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;
247 struct btf *btf_vmlinux;
249 static DEFINE_MUTEX(bpf_verifier_lock);
251 static const struct bpf_line_info *
252 find_linfo(const struct bpf_verifier_env *env, u32 insn_off)
254 const struct bpf_line_info *linfo;
255 const struct bpf_prog *prog;
259 nr_linfo = prog->aux->nr_linfo;
261 if (!nr_linfo || insn_off >= prog->len)
264 linfo = prog->aux->linfo;
265 for (i = 1; i < nr_linfo; i++)
266 if (insn_off < linfo[i].insn_off)
269 return &linfo[i - 1];
272 void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
277 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
279 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
280 "verifier log line truncated - local buffer too short\n");
282 n = min(log->len_total - log->len_used - 1, n);
285 if (log->level == BPF_LOG_KERNEL) {
286 pr_err("BPF:%s\n", log->kbuf);
289 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
295 static void bpf_vlog_reset(struct bpf_verifier_log *log, u32 new_pos)
299 if (!bpf_verifier_log_needed(log))
302 log->len_used = new_pos;
303 if (put_user(zero, log->ubuf + new_pos))
307 /* log_level controls verbosity level of eBPF verifier.
308 * bpf_verifier_log_write() is used to dump the verification trace to the log,
309 * so the user can figure out what's wrong with the program
311 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
312 const char *fmt, ...)
316 if (!bpf_verifier_log_needed(&env->log))
320 bpf_verifier_vlog(&env->log, fmt, args);
323 EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
325 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
327 struct bpf_verifier_env *env = private_data;
330 if (!bpf_verifier_log_needed(&env->log))
334 bpf_verifier_vlog(&env->log, fmt, args);
338 __printf(2, 3) void bpf_log(struct bpf_verifier_log *log,
339 const char *fmt, ...)
343 if (!bpf_verifier_log_needed(log))
347 bpf_verifier_vlog(log, fmt, args);
351 static const char *ltrim(const char *s)
359 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env,
361 const char *prefix_fmt, ...)
363 const struct bpf_line_info *linfo;
365 if (!bpf_verifier_log_needed(&env->log))
368 linfo = find_linfo(env, insn_off);
369 if (!linfo || linfo == env->prev_linfo)
375 va_start(args, prefix_fmt);
376 bpf_verifier_vlog(&env->log, prefix_fmt, args);
381 ltrim(btf_name_by_offset(env->prog->aux->btf,
384 env->prev_linfo = linfo;
387 static bool type_is_pkt_pointer(enum bpf_reg_type type)
389 return type == PTR_TO_PACKET ||
390 type == PTR_TO_PACKET_META;
393 static bool type_is_sk_pointer(enum bpf_reg_type type)
395 return type == PTR_TO_SOCKET ||
396 type == PTR_TO_SOCK_COMMON ||
397 type == PTR_TO_TCP_SOCK ||
398 type == PTR_TO_XDP_SOCK;
401 static bool reg_type_not_null(enum bpf_reg_type type)
403 return type == PTR_TO_SOCKET ||
404 type == PTR_TO_TCP_SOCK ||
405 type == PTR_TO_MAP_VALUE ||
406 type == PTR_TO_SOCK_COMMON;
409 static bool reg_type_may_be_null(enum bpf_reg_type type)
411 return type == PTR_TO_MAP_VALUE_OR_NULL ||
412 type == PTR_TO_SOCKET_OR_NULL ||
413 type == PTR_TO_SOCK_COMMON_OR_NULL ||
414 type == PTR_TO_TCP_SOCK_OR_NULL ||
415 type == PTR_TO_BTF_ID_OR_NULL ||
416 type == PTR_TO_MEM_OR_NULL ||
417 type == PTR_TO_RDONLY_BUF_OR_NULL ||
418 type == PTR_TO_RDWR_BUF_OR_NULL;
421 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
423 return reg->type == PTR_TO_MAP_VALUE &&
424 map_value_has_spin_lock(reg->map_ptr);
427 static bool reg_type_may_be_refcounted_or_null(enum bpf_reg_type type)
429 return type == PTR_TO_SOCKET ||
430 type == PTR_TO_SOCKET_OR_NULL ||
431 type == PTR_TO_TCP_SOCK ||
432 type == PTR_TO_TCP_SOCK_OR_NULL ||
433 type == PTR_TO_MEM ||
434 type == PTR_TO_MEM_OR_NULL;
437 static bool arg_type_may_be_refcounted(enum bpf_arg_type type)
439 return type == ARG_PTR_TO_SOCK_COMMON;
442 static bool arg_type_may_be_null(enum bpf_arg_type type)
444 return type == ARG_PTR_TO_MAP_VALUE_OR_NULL ||
445 type == ARG_PTR_TO_MEM_OR_NULL ||
446 type == ARG_PTR_TO_CTX_OR_NULL ||
447 type == ARG_PTR_TO_SOCKET_OR_NULL ||
448 type == ARG_PTR_TO_ALLOC_MEM_OR_NULL;
451 /* Determine whether the function releases some resources allocated by another
452 * function call. The first reference type argument will be assumed to be
453 * released by release_reference().
455 static bool is_release_function(enum bpf_func_id func_id)
457 return func_id == BPF_FUNC_sk_release ||
458 func_id == BPF_FUNC_ringbuf_submit ||
459 func_id == BPF_FUNC_ringbuf_discard;
462 static bool may_be_acquire_function(enum bpf_func_id func_id)
464 return func_id == BPF_FUNC_sk_lookup_tcp ||
465 func_id == BPF_FUNC_sk_lookup_udp ||
466 func_id == BPF_FUNC_skc_lookup_tcp ||
467 func_id == BPF_FUNC_map_lookup_elem ||
468 func_id == BPF_FUNC_ringbuf_reserve;
471 static bool is_acquire_function(enum bpf_func_id func_id,
472 const struct bpf_map *map)
474 enum bpf_map_type map_type = map ? map->map_type : BPF_MAP_TYPE_UNSPEC;
476 if (func_id == BPF_FUNC_sk_lookup_tcp ||
477 func_id == BPF_FUNC_sk_lookup_udp ||
478 func_id == BPF_FUNC_skc_lookup_tcp ||
479 func_id == BPF_FUNC_ringbuf_reserve)
482 if (func_id == BPF_FUNC_map_lookup_elem &&
483 (map_type == BPF_MAP_TYPE_SOCKMAP ||
484 map_type == BPF_MAP_TYPE_SOCKHASH))
490 static bool is_ptr_cast_function(enum bpf_func_id func_id)
492 return func_id == BPF_FUNC_tcp_sock ||
493 func_id == BPF_FUNC_sk_fullsock ||
494 func_id == BPF_FUNC_skc_to_tcp_sock ||
495 func_id == BPF_FUNC_skc_to_tcp6_sock ||
496 func_id == BPF_FUNC_skc_to_udp6_sock ||
497 func_id == BPF_FUNC_skc_to_tcp_timewait_sock ||
498 func_id == BPF_FUNC_skc_to_tcp_request_sock;
501 /* string representation of 'enum bpf_reg_type' */
502 static const char * const reg_type_str[] = {
504 [SCALAR_VALUE] = "inv",
505 [PTR_TO_CTX] = "ctx",
506 [CONST_PTR_TO_MAP] = "map_ptr",
507 [PTR_TO_MAP_VALUE] = "map_value",
508 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
509 [PTR_TO_STACK] = "fp",
510 [PTR_TO_PACKET] = "pkt",
511 [PTR_TO_PACKET_META] = "pkt_meta",
512 [PTR_TO_PACKET_END] = "pkt_end",
513 [PTR_TO_FLOW_KEYS] = "flow_keys",
514 [PTR_TO_SOCKET] = "sock",
515 [PTR_TO_SOCKET_OR_NULL] = "sock_or_null",
516 [PTR_TO_SOCK_COMMON] = "sock_common",
517 [PTR_TO_SOCK_COMMON_OR_NULL] = "sock_common_or_null",
518 [PTR_TO_TCP_SOCK] = "tcp_sock",
519 [PTR_TO_TCP_SOCK_OR_NULL] = "tcp_sock_or_null",
520 [PTR_TO_TP_BUFFER] = "tp_buffer",
521 [PTR_TO_XDP_SOCK] = "xdp_sock",
522 [PTR_TO_BTF_ID] = "ptr_",
523 [PTR_TO_BTF_ID_OR_NULL] = "ptr_or_null_",
524 [PTR_TO_PERCPU_BTF_ID] = "percpu_ptr_",
525 [PTR_TO_MEM] = "mem",
526 [PTR_TO_MEM_OR_NULL] = "mem_or_null",
527 [PTR_TO_RDONLY_BUF] = "rdonly_buf",
528 [PTR_TO_RDONLY_BUF_OR_NULL] = "rdonly_buf_or_null",
529 [PTR_TO_RDWR_BUF] = "rdwr_buf",
530 [PTR_TO_RDWR_BUF_OR_NULL] = "rdwr_buf_or_null",
533 static char slot_type_char[] = {
534 [STACK_INVALID] = '?',
540 static void print_liveness(struct bpf_verifier_env *env,
541 enum bpf_reg_liveness live)
543 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
545 if (live & REG_LIVE_READ)
547 if (live & REG_LIVE_WRITTEN)
549 if (live & REG_LIVE_DONE)
553 static struct bpf_func_state *func(struct bpf_verifier_env *env,
554 const struct bpf_reg_state *reg)
556 struct bpf_verifier_state *cur = env->cur_state;
558 return cur->frame[reg->frameno];
561 static const char *kernel_type_name(const struct btf* btf, u32 id)
563 return btf_name_by_offset(btf, btf_type_by_id(btf, id)->name_off);
566 static void print_verifier_state(struct bpf_verifier_env *env,
567 const struct bpf_func_state *state)
569 const struct bpf_reg_state *reg;
574 verbose(env, " frame%d:", state->frameno);
575 for (i = 0; i < MAX_BPF_REG; i++) {
576 reg = &state->regs[i];
580 verbose(env, " R%d", i);
581 print_liveness(env, reg->live);
582 verbose(env, "=%s", reg_type_str[t]);
583 if (t == SCALAR_VALUE && reg->precise)
585 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
586 tnum_is_const(reg->var_off)) {
587 /* reg->off should be 0 for SCALAR_VALUE */
588 verbose(env, "%lld", reg->var_off.value + reg->off);
590 if (t == PTR_TO_BTF_ID ||
591 t == PTR_TO_BTF_ID_OR_NULL ||
592 t == PTR_TO_PERCPU_BTF_ID)
593 verbose(env, "%s", kernel_type_name(reg->btf, reg->btf_id));
594 verbose(env, "(id=%d", reg->id);
595 if (reg_type_may_be_refcounted_or_null(t))
596 verbose(env, ",ref_obj_id=%d", reg->ref_obj_id);
597 if (t != SCALAR_VALUE)
598 verbose(env, ",off=%d", reg->off);
599 if (type_is_pkt_pointer(t))
600 verbose(env, ",r=%d", reg->range);
601 else if (t == CONST_PTR_TO_MAP ||
602 t == PTR_TO_MAP_VALUE ||
603 t == PTR_TO_MAP_VALUE_OR_NULL)
604 verbose(env, ",ks=%d,vs=%d",
605 reg->map_ptr->key_size,
606 reg->map_ptr->value_size);
607 if (tnum_is_const(reg->var_off)) {
608 /* Typically an immediate SCALAR_VALUE, but
609 * could be a pointer whose offset is too big
612 verbose(env, ",imm=%llx", reg->var_off.value);
614 if (reg->smin_value != reg->umin_value &&
615 reg->smin_value != S64_MIN)
616 verbose(env, ",smin_value=%lld",
617 (long long)reg->smin_value);
618 if (reg->smax_value != reg->umax_value &&
619 reg->smax_value != S64_MAX)
620 verbose(env, ",smax_value=%lld",
621 (long long)reg->smax_value);
622 if (reg->umin_value != 0)
623 verbose(env, ",umin_value=%llu",
624 (unsigned long long)reg->umin_value);
625 if (reg->umax_value != U64_MAX)
626 verbose(env, ",umax_value=%llu",
627 (unsigned long long)reg->umax_value);
628 if (!tnum_is_unknown(reg->var_off)) {
631 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
632 verbose(env, ",var_off=%s", tn_buf);
634 if (reg->s32_min_value != reg->smin_value &&
635 reg->s32_min_value != S32_MIN)
636 verbose(env, ",s32_min_value=%d",
637 (int)(reg->s32_min_value));
638 if (reg->s32_max_value != reg->smax_value &&
639 reg->s32_max_value != S32_MAX)
640 verbose(env, ",s32_max_value=%d",
641 (int)(reg->s32_max_value));
642 if (reg->u32_min_value != reg->umin_value &&
643 reg->u32_min_value != U32_MIN)
644 verbose(env, ",u32_min_value=%d",
645 (int)(reg->u32_min_value));
646 if (reg->u32_max_value != reg->umax_value &&
647 reg->u32_max_value != U32_MAX)
648 verbose(env, ",u32_max_value=%d",
649 (int)(reg->u32_max_value));
654 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
655 char types_buf[BPF_REG_SIZE + 1];
659 for (j = 0; j < BPF_REG_SIZE; j++) {
660 if (state->stack[i].slot_type[j] != STACK_INVALID)
662 types_buf[j] = slot_type_char[
663 state->stack[i].slot_type[j]];
665 types_buf[BPF_REG_SIZE] = 0;
668 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
669 print_liveness(env, state->stack[i].spilled_ptr.live);
670 if (state->stack[i].slot_type[0] == STACK_SPILL) {
671 reg = &state->stack[i].spilled_ptr;
673 verbose(env, "=%s", reg_type_str[t]);
674 if (t == SCALAR_VALUE && reg->precise)
676 if (t == SCALAR_VALUE && tnum_is_const(reg->var_off))
677 verbose(env, "%lld", reg->var_off.value + reg->off);
679 verbose(env, "=%s", types_buf);
682 if (state->acquired_refs && state->refs[0].id) {
683 verbose(env, " refs=%d", state->refs[0].id);
684 for (i = 1; i < state->acquired_refs; i++)
685 if (state->refs[i].id)
686 verbose(env, ",%d", state->refs[i].id);
691 #define COPY_STATE_FN(NAME, COUNT, FIELD, SIZE) \
692 static int copy_##NAME##_state(struct bpf_func_state *dst, \
693 const struct bpf_func_state *src) \
697 if (WARN_ON_ONCE(dst->COUNT < src->COUNT)) { \
698 /* internal bug, make state invalid to reject the program */ \
699 memset(dst, 0, sizeof(*dst)); \
702 memcpy(dst->FIELD, src->FIELD, \
703 sizeof(*src->FIELD) * (src->COUNT / SIZE)); \
706 /* copy_reference_state() */
707 COPY_STATE_FN(reference, acquired_refs, refs, 1)
708 /* copy_stack_state() */
709 COPY_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
712 #define REALLOC_STATE_FN(NAME, COUNT, FIELD, SIZE) \
713 static int realloc_##NAME##_state(struct bpf_func_state *state, int size, \
716 u32 old_size = state->COUNT; \
717 struct bpf_##NAME##_state *new_##FIELD; \
718 int slot = size / SIZE; \
720 if (size <= old_size || !size) { \
723 state->COUNT = slot * SIZE; \
724 if (!size && old_size) { \
725 kfree(state->FIELD); \
726 state->FIELD = NULL; \
730 new_##FIELD = kmalloc_array(slot, sizeof(struct bpf_##NAME##_state), \
736 memcpy(new_##FIELD, state->FIELD, \
737 sizeof(*new_##FIELD) * (old_size / SIZE)); \
738 memset(new_##FIELD + old_size / SIZE, 0, \
739 sizeof(*new_##FIELD) * (size - old_size) / SIZE); \
741 state->COUNT = slot * SIZE; \
742 kfree(state->FIELD); \
743 state->FIELD = new_##FIELD; \
746 /* realloc_reference_state() */
747 REALLOC_STATE_FN(reference, acquired_refs, refs, 1)
748 /* realloc_stack_state() */
749 REALLOC_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
750 #undef REALLOC_STATE_FN
752 /* do_check() starts with zero-sized stack in struct bpf_verifier_state to
753 * make it consume minimal amount of memory. check_stack_write() access from
754 * the program calls into realloc_func_state() to grow the stack size.
755 * Note there is a non-zero 'parent' pointer inside bpf_verifier_state
756 * which realloc_stack_state() copies over. It points to previous
757 * bpf_verifier_state which is never reallocated.
759 static int realloc_func_state(struct bpf_func_state *state, int stack_size,
760 int refs_size, bool copy_old)
762 int err = realloc_reference_state(state, refs_size, copy_old);
765 return realloc_stack_state(state, stack_size, copy_old);
768 /* Acquire a pointer id from the env and update the state->refs to include
769 * this new pointer reference.
770 * On success, returns a valid pointer id to associate with the register
771 * On failure, returns a negative errno.
773 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
775 struct bpf_func_state *state = cur_func(env);
776 int new_ofs = state->acquired_refs;
779 err = realloc_reference_state(state, state->acquired_refs + 1, true);
783 state->refs[new_ofs].id = id;
784 state->refs[new_ofs].insn_idx = insn_idx;
789 /* release function corresponding to acquire_reference_state(). Idempotent. */
790 static int release_reference_state(struct bpf_func_state *state, int ptr_id)
794 last_idx = state->acquired_refs - 1;
795 for (i = 0; i < state->acquired_refs; i++) {
796 if (state->refs[i].id == ptr_id) {
797 if (last_idx && i != last_idx)
798 memcpy(&state->refs[i], &state->refs[last_idx],
799 sizeof(*state->refs));
800 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
801 state->acquired_refs--;
808 static int transfer_reference_state(struct bpf_func_state *dst,
809 struct bpf_func_state *src)
811 int err = realloc_reference_state(dst, src->acquired_refs, false);
814 err = copy_reference_state(dst, src);
820 static void free_func_state(struct bpf_func_state *state)
829 static void clear_jmp_history(struct bpf_verifier_state *state)
831 kfree(state->jmp_history);
832 state->jmp_history = NULL;
833 state->jmp_history_cnt = 0;
836 static void free_verifier_state(struct bpf_verifier_state *state,
841 for (i = 0; i <= state->curframe; i++) {
842 free_func_state(state->frame[i]);
843 state->frame[i] = NULL;
845 clear_jmp_history(state);
850 /* copy verifier state from src to dst growing dst stack space
851 * when necessary to accommodate larger src stack
853 static int copy_func_state(struct bpf_func_state *dst,
854 const struct bpf_func_state *src)
858 err = realloc_func_state(dst, src->allocated_stack, src->acquired_refs,
862 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
863 err = copy_reference_state(dst, src);
866 return copy_stack_state(dst, src);
869 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
870 const struct bpf_verifier_state *src)
872 struct bpf_func_state *dst;
873 u32 jmp_sz = sizeof(struct bpf_idx_pair) * src->jmp_history_cnt;
876 if (dst_state->jmp_history_cnt < src->jmp_history_cnt) {
877 kfree(dst_state->jmp_history);
878 dst_state->jmp_history = kmalloc(jmp_sz, GFP_USER);
879 if (!dst_state->jmp_history)
882 memcpy(dst_state->jmp_history, src->jmp_history, jmp_sz);
883 dst_state->jmp_history_cnt = src->jmp_history_cnt;
885 /* if dst has more stack frames then src frame, free them */
886 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
887 free_func_state(dst_state->frame[i]);
888 dst_state->frame[i] = NULL;
890 dst_state->speculative = src->speculative;
891 dst_state->curframe = src->curframe;
892 dst_state->active_spin_lock = src->active_spin_lock;
893 dst_state->branches = src->branches;
894 dst_state->parent = src->parent;
895 dst_state->first_insn_idx = src->first_insn_idx;
896 dst_state->last_insn_idx = src->last_insn_idx;
897 for (i = 0; i <= src->curframe; i++) {
898 dst = dst_state->frame[i];
900 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
903 dst_state->frame[i] = dst;
905 err = copy_func_state(dst, src->frame[i]);
912 static void update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
915 u32 br = --st->branches;
917 /* WARN_ON(br > 1) technically makes sense here,
918 * but see comment in push_stack(), hence:
920 WARN_ONCE((int)br < 0,
921 "BUG update_branch_counts:branches_to_explore=%d\n",
929 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
930 int *insn_idx, bool pop_log)
932 struct bpf_verifier_state *cur = env->cur_state;
933 struct bpf_verifier_stack_elem *elem, *head = env->head;
936 if (env->head == NULL)
940 err = copy_verifier_state(cur, &head->st);
945 bpf_vlog_reset(&env->log, head->log_pos);
947 *insn_idx = head->insn_idx;
949 *prev_insn_idx = head->prev_insn_idx;
951 free_verifier_state(&head->st, false);
958 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
959 int insn_idx, int prev_insn_idx,
962 struct bpf_verifier_state *cur = env->cur_state;
963 struct bpf_verifier_stack_elem *elem;
966 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
970 elem->insn_idx = insn_idx;
971 elem->prev_insn_idx = prev_insn_idx;
972 elem->next = env->head;
973 elem->log_pos = env->log.len_used;
976 err = copy_verifier_state(&elem->st, cur);
979 elem->st.speculative |= speculative;
980 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
981 verbose(env, "The sequence of %d jumps is too complex.\n",
985 if (elem->st.parent) {
986 ++elem->st.parent->branches;
987 /* WARN_ON(branches > 2) technically makes sense here,
989 * 1. speculative states will bump 'branches' for non-branch
991 * 2. is_state_visited() heuristics may decide not to create
992 * a new state for a sequence of branches and all such current
993 * and cloned states will be pointing to a single parent state
994 * which might have large 'branches' count.
999 free_verifier_state(env->cur_state, true);
1000 env->cur_state = NULL;
1001 /* pop all elements and return */
1002 while (!pop_stack(env, NULL, NULL, false));
1006 #define CALLER_SAVED_REGS 6
1007 static const int caller_saved[CALLER_SAVED_REGS] = {
1008 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
1011 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1012 struct bpf_reg_state *reg);
1014 /* This helper doesn't clear reg->id */
1015 static void ___mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1017 reg->var_off = tnum_const(imm);
1018 reg->smin_value = (s64)imm;
1019 reg->smax_value = (s64)imm;
1020 reg->umin_value = imm;
1021 reg->umax_value = imm;
1023 reg->s32_min_value = (s32)imm;
1024 reg->s32_max_value = (s32)imm;
1025 reg->u32_min_value = (u32)imm;
1026 reg->u32_max_value = (u32)imm;
1029 /* Mark the unknown part of a register (variable offset or scalar value) as
1030 * known to have the value @imm.
1032 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1034 /* Clear id, off, and union(map_ptr, range) */
1035 memset(((u8 *)reg) + sizeof(reg->type), 0,
1036 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
1037 ___mark_reg_known(reg, imm);
1040 static void __mark_reg32_known(struct bpf_reg_state *reg, u64 imm)
1042 reg->var_off = tnum_const_subreg(reg->var_off, imm);
1043 reg->s32_min_value = (s32)imm;
1044 reg->s32_max_value = (s32)imm;
1045 reg->u32_min_value = (u32)imm;
1046 reg->u32_max_value = (u32)imm;
1049 /* Mark the 'variable offset' part of a register as zero. This should be
1050 * used only on registers holding a pointer type.
1052 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
1054 __mark_reg_known(reg, 0);
1057 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
1059 __mark_reg_known(reg, 0);
1060 reg->type = SCALAR_VALUE;
1063 static void mark_reg_known_zero(struct bpf_verifier_env *env,
1064 struct bpf_reg_state *regs, u32 regno)
1066 if (WARN_ON(regno >= MAX_BPF_REG)) {
1067 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
1068 /* Something bad happened, let's kill all regs */
1069 for (regno = 0; regno < MAX_BPF_REG; regno++)
1070 __mark_reg_not_init(env, regs + regno);
1073 __mark_reg_known_zero(regs + regno);
1076 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
1078 return type_is_pkt_pointer(reg->type);
1081 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
1083 return reg_is_pkt_pointer(reg) ||
1084 reg->type == PTR_TO_PACKET_END;
1087 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
1088 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
1089 enum bpf_reg_type which)
1091 /* The register can already have a range from prior markings.
1092 * This is fine as long as it hasn't been advanced from its
1095 return reg->type == which &&
1098 tnum_equals_const(reg->var_off, 0);
1101 /* Reset the min/max bounds of a register */
1102 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
1104 reg->smin_value = S64_MIN;
1105 reg->smax_value = S64_MAX;
1106 reg->umin_value = 0;
1107 reg->umax_value = U64_MAX;
1109 reg->s32_min_value = S32_MIN;
1110 reg->s32_max_value = S32_MAX;
1111 reg->u32_min_value = 0;
1112 reg->u32_max_value = U32_MAX;
1115 static void __mark_reg64_unbounded(struct bpf_reg_state *reg)
1117 reg->smin_value = S64_MIN;
1118 reg->smax_value = S64_MAX;
1119 reg->umin_value = 0;
1120 reg->umax_value = U64_MAX;
1123 static void __mark_reg32_unbounded(struct bpf_reg_state *reg)
1125 reg->s32_min_value = S32_MIN;
1126 reg->s32_max_value = S32_MAX;
1127 reg->u32_min_value = 0;
1128 reg->u32_max_value = U32_MAX;
1131 static void __update_reg32_bounds(struct bpf_reg_state *reg)
1133 struct tnum var32_off = tnum_subreg(reg->var_off);
1135 /* min signed is max(sign bit) | min(other bits) */
1136 reg->s32_min_value = max_t(s32, reg->s32_min_value,
1137 var32_off.value | (var32_off.mask & S32_MIN));
1138 /* max signed is min(sign bit) | max(other bits) */
1139 reg->s32_max_value = min_t(s32, reg->s32_max_value,
1140 var32_off.value | (var32_off.mask & S32_MAX));
1141 reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)var32_off.value);
1142 reg->u32_max_value = min(reg->u32_max_value,
1143 (u32)(var32_off.value | var32_off.mask));
1146 static void __update_reg64_bounds(struct bpf_reg_state *reg)
1148 /* min signed is max(sign bit) | min(other bits) */
1149 reg->smin_value = max_t(s64, reg->smin_value,
1150 reg->var_off.value | (reg->var_off.mask & S64_MIN));
1151 /* max signed is min(sign bit) | max(other bits) */
1152 reg->smax_value = min_t(s64, reg->smax_value,
1153 reg->var_off.value | (reg->var_off.mask & S64_MAX));
1154 reg->umin_value = max(reg->umin_value, reg->var_off.value);
1155 reg->umax_value = min(reg->umax_value,
1156 reg->var_off.value | reg->var_off.mask);
1159 static void __update_reg_bounds(struct bpf_reg_state *reg)
1161 __update_reg32_bounds(reg);
1162 __update_reg64_bounds(reg);
1165 /* Uses signed min/max values to inform unsigned, and vice-versa */
1166 static void __reg32_deduce_bounds(struct bpf_reg_state *reg)
1168 /* Learn sign from signed bounds.
1169 * If we cannot cross the sign boundary, then signed and unsigned bounds
1170 * are the same, so combine. This works even in the negative case, e.g.
1171 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1173 if (reg->s32_min_value >= 0 || reg->s32_max_value < 0) {
1174 reg->s32_min_value = reg->u32_min_value =
1175 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1176 reg->s32_max_value = reg->u32_max_value =
1177 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1180 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1181 * boundary, so we must be careful.
1183 if ((s32)reg->u32_max_value >= 0) {
1184 /* Positive. We can't learn anything from the smin, but smax
1185 * is positive, hence safe.
1187 reg->s32_min_value = reg->u32_min_value;
1188 reg->s32_max_value = reg->u32_max_value =
1189 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1190 } else if ((s32)reg->u32_min_value < 0) {
1191 /* Negative. We can't learn anything from the smax, but smin
1192 * is negative, hence safe.
1194 reg->s32_min_value = reg->u32_min_value =
1195 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1196 reg->s32_max_value = reg->u32_max_value;
1200 static void __reg64_deduce_bounds(struct bpf_reg_state *reg)
1202 /* Learn sign from signed bounds.
1203 * If we cannot cross the sign boundary, then signed and unsigned bounds
1204 * are the same, so combine. This works even in the negative case, e.g.
1205 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1207 if (reg->smin_value >= 0 || reg->smax_value < 0) {
1208 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1210 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1214 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1215 * boundary, so we must be careful.
1217 if ((s64)reg->umax_value >= 0) {
1218 /* Positive. We can't learn anything from the smin, but smax
1219 * is positive, hence safe.
1221 reg->smin_value = reg->umin_value;
1222 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1224 } else if ((s64)reg->umin_value < 0) {
1225 /* Negative. We can't learn anything from the smax, but smin
1226 * is negative, hence safe.
1228 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1230 reg->smax_value = reg->umax_value;
1234 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
1236 __reg32_deduce_bounds(reg);
1237 __reg64_deduce_bounds(reg);
1240 /* Attempts to improve var_off based on unsigned min/max information */
1241 static void __reg_bound_offset(struct bpf_reg_state *reg)
1243 struct tnum var64_off = tnum_intersect(reg->var_off,
1244 tnum_range(reg->umin_value,
1246 struct tnum var32_off = tnum_intersect(tnum_subreg(reg->var_off),
1247 tnum_range(reg->u32_min_value,
1248 reg->u32_max_value));
1250 reg->var_off = tnum_or(tnum_clear_subreg(var64_off), var32_off);
1253 static void __reg_assign_32_into_64(struct bpf_reg_state *reg)
1255 reg->umin_value = reg->u32_min_value;
1256 reg->umax_value = reg->u32_max_value;
1257 /* Attempt to pull 32-bit signed bounds into 64-bit bounds
1258 * but must be positive otherwise set to worse case bounds
1259 * and refine later from tnum.
1261 if (reg->s32_min_value >= 0 && reg->s32_max_value >= 0)
1262 reg->smax_value = reg->s32_max_value;
1264 reg->smax_value = U32_MAX;
1265 if (reg->s32_min_value >= 0)
1266 reg->smin_value = reg->s32_min_value;
1268 reg->smin_value = 0;
1271 static void __reg_combine_32_into_64(struct bpf_reg_state *reg)
1273 /* special case when 64-bit register has upper 32-bit register
1274 * zeroed. Typically happens after zext or <<32, >>32 sequence
1275 * allowing us to use 32-bit bounds directly,
1277 if (tnum_equals_const(tnum_clear_subreg(reg->var_off), 0)) {
1278 __reg_assign_32_into_64(reg);
1280 /* Otherwise the best we can do is push lower 32bit known and
1281 * unknown bits into register (var_off set from jmp logic)
1282 * then learn as much as possible from the 64-bit tnum
1283 * known and unknown bits. The previous smin/smax bounds are
1284 * invalid here because of jmp32 compare so mark them unknown
1285 * so they do not impact tnum bounds calculation.
1287 __mark_reg64_unbounded(reg);
1288 __update_reg_bounds(reg);
1291 /* Intersecting with the old var_off might have improved our bounds
1292 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1293 * then new var_off is (0; 0x7f...fc) which improves our umax.
1295 __reg_deduce_bounds(reg);
1296 __reg_bound_offset(reg);
1297 __update_reg_bounds(reg);
1300 static bool __reg64_bound_s32(s64 a)
1302 return a > S32_MIN && a < S32_MAX;
1305 static bool __reg64_bound_u32(u64 a)
1307 if (a > U32_MIN && a < U32_MAX)
1312 static void __reg_combine_64_into_32(struct bpf_reg_state *reg)
1314 __mark_reg32_unbounded(reg);
1316 if (__reg64_bound_s32(reg->smin_value) && __reg64_bound_s32(reg->smax_value)) {
1317 reg->s32_min_value = (s32)reg->smin_value;
1318 reg->s32_max_value = (s32)reg->smax_value;
1320 if (__reg64_bound_u32(reg->umin_value))
1321 reg->u32_min_value = (u32)reg->umin_value;
1322 if (__reg64_bound_u32(reg->umax_value))
1323 reg->u32_max_value = (u32)reg->umax_value;
1325 /* Intersecting with the old var_off might have improved our bounds
1326 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1327 * then new var_off is (0; 0x7f...fc) which improves our umax.
1329 __reg_deduce_bounds(reg);
1330 __reg_bound_offset(reg);
1331 __update_reg_bounds(reg);
1334 /* Mark a register as having a completely unknown (scalar) value. */
1335 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
1336 struct bpf_reg_state *reg)
1339 * Clear type, id, off, and union(map_ptr, range) and
1340 * padding between 'type' and union
1342 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
1343 reg->type = SCALAR_VALUE;
1344 reg->var_off = tnum_unknown;
1346 reg->precise = env->subprog_cnt > 1 || !env->bpf_capable;
1347 __mark_reg_unbounded(reg);
1350 static void mark_reg_unknown(struct bpf_verifier_env *env,
1351 struct bpf_reg_state *regs, u32 regno)
1353 if (WARN_ON(regno >= MAX_BPF_REG)) {
1354 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
1355 /* Something bad happened, let's kill all regs except FP */
1356 for (regno = 0; regno < BPF_REG_FP; regno++)
1357 __mark_reg_not_init(env, regs + regno);
1360 __mark_reg_unknown(env, regs + regno);
1363 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1364 struct bpf_reg_state *reg)
1366 __mark_reg_unknown(env, reg);
1367 reg->type = NOT_INIT;
1370 static void mark_reg_not_init(struct bpf_verifier_env *env,
1371 struct bpf_reg_state *regs, u32 regno)
1373 if (WARN_ON(regno >= MAX_BPF_REG)) {
1374 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
1375 /* Something bad happened, let's kill all regs except FP */
1376 for (regno = 0; regno < BPF_REG_FP; regno++)
1377 __mark_reg_not_init(env, regs + regno);
1380 __mark_reg_not_init(env, regs + regno);
1383 static void mark_btf_ld_reg(struct bpf_verifier_env *env,
1384 struct bpf_reg_state *regs, u32 regno,
1385 enum bpf_reg_type reg_type,
1386 struct btf *btf, u32 btf_id)
1388 if (reg_type == SCALAR_VALUE) {
1389 mark_reg_unknown(env, regs, regno);
1392 mark_reg_known_zero(env, regs, regno);
1393 regs[regno].type = PTR_TO_BTF_ID;
1394 regs[regno].btf = btf;
1395 regs[regno].btf_id = btf_id;
1398 #define DEF_NOT_SUBREG (0)
1399 static void init_reg_state(struct bpf_verifier_env *env,
1400 struct bpf_func_state *state)
1402 struct bpf_reg_state *regs = state->regs;
1405 for (i = 0; i < MAX_BPF_REG; i++) {
1406 mark_reg_not_init(env, regs, i);
1407 regs[i].live = REG_LIVE_NONE;
1408 regs[i].parent = NULL;
1409 regs[i].subreg_def = DEF_NOT_SUBREG;
1413 regs[BPF_REG_FP].type = PTR_TO_STACK;
1414 mark_reg_known_zero(env, regs, BPF_REG_FP);
1415 regs[BPF_REG_FP].frameno = state->frameno;
1418 #define BPF_MAIN_FUNC (-1)
1419 static void init_func_state(struct bpf_verifier_env *env,
1420 struct bpf_func_state *state,
1421 int callsite, int frameno, int subprogno)
1423 state->callsite = callsite;
1424 state->frameno = frameno;
1425 state->subprogno = subprogno;
1426 init_reg_state(env, state);
1430 SRC_OP, /* register is used as source operand */
1431 DST_OP, /* register is used as destination operand */
1432 DST_OP_NO_MARK /* same as above, check only, don't mark */
1435 static int cmp_subprogs(const void *a, const void *b)
1437 return ((struct bpf_subprog_info *)a)->start -
1438 ((struct bpf_subprog_info *)b)->start;
1441 static int find_subprog(struct bpf_verifier_env *env, int off)
1443 struct bpf_subprog_info *p;
1445 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
1446 sizeof(env->subprog_info[0]), cmp_subprogs);
1449 return p - env->subprog_info;
1453 static int add_subprog(struct bpf_verifier_env *env, int off)
1455 int insn_cnt = env->prog->len;
1458 if (off >= insn_cnt || off < 0) {
1459 verbose(env, "call to invalid destination\n");
1462 ret = find_subprog(env, off);
1465 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
1466 verbose(env, "too many subprograms\n");
1469 env->subprog_info[env->subprog_cnt++].start = off;
1470 sort(env->subprog_info, env->subprog_cnt,
1471 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
1475 static int check_subprogs(struct bpf_verifier_env *env)
1477 int i, ret, subprog_start, subprog_end, off, cur_subprog = 0;
1478 struct bpf_subprog_info *subprog = env->subprog_info;
1479 struct bpf_insn *insn = env->prog->insnsi;
1480 int insn_cnt = env->prog->len;
1482 /* Add entry function. */
1483 ret = add_subprog(env, 0);
1487 /* determine subprog starts. The end is one before the next starts */
1488 for (i = 0; i < insn_cnt; i++) {
1489 if (insn[i].code != (BPF_JMP | BPF_CALL))
1491 if (insn[i].src_reg != BPF_PSEUDO_CALL)
1493 if (!env->bpf_capable) {
1495 "function calls to other bpf functions are allowed for CAP_BPF and CAP_SYS_ADMIN\n");
1498 ret = add_subprog(env, i + insn[i].imm + 1);
1503 /* Add a fake 'exit' subprog which could simplify subprog iteration
1504 * logic. 'subprog_cnt' should not be increased.
1506 subprog[env->subprog_cnt].start = insn_cnt;
1508 if (env->log.level & BPF_LOG_LEVEL2)
1509 for (i = 0; i < env->subprog_cnt; i++)
1510 verbose(env, "func#%d @%d\n", i, subprog[i].start);
1512 /* now check that all jumps are within the same subprog */
1513 subprog_start = subprog[cur_subprog].start;
1514 subprog_end = subprog[cur_subprog + 1].start;
1515 for (i = 0; i < insn_cnt; i++) {
1516 u8 code = insn[i].code;
1518 if (code == (BPF_JMP | BPF_CALL) &&
1519 insn[i].imm == BPF_FUNC_tail_call &&
1520 insn[i].src_reg != BPF_PSEUDO_CALL)
1521 subprog[cur_subprog].has_tail_call = true;
1522 if (BPF_CLASS(code) == BPF_LD &&
1523 (BPF_MODE(code) == BPF_ABS || BPF_MODE(code) == BPF_IND))
1524 subprog[cur_subprog].has_ld_abs = true;
1525 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
1527 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
1529 off = i + insn[i].off + 1;
1530 if (off < subprog_start || off >= subprog_end) {
1531 verbose(env, "jump out of range from insn %d to %d\n", i, off);
1535 if (i == subprog_end - 1) {
1536 /* to avoid fall-through from one subprog into another
1537 * the last insn of the subprog should be either exit
1538 * or unconditional jump back
1540 if (code != (BPF_JMP | BPF_EXIT) &&
1541 code != (BPF_JMP | BPF_JA)) {
1542 verbose(env, "last insn is not an exit or jmp\n");
1545 subprog_start = subprog_end;
1547 if (cur_subprog < env->subprog_cnt)
1548 subprog_end = subprog[cur_subprog + 1].start;
1554 /* Parentage chain of this register (or stack slot) should take care of all
1555 * issues like callee-saved registers, stack slot allocation time, etc.
1557 static int mark_reg_read(struct bpf_verifier_env *env,
1558 const struct bpf_reg_state *state,
1559 struct bpf_reg_state *parent, u8 flag)
1561 bool writes = parent == state->parent; /* Observe write marks */
1565 /* if read wasn't screened by an earlier write ... */
1566 if (writes && state->live & REG_LIVE_WRITTEN)
1568 if (parent->live & REG_LIVE_DONE) {
1569 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
1570 reg_type_str[parent->type],
1571 parent->var_off.value, parent->off);
1574 /* The first condition is more likely to be true than the
1575 * second, checked it first.
1577 if ((parent->live & REG_LIVE_READ) == flag ||
1578 parent->live & REG_LIVE_READ64)
1579 /* The parentage chain never changes and
1580 * this parent was already marked as LIVE_READ.
1581 * There is no need to keep walking the chain again and
1582 * keep re-marking all parents as LIVE_READ.
1583 * This case happens when the same register is read
1584 * multiple times without writes into it in-between.
1585 * Also, if parent has the stronger REG_LIVE_READ64 set,
1586 * then no need to set the weak REG_LIVE_READ32.
1589 /* ... then we depend on parent's value */
1590 parent->live |= flag;
1591 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
1592 if (flag == REG_LIVE_READ64)
1593 parent->live &= ~REG_LIVE_READ32;
1595 parent = state->parent;
1600 if (env->longest_mark_read_walk < cnt)
1601 env->longest_mark_read_walk = cnt;
1605 /* This function is supposed to be used by the following 32-bit optimization
1606 * code only. It returns TRUE if the source or destination register operates
1607 * on 64-bit, otherwise return FALSE.
