
2183 return __mark_chain_precision(env, regno, -1);
2186 static int mark_chain_precision_stack(struct bpf_verifier_env *env, int spi)
2188 return __mark_chain_precision(env, -1, spi);
2191 static bool is_spillable_regtype(enum bpf_reg_type type)
2194 case PTR_TO_MAP_VALUE:
2195 case PTR_TO_MAP_VALUE_OR_NULL:
2199 case PTR_TO_PACKET_META:
2200 case PTR_TO_PACKET_END:
2201 case PTR_TO_FLOW_KEYS:
2202 case CONST_PTR_TO_MAP:
2204 case PTR_TO_SOCKET_OR_NULL:
2205 case PTR_TO_SOCK_COMMON:
2206 case PTR_TO_SOCK_COMMON_OR_NULL:
2207 case PTR_TO_TCP_SOCK:
2208 case PTR_TO_TCP_SOCK_OR_NULL:
2209 case PTR_TO_XDP_SOCK:
2211 case PTR_TO_BTF_ID_OR_NULL:
2212 case PTR_TO_RDONLY_BUF:
2213 case PTR_TO_RDONLY_BUF_OR_NULL:
2214 case PTR_TO_RDWR_BUF:
2215 case PTR_TO_RDWR_BUF_OR_NULL:
2216 case PTR_TO_PERCPU_BTF_ID:
2218 case PTR_TO_MEM_OR_NULL:
2225 /* Does this register contain a constant zero? */
2226 static bool register_is_null(struct bpf_reg_state *reg)
2228 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
2231 static bool register_is_const(struct bpf_reg_state *reg)
2233 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
2236 static bool __is_scalar_unbounded(struct bpf_reg_state *reg)
2238 return tnum_is_unknown(reg->var_off) &&
2239 reg->smin_value == S64_MIN && reg->smax_value == S64_MAX &&
2240 reg->umin_value == 0 && reg->umax_value == U64_MAX &&
2241 reg->s32_min_value == S32_MIN && reg->s32_max_value == S32_MAX &&
2242 reg->u32_min_value == 0 && reg->u32_max_value == U32_MAX;
2245 static bool register_is_bounded(struct bpf_reg_state *reg)
2247 return reg->type == SCALAR_VALUE && !__is_scalar_unbounded(reg);
2250 static bool __is_pointer_value(bool allow_ptr_leaks,
2251 const struct bpf_reg_state *reg)
2253 if (allow_ptr_leaks)
2256 return reg->type != SCALAR_VALUE;
2259 static void save_register_state(struct bpf_func_state *state,
2260 int spi, struct bpf_reg_state *reg)
2264 state->stack[spi].spilled_ptr = *reg;
2265 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2267 for (i = 0; i < BPF_REG_SIZE; i++)
2268 state->stack[spi].slot_type[i] = STACK_SPILL;
2271 /* check_stack_read/write functions track spill/fill of registers,
2272 * stack boundary and alignment are checked in check_mem_access()
2274 static int check_stack_write(struct bpf_verifier_env *env,
2275 struct bpf_func_state *state, /* func where register points to */
2276 int off, int size, int value_regno, int insn_idx)
2278 struct bpf_func_state *cur; /* state of the current function */
2279 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
2280 u32 dst_reg = env->prog->insnsi[insn_idx].dst_reg;
2281 struct bpf_reg_state *reg = NULL;
2283 err = realloc_func_state(state, round_up(slot + 1, BPF_REG_SIZE),
2284 state->acquired_refs, true);
2287 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
2288 * so it's aligned access and [off, off + size) are within stack limits
2290 if (!env->allow_ptr_leaks &&
2291 state->stack[spi].slot_type[0] == STACK_SPILL &&
2292 size != BPF_REG_SIZE) {
2293 verbose(env, "attempt to corrupt spilled pointer on stack\n");
2297 cur = env->cur_state->frame[env->cur_state->curframe];
2298 if (value_regno >= 0)
2299 reg = &cur->regs[value_regno];
2301 if (reg && size == BPF_REG_SIZE && register_is_bounded(reg) &&
2302 !register_is_null(reg) && env->bpf_capable) {
2303 if (dst_reg != BPF_REG_FP) {
2304 /* The backtracking logic can only recognize explicit
2305 * stack slot address like [fp - 8]. Other spill of
2306 * scalar via different register has to be conervative.
2307 * Backtrack from here and mark all registers as precise
2308 * that contributed into 'reg' being a constant.
2310 err = mark_chain_precision(env, value_regno);
2314 save_register_state(state, spi, reg);
2315 } else if (reg && is_spillable_regtype(reg->type)) {
2316 /* register containing pointer is being spilled into stack */
2317 if (size != BPF_REG_SIZE) {
2318 verbose_linfo(env, insn_idx, "; ");
2319 verbose(env, "invalid size of register spill\n");
2323 if (state != cur && reg->type == PTR_TO_STACK) {
2324 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
2328 if (!env->bypass_spec_v4) {
2329 bool sanitize = false;
2331 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
2332 register_is_const(&state->stack[spi].spilled_ptr))
2334 for (i = 0; i < BPF_REG_SIZE; i++)
2335 if (state->stack[spi].slot_type[i] == STACK_MISC) {
2340 int *poff = &env->insn_aux_data[insn_idx].sanitize_stack_off;
2341 int soff = (-spi - 1) * BPF_REG_SIZE;
2343 /* detected reuse of integer stack slot with a pointer
2344 * which means either llvm is reusing stack slot or
2345 * an attacker is trying to exploit CVE-2018-3639
2346 * (speculative store bypass)
2347 * Have to sanitize that slot with preemptive
2350 if (*poff && *poff != soff) {
2351 /* disallow programs where single insn stores
2352 * into two different stack slots, since verifier
2353 * cannot sanitize them
2356 "insn %d cannot access two stack slots fp%d and fp%d",
2357 insn_idx, *poff, soff);
2363 save_register_state(state, spi, reg);
2365 u8 type = STACK_MISC;
2367 /* regular write of data into stack destroys any spilled ptr */
2368 state->stack[spi].spilled_ptr.type = NOT_INIT;
2369 /* Mark slots as STACK_MISC if they belonged to spilled ptr. */
2370 if (state->stack[spi].slot_type[0] == STACK_SPILL)
2371 for (i = 0; i < BPF_REG_SIZE; i++)
2372 state->stack[spi].slot_type[i] = STACK_MISC;
2374 /* only mark the slot as written if all 8 bytes were written
2375 * otherwise read propagation may incorrectly stop too soon
2376 * when stack slots are partially written.
2377 * This heuristic means that read propagation will be
2378 * conservative, since it will add reg_live_read marks
2379 * to stack slots all the way to first state when programs
2380 * writes+reads less than 8 bytes
2382 if (size == BPF_REG_SIZE)
2383 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2385 /* when we zero initialize stack slots mark them as such */
2386 if (reg && register_is_null(reg)) {
2387 /* backtracking doesn't work for STACK_ZERO yet. */
2388 err = mark_chain_precision(env, value_regno);
2394 /* Mark slots affected by this stack write. */
2395 for (i = 0; i < size; i++)
2396 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
2402 static int check_stack_read(struct bpf_verifier_env *env,
2403 struct bpf_func_state *reg_state /* func where register points to */,
2404 int off, int size, int value_regno)
2406 struct bpf_verifier_state *vstate = env->cur_state;
2407 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2408 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
2409 struct bpf_reg_state *reg;
2412 if (reg_state->allocated_stack <= slot) {
2413 verbose(env, "invalid read from stack off %d+0 size %d\n",
2417 stype = reg_state->stack[spi].slot_type;
2418 reg = ®_state->stack[spi].spilled_ptr;
2420 if (stype[0] == STACK_SPILL) {
2421 if (size != BPF_REG_SIZE) {
2422 if (reg->type != SCALAR_VALUE) {
2423 verbose_linfo(env, env->insn_idx, "; ");
2424 verbose(env, "invalid size of register fill\n");
2427 if (value_regno >= 0) {
2428 mark_reg_unknown(env, state->regs, value_regno);
2429 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
2431 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2434 for (i = 1; i < BPF_REG_SIZE; i++) {
2435 if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) {
2436 verbose(env, "corrupted spill memory\n");
2441 if (value_regno >= 0) {
2442 /* restore register state from stack */
2443 state->regs[value_regno] = *reg;
2444 /* mark reg as written since spilled pointer state likely
2445 * has its liveness marks cleared by is_state_visited()
2446 * which resets stack/reg liveness for state transitions
2448 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
2449 } else if (__is_pointer_value(env->allow_ptr_leaks, reg)) {
2450 /* If value_regno==-1, the caller is asking us whether
2451 * it is acceptable to use this value as a SCALAR_VALUE
2453 * We must not allow unprivileged callers to do that
2454 * with spilled pointers.
2456 verbose(env, "leaking pointer from stack off %d\n",
2460 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2464 for (i = 0; i < size; i++) {
2465 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_MISC)
2467 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_ZERO) {
2471 verbose(env, "invalid read from stack off %d+%d size %d\n",
2475 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2476 if (value_regno >= 0) {
2477 if (zeros == size) {
2478 /* any size read into register is zero extended,
2479 * so the whole register == const_zero
2481 __mark_reg_const_zero(&state->regs[value_regno]);
2482 /* backtracking doesn't support STACK_ZERO yet,
2483 * so mark it precise here, so that later
2484 * backtracking can stop here.
2485 * Backtracking may not need this if this register
2486 * doesn't participate in pointer adjustment.
2487 * Forward propagation of precise flag is not
2488 * necessary either. This mark is only to stop
2489 * backtracking. Any register that contributed
2490 * to const 0 was marked precise before spill.
2492 state->regs[value_regno].precise = true;
2494 /* have read misc data from the stack */
2495 mark_reg_unknown(env, state->regs, value_regno);
2497 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
2503 static int check_stack_access(struct bpf_verifier_env *env,
2504 const struct bpf_reg_state *reg,
2507 /* Stack accesses must be at a fixed offset, so that we
2508 * can determine what type of data were returned. See
2509 * check_stack_read().
2511 if (!tnum_is_const(reg->var_off)) {
2514 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2515 verbose(env, "variable stack access var_off=%s off=%d size=%d\n",
2520 if (off >= 0 || off < -MAX_BPF_STACK) {
2521 verbose(env, "invalid stack off=%d size=%d\n", off, size);
2528 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
2529 int off, int size, enum bpf_access_type type)
2531 struct bpf_reg_state *regs = cur_regs(env);
2532 struct bpf_map *map = regs[regno].map_ptr;
2533 u32 cap = bpf_map_flags_to_cap(map);
2535 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
2536 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
2537 map->value_size, off, size);
2541 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
2542 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
2543 map->value_size, off, size);
2550 /* check read/write into memory region (e.g., map value, ringbuf sample, etc) */
2551 static int __check_mem_access(struct bpf_verifier_env *env, int regno,
2552 int off, int size, u32 mem_size,
2553 bool zero_size_allowed)
2555 bool size_ok = size > 0 || (size == 0 && zero_size_allowed);
2556 struct bpf_reg_state *reg;
2558 if (off >= 0 && size_ok && (u64)off + size <= mem_size)
2561 reg = &cur_regs(env)[regno];
2562 switch (reg->type) {
2563 case PTR_TO_MAP_VALUE:
2564 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
2565 mem_size, off, size);
2568 case PTR_TO_PACKET_META:
2569 case PTR_TO_PACKET_END:
2570 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
2571 off, size, regno, reg->id, off, mem_size);
2575 verbose(env, "invalid access to memory, mem_size=%u off=%d size=%d\n",
2576 mem_size, off, size);
2582 /* check read/write into a memory region with possible variable offset */
2583 static int check_mem_region_access(struct bpf_verifier_env *env, u32 regno,
2584 int off, int size, u32 mem_size,
2585 bool zero_size_allowed)
2587 struct bpf_verifier_state *vstate = env->cur_state;
2588 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2589 struct bpf_reg_state *reg = &state->regs[regno];
2592 /* We may have adjusted the register pointing to memory region, so we
2593 * need to try adding each of min_value and max_value to off
2594 * to make sure our theoretical access will be safe.
2596 if (env->log.level & BPF_LOG_LEVEL)
2597 print_verifier_state(env, state);
2599 /* The minimum value is only important with signed
2600 * comparisons where we can't assume the floor of a
2601 * value is 0. If we are using signed variables for our
2602 * index'es we need to make sure that whatever we use
2603 * will have a set floor within our range.
2605 if (reg->smin_value < 0 &&
2606 (reg->smin_value == S64_MIN ||
2607 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
2608 reg->smin_value + off < 0)) {
2609 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2613 err = __check_mem_access(env, regno, reg->smin_value + off, size,
2614 mem_size, zero_size_allowed);
2616 verbose(env, "R%d min value is outside of the allowed memory range\n",
2621 /* If we haven't set a max value then we need to bail since we can't be
2622 * sure we won't do bad things.
2623 * If reg->umax_value + off could overflow, treat that as unbounded too.
2625 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
2626 verbose(env, "R%d unbounded memory access, make sure to bounds check any such access\n",
2630 err = __check_mem_access(env, regno, reg->umax_value + off, size,
2631 mem_size, zero_size_allowed);
2633 verbose(env, "R%d max value is outside of the allowed memory range\n",
2641 /* check read/write into a map element with possible variable offset */
2642 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
2643 int off, int size, bool zero_size_allowed)
2645 struct bpf_verifier_state *vstate = env->cur_state;
2646 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2647 struct bpf_reg_state *reg = &state->regs[regno];
2648 struct bpf_map *map = reg->map_ptr;
2651 err = check_mem_region_access(env, regno, off, size, map->value_size,
2656 if (map_value_has_spin_lock(map)) {
2657 u32 lock = map->spin_lock_off;
2659 /* if any part of struct bpf_spin_lock can be touched by
2660 * load/store reject this program.
2661 * To check that [x1, x2) overlaps with [y1, y2)
2662 * it is sufficient to check x1 < y2 && y1 < x2.
2664 if (reg->smin_value + off < lock + sizeof(struct bpf_spin_lock) &&
2665 lock < reg->umax_value + off + size) {
2666 verbose(env, "bpf_spin_lock cannot be accessed directly by load/store\n");
2673 #define MAX_PACKET_OFF 0xffff
2675 static enum bpf_prog_type resolve_prog_type(struct bpf_prog *prog)
2677 return prog->aux->dst_prog ? prog->aux->dst_prog->type : prog->type;
2680 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
2681 const struct bpf_call_arg_meta *meta,
2682 enum bpf_access_type t)
2684 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
2686 switch (prog_type) {
2687 /* Program types only with direct read access go here! */
2688 case BPF_PROG_TYPE_LWT_IN:
2689 case BPF_PROG_TYPE_LWT_OUT:
2690 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
2691 case BPF_PROG_TYPE_SK_REUSEPORT:
2692 case BPF_PROG_TYPE_FLOW_DISSECTOR:
2693 case BPF_PROG_TYPE_CGROUP_SKB:
2698 /* Program types with direct read + write access go here! */
2699 case BPF_PROG_TYPE_SCHED_CLS:
2700 case BPF_PROG_TYPE_SCHED_ACT:
2701 case BPF_PROG_TYPE_XDP:
2702 case BPF_PROG_TYPE_LWT_XMIT:
2703 case BPF_PROG_TYPE_SK_SKB:
2704 case BPF_PROG_TYPE_SK_MSG:
2706 return meta->pkt_access;
2708 env->seen_direct_write = true;
2711 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
2713 env->seen_direct_write = true;
2722 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
2723 int size, bool zero_size_allowed)
2725 struct bpf_reg_state *regs = cur_regs(env);
2726 struct bpf_reg_state *reg = ®s[regno];
2729 /* We may have added a variable offset to the packet pointer; but any
2730 * reg->range we have comes after that. We are only checking the fixed
2734 /* We don't allow negative numbers, because we aren't tracking enough
2735 * detail to prove they're safe.
2737 if (reg->smin_value < 0) {
2738 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2742 err = __check_mem_access(env, regno, off, size, reg->range,
2745 verbose(env, "R%d offset is outside of the packet\n", regno);
2749 /* __check_mem_access has made sure "off + size - 1" is within u16.
2750 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
2751 * otherwise find_good_pkt_pointers would have refused to set range info
2752 * that __check_mem_access would have rejected this pkt access.
2753 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
2755 env->prog->aux->max_pkt_offset =
2756 max_t(u32, env->prog->aux->max_pkt_offset,
2757 off + reg->umax_value + size - 1);
2762 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
2763 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
2764 enum bpf_access_type t, enum bpf_reg_type *reg_type,
2767 struct bpf_insn_access_aux info = {
2768 .reg_type = *reg_type,
2772 if (env->ops->is_valid_access &&
2773 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
2774 /* A non zero info.ctx_field_size indicates that this field is a
2775 * candidate for later verifier transformation to load the whole
2776 * field and then apply a mask when accessed with a narrower
2777 * access than actual ctx access size. A zero info.ctx_field_size
2778 * will only allow for whole field access and rejects any other
2779 * type of narrower access.
2781 *reg_type = info.reg_type;
2783 if (*reg_type == PTR_TO_BTF_ID || *reg_type == PTR_TO_BTF_ID_OR_NULL)
2784 *btf_id = info.btf_id;
2786 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
2787 /* remember the offset of last byte accessed in ctx */
2788 if (env->prog->aux->max_ctx_offset < off + size)
2789 env->prog->aux->max_ctx_offset = off + size;
2793 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
2797 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
2800 if (size < 0 || off < 0 ||
2801 (u64)off + size > sizeof(struct bpf_flow_keys)) {
2802 verbose(env, "invalid access to flow keys off=%d size=%d\n",
2809 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
2810 u32 regno, int off, int size,
2811 enum bpf_access_type t)
2813 struct bpf_reg_state *regs = cur_regs(env);
2814 struct bpf_reg_state *reg = ®s[regno];
2815 struct bpf_insn_access_aux info = {};
2818 if (reg->smin_value < 0) {
2819 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2824 switch (reg->type) {
2825 case PTR_TO_SOCK_COMMON:
2826 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
2829 valid = bpf_sock_is_valid_access(off, size, t, &info);
2831 case PTR_TO_TCP_SOCK:
2832 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
2834 case PTR_TO_XDP_SOCK:
2835 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
2843 env->insn_aux_data[insn_idx].ctx_field_size =
2844 info.ctx_field_size;
2848 verbose(env, "R%d invalid %s access off=%d size=%d\n",
2849 regno, reg_type_str[reg->type], off, size);
2854 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
2856 return cur_regs(env) + regno;
2859 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
2861 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
2864 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
2866 const struct bpf_reg_state *reg = reg_state(env, regno);
2868 return reg->type == PTR_TO_CTX;
2871 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
2873 const struct bpf_reg_state *reg = reg_state(env, regno);
2875 return type_is_sk_pointer(reg->type);
2878 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
2880 const struct bpf_reg_state *reg = reg_state(env, regno);
2882 return type_is_pkt_pointer(reg->type);
2885 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
2887 const struct bpf_reg_state *reg = reg_state(env, regno);
2889 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
2890 return reg->type == PTR_TO_FLOW_KEYS;
2893 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
2894 const struct bpf_reg_state *reg,
2895 int off, int size, bool strict)
2897 struct tnum reg_off;
2900 /* Byte size accesses are always allowed. */
2901 if (!strict || size == 1)
2904 /* For platforms that do not have a Kconfig enabling
2905 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
2906 * NET_IP_ALIGN is universally set to '2'. And on platforms
2907 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
2908 * to this code only in strict mode where we want to emulate
2909 * the NET_IP_ALIGN==2 checking. Therefore use an
2910 * unconditional IP align value of '2'.
2914 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
2915 if (!tnum_is_aligned(reg_off, size)) {
2918 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2920 "misaligned packet access off %d+%s+%d+%d size %d\n",
2921 ip_align, tn_buf, reg->off, off, size);
2928 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
2929 const struct bpf_reg_state *reg,
2930 const char *pointer_desc,
2931 int off, int size, bool strict)
2933 struct tnum reg_off;
2935 /* Byte size accesses are always allowed. */
2936 if (!strict || size == 1)
2939 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
2940 if (!tnum_is_aligned(reg_off, size)) {
2943 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2944 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
2945 pointer_desc, tn_buf, reg->off, off, size);
2952 static int check_ptr_alignment(struct bpf_verifier_env *env,
2953 const struct bpf_reg_state *reg, int off,
2954 int size, bool strict_alignment_once)
2956 bool strict = env->strict_alignment || strict_alignment_once;
2957 const char *pointer_desc = "";
2959 switch (reg->type) {
2961 case PTR_TO_PACKET_META:
2962 /* Special case, because of NET_IP_ALIGN. Given metadata sits
2963 * right in front, treat it the very same way.
2965 return check_pkt_ptr_alignment(env, reg, off, size, strict);
2966 case PTR_TO_FLOW_KEYS:
2967 pointer_desc = "flow keys ";
2969 case PTR_TO_MAP_VALUE:
2970 pointer_desc = "value ";
2973 pointer_desc = "context ";
2976 pointer_desc = "stack ";
2977 /* The stack spill tracking logic in check_stack_write()
2978 * and check_stack_read() relies on stack accesses being
2984 pointer_desc = "sock ";
2986 case PTR_TO_SOCK_COMMON:
2987 pointer_desc = "sock_common ";
2989 case PTR_TO_TCP_SOCK:
2990 pointer_desc = "tcp_sock ";
2992 case PTR_TO_XDP_SOCK:
2993 pointer_desc = "xdp_sock ";
2998 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
3002 static int update_stack_depth(struct bpf_verifier_env *env,
3003 const struct bpf_func_state *func,
3006 u16 stack = env->subprog_info[func->subprogno].stack_depth;
3011 /* update known max for given subprogram */
3012 env->subprog_info[func->subprogno].stack_depth = -off;
3016 /* starting from main bpf function walk all instructions of the function
3017 * and recursively walk all callees that given function can call.
3018 * Ignore jump and exit insns.
3019 * Since recursion is prevented by check_cfg() this algorithm
3020 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
3022 static int check_max_stack_depth(struct bpf_verifier_env *env)
3024 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
3025 struct bpf_subprog_info *subprog = env->subprog_info;
3026 struct bpf_insn *insn = env->prog->insnsi;
3027 bool tail_call_reachable = false;
3028 int ret_insn[MAX_CALL_FRAMES];
3029 int ret_prog[MAX_CALL_FRAMES];
3033 /* protect against potential stack overflow that might happen when
3034 * bpf2bpf calls get combined with tailcalls. Limit the caller's stack
3035 * depth for such case down to 256 so that the worst case scenario
3036 * would result in 8k stack size (32 which is tailcall limit * 256 =
3039 * To get the idea what might happen, see an example:
3040 * func1 -> sub rsp, 128
3041 * subfunc1 -> sub rsp, 256
3042 * tailcall1 -> add rsp, 256
3043 * func2 -> sub rsp, 192 (total stack size = 128 + 192 = 320)
3044 * subfunc2 -> sub rsp, 64
3045 * subfunc22 -> sub rsp, 128
3046 * tailcall2 -> add rsp, 128
3047 * func3 -> sub rsp, 32 (total stack size 128 + 192 + 64 + 32 = 416)
3049 * tailcall will unwind the current stack frame but it will not get rid
3050 * of caller's stack as shown on the example above.
3052 if (idx && subprog[idx].has_tail_call && depth >= 256) {
3054 "tail_calls are not allowed when call stack of previous frames is %d bytes. Too large\n",
3058 /* round up to 32-bytes, since this is granularity
3059 * of interpreter stack size
3061 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
3062 if (depth > MAX_BPF_STACK) {
3063 verbose(env, "combined stack size of %d calls is %d. Too large\n",
3068 subprog_end = subprog[idx + 1].start;
3069 for (; i < subprog_end; i++) {
3070 if (insn[i].code != (BPF_JMP | BPF_CALL))
3072 if (insn[i].src_reg != BPF_PSEUDO_CALL)
3074 /* remember insn and function to return to */
3075 ret_insn[frame] = i + 1;
3076 ret_prog[frame] = idx;
3078 /* find the callee */
3079 i = i + insn[i].imm + 1;
3080 idx = find_subprog(env, i);
3082 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3087 if (subprog[idx].has_tail_call)
3088 tail_call_reachable = true;
3091 if (frame >= MAX_CALL_FRAMES) {
3092 verbose(env, "the call stack of %d frames is too deep !\n",
3098 /* if tail call got detected across bpf2bpf calls then mark each of the
3099 * currently present subprog frames as tail call reachable subprogs;
3100 * this info will be utilized by JIT so that we will be preserving the
3101 * tail call counter throughout bpf2bpf calls combined with tailcalls
3103 if (tail_call_reachable)
3104 for (j = 0; j < frame; j++)
3105 subprog[ret_prog[j]].tail_call_reachable = true;
3107 /* end of for() loop means the last insn of the 'subprog'
3108 * was reached. Doesn't matter whether it was JA or EXIT
3112 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
3114 i = ret_insn[frame];
3115 idx = ret_prog[frame];
3119 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
3120 static int get_callee_stack_depth(struct bpf_verifier_env *env,
3121 const struct bpf_insn *insn, int idx)
3123 int start = idx + insn->imm + 1, subprog;
3125 subprog = find_subprog(env, start);
3127 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3131 return env->subprog_info[subprog].stack_depth;
3135 int check_ctx_reg(struct bpf_verifier_env *env,
3136 const struct bpf_reg_state *reg, int regno)
3138 /* Access to ctx or passing it to a helper is only allowed in
3139 * its original, unmodified form.
3143 verbose(env, "dereference of modified ctx ptr R%d off=%d disallowed\n",
3148 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3151 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3152 verbose(env, "variable ctx access var_off=%s disallowed\n", tn_buf);
3159 static int __check_buffer_access(struct bpf_verifier_env *env,
3160 const char *buf_info,
3161 const struct bpf_reg_state *reg,
3162 int regno, int off, int size)
3166 "R%d invalid %s buffer access: off=%d, size=%d\n",
3167 regno, buf_info, off, size);
3170 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3173 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3175 "R%d invalid variable buffer offset: off=%d, var_off=%s\n",
3176 regno, off, tn_buf);
3183 static int check_tp_buffer_access(struct bpf_verifier_env *env,
3184 const struct bpf_reg_state *reg,
3185 int regno, int off, int size)
3189 err = __check_buffer_access(env, "tracepoint", reg, regno, off, size);
3193 if (off + size > env->prog->aux->max_tp_access)
3194 env->prog->aux->max_tp_access = off + size;
3199 static int check_buffer_access(struct bpf_verifier_env *env,
3200 const struct bpf_reg_state *reg,
3201 int regno, int off, int size,
3202 bool zero_size_allowed,
3203 const char *buf_info,
3208 err = __check_buffer_access(env, buf_info, reg, regno, off, size);
3212 if (off + size > *max_access)
3213 *max_access = off + size;
3218 /* BPF architecture zero extends alu32 ops into 64-bit registesr */
3219 static void zext_32_to_64(struct bpf_reg_state *reg)
3221 reg->var_off = tnum_subreg(reg->var_off);
3222 __reg_assign_32_into_64(reg);
3225 /* truncate register to smaller size (in bytes)
3226 * must be called with size < BPF_REG_SIZE
3228 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
3232 /* clear high bits in bit representation */
3233 reg->var_off = tnum_cast(reg->var_off, size);
3235 /* fix arithmetic bounds */
3236 mask = ((u64)1 << (size * 8)) - 1;
3237 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
3238 reg->umin_value &= mask;
3239 reg->umax_value &= mask;
3241 reg->umin_value = 0;
3242 reg->umax_value = mask;
3244 reg->smin_value = reg->umin_value;
3245 reg->smax_value = reg->umax_value;
3247 /* If size is smaller than 32bit register the 32bit register
3248 * values are also truncated so we push 64-bit bounds into
3249 * 32-bit bounds. Above were truncated < 32-bits already.
3253 __reg_combine_64_into_32(reg);
3256 static bool bpf_map_is_rdonly(const struct bpf_map *map)
3258 return (map->map_flags & BPF_F_RDONLY_PROG) && map->frozen;
3261 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val)
3267 err = map->ops->map_direct_value_addr(map, &addr, off);
3270 ptr = (void *)(long)addr + off;
3274 *val = (u64)*(u8 *)ptr;
3277 *val = (u64)*(u16 *)ptr;
3280 *val = (u64)*(u32 *)ptr;
3291 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
3292 struct bpf_reg_state *regs,
3293 int regno, int off, int size,
3294 enum bpf_access_type atype,
3297 struct bpf_reg_state *reg = regs + regno;
3298 const struct btf_type *t = btf_type_by_id(btf_vmlinux, reg->btf_id);
3299 const char *tname = btf_name_by_offset(btf_vmlinux, t->name_off);
3305 "R%d is ptr_%s invalid negative access: off=%d\n",
3309 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3312 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3314 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
3315 regno, tname, off, tn_buf);
3319 if (env->ops->btf_struct_access) {
3320 ret = env->ops->btf_struct_access(&env->log, t, off, size,
3323 if (atype != BPF_READ) {
3324 verbose(env, "only read is supported\n");
3328 ret = btf_struct_access(&env->log, t, off, size, atype,
3335 if (atype == BPF_READ && value_regno >= 0)
3336 mark_btf_ld_reg(env, regs, value_regno, ret, btf_id);
3341 static int check_ptr_to_map_access(struct bpf_verifier_env *env,
3342 struct bpf_reg_state *regs,
3343 int regno, int off, int size,
3344 enum bpf_access_type atype,
3347 struct bpf_reg_state *reg = regs + regno;
3348 struct bpf_map *map = reg->map_ptr;
3349 const struct btf_type *t;
3355 verbose(env, "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n");
3359 if (!map->ops->map_btf_id || !*map->ops->map_btf_id) {
3360 verbose(env, "map_ptr access not supported for map type %d\n",
3365 t = btf_type_by_id(btf_vmlinux, *map->ops->map_btf_id);
3366 tname = btf_name_by_offset(btf_vmlinux, t->name_off);
3368 if (!env->allow_ptr_to_map_access) {
3370 "%s access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
3376 verbose(env, "R%d is %s invalid negative access: off=%d\n",
3381 if (atype != BPF_READ) {
3382 verbose(env, "only read from %s is supported\n", tname);
3386 ret = btf_struct_access(&env->log, t, off, size, atype, &btf_id);
3390 if (value_regno >= 0)
3391 mark_btf_ld_reg(env, regs, value_regno, ret, btf_id);
3397 /* check whether memory at (regno + off) is accessible for t = (read | write)
3398 * if t==write, value_regno is a register which value is stored into memory
3399 * if t==read, value_regno is a register which will receive the value from memory
3400 * if t==write && value_regno==-1, some unknown value is stored into memory
3401 * if t==read && value_regno==-1, don't care what we read from memory
3403 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
3404 int off, int bpf_size, enum bpf_access_type t,
3405 int value_regno, bool strict_alignment_once)
3407 struct bpf_reg_state *regs = cur_regs(env);
3408 struct bpf_reg_state *reg = regs + regno;
3409 struct bpf_func_state *state;
3412 size = bpf_size_to_bytes(bpf_size);
3416 /* alignment checks will add in reg->off themselves */
3417 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
3421 /* for access checks, reg->off is just part of off */
3424 if (reg->type == PTR_TO_MAP_VALUE) {
3425 if (t == BPF_WRITE && value_regno >= 0 &&
3426 is_pointer_value(env, value_regno)) {
3427 verbose(env, "R%d leaks addr into map\n", value_regno);
3430 err = check_map_access_type(env, regno, off, size, t);
3433 err = check_map_access(env, regno, off, size, false);
3434 if (!err && t == BPF_READ && value_regno >= 0) {
3435 struct bpf_map *map = reg->map_ptr;
3437 /* if map is read-only, track its contents as scalars */
3438 if (tnum_is_const(reg->var_off) &&
3439 bpf_map_is_rdonly(map) &&
3440 map->ops->map_direct_value_addr) {
3441 int map_off = off + reg->var_off.value;
3444 err = bpf_map_direct_read(map, map_off, size,
3449 regs[value_regno].type = SCALAR_VALUE;
3450 __mark_reg_known(®s[value_regno], val);
3452 mark_reg_unknown(env, regs, value_regno);
3455 } else if (reg->type == PTR_TO_MEM) {
3456 if (t == BPF_WRITE && value_regno >= 0 &&
3457 is_pointer_value(env, value_regno)) {
3458 verbose(env, "R%d leaks addr into mem\n", value_regno);
3461 err = check_mem_region_access(env, regno, off, size,
3462 reg->mem_size, false);
3463 if (!err && t == BPF_READ && value_regno >= 0)
3464 mark_reg_unknown(env, regs, value_regno);
3465 } else if (reg->type == PTR_TO_CTX) {
3466 enum bpf_reg_type reg_type = SCALAR_VALUE;
3469 if (t == BPF_WRITE && value_regno >= 0 &&
3470 is_pointer_value(env, value_regno)) {
3471 verbose(env, "R%d leaks addr into ctx\n", value_regno);
3475 err = check_ctx_reg(env, reg, regno);
3479 err = check_ctx_access(env, insn_idx, off, size, t, ®_type, &btf_id);
3481 verbose_linfo(env, insn_idx, "; ");
3482 if (!err && t == BPF_READ && value_regno >= 0) {
3483 /* ctx access returns either a scalar, or a
3484 * PTR_TO_PACKET[_META,_END]. In the latter
3485 * case, we know the offset is zero.
