1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
3 * Copyright (c) 2016 Facebook
4 * Copyright (c) 2018 Covalent IO, Inc. http://covalent.io
6 #include <uapi/linux/btf.h>
7 #include <linux/kernel.h>
8 #include <linux/types.h>
9 #include <linux/slab.h>
10 #include <linux/bpf.h>
11 #include <linux/btf.h>
12 #include <linux/bpf_verifier.h>
13 #include <linux/filter.h>
14 #include <net/netlink.h>
15 #include <linux/file.h>
16 #include <linux/vmalloc.h>
17 #include <linux/stringify.h>
18 #include <linux/bsearch.h>
19 #include <linux/sort.h>
20 #include <linux/perf_event.h>
21 #include <linux/ctype.h>
22 #include <linux/error-injection.h>
23 #include <linux/bpf_lsm.h>
24 #include <linux/btf_ids.h>
28 static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
29 #define BPF_PROG_TYPE(_id, _name, prog_ctx_type, kern_ctx_type) \
30 [_id] = & _name ## _verifier_ops,
31 #define BPF_MAP_TYPE(_id, _ops)
32 #define BPF_LINK_TYPE(_id, _name)
33 #include <linux/bpf_types.h>
39 /* bpf_check() is a static code analyzer that walks eBPF program
40 * instruction by instruction and updates register/stack state.
41 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
43 * The first pass is depth-first-search to check that the program is a DAG.
44 * It rejects the following programs:
45 * - larger than BPF_MAXINSNS insns
46 * - if loop is present (detected via back-edge)
47 * - unreachable insns exist (shouldn't be a forest. program = one function)
48 * - out of bounds or malformed jumps
49 * The second pass is all possible path descent from the 1st insn.
50 * Since it's analyzing all paths through the program, the length of the
51 * analysis is limited to 64k insn, which may be hit even if total number of
52 * insn is less then 4K, but there are too many branches that change stack/regs.
53 * Number of 'branches to be analyzed' is limited to 1k
55 * On entry to each instruction, each register has a type, and the instruction
56 * changes the types of the registers depending on instruction semantics.
57 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
60 * All registers are 64-bit.
61 * R0 - return register
62 * R1-R5 argument passing registers
63 * R6-R9 callee saved registers
64 * R10 - frame pointer read-only
66 * At the start of BPF program the register R1 contains a pointer to bpf_context
67 * and has type PTR_TO_CTX.
69 * Verifier tracks arithmetic operations on pointers in case:
70 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
71 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
72 * 1st insn copies R10 (which has FRAME_PTR) type into R1
73 * and 2nd arithmetic instruction is pattern matched to recognize
74 * that it wants to construct a pointer to some element within stack.
75 * So after 2nd insn, the register R1 has type PTR_TO_STACK
76 * (and -20 constant is saved for further stack bounds checking).
77 * Meaning that this reg is a pointer to stack plus known immediate constant.
79 * Most of the time the registers have SCALAR_VALUE type, which
80 * means the register has some value, but it's not a valid pointer.
81 * (like pointer plus pointer becomes SCALAR_VALUE type)
83 * When verifier sees load or store instructions the type of base register
84 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK, PTR_TO_SOCKET. These are
85 * four pointer types recognized by check_mem_access() function.
87 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
88 * and the range of [ptr, ptr + map's value_size) is accessible.
90 * registers used to pass values to function calls are checked against
91 * function argument constraints.
93 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
94 * It means that the register type passed to this function must be
95 * PTR_TO_STACK and it will be used inside the function as
96 * 'pointer to map element key'
98 * For example the argument constraints for bpf_map_lookup_elem():
99 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
100 * .arg1_type = ARG_CONST_MAP_PTR,
101 * .arg2_type = ARG_PTR_TO_MAP_KEY,
103 * ret_type says that this function returns 'pointer to map elem value or null'
104 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
105 * 2nd argument should be a pointer to stack, which will be used inside
106 * the helper function as a pointer to map element key.
108 * On the kernel side the helper function looks like:
109 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
111 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
112 * void *key = (void *) (unsigned long) r2;
115 * here kernel can access 'key' and 'map' pointers safely, knowing that
116 * [key, key + map->key_size) bytes are valid and were initialized on
117 * the stack of eBPF program.
120 * Corresponding eBPF program may look like:
121 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
122 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
123 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
124 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
125 * here verifier looks at prototype of map_lookup_elem() and sees:
126 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
127 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
129 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
130 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
131 * and were initialized prior to this call.
132 * If it's ok, then verifier allows this BPF_CALL insn and looks at
133 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
134 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
135 * returns either pointer to map value or NULL.
137 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
138 * insn, the register holding that pointer in the true branch changes state to
139 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
140 * branch. See check_cond_jmp_op().
142 * After the call R0 is set to return type of the function and registers R1-R5
143 * are set to NOT_INIT to indicate that they are no longer readable.
145 * The following reference types represent a potential reference to a kernel
146 * resource which, after first being allocated, must be checked and freed by
148 * - PTR_TO_SOCKET_OR_NULL, PTR_TO_SOCKET
150 * When the verifier sees a helper call return a reference type, it allocates a
151 * pointer id for the reference and stores it in the current function state.