1609 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
1610 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
1615 class = BPF_CLASS(code);
1617 if (class == BPF_JMP) {
1618 /* BPF_EXIT for "main" will reach here. Return TRUE
1623 if (op == BPF_CALL) {
1624 /* BPF to BPF call will reach here because of marking
1625 * caller saved clobber with DST_OP_NO_MARK for which we
1626 * don't care the register def because they are anyway
1627 * marked as NOT_INIT already.
1629 if (insn->src_reg == BPF_PSEUDO_CALL)
1631 /* Helper call will reach here because of arg type
1632 * check, conservatively return TRUE.
1641 if (class == BPF_ALU64 || class == BPF_JMP ||
1642 /* BPF_END always use BPF_ALU class. */
1643 (class == BPF_ALU && op == BPF_END && insn->imm == 64))
1646 if (class == BPF_ALU || class == BPF_JMP32)
1649 if (class == BPF_LDX) {
1651 return BPF_SIZE(code) == BPF_DW;
1652 /* LDX source must be ptr. */
1656 if (class == BPF_STX) {
1657 if (reg->type != SCALAR_VALUE)
1659 return BPF_SIZE(code) == BPF_DW;
1662 if (class == BPF_LD) {
1663 u8 mode = BPF_MODE(code);
1666 if (mode == BPF_IMM)
1669 /* Both LD_IND and LD_ABS return 32-bit data. */
1673 /* Implicit ctx ptr. */
1674 if (regno == BPF_REG_6)
1677 /* Explicit source could be any width. */
1681 if (class == BPF_ST)
1682 /* The only source register for BPF_ST is a ptr. */
1685 /* Conservatively return true at default. */
1689 /* Return TRUE if INSN doesn't have explicit value define. */
1690 static bool insn_no_def(struct bpf_insn *insn)
1692 u8 class = BPF_CLASS(insn->code);
1694 return (class == BPF_JMP || class == BPF_JMP32 ||
1695 class == BPF_STX || class == BPF_ST);
1698 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
1699 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
1701 if (insn_no_def(insn))
1704 return !is_reg64(env, insn, insn->dst_reg, NULL, DST_OP);
1707 static void mark_insn_zext(struct bpf_verifier_env *env,
1708 struct bpf_reg_state *reg)
1710 s32 def_idx = reg->subreg_def;
1712 if (def_idx == DEF_NOT_SUBREG)
1715 env->insn_aux_data[def_idx - 1].zext_dst = true;
1716 /* The dst will be zero extended, so won't be sub-register anymore. */
1717 reg->subreg_def = DEF_NOT_SUBREG;
1720 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
1721 enum reg_arg_type t)
1723 struct bpf_verifier_state *vstate = env->cur_state;
1724 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1725 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
1726 struct bpf_reg_state *reg, *regs = state->regs;
1729 if (regno >= MAX_BPF_REG) {
1730 verbose(env, "R%d is invalid\n", regno);
1735 rw64 = is_reg64(env, insn, regno, reg, t);
1737 /* check whether register used as source operand can be read */
1738 if (reg->type == NOT_INIT) {
1739 verbose(env, "R%d !read_ok\n", regno);
1742 /* We don't need to worry about FP liveness because it's read-only */
1743 if (regno == BPF_REG_FP)
1747 mark_insn_zext(env, reg);
1749 return mark_reg_read(env, reg, reg->parent,
1750 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
1752 /* check whether register used as dest operand can be written to */
1753 if (regno == BPF_REG_FP) {
1754 verbose(env, "frame pointer is read only\n");
1757 reg->live |= REG_LIVE_WRITTEN;
1758 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
1760 mark_reg_unknown(env, regs, regno);
1765 /* for any branch, call, exit record the history of jmps in the given state */
1766 static int push_jmp_history(struct bpf_verifier_env *env,
1767 struct bpf_verifier_state *cur)
1769 u32 cnt = cur->jmp_history_cnt;
1770 struct bpf_idx_pair *p;
1773 p = krealloc(cur->jmp_history, cnt * sizeof(*p), GFP_USER);
1776 p[cnt - 1].idx = env->insn_idx;
1777 p[cnt - 1].prev_idx = env->prev_insn_idx;
1778 cur->jmp_history = p;
1779 cur->jmp_history_cnt = cnt;
1783 /* Backtrack one insn at a time. If idx is not at the top of recorded
1784 * history then previous instruction came from straight line execution.
1786 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
1791 if (cnt && st->jmp_history[cnt - 1].idx == i) {
1792 i = st->jmp_history[cnt - 1].prev_idx;
1800 /* For given verifier state backtrack_insn() is called from the last insn to
1801 * the first insn. Its purpose is to compute a bitmask of registers and
1802 * stack slots that needs precision in the parent verifier state.
1804 static int backtrack_insn(struct bpf_verifier_env *env, int idx,
1805 u32 *reg_mask, u64 *stack_mask)
1807 const struct bpf_insn_cbs cbs = {
1808 .cb_print = verbose,
1809 .private_data = env,
1811 struct bpf_insn *insn = env->prog->insnsi + idx;
1812 u8 class = BPF_CLASS(insn->code);
1813 u8 opcode = BPF_OP(insn->code);
1814 u8 mode = BPF_MODE(insn->code);
1815 u32 dreg = 1u << insn->dst_reg;
1816 u32 sreg = 1u << insn->src_reg;
1819 if (insn->code == 0)
1821 if (env->log.level & BPF_LOG_LEVEL) {
1822 verbose(env, "regs=%x stack=%llx before ", *reg_mask, *stack_mask);
1823 verbose(env, "%d: ", idx);
1824 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
1827 if (class == BPF_ALU || class == BPF_ALU64) {
1828 if (!(*reg_mask & dreg))
1830 if (opcode == BPF_MOV) {
1831 if (BPF_SRC(insn->code) == BPF_X) {
1833 * dreg needs precision after this insn
1834 * sreg needs precision before this insn
1840 * dreg needs precision after this insn.
1841 * Corresponding register is already marked
1842 * as precise=true in this verifier state.
1843 * No further markings in parent are necessary
1848 if (BPF_SRC(insn->code) == BPF_X) {
1850 * both dreg and sreg need precision
1855 * dreg still needs precision before this insn
1858 } else if (class == BPF_LDX) {
1859 if (!(*reg_mask & dreg))
1863 /* scalars can only be spilled into stack w/o losing precision.
1864 * Load from any other memory can be zero extended.
1865 * The desire to keep that precision is already indicated
1866 * by 'precise' mark in corresponding register of this state.
1867 * No further tracking necessary.
1869 if (insn->src_reg != BPF_REG_FP)
1871 if (BPF_SIZE(insn->code) != BPF_DW)
1874 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
1875 * that [fp - off] slot contains scalar that needs to be
1876 * tracked with precision
1878 spi = (-insn->off - 1) / BPF_REG_SIZE;
1880 verbose(env, "BUG spi %d\n", spi);
1881 WARN_ONCE(1, "verifier backtracking bug");
1884 *stack_mask |= 1ull << spi;
1885 } else if (class == BPF_STX || class == BPF_ST) {
1886 if (*reg_mask & dreg)
1887 /* stx & st shouldn't be using _scalar_ dst_reg
1888 * to access memory. It means backtracking
1889 * encountered a case of pointer subtraction.
1892 /* scalars can only be spilled into stack */
1893 if (insn->dst_reg != BPF_REG_FP)
1895 if (BPF_SIZE(insn->code) != BPF_DW)
1897 spi = (-insn->off - 1) / BPF_REG_SIZE;
1899 verbose(env, "BUG spi %d\n", spi);
1900 WARN_ONCE(1, "verifier backtracking bug");
1903 if (!(*stack_mask & (1ull << spi)))
1905 *stack_mask &= ~(1ull << spi);
1906 if (class == BPF_STX)
1908 } else if (class == BPF_JMP || class == BPF_JMP32) {
1909 if (opcode == BPF_CALL) {
1910 if (insn->src_reg == BPF_PSEUDO_CALL)
1912 /* regular helper call sets R0 */
1914 if (*reg_mask & 0x3f) {
1915 /* if backtracing was looking for registers R1-R5
1916 * they should have been found already.
1918 verbose(env, "BUG regs %x\n", *reg_mask);
1919 WARN_ONCE(1, "verifier backtracking bug");
1922 } else if (opcode == BPF_EXIT) {
1925 } else if (class == BPF_LD) {
1926 if (!(*reg_mask & dreg))
1929 /* It's ld_imm64 or ld_abs or ld_ind.
1930 * For ld_imm64 no further tracking of precision
1931 * into parent is necessary
1933 if (mode == BPF_IND || mode == BPF_ABS)
1934 /* to be analyzed */
1940 /* the scalar precision tracking algorithm:
1941 * . at the start all registers have precise=false.
1942 * . scalar ranges are tracked as normal through alu and jmp insns.
1943 * . once precise value of the scalar register is used in:
1944 * . ptr + scalar alu
1945 * . if (scalar cond K|scalar)
1946 * . helper_call(.., scalar, ...) where ARG_CONST is expected
1947 * backtrack through the verifier states and mark all registers and
1948 * stack slots with spilled constants that these scalar regisers
1949 * should be precise.
1950 * . during state pruning two registers (or spilled stack slots)
1951 * are equivalent if both are not precise.
1953 * Note the verifier cannot simply walk register parentage chain,
1954 * since many different registers and stack slots could have been
1955 * used to compute single precise scalar.
1957 * The approach of starting with precise=true for all registers and then
1958 * backtrack to mark a register as not precise when the verifier detects
1959 * that program doesn't care about specific value (e.g., when helper
1960 * takes register as ARG_ANYTHING parameter) is not safe.
1962 * It's ok to walk single parentage chain of the verifier states.
1963 * It's possible that this backtracking will go all the way till 1st insn.
1964 * All other branches will be explored for needing precision later.
1966 * The backtracking needs to deal with cases like:
1967 * 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)
1970 * if r5 > 0x79f goto pc+7
1971 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
1974 * call bpf_perf_event_output#25
1975 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
1979 * call foo // uses callee's r6 inside to compute r0
1983 * to track above reg_mask/stack_mask needs to be independent for each frame.
1985 * Also if parent's curframe > frame where backtracking started,
1986 * the verifier need to mark registers in both frames, otherwise callees
1987 * may incorrectly prune callers. This is similar to
1988 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
1990 * For now backtracking falls back into conservative marking.
1992 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
1993 struct bpf_verifier_state *st)
1995 struct bpf_func_state *func;
1996 struct bpf_reg_state *reg;
1999 /* big hammer: mark all scalars precise in this path.
2000 * pop_stack may still get !precise scalars.
2002 for (; st; st = st->parent)
2003 for (i = 0; i <= st->curframe; i++) {
2004 func = st->frame[i];
2005 for (j = 0; j < BPF_REG_FP; j++) {
2006 reg = &func->regs[j];
2007 if (reg->type != SCALAR_VALUE)
2009 reg->precise = true;
2011 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
2012 if (func->stack[j].slot_type[0] != STACK_SPILL)
2014 reg = &func->stack[j].spilled_ptr;
2015 if (reg->type != SCALAR_VALUE)
2017 reg->precise = true;
2022 static int __mark_chain_precision(struct bpf_verifier_env *env, int regno,
2025 struct bpf_verifier_state *st = env->cur_state;
2026 int first_idx = st->first_insn_idx;
2027 int last_idx = env->insn_idx;
2028 struct bpf_func_state *func;
2029 struct bpf_reg_state *reg;
2030 u32 reg_mask = regno >= 0 ? 1u << regno : 0;
2031 u64 stack_mask = spi >= 0 ? 1ull << spi : 0;
2032 bool skip_first = true;
2033 bool new_marks = false;
2036 if (!env->bpf_capable)
2039 func = st->frame[st->curframe];
2041 reg = &func->regs[regno];
2042 if (reg->type != SCALAR_VALUE) {
2043 WARN_ONCE(1, "backtracing misuse");
2050 reg->precise = true;
2054 if (func->stack[spi].slot_type[0] != STACK_SPILL) {
2058 reg = &func->stack[spi].spilled_ptr;
2059 if (reg->type != SCALAR_VALUE) {
2067 reg->precise = true;
2073 if (!reg_mask && !stack_mask)
2076 DECLARE_BITMAP(mask, 64);
2077 u32 history = st->jmp_history_cnt;
2079 if (env->log.level & BPF_LOG_LEVEL)
2080 verbose(env, "last_idx %d first_idx %d\n", last_idx, first_idx);
2081 for (i = last_idx;;) {
2086 err = backtrack_insn(env, i, ®_mask, &stack_mask);
2088 if (err == -ENOTSUPP) {
2089 mark_all_scalars_precise(env, st);
2094 if (!reg_mask && !stack_mask)
2095 /* Found assignment(s) into tracked register in this state.
2096 * Since this state is already marked, just return.
2097 * Nothing to be tracked further in the parent state.
2102 i = get_prev_insn_idx(st, i, &history);
2103 if (i >= env->prog->len) {
2104 /* This can happen if backtracking reached insn 0
2105 * and there are still reg_mask or stack_mask
2107 * It means the backtracking missed the spot where
2108 * particular register was initialized with a constant.
2110 verbose(env, "BUG backtracking idx %d\n", i);
2111 WARN_ONCE(1, "verifier backtracking bug");
2120 func = st->frame[st->curframe];
2121 bitmap_from_u64(mask, reg_mask);
2122 for_each_set_bit(i, mask, 32) {
2123 reg = &func->regs[i];
2124 if (reg->type != SCALAR_VALUE) {
2125 reg_mask &= ~(1u << i);
2130 reg->precise = true;
2133 bitmap_from_u64(mask, stack_mask);
2134 for_each_set_bit(i, mask, 64) {
2135 if (i >= func->allocated_stack / BPF_REG_SIZE) {
2136 /* the sequence of instructions:
2138 * 3: (7b) *(u64 *)(r3 -8) = r0
2139 * 4: (79) r4 = *(u64 *)(r10 -8)
2140 * doesn't contain jmps. It's backtracked
2141 * as a single block.
2142 * During backtracking insn 3 is not recognized as
2143 * stack access, so at the end of backtracking
2144 * stack slot fp-8 is still marked in stack_mask.
2145 * However the parent state may not have accessed
2146 * fp-8 and it's "unallocated" stack space.
2147 * In such case fallback to conservative.
2149 mark_all_scalars_precise(env, st);
2153 if (func->stack[i].slot_type[0] != STACK_SPILL) {
2154 stack_mask &= ~(1ull << i);
2157 reg = &func->stack[i].spilled_ptr;
2158 if (reg->type != SCALAR_VALUE) {
2159 stack_mask &= ~(1ull << i);
2164 reg->precise = true;
2166 if (env->log.level & BPF_LOG_LEVEL) {
2167 print_verifier_state(env, func);
2168 verbose(env, "parent %s regs=%x stack=%llx marks\n",
2169 new_marks ? "didn't have" : "already had",
2170 reg_mask, stack_mask);
2173 if (!reg_mask && !stack_mask)
2178 last_idx = st->last_insn_idx;
2179 first_idx = st->first_insn_idx;
2184 static int mark_chain_precision(struct bpf_verifier_env *env, int regno)
2186 return __mark_chain_precision(env, regno, -1);
2189 static int mark_chain_precision_stack(struct bpf_verifier_env *env, int spi)
2191 return __mark_chain_precision(env, -1, spi);
2194 static bool is_spillable_regtype(enum bpf_reg_type type)
2197 case PTR_TO_MAP_VALUE:
2198 case PTR_TO_MAP_VALUE_OR_NULL:
2202 case PTR_TO_PACKET_META:
2203 case PTR_TO_PACKET_END:
2204 case PTR_TO_FLOW_KEYS:
2205 case CONST_PTR_TO_MAP:
2207 case PTR_TO_SOCKET_OR_NULL:
2208 case PTR_TO_SOCK_COMMON:
2209 case PTR_TO_SOCK_COMMON_OR_NULL:
2210 case PTR_TO_TCP_SOCK:
2211 case PTR_TO_TCP_SOCK_OR_NULL:
2212 case PTR_TO_XDP_SOCK:
2214 case PTR_TO_BTF_ID_OR_NULL:
2215 case PTR_TO_RDONLY_BUF:
2216 case PTR_TO_RDONLY_BUF_OR_NULL:
2217 case PTR_TO_RDWR_BUF:
2218 case PTR_TO_RDWR_BUF_OR_NULL:
2219 case PTR_TO_PERCPU_BTF_ID:
2221 case PTR_TO_MEM_OR_NULL:
2228 /* Does this register contain a constant zero? */
2229 static bool register_is_null(struct bpf_reg_state *reg)
2231 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
2234 static bool register_is_const(struct bpf_reg_state *reg)
2236 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
2239 static bool __is_scalar_unbounded(struct bpf_reg_state *reg)
2241 return tnum_is_unknown(reg->var_off) &&
2242 reg->smin_value == S64_MIN && reg->smax_value == S64_MAX &&
2243 reg->umin_value == 0 && reg->umax_value == U64_MAX &&
2244 reg->s32_min_value == S32_MIN && reg->s32_max_value == S32_MAX &&
2245 reg->u32_min_value == 0 && reg->u32_max_value == U32_MAX;
2248 static bool register_is_bounded(struct bpf_reg_state *reg)
2250 return reg->type == SCALAR_VALUE && !__is_scalar_unbounded(reg);
2253 static bool __is_pointer_value(bool allow_ptr_leaks,
2254 const struct bpf_reg_state *reg)
2256 if (allow_ptr_leaks)
2259 return reg->type != SCALAR_VALUE;
2262 static void save_register_state(struct bpf_func_state *state,
2263 int spi, struct bpf_reg_state *reg)
2267 state->stack[spi].spilled_ptr = *reg;
2268 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2270 for (i = 0; i < BPF_REG_SIZE; i++)
2271 state->stack[spi].slot_type[i] = STACK_SPILL;
2274 /* check_stack_read/write functions track spill/fill of registers,
2275 * stack boundary and alignment are checked in check_mem_access()
2277 static int check_stack_write(struct bpf_verifier_env *env,
2278 struct bpf_func_state *state, /* func where register points to */
2279 int off, int size, int value_regno, int insn_idx)
2281 struct bpf_func_state *cur; /* state of the current function */
2282 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
2283 u32 dst_reg = env->prog->insnsi[insn_idx].dst_reg;
2284 struct bpf_reg_state *reg = NULL;
2286 err = realloc_func_state(state, round_up(slot + 1, BPF_REG_SIZE),
2287 state->acquired_refs, true);
2290 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
2291 * so it's aligned access and [off, off + size) are within stack limits
2293 if (!env->allow_ptr_leaks &&
2294 state->stack[spi].slot_type[0] == STACK_SPILL &&
2295 size != BPF_REG_SIZE) {
2296 verbose(env, "attempt to corrupt spilled pointer on stack\n");
2300 cur = env->cur_state->frame[env->cur_state->curframe];
2301 if (value_regno >= 0)
2302 reg = &cur->regs[value_regno];
2304 if (reg && size == BPF_REG_SIZE && register_is_bounded(reg) &&
2305 !register_is_null(reg) && env->bpf_capable) {
2306 if (dst_reg != BPF_REG_FP) {
2307 /* The backtracking logic can only recognize explicit
2308 * stack slot address like [fp - 8]. Other spill of
2309 * scalar via different register has to be conervative.
2310 * Backtrack from here and mark all registers as precise
2311 * that contributed into 'reg' being a constant.
2313 err = mark_chain_precision(env, value_regno);
2317 save_register_state(state, spi, reg);
2318 } else if (reg && is_spillable_regtype(reg->type)) {
2319 /* register containing pointer is being spilled into stack */
2320 if (size != BPF_REG_SIZE) {
2321 verbose_linfo(env, insn_idx, "; ");
2322 verbose(env, "invalid size of register spill\n");
2326 if (state != cur && reg->type == PTR_TO_STACK) {
2327 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
2331 if (!env->bypass_spec_v4) {
2332 bool sanitize = false;
2334 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
2335 register_is_const(&state->stack[spi].spilled_ptr))
2337 for (i = 0; i < BPF_REG_SIZE; i++)
2338 if (state->stack[spi].slot_type[i] == STACK_MISC) {
2343 int *poff = &env->insn_aux_data[insn_idx].sanitize_stack_off;
2344 int soff = (-spi - 1) * BPF_REG_SIZE;
2346 /* detected reuse of integer stack slot with a pointer
2347 * which means either llvm is reusing stack slot or
2348 * an attacker is trying to exploit CVE-2018-3639
2349 * (speculative store bypass)
2350 * Have to sanitize that slot with preemptive
2353 if (*poff && *poff != soff) {
2354 /* disallow programs where single insn stores
2355 * into two different stack slots, since verifier
2356 * cannot sanitize them
2359 "insn %d cannot access two stack slots fp%d and fp%d",
2360 insn_idx, *poff, soff);
2366 save_register_state(state, spi, reg);
2368 u8 type = STACK_MISC;
2370 /* regular write of data into stack destroys any spilled ptr */
2371 state->stack[spi].spilled_ptr.type = NOT_INIT;
2372 /* Mark slots as STACK_MISC if they belonged to spilled ptr. */
2373 if (state->stack[spi].slot_type[0] == STACK_SPILL)
2374 for (i = 0; i < BPF_REG_SIZE; i++)
2375 state->stack[spi].slot_type[i] = STACK_MISC;
2377 /* only mark the slot as written if all 8 bytes were written
2378 * otherwise read propagation may incorrectly stop too soon
2379 * when stack slots are partially written.
2380 * This heuristic means that read propagation will be
2381 * conservative, since it will add reg_live_read marks
2382 * to stack slots all the way to first state when programs
2383 * writes+reads less than 8 bytes
2385 if (size == BPF_REG_SIZE)
2386 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2388 /* when we zero initialize stack slots mark them as such */
2389 if (reg && register_is_null(reg)) {
2390 /* backtracking doesn't work for STACK_ZERO yet. */
2391 err = mark_chain_precision(env, value_regno);
2397 /* Mark slots affected by this stack write. */
2398 for (i = 0; i < size; i++)
2399 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
2405 static int check_stack_read(struct bpf_verifier_env *env,
2406 struct bpf_func_state *reg_state /* func where register points to */,
2407 int off, int size, int value_regno)
2409 struct bpf_verifier_state *vstate = env->cur_state;
2410 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2411 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
2412 struct bpf_reg_state *reg;
2415 if (reg_state->allocated_stack <= slot) {
2416 verbose(env, "invalid read from stack off %d+0 size %d\n",
2420 stype = reg_state->stack[spi].slot_type;
2421 reg = ®_state->stack[spi].spilled_ptr;
2423 if (stype[0] == STACK_SPILL) {
2424 if (size != BPF_REG_SIZE) {
2425 if (reg->type != SCALAR_VALUE) {
2426 verbose_linfo(env, env->insn_idx, "; ");
2427 verbose(env, "invalid size of register fill\n");
2430 if (value_regno >= 0) {
2431 mark_reg_unknown(env, state->regs, value_regno);
2432 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
2434 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2437 for (i = 1; i < BPF_REG_SIZE; i++) {
2438 if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) {
2439 verbose(env, "corrupted spill memory\n");
2444 if (value_regno >= 0) {
2445 /* restore register state from stack */
2446 state->regs[value_regno] = *reg;
2447 /* mark reg as written since spilled pointer state likely
2448 * has its liveness marks cleared by is_state_visited()
2449 * which resets stack/reg liveness for state transitions
2451 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
2452 } else if (__is_pointer_value(env->allow_ptr_leaks, reg)) {
2453 /* If value_regno==-1, the caller is asking us whether
2454 * it is acceptable to use this value as a SCALAR_VALUE
2456 * We must not allow unprivileged callers to do that
2457 * with spilled pointers.
2459 verbose(env, "leaking pointer from stack off %d\n",
2463 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2467 for (i = 0; i < size; i++) {
2468 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_MISC)
2470 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_ZERO) {
2474 verbose(env, "invalid read from stack off %d+%d size %d\n",
2478 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2479 if (value_regno >= 0) {
2480 if (zeros == size) {
2481 /* any size read into register is zero extended,
2482 * so the whole register == const_zero
2484 __mark_reg_const_zero(&state->regs[value_regno]);
2485 /* backtracking doesn't support STACK_ZERO yet,
2486 * so mark it precise here, so that later
2487 * backtracking can stop here.
2488 * Backtracking may not need this if this register
2489 * doesn't participate in pointer adjustment.
2490 * Forward propagation of precise flag is not
2491 * necessary either. This mark is only to stop
2492 * backtracking. Any register that contributed
2493 * to const 0 was marked precise before spill.
2495 state->regs[value_regno].precise = true;
2497 /* have read misc data from the stack */
2498 mark_reg_unknown(env, state->regs, value_regno);
2500 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
2506 static int check_stack_access(struct bpf_verifier_env *env,
2507 const struct bpf_reg_state *reg,
2510 /* Stack accesses must be at a fixed offset, so that we
2511 * can determine what type of data were returned. See
2512 * check_stack_read().
2514 if (!tnum_is_const(reg->var_off)) {
2517 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2518 verbose(env, "variable stack access var_off=%s off=%d size=%d\n",
2523 if (off >= 0 || off < -MAX_BPF_STACK) {
2524 verbose(env, "invalid stack off=%d size=%d\n", off, size);
2531 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
2532 int off, int size, enum bpf_access_type type)
2534 struct bpf_reg_state *regs = cur_regs(env);
2535 struct bpf_map *map = regs[regno].map_ptr;
2536 u32 cap = bpf_map_flags_to_cap(map);
2538 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
2539 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
2540 map->value_size, off, size);
2544 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
2545 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
2546 map->value_size, off, size);
2553 /* check read/write into memory region (e.g., map value, ringbuf sample, etc) */
2554 static int __check_mem_access(struct bpf_verifier_env *env, int regno,
2555 int off, int size, u32 mem_size,
2556 bool zero_size_allowed)
2558 bool size_ok = size > 0 || (size == 0 && zero_size_allowed);
2559 struct bpf_reg_state *reg;
2561 if (off >= 0 && size_ok && (u64)off + size <= mem_size)
2564 reg = &cur_regs(env)[regno];
2565 switch (reg->type) {
2566 case PTR_TO_MAP_VALUE:
2567 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
2568 mem_size, off, size);
2571 case PTR_TO_PACKET_META:
2572 case PTR_TO_PACKET_END:
2573 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
2574 off, size, regno, reg->id, off, mem_size);
2578 verbose(env, "invalid access to memory, mem_size=%u off=%d size=%d\n",
2579 mem_size, off, size);
2585 /* check read/write into a memory region with possible variable offset */
2586 static int check_mem_region_access(struct bpf_verifier_env *env, u32 regno,
2587 int off, int size, u32 mem_size,
2588 bool zero_size_allowed)
2590 struct bpf_verifier_state *vstate = env->cur_state;
2591 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2592 struct bpf_reg_state *reg = &state->regs[regno];
2595 /* We may have adjusted the register pointing to memory region, so we
2596 * need to try adding each of min_value and max_value to off
2597 * to make sure our theoretical access will be safe.
2599 if (env->log.level & BPF_LOG_LEVEL)
2600 print_verifier_state(env, state);
2602 /* The minimum value is only important with signed
2603 * comparisons where we can't assume the floor of a
2604 * value is 0. If we are using signed variables for our
2605 * index'es we need to make sure that whatever we use
2606 * will have a set floor within our range.
2608 if (reg->smin_value < 0 &&
2609 (reg->smin_value == S64_MIN ||
2610 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
2611 reg->smin_value + off < 0)) {
2612 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2616 err = __check_mem_access(env, regno, reg->smin_value + off, size,
2617 mem_size, zero_size_allowed);
2619 verbose(env, "R%d min value is outside of the allowed memory range\n",
2624 /* If we haven't set a max value then we need to bail since we can't be
2625 * sure we won't do bad things.
2626 * If reg->umax_value + off could overflow, treat that as unbounded too.
2628 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
2629 verbose(env, "R%d unbounded memory access, make sure to bounds check any such access\n",
2633 err = __check_mem_access(env, regno, reg->umax_value + off, size,
2634 mem_size, zero_size_allowed);
2636 verbose(env, "R%d max value is outside of the allowed memory range\n",
2644 /* check read/write into a map element with possible variable offset */
2645 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
2646 int off, int size, bool zero_size_allowed)
2648 struct bpf_verifier_state *vstate = env->cur_state;
2649 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2650 struct bpf_reg_state *reg = &state->regs[regno];
2651 struct bpf_map *map = reg->map_ptr;
2654 err = check_mem_region_access(env, regno, off, size, map->value_size,
2659 if (map_value_has_spin_lock(map)) {
2660 u32 lock = map->spin_lock_off;
2662 /* if any part of struct bpf_spin_lock can be touched by
2663 * load/store reject this program.
2664 * To check that [x1, x2) overlaps with [y1, y2)
2665 * it is sufficient to check x1 < y2 && y1 < x2.
2667 if (reg->smin_value + off < lock + sizeof(struct bpf_spin_lock) &&
2668 lock < reg->umax_value + off + size) {
2669 verbose(env, "bpf_spin_lock cannot be accessed directly by load/store\n");
2676 #define MAX_PACKET_OFF 0xffff
2678 static enum bpf_prog_type resolve_prog_type(struct bpf_prog *prog)
2680 return prog->aux->dst_prog ? prog->aux->dst_prog->type : prog->type;
2683 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
2684 const struct bpf_call_arg_meta *meta,
2685 enum bpf_access_type t)
2687 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
2689 switch (prog_type) {
2690 /* Program types only with direct read access go here! */
2691 case BPF_PROG_TYPE_LWT_IN:
2692 case BPF_PROG_TYPE_LWT_OUT:
2693 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
2694 case BPF_PROG_TYPE_SK_REUSEPORT:
2695 case BPF_PROG_TYPE_FLOW_DISSECTOR:
2696 case BPF_PROG_TYPE_CGROUP_SKB:
2701 /* Program types with direct read + write access go here! */
2702 case BPF_PROG_TYPE_SCHED_CLS:
2703 case BPF_PROG_TYPE_SCHED_ACT:
2704 case BPF_PROG_TYPE_XDP:
2705 case BPF_PROG_TYPE_LWT_XMIT:
2706 case BPF_PROG_TYPE_SK_SKB:
2707 case BPF_PROG_TYPE_SK_MSG:
2709 return meta->pkt_access;
2711 env->seen_direct_write = true;
2714 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
2716 env->seen_direct_write = true;
2725 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
2726 int size, bool zero_size_allowed)
2728 struct bpf_reg_state *regs = cur_regs(env);
2729 struct bpf_reg_state *reg = ®s[regno];
2732 /* We may have added a variable offset to the packet pointer; but any
2733 * reg->range we have comes after that. We are only checking the fixed
2737 /* We don't allow negative numbers, because we aren't tracking enough
2738 * detail to prove they're safe.
2740 if (reg->smin_value < 0) {
2741 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2746 err = reg->range < 0 ? -EINVAL :
2747 __check_mem_access(env, regno, off, size, reg->range,
2750 verbose(env, "R%d offset is outside of the packet\n", regno);
2754 /* __check_mem_access has made sure "off + size - 1" is within u16.
2755 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
2756 * otherwise find_good_pkt_pointers would have refused to set range info
2757 * that __check_mem_access would have rejected this pkt access.
2758 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
2760 env->prog->aux->max_pkt_offset =
2761 max_t(u32, env->prog->aux->max_pkt_offset,
2762 off + reg->umax_value + size - 1);
2767 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
2768 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
2769 enum bpf_access_type t, enum bpf_reg_type *reg_type,
2770 struct btf **btf, u32 *btf_id)
2772 struct bpf_insn_access_aux info = {
2773 .reg_type = *reg_type,
2777 if (env->ops->is_valid_access &&
2778 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
2779 /* A non zero info.ctx_field_size indicates that this field is a
2780 * candidate for later verifier transformation to load the whole
2781 * field and then apply a mask when accessed with a narrower
2782 * access than actual ctx access size. A zero info.ctx_field_size
2783 * will only allow for whole field access and rejects any other
2784 * type of narrower access.
2786 *reg_type = info.reg_type;
2788 if (*reg_type == PTR_TO_BTF_ID || *reg_type == PTR_TO_BTF_ID_OR_NULL) {
2790 *btf_id = info.btf_id;
2792 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
2794 /* remember the offset of last byte accessed in ctx */
2795 if (env->prog->aux->max_ctx_offset < off + size)
2796 env->prog->aux->max_ctx_offset = off + size;
2800 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
2804 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
2807 if (size < 0 || off < 0 ||
2808 (u64)off + size > sizeof(struct bpf_flow_keys)) {
2809 verbose(env, "invalid access to flow keys off=%d size=%d\n",
2816 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
2817 u32 regno, int off, int size,
2818 enum bpf_access_type t)
2820 struct bpf_reg_state *regs = cur_regs(env);
2821 struct bpf_reg_state *reg = ®s[regno];
2822 struct bpf_insn_access_aux info = {};
2825 if (reg->smin_value < 0) {
2826 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2831 switch (reg->type) {
2832 case PTR_TO_SOCK_COMMON:
2833 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
2836 valid = bpf_sock_is_valid_access(off, size, t, &info);
2838 case PTR_TO_TCP_SOCK:
2839 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
2841 case PTR_TO_XDP_SOCK:
2842 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
2850 env->insn_aux_data[insn_idx].ctx_field_size =
2851 info.ctx_field_size;
2855 verbose(env, "R%d invalid %s access off=%d size=%d\n",
2856 regno, reg_type_str[reg->type], off, size);
2861 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
2863 return cur_regs(env) + regno;
2866 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
2868 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
2871 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
2873 const struct bpf_reg_state *reg = reg_state(env, regno);
2875 return reg->type == PTR_TO_CTX;
2878 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
2880 const struct bpf_reg_state *reg = reg_state(env, regno);
2882 return type_is_sk_pointer(reg->type);
2885 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
2887 const struct bpf_reg_state *reg = reg_state(env, regno);
2889 return type_is_pkt_pointer(reg->type);
2892 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
2894 const struct bpf_reg_state *reg = reg_state(env, regno);
2896 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
2897 return reg->type == PTR_TO_FLOW_KEYS;
2900 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
2901 const struct bpf_reg_state *reg,
2902 int off, int size, bool strict)
2904 struct tnum reg_off;
2907 /* Byte size accesses are always allowed. */
2908 if (!strict || size == 1)
2911 /* For platforms that do not have a Kconfig enabling
2912 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
2913 * NET_IP_ALIGN is universally set to '2'. And on platforms
2914 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
2915 * to this code only in strict mode where we want to emulate
2916 * the NET_IP_ALIGN==2 checking. Therefore use an
2917 * unconditional IP align value of '2'.
2921 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
2922 if (!tnum_is_aligned(reg_off, size)) {
2925 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2927 "misaligned packet access off %d+%s+%d+%d size %d\n",
2928 ip_align, tn_buf, reg->off, off, size);
2935 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
2936 const struct bpf_reg_state *reg,
2937 const char *pointer_desc,
2938 int off, int size, bool strict)
2940 struct tnum reg_off;
2942 /* Byte size accesses are always allowed. */
2943 if (!strict || size == 1)
2946 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
2947 if (!tnum_is_aligned(reg_off, size)) {
2950 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2951 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
2952 pointer_desc, tn_buf, reg->off, off, size);
2959 static int check_ptr_alignment(struct bpf_verifier_env *env,
2960 const struct bpf_reg_state *reg, int off,
2961 int size, bool strict_alignment_once)
2963 bool strict = env->strict_alignment || strict_alignment_once;
2964 const char *pointer_desc = "";
2966 switch (reg->type) {
2968 case PTR_TO_PACKET_META:
2969 /* Special case, because of NET_IP_ALIGN. Given metadata sits
2970 * right in front, treat it the very same way.
2972 return check_pkt_ptr_alignment(env, reg, off, size, strict);
2973 case PTR_TO_FLOW_KEYS:
2974 pointer_desc = "flow keys ";
2976 case PTR_TO_MAP_VALUE:
2977 pointer_desc = "value ";
2980 pointer_desc = "context ";
2983 pointer_desc = "stack ";
2984 /* The stack spill tracking logic in check_stack_write()
2985 * and check_stack_read() relies on stack accesses being
2991 pointer_desc = "sock ";
2993 case PTR_TO_SOCK_COMMON:
2994 pointer_desc = "sock_common ";
2996 case PTR_TO_TCP_SOCK:
2997 pointer_desc = "tcp_sock ";
2999 case PTR_TO_XDP_SOCK:
3000 pointer_desc = "xdp_sock ";
3005 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
3009 static int update_stack_depth(struct bpf_verifier_env *env,
3010 const struct bpf_func_state *func,
3013 u16 stack = env->subprog_info[func->subprogno].stack_depth;
3018 /* update known max for given subprogram */
3019 env->subprog_info[func->subprogno].stack_depth = -off;
3023 /* starting from main bpf function walk all instructions of the function
3024 * and recursively walk all callees that given function can call.
3025 * Ignore jump and exit insns.
3026 * Since recursion is prevented by check_cfg() this algorithm
3027 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
3029 static int check_max_stack_depth(struct bpf_verifier_env *env)
3031 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
3032 struct bpf_subprog_info *subprog = env->subprog_info;
3033 struct bpf_insn *insn = env->prog->insnsi;
3034 bool tail_call_reachable = false;
3035 int ret_insn[MAX_CALL_FRAMES];
3036 int ret_prog[MAX_CALL_FRAMES];
3040 /* protect against potential stack overflow that might happen when
3041 * bpf2bpf calls get combined with tailcalls. Limit the caller's stack
3042 * depth for such case down to 256 so that the worst case scenario
3043 * would result in 8k stack size (32 which is tailcall limit * 256 =
3046 * To get the idea what might happen, see an example:
3047 * func1 -> sub rsp, 128
3048 * subfunc1 -> sub rsp, 256
3049 * tailcall1 -> add rsp, 256
3050 * func2 -> sub rsp, 192 (total stack size = 128 + 192 = 320)
3051 * subfunc2 -> sub rsp, 64
3052 * subfunc22 -> sub rsp, 128
3053 * tailcall2 -> add rsp, 128
3054 * func3 -> sub rsp, 32 (total stack size 128 + 192 + 64 + 32 = 416)
3056 * tailcall will unwind the current stack frame but it will not get rid
3057 * of caller's stack as shown on the example above.
3059 if (idx && subprog[idx].has_tail_call && depth >= 256) {
3061 "tail_calls are not allowed when call stack of previous frames is %d bytes. Too large\n",
3065 /* round up to 32-bytes, since this is granularity
3066 * of interpreter stack size
3068 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
3069 if (depth > MAX_BPF_STACK) {
3070 verbose(env, "combined stack size of %d calls is %d. Too large\n",
3075 subprog_end = subprog[idx + 1].start;
3076 for (; i < subprog_end; i++) {
3077 if (insn[i].code != (BPF_JMP | BPF_CALL))
3079 if (insn[i].src_reg != BPF_PSEUDO_CALL)
3081 /* remember insn and function to return to */
3082 ret_insn[frame] = i + 1;
3083 ret_prog[frame] = idx;
3085 /* find the callee */
3086 i = i + insn[i].imm + 1;
3087 idx = find_subprog(env, i);
3089 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3094 if (subprog[idx].has_tail_call)
3095 tail_call_reachable = true;
3098 if (frame >= MAX_CALL_FRAMES) {
3099 verbose(env, "the call stack of %d frames is too deep !\n",
3105 /* if tail call got detected across bpf2bpf calls then mark each of the
3106 * currently present subprog frames as tail call reachable subprogs;
3107 * this info will be utilized by JIT so that we will be preserving the
3108 * tail call counter throughout bpf2bpf calls combined with tailcalls
3110 if (tail_call_reachable)
3111 for (j = 0; j < frame; j++)
3112 subprog[ret_prog[j]].tail_call_reachable = true;
3114 /* end of for() loop means the last insn of the 'subprog'
3115 * was reached. Doesn't matter whether it was JA or EXIT
3119 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
3121 i = ret_insn[frame];
3122 idx = ret_prog[frame];
3126 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
3127 static int get_callee_stack_depth(struct bpf_verifier_env *env,
3128 const struct bpf_insn *insn, int idx)
3130 int start = idx + insn->imm + 1, subprog;
3132 subprog = find_subprog(env, start);
3134 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3138 return env->subprog_info[subprog].stack_depth;
3142 int check_ctx_reg(struct bpf_verifier_env *env,
3143 const struct bpf_reg_state *reg, int regno)
3145 /* Access to ctx or passing it to a helper is only allowed in
3146 * its original, unmodified form.