3487 if (reg_type == SCALAR_VALUE) {
3488 mark_reg_unknown(env, regs, value_regno);
3490 mark_reg_known_zero(env, regs,
3492 if (reg_type_may_be_null(reg_type))
3493 regs[value_regno].id = ++env->id_gen;
3494 /* A load of ctx field could have different
3495 * actual load size with the one encoded in the
3496 * insn. When the dst is PTR, it is for sure not
3499 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
3500 if (reg_type == PTR_TO_BTF_ID ||
3501 reg_type == PTR_TO_BTF_ID_OR_NULL)
3502 regs[value_regno].btf_id = btf_id;
3504 regs[value_regno].type = reg_type;
3507 } else if (reg->type == PTR_TO_STACK) {
3508 off += reg->var_off.value;
3509 err = check_stack_access(env, reg, off, size);
3513 state = func(env, reg);
3514 err = update_stack_depth(env, state, off);
3519 err = check_stack_write(env, state, off, size,
3520 value_regno, insn_idx);
3522 err = check_stack_read(env, state, off, size,
3524 } else if (reg_is_pkt_pointer(reg)) {
3525 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
3526 verbose(env, "cannot write into packet\n");
3529 if (t == BPF_WRITE && value_regno >= 0 &&
3530 is_pointer_value(env, value_regno)) {
3531 verbose(env, "R%d leaks addr into packet\n",
3535 err = check_packet_access(env, regno, off, size, false);
3536 if (!err && t == BPF_READ && value_regno >= 0)
3537 mark_reg_unknown(env, regs, value_regno);
3538 } else if (reg->type == PTR_TO_FLOW_KEYS) {
3539 if (t == BPF_WRITE && value_regno >= 0 &&
3540 is_pointer_value(env, value_regno)) {
3541 verbose(env, "R%d leaks addr into flow keys\n",
3546 err = check_flow_keys_access(env, off, size);
3547 if (!err && t == BPF_READ && value_regno >= 0)
3548 mark_reg_unknown(env, regs, value_regno);
3549 } else if (type_is_sk_pointer(reg->type)) {
3550 if (t == BPF_WRITE) {
3551 verbose(env, "R%d cannot write into %s\n",
3552 regno, reg_type_str[reg->type]);
3555 err = check_sock_access(env, insn_idx, regno, off, size, t);
3556 if (!err && value_regno >= 0)
3557 mark_reg_unknown(env, regs, value_regno);
3558 } else if (reg->type == PTR_TO_TP_BUFFER) {
3559 err = check_tp_buffer_access(env, reg, regno, off, size);
3560 if (!err && t == BPF_READ && value_regno >= 0)
3561 mark_reg_unknown(env, regs, value_regno);
3562 } else if (reg->type == PTR_TO_BTF_ID) {
3563 err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
3565 } else if (reg->type == CONST_PTR_TO_MAP) {
3566 err = check_ptr_to_map_access(env, regs, regno, off, size, t,
3568 } else if (reg->type == PTR_TO_RDONLY_BUF) {
3569 if (t == BPF_WRITE) {
3570 verbose(env, "R%d cannot write into %s\n",
3571 regno, reg_type_str[reg->type]);
3574 err = check_buffer_access(env, reg, regno, off, size, false,
3576 &env->prog->aux->max_rdonly_access);
3577 if (!err && value_regno >= 0)
3578 mark_reg_unknown(env, regs, value_regno);
3579 } else if (reg->type == PTR_TO_RDWR_BUF) {
3580 err = check_buffer_access(env, reg, regno, off, size, false,
3582 &env->prog->aux->max_rdwr_access);
3583 if (!err && t == BPF_READ && value_regno >= 0)
3584 mark_reg_unknown(env, regs, value_regno);
3586 verbose(env, "R%d invalid mem access '%s'\n", regno,
3587 reg_type_str[reg->type]);
3591 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
3592 regs[value_regno].type == SCALAR_VALUE) {
3593 /* b/h/w load zero-extends, mark upper bits as known 0 */
3594 coerce_reg_to_size(®s[value_regno], size);
3599 static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
3603 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
3605 verbose(env, "BPF_XADD uses reserved fields\n");
3609 /* check src1 operand */
3610 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3614 /* check src2 operand */
3615 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3619 if (is_pointer_value(env, insn->src_reg)) {
3620 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
3624 if (is_ctx_reg(env, insn->dst_reg) ||
3625 is_pkt_reg(env, insn->dst_reg) ||
3626 is_flow_key_reg(env, insn->dst_reg) ||
3627 is_sk_reg(env, insn->dst_reg)) {
3628 verbose(env, "BPF_XADD stores into R%d %s is not allowed\n",
3630 reg_type_str[reg_state(env, insn->dst_reg)->type]);
3634 /* check whether atomic_add can read the memory */
3635 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3636 BPF_SIZE(insn->code), BPF_READ, -1, true);
3640 /* check whether atomic_add can write into the same memory */
3641 return check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3642 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
3645 static int __check_stack_boundary(struct bpf_verifier_env *env, u32 regno,
3646 int off, int access_size,
3647 bool zero_size_allowed)
3649 struct bpf_reg_state *reg = reg_state(env, regno);
3651 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
3652 access_size < 0 || (access_size == 0 && !zero_size_allowed)) {
3653 if (tnum_is_const(reg->var_off)) {
3654 verbose(env, "invalid stack type R%d off=%d access_size=%d\n",
3655 regno, off, access_size);
3659 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3660 verbose(env, "invalid stack type R%d var_off=%s access_size=%d\n",
3661 regno, tn_buf, access_size);
3668 /* when register 'regno' is passed into function that will read 'access_size'
3669 * bytes from that pointer, make sure that it's within stack boundary
3670 * and all elements of stack are initialized.
3671 * Unlike most pointer bounds-checking functions, this one doesn't take an
3672 * 'off' argument, so it has to add in reg->off itself.
3674 static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
3675 int access_size, bool zero_size_allowed,
3676 struct bpf_call_arg_meta *meta)
3678 struct bpf_reg_state *reg = reg_state(env, regno);
3679 struct bpf_func_state *state = func(env, reg);
3680 int err, min_off, max_off, i, j, slot, spi;
3682 if (tnum_is_const(reg->var_off)) {
3683 min_off = max_off = reg->var_off.value + reg->off;
3684 err = __check_stack_boundary(env, regno, min_off, access_size,
3689 /* Variable offset is prohibited for unprivileged mode for
3690 * simplicity since it requires corresponding support in
3691 * Spectre masking for stack ALU.
3692 * See also retrieve_ptr_limit().
3694 if (!env->bypass_spec_v1) {
3697 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3698 verbose(env, "R%d indirect variable offset stack access prohibited for !root, var_off=%s\n",
3702 /* Only initialized buffer on stack is allowed to be accessed
3703 * with variable offset. With uninitialized buffer it's hard to
3704 * guarantee that whole memory is marked as initialized on
3705 * helper return since specific bounds are unknown what may
3706 * cause uninitialized stack leaking.
3708 if (meta && meta->raw_mode)
3711 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
3712 reg->smax_value <= -BPF_MAX_VAR_OFF) {
3713 verbose(env, "R%d unbounded indirect variable offset stack access\n",
3717 min_off = reg->smin_value + reg->off;
3718 max_off = reg->smax_value + reg->off;
3719 err = __check_stack_boundary(env, regno, min_off, access_size,
3722 verbose(env, "R%d min value is outside of stack bound\n",
3726 err = __check_stack_boundary(env, regno, max_off, access_size,
3729 verbose(env, "R%d max value is outside of stack bound\n",
3735 if (meta && meta->raw_mode) {
3736 meta->access_size = access_size;
3737 meta->regno = regno;
3741 for (i = min_off; i < max_off + access_size; i++) {
3745 spi = slot / BPF_REG_SIZE;
3746 if (state->allocated_stack <= slot)
3748 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
3749 if (*stype == STACK_MISC)
3751 if (*stype == STACK_ZERO) {
3752 /* helper can write anything into the stack */
3753 *stype = STACK_MISC;
3757 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
3758 state->stack[spi].spilled_ptr.type == PTR_TO_BTF_ID)
3761 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
3762 state->stack[spi].spilled_ptr.type == SCALAR_VALUE) {
3763 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
3764 for (j = 0; j < BPF_REG_SIZE; j++)
3765 state->stack[spi].slot_type[j] = STACK_MISC;
3770 if (tnum_is_const(reg->var_off)) {
3771 verbose(env, "invalid indirect read from stack off %d+%d size %d\n",
3772 min_off, i - min_off, access_size);
3776 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3777 verbose(env, "invalid indirect read from stack var_off %s+%d size %d\n",
3778 tn_buf, i - min_off, access_size);
3782 /* reading any byte out of 8-byte 'spill_slot' will cause
3783 * the whole slot to be marked as 'read'
3785 mark_reg_read(env, &state->stack[spi].spilled_ptr,
3786 state->stack[spi].spilled_ptr.parent,
3789 return update_stack_depth(env, state, min_off);
3792 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
3793 int access_size, bool zero_size_allowed,
3794 struct bpf_call_arg_meta *meta)
3796 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
3798 switch (reg->type) {
3800 case PTR_TO_PACKET_META:
3801 return check_packet_access(env, regno, reg->off, access_size,
3803 case PTR_TO_MAP_VALUE:
3804 if (check_map_access_type(env, regno, reg->off, access_size,
3805 meta && meta->raw_mode ? BPF_WRITE :
3808 return check_map_access(env, regno, reg->off, access_size,
3811 return check_mem_region_access(env, regno, reg->off,
3812 access_size, reg->mem_size,
3814 case PTR_TO_RDONLY_BUF:
3815 if (meta && meta->raw_mode)
3817 return check_buffer_access(env, reg, regno, reg->off,
3818 access_size, zero_size_allowed,
3820 &env->prog->aux->max_rdonly_access);
3821 case PTR_TO_RDWR_BUF:
3822 return check_buffer_access(env, reg, regno, reg->off,
3823 access_size, zero_size_allowed,
3825 &env->prog->aux->max_rdwr_access);
3827 return check_stack_boundary(env, regno, access_size,
3828 zero_size_allowed, meta);
3829 default: /* scalar_value or invalid ptr */
3830 /* Allow zero-byte read from NULL, regardless of pointer type */
3831 if (zero_size_allowed && access_size == 0 &&
3832 register_is_null(reg))
3835 verbose(env, "R%d type=%s expected=%s\n", regno,
3836 reg_type_str[reg->type],
3837 reg_type_str[PTR_TO_STACK]);
3842 /* Implementation details:
3843 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL
3844 * Two bpf_map_lookups (even with the same key) will have different reg->id.
3845 * For traditional PTR_TO_MAP_VALUE the verifier clears reg->id after
3846 * value_or_null->value transition, since the verifier only cares about
3847 * the range of access to valid map value pointer and doesn't care about actual
3848 * address of the map element.
3849 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
3850 * reg->id > 0 after value_or_null->value transition. By doing so
3851 * two bpf_map_lookups will be considered two different pointers that
3852 * point to different bpf_spin_locks.
3853 * The verifier allows taking only one bpf_spin_lock at a time to avoid
3855 * Since only one bpf_spin_lock is allowed the checks are simpler than
3856 * reg_is_refcounted() logic. The verifier needs to remember only
3857 * one spin_lock instead of array of acquired_refs.
3858 * cur_state->active_spin_lock remembers which map value element got locked
3859 * and clears it after bpf_spin_unlock.
3861 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
3864 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
3865 struct bpf_verifier_state *cur = env->cur_state;
3866 bool is_const = tnum_is_const(reg->var_off);
3867 struct bpf_map *map = reg->map_ptr;
3868 u64 val = reg->var_off.value;
3872 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
3878 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
3882 if (!map_value_has_spin_lock(map)) {
3883 if (map->spin_lock_off == -E2BIG)
3885 "map '%s' has more than one 'struct bpf_spin_lock'\n",
3887 else if (map->spin_lock_off == -ENOENT)
3889 "map '%s' doesn't have 'struct bpf_spin_lock'\n",
3893 "map '%s' is not a struct type or bpf_spin_lock is mangled\n",
3897 if (map->spin_lock_off != val + reg->off) {
3898 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock'\n",
3903 if (cur->active_spin_lock) {
3905 "Locking two bpf_spin_locks are not allowed\n");
3908 cur->active_spin_lock = reg->id;
3910 if (!cur->active_spin_lock) {
3911 verbose(env, "bpf_spin_unlock without taking a lock\n");
3914 if (cur->active_spin_lock != reg->id) {
3915 verbose(env, "bpf_spin_unlock of different lock\n");
3918 cur->active_spin_lock = 0;
3923 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
3925 return type == ARG_PTR_TO_MEM ||
3926 type == ARG_PTR_TO_MEM_OR_NULL ||
3927 type == ARG_PTR_TO_UNINIT_MEM;
3930 static bool arg_type_is_mem_size(enum bpf_arg_type type)
3932 return type == ARG_CONST_SIZE ||
3933 type == ARG_CONST_SIZE_OR_ZERO;
3936 static bool arg_type_is_alloc_size(enum bpf_arg_type type)
3938 return type == ARG_CONST_ALLOC_SIZE_OR_ZERO;
3941 static bool arg_type_is_int_ptr(enum bpf_arg_type type)
3943 return type == ARG_PTR_TO_INT ||
3944 type == ARG_PTR_TO_LONG;
3947 static int int_ptr_type_to_size(enum bpf_arg_type type)
3949 if (type == ARG_PTR_TO_INT)
3951 else if (type == ARG_PTR_TO_LONG)
3957 static int resolve_map_arg_type(struct bpf_verifier_env *env,
3958 const struct bpf_call_arg_meta *meta,
3959 enum bpf_arg_type *arg_type)
3961 if (!meta->map_ptr) {
3962 /* kernel subsystem misconfigured verifier */
3963 verbose(env, "invalid map_ptr to access map->type\n");
3967 switch (meta->map_ptr->map_type) {
3968 case BPF_MAP_TYPE_SOCKMAP:
3969 case BPF_MAP_TYPE_SOCKHASH:
3970 if (*arg_type == ARG_PTR_TO_MAP_VALUE) {
3971 *arg_type = ARG_PTR_TO_BTF_ID_SOCK_COMMON;
3973 verbose(env, "invalid arg_type for sockmap/sockhash\n");
3984 struct bpf_reg_types {
3985 const enum bpf_reg_type types[10];
3989 static const struct bpf_reg_types map_key_value_types = {
3998 static const struct bpf_reg_types sock_types = {
4008 static const struct bpf_reg_types btf_id_sock_common_types = {
4016 .btf_id = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
4020 static const struct bpf_reg_types mem_types = {
4032 static const struct bpf_reg_types int_ptr_types = {
4041 static const struct bpf_reg_types fullsock_types = { .types = { PTR_TO_SOCKET } };
4042 static const struct bpf_reg_types scalar_types = { .types = { SCALAR_VALUE } };
4043 static const struct bpf_reg_types context_types = { .types = { PTR_TO_CTX } };
4044 static const struct bpf_reg_types alloc_mem_types = { .types = { PTR_TO_MEM } };
4045 static const struct bpf_reg_types const_map_ptr_types = { .types = { CONST_PTR_TO_MAP } };
4046 static const struct bpf_reg_types btf_ptr_types = { .types = { PTR_TO_BTF_ID } };
4047 static const struct bpf_reg_types spin_lock_types = { .types = { PTR_TO_MAP_VALUE } };
4048 static const struct bpf_reg_types percpu_btf_ptr_types = { .types = { PTR_TO_PERCPU_BTF_ID } };
4050 static const struct bpf_reg_types *compatible_reg_types[__BPF_ARG_TYPE_MAX] = {
4051 [ARG_PTR_TO_MAP_KEY] = &map_key_value_types,
4052 [ARG_PTR_TO_MAP_VALUE] = &map_key_value_types,
4053 [ARG_PTR_TO_UNINIT_MAP_VALUE] = &map_key_value_types,
4054 [ARG_PTR_TO_MAP_VALUE_OR_NULL] = &map_key_value_types,
4055 [ARG_CONST_SIZE] = &scalar_types,
4056 [ARG_CONST_SIZE_OR_ZERO] = &scalar_types,
4057 [ARG_CONST_ALLOC_SIZE_OR_ZERO] = &scalar_types,
4058 [ARG_CONST_MAP_PTR] = &const_map_ptr_types,
4059 [ARG_PTR_TO_CTX] = &context_types,
4060 [ARG_PTR_TO_CTX_OR_NULL] = &context_types,
4061 [ARG_PTR_TO_SOCK_COMMON] = &sock_types,
4063 [ARG_PTR_TO_BTF_ID_SOCK_COMMON] = &btf_id_sock_common_types,
4065 [ARG_PTR_TO_SOCKET] = &fullsock_types,
4066 [ARG_PTR_TO_SOCKET_OR_NULL] = &fullsock_types,
4067 [ARG_PTR_TO_BTF_ID] = &btf_ptr_types,
4068 [ARG_PTR_TO_SPIN_LOCK] = &spin_lock_types,
4069 [ARG_PTR_TO_MEM] = &mem_types,
4070 [ARG_PTR_TO_MEM_OR_NULL] = &mem_types,
4071 [ARG_PTR_TO_UNINIT_MEM] = &mem_types,
4072 [ARG_PTR_TO_ALLOC_MEM] = &alloc_mem_types,
4073 [ARG_PTR_TO_ALLOC_MEM_OR_NULL] = &alloc_mem_types,
4074 [ARG_PTR_TO_INT] = &int_ptr_types,
4075 [ARG_PTR_TO_LONG] = &int_ptr_types,
4076 [ARG_PTR_TO_PERCPU_BTF_ID] = &percpu_btf_ptr_types,
4079 static int check_reg_type(struct bpf_verifier_env *env, u32 regno,
4080 enum bpf_arg_type arg_type,
4081 const u32 *arg_btf_id)
4083 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4084 enum bpf_reg_type expected, type = reg->type;
4085 const struct bpf_reg_types *compatible;
4088 compatible = compatible_reg_types[arg_type];
4090 verbose(env, "verifier internal error: unsupported arg type %d\n", arg_type);
4094 for (i = 0; i < ARRAY_SIZE(compatible->types); i++) {
4095 expected = compatible->types[i];
4096 if (expected == NOT_INIT)
4099 if (type == expected)
4103 verbose(env, "R%d type=%s expected=", regno, reg_type_str[type]);
4104 for (j = 0; j + 1 < i; j++)
4105 verbose(env, "%s, ", reg_type_str[compatible->types[j]]);
4106 verbose(env, "%s\n", reg_type_str[compatible->types[j]]);
4110 if (type == PTR_TO_BTF_ID) {
4112 if (!compatible->btf_id) {
4113 verbose(env, "verifier internal error: missing arg compatible BTF ID\n");
4116 arg_btf_id = compatible->btf_id;
4119 if (!btf_struct_ids_match(&env->log, reg->off, reg->btf_id,
4121 verbose(env, "R%d is of type %s but %s is expected\n",
4122 regno, kernel_type_name(reg->btf_id),
4123 kernel_type_name(*arg_btf_id));
4127 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
4128 verbose(env, "R%d is a pointer to in-kernel struct with non-zero offset\n",
4137 static int check_func_arg(struct bpf_verifier_env *env, u32 arg,
4138 struct bpf_call_arg_meta *meta,
4139 const struct bpf_func_proto *fn)
4141 u32 regno = BPF_REG_1 + arg;
4142 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4143 enum bpf_arg_type arg_type = fn->arg_type[arg];
4144 enum bpf_reg_type type = reg->type;
4147 if (arg_type == ARG_DONTCARE)
4150 err = check_reg_arg(env, regno, SRC_OP);
4154 if (arg_type == ARG_ANYTHING) {
4155 if (is_pointer_value(env, regno)) {
4156 verbose(env, "R%d leaks addr into helper function\n",
4163 if (type_is_pkt_pointer(type) &&
4164 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
4165 verbose(env, "helper access to the packet is not allowed\n");
4169 if (arg_type == ARG_PTR_TO_MAP_VALUE ||
4170 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE ||
4171 arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL) {
4172 err = resolve_map_arg_type(env, meta, &arg_type);
4177 if (register_is_null(reg) && arg_type_may_be_null(arg_type))
4178 /* A NULL register has a SCALAR_VALUE type, so skip
4181 goto skip_type_check;
4183 err = check_reg_type(env, regno, arg_type, fn->arg_btf_id[arg]);
4187 if (type == PTR_TO_CTX) {
4188 err = check_ctx_reg(env, reg, regno);
4194 if (reg->ref_obj_id) {
4195 if (meta->ref_obj_id) {
4196 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
4197 regno, reg->ref_obj_id,
4201 meta->ref_obj_id = reg->ref_obj_id;
4204 if (arg_type == ARG_CONST_MAP_PTR) {
4205 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
4206 meta->map_ptr = reg->map_ptr;
4207 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
4208 /* bpf_map_xxx(..., map_ptr, ..., key) call:
4209 * check that [key, key + map->key_size) are within
4210 * stack limits and initialized
4212 if (!meta->map_ptr) {
4213 /* in function declaration map_ptr must come before
4214 * map_key, so that it's verified and known before
4215 * we have to check map_key here. Otherwise it means
4216 * that kernel subsystem misconfigured verifier
4218 verbose(env, "invalid map_ptr to access map->key\n");
4221 err = check_helper_mem_access(env, regno,
4222 meta->map_ptr->key_size, false,
4224 } else if (arg_type == ARG_PTR_TO_MAP_VALUE ||
4225 (arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL &&
4226 !register_is_null(reg)) ||
4227 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
4228 /* bpf_map_xxx(..., map_ptr, ..., value) call:
4229 * check [value, value + map->value_size) validity
4231 if (!meta->map_ptr) {
4232 /* kernel subsystem misconfigured verifier */
4233 verbose(env, "invalid map_ptr to access map->value\n");
4236 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE);
4237 err = check_helper_mem_access(env, regno,
4238 meta->map_ptr->value_size, false,
4240 } else if (arg_type == ARG_PTR_TO_PERCPU_BTF_ID) {
4242 verbose(env, "Helper has invalid btf_id in R%d\n", regno);
4245 meta->ret_btf_id = reg->btf_id;
4246 } else if (arg_type == ARG_PTR_TO_SPIN_LOCK) {
4247 if (meta->func_id == BPF_FUNC_spin_lock) {
4248 if (process_spin_lock(env, regno, true))
4250 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
4251 if (process_spin_lock(env, regno, false))
4254 verbose(env, "verifier internal error\n");
4257 } else if (arg_type_is_mem_ptr(arg_type)) {
4258 /* The access to this pointer is only checked when we hit the
4259 * next is_mem_size argument below.
4261 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MEM);
4262 } else if (arg_type_is_mem_size(arg_type)) {
4263 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
4265 /* This is used to refine r0 return value bounds for helpers
4266 * that enforce this value as an upper bound on return values.
4267 * See do_refine_retval_range() for helpers that can refine
4268 * the return value. C type of helper is u32 so we pull register
4269 * bound from umax_value however, if negative verifier errors
4270 * out. Only upper bounds can be learned because retval is an
4271 * int type and negative retvals are allowed.
4273 meta->msize_max_value = reg->umax_value;
4275 /* The register is SCALAR_VALUE; the access check
4276 * happens using its boundaries.
4278 if (!tnum_is_const(reg->var_off))
4279 /* For unprivileged variable accesses, disable raw
4280 * mode so that the program is required to
4281 * initialize all the memory that the helper could
4282 * just partially fill up.
4286 if (reg->smin_value < 0) {
4287 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
4292 if (reg->umin_value == 0) {
4293 err = check_helper_mem_access(env, regno - 1, 0,
4300 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
4301 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
4305 err = check_helper_mem_access(env, regno - 1,
4307 zero_size_allowed, meta);
4309 err = mark_chain_precision(env, regno);
4310 } else if (arg_type_is_alloc_size(arg_type)) {
4311 if (!tnum_is_const(reg->var_off)) {
4312 verbose(env, "R%d unbounded size, use 'var &= const' or 'if (var < const)'\n",
4316 meta->mem_size = reg->var_off.value;
4317 } else if (arg_type_is_int_ptr(arg_type)) {
4318 int size = int_ptr_type_to_size(arg_type);
4320 err = check_helper_mem_access(env, regno, size, false, meta);
4323 err = check_ptr_alignment(env, reg, 0, size, true);
4329 static bool may_update_sockmap(struct bpf_verifier_env *env, int func_id)
4331 enum bpf_attach_type eatype = env->prog->expected_attach_type;
4332 enum bpf_prog_type type = resolve_prog_type(env->prog);
4334 if (func_id != BPF_FUNC_map_update_elem)
4337 /* It's not possible to get access to a locked struct sock in these
4338 * contexts, so updating is safe.
4341 case BPF_PROG_TYPE_TRACING:
4342 if (eatype == BPF_TRACE_ITER)
4345 case BPF_PROG_TYPE_SOCKET_FILTER:
4346 case BPF_PROG_TYPE_SCHED_CLS:
4347 case BPF_PROG_TYPE_SCHED_ACT:
4348 case BPF_PROG_TYPE_XDP:
4349 case BPF_PROG_TYPE_SK_REUSEPORT:
4350 case BPF_PROG_TYPE_FLOW_DISSECTOR:
4351 case BPF_PROG_TYPE_SK_LOOKUP:
4357 verbose(env, "cannot update sockmap in this context\n");
4361 static bool allow_tail_call_in_subprogs(struct bpf_verifier_env *env)
4363 return env->prog->jit_requested && IS_ENABLED(CONFIG_X86_64);
4366 static int check_map_func_compatibility(struct bpf_verifier_env *env,
4367 struct bpf_map *map, int func_id)
4372 /* We need a two way check, first is from map perspective ... */
4373 switch (map->map_type) {
4374 case BPF_MAP_TYPE_PROG_ARRAY:
4375 if (func_id != BPF_FUNC_tail_call)
4378 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
4379 if (func_id != BPF_FUNC_perf_event_read &&
4380 func_id != BPF_FUNC_perf_event_output &&
4381 func_id != BPF_FUNC_skb_output &&
4382 func_id != BPF_FUNC_perf_event_read_value &&
4383 func_id != BPF_FUNC_xdp_output)
4386 case BPF_MAP_TYPE_RINGBUF:
4387 if (func_id != BPF_FUNC_ringbuf_output &&
4388 func_id != BPF_FUNC_ringbuf_reserve &&
4389 func_id != BPF_FUNC_ringbuf_submit &&
4390 func_id != BPF_FUNC_ringbuf_discard &&
4391 func_id != BPF_FUNC_ringbuf_query)
4394 case BPF_MAP_TYPE_STACK_TRACE:
4395 if (func_id != BPF_FUNC_get_stackid)
4398 case BPF_MAP_TYPE_CGROUP_ARRAY:
4399 if (func_id != BPF_FUNC_skb_under_cgroup &&
4400 func_id != BPF_FUNC_current_task_under_cgroup)
4403 case BPF_MAP_TYPE_CGROUP_STORAGE:
4404 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
4405 if (func_id != BPF_FUNC_get_local_storage)
4408 case BPF_MAP_TYPE_DEVMAP:
4409 case BPF_MAP_TYPE_DEVMAP_HASH:
4410 if (func_id != BPF_FUNC_redirect_map &&
4411 func_id != BPF_FUNC_map_lookup_elem)
4414 /* Restrict bpf side of cpumap and xskmap, open when use-cases
4417 case BPF_MAP_TYPE_CPUMAP:
4418 if (func_id != BPF_FUNC_redirect_map)
4421 case BPF_MAP_TYPE_XSKMAP:
4422 if (func_id != BPF_FUNC_redirect_map &&
4423 func_id != BPF_FUNC_map_lookup_elem)
4426 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
4427 case BPF_MAP_TYPE_HASH_OF_MAPS:
4428 if (func_id != BPF_FUNC_map_lookup_elem)
4431 case BPF_MAP_TYPE_SOCKMAP:
4432 if (func_id != BPF_FUNC_sk_redirect_map &&
4433 func_id != BPF_FUNC_sock_map_update &&
4434 func_id != BPF_FUNC_map_delete_elem &&
4435 func_id != BPF_FUNC_msg_redirect_map &&
4436 func_id != BPF_FUNC_sk_select_reuseport &&
4437 func_id != BPF_FUNC_map_lookup_elem &&
4438 !may_update_sockmap(env, func_id))
4441 case BPF_MAP_TYPE_SOCKHASH:
4442 if (func_id != BPF_FUNC_sk_redirect_hash &&
4443 func_id != BPF_FUNC_sock_hash_update &&
4444 func_id != BPF_FUNC_map_delete_elem &&
4445 func_id != BPF_FUNC_msg_redirect_hash &&
4446 func_id != BPF_FUNC_sk_select_reuseport &&
4447 func_id != BPF_FUNC_map_lookup_elem &&
4448 !may_update_sockmap(env, func_id))
4451 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
4452 if (func_id != BPF_FUNC_sk_select_reuseport)
4455 case BPF_MAP_TYPE_QUEUE:
4456 case BPF_MAP_TYPE_STACK:
4457 if (func_id != BPF_FUNC_map_peek_elem &&
4458 func_id != BPF_FUNC_map_pop_elem &&
4459 func_id != BPF_FUNC_map_push_elem)
4462 case BPF_MAP_TYPE_SK_STORAGE:
4463 if (func_id != BPF_FUNC_sk_storage_get &&
4464 func_id != BPF_FUNC_sk_storage_delete)
4467 case BPF_MAP_TYPE_INODE_STORAGE:
4468 if (func_id != BPF_FUNC_inode_storage_get &&
4469 func_id != BPF_FUNC_inode_storage_delete)
4476 /* ... and second from the function itself. */
4478 case BPF_FUNC_tail_call:
4479 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
4481 if (env->subprog_cnt > 1 && !allow_tail_call_in_subprogs(env)) {
4482 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
4486 case BPF_FUNC_perf_event_read:
4487 case BPF_FUNC_perf_event_output:
4488 case BPF_FUNC_perf_event_read_value:
4489 case BPF_FUNC_skb_output:
4490 case BPF_FUNC_xdp_output:
4491 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
4494 case BPF_FUNC_get_stackid:
4495 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
4498 case BPF_FUNC_current_task_under_cgroup:
4499 case BPF_FUNC_skb_under_cgroup:
4500 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
4503 case BPF_FUNC_redirect_map:
4504 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
4505 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
4506 map->map_type != BPF_MAP_TYPE_CPUMAP &&
4507 map->map_type != BPF_MAP_TYPE_XSKMAP)
4510 case BPF_FUNC_sk_redirect_map:
4511 case BPF_FUNC_msg_redirect_map:
4512 case BPF_FUNC_sock_map_update:
4513 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
4516 case BPF_FUNC_sk_redirect_hash:
4517 case BPF_FUNC_msg_redirect_hash:
4518 case BPF_FUNC_sock_hash_update:
4519 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
4522 case BPF_FUNC_get_local_storage:
4523 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
4524 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
4527 case BPF_FUNC_sk_select_reuseport:
4528 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
4529 map->map_type != BPF_MAP_TYPE_SOCKMAP &&
4530 map->map_type != BPF_MAP_TYPE_SOCKHASH)
4533 case BPF_FUNC_map_peek_elem:
4534 case BPF_FUNC_map_pop_elem:
4535 case BPF_FUNC_map_push_elem:
4536 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
4537 map->map_type != BPF_MAP_TYPE_STACK)
4540 case BPF_FUNC_sk_storage_get:
4541 case BPF_FUNC_sk_storage_delete:
4542 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
4545 case BPF_FUNC_inode_storage_get:
4546 case BPF_FUNC_inode_storage_delete:
4547 if (map->map_type != BPF_MAP_TYPE_INODE_STORAGE)
4556 verbose(env, "cannot pass map_type %d into func %s#%d\n",
4557 map->map_type, func_id_name(func_id), func_id);
4561 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
4565 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
4567 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
4569 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
4571 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
4573 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
4576 /* We only support one arg being in raw mode at the moment,
4577 * which is sufficient for the helper functions we have
4583 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
4584 enum bpf_arg_type arg_next)
4586 return (arg_type_is_mem_ptr(arg_curr) &&
4587 !arg_type_is_mem_size(arg_next)) ||
4588 (!arg_type_is_mem_ptr(arg_curr) &&
4589 arg_type_is_mem_size(arg_next));
4592 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
4594 /* bpf_xxx(..., buf, len) call will access 'len'
4595 * bytes from memory 'buf'. Both arg types need
4596 * to be paired, so make sure there's no buggy
4597 * helper function specification.
4599 if (arg_type_is_mem_size(fn->arg1_type) ||
4600 arg_type_is_mem_ptr(fn->arg5_type) ||
4601 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
4602 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
4603 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
4604 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
4610 static bool check_refcount_ok(const struct bpf_func_proto *fn, int func_id)
4614 if (arg_type_may_be_refcounted(fn->arg1_type))
4616 if (arg_type_may_be_refcounted(fn->arg2_type))
4618 if (arg_type_may_be_refcounted(fn->arg3_type))
4620 if (arg_type_may_be_refcounted(fn->arg4_type))
4622 if (arg_type_may_be_refcounted(fn->arg5_type))
4625 /* A reference acquiring function cannot acquire
4626 * another refcounted ptr.
4628 if (may_be_acquire_function(func_id) && count)
4631 /* We only support one arg being unreferenced at the moment,
4632 * which is sufficient for the helper functions we have right now.
4637 static bool check_btf_id_ok(const struct bpf_func_proto *fn)
4641 for (i = 0; i < ARRAY_SIZE(fn->arg_type); i++) {
4642 if (fn->arg_type[i] == ARG_PTR_TO_BTF_ID && !fn->arg_btf_id[i])
4645 if (fn->arg_type[i] != ARG_PTR_TO_BTF_ID && fn->arg_btf_id[i])
4652 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
4654 return check_raw_mode_ok(fn) &&
4655 check_arg_pair_ok(fn) &&
4656 check_btf_id_ok(fn) &&
4657 check_refcount_ok(fn, func_id) ? 0 : -EINVAL;
4660 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
4661 * are now invalid, so turn them into unknown SCALAR_VALUE.