152 * Similar to the way that PTR_TO_MAP_VALUE_OR_NULL is converted into
153 * PTR_TO_MAP_VALUE, PTR_TO_SOCKET_OR_NULL becomes PTR_TO_SOCKET when the type
154 * passes through a NULL-check conditional. For the branch wherein the state is
155 * changed to CONST_IMM, the verifier releases the reference.
157 * For each helper function that allocates a reference, such as
158 * bpf_sk_lookup_tcp(), there is a corresponding release function, such as
159 * bpf_sk_release(). When a reference type passes into the release function,
160 * the verifier also releases the reference. If any unchecked or unreleased
161 * reference remains at the end of the program, the verifier rejects it.
164 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
165 struct bpf_verifier_stack_elem {
166 /* verifer state is 'st'
167 * before processing instruction 'insn_idx'
168 * and after processing instruction 'prev_insn_idx'
170 struct bpf_verifier_state st;
173 struct bpf_verifier_stack_elem *next;
174 /* length of verifier log at the time this state was pushed on stack */
178 #define BPF_COMPLEXITY_LIMIT_JMP_SEQ 8192
179 #define BPF_COMPLEXITY_LIMIT_STATES 64
181 #define BPF_MAP_KEY_POISON (1ULL << 63)
182 #define BPF_MAP_KEY_SEEN (1ULL << 62)
184 #define BPF_MAP_PTR_UNPRIV 1UL
185 #define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \
186 POISON_POINTER_DELTA))
187 #define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
189 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
191 return BPF_MAP_PTR(aux->map_ptr_state) == BPF_MAP_PTR_POISON;
194 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
196 return aux->map_ptr_state & BPF_MAP_PTR_UNPRIV;
199 static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
200 const struct bpf_map *map, bool unpriv)
202 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON & BPF_MAP_PTR_UNPRIV);
203 unpriv |= bpf_map_ptr_unpriv(aux);
204 aux->map_ptr_state = (unsigned long)map |
205 (unpriv ? BPF_MAP_PTR_UNPRIV : 0UL);
208 static bool bpf_map_key_poisoned(const struct bpf_insn_aux_data *aux)
210 return aux->map_key_state & BPF_MAP_KEY_POISON;
213 static bool bpf_map_key_unseen(const struct bpf_insn_aux_data *aux)
215 return !(aux->map_key_state & BPF_MAP_KEY_SEEN);
218 static u64 bpf_map_key_immediate(const struct bpf_insn_aux_data *aux)
220 return aux->map_key_state & ~(BPF_MAP_KEY_SEEN | BPF_MAP_KEY_POISON);
223 static void bpf_map_key_store(struct bpf_insn_aux_data *aux, u64 state)
225 bool poisoned = bpf_map_key_poisoned(aux);
227 aux->map_key_state = state | BPF_MAP_KEY_SEEN |
228 (poisoned ? BPF_MAP_KEY_POISON : 0ULL);
231 static bool bpf_pseudo_call(const struct bpf_insn *insn)
233 return insn->code == (BPF_JMP | BPF_CALL) &&
234 insn->src_reg == BPF_PSEUDO_CALL;
237 static bool bpf_pseudo_kfunc_call(const struct bpf_insn *insn)
239 return insn->code == (BPF_JMP | BPF_CALL) &&
240 insn->src_reg == BPF_PSEUDO_KFUNC_CALL;
243 struct bpf_call_arg_meta {
244 struct bpf_map *map_ptr;
261 struct btf *btf_vmlinux;
263 static DEFINE_MUTEX(bpf_verifier_lock);
265 static const struct bpf_line_info *
266 find_linfo(const struct bpf_verifier_env *env, u32 insn_off)
268 const struct bpf_line_info *linfo;
269 const struct bpf_prog *prog;
273 nr_linfo = prog->aux->nr_linfo;
275 if (!nr_linfo || insn_off >= prog->len)
278 linfo = prog->aux->linfo;
279 for (i = 1; i < nr_linfo; i++)
280 if (insn_off < linfo[i].insn_off)
283 return &linfo[i - 1];
286 void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
291 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
293 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
294 "verifier log line truncated - local buffer too short\n");
296 if (log->level == BPF_LOG_KERNEL) {
297 bool newline = n > 0 && log->kbuf[n - 1] == '\n';
299 pr_err("BPF: %s%s", log->kbuf, newline ? "" : "\n");
303 n = min(log->len_total - log->len_used - 1, n);
305 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
311 static void bpf_vlog_reset(struct bpf_verifier_log *log, u32 new_pos)
315 if (!bpf_verifier_log_needed(log))
318 log->len_used = new_pos;
319 if (put_user(zero, log->ubuf + new_pos))
323 /* log_level controls verbosity level of eBPF verifier.
324 * bpf_verifier_log_write() is used to dump the verification trace to the log,
325 * so the user can figure out what's wrong with the program
327 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
328 const char *fmt, ...)