3150 verbose(env, "dereference of modified ctx ptr R%d off=%d disallowed\n",
3155 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3158 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3159 verbose(env, "variable ctx access var_off=%s disallowed\n", tn_buf);
3166 static int __check_buffer_access(struct bpf_verifier_env *env,
3167 const char *buf_info,
3168 const struct bpf_reg_state *reg,
3169 int regno, int off, int size)
3173 "R%d invalid %s buffer access: off=%d, size=%d\n",
3174 regno, buf_info, off, size);
3177 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3180 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3182 "R%d invalid variable buffer offset: off=%d, var_off=%s\n",
3183 regno, off, tn_buf);
3190 static int check_tp_buffer_access(struct bpf_verifier_env *env,
3191 const struct bpf_reg_state *reg,
3192 int regno, int off, int size)
3196 err = __check_buffer_access(env, "tracepoint", reg, regno, off, size);
3200 if (off + size > env->prog->aux->max_tp_access)
3201 env->prog->aux->max_tp_access = off + size;
3206 static int check_buffer_access(struct bpf_verifier_env *env,
3207 const struct bpf_reg_state *reg,
3208 int regno, int off, int size,
3209 bool zero_size_allowed,
3210 const char *buf_info,
3215 err = __check_buffer_access(env, buf_info, reg, regno, off, size);
3219 if (off + size > *max_access)
3220 *max_access = off + size;
3225 /* BPF architecture zero extends alu32 ops into 64-bit registesr */
3226 static void zext_32_to_64(struct bpf_reg_state *reg)
3228 reg->var_off = tnum_subreg(reg->var_off);
3229 __reg_assign_32_into_64(reg);
3232 /* truncate register to smaller size (in bytes)
3233 * must be called with size < BPF_REG_SIZE
3235 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
3239 /* clear high bits in bit representation */
3240 reg->var_off = tnum_cast(reg->var_off, size);
3242 /* fix arithmetic bounds */
3243 mask = ((u64)1 << (size * 8)) - 1;
3244 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
3245 reg->umin_value &= mask;
3246 reg->umax_value &= mask;
3248 reg->umin_value = 0;
3249 reg->umax_value = mask;
3251 reg->smin_value = reg->umin_value;
3252 reg->smax_value = reg->umax_value;
3254 /* If size is smaller than 32bit register the 32bit register
3255 * values are also truncated so we push 64-bit bounds into
3256 * 32-bit bounds. Above were truncated < 32-bits already.
3260 __reg_combine_64_into_32(reg);
3263 static bool bpf_map_is_rdonly(const struct bpf_map *map)
3265 return (map->map_flags & BPF_F_RDONLY_PROG) && map->frozen;
3268 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val)
3274 err = map->ops->map_direct_value_addr(map, &addr, off);
3277 ptr = (void *)(long)addr + off;
3281 *val = (u64)*(u8 *)ptr;
3284 *val = (u64)*(u16 *)ptr;
3287 *val = (u64)*(u32 *)ptr;
3298 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
3299 struct bpf_reg_state *regs,
3300 int regno, int off, int size,
3301 enum bpf_access_type atype,
3304 struct bpf_reg_state *reg = regs + regno;
3305 const struct btf_type *t = btf_type_by_id(reg->btf, reg->btf_id);
3306 const char *tname = btf_name_by_offset(reg->btf, t->name_off);
3312 "R%d is ptr_%s invalid negative access: off=%d\n",
3316 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3319 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3321 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
3322 regno, tname, off, tn_buf);
3326 if (env->ops->btf_struct_access) {
3327 ret = env->ops->btf_struct_access(&env->log, reg->btf, t,
3328 off, size, atype, &btf_id);
3330 if (atype != BPF_READ) {
3331 verbose(env, "only read is supported\n");
3335 ret = btf_struct_access(&env->log, reg->btf, t, off, size,
3342 if (atype == BPF_READ && value_regno >= 0)
3343 mark_btf_ld_reg(env, regs, value_regno, ret, reg->btf, btf_id);
3348 static int check_ptr_to_map_access(struct bpf_verifier_env *env,
3349 struct bpf_reg_state *regs,
3350 int regno, int off, int size,
3351 enum bpf_access_type atype,
3354 struct bpf_reg_state *reg = regs + regno;
3355 struct bpf_map *map = reg->map_ptr;
3356 const struct btf_type *t;
3362 verbose(env, "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n");
3366 if (!map->ops->map_btf_id || !*map->ops->map_btf_id) {
3367 verbose(env, "map_ptr access not supported for map type %d\n",
3372 t = btf_type_by_id(btf_vmlinux, *map->ops->map_btf_id);
3373 tname = btf_name_by_offset(btf_vmlinux, t->name_off);
3375 if (!env->allow_ptr_to_map_access) {
3377 "%s access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
3383 verbose(env, "R%d is %s invalid negative access: off=%d\n",
3388 if (atype != BPF_READ) {
3389 verbose(env, "only read from %s is supported\n", tname);
3393 ret = btf_struct_access(&env->log, btf_vmlinux, t, off, size, atype, &btf_id);
3397 if (value_regno >= 0)
3398 mark_btf_ld_reg(env, regs, value_regno, ret, btf_vmlinux, btf_id);
3404 /* check whether memory at (regno + off) is accessible for t = (read | write)
3405 * if t==write, value_regno is a register which value is stored into memory
3406 * if t==read, value_regno is a register which will receive the value from memory
3407 * if t==write && value_regno==-1, some unknown value is stored into memory
3408 * if t==read && value_regno==-1, don't care what we read from memory
3410 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
3411 int off, int bpf_size, enum bpf_access_type t,
3412 int value_regno, bool strict_alignment_once)
3414 struct bpf_reg_state *regs = cur_regs(env);
3415 struct bpf_reg_state *reg = regs + regno;
3416 struct bpf_func_state *state;
3419 size = bpf_size_to_bytes(bpf_size);
3423 /* alignment checks will add in reg->off themselves */
3424 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
3428 /* for access checks, reg->off is just part of off */
3431 if (reg->type == PTR_TO_MAP_VALUE) {
3432 if (t == BPF_WRITE && value_regno >= 0 &&
3433 is_pointer_value(env, value_regno)) {
3434 verbose(env, "R%d leaks addr into map\n", value_regno);
3437 err = check_map_access_type(env, regno, off, size, t);
3440 err = check_map_access(env, regno, off, size, false);
3441 if (!err && t == BPF_READ && value_regno >= 0) {
3442 struct bpf_map *map = reg->map_ptr;
3444 /* if map is read-only, track its contents as scalars */
3445 if (tnum_is_const(reg->var_off) &&
3446 bpf_map_is_rdonly(map) &&
3447 map->ops->map_direct_value_addr) {
3448 int map_off = off + reg->var_off.value;
3451 err = bpf_map_direct_read(map, map_off, size,
3456 regs[value_regno].type = SCALAR_VALUE;
3457 __mark_reg_known(®s[value_regno], val);
3459 mark_reg_unknown(env, regs, value_regno);
3462 } else if (reg->type == PTR_TO_MEM) {
3463 if (t == BPF_WRITE && value_regno >= 0 &&
3464 is_pointer_value(env, value_regno)) {
3465 verbose(env, "R%d leaks addr into mem\n", value_regno);
3468 err = check_mem_region_access(env, regno, off, size,
3469 reg->mem_size, false);
3470 if (!err && t == BPF_READ && value_regno >= 0)
3471 mark_reg_unknown(env, regs, value_regno);
3472 } else if (reg->type == PTR_TO_CTX) {
3473 enum bpf_reg_type reg_type = SCALAR_VALUE;
3474 struct btf *btf = NULL;
3477 if (t == BPF_WRITE && value_regno >= 0 &&
3478 is_pointer_value(env, value_regno)) {
3479 verbose(env, "R%d leaks addr into ctx\n", value_regno);
3483 err = check_ctx_reg(env, reg, regno);
3487 err = check_ctx_access(env, insn_idx, off, size, t, ®_type, &btf, &btf_id);
3489 verbose_linfo(env, insn_idx, "; ");
3490 if (!err && t == BPF_READ && value_regno >= 0) {
3491 /* ctx access returns either a scalar, or a
3492 * PTR_TO_PACKET[_META,_END]. In the latter
3493 * case, we know the offset is zero.
3495 if (reg_type == SCALAR_VALUE) {
3496 mark_reg_unknown(env, regs, value_regno);
3498 mark_reg_known_zero(env, regs,
3500 if (reg_type_may_be_null(reg_type))
3501 regs[value_regno].id = ++env->id_gen;
3502 /* A load of ctx field could have different
3503 * actual load size with the one encoded in the
3504 * insn. When the dst is PTR, it is for sure not
3507 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
3508 if (reg_type == PTR_TO_BTF_ID ||
3509 reg_type == PTR_TO_BTF_ID_OR_NULL) {
3510 regs[value_regno].btf = btf;
3511 regs[value_regno].btf_id = btf_id;
3514 regs[value_regno].type = reg_type;
3517 } else if (reg->type == PTR_TO_STACK) {
3518 off += reg->var_off.value;
3519 err = check_stack_access(env, reg, off, size);
3523 state = func(env, reg);
3524 err = update_stack_depth(env, state, off);
3529 err = check_stack_write(env, state, off, size,
3530 value_regno, insn_idx);
3532 err = check_stack_read(env, state, off, size,
3534 } else if (reg_is_pkt_pointer(reg)) {
3535 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
3536 verbose(env, "cannot write into packet\n");
3539 if (t == BPF_WRITE && value_regno >= 0 &&
3540 is_pointer_value(env, value_regno)) {
3541 verbose(env, "R%d leaks addr into packet\n",
3545 err = check_packet_access(env, regno, off, size, false);
3546 if (!err && t == BPF_READ && value_regno >= 0)
3547 mark_reg_unknown(env, regs, value_regno);
3548 } else if (reg->type == PTR_TO_FLOW_KEYS) {
3549 if (t == BPF_WRITE && value_regno >= 0 &&
3550 is_pointer_value(env, value_regno)) {
3551 verbose(env, "R%d leaks addr into flow keys\n",
3556 err = check_flow_keys_access(env, off, size);
3557 if (!err && t == BPF_READ && value_regno >= 0)
3558 mark_reg_unknown(env, regs, value_regno);
3559 } else if (type_is_sk_pointer(reg->type)) {
3560 if (t == BPF_WRITE) {
3561 verbose(env, "R%d cannot write into %s\n",
3562 regno, reg_type_str[reg->type]);
3565 err = check_sock_access(env, insn_idx, regno, off, size, t);
3566 if (!err && value_regno >= 0)
3567 mark_reg_unknown(env, regs, value_regno);
3568 } else if (reg->type == PTR_TO_TP_BUFFER) {
3569 err = check_tp_buffer_access(env, reg, regno, off, size);
3570 if (!err && t == BPF_READ && value_regno >= 0)
3571 mark_reg_unknown(env, regs, value_regno);
3572 } else if (reg->type == PTR_TO_BTF_ID) {
3573 err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
3575 } else if (reg->type == CONST_PTR_TO_MAP) {
3576 err = check_ptr_to_map_access(env, regs, regno, off, size, t,
3578 } else if (reg->type == PTR_TO_RDONLY_BUF) {
3579 if (t == BPF_WRITE) {
3580 verbose(env, "R%d cannot write into %s\n",
3581 regno, reg_type_str[reg->type]);
3584 err = check_buffer_access(env, reg, regno, off, size, false,
3586 &env->prog->aux->max_rdonly_access);
3587 if (!err && value_regno >= 0)
3588 mark_reg_unknown(env, regs, value_regno);
3589 } else if (reg->type == PTR_TO_RDWR_BUF) {
3590 err = check_buffer_access(env, reg, regno, off, size, false,
3592 &env->prog->aux->max_rdwr_access);
3593 if (!err && t == BPF_READ && value_regno >= 0)
3594 mark_reg_unknown(env, regs, value_regno);
3596 verbose(env, "R%d invalid mem access '%s'\n", regno,
3597 reg_type_str[reg->type]);
3601 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
3602 regs[value_regno].type == SCALAR_VALUE) {
3603 /* b/h/w load zero-extends, mark upper bits as known 0 */
3604 coerce_reg_to_size(®s[value_regno], size);
3609 static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
3613 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
3615 verbose(env, "BPF_XADD uses reserved fields\n");
3619 /* check src1 operand */
3620 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3624 /* check src2 operand */
3625 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3629 if (is_pointer_value(env, insn->src_reg)) {
3630 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
3634 if (is_ctx_reg(env, insn->dst_reg) ||
3635 is_pkt_reg(env, insn->dst_reg) ||
3636 is_flow_key_reg(env, insn->dst_reg) ||
3637 is_sk_reg(env, insn->dst_reg)) {
3638 verbose(env, "BPF_XADD stores into R%d %s is not allowed\n",
3640 reg_type_str[reg_state(env, insn->dst_reg)->type]);
3644 /* check whether atomic_add can read the memory */
3645 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3646 BPF_SIZE(insn->code), BPF_READ, -1, true);
3650 /* check whether atomic_add can write into the same memory */
3651 return check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3652 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
3655 static int __check_stack_boundary(struct bpf_verifier_env *env, u32 regno,
3656 int off, int access_size,
3657 bool zero_size_allowed)
3659 struct bpf_reg_state *reg = reg_state(env, regno);
3661 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
3662 access_size < 0 || (access_size == 0 && !zero_size_allowed)) {
3663 if (tnum_is_const(reg->var_off)) {
3664 verbose(env, "invalid stack type R%d off=%d access_size=%d\n",
3665 regno, off, access_size);
3669 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3670 verbose(env, "invalid stack type R%d var_off=%s access_size=%d\n",
3671 regno, tn_buf, access_size);
3678 /* when register 'regno' is passed into function that will read 'access_size'
3679 * bytes from that pointer, make sure that it's within stack boundary
3680 * and all elements of stack are initialized.
3681 * Unlike most pointer bounds-checking functions, this one doesn't take an
3682 * 'off' argument, so it has to add in reg->off itself.
3684 static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
3685 int access_size, bool zero_size_allowed,
3686 struct bpf_call_arg_meta *meta)
3688 struct bpf_reg_state *reg = reg_state(env, regno);
3689 struct bpf_func_state *state = func(env, reg);
3690 int err, min_off, max_off, i, j, slot, spi;
3692 if (tnum_is_const(reg->var_off)) {
3693 min_off = max_off = reg->var_off.value + reg->off;
3694 err = __check_stack_boundary(env, regno, min_off, access_size,
3699 /* Variable offset is prohibited for unprivileged mode for
3700 * simplicity since it requires corresponding support in
3701 * Spectre masking for stack ALU.
3702 * See also retrieve_ptr_limit().
3704 if (!env->bypass_spec_v1) {
3707 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3708 verbose(env, "R%d indirect variable offset stack access prohibited for !root, var_off=%s\n",
3712 /* Only initialized buffer on stack is allowed to be accessed
3713 * with variable offset. With uninitialized buffer it's hard to
3714 * guarantee that whole memory is marked as initialized on
3715 * helper return since specific bounds are unknown what may
3716 * cause uninitialized stack leaking.
3718 if (meta && meta->raw_mode)
3721 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
3722 reg->smax_value <= -BPF_MAX_VAR_OFF) {
3723 verbose(env, "R%d unbounded indirect variable offset stack access\n",
3727 min_off = reg->smin_value + reg->off;
3728 max_off = reg->smax_value + reg->off;
3729 err = __check_stack_boundary(env, regno, min_off, access_size,
3732 verbose(env, "R%d min value is outside of stack bound\n",
3736 err = __check_stack_boundary(env, regno, max_off, access_size,
3739 verbose(env, "R%d max value is outside of stack bound\n",
3745 if (meta && meta->raw_mode) {
3746 meta->access_size = access_size;
3747 meta->regno = regno;
3751 for (i = min_off; i < max_off + access_size; i++) {
3755 spi = slot / BPF_REG_SIZE;
3756 if (state->allocated_stack <= slot)
3758 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
3759 if (*stype == STACK_MISC)
3761 if (*stype == STACK_ZERO) {
3762 /* helper can write anything into the stack */
3763 *stype = STACK_MISC;
3767 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
3768 state->stack[spi].spilled_ptr.type == PTR_TO_BTF_ID)
3771 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
3772 (state->stack[spi].spilled_ptr.type == SCALAR_VALUE ||
3773 env->allow_ptr_leaks)) {
3774 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
3775 for (j = 0; j < BPF_REG_SIZE; j++)
3776 state->stack[spi].slot_type[j] = STACK_MISC;
3781 if (tnum_is_const(reg->var_off)) {
3782 verbose(env, "invalid indirect read from stack off %d+%d size %d\n",
3783 min_off, i - min_off, access_size);
3787 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3788 verbose(env, "invalid indirect read from stack var_off %s+%d size %d\n",
3789 tn_buf, i - min_off, access_size);
3793 /* reading any byte out of 8-byte 'spill_slot' will cause
3794 * the whole slot to be marked as 'read'
3796 mark_reg_read(env, &state->stack[spi].spilled_ptr,
3797 state->stack[spi].spilled_ptr.parent,
3800 return update_stack_depth(env, state, min_off);
3803 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
3804 int access_size, bool zero_size_allowed,
3805 struct bpf_call_arg_meta *meta)
3807 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
3809 switch (reg->type) {
3811 case PTR_TO_PACKET_META:
3812 return check_packet_access(env, regno, reg->off, access_size,
3814 case PTR_TO_MAP_VALUE:
3815 if (check_map_access_type(env, regno, reg->off, access_size,
3816 meta && meta->raw_mode ? BPF_WRITE :
3819 return check_map_access(env, regno, reg->off, access_size,
3822 return check_mem_region_access(env, regno, reg->off,
3823 access_size, reg->mem_size,
3825 case PTR_TO_RDONLY_BUF:
3826 if (meta && meta->raw_mode)
3828 return check_buffer_access(env, reg, regno, reg->off,
3829 access_size, zero_size_allowed,
3831 &env->prog->aux->max_rdonly_access);
3832 case PTR_TO_RDWR_BUF:
3833 return check_buffer_access(env, reg, regno, reg->off,
3834 access_size, zero_size_allowed,
3836 &env->prog->aux->max_rdwr_access);
3838 return check_stack_boundary(env, regno, access_size,
3839 zero_size_allowed, meta);
3840 default: /* scalar_value or invalid ptr */
3841 /* Allow zero-byte read from NULL, regardless of pointer type */
3842 if (zero_size_allowed && access_size == 0 &&
3843 register_is_null(reg))
3846 verbose(env, "R%d type=%s expected=%s\n", regno,
3847 reg_type_str[reg->type],
3848 reg_type_str[PTR_TO_STACK]);
3853 /* Implementation details:
3854 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL
3855 * Two bpf_map_lookups (even with the same key) will have different reg->id.
3856 * For traditional PTR_TO_MAP_VALUE the verifier clears reg->id after
3857 * value_or_null->value transition, since the verifier only cares about
3858 * the range of access to valid map value pointer and doesn't care about actual
3859 * address of the map element.
3860 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
3861 * reg->id > 0 after value_or_null->value transition. By doing so
3862 * two bpf_map_lookups will be considered two different pointers that
3863 * point to different bpf_spin_locks.
3864 * The verifier allows taking only one bpf_spin_lock at a time to avoid
3866 * Since only one bpf_spin_lock is allowed the checks are simpler than
3867 * reg_is_refcounted() logic. The verifier needs to remember only
3868 * one spin_lock instead of array of acquired_refs.
3869 * cur_state->active_spin_lock remembers which map value element got locked
3870 * and clears it after bpf_spin_unlock.
3872 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
3875 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
3876 struct bpf_verifier_state *cur = env->cur_state;
3877 bool is_const = tnum_is_const(reg->var_off);
3878 struct bpf_map *map = reg->map_ptr;
3879 u64 val = reg->var_off.value;
3883 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
3889 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
3893 if (!map_value_has_spin_lock(map)) {
3894 if (map->spin_lock_off == -E2BIG)
3896 "map '%s' has more than one 'struct bpf_spin_lock'\n",
3898 else if (map->spin_lock_off == -ENOENT)
3900 "map '%s' doesn't have 'struct bpf_spin_lock'\n",
3904 "map '%s' is not a struct type or bpf_spin_lock is mangled\n",
3908 if (map->spin_lock_off != val + reg->off) {
3909 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock'\n",
3914 if (cur->active_spin_lock) {
3916 "Locking two bpf_spin_locks are not allowed\n");
3919 cur->active_spin_lock = reg->id;
3921 if (!cur->active_spin_lock) {
3922 verbose(env, "bpf_spin_unlock without taking a lock\n");
3925 if (cur->active_spin_lock != reg->id) {
3926 verbose(env, "bpf_spin_unlock of different lock\n");
3929 cur->active_spin_lock = 0;
3934 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
3936 return type == ARG_PTR_TO_MEM ||
3937 type == ARG_PTR_TO_MEM_OR_NULL ||
3938 type == ARG_PTR_TO_UNINIT_MEM;
3941 static bool arg_type_is_mem_size(enum bpf_arg_type type)
3943 return type == ARG_CONST_SIZE ||
3944 type == ARG_CONST_SIZE_OR_ZERO;
3947 static bool arg_type_is_alloc_size(enum bpf_arg_type type)
3949 return type == ARG_CONST_ALLOC_SIZE_OR_ZERO;
3952 static bool arg_type_is_int_ptr(enum bpf_arg_type type)
3954 return type == ARG_PTR_TO_INT ||
3955 type == ARG_PTR_TO_LONG;
3958 static int int_ptr_type_to_size(enum bpf_arg_type type)
3960 if (type == ARG_PTR_TO_INT)
3962 else if (type == ARG_PTR_TO_LONG)
3968 static int resolve_map_arg_type(struct bpf_verifier_env *env,
3969 const struct bpf_call_arg_meta *meta,
3970 enum bpf_arg_type *arg_type)
3972 if (!meta->map_ptr) {
3973 /* kernel subsystem misconfigured verifier */
3974 verbose(env, "invalid map_ptr to access map->type\n");
3978 switch (meta->map_ptr->map_type) {
3979 case BPF_MAP_TYPE_SOCKMAP:
3980 case BPF_MAP_TYPE_SOCKHASH:
3981 if (*arg_type == ARG_PTR_TO_MAP_VALUE) {
3982 *arg_type = ARG_PTR_TO_BTF_ID_SOCK_COMMON;
3984 verbose(env, "invalid arg_type for sockmap/sockhash\n");
3995 struct bpf_reg_types {
3996 const enum bpf_reg_type types[10];
4000 static const struct bpf_reg_types map_key_value_types = {
4009 static const struct bpf_reg_types sock_types = {
4019 static const struct bpf_reg_types btf_id_sock_common_types = {
4027 .btf_id = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
4031 static const struct bpf_reg_types mem_types = {
4043 static const struct bpf_reg_types int_ptr_types = {
4052 static const struct bpf_reg_types fullsock_types = { .types = { PTR_TO_SOCKET } };
4053 static const struct bpf_reg_types scalar_types = { .types = { SCALAR_VALUE } };
4054 static const struct bpf_reg_types context_types = { .types = { PTR_TO_CTX } };
4055 static const struct bpf_reg_types alloc_mem_types = { .types = { PTR_TO_MEM } };
4056 static const struct bpf_reg_types const_map_ptr_types = { .types = { CONST_PTR_TO_MAP } };
4057 static const struct bpf_reg_types btf_ptr_types = { .types = { PTR_TO_BTF_ID } };
4058 static const struct bpf_reg_types spin_lock_types = { .types = { PTR_TO_MAP_VALUE } };
4059 static const struct bpf_reg_types percpu_btf_ptr_types = { .types = { PTR_TO_PERCPU_BTF_ID } };
4061 static const struct bpf_reg_types *compatible_reg_types[__BPF_ARG_TYPE_MAX] = {
4062 [ARG_PTR_TO_MAP_KEY] = &map_key_value_types,
4063 [ARG_PTR_TO_MAP_VALUE] = &map_key_value_types,
4064 [ARG_PTR_TO_UNINIT_MAP_VALUE] = &map_key_value_types,
4065 [ARG_PTR_TO_MAP_VALUE_OR_NULL] = &map_key_value_types,
4066 [ARG_CONST_SIZE] = &scalar_types,
4067 [ARG_CONST_SIZE_OR_ZERO] = &scalar_types,
4068 [ARG_CONST_ALLOC_SIZE_OR_ZERO] = &scalar_types,
4069 [ARG_CONST_MAP_PTR] = &const_map_ptr_types,
4070 [ARG_PTR_TO_CTX] = &context_types,
4071 [ARG_PTR_TO_CTX_OR_NULL] = &context_types,
4072 [ARG_PTR_TO_SOCK_COMMON] = &sock_types,
4074 [ARG_PTR_TO_BTF_ID_SOCK_COMMON] = &btf_id_sock_common_types,
4076 [ARG_PTR_TO_SOCKET] = &fullsock_types,
4077 [ARG_PTR_TO_SOCKET_OR_NULL] = &fullsock_types,
4078 [ARG_PTR_TO_BTF_ID] = &btf_ptr_types,
4079 [ARG_PTR_TO_SPIN_LOCK] = &spin_lock_types,
4080 [ARG_PTR_TO_MEM] = &mem_types,
4081 [ARG_PTR_TO_MEM_OR_NULL] = &mem_types,
4082 [ARG_PTR_TO_UNINIT_MEM] = &mem_types,
4083 [ARG_PTR_TO_ALLOC_MEM] = &alloc_mem_types,
4084 [ARG_PTR_TO_ALLOC_MEM_OR_NULL] = &alloc_mem_types,
4085 [ARG_PTR_TO_INT] = &int_ptr_types,
4086 [ARG_PTR_TO_LONG] = &int_ptr_types,
4087 [ARG_PTR_TO_PERCPU_BTF_ID] = &percpu_btf_ptr_types,
4090 static int check_reg_type(struct bpf_verifier_env *env, u32 regno,
4091 enum bpf_arg_type arg_type,
4092 const u32 *arg_btf_id)
4094 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4095 enum bpf_reg_type expected, type = reg->type;
4096 const struct bpf_reg_types *compatible;
4099 compatible = compatible_reg_types[arg_type];
4101 verbose(env, "verifier internal error: unsupported arg type %d\n", arg_type);
4105 for (i = 0; i < ARRAY_SIZE(compatible->types); i++) {
4106 expected = compatible->types[i];
4107 if (expected == NOT_INIT)
4110 if (type == expected)
4114 verbose(env, "R%d type=%s expected=", regno, reg_type_str[type]);
4115 for (j = 0; j + 1 < i; j++)
4116 verbose(env, "%s, ", reg_type_str[compatible->types[j]]);
4117 verbose(env, "%s\n", reg_type_str[compatible->types[j]]);
4121 if (type == PTR_TO_BTF_ID) {
4123 if (!compatible->btf_id) {
4124 verbose(env, "verifier internal error: missing arg compatible BTF ID\n");
4127 arg_btf_id = compatible->btf_id;
4130 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
4131 btf_vmlinux, *arg_btf_id)) {
4132 verbose(env, "R%d is of type %s but %s is expected\n",
4133 regno, kernel_type_name(reg->btf, reg->btf_id),
4134 kernel_type_name(btf_vmlinux, *arg_btf_id));
4138 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
4139 verbose(env, "R%d is a pointer to in-kernel struct with non-zero offset\n",
4148 static int check_func_arg(struct bpf_verifier_env *env, u32 arg,
4149 struct bpf_call_arg_meta *meta,
4150 const struct bpf_func_proto *fn)
4152 u32 regno = BPF_REG_1 + arg;
4153 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4154 enum bpf_arg_type arg_type = fn->arg_type[arg];
4155 enum bpf_reg_type type = reg->type;
4158 if (arg_type == ARG_DONTCARE)
4161 err = check_reg_arg(env, regno, SRC_OP);
4165 if (arg_type == ARG_ANYTHING) {
4166 if (is_pointer_value(env, regno)) {
4167 verbose(env, "R%d leaks addr into helper function\n",
4174 if (type_is_pkt_pointer(type) &&
4175 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
4176 verbose(env, "helper access to the packet is not allowed\n");
4180 if (arg_type == ARG_PTR_TO_MAP_VALUE ||
4181 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE ||
4182 arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL) {
4183 err = resolve_map_arg_type(env, meta, &arg_type);
4188 if (register_is_null(reg) && arg_type_may_be_null(arg_type))
4189 /* A NULL register has a SCALAR_VALUE type, so skip
4192 goto skip_type_check;
4194 err = check_reg_type(env, regno, arg_type, fn->arg_btf_id[arg]);
4198 if (type == PTR_TO_CTX) {
4199 err = check_ctx_reg(env, reg, regno);
4205 if (reg->ref_obj_id) {
4206 if (meta->ref_obj_id) {
4207 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
4208 regno, reg->ref_obj_id,
4212 meta->ref_obj_id = reg->ref_obj_id;
4215 if (arg_type == ARG_CONST_MAP_PTR) {
4216 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
4217 meta->map_ptr = reg->map_ptr;
4218 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
4219 /* bpf_map_xxx(..., map_ptr, ..., key) call:
4220 * check that [key, key + map->key_size) are within
4221 * stack limits and initialized
4223 if (!meta->map_ptr) {
4224 /* in function declaration map_ptr must come before
4225 * map_key, so that it's verified and known before
4226 * we have to check map_key here. Otherwise it means
4227 * that kernel subsystem misconfigured verifier
4229 verbose(env, "invalid map_ptr to access map->key\n");
4232 err = check_helper_mem_access(env, regno,
4233 meta->map_ptr->key_size, false,
4235 } else if (arg_type == ARG_PTR_TO_MAP_VALUE ||
4236 (arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL &&
4237 !register_is_null(reg)) ||
4238 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
4239 /* bpf_map_xxx(..., map_ptr, ..., value) call:
4240 * check [value, value + map->value_size) validity
4242 if (!meta->map_ptr) {
4243 /* kernel subsystem misconfigured verifier */
4244 verbose(env, "invalid map_ptr to access map->value\n");
4247 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE);
4248 err = check_helper_mem_access(env, regno,
4249 meta->map_ptr->value_size, false,
4251 } else if (arg_type == ARG_PTR_TO_PERCPU_BTF_ID) {
4253 verbose(env, "Helper has invalid btf_id in R%d\n", regno);
4256 meta->ret_btf = reg->btf;
4257 meta->ret_btf_id = reg->btf_id;
4258 } else if (arg_type == ARG_PTR_TO_SPIN_LOCK) {
4259 if (meta->func_id == BPF_FUNC_spin_lock) {
4260 if (process_spin_lock(env, regno, true))
4262 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
4263 if (process_spin_lock(env, regno, false))
4266 verbose(env, "verifier internal error\n");
4269 } else if (arg_type_is_mem_ptr(arg_type)) {
4270 /* The access to this pointer is only checked when we hit the
4271 * next is_mem_size argument below.
4273 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MEM);
4274 } else if (arg_type_is_mem_size(arg_type)) {
4275 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
4277 /* This is used to refine r0 return value bounds for helpers
4278 * that enforce this value as an upper bound on return values.
4279 * See do_refine_retval_range() for helpers that can refine
4280 * the return value. C type of helper is u32 so we pull register
4281 * bound from umax_value however, if negative verifier errors
4282 * out. Only upper bounds can be learned because retval is an
4283 * int type and negative retvals are allowed.
4285 meta->msize_max_value = reg->umax_value;
4287 /* The register is SCALAR_VALUE; the access check
4288 * happens using its boundaries.
4290 if (!tnum_is_const(reg->var_off))
4291 /* For unprivileged variable accesses, disable raw
4292 * mode so that the program is required to
4293 * initialize all the memory that the helper could
4294 * just partially fill up.
4298 if (reg->smin_value < 0) {
4299 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
4304 if (reg->umin_value == 0) {
4305 err = check_helper_mem_access(env, regno - 1, 0,
4312 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
4313 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
4317 err = check_helper_mem_access(env, regno - 1,
4319 zero_size_allowed, meta);
4321 err = mark_chain_precision(env, regno);
4322 } else if (arg_type_is_alloc_size(arg_type)) {
4323 if (!tnum_is_const(reg->var_off)) {
4324 verbose(env, "R%d unbounded size, use 'var &= const' or 'if (var < const)'\n",
4328 meta->mem_size = reg->var_off.value;
4329 } else if (arg_type_is_int_ptr(arg_type)) {
4330 int size = int_ptr_type_to_size(arg_type);
4332 err = check_helper_mem_access(env, regno, size, false, meta);
4335 err = check_ptr_alignment(env, reg, 0, size, true);
4341 static bool may_update_sockmap(struct bpf_verifier_env *env, int func_id)
4343 enum bpf_attach_type eatype = env->prog->expected_attach_type;
4344 enum bpf_prog_type type = resolve_prog_type(env->prog);
4346 if (func_id != BPF_FUNC_map_update_elem)
4349 /* It's not possible to get access to a locked struct sock in these
4350 * contexts, so updating is safe.
4353 case BPF_PROG_TYPE_TRACING:
4354 if (eatype == BPF_TRACE_ITER)
4357 case BPF_PROG_TYPE_SOCKET_FILTER:
4358 case BPF_PROG_TYPE_SCHED_CLS:
4359 case BPF_PROG_TYPE_SCHED_ACT:
4360 case BPF_PROG_TYPE_XDP:
4361 case BPF_PROG_TYPE_SK_REUSEPORT:
4362 case BPF_PROG_TYPE_FLOW_DISSECTOR:
4363 case BPF_PROG_TYPE_SK_LOOKUP:
4369 verbose(env, "cannot update sockmap in this context\n");
4373 static bool allow_tail_call_in_subprogs(struct bpf_verifier_env *env)
4375 return env->prog->jit_requested && IS_ENABLED(CONFIG_X86_64);
4378 static int check_map_func_compatibility(struct bpf_verifier_env *env,
4379 struct bpf_map *map, int func_id)
4384 /* We need a two way check, first is from map perspective ... */
4385 switch (map->map_type) {
4386 case BPF_MAP_TYPE_PROG_ARRAY:
4387 if (func_id != BPF_FUNC_tail_call)
4390 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
4391 if (func_id != BPF_FUNC_perf_event_read &&
4392 func_id != BPF_FUNC_perf_event_output &&
4393 func_id != BPF_FUNC_skb_output &&
4394 func_id != BPF_FUNC_perf_event_read_value &&
4395 func_id != BPF_FUNC_xdp_output)
4398 case BPF_MAP_TYPE_RINGBUF:
4399 if (func_id != BPF_FUNC_ringbuf_output &&
4400 func_id != BPF_FUNC_ringbuf_reserve &&
4401 func_id != BPF_FUNC_ringbuf_submit &&
4402 func_id != BPF_FUNC_ringbuf_discard &&
4403 func_id != BPF_FUNC_ringbuf_query)
4406 case BPF_MAP_TYPE_STACK_TRACE:
4407 if (func_id != BPF_FUNC_get_stackid)
4410 case BPF_MAP_TYPE_CGROUP_ARRAY:
4411 if (func_id != BPF_FUNC_skb_under_cgroup &&
4412 func_id != BPF_FUNC_current_task_under_cgroup)
4415 case BPF_MAP_TYPE_CGROUP_STORAGE:
4416 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
4417 if (func_id != BPF_FUNC_get_local_storage)
4420 case BPF_MAP_TYPE_DEVMAP:
4421 case BPF_MAP_TYPE_DEVMAP_HASH:
4422 if (func_id != BPF_FUNC_redirect_map &&
4423 func_id != BPF_FUNC_map_lookup_elem)
4426 /* Restrict bpf side of cpumap and xskmap, open when use-cases
4429 case BPF_MAP_TYPE_CPUMAP:
4430 if (func_id != BPF_FUNC_redirect_map)
4433 case BPF_MAP_TYPE_XSKMAP:
4434 if (func_id != BPF_FUNC_redirect_map &&
4435 func_id != BPF_FUNC_map_lookup_elem)
4438 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
4439 case BPF_MAP_TYPE_HASH_OF_MAPS:
4440 if (func_id != BPF_FUNC_map_lookup_elem)
4443 case BPF_MAP_TYPE_SOCKMAP:
4444 if (func_id != BPF_FUNC_sk_redirect_map &&
4445 func_id != BPF_FUNC_sock_map_update &&
4446 func_id != BPF_FUNC_map_delete_elem &&
4447 func_id != BPF_FUNC_msg_redirect_map &&
4448 func_id != BPF_FUNC_sk_select_reuseport &&
4449 func_id != BPF_FUNC_map_lookup_elem &&
4450 !may_update_sockmap(env, func_id))
4453 case BPF_MAP_TYPE_SOCKHASH:
4454 if (func_id != BPF_FUNC_sk_redirect_hash &&
4455 func_id != BPF_FUNC_sock_hash_update &&
4456 func_id != BPF_FUNC_map_delete_elem &&
4457 func_id != BPF_FUNC_msg_redirect_hash &&
4458 func_id != BPF_FUNC_sk_select_reuseport &&
4459 func_id != BPF_FUNC_map_lookup_elem &&
4460 !may_update_sockmap(env, func_id))
4463 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
4464 if (func_id != BPF_FUNC_sk_select_reuseport)
4467 case BPF_MAP_TYPE_QUEUE:
4468 case BPF_MAP_TYPE_STACK:
4469 if (func_id != BPF_FUNC_map_peek_elem &&
4470 func_id != BPF_FUNC_map_pop_elem &&
4471 func_id != BPF_FUNC_map_push_elem)
4474 case BPF_MAP_TYPE_SK_STORAGE:
4475 if (func_id != BPF_FUNC_sk_storage_get &&
4476 func_id != BPF_FUNC_sk_storage_delete)
4479 case BPF_MAP_TYPE_INODE_STORAGE:
4480 if (func_id != BPF_FUNC_inode_storage_get &&
4481 func_id != BPF_FUNC_inode_storage_delete)
4484 case BPF_MAP_TYPE_TASK_STORAGE:
4485 if (func_id != BPF_FUNC_task_storage_get &&
4486 func_id != BPF_FUNC_task_storage_delete)
4493 /* ... and second from the function itself. */
4495 case BPF_FUNC_tail_call:
4496 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
4498 if (env->subprog_cnt > 1 && !allow_tail_call_in_subprogs(env)) {
4499 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
4503 case BPF_FUNC_perf_event_read:
4504 case BPF_FUNC_perf_event_output:
4505 case BPF_FUNC_perf_event_read_value:
4506 case BPF_FUNC_skb_output:
4507 case BPF_FUNC_xdp_output:
4508 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
4511 case BPF_FUNC_get_stackid:
4512 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
4515 case BPF_FUNC_current_task_under_cgroup:
4516 case BPF_FUNC_skb_under_cgroup:
4517 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
4520 case BPF_FUNC_redirect_map:
4521 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
4522 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
4523 map->map_type != BPF_MAP_TYPE_CPUMAP &&
4524 map->map_type != BPF_MAP_TYPE_XSKMAP)
4527 case BPF_FUNC_sk_redirect_map:
4528 case BPF_FUNC_msg_redirect_map:
4529 case BPF_FUNC_sock_map_update:
4530 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
4533 case BPF_FUNC_sk_redirect_hash:
4534 case BPF_FUNC_msg_redirect_hash:
4535 case BPF_FUNC_sock_hash_update:
4536 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
4539 case BPF_FUNC_get_local_storage:
4540 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
4541 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
4544 case BPF_FUNC_sk_select_reuseport:
4545 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
4546 map->map_type != BPF_MAP_TYPE_SOCKMAP &&
4547 map->map_type != BPF_MAP_TYPE_SOCKHASH)
4550 case BPF_FUNC_map_peek_elem:
4551 case BPF_FUNC_map_pop_elem:
4552 case BPF_FUNC_map_push_elem:
4553 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
4554 map->map_type != BPF_MAP_TYPE_STACK)
4557 case BPF_FUNC_sk_storage_get:
4558 case BPF_FUNC_sk_storage_delete:
4559 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
4562 case BPF_FUNC_inode_storage_get:
4563 case BPF_FUNC_inode_storage_delete:
4564 if (map->map_type != BPF_MAP_TYPE_INODE_STORAGE)
4567 case BPF_FUNC_task_storage_get:
4568 case BPF_FUNC_task_storage_delete:
4569 if (map->map_type != BPF_MAP_TYPE_TASK_STORAGE)
4578 verbose(env, "cannot pass map_type %d into func %s#%d\n",
4579 map->map_type, func_id_name(func_id), func_id);
4583 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
4587 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
4589 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
4591 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
4593 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
4595 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
4598 /* We only support one arg being in raw mode at the moment,
4599 * which is sufficient for the helper functions we have
4605 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
4606 enum bpf_arg_type arg_next)
4608 return (arg_type_is_mem_ptr(arg_curr) &&
4609 !arg_type_is_mem_size(arg_next)) ||
4610 (!arg_type_is_mem_ptr(arg_curr) &&
4611 arg_type_is_mem_size(arg_next));
4614 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
4616 /* bpf_xxx(..., buf, len) call will access 'len'
4617 * bytes from memory 'buf'. Both arg types need
4618 * to be paired, so make sure there's no buggy
4619 * helper function specification.
4621 if (arg_type_is_mem_size(fn->arg1_type) ||
4622 arg_type_is_mem_ptr(fn->arg5_type) ||
4623 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
4624 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
4625 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
4626 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
4632 static bool check_refcount_ok(const struct bpf_func_proto *fn, int func_id)
4636 if (arg_type_may_be_refcounted(fn->arg1_type))
4638 if (arg_type_may_be_refcounted(fn->arg2_type))
4640 if (arg_type_may_be_refcounted(fn->arg3_type))
4642 if (arg_type_may_be_refcounted(fn->arg4_type))
4644 if (arg_type_may_be_refcounted(fn->arg5_type))
4647 /* A reference acquiring function cannot acquire
4648 * another refcounted ptr.