4663 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
4664 struct bpf_func_state *state)
4666 struct bpf_reg_state *regs = state->regs, *reg;
4669 for (i = 0; i < MAX_BPF_REG; i++)
4670 if (reg_is_pkt_pointer_any(®s[i]))
4671 mark_reg_unknown(env, regs, i);
4673 bpf_for_each_spilled_reg(i, state, reg) {
4676 if (reg_is_pkt_pointer_any(reg))
4677 __mark_reg_unknown(env, reg);
4681 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
4683 struct bpf_verifier_state *vstate = env->cur_state;
4686 for (i = 0; i <= vstate->curframe; i++)
4687 __clear_all_pkt_pointers(env, vstate->frame[i]);
4690 static void release_reg_references(struct bpf_verifier_env *env,
4691 struct bpf_func_state *state,
4694 struct bpf_reg_state *regs = state->regs, *reg;
4697 for (i = 0; i < MAX_BPF_REG; i++)
4698 if (regs[i].ref_obj_id == ref_obj_id)
4699 mark_reg_unknown(env, regs, i);
4701 bpf_for_each_spilled_reg(i, state, reg) {
4704 if (reg->ref_obj_id == ref_obj_id)
4705 __mark_reg_unknown(env, reg);
4709 /* The pointer with the specified id has released its reference to kernel
4710 * resources. Identify all copies of the same pointer and clear the reference.
4712 static int release_reference(struct bpf_verifier_env *env,
4715 struct bpf_verifier_state *vstate = env->cur_state;
4719 err = release_reference_state(cur_func(env), ref_obj_id);
4723 for (i = 0; i <= vstate->curframe; i++)
4724 release_reg_references(env, vstate->frame[i], ref_obj_id);
4729 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
4730 struct bpf_reg_state *regs)
4734 /* after the call registers r0 - r5 were scratched */
4735 for (i = 0; i < CALLER_SAVED_REGS; i++) {
4736 mark_reg_not_init(env, regs, caller_saved[i]);
4737 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
4741 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
4744 struct bpf_verifier_state *state = env->cur_state;
4745 struct bpf_func_info_aux *func_info_aux;
4746 struct bpf_func_state *caller, *callee;
4747 int i, err, subprog, target_insn;
4748 bool is_global = false;
4750 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
4751 verbose(env, "the call stack of %d frames is too deep\n",
4752 state->curframe + 2);
4756 target_insn = *insn_idx + insn->imm;
4757 subprog = find_subprog(env, target_insn + 1);
4759 verbose(env, "verifier bug. No program starts at insn %d\n",
4764 caller = state->frame[state->curframe];
4765 if (state->frame[state->curframe + 1]) {
4766 verbose(env, "verifier bug. Frame %d already allocated\n",
4767 state->curframe + 1);
4771 func_info_aux = env->prog->aux->func_info_aux;
4773 is_global = func_info_aux[subprog].linkage == BTF_FUNC_GLOBAL;
4774 err = btf_check_func_arg_match(env, subprog, caller->regs);
4779 verbose(env, "Caller passes invalid args into func#%d\n",
4783 if (env->log.level & BPF_LOG_LEVEL)
4785 "Func#%d is global and valid. Skipping.\n",
4787 clear_caller_saved_regs(env, caller->regs);
4789 /* All global functions return a 64-bit SCALAR_VALUE */
4790 mark_reg_unknown(env, caller->regs, BPF_REG_0);
4791 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
4793 /* continue with next insn after call */
4798 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
4801 state->frame[state->curframe + 1] = callee;
4803 /* callee cannot access r0, r6 - r9 for reading and has to write
4804 * into its own stack before reading from it.
4805 * callee can read/write into caller's stack
4807 init_func_state(env, callee,
4808 /* remember the callsite, it will be used by bpf_exit */
4809 *insn_idx /* callsite */,
4810 state->curframe + 1 /* frameno within this callchain */,
4811 subprog /* subprog number within this prog */);
4813 /* Transfer references to the callee */
4814 err = transfer_reference_state(callee, caller);
4818 /* copy r1 - r5 args that callee can access. The copy includes parent
4819 * pointers, which connects us up to the liveness chain
4821 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
4822 callee->regs[i] = caller->regs[i];
4824 clear_caller_saved_regs(env, caller->regs);
4826 /* only increment it after check_reg_arg() finished */
4829 /* and go analyze first insn of the callee */
4830 *insn_idx = target_insn;
4832 if (env->log.level & BPF_LOG_LEVEL) {
4833 verbose(env, "caller:\n");
4834 print_verifier_state(env, caller);
4835 verbose(env, "callee:\n");
4836 print_verifier_state(env, callee);
4841 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
4843 struct bpf_verifier_state *state = env->cur_state;
4844 struct bpf_func_state *caller, *callee;
4845 struct bpf_reg_state *r0;
4848 callee = state->frame[state->curframe];
4849 r0 = &callee->regs[BPF_REG_0];
4850 if (r0->type == PTR_TO_STACK) {
4851 /* technically it's ok to return caller's stack pointer
4852 * (or caller's caller's pointer) back to the caller,
4853 * since these pointers are valid. Only current stack
4854 * pointer will be invalid as soon as function exits,
4855 * but let's be conservative
4857 verbose(env, "cannot return stack pointer to the caller\n");
4862 caller = state->frame[state->curframe];
4863 /* return to the caller whatever r0 had in the callee */
4864 caller->regs[BPF_REG_0] = *r0;
4866 /* Transfer references to the caller */
4867 err = transfer_reference_state(caller, callee);
4871 *insn_idx = callee->callsite + 1;
4872 if (env->log.level & BPF_LOG_LEVEL) {
4873 verbose(env, "returning from callee:\n");
4874 print_verifier_state(env, callee);
4875 verbose(env, "to caller at %d:\n", *insn_idx);
4876 print_verifier_state(env, caller);
4878 /* clear everything in the callee */
4879 free_func_state(callee);
4880 state->frame[state->curframe + 1] = NULL;
4884 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
4886 struct bpf_call_arg_meta *meta)
4888 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
4890 if (ret_type != RET_INTEGER ||
4891 (func_id != BPF_FUNC_get_stack &&
4892 func_id != BPF_FUNC_probe_read_str &&
4893 func_id != BPF_FUNC_probe_read_kernel_str &&
4894 func_id != BPF_FUNC_probe_read_user_str))
4897 ret_reg->smax_value = meta->msize_max_value;
4898 ret_reg->s32_max_value = meta->msize_max_value;
4899 ret_reg->smin_value = -MAX_ERRNO;
4900 ret_reg->s32_min_value = -MAX_ERRNO;
4901 __reg_deduce_bounds(ret_reg);
4902 __reg_bound_offset(ret_reg);
4903 __update_reg_bounds(ret_reg);
4907 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
4908 int func_id, int insn_idx)
4910 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
4911 struct bpf_map *map = meta->map_ptr;
4913 if (func_id != BPF_FUNC_tail_call &&
4914 func_id != BPF_FUNC_map_lookup_elem &&
4915 func_id != BPF_FUNC_map_update_elem &&
4916 func_id != BPF_FUNC_map_delete_elem &&
4917 func_id != BPF_FUNC_map_push_elem &&
4918 func_id != BPF_FUNC_map_pop_elem &&
4919 func_id != BPF_FUNC_map_peek_elem)
4923 verbose(env, "kernel subsystem misconfigured verifier\n");
4927 /* In case of read-only, some additional restrictions
4928 * need to be applied in order to prevent altering the
4929 * state of the map from program side.
4931 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
4932 (func_id == BPF_FUNC_map_delete_elem ||
4933 func_id == BPF_FUNC_map_update_elem ||
4934 func_id == BPF_FUNC_map_push_elem ||
4935 func_id == BPF_FUNC_map_pop_elem)) {
4936 verbose(env, "write into map forbidden\n");
4940 if (!BPF_MAP_PTR(aux->map_ptr_state))
4941 bpf_map_ptr_store(aux, meta->map_ptr,
4942 !meta->map_ptr->bypass_spec_v1);
4943 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
4944 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
4945 !meta->map_ptr->bypass_spec_v1);
4950 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
4951 int func_id, int insn_idx)
4953 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
4954 struct bpf_reg_state *regs = cur_regs(env), *reg;
4955 struct bpf_map *map = meta->map_ptr;
4960 if (func_id != BPF_FUNC_tail_call)
4962 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
4963 verbose(env, "kernel subsystem misconfigured verifier\n");
4967 range = tnum_range(0, map->max_entries - 1);
4968 reg = ®s[BPF_REG_3];
4970 if (!register_is_const(reg) || !tnum_in(range, reg->var_off)) {
4971 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
4975 err = mark_chain_precision(env, BPF_REG_3);
4979 val = reg->var_off.value;
4980 if (bpf_map_key_unseen(aux))
4981 bpf_map_key_store(aux, val);
4982 else if (!bpf_map_key_poisoned(aux) &&
4983 bpf_map_key_immediate(aux) != val)
4984 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
4988 static int check_reference_leak(struct bpf_verifier_env *env)
4990 struct bpf_func_state *state = cur_func(env);
4993 for (i = 0; i < state->acquired_refs; i++) {
4994 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
4995 state->refs[i].id, state->refs[i].insn_idx);
4997 return state->acquired_refs ? -EINVAL : 0;
5000 static int check_helper_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
5002 const struct bpf_func_proto *fn = NULL;
5003 struct bpf_reg_state *regs;
5004 struct bpf_call_arg_meta meta;
5008 /* find function prototype */
5009 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
5010 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
5015 if (env->ops->get_func_proto)
5016 fn = env->ops->get_func_proto(func_id, env->prog);
5018 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
5023 /* eBPF programs must be GPL compatible to use GPL-ed functions */
5024 if (!env->prog->gpl_compatible && fn->gpl_only) {
5025 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
5029 if (fn->allowed && !fn->allowed(env->prog)) {
5030 verbose(env, "helper call is not allowed in probe\n");
5034 /* With LD_ABS/IND some JITs save/restore skb from r1. */
5035 changes_data = bpf_helper_changes_pkt_data(fn->func);
5036 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
5037 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
5038 func_id_name(func_id), func_id);
5042 memset(&meta, 0, sizeof(meta));
5043 meta.pkt_access = fn->pkt_access;
5045 err = check_func_proto(fn, func_id);
5047 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
5048 func_id_name(func_id), func_id);
5052 meta.func_id = func_id;
5054 for (i = 0; i < 5; i++) {
5055 err = check_func_arg(env, i, &meta, fn);
5060 err = record_func_map(env, &meta, func_id, insn_idx);
5064 err = record_func_key(env, &meta, func_id, insn_idx);
5068 /* Mark slots with STACK_MISC in case of raw mode, stack offset
5069 * is inferred from register state.
5071 for (i = 0; i < meta.access_size; i++) {
5072 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
5073 BPF_WRITE, -1, false);
5078 if (func_id == BPF_FUNC_tail_call) {
5079 err = check_reference_leak(env);
5081 verbose(env, "tail_call would lead to reference leak\n");
5084 } else if (is_release_function(func_id)) {
5085 err = release_reference(env, meta.ref_obj_id);
5087 verbose(env, "func %s#%d reference has not been acquired before\n",
5088 func_id_name(func_id), func_id);
5093 regs = cur_regs(env);
5095 /* check that flags argument in get_local_storage(map, flags) is 0,
5096 * this is required because get_local_storage() can't return an error.
5098 if (func_id == BPF_FUNC_get_local_storage &&
5099 !register_is_null(®s[BPF_REG_2])) {
5100 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
5104 /* reset caller saved regs */
5105 for (i = 0; i < CALLER_SAVED_REGS; i++) {
5106 mark_reg_not_init(env, regs, caller_saved[i]);
5107 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
5110 /* helper call returns 64-bit value. */
5111 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
5113 /* update return register (already marked as written above) */
5114 if (fn->ret_type == RET_INTEGER) {
5115 /* sets type to SCALAR_VALUE */
5116 mark_reg_unknown(env, regs, BPF_REG_0);
5117 } else if (fn->ret_type == RET_VOID) {
5118 regs[BPF_REG_0].type = NOT_INIT;
5119 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL ||
5120 fn->ret_type == RET_PTR_TO_MAP_VALUE) {
5121 /* There is no offset yet applied, variable or fixed */
5122 mark_reg_known_zero(env, regs, BPF_REG_0);
5123 /* remember map_ptr, so that check_map_access()
5124 * can check 'value_size' boundary of memory access
5125 * to map element returned from bpf_map_lookup_elem()
5127 if (meta.map_ptr == NULL) {
5129 "kernel subsystem misconfigured verifier\n");
5132 regs[BPF_REG_0].map_ptr = meta.map_ptr;
5133 if (fn->ret_type == RET_PTR_TO_MAP_VALUE) {
5134 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE;
5135 if (map_value_has_spin_lock(meta.map_ptr))
5136 regs[BPF_REG_0].id = ++env->id_gen;
5138 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
5140 } else if (fn->ret_type == RET_PTR_TO_SOCKET_OR_NULL) {
5141 mark_reg_known_zero(env, regs, BPF_REG_0);
5142 regs[BPF_REG_0].type = PTR_TO_SOCKET_OR_NULL;
5143 } else if (fn->ret_type == RET_PTR_TO_SOCK_COMMON_OR_NULL) {
5144 mark_reg_known_zero(env, regs, BPF_REG_0);
5145 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON_OR_NULL;
5146 } else if (fn->ret_type == RET_PTR_TO_TCP_SOCK_OR_NULL) {
5147 mark_reg_known_zero(env, regs, BPF_REG_0);
5148 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK_OR_NULL;
5149 } else if (fn->ret_type == RET_PTR_TO_ALLOC_MEM_OR_NULL) {
5150 mark_reg_known_zero(env, regs, BPF_REG_0);
5151 regs[BPF_REG_0].type = PTR_TO_MEM_OR_NULL;
5152 regs[BPF_REG_0].mem_size = meta.mem_size;
5153 } else if (fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID_OR_NULL ||
5154 fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID) {
5155 const struct btf_type *t;
5157 mark_reg_known_zero(env, regs, BPF_REG_0);
5158 t = btf_type_skip_modifiers(btf_vmlinux, meta.ret_btf_id, NULL);
5159 if (!btf_type_is_struct(t)) {
5161 const struct btf_type *ret;
5164 /* resolve the type size of ksym. */
5165 ret = btf_resolve_size(btf_vmlinux, t, &tsize);
5167 tname = btf_name_by_offset(btf_vmlinux, t->name_off);
5168 verbose(env, "unable to resolve the size of type '%s': %ld\n",
5169 tname, PTR_ERR(ret));
5172 regs[BPF_REG_0].type =
5173 fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID ?
5174 PTR_TO_MEM : PTR_TO_MEM_OR_NULL;
5175 regs[BPF_REG_0].mem_size = tsize;
5177 regs[BPF_REG_0].type =
5178 fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID ?
5179 PTR_TO_BTF_ID : PTR_TO_BTF_ID_OR_NULL;
5180 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
5182 } else if (fn->ret_type == RET_PTR_TO_BTF_ID_OR_NULL) {
5185 mark_reg_known_zero(env, regs, BPF_REG_0);
5186 regs[BPF_REG_0].type = PTR_TO_BTF_ID_OR_NULL;
5187 ret_btf_id = *fn->ret_btf_id;
5188 if (ret_btf_id == 0) {
5189 verbose(env, "invalid return type %d of func %s#%d\n",
5190 fn->ret_type, func_id_name(func_id), func_id);
5193 regs[BPF_REG_0].btf_id = ret_btf_id;
5195 verbose(env, "unknown return type %d of func %s#%d\n",
5196 fn->ret_type, func_id_name(func_id), func_id);
5200 if (reg_type_may_be_null(regs[BPF_REG_0].type))
5201 regs[BPF_REG_0].id = ++env->id_gen;
5203 if (is_ptr_cast_function(func_id)) {
5204 /* For release_reference() */
5205 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
5206 } else if (is_acquire_function(func_id, meta.map_ptr)) {
5207 int id = acquire_reference_state(env, insn_idx);
5211 /* For mark_ptr_or_null_reg() */
5212 regs[BPF_REG_0].id = id;
5213 /* For release_reference() */
5214 regs[BPF_REG_0].ref_obj_id = id;
5217 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
5219 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
5223 if ((func_id == BPF_FUNC_get_stack ||
5224 func_id == BPF_FUNC_get_task_stack) &&
5225 !env->prog->has_callchain_buf) {
5226 const char *err_str;
5228 #ifdef CONFIG_PERF_EVENTS
5229 err = get_callchain_buffers(sysctl_perf_event_max_stack);
5230 err_str = "cannot get callchain buffer for func %s#%d\n";
5233 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
5236 verbose(env, err_str, func_id_name(func_id), func_id);
5240 env->prog->has_callchain_buf = true;
5243 if (func_id == BPF_FUNC_get_stackid || func_id == BPF_FUNC_get_stack)
5244 env->prog->call_get_stack = true;
5247 clear_all_pkt_pointers(env);
5251 static bool signed_add_overflows(s64 a, s64 b)
5253 /* Do the add in u64, where overflow is well-defined */
5254 s64 res = (s64)((u64)a + (u64)b);
5261 static bool signed_add32_overflows(s32 a, s32 b)
5263 /* Do the add in u32, where overflow is well-defined */
5264 s32 res = (s32)((u32)a + (u32)b);
5271 static bool signed_sub_overflows(s64 a, s64 b)
5273 /* Do the sub in u64, where overflow is well-defined */
5274 s64 res = (s64)((u64)a - (u64)b);
5281 static bool signed_sub32_overflows(s32 a, s32 b)
5283 /* Do the sub in u32, where overflow is well-defined */
5284 s32 res = (s32)((u32)a - (u32)b);
5291 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
5292 const struct bpf_reg_state *reg,
5293 enum bpf_reg_type type)
5295 bool known = tnum_is_const(reg->var_off);
5296 s64 val = reg->var_off.value;
5297 s64 smin = reg->smin_value;
5299 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
5300 verbose(env, "math between %s pointer and %lld is not allowed\n",
5301 reg_type_str[type], val);
5305 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
5306 verbose(env, "%s pointer offset %d is not allowed\n",
5307 reg_type_str[type], reg->off);
5311 if (smin == S64_MIN) {
5312 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
5313 reg_type_str[type]);
5317 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
5318 verbose(env, "value %lld makes %s pointer be out of bounds\n",
5319 smin, reg_type_str[type]);
5326 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
5328 return &env->insn_aux_data[env->insn_idx];
5331 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
5332 const struct bpf_reg_state *off_reg,
5333 u32 *ptr_limit, u8 opcode)
5335 bool off_is_neg = off_reg->smin_value < 0;
5336 bool mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
5337 (opcode == BPF_SUB && !off_is_neg);
5340 if (!tnum_is_const(off_reg->var_off) &&
5341 (off_reg->smin_value < 0) != (off_reg->smax_value < 0))
5344 switch (ptr_reg->type) {
5346 /* Offset 0 is out-of-bounds, but acceptable start for the
5347 * left direction, see BPF_REG_FP.
5349 max = MAX_BPF_STACK + mask_to_left;
5350 /* Indirect variable offset stack access is prohibited in
5351 * unprivileged mode so it's not handled here.
5353 off = ptr_reg->off + ptr_reg->var_off.value;
5355 *ptr_limit = MAX_BPF_STACK + off;
5357 *ptr_limit = -off - 1;
5358 return *ptr_limit >= max ? -ERANGE : 0;
5359 case PTR_TO_MAP_VALUE:
5360 max = ptr_reg->map_ptr->value_size;
5362 *ptr_limit = ptr_reg->umax_value + ptr_reg->off;
5364 off = ptr_reg->smin_value + ptr_reg->off;
5365 *ptr_limit = ptr_reg->map_ptr->value_size - off - 1;
5367 return *ptr_limit >= max ? -ERANGE : 0;
5373 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
5374 const struct bpf_insn *insn)
5376 return env->bypass_spec_v1 || BPF_SRC(insn->code) == BPF_K;
5379 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
5380 u32 alu_state, u32 alu_limit)
5382 /* If we arrived here from different branches with different
5383 * state or limits to sanitize, then this won't work.
5385 if (aux->alu_state &&
5386 (aux->alu_state != alu_state ||
5387 aux->alu_limit != alu_limit))
5390 /* Corresponding fixup done in fixup_bpf_calls(). */
5391 aux->alu_state = alu_state;
5392 aux->alu_limit = alu_limit;
5396 static int sanitize_val_alu(struct bpf_verifier_env *env,
5397 struct bpf_insn *insn)
5399 struct bpf_insn_aux_data *aux = cur_aux(env);
5401 if (can_skip_alu_sanitation(env, insn))
5404 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
5407 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
5408 struct bpf_insn *insn,
5409 const struct bpf_reg_state *ptr_reg,
5410 const struct bpf_reg_state *off_reg,
5411 struct bpf_reg_state *dst_reg)
5413 struct bpf_verifier_state *vstate = env->cur_state;
5414 struct bpf_insn_aux_data *aux = cur_aux(env);
5415 bool off_is_neg = off_reg->smin_value < 0;
5416 bool ptr_is_dst_reg = ptr_reg == dst_reg;
5417 u8 opcode = BPF_OP(insn->code);
5418 u32 alu_state, alu_limit;
5419 struct bpf_reg_state tmp;
5423 if (can_skip_alu_sanitation(env, insn))
5426 /* We already marked aux for masking from non-speculative
5427 * paths, thus we got here in the first place. We only care
5428 * to explore bad access from here.
5430 if (vstate->speculative)
5433 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
5434 alu_state |= ptr_is_dst_reg ?
5435 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
5437 err = retrieve_ptr_limit(ptr_reg, off_reg, &alu_limit, opcode);
5441 err = update_alu_sanitation_state(aux, alu_state, alu_limit);
5445 /* Simulate and find potential out-of-bounds access under
5446 * speculative execution from truncation as a result of
5447 * masking when off was not within expected range. If off
5448 * sits in dst, then we temporarily need to move ptr there
5449 * to simulate dst (== 0) +/-= ptr. Needed, for example,
5450 * for cases where we use K-based arithmetic in one direction
5451 * and truncated reg-based in the other in order to explore
5454 if (!ptr_is_dst_reg) {
5456 *dst_reg = *ptr_reg;
5458 ret = push_stack(env, env->insn_idx + 1, env->insn_idx, true);
5459 if (!ptr_is_dst_reg && ret)
5461 return !ret ? -EFAULT : 0;
5464 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
5465 * Caller should also handle BPF_MOV case separately.
5466 * If we return -EACCES, caller may want to try again treating pointer as a
5467 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
5469 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
5470 struct bpf_insn *insn,
5471 const struct bpf_reg_state *ptr_reg,
5472 const struct bpf_reg_state *off_reg)
5474 struct bpf_verifier_state *vstate = env->cur_state;
5475 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5476 struct bpf_reg_state *regs = state->regs, *dst_reg;
5477 bool known = tnum_is_const(off_reg->var_off);
5478 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
5479 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
5480 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
5481 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
5482 u8 opcode = BPF_OP(insn->code);
5483 u32 dst = insn->dst_reg;
5486 dst_reg = ®s[dst];
5488 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
5489 smin_val > smax_val || umin_val > umax_val) {
5490 /* Taint dst register if offset had invalid bounds derived from
5491 * e.g. dead branches.
5493 __mark_reg_unknown(env, dst_reg);
5497 if (BPF_CLASS(insn->code) != BPF_ALU64) {
5498 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
5499 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
5500 __mark_reg_unknown(env, dst_reg);
5505 "R%d 32-bit pointer arithmetic prohibited\n",
5510 switch (ptr_reg->type) {
5511 case PTR_TO_MAP_VALUE_OR_NULL:
5512 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
5513 dst, reg_type_str[ptr_reg->type]);
5515 case CONST_PTR_TO_MAP:
5516 /* smin_val represents the known value */
5517 if (known && smin_val == 0 && opcode == BPF_ADD)
5520 case PTR_TO_PACKET_END:
5522 case PTR_TO_SOCKET_OR_NULL:
5523 case PTR_TO_SOCK_COMMON:
5524 case PTR_TO_SOCK_COMMON_OR_NULL:
5525 case PTR_TO_TCP_SOCK:
5526 case PTR_TO_TCP_SOCK_OR_NULL:
5527 case PTR_TO_XDP_SOCK:
5528 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
5529 dst, reg_type_str[ptr_reg->type]);
5535 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
5536 * The id may be overwritten later if we create a new variable offset.
5538 dst_reg->type = ptr_reg->type;
5539 dst_reg->id = ptr_reg->id;
5541 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
5542 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
5545 /* pointer types do not carry 32-bit bounds at the moment. */
5546 __mark_reg32_unbounded(dst_reg);
5550 ret = sanitize_ptr_alu(env, insn, ptr_reg, off_reg, dst_reg);
5552 verbose(env, "R%d tried to add from different maps, paths, or prohibited types\n", dst);
5555 /* We can take a fixed offset as long as it doesn't overflow
5556 * the s32 'off' field
5558 if (known && (ptr_reg->off + smin_val ==
5559 (s64)(s32)(ptr_reg->off + smin_val))) {
5560 /* pointer += K. Accumulate it into fixed offset */
5561 dst_reg->smin_value = smin_ptr;
5562 dst_reg->smax_value = smax_ptr;
5563 dst_reg->umin_value = umin_ptr;
5564 dst_reg->umax_value = umax_ptr;
5565 dst_reg->var_off = ptr_reg->var_off;
5566 dst_reg->off = ptr_reg->off + smin_val;
5567 dst_reg->raw = ptr_reg->raw;
5570 /* A new variable offset is created. Note that off_reg->off
5571 * == 0, since it's a scalar.
5572 * dst_reg gets the pointer type and since some positive
5573 * integer value was added to the pointer, give it a new 'id'
5574 * if it's a PTR_TO_PACKET.
5575 * this creates a new 'base' pointer, off_reg (variable) gets
5576 * added into the variable offset, and we copy the fixed offset
5579 if (signed_add_overflows(smin_ptr, smin_val) ||
5580 signed_add_overflows(smax_ptr, smax_val)) {
5581 dst_reg->smin_value = S64_MIN;
5582 dst_reg->smax_value = S64_MAX;
5584 dst_reg->smin_value = smin_ptr + smin_val;
5585 dst_reg->smax_value = smax_ptr + smax_val;
5587 if (umin_ptr + umin_val < umin_ptr ||
5588 umax_ptr + umax_val < umax_ptr) {
5589 dst_reg->umin_value = 0;
5590 dst_reg->umax_value = U64_MAX;
5592 dst_reg->umin_value = umin_ptr + umin_val;
5593 dst_reg->umax_value = umax_ptr + umax_val;
5595 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
5596 dst_reg->off = ptr_reg->off;
5597 dst_reg->raw = ptr_reg->raw;
5598 if (reg_is_pkt_pointer(ptr_reg)) {
5599 dst_reg->id = ++env->id_gen;
5600 /* something was added to pkt_ptr, set range to zero */
5605 ret = sanitize_ptr_alu(env, insn, ptr_reg, off_reg, dst_reg);
5607 verbose(env, "R%d tried to sub from different maps, paths, or prohibited types\n", dst);
5610 if (dst_reg == off_reg) {
5611 /* scalar -= pointer. Creates an unknown scalar */
5612 verbose(env, "R%d tried to subtract pointer from scalar\n",
5616 /* We don't allow subtraction from FP, because (according to
5617 * test_verifier.c test "invalid fp arithmetic", JITs might not
5618 * be able to deal with it.
5620 if (ptr_reg->type == PTR_TO_STACK) {
5621 verbose(env, "R%d subtraction from stack pointer prohibited\n",
5625 if (known && (ptr_reg->off - smin_val ==
5626 (s64)(s32)(ptr_reg->off - smin_val))) {
5627 /* pointer -= K. Subtract it from fixed offset */
5628 dst_reg->smin_value = smin_ptr;
5629 dst_reg->smax_value = smax_ptr;
5630 dst_reg->umin_value = umin_ptr;
5631 dst_reg->umax_value = umax_ptr;
5632 dst_reg->var_off = ptr_reg->var_off;
5633 dst_reg->id = ptr_reg->id;
5634 dst_reg->off = ptr_reg->off - smin_val;
5635 dst_reg->raw = ptr_reg->raw;
5638 /* A new variable offset is created. If the subtrahend is known
5639 * nonnegative, then any reg->range we had before is still good.
5641 if (signed_sub_overflows(smin_ptr, smax_val) ||
5642 signed_sub_overflows(smax_ptr, smin_val)) {
5643 /* Overflow possible, we know nothing */
5644 dst_reg->smin_value = S64_MIN;
5645 dst_reg->smax_value = S64_MAX;
5647 dst_reg->smin_value = smin_ptr - smax_val;
5648 dst_reg->smax_value = smax_ptr - smin_val;
5650 if (umin_ptr < umax_val) {
5651 /* Overflow possible, we know nothing */
5652 dst_reg->umin_value = 0;
5653 dst_reg->umax_value = U64_MAX;
5655 /* Cannot overflow (as long as bounds are consistent) */
5656 dst_reg->umin_value = umin_ptr - umax_val;
5657 dst_reg->umax_value = umax_ptr - umin_val;
5659 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
5660 dst_reg->off = ptr_reg->off;
5661 dst_reg->raw = ptr_reg->raw;
5662 if (reg_is_pkt_pointer(ptr_reg)) {
5663 dst_reg->id = ++env->id_gen;
5664 /* something was added to pkt_ptr, set range to zero */
5672 /* bitwise ops on pointers are troublesome, prohibit. */
5673 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
5674 dst, bpf_alu_string[opcode >> 4]);
5677 /* other operators (e.g. MUL,LSH) produce non-pointer results */
5678 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
5679 dst, bpf_alu_string[opcode >> 4]);
5683 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
5686 __update_reg_bounds(dst_reg);
5687 __reg_deduce_bounds(dst_reg);
5688 __reg_bound_offset(dst_reg);
5690 /* For unprivileged we require that resulting offset must be in bounds
5691 * in order to be able to sanitize access later on.
5693 if (!env->bypass_spec_v1) {
5694 if (dst_reg->type == PTR_TO_MAP_VALUE &&
5695 check_map_access(env, dst, dst_reg->off, 1, false)) {
5696 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
5697 "prohibited for !root\n", dst);
5699 } else if (dst_reg->type == PTR_TO_STACK &&
5700 check_stack_access(env, dst_reg, dst_reg->off +
5701 dst_reg->var_off.value, 1)) {
5702 verbose(env, "R%d stack pointer arithmetic goes out of range, "
5703 "prohibited for !root\n", dst);
5711 static void scalar32_min_max_add(struct bpf_reg_state *dst_reg,
5712 struct bpf_reg_state *src_reg)
5714 s32 smin_val = src_reg->s32_min_value;
5715 s32 smax_val = src_reg->s32_max_value;
5716 u32 umin_val = src_reg->u32_min_value;
5717 u32 umax_val = src_reg->u32_max_value;
5719 if (signed_add32_overflows(dst_reg->s32_min_value, smin_val) ||
5720 signed_add32_overflows(dst_reg->s32_max_value, smax_val)) {
5721 dst_reg->s32_min_value = S32_MIN;
5722 dst_reg->s32_max_value = S32_MAX;
5724 dst_reg->s32_min_value += smin_val;
5725 dst_reg->s32_max_value += smax_val;
5727 if (dst_reg->u32_min_value + umin_val < umin_val ||
5728 dst_reg->u32_max_value + umax_val < umax_val) {
5729 dst_reg->u32_min_value = 0;
5730 dst_reg->u32_max_value = U32_MAX;
5732 dst_reg->u32_min_value += umin_val;
5733 dst_reg->u32_max_value += umax_val;
5737 static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
5738 struct bpf_reg_state *src_reg)
5740 s64 smin_val = src_reg->smin_value;
5741 s64 smax_val = src_reg->smax_value;
5742 u64 umin_val = src_reg->umin_value;
5743 u64 umax_val = src_reg->umax_value;
5745 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
5746 signed_add_overflows(dst_reg->smax_value, smax_val)) {
5747 dst_reg->smin_value = S64_MIN;
5748 dst_reg->smax_value = S64_MAX;
5750 dst_reg->smin_value += smin_val;
5751 dst_reg->smax_value += smax_val;
5753 if (dst_reg->umin_value + umin_val < umin_val ||
5754 dst_reg->umax_value + umax_val < umax_val) {
5755 dst_reg->umin_value = 0;
5756 dst_reg->umax_value = U64_MAX;
5758 dst_reg->umin_value += umin_val;
5759 dst_reg->umax_value += umax_val;
5763 static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg,
5764 struct bpf_reg_state *src_reg)
5766 s32 smin_val = src_reg->s32_min_value;
5767 s32 smax_val = src_reg->s32_max_value;
5768 u32 umin_val = src_reg->u32_min_value;
5769 u32 umax_val = src_reg->u32_max_value;
5771 if (signed_sub32_overflows(dst_reg->s32_min_value, smax_val) ||
5772 signed_sub32_overflows(dst_reg->s32_max_value, smin_val)) {
5773 /* Overflow possible, we know nothing */
5774 dst_reg->s32_min_value = S32_MIN;
5775 dst_reg->s32_max_value = S32_MAX;
5777 dst_reg->s32_min_value -= smax_val;
5778 dst_reg->s32_max_value -= smin_val;
5780 if (dst_reg->u32_min_value < umax_val) {
5781 /* Overflow possible, we know nothing */
5782 dst_reg->u32_min_value = 0;
5783 dst_reg->u32_max_value = U32_MAX;
5785 /* Cannot overflow (as long as bounds are consistent) */
5786 dst_reg->u32_min_value -= umax_val;
5787 dst_reg->u32_max_value -= umin_val;
5791 static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
5792 struct bpf_reg_state *src_reg)
5794 s64 smin_val = src_reg->smin_value;
5795 s64 smax_val = src_reg->smax_value;
5796 u64 umin_val = src_reg->umin_value;
5797 u64 umax_val = src_reg->umax_value;
5799 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
5800 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
5801 /* Overflow possible, we know nothing */
5802 dst_reg->smin_value = S64_MIN;
5803 dst_reg->smax_value = S64_MAX;
5805 dst_reg->smin_value -= smax_val;
5806 dst_reg->smax_value -= smin_val;
5808 if (dst_reg->umin_value < umax_val) {
5809 /* Overflow possible, we know nothing */
5810 dst_reg->umin_value = 0;
5811 dst_reg->umax_value = U64_MAX;
5813 /* Cannot overflow (as long as bounds are consistent) */
5814 dst_reg->umin_value -= umax_val;
5815 dst_reg->umax_value -= umin_val;
5819 static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg,
5820 struct bpf_reg_state *src_reg)
5822 s32 smin_val = src_reg->s32_min_value;
5823 u32 umin_val = src_reg->u32_min_value;
5824 u32 umax_val = src_reg->u32_max_value;
5826 if (smin_val < 0 || dst_reg->s32_min_value < 0) {
5827 /* Ain't nobody got time to multiply that sign */
5828 __mark_reg32_unbounded(dst_reg);
5831 /* Both values are positive, so we can work with unsigned and
5832 * copy the result to signed (unless it exceeds S32_MAX).