332 if (!bpf_verifier_log_needed(&env->log))
336 bpf_verifier_vlog(&env->log, fmt, args);
339 EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
341 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
343 struct bpf_verifier_env *env = private_data;
346 if (!bpf_verifier_log_needed(&env->log))
350 bpf_verifier_vlog(&env->log, fmt, args);
354 __printf(2, 3) void bpf_log(struct bpf_verifier_log *log,
355 const char *fmt, ...)
359 if (!bpf_verifier_log_needed(log))
363 bpf_verifier_vlog(log, fmt, args);
367 static const char *ltrim(const char *s)
375 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env,
377 const char *prefix_fmt, ...)
379 const struct bpf_line_info *linfo;
381 if (!bpf_verifier_log_needed(&env->log))
384 linfo = find_linfo(env, insn_off);
385 if (!linfo || linfo == env->prev_linfo)
391 va_start(args, prefix_fmt);
392 bpf_verifier_vlog(&env->log, prefix_fmt, args);
397 ltrim(btf_name_by_offset(env->prog->aux->btf,
400 env->prev_linfo = linfo;
403 static void verbose_invalid_scalar(struct bpf_verifier_env *env,
404 struct bpf_reg_state *reg,
405 struct tnum *range, const char *ctx,
406 const char *reg_name)
410 verbose(env, "At %s the register %s ", ctx, reg_name);
411 if (!tnum_is_unknown(reg->var_off)) {
412 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
413 verbose(env, "has value %s", tn_buf);
415 verbose(env, "has unknown scalar value");
417 tnum_strn(tn_buf, sizeof(tn_buf), *range);
418 verbose(env, " should have been in %s\n", tn_buf);
421 static bool type_is_pkt_pointer(enum bpf_reg_type type)
423 return type == PTR_TO_PACKET ||
424 type == PTR_TO_PACKET_META;
427 static bool type_is_sk_pointer(enum bpf_reg_type type)
429 return type == PTR_TO_SOCKET ||
430 type == PTR_TO_SOCK_COMMON ||
431 type == PTR_TO_TCP_SOCK ||
432 type == PTR_TO_XDP_SOCK;
435 static bool reg_type_not_null(enum bpf_reg_type type)
437 return type == PTR_TO_SOCKET ||
438 type == PTR_TO_TCP_SOCK ||
439 type == PTR_TO_MAP_VALUE ||
440 type == PTR_TO_MAP_KEY ||
441 type == PTR_TO_SOCK_COMMON;
444 static bool reg_type_may_be_null(enum bpf_reg_type type)
446 return type == PTR_TO_MAP_VALUE_OR_NULL ||
447 type == PTR_TO_SOCKET_OR_NULL ||
448 type == PTR_TO_SOCK_COMMON_OR_NULL ||
449 type == PTR_TO_TCP_SOCK_OR_NULL ||
450 type == PTR_TO_BTF_ID_OR_NULL ||
451 type == PTR_TO_MEM_OR_NULL ||
452 type == PTR_TO_RDONLY_BUF_OR_NULL ||
453 type == PTR_TO_RDWR_BUF_OR_NULL;
456 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
458 return reg->type == PTR_TO_MAP_VALUE &&
459 map_value_has_spin_lock(reg->map_ptr);
462 static bool reg_type_may_be_refcounted_or_null(enum bpf_reg_type type)
464 return type == PTR_TO_SOCKET ||
465 type == PTR_TO_SOCKET_OR_NULL ||
466 type == PTR_TO_TCP_SOCK ||
467 type == PTR_TO_TCP_SOCK_OR_NULL ||
468 type == PTR_TO_MEM ||
469 type == PTR_TO_MEM_OR_NULL;
472 static bool arg_type_may_be_refcounted(enum bpf_arg_type type)
474 return type == ARG_PTR_TO_SOCK_COMMON;
477 static bool arg_type_may_be_null(enum bpf_arg_type type)
479 return type == ARG_PTR_TO_MAP_VALUE_OR_NULL ||
480 type == ARG_PTR_TO_MEM_OR_NULL ||
481 type == ARG_PTR_TO_CTX_OR_NULL ||
482 type == ARG_PTR_TO_SOCKET_OR_NULL ||
483 type == ARG_PTR_TO_ALLOC_MEM_OR_NULL ||
484 type == ARG_PTR_TO_STACK_OR_NULL;
487 /* Determine whether the function releases some resources allocated by another
488 * function call. The first reference type argument will be assumed to be
489 * released by release_reference().