4650 if (may_be_acquire_function(func_id) && count)
4653 /* We only support one arg being unreferenced at the moment,
4654 * which is sufficient for the helper functions we have right now.
4659 static bool check_btf_id_ok(const struct bpf_func_proto *fn)
4663 for (i = 0; i < ARRAY_SIZE(fn->arg_type); i++) {
4664 if (fn->arg_type[i] == ARG_PTR_TO_BTF_ID && !fn->arg_btf_id[i])
4667 if (fn->arg_type[i] != ARG_PTR_TO_BTF_ID && fn->arg_btf_id[i])
4674 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
4676 return check_raw_mode_ok(fn) &&
4677 check_arg_pair_ok(fn) &&
4678 check_btf_id_ok(fn) &&
4679 check_refcount_ok(fn, func_id) ? 0 : -EINVAL;
4682 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
4683 * are now invalid, so turn them into unknown SCALAR_VALUE.
4685 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
4686 struct bpf_func_state *state)
4688 struct bpf_reg_state *regs = state->regs, *reg;
4691 for (i = 0; i < MAX_BPF_REG; i++)
4692 if (reg_is_pkt_pointer_any(®s[i]))
4693 mark_reg_unknown(env, regs, i);
4695 bpf_for_each_spilled_reg(i, state, reg) {
4698 if (reg_is_pkt_pointer_any(reg))
4699 __mark_reg_unknown(env, reg);
4703 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
4705 struct bpf_verifier_state *vstate = env->cur_state;
4708 for (i = 0; i <= vstate->curframe; i++)
4709 __clear_all_pkt_pointers(env, vstate->frame[i]);
4714 BEYOND_PKT_END = -2,
4717 static void mark_pkt_end(struct bpf_verifier_state *vstate, int regn, bool range_open)
4719 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4720 struct bpf_reg_state *reg = &state->regs[regn];
4722 if (reg->type != PTR_TO_PACKET)
4723 /* PTR_TO_PACKET_META is not supported yet */
4726 /* The 'reg' is pkt > pkt_end or pkt >= pkt_end.
4727 * How far beyond pkt_end it goes is unknown.
4728 * if (!range_open) it's the case of pkt >= pkt_end
4729 * if (range_open) it's the case of pkt > pkt_end
4730 * hence this pointer is at least 1 byte bigger than pkt_end
4733 reg->range = BEYOND_PKT_END;
4735 reg->range = AT_PKT_END;
4738 static void release_reg_references(struct bpf_verifier_env *env,
4739 struct bpf_func_state *state,
4742 struct bpf_reg_state *regs = state->regs, *reg;
4745 for (i = 0; i < MAX_BPF_REG; i++)
4746 if (regs[i].ref_obj_id == ref_obj_id)
4747 mark_reg_unknown(env, regs, i);
4749 bpf_for_each_spilled_reg(i, state, reg) {
4752 if (reg->ref_obj_id == ref_obj_id)
4753 __mark_reg_unknown(env, reg);
4757 /* The pointer with the specified id has released its reference to kernel
4758 * resources. Identify all copies of the same pointer and clear the reference.
4760 static int release_reference(struct bpf_verifier_env *env,
4763 struct bpf_verifier_state *vstate = env->cur_state;
4767 err = release_reference_state(cur_func(env), ref_obj_id);
4771 for (i = 0; i <= vstate->curframe; i++)
4772 release_reg_references(env, vstate->frame[i], ref_obj_id);
4777 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
4778 struct bpf_reg_state *regs)
4782 /* after the call registers r0 - r5 were scratched */
4783 for (i = 0; i < CALLER_SAVED_REGS; i++) {
4784 mark_reg_not_init(env, regs, caller_saved[i]);
4785 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
4789 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
4792 struct bpf_verifier_state *state = env->cur_state;
4793 struct bpf_func_info_aux *func_info_aux;
4794 struct bpf_func_state *caller, *callee;
4795 int i, err, subprog, target_insn;
4796 bool is_global = false;
4798 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
4799 verbose(env, "the call stack of %d frames is too deep\n",
4800 state->curframe + 2);
4804 target_insn = *insn_idx + insn->imm;
4805 subprog = find_subprog(env, target_insn + 1);
4807 verbose(env, "verifier bug. No program starts at insn %d\n",
4812 caller = state->frame[state->curframe];
4813 if (state->frame[state->curframe + 1]) {
4814 verbose(env, "verifier bug. Frame %d already allocated\n",
4815 state->curframe + 1);
4819 func_info_aux = env->prog->aux->func_info_aux;
4821 is_global = func_info_aux[subprog].linkage == BTF_FUNC_GLOBAL;
4822 err = btf_check_func_arg_match(env, subprog, caller->regs);
4827 verbose(env, "Caller passes invalid args into func#%d\n",
4831 if (env->log.level & BPF_LOG_LEVEL)
4833 "Func#%d is global and valid. Skipping.\n",
4835 clear_caller_saved_regs(env, caller->regs);
4837 /* All global functions return SCALAR_VALUE */
4838 mark_reg_unknown(env, caller->regs, BPF_REG_0);
4840 /* continue with next insn after call */
4845 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
4848 state->frame[state->curframe + 1] = callee;
4850 /* callee cannot access r0, r6 - r9 for reading and has to write
4851 * into its own stack before reading from it.
4852 * callee can read/write into caller's stack
4854 init_func_state(env, callee,
4855 /* remember the callsite, it will be used by bpf_exit */
4856 *insn_idx /* callsite */,
4857 state->curframe + 1 /* frameno within this callchain */,
4858 subprog /* subprog number within this prog */);
4860 /* Transfer references to the callee */
4861 err = transfer_reference_state(callee, caller);
4865 /* copy r1 - r5 args that callee can access. The copy includes parent
4866 * pointers, which connects us up to the liveness chain
4868 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
4869 callee->regs[i] = caller->regs[i];
4871 clear_caller_saved_regs(env, caller->regs);
4873 /* only increment it after check_reg_arg() finished */
4876 /* and go analyze first insn of the callee */
4877 *insn_idx = target_insn;
4879 if (env->log.level & BPF_LOG_LEVEL) {
4880 verbose(env, "caller:\n");
4881 print_verifier_state(env, caller);
4882 verbose(env, "callee:\n");
4883 print_verifier_state(env, callee);
4888 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
4890 struct bpf_verifier_state *state = env->cur_state;
4891 struct bpf_func_state *caller, *callee;
4892 struct bpf_reg_state *r0;
4895 callee = state->frame[state->curframe];
4896 r0 = &callee->regs[BPF_REG_0];
4897 if (r0->type == PTR_TO_STACK) {
4898 /* technically it's ok to return caller's stack pointer
4899 * (or caller's caller's pointer) back to the caller,
4900 * since these pointers are valid. Only current stack
4901 * pointer will be invalid as soon as function exits,
4902 * but let's be conservative
4904 verbose(env, "cannot return stack pointer to the caller\n");
4909 caller = state->frame[state->curframe];
4910 /* return to the caller whatever r0 had in the callee */
4911 caller->regs[BPF_REG_0] = *r0;
4913 /* Transfer references to the caller */
4914 err = transfer_reference_state(caller, callee);
4918 *insn_idx = callee->callsite + 1;
4919 if (env->log.level & BPF_LOG_LEVEL) {
4920 verbose(env, "returning from callee:\n");
4921 print_verifier_state(env, callee);
4922 verbose(env, "to caller at %d:\n", *insn_idx);
4923 print_verifier_state(env, caller);
4925 /* clear everything in the callee */
4926 free_func_state(callee);
4927 state->frame[state->curframe + 1] = NULL;
4931 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
4933 struct bpf_call_arg_meta *meta)
4935 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
4937 if (ret_type != RET_INTEGER ||
4938 (func_id != BPF_FUNC_get_stack &&
4939 func_id != BPF_FUNC_probe_read_str &&
4940 func_id != BPF_FUNC_probe_read_kernel_str &&
4941 func_id != BPF_FUNC_probe_read_user_str))
4944 ret_reg->smax_value = meta->msize_max_value;
4945 ret_reg->s32_max_value = meta->msize_max_value;
4946 ret_reg->smin_value = -MAX_ERRNO;
4947 ret_reg->s32_min_value = -MAX_ERRNO;
4948 __reg_deduce_bounds(ret_reg);
4949 __reg_bound_offset(ret_reg);
4950 __update_reg_bounds(ret_reg);
4954 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
4955 int func_id, int insn_idx)
4957 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
4958 struct bpf_map *map = meta->map_ptr;
4960 if (func_id != BPF_FUNC_tail_call &&
4961 func_id != BPF_FUNC_map_lookup_elem &&
4962 func_id != BPF_FUNC_map_update_elem &&
4963 func_id != BPF_FUNC_map_delete_elem &&
4964 func_id != BPF_FUNC_map_push_elem &&
4965 func_id != BPF_FUNC_map_pop_elem &&
4966 func_id != BPF_FUNC_map_peek_elem)
4970 verbose(env, "kernel subsystem misconfigured verifier\n");
4974 /* In case of read-only, some additional restrictions
4975 * need to be applied in order to prevent altering the
4976 * state of the map from program side.
4978 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
4979 (func_id == BPF_FUNC_map_delete_elem ||
4980 func_id == BPF_FUNC_map_update_elem ||
4981 func_id == BPF_FUNC_map_push_elem ||
4982 func_id == BPF_FUNC_map_pop_elem)) {
4983 verbose(env, "write into map forbidden\n");
4987 if (!BPF_MAP_PTR(aux->map_ptr_state))
4988 bpf_map_ptr_store(aux, meta->map_ptr,
4989 !meta->map_ptr->bypass_spec_v1);
4990 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
4991 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
4992 !meta->map_ptr->bypass_spec_v1);
4997 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
4998 int func_id, int insn_idx)
5000 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
5001 struct bpf_reg_state *regs = cur_regs(env), *reg;
5002 struct bpf_map *map = meta->map_ptr;
5007 if (func_id != BPF_FUNC_tail_call)
5009 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
5010 verbose(env, "kernel subsystem misconfigured verifier\n");
5014 range = tnum_range(0, map->max_entries - 1);
5015 reg = ®s[BPF_REG_3];
5017 if (!register_is_const(reg) || !tnum_in(range, reg->var_off)) {
5018 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
5022 err = mark_chain_precision(env, BPF_REG_3);
5026 val = reg->var_off.value;
5027 if (bpf_map_key_unseen(aux))
5028 bpf_map_key_store(aux, val);
5029 else if (!bpf_map_key_poisoned(aux) &&
5030 bpf_map_key_immediate(aux) != val)
5031 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
5035 static int check_reference_leak(struct bpf_verifier_env *env)
5037 struct bpf_func_state *state = cur_func(env);
5040 for (i = 0; i < state->acquired_refs; i++) {
5041 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
5042 state->refs[i].id, state->refs[i].insn_idx);
5044 return state->acquired_refs ? -EINVAL : 0;
5047 static int check_helper_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
5049 const struct bpf_func_proto *fn = NULL;
5050 struct bpf_reg_state *regs;
5051 struct bpf_call_arg_meta meta;
5055 /* find function prototype */
5056 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
5057 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
5062 if (env->ops->get_func_proto)
5063 fn = env->ops->get_func_proto(func_id, env->prog);
5065 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
5070 /* eBPF programs must be GPL compatible to use GPL-ed functions */
5071 if (!env->prog->gpl_compatible && fn->gpl_only) {
5072 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
5076 if (fn->allowed && !fn->allowed(env->prog)) {
5077 verbose(env, "helper call is not allowed in probe\n");
5081 /* With LD_ABS/IND some JITs save/restore skb from r1. */
5082 changes_data = bpf_helper_changes_pkt_data(fn->func);
5083 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
5084 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
5085 func_id_name(func_id), func_id);
5089 memset(&meta, 0, sizeof(meta));
5090 meta.pkt_access = fn->pkt_access;
5092 err = check_func_proto(fn, func_id);
5094 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
5095 func_id_name(func_id), func_id);
5099 meta.func_id = func_id;
5101 for (i = 0; i < 5; i++) {
5102 err = check_func_arg(env, i, &meta, fn);
5107 err = record_func_map(env, &meta, func_id, insn_idx);
5111 err = record_func_key(env, &meta, func_id, insn_idx);
5115 /* Mark slots with STACK_MISC in case of raw mode, stack offset
5116 * is inferred from register state.
5118 for (i = 0; i < meta.access_size; i++) {
5119 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
5120 BPF_WRITE, -1, false);
5125 if (func_id == BPF_FUNC_tail_call) {
5126 err = check_reference_leak(env);
5128 verbose(env, "tail_call would lead to reference leak\n");
5131 } else if (is_release_function(func_id)) {
5132 err = release_reference(env, meta.ref_obj_id);
5134 verbose(env, "func %s#%d reference has not been acquired before\n",
5135 func_id_name(func_id), func_id);
5140 regs = cur_regs(env);
5142 /* check that flags argument in get_local_storage(map, flags) is 0,
5143 * this is required because get_local_storage() can't return an error.
5145 if (func_id == BPF_FUNC_get_local_storage &&
5146 !register_is_null(®s[BPF_REG_2])) {
5147 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
5151 /* reset caller saved regs */
5152 for (i = 0; i < CALLER_SAVED_REGS; i++) {
5153 mark_reg_not_init(env, regs, caller_saved[i]);
5154 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
5157 /* helper call returns 64-bit value. */
5158 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
5160 /* update return register (already marked as written above) */
5161 if (fn->ret_type == RET_INTEGER) {
5162 /* sets type to SCALAR_VALUE */
5163 mark_reg_unknown(env, regs, BPF_REG_0);
5164 } else if (fn->ret_type == RET_VOID) {
5165 regs[BPF_REG_0].type = NOT_INIT;
5166 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL ||
5167 fn->ret_type == RET_PTR_TO_MAP_VALUE) {
5168 /* There is no offset yet applied, variable or fixed */
5169 mark_reg_known_zero(env, regs, BPF_REG_0);
5170 /* remember map_ptr, so that check_map_access()
5171 * can check 'value_size' boundary of memory access
5172 * to map element returned from bpf_map_lookup_elem()
5174 if (meta.map_ptr == NULL) {
5176 "kernel subsystem misconfigured verifier\n");
5179 regs[BPF_REG_0].map_ptr = meta.map_ptr;
5180 if (fn->ret_type == RET_PTR_TO_MAP_VALUE) {
5181 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE;
5182 if (map_value_has_spin_lock(meta.map_ptr))
5183 regs[BPF_REG_0].id = ++env->id_gen;
5185 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
5187 } else if (fn->ret_type == RET_PTR_TO_SOCKET_OR_NULL) {
5188 mark_reg_known_zero(env, regs, BPF_REG_0);
5189 regs[BPF_REG_0].type = PTR_TO_SOCKET_OR_NULL;
5190 } else if (fn->ret_type == RET_PTR_TO_SOCK_COMMON_OR_NULL) {
5191 mark_reg_known_zero(env, regs, BPF_REG_0);
5192 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON_OR_NULL;
5193 } else if (fn->ret_type == RET_PTR_TO_TCP_SOCK_OR_NULL) {
5194 mark_reg_known_zero(env, regs, BPF_REG_0);
5195 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK_OR_NULL;
5196 } else if (fn->ret_type == RET_PTR_TO_ALLOC_MEM_OR_NULL) {
5197 mark_reg_known_zero(env, regs, BPF_REG_0);
5198 regs[BPF_REG_0].type = PTR_TO_MEM_OR_NULL;
5199 regs[BPF_REG_0].mem_size = meta.mem_size;
5200 } else if (fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID_OR_NULL ||
5201 fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID) {
5202 const struct btf_type *t;
5204 mark_reg_known_zero(env, regs, BPF_REG_0);
5205 t = btf_type_skip_modifiers(meta.ret_btf, meta.ret_btf_id, NULL);
5206 if (!btf_type_is_struct(t)) {
5208 const struct btf_type *ret;
5211 /* resolve the type size of ksym. */
5212 ret = btf_resolve_size(meta.ret_btf, t, &tsize);
5214 tname = btf_name_by_offset(meta.ret_btf, t->name_off);
5215 verbose(env, "unable to resolve the size of type '%s': %ld\n",
5216 tname, PTR_ERR(ret));
5219 regs[BPF_REG_0].type =
5220 fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID ?
5221 PTR_TO_MEM : PTR_TO_MEM_OR_NULL;
5222 regs[BPF_REG_0].mem_size = tsize;
5224 regs[BPF_REG_0].type =
5225 fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID ?
5226 PTR_TO_BTF_ID : PTR_TO_BTF_ID_OR_NULL;
5227 regs[BPF_REG_0].btf = meta.ret_btf;
5228 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
5230 } else if (fn->ret_type == RET_PTR_TO_BTF_ID_OR_NULL ||
5231 fn->ret_type == RET_PTR_TO_BTF_ID) {
5234 mark_reg_known_zero(env, regs, BPF_REG_0);
5235 regs[BPF_REG_0].type = fn->ret_type == RET_PTR_TO_BTF_ID ?
5237 PTR_TO_BTF_ID_OR_NULL;
5238 ret_btf_id = *fn->ret_btf_id;
5239 if (ret_btf_id == 0) {
5240 verbose(env, "invalid return type %d of func %s#%d\n",
5241 fn->ret_type, func_id_name(func_id), func_id);
5244 /* current BPF helper definitions are only coming from
5245 * built-in code with type IDs from vmlinux BTF
5247 regs[BPF_REG_0].btf = btf_vmlinux;
5248 regs[BPF_REG_0].btf_id = ret_btf_id;
5250 verbose(env, "unknown return type %d of func %s#%d\n",
5251 fn->ret_type, func_id_name(func_id), func_id);
5255 if (reg_type_may_be_null(regs[BPF_REG_0].type))
5256 regs[BPF_REG_0].id = ++env->id_gen;
5258 if (is_ptr_cast_function(func_id)) {
5259 /* For release_reference() */
5260 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
5261 } else if (is_acquire_function(func_id, meta.map_ptr)) {
5262 int id = acquire_reference_state(env, insn_idx);
5266 /* For mark_ptr_or_null_reg() */
5267 regs[BPF_REG_0].id = id;
5268 /* For release_reference() */
5269 regs[BPF_REG_0].ref_obj_id = id;
5272 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
5274 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
5278 if ((func_id == BPF_FUNC_get_stack ||
5279 func_id == BPF_FUNC_get_task_stack) &&
5280 !env->prog->has_callchain_buf) {
5281 const char *err_str;
5283 #ifdef CONFIG_PERF_EVENTS
5284 err = get_callchain_buffers(sysctl_perf_event_max_stack);
5285 err_str = "cannot get callchain buffer for func %s#%d\n";
5288 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
5291 verbose(env, err_str, func_id_name(func_id), func_id);
5295 env->prog->has_callchain_buf = true;
5298 if (func_id == BPF_FUNC_get_stackid || func_id == BPF_FUNC_get_stack)
5299 env->prog->call_get_stack = true;
5302 clear_all_pkt_pointers(env);
5306 static bool signed_add_overflows(s64 a, s64 b)
5308 /* Do the add in u64, where overflow is well-defined */
5309 s64 res = (s64)((u64)a + (u64)b);
5316 static bool signed_add32_overflows(s32 a, s32 b)
5318 /* Do the add in u32, where overflow is well-defined */
5319 s32 res = (s32)((u32)a + (u32)b);
5326 static bool signed_sub_overflows(s64 a, s64 b)
5328 /* Do the sub in u64, where overflow is well-defined */
5329 s64 res = (s64)((u64)a - (u64)b);
5336 static bool signed_sub32_overflows(s32 a, s32 b)
5338 /* Do the sub in u32, where overflow is well-defined */
5339 s32 res = (s32)((u32)a - (u32)b);
5346 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
5347 const struct bpf_reg_state *reg,
5348 enum bpf_reg_type type)
5350 bool known = tnum_is_const(reg->var_off);
5351 s64 val = reg->var_off.value;
5352 s64 smin = reg->smin_value;
5354 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
5355 verbose(env, "math between %s pointer and %lld is not allowed\n",
5356 reg_type_str[type], val);
5360 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
5361 verbose(env, "%s pointer offset %d is not allowed\n",
5362 reg_type_str[type], reg->off);
5366 if (smin == S64_MIN) {
5367 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
5368 reg_type_str[type]);
5372 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
5373 verbose(env, "value %lld makes %s pointer be out of bounds\n",
5374 smin, reg_type_str[type]);
5381 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
5383 return &env->insn_aux_data[env->insn_idx];
5386 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
5387 u32 *ptr_limit, u8 opcode, bool off_is_neg)
5389 bool mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
5390 (opcode == BPF_SUB && !off_is_neg);
5393 switch (ptr_reg->type) {
5395 /* Indirect variable offset stack access is prohibited in
5396 * unprivileged mode so it's not handled here.
5398 off = ptr_reg->off + ptr_reg->var_off.value;
5400 *ptr_limit = MAX_BPF_STACK + off;
5404 case PTR_TO_MAP_VALUE:
5406 *ptr_limit = ptr_reg->umax_value + ptr_reg->off;
5408 off = ptr_reg->smin_value + ptr_reg->off;
5409 *ptr_limit = ptr_reg->map_ptr->value_size - off;
5417 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
5418 const struct bpf_insn *insn)
5420 return env->bypass_spec_v1 || BPF_SRC(insn->code) == BPF_K;
5423 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
5424 u32 alu_state, u32 alu_limit)
5426 /* If we arrived here from different branches with different
5427 * state or limits to sanitize, then this won't work.
5429 if (aux->alu_state &&
5430 (aux->alu_state != alu_state ||
5431 aux->alu_limit != alu_limit))
5434 /* Corresponding fixup done in fixup_bpf_calls(). */
5435 aux->alu_state = alu_state;
5436 aux->alu_limit = alu_limit;
5440 static int sanitize_val_alu(struct bpf_verifier_env *env,
5441 struct bpf_insn *insn)
5443 struct bpf_insn_aux_data *aux = cur_aux(env);
5445 if (can_skip_alu_sanitation(env, insn))
5448 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
5451 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
5452 struct bpf_insn *insn,
5453 const struct bpf_reg_state *ptr_reg,
5454 struct bpf_reg_state *dst_reg,
5457 struct bpf_verifier_state *vstate = env->cur_state;
5458 struct bpf_insn_aux_data *aux = cur_aux(env);
5459 bool ptr_is_dst_reg = ptr_reg == dst_reg;
5460 u8 opcode = BPF_OP(insn->code);
5461 u32 alu_state, alu_limit;
5462 struct bpf_reg_state tmp;
5465 if (can_skip_alu_sanitation(env, insn))
5468 /* We already marked aux for masking from non-speculative
5469 * paths, thus we got here in the first place. We only care
5470 * to explore bad access from here.
5472 if (vstate->speculative)
5475 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
5476 alu_state |= ptr_is_dst_reg ?
5477 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
5479 if (retrieve_ptr_limit(ptr_reg, &alu_limit, opcode, off_is_neg))
5481 if (update_alu_sanitation_state(aux, alu_state, alu_limit))
5484 /* Simulate and find potential out-of-bounds access under
5485 * speculative execution from truncation as a result of
5486 * masking when off was not within expected range. If off
5487 * sits in dst, then we temporarily need to move ptr there
5488 * to simulate dst (== 0) +/-= ptr. Needed, for example,
5489 * for cases where we use K-based arithmetic in one direction
5490 * and truncated reg-based in the other in order to explore
5493 if (!ptr_is_dst_reg) {
5495 *dst_reg = *ptr_reg;
5497 ret = push_stack(env, env->insn_idx + 1, env->insn_idx, true);
5498 if (!ptr_is_dst_reg && ret)
5500 return !ret ? -EFAULT : 0;
5503 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
5504 * Caller should also handle BPF_MOV case separately.
5505 * If we return -EACCES, caller may want to try again treating pointer as a
5506 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
5508 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
5509 struct bpf_insn *insn,
5510 const struct bpf_reg_state *ptr_reg,
5511 const struct bpf_reg_state *off_reg)
5513 struct bpf_verifier_state *vstate = env->cur_state;
5514 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5515 struct bpf_reg_state *regs = state->regs, *dst_reg;
5516 bool known = tnum_is_const(off_reg->var_off);
5517 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
5518 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
5519 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
5520 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
5521 u32 dst = insn->dst_reg, src = insn->src_reg;
5522 u8 opcode = BPF_OP(insn->code);
5525 dst_reg = ®s[dst];
5527 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
5528 smin_val > smax_val || umin_val > umax_val) {
5529 /* Taint dst register if offset had invalid bounds derived from
5530 * e.g. dead branches.
5532 __mark_reg_unknown(env, dst_reg);
5536 if (BPF_CLASS(insn->code) != BPF_ALU64) {
5537 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
5538 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
5539 __mark_reg_unknown(env, dst_reg);
5544 "R%d 32-bit pointer arithmetic prohibited\n",
5549 switch (ptr_reg->type) {
5550 case PTR_TO_MAP_VALUE_OR_NULL:
5551 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
5552 dst, reg_type_str[ptr_reg->type]);
5554 case CONST_PTR_TO_MAP:
5555 /* smin_val represents the known value */
5556 if (known && smin_val == 0 && opcode == BPF_ADD)
5559 case PTR_TO_PACKET_END:
5561 case PTR_TO_SOCKET_OR_NULL:
5562 case PTR_TO_SOCK_COMMON:
5563 case PTR_TO_SOCK_COMMON_OR_NULL:
5564 case PTR_TO_TCP_SOCK:
5565 case PTR_TO_TCP_SOCK_OR_NULL:
5566 case PTR_TO_XDP_SOCK:
5567 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
5568 dst, reg_type_str[ptr_reg->type]);
5570 case PTR_TO_MAP_VALUE:
5571 if (!env->allow_ptr_leaks && !known && (smin_val < 0) != (smax_val < 0)) {
5572 verbose(env, "R%d has unknown scalar with mixed signed bounds, pointer arithmetic with it prohibited for !root\n",
5573 off_reg == dst_reg ? dst : src);
5581 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
5582 * The id may be overwritten later if we create a new variable offset.
5584 dst_reg->type = ptr_reg->type;
5585 dst_reg->id = ptr_reg->id;
5587 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
5588 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
5591 /* pointer types do not carry 32-bit bounds at the moment. */
5592 __mark_reg32_unbounded(dst_reg);
5596 ret = sanitize_ptr_alu(env, insn, ptr_reg, dst_reg, smin_val < 0);
5598 verbose(env, "R%d tried to add from different maps or paths\n", dst);
5601 /* We can take a fixed offset as long as it doesn't overflow
5602 * the s32 'off' field
5604 if (known && (ptr_reg->off + smin_val ==
5605 (s64)(s32)(ptr_reg->off + smin_val))) {
5606 /* pointer += K. Accumulate it into fixed offset */
5607 dst_reg->smin_value = smin_ptr;
5608 dst_reg->smax_value = smax_ptr;
5609 dst_reg->umin_value = umin_ptr;
5610 dst_reg->umax_value = umax_ptr;
5611 dst_reg->var_off = ptr_reg->var_off;
5612 dst_reg->off = ptr_reg->off + smin_val;
5613 dst_reg->raw = ptr_reg->raw;
5616 /* A new variable offset is created. Note that off_reg->off
5617 * == 0, since it's a scalar.
5618 * dst_reg gets the pointer type and since some positive
5619 * integer value was added to the pointer, give it a new 'id'
5620 * if it's a PTR_TO_PACKET.
5621 * this creates a new 'base' pointer, off_reg (variable) gets
5622 * added into the variable offset, and we copy the fixed offset
5625 if (signed_add_overflows(smin_ptr, smin_val) ||
5626 signed_add_overflows(smax_ptr, smax_val)) {
5627 dst_reg->smin_value = S64_MIN;
5628 dst_reg->smax_value = S64_MAX;
5630 dst_reg->smin_value = smin_ptr + smin_val;
5631 dst_reg->smax_value = smax_ptr + smax_val;
5633 if (umin_ptr + umin_val < umin_ptr ||
5634 umax_ptr + umax_val < umax_ptr) {
5635 dst_reg->umin_value = 0;
5636 dst_reg->umax_value = U64_MAX;
5638 dst_reg->umin_value = umin_ptr + umin_val;
5639 dst_reg->umax_value = umax_ptr + umax_val;
5641 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
5642 dst_reg->off = ptr_reg->off;
5643 dst_reg->raw = ptr_reg->raw;
5644 if (reg_is_pkt_pointer(ptr_reg)) {
5645 dst_reg->id = ++env->id_gen;
5646 /* something was added to pkt_ptr, set range to zero */
5647 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
5651 ret = sanitize_ptr_alu(env, insn, ptr_reg, dst_reg, smin_val < 0);
5653 verbose(env, "R%d tried to sub from different maps or paths\n", dst);
5656 if (dst_reg == off_reg) {
5657 /* scalar -= pointer. Creates an unknown scalar */
5658 verbose(env, "R%d tried to subtract pointer from scalar\n",
5662 /* We don't allow subtraction from FP, because (according to
5663 * test_verifier.c test "invalid fp arithmetic", JITs might not
5664 * be able to deal with it.
5666 if (ptr_reg->type == PTR_TO_STACK) {
5667 verbose(env, "R%d subtraction from stack pointer prohibited\n",
5671 if (known && (ptr_reg->off - smin_val ==
5672 (s64)(s32)(ptr_reg->off - smin_val))) {
5673 /* pointer -= K. Subtract it from fixed offset */
5674 dst_reg->smin_value = smin_ptr;
5675 dst_reg->smax_value = smax_ptr;
5676 dst_reg->umin_value = umin_ptr;
5677 dst_reg->umax_value = umax_ptr;
5678 dst_reg->var_off = ptr_reg->var_off;
5679 dst_reg->id = ptr_reg->id;
5680 dst_reg->off = ptr_reg->off - smin_val;
5681 dst_reg->raw = ptr_reg->raw;
5684 /* A new variable offset is created. If the subtrahend is known
5685 * nonnegative, then any reg->range we had before is still good.
5687 if (signed_sub_overflows(smin_ptr, smax_val) ||
5688 signed_sub_overflows(smax_ptr, smin_val)) {
5689 /* Overflow possible, we know nothing */
5690 dst_reg->smin_value = S64_MIN;
5691 dst_reg->smax_value = S64_MAX;
5693 dst_reg->smin_value = smin_ptr - smax_val;
5694 dst_reg->smax_value = smax_ptr - smin_val;
5696 if (umin_ptr < umax_val) {
5697 /* Overflow possible, we know nothing */
5698 dst_reg->umin_value = 0;
5699 dst_reg->umax_value = U64_MAX;
5701 /* Cannot overflow (as long as bounds are consistent) */
5702 dst_reg->umin_value = umin_ptr - umax_val;
5703 dst_reg->umax_value = umax_ptr - umin_val;
5705 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
5706 dst_reg->off = ptr_reg->off;
5707 dst_reg->raw = ptr_reg->raw;
5708 if (reg_is_pkt_pointer(ptr_reg)) {
5709 dst_reg->id = ++env->id_gen;
5710 /* something was added to pkt_ptr, set range to zero */
5712 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
5718 /* bitwise ops on pointers are troublesome, prohibit. */
5719 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
5720 dst, bpf_alu_string[opcode >> 4]);
5723 /* other operators (e.g. MUL,LSH) produce non-pointer results */
5724 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
5725 dst, bpf_alu_string[opcode >> 4]);
5729 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
5732 __update_reg_bounds(dst_reg);
5733 __reg_deduce_bounds(dst_reg);
5734 __reg_bound_offset(dst_reg);
5736 /* For unprivileged we require that resulting offset must be in bounds
5737 * in order to be able to sanitize access later on.
5739 if (!env->bypass_spec_v1) {
5740 if (dst_reg->type == PTR_TO_MAP_VALUE &&
5741 check_map_access(env, dst, dst_reg->off, 1, false)) {
5742 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
5743 "prohibited for !root\n", dst);
5745 } else if (dst_reg->type == PTR_TO_STACK &&
5746 check_stack_access(env, dst_reg, dst_reg->off +
5747 dst_reg->var_off.value, 1)) {
5748 verbose(env, "R%d stack pointer arithmetic goes out of range, "
5749 "prohibited for !root\n", dst);
5757 static void scalar32_min_max_add(struct bpf_reg_state *dst_reg,
5758 struct bpf_reg_state *src_reg)
5760 s32 smin_val = src_reg->s32_min_value;
5761 s32 smax_val = src_reg->s32_max_value;
5762 u32 umin_val = src_reg->u32_min_value;
5763 u32 umax_val = src_reg->u32_max_value;
5765 if (signed_add32_overflows(dst_reg->s32_min_value, smin_val) ||
5766 signed_add32_overflows(dst_reg->s32_max_value, smax_val)) {
5767 dst_reg->s32_min_value = S32_MIN;
5768 dst_reg->s32_max_value = S32_MAX;
5770 dst_reg->s32_min_value += smin_val;
5771 dst_reg->s32_max_value += smax_val;
5773 if (dst_reg->u32_min_value + umin_val < umin_val ||
5774 dst_reg->u32_max_value + umax_val < umax_val) {
5775 dst_reg->u32_min_value = 0;
5776 dst_reg->u32_max_value = U32_MAX;
5778 dst_reg->u32_min_value += umin_val;
5779 dst_reg->u32_max_value += umax_val;
5783 static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
5784 struct bpf_reg_state *src_reg)
5786 s64 smin_val = src_reg->smin_value;
5787 s64 smax_val = src_reg->smax_value;
5788 u64 umin_val = src_reg->umin_value;
5789 u64 umax_val = src_reg->umax_value;
5791 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
5792 signed_add_overflows(dst_reg->smax_value, smax_val)) {
5793 dst_reg->smin_value = S64_MIN;
5794 dst_reg->smax_value = S64_MAX;
5796 dst_reg->smin_value += smin_val;
5797 dst_reg->smax_value += smax_val;
5799 if (dst_reg->umin_value + umin_val < umin_val ||
5800 dst_reg->umax_value + umax_val < umax_val) {
5801 dst_reg->umin_value = 0;
5802 dst_reg->umax_value = U64_MAX;
5804 dst_reg->umin_value += umin_val;
5805 dst_reg->umax_value += umax_val;
5809 static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg,
5810 struct bpf_reg_state *src_reg)
5812 s32 smin_val = src_reg->s32_min_value;
5813 s32 smax_val = src_reg->s32_max_value;
5814 u32 umin_val = src_reg->u32_min_value;
5815 u32 umax_val = src_reg->u32_max_value;
5817 if (signed_sub32_overflows(dst_reg->s32_min_value, smax_val) ||
5818 signed_sub32_overflows(dst_reg->s32_max_value, smin_val)) {
5819 /* Overflow possible, we know nothing */
5820 dst_reg->s32_min_value = S32_MIN;
5821 dst_reg->s32_max_value = S32_MAX;
5823 dst_reg->s32_min_value -= smax_val;
5824 dst_reg->s32_max_value -= smin_val;
5826 if (dst_reg->u32_min_value < umax_val) {
5827 /* Overflow possible, we know nothing */
5828 dst_reg->u32_min_value = 0;
5829 dst_reg->u32_max_value = U32_MAX;
5831 /* Cannot overflow (as long as bounds are consistent) */
5832 dst_reg->u32_min_value -= umax_val;
5833 dst_reg->u32_max_value -= umin_val;
5837 static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
5838 struct bpf_reg_state *src_reg)
5840 s64 smin_val = src_reg->smin_value;
5841 s64 smax_val = src_reg->smax_value;
5842 u64 umin_val = src_reg->umin_value;
5843 u64 umax_val = src_reg->umax_value;
5845 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
5846 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
5847 /* Overflow possible, we know nothing */
5848 dst_reg->smin_value = S64_MIN;
5849 dst_reg->smax_value = S64_MAX;
5851 dst_reg->smin_value -= smax_val;
5852 dst_reg->smax_value -= smin_val;
5854 if (dst_reg->umin_value < umax_val) {
5855 /* Overflow possible, we know nothing */
5856 dst_reg->umin_value = 0;
5857 dst_reg->umax_value = U64_MAX;
5859 /* Cannot overflow (as long as bounds are consistent) */
5860 dst_reg->umin_value -= umax_val;
5861 dst_reg->umax_value -= umin_val;
5865 static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg,
5866 struct bpf_reg_state *src_reg)
5868 s32 smin_val = src_reg->s32_min_value;
5869 u32 umin_val = src_reg->u32_min_value;
5870 u32 umax_val = src_reg->u32_max_value;
5872 if (smin_val < 0 || dst_reg->s32_min_value < 0) {
5873 /* Ain't nobody got time to multiply that sign */
5874 __mark_reg32_unbounded(dst_reg);
5877 /* Both values are positive, so we can work with unsigned and
5878 * copy the result to signed (unless it exceeds S32_MAX).
5880 if (umax_val > U16_MAX || dst_reg->u32_max_value > U16_MAX) {
5881 /* Potential overflow, we know nothing */
5882 __mark_reg32_unbounded(dst_reg);
5885 dst_reg->u32_min_value *= umin_val;
5886 dst_reg->u32_max_value *= umax_val;
5887 if (dst_reg->u32_max_value > S32_MAX) {
5888 /* Overflow possible, we know nothing */
5889 dst_reg->s32_min_value = S32_MIN;
5890 dst_reg->s32_max_value = S32_MAX;
5892 dst_reg->s32_min_value = dst_reg->u32_min_value;
5893 dst_reg->s32_max_value = dst_reg->u32_max_value;
5897 static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
5898 struct bpf_reg_state *src_reg)
5900 s64 smin_val = src_reg->smin_value;
5901 u64 umin_val = src_reg->umin_value;
5902 u64 umax_val = src_reg->umax_value;
5904 if (smin_val < 0 || dst_reg->smin_value < 0) {
5905 /* Ain't nobody got time to multiply that sign */
5906 __mark_reg64_unbounded(dst_reg);
5909 /* Both values are positive, so we can work with unsigned and
5910 * copy the result to signed (unless it exceeds S64_MAX).
5912 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
5913 /* Potential overflow, we know nothing */
5914 __mark_reg64_unbounded(dst_reg);
5917 dst_reg->umin_value *= umin_val;
5918 dst_reg->umax_value *= umax_val;
5919 if (dst_reg->umax_value > S64_MAX) {
5920 /* Overflow possible, we know nothing */
5921 dst_reg->smin_value = S64_MIN;
5922 dst_reg->smax_value = S64_MAX;
5924 dst_reg->smin_value = dst_reg->umin_value;
5925 dst_reg->smax_value = dst_reg->umax_value;
5929 static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
5930 struct bpf_reg_state *src_reg)
5932 bool src_known = tnum_subreg_is_const(src_reg->var_off);
5933 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
5934 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
5935 s32 smin_val = src_reg->s32_min_value;
5936 u32 umax_val = src_reg->u32_max_value;
5938 /* Assuming scalar64_min_max_and will be called so its safe
5939 * to skip updating register for known 32-bit case.
5941 if (src_known && dst_known)
5944 /* We get our minimum from the var_off, since that's inherently
5945 * bitwise. Our maximum is the minimum of the operands' maxima.
5947 dst_reg->u32_min_value = var32_off.value;
5948 dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val);
5949 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
5950 /* Lose signed bounds when ANDing negative numbers,
5951 * ain't nobody got time for that.
5953 dst_reg->s32_min_value = S32_MIN;
5954 dst_reg->s32_max_value = S32_MAX;
5956 /* ANDing two positives gives a positive, so safe to
5957 * cast result into s64.
5959 dst_reg->s32_min_value = dst_reg->u32_min_value;
5960 dst_reg->s32_max_value = dst_reg->u32_max_value;
5965 static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
5966 struct bpf_reg_state *src_reg)
5968 bool src_known = tnum_is_const(src_reg->var_off);
5969 bool dst_known = tnum_is_const(dst_reg->var_off);
5970 s64 smin_val = src_reg->smin_value;
5971 u64 umax_val = src_reg->umax_value;
5973 if (src_known && dst_known) {
5974 __mark_reg_known(dst_reg, dst_reg->var_off.value);
5978 /* We get our minimum from the var_off, since that's inherently
5979 * bitwise. Our maximum is the minimum of the operands' maxima.
5981 dst_reg->umin_value = dst_reg->var_off.value;
5982 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
5983 if (dst_reg->smin_value < 0 || smin_val < 0) {
5984 /* Lose signed bounds when ANDing negative numbers,
5985 * ain't nobody got time for that.
5987 dst_reg->smin_value = S64_MIN;
5988 dst_reg->smax_value = S64_MAX;
5990 /* ANDing two positives gives a positive, so safe to
5991 * cast result into s64.
5993 dst_reg->smin_value = dst_reg->umin_value;
5994 dst_reg->smax_value = dst_reg->umax_value;
5996 /* We may learn something more from the var_off */
5997 __update_reg_bounds(dst_reg);
6000 static void scalar32_min_max_or(struct bpf_reg_state *dst_reg,
6001 struct bpf_reg_state *src_reg)
6003 bool src_known = tnum_subreg_is_const(src_reg->var_off);
6004 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
6005 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
6006 s32 smin_val = src_reg->s32_min_value;
6007 u32 umin_val = src_reg->u32_min_value;
6009 /* Assuming scalar64_min_max_or will be called so it is safe
6010 * to skip updating register for known case.