5834 if (umax_val > U16_MAX || dst_reg->u32_max_value > U16_MAX) {
5835 /* Potential overflow, we know nothing */
5836 __mark_reg32_unbounded(dst_reg);
5839 dst_reg->u32_min_value *= umin_val;
5840 dst_reg->u32_max_value *= umax_val;
5841 if (dst_reg->u32_max_value > S32_MAX) {
5842 /* Overflow possible, we know nothing */
5843 dst_reg->s32_min_value = S32_MIN;
5844 dst_reg->s32_max_value = S32_MAX;
5846 dst_reg->s32_min_value = dst_reg->u32_min_value;
5847 dst_reg->s32_max_value = dst_reg->u32_max_value;
5851 static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
5852 struct bpf_reg_state *src_reg)
5854 s64 smin_val = src_reg->smin_value;
5855 u64 umin_val = src_reg->umin_value;
5856 u64 umax_val = src_reg->umax_value;
5858 if (smin_val < 0 || dst_reg->smin_value < 0) {
5859 /* Ain't nobody got time to multiply that sign */
5860 __mark_reg64_unbounded(dst_reg);
5863 /* Both values are positive, so we can work with unsigned and
5864 * copy the result to signed (unless it exceeds S64_MAX).
5866 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
5867 /* Potential overflow, we know nothing */
5868 __mark_reg64_unbounded(dst_reg);
5871 dst_reg->umin_value *= umin_val;
5872 dst_reg->umax_value *= umax_val;
5873 if (dst_reg->umax_value > S64_MAX) {
5874 /* Overflow possible, we know nothing */
5875 dst_reg->smin_value = S64_MIN;
5876 dst_reg->smax_value = S64_MAX;
5878 dst_reg->smin_value = dst_reg->umin_value;
5879 dst_reg->smax_value = dst_reg->umax_value;
5883 static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
5884 struct bpf_reg_state *src_reg)
5886 bool src_known = tnum_subreg_is_const(src_reg->var_off);
5887 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
5888 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
5889 s32 smin_val = src_reg->s32_min_value;
5890 u32 umax_val = src_reg->u32_max_value;
5892 /* Assuming scalar64_min_max_and will be called so its safe
5893 * to skip updating register for known 32-bit case.
5895 if (src_known && dst_known)
5898 /* We get our minimum from the var_off, since that's inherently
5899 * bitwise. Our maximum is the minimum of the operands' maxima.
5901 dst_reg->u32_min_value = var32_off.value;
5902 dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val);
5903 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
5904 /* Lose signed bounds when ANDing negative numbers,
5905 * ain't nobody got time for that.
5907 dst_reg->s32_min_value = S32_MIN;
5908 dst_reg->s32_max_value = S32_MAX;
5910 /* ANDing two positives gives a positive, so safe to
5911 * cast result into s64.
5913 dst_reg->s32_min_value = dst_reg->u32_min_value;
5914 dst_reg->s32_max_value = dst_reg->u32_max_value;
5919 static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
5920 struct bpf_reg_state *src_reg)
5922 bool src_known = tnum_is_const(src_reg->var_off);
5923 bool dst_known = tnum_is_const(dst_reg->var_off);
5924 s64 smin_val = src_reg->smin_value;
5925 u64 umax_val = src_reg->umax_value;
5927 if (src_known && dst_known) {
5928 __mark_reg_known(dst_reg, dst_reg->var_off.value);
5932 /* We get our minimum from the var_off, since that's inherently
5933 * bitwise. Our maximum is the minimum of the operands' maxima.
5935 dst_reg->umin_value = dst_reg->var_off.value;
5936 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
5937 if (dst_reg->smin_value < 0 || smin_val < 0) {
5938 /* Lose signed bounds when ANDing negative numbers,
5939 * ain't nobody got time for that.
5941 dst_reg->smin_value = S64_MIN;
5942 dst_reg->smax_value = S64_MAX;
5944 /* ANDing two positives gives a positive, so safe to
5945 * cast result into s64.
5947 dst_reg->smin_value = dst_reg->umin_value;
5948 dst_reg->smax_value = dst_reg->umax_value;
5950 /* We may learn something more from the var_off */
5951 __update_reg_bounds(dst_reg);
5954 static void scalar32_min_max_or(struct bpf_reg_state *dst_reg,
5955 struct bpf_reg_state *src_reg)
5957 bool src_known = tnum_subreg_is_const(src_reg->var_off);
5958 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
5959 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
5960 s32 smin_val = src_reg->s32_min_value;
5961 u32 umin_val = src_reg->u32_min_value;
5963 /* Assuming scalar64_min_max_or will be called so it is safe
5964 * to skip updating register for known case.
5966 if (src_known && dst_known)
5969 /* We get our maximum from the var_off, and our minimum is the
5970 * maximum of the operands' minima
5972 dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val);
5973 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
5974 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
5975 /* Lose signed bounds when ORing negative numbers,
5976 * ain't nobody got time for that.
5978 dst_reg->s32_min_value = S32_MIN;
5979 dst_reg->s32_max_value = S32_MAX;
5981 /* ORing two positives gives a positive, so safe to
5982 * cast result into s64.
5984 dst_reg->s32_min_value = dst_reg->u32_min_value;
5985 dst_reg->s32_max_value = dst_reg->u32_max_value;
5989 static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
5990 struct bpf_reg_state *src_reg)
5992 bool src_known = tnum_is_const(src_reg->var_off);
5993 bool dst_known = tnum_is_const(dst_reg->var_off);
5994 s64 smin_val = src_reg->smin_value;
5995 u64 umin_val = src_reg->umin_value;
5997 if (src_known && dst_known) {
5998 __mark_reg_known(dst_reg, dst_reg->var_off.value);
6002 /* We get our maximum from the var_off, and our minimum is the
6003 * maximum of the operands' minima
6005 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
6006 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
6007 if (dst_reg->smin_value < 0 || smin_val < 0) {
6008 /* Lose signed bounds when ORing negative numbers,
6009 * ain't nobody got time for that.
6011 dst_reg->smin_value = S64_MIN;
6012 dst_reg->smax_value = S64_MAX;
6014 /* ORing two positives gives a positive, so safe to
6015 * cast result into s64.
6017 dst_reg->smin_value = dst_reg->umin_value;
6018 dst_reg->smax_value = dst_reg->umax_value;
6020 /* We may learn something more from the var_off */
6021 __update_reg_bounds(dst_reg);
6024 static void scalar32_min_max_xor(struct bpf_reg_state *dst_reg,
6025 struct bpf_reg_state *src_reg)
6027 bool src_known = tnum_subreg_is_const(src_reg->var_off);
6028 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
6029 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
6030 s32 smin_val = src_reg->s32_min_value;
6032 /* Assuming scalar64_min_max_xor will be called so it is safe
6033 * to skip updating register for known case.
6035 if (src_known && dst_known)
6038 /* We get both minimum and maximum from the var32_off. */
6039 dst_reg->u32_min_value = var32_off.value;
6040 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
6042 if (dst_reg->s32_min_value >= 0 && smin_val >= 0) {
6043 /* XORing two positive sign numbers gives a positive,
6044 * so safe to cast u32 result into s32.
6046 dst_reg->s32_min_value = dst_reg->u32_min_value;
6047 dst_reg->s32_max_value = dst_reg->u32_max_value;
6049 dst_reg->s32_min_value = S32_MIN;
6050 dst_reg->s32_max_value = S32_MAX;
6054 static void scalar_min_max_xor(struct bpf_reg_state *dst_reg,
6055 struct bpf_reg_state *src_reg)
6057 bool src_known = tnum_is_const(src_reg->var_off);
6058 bool dst_known = tnum_is_const(dst_reg->var_off);
6059 s64 smin_val = src_reg->smin_value;
6061 if (src_known && dst_known) {
6062 /* dst_reg->var_off.value has been updated earlier */
6063 __mark_reg_known(dst_reg, dst_reg->var_off.value);
6067 /* We get both minimum and maximum from the var_off. */
6068 dst_reg->umin_value = dst_reg->var_off.value;
6069 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
6071 if (dst_reg->smin_value >= 0 && smin_val >= 0) {
6072 /* XORing two positive sign numbers gives a positive,
6073 * so safe to cast u64 result into s64.
6075 dst_reg->smin_value = dst_reg->umin_value;
6076 dst_reg->smax_value = dst_reg->umax_value;
6078 dst_reg->smin_value = S64_MIN;
6079 dst_reg->smax_value = S64_MAX;
6082 __update_reg_bounds(dst_reg);
6085 static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
6086 u64 umin_val, u64 umax_val)
6088 /* We lose all sign bit information (except what we can pick
6091 dst_reg->s32_min_value = S32_MIN;
6092 dst_reg->s32_max_value = S32_MAX;
6093 /* If we might shift our top bit out, then we know nothing */
6094 if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) {
6095 dst_reg->u32_min_value = 0;
6096 dst_reg->u32_max_value = U32_MAX;
6098 dst_reg->u32_min_value <<= umin_val;
6099 dst_reg->u32_max_value <<= umax_val;
6103 static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
6104 struct bpf_reg_state *src_reg)
6106 u32 umax_val = src_reg->u32_max_value;
6107 u32 umin_val = src_reg->u32_min_value;
6108 /* u32 alu operation will zext upper bits */
6109 struct tnum subreg = tnum_subreg(dst_reg->var_off);
6111 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
6112 dst_reg->var_off = tnum_subreg(tnum_lshift(subreg, umin_val));
6113 /* Not required but being careful mark reg64 bounds as unknown so
6114 * that we are forced to pick them up from tnum and zext later and
6115 * if some path skips this step we are still safe.
6117 __mark_reg64_unbounded(dst_reg);
6118 __update_reg32_bounds(dst_reg);
6121 static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg,
6122 u64 umin_val, u64 umax_val)
6124 /* Special case <<32 because it is a common compiler pattern to sign
6125 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
6126 * positive we know this shift will also be positive so we can track
6127 * bounds correctly. Otherwise we lose all sign bit information except
6128 * what we can pick up from var_off. Perhaps we can generalize this
6129 * later to shifts of any length.
6131 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0)
6132 dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32;
6134 dst_reg->smax_value = S64_MAX;
6136 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0)
6137 dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32;
6139 dst_reg->smin_value = S64_MIN;
6141 /* If we might shift our top bit out, then we know nothing */
6142 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
6143 dst_reg->umin_value = 0;
6144 dst_reg->umax_value = U64_MAX;
6146 dst_reg->umin_value <<= umin_val;
6147 dst_reg->umax_value <<= umax_val;
6151 static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
6152 struct bpf_reg_state *src_reg)
6154 u64 umax_val = src_reg->umax_value;
6155 u64 umin_val = src_reg->umin_value;
6157 /* scalar64 calc uses 32bit unshifted bounds so must be called first */
6158 __scalar64_min_max_lsh(dst_reg, umin_val, umax_val);
6159 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
6161 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
6162 /* We may learn something more from the var_off */
6163 __update_reg_bounds(dst_reg);
6166 static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg,
6167 struct bpf_reg_state *src_reg)
6169 struct tnum subreg = tnum_subreg(dst_reg->var_off);
6170 u32 umax_val = src_reg->u32_max_value;
6171 u32 umin_val = src_reg->u32_min_value;
6173 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
6174 * be negative, then either:
6175 * 1) src_reg might be zero, so the sign bit of the result is
6176 * unknown, so we lose our signed bounds
6177 * 2) it's known negative, thus the unsigned bounds capture the
6179 * 3) the signed bounds cross zero, so they tell us nothing
6181 * If the value in dst_reg is known nonnegative, then again the
6182 * unsigned bounts capture the signed bounds.
6183 * Thus, in all cases it suffices to blow away our signed bounds
6184 * and rely on inferring new ones from the unsigned bounds and
6185 * var_off of the result.
6187 dst_reg->s32_min_value = S32_MIN;
6188 dst_reg->s32_max_value = S32_MAX;
6190 dst_reg->var_off = tnum_rshift(subreg, umin_val);
6191 dst_reg->u32_min_value >>= umax_val;
6192 dst_reg->u32_max_value >>= umin_val;
6194 __mark_reg64_unbounded(dst_reg);
6195 __update_reg32_bounds(dst_reg);
6198 static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
6199 struct bpf_reg_state *src_reg)
6201 u64 umax_val = src_reg->umax_value;
6202 u64 umin_val = src_reg->umin_value;
6204 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
6205 * be negative, then either:
6206 * 1) src_reg might be zero, so the sign bit of the result is
6207 * unknown, so we lose our signed bounds
6208 * 2) it's known negative, thus the unsigned bounds capture the
6210 * 3) the signed bounds cross zero, so they tell us nothing
6212 * If the value in dst_reg is known nonnegative, then again the
6213 * unsigned bounts capture the signed bounds.
6214 * Thus, in all cases it suffices to blow away our signed bounds
6215 * and rely on inferring new ones from the unsigned bounds and
6216 * var_off of the result.
6218 dst_reg->smin_value = S64_MIN;
6219 dst_reg->smax_value = S64_MAX;
6220 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
6221 dst_reg->umin_value >>= umax_val;
6222 dst_reg->umax_value >>= umin_val;
6224 /* Its not easy to operate on alu32 bounds here because it depends
6225 * on bits being shifted in. Take easy way out and mark unbounded
6226 * so we can recalculate later from tnum.
6228 __mark_reg32_unbounded(dst_reg);
6229 __update_reg_bounds(dst_reg);
6232 static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg,
6233 struct bpf_reg_state *src_reg)
6235 u64 umin_val = src_reg->u32_min_value;
6237 /* Upon reaching here, src_known is true and
6238 * umax_val is equal to umin_val.
6240 dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val);
6241 dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val);
6243 dst_reg->var_off = tnum_arshift(tnum_subreg(dst_reg->var_off), umin_val, 32);
6245 /* blow away the dst_reg umin_value/umax_value and rely on
6246 * dst_reg var_off to refine the result.
6248 dst_reg->u32_min_value = 0;
6249 dst_reg->u32_max_value = U32_MAX;
6251 __mark_reg64_unbounded(dst_reg);
6252 __update_reg32_bounds(dst_reg);
6255 static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
6256 struct bpf_reg_state *src_reg)
6258 u64 umin_val = src_reg->umin_value;
6260 /* Upon reaching here, src_known is true and umax_val is equal
6263 dst_reg->smin_value >>= umin_val;
6264 dst_reg->smax_value >>= umin_val;
6266 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val, 64);
6268 /* blow away the dst_reg umin_value/umax_value and rely on
6269 * dst_reg var_off to refine the result.
6271 dst_reg->umin_value = 0;
6272 dst_reg->umax_value = U64_MAX;
6274 /* Its not easy to operate on alu32 bounds here because it depends
6275 * on bits being shifted in from upper 32-bits. Take easy way out
6276 * and mark unbounded so we can recalculate later from tnum.
6278 __mark_reg32_unbounded(dst_reg);
6279 __update_reg_bounds(dst_reg);
6282 /* WARNING: This function does calculations on 64-bit values, but the actual
6283 * execution may occur on 32-bit values. Therefore, things like bitshifts
6284 * need extra checks in the 32-bit case.
6286 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
6287 struct bpf_insn *insn,
6288 struct bpf_reg_state *dst_reg,
6289 struct bpf_reg_state src_reg)
6291 struct bpf_reg_state *regs = cur_regs(env);
6292 u8 opcode = BPF_OP(insn->code);
6294 s64 smin_val, smax_val;
6295 u64 umin_val, umax_val;
6296 s32 s32_min_val, s32_max_val;
6297 u32 u32_min_val, u32_max_val;
6298 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
6299 u32 dst = insn->dst_reg;
6301 bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
6303 smin_val = src_reg.smin_value;
6304 smax_val = src_reg.smax_value;
6305 umin_val = src_reg.umin_value;
6306 umax_val = src_reg.umax_value;
6308 s32_min_val = src_reg.s32_min_value;
6309 s32_max_val = src_reg.s32_max_value;
6310 u32_min_val = src_reg.u32_min_value;
6311 u32_max_val = src_reg.u32_max_value;
6314 src_known = tnum_subreg_is_const(src_reg.var_off);
6316 (s32_min_val != s32_max_val || u32_min_val != u32_max_val)) ||
6317 s32_min_val > s32_max_val || u32_min_val > u32_max_val) {
6318 /* Taint dst register if offset had invalid bounds
6319 * derived from e.g. dead branches.
6321 __mark_reg_unknown(env, dst_reg);
6325 src_known = tnum_is_const(src_reg.var_off);
6327 (smin_val != smax_val || umin_val != umax_val)) ||
6328 smin_val > smax_val || umin_val > umax_val) {
6329 /* Taint dst register if offset had invalid bounds
6330 * derived from e.g. dead branches.
6332 __mark_reg_unknown(env, dst_reg);
6338 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
6339 __mark_reg_unknown(env, dst_reg);
6343 /* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
6344 * There are two classes of instructions: The first class we track both
6345 * alu32 and alu64 sign/unsigned bounds independently this provides the
6346 * greatest amount of precision when alu operations are mixed with jmp32
6347 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
6348 * and BPF_OR. This is possible because these ops have fairly easy to
6349 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
6350 * See alu32 verifier tests for examples. The second class of
6351 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
6352 * with regards to tracking sign/unsigned bounds because the bits may
6353 * cross subreg boundaries in the alu64 case. When this happens we mark
6354 * the reg unbounded in the subreg bound space and use the resulting
6355 * tnum to calculate an approximation of the sign/unsigned bounds.
6359 ret = sanitize_val_alu(env, insn);
6361 verbose(env, "R%d tried to add from different pointers or scalars\n", dst);
6364 scalar32_min_max_add(dst_reg, &src_reg);
6365 scalar_min_max_add(dst_reg, &src_reg);
6366 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
6369 ret = sanitize_val_alu(env, insn);
6371 verbose(env, "R%d tried to sub from different pointers or scalars\n", dst);
6374 scalar32_min_max_sub(dst_reg, &src_reg);
6375 scalar_min_max_sub(dst_reg, &src_reg);
6376 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
6379 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
6380 scalar32_min_max_mul(dst_reg, &src_reg);
6381 scalar_min_max_mul(dst_reg, &src_reg);
6384 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
6385 scalar32_min_max_and(dst_reg, &src_reg);
6386 scalar_min_max_and(dst_reg, &src_reg);
6389 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
6390 scalar32_min_max_or(dst_reg, &src_reg);
6391 scalar_min_max_or(dst_reg, &src_reg);
6394 dst_reg->var_off = tnum_xor(dst_reg->var_off, src_reg.var_off);
6395 scalar32_min_max_xor(dst_reg, &src_reg);
6396 scalar_min_max_xor(dst_reg, &src_reg);
6399 if (umax_val >= insn_bitness) {
6400 /* Shifts greater than 31 or 63 are undefined.
6401 * This includes shifts by a negative number.
6403 mark_reg_unknown(env, regs, insn->dst_reg);
6407 scalar32_min_max_lsh(dst_reg, &src_reg);
6409 scalar_min_max_lsh(dst_reg, &src_reg);
6412 if (umax_val >= insn_bitness) {
6413 /* Shifts greater than 31 or 63 are undefined.
6414 * This includes shifts by a negative number.
6416 mark_reg_unknown(env, regs, insn->dst_reg);
6420 scalar32_min_max_rsh(dst_reg, &src_reg);
6422 scalar_min_max_rsh(dst_reg, &src_reg);
6425 if (umax_val >= insn_bitness) {
6426 /* Shifts greater than 31 or 63 are undefined.
6427 * This includes shifts by a negative number.
6429 mark_reg_unknown(env, regs, insn->dst_reg);
6433 scalar32_min_max_arsh(dst_reg, &src_reg);
6435 scalar_min_max_arsh(dst_reg, &src_reg);
6438 mark_reg_unknown(env, regs, insn->dst_reg);
6442 /* ALU32 ops are zero extended into 64bit register */
6444 zext_32_to_64(dst_reg);
6446 __update_reg_bounds(dst_reg);
6447 __reg_deduce_bounds(dst_reg);
6448 __reg_bound_offset(dst_reg);
6452 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
6455 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
6456 struct bpf_insn *insn)
6458 struct bpf_verifier_state *vstate = env->cur_state;
6459 struct bpf_func_state *state = vstate->frame[vstate->curframe];
6460 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
6461 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
6462 u8 opcode = BPF_OP(insn->code);
6465 dst_reg = ®s[insn->dst_reg];
6467 if (dst_reg->type != SCALAR_VALUE)
6470 /* Make sure ID is cleared otherwise dst_reg min/max could be
6471 * incorrectly propagated into other registers by find_equal_scalars()
6474 if (BPF_SRC(insn->code) == BPF_X) {
6475 src_reg = ®s[insn->src_reg];
6476 if (src_reg->type != SCALAR_VALUE) {
6477 if (dst_reg->type != SCALAR_VALUE) {
6478 /* Combining two pointers by any ALU op yields
6479 * an arbitrary scalar. Disallow all math except
6480 * pointer subtraction
6482 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
6483 mark_reg_unknown(env, regs, insn->dst_reg);
6486 verbose(env, "R%d pointer %s pointer prohibited\n",
6488 bpf_alu_string[opcode >> 4]);
6491 /* scalar += pointer
6492 * This is legal, but we have to reverse our
6493 * src/dest handling in computing the range
6495 err = mark_chain_precision(env, insn->dst_reg);
6498 return adjust_ptr_min_max_vals(env, insn,
6501 } else if (ptr_reg) {
6502 /* pointer += scalar */
6503 err = mark_chain_precision(env, insn->src_reg);
6506 return adjust_ptr_min_max_vals(env, insn,
6510 /* Pretend the src is a reg with a known value, since we only
6511 * need to be able to read from this state.
6513 off_reg.type = SCALAR_VALUE;
6514 __mark_reg_known(&off_reg, insn->imm);
6516 if (ptr_reg) /* pointer += K */
6517 return adjust_ptr_min_max_vals(env, insn,
6521 /* Got here implies adding two SCALAR_VALUEs */
6522 if (WARN_ON_ONCE(ptr_reg)) {
6523 print_verifier_state(env, state);
6524 verbose(env, "verifier internal error: unexpected ptr_reg\n");
6527 if (WARN_ON(!src_reg)) {
6528 print_verifier_state(env, state);
6529 verbose(env, "verifier internal error: no src_reg\n");
6532 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
6535 /* check validity of 32-bit and 64-bit arithmetic operations */
6536 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
6538 struct bpf_reg_state *regs = cur_regs(env);
6539 u8 opcode = BPF_OP(insn->code);
6542 if (opcode == BPF_END || opcode == BPF_NEG) {
6543 if (opcode == BPF_NEG) {
6544 if (BPF_SRC(insn->code) != 0 ||
6545 insn->src_reg != BPF_REG_0 ||
6546 insn->off != 0 || insn->imm != 0) {
6547 verbose(env, "BPF_NEG uses reserved fields\n");
6551 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
6552 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
6553 BPF_CLASS(insn->code) == BPF_ALU64) {
6554 verbose(env, "BPF_END uses reserved fields\n");
6559 /* check src operand */
6560 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
6564 if (is_pointer_value(env, insn->dst_reg)) {
6565 verbose(env, "R%d pointer arithmetic prohibited\n",
6570 /* check dest operand */
6571 err = check_reg_arg(env, insn->dst_reg, DST_OP);
6575 } else if (opcode == BPF_MOV) {
6577 if (BPF_SRC(insn->code) == BPF_X) {
6578 if (insn->imm != 0 || insn->off != 0) {
6579 verbose(env, "BPF_MOV uses reserved fields\n");
6583 /* check src operand */
6584 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6588 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
6589 verbose(env, "BPF_MOV uses reserved fields\n");
6594 /* check dest operand, mark as required later */
6595 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
6599 if (BPF_SRC(insn->code) == BPF_X) {
6600 struct bpf_reg_state *src_reg = regs + insn->src_reg;
6601 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
6603 if (BPF_CLASS(insn->code) == BPF_ALU64) {
6605 * copy register state to dest reg
6607 if (src_reg->type == SCALAR_VALUE && !src_reg->id)
6608 /* Assign src and dst registers the same ID
6609 * that will be used by find_equal_scalars()
6610 * to propagate min/max range.
6612 src_reg->id = ++env->id_gen;
6613 *dst_reg = *src_reg;
6614 dst_reg->live |= REG_LIVE_WRITTEN;
6615 dst_reg->subreg_def = DEF_NOT_SUBREG;
6618 if (is_pointer_value(env, insn->src_reg)) {
6620 "R%d partial copy of pointer\n",
6623 } else if (src_reg->type == SCALAR_VALUE) {
6624 *dst_reg = *src_reg;
6625 /* Make sure ID is cleared otherwise
6626 * dst_reg min/max could be incorrectly
6627 * propagated into src_reg by find_equal_scalars()
6630 dst_reg->live |= REG_LIVE_WRITTEN;
6631 dst_reg->subreg_def = env->insn_idx + 1;
6633 mark_reg_unknown(env, regs,
6636 zext_32_to_64(dst_reg);
6640 * remember the value we stored into this reg
6642 /* clear any state __mark_reg_known doesn't set */
6643 mark_reg_unknown(env, regs, insn->dst_reg);
6644 regs[insn->dst_reg].type = SCALAR_VALUE;
6645 if (BPF_CLASS(insn->code) == BPF_ALU64) {
6646 __mark_reg_known(regs + insn->dst_reg,
6649 __mark_reg_known(regs + insn->dst_reg,
6654 } else if (opcode > BPF_END) {
6655 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
6658 } else { /* all other ALU ops: and, sub, xor, add, ... */
6660 if (BPF_SRC(insn->code) == BPF_X) {
6661 if (insn->imm != 0 || insn->off != 0) {
6662 verbose(env, "BPF_ALU uses reserved fields\n");
6665 /* check src1 operand */
6666 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6670 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
6671 verbose(env, "BPF_ALU uses reserved fields\n");
6676 /* check src2 operand */
6677 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
6681 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
6682 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
6683 verbose(env, "div by zero\n");
6687 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
6688 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
6689 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
6691 if (insn->imm < 0 || insn->imm >= size) {
6692 verbose(env, "invalid shift %d\n", insn->imm);
6697 /* check dest operand */
6698 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
6702 return adjust_reg_min_max_vals(env, insn);
6708 static void __find_good_pkt_pointers(struct bpf_func_state *state,
6709 struct bpf_reg_state *dst_reg,
6710 enum bpf_reg_type type, u16 new_range)
6712 struct bpf_reg_state *reg;
6715 for (i = 0; i < MAX_BPF_REG; i++) {
6716 reg = &state->regs[i];
6717 if (reg->type == type && reg->id == dst_reg->id)
6718 /* keep the maximum range already checked */
6719 reg->range = max(reg->range, new_range);
6722 bpf_for_each_spilled_reg(i, state, reg) {
6725 if (reg->type == type && reg->id == dst_reg->id)
6726 reg->range = max(reg->range, new_range);
6730 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
6731 struct bpf_reg_state *dst_reg,
6732 enum bpf_reg_type type,
6733 bool range_right_open)
6738 if (dst_reg->off < 0 ||
6739 (dst_reg->off == 0 && range_right_open))
6740 /* This doesn't give us any range */
6743 if (dst_reg->umax_value > MAX_PACKET_OFF ||
6744 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
6745 /* Risk of overflow. For instance, ptr + (1<<63) may be less
6746 * than pkt_end, but that's because it's also less than pkt.
6750 new_range = dst_reg->off;
6751 if (range_right_open)
6754 /* Examples for register markings:
6756 * pkt_data in dst register:
6760 * if (r2 > pkt_end) goto <handle exception>
6765 * if (r2 < pkt_end) goto <access okay>
6766 * <handle exception>
6769 * r2 == dst_reg, pkt_end == src_reg
6770 * r2=pkt(id=n,off=8,r=0)
6771 * r3=pkt(id=n,off=0,r=0)
6773 * pkt_data in src register:
6777 * if (pkt_end >= r2) goto <access okay>
6778 * <handle exception>
6782 * if (pkt_end <= r2) goto <handle exception>
6786 * pkt_end == dst_reg, r2 == src_reg
6787 * r2=pkt(id=n,off=8,r=0)
6788 * r3=pkt(id=n,off=0,r=0)
6790 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
6791 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
6792 * and [r3, r3 + 8-1) respectively is safe to access depending on
6796 /* If our ids match, then we must have the same max_value. And we
6797 * don't care about the other reg's fixed offset, since if it's too big
6798 * the range won't allow anything.
6799 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
6801 for (i = 0; i <= vstate->curframe; i++)
6802 __find_good_pkt_pointers(vstate->frame[i], dst_reg, type,
6806 static int is_branch32_taken(struct bpf_reg_state *reg, u32 val, u8 opcode)
6808 struct tnum subreg = tnum_subreg(reg->var_off);
6809 s32 sval = (s32)val;
6813 if (tnum_is_const(subreg))
6814 return !!tnum_equals_const(subreg, val);
6817 if (tnum_is_const(subreg))
6818 return !tnum_equals_const(subreg, val);
6821 if ((~subreg.mask & subreg.value) & val)
6823 if (!((subreg.mask | subreg.value) & val))
6827 if (reg->u32_min_value > val)
6829 else if (reg->u32_max_value <= val)
6833 if (reg->s32_min_value > sval)
6835 else if (reg->s32_max_value <= sval)
6839 if (reg->u32_max_value < val)
6841 else if (reg->u32_min_value >= val)
6845 if (reg->s32_max_value < sval)
6847 else if (reg->s32_min_value >= sval)
6851 if (reg->u32_min_value >= val)
6853 else if (reg->u32_max_value < val)
6857 if (reg->s32_min_value >= sval)
6859 else if (reg->s32_max_value < sval)
6863 if (reg->u32_max_value <= val)
6865 else if (reg->u32_min_value > val)
6869 if (reg->s32_max_value <= sval)
6871 else if (reg->s32_min_value > sval)
6880 static int is_branch64_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
6882 s64 sval = (s64)val;
6886 if (tnum_is_const(reg->var_off))
6887 return !!tnum_equals_const(reg->var_off, val);
6890 if (tnum_is_const(reg->var_off))
6891 return !tnum_equals_const(reg->var_off, val);
6894 if ((~reg->var_off.mask & reg->var_off.value) & val)
6896 if (!((reg->var_off.mask | reg->var_off.value) & val))
6900 if (reg->umin_value > val)
6902 else if (reg->umax_value <= val)
6906 if (reg->smin_value > sval)
6908 else if (reg->smax_value <= sval)
6912 if (reg->umax_value < val)
6914 else if (reg->umin_value >= val)
6918 if (reg->smax_value < sval)
6920 else if (reg->smin_value >= sval)
6924 if (reg->umin_value >= val)
6926 else if (reg->umax_value < val)
6930 if (reg->smin_value >= sval)
6932 else if (reg->smax_value < sval)
6936 if (reg->umax_value <= val)
6938 else if (reg->umin_value > val)
6942 if (reg->smax_value <= sval)
6944 else if (reg->smin_value > sval)
6952 /* compute branch direction of the expression "if (reg opcode val) goto target;"
6954 * 1 - branch will be taken and "goto target" will be executed
6955 * 0 - branch will not be taken and fall-through to next insn
6956 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value
6959 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
6962 if (__is_pointer_value(false, reg)) {
6963 if (!reg_type_not_null(reg->type))
6966 /* If pointer is valid tests against zero will fail so we can
6967 * use this to direct branch taken.
6983 return is_branch32_taken(reg, val, opcode);
6984 return is_branch64_taken(reg, val, opcode);
6987 /* Adjusts the register min/max values in the case that the dst_reg is the
6988 * variable register that we are working on, and src_reg is a constant or we're
6989 * simply doing a BPF_K check.
6990 * In JEQ/JNE cases we also adjust the var_off values.