491 static bool is_release_function(enum bpf_func_id func_id)
493 return func_id == BPF_FUNC_sk_release ||
494 func_id == BPF_FUNC_ringbuf_submit ||
495 func_id == BPF_FUNC_ringbuf_discard;
498 static bool may_be_acquire_function(enum bpf_func_id func_id)
500 return func_id == BPF_FUNC_sk_lookup_tcp ||
501 func_id == BPF_FUNC_sk_lookup_udp ||
502 func_id == BPF_FUNC_skc_lookup_tcp ||
503 func_id == BPF_FUNC_map_lookup_elem ||
504 func_id == BPF_FUNC_ringbuf_reserve;
507 static bool is_acquire_function(enum bpf_func_id func_id,
508 const struct bpf_map *map)
510 enum bpf_map_type map_type = map ? map->map_type : BPF_MAP_TYPE_UNSPEC;
512 if (func_id == BPF_FUNC_sk_lookup_tcp ||
513 func_id == BPF_FUNC_sk_lookup_udp ||
514 func_id == BPF_FUNC_skc_lookup_tcp ||
515 func_id == BPF_FUNC_ringbuf_reserve)
518 if (func_id == BPF_FUNC_map_lookup_elem &&
519 (map_type == BPF_MAP_TYPE_SOCKMAP ||
520 map_type == BPF_MAP_TYPE_SOCKHASH))
526 static bool is_ptr_cast_function(enum bpf_func_id func_id)
528 return func_id == BPF_FUNC_tcp_sock ||
529 func_id == BPF_FUNC_sk_fullsock ||
530 func_id == BPF_FUNC_skc_to_tcp_sock ||
531 func_id == BPF_FUNC_skc_to_tcp6_sock ||
532 func_id == BPF_FUNC_skc_to_udp6_sock ||
533 func_id == BPF_FUNC_skc_to_tcp_timewait_sock ||
534 func_id == BPF_FUNC_skc_to_tcp_request_sock;
537 static bool is_cmpxchg_insn(const struct bpf_insn *insn)
539 return BPF_CLASS(insn->code) == BPF_STX &&
540 BPF_MODE(insn->code) == BPF_ATOMIC &&
541 insn->imm == BPF_CMPXCHG;
544 /* string representation of 'enum bpf_reg_type' */
545 static const char * const reg_type_str[] = {
547 [SCALAR_VALUE] = "inv",
548 [PTR_TO_CTX] = "ctx",
549 [CONST_PTR_TO_MAP] = "map_ptr",
550 [PTR_TO_MAP_VALUE] = "map_value",
551 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
552 [PTR_TO_STACK] = "fp",
553 [PTR_TO_PACKET] = "pkt",
554 [PTR_TO_PACKET_META] = "pkt_meta",
555 [PTR_TO_PACKET_END] = "pkt_end",
556 [PTR_TO_FLOW_KEYS] = "flow_keys",
557 [PTR_TO_SOCKET] = "sock",
558 [PTR_TO_SOCKET_OR_NULL] = "sock_or_null",
559 [PTR_TO_SOCK_COMMON] = "sock_common",
560 [PTR_TO_SOCK_COMMON_OR_NULL] = "sock_common_or_null",
561 [PTR_TO_TCP_SOCK] = "tcp_sock",
562 [PTR_TO_TCP_SOCK_OR_NULL] = "tcp_sock_or_null",
563 [PTR_TO_TP_BUFFER] = "tp_buffer",
564 [PTR_TO_XDP_SOCK] = "xdp_sock",
565 [PTR_TO_BTF_ID] = "ptr_",
566 [PTR_TO_BTF_ID_OR_NULL] = "ptr_or_null_",
567 [PTR_TO_PERCPU_BTF_ID] = "percpu_ptr_",
568 [PTR_TO_MEM] = "mem",
569 [PTR_TO_MEM_OR_NULL] = "mem_or_null",
570 [PTR_TO_RDONLY_BUF] = "rdonly_buf",
571 [PTR_TO_RDONLY_BUF_OR_NULL] = "rdonly_buf_or_null",
572 [PTR_TO_RDWR_BUF] = "rdwr_buf",
573 [PTR_TO_RDWR_BUF_OR_NULL] = "rdwr_buf_or_null",
574 [PTR_TO_FUNC] = "func",
575 [PTR_TO_MAP_KEY] = "map_key",
578 static char slot_type_char[] = {
579 [STACK_INVALID] = '?',
585 static void print_liveness(struct bpf_verifier_env *env,
586 enum bpf_reg_liveness live)
588 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
590 if (live & REG_LIVE_READ)
592 if (live & REG_LIVE_WRITTEN)
594 if (live & REG_LIVE_DONE)
598 static struct bpf_func_state *func(struct bpf_verifier_env *env,
599 const struct bpf_reg_state *reg)
601 struct bpf_verifier_state *cur = env->cur_state;
603 return cur->frame[reg->frameno];
606 static const char *kernel_type_name(const struct btf* btf, u32 id)
608 return btf_name_by_offset(btf, btf_type_by_id(btf, id)->name_off);
611 /* The reg state of a pointer or a bounded scalar was saved when
612 * it was spilled to the stack.