6012 if (src_known && dst_known)
6015 /* We get our maximum from the var_off, and our minimum is the
6016 * maximum of the operands' minima
6018 dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val);
6019 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
6020 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
6021 /* Lose signed bounds when ORing negative numbers,
6022 * ain't nobody got time for that.
6024 dst_reg->s32_min_value = S32_MIN;
6025 dst_reg->s32_max_value = S32_MAX;
6027 /* ORing two positives gives a positive, so safe to
6028 * cast result into s64.
6030 dst_reg->s32_min_value = dst_reg->u32_min_value;
6031 dst_reg->s32_max_value = dst_reg->u32_max_value;
6035 static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
6036 struct bpf_reg_state *src_reg)
6038 bool src_known = tnum_is_const(src_reg->var_off);
6039 bool dst_known = tnum_is_const(dst_reg->var_off);
6040 s64 smin_val = src_reg->smin_value;
6041 u64 umin_val = src_reg->umin_value;
6043 if (src_known && dst_known) {
6044 __mark_reg_known(dst_reg, dst_reg->var_off.value);
6048 /* We get our maximum from the var_off, and our minimum is the
6049 * maximum of the operands' minima
6051 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
6052 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
6053 if (dst_reg->smin_value < 0 || smin_val < 0) {
6054 /* Lose signed bounds when ORing negative numbers,
6055 * ain't nobody got time for that.
6057 dst_reg->smin_value = S64_MIN;
6058 dst_reg->smax_value = S64_MAX;
6060 /* ORing two positives gives a positive, so safe to
6061 * cast result into s64.
6063 dst_reg->smin_value = dst_reg->umin_value;
6064 dst_reg->smax_value = dst_reg->umax_value;
6066 /* We may learn something more from the var_off */
6067 __update_reg_bounds(dst_reg);
6070 static void scalar32_min_max_xor(struct bpf_reg_state *dst_reg,
6071 struct bpf_reg_state *src_reg)
6073 bool src_known = tnum_subreg_is_const(src_reg->var_off);
6074 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
6075 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
6076 s32 smin_val = src_reg->s32_min_value;
6078 /* Assuming scalar64_min_max_xor will be called so it is safe
6079 * to skip updating register for known case.
6081 if (src_known && dst_known)
6084 /* We get both minimum and maximum from the var32_off. */
6085 dst_reg->u32_min_value = var32_off.value;
6086 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
6088 if (dst_reg->s32_min_value >= 0 && smin_val >= 0) {
6089 /* XORing two positive sign numbers gives a positive,
6090 * so safe to cast u32 result into s32.
6092 dst_reg->s32_min_value = dst_reg->u32_min_value;
6093 dst_reg->s32_max_value = dst_reg->u32_max_value;
6095 dst_reg->s32_min_value = S32_MIN;
6096 dst_reg->s32_max_value = S32_MAX;
6100 static void scalar_min_max_xor(struct bpf_reg_state *dst_reg,
6101 struct bpf_reg_state *src_reg)
6103 bool src_known = tnum_is_const(src_reg->var_off);
6104 bool dst_known = tnum_is_const(dst_reg->var_off);
6105 s64 smin_val = src_reg->smin_value;
6107 if (src_known && dst_known) {
6108 /* dst_reg->var_off.value has been updated earlier */
6109 __mark_reg_known(dst_reg, dst_reg->var_off.value);
6113 /* We get both minimum and maximum from the var_off. */
6114 dst_reg->umin_value = dst_reg->var_off.value;
6115 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
6117 if (dst_reg->smin_value >= 0 && smin_val >= 0) {
6118 /* XORing two positive sign numbers gives a positive,
6119 * so safe to cast u64 result into s64.
6121 dst_reg->smin_value = dst_reg->umin_value;
6122 dst_reg->smax_value = dst_reg->umax_value;
6124 dst_reg->smin_value = S64_MIN;
6125 dst_reg->smax_value = S64_MAX;
6128 __update_reg_bounds(dst_reg);
6131 static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
6132 u64 umin_val, u64 umax_val)
6134 /* We lose all sign bit information (except what we can pick
6137 dst_reg->s32_min_value = S32_MIN;
6138 dst_reg->s32_max_value = S32_MAX;
6139 /* If we might shift our top bit out, then we know nothing */
6140 if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) {
6141 dst_reg->u32_min_value = 0;
6142 dst_reg->u32_max_value = U32_MAX;
6144 dst_reg->u32_min_value <<= umin_val;
6145 dst_reg->u32_max_value <<= umax_val;
6149 static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
6150 struct bpf_reg_state *src_reg)
6152 u32 umax_val = src_reg->u32_max_value;
6153 u32 umin_val = src_reg->u32_min_value;
6154 /* u32 alu operation will zext upper bits */
6155 struct tnum subreg = tnum_subreg(dst_reg->var_off);
6157 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
6158 dst_reg->var_off = tnum_subreg(tnum_lshift(subreg, umin_val));
6159 /* Not required but being careful mark reg64 bounds as unknown so
6160 * that we are forced to pick them up from tnum and zext later and
6161 * if some path skips this step we are still safe.
6163 __mark_reg64_unbounded(dst_reg);
6164 __update_reg32_bounds(dst_reg);
6167 static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg,
6168 u64 umin_val, u64 umax_val)
6170 /* Special case <<32 because it is a common compiler pattern to sign
6171 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
6172 * positive we know this shift will also be positive so we can track
6173 * bounds correctly. Otherwise we lose all sign bit information except
6174 * what we can pick up from var_off. Perhaps we can generalize this
6175 * later to shifts of any length.
6177 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0)
6178 dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32;
6180 dst_reg->smax_value = S64_MAX;
6182 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0)
6183 dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32;
6185 dst_reg->smin_value = S64_MIN;
6187 /* If we might shift our top bit out, then we know nothing */
6188 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
6189 dst_reg->umin_value = 0;
6190 dst_reg->umax_value = U64_MAX;
6192 dst_reg->umin_value <<= umin_val;
6193 dst_reg->umax_value <<= umax_val;
6197 static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
6198 struct bpf_reg_state *src_reg)
6200 u64 umax_val = src_reg->umax_value;
6201 u64 umin_val = src_reg->umin_value;
6203 /* scalar64 calc uses 32bit unshifted bounds so must be called first */
6204 __scalar64_min_max_lsh(dst_reg, umin_val, umax_val);
6205 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
6207 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
6208 /* We may learn something more from the var_off */
6209 __update_reg_bounds(dst_reg);
6212 static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg,
6213 struct bpf_reg_state *src_reg)
6215 struct tnum subreg = tnum_subreg(dst_reg->var_off);
6216 u32 umax_val = src_reg->u32_max_value;
6217 u32 umin_val = src_reg->u32_min_value;
6219 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
6220 * be negative, then either:
6221 * 1) src_reg might be zero, so the sign bit of the result is
6222 * unknown, so we lose our signed bounds
6223 * 2) it's known negative, thus the unsigned bounds capture the
6225 * 3) the signed bounds cross zero, so they tell us nothing
6227 * If the value in dst_reg is known nonnegative, then again the
6228 * unsigned bounts capture the signed bounds.
6229 * Thus, in all cases it suffices to blow away our signed bounds
6230 * and rely on inferring new ones from the unsigned bounds and
6231 * var_off of the result.
6233 dst_reg->s32_min_value = S32_MIN;
6234 dst_reg->s32_max_value = S32_MAX;
6236 dst_reg->var_off = tnum_rshift(subreg, umin_val);
6237 dst_reg->u32_min_value >>= umax_val;
6238 dst_reg->u32_max_value >>= umin_val;
6240 __mark_reg64_unbounded(dst_reg);
6241 __update_reg32_bounds(dst_reg);
6244 static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
6245 struct bpf_reg_state *src_reg)
6247 u64 umax_val = src_reg->umax_value;
6248 u64 umin_val = src_reg->umin_value;
6250 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
6251 * be negative, then either:
6252 * 1) src_reg might be zero, so the sign bit of the result is
6253 * unknown, so we lose our signed bounds
6254 * 2) it's known negative, thus the unsigned bounds capture the
6256 * 3) the signed bounds cross zero, so they tell us nothing
6258 * If the value in dst_reg is known nonnegative, then again the
6259 * unsigned bounts capture the signed bounds.
6260 * Thus, in all cases it suffices to blow away our signed bounds
6261 * and rely on inferring new ones from the unsigned bounds and
6262 * var_off of the result.
6264 dst_reg->smin_value = S64_MIN;
6265 dst_reg->smax_value = S64_MAX;
6266 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
6267 dst_reg->umin_value >>= umax_val;
6268 dst_reg->umax_value >>= umin_val;
6270 /* Its not easy to operate on alu32 bounds here because it depends
6271 * on bits being shifted in. Take easy way out and mark unbounded
6272 * so we can recalculate later from tnum.
6274 __mark_reg32_unbounded(dst_reg);
6275 __update_reg_bounds(dst_reg);
6278 static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg,
6279 struct bpf_reg_state *src_reg)
6281 u64 umin_val = src_reg->u32_min_value;
6283 /* Upon reaching here, src_known is true and
6284 * umax_val is equal to umin_val.
6286 dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val);
6287 dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val);
6289 dst_reg->var_off = tnum_arshift(tnum_subreg(dst_reg->var_off), umin_val, 32);
6291 /* blow away the dst_reg umin_value/umax_value and rely on
6292 * dst_reg var_off to refine the result.
6294 dst_reg->u32_min_value = 0;
6295 dst_reg->u32_max_value = U32_MAX;
6297 __mark_reg64_unbounded(dst_reg);
6298 __update_reg32_bounds(dst_reg);
6301 static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
6302 struct bpf_reg_state *src_reg)
6304 u64 umin_val = src_reg->umin_value;
6306 /* Upon reaching here, src_known is true and umax_val is equal
6309 dst_reg->smin_value >>= umin_val;
6310 dst_reg->smax_value >>= umin_val;
6312 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val, 64);
6314 /* blow away the dst_reg umin_value/umax_value and rely on
6315 * dst_reg var_off to refine the result.
6317 dst_reg->umin_value = 0;
6318 dst_reg->umax_value = U64_MAX;
6320 /* Its not easy to operate on alu32 bounds here because it depends
6321 * on bits being shifted in from upper 32-bits. Take easy way out
6322 * and mark unbounded so we can recalculate later from tnum.
6324 __mark_reg32_unbounded(dst_reg);
6325 __update_reg_bounds(dst_reg);
6328 /* WARNING: This function does calculations on 64-bit values, but the actual
6329 * execution may occur on 32-bit values. Therefore, things like bitshifts
6330 * need extra checks in the 32-bit case.
6332 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
6333 struct bpf_insn *insn,
6334 struct bpf_reg_state *dst_reg,
6335 struct bpf_reg_state src_reg)
6337 struct bpf_reg_state *regs = cur_regs(env);
6338 u8 opcode = BPF_OP(insn->code);
6340 s64 smin_val, smax_val;
6341 u64 umin_val, umax_val;
6342 s32 s32_min_val, s32_max_val;
6343 u32 u32_min_val, u32_max_val;
6344 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
6345 u32 dst = insn->dst_reg;
6347 bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
6349 smin_val = src_reg.smin_value;
6350 smax_val = src_reg.smax_value;
6351 umin_val = src_reg.umin_value;
6352 umax_val = src_reg.umax_value;
6354 s32_min_val = src_reg.s32_min_value;
6355 s32_max_val = src_reg.s32_max_value;
6356 u32_min_val = src_reg.u32_min_value;
6357 u32_max_val = src_reg.u32_max_value;
6360 src_known = tnum_subreg_is_const(src_reg.var_off);
6362 (s32_min_val != s32_max_val || u32_min_val != u32_max_val)) ||
6363 s32_min_val > s32_max_val || u32_min_val > u32_max_val) {
6364 /* Taint dst register if offset had invalid bounds
6365 * derived from e.g. dead branches.
6367 __mark_reg_unknown(env, dst_reg);
6371 src_known = tnum_is_const(src_reg.var_off);
6373 (smin_val != smax_val || umin_val != umax_val)) ||
6374 smin_val > smax_val || umin_val > umax_val) {
6375 /* Taint dst register if offset had invalid bounds
6376 * derived from e.g. dead branches.
6378 __mark_reg_unknown(env, dst_reg);
6384 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
6385 __mark_reg_unknown(env, dst_reg);
6389 /* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
6390 * There are two classes of instructions: The first class we track both
6391 * alu32 and alu64 sign/unsigned bounds independently this provides the
6392 * greatest amount of precision when alu operations are mixed with jmp32
6393 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
6394 * and BPF_OR. This is possible because these ops have fairly easy to
6395 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
6396 * See alu32 verifier tests for examples. The second class of
6397 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
6398 * with regards to tracking sign/unsigned bounds because the bits may
6399 * cross subreg boundaries in the alu64 case. When this happens we mark
6400 * the reg unbounded in the subreg bound space and use the resulting
6401 * tnum to calculate an approximation of the sign/unsigned bounds.
6405 ret = sanitize_val_alu(env, insn);
6407 verbose(env, "R%d tried to add from different pointers or scalars\n", dst);
6410 scalar32_min_max_add(dst_reg, &src_reg);
6411 scalar_min_max_add(dst_reg, &src_reg);
6412 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
6415 ret = sanitize_val_alu(env, insn);
6417 verbose(env, "R%d tried to sub from different pointers or scalars\n", dst);
6420 scalar32_min_max_sub(dst_reg, &src_reg);
6421 scalar_min_max_sub(dst_reg, &src_reg);
6422 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
6425 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
6426 scalar32_min_max_mul(dst_reg, &src_reg);
6427 scalar_min_max_mul(dst_reg, &src_reg);
6430 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
6431 scalar32_min_max_and(dst_reg, &src_reg);
6432 scalar_min_max_and(dst_reg, &src_reg);
6435 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
6436 scalar32_min_max_or(dst_reg, &src_reg);
6437 scalar_min_max_or(dst_reg, &src_reg);
6440 dst_reg->var_off = tnum_xor(dst_reg->var_off, src_reg.var_off);
6441 scalar32_min_max_xor(dst_reg, &src_reg);
6442 scalar_min_max_xor(dst_reg, &src_reg);
6445 if (umax_val >= insn_bitness) {
6446 /* Shifts greater than 31 or 63 are undefined.
6447 * This includes shifts by a negative number.
6449 mark_reg_unknown(env, regs, insn->dst_reg);
6453 scalar32_min_max_lsh(dst_reg, &src_reg);
6455 scalar_min_max_lsh(dst_reg, &src_reg);
6458 if (umax_val >= insn_bitness) {
6459 /* Shifts greater than 31 or 63 are undefined.
6460 * This includes shifts by a negative number.
6462 mark_reg_unknown(env, regs, insn->dst_reg);
6466 scalar32_min_max_rsh(dst_reg, &src_reg);
6468 scalar_min_max_rsh(dst_reg, &src_reg);
6471 if (umax_val >= insn_bitness) {
6472 /* Shifts greater than 31 or 63 are undefined.
6473 * This includes shifts by a negative number.
6475 mark_reg_unknown(env, regs, insn->dst_reg);
6479 scalar32_min_max_arsh(dst_reg, &src_reg);
6481 scalar_min_max_arsh(dst_reg, &src_reg);
6484 mark_reg_unknown(env, regs, insn->dst_reg);
6488 /* ALU32 ops are zero extended into 64bit register */
6490 zext_32_to_64(dst_reg);
6492 __update_reg_bounds(dst_reg);
6493 __reg_deduce_bounds(dst_reg);
6494 __reg_bound_offset(dst_reg);
6498 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
6501 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
6502 struct bpf_insn *insn)
6504 struct bpf_verifier_state *vstate = env->cur_state;
6505 struct bpf_func_state *state = vstate->frame[vstate->curframe];
6506 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
6507 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
6508 u8 opcode = BPF_OP(insn->code);
6511 dst_reg = ®s[insn->dst_reg];
6513 if (dst_reg->type != SCALAR_VALUE)
6516 /* Make sure ID is cleared otherwise dst_reg min/max could be
6517 * incorrectly propagated into other registers by find_equal_scalars()
6520 if (BPF_SRC(insn->code) == BPF_X) {
6521 src_reg = ®s[insn->src_reg];
6522 if (src_reg->type != SCALAR_VALUE) {
6523 if (dst_reg->type != SCALAR_VALUE) {
6524 /* Combining two pointers by any ALU op yields
6525 * an arbitrary scalar. Disallow all math except
6526 * pointer subtraction
6528 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
6529 mark_reg_unknown(env, regs, insn->dst_reg);
6532 verbose(env, "R%d pointer %s pointer prohibited\n",
6534 bpf_alu_string[opcode >> 4]);
6537 /* scalar += pointer
6538 * This is legal, but we have to reverse our
6539 * src/dest handling in computing the range
6541 err = mark_chain_precision(env, insn->dst_reg);
6544 return adjust_ptr_min_max_vals(env, insn,
6547 } else if (ptr_reg) {
6548 /* pointer += scalar */
6549 err = mark_chain_precision(env, insn->src_reg);
6552 return adjust_ptr_min_max_vals(env, insn,
6556 /* Pretend the src is a reg with a known value, since we only
6557 * need to be able to read from this state.
6559 off_reg.type = SCALAR_VALUE;
6560 __mark_reg_known(&off_reg, insn->imm);
6562 if (ptr_reg) /* pointer += K */
6563 return adjust_ptr_min_max_vals(env, insn,
6567 /* Got here implies adding two SCALAR_VALUEs */
6568 if (WARN_ON_ONCE(ptr_reg)) {
6569 print_verifier_state(env, state);
6570 verbose(env, "verifier internal error: unexpected ptr_reg\n");
6573 if (WARN_ON(!src_reg)) {
6574 print_verifier_state(env, state);
6575 verbose(env, "verifier internal error: no src_reg\n");
6578 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
6581 /* check validity of 32-bit and 64-bit arithmetic operations */
6582 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
6584 struct bpf_reg_state *regs = cur_regs(env);
6585 u8 opcode = BPF_OP(insn->code);
6588 if (opcode == BPF_END || opcode == BPF_NEG) {
6589 if (opcode == BPF_NEG) {
6590 if (BPF_SRC(insn->code) != 0 ||
6591 insn->src_reg != BPF_REG_0 ||
6592 insn->off != 0 || insn->imm != 0) {
6593 verbose(env, "BPF_NEG uses reserved fields\n");
6597 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
6598 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
6599 BPF_CLASS(insn->code) == BPF_ALU64) {
6600 verbose(env, "BPF_END uses reserved fields\n");
6605 /* check src operand */
6606 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
6610 if (is_pointer_value(env, insn->dst_reg)) {
6611 verbose(env, "R%d pointer arithmetic prohibited\n",
6616 /* check dest operand */
6617 err = check_reg_arg(env, insn->dst_reg, DST_OP);
6621 } else if (opcode == BPF_MOV) {
6623 if (BPF_SRC(insn->code) == BPF_X) {
6624 if (insn->imm != 0 || insn->off != 0) {
6625 verbose(env, "BPF_MOV uses reserved fields\n");
6629 /* check src operand */
6630 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6634 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
6635 verbose(env, "BPF_MOV uses reserved fields\n");
6640 /* check dest operand, mark as required later */
6641 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
6645 if (BPF_SRC(insn->code) == BPF_X) {
6646 struct bpf_reg_state *src_reg = regs + insn->src_reg;
6647 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
6649 if (BPF_CLASS(insn->code) == BPF_ALU64) {
6651 * copy register state to dest reg
6653 if (src_reg->type == SCALAR_VALUE && !src_reg->id)
6654 /* Assign src and dst registers the same ID
6655 * that will be used by find_equal_scalars()
6656 * to propagate min/max range.
6658 src_reg->id = ++env->id_gen;
6659 *dst_reg = *src_reg;
6660 dst_reg->live |= REG_LIVE_WRITTEN;
6661 dst_reg->subreg_def = DEF_NOT_SUBREG;
6664 if (is_pointer_value(env, insn->src_reg)) {
6666 "R%d partial copy of pointer\n",
6669 } else if (src_reg->type == SCALAR_VALUE) {
6670 *dst_reg = *src_reg;
6671 /* Make sure ID is cleared otherwise
6672 * dst_reg min/max could be incorrectly
6673 * propagated into src_reg by find_equal_scalars()
6676 dst_reg->live |= REG_LIVE_WRITTEN;
6677 dst_reg->subreg_def = env->insn_idx + 1;
6679 mark_reg_unknown(env, regs,
6682 zext_32_to_64(dst_reg);
6686 * remember the value we stored into this reg
6688 /* clear any state __mark_reg_known doesn't set */
6689 mark_reg_unknown(env, regs, insn->dst_reg);
6690 regs[insn->dst_reg].type = SCALAR_VALUE;
6691 if (BPF_CLASS(insn->code) == BPF_ALU64) {
6692 __mark_reg_known(regs + insn->dst_reg,
6695 __mark_reg_known(regs + insn->dst_reg,
6700 } else if (opcode > BPF_END) {
6701 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
6704 } else { /* all other ALU ops: and, sub, xor, add, ... */
6706 if (BPF_SRC(insn->code) == BPF_X) {
6707 if (insn->imm != 0 || insn->off != 0) {
6708 verbose(env, "BPF_ALU uses reserved fields\n");
6711 /* check src1 operand */
6712 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6716 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
6717 verbose(env, "BPF_ALU uses reserved fields\n");
6722 /* check src2 operand */
6723 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
6727 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
6728 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
6729 verbose(env, "div by zero\n");
6733 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
6734 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
6735 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
6737 if (insn->imm < 0 || insn->imm >= size) {
6738 verbose(env, "invalid shift %d\n", insn->imm);
6743 /* check dest operand */
6744 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
6748 return adjust_reg_min_max_vals(env, insn);
6754 static void __find_good_pkt_pointers(struct bpf_func_state *state,
6755 struct bpf_reg_state *dst_reg,
6756 enum bpf_reg_type type, int new_range)
6758 struct bpf_reg_state *reg;
6761 for (i = 0; i < MAX_BPF_REG; i++) {
6762 reg = &state->regs[i];
6763 if (reg->type == type && reg->id == dst_reg->id)
6764 /* keep the maximum range already checked */
6765 reg->range = max(reg->range, new_range);
6768 bpf_for_each_spilled_reg(i, state, reg) {
6771 if (reg->type == type && reg->id == dst_reg->id)
6772 reg->range = max(reg->range, new_range);
6776 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
6777 struct bpf_reg_state *dst_reg,
6778 enum bpf_reg_type type,
6779 bool range_right_open)
6783 if (dst_reg->off < 0 ||
6784 (dst_reg->off == 0 && range_right_open))
6785 /* This doesn't give us any range */
6788 if (dst_reg->umax_value > MAX_PACKET_OFF ||
6789 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
6790 /* Risk of overflow. For instance, ptr + (1<<63) may be less
6791 * than pkt_end, but that's because it's also less than pkt.
6795 new_range = dst_reg->off;
6796 if (range_right_open)
6799 /* Examples for register markings:
6801 * pkt_data in dst register:
6805 * if (r2 > pkt_end) goto <handle exception>
6810 * if (r2 < pkt_end) goto <access okay>
6811 * <handle exception>
6814 * r2 == dst_reg, pkt_end == src_reg
6815 * r2=pkt(id=n,off=8,r=0)
6816 * r3=pkt(id=n,off=0,r=0)
6818 * pkt_data in src register:
6822 * if (pkt_end >= r2) goto <access okay>
6823 * <handle exception>
6827 * if (pkt_end <= r2) goto <handle exception>
6831 * pkt_end == dst_reg, r2 == src_reg
6832 * r2=pkt(id=n,off=8,r=0)
6833 * r3=pkt(id=n,off=0,r=0)
6835 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
6836 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
6837 * and [r3, r3 + 8-1) respectively is safe to access depending on
6841 /* If our ids match, then we must have the same max_value. And we
6842 * don't care about the other reg's fixed offset, since if it's too big
6843 * the range won't allow anything.
6844 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
6846 for (i = 0; i <= vstate->curframe; i++)
6847 __find_good_pkt_pointers(vstate->frame[i], dst_reg, type,
6851 static int is_branch32_taken(struct bpf_reg_state *reg, u32 val, u8 opcode)
6853 struct tnum subreg = tnum_subreg(reg->var_off);
6854 s32 sval = (s32)val;
6858 if (tnum_is_const(subreg))
6859 return !!tnum_equals_const(subreg, val);
6862 if (tnum_is_const(subreg))
6863 return !tnum_equals_const(subreg, val);
6866 if ((~subreg.mask & subreg.value) & val)
6868 if (!((subreg.mask | subreg.value) & val))
6872 if (reg->u32_min_value > val)
6874 else if (reg->u32_max_value <= val)
6878 if (reg->s32_min_value > sval)
6880 else if (reg->s32_max_value < sval)
6884 if (reg->u32_max_value < val)
6886 else if (reg->u32_min_value >= val)
6890 if (reg->s32_max_value < sval)
6892 else if (reg->s32_min_value >= sval)
6896 if (reg->u32_min_value >= val)
6898 else if (reg->u32_max_value < val)
6902 if (reg->s32_min_value >= sval)
6904 else if (reg->s32_max_value < sval)
6908 if (reg->u32_max_value <= val)
6910 else if (reg->u32_min_value > val)
6914 if (reg->s32_max_value <= sval)
6916 else if (reg->s32_min_value > sval)
6925 static int is_branch64_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
6927 s64 sval = (s64)val;
6931 if (tnum_is_const(reg->var_off))
6932 return !!tnum_equals_const(reg->var_off, val);
6935 if (tnum_is_const(reg->var_off))
6936 return !tnum_equals_const(reg->var_off, val);
6939 if ((~reg->var_off.mask & reg->var_off.value) & val)
6941 if (!((reg->var_off.mask | reg->var_off.value) & val))
6945 if (reg->umin_value > val)
6947 else if (reg->umax_value <= val)
6951 if (reg->smin_value > sval)
6953 else if (reg->smax_value < sval)
6957 if (reg->umax_value < val)
6959 else if (reg->umin_value >= val)
6963 if (reg->smax_value < sval)
6965 else if (reg->smin_value >= sval)
6969 if (reg->umin_value >= val)
6971 else if (reg->umax_value < val)
6975 if (reg->smin_value >= sval)
6977 else if (reg->smax_value < sval)
6981 if (reg->umax_value <= val)
6983 else if (reg->umin_value > val)
6987 if (reg->smax_value <= sval)
6989 else if (reg->smin_value > sval)
6997 /* compute branch direction of the expression "if (reg opcode val) goto target;"
6999 * 1 - branch will be taken and "goto target" will be executed
7000 * 0 - branch will not be taken and fall-through to next insn
7001 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value
7004 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
7007 if (__is_pointer_value(false, reg)) {
7008 if (!reg_type_not_null(reg->type))
7011 /* If pointer is valid tests against zero will fail so we can
7012 * use this to direct branch taken.
7028 return is_branch32_taken(reg, val, opcode);
7029 return is_branch64_taken(reg, val, opcode);
7032 static int flip_opcode(u32 opcode)
7034 /* How can we transform "a <op> b" into "b <op> a"? */
7035 static const u8 opcode_flip[16] = {
7036 /* these stay the same */
7037 [BPF_JEQ >> 4] = BPF_JEQ,
7038 [BPF_JNE >> 4] = BPF_JNE,
7039 [BPF_JSET >> 4] = BPF_JSET,
7040 /* these swap "lesser" and "greater" (L and G in the opcodes) */
7041 [BPF_JGE >> 4] = BPF_JLE,
7042 [BPF_JGT >> 4] = BPF_JLT,
7043 [BPF_JLE >> 4] = BPF_JGE,
7044 [BPF_JLT >> 4] = BPF_JGT,
7045 [BPF_JSGE >> 4] = BPF_JSLE,
7046 [BPF_JSGT >> 4] = BPF_JSLT,
7047 [BPF_JSLE >> 4] = BPF_JSGE,
7048 [BPF_JSLT >> 4] = BPF_JSGT
7050 return opcode_flip[opcode >> 4];
7053 static int is_pkt_ptr_branch_taken(struct bpf_reg_state *dst_reg,
7054 struct bpf_reg_state *src_reg,
7057 struct bpf_reg_state *pkt;
7059 if (src_reg->type == PTR_TO_PACKET_END) {
7061 } else if (dst_reg->type == PTR_TO_PACKET_END) {
7063 opcode = flip_opcode(opcode);
7068 if (pkt->range >= 0)
7073 /* pkt <= pkt_end */
7077 if (pkt->range == BEYOND_PKT_END)
7078 /* pkt has at last one extra byte beyond pkt_end */
7079 return opcode == BPF_JGT;
7085 /* pkt >= pkt_end */
7086 if (pkt->range == BEYOND_PKT_END || pkt->range == AT_PKT_END)
7087 return opcode == BPF_JGE;
7093 /* Adjusts the register min/max values in the case that the dst_reg is the
7094 * variable register that we are working on, and src_reg is a constant or we're
7095 * simply doing a BPF_K check.
7096 * In JEQ/JNE cases we also adjust the var_off values.
7098 static void reg_set_min_max(struct bpf_reg_state *true_reg,
7099 struct bpf_reg_state *false_reg,
7101 u8 opcode, bool is_jmp32)
7103 struct tnum false_32off = tnum_subreg(false_reg->var_off);
7104 struct tnum false_64off = false_reg->var_off;
7105 struct tnum true_32off = tnum_subreg(true_reg->var_off);
7106 struct tnum true_64off = true_reg->var_off;
7107 s64 sval = (s64)val;
7108 s32 sval32 = (s32)val32;
7110 /* If the dst_reg is a pointer, we can't learn anything about its
7111 * variable offset from the compare (unless src_reg were a pointer into
7112 * the same object, but we don't bother with that.
7113 * Since false_reg and true_reg have the same type by construction, we
7114 * only need to check one of them for pointerness.
7116 if (__is_pointer_value(false, false_reg))
7123 struct bpf_reg_state *reg =
7124 opcode == BPF_JEQ ? true_reg : false_reg;
7126 /* JEQ/JNE comparison doesn't change the register equivalence.
7128 * if (r1 == 42) goto label;
7130 * label: // here both r1 and r2 are known to be 42.
7132 * Hence when marking register as known preserve it's ID.
7135 __mark_reg32_known(reg, val32);
7137 ___mark_reg_known(reg, val);
7142 false_32off = tnum_and(false_32off, tnum_const(~val32));
7143 if (is_power_of_2(val32))
7144 true_32off = tnum_or(true_32off,
7147 false_64off = tnum_and(false_64off, tnum_const(~val));
7148 if (is_power_of_2(val))
7149 true_64off = tnum_or(true_64off,
7157 u32 false_umax = opcode == BPF_JGT ? val32 : val32 - 1;
7158 u32 true_umin = opcode == BPF_JGT ? val32 + 1 : val32;
7160 false_reg->u32_max_value = min(false_reg->u32_max_value,
7162 true_reg->u32_min_value = max(true_reg->u32_min_value,
7165 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
7166 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
7168 false_reg->umax_value = min(false_reg->umax_value, false_umax);
7169 true_reg->umin_value = max(true_reg->umin_value, true_umin);
7177 s32 false_smax = opcode == BPF_JSGT ? sval32 : sval32 - 1;
7178 s32 true_smin = opcode == BPF_JSGT ? sval32 + 1 : sval32;
7180 false_reg->s32_max_value = min(false_reg->s32_max_value, false_smax);
7181 true_reg->s32_min_value = max(true_reg->s32_min_value, true_smin);
7183 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
7184 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
7186 false_reg->smax_value = min(false_reg->smax_value, false_smax);
7187 true_reg->smin_value = max(true_reg->smin_value, true_smin);
7195 u32 false_umin = opcode == BPF_JLT ? val32 : val32 + 1;
7196 u32 true_umax = opcode == BPF_JLT ? val32 - 1 : val32;
7198 false_reg->u32_min_value = max(false_reg->u32_min_value,
7200 true_reg->u32_max_value = min(true_reg->u32_max_value,
7203 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
7204 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
7206 false_reg->umin_value = max(false_reg->umin_value, false_umin);
7207 true_reg->umax_value = min(true_reg->umax_value, true_umax);
7215 s32 false_smin = opcode == BPF_JSLT ? sval32 : sval32 + 1;
7216 s32 true_smax = opcode == BPF_JSLT ? sval32 - 1 : sval32;
7218 false_reg->s32_min_value = max(false_reg->s32_min_value, false_smin);
7219 true_reg->s32_max_value = min(true_reg->s32_max_value, true_smax);
7221 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
7222 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
7224 false_reg->smin_value = max(false_reg->smin_value, false_smin);
7225 true_reg->smax_value = min(true_reg->smax_value, true_smax);
7234 false_reg->var_off = tnum_or(tnum_clear_subreg(false_64off),
7235 tnum_subreg(false_32off));
7236 true_reg->var_off = tnum_or(tnum_clear_subreg(true_64off),
7237 tnum_subreg(true_32off));
7238 __reg_combine_32_into_64(false_reg);
7239 __reg_combine_32_into_64(true_reg);
7241 false_reg->var_off = false_64off;
7242 true_reg->var_off = true_64off;
7243 __reg_combine_64_into_32(false_reg);
7244 __reg_combine_64_into_32(true_reg);
7248 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
7251 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
7252 struct bpf_reg_state *false_reg,
7254 u8 opcode, bool is_jmp32)
7256 opcode = flip_opcode(opcode);
7257 /* This uses zero as "not present in table"; luckily the zero opcode,
7258 * BPF_JA, can't get here.
7261 reg_set_min_max(true_reg, false_reg, val, val32, opcode, is_jmp32);
7264 /* Regs are known to be equal, so intersect their min/max/var_off */
7265 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
7266 struct bpf_reg_state *dst_reg)
7268 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
7269 dst_reg->umin_value);
7270 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
7271 dst_reg->umax_value);
7272 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
7273 dst_reg->smin_value);
7274 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
7275 dst_reg->smax_value);
7276 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
7278 /* We might have learned new bounds from the var_off. */
7279 __update_reg_bounds(src_reg);
7280 __update_reg_bounds(dst_reg);
7281 /* We might have learned something about the sign bit. */
7282 __reg_deduce_bounds(src_reg);
7283 __reg_deduce_bounds(dst_reg);
7284 /* We might have learned some bits from the bounds. */
7285 __reg_bound_offset(src_reg);
7286 __reg_bound_offset(dst_reg);
7287 /* Intersecting with the old var_off might have improved our bounds
7288 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
7289 * then new var_off is (0; 0x7f...fc) which improves our umax.
7291 __update_reg_bounds(src_reg);
7292 __update_reg_bounds(dst_reg);
7295 static void reg_combine_min_max(struct bpf_reg_state *true_src,
7296 struct bpf_reg_state *true_dst,
7297 struct bpf_reg_state *false_src,
7298 struct bpf_reg_state *false_dst,
7303 __reg_combine_min_max(true_src, true_dst);
7306 __reg_combine_min_max(false_src, false_dst);
7311 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
7312 struct bpf_reg_state *reg, u32 id,
7315 if (reg_type_may_be_null(reg->type) && reg->id == id &&
7316 !WARN_ON_ONCE(!reg->id)) {
7317 /* Old offset (both fixed and variable parts) should
7318 * have been known-zero, because we don't allow pointer
7319 * arithmetic on pointers that might be NULL.
7321 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
7322 !tnum_equals_const(reg->var_off, 0) ||
7324 __mark_reg_known_zero(reg);
7328 reg->type = SCALAR_VALUE;
7329 } else if (reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
7330 const struct bpf_map *map = reg->map_ptr;
7332 if (map->inner_map_meta) {
7333 reg->type = CONST_PTR_TO_MAP;
7334 reg->map_ptr = map->inner_map_meta;
7335 } else if (map->map_type == BPF_MAP_TYPE_XSKMAP) {
7336 reg->type = PTR_TO_XDP_SOCK;
7337 } else if (map->map_type == BPF_MAP_TYPE_SOCKMAP ||
7338 map->map_type == BPF_MAP_TYPE_SOCKHASH) {
7339 reg->type = PTR_TO_SOCKET;
7341 reg->type = PTR_TO_MAP_VALUE;
7343 } else if (reg->type == PTR_TO_SOCKET_OR_NULL) {
7344 reg->type = PTR_TO_SOCKET;
7345 } else if (reg->type == PTR_TO_SOCK_COMMON_OR_NULL) {
7346 reg->type = PTR_TO_SOCK_COMMON;
7347 } else if (reg->type == PTR_TO_TCP_SOCK_OR_NULL) {
7348 reg->type = PTR_TO_TCP_SOCK;
7349 } else if (reg->type == PTR_TO_BTF_ID_OR_NULL) {
7350 reg->type = PTR_TO_BTF_ID;
7351 } else if (reg->type == PTR_TO_MEM_OR_NULL) {
7352 reg->type = PTR_TO_MEM;
7353 } else if (reg->type == PTR_TO_RDONLY_BUF_OR_NULL) {
7354 reg->type = PTR_TO_RDONLY_BUF;
7355 } else if (reg->type == PTR_TO_RDWR_BUF_OR_NULL) {
7356 reg->type = PTR_TO_RDWR_BUF;
7359 /* We don't need id and ref_obj_id from this point
7360 * onwards anymore, thus we should better reset it,
7361 * so that state pruning has chances to take effect.
7364 reg->ref_obj_id = 0;
7365 } else if (!reg_may_point_to_spin_lock(reg)) {
7366 /* For not-NULL ptr, reg->ref_obj_id will be reset
7367 * in release_reg_references().
7369 * reg->id is still used by spin_lock ptr. Other
7370 * than spin_lock ptr type, reg->id can be reset.
7377 static void __mark_ptr_or_null_regs(struct bpf_func_state *state, u32 id,
7380 struct bpf_reg_state *reg;
7383 for (i = 0; i < MAX_BPF_REG; i++)
7384 mark_ptr_or_null_reg(state, &state->regs[i], id, is_null);
7386 bpf_for_each_spilled_reg(i, state, reg) {
7389 mark_ptr_or_null_reg(state, reg, id, is_null);
7393 /* The logic is similar to find_good_pkt_pointers(), both could eventually
7394 * be folded together at some point.
7396 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
7399 struct bpf_func_state *state = vstate->frame[vstate->curframe];
7400 struct bpf_reg_state *regs = state->regs;
7401 u32 ref_obj_id = regs[regno].ref_obj_id;
7402 u32 id = regs[regno].id;
7405 if (ref_obj_id && ref_obj_id == id && is_null)
7406 /* regs[regno] is in the " == NULL" branch.
7407 * No one could have freed the reference state before
7408 * doing the NULL check.