6992 static void reg_set_min_max(struct bpf_reg_state *true_reg,
6993 struct bpf_reg_state *false_reg,
6995 u8 opcode, bool is_jmp32)
6997 struct tnum false_32off = tnum_subreg(false_reg->var_off);
6998 struct tnum false_64off = false_reg->var_off;
6999 struct tnum true_32off = tnum_subreg(true_reg->var_off);
7000 struct tnum true_64off = true_reg->var_off;
7001 s64 sval = (s64)val;
7002 s32 sval32 = (s32)val32;
7004 /* If the dst_reg is a pointer, we can't learn anything about its
7005 * variable offset from the compare (unless src_reg were a pointer into
7006 * the same object, but we don't bother with that.
7007 * Since false_reg and true_reg have the same type by construction, we
7008 * only need to check one of them for pointerness.
7010 if (__is_pointer_value(false, false_reg))
7017 struct bpf_reg_state *reg =
7018 opcode == BPF_JEQ ? true_reg : false_reg;
7020 /* JEQ/JNE comparison doesn't change the register equivalence.
7022 * if (r1 == 42) goto label;
7024 * label: // here both r1 and r2 are known to be 42.
7026 * Hence when marking register as known preserve it's ID.
7029 __mark_reg32_known(reg, val32);
7031 ___mark_reg_known(reg, val);
7036 false_32off = tnum_and(false_32off, tnum_const(~val32));
7037 if (is_power_of_2(val32))
7038 true_32off = tnum_or(true_32off,
7041 false_64off = tnum_and(false_64off, tnum_const(~val));
7042 if (is_power_of_2(val))
7043 true_64off = tnum_or(true_64off,
7051 u32 false_umax = opcode == BPF_JGT ? val32 : val32 - 1;
7052 u32 true_umin = opcode == BPF_JGT ? val32 + 1 : val32;
7054 false_reg->u32_max_value = min(false_reg->u32_max_value,
7056 true_reg->u32_min_value = max(true_reg->u32_min_value,
7059 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
7060 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
7062 false_reg->umax_value = min(false_reg->umax_value, false_umax);
7063 true_reg->umin_value = max(true_reg->umin_value, true_umin);
7071 s32 false_smax = opcode == BPF_JSGT ? sval32 : sval32 - 1;
7072 s32 true_smin = opcode == BPF_JSGT ? sval32 + 1 : sval32;
7074 false_reg->s32_max_value = min(false_reg->s32_max_value, false_smax);
7075 true_reg->s32_min_value = max(true_reg->s32_min_value, true_smin);
7077 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
7078 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
7080 false_reg->smax_value = min(false_reg->smax_value, false_smax);
7081 true_reg->smin_value = max(true_reg->smin_value, true_smin);
7089 u32 false_umin = opcode == BPF_JLT ? val32 : val32 + 1;
7090 u32 true_umax = opcode == BPF_JLT ? val32 - 1 : val32;
7092 false_reg->u32_min_value = max(false_reg->u32_min_value,
7094 true_reg->u32_max_value = min(true_reg->u32_max_value,
7097 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
7098 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
7100 false_reg->umin_value = max(false_reg->umin_value, false_umin);
7101 true_reg->umax_value = min(true_reg->umax_value, true_umax);
7109 s32 false_smin = opcode == BPF_JSLT ? sval32 : sval32 + 1;
7110 s32 true_smax = opcode == BPF_JSLT ? sval32 - 1 : sval32;
7112 false_reg->s32_min_value = max(false_reg->s32_min_value, false_smin);
7113 true_reg->s32_max_value = min(true_reg->s32_max_value, true_smax);
7115 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
7116 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
7118 false_reg->smin_value = max(false_reg->smin_value, false_smin);
7119 true_reg->smax_value = min(true_reg->smax_value, true_smax);
7128 false_reg->var_off = tnum_or(tnum_clear_subreg(false_64off),
7129 tnum_subreg(false_32off));
7130 true_reg->var_off = tnum_or(tnum_clear_subreg(true_64off),
7131 tnum_subreg(true_32off));
7132 __reg_combine_32_into_64(false_reg);
7133 __reg_combine_32_into_64(true_reg);
7135 false_reg->var_off = false_64off;
7136 true_reg->var_off = true_64off;
7137 __reg_combine_64_into_32(false_reg);
7138 __reg_combine_64_into_32(true_reg);
7142 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
7145 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
7146 struct bpf_reg_state *false_reg,
7148 u8 opcode, bool is_jmp32)
7150 /* How can we transform "a <op> b" into "b <op> a"? */
7151 static const u8 opcode_flip[16] = {
7152 /* these stay the same */
7153 [BPF_JEQ >> 4] = BPF_JEQ,
7154 [BPF_JNE >> 4] = BPF_JNE,
7155 [BPF_JSET >> 4] = BPF_JSET,
7156 /* these swap "lesser" and "greater" (L and G in the opcodes) */
7157 [BPF_JGE >> 4] = BPF_JLE,
7158 [BPF_JGT >> 4] = BPF_JLT,
7159 [BPF_JLE >> 4] = BPF_JGE,
7160 [BPF_JLT >> 4] = BPF_JGT,
7161 [BPF_JSGE >> 4] = BPF_JSLE,
7162 [BPF_JSGT >> 4] = BPF_JSLT,
7163 [BPF_JSLE >> 4] = BPF_JSGE,
7164 [BPF_JSLT >> 4] = BPF_JSGT
7166 opcode = opcode_flip[opcode >> 4];
7167 /* This uses zero as "not present in table"; luckily the zero opcode,
7168 * BPF_JA, can't get here.
7171 reg_set_min_max(true_reg, false_reg, val, val32, opcode, is_jmp32);
7174 /* Regs are known to be equal, so intersect their min/max/var_off */
7175 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
7176 struct bpf_reg_state *dst_reg)
7178 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
7179 dst_reg->umin_value);
7180 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
7181 dst_reg->umax_value);
7182 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
7183 dst_reg->smin_value);
7184 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
7185 dst_reg->smax_value);
7186 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
7188 /* We might have learned new bounds from the var_off. */
7189 __update_reg_bounds(src_reg);
7190 __update_reg_bounds(dst_reg);
7191 /* We might have learned something about the sign bit. */
7192 __reg_deduce_bounds(src_reg);
7193 __reg_deduce_bounds(dst_reg);
7194 /* We might have learned some bits from the bounds. */
7195 __reg_bound_offset(src_reg);
7196 __reg_bound_offset(dst_reg);
7197 /* Intersecting with the old var_off might have improved our bounds
7198 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
7199 * then new var_off is (0; 0x7f...fc) which improves our umax.
7201 __update_reg_bounds(src_reg);
7202 __update_reg_bounds(dst_reg);
7205 static void reg_combine_min_max(struct bpf_reg_state *true_src,
7206 struct bpf_reg_state *true_dst,
7207 struct bpf_reg_state *false_src,
7208 struct bpf_reg_state *false_dst,
7213 __reg_combine_min_max(true_src, true_dst);
7216 __reg_combine_min_max(false_src, false_dst);
7221 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
7222 struct bpf_reg_state *reg, u32 id,
7225 if (reg_type_may_be_null(reg->type) && reg->id == id &&
7226 !WARN_ON_ONCE(!reg->id)) {
7227 /* Old offset (both fixed and variable parts) should
7228 * have been known-zero, because we don't allow pointer
7229 * arithmetic on pointers that might be NULL.
7231 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
7232 !tnum_equals_const(reg->var_off, 0) ||
7234 __mark_reg_known_zero(reg);
7238 reg->type = SCALAR_VALUE;
7239 } else if (reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
7240 const struct bpf_map *map = reg->map_ptr;
7242 if (map->inner_map_meta) {
7243 reg->type = CONST_PTR_TO_MAP;
7244 reg->map_ptr = map->inner_map_meta;
7245 } else if (map->map_type == BPF_MAP_TYPE_XSKMAP) {
7246 reg->type = PTR_TO_XDP_SOCK;
7247 } else if (map->map_type == BPF_MAP_TYPE_SOCKMAP ||
7248 map->map_type == BPF_MAP_TYPE_SOCKHASH) {
7249 reg->type = PTR_TO_SOCKET;
7251 reg->type = PTR_TO_MAP_VALUE;
7253 } else if (reg->type == PTR_TO_SOCKET_OR_NULL) {
7254 reg->type = PTR_TO_SOCKET;
7255 } else if (reg->type == PTR_TO_SOCK_COMMON_OR_NULL) {
7256 reg->type = PTR_TO_SOCK_COMMON;
7257 } else if (reg->type == PTR_TO_TCP_SOCK_OR_NULL) {
7258 reg->type = PTR_TO_TCP_SOCK;
7259 } else if (reg->type == PTR_TO_BTF_ID_OR_NULL) {
7260 reg->type = PTR_TO_BTF_ID;
7261 } else if (reg->type == PTR_TO_MEM_OR_NULL) {
7262 reg->type = PTR_TO_MEM;
7263 } else if (reg->type == PTR_TO_RDONLY_BUF_OR_NULL) {
7264 reg->type = PTR_TO_RDONLY_BUF;
7265 } else if (reg->type == PTR_TO_RDWR_BUF_OR_NULL) {
7266 reg->type = PTR_TO_RDWR_BUF;
7269 /* We don't need id and ref_obj_id from this point
7270 * onwards anymore, thus we should better reset it,
7271 * so that state pruning has chances to take effect.
7274 reg->ref_obj_id = 0;
7275 } else if (!reg_may_point_to_spin_lock(reg)) {
7276 /* For not-NULL ptr, reg->ref_obj_id will be reset
7277 * in release_reg_references().
7279 * reg->id is still used by spin_lock ptr. Other
7280 * than spin_lock ptr type, reg->id can be reset.
7287 static void __mark_ptr_or_null_regs(struct bpf_func_state *state, u32 id,
7290 struct bpf_reg_state *reg;
7293 for (i = 0; i < MAX_BPF_REG; i++)
7294 mark_ptr_or_null_reg(state, &state->regs[i], id, is_null);
7296 bpf_for_each_spilled_reg(i, state, reg) {
7299 mark_ptr_or_null_reg(state, reg, id, is_null);
7303 /* The logic is similar to find_good_pkt_pointers(), both could eventually
7304 * be folded together at some point.
7306 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
7309 struct bpf_func_state *state = vstate->frame[vstate->curframe];
7310 struct bpf_reg_state *regs = state->regs;
7311 u32 ref_obj_id = regs[regno].ref_obj_id;
7312 u32 id = regs[regno].id;
7315 if (ref_obj_id && ref_obj_id == id && is_null)
7316 /* regs[regno] is in the " == NULL" branch.
7317 * No one could have freed the reference state before
7318 * doing the NULL check.
7320 WARN_ON_ONCE(release_reference_state(state, id));
7322 for (i = 0; i <= vstate->curframe; i++)
7323 __mark_ptr_or_null_regs(vstate->frame[i], id, is_null);
7326 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
7327 struct bpf_reg_state *dst_reg,
7328 struct bpf_reg_state *src_reg,
7329 struct bpf_verifier_state *this_branch,
7330 struct bpf_verifier_state *other_branch)
7332 if (BPF_SRC(insn->code) != BPF_X)
7335 /* Pointers are always 64-bit. */
7336 if (BPF_CLASS(insn->code) == BPF_JMP32)
7339 switch (BPF_OP(insn->code)) {
7341 if ((dst_reg->type == PTR_TO_PACKET &&
7342 src_reg->type == PTR_TO_PACKET_END) ||
7343 (dst_reg->type == PTR_TO_PACKET_META &&
7344 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
7345 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
7346 find_good_pkt_pointers(this_branch, dst_reg,
7347 dst_reg->type, false);
7348 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
7349 src_reg->type == PTR_TO_PACKET) ||
7350 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
7351 src_reg->type == PTR_TO_PACKET_META)) {
7352 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
7353 find_good_pkt_pointers(other_branch, src_reg,
7354 src_reg->type, true);
7360 if ((dst_reg->type == PTR_TO_PACKET &&
7361 src_reg->type == PTR_TO_PACKET_END) ||
7362 (dst_reg->type == PTR_TO_PACKET_META &&
7363 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
7364 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
7365 find_good_pkt_pointers(other_branch, dst_reg,
7366 dst_reg->type, true);
7367 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
7368 src_reg->type == PTR_TO_PACKET) ||
7369 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
7370 src_reg->type == PTR_TO_PACKET_META)) {
7371 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
7372 find_good_pkt_pointers(this_branch, src_reg,
7373 src_reg->type, false);
7379 if ((dst_reg->type == PTR_TO_PACKET &&
7380 src_reg->type == PTR_TO_PACKET_END) ||
7381 (dst_reg->type == PTR_TO_PACKET_META &&
7382 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
7383 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
7384 find_good_pkt_pointers(this_branch, dst_reg,
7385 dst_reg->type, true);
7386 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
7387 src_reg->type == PTR_TO_PACKET) ||
7388 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
7389 src_reg->type == PTR_TO_PACKET_META)) {
7390 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
7391 find_good_pkt_pointers(other_branch, src_reg,
7392 src_reg->type, false);
7398 if ((dst_reg->type == PTR_TO_PACKET &&
7399 src_reg->type == PTR_TO_PACKET_END) ||
7400 (dst_reg->type == PTR_TO_PACKET_META &&
7401 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
7402 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
7403 find_good_pkt_pointers(other_branch, dst_reg,
7404 dst_reg->type, false);
7405 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
7406 src_reg->type == PTR_TO_PACKET) ||
7407 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
7408 src_reg->type == PTR_TO_PACKET_META)) {
7409 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
7410 find_good_pkt_pointers(this_branch, src_reg,
7411 src_reg->type, true);
7423 static void find_equal_scalars(struct bpf_verifier_state *vstate,
7424 struct bpf_reg_state *known_reg)
7426 struct bpf_func_state *state;
7427 struct bpf_reg_state *reg;
7430 for (i = 0; i <= vstate->curframe; i++) {
7431 state = vstate->frame[i];
7432 for (j = 0; j < MAX_BPF_REG; j++) {
7433 reg = &state->regs[j];
7434 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
7438 bpf_for_each_spilled_reg(j, state, reg) {
7441 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
7447 static int check_cond_jmp_op(struct bpf_verifier_env *env,
7448 struct bpf_insn *insn, int *insn_idx)
7450 struct bpf_verifier_state *this_branch = env->cur_state;
7451 struct bpf_verifier_state *other_branch;
7452 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
7453 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
7454 u8 opcode = BPF_OP(insn->code);
7459 /* Only conditional jumps are expected to reach here. */
7460 if (opcode == BPF_JA || opcode > BPF_JSLE) {
7461 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
7465 if (BPF_SRC(insn->code) == BPF_X) {
7466 if (insn->imm != 0) {
7467 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
7471 /* check src1 operand */
7472 err = check_reg_arg(env, insn->src_reg, SRC_OP);
7476 if (is_pointer_value(env, insn->src_reg)) {
7477 verbose(env, "R%d pointer comparison prohibited\n",
7481 src_reg = ®s[insn->src_reg];
7483 if (insn->src_reg != BPF_REG_0) {
7484 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
7489 /* check src2 operand */
7490 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
7494 dst_reg = ®s[insn->dst_reg];
7495 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
7497 if (BPF_SRC(insn->code) == BPF_K) {
7498 pred = is_branch_taken(dst_reg, insn->imm, opcode, is_jmp32);
7499 } else if (src_reg->type == SCALAR_VALUE &&
7500 is_jmp32 && tnum_is_const(tnum_subreg(src_reg->var_off))) {
7501 pred = is_branch_taken(dst_reg,
7502 tnum_subreg(src_reg->var_off).value,
7505 } else if (src_reg->type == SCALAR_VALUE &&
7506 !is_jmp32 && tnum_is_const(src_reg->var_off)) {
7507 pred = is_branch_taken(dst_reg,
7508 src_reg->var_off.value,
7514 /* If we get here with a dst_reg pointer type it is because
7515 * above is_branch_taken() special cased the 0 comparison.
7517 if (!__is_pointer_value(false, dst_reg))
7518 err = mark_chain_precision(env, insn->dst_reg);
7519 if (BPF_SRC(insn->code) == BPF_X && !err)
7520 err = mark_chain_precision(env, insn->src_reg);
7525 /* only follow the goto, ignore fall-through */
7526 *insn_idx += insn->off;
7528 } else if (pred == 0) {
7529 /* only follow fall-through branch, since
7530 * that's where the program will go
7535 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
7539 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
7541 /* detect if we are comparing against a constant value so we can adjust
7542 * our min/max values for our dst register.
7543 * this is only legit if both are scalars (or pointers to the same
7544 * object, I suppose, but we don't support that right now), because
7545 * otherwise the different base pointers mean the offsets aren't
7548 if (BPF_SRC(insn->code) == BPF_X) {
7549 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
7551 if (dst_reg->type == SCALAR_VALUE &&
7552 src_reg->type == SCALAR_VALUE) {
7553 if (tnum_is_const(src_reg->var_off) ||
7555 tnum_is_const(tnum_subreg(src_reg->var_off))))
7556 reg_set_min_max(&other_branch_regs[insn->dst_reg],
7558 src_reg->var_off.value,
7559 tnum_subreg(src_reg->var_off).value,
7561 else if (tnum_is_const(dst_reg->var_off) ||
7563 tnum_is_const(tnum_subreg(dst_reg->var_off))))
7564 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
7566 dst_reg->var_off.value,
7567 tnum_subreg(dst_reg->var_off).value,
7569 else if (!is_jmp32 &&
7570 (opcode == BPF_JEQ || opcode == BPF_JNE))
7571 /* Comparing for equality, we can combine knowledge */
7572 reg_combine_min_max(&other_branch_regs[insn->src_reg],
7573 &other_branch_regs[insn->dst_reg],
7574 src_reg, dst_reg, opcode);
7576 !WARN_ON_ONCE(src_reg->id != other_branch_regs[insn->src_reg].id)) {
7577 find_equal_scalars(this_branch, src_reg);
7578 find_equal_scalars(other_branch, &other_branch_regs[insn->src_reg]);
7582 } else if (dst_reg->type == SCALAR_VALUE) {
7583 reg_set_min_max(&other_branch_regs[insn->dst_reg],
7584 dst_reg, insn->imm, (u32)insn->imm,
7588 if (dst_reg->type == SCALAR_VALUE && dst_reg->id &&
7589 !WARN_ON_ONCE(dst_reg->id != other_branch_regs[insn->dst_reg].id)) {
7590 find_equal_scalars(this_branch, dst_reg);
7591 find_equal_scalars(other_branch, &other_branch_regs[insn->dst_reg]);
7594 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
7595 * NOTE: these optimizations below are related with pointer comparison
7596 * which will never be JMP32.
7598 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
7599 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
7600 reg_type_may_be_null(dst_reg->type)) {
7601 /* Mark all identical registers in each branch as either
7602 * safe or unknown depending R == 0 or R != 0 conditional.
7604 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
7606 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
7608 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
7609 this_branch, other_branch) &&
7610 is_pointer_value(env, insn->dst_reg)) {
7611 verbose(env, "R%d pointer comparison prohibited\n",
7615 if (env->log.level & BPF_LOG_LEVEL)
7616 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
7620 /* verify BPF_LD_IMM64 instruction */
7621 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
7623 struct bpf_insn_aux_data *aux = cur_aux(env);
7624 struct bpf_reg_state *regs = cur_regs(env);
7625 struct bpf_reg_state *dst_reg;
7626 struct bpf_map *map;
7629 if (BPF_SIZE(insn->code) != BPF_DW) {
7630 verbose(env, "invalid BPF_LD_IMM insn\n");
7633 if (insn->off != 0) {
7634 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
7638 err = check_reg_arg(env, insn->dst_reg, DST_OP);
7642 dst_reg = ®s[insn->dst_reg];
7643 if (insn->src_reg == 0) {
7644 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
7646 dst_reg->type = SCALAR_VALUE;
7647 __mark_reg_known(®s[insn->dst_reg], imm);
7651 if (insn->src_reg == BPF_PSEUDO_BTF_ID) {
7652 mark_reg_known_zero(env, regs, insn->dst_reg);
7654 dst_reg->type = aux->btf_var.reg_type;
7655 switch (dst_reg->type) {
7657 dst_reg->mem_size = aux->btf_var.mem_size;
7660 case PTR_TO_PERCPU_BTF_ID:
7661 dst_reg->btf_id = aux->btf_var.btf_id;
7664 verbose(env, "bpf verifier is misconfigured\n");
7670 map = env->used_maps[aux->map_index];
7671 mark_reg_known_zero(env, regs, insn->dst_reg);
7672 dst_reg->map_ptr = map;
7674 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE) {
7675 dst_reg->type = PTR_TO_MAP_VALUE;
7676 dst_reg->off = aux->map_off;
7677 if (map_value_has_spin_lock(map))
7678 dst_reg->id = ++env->id_gen;
7679 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD) {
7680 dst_reg->type = CONST_PTR_TO_MAP;
7682 verbose(env, "bpf verifier is misconfigured\n");
7689 static bool may_access_skb(enum bpf_prog_type type)
7692 case BPF_PROG_TYPE_SOCKET_FILTER:
7693 case BPF_PROG_TYPE_SCHED_CLS:
7694 case BPF_PROG_TYPE_SCHED_ACT:
7701 /* verify safety of LD_ABS|LD_IND instructions:
7702 * - they can only appear in the programs where ctx == skb
7703 * - since they are wrappers of function calls, they scratch R1-R5 registers,
7704 * preserve R6-R9, and store return value into R0
7707 * ctx == skb == R6 == CTX
7710 * SRC == any register
7711 * IMM == 32-bit immediate
7714 * R0 - 8/16/32-bit skb data converted to cpu endianness
7716 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
7718 struct bpf_reg_state *regs = cur_regs(env);
7719 static const int ctx_reg = BPF_REG_6;
7720 u8 mode = BPF_MODE(insn->code);
7723 if (!may_access_skb(resolve_prog_type(env->prog))) {
7724 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
7728 if (!env->ops->gen_ld_abs) {
7729 verbose(env, "bpf verifier is misconfigured\n");
7733 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
7734 BPF_SIZE(insn->code) == BPF_DW ||
7735 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
7736 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
7740 /* check whether implicit source operand (register R6) is readable */
7741 err = check_reg_arg(env, ctx_reg, SRC_OP);
7745 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
7746 * gen_ld_abs() may terminate the program at runtime, leading to
7749 err = check_reference_leak(env);
7751 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
7755 if (env->cur_state->active_spin_lock) {
7756 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
7760 if (regs[ctx_reg].type != PTR_TO_CTX) {
7762 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
7766 if (mode == BPF_IND) {
7767 /* check explicit source operand */
7768 err = check_reg_arg(env, insn->src_reg, SRC_OP);
7773 err = check_ctx_reg(env, ®s[ctx_reg], ctx_reg);
7777 /* reset caller saved regs to unreadable */
7778 for (i = 0; i < CALLER_SAVED_REGS; i++) {
7779 mark_reg_not_init(env, regs, caller_saved[i]);
7780 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
7783 /* mark destination R0 register as readable, since it contains
7784 * the value fetched from the packet.
7785 * Already marked as written above.
7787 mark_reg_unknown(env, regs, BPF_REG_0);
7788 /* ld_abs load up to 32-bit skb data. */
7789 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
7793 static int check_return_code(struct bpf_verifier_env *env)
7795 struct tnum enforce_attach_type_range = tnum_unknown;
7796 const struct bpf_prog *prog = env->prog;
7797 struct bpf_reg_state *reg;
7798 struct tnum range = tnum_range(0, 1);
7799 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
7801 const bool is_subprog = env->cur_state->frame[0]->subprogno;
7803 /* LSM and struct_ops func-ptr's return type could be "void" */
7805 (prog_type == BPF_PROG_TYPE_STRUCT_OPS ||
7806 prog_type == BPF_PROG_TYPE_LSM) &&
7807 !prog->aux->attach_func_proto->type)
7810 /* eBPF calling convetion is such that R0 is used
7811 * to return the value from eBPF program.
7812 * Make sure that it's readable at this time
7813 * of bpf_exit, which means that program wrote
7814 * something into it earlier
7816 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
7820 if (is_pointer_value(env, BPF_REG_0)) {
7821 verbose(env, "R0 leaks addr as return value\n");
7825 reg = cur_regs(env) + BPF_REG_0;
7827 if (reg->type != SCALAR_VALUE) {
7828 verbose(env, "At subprogram exit the register R0 is not a scalar value (%s)\n",
7829 reg_type_str[reg->type]);
7835 switch (prog_type) {
7836 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
7837 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
7838 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG ||
7839 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETPEERNAME ||
7840 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETPEERNAME ||
7841 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETSOCKNAME ||
7842 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETSOCKNAME)
7843 range = tnum_range(1, 1);
7845 case BPF_PROG_TYPE_CGROUP_SKB:
7846 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
7847 range = tnum_range(0, 3);
7848 enforce_attach_type_range = tnum_range(2, 3);
7851 case BPF_PROG_TYPE_CGROUP_SOCK:
7852 case BPF_PROG_TYPE_SOCK_OPS:
7853 case BPF_PROG_TYPE_CGROUP_DEVICE:
7854 case BPF_PROG_TYPE_CGROUP_SYSCTL:
7855 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
7857 case BPF_PROG_TYPE_RAW_TRACEPOINT:
7858 if (!env->prog->aux->attach_btf_id)
7860 range = tnum_const(0);
7862 case BPF_PROG_TYPE_TRACING:
7863 switch (env->prog->expected_attach_type) {
7864 case BPF_TRACE_FENTRY:
7865 case BPF_TRACE_FEXIT:
7866 range = tnum_const(0);
7868 case BPF_TRACE_RAW_TP:
7869 case BPF_MODIFY_RETURN:
7871 case BPF_TRACE_ITER:
7877 case BPF_PROG_TYPE_SK_LOOKUP:
7878 range = tnum_range(SK_DROP, SK_PASS);
7880 case BPF_PROG_TYPE_EXT:
7881 /* freplace program can return anything as its return value
7882 * depends on the to-be-replaced kernel func or bpf program.
7888 if (reg->type != SCALAR_VALUE) {
7889 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
7890 reg_type_str[reg->type]);
7894 if (!tnum_in(range, reg->var_off)) {
7897 verbose(env, "At program exit the register R0 ");
7898 if (!tnum_is_unknown(reg->var_off)) {
7899 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
7900 verbose(env, "has value %s", tn_buf);
7902 verbose(env, "has unknown scalar value");
7904 tnum_strn(tn_buf, sizeof(tn_buf), range);
7905 verbose(env, " should have been in %s\n", tn_buf);
7909 if (!tnum_is_unknown(enforce_attach_type_range) &&
7910 tnum_in(enforce_attach_type_range, reg->var_off))
7911 env->prog->enforce_expected_attach_type = 1;
7915 /* non-recursive DFS pseudo code
7916 * 1 procedure DFS-iterative(G,v):
7917 * 2 label v as discovered
7918 * 3 let S be a stack
7920 * 5 while S is not empty
7922 * 7 if t is what we're looking for:
7924 * 9 for all edges e in G.adjacentEdges(t) do
7925 * 10 if edge e is already labelled
7926 * 11 continue with the next edge
7927 * 12 w <- G.adjacentVertex(t,e)
7928 * 13 if vertex w is not discovered and not explored
7929 * 14 label e as tree-edge
7930 * 15 label w as discovered
7933 * 18 else if vertex w is discovered
7934 * 19 label e as back-edge
7936 * 21 // vertex w is explored
7937 * 22 label e as forward- or cross-edge
7938 * 23 label t as explored
7943 * 0x11 - discovered and fall-through edge labelled
7944 * 0x12 - discovered and fall-through and branch edges labelled
7955 static u32 state_htab_size(struct bpf_verifier_env *env)
7957 return env->prog->len;
7960 static struct bpf_verifier_state_list **explored_state(
7961 struct bpf_verifier_env *env,
7964 struct bpf_verifier_state *cur = env->cur_state;
7965 struct bpf_func_state *state = cur->frame[cur->curframe];
7967 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
7970 static void init_explored_state(struct bpf_verifier_env *env, int idx)
7972 env->insn_aux_data[idx].prune_point = true;
7975 /* t, w, e - match pseudo-code above:
7976 * t - index of current instruction
7977 * w - next instruction
7980 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env,
7983 int *insn_stack = env->cfg.insn_stack;
7984 int *insn_state = env->cfg.insn_state;
7986 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
7989 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
7992 if (w < 0 || w >= env->prog->len) {
7993 verbose_linfo(env, t, "%d: ", t);
7994 verbose(env, "jump out of range from insn %d to %d\n", t, w);
7999 /* mark branch target for state pruning */
8000 init_explored_state(env, w);
8002 if (insn_state[w] == 0) {
8004 insn_state[t] = DISCOVERED | e;
8005 insn_state[w] = DISCOVERED;
8006 if (env->cfg.cur_stack >= env->prog->len)
8008 insn_stack[env->cfg.cur_stack++] = w;
8010 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
8011 if (loop_ok && env->bpf_capable)
8013 verbose_linfo(env, t, "%d: ", t);
8014 verbose_linfo(env, w, "%d: ", w);
8015 verbose(env, "back-edge from insn %d to %d\n", t, w);
8017 } else if (insn_state[w] == EXPLORED) {
8018 /* forward- or cross-edge */
8019 insn_state[t] = DISCOVERED | e;
8021 verbose(env, "insn state internal bug\n");
8027 /* non-recursive depth-first-search to detect loops in BPF program
8028 * loop == back-edge in directed graph
8030 static int check_cfg(struct bpf_verifier_env *env)
8032 struct bpf_insn *insns = env->prog->insnsi;
8033 int insn_cnt = env->prog->len;
8034 int *insn_stack, *insn_state;
8038 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
8042 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
8048 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
8049 insn_stack[0] = 0; /* 0 is the first instruction */
8050 env->cfg.cur_stack = 1;
8053 if (env->cfg.cur_stack == 0)
8055 t = insn_stack[env->cfg.cur_stack - 1];
8057 if (BPF_CLASS(insns[t].code) == BPF_JMP ||
8058 BPF_CLASS(insns[t].code) == BPF_JMP32) {
8059 u8 opcode = BPF_OP(insns[t].code);
8061 if (opcode == BPF_EXIT) {
8063 } else if (opcode == BPF_CALL) {
8064 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
8069 if (t + 1 < insn_cnt)
8070 init_explored_state(env, t + 1);
8071 if (insns[t].src_reg == BPF_PSEUDO_CALL) {
8072 init_explored_state(env, t);
8073 ret = push_insn(t, t + insns[t].imm + 1, BRANCH,
8080 } else if (opcode == BPF_JA) {
8081 if (BPF_SRC(insns[t].code) != BPF_K) {
8085 /* unconditional jump with single edge */
8086 ret = push_insn(t, t + insns[t].off + 1,
8087 FALLTHROUGH, env, true);
8092 /* unconditional jmp is not a good pruning point,
8093 * but it's marked, since backtracking needs
8094 * to record jmp history in is_state_visited().
8096 init_explored_state(env, t + insns[t].off + 1);
8097 /* tell verifier to check for equivalent states
8098 * after every call and jump
8100 if (t + 1 < insn_cnt)
8101 init_explored_state(env, t + 1);
8103 /* conditional jump with two edges */
8104 init_explored_state(env, t);
8105 ret = push_insn(t, t + 1, FALLTHROUGH, env, true);
8111 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env, true);
8118 /* all other non-branch instructions with single
8121 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
8129 insn_state[t] = EXPLORED;
8130 if (env->cfg.cur_stack-- <= 0) {
8131 verbose(env, "pop stack internal bug\n");
8138 for (i = 0; i < insn_cnt; i++) {
8139 if (insn_state[i] != EXPLORED) {
8140 verbose(env, "unreachable insn %d\n", i);
8145 ret = 0; /* cfg looks good */
8150 env->cfg.insn_state = env->cfg.insn_stack = NULL;
8154 static int check_abnormal_return(struct bpf_verifier_env *env)
8158 for (i = 1; i < env->subprog_cnt; i++) {
8159 if (env->subprog_info[i].has_ld_abs) {
8160 verbose(env, "LD_ABS is not allowed in subprogs without BTF\n");
8163 if (env->subprog_info[i].has_tail_call) {
8164 verbose(env, "tail_call is not allowed in subprogs without BTF\n");
8171 /* The minimum supported BTF func info size */
8172 #define MIN_BPF_FUNCINFO_SIZE 8
8173 #define MAX_FUNCINFO_REC_SIZE 252
8175 static int check_btf_func(struct bpf_verifier_env *env,
8176 const union bpf_attr *attr,
8177 union bpf_attr __user *uattr)
8179 const struct btf_type *type, *func_proto, *ret_type;
8180 u32 i, nfuncs, urec_size, min_size;
8181 u32 krec_size = sizeof(struct bpf_func_info);
8182 struct bpf_func_info *krecord;
8183 struct bpf_func_info_aux *info_aux = NULL;
8184 struct bpf_prog *prog;
8185 const struct btf *btf;
8186 void __user *urecord;
8187 u32 prev_offset = 0;
8191 nfuncs = attr->func_info_cnt;
8193 if (check_abnormal_return(env))
8198 if (nfuncs != env->subprog_cnt) {
8199 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
8203 urec_size = attr->func_info_rec_size;
8204 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
8205 urec_size > MAX_FUNCINFO_REC_SIZE ||
8206 urec_size % sizeof(u32)) {
8207 verbose(env, "invalid func info rec size %u\n", urec_size);
8212 btf = prog->aux->btf;
8214 urecord = u64_to_user_ptr(attr->func_info);
8215 min_size = min_t(u32, krec_size, urec_size);
8217 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
8220 info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN);
8224 for (i = 0; i < nfuncs; i++) {
8225 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
8227 if (ret == -E2BIG) {
8228 verbose(env, "nonzero tailing record in func info");
8229 /* set the size kernel expects so loader can zero
8230 * out the rest of the record.