614 static bool is_spilled_reg(const struct bpf_stack_state *stack)
616 return stack->slot_type[BPF_REG_SIZE - 1] == STACK_SPILL;
619 static void scrub_spilled_slot(u8 *stype)
621 if (*stype != STACK_INVALID)
625 static void print_verifier_state(struct bpf_verifier_env *env,
626 const struct bpf_func_state *state)
628 const struct bpf_reg_state *reg;
633 verbose(env, " frame%d:", state->frameno);
634 for (i = 0; i < MAX_BPF_REG; i++) {
635 reg = &state->regs[i];
639 verbose(env, " R%d", i);
640 print_liveness(env, reg->live);
641 verbose(env, "=%s", reg_type_str[t]);
642 if (t == SCALAR_VALUE && reg->precise)
644 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
645 tnum_is_const(reg->var_off)) {
646 /* reg->off should be 0 for SCALAR_VALUE */
647 verbose(env, "%lld", reg->var_off.value + reg->off);
649 if (t == PTR_TO_BTF_ID ||
650 t == PTR_TO_BTF_ID_OR_NULL ||
651 t == PTR_TO_PERCPU_BTF_ID)
652 verbose(env, "%s", kernel_type_name(reg->btf, reg->btf_id));
653 verbose(env, "(id=%d", reg->id);
654 if (reg_type_may_be_refcounted_or_null(t))
655 verbose(env, ",ref_obj_id=%d", reg->ref_obj_id);
656 if (t != SCALAR_VALUE)
657 verbose(env, ",off=%d", reg->off);
658 if (type_is_pkt_pointer(t))
659 verbose(env, ",r=%d", reg->range);
660 else if (t == CONST_PTR_TO_MAP ||
661 t == PTR_TO_MAP_KEY ||
662 t == PTR_TO_MAP_VALUE ||
663 t == PTR_TO_MAP_VALUE_OR_NULL)
664 verbose(env, ",ks=%d,vs=%d",
665 reg->map_ptr->key_size,
666 reg->map_ptr->value_size);
667 if (tnum_is_const(reg->var_off)) {
668 /* Typically an immediate SCALAR_VALUE, but
669 * could be a pointer whose offset is too big
672 verbose(env, ",imm=%llx", reg->var_off.value);
674 if (reg->smin_value != reg->umin_value &&
675 reg->smin_value != S64_MIN)
676 verbose(env, ",smin_value=%lld",
677 (long long)reg->smin_value);
678 if (reg->smax_value != reg->umax_value &&
679 reg->smax_value != S64_MAX)
680 verbose(env, ",smax_value=%lld",
681 (long long)reg->smax_value);
682 if (reg->umin_value != 0)
683 verbose(env, ",umin_value=%llu",
684 (unsigned long long)reg->umin_value);
685 if (reg->umax_value != U64_MAX)
686 verbose(env, ",umax_value=%llu",
687 (unsigned long long)reg->umax_value);
688 if (!tnum_is_unknown(reg->var_off)) {
691 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
692 verbose(env, ",var_off=%s", tn_buf);
694 if (reg->s32_min_value != reg->smin_value &&
695 reg->s32_min_value != S32_MIN)
696 verbose(env, ",s32_min_value=%d",
697 (int)(reg->s32_min_value));
698 if (reg->s32_max_value != reg->smax_value &&
699 reg->s32_max_value != S32_MAX)
700 verbose(env, ",s32_max_value=%d",
701 (int)(reg->s32_max_value));
702 if (reg->u32_min_value != reg->umin_value &&
703 reg->u32_min_value != U32_MIN)
704 verbose(env, ",u32_min_value=%d",
705 (int)(reg->u32_min_value));
706 if (reg->u32_max_value != reg->umax_value &&
707 reg->u32_max_value != U32_MAX)
708 verbose(env, ",u32_max_value=%d",
709 (int)(reg->u32_max_value));
714 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
715 char types_buf[BPF_REG_SIZE + 1];
719 for (j = 0; j < BPF_REG_SIZE; j++) {
720 if (state->stack[i].slot_type[j] != STACK_INVALID)
722 types_buf[j] = slot_type_char[
723 state->stack[i].slot_type[j]];
725 types_buf[BPF_REG_SIZE] = 0;
728 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
729 print_liveness(env, state->stack[i].spilled_ptr.live);
730 if (is_spilled_reg(&state->stack[i])) {
731 reg = &state->stack[i].spilled_ptr;
733 verbose(env, "=%s", reg_type_str[t]);
734 if (t == SCALAR_VALUE && reg->precise)
736 if (t == SCALAR_VALUE && tnum_is_const(reg->var_off))
737 verbose(env, "%lld", reg->var_off.value + reg->off);
739 verbose(env, "=%s", types_buf);
742 if (state->acquired_refs && state->refs[0].id) {
743 verbose(env, " refs=%d", state->refs[0].id);
744 for (i = 1; i < state->acquired_refs; i++)
745 if (state->refs[i].id)
746 verbose(env, ",%d", state->refs[i].id);
748 if (state->in_callback_fn)
750 if (state->in_async_callback_fn)
751 verbose(env, " async_cb");
755 /* copy array src of length n * size bytes to dst. dst is reallocated if it's too
756 * small to hold src. This is different from krealloc since we don't want to preserve
757 * the contents of dst.