7410 WARN_ON_ONCE(release_reference_state(state, id));
7412 for (i = 0; i <= vstate->curframe; i++)
7413 __mark_ptr_or_null_regs(vstate->frame[i], id, is_null);
7416 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
7417 struct bpf_reg_state *dst_reg,
7418 struct bpf_reg_state *src_reg,
7419 struct bpf_verifier_state *this_branch,
7420 struct bpf_verifier_state *other_branch)
7422 if (BPF_SRC(insn->code) != BPF_X)
7425 /* Pointers are always 64-bit. */
7426 if (BPF_CLASS(insn->code) == BPF_JMP32)
7429 switch (BPF_OP(insn->code)) {
7431 if ((dst_reg->type == PTR_TO_PACKET &&
7432 src_reg->type == PTR_TO_PACKET_END) ||
7433 (dst_reg->type == PTR_TO_PACKET_META &&
7434 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
7435 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
7436 find_good_pkt_pointers(this_branch, dst_reg,
7437 dst_reg->type, false);
7438 mark_pkt_end(other_branch, insn->dst_reg, true);
7439 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
7440 src_reg->type == PTR_TO_PACKET) ||
7441 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
7442 src_reg->type == PTR_TO_PACKET_META)) {
7443 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
7444 find_good_pkt_pointers(other_branch, src_reg,
7445 src_reg->type, true);
7446 mark_pkt_end(this_branch, insn->src_reg, false);
7452 if ((dst_reg->type == PTR_TO_PACKET &&
7453 src_reg->type == PTR_TO_PACKET_END) ||
7454 (dst_reg->type == PTR_TO_PACKET_META &&
7455 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
7456 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
7457 find_good_pkt_pointers(other_branch, dst_reg,
7458 dst_reg->type, true);
7459 mark_pkt_end(this_branch, insn->dst_reg, false);
7460 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
7461 src_reg->type == PTR_TO_PACKET) ||
7462 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
7463 src_reg->type == PTR_TO_PACKET_META)) {
7464 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
7465 find_good_pkt_pointers(this_branch, src_reg,
7466 src_reg->type, false);
7467 mark_pkt_end(other_branch, insn->src_reg, true);
7473 if ((dst_reg->type == PTR_TO_PACKET &&
7474 src_reg->type == PTR_TO_PACKET_END) ||
7475 (dst_reg->type == PTR_TO_PACKET_META &&
7476 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
7477 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
7478 find_good_pkt_pointers(this_branch, dst_reg,
7479 dst_reg->type, true);
7480 mark_pkt_end(other_branch, insn->dst_reg, false);
7481 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
7482 src_reg->type == PTR_TO_PACKET) ||
7483 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
7484 src_reg->type == PTR_TO_PACKET_META)) {
7485 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
7486 find_good_pkt_pointers(other_branch, src_reg,
7487 src_reg->type, false);
7488 mark_pkt_end(this_branch, insn->src_reg, true);
7494 if ((dst_reg->type == PTR_TO_PACKET &&
7495 src_reg->type == PTR_TO_PACKET_END) ||
7496 (dst_reg->type == PTR_TO_PACKET_META &&
7497 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
7498 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
7499 find_good_pkt_pointers(other_branch, dst_reg,
7500 dst_reg->type, false);
7501 mark_pkt_end(this_branch, insn->dst_reg, true);
7502 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
7503 src_reg->type == PTR_TO_PACKET) ||
7504 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
7505 src_reg->type == PTR_TO_PACKET_META)) {
7506 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
7507 find_good_pkt_pointers(this_branch, src_reg,
7508 src_reg->type, true);
7509 mark_pkt_end(other_branch, insn->src_reg, false);
7521 static void find_equal_scalars(struct bpf_verifier_state *vstate,
7522 struct bpf_reg_state *known_reg)
7524 struct bpf_func_state *state;
7525 struct bpf_reg_state *reg;
7528 for (i = 0; i <= vstate->curframe; i++) {
7529 state = vstate->frame[i];
7530 for (j = 0; j < MAX_BPF_REG; j++) {
7531 reg = &state->regs[j];
7532 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
7536 bpf_for_each_spilled_reg(j, state, reg) {
7539 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
7545 static int check_cond_jmp_op(struct bpf_verifier_env *env,
7546 struct bpf_insn *insn, int *insn_idx)
7548 struct bpf_verifier_state *this_branch = env->cur_state;
7549 struct bpf_verifier_state *other_branch;
7550 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
7551 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
7552 u8 opcode = BPF_OP(insn->code);
7557 /* Only conditional jumps are expected to reach here. */
7558 if (opcode == BPF_JA || opcode > BPF_JSLE) {
7559 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
7563 if (BPF_SRC(insn->code) == BPF_X) {
7564 if (insn->imm != 0) {
7565 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
7569 /* check src1 operand */
7570 err = check_reg_arg(env, insn->src_reg, SRC_OP);
7574 if (is_pointer_value(env, insn->src_reg)) {
7575 verbose(env, "R%d pointer comparison prohibited\n",
7579 src_reg = ®s[insn->src_reg];
7581 if (insn->src_reg != BPF_REG_0) {
7582 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
7587 /* check src2 operand */
7588 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
7592 dst_reg = ®s[insn->dst_reg];
7593 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
7595 if (BPF_SRC(insn->code) == BPF_K) {
7596 pred = is_branch_taken(dst_reg, insn->imm, opcode, is_jmp32);
7597 } else if (src_reg->type == SCALAR_VALUE &&
7598 is_jmp32 && tnum_is_const(tnum_subreg(src_reg->var_off))) {
7599 pred = is_branch_taken(dst_reg,
7600 tnum_subreg(src_reg->var_off).value,
7603 } else if (src_reg->type == SCALAR_VALUE &&
7604 !is_jmp32 && tnum_is_const(src_reg->var_off)) {
7605 pred = is_branch_taken(dst_reg,
7606 src_reg->var_off.value,
7609 } else if (reg_is_pkt_pointer_any(dst_reg) &&
7610 reg_is_pkt_pointer_any(src_reg) &&
7612 pred = is_pkt_ptr_branch_taken(dst_reg, src_reg, opcode);
7616 /* If we get here with a dst_reg pointer type it is because
7617 * above is_branch_taken() special cased the 0 comparison.
7619 if (!__is_pointer_value(false, dst_reg))
7620 err = mark_chain_precision(env, insn->dst_reg);
7621 if (BPF_SRC(insn->code) == BPF_X && !err &&
7622 !__is_pointer_value(false, src_reg))
7623 err = mark_chain_precision(env, insn->src_reg);
7628 /* only follow the goto, ignore fall-through */
7629 *insn_idx += insn->off;
7631 } else if (pred == 0) {
7632 /* only follow fall-through branch, since
7633 * that's where the program will go
7638 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
7642 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
7644 /* detect if we are comparing against a constant value so we can adjust
7645 * our min/max values for our dst register.
7646 * this is only legit if both are scalars (or pointers to the same
7647 * object, I suppose, but we don't support that right now), because
7648 * otherwise the different base pointers mean the offsets aren't
7651 if (BPF_SRC(insn->code) == BPF_X) {
7652 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
7654 if (dst_reg->type == SCALAR_VALUE &&
7655 src_reg->type == SCALAR_VALUE) {
7656 if (tnum_is_const(src_reg->var_off) ||
7658 tnum_is_const(tnum_subreg(src_reg->var_off))))
7659 reg_set_min_max(&other_branch_regs[insn->dst_reg],
7661 src_reg->var_off.value,
7662 tnum_subreg(src_reg->var_off).value,
7664 else if (tnum_is_const(dst_reg->var_off) ||
7666 tnum_is_const(tnum_subreg(dst_reg->var_off))))
7667 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
7669 dst_reg->var_off.value,
7670 tnum_subreg(dst_reg->var_off).value,
7672 else if (!is_jmp32 &&
7673 (opcode == BPF_JEQ || opcode == BPF_JNE))
7674 /* Comparing for equality, we can combine knowledge */
7675 reg_combine_min_max(&other_branch_regs[insn->src_reg],
7676 &other_branch_regs[insn->dst_reg],
7677 src_reg, dst_reg, opcode);
7679 !WARN_ON_ONCE(src_reg->id != other_branch_regs[insn->src_reg].id)) {
7680 find_equal_scalars(this_branch, src_reg);
7681 find_equal_scalars(other_branch, &other_branch_regs[insn->src_reg]);
7685 } else if (dst_reg->type == SCALAR_VALUE) {
7686 reg_set_min_max(&other_branch_regs[insn->dst_reg],
7687 dst_reg, insn->imm, (u32)insn->imm,
7691 if (dst_reg->type == SCALAR_VALUE && dst_reg->id &&
7692 !WARN_ON_ONCE(dst_reg->id != other_branch_regs[insn->dst_reg].id)) {
7693 find_equal_scalars(this_branch, dst_reg);
7694 find_equal_scalars(other_branch, &other_branch_regs[insn->dst_reg]);
7697 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
7698 * NOTE: these optimizations below are related with pointer comparison
7699 * which will never be JMP32.
7701 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
7702 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
7703 reg_type_may_be_null(dst_reg->type)) {
7704 /* Mark all identical registers in each branch as either
7705 * safe or unknown depending R == 0 or R != 0 conditional.
7707 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
7709 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
7711 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
7712 this_branch, other_branch) &&
7713 is_pointer_value(env, insn->dst_reg)) {
7714 verbose(env, "R%d pointer comparison prohibited\n",
7718 if (env->log.level & BPF_LOG_LEVEL)
7719 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
7723 /* verify BPF_LD_IMM64 instruction */
7724 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
7726 struct bpf_insn_aux_data *aux = cur_aux(env);
7727 struct bpf_reg_state *regs = cur_regs(env);
7728 struct bpf_reg_state *dst_reg;
7729 struct bpf_map *map;
7732 if (BPF_SIZE(insn->code) != BPF_DW) {
7733 verbose(env, "invalid BPF_LD_IMM insn\n");
7736 if (insn->off != 0) {
7737 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
7741 err = check_reg_arg(env, insn->dst_reg, DST_OP);
7745 dst_reg = ®s[insn->dst_reg];
7746 if (insn->src_reg == 0) {
7747 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
7749 dst_reg->type = SCALAR_VALUE;
7750 __mark_reg_known(®s[insn->dst_reg], imm);
7754 if (insn->src_reg == BPF_PSEUDO_BTF_ID) {
7755 mark_reg_known_zero(env, regs, insn->dst_reg);
7757 dst_reg->type = aux->btf_var.reg_type;
7758 switch (dst_reg->type) {
7760 dst_reg->mem_size = aux->btf_var.mem_size;
7763 case PTR_TO_PERCPU_BTF_ID:
7764 dst_reg->btf = aux->btf_var.btf;
7765 dst_reg->btf_id = aux->btf_var.btf_id;
7768 verbose(env, "bpf verifier is misconfigured\n");
7774 map = env->used_maps[aux->map_index];
7775 mark_reg_known_zero(env, regs, insn->dst_reg);
7776 dst_reg->map_ptr = map;
7778 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE) {
7779 dst_reg->type = PTR_TO_MAP_VALUE;
7780 dst_reg->off = aux->map_off;
7781 if (map_value_has_spin_lock(map))
7782 dst_reg->id = ++env->id_gen;
7783 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD) {
7784 dst_reg->type = CONST_PTR_TO_MAP;
7786 verbose(env, "bpf verifier is misconfigured\n");
7793 static bool may_access_skb(enum bpf_prog_type type)
7796 case BPF_PROG_TYPE_SOCKET_FILTER:
7797 case BPF_PROG_TYPE_SCHED_CLS:
7798 case BPF_PROG_TYPE_SCHED_ACT:
7805 /* verify safety of LD_ABS|LD_IND instructions:
7806 * - they can only appear in the programs where ctx == skb
7807 * - since they are wrappers of function calls, they scratch R1-R5 registers,
7808 * preserve R6-R9, and store return value into R0
7811 * ctx == skb == R6 == CTX
7814 * SRC == any register
7815 * IMM == 32-bit immediate
7818 * R0 - 8/16/32-bit skb data converted to cpu endianness
7820 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
7822 struct bpf_reg_state *regs = cur_regs(env);
7823 static const int ctx_reg = BPF_REG_6;
7824 u8 mode = BPF_MODE(insn->code);
7827 if (!may_access_skb(resolve_prog_type(env->prog))) {
7828 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
7832 if (!env->ops->gen_ld_abs) {
7833 verbose(env, "bpf verifier is misconfigured\n");
7837 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
7838 BPF_SIZE(insn->code) == BPF_DW ||
7839 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
7840 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
7844 /* check whether implicit source operand (register R6) is readable */
7845 err = check_reg_arg(env, ctx_reg, SRC_OP);
7849 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
7850 * gen_ld_abs() may terminate the program at runtime, leading to
7853 err = check_reference_leak(env);
7855 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
7859 if (env->cur_state->active_spin_lock) {
7860 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
7864 if (regs[ctx_reg].type != PTR_TO_CTX) {
7866 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
7870 if (mode == BPF_IND) {
7871 /* check explicit source operand */
7872 err = check_reg_arg(env, insn->src_reg, SRC_OP);
7877 err = check_ctx_reg(env, ®s[ctx_reg], ctx_reg);
7881 /* reset caller saved regs to unreadable */
7882 for (i = 0; i < CALLER_SAVED_REGS; i++) {
7883 mark_reg_not_init(env, regs, caller_saved[i]);
7884 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
7887 /* mark destination R0 register as readable, since it contains
7888 * the value fetched from the packet.
7889 * Already marked as written above.
7891 mark_reg_unknown(env, regs, BPF_REG_0);
7892 /* ld_abs load up to 32-bit skb data. */
7893 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
7897 static int check_return_code(struct bpf_verifier_env *env)
7899 struct tnum enforce_attach_type_range = tnum_unknown;
7900 const struct bpf_prog *prog = env->prog;
7901 struct bpf_reg_state *reg;
7902 struct tnum range = tnum_range(0, 1);
7903 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
7905 const bool is_subprog = env->cur_state->frame[0]->subprogno;
7907 /* LSM and struct_ops func-ptr's return type could be "void" */
7909 (prog_type == BPF_PROG_TYPE_STRUCT_OPS ||
7910 prog_type == BPF_PROG_TYPE_LSM) &&
7911 !prog->aux->attach_func_proto->type)
7914 /* eBPF calling convetion is such that R0 is used
7915 * to return the value from eBPF program.
7916 * Make sure that it's readable at this time
7917 * of bpf_exit, which means that program wrote
7918 * something into it earlier
7920 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
7924 if (is_pointer_value(env, BPF_REG_0)) {
7925 verbose(env, "R0 leaks addr as return value\n");
7929 reg = cur_regs(env) + BPF_REG_0;
7931 if (reg->type != SCALAR_VALUE) {
7932 verbose(env, "At subprogram exit the register R0 is not a scalar value (%s)\n",
7933 reg_type_str[reg->type]);
7939 switch (prog_type) {
7940 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
7941 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
7942 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG ||
7943 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETPEERNAME ||
7944 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETPEERNAME ||
7945 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETSOCKNAME ||
7946 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETSOCKNAME)
7947 range = tnum_range(1, 1);
7949 case BPF_PROG_TYPE_CGROUP_SKB:
7950 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
7951 range = tnum_range(0, 3);
7952 enforce_attach_type_range = tnum_range(2, 3);
7955 case BPF_PROG_TYPE_CGROUP_SOCK:
7956 case BPF_PROG_TYPE_SOCK_OPS:
7957 case BPF_PROG_TYPE_CGROUP_DEVICE:
7958 case BPF_PROG_TYPE_CGROUP_SYSCTL:
7959 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
7961 case BPF_PROG_TYPE_RAW_TRACEPOINT:
7962 if (!env->prog->aux->attach_btf_id)
7964 range = tnum_const(0);
7966 case BPF_PROG_TYPE_TRACING:
7967 switch (env->prog->expected_attach_type) {
7968 case BPF_TRACE_FENTRY:
7969 case BPF_TRACE_FEXIT:
7970 range = tnum_const(0);
7972 case BPF_TRACE_RAW_TP:
7973 case BPF_MODIFY_RETURN:
7975 case BPF_TRACE_ITER:
7981 case BPF_PROG_TYPE_SK_LOOKUP:
7982 range = tnum_range(SK_DROP, SK_PASS);
7984 case BPF_PROG_TYPE_EXT:
7985 /* freplace program can return anything as its return value
7986 * depends on the to-be-replaced kernel func or bpf program.
7992 if (reg->type != SCALAR_VALUE) {
7993 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
7994 reg_type_str[reg->type]);
7998 if (!tnum_in(range, reg->var_off)) {
8001 verbose(env, "At program exit the register R0 ");
8002 if (!tnum_is_unknown(reg->var_off)) {
8003 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
8004 verbose(env, "has value %s", tn_buf);
8006 verbose(env, "has unknown scalar value");
8008 tnum_strn(tn_buf, sizeof(tn_buf), range);
8009 verbose(env, " should have been in %s\n", tn_buf);
8013 if (!tnum_is_unknown(enforce_attach_type_range) &&
8014 tnum_in(enforce_attach_type_range, reg->var_off))
8015 env->prog->enforce_expected_attach_type = 1;
8019 /* non-recursive DFS pseudo code
8020 * 1 procedure DFS-iterative(G,v):
8021 * 2 label v as discovered
8022 * 3 let S be a stack
8024 * 5 while S is not empty
8026 * 7 if t is what we're looking for:
8028 * 9 for all edges e in G.adjacentEdges(t) do
8029 * 10 if edge e is already labelled
8030 * 11 continue with the next edge
8031 * 12 w <- G.adjacentVertex(t,e)
8032 * 13 if vertex w is not discovered and not explored
8033 * 14 label e as tree-edge
8034 * 15 label w as discovered
8037 * 18 else if vertex w is discovered
8038 * 19 label e as back-edge
8040 * 21 // vertex w is explored
8041 * 22 label e as forward- or cross-edge
8042 * 23 label t as explored
8047 * 0x11 - discovered and fall-through edge labelled
8048 * 0x12 - discovered and fall-through and branch edges labelled
8059 static u32 state_htab_size(struct bpf_verifier_env *env)
8061 return env->prog->len;
8064 static struct bpf_verifier_state_list **explored_state(
8065 struct bpf_verifier_env *env,
8068 struct bpf_verifier_state *cur = env->cur_state;
8069 struct bpf_func_state *state = cur->frame[cur->curframe];
8071 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
8074 static void init_explored_state(struct bpf_verifier_env *env, int idx)
8076 env->insn_aux_data[idx].prune_point = true;
8084 /* t, w, e - match pseudo-code above:
8085 * t - index of current instruction
8086 * w - next instruction
8089 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env,
8092 int *insn_stack = env->cfg.insn_stack;
8093 int *insn_state = env->cfg.insn_state;
8095 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
8096 return DONE_EXPLORING;
8098 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
8099 return DONE_EXPLORING;
8101 if (w < 0 || w >= env->prog->len) {
8102 verbose_linfo(env, t, "%d: ", t);
8103 verbose(env, "jump out of range from insn %d to %d\n", t, w);
8108 /* mark branch target for state pruning */
8109 init_explored_state(env, w);
8111 if (insn_state[w] == 0) {
8113 insn_state[t] = DISCOVERED | e;
8114 insn_state[w] = DISCOVERED;
8115 if (env->cfg.cur_stack >= env->prog->len)
8117 insn_stack[env->cfg.cur_stack++] = w;
8118 return KEEP_EXPLORING;
8119 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
8120 if (loop_ok && env->bpf_capable)
8121 return DONE_EXPLORING;
8122 verbose_linfo(env, t, "%d: ", t);
8123 verbose_linfo(env, w, "%d: ", w);
8124 verbose(env, "back-edge from insn %d to %d\n", t, w);
8126 } else if (insn_state[w] == EXPLORED) {
8127 /* forward- or cross-edge */
8128 insn_state[t] = DISCOVERED | e;
8130 verbose(env, "insn state internal bug\n");
8133 return DONE_EXPLORING;
8136 /* Visits the instruction at index t and returns one of the following:
8137 * < 0 - an error occurred
8138 * DONE_EXPLORING - the instruction was fully explored
8139 * KEEP_EXPLORING - there is still work to be done before it is fully explored
8141 static int visit_insn(int t, int insn_cnt, struct bpf_verifier_env *env)
8143 struct bpf_insn *insns = env->prog->insnsi;
8146 /* All non-branch instructions have a single fall-through edge. */
8147 if (BPF_CLASS(insns[t].code) != BPF_JMP &&
8148 BPF_CLASS(insns[t].code) != BPF_JMP32)
8149 return push_insn(t, t + 1, FALLTHROUGH, env, false);
8151 switch (BPF_OP(insns[t].code)) {
8153 return DONE_EXPLORING;
8156 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
8160 if (t + 1 < insn_cnt)
8161 init_explored_state(env, t + 1);
8162 if (insns[t].src_reg == BPF_PSEUDO_CALL) {
8163 init_explored_state(env, t);
8164 ret = push_insn(t, t + insns[t].imm + 1, BRANCH,
8170 if (BPF_SRC(insns[t].code) != BPF_K)
8173 /* unconditional jump with single edge */
8174 ret = push_insn(t, t + insns[t].off + 1, FALLTHROUGH, env,
8179 /* unconditional jmp is not a good pruning point,
8180 * but it's marked, since backtracking needs
8181 * to record jmp history in is_state_visited().
8183 init_explored_state(env, t + insns[t].off + 1);
8184 /* tell verifier to check for equivalent states
8185 * after every call and jump
8187 if (t + 1 < insn_cnt)
8188 init_explored_state(env, t + 1);
8193 /* conditional jump with two edges */
8194 init_explored_state(env, t);
8195 ret = push_insn(t, t + 1, FALLTHROUGH, env, true);
8199 return push_insn(t, t + insns[t].off + 1, BRANCH, env, true);
8203 /* non-recursive depth-first-search to detect loops in BPF program
8204 * loop == back-edge in directed graph
8206 static int check_cfg(struct bpf_verifier_env *env)
8208 int insn_cnt = env->prog->len;
8209 int *insn_stack, *insn_state;
8213 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
8217 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
8223 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
8224 insn_stack[0] = 0; /* 0 is the first instruction */
8225 env->cfg.cur_stack = 1;
8227 while (env->cfg.cur_stack > 0) {
8228 int t = insn_stack[env->cfg.cur_stack - 1];
8230 ret = visit_insn(t, insn_cnt, env);
8232 case DONE_EXPLORING:
8233 insn_state[t] = EXPLORED;
8234 env->cfg.cur_stack--;
8236 case KEEP_EXPLORING:
8240 verbose(env, "visit_insn internal bug\n");
8247 if (env->cfg.cur_stack < 0) {
8248 verbose(env, "pop stack internal bug\n");
8253 for (i = 0; i < insn_cnt; i++) {
8254 if (insn_state[i] != EXPLORED) {
8255 verbose(env, "unreachable insn %d\n", i);
8260 ret = 0; /* cfg looks good */
8265 env->cfg.insn_state = env->cfg.insn_stack = NULL;
8269 static int check_abnormal_return(struct bpf_verifier_env *env)
8273 for (i = 1; i < env->subprog_cnt; i++) {
8274 if (env->subprog_info[i].has_ld_abs) {
8275 verbose(env, "LD_ABS is not allowed in subprogs without BTF\n");
8278 if (env->subprog_info[i].has_tail_call) {
8279 verbose(env, "tail_call is not allowed in subprogs without BTF\n");
8286 /* The minimum supported BTF func info size */
8287 #define MIN_BPF_FUNCINFO_SIZE 8
8288 #define MAX_FUNCINFO_REC_SIZE 252
8290 static int check_btf_func(struct bpf_verifier_env *env,
8291 const union bpf_attr *attr,
8292 union bpf_attr __user *uattr)
8294 const struct btf_type *type, *func_proto, *ret_type;
8295 u32 i, nfuncs, urec_size, min_size;
8296 u32 krec_size = sizeof(struct bpf_func_info);
8297 struct bpf_func_info *krecord;
8298 struct bpf_func_info_aux *info_aux = NULL;
8299 struct bpf_prog *prog;
8300 const struct btf *btf;
8301 void __user *urecord;
8302 u32 prev_offset = 0;
8306 nfuncs = attr->func_info_cnt;
8308 if (check_abnormal_return(env))
8313 if (nfuncs != env->subprog_cnt) {
8314 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
8318 urec_size = attr->func_info_rec_size;
8319 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
8320 urec_size > MAX_FUNCINFO_REC_SIZE ||
8321 urec_size % sizeof(u32)) {
8322 verbose(env, "invalid func info rec size %u\n", urec_size);
8327 btf = prog->aux->btf;
8329 urecord = u64_to_user_ptr(attr->func_info);
8330 min_size = min_t(u32, krec_size, urec_size);
8332 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
8335 info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN);
8339 for (i = 0; i < nfuncs; i++) {
8340 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
8342 if (ret == -E2BIG) {
8343 verbose(env, "nonzero tailing record in func info");
8344 /* set the size kernel expects so loader can zero
8345 * out the rest of the record.
8347 if (put_user(min_size, &uattr->func_info_rec_size))
8353 if (copy_from_user(&krecord[i], urecord, min_size)) {
8358 /* check insn_off */
8361 if (krecord[i].insn_off) {
8363 "nonzero insn_off %u for the first func info record",
8364 krecord[i].insn_off);
8367 } else if (krecord[i].insn_off <= prev_offset) {
8369 "same or smaller insn offset (%u) than previous func info record (%u)",
8370 krecord[i].insn_off, prev_offset);
8374 if (env->subprog_info[i].start != krecord[i].insn_off) {
8375 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
8380 type = btf_type_by_id(btf, krecord[i].type_id);
8381 if (!type || !btf_type_is_func(type)) {
8382 verbose(env, "invalid type id %d in func info",
8383 krecord[i].type_id);
8386 info_aux[i].linkage = BTF_INFO_VLEN(type->info);
8388 func_proto = btf_type_by_id(btf, type->type);
8389 if (unlikely(!func_proto || !btf_type_is_func_proto(func_proto)))
8390 /* btf_func_check() already verified it during BTF load */
8392 ret_type = btf_type_skip_modifiers(btf, func_proto->type, NULL);
8394 btf_type_is_small_int(ret_type) || btf_type_is_enum(ret_type);
8395 if (i && !scalar_return && env->subprog_info[i].has_ld_abs) {
8396 verbose(env, "LD_ABS is only allowed in functions that return 'int'.\n");
8399 if (i && !scalar_return && env->subprog_info[i].has_tail_call) {
8400 verbose(env, "tail_call is only allowed in functions that return 'int'.\n");
8404 prev_offset = krecord[i].insn_off;
8405 urecord += urec_size;
8408 prog->aux->func_info = krecord;
8409 prog->aux->func_info_cnt = nfuncs;
8410 prog->aux->func_info_aux = info_aux;
8419 static void adjust_btf_func(struct bpf_verifier_env *env)
8421 struct bpf_prog_aux *aux = env->prog->aux;
8424 if (!aux->func_info)
8427 for (i = 0; i < env->subprog_cnt; i++)
8428 aux->func_info[i].insn_off = env->subprog_info[i].start;
8431 #define MIN_BPF_LINEINFO_SIZE (offsetof(struct bpf_line_info, line_col) + \
8432 sizeof(((struct bpf_line_info *)(0))->line_col))
8433 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
8435 static int check_btf_line(struct bpf_verifier_env *env,
8436 const union bpf_attr *attr,
8437 union bpf_attr __user *uattr)
8439 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
8440 struct bpf_subprog_info *sub;
8441 struct bpf_line_info *linfo;
8442 struct bpf_prog *prog;
8443 const struct btf *btf;
8444 void __user *ulinfo;
8447 nr_linfo = attr->line_info_cnt;
8451 rec_size = attr->line_info_rec_size;
8452 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
8453 rec_size > MAX_LINEINFO_REC_SIZE ||
8454 rec_size & (sizeof(u32) - 1))
8457 /* Need to zero it in case the userspace may
8458 * pass in a smaller bpf_line_info object.
8460 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
8461 GFP_KERNEL | __GFP_NOWARN);
8466 btf = prog->aux->btf;
8469 sub = env->subprog_info;
8470 ulinfo = u64_to_user_ptr(attr->line_info);
8471 expected_size = sizeof(struct bpf_line_info);
8472 ncopy = min_t(u32, expected_size, rec_size);
8473 for (i = 0; i < nr_linfo; i++) {
8474 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
8476 if (err == -E2BIG) {
8477 verbose(env, "nonzero tailing record in line_info");
8478 if (put_user(expected_size,
8479 &uattr->line_info_rec_size))
8485 if (copy_from_user(&linfo[i], ulinfo, ncopy)) {
8491 * Check insn_off to ensure
8492 * 1) strictly increasing AND
8493 * 2) bounded by prog->len
8495 * The linfo[0].insn_off == 0 check logically falls into
8496 * the later "missing bpf_line_info for func..." case
8497 * because the first linfo[0].insn_off must be the
8498 * first sub also and the first sub must have
8499 * subprog_info[0].start == 0.
8501 if ((i && linfo[i].insn_off <= prev_offset) ||
8502 linfo[i].insn_off >= prog->len) {
8503 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
8504 i, linfo[i].insn_off, prev_offset,
8510 if (!prog->insnsi[linfo[i].insn_off].code) {
8512 "Invalid insn code at line_info[%u].insn_off\n",
8518 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
8519 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
8520 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
8525 if (s != env->subprog_cnt) {
8526 if (linfo[i].insn_off == sub[s].start) {
8527 sub[s].linfo_idx = i;
8529 } else if (sub[s].start < linfo[i].insn_off) {
8530 verbose(env, "missing bpf_line_info for func#%u\n", s);
8536 prev_offset = linfo[i].insn_off;
8540 if (s != env->subprog_cnt) {
8541 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
8542 env->subprog_cnt - s, s);
8547 prog->aux->linfo = linfo;
8548 prog->aux->nr_linfo = nr_linfo;
8557 static int check_btf_info(struct bpf_verifier_env *env,
8558 const union bpf_attr *attr,
8559 union bpf_attr __user *uattr)
8564 if (!attr->func_info_cnt && !attr->line_info_cnt) {
8565 if (check_abnormal_return(env))
8570 btf = btf_get_by_fd(attr->prog_btf_fd);
8572 return PTR_ERR(btf);
8573 env->prog->aux->btf = btf;
8575 err = check_btf_func(env, attr, uattr);
8579 err = check_btf_line(env, attr, uattr);
8586 /* check %cur's range satisfies %old's */
8587 static bool range_within(struct bpf_reg_state *old,
8588 struct bpf_reg_state *cur)
8590 return old->umin_value <= cur->umin_value &&
8591 old->umax_value >= cur->umax_value &&
8592 old->smin_value <= cur->smin_value &&
8593 old->smax_value >= cur->smax_value;
8596 /* Maximum number of register states that can exist at once */
8597 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
8603 /* If in the old state two registers had the same id, then they need to have
8604 * the same id in the new state as well. But that id could be different from
8605 * the old state, so we need to track the mapping from old to new ids.
8606 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
8607 * regs with old id 5 must also have new id 9 for the new state to be safe. But
8608 * regs with a different old id could still have new id 9, we don't care about
8610 * So we look through our idmap to see if this old id has been seen before. If
8611 * so, we require the new id to match; otherwise, we add the id pair to the map.
8613 static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
8617 for (i = 0; i < ID_MAP_SIZE; i++) {
8618 if (!idmap[i].old) {
8619 /* Reached an empty slot; haven't seen this id before */
8620 idmap[i].old = old_id;
8621 idmap[i].cur = cur_id;
8624 if (idmap[i].old == old_id)
8625 return idmap[i].cur == cur_id;
8627 /* We ran out of idmap slots, which should be impossible */
8632 static void clean_func_state(struct bpf_verifier_env *env,
8633 struct bpf_func_state *st)
8635 enum bpf_reg_liveness live;
8638 for (i = 0; i < BPF_REG_FP; i++) {
8639 live = st->regs[i].live;
8640 /* liveness must not touch this register anymore */
8641 st->regs[i].live |= REG_LIVE_DONE;
8642 if (!(live & REG_LIVE_READ))
8643 /* since the register is unused, clear its state
8644 * to make further comparison simpler
8646 __mark_reg_not_init(env, &st->regs[i]);
8649 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
8650 live = st->stack[i].spilled_ptr.live;
8651 /* liveness must not touch this stack slot anymore */
8652 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
8653 if (!(live & REG_LIVE_READ)) {
8654 __mark_reg_not_init(env, &st->stack[i].spilled_ptr);
8655 for (j = 0; j < BPF_REG_SIZE; j++)
8656 st->stack[i].slot_type[j] = STACK_INVALID;
8661 static void clean_verifier_state(struct bpf_verifier_env *env,
8662 struct bpf_verifier_state *st)
8666 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
8667 /* all regs in this state in all frames were already marked */
8670 for (i = 0; i <= st->curframe; i++)
8671 clean_func_state(env, st->frame[i]);
8674 /* the parentage chains form a tree.
8675 * the verifier states are added to state lists at given insn and
8676 * pushed into state stack for future exploration.
8677 * when the verifier reaches bpf_exit insn some of the verifer states
8678 * stored in the state lists have their final liveness state already,
8679 * but a lot of states will get revised from liveness point of view when
8680 * the verifier explores other branches.
8683 * 2: if r1 == 100 goto pc+1
8686 * when the verifier reaches exit insn the register r0 in the state list of
8687 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
8688 * of insn 2 and goes exploring further. At the insn 4 it will walk the
8689 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
8691 * Since the verifier pushes the branch states as it sees them while exploring
8692 * the program the condition of walking the branch instruction for the second
8693 * time means that all states below this branch were already explored and
8694 * their final liveness markes are already propagated.
8695 * Hence when the verifier completes the search of state list in is_state_visited()
8696 * we can call this clean_live_states() function to mark all liveness states
8697 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
8699 * This function also clears the registers and stack for states that !READ
8700 * to simplify state merging.
8702 * Important note here that walking the same branch instruction in the callee
8703 * doesn't meant that the states are DONE. The verifier has to compare
8706 static void clean_live_states(struct bpf_verifier_env *env, int insn,
8707 struct bpf_verifier_state *cur)
8709 struct bpf_verifier_state_list *sl;
8712 sl = *explored_state(env, insn);
8714 if (sl->state.branches)
8716 if (sl->state.insn_idx != insn ||
8717 sl->state.curframe != cur->curframe)
8719 for (i = 0; i <= cur->curframe; i++)
8720 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
8722 clean_verifier_state(env, &sl->state);
8728 /* Returns true if (rold safe implies rcur safe) */
8729 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
8730 struct idpair *idmap)
8734 if (!(rold->live & REG_LIVE_READ))
8735 /* explored state didn't use this */
8738 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
8740 if (rold->type == PTR_TO_STACK)
8741 /* two stack pointers are equal only if they're pointing to
8742 * the same stack frame, since fp-8 in foo != fp-8 in bar
8744 return equal && rold->frameno == rcur->frameno;
8749 if (rold->type == NOT_INIT)
8750 /* explored state can't have used this */
8752 if (rcur->type == NOT_INIT)
8754 switch (rold->type) {
8756 if (rcur->type == SCALAR_VALUE) {
8757 if (!rold->precise && !rcur->precise)
8759 /* new val must satisfy old val knowledge */
8760 return range_within(rold, rcur) &&
8761 tnum_in(rold->var_off, rcur->var_off);
8763 /* We're trying to use a pointer in place of a scalar.
8764 * Even if the scalar was unbounded, this could lead to
8765 * pointer leaks because scalars are allowed to leak
8766 * while pointers are not. We could make this safe in
8767 * special cases if root is calling us, but it's
8768 * probably not worth the hassle.
8772 case PTR_TO_MAP_VALUE:
8773 /* If the new min/max/var_off satisfy the old ones and
8774 * everything else matches, we are OK.
8775 * 'id' is not compared, since it's only used for maps with
8776 * bpf_spin_lock inside map element and in such cases if
8777 * the rest of the prog is valid for one map element then
8778 * it's valid for all map elements regardless of the key
8779 * used in bpf_map_lookup()
8781 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
8782 range_within(rold, rcur) &&
8783 tnum_in(rold->var_off, rcur->var_off);
8784 case PTR_TO_MAP_VALUE_OR_NULL:
8785 /* a PTR_TO_MAP_VALUE could be safe to use as a
8786 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
8787 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
8788 * checked, doing so could have affected others with the same
8789 * id, and we can't check for that because we lost the id when
8790 * we converted to a PTR_TO_MAP_VALUE.
8792 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
8794 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
8796 /* Check our ids match any regs they're supposed to */
8797 return check_ids(rold->id, rcur->id, idmap);
8798 case PTR_TO_PACKET_META:
8800 if (rcur->type != rold->type)
8802 /* We must have at least as much range as the old ptr
8803 * did, so that any accesses which were safe before are
8804 * still safe. This is true even if old range < old off,
8805 * since someone could have accessed through (ptr - k), or
8806 * even done ptr -= k in a register, to get a safe access.
8808 if (rold->range > rcur->range)
8810 /* If the offsets don't match, we can't trust our alignment;
8811 * nor can we be sure that we won't fall out of range.
8813 if (rold->off != rcur->off)
8815 /* id relations must be preserved */
8816 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
8818 /* new val must satisfy old val knowledge */
8819 return range_within(rold, rcur) &&
8820 tnum_in(rold->var_off, rcur->var_off);
8822 case CONST_PTR_TO_MAP:
8823 case PTR_TO_PACKET_END:
8824 case PTR_TO_FLOW_KEYS:
8826 case PTR_TO_SOCKET_OR_NULL:
8827 case PTR_TO_SOCK_COMMON:
8828 case PTR_TO_SOCK_COMMON_OR_NULL:
8829 case PTR_TO_TCP_SOCK:
8830 case PTR_TO_TCP_SOCK_OR_NULL:
8831 case PTR_TO_XDP_SOCK:
8832 /* Only valid matches are exact, which memcmp() above
8833 * would have accepted
8836 /* Don't know what's going on, just say it's not safe */
8840 /* Shouldn't get here; if we do, say it's not safe */
8845 static bool stacksafe(struct bpf_func_state *old,
8846 struct bpf_func_state *cur,
8847 struct idpair *idmap)
8851 /* walk slots of the explored stack and ignore any additional
8852 * slots in the current stack, since explored(safe) state
8855 for (i = 0; i < old->allocated_stack; i++) {
8856 spi = i / BPF_REG_SIZE;
8858 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
8859 i += BPF_REG_SIZE - 1;
8860 /* explored state didn't use this */
8864 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
8867 /* explored stack has more populated slots than current stack
8868 * and these slots were used
8870 if (i >= cur->allocated_stack)
8873 /* if old state was safe with misc data in the stack
8874 * it will be safe with zero-initialized stack.
8875 * The opposite is not true
8877 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
8878 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
8880 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
8881 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
8882 /* Ex: old explored (safe) state has STACK_SPILL in
8883 * this stack slot, but current has STACK_MISC ->
8884 * this verifier states are not equivalent,
8885 * return false to continue verification of this path
8888 if (i % BPF_REG_SIZE)
8890 if (old->stack[spi].slot_type[0] != STACK_SPILL)
8892 if (!regsafe(&old->stack[spi].spilled_ptr,
8893 &cur->stack[spi].spilled_ptr,
8895 /* when explored and current stack slot are both storing
8896 * spilled registers, check that stored pointers types
8897 * are the same as well.
8898 * Ex: explored safe path could have stored
8899 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
8900 * but current path has stored:
8901 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
8902 * such verifier states are not equivalent.
8903 * return false to continue verification of this path
8910 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
8912 if (old->acquired_refs != cur->acquired_refs)
8914 return !memcmp(old->refs, cur->refs,
8915 sizeof(*old->refs) * old->acquired_refs);
8918 /* compare two verifier states
8920 * all states stored in state_list are known to be valid, since
8921 * verifier reached 'bpf_exit' instruction through them
8923 * this function is called when verifier exploring different branches of
8924 * execution popped from the state stack. If it sees an old state that has
8925 * more strict register state and more strict stack state then this execution
8926 * branch doesn't need to be explored further, since verifier already
8927 * concluded that more strict state leads to valid finish.
8929 * Therefore two states are equivalent if register state is more conservative
8930 * and explored stack state is more conservative than the current one.
8933 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
8934 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
8936 * In other words if current stack state (one being explored) has more
8937 * valid slots than old one that already passed validation, it means
8938 * the verifier can stop exploring and conclude that current state is valid too
8940 * Similarly with registers. If explored state has register type as invalid
8941 * whereas register type in current state is meaningful, it means that
8942 * the current state will reach 'bpf_exit' instruction safely
8944 static bool func_states_equal(struct bpf_func_state *old,
8945 struct bpf_func_state *cur)
8947 struct idpair *idmap;
8951 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
8952 /* If we failed to allocate the idmap, just say it's not safe */
8956 for (i = 0; i < MAX_BPF_REG; i++) {
8957 if (!regsafe(&old->regs[i], &cur->regs[i], idmap))
8961 if (!stacksafe(old, cur, idmap))
8964 if (!refsafe(old, cur))
8972 static bool states_equal(struct bpf_verifier_env *env,
8973 struct bpf_verifier_state *old,
8974 struct bpf_verifier_state *cur)
8978 if (old->curframe != cur->curframe)
8981 /* Verification state from speculative execution simulation
8982 * must never prune a non-speculative execution one.
8984 if (old->speculative && !cur->speculative)
8987 if (old->active_spin_lock != cur->active_spin_lock)
8990 /* for states to be equal callsites have to be the same
8991 * and all frame states need to be equivalent
8993 for (i = 0; i <= old->curframe; i++) {
8994 if (old->frame[i]->callsite != cur->frame[i]->callsite)
8996 if (!func_states_equal(old->frame[i], cur->frame[i]))
9002 /* Return 0 if no propagation happened. Return negative error code if error
9003 * happened. Otherwise, return the propagated bit.
9005 static int propagate_liveness_reg(struct bpf_verifier_env *env,
9006 struct bpf_reg_state *reg,
9007 struct bpf_reg_state *parent_reg)
9009 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
9010 u8 flag = reg->live & REG_LIVE_READ;
9013 /* When comes here, read flags of PARENT_REG or REG could be any of
9014 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
9015 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
9017 if (parent_flag == REG_LIVE_READ64 ||
9018 /* Or if there is no read flag from REG. */
9020 /* Or if the read flag from REG is the same as PARENT_REG. */
9021 parent_flag == flag)
9024 err = mark_reg_read(env, reg, parent_reg, flag);
9031 /* A write screens off any subsequent reads; but write marks come from the
9032 * straight-line code between a state and its parent. When we arrive at an
9033 * equivalent state (jump target or such) we didn't arrive by the straight-line
9034 * code, so read marks in the state must propagate to the parent regardless
9035 * of the state's write marks. That's what 'parent == state->parent' comparison
9036 * in mark_reg_read() is for.