8232 if (put_user(min_size, &uattr->func_info_rec_size))
8238 if (copy_from_user(&krecord[i], urecord, min_size)) {
8243 /* check insn_off */
8246 if (krecord[i].insn_off) {
8248 "nonzero insn_off %u for the first func info record",
8249 krecord[i].insn_off);
8252 } else if (krecord[i].insn_off <= prev_offset) {
8254 "same or smaller insn offset (%u) than previous func info record (%u)",
8255 krecord[i].insn_off, prev_offset);
8259 if (env->subprog_info[i].start != krecord[i].insn_off) {
8260 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
8265 type = btf_type_by_id(btf, krecord[i].type_id);
8266 if (!type || !btf_type_is_func(type)) {
8267 verbose(env, "invalid type id %d in func info",
8268 krecord[i].type_id);
8271 info_aux[i].linkage = BTF_INFO_VLEN(type->info);
8273 func_proto = btf_type_by_id(btf, type->type);
8274 if (unlikely(!func_proto || !btf_type_is_func_proto(func_proto)))
8275 /* btf_func_check() already verified it during BTF load */
8277 ret_type = btf_type_skip_modifiers(btf, func_proto->type, NULL);
8279 btf_type_is_small_int(ret_type) || btf_type_is_enum(ret_type);
8280 if (i && !scalar_return && env->subprog_info[i].has_ld_abs) {
8281 verbose(env, "LD_ABS is only allowed in functions that return 'int'.\n");
8284 if (i && !scalar_return && env->subprog_info[i].has_tail_call) {
8285 verbose(env, "tail_call is only allowed in functions that return 'int'.\n");
8289 prev_offset = krecord[i].insn_off;
8290 urecord += urec_size;
8293 prog->aux->func_info = krecord;
8294 prog->aux->func_info_cnt = nfuncs;
8295 prog->aux->func_info_aux = info_aux;
8304 static void adjust_btf_func(struct bpf_verifier_env *env)
8306 struct bpf_prog_aux *aux = env->prog->aux;
8309 if (!aux->func_info)
8312 for (i = 0; i < env->subprog_cnt; i++)
8313 aux->func_info[i].insn_off = env->subprog_info[i].start;
8316 #define MIN_BPF_LINEINFO_SIZE (offsetof(struct bpf_line_info, line_col) + \
8317 sizeof(((struct bpf_line_info *)(0))->line_col))
8318 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
8320 static int check_btf_line(struct bpf_verifier_env *env,
8321 const union bpf_attr *attr,
8322 union bpf_attr __user *uattr)
8324 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
8325 struct bpf_subprog_info *sub;
8326 struct bpf_line_info *linfo;
8327 struct bpf_prog *prog;
8328 const struct btf *btf;
8329 void __user *ulinfo;
8332 nr_linfo = attr->line_info_cnt;
8336 rec_size = attr->line_info_rec_size;
8337 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
8338 rec_size > MAX_LINEINFO_REC_SIZE ||
8339 rec_size & (sizeof(u32) - 1))
8342 /* Need to zero it in case the userspace may
8343 * pass in a smaller bpf_line_info object.
8345 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
8346 GFP_KERNEL | __GFP_NOWARN);
8351 btf = prog->aux->btf;
8354 sub = env->subprog_info;
8355 ulinfo = u64_to_user_ptr(attr->line_info);
8356 expected_size = sizeof(struct bpf_line_info);
8357 ncopy = min_t(u32, expected_size, rec_size);
8358 for (i = 0; i < nr_linfo; i++) {
8359 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
8361 if (err == -E2BIG) {
8362 verbose(env, "nonzero tailing record in line_info");
8363 if (put_user(expected_size,
8364 &uattr->line_info_rec_size))
8370 if (copy_from_user(&linfo[i], ulinfo, ncopy)) {
8376 * Check insn_off to ensure
8377 * 1) strictly increasing AND
8378 * 2) bounded by prog->len
8380 * The linfo[0].insn_off == 0 check logically falls into
8381 * the later "missing bpf_line_info for func..." case
8382 * because the first linfo[0].insn_off must be the
8383 * first sub also and the first sub must have
8384 * subprog_info[0].start == 0.
8386 if ((i && linfo[i].insn_off <= prev_offset) ||
8387 linfo[i].insn_off >= prog->len) {
8388 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
8389 i, linfo[i].insn_off, prev_offset,
8395 if (!prog->insnsi[linfo[i].insn_off].code) {
8397 "Invalid insn code at line_info[%u].insn_off\n",
8403 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
8404 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
8405 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
8410 if (s != env->subprog_cnt) {
8411 if (linfo[i].insn_off == sub[s].start) {
8412 sub[s].linfo_idx = i;
8414 } else if (sub[s].start < linfo[i].insn_off) {
8415 verbose(env, "missing bpf_line_info for func#%u\n", s);
8421 prev_offset = linfo[i].insn_off;
8425 if (s != env->subprog_cnt) {
8426 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
8427 env->subprog_cnt - s, s);
8432 prog->aux->linfo = linfo;
8433 prog->aux->nr_linfo = nr_linfo;
8442 static int check_btf_info(struct bpf_verifier_env *env,
8443 const union bpf_attr *attr,
8444 union bpf_attr __user *uattr)
8449 if (!attr->func_info_cnt && !attr->line_info_cnt) {
8450 if (check_abnormal_return(env))
8455 btf = btf_get_by_fd(attr->prog_btf_fd);
8457 return PTR_ERR(btf);
8458 env->prog->aux->btf = btf;
8460 err = check_btf_func(env, attr, uattr);
8464 err = check_btf_line(env, attr, uattr);
8471 /* check %cur's range satisfies %old's */
8472 static bool range_within(struct bpf_reg_state *old,
8473 struct bpf_reg_state *cur)
8475 return old->umin_value <= cur->umin_value &&
8476 old->umax_value >= cur->umax_value &&
8477 old->smin_value <= cur->smin_value &&
8478 old->smax_value >= cur->smax_value &&
8479 old->u32_min_value <= cur->u32_min_value &&
8480 old->u32_max_value >= cur->u32_max_value &&
8481 old->s32_min_value <= cur->s32_min_value &&
8482 old->s32_max_value >= cur->s32_max_value;
8485 /* Maximum number of register states that can exist at once */
8486 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
8492 /* If in the old state two registers had the same id, then they need to have
8493 * the same id in the new state as well. But that id could be different from
8494 * the old state, so we need to track the mapping from old to new ids.
8495 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
8496 * regs with old id 5 must also have new id 9 for the new state to be safe. But
8497 * regs with a different old id could still have new id 9, we don't care about
8499 * So we look through our idmap to see if this old id has been seen before. If
8500 * so, we require the new id to match; otherwise, we add the id pair to the map.
8502 static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
8506 for (i = 0; i < ID_MAP_SIZE; i++) {
8507 if (!idmap[i].old) {
8508 /* Reached an empty slot; haven't seen this id before */
8509 idmap[i].old = old_id;
8510 idmap[i].cur = cur_id;
8513 if (idmap[i].old == old_id)
8514 return idmap[i].cur == cur_id;
8516 /* We ran out of idmap slots, which should be impossible */
8521 static void clean_func_state(struct bpf_verifier_env *env,
8522 struct bpf_func_state *st)
8524 enum bpf_reg_liveness live;
8527 for (i = 0; i < BPF_REG_FP; i++) {
8528 live = st->regs[i].live;
8529 /* liveness must not touch this register anymore */
8530 st->regs[i].live |= REG_LIVE_DONE;
8531 if (!(live & REG_LIVE_READ))
8532 /* since the register is unused, clear its state
8533 * to make further comparison simpler
8535 __mark_reg_not_init(env, &st->regs[i]);
8538 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
8539 live = st->stack[i].spilled_ptr.live;
8540 /* liveness must not touch this stack slot anymore */
8541 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
8542 if (!(live & REG_LIVE_READ)) {
8543 __mark_reg_not_init(env, &st->stack[i].spilled_ptr);
8544 for (j = 0; j < BPF_REG_SIZE; j++)
8545 st->stack[i].slot_type[j] = STACK_INVALID;
8550 static void clean_verifier_state(struct bpf_verifier_env *env,
8551 struct bpf_verifier_state *st)
8555 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
8556 /* all regs in this state in all frames were already marked */
8559 for (i = 0; i <= st->curframe; i++)
8560 clean_func_state(env, st->frame[i]);
8563 /* the parentage chains form a tree.
8564 * the verifier states are added to state lists at given insn and
8565 * pushed into state stack for future exploration.
8566 * when the verifier reaches bpf_exit insn some of the verifer states
8567 * stored in the state lists have their final liveness state already,
8568 * but a lot of states will get revised from liveness point of view when
8569 * the verifier explores other branches.
8572 * 2: if r1 == 100 goto pc+1
8575 * when the verifier reaches exit insn the register r0 in the state list of
8576 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
8577 * of insn 2 and goes exploring further. At the insn 4 it will walk the
8578 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
8580 * Since the verifier pushes the branch states as it sees them while exploring
8581 * the program the condition of walking the branch instruction for the second
8582 * time means that all states below this branch were already explored and
8583 * their final liveness markes are already propagated.
8584 * Hence when the verifier completes the search of state list in is_state_visited()
8585 * we can call this clean_live_states() function to mark all liveness states
8586 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
8588 * This function also clears the registers and stack for states that !READ
8589 * to simplify state merging.
8591 * Important note here that walking the same branch instruction in the callee
8592 * doesn't meant that the states are DONE. The verifier has to compare
8595 static void clean_live_states(struct bpf_verifier_env *env, int insn,
8596 struct bpf_verifier_state *cur)
8598 struct bpf_verifier_state_list *sl;
8601 sl = *explored_state(env, insn);
8603 if (sl->state.branches)
8605 if (sl->state.insn_idx != insn ||
8606 sl->state.curframe != cur->curframe)
8608 for (i = 0; i <= cur->curframe; i++)
8609 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
8611 clean_verifier_state(env, &sl->state);
8617 /* Returns true if (rold safe implies rcur safe) */
8618 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
8619 struct idpair *idmap)
8623 if (!(rold->live & REG_LIVE_READ))
8624 /* explored state didn't use this */
8627 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
8629 if (rold->type == PTR_TO_STACK)
8630 /* two stack pointers are equal only if they're pointing to
8631 * the same stack frame, since fp-8 in foo != fp-8 in bar
8633 return equal && rold->frameno == rcur->frameno;
8638 if (rold->type == NOT_INIT)
8639 /* explored state can't have used this */
8641 if (rcur->type == NOT_INIT)
8643 switch (rold->type) {
8645 if (rcur->type == SCALAR_VALUE) {
8646 if (!rold->precise && !rcur->precise)
8648 /* new val must satisfy old val knowledge */
8649 return range_within(rold, rcur) &&
8650 tnum_in(rold->var_off, rcur->var_off);
8652 /* We're trying to use a pointer in place of a scalar.
8653 * Even if the scalar was unbounded, this could lead to
8654 * pointer leaks because scalars are allowed to leak
8655 * while pointers are not. We could make this safe in
8656 * special cases if root is calling us, but it's
8657 * probably not worth the hassle.
8661 case PTR_TO_MAP_VALUE:
8662 /* If the new min/max/var_off satisfy the old ones and
8663 * everything else matches, we are OK.
8664 * 'id' is not compared, since it's only used for maps with
8665 * bpf_spin_lock inside map element and in such cases if
8666 * the rest of the prog is valid for one map element then
8667 * it's valid for all map elements regardless of the key
8668 * used in bpf_map_lookup()
8670 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
8671 range_within(rold, rcur) &&
8672 tnum_in(rold->var_off, rcur->var_off);
8673 case PTR_TO_MAP_VALUE_OR_NULL:
8674 /* a PTR_TO_MAP_VALUE could be safe to use as a
8675 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
8676 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
8677 * checked, doing so could have affected others with the same
8678 * id, and we can't check for that because we lost the id when
8679 * we converted to a PTR_TO_MAP_VALUE.
8681 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
8683 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
8685 /* Check our ids match any regs they're supposed to */
8686 return check_ids(rold->id, rcur->id, idmap);
8687 case PTR_TO_PACKET_META:
8689 if (rcur->type != rold->type)
8691 /* We must have at least as much range as the old ptr
8692 * did, so that any accesses which were safe before are
8693 * still safe. This is true even if old range < old off,
8694 * since someone could have accessed through (ptr - k), or
8695 * even done ptr -= k in a register, to get a safe access.
8697 if (rold->range > rcur->range)
8699 /* If the offsets don't match, we can't trust our alignment;
8700 * nor can we be sure that we won't fall out of range.
8702 if (rold->off != rcur->off)
8704 /* id relations must be preserved */
8705 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
8707 /* new val must satisfy old val knowledge */
8708 return range_within(rold, rcur) &&
8709 tnum_in(rold->var_off, rcur->var_off);
8711 case CONST_PTR_TO_MAP:
8712 case PTR_TO_PACKET_END:
8713 case PTR_TO_FLOW_KEYS:
8715 case PTR_TO_SOCKET_OR_NULL:
8716 case PTR_TO_SOCK_COMMON:
8717 case PTR_TO_SOCK_COMMON_OR_NULL:
8718 case PTR_TO_TCP_SOCK:
8719 case PTR_TO_TCP_SOCK_OR_NULL:
8720 case PTR_TO_XDP_SOCK:
8721 /* Only valid matches are exact, which memcmp() above
8722 * would have accepted
8725 /* Don't know what's going on, just say it's not safe */
8729 /* Shouldn't get here; if we do, say it's not safe */
8734 static bool stacksafe(struct bpf_func_state *old,
8735 struct bpf_func_state *cur,
8736 struct idpair *idmap)
8740 /* walk slots of the explored stack and ignore any additional
8741 * slots in the current stack, since explored(safe) state
8744 for (i = 0; i < old->allocated_stack; i++) {
8745 spi = i / BPF_REG_SIZE;
8747 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
8748 i += BPF_REG_SIZE - 1;
8749 /* explored state didn't use this */
8753 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
8756 /* explored stack has more populated slots than current stack
8757 * and these slots were used
8759 if (i >= cur->allocated_stack)
8762 /* if old state was safe with misc data in the stack
8763 * it will be safe with zero-initialized stack.
8764 * The opposite is not true
8766 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
8767 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
8769 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
8770 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
8771 /* Ex: old explored (safe) state has STACK_SPILL in
8772 * this stack slot, but current has STACK_MISC ->
8773 * this verifier states are not equivalent,
8774 * return false to continue verification of this path
8777 if (i % BPF_REG_SIZE)
8779 if (old->stack[spi].slot_type[0] != STACK_SPILL)
8781 if (!regsafe(&old->stack[spi].spilled_ptr,
8782 &cur->stack[spi].spilled_ptr,
8784 /* when explored and current stack slot are both storing
8785 * spilled registers, check that stored pointers types
8786 * are the same as well.
8787 * Ex: explored safe path could have stored
8788 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
8789 * but current path has stored:
8790 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
8791 * such verifier states are not equivalent.
8792 * return false to continue verification of this path
8799 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
8801 if (old->acquired_refs != cur->acquired_refs)
8803 return !memcmp(old->refs, cur->refs,
8804 sizeof(*old->refs) * old->acquired_refs);
8807 /* compare two verifier states
8809 * all states stored in state_list are known to be valid, since
8810 * verifier reached 'bpf_exit' instruction through them
8812 * this function is called when verifier exploring different branches of
8813 * execution popped from the state stack. If it sees an old state that has
8814 * more strict register state and more strict stack state then this execution
8815 * branch doesn't need to be explored further, since verifier already
8816 * concluded that more strict state leads to valid finish.
8818 * Therefore two states are equivalent if register state is more conservative
8819 * and explored stack state is more conservative than the current one.
8822 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
8823 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
8825 * In other words if current stack state (one being explored) has more
8826 * valid slots than old one that already passed validation, it means
8827 * the verifier can stop exploring and conclude that current state is valid too
8829 * Similarly with registers. If explored state has register type as invalid
8830 * whereas register type in current state is meaningful, it means that
8831 * the current state will reach 'bpf_exit' instruction safely
8833 static bool func_states_equal(struct bpf_func_state *old,
8834 struct bpf_func_state *cur)
8836 struct idpair *idmap;
8840 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
8841 /* If we failed to allocate the idmap, just say it's not safe */
8845 for (i = 0; i < MAX_BPF_REG; i++) {
8846 if (!regsafe(&old->regs[i], &cur->regs[i], idmap))
8850 if (!stacksafe(old, cur, idmap))
8853 if (!refsafe(old, cur))
8861 static bool states_equal(struct bpf_verifier_env *env,
8862 struct bpf_verifier_state *old,
8863 struct bpf_verifier_state *cur)
8867 if (old->curframe != cur->curframe)
8870 /* Verification state from speculative execution simulation
8871 * must never prune a non-speculative execution one.
8873 if (old->speculative && !cur->speculative)
8876 if (old->active_spin_lock != cur->active_spin_lock)
8879 /* for states to be equal callsites have to be the same
8880 * and all frame states need to be equivalent
8882 for (i = 0; i <= old->curframe; i++) {
8883 if (old->frame[i]->callsite != cur->frame[i]->callsite)
8885 if (!func_states_equal(old->frame[i], cur->frame[i]))
8891 /* Return 0 if no propagation happened. Return negative error code if error
8892 * happened. Otherwise, return the propagated bit.
8894 static int propagate_liveness_reg(struct bpf_verifier_env *env,
8895 struct bpf_reg_state *reg,
8896 struct bpf_reg_state *parent_reg)
8898 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
8899 u8 flag = reg->live & REG_LIVE_READ;
8902 /* When comes here, read flags of PARENT_REG or REG could be any of
8903 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
8904 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
8906 if (parent_flag == REG_LIVE_READ64 ||
8907 /* Or if there is no read flag from REG. */
8909 /* Or if the read flag from REG is the same as PARENT_REG. */
8910 parent_flag == flag)
8913 err = mark_reg_read(env, reg, parent_reg, flag);
8920 /* A write screens off any subsequent reads; but write marks come from the
8921 * straight-line code between a state and its parent. When we arrive at an
8922 * equivalent state (jump target or such) we didn't arrive by the straight-line
8923 * code, so read marks in the state must propagate to the parent regardless
8924 * of the state's write marks. That's what 'parent == state->parent' comparison
8925 * in mark_reg_read() is for.
8927 static int propagate_liveness(struct bpf_verifier_env *env,
8928 const struct bpf_verifier_state *vstate,
8929 struct bpf_verifier_state *vparent)
8931 struct bpf_reg_state *state_reg, *parent_reg;
8932 struct bpf_func_state *state, *parent;
8933 int i, frame, err = 0;
8935 if (vparent->curframe != vstate->curframe) {
8936 WARN(1, "propagate_live: parent frame %d current frame %d\n",
8937 vparent->curframe, vstate->curframe);
8940 /* Propagate read liveness of registers... */
8941 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
8942 for (frame = 0; frame <= vstate->curframe; frame++) {
8943 parent = vparent->frame[frame];
8944 state = vstate->frame[frame];
8945 parent_reg = parent->regs;
8946 state_reg = state->regs;
8947 /* We don't need to worry about FP liveness, it's read-only */
8948 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
8949 err = propagate_liveness_reg(env, &state_reg[i],
8953 if (err == REG_LIVE_READ64)
8954 mark_insn_zext(env, &parent_reg[i]);
8957 /* Propagate stack slots. */
8958 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
8959 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
8960 parent_reg = &parent->stack[i].spilled_ptr;
8961 state_reg = &state->stack[i].spilled_ptr;
8962 err = propagate_liveness_reg(env, state_reg,
8971 /* find precise scalars in the previous equivalent state and
8972 * propagate them into the current state
8974 static int propagate_precision(struct bpf_verifier_env *env,
8975 const struct bpf_verifier_state *old)
8977 struct bpf_reg_state *state_reg;
8978 struct bpf_func_state *state;
8981 state = old->frame[old->curframe];
8982 state_reg = state->regs;
8983 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
8984 if (state_reg->type != SCALAR_VALUE ||
8985 !state_reg->precise)
8987 if (env->log.level & BPF_LOG_LEVEL2)
8988 verbose(env, "propagating r%d\n", i);
8989 err = mark_chain_precision(env, i);
8994 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
8995 if (state->stack[i].slot_type[0] != STACK_SPILL)
8997 state_reg = &state->stack[i].spilled_ptr;
8998 if (state_reg->type != SCALAR_VALUE ||
8999 !state_reg->precise)
9001 if (env->log.level & BPF_LOG_LEVEL2)
9002 verbose(env, "propagating fp%d\n",
9003 (-i - 1) * BPF_REG_SIZE);
9004 err = mark_chain_precision_stack(env, i);
9011 static bool states_maybe_looping(struct bpf_verifier_state *old,
9012 struct bpf_verifier_state *cur)
9014 struct bpf_func_state *fold, *fcur;
9015 int i, fr = cur->curframe;
9017 if (old->curframe != fr)
9020 fold = old->frame[fr];
9021 fcur = cur->frame[fr];
9022 for (i = 0; i < MAX_BPF_REG; i++)
9023 if (memcmp(&fold->regs[i], &fcur->regs[i],
9024 offsetof(struct bpf_reg_state, parent)))
9030 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
9032 struct bpf_verifier_state_list *new_sl;
9033 struct bpf_verifier_state_list *sl, **pprev;
9034 struct bpf_verifier_state *cur = env->cur_state, *new;
9035 int i, j, err, states_cnt = 0;
9036 bool add_new_state = env->test_state_freq ? true : false;
9038 cur->last_insn_idx = env->prev_insn_idx;
9039 if (!env->insn_aux_data[insn_idx].prune_point)
9040 /* this 'insn_idx' instruction wasn't marked, so we will not
9041 * be doing state search here
9045 /* bpf progs typically have pruning point every 4 instructions
9046 * http://vger.kernel.org/bpfconf2019.html#session-1
9047 * Do not add new state for future pruning if the verifier hasn't seen
9048 * at least 2 jumps and at least 8 instructions.
9049 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
9050 * In tests that amounts to up to 50% reduction into total verifier
9051 * memory consumption and 20% verifier time speedup.
9053 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
9054 env->insn_processed - env->prev_insn_processed >= 8)
9055 add_new_state = true;
9057 pprev = explored_state(env, insn_idx);
9060 clean_live_states(env, insn_idx, cur);
9064 if (sl->state.insn_idx != insn_idx)
9066 if (sl->state.branches) {
9067 if (states_maybe_looping(&sl->state, cur) &&
9068 states_equal(env, &sl->state, cur)) {
9069 verbose_linfo(env, insn_idx, "; ");
9070 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
9073 /* if the verifier is processing a loop, avoid adding new state
9074 * too often, since different loop iterations have distinct
9075 * states and may not help future pruning.
9076 * This threshold shouldn't be too low to make sure that
9077 * a loop with large bound will be rejected quickly.
9078 * The most abusive loop will be:
9080 * if r1 < 1000000 goto pc-2
9081 * 1M insn_procssed limit / 100 == 10k peak states.
9082 * This threshold shouldn't be too high either, since states
9083 * at the end of the loop are likely to be useful in pruning.
9085 if (env->jmps_processed - env->prev_jmps_processed < 20 &&
9086 env->insn_processed - env->prev_insn_processed < 100)
9087 add_new_state = false;
9090 if (states_equal(env, &sl->state, cur)) {
9092 /* reached equivalent register/stack state,
9094 * Registers read by the continuation are read by us.
9095 * If we have any write marks in env->cur_state, they
9096 * will prevent corresponding reads in the continuation
9097 * from reaching our parent (an explored_state). Our
9098 * own state will get the read marks recorded, but
9099 * they'll be immediately forgotten as we're pruning
9100 * this state and will pop a new one.
9102 err = propagate_liveness(env, &sl->state, cur);
9104 /* if previous state reached the exit with precision and
9105 * current state is equivalent to it (except precsion marks)
9106 * the precision needs to be propagated back in
9107 * the current state.
9109 err = err ? : push_jmp_history(env, cur);
9110 err = err ? : propagate_precision(env, &sl->state);
9116 /* when new state is not going to be added do not increase miss count.
9117 * Otherwise several loop iterations will remove the state
9118 * recorded earlier. The goal of these heuristics is to have
9119 * states from some iterations of the loop (some in the beginning
9120 * and some at the end) to help pruning.
9124 /* heuristic to determine whether this state is beneficial
9125 * to keep checking from state equivalence point of view.
9126 * Higher numbers increase max_states_per_insn and verification time,
9127 * but do not meaningfully decrease insn_processed.
9129 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
9130 /* the state is unlikely to be useful. Remove it to
9131 * speed up verification
9134 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
9135 u32 br = sl->state.branches;
9138 "BUG live_done but branches_to_explore %d\n",
9140 free_verifier_state(&sl->state, false);
9144 /* cannot free this state, since parentage chain may
9145 * walk it later. Add it for free_list instead to
9146 * be freed at the end of verification
9148 sl->next = env->free_list;
9149 env->free_list = sl;
9159 if (env->max_states_per_insn < states_cnt)
9160 env->max_states_per_insn = states_cnt;
9162 if (!env->bpf_capable && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
9163 return push_jmp_history(env, cur);
9166 return push_jmp_history(env, cur);
9168 /* There were no equivalent states, remember the current one.
9169 * Technically the current state is not proven to be safe yet,
9170 * but it will either reach outer most bpf_exit (which means it's safe)
9171 * or it will be rejected. When there are no loops the verifier won't be
9172 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
9173 * again on the way to bpf_exit.
9174 * When looping the sl->state.branches will be > 0 and this state
9175 * will not be considered for equivalence until branches == 0.
9177 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
9180 env->total_states++;
9182 env->prev_jmps_processed = env->jmps_processed;
9183 env->prev_insn_processed = env->insn_processed;
9185 /* add new state to the head of linked list */
9186 new = &new_sl->state;
9187 err = copy_verifier_state(new, cur);
9189 free_verifier_state(new, false);
9193 new->insn_idx = insn_idx;
9194 WARN_ONCE(new->branches != 1,
9195 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
9198 cur->first_insn_idx = insn_idx;
9199 clear_jmp_history(cur);
9200 new_sl->next = *explored_state(env, insn_idx);
9201 *explored_state(env, insn_idx) = new_sl;
9202 /* connect new state to parentage chain. Current frame needs all
9203 * registers connected. Only r6 - r9 of the callers are alive (pushed
9204 * to the stack implicitly by JITs) so in callers' frames connect just
9205 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
9206 * the state of the call instruction (with WRITTEN set), and r0 comes
9207 * from callee with its full parentage chain, anyway.
9209 /* clear write marks in current state: the writes we did are not writes
9210 * our child did, so they don't screen off its reads from us.
9211 * (There are no read marks in current state, because reads always mark
9212 * their parent and current state never has children yet. Only
9213 * explored_states can get read marks.)
9215 for (j = 0; j <= cur->curframe; j++) {
9216 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
9217 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
9218 for (i = 0; i < BPF_REG_FP; i++)
9219 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
9222 /* all stack frames are accessible from callee, clear them all */
9223 for (j = 0; j <= cur->curframe; j++) {
9224 struct bpf_func_state *frame = cur->frame[j];
9225 struct bpf_func_state *newframe = new->frame[j];
9227 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
9228 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
9229 frame->stack[i].spilled_ptr.parent =
9230 &newframe->stack[i].spilled_ptr;
9236 /* Return true if it's OK to have the same insn return a different type. */
9237 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
9242 case PTR_TO_SOCKET_OR_NULL:
9243 case PTR_TO_SOCK_COMMON:
9244 case PTR_TO_SOCK_COMMON_OR_NULL:
9245 case PTR_TO_TCP_SOCK:
9246 case PTR_TO_TCP_SOCK_OR_NULL:
9247 case PTR_TO_XDP_SOCK:
9249 case PTR_TO_BTF_ID_OR_NULL:
9256 /* If an instruction was previously used with particular pointer types, then we
9257 * need to be careful to avoid cases such as the below, where it may be ok
9258 * for one branch accessing the pointer, but not ok for the other branch:
9263 * R1 = some_other_valid_ptr;
9266 * R2 = *(u32 *)(R1 + 0);
9268 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
9270 return src != prev && (!reg_type_mismatch_ok(src) ||
9271 !reg_type_mismatch_ok(prev));
9274 static int do_check(struct bpf_verifier_env *env)
9276 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
9277 struct bpf_verifier_state *state = env->cur_state;
9278 struct bpf_insn *insns = env->prog->insnsi;
9279 struct bpf_reg_state *regs;
9280 int insn_cnt = env->prog->len;
9281 bool do_print_state = false;
9282 int prev_insn_idx = -1;
9285 struct bpf_insn *insn;
9289 env->prev_insn_idx = prev_insn_idx;
9290 if (env->insn_idx >= insn_cnt) {
9291 verbose(env, "invalid insn idx %d insn_cnt %d\n",
9292 env->insn_idx, insn_cnt);
9296 insn = &insns[env->insn_idx];
9297 class = BPF_CLASS(insn->code);
9299 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
9301 "BPF program is too large. Processed %d insn\n",
9302 env->insn_processed);
9306 err = is_state_visited(env, env->insn_idx);
9310 /* found equivalent state, can prune the search */
9311 if (env->log.level & BPF_LOG_LEVEL) {
9313 verbose(env, "\nfrom %d to %d%s: safe\n",
9314 env->prev_insn_idx, env->insn_idx,
9315 env->cur_state->speculative ?
9316 " (speculative execution)" : "");
9318 verbose(env, "%d: safe\n", env->insn_idx);
9320 goto process_bpf_exit;
9323 if (signal_pending(current))
9329 if (env->log.level & BPF_LOG_LEVEL2 ||
9330 (env->log.level & BPF_LOG_LEVEL && do_print_state)) {
9331 if (env->log.level & BPF_LOG_LEVEL2)
9332 verbose(env, "%d:", env->insn_idx);
9334 verbose(env, "\nfrom %d to %d%s:",
9335 env->prev_insn_idx, env->insn_idx,
9336 env->cur_state->speculative ?
9337 " (speculative execution)" : "");
9338 print_verifier_state(env, state->frame[state->curframe]);
9339 do_print_state = false;
9342 if (env->log.level & BPF_LOG_LEVEL) {
9343 const struct bpf_insn_cbs cbs = {
9344 .cb_print = verbose,
9345 .private_data = env,
9348 verbose_linfo(env, env->insn_idx, "; ");
9349 verbose(env, "%d: ", env->insn_idx);
9350 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
9353 if (bpf_prog_is_dev_bound(env->prog->aux)) {
9354 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
9355 env->prev_insn_idx);
9360 regs = cur_regs(env);
9361 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
9362 prev_insn_idx = env->insn_idx;
9364 if (class == BPF_ALU || class == BPF_ALU64) {
9365 err = check_alu_op(env, insn);
9369 } else if (class == BPF_LDX) {
9370 enum bpf_reg_type *prev_src_type, src_reg_type;
9372 /* check for reserved fields is already done */
9374 /* check src operand */
9375 err = check_reg_arg(env, insn->src_reg, SRC_OP);
9379 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
9383 src_reg_type = regs[insn->src_reg].type;
9385 /* check that memory (src_reg + off) is readable,
9386 * the state of dst_reg will be updated by this func
9388 err = check_mem_access(env, env->insn_idx, insn->src_reg,
9389 insn->off, BPF_SIZE(insn->code),
9390 BPF_READ, insn->dst_reg, false);
9394 prev_src_type = &env->insn_aux_data[env->insn_idx].ptr_type;
9396 if (*prev_src_type == NOT_INIT) {
9398 * dst_reg = *(u32 *)(src_reg + off)
9399 * save type to validate intersecting paths
9401 *prev_src_type = src_reg_type;
9403 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
9404 /* ABuser program is trying to use the same insn
9405 * dst_reg = *(u32*) (src_reg + off)
9406 * with different pointer types:
9407 * src_reg == ctx in one branch and
9408 * src_reg == stack|map in some other branch.