759 * Leaves dst untouched if src is NULL or length is zero. Returns NULL if memory could
762 static void *copy_array(void *dst, const void *src, size_t n, size_t size, gfp_t flags)
766 if (ZERO_OR_NULL_PTR(src))
769 if (unlikely(check_mul_overflow(n, size, &bytes)))
772 if (ksize(dst) < bytes) {
774 dst = kmalloc_track_caller(bytes, flags);
779 memcpy(dst, src, bytes);
781 return dst ? dst : ZERO_SIZE_PTR;
784 /* resize an array from old_n items to new_n items. the array is reallocated if it's too
785 * small to hold new_n items. new items are zeroed out if the array grows.
787 * Contrary to krealloc_array, does not free arr if new_n is zero.
789 static void *realloc_array(void *arr, size_t old_n, size_t new_n, size_t size)
791 if (!new_n || old_n == new_n)
794 arr = krealloc_array(arr, new_n, size, GFP_KERNEL);
799 memset(arr + old_n * size, 0, (new_n - old_n) * size);
802 return arr ? arr : ZERO_SIZE_PTR;
805 static int copy_reference_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
807 dst->refs = copy_array(dst->refs, src->refs, src->acquired_refs,
808 sizeof(struct bpf_reference_state), GFP_KERNEL);
812 dst->acquired_refs = src->acquired_refs;
816 static int copy_stack_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
818 size_t n = src->allocated_stack / BPF_REG_SIZE;
820 dst->stack = copy_array(dst->stack, src->stack, n, sizeof(struct bpf_stack_state),
825 dst->allocated_stack = src->allocated_stack;
829 static int resize_reference_state(struct bpf_func_state *state, size_t n)
831 state->refs = realloc_array(state->refs, state->acquired_refs, n,
832 sizeof(struct bpf_reference_state));
836 state->acquired_refs = n;
840 static int grow_stack_state(struct bpf_func_state *state, int size)
842 size_t old_n = state->allocated_stack / BPF_REG_SIZE, n = size / BPF_REG_SIZE;
847 state->stack = realloc_array(state->stack, old_n, n, sizeof(struct bpf_stack_state));
851 state->allocated_stack = size;
855 /* Acquire a pointer id from the env and update the state->refs to include
856 * this new pointer reference.
857 * On success, returns a valid pointer id to associate with the register
858 * On failure, returns a negative errno.
860 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
862 struct bpf_func_state *state = cur_func(env);
863 int new_ofs = state->acquired_refs;
866 err = resize_reference_state(state, state->acquired_refs + 1);
870 state->refs[new_ofs].id = id;
871 state->refs[new_ofs].insn_idx = insn_idx;
876 /* release function corresponding to acquire_reference_state(). Idempotent. */
877 static int release_reference_state(struct bpf_func_state *state, int ptr_id)
881 last_idx = state->acquired_refs - 1;
882 for (i = 0; i < state->acquired_refs; i++) {
883 if (state->refs[i].id == ptr_id) {
884 if (last_idx && i != last_idx)
885 memcpy(&state->refs[i], &state->refs[last_idx],
886 sizeof(*state->refs));
887 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
888 state->acquired_refs--;
895 static void free_func_state(struct bpf_func_state *state)
904 static void clear_jmp_history(struct bpf_verifier_state *state)
906 kfree(state->jmp_history);
907 state->jmp_history = NULL;
908 state->jmp_history_cnt = 0;
911 static void free_verifier_state(struct bpf_verifier_state *state,
916 for (i = 0; i <= state->curframe; i++) {
917 free_func_state(state->frame[i]);
918 state->frame[i] = NULL;
920 clear_jmp_history(state);
925 /* copy verifier state from src to dst growing dst stack space
926 * when necessary to accommodate larger src stack
928 static int copy_func_state(struct bpf_func_state *dst,
929 const struct bpf_func_state *src)
933 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
934 err = copy_reference_state(dst, src);
937 return copy_stack_state(dst, src);
940 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
941 const struct bpf_verifier_state *src)
943 struct bpf_func_state *dst;
946 dst_state->jmp_history = copy_array(dst_state->jmp_history, src->jmp_history,
947 src->jmp_history_cnt, sizeof(struct bpf_idx_pair),
949 if (!dst_state->jmp_history)
951 dst_state->jmp_history_cnt = src->jmp_history_cnt;
953 /* if dst has more stack frames then src frame, free them */
954 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
955 free_func_state(dst_state->frame[i]);
956 dst_state->frame[i] = NULL;
958 dst_state->speculative = src->speculative;
959 dst_state->curframe = src->curframe;
960 dst_state->active_spin_lock = src->active_spin_lock;
961 dst_state->branches = src->branches;
962 dst_state->parent = src->parent;
963 dst_state->first_insn_idx = src->first_insn_idx;
964 dst_state->last_insn_idx = src->last_insn_idx;