9038 static int propagate_liveness(struct bpf_verifier_env *env,
9039 const struct bpf_verifier_state *vstate,
9040 struct bpf_verifier_state *vparent)
9042 struct bpf_reg_state *state_reg, *parent_reg;
9043 struct bpf_func_state *state, *parent;
9044 int i, frame, err = 0;
9046 if (vparent->curframe != vstate->curframe) {
9047 WARN(1, "propagate_live: parent frame %d current frame %d\n",
9048 vparent->curframe, vstate->curframe);
9051 /* Propagate read liveness of registers... */
9052 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
9053 for (frame = 0; frame <= vstate->curframe; frame++) {
9054 parent = vparent->frame[frame];
9055 state = vstate->frame[frame];
9056 parent_reg = parent->regs;
9057 state_reg = state->regs;
9058 /* We don't need to worry about FP liveness, it's read-only */
9059 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
9060 err = propagate_liveness_reg(env, &state_reg[i],
9064 if (err == REG_LIVE_READ64)
9065 mark_insn_zext(env, &parent_reg[i]);
9068 /* Propagate stack slots. */
9069 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
9070 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
9071 parent_reg = &parent->stack[i].spilled_ptr;
9072 state_reg = &state->stack[i].spilled_ptr;
9073 err = propagate_liveness_reg(env, state_reg,
9082 /* find precise scalars in the previous equivalent state and
9083 * propagate them into the current state
9085 static int propagate_precision(struct bpf_verifier_env *env,
9086 const struct bpf_verifier_state *old)
9088 struct bpf_reg_state *state_reg;
9089 struct bpf_func_state *state;
9092 state = old->frame[old->curframe];
9093 state_reg = state->regs;
9094 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
9095 if (state_reg->type != SCALAR_VALUE ||
9096 !state_reg->precise)
9098 if (env->log.level & BPF_LOG_LEVEL2)
9099 verbose(env, "propagating r%d\n", i);
9100 err = mark_chain_precision(env, i);
9105 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
9106 if (state->stack[i].slot_type[0] != STACK_SPILL)
9108 state_reg = &state->stack[i].spilled_ptr;
9109 if (state_reg->type != SCALAR_VALUE ||
9110 !state_reg->precise)
9112 if (env->log.level & BPF_LOG_LEVEL2)
9113 verbose(env, "propagating fp%d\n",
9114 (-i - 1) * BPF_REG_SIZE);
9115 err = mark_chain_precision_stack(env, i);
9122 static bool states_maybe_looping(struct bpf_verifier_state *old,
9123 struct bpf_verifier_state *cur)
9125 struct bpf_func_state *fold, *fcur;
9126 int i, fr = cur->curframe;
9128 if (old->curframe != fr)
9131 fold = old->frame[fr];
9132 fcur = cur->frame[fr];
9133 for (i = 0; i < MAX_BPF_REG; i++)
9134 if (memcmp(&fold->regs[i], &fcur->regs[i],
9135 offsetof(struct bpf_reg_state, parent)))
9141 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
9143 struct bpf_verifier_state_list *new_sl;
9144 struct bpf_verifier_state_list *sl, **pprev;
9145 struct bpf_verifier_state *cur = env->cur_state, *new;
9146 int i, j, err, states_cnt = 0;
9147 bool add_new_state = env->test_state_freq ? true : false;
9149 cur->last_insn_idx = env->prev_insn_idx;
9150 if (!env->insn_aux_data[insn_idx].prune_point)
9151 /* this 'insn_idx' instruction wasn't marked, so we will not
9152 * be doing state search here
9156 /* bpf progs typically have pruning point every 4 instructions
9157 * http://vger.kernel.org/bpfconf2019.html#session-1
9158 * Do not add new state for future pruning if the verifier hasn't seen
9159 * at least 2 jumps and at least 8 instructions.
9160 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
9161 * In tests that amounts to up to 50% reduction into total verifier
9162 * memory consumption and 20% verifier time speedup.
9164 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
9165 env->insn_processed - env->prev_insn_processed >= 8)
9166 add_new_state = true;
9168 pprev = explored_state(env, insn_idx);
9171 clean_live_states(env, insn_idx, cur);
9175 if (sl->state.insn_idx != insn_idx)
9177 if (sl->state.branches) {
9178 if (states_maybe_looping(&sl->state, cur) &&
9179 states_equal(env, &sl->state, cur)) {
9180 verbose_linfo(env, insn_idx, "; ");
9181 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
9184 /* if the verifier is processing a loop, avoid adding new state
9185 * too often, since different loop iterations have distinct
9186 * states and may not help future pruning.
9187 * This threshold shouldn't be too low to make sure that
9188 * a loop with large bound will be rejected quickly.
9189 * The most abusive loop will be:
9191 * if r1 < 1000000 goto pc-2
9192 * 1M insn_procssed limit / 100 == 10k peak states.
9193 * This threshold shouldn't be too high either, since states
9194 * at the end of the loop are likely to be useful in pruning.
9196 if (env->jmps_processed - env->prev_jmps_processed < 20 &&
9197 env->insn_processed - env->prev_insn_processed < 100)
9198 add_new_state = false;
9201 if (states_equal(env, &sl->state, cur)) {
9203 /* reached equivalent register/stack state,
9205 * Registers read by the continuation are read by us.
9206 * If we have any write marks in env->cur_state, they
9207 * will prevent corresponding reads in the continuation
9208 * from reaching our parent (an explored_state). Our
9209 * own state will get the read marks recorded, but
9210 * they'll be immediately forgotten as we're pruning
9211 * this state and will pop a new one.
9213 err = propagate_liveness(env, &sl->state, cur);
9215 /* if previous state reached the exit with precision and
9216 * current state is equivalent to it (except precsion marks)
9217 * the precision needs to be propagated back in
9218 * the current state.
9220 err = err ? : push_jmp_history(env, cur);
9221 err = err ? : propagate_precision(env, &sl->state);
9227 /* when new state is not going to be added do not increase miss count.
9228 * Otherwise several loop iterations will remove the state
9229 * recorded earlier. The goal of these heuristics is to have
9230 * states from some iterations of the loop (some in the beginning
9231 * and some at the end) to help pruning.
9235 /* heuristic to determine whether this state is beneficial
9236 * to keep checking from state equivalence point of view.
9237 * Higher numbers increase max_states_per_insn and verification time,
9238 * but do not meaningfully decrease insn_processed.
9240 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
9241 /* the state is unlikely to be useful. Remove it to
9242 * speed up verification
9245 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
9246 u32 br = sl->state.branches;
9249 "BUG live_done but branches_to_explore %d\n",
9251 free_verifier_state(&sl->state, false);
9255 /* cannot free this state, since parentage chain may
9256 * walk it later. Add it for free_list instead to
9257 * be freed at the end of verification
9259 sl->next = env->free_list;
9260 env->free_list = sl;
9270 if (env->max_states_per_insn < states_cnt)
9271 env->max_states_per_insn = states_cnt;
9273 if (!env->bpf_capable && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
9274 return push_jmp_history(env, cur);
9277 return push_jmp_history(env, cur);
9279 /* There were no equivalent states, remember the current one.
9280 * Technically the current state is not proven to be safe yet,
9281 * but it will either reach outer most bpf_exit (which means it's safe)
9282 * or it will be rejected. When there are no loops the verifier won't be
9283 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
9284 * again on the way to bpf_exit.
9285 * When looping the sl->state.branches will be > 0 and this state
9286 * will not be considered for equivalence until branches == 0.
9288 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
9291 env->total_states++;
9293 env->prev_jmps_processed = env->jmps_processed;
9294 env->prev_insn_processed = env->insn_processed;
9296 /* add new state to the head of linked list */
9297 new = &new_sl->state;
9298 err = copy_verifier_state(new, cur);
9300 free_verifier_state(new, false);
9304 new->insn_idx = insn_idx;
9305 WARN_ONCE(new->branches != 1,
9306 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
9309 cur->first_insn_idx = insn_idx;
9310 clear_jmp_history(cur);
9311 new_sl->next = *explored_state(env, insn_idx);
9312 *explored_state(env, insn_idx) = new_sl;
9313 /* connect new state to parentage chain. Current frame needs all
9314 * registers connected. Only r6 - r9 of the callers are alive (pushed
9315 * to the stack implicitly by JITs) so in callers' frames connect just
9316 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
9317 * the state of the call instruction (with WRITTEN set), and r0 comes
9318 * from callee with its full parentage chain, anyway.
9320 /* clear write marks in current state: the writes we did are not writes
9321 * our child did, so they don't screen off its reads from us.
9322 * (There are no read marks in current state, because reads always mark
9323 * their parent and current state never has children yet. Only
9324 * explored_states can get read marks.)
9326 for (j = 0; j <= cur->curframe; j++) {
9327 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
9328 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
9329 for (i = 0; i < BPF_REG_FP; i++)
9330 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
9333 /* all stack frames are accessible from callee, clear them all */
9334 for (j = 0; j <= cur->curframe; j++) {
9335 struct bpf_func_state *frame = cur->frame[j];
9336 struct bpf_func_state *newframe = new->frame[j];
9338 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
9339 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
9340 frame->stack[i].spilled_ptr.parent =
9341 &newframe->stack[i].spilled_ptr;
9347 /* Return true if it's OK to have the same insn return a different type. */
9348 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
9353 case PTR_TO_SOCKET_OR_NULL:
9354 case PTR_TO_SOCK_COMMON:
9355 case PTR_TO_SOCK_COMMON_OR_NULL:
9356 case PTR_TO_TCP_SOCK:
9357 case PTR_TO_TCP_SOCK_OR_NULL:
9358 case PTR_TO_XDP_SOCK:
9360 case PTR_TO_BTF_ID_OR_NULL:
9367 /* If an instruction was previously used with particular pointer types, then we
9368 * need to be careful to avoid cases such as the below, where it may be ok
9369 * for one branch accessing the pointer, but not ok for the other branch:
9374 * R1 = some_other_valid_ptr;
9377 * R2 = *(u32 *)(R1 + 0);
9379 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
9381 return src != prev && (!reg_type_mismatch_ok(src) ||
9382 !reg_type_mismatch_ok(prev));
9385 static int do_check(struct bpf_verifier_env *env)
9387 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
9388 struct bpf_verifier_state *state = env->cur_state;
9389 struct bpf_insn *insns = env->prog->insnsi;
9390 struct bpf_reg_state *regs;
9391 int insn_cnt = env->prog->len;
9392 bool do_print_state = false;
9393 int prev_insn_idx = -1;
9396 struct bpf_insn *insn;
9400 env->prev_insn_idx = prev_insn_idx;
9401 if (env->insn_idx >= insn_cnt) {
9402 verbose(env, "invalid insn idx %d insn_cnt %d\n",
9403 env->insn_idx, insn_cnt);
9407 insn = &insns[env->insn_idx];
9408 class = BPF_CLASS(insn->code);
9410 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
9412 "BPF program is too large. Processed %d insn\n",
9413 env->insn_processed);
9417 err = is_state_visited(env, env->insn_idx);
9421 /* found equivalent state, can prune the search */
9422 if (env->log.level & BPF_LOG_LEVEL) {
9424 verbose(env, "\nfrom %d to %d%s: safe\n",
9425 env->prev_insn_idx, env->insn_idx,
9426 env->cur_state->speculative ?
9427 " (speculative execution)" : "");
9429 verbose(env, "%d: safe\n", env->insn_idx);
9431 goto process_bpf_exit;
9434 if (signal_pending(current))
9440 if (env->log.level & BPF_LOG_LEVEL2 ||
9441 (env->log.level & BPF_LOG_LEVEL && do_print_state)) {
9442 if (env->log.level & BPF_LOG_LEVEL2)
9443 verbose(env, "%d:", env->insn_idx);
9445 verbose(env, "\nfrom %d to %d%s:",
9446 env->prev_insn_idx, env->insn_idx,
9447 env->cur_state->speculative ?
9448 " (speculative execution)" : "");
9449 print_verifier_state(env, state->frame[state->curframe]);
9450 do_print_state = false;
9453 if (env->log.level & BPF_LOG_LEVEL) {
9454 const struct bpf_insn_cbs cbs = {
9455 .cb_print = verbose,
9456 .private_data = env,
9459 verbose_linfo(env, env->insn_idx, "; ");
9460 verbose(env, "%d: ", env->insn_idx);
9461 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
9464 if (bpf_prog_is_dev_bound(env->prog->aux)) {
9465 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
9466 env->prev_insn_idx);
9471 regs = cur_regs(env);
9472 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
9473 prev_insn_idx = env->insn_idx;
9475 if (class == BPF_ALU || class == BPF_ALU64) {
9476 err = check_alu_op(env, insn);
9480 } else if (class == BPF_LDX) {
9481 enum bpf_reg_type *prev_src_type, src_reg_type;
9483 /* check for reserved fields is already done */
9485 /* check src operand */
9486 err = check_reg_arg(env, insn->src_reg, SRC_OP);
9490 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
9494 src_reg_type = regs[insn->src_reg].type;
9496 /* check that memory (src_reg + off) is readable,
9497 * the state of dst_reg will be updated by this func
9499 err = check_mem_access(env, env->insn_idx, insn->src_reg,
9500 insn->off, BPF_SIZE(insn->code),
9501 BPF_READ, insn->dst_reg, false);
9505 prev_src_type = &env->insn_aux_data[env->insn_idx].ptr_type;
9507 if (*prev_src_type == NOT_INIT) {
9509 * dst_reg = *(u32 *)(src_reg + off)
9510 * save type to validate intersecting paths
9512 *prev_src_type = src_reg_type;
9514 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
9515 /* ABuser program is trying to use the same insn
9516 * dst_reg = *(u32*) (src_reg + off)
9517 * with different pointer types:
9518 * src_reg == ctx in one branch and
9519 * src_reg == stack|map in some other branch.
9522 verbose(env, "same insn cannot be used with different pointers\n");
9526 } else if (class == BPF_STX) {
9527 enum bpf_reg_type *prev_dst_type, dst_reg_type;
9529 if (BPF_MODE(insn->code) == BPF_XADD) {
9530 err = check_xadd(env, env->insn_idx, insn);
9537 /* check src1 operand */
9538 err = check_reg_arg(env, insn->src_reg, SRC_OP);
9541 /* check src2 operand */
9542 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
9546 dst_reg_type = regs[insn->dst_reg].type;
9548 /* check that memory (dst_reg + off) is writeable */
9549 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
9550 insn->off, BPF_SIZE(insn->code),
9551 BPF_WRITE, insn->src_reg, false);
9555 prev_dst_type = &env->insn_aux_data[env->insn_idx].ptr_type;
9557 if (*prev_dst_type == NOT_INIT) {
9558 *prev_dst_type = dst_reg_type;
9559 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
9560 verbose(env, "same insn cannot be used with different pointers\n");
9564 } else if (class == BPF_ST) {
9565 if (BPF_MODE(insn->code) != BPF_MEM ||
9566 insn->src_reg != BPF_REG_0) {
9567 verbose(env, "BPF_ST uses reserved fields\n");
9570 /* check src operand */
9571 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
9575 if (is_ctx_reg(env, insn->dst_reg)) {
9576 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
9578 reg_type_str[reg_state(env, insn->dst_reg)->type]);
9582 /* check that memory (dst_reg + off) is writeable */
9583 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
9584 insn->off, BPF_SIZE(insn->code),
9585 BPF_WRITE, -1, false);
9589 } else if (class == BPF_JMP || class == BPF_JMP32) {
9590 u8 opcode = BPF_OP(insn->code);
9592 env->jmps_processed++;
9593 if (opcode == BPF_CALL) {
9594 if (BPF_SRC(insn->code) != BPF_K ||
9596 (insn->src_reg != BPF_REG_0 &&
9597 insn->src_reg != BPF_PSEUDO_CALL) ||
9598 insn->dst_reg != BPF_REG_0 ||
9599 class == BPF_JMP32) {
9600 verbose(env, "BPF_CALL uses reserved fields\n");
9604 if (env->cur_state->active_spin_lock &&
9605 (insn->src_reg == BPF_PSEUDO_CALL ||
9606 insn->imm != BPF_FUNC_spin_unlock)) {
9607 verbose(env, "function calls are not allowed while holding a lock\n");
9610 if (insn->src_reg == BPF_PSEUDO_CALL)
9611 err = check_func_call(env, insn, &env->insn_idx);
9613 err = check_helper_call(env, insn->imm, env->insn_idx);
9617 } else if (opcode == BPF_JA) {
9618 if (BPF_SRC(insn->code) != BPF_K ||
9620 insn->src_reg != BPF_REG_0 ||
9621 insn->dst_reg != BPF_REG_0 ||
9622 class == BPF_JMP32) {
9623 verbose(env, "BPF_JA uses reserved fields\n");
9627 env->insn_idx += insn->off + 1;
9630 } else if (opcode == BPF_EXIT) {
9631 if (BPF_SRC(insn->code) != BPF_K ||
9633 insn->src_reg != BPF_REG_0 ||
9634 insn->dst_reg != BPF_REG_0 ||
9635 class == BPF_JMP32) {
9636 verbose(env, "BPF_EXIT uses reserved fields\n");
9640 if (env->cur_state->active_spin_lock) {
9641 verbose(env, "bpf_spin_unlock is missing\n");
9645 if (state->curframe) {
9646 /* exit from nested function */
9647 err = prepare_func_exit(env, &env->insn_idx);
9650 do_print_state = true;
9654 err = check_reference_leak(env);
9658 err = check_return_code(env);
9662 update_branch_counts(env, env->cur_state);
9663 err = pop_stack(env, &prev_insn_idx,
9664 &env->insn_idx, pop_log);
9670 do_print_state = true;
9674 err = check_cond_jmp_op(env, insn, &env->insn_idx);
9678 } else if (class == BPF_LD) {
9679 u8 mode = BPF_MODE(insn->code);
9681 if (mode == BPF_ABS || mode == BPF_IND) {
9682 err = check_ld_abs(env, insn);
9686 } else if (mode == BPF_IMM) {
9687 err = check_ld_imm(env, insn);
9692 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
9694 verbose(env, "invalid BPF_LD mode\n");
9698 verbose(env, "unknown insn class %d\n", class);
9708 /* replace pseudo btf_id with kernel symbol address */
9709 static int check_pseudo_btf_id(struct bpf_verifier_env *env,
9710 struct bpf_insn *insn,
9711 struct bpf_insn_aux_data *aux)
9713 const struct btf_var_secinfo *vsi;
9714 const struct btf_type *datasec;
9715 const struct btf_type *t;
9716 const char *sym_name;
9717 bool percpu = false;
9718 u32 type, id = insn->imm;
9724 verbose(env, "kernel is missing BTF, make sure CONFIG_DEBUG_INFO_BTF=y is specified in Kconfig.\n");
9728 if (insn[1].imm != 0) {
9729 verbose(env, "reserved field (insn[1].imm) is used in pseudo_btf_id ldimm64 insn.\n");
9733 t = btf_type_by_id(btf_vmlinux, id);
9735 verbose(env, "ldimm64 insn specifies invalid btf_id %d.\n", id);
9739 if (!btf_type_is_var(t)) {
9740 verbose(env, "pseudo btf_id %d in ldimm64 isn't KIND_VAR.\n",
9745 sym_name = btf_name_by_offset(btf_vmlinux, t->name_off);
9746 addr = kallsyms_lookup_name(sym_name);
9748 verbose(env, "ldimm64 failed to find the address for kernel symbol '%s'.\n",
9753 datasec_id = btf_find_by_name_kind(btf_vmlinux, ".data..percpu",
9755 if (datasec_id > 0) {
9756 datasec = btf_type_by_id(btf_vmlinux, datasec_id);
9757 for_each_vsi(i, datasec, vsi) {
9758 if (vsi->type == id) {
9765 insn[0].imm = (u32)addr;
9766 insn[1].imm = addr >> 32;
9769 t = btf_type_skip_modifiers(btf_vmlinux, type, NULL);
9771 aux->btf_var.reg_type = PTR_TO_PERCPU_BTF_ID;
9772 aux->btf_var.btf = btf_vmlinux;
9773 aux->btf_var.btf_id = type;
9774 } else if (!btf_type_is_struct(t)) {
9775 const struct btf_type *ret;
9779 /* resolve the type size of ksym. */
9780 ret = btf_resolve_size(btf_vmlinux, t, &tsize);
9782 tname = btf_name_by_offset(btf_vmlinux, t->name_off);
9783 verbose(env, "ldimm64 unable to resolve the size of type '%s': %ld\n",
9784 tname, PTR_ERR(ret));
9787 aux->btf_var.reg_type = PTR_TO_MEM;
9788 aux->btf_var.mem_size = tsize;
9790 aux->btf_var.reg_type = PTR_TO_BTF_ID;
9791 aux->btf_var.btf = btf_vmlinux;
9792 aux->btf_var.btf_id = type;
9797 static int check_map_prealloc(struct bpf_map *map)
9799 return (map->map_type != BPF_MAP_TYPE_HASH &&
9800 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
9801 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
9802 !(map->map_flags & BPF_F_NO_PREALLOC);
9805 static bool is_tracing_prog_type(enum bpf_prog_type type)
9808 case BPF_PROG_TYPE_KPROBE:
9809 case BPF_PROG_TYPE_TRACEPOINT:
9810 case BPF_PROG_TYPE_PERF_EVENT:
9811 case BPF_PROG_TYPE_RAW_TRACEPOINT:
9818 static bool is_preallocated_map(struct bpf_map *map)
9820 if (!check_map_prealloc(map))
9822 if (map->inner_map_meta && !check_map_prealloc(map->inner_map_meta))
9827 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
9828 struct bpf_map *map,
9829 struct bpf_prog *prog)
9832 enum bpf_prog_type prog_type = resolve_prog_type(prog);
9834 * Validate that trace type programs use preallocated hash maps.
9836 * For programs attached to PERF events this is mandatory as the
9837 * perf NMI can hit any arbitrary code sequence.
9839 * All other trace types using preallocated hash maps are unsafe as
9840 * well because tracepoint or kprobes can be inside locked regions
9841 * of the memory allocator or at a place where a recursion into the
9842 * memory allocator would see inconsistent state.
9844 * On RT enabled kernels run-time allocation of all trace type
9845 * programs is strictly prohibited due to lock type constraints. On
9846 * !RT kernels it is allowed for backwards compatibility reasons for
9847 * now, but warnings are emitted so developers are made aware of
9848 * the unsafety and can fix their programs before this is enforced.
9850 if (is_tracing_prog_type(prog_type) && !is_preallocated_map(map)) {
9851 if (prog_type == BPF_PROG_TYPE_PERF_EVENT) {
9852 verbose(env, "perf_event programs can only use preallocated hash map\n");
9855 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
9856 verbose(env, "trace type programs can only use preallocated hash map\n");
9859 WARN_ONCE(1, "trace type BPF program uses run-time allocation\n");
9860 verbose(env, "trace type programs with run-time allocated hash maps are unsafe. Switch to preallocated hash maps.\n");
9863 if (map_value_has_spin_lock(map)) {
9864 if (prog_type == BPF_PROG_TYPE_SOCKET_FILTER) {
9865 verbose(env, "socket filter progs cannot use bpf_spin_lock yet\n");
9869 if (is_tracing_prog_type(prog_type)) {
9870 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
9874 if (prog->aux->sleepable) {
9875 verbose(env, "sleepable progs cannot use bpf_spin_lock yet\n");
9880 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
9881 !bpf_offload_prog_map_match(prog, map)) {
9882 verbose(env, "offload device mismatch between prog and map\n");
9886 if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
9887 verbose(env, "bpf_struct_ops map cannot be used in prog\n");
9891 if (prog->aux->sleepable)
9892 switch (map->map_type) {
9893 case BPF_MAP_TYPE_HASH:
9894 case BPF_MAP_TYPE_LRU_HASH:
9895 case BPF_MAP_TYPE_ARRAY:
9896 if (!is_preallocated_map(map)) {
9898 "Sleepable programs can only use preallocated hash maps\n");
9904 "Sleepable programs can only use array and hash maps\n");
9911 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
9913 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
9914 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
9917 /* find and rewrite pseudo imm in ld_imm64 instructions:
9919 * 1. if it accesses map FD, replace it with actual map pointer.
9920 * 2. if it accesses btf_id of a VAR, replace it with pointer to the var.
9922 * NOTE: btf_vmlinux is required for converting pseudo btf_id.
9924 static int resolve_pseudo_ldimm64(struct bpf_verifier_env *env)
9926 struct bpf_insn *insn = env->prog->insnsi;
9927 int insn_cnt = env->prog->len;
9930 err = bpf_prog_calc_tag(env->prog);
9934 for (i = 0; i < insn_cnt; i++, insn++) {
9935 if (BPF_CLASS(insn->code) == BPF_LDX &&
9936 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
9937 verbose(env, "BPF_LDX uses reserved fields\n");
9941 if (BPF_CLASS(insn->code) == BPF_STX &&
9942 ((BPF_MODE(insn->code) != BPF_MEM &&
9943 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
9944 verbose(env, "BPF_STX uses reserved fields\n");
9948 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
9949 struct bpf_insn_aux_data *aux;
9950 struct bpf_map *map;
9954 if (i == insn_cnt - 1 || insn[1].code != 0 ||
9955 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
9957 verbose(env, "invalid bpf_ld_imm64 insn\n");
9961 if (insn[0].src_reg == 0)
9962 /* valid generic load 64-bit imm */
9965 if (insn[0].src_reg == BPF_PSEUDO_BTF_ID) {
9966 aux = &env->insn_aux_data[i];
9967 err = check_pseudo_btf_id(env, insn, aux);
9973 /* In final convert_pseudo_ld_imm64() step, this is
9974 * converted into regular 64-bit imm load insn.
9976 if ((insn[0].src_reg != BPF_PSEUDO_MAP_FD &&
9977 insn[0].src_reg != BPF_PSEUDO_MAP_VALUE) ||
9978 (insn[0].src_reg == BPF_PSEUDO_MAP_FD &&
9979 insn[1].imm != 0)) {
9981 "unrecognized bpf_ld_imm64 insn\n");
9985 f = fdget(insn[0].imm);
9986 map = __bpf_map_get(f);
9988 verbose(env, "fd %d is not pointing to valid bpf_map\n",
9990 return PTR_ERR(map);
9993 err = check_map_prog_compatibility(env, map, env->prog);
9999 aux = &env->insn_aux_data[i];
10000 if (insn->src_reg == BPF_PSEUDO_MAP_FD) {
10001 addr = (unsigned long)map;
10003 u32 off = insn[1].imm;
10005 if (off >= BPF_MAX_VAR_OFF) {
10006 verbose(env, "direct value offset of %u is not allowed\n", off);
10011 if (!map->ops->map_direct_value_addr) {
10012 verbose(env, "no direct value access support for this map type\n");
10017 err = map->ops->map_direct_value_addr(map, &addr, off);
10019 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
10020 map->value_size, off);
10025 aux->map_off = off;
10029 insn[0].imm = (u32)addr;
10030 insn[1].imm = addr >> 32;
10032 /* check whether we recorded this map already */
10033 for (j = 0; j < env->used_map_cnt; j++) {
10034 if (env->used_maps[j] == map) {
10035 aux->map_index = j;
10041 if (env->used_map_cnt >= MAX_USED_MAPS) {
10046 /* hold the map. If the program is rejected by verifier,
10047 * the map will be released by release_maps() or it
10048 * will be used by the valid program until it's unloaded
10049 * and all maps are released in free_used_maps()
10053 aux->map_index = env->used_map_cnt;
10054 env->used_maps[env->used_map_cnt++] = map;
10056 if (bpf_map_is_cgroup_storage(map) &&
10057 bpf_cgroup_storage_assign(env->prog->aux, map)) {
10058 verbose(env, "only one cgroup storage of each type is allowed\n");
10070 /* Basic sanity check before we invest more work here. */
10071 if (!bpf_opcode_in_insntable(insn->code)) {
10072 verbose(env, "unknown opcode %02x\n", insn->code);
10077 /* now all pseudo BPF_LD_IMM64 instructions load valid
10078 * 'struct bpf_map *' into a register instead of user map_fd.
10079 * These pointers will be used later by verifier to validate map access.
10084 /* drop refcnt of maps used by the rejected program */
10085 static void release_maps(struct bpf_verifier_env *env)
10087 __bpf_free_used_maps(env->prog->aux, env->used_maps,
10088 env->used_map_cnt);
10091 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
10092 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
10094 struct bpf_insn *insn = env->prog->insnsi;
10095 int insn_cnt = env->prog->len;
10098 for (i = 0; i < insn_cnt; i++, insn++)
10099 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
10103 /* single env->prog->insni[off] instruction was replaced with the range
10104 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
10105 * [0, off) and [off, end) to new locations, so the patched range stays zero
10107 static int adjust_insn_aux_data(struct bpf_verifier_env *env,
10108 struct bpf_prog *new_prog, u32 off, u32 cnt)
10110 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
10111 struct bpf_insn *insn = new_prog->insnsi;
10115 /* aux info at OFF always needs adjustment, no matter fast path
10116 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
10117 * original insn at old prog.
10119 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
10123 prog_len = new_prog->len;
10124 new_data = vzalloc(array_size(prog_len,
10125 sizeof(struct bpf_insn_aux_data)));
10128 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
10129 memcpy(new_data + off + cnt - 1, old_data + off,
10130 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
10131 for (i = off; i < off + cnt - 1; i++) {
10132 new_data[i].seen = env->pass_cnt;
10133 new_data[i].zext_dst = insn_has_def32(env, insn + i);
10135 env->insn_aux_data = new_data;
10140 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
10146 /* NOTE: fake 'exit' subprog should be updated as well. */
10147 for (i = 0; i <= env->subprog_cnt; i++) {
10148 if (env->subprog_info[i].start <= off)
10150 env->subprog_info[i].start += len - 1;
10154 static void adjust_poke_descs(struct bpf_prog *prog, u32 len)
10156 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
10157 int i, sz = prog->aux->size_poke_tab;
10158 struct bpf_jit_poke_descriptor *desc;
10160 for (i = 0; i < sz; i++) {
10162 desc->insn_idx += len - 1;
10166 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
10167 const struct bpf_insn *patch, u32 len)
10169 struct bpf_prog *new_prog;
10171 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
10172 if (IS_ERR(new_prog)) {
10173 if (PTR_ERR(new_prog) == -ERANGE)
10175 "insn %d cannot be patched due to 16-bit range\n",
10176 env->insn_aux_data[off].orig_idx);
10179 if (adjust_insn_aux_data(env, new_prog, off, len))
10181 adjust_subprog_starts(env, off, len);
10182 adjust_poke_descs(new_prog, len);
10186 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
10191 /* find first prog starting at or after off (first to remove) */
10192 for (i = 0; i < env->subprog_cnt; i++)
10193 if (env->subprog_info[i].start >= off)
10195 /* find first prog starting at or after off + cnt (first to stay) */
10196 for (j = i; j < env->subprog_cnt; j++)
10197 if (env->subprog_info[j].start >= off + cnt)
10199 /* if j doesn't start exactly at off + cnt, we are just removing
10200 * the front of previous prog
10202 if (env->subprog_info[j].start != off + cnt)
10206 struct bpf_prog_aux *aux = env->prog->aux;
10209 /* move fake 'exit' subprog as well */
10210 move = env->subprog_cnt + 1 - j;
10212 memmove(env->subprog_info + i,
10213 env->subprog_info + j,
10214 sizeof(*env->subprog_info) * move);
10215 env->subprog_cnt -= j - i;
10217 /* remove func_info */
10218 if (aux->func_info) {
10219 move = aux->func_info_cnt - j;
10221 memmove(aux->func_info + i,
10222 aux->func_info + j,
10223 sizeof(*aux->func_info) * move);
10224 aux->func_info_cnt -= j - i;
10225 /* func_info->insn_off is set after all code rewrites,
10226 * in adjust_btf_func() - no need to adjust
10230 /* convert i from "first prog to remove" to "first to adjust" */
10231 if (env->subprog_info[i].start == off)
10235 /* update fake 'exit' subprog as well */
10236 for (; i <= env->subprog_cnt; i++)
10237 env->subprog_info[i].start -= cnt;
10242 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
10245 struct bpf_prog *prog = env->prog;
10246 u32 i, l_off, l_cnt, nr_linfo;
10247 struct bpf_line_info *linfo;
10249 nr_linfo = prog->aux->nr_linfo;
10253 linfo = prog->aux->linfo;
10255 /* find first line info to remove, count lines to be removed */
10256 for (i = 0; i < nr_linfo; i++)
10257 if (linfo[i].insn_off >= off)
10262 for (; i < nr_linfo; i++)
10263 if (linfo[i].insn_off < off + cnt)
10268 /* First live insn doesn't match first live linfo, it needs to "inherit"
10269 * last removed linfo. prog is already modified, so prog->len == off
10270 * means no live instructions after (tail of the program was removed).
10272 if (prog->len != off && l_cnt &&
10273 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
10275 linfo[--i].insn_off = off + cnt;
10278 /* remove the line info which refer to the removed instructions */
10280 memmove(linfo + l_off, linfo + i,
10281 sizeof(*linfo) * (nr_linfo - i));
10283 prog->aux->nr_linfo -= l_cnt;
10284 nr_linfo = prog->aux->nr_linfo;
10287 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
10288 for (i = l_off; i < nr_linfo; i++)
10289 linfo[i].insn_off -= cnt;
10291 /* fix up all subprogs (incl. 'exit') which start >= off */
10292 for (i = 0; i <= env->subprog_cnt; i++)
10293 if (env->subprog_info[i].linfo_idx > l_off) {
10294 /* program may have started in the removed region but
10295 * may not be fully removed
10297 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
10298 env->subprog_info[i].linfo_idx -= l_cnt;
10300 env->subprog_info[i].linfo_idx = l_off;
10306 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
10308 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
10309 unsigned int orig_prog_len = env->prog->len;
10312 if (bpf_prog_is_dev_bound(env->prog->aux))
10313 bpf_prog_offload_remove_insns(env, off, cnt);
10315 err = bpf_remove_insns(env->prog, off, cnt);
10319 err = adjust_subprog_starts_after_remove(env, off, cnt);
10323 err = bpf_adj_linfo_after_remove(env, off, cnt);
10327 memmove(aux_data + off, aux_data + off + cnt,
10328 sizeof(*aux_data) * (orig_prog_len - off - cnt));
10333 /* The verifier does more data flow analysis than llvm and will not
10334 * explore branches that are dead at run time. Malicious programs can
10335 * have dead code too. Therefore replace all dead at-run-time code
10338 * Just nops are not optimal, e.g. if they would sit at the end of the
10339 * program and through another bug we would manage to jump there, then
10340 * we'd execute beyond program memory otherwise. Returning exception
10341 * code also wouldn't work since we can have subprogs where the dead
10342 * code could be located.
10344 static void sanitize_dead_code(struct bpf_verifier_env *env)
10346 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
10347 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
10348 struct bpf_insn *insn = env->prog->insnsi;
10349 const int insn_cnt = env->prog->len;
10352 for (i = 0; i < insn_cnt; i++) {
10353 if (aux_data[i].seen)
10355 memcpy(insn + i, &trap, sizeof(trap));
10359 static bool insn_is_cond_jump(u8 code)
10363 if (BPF_CLASS(code) == BPF_JMP32)
10366 if (BPF_CLASS(code) != BPF_JMP)
10370 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
10373 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
10375 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
10376 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
10377 struct bpf_insn *insn = env->prog->insnsi;
10378 const int insn_cnt = env->prog->len;
10381 for (i = 0; i < insn_cnt; i++, insn++) {
10382 if (!insn_is_cond_jump(insn->code))
10385 if (!aux_data[i + 1].seen)
10386 ja.off = insn->off;
10387 else if (!aux_data[i + 1 + insn->off].seen)
10392 if (bpf_prog_is_dev_bound(env->prog->aux))
10393 bpf_prog_offload_replace_insn(env, i, &ja);
10395 memcpy(insn, &ja, sizeof(ja));
10399 static int opt_remove_dead_code(struct bpf_verifier_env *env)
10401 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
10402 int insn_cnt = env->prog->len;
10405 for (i = 0; i < insn_cnt; i++) {
10409 while (i + j < insn_cnt && !aux_data[i + j].seen)
10414 err = verifier_remove_insns(env, i, j);
10417 insn_cnt = env->prog->len;
10423 static int opt_remove_nops(struct bpf_verifier_env *env)
10425 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
10426 struct bpf_insn *insn = env->prog->insnsi;
10427 int insn_cnt = env->prog->len;
10430 for (i = 0; i < insn_cnt; i++) {
10431 if (memcmp(&insn[i], &ja, sizeof(ja)))
10434 err = verifier_remove_insns(env, i, 1);
10444 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
10445 const union bpf_attr *attr)
10447 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
10448 struct bpf_insn_aux_data *aux = env->insn_aux_data;
10449 int i, patch_len, delta = 0, len = env->prog->len;
10450 struct bpf_insn *insns = env->prog->insnsi;
10451 struct bpf_prog *new_prog;
10454 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
10455 zext_patch[1] = BPF_ZEXT_REG(0);
10456 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
10457 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
10458 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
10459 for (i = 0; i < len; i++) {
10460 int adj_idx = i + delta;
10461 struct bpf_insn insn;
10463 insn = insns[adj_idx];
10464 if (!aux[adj_idx].zext_dst) {
10472 class = BPF_CLASS(code);
10473 if (insn_no_def(&insn))
10476 /* NOTE: arg "reg" (the fourth one) is only used for
10477 * BPF_STX which has been ruled out in above
10478 * check, it is safe to pass NULL here.
10480 if (is_reg64(env, &insn, insn.dst_reg, NULL, DST_OP)) {
10481 if (class == BPF_LD &&
10482 BPF_MODE(code) == BPF_IMM)
10487 /* ctx load could be transformed into wider load. */
10488 if (class == BPF_LDX &&
10489 aux[adj_idx].ptr_type == PTR_TO_CTX)
10492 imm_rnd = get_random_int();
10493 rnd_hi32_patch[0] = insn;
10494 rnd_hi32_patch[1].imm = imm_rnd;
10495 rnd_hi32_patch[3].dst_reg = insn.dst_reg;
10496 patch = rnd_hi32_patch;
10498 goto apply_patch_buffer;
10501 if (!bpf_jit_needs_zext())
10504 zext_patch[0] = insn;
10505 zext_patch[1].dst_reg = insn.dst_reg;
10506 zext_patch[1].src_reg = insn.dst_reg;
10507 patch = zext_patch;
10509 apply_patch_buffer:
10510 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
10513 env->prog = new_prog;
10514 insns = new_prog->insnsi;
10515 aux = env->insn_aux_data;
10516 delta += patch_len - 1;
10522 /* convert load instructions that access fields of a context type into a
10523 * sequence of instructions that access fields of the underlying structure:
10524 * struct __sk_buff -> struct sk_buff
10525 * struct bpf_sock_ops -> struct sock
10527 static int convert_ctx_accesses(struct bpf_verifier_env *env)
10529 const struct bpf_verifier_ops *ops = env->ops;
10530 int i, cnt, size, ctx_field_size, delta = 0;
10531 const int insn_cnt = env->prog->len;
10532 struct bpf_insn insn_buf[16], *insn;
10533 u32 target_size, size_default, off;
10534 struct bpf_prog *new_prog;
10535 enum bpf_access_type type;
10536 bool is_narrower_load;
10538 if (ops->gen_prologue || env->seen_direct_write) {
10539 if (!ops->gen_prologue) {
10540 verbose(env, "bpf verifier is misconfigured\n");
10543 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
10545 if (cnt >= ARRAY_SIZE(insn_buf)) {
10546 verbose(env, "bpf verifier is misconfigured\n");
10549 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
10553 env->prog = new_prog;
10558 if (bpf_prog_is_dev_bound(env->prog->aux))
10561 insn = env->prog->insnsi + delta;
10563 for (i = 0; i < insn_cnt; i++, insn++) {
10564 bpf_convert_ctx_access_t convert_ctx_access;
10566 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
10567 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
10568 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
10569 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
10571 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
10572 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
10573 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
10574 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
10579 if (type == BPF_WRITE &&
10580 env->insn_aux_data[i + delta].sanitize_stack_off) {
10581 struct bpf_insn patch[] = {
10582 /* Sanitize suspicious stack slot with zero.
10583 * There are no memory dependencies for this store,
10584 * since it's only using frame pointer and immediate
10587 BPF_ST_MEM(BPF_DW, BPF_REG_FP,
10588 env->insn_aux_data[i + delta].sanitize_stack_off,
10590 /* the original STX instruction will immediately
10591 * overwrite the same stack slot with appropriate value
10596 cnt = ARRAY_SIZE(patch);
10597 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
10602 env->prog = new_prog;
10603 insn = new_prog->insnsi + i + delta;
10607 switch (env->insn_aux_data[i + delta].ptr_type) {
10609 if (!ops->convert_ctx_access)
10611 convert_ctx_access = ops->convert_ctx_access;
10613 case PTR_TO_SOCKET:
10614 case PTR_TO_SOCK_COMMON:
10615 convert_ctx_access = bpf_sock_convert_ctx_access;
10617 case PTR_TO_TCP_SOCK:
10618 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
10620 case PTR_TO_XDP_SOCK:
10621 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
10623 case PTR_TO_BTF_ID:
10624 if (type == BPF_READ) {
10625 insn->code = BPF_LDX | BPF_PROBE_MEM |
10626 BPF_SIZE((insn)->code);
10627 env->prog->aux->num_exentries++;
10628 } else if (resolve_prog_type(env->prog) != BPF_PROG_TYPE_STRUCT_OPS) {
10629 verbose(env, "Writes through BTF pointers are not allowed\n");
10637 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
10638 size = BPF_LDST_BYTES(insn);
10640 /* If the read access is a narrower load of the field,
10641 * convert to a 4/8-byte load, to minimum program type specific
10642 * convert_ctx_access changes. If conversion is successful,
10643 * we will apply proper mask to the result.