9411 verbose(env, "same insn cannot be used with different pointers\n");
9415 } else if (class == BPF_STX) {
9416 enum bpf_reg_type *prev_dst_type, dst_reg_type;
9418 if (BPF_MODE(insn->code) == BPF_XADD) {
9419 err = check_xadd(env, env->insn_idx, insn);
9426 /* check src1 operand */
9427 err = check_reg_arg(env, insn->src_reg, SRC_OP);
9430 /* check src2 operand */
9431 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
9435 dst_reg_type = regs[insn->dst_reg].type;
9437 /* check that memory (dst_reg + off) is writeable */
9438 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
9439 insn->off, BPF_SIZE(insn->code),
9440 BPF_WRITE, insn->src_reg, false);
9444 prev_dst_type = &env->insn_aux_data[env->insn_idx].ptr_type;
9446 if (*prev_dst_type == NOT_INIT) {
9447 *prev_dst_type = dst_reg_type;
9448 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
9449 verbose(env, "same insn cannot be used with different pointers\n");
9453 } else if (class == BPF_ST) {
9454 if (BPF_MODE(insn->code) != BPF_MEM ||
9455 insn->src_reg != BPF_REG_0) {
9456 verbose(env, "BPF_ST uses reserved fields\n");
9459 /* check src operand */
9460 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
9464 if (is_ctx_reg(env, insn->dst_reg)) {
9465 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
9467 reg_type_str[reg_state(env, insn->dst_reg)->type]);
9471 /* check that memory (dst_reg + off) is writeable */
9472 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
9473 insn->off, BPF_SIZE(insn->code),
9474 BPF_WRITE, -1, false);
9478 } else if (class == BPF_JMP || class == BPF_JMP32) {
9479 u8 opcode = BPF_OP(insn->code);
9481 env->jmps_processed++;
9482 if (opcode == BPF_CALL) {
9483 if (BPF_SRC(insn->code) != BPF_K ||
9485 (insn->src_reg != BPF_REG_0 &&
9486 insn->src_reg != BPF_PSEUDO_CALL) ||
9487 insn->dst_reg != BPF_REG_0 ||
9488 class == BPF_JMP32) {
9489 verbose(env, "BPF_CALL uses reserved fields\n");
9493 if (env->cur_state->active_spin_lock &&
9494 (insn->src_reg == BPF_PSEUDO_CALL ||
9495 insn->imm != BPF_FUNC_spin_unlock)) {
9496 verbose(env, "function calls are not allowed while holding a lock\n");
9499 if (insn->src_reg == BPF_PSEUDO_CALL)
9500 err = check_func_call(env, insn, &env->insn_idx);
9502 err = check_helper_call(env, insn->imm, env->insn_idx);
9506 } else if (opcode == BPF_JA) {
9507 if (BPF_SRC(insn->code) != BPF_K ||
9509 insn->src_reg != BPF_REG_0 ||
9510 insn->dst_reg != BPF_REG_0 ||
9511 class == BPF_JMP32) {
9512 verbose(env, "BPF_JA uses reserved fields\n");
9516 env->insn_idx += insn->off + 1;
9519 } else if (opcode == BPF_EXIT) {
9520 if (BPF_SRC(insn->code) != BPF_K ||
9522 insn->src_reg != BPF_REG_0 ||
9523 insn->dst_reg != BPF_REG_0 ||
9524 class == BPF_JMP32) {
9525 verbose(env, "BPF_EXIT uses reserved fields\n");
9529 if (env->cur_state->active_spin_lock) {
9530 verbose(env, "bpf_spin_unlock is missing\n");
9534 if (state->curframe) {
9535 /* exit from nested function */
9536 err = prepare_func_exit(env, &env->insn_idx);
9539 do_print_state = true;
9543 err = check_reference_leak(env);
9547 err = check_return_code(env);
9551 update_branch_counts(env, env->cur_state);
9552 err = pop_stack(env, &prev_insn_idx,
9553 &env->insn_idx, pop_log);
9559 do_print_state = true;
9563 err = check_cond_jmp_op(env, insn, &env->insn_idx);
9567 } else if (class == BPF_LD) {
9568 u8 mode = BPF_MODE(insn->code);
9570 if (mode == BPF_ABS || mode == BPF_IND) {
9571 err = check_ld_abs(env, insn);
9575 } else if (mode == BPF_IMM) {
9576 err = check_ld_imm(env, insn);
9581 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
9583 verbose(env, "invalid BPF_LD mode\n");
9587 verbose(env, "unknown insn class %d\n", class);
9597 /* replace pseudo btf_id with kernel symbol address */
9598 static int check_pseudo_btf_id(struct bpf_verifier_env *env,
9599 struct bpf_insn *insn,
9600 struct bpf_insn_aux_data *aux)
9602 const struct btf_var_secinfo *vsi;
9603 const struct btf_type *datasec;
9604 const struct btf_type *t;
9605 const char *sym_name;
9606 bool percpu = false;
9607 u32 type, id = insn->imm;
9613 verbose(env, "kernel is missing BTF, make sure CONFIG_DEBUG_INFO_BTF=y is specified in Kconfig.\n");
9617 if (insn[1].imm != 0) {
9618 verbose(env, "reserved field (insn[1].imm) is used in pseudo_btf_id ldimm64 insn.\n");
9622 t = btf_type_by_id(btf_vmlinux, id);
9624 verbose(env, "ldimm64 insn specifies invalid btf_id %d.\n", id);
9628 if (!btf_type_is_var(t)) {
9629 verbose(env, "pseudo btf_id %d in ldimm64 isn't KIND_VAR.\n",
9634 sym_name = btf_name_by_offset(btf_vmlinux, t->name_off);
9635 addr = kallsyms_lookup_name(sym_name);
9637 verbose(env, "ldimm64 failed to find the address for kernel symbol '%s'.\n",
9642 datasec_id = btf_find_by_name_kind(btf_vmlinux, ".data..percpu",
9644 if (datasec_id > 0) {
9645 datasec = btf_type_by_id(btf_vmlinux, datasec_id);
9646 for_each_vsi(i, datasec, vsi) {
9647 if (vsi->type == id) {
9654 insn[0].imm = (u32)addr;
9655 insn[1].imm = addr >> 32;
9658 t = btf_type_skip_modifiers(btf_vmlinux, type, NULL);
9660 aux->btf_var.reg_type = PTR_TO_PERCPU_BTF_ID;
9661 aux->btf_var.btf_id = type;
9662 } else if (!btf_type_is_struct(t)) {
9663 const struct btf_type *ret;
9667 /* resolve the type size of ksym. */
9668 ret = btf_resolve_size(btf_vmlinux, t, &tsize);
9670 tname = btf_name_by_offset(btf_vmlinux, t->name_off);
9671 verbose(env, "ldimm64 unable to resolve the size of type '%s': %ld\n",
9672 tname, PTR_ERR(ret));
9675 aux->btf_var.reg_type = PTR_TO_MEM;
9676 aux->btf_var.mem_size = tsize;
9678 aux->btf_var.reg_type = PTR_TO_BTF_ID;
9679 aux->btf_var.btf_id = type;
9684 static int check_map_prealloc(struct bpf_map *map)
9686 return (map->map_type != BPF_MAP_TYPE_HASH &&
9687 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
9688 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
9689 !(map->map_flags & BPF_F_NO_PREALLOC);
9692 static bool is_tracing_prog_type(enum bpf_prog_type type)
9695 case BPF_PROG_TYPE_KPROBE:
9696 case BPF_PROG_TYPE_TRACEPOINT:
9697 case BPF_PROG_TYPE_PERF_EVENT:
9698 case BPF_PROG_TYPE_RAW_TRACEPOINT:
9705 static bool is_preallocated_map(struct bpf_map *map)
9707 if (!check_map_prealloc(map))
9709 if (map->inner_map_meta && !check_map_prealloc(map->inner_map_meta))
9714 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
9715 struct bpf_map *map,
9716 struct bpf_prog *prog)
9719 enum bpf_prog_type prog_type = resolve_prog_type(prog);
9721 * Validate that trace type programs use preallocated hash maps.
9723 * For programs attached to PERF events this is mandatory as the
9724 * perf NMI can hit any arbitrary code sequence.
9726 * All other trace types using preallocated hash maps are unsafe as
9727 * well because tracepoint or kprobes can be inside locked regions
9728 * of the memory allocator or at a place where a recursion into the
9729 * memory allocator would see inconsistent state.
9731 * On RT enabled kernels run-time allocation of all trace type
9732 * programs is strictly prohibited due to lock type constraints. On
9733 * !RT kernels it is allowed for backwards compatibility reasons for
9734 * now, but warnings are emitted so developers are made aware of
9735 * the unsafety and can fix their programs before this is enforced.
9737 if (is_tracing_prog_type(prog_type) && !is_preallocated_map(map)) {
9738 if (prog_type == BPF_PROG_TYPE_PERF_EVENT) {
9739 verbose(env, "perf_event programs can only use preallocated hash map\n");
9742 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
9743 verbose(env, "trace type programs can only use preallocated hash map\n");
9746 WARN_ONCE(1, "trace type BPF program uses run-time allocation\n");
9747 verbose(env, "trace type programs with run-time allocated hash maps are unsafe. Switch to preallocated hash maps.\n");
9750 if ((is_tracing_prog_type(prog_type) ||
9751 prog_type == BPF_PROG_TYPE_SOCKET_FILTER) &&
9752 map_value_has_spin_lock(map)) {
9753 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
9757 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
9758 !bpf_offload_prog_map_match(prog, map)) {
9759 verbose(env, "offload device mismatch between prog and map\n");
9763 if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
9764 verbose(env, "bpf_struct_ops map cannot be used in prog\n");
9768 if (prog->aux->sleepable)
9769 switch (map->map_type) {
9770 case BPF_MAP_TYPE_HASH:
9771 case BPF_MAP_TYPE_LRU_HASH:
9772 case BPF_MAP_TYPE_ARRAY:
9773 if (!is_preallocated_map(map)) {
9775 "Sleepable programs can only use preallocated hash maps\n");
9781 "Sleepable programs can only use array and hash maps\n");
9788 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
9790 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
9791 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
9794 /* find and rewrite pseudo imm in ld_imm64 instructions:
9796 * 1. if it accesses map FD, replace it with actual map pointer.
9797 * 2. if it accesses btf_id of a VAR, replace it with pointer to the var.
9799 * NOTE: btf_vmlinux is required for converting pseudo btf_id.
9801 static int resolve_pseudo_ldimm64(struct bpf_verifier_env *env)
9803 struct bpf_insn *insn = env->prog->insnsi;
9804 int insn_cnt = env->prog->len;
9807 err = bpf_prog_calc_tag(env->prog);
9811 for (i = 0; i < insn_cnt; i++, insn++) {
9812 if (BPF_CLASS(insn->code) == BPF_LDX &&
9813 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
9814 verbose(env, "BPF_LDX uses reserved fields\n");
9818 if (BPF_CLASS(insn->code) == BPF_STX &&
9819 ((BPF_MODE(insn->code) != BPF_MEM &&
9820 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
9821 verbose(env, "BPF_STX uses reserved fields\n");
9825 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
9826 struct bpf_insn_aux_data *aux;
9827 struct bpf_map *map;
9831 if (i == insn_cnt - 1 || insn[1].code != 0 ||
9832 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
9834 verbose(env, "invalid bpf_ld_imm64 insn\n");
9838 if (insn[0].src_reg == 0)
9839 /* valid generic load 64-bit imm */
9842 if (insn[0].src_reg == BPF_PSEUDO_BTF_ID) {
9843 aux = &env->insn_aux_data[i];
9844 err = check_pseudo_btf_id(env, insn, aux);
9850 /* In final convert_pseudo_ld_imm64() step, this is
9851 * converted into regular 64-bit imm load insn.
9853 if ((insn[0].src_reg != BPF_PSEUDO_MAP_FD &&
9854 insn[0].src_reg != BPF_PSEUDO_MAP_VALUE) ||
9855 (insn[0].src_reg == BPF_PSEUDO_MAP_FD &&
9856 insn[1].imm != 0)) {
9858 "unrecognized bpf_ld_imm64 insn\n");
9862 f = fdget(insn[0].imm);
9863 map = __bpf_map_get(f);
9865 verbose(env, "fd %d is not pointing to valid bpf_map\n",
9867 return PTR_ERR(map);
9870 err = check_map_prog_compatibility(env, map, env->prog);
9876 aux = &env->insn_aux_data[i];
9877 if (insn->src_reg == BPF_PSEUDO_MAP_FD) {
9878 addr = (unsigned long)map;
9880 u32 off = insn[1].imm;
9882 if (off >= BPF_MAX_VAR_OFF) {
9883 verbose(env, "direct value offset of %u is not allowed\n", off);
9888 if (!map->ops->map_direct_value_addr) {
9889 verbose(env, "no direct value access support for this map type\n");
9894 err = map->ops->map_direct_value_addr(map, &addr, off);
9896 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
9897 map->value_size, off);
9906 insn[0].imm = (u32)addr;
9907 insn[1].imm = addr >> 32;
9909 /* check whether we recorded this map already */
9910 for (j = 0; j < env->used_map_cnt; j++) {
9911 if (env->used_maps[j] == map) {
9918 if (env->used_map_cnt >= MAX_USED_MAPS) {
9923 /* hold the map. If the program is rejected by verifier,
9924 * the map will be released by release_maps() or it
9925 * will be used by the valid program until it's unloaded
9926 * and all maps are released in free_used_maps()
9930 aux->map_index = env->used_map_cnt;
9931 env->used_maps[env->used_map_cnt++] = map;
9933 if (bpf_map_is_cgroup_storage(map) &&
9934 bpf_cgroup_storage_assign(env->prog->aux, map)) {
9935 verbose(env, "only one cgroup storage of each type is allowed\n");
9947 /* Basic sanity check before we invest more work here. */
9948 if (!bpf_opcode_in_insntable(insn->code)) {
9949 verbose(env, "unknown opcode %02x\n", insn->code);
9954 /* now all pseudo BPF_LD_IMM64 instructions load valid
9955 * 'struct bpf_map *' into a register instead of user map_fd.
9956 * These pointers will be used later by verifier to validate map access.
9961 /* drop refcnt of maps used by the rejected program */
9962 static void release_maps(struct bpf_verifier_env *env)
9964 __bpf_free_used_maps(env->prog->aux, env->used_maps,
9968 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
9969 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
9971 struct bpf_insn *insn = env->prog->insnsi;
9972 int insn_cnt = env->prog->len;
9975 for (i = 0; i < insn_cnt; i++, insn++)
9976 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
9980 /* single env->prog->insni[off] instruction was replaced with the range
9981 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
9982 * [0, off) and [off, end) to new locations, so the patched range stays zero
9984 static int adjust_insn_aux_data(struct bpf_verifier_env *env,
9985 struct bpf_prog *new_prog, u32 off, u32 cnt)
9987 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
9988 struct bpf_insn *insn = new_prog->insnsi;
9992 /* aux info at OFF always needs adjustment, no matter fast path
9993 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
9994 * original insn at old prog.
9996 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
10000 prog_len = new_prog->len;
10001 new_data = vzalloc(array_size(prog_len,
10002 sizeof(struct bpf_insn_aux_data)));
10005 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
10006 memcpy(new_data + off + cnt - 1, old_data + off,
10007 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
10008 for (i = off; i < off + cnt - 1; i++) {
10009 new_data[i].seen = env->pass_cnt;
10010 new_data[i].zext_dst = insn_has_def32(env, insn + i);
10012 env->insn_aux_data = new_data;
10017 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
10023 /* NOTE: fake 'exit' subprog should be updated as well. */
10024 for (i = 0; i <= env->subprog_cnt; i++) {
10025 if (env->subprog_info[i].start <= off)
10027 env->subprog_info[i].start += len - 1;
10031 static void adjust_poke_descs(struct bpf_prog *prog, u32 len)
10033 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
10034 int i, sz = prog->aux->size_poke_tab;
10035 struct bpf_jit_poke_descriptor *desc;
10037 for (i = 0; i < sz; i++) {
10039 desc->insn_idx += len - 1;
10043 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
10044 const struct bpf_insn *patch, u32 len)
10046 struct bpf_prog *new_prog;
10048 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
10049 if (IS_ERR(new_prog)) {
10050 if (PTR_ERR(new_prog) == -ERANGE)
10052 "insn %d cannot be patched due to 16-bit range\n",
10053 env->insn_aux_data[off].orig_idx);
10056 if (adjust_insn_aux_data(env, new_prog, off, len))
10058 adjust_subprog_starts(env, off, len);
10059 adjust_poke_descs(new_prog, len);
10063 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
10068 /* find first prog starting at or after off (first to remove) */
10069 for (i = 0; i < env->subprog_cnt; i++)
10070 if (env->subprog_info[i].start >= off)
10072 /* find first prog starting at or after off + cnt (first to stay) */
10073 for (j = i; j < env->subprog_cnt; j++)
10074 if (env->subprog_info[j].start >= off + cnt)
10076 /* if j doesn't start exactly at off + cnt, we are just removing
10077 * the front of previous prog
10079 if (env->subprog_info[j].start != off + cnt)
10083 struct bpf_prog_aux *aux = env->prog->aux;
10086 /* move fake 'exit' subprog as well */
10087 move = env->subprog_cnt + 1 - j;
10089 memmove(env->subprog_info + i,
10090 env->subprog_info + j,
10091 sizeof(*env->subprog_info) * move);
10092 env->subprog_cnt -= j - i;
10094 /* remove func_info */
10095 if (aux->func_info) {
10096 move = aux->func_info_cnt - j;
10098 memmove(aux->func_info + i,
10099 aux->func_info + j,
10100 sizeof(*aux->func_info) * move);
10101 aux->func_info_cnt -= j - i;
10102 /* func_info->insn_off is set after all code rewrites,
10103 * in adjust_btf_func() - no need to adjust
10107 /* convert i from "first prog to remove" to "first to adjust" */
10108 if (env->subprog_info[i].start == off)
10112 /* update fake 'exit' subprog as well */
10113 for (; i <= env->subprog_cnt; i++)
10114 env->subprog_info[i].start -= cnt;
10119 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
10122 struct bpf_prog *prog = env->prog;
10123 u32 i, l_off, l_cnt, nr_linfo;
10124 struct bpf_line_info *linfo;
10126 nr_linfo = prog->aux->nr_linfo;
10130 linfo = prog->aux->linfo;
10132 /* find first line info to remove, count lines to be removed */
10133 for (i = 0; i < nr_linfo; i++)
10134 if (linfo[i].insn_off >= off)
10139 for (; i < nr_linfo; i++)
10140 if (linfo[i].insn_off < off + cnt)
10145 /* First live insn doesn't match first live linfo, it needs to "inherit"
10146 * last removed linfo. prog is already modified, so prog->len == off
10147 * means no live instructions after (tail of the program was removed).
10149 if (prog->len != off && l_cnt &&
10150 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
10152 linfo[--i].insn_off = off + cnt;
10155 /* remove the line info which refer to the removed instructions */
10157 memmove(linfo + l_off, linfo + i,
10158 sizeof(*linfo) * (nr_linfo - i));
10160 prog->aux->nr_linfo -= l_cnt;
10161 nr_linfo = prog->aux->nr_linfo;
10164 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
10165 for (i = l_off; i < nr_linfo; i++)
10166 linfo[i].insn_off -= cnt;
10168 /* fix up all subprogs (incl. 'exit') which start >= off */
10169 for (i = 0; i <= env->subprog_cnt; i++)
10170 if (env->subprog_info[i].linfo_idx > l_off) {
10171 /* program may have started in the removed region but
10172 * may not be fully removed
10174 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
10175 env->subprog_info[i].linfo_idx -= l_cnt;
10177 env->subprog_info[i].linfo_idx = l_off;
10183 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
10185 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
10186 unsigned int orig_prog_len = env->prog->len;
10189 if (bpf_prog_is_dev_bound(env->prog->aux))
10190 bpf_prog_offload_remove_insns(env, off, cnt);
10192 err = bpf_remove_insns(env->prog, off, cnt);
10196 err = adjust_subprog_starts_after_remove(env, off, cnt);
10200 err = bpf_adj_linfo_after_remove(env, off, cnt);
10204 memmove(aux_data + off, aux_data + off + cnt,
10205 sizeof(*aux_data) * (orig_prog_len - off - cnt));
10210 /* The verifier does more data flow analysis than llvm and will not
10211 * explore branches that are dead at run time. Malicious programs can
10212 * have dead code too. Therefore replace all dead at-run-time code
10215 * Just nops are not optimal, e.g. if they would sit at the end of the
10216 * program and through another bug we would manage to jump there, then
10217 * we'd execute beyond program memory otherwise. Returning exception
10218 * code also wouldn't work since we can have subprogs where the dead
10219 * code could be located.
10221 static void sanitize_dead_code(struct bpf_verifier_env *env)
10223 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
10224 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
10225 struct bpf_insn *insn = env->prog->insnsi;
10226 const int insn_cnt = env->prog->len;
10229 for (i = 0; i < insn_cnt; i++) {
10230 if (aux_data[i].seen)
10232 memcpy(insn + i, &trap, sizeof(trap));
10236 static bool insn_is_cond_jump(u8 code)
10240 if (BPF_CLASS(code) == BPF_JMP32)
10243 if (BPF_CLASS(code) != BPF_JMP)
10247 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
10250 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
10252 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
10253 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
10254 struct bpf_insn *insn = env->prog->insnsi;
10255 const int insn_cnt = env->prog->len;
10258 for (i = 0; i < insn_cnt; i++, insn++) {
10259 if (!insn_is_cond_jump(insn->code))
10262 if (!aux_data[i + 1].seen)
10263 ja.off = insn->off;
10264 else if (!aux_data[i + 1 + insn->off].seen)
10269 if (bpf_prog_is_dev_bound(env->prog->aux))
10270 bpf_prog_offload_replace_insn(env, i, &ja);
10272 memcpy(insn, &ja, sizeof(ja));
10276 static int opt_remove_dead_code(struct bpf_verifier_env *env)
10278 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
10279 int insn_cnt = env->prog->len;
10282 for (i = 0; i < insn_cnt; i++) {
10286 while (i + j < insn_cnt && !aux_data[i + j].seen)
10291 err = verifier_remove_insns(env, i, j);
10294 insn_cnt = env->prog->len;
10300 static int opt_remove_nops(struct bpf_verifier_env *env)
10302 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
10303 struct bpf_insn *insn = env->prog->insnsi;
10304 int insn_cnt = env->prog->len;
10307 for (i = 0; i < insn_cnt; i++) {
10308 if (memcmp(&insn[i], &ja, sizeof(ja)))
10311 err = verifier_remove_insns(env, i, 1);
10321 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
10322 const union bpf_attr *attr)
10324 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
10325 struct bpf_insn_aux_data *aux = env->insn_aux_data;
10326 int i, patch_len, delta = 0, len = env->prog->len;
10327 struct bpf_insn *insns = env->prog->insnsi;
10328 struct bpf_prog *new_prog;
10331 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
10332 zext_patch[1] = BPF_ZEXT_REG(0);
10333 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
10334 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
10335 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
10336 for (i = 0; i < len; i++) {
10337 int adj_idx = i + delta;
10338 struct bpf_insn insn;
10340 insn = insns[adj_idx];
10341 if (!aux[adj_idx].zext_dst) {
10349 class = BPF_CLASS(code);
10350 if (insn_no_def(&insn))
10353 /* NOTE: arg "reg" (the fourth one) is only used for
10354 * BPF_STX which has been ruled out in above
10355 * check, it is safe to pass NULL here.
10357 if (is_reg64(env, &insn, insn.dst_reg, NULL, DST_OP)) {
10358 if (class == BPF_LD &&
10359 BPF_MODE(code) == BPF_IMM)
10364 /* ctx load could be transformed into wider load. */
10365 if (class == BPF_LDX &&
10366 aux[adj_idx].ptr_type == PTR_TO_CTX)
10369 imm_rnd = get_random_int();
10370 rnd_hi32_patch[0] = insn;
10371 rnd_hi32_patch[1].imm = imm_rnd;
10372 rnd_hi32_patch[3].dst_reg = insn.dst_reg;
10373 patch = rnd_hi32_patch;
10375 goto apply_patch_buffer;
10378 if (!bpf_jit_needs_zext())
10381 zext_patch[0] = insn;
10382 zext_patch[1].dst_reg = insn.dst_reg;
10383 zext_patch[1].src_reg = insn.dst_reg;
10384 patch = zext_patch;
10386 apply_patch_buffer:
10387 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
10390 env->prog = new_prog;
10391 insns = new_prog->insnsi;
10392 aux = env->insn_aux_data;
10393 delta += patch_len - 1;
10399 /* convert load instructions that access fields of a context type into a
10400 * sequence of instructions that access fields of the underlying structure:
10401 * struct __sk_buff -> struct sk_buff
10402 * struct bpf_sock_ops -> struct sock
10404 static int convert_ctx_accesses(struct bpf_verifier_env *env)
10406 const struct bpf_verifier_ops *ops = env->ops;
10407 int i, cnt, size, ctx_field_size, delta = 0;
10408 const int insn_cnt = env->prog->len;
10409 struct bpf_insn insn_buf[16], *insn;
10410 u32 target_size, size_default, off;
10411 struct bpf_prog *new_prog;
10412 enum bpf_access_type type;
10413 bool is_narrower_load;
10415 if (ops->gen_prologue || env->seen_direct_write) {
10416 if (!ops->gen_prologue) {
10417 verbose(env, "bpf verifier is misconfigured\n");
10420 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
10422 if (cnt >= ARRAY_SIZE(insn_buf)) {
10423 verbose(env, "bpf verifier is misconfigured\n");
10426 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
10430 env->prog = new_prog;
10435 if (bpf_prog_is_dev_bound(env->prog->aux))
10438 insn = env->prog->insnsi + delta;
10440 for (i = 0; i < insn_cnt; i++, insn++) {
10441 bpf_convert_ctx_access_t convert_ctx_access;
10443 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
10444 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
10445 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
10446 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
10448 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
10449 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
10450 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
10451 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
10456 if (type == BPF_WRITE &&
10457 env->insn_aux_data[i + delta].sanitize_stack_off) {
10458 struct bpf_insn patch[] = {
10459 /* Sanitize suspicious stack slot with zero.
10460 * There are no memory dependencies for this store,
10461 * since it's only using frame pointer and immediate
10464 BPF_ST_MEM(BPF_DW, BPF_REG_FP,
10465 env->insn_aux_data[i + delta].sanitize_stack_off,
10467 /* the original STX instruction will immediately
10468 * overwrite the same stack slot with appropriate value
10473 cnt = ARRAY_SIZE(patch);
10474 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
10479 env->prog = new_prog;
10480 insn = new_prog->insnsi + i + delta;
10484 switch (env->insn_aux_data[i + delta].ptr_type) {
10486 if (!ops->convert_ctx_access)
10488 convert_ctx_access = ops->convert_ctx_access;
10490 case PTR_TO_SOCKET:
10491 case PTR_TO_SOCK_COMMON:
10492 convert_ctx_access = bpf_sock_convert_ctx_access;
10494 case PTR_TO_TCP_SOCK:
10495 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
10497 case PTR_TO_XDP_SOCK:
10498 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
10500 case PTR_TO_BTF_ID:
10501 if (type == BPF_READ) {
10502 insn->code = BPF_LDX | BPF_PROBE_MEM |
10503 BPF_SIZE((insn)->code);
10504 env->prog->aux->num_exentries++;
10505 } else if (resolve_prog_type(env->prog) != BPF_PROG_TYPE_STRUCT_OPS) {
10506 verbose(env, "Writes through BTF pointers are not allowed\n");
10514 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
10515 size = BPF_LDST_BYTES(insn);
10517 /* If the read access is a narrower load of the field,
10518 * convert to a 4/8-byte load, to minimum program type specific
10519 * convert_ctx_access changes. If conversion is successful,
10520 * we will apply proper mask to the result.
10522 is_narrower_load = size < ctx_field_size;
10523 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
10525 if (is_narrower_load) {
10528 if (type == BPF_WRITE) {
10529 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
10534 if (ctx_field_size == 4)
10536 else if (ctx_field_size == 8)
10537 size_code = BPF_DW;
10539 insn->off = off & ~(size_default - 1);
10540 insn->code = BPF_LDX | BPF_MEM | size_code;
10544 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
10546 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
10547 (ctx_field_size && !target_size)) {
10548 verbose(env, "bpf verifier is misconfigured\n");
10552 if (is_narrower_load && size < target_size) {
10553 u8 shift = bpf_ctx_narrow_access_offset(
10554 off, size, size_default) * 8;
10555 if (ctx_field_size <= 4) {
10557 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
10560 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
10561 (1 << size * 8) - 1);
10564 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
10567 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
10568 (1ULL << size * 8) - 1);
10572 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
10578 /* keep walking new program and skip insns we just inserted */
10579 env->prog = new_prog;
10580 insn = new_prog->insnsi + i + delta;
10586 static int jit_subprogs(struct bpf_verifier_env *env)
10588 struct bpf_prog *prog = env->prog, **func, *tmp;
10589 int i, j, subprog_start, subprog_end = 0, len, subprog;
10590 struct bpf_map *map_ptr;
10591 struct bpf_insn *insn;
10592 void *old_bpf_func;
10593 int err, num_exentries;
10595 if (env->subprog_cnt <= 1)
10598 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
10599 if (insn->code != (BPF_JMP | BPF_CALL) ||
10600 insn->src_reg != BPF_PSEUDO_CALL)
10602 /* Upon error here we cannot fall back to interpreter but
10603 * need a hard reject of the program. Thus -EFAULT is
10604 * propagated in any case.
10606 subprog = find_subprog(env, i + insn->imm + 1);
10608 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
10609 i + insn->imm + 1);
10612 /* temporarily remember subprog id inside insn instead of
10613 * aux_data, since next loop will split up all insns into funcs
10615 insn->off = subprog;
10616 /* remember original imm in case JIT fails and fallback
10617 * to interpreter will be needed
10619 env->insn_aux_data[i].call_imm = insn->imm;
10620 /* point imm to __bpf_call_base+1 from JITs point of view */
10624 err = bpf_prog_alloc_jited_linfo(prog);
10626 goto out_undo_insn;
10629 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
10631 goto out_undo_insn;
10633 for (i = 0; i < env->subprog_cnt; i++) {
10634 subprog_start = subprog_end;
10635 subprog_end = env->subprog_info[i + 1].start;
10637 len = subprog_end - subprog_start;
10638 /* BPF_PROG_RUN doesn't call subprogs directly,
10639 * hence main prog stats include the runtime of subprogs.
10640 * subprogs don't have IDs and not reachable via prog_get_next_id
10641 * func[i]->aux->stats will never be accessed and stays NULL
10643 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
10646 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
10647 len * sizeof(struct bpf_insn));
10648 func[i]->type = prog->type;
10649 func[i]->len = len;
10650 if (bpf_prog_calc_tag(func[i]))
10652 func[i]->is_func = 1;
10653 func[i]->aux->func_idx = i;
10654 /* the btf and func_info will be freed only at prog->aux */
10655 func[i]->aux->btf = prog->aux->btf;
10656 func[i]->aux->func_info = prog->aux->func_info;
10658 for (j = 0; j < prog->aux->size_poke_tab; j++) {
10659 u32 insn_idx = prog->aux->poke_tab[j].insn_idx;
10662 if (!(insn_idx >= subprog_start &&
10663 insn_idx <= subprog_end))
10666 ret = bpf_jit_add_poke_descriptor(func[i],
10667 &prog->aux->poke_tab[j]);
10669 verbose(env, "adding tail call poke descriptor failed\n");
10673 func[i]->insnsi[insn_idx - subprog_start].imm = ret + 1;
10675 map_ptr = func[i]->aux->poke_tab[ret].tail_call.map;
10676 ret = map_ptr->ops->map_poke_track(map_ptr, func[i]->aux);
10678 verbose(env, "tracking tail call prog failed\n");
10683 /* Use bpf_prog_F_tag to indicate functions in stack traces.
10684 * Long term would need debug info to populate names
10686 func[i]->aux->name[0] = 'F';
10687 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
10688 func[i]->jit_requested = 1;
10689 func[i]->aux->linfo = prog->aux->linfo;
10690 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
10691 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
10692 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
10694 insn = func[i]->insnsi;
10695 for (j = 0; j < func[i]->len; j++, insn++) {
10696 if (BPF_CLASS(insn->code) == BPF_LDX &&
10697 BPF_MODE(insn->code) == BPF_PROBE_MEM)
10700 func[i]->aux->num_exentries = num_exentries;
10701 func[i]->aux->tail_call_reachable = env->subprog_info[i].tail_call_reachable;
10702 func[i] = bpf_int_jit_compile(func[i]);
10703 if (!func[i]->jited) {
10710 /* Untrack main program's aux structs so that during map_poke_run()
10711 * we will not stumble upon the unfilled poke descriptors; each
10712 * of the main program's poke descs got distributed across subprogs
10713 * and got tracked onto map, so we are sure that none of them will
10714 * be missed after the operation below
10716 for (i = 0; i < prog->aux->size_poke_tab; i++) {
10717 map_ptr = prog->aux->poke_tab[i].tail_call.map;
10719 map_ptr->ops->map_poke_untrack(map_ptr, prog->aux);
10722 /* at this point all bpf functions were successfully JITed
10723 * now populate all bpf_calls with correct addresses and
10724 * run last pass of JIT
10726 for (i = 0; i < env->subprog_cnt; i++) {
10727 insn = func[i]->insnsi;
10728 for (j = 0; j < func[i]->len; j++, insn++) {
10729 if (insn->code != (BPF_JMP | BPF_CALL) ||
10730 insn->src_reg != BPF_PSEUDO_CALL)
10732 subprog = insn->off;
10733 insn->imm = BPF_CAST_CALL(func[subprog]->bpf_func) -
10737 /* we use the aux data to keep a list of the start addresses
10738 * of the JITed images for each function in the program
10740 * for some architectures, such as powerpc64, the imm field
10741 * might not be large enough to hold the offset of the start
10742 * address of the callee's JITed image from __bpf_call_base
10744 * in such cases, we can lookup the start address of a callee
10745 * by using its subprog id, available from the off field of
10746 * the call instruction, as an index for this list
10748 func[i]->aux->func = func;
10749 func[i]->aux->func_cnt = env->subprog_cnt;
10751 for (i = 0; i < env->subprog_cnt; i++) {
10752 old_bpf_func = func[i]->bpf_func;
10753 tmp = bpf_int_jit_compile(func[i]);
10754 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
10755 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
10762 /* finally lock prog and jit images for all functions and
10763 * populate kallsysm
10765 for (i = 0; i < env->subprog_cnt; i++) {
10766 bpf_prog_lock_ro(func[i]);
10767 bpf_prog_kallsyms_add(func[i]);
10770 /* Last step: make now unused interpreter insns from main
10771 * prog consistent for later dump requests, so they can
10772 * later look the same as if they were interpreted only.