965 for (i = 0; i <= src->curframe; i++) {
966 dst = dst_state->frame[i];
968 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
971 dst_state->frame[i] = dst;
973 err = copy_func_state(dst, src->frame[i]);
980 static void update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
983 u32 br = --st->branches;
985 /* WARN_ON(br > 1) technically makes sense here,
986 * but see comment in push_stack(), hence:
988 WARN_ONCE((int)br < 0,
989 "BUG update_branch_counts:branches_to_explore=%d\n",
997 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
998 int *insn_idx, bool pop_log)
1000 struct bpf_verifier_state *cur = env->cur_state;
1001 struct bpf_verifier_stack_elem *elem, *head = env->head;
1004 if (env->head == NULL)
1008 err = copy_verifier_state(cur, &head->st);
1013 bpf_vlog_reset(&env->log, head->log_pos);
1015 *insn_idx = head->insn_idx;
1017 *prev_insn_idx = head->prev_insn_idx;
1019 free_verifier_state(&head->st, false);
1026 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
1027 int insn_idx, int prev_insn_idx,
1030 struct bpf_verifier_state *cur = env->cur_state;
1031 struct bpf_verifier_stack_elem *elem;
1034 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
1038 elem->insn_idx = insn_idx;
1039 elem->prev_insn_idx = prev_insn_idx;
1040 elem->next = env->head;
1041 elem->log_pos = env->log.len_used;
1044 err = copy_verifier_state(&elem->st, cur);
1047 elem->st.speculative |= speculative;
1048 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
1049 verbose(env, "The sequence of %d jumps is too complex.\n",
1053 if (elem->st.parent) {
1054 ++elem->st.parent->branches;
1055 /* WARN_ON(branches > 2) technically makes sense here,
1057 * 1. speculative states will bump 'branches' for non-branch
1059 * 2. is_state_visited() heuristics may decide not to create
1060 * a new state for a sequence of branches and all such current
1061 * and cloned states will be pointing to a single parent state
1062 * which might have large 'branches' count.
1067 free_verifier_state(env->cur_state, true);
1068 env->cur_state = NULL;
1069 /* pop all elements and return */
1070 while (!pop_stack(env, NULL, NULL, false));
1074 #define CALLER_SAVED_REGS 6
1075 static const int caller_saved[CALLER_SAVED_REGS] = {
1076 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
1079 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1080 struct bpf_reg_state *reg);
1082 /* This helper doesn't clear reg->id */
1083 static void ___mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1085 reg->var_off = tnum_const(imm);
1086 reg->smin_value = (s64)imm;
1087 reg->smax_value = (s64)imm;
1088 reg->umin_value = imm;
1089 reg->umax_value = imm;
1091 reg->s32_min_value = (s32)imm;
1092 reg->s32_max_value = (s32)imm;
1093 reg->u32_min_value = (u32)imm;
1094 reg->u32_max_value = (u32)imm;
1097 /* Mark the unknown part of a register (variable offset or scalar value) as
1098 * known to have the value @imm.
1100 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1102 /* Clear id, off, and union(map_ptr, range) */
1103 memset(((u8 *)reg) + sizeof(reg->type), 0,
1104 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
1105 ___mark_reg_known(reg, imm);
1108 static void __mark_reg32_known(struct bpf_reg_state *reg, u64 imm)
1110 reg->var_off = tnum_const_subreg(reg->var_off, imm);
1111 reg->s32_min_value = (s32)imm;
1112 reg->s32_max_value = (s32)imm;
1113 reg->u32_min_value = (u32)imm;
1114 reg->u32_max_value = (u32)imm;
1117 /* Mark the 'variable offset' part of a register as zero. This should be
1118 * used only on registers holding a pointer type.
1120 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
1122 __mark_reg_known(reg, 0);
1125 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
1127 __mark_reg_known(reg, 0);
1128 reg->type = SCALAR_VALUE;
1131 static void mark_reg_known_zero(struct bpf_verifier_env *env,
1132 struct bpf_reg_state *regs, u32 regno)
1134 if (WARN_ON(regno >= MAX_BPF_REG)) {
1135 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
1136 /* Something bad happened, let's kill all regs */
1137 for (regno = 0; regno < MAX_BPF_REG; regno++)
1138 __mark_reg_not_init(env, regs + regno);
1141 __mark_reg_known_zero(regs + regno);
1144 static void mark_ptr_not_null_reg(struct bpf_reg_state *reg)
1146 switch (reg->type) {
1147 case PTR_TO_MAP_VALUE_OR_NULL: {
1148 const struct bpf_map *map = reg->map_ptr;
1150 if (map->inner_map_meta) {
1151 reg->type = CONST_PTR_TO_MAP;
1152 reg->map_ptr = map->inner_map_meta;
1153 /* transfer reg's id which is unique for every map_lookup_elem
1154 * as UID of the inner map.
1156 if (map_value_has_timer(map->inner_map_meta))
1157 reg->map_uid = reg->id;
1158 } else if (map->map_type == BPF_MAP_TYPE_XSKMAP) {
1159 reg->type = PTR_TO_XDP_SOCK;
1160 } else if (map->map_type == BPF_MAP_TYPE_SOCKMAP ||
1161 map->map_type == BPF_MAP_TYPE_SOCKHASH) {
1162 reg->type = PTR_TO_SOCKET;
1164 reg->type = PTR_TO_MAP_VALUE;
1168 case PTR_TO_SOCKET_OR_NULL:
1169 reg->type = PTR_TO_SOCKET;
1171 case PTR_TO_SOCK_COMMON_OR_NULL:
1172 reg->type = PTR_TO_SOCK_COMMON;
1174 case PTR_TO_TCP_SOCK_OR_NULL:
1175 reg->type = PTR_TO_TCP_SOCK;
1177 case PTR_TO_BTF_ID_OR_NULL:
1178 reg->type = PTR_TO_BTF_ID;
1180 case PTR_TO_MEM_OR_NULL:
1181 reg->type = PTR_TO_MEM;
1183 case PTR_TO_RDONLY_BUF_OR_NULL:
1184 reg->type = PTR_TO_RDONLY_BUF;
1186 case PTR_TO_RDWR_BUF_OR_NULL:
1187 reg->type = PTR_TO_RDWR_BUF;
1190 WARN_ONCE(1, "unknown nullable register type");
1194 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
1196 return type_is_pkt_pointer(reg->type);
1199 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
1201 return reg_is_pkt_pointer(reg) ||
1202 reg->type == PTR_TO_PACKET_END;
1205 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
1206 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
1207 enum bpf_reg_type which)
1209 /* The register can already have a range from prior markings.
1210 * This is fine as long as it hasn't been advanced from its
1213 return reg->type == which &&
1216 tnum_equals_const(reg->var_off, 0);
1219 /* Reset the min/max bounds of a register */
1220 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
1222 reg->smin_value = S64_MIN;
1223 reg->smax_value = S64_MAX;
1224 reg->umin_value = 0;
1225 reg->umax_value = U64_MAX;
1227 reg->s32_min_value = S32_MIN;
1228 reg->s32_max_value = S32_MAX;
1229 reg->u32_min_value = 0;
1230 reg->u32_max_value = U32_MAX;
1233 static void __mark_reg64_unbounded(struct bpf_reg_state *reg)
1235 reg->smin_value = S64_MIN;
1236 reg->smax_value = S64_MAX;
1237 reg->umin_value = 0;
1238 reg->umax_value = U64_MAX;
1241 static void __mark_reg32_unbounded(struct bpf_reg_state *reg)
1243 reg->s32_min_value = S32_MIN;
1244 reg->s32_max_value = S32_MAX;
1245 reg->u32_min_value = 0;
1246 reg->u32_max_value = U32_MAX;
1249 static void __update_reg32_bounds(struct bpf_reg_state *reg)
1251 struct tnum var32_off = tnum_subreg(reg->var_off);
1253 /* min signed is max(sign bit) | min(other bits) */
1254 reg->s32_min_value = max_t(s32, reg->s32_min_value,
1255 var32_off.value | (var32_off.mask & S32_MIN));
1256 /* max signed is min(sign bit) | max(other bits) */
1257 reg->s32_max_value = min_t(s32, reg->s32_max_value,
1258 var32_off.value | (var32_off.mask & S32_MAX));
1259 reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)var32_off.value);
1260 reg->u32_max_value = min(reg->u32_max_value,
1261 (u32)(var32_off.value | var32_off.mask));
1264 static void __update_reg64_bounds(struct bpf_reg_state *reg)
1266 /* min signed is max(sign bit) | min(other bits) */
1267 reg->smin_value = max_t(s64, reg->smin_value,
1268 reg->var_off.value | (reg->var_off.mask & S64_MIN));
1269 /* max signed is min(sign bit) | max(other bits) */
1270 reg->smax_value = min_t(s64, reg->smax_value,
1271 reg->var_off.value | (reg->var_off.mask & S64_MAX));
1272 reg->umin_value = max(reg->umin_value, reg->var_off.value);
1273 reg->umax_value = min(reg->umax_value,
1274 reg->var_off.value | reg->var_off.mask);
1277 static void __update_reg_bounds(struct bpf_reg_state *reg)
1279 __update_reg32_bounds(reg);
1280 __update_reg64_bounds(reg);
1283 /* Uses signed min/max values to inform unsigned, and vice-versa */
1284 static void __reg32_deduce_bounds(struct bpf_reg_state *reg)
1286 /* Learn sign from signed bounds.
1287 * If we cannot cross the sign boundary, then signed and unsigned bounds
1288 * are the same, so combine. This works even in the negative case, e.g.
1289 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1291 if (reg->s32_min_value >= 0 || reg->s32_max_value < 0) {
1292 reg->s32_min_value = reg->u32_min_value =
1293 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1294 reg->s32_max_value = reg->u32_max_value =
1295 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1298 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1299 * boundary, so we must be careful.
1301 if ((s32)reg->u32_max_value >= 0) {
1302 /* Positive. We can't learn anything from the smin, but smax
1303 * is positive, hence safe.
1305 reg->s32_min_value = reg->u32_min_value;
1306 reg->s32_max_value = reg->u32_max_value =
1307 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1308 } else if ((s32)reg->u32_min_value < 0) {
1309 /* Negative. We can't learn anything from the smax, but smin
1310 * is negative, hence safe.
1312 reg->s32_min_value = reg->u32_min_value =
1313 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1314 reg->