10645 is_narrower_load = size < ctx_field_size;
10646 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
10648 if (is_narrower_load) {
10651 if (type == BPF_WRITE) {
10652 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
10657 if (ctx_field_size == 4)
10659 else if (ctx_field_size == 8)
10660 size_code = BPF_DW;
10662 insn->off = off & ~(size_default - 1);
10663 insn->code = BPF_LDX | BPF_MEM | size_code;
10667 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
10669 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
10670 (ctx_field_size && !target_size)) {
10671 verbose(env, "bpf verifier is misconfigured\n");
10675 if (is_narrower_load && size < target_size) {
10676 u8 shift = bpf_ctx_narrow_access_offset(
10677 off, size, size_default) * 8;
10678 if (ctx_field_size <= 4) {
10680 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
10683 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
10684 (1 << size * 8) - 1);
10687 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
10690 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
10691 (1ULL << size * 8) - 1);
10695 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
10701 /* keep walking new program and skip insns we just inserted */
10702 env->prog = new_prog;
10703 insn = new_prog->insnsi + i + delta;
10709 static int jit_subprogs(struct bpf_verifier_env *env)
10711 struct bpf_prog *prog = env->prog, **func, *tmp;
10712 int i, j, subprog_start, subprog_end = 0, len, subprog;
10713 struct bpf_map *map_ptr;
10714 struct bpf_insn *insn;
10715 void *old_bpf_func;
10716 int err, num_exentries;
10718 if (env->subprog_cnt <= 1)
10721 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
10722 if (insn->code != (BPF_JMP | BPF_CALL) ||
10723 insn->src_reg != BPF_PSEUDO_CALL)
10725 /* Upon error here we cannot fall back to interpreter but
10726 * need a hard reject of the program. Thus -EFAULT is
10727 * propagated in any case.
10729 subprog = find_subprog(env, i + insn->imm + 1);
10731 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
10732 i + insn->imm + 1);
10735 /* temporarily remember subprog id inside insn instead of
10736 * aux_data, since next loop will split up all insns into funcs
10738 insn->off = subprog;
10739 /* remember original imm in case JIT fails and fallback
10740 * to interpreter will be needed
10742 env->insn_aux_data[i].call_imm = insn->imm;
10743 /* point imm to __bpf_call_base+1 from JITs point of view */
10747 err = bpf_prog_alloc_jited_linfo(prog);
10749 goto out_undo_insn;
10752 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
10754 goto out_undo_insn;
10756 for (i = 0; i < env->subprog_cnt; i++) {
10757 subprog_start = subprog_end;
10758 subprog_end = env->subprog_info[i + 1].start;
10760 len = subprog_end - subprog_start;
10761 /* BPF_PROG_RUN doesn't call subprogs directly,
10762 * hence main prog stats include the runtime of subprogs.
10763 * subprogs don't have IDs and not reachable via prog_get_next_id
10764 * func[i]->aux->stats will never be accessed and stays NULL
10766 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
10769 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
10770 len * sizeof(struct bpf_insn));
10771 func[i]->type = prog->type;
10772 func[i]->len = len;
10773 if (bpf_prog_calc_tag(func[i]))
10775 func[i]->is_func = 1;
10776 func[i]->aux->func_idx = i;
10777 /* the btf and func_info will be freed only at prog->aux */
10778 func[i]->aux->btf = prog->aux->btf;
10779 func[i]->aux->func_info = prog->aux->func_info;
10781 for (j = 0; j < prog->aux->size_poke_tab; j++) {
10782 u32 insn_idx = prog->aux->poke_tab[j].insn_idx;
10785 if (!(insn_idx >= subprog_start &&
10786 insn_idx <= subprog_end))
10789 ret = bpf_jit_add_poke_descriptor(func[i],
10790 &prog->aux->poke_tab[j]);
10792 verbose(env, "adding tail call poke descriptor failed\n");
10796 func[i]->insnsi[insn_idx - subprog_start].imm = ret + 1;
10798 map_ptr = func[i]->aux->poke_tab[ret].tail_call.map;
10799 ret = map_ptr->ops->map_poke_track(map_ptr, func[i]->aux);
10801 verbose(env, "tracking tail call prog failed\n");
10806 /* Use bpf_prog_F_tag to indicate functions in stack traces.
10807 * Long term would need debug info to populate names
10809 func[i]->aux->name[0] = 'F';
10810 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
10811 func[i]->jit_requested = 1;
10812 func[i]->aux->linfo = prog->aux->linfo;
10813 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
10814 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
10815 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
10817 insn = func[i]->insnsi;
10818 for (j = 0; j < func[i]->len; j++, insn++) {
10819 if (BPF_CLASS(insn->code) == BPF_LDX &&
10820 BPF_MODE(insn->code) == BPF_PROBE_MEM)
10823 func[i]->aux->num_exentries = num_exentries;
10824 func[i]->aux->tail_call_reachable = env->subprog_info[i].tail_call_reachable;
10825 func[i] = bpf_int_jit_compile(func[i]);
10826 if (!func[i]->jited) {
10833 /* Untrack main program's aux structs so that during map_poke_run()
10834 * we will not stumble upon the unfilled poke descriptors; each
10835 * of the main program's poke descs got distributed across subprogs
10836 * and got tracked onto map, so we are sure that none of them will
10837 * be missed after the operation below
10839 for (i = 0; i < prog->aux->size_poke_tab; i++) {
10840 map_ptr = prog->aux->poke_tab[i].tail_call.map;
10842 map_ptr->ops->map_poke_untrack(map_ptr, prog->aux);
10845 /* at this point all bpf functions were successfully JITed
10846 * now populate all bpf_calls with correct addresses and
10847 * run last pass of JIT
10849 for (i = 0; i < env->subprog_cnt; i++) {
10850 insn = func[i]->insnsi;
10851 for (j = 0; j < func[i]->len; j++, insn++) {
10852 if (insn->code != (BPF_JMP | BPF_CALL) ||
10853 insn->src_reg != BPF_PSEUDO_CALL)
10855 subprog = insn->off;
10856 insn->imm = BPF_CAST_CALL(func[subprog]->bpf_func) -
10860 /* we use the aux data to keep a list of the start addresses
10861 * of the JITed images for each function in the program
10863 * for some architectures, such as powerpc64, the imm field
10864 * might not be large enough to hold the offset of the start
10865 * address of the callee's JITed image from __bpf_call_base
10867 * in such cases, we can lookup the start address of a callee
10868 * by using its subprog id, available from the off field of
10869 * the call instruction, as an index for this list
10871 func[i]->aux->func = func;
10872 func[i]->aux->func_cnt = env->subprog_cnt;
10874 for (i = 0; i < env->subprog_cnt; i++) {
10875 old_bpf_func = func[i]->bpf_func;
10876 tmp = bpf_int_jit_compile(func[i]);
10877 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
10878 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
10885 /* finally lock prog and jit images for all functions and
10886 * populate kallsysm
10888 for (i = 0; i < env->subprog_cnt; i++) {
10889 bpf_prog_lock_ro(func[i]);
10890 bpf_prog_kallsyms_add(func[i]);
10893 /* Last step: make now unused interpreter insns from main
10894 * prog consistent for later dump requests, so they can
10895 * later look the same as if they were interpreted only.
10897 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
10898 if (insn->code != (BPF_JMP | BPF_CALL) ||
10899 insn->src_reg != BPF_PSEUDO_CALL)
10901 insn->off = env->insn_aux_data[i].call_imm;
10902 subprog = find_subprog(env, i + insn->off + 1);
10903 insn->imm = subprog;
10907 prog->bpf_func = func[0]->bpf_func;
10908 prog->aux->func = func;
10909 prog->aux->func_cnt = env->subprog_cnt;
10910 bpf_prog_free_unused_jited_linfo(prog);
10913 for (i = 0; i < env->subprog_cnt; i++) {
10917 for (j = 0; j < func[i]->aux->size_poke_tab; j++) {
10918 map_ptr = func[i]->aux->poke_tab[j].tail_call.map;
10919 map_ptr->ops->map_poke_untrack(map_ptr, func[i]->aux);
10921 bpf_jit_free(func[i]);
10925 /* cleanup main prog to be interpreted */
10926 prog->jit_requested = 0;
10927 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
10928 if (insn->code != (BPF_JMP | BPF_CALL) ||
10929 insn->src_reg != BPF_PSEUDO_CALL)
10932 insn->imm = env->insn_aux_data[i].call_imm;
10934 bpf_prog_free_jited_linfo(prog);
10938 static int fixup_call_args(struct bpf_verifier_env *env)
10940 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
10941 struct bpf_prog *prog = env->prog;
10942 struct bpf_insn *insn = prog->insnsi;
10947 if (env->prog->jit_requested &&
10948 !bpf_prog_is_dev_bound(env->prog->aux)) {
10949 err = jit_subprogs(env);
10952 if (err == -EFAULT)
10955 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
10956 if (env->subprog_cnt > 1 && env->prog->aux->tail_call_reachable) {
10957 /* When JIT fails the progs with bpf2bpf calls and tail_calls
10958 * have to be rejected, since interpreter doesn't support them yet.
10960 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
10963 for (i = 0; i < prog->len; i++, insn++) {
10964 if (insn->code != (BPF_JMP | BPF_CALL) ||
10965 insn->src_reg != BPF_PSEUDO_CALL)
10967 depth = get_callee_stack_depth(env, insn, i);
10970 bpf_patch_call_args(insn, depth);
10977 /* fixup insn->imm field of bpf_call instructions
10978 * and inline eligible helpers as explicit sequence of BPF instructions
10980 * this function is called after eBPF program passed verification
10982 static int fixup_bpf_calls(struct bpf_verifier_env *env)
10984 struct bpf_prog *prog = env->prog;
10985 bool expect_blinding = bpf_jit_blinding_enabled(prog);
10986 struct bpf_insn *insn = prog->insnsi;
10987 const struct bpf_func_proto *fn;
10988 const int insn_cnt = prog->len;
10989 const struct bpf_map_ops *ops;
10990 struct bpf_insn_aux_data *aux;
10991 struct bpf_insn insn_buf[16];
10992 struct bpf_prog *new_prog;
10993 struct bpf_map *map_ptr;
10994 int i, ret, cnt, delta = 0;
10996 for (i = 0; i < insn_cnt; i++, insn++) {
10997 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
10998 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
10999 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
11000 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
11001 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
11002 struct bpf_insn mask_and_div[] = {
11003 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
11004 /* Rx div 0 -> 0 */
11005 BPF_JMP_IMM(BPF_JNE, insn->src_reg, 0, 2),
11006 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
11007 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
11010 struct bpf_insn mask_and_mod[] = {
11011 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
11012 /* Rx mod 0 -> Rx */
11013 BPF_JMP_IMM(BPF_JEQ, insn->src_reg, 0, 1),
11016 struct bpf_insn *patchlet;
11018 if (insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
11019 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
11020 patchlet = mask_and_div + (is64 ? 1 : 0);
11021 cnt = ARRAY_SIZE(mask_and_div) - (is64 ? 1 : 0);
11023 patchlet = mask_and_mod + (is64 ? 1 : 0);
11024 cnt = ARRAY_SIZE(mask_and_mod) - (is64 ? 1 : 0);
11027 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
11032 env->prog = prog = new_prog;
11033 insn = new_prog->insnsi + i + delta;
11037 if (BPF_CLASS(insn->code) == BPF_LD &&
11038 (BPF_MODE(insn->code) == BPF_ABS ||
11039 BPF_MODE(insn->code) == BPF_IND)) {
11040 cnt = env->ops->gen_ld_abs(insn, insn_buf);
11041 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
11042 verbose(env, "bpf verifier is misconfigured\n");
11046 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
11051 env->prog = prog = new_prog;
11052 insn = new_prog->insnsi + i + delta;
11056 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
11057 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
11058 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
11059 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
11060 struct bpf_insn insn_buf[16];
11061 struct bpf_insn *patch = &insn_buf[0];
11065 aux = &env->insn_aux_data[i + delta];
11066 if (!aux->alu_state ||
11067 aux->alu_state == BPF_ALU_NON_POINTER)
11070 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
11071 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
11072 BPF_ALU_SANITIZE_SRC;
11074 off_reg = issrc ? insn->src_reg : insn->dst_reg;
11076 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
11077 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit - 1);
11078 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
11079 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
11080 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
11081 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
11083 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX,
11085 insn->src_reg = BPF_REG_AX;
11087 *patch++ = BPF_ALU64_REG(BPF_AND, off_reg,
11091 insn->code = insn->code == code_add ?
11092 code_sub : code_add;
11094 if (issrc && isneg)
11095 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
11096 cnt = patch - insn_buf;
11098 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
11103 env->prog = prog = new_prog;
11104 insn = new_prog->insnsi + i + delta;
11108 if (insn->code != (BPF_JMP | BPF_CALL))
11110 if (insn->src_reg == BPF_PSEUDO_CALL)
11113 if (insn->imm == BPF_FUNC_get_route_realm)
11114 prog->dst_needed = 1;
11115 if (insn->imm == BPF_FUNC_get_prandom_u32)
11116 bpf_user_rnd_init_once();
11117 if (insn->imm == BPF_FUNC_override_return)
11118 prog->kprobe_override = 1;
11119 if (insn->imm == BPF_FUNC_tail_call) {
11120 /* If we tail call into other programs, we
11121 * cannot make any assumptions since they can
11122 * be replaced dynamically during runtime in
11123 * the program array.
11125 prog->cb_access = 1;
11126 if (!allow_tail_call_in_subprogs(env))
11127 prog->aux->stack_depth = MAX_BPF_STACK;
11128 prog->aux->max_pkt_offset = MAX_PACKET_OFF;
11130 /* mark bpf_tail_call as different opcode to avoid
11131 * conditional branch in the interpeter for every normal
11132 * call and to prevent accidental JITing by JIT compiler
11133 * that doesn't support bpf_tail_call yet
11136 insn->code = BPF_JMP | BPF_TAIL_CALL;
11138 aux = &env->insn_aux_data[i + delta];
11139 if (env->bpf_capable && !expect_blinding &&
11140 prog->jit_requested &&
11141 !bpf_map_key_poisoned(aux) &&
11142 !bpf_map_ptr_poisoned(aux) &&
11143 !bpf_map_ptr_unpriv(aux)) {
11144 struct bpf_jit_poke_descriptor desc = {
11145 .reason = BPF_POKE_REASON_TAIL_CALL,
11146 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
11147 .tail_call.key = bpf_map_key_immediate(aux),
11148 .insn_idx = i + delta,
11151 ret = bpf_jit_add_poke_descriptor(prog, &desc);
11153 verbose(env, "adding tail call poke descriptor failed\n");
11157 insn->imm = ret + 1;
11161 if (!bpf_map_ptr_unpriv(aux))
11164 /* instead of changing every JIT dealing with tail_call
11165 * emit two extra insns:
11166 * if (index >= max_entries) goto out;
11167 * index &= array->index_mask;
11168 * to avoid out-of-bounds cpu speculation
11170 if (bpf_map_ptr_poisoned(aux)) {
11171 verbose(env, "tail_call abusing map_ptr\n");
11175 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
11176 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
11177 map_ptr->max_entries, 2);
11178 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
11179 container_of(map_ptr,
11182 insn_buf[2] = *insn;
11184 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
11189 env->prog = prog = new_prog;
11190 insn = new_prog->insnsi + i + delta;
11194 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
11195 * and other inlining handlers are currently limited to 64 bit
11198 if (prog->jit_requested && BITS_PER_LONG == 64 &&
11199 (insn->imm == BPF_FUNC_map_lookup_elem ||
11200 insn->imm == BPF_FUNC_map_update_elem ||
11201 insn->imm == BPF_FUNC_map_delete_elem ||
11202 insn->imm == BPF_FUNC_map_push_elem ||
11203 insn->imm == BPF_FUNC_map_pop_elem ||
11204 insn->imm == BPF_FUNC_map_peek_elem)) {
11205 aux = &env->insn_aux_data[i + delta];
11206 if (bpf_map_ptr_poisoned(aux))
11207 goto patch_call_imm;
11209 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
11210 ops = map_ptr->ops;
11211 if (insn->imm == BPF_FUNC_map_lookup_elem &&
11212 ops->map_gen_lookup) {
11213 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
11214 if (cnt == -EOPNOTSUPP)
11215 goto patch_map_ops_generic;
11216 if (cnt <= 0 || cnt >= ARRAY_SIZE(insn_buf)) {
11217 verbose(env, "bpf verifier is misconfigured\n");
11221 new_prog = bpf_patch_insn_data(env, i + delta,
11227 env->prog = prog = new_prog;
11228 insn = new_prog->insnsi + i + delta;
11232 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
11233 (void *(*)(struct bpf_map *map, void *key))NULL));
11234 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
11235 (int (*)(struct bpf_map *map, void *key))NULL));
11236 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
11237 (int (*)(struct bpf_map *map, void *key, void *value,
11239 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
11240 (int (*)(struct bpf_map *map, void *value,
11242 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
11243 (int (*)(struct bpf_map *map, void *value))NULL));
11244 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
11245 (int (*)(struct bpf_map *map, void *value))NULL));
11246 patch_map_ops_generic:
11247 switch (insn->imm) {
11248 case BPF_FUNC_map_lookup_elem:
11249 insn->imm = BPF_CAST_CALL(ops->map_lookup_elem) -
11252 case BPF_FUNC_map_update_elem:
11253 insn->imm = BPF_CAST_CALL(ops->map_update_elem) -
11256 case BPF_FUNC_map_delete_elem:
11257 insn->imm = BPF_CAST_CALL(ops->map_delete_elem) -
11260 case BPF_FUNC_map_push_elem:
11261 insn->imm = BPF_CAST_CALL(ops->map_push_elem) -
11264 case BPF_FUNC_map_pop_elem:
11265 insn->imm = BPF_CAST_CALL(ops->map_pop_elem) -
11268 case BPF_FUNC_map_peek_elem:
11269 insn->imm = BPF_CAST_CALL(ops->map_peek_elem) -
11274 goto patch_call_imm;
11277 if (prog->jit_requested && BITS_PER_LONG == 64 &&
11278 insn->imm == BPF_FUNC_jiffies64) {
11279 struct bpf_insn ld_jiffies_addr[2] = {
11280 BPF_LD_IMM64(BPF_REG_0,
11281 (unsigned long)&jiffies),
11284 insn_buf[0] = ld_jiffies_addr[0];
11285 insn_buf[1] = ld_jiffies_addr[1];
11286 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
11290 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
11296 env->prog = prog = new_prog;
11297 insn = new_prog->insnsi + i + delta;
11302 fn = env->ops->get_func_proto(insn->imm, env->prog);
11303 /* all functions that have prototype and verifier allowed
11304 * programs to call them, must be real in-kernel functions
11308 "kernel subsystem misconfigured func %s#%d\n",
11309 func_id_name(insn->imm), insn->imm);
11312 insn->imm = fn->func - __bpf_call_base;
11315 /* Since poke tab is now finalized, publish aux to tracker. */
11316 for (i = 0; i < prog->aux->size_poke_tab; i++) {
11317 map_ptr = prog->aux->poke_tab[i].tail_call.map;
11318 if (!map_ptr->ops->map_poke_track ||
11319 !map_ptr->ops->map_poke_untrack ||
11320 !map_ptr->ops->map_poke_run) {
11321 verbose(env, "bpf verifier is misconfigured\n");
11325 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
11327 verbose(env, "tracking tail call prog failed\n");
11335 static void free_states(struct bpf_verifier_env *env)
11337 struct bpf_verifier_state_list *sl, *sln;
11340 sl = env->free_list;
11343 free_verifier_state(&sl->state, false);
11347 env->free_list = NULL;
11349 if (!env->explored_states)
11352 for (i = 0; i < state_htab_size(env); i++) {
11353 sl = env->explored_states[i];
11357 free_verifier_state(&sl->state, false);
11361 env->explored_states[i] = NULL;
11365 /* The verifier is using insn_aux_data[] to store temporary data during
11366 * verification and to store information for passes that run after the
11367 * verification like dead code sanitization. do_check_common() for subprogram N
11368 * may analyze many other subprograms. sanitize_insn_aux_data() clears all
11369 * temporary data after do_check_common() finds that subprogram N cannot be
11370 * verified independently. pass_cnt counts the number of times
11371 * do_check_common() was run and insn->aux->seen tells the pass number
11372 * insn_aux_data was touched. These variables are compared to clear temporary
11373 * data from failed pass. For testing and experiments do_check_common() can be
11374 * run multiple times even when prior attempt to verify is unsuccessful.
11376 static void sanitize_insn_aux_data(struct bpf_verifier_env *env)
11378 struct bpf_insn *insn = env->prog->insnsi;
11379 struct bpf_insn_aux_data *aux;
11382 for (i = 0; i < env->prog->len; i++) {
11383 class = BPF_CLASS(insn[i].code);
11384 if (class != BPF_LDX && class != BPF_STX)
11386 aux = &env->insn_aux_data[i];
11387 if (aux->seen != env->pass_cnt)
11389 memset(aux, 0, offsetof(typeof(*aux), orig_idx));
11393 static int do_check_common(struct bpf_verifier_env *env, int subprog)
11395 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
11396 struct bpf_verifier_state *state;
11397 struct bpf_reg_state *regs;
11400 env->prev_linfo = NULL;
11403 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
11406 state->curframe = 0;
11407 state->speculative = false;
11408 state->branches = 1;
11409 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
11410 if (!state->frame[0]) {
11414 env->cur_state = state;
11415 init_func_state(env, state->frame[0],
11416 BPF_MAIN_FUNC /* callsite */,
11420 regs = state->frame[state->curframe]->regs;
11421 if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
11422 ret = btf_prepare_func_args(env, subprog, regs);
11425 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
11426 if (regs[i].type == PTR_TO_CTX)
11427 mark_reg_known_zero(env, regs, i);
11428 else if (regs[i].type == SCALAR_VALUE)
11429 mark_reg_unknown(env, regs, i);
11432 /* 1st arg to a function */
11433 regs[BPF_REG_1].type = PTR_TO_CTX;
11434 mark_reg_known_zero(env, regs, BPF_REG_1);
11435 ret = btf_check_func_arg_match(env, subprog, regs);
11436 if (ret == -EFAULT)
11437 /* unlikely verifier bug. abort.
11438 * ret == 0 and ret < 0 are sadly acceptable for
11439 * main() function due to backward compatibility.
11440 * Like socket filter program may be written as:
11441 * int bpf_prog(struct pt_regs *ctx)
11442 * and never dereference that ctx in the program.
11443 * 'struct pt_regs' is a type mismatch for socket
11444 * filter that should be using 'struct __sk_buff'.
11449 ret = do_check(env);
11451 /* check for NULL is necessary, since cur_state can be freed inside
11452 * do_check() under memory pressure.
11454 if (env->cur_state) {
11455 free_verifier_state(env->cur_state, true);
11456 env->cur_state = NULL;
11458 while (!pop_stack(env, NULL, NULL, false));
11459 if (!ret && pop_log)
11460 bpf_vlog_reset(&env->log, 0);
11463 /* clean aux data in case subprog was rejected */
11464 sanitize_insn_aux_data(env);
11468 /* Verify all global functions in a BPF program one by one based on their BTF.
11469 * All global functions must pass verification. Otherwise the whole program is rejected.
11480 * foo() will be verified first for R1=any_scalar_value. During verification it
11481 * will be assumed that bar() already verified successfully and call to bar()
11482 * from foo() will be checked for type match only. Later bar() will be verified
11483 * independently to check that it's safe for R1=any_scalar_value.
11485 static int do_check_subprogs(struct bpf_verifier_env *env)
11487 struct bpf_prog_aux *aux = env->prog->aux;
11490 if (!aux->func_info)
11493 for (i = 1; i < env->subprog_cnt; i++) {
11494 if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL)
11496 env->insn_idx = env->subprog_info[i].start;
11497 WARN_ON_ONCE(env->insn_idx == 0);
11498 ret = do_check_common(env, i);
11501 } else if (env->log.level & BPF_LOG_LEVEL) {
11503 "Func#%d is safe for any args that match its prototype\n",
11510 static int do_check_main(struct bpf_verifier_env *env)
11515 ret = do_check_common(env, 0);
11517 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
11522 static void print_verification_stats(struct bpf_verifier_env *env)
11526 if (env->log.level & BPF_LOG_STATS) {
11527 verbose(env, "verification time %lld usec\n",
11528 div_u64(env->verification_time, 1000));
11529 verbose(env, "stack depth ");
11530 for (i = 0; i < env->subprog_cnt; i++) {
11531 u32 depth = env->subprog_info[i].stack_depth;
11533 verbose(env, "%d", depth);
11534 if (i + 1 < env->subprog_cnt)
11537 verbose(env, "\n");
11539 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
11540 "total_states %d peak_states %d mark_read %d\n",
11541 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
11542 env->max_states_per_insn, env->total_states,
11543 env->peak_states, env->longest_mark_read_walk);
11546 static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
11548 const struct btf_type *t, *func_proto;
11549 const struct bpf_struct_ops *st_ops;
11550 const struct btf_member *member;
11551 struct bpf_prog *prog = env->prog;
11552 u32 btf_id, member_idx;
11555 btf_id = prog->aux->attach_btf_id;
11556 st_ops = bpf_struct_ops_find(btf_id);
11558 verbose(env, "attach_btf_id %u is not a supported struct\n",
11564 member_idx = prog->expected_attach_type;
11565 if (member_idx >= btf_type_vlen(t)) {
11566 verbose(env, "attach to invalid member idx %u of struct %s\n",
11567 member_idx, st_ops->name);
11571 member = &btf_type_member(t)[member_idx];
11572 mname = btf_name_by_offset(btf_vmlinux, member->name_off);
11573 func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type,
11576 verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
11577 mname, member_idx, st_ops->name);
11581 if (st_ops->check_member) {
11582 int err = st_ops->check_member(t, member);
11585 verbose(env, "attach to unsupported member %s of struct %s\n",
11586 mname, st_ops->name);
11591 prog->aux->attach_func_proto = func_proto;
11592 prog->aux->attach_func_name = mname;
11593 env->ops = st_ops->verifier_ops;
11597 #define SECURITY_PREFIX "security_"
11599 static int check_attach_modify_return(unsigned long addr, const char *func_name)
11601 if (within_error_injection_list(addr) ||
11602 !strncmp(SECURITY_PREFIX, func_name, sizeof(SECURITY_PREFIX) - 1))
11608 /* list of non-sleepable functions that are otherwise on
11609 * ALLOW_ERROR_INJECTION list
11611 BTF_SET_START(btf_non_sleepable_error_inject)
11612 /* Three functions below can be called from sleepable and non-sleepable context.
11613 * Assume non-sleepable from bpf safety point of view.
11615 BTF_ID(func, __add_to_page_cache_locked)
11616 BTF_ID(func, should_fail_alloc_page)
11617 BTF_ID(func, should_failslab)
11618 BTF_SET_END(btf_non_sleepable_error_inject)
11620 static int check_non_sleepable_error_inject(u32 btf_id)
11622 return btf_id_set_contains(&btf_non_sleepable_error_inject, btf_id);
11625 int bpf_check_attach_target(struct bpf_verifier_log *log,
11626 const struct bpf_prog *prog,
11627 const struct bpf_prog *tgt_prog,
11629 struct bpf_attach_target_info *tgt_info)
11631 bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
11632 const char prefix[] = "btf_trace_";
11633 int ret = 0, subprog = -1, i;
11634 const struct btf_type *t;
11635 bool conservative = true;
11641 bpf_log(log, "Tracing programs must provide btf_id\n");
11644 btf = tgt_prog ? tgt_prog->aux->btf : prog->aux->attach_btf;
11647 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
11650 t = btf_type_by_id(btf, btf_id);
11652 bpf_log(log, "attach_btf_id %u is invalid\n", btf_id);
11655 tname = btf_name_by_offset(btf, t->name_off);
11657 bpf_log(log, "attach_btf_id %u doesn't have a name\n", btf_id);
11661 struct bpf_prog_aux *aux = tgt_prog->aux;
11663 for (i = 0; i < aux->func_info_cnt; i++)
11664 if (aux->func_info[i].type_id == btf_id) {
11668 if (subprog == -1) {
11669 bpf_log(log, "Subprog %s doesn't exist\n", tname);
11672 conservative = aux->func_info_aux[subprog].unreliable;
11673 if (prog_extension) {
11674 if (conservative) {
11676 "Cannot replace static functions\n");
11679 if (!prog->jit_requested) {
11681 "Extension programs should be JITed\n");
11685 if (!tgt_prog->jited) {
11686 bpf_log(log, "Can attach to only JITed progs\n");
11689 if (tgt_prog->type == prog->type) {
11690 /* Cannot fentry/fexit another fentry/fexit program.
11691 * Cannot attach program extension to another extension.
11692 * It's ok to attach fentry/fexit to extension program.
11694 bpf_log(log, "Cannot recursively attach\n");
11697 if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
11699 (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
11700 tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
11701 /* Program extensions can extend all program types
11702 * except fentry/fexit. The reason is the following.
11703 * The fentry/fexit programs are used for performance
11704 * analysis, stats and can be attached to any program
11705 * type except themselves. When extension program is
11706 * replacing XDP function it is necessary to allow
11707 * performance analysis of all functions. Both original
11708 * XDP program and its program extension. Hence
11709 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
11710 * allowed. If extending of fentry/fexit was allowed it
11711 * would be possible to create long call chain
11712 * fentry->extension->fentry->extension beyond
11713 * reasonable stack size. Hence extending fentry is not
11716 bpf_log(log, "Cannot extend fentry/fexit\n");
11720 if (prog_extension) {
11721 bpf_log(log, "Cannot replace kernel functions\n");
11726 switch (prog->expected_attach_type) {
11727 case BPF_TRACE_RAW_TP:
11730 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
11733 if (!btf_type_is_typedef(t)) {
11734 bpf_log(log, "attach_btf_id %u is not a typedef\n",
11738 if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
11739 bpf_log(log, "attach_btf_id %u points to wrong type name %s\n",
11743 tname += sizeof(prefix) - 1;
11744 t = btf_type_by_id(btf, t->type);
11745 if (!btf_type_is_ptr(t))
11746 /* should never happen in valid vmlinux build */
11748 t = btf_type_by_id(btf, t->type);
11749 if (!btf_type_is_func_proto(t))
11750 /* should never happen in valid vmlinux build */
11754 case BPF_TRACE_ITER:
11755 if (!btf_type_is_func(t)) {
11756 bpf_log(log, "attach_btf_id %u is not a function\n",
11760 t = btf_type_by_id(btf, t->type);
11761 if (!btf_type_is_func_proto(t))
11763 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
11768 if (!prog_extension)
11771 case BPF_MODIFY_RETURN:
11773 case BPF_TRACE_FENTRY:
11774 case BPF_TRACE_FEXIT:
11775 if (!btf_type_is_func(t)) {
11776 bpf_log(log, "attach_btf_id %u is not a function\n",
11780 if (prog_extension &&
11781 btf_check_type_match(log, prog, btf, t))
11783 t = btf_type_by_id(btf, t->type);
11784 if (!btf_type_is_func_proto(t))
11787 if ((prog->aux->saved_dst_prog_type || prog->aux->saved_dst_attach_type) &&
11788 (!tgt_prog || prog->aux->saved_dst_prog_type != tgt_prog->type ||
11789 prog->aux->saved_dst_attach_type != tgt_prog->expected_attach_type))
11792 if (tgt_prog && conservative)
11795 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
11801 addr = (long) tgt_prog->bpf_func;
11803 addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
11805 addr = kallsyms_lookup_name(tname);
11808 "The address of function %s cannot be found\n",
11814 if (prog->aux->sleepable) {
11816 switch (prog->type) {
11817 case BPF_PROG_TYPE_TRACING:
11818 /* fentry/fexit/fmod_ret progs can be sleepable only if they are
11819 * attached to ALLOW_ERROR_INJECTION and are not in denylist.
11821 if (!check_non_sleepable_error_inject(btf_id) &&
11822 within_error_injection_list(addr))
11825 case BPF_PROG_TYPE_LSM:
11826 /* LSM progs check that they are attached to bpf_lsm_*() funcs.
11827 * Only some of them are sleepable.
11829 if (bpf_lsm_is_sleepable_hook(btf_id))
11836 bpf_log(log, "%s is not sleepable\n", tname);
11839 } else if (prog->expected_attach_type == BPF_MODIFY_RETURN) {
11841 bpf_log(log, "can't modify return codes of BPF programs\n");
11844 ret = check_attach_modify_return(addr, tname);
11846 bpf_log(log, "%s() is not modifiable\n", tname);
11853 tgt_info->tgt_addr = addr;
11854 tgt_info->tgt_name = tname;
11855 tgt_info->tgt_type = t;
11859 static int check_attach_btf_id(struct bpf_verifier_env *env)
11861 struct bpf_prog *prog = env->prog;
11862 struct bpf_prog *tgt_prog = prog->aux->dst_prog;
11863 struct bpf_attach_target_info tgt_info = {};
11864 u32 btf_id = prog->aux->attach_btf_id;
11865 struct bpf_trampoline *tr;
11869 if (prog->aux->sleepable && prog->type != BPF_PROG_TYPE_TRACING &&
11870 prog->type != BPF_PROG_TYPE_LSM) {
11871 verbose(env, "Only fentry/fexit/fmod_ret and lsm programs can be sleepable\n");
11875 if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
11876 return check_struct_ops_btf_id(env);
11878 if (prog->type != BPF_PROG_TYPE_TRACING &&
11879 prog->type != BPF_PROG_TYPE_LSM &&
11880 prog->type != BPF_PROG_TYPE_EXT)
11883 ret = bpf_check_attach_target(&env->log, prog, tgt_prog, btf_id, &tgt_info);
11887 if (tgt_prog && prog->type == BPF_PROG_TYPE_EXT) {
11888 /* to make freplace equivalent to their targets, they need to
11889 * inherit env->ops and expected_attach_type for the rest of the
11892 env->ops = bpf_verifier_ops[tgt_prog->type];
11893 prog->expected_attach_type = tgt_prog->expected_attach_type;
11896 /* store info about the attachment target that will be used later */
11897 prog->aux->attach_func_proto = tgt_info.tgt_type;
11898 prog->aux->attach_func_name = tgt_info.tgt_name;
11901 prog->aux->saved_dst_prog_type = tgt_prog->type;
11902 prog->aux->saved_dst_attach_type = tgt_prog->expected_attach_type;
11905 if (prog->expected_attach_type == BPF_TRACE_RAW_TP) {
11906 prog->aux->attach_btf_trace = true;
11908 } else if (prog->expected_attach_type == BPF_TRACE_ITER) {
11909 if (!bpf_iter_prog_supported(prog))
11914 if (prog->type == BPF_PROG_TYPE_LSM) {
11915 ret = bpf_lsm_verify_prog(&env->log, prog);
11920 key = bpf_trampoline_compute_key(tgt_prog, prog->aux->attach_btf, btf_id);
11921 tr = bpf_trampoline_get(key, &tgt_info);
11925 prog->aux->dst_trampoline = tr;
11929 struct btf *bpf_get_btf_vmlinux(void)
11931 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
11932 mutex_lock(&bpf_verifier_lock);
11934 btf_vmlinux = btf_parse_vmlinux();
11935 mutex_unlock(&bpf_verifier_lock);
11937 return btf_vmlinux;
11940 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr,
11941 union bpf_attr __user *uattr)
11943 u64 start_time = ktime_get_ns();
11944 struct bpf_verifier_env *env;
11945 struct bpf_verifier_log *log;
11946 int i, len, ret = -EINVAL;
11949 /* no program is valid */
11950 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
11953 /* 'struct bpf_verifier_env' can be global, but since it's not small,
11954 * allocate/free it every time bpf_check() is called
11956 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
11961 len = (*prog)->len;
11962 env->insn_aux_data =
11963 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
11965 if (!env->insn_aux_data)
11967 for (i = 0; i < len; i++)
11968 env->insn_aux_data[i].orig_idx = i;
11970 env->ops = bpf_verifier_ops[env->prog->type];
11971 is_priv = bpf_capable();
11973 bpf_get_btf_vmlinux();
11975 /* grab the mutex to protect few globals used by verifier */
11977 mutex_lock(&bpf_verifier_lock);
11979 if (attr->log_level || attr->log_buf || attr->log_size) {
11980 /* user requested verbose verifier output
11981 * and supplied buffer to store the verification trace
11983 log->level = attr->log_level;
11984 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
11985 log->len_total = attr->log_size;
11988 /* log attributes have to be sane */
11989 if (log->len_total < 128 || log->len_total > UINT_MAX >> 2 ||
11990 !log->level || !log->ubuf || log->level & ~BPF_LOG_MASK)
11994 if (IS_ERR(btf_vmlinux)) {
11995 /* Either gcc or pahole or kernel are broken. */
11996 verbose(env, "in-kernel BTF is malformed\n");
11997 ret = PTR_ERR(btf_vmlinux);
11998 goto skip_full_check;
12001 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
12002 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
12003 env->strict_alignment = true;
12004 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
12005 env->strict_alignment = false;
12007 env->allow_ptr_leaks = bpf_allow_ptr_leaks();
12008 env->allow_ptr_to_map_access = bpf_allow_ptr_to_map_access();
12009 env->bypass_spec_v1 = bpf_bypass_spec_v1();
12010 env->bypass_spec_v4 = bpf_bypass_spec_v4();
12011 env->bpf_capable = bpf_capable();
12014 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
12016 if (bpf_prog_is_dev_bound(env->prog->aux)) {
12017 ret = bpf_prog_offload_verifier_prep(env->prog);
12019 goto skip_full_check;
12022 env->explored_states = kvcalloc(state_htab_size(env),
12023 sizeof(struct bpf_verifier_state_list *),
12026 if (!env->explored_states)
12027 goto skip_full_check;
12029 ret = check_subprogs(env);
12031 goto skip_full_check;
12033 ret = check_btf_info(env, attr, uattr);
12035 goto skip_full_check;
12037 ret = check_attach_btf_id(env);
12039 goto skip_full_check;
12041 ret = resolve_pseudo_ldimm64(env);
12043 goto skip_full_check;
12045 ret = check_cfg(env);
12047 goto skip_full_check;
12049 ret = do_check_subprogs(env);
12050 ret = ret ?: do_check_main(env);
12052 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
12053 ret = bpf_prog_offload_finalize(env);
12056 kvfree(env->explored_states);
12059 ret = check_max_stack_depth(env);
12061 /* instruction rewrites happen after this point */
12064 opt_hard_wire_dead_code_branches(env);
12066 ret = opt_remove_dead_code(env);
12068 ret = opt_remove_nops(env);
12071 sanitize_dead_code(env);
12075 /* program is valid, convert *(u32*)(ctx + off) accesses */
12076 ret = convert_ctx_accesses(env);
12079 ret = fixup_bpf_calls(env);
12081 /* do 32-bit optimization after insn patching has done so those patched
12082 * insns could be handled correctly.
12084 if (ret == 0 && !bpf_prog_is_dev_bound(env->prog->aux)) {
12085 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
12086 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
12091 ret = fixup_call_args(env);
12093 env->verification_time = ktime_get_ns() - start_time;
12094 print_verification_stats(env);
12096 if (log->level && bpf_verifier_log_full(log))
12098 if (log->level && !log->ubuf) {
12100 goto err_release_maps;
12103 if (ret == 0 && env->used_map_cnt) {
12104 /* if program passed verifier, update used_maps in bpf_prog_info */
12105 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
12106 sizeof(env->used_maps[0]),
12109 if (!env->prog->aux->used_maps) {
12111 goto err_release_maps;
12114 memcpy(env->prog->aux->used_maps, env->used_maps,
12115 sizeof(env->used_maps[0]) * env->used_map_cnt);
12116 env->prog->aux->used_map_cnt = env->used_map_cnt;
12118 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
12119 * bpf_ld_imm64 instructions
12121 convert_pseudo_ld_imm64(env);
12125 adjust_btf_func(env);
12128 if (!env->prog->aux->used_maps)
12129 /* if we didn't copy map pointers into bpf_prog_info, release
12130 * them now. Otherwise free_used_maps() will release them.
12134 /* extension progs temporarily inherit the attach_type of their targets
12135 for verification purposes, so set it back to zero before returning
12137 if (env->prog->type == BPF_PROG_TYPE_EXT)
12138 env->prog->expected_attach_type = 0;
12143 mutex_unlock(&bpf_verifier_lock);
12144 vfree(env->insn_aux_data);