10774 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
10775 if (insn->code != (BPF_JMP | BPF_CALL) ||
10776 insn->src_reg != BPF_PSEUDO_CALL)
10778 insn->off = env->insn_aux_data[i].call_imm;
10779 subprog = find_subprog(env, i + insn->off + 1);
10780 insn->imm = subprog;
10784 prog->bpf_func = func[0]->bpf_func;
10785 prog->aux->func = func;
10786 prog->aux->func_cnt = env->subprog_cnt;
10787 bpf_prog_free_unused_jited_linfo(prog);
10790 for (i = 0; i < env->subprog_cnt; i++) {
10794 for (j = 0; j < func[i]->aux->size_poke_tab; j++) {
10795 map_ptr = func[i]->aux->poke_tab[j].tail_call.map;
10796 map_ptr->ops->map_poke_untrack(map_ptr, func[i]->aux);
10798 bpf_jit_free(func[i]);
10802 /* cleanup main prog to be interpreted */
10803 prog->jit_requested = 0;
10804 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
10805 if (insn->code != (BPF_JMP | BPF_CALL) ||
10806 insn->src_reg != BPF_PSEUDO_CALL)
10809 insn->imm = env->insn_aux_data[i].call_imm;
10811 bpf_prog_free_jited_linfo(prog);
10815 static int fixup_call_args(struct bpf_verifier_env *env)
10817 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
10818 struct bpf_prog *prog = env->prog;
10819 struct bpf_insn *insn = prog->insnsi;
10824 if (env->prog->jit_requested &&
10825 !bpf_prog_is_dev_bound(env->prog->aux)) {
10826 err = jit_subprogs(env);
10829 if (err == -EFAULT)
10832 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
10833 if (env->subprog_cnt > 1 && env->prog->aux->tail_call_reachable) {
10834 /* When JIT fails the progs with bpf2bpf calls and tail_calls
10835 * have to be rejected, since interpreter doesn't support them yet.
10837 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
10840 for (i = 0; i < prog->len; i++, insn++) {
10841 if (insn->code != (BPF_JMP | BPF_CALL) ||
10842 insn->src_reg != BPF_PSEUDO_CALL)
10844 depth = get_callee_stack_depth(env, insn, i);
10847 bpf_patch_call_args(insn, depth);
10854 /* fixup insn->imm field of bpf_call instructions
10855 * and inline eligible helpers as explicit sequence of BPF instructions
10857 * this function is called after eBPF program passed verification
10859 static int fixup_bpf_calls(struct bpf_verifier_env *env)
10861 struct bpf_prog *prog = env->prog;
10862 bool expect_blinding = bpf_jit_blinding_enabled(prog);
10863 struct bpf_insn *insn = prog->insnsi;
10864 const struct bpf_func_proto *fn;
10865 const int insn_cnt = prog->len;
10866 const struct bpf_map_ops *ops;
10867 struct bpf_insn_aux_data *aux;
10868 struct bpf_insn insn_buf[16];
10869 struct bpf_prog *new_prog;
10870 struct bpf_map *map_ptr;
10871 int i, ret, cnt, delta = 0;
10873 for (i = 0; i < insn_cnt; i++, insn++) {
10874 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
10875 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
10876 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
10877 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
10878 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
10879 bool isdiv = BPF_OP(insn->code) == BPF_DIV;
10880 struct bpf_insn *patchlet;
10881 struct bpf_insn chk_and_div[] = {
10882 /* [R,W]x div 0 -> 0 */
10883 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
10884 BPF_JNE | BPF_K, insn->src_reg,
10886 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
10887 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
10890 struct bpf_insn chk_and_mod[] = {
10891 /* [R,W]x mod 0 -> [R,W]x */
10892 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
10893 BPF_JEQ | BPF_K, insn->src_reg,
10894 0, 1 + (is64 ? 0 : 1), 0),
10896 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
10897 BPF_MOV32_REG(insn->dst_reg, insn->dst_reg),
10900 patchlet = isdiv ? chk_and_div : chk_and_mod;
10901 cnt = isdiv ? ARRAY_SIZE(chk_and_div) :
10902 ARRAY_SIZE(chk_and_mod) - (is64 ? 2 : 0);
10904 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
10909 env->prog = prog = new_prog;
10910 insn = new_prog->insnsi + i + delta;
10914 if (BPF_CLASS(insn->code) == BPF_LD &&
10915 (BPF_MODE(insn->code) == BPF_ABS ||
10916 BPF_MODE(insn->code) == BPF_IND)) {
10917 cnt = env->ops->gen_ld_abs(insn, insn_buf);
10918 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
10919 verbose(env, "bpf verifier is misconfigured\n");
10923 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
10928 env->prog = prog = new_prog;
10929 insn = new_prog->insnsi + i + delta;
10933 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
10934 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
10935 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
10936 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
10937 struct bpf_insn insn_buf[16];
10938 struct bpf_insn *patch = &insn_buf[0];
10942 aux = &env->insn_aux_data[i + delta];
10943 if (!aux->alu_state ||
10944 aux->alu_state == BPF_ALU_NON_POINTER)
10947 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
10948 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
10949 BPF_ALU_SANITIZE_SRC;
10951 off_reg = issrc ? insn->src_reg : insn->dst_reg;
10953 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
10954 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
10955 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
10956 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
10957 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
10958 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
10960 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX,
10962 insn->src_reg = BPF_REG_AX;
10964 *patch++ = BPF_ALU64_REG(BPF_AND, off_reg,
10968 insn->code = insn->code == code_add ?
10969 code_sub : code_add;
10971 if (issrc && isneg)
10972 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
10973 cnt = patch - insn_buf;
10975 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
10980 env->prog = prog = new_prog;
10981 insn = new_prog->insnsi + i + delta;
10985 if (insn->code != (BPF_JMP | BPF_CALL))
10987 if (insn->src_reg == BPF_PSEUDO_CALL)
10990 if (insn->imm == BPF_FUNC_get_route_realm)
10991 prog->dst_needed = 1;
10992 if (insn->imm == BPF_FUNC_get_prandom_u32)
10993 bpf_user_rnd_init_once();
10994 if (insn->imm == BPF_FUNC_override_return)
10995 prog->kprobe_override = 1;
10996 if (insn->imm == BPF_FUNC_tail_call) {
10997 /* If we tail call into other programs, we
10998 * cannot make any assumptions since they can
10999 * be replaced dynamically during runtime in
11000 * the program array.
11002 prog->cb_access = 1;
11003 if (!allow_tail_call_in_subprogs(env))
11004 prog->aux->stack_depth = MAX_BPF_STACK;
11005 prog->aux->max_pkt_offset = MAX_PACKET_OFF;
11007 /* mark bpf_tail_call as different opcode to avoid
11008 * conditional branch in the interpeter for every normal
11009 * call and to prevent accidental JITing by JIT compiler
11010 * that doesn't support bpf_tail_call yet
11013 insn->code = BPF_JMP | BPF_TAIL_CALL;
11015 aux = &env->insn_aux_data[i + delta];
11016 if (env->bpf_capable && !expect_blinding &&
11017 prog->jit_requested &&
11018 !bpf_map_key_poisoned(aux) &&
11019 !bpf_map_ptr_poisoned(aux) &&
11020 !bpf_map_ptr_unpriv(aux)) {
11021 struct bpf_jit_poke_descriptor desc = {
11022 .reason = BPF_POKE_REASON_TAIL_CALL,
11023 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
11024 .tail_call.key = bpf_map_key_immediate(aux),
11025 .insn_idx = i + delta,
11028 ret = bpf_jit_add_poke_descriptor(prog, &desc);
11030 verbose(env, "adding tail call poke descriptor failed\n");
11034 insn->imm = ret + 1;
11038 if (!bpf_map_ptr_unpriv(aux))
11041 /* instead of changing every JIT dealing with tail_call
11042 * emit two extra insns:
11043 * if (index >= max_entries) goto out;
11044 * index &= array->index_mask;
11045 * to avoid out-of-bounds cpu speculation
11047 if (bpf_map_ptr_poisoned(aux)) {
11048 verbose(env, "tail_call abusing map_ptr\n");
11052 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
11053 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
11054 map_ptr->max_entries, 2);
11055 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
11056 container_of(map_ptr,
11059 insn_buf[2] = *insn;
11061 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
11066 env->prog = prog = new_prog;
11067 insn = new_prog->insnsi + i + delta;
11071 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
11072 * and other inlining handlers are currently limited to 64 bit
11075 if (prog->jit_requested && BITS_PER_LONG == 64 &&
11076 (insn->imm == BPF_FUNC_map_lookup_elem ||
11077 insn->imm == BPF_FUNC_map_update_elem ||
11078 insn->imm == BPF_FUNC_map_delete_elem ||
11079 insn->imm == BPF_FUNC_map_push_elem ||
11080 insn->imm == BPF_FUNC_map_pop_elem ||
11081 insn->imm == BPF_FUNC_map_peek_elem)) {
11082 aux = &env->insn_aux_data[i + delta];
11083 if (bpf_map_ptr_poisoned(aux))
11084 goto patch_call_imm;
11086 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
11087 ops = map_ptr->ops;
11088 if (insn->imm == BPF_FUNC_map_lookup_elem &&
11089 ops->map_gen_lookup) {
11090 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
11091 if (cnt == -EOPNOTSUPP)
11092 goto patch_map_ops_generic;
11093 if (cnt <= 0 || cnt >= ARRAY_SIZE(insn_buf)) {
11094 verbose(env, "bpf verifier is misconfigured\n");
11098 new_prog = bpf_patch_insn_data(env, i + delta,
11104 env->prog = prog = new_prog;
11105 insn = new_prog->insnsi + i + delta;
11109 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
11110 (void *(*)(struct bpf_map *map, void *key))NULL));
11111 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
11112 (int (*)(struct bpf_map *map, void *key))NULL));
11113 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
11114 (int (*)(struct bpf_map *map, void *key, void *value,
11116 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
11117 (int (*)(struct bpf_map *map, void *value,
11119 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
11120 (int (*)(struct bpf_map *map, void *value))NULL));
11121 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
11122 (int (*)(struct bpf_map *map, void *value))NULL));
11123 patch_map_ops_generic:
11124 switch (insn->imm) {
11125 case BPF_FUNC_map_lookup_elem:
11126 insn->imm = BPF_CAST_CALL(ops->map_lookup_elem) -
11129 case BPF_FUNC_map_update_elem:
11130 insn->imm = BPF_CAST_CALL(ops->map_update_elem) -
11133 case BPF_FUNC_map_delete_elem:
11134 insn->imm = BPF_CAST_CALL(ops->map_delete_elem) -
11137 case BPF_FUNC_map_push_elem:
11138 insn->imm = BPF_CAST_CALL(ops->map_push_elem) -
11141 case BPF_FUNC_map_pop_elem:
11142 insn->imm = BPF_CAST_CALL(ops->map_pop_elem) -
11145 case BPF_FUNC_map_peek_elem:
11146 insn->imm = BPF_CAST_CALL(ops->map_peek_elem) -
11151 goto patch_call_imm;
11154 if (prog->jit_requested && BITS_PER_LONG == 64 &&
11155 insn->imm == BPF_FUNC_jiffies64) {
11156 struct bpf_insn ld_jiffies_addr[2] = {
11157 BPF_LD_IMM64(BPF_REG_0,
11158 (unsigned long)&jiffies),
11161 insn_buf[0] = ld_jiffies_addr[0];
11162 insn_buf[1] = ld_jiffies_addr[1];
11163 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
11167 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
11173 env->prog = prog = new_prog;
11174 insn = new_prog->insnsi + i + delta;
11179 fn = env->ops->get_func_proto(insn->imm, env->prog);
11180 /* all functions that have prototype and verifier allowed
11181 * programs to call them, must be real in-kernel functions
11185 "kernel subsystem misconfigured func %s#%d\n",
11186 func_id_name(insn->imm), insn->imm);
11189 insn->imm = fn->func - __bpf_call_base;
11192 /* Since poke tab is now finalized, publish aux to tracker. */
11193 for (i = 0; i < prog->aux->size_poke_tab; i++) {
11194 map_ptr = prog->aux->poke_tab[i].tail_call.map;
11195 if (!map_ptr->ops->map_poke_track ||
11196 !map_ptr->ops->map_poke_untrack ||
11197 !map_ptr->ops->map_poke_run) {
11198 verbose(env, "bpf verifier is misconfigured\n");
11202 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
11204 verbose(env, "tracking tail call prog failed\n");
11212 static void free_states(struct bpf_verifier_env *env)
11214 struct bpf_verifier_state_list *sl, *sln;
11217 sl = env->free_list;
11220 free_verifier_state(&sl->state, false);
11224 env->free_list = NULL;
11226 if (!env->explored_states)
11229 for (i = 0; i < state_htab_size(env); i++) {
11230 sl = env->explored_states[i];
11234 free_verifier_state(&sl->state, false);
11238 env->explored_states[i] = NULL;
11242 /* The verifier is using insn_aux_data[] to store temporary data during
11243 * verification and to store information for passes that run after the
11244 * verification like dead code sanitization. do_check_common() for subprogram N
11245 * may analyze many other subprograms. sanitize_insn_aux_data() clears all
11246 * temporary data after do_check_common() finds that subprogram N cannot be
11247 * verified independently. pass_cnt counts the number of times
11248 * do_check_common() was run and insn->aux->seen tells the pass number
11249 * insn_aux_data was touched. These variables are compared to clear temporary
11250 * data from failed pass. For testing and experiments do_check_common() can be
11251 * run multiple times even when prior attempt to verify is unsuccessful.
11253 static void sanitize_insn_aux_data(struct bpf_verifier_env *env)
11255 struct bpf_insn *insn = env->prog->insnsi;
11256 struct bpf_insn_aux_data *aux;
11259 for (i = 0; i < env->prog->len; i++) {
11260 class = BPF_CLASS(insn[i].code);
11261 if (class != BPF_LDX && class != BPF_STX)
11263 aux = &env->insn_aux_data[i];
11264 if (aux->seen != env->pass_cnt)
11266 memset(aux, 0, offsetof(typeof(*aux), orig_idx));
11270 static int do_check_common(struct bpf_verifier_env *env, int subprog)
11272 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
11273 struct bpf_verifier_state *state;
11274 struct bpf_reg_state *regs;
11277 env->prev_linfo = NULL;
11280 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
11283 state->curframe = 0;
11284 state->speculative = false;
11285 state->branches = 1;
11286 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
11287 if (!state->frame[0]) {
11291 env->cur_state = state;
11292 init_func_state(env, state->frame[0],
11293 BPF_MAIN_FUNC /* callsite */,
11297 regs = state->frame[state->curframe]->regs;
11298 if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
11299 ret = btf_prepare_func_args(env, subprog, regs);
11302 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
11303 if (regs[i].type == PTR_TO_CTX)
11304 mark_reg_known_zero(env, regs, i);
11305 else if (regs[i].type == SCALAR_VALUE)
11306 mark_reg_unknown(env, regs, i);
11309 /* 1st arg to a function */
11310 regs[BPF_REG_1].type = PTR_TO_CTX;
11311 mark_reg_known_zero(env, regs, BPF_REG_1);
11312 ret = btf_check_func_arg_match(env, subprog, regs);
11313 if (ret == -EFAULT)
11314 /* unlikely verifier bug. abort.
11315 * ret == 0 and ret < 0 are sadly acceptable for
11316 * main() function due to backward compatibility.
11317 * Like socket filter program may be written as:
11318 * int bpf_prog(struct pt_regs *ctx)
11319 * and never dereference that ctx in the program.
11320 * 'struct pt_regs' is a type mismatch for socket
11321 * filter that should be using 'struct __sk_buff'.
11326 ret = do_check(env);
11328 /* check for NULL is necessary, since cur_state can be freed inside
11329 * do_check() under memory pressure.
11331 if (env->cur_state) {
11332 free_verifier_state(env->cur_state, true);
11333 env->cur_state = NULL;
11335 while (!pop_stack(env, NULL, NULL, false));
11336 if (!ret && pop_log)
11337 bpf_vlog_reset(&env->log, 0);
11340 /* clean aux data in case subprog was rejected */
11341 sanitize_insn_aux_data(env);
11345 /* Verify all global functions in a BPF program one by one based on their BTF.
11346 * All global functions must pass verification. Otherwise the whole program is rejected.
11357 * foo() will be verified first for R1=any_scalar_value. During verification it
11358 * will be assumed that bar() already verified successfully and call to bar()
11359 * from foo() will be checked for type match only. Later bar() will be verified
11360 * independently to check that it's safe for R1=any_scalar_value.
11362 static int do_check_subprogs(struct bpf_verifier_env *env)
11364 struct bpf_prog_aux *aux = env->prog->aux;
11367 if (!aux->func_info)
11370 for (i = 1; i < env->subprog_cnt; i++) {
11371 if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL)
11373 env->insn_idx = env->subprog_info[i].start;
11374 WARN_ON_ONCE(env->insn_idx == 0);
11375 ret = do_check_common(env, i);
11378 } else if (env->log.level & BPF_LOG_LEVEL) {
11380 "Func#%d is safe for any args that match its prototype\n",
11387 static int do_check_main(struct bpf_verifier_env *env)
11392 ret = do_check_common(env, 0);
11394 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
11399 static void print_verification_stats(struct bpf_verifier_env *env)
11403 if (env->log.level & BPF_LOG_STATS) {
11404 verbose(env, "verification time %lld usec\n",
11405 div_u64(env->verification_time, 1000));
11406 verbose(env, "stack depth ");
11407 for (i = 0; i < env->subprog_cnt; i++) {
11408 u32 depth = env->subprog_info[i].stack_depth;
11410 verbose(env, "%d", depth);
11411 if (i + 1 < env->subprog_cnt)
11414 verbose(env, "\n");
11416 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
11417 "total_states %d peak_states %d mark_read %d\n",
11418 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
11419 env->max_states_per_insn, env->total_states,
11420 env->peak_states, env->longest_mark_read_walk);
11423 static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
11425 const struct btf_type *t, *func_proto;
11426 const struct bpf_struct_ops *st_ops;
11427 const struct btf_member *member;
11428 struct bpf_prog *prog = env->prog;
11429 u32 btf_id, member_idx;
11432 if (!prog->gpl_compatible) {
11433 verbose(env, "struct ops programs must have a GPL compatible license\n");
11437 btf_id = prog->aux->attach_btf_id;
11438 st_ops = bpf_struct_ops_find(btf_id);
11440 verbose(env, "attach_btf_id %u is not a supported struct\n",
11446 member_idx = prog->expected_attach_type;
11447 if (member_idx >= btf_type_vlen(t)) {
11448 verbose(env, "attach to invalid member idx %u of struct %s\n",
11449 member_idx, st_ops->name);
11453 member = &btf_type_member(t)[member_idx];
11454 mname = btf_name_by_offset(btf_vmlinux, member->name_off);
11455 func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type,
11458 verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
11459 mname, member_idx, st_ops->name);
11463 if (st_ops->check_member) {
11464 int err = st_ops->check_member(t, member);
11467 verbose(env, "attach to unsupported member %s of struct %s\n",
11468 mname, st_ops->name);
11473 prog->aux->attach_func_proto = func_proto;
11474 prog->aux->attach_func_name = mname;
11475 env->ops = st_ops->verifier_ops;
11479 #define SECURITY_PREFIX "security_"
11481 static int check_attach_modify_return(unsigned long addr, const char *func_name)
11483 if (within_error_injection_list(addr) ||
11484 !strncmp(SECURITY_PREFIX, func_name, sizeof(SECURITY_PREFIX) - 1))
11490 /* non exhaustive list of sleepable bpf_lsm_*() functions */
11491 BTF_SET_START(btf_sleepable_lsm_hooks)
11492 #ifdef CONFIG_BPF_LSM
11493 BTF_ID(func, bpf_lsm_bprm_committed_creds)
11497 BTF_SET_END(btf_sleepable_lsm_hooks)
11499 static int check_sleepable_lsm_hook(u32 btf_id)
11501 return btf_id_set_contains(&btf_sleepable_lsm_hooks, btf_id);
11504 /* list of non-sleepable functions that are otherwise on
11505 * ALLOW_ERROR_INJECTION list
11507 BTF_SET_START(btf_non_sleepable_error_inject)
11508 /* Three functions below can be called from sleepable and non-sleepable context.
11509 * Assume non-sleepable from bpf safety point of view.
11511 BTF_ID(func, __add_to_page_cache_locked)
11512 BTF_ID(func, should_fail_alloc_page)
11513 BTF_ID(func, should_failslab)
11514 BTF_SET_END(btf_non_sleepable_error_inject)
11516 static int check_non_sleepable_error_inject(u32 btf_id)
11518 return btf_id_set_contains(&btf_non_sleepable_error_inject, btf_id);
11521 int bpf_check_attach_target(struct bpf_verifier_log *log,
11522 const struct bpf_prog *prog,
11523 const struct bpf_prog *tgt_prog,
11525 struct bpf_attach_target_info *tgt_info)
11527 bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
11528 const char prefix[] = "btf_trace_";
11529 int ret = 0, subprog = -1, i;
11530 const struct btf_type *t;
11531 bool conservative = true;
11537 bpf_log(log, "Tracing programs must provide btf_id\n");
11540 btf = tgt_prog ? tgt_prog->aux->btf : btf_vmlinux;
11543 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
11546 t = btf_type_by_id(btf, btf_id);
11548 bpf_log(log, "attach_btf_id %u is invalid\n", btf_id);
11551 tname = btf_name_by_offset(btf, t->name_off);
11553 bpf_log(log, "attach_btf_id %u doesn't have a name\n", btf_id);
11557 struct bpf_prog_aux *aux = tgt_prog->aux;
11559 for (i = 0; i < aux->func_info_cnt; i++)
11560 if (aux->func_info[i].type_id == btf_id) {
11564 if (subprog == -1) {
11565 bpf_log(log, "Subprog %s doesn't exist\n", tname);
11568 conservative = aux->func_info_aux[subprog].unreliable;
11569 if (prog_extension) {
11570 if (conservative) {
11572 "Cannot replace static functions\n");
11575 if (!prog->jit_requested) {
11577 "Extension programs should be JITed\n");
11581 if (!tgt_prog->jited) {
11582 bpf_log(log, "Can attach to only JITed progs\n");
11585 if (tgt_prog->type == prog->type) {
11586 /* Cannot fentry/fexit another fentry/fexit program.
11587 * Cannot attach program extension to another extension.
11588 * It's ok to attach fentry/fexit to extension program.
11590 bpf_log(log, "Cannot recursively attach\n");
11593 if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
11595 (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
11596 tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
11597 /* Program extensions can extend all program types
11598 * except fentry/fexit. The reason is the following.
11599 * The fentry/fexit programs are used for performance
11600 * analysis, stats and can be attached to any program
11601 * type except themselves. When extension program is
11602 * replacing XDP function it is necessary to allow
11603 * performance analysis of all functions. Both original
11604 * XDP program and its program extension. Hence
11605 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
11606 * allowed. If extending of fentry/fexit was allowed it
11607 * would be possible to create long call chain
11608 * fentry->extension->fentry->extension beyond
11609 * reasonable stack size. Hence extending fentry is not
11612 bpf_log(log, "Cannot extend fentry/fexit\n");
11616 if (prog_extension) {
11617 bpf_log(log, "Cannot replace kernel functions\n");
11622 switch (prog->expected_attach_type) {
11623 case BPF_TRACE_RAW_TP:
11626 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
11629 if (!btf_type_is_typedef(t)) {
11630 bpf_log(log, "attach_btf_id %u is not a typedef\n",
11634 if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
11635 bpf_log(log, "attach_btf_id %u points to wrong type name %s\n",
11639 tname += sizeof(prefix) - 1;
11640 t = btf_type_by_id(btf, t->type);
11641 if (!btf_type_is_ptr(t))
11642 /* should never happen in valid vmlinux build */
11644 t = btf_type_by_id(btf, t->type);
11645 if (!btf_type_is_func_proto(t))
11646 /* should never happen in valid vmlinux build */
11650 case BPF_TRACE_ITER:
11651 if (!btf_type_is_func(t)) {
11652 bpf_log(log, "attach_btf_id %u is not a function\n",
11656 t = btf_type_by_id(btf, t->type);
11657 if (!btf_type_is_func_proto(t))
11659 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
11664 if (!prog_extension)
11667 case BPF_MODIFY_RETURN:
11669 case BPF_TRACE_FENTRY:
11670 case BPF_TRACE_FEXIT:
11671 if (!btf_type_is_func(t)) {
11672 bpf_log(log, "attach_btf_id %u is not a function\n",
11676 if (prog_extension &&
11677 btf_check_type_match(log, prog, btf, t))
11679 t = btf_type_by_id(btf, t->type);
11680 if (!btf_type_is_func_proto(t))
11683 if ((prog->aux->saved_dst_prog_type || prog->aux->saved_dst_attach_type) &&
11684 (!tgt_prog || prog->aux->saved_dst_prog_type != tgt_prog->type ||
11685 prog->aux->saved_dst_attach_type != tgt_prog->expected_attach_type))
11688 if (tgt_prog && conservative)
11691 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
11697 addr = (long) tgt_prog->bpf_func;
11699 addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
11701 addr = kallsyms_lookup_name(tname);
11704 "The address of function %s cannot be found\n",
11710 if (prog->aux->sleepable) {
11712 switch (prog->type) {
11713 case BPF_PROG_TYPE_TRACING:
11714 /* fentry/fexit/fmod_ret progs can be sleepable only if they are
11715 * attached to ALLOW_ERROR_INJECTION and are not in denylist.
11717 if (!check_non_sleepable_error_inject(btf_id) &&
11718 within_error_injection_list(addr))
11721 case BPF_PROG_TYPE_LSM:
11722 /* LSM progs check that they are attached to bpf_lsm_*() funcs.
11723 * Only some of them are sleepable.
11725 if (check_sleepable_lsm_hook(btf_id))
11732 bpf_log(log, "%s is not sleepable\n", tname);
11735 } else if (prog->expected_attach_type == BPF_MODIFY_RETURN) {
11737 bpf_log(log, "can't modify return codes of BPF programs\n");
11740 ret = check_attach_modify_return(addr, tname);
11742 bpf_log(log, "%s() is not modifiable\n", tname);
11749 tgt_info->tgt_addr = addr;
11750 tgt_info->tgt_name = tname;
11751 tgt_info->tgt_type = t;
11755 static int check_attach_btf_id(struct bpf_verifier_env *env)
11757 struct bpf_prog *prog = env->prog;
11758 struct bpf_prog *tgt_prog = prog->aux->dst_prog;
11759 struct bpf_attach_target_info tgt_info = {};
11760 u32 btf_id = prog->aux->attach_btf_id;
11761 struct bpf_trampoline *tr;
11765 if (prog->aux->sleepable && prog->type != BPF_PROG_TYPE_TRACING &&
11766 prog->type != BPF_PROG_TYPE_LSM) {
11767 verbose(env, "Only fentry/fexit/fmod_ret and lsm programs can be sleepable\n");
11771 if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
11772 return check_struct_ops_btf_id(env);
11774 if (prog->type != BPF_PROG_TYPE_TRACING &&
11775 prog->type != BPF_PROG_TYPE_LSM &&
11776 prog->type != BPF_PROG_TYPE_EXT)
11779 ret = bpf_check_attach_target(&env->log, prog, tgt_prog, btf_id, &tgt_info);
11783 if (tgt_prog && prog->type == BPF_PROG_TYPE_EXT) {
11784 /* to make freplace equivalent to their targets, they need to
11785 * inherit env->ops and expected_attach_type for the rest of the
11788 env->ops = bpf_verifier_ops[tgt_prog->type];
11789 prog->expected_attach_type = tgt_prog->expected_attach_type;
11792 /* store info about the attachment target that will be used later */
11793 prog->aux->attach_func_proto = tgt_info.tgt_type;
11794 prog->aux->attach_func_name = tgt_info.tgt_name;
11797 prog->aux->saved_dst_prog_type = tgt_prog->type;
11798 prog->aux->saved_dst_attach_type = tgt_prog->expected_attach_type;
11801 if (prog->expected_attach_type == BPF_TRACE_RAW_TP) {
11802 prog->aux->attach_btf_trace = true;
11804 } else if (prog->expected_attach_type == BPF_TRACE_ITER) {
11805 if (!bpf_iter_prog_supported(prog))
11810 if (prog->type == BPF_PROG_TYPE_LSM) {
11811 ret = bpf_lsm_verify_prog(&env->log, prog);
11816 key = bpf_trampoline_compute_key(tgt_prog, btf_id);
11817 tr = bpf_trampoline_get(key, &tgt_info);
11821 prog->aux->dst_trampoline = tr;
11825 struct btf *bpf_get_btf_vmlinux(void)
11827 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
11828 mutex_lock(&bpf_verifier_lock);
11830 btf_vmlinux = btf_parse_vmlinux();
11831 mutex_unlock(&bpf_verifier_lock);
11833 return btf_vmlinux;
11836 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr,
11837 union bpf_attr __user *uattr)
11839 u64 start_time = ktime_get_ns();
11840 struct bpf_verifier_env *env;
11841 struct bpf_verifier_log *log;
11842 int i, len, ret = -EINVAL;
11845 /* no program is valid */
11846 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
11849 /* 'struct bpf_verifier_env' can be global, but since it's not small,
11850 * allocate/free it every time bpf_check() is called
11852 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
11857 len = (*prog)->len;
11858 env->insn_aux_data =
11859 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
11861 if (!env->insn_aux_data)
11863 for (i = 0; i < len; i++)
11864 env->insn_aux_data[i].orig_idx = i;
11866 env->ops = bpf_verifier_ops[env->prog->type];
11867 is_priv = bpf_capable();
11869 bpf_get_btf_vmlinux();
11871 /* grab the mutex to protect few globals used by verifier */
11873 mutex_lock(&bpf_verifier_lock);
11875 if (attr->log_level || attr->log_buf || attr->log_size) {
11876 /* user requested verbose verifier output
11877 * and supplied buffer to store the verification trace
11879 log->level = attr->log_level;
11880 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
11881 log->len_total = attr->log_size;
11884 /* log attributes have to be sane */
11885 if (log->len_total < 128 || log->len_total > UINT_MAX >> 2 ||
11886 !log->level || !log->ubuf || log->level & ~BPF_LOG_MASK)
11890 if (IS_ERR(btf_vmlinux)) {
11891 /* Either gcc or pahole or kernel are broken. */
11892 verbose(env, "in-kernel BTF is malformed\n");
11893 ret = PTR_ERR(btf_vmlinux);
11894 goto skip_full_check;
11897 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
11898 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
11899 env->strict_alignment = true;
11900 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
11901 env->strict_alignment = false;
11903 env->allow_ptr_leaks = bpf_allow_ptr_leaks();
11904 env->allow_ptr_to_map_access = bpf_allow_ptr_to_map_access();
11905 env->bypass_spec_v1 = bpf_bypass_spec_v1();
11906 env->bypass_spec_v4 = bpf_bypass_spec_v4();
11907 env->bpf_capable = bpf_capable();
11910 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
11912 if (bpf_prog_is_dev_bound(env->prog->aux)) {
11913 ret = bpf_prog_offload_verifier_prep(env->prog);
11915 goto skip_full_check;
11918 env->explored_states = kvcalloc(state_htab_size(env),
11919 sizeof(struct bpf_verifier_state_list *),
11922 if (!env->explored_states)
11923 goto skip_full_check;
11925 ret = check_subprogs(env);
11927 goto skip_full_check;
11929 ret = check_btf_info(env, attr, uattr);
11931 goto skip_full_check;
11933 ret = check_attach_btf_id(env);
11935 goto skip_full_check;
11937 ret = resolve_pseudo_ldimm64(env);
11939 goto skip_full_check;
11941 ret = check_cfg(env);
11943 goto skip_full_check;
11945 ret = do_check_subprogs(env);
11946 ret = ret ?: do_check_main(env);
11948 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
11949 ret = bpf_prog_offload_finalize(env);
11952 kvfree(env->explored_states);
11955 ret = check_max_stack_depth(env);
11957 /* instruction rewrites happen after this point */
11960 opt_hard_wire_dead_code_branches(env);
11962 ret = opt_remove_dead_code(env);
11964 ret = opt_remove_nops(env);
11967 sanitize_dead_code(env);
11971 /* program is valid, convert *(u32*)(ctx + off) accesses */
11972 ret = convert_ctx_accesses(env);
11975 ret = fixup_bpf_calls(env);
11977 /* do 32-bit optimization after insn patching has done so those patched
11978 * insns could be handled correctly.
11980 if (ret == 0 && !bpf_prog_is_dev_bound(env->prog->aux)) {
11981 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
11982 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
11987 ret = fixup_call_args(env);
11989 env->verification_time = ktime_get_ns() - start_time;
11990 print_verification_stats(env);
11992 if (log->level && bpf_verifier_log_full(log))
11994 if (log->level && !log->ubuf) {
11996 goto err_release_maps;
11999 if (ret == 0 && env->used_map_cnt) {
12000 /* if program passed verifier, update used_maps in bpf_prog_info */
12001 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
12002 sizeof(env->used_maps[0]),
12005 if (!env->prog->aux->used_maps) {
12007 goto err_release_maps;
12010 memcpy(env->prog->aux->used_maps, env->used_maps,
12011 sizeof(env->used_maps[0]) * env->used_map_cnt);
12012 env->prog->aux->used_map_cnt = env->used_map_cnt;
12014 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
12015 * bpf_ld_imm64 instructions
12017 convert_pseudo_ld_imm64(env);
12021 adjust_btf_func(env);
12024 if (!env->prog->aux->used_maps)
12025 /* if we didn't copy map pointers into bpf_prog_info, release
12026 * them now. Otherwise free_used_maps() will release them.
12030 /* extension progs temporarily inherit the attach_type of their targets
12031 for verification purposes, so set it back to zero before returning
12033 if (env->prog->type == BPF_PROG_TYPE_EXT)
12034 env->prog->expected_attach_type = 0;
12039 mutex_unlock(&bpf_verifier_lock);
12040 vfree(env->insn_aux_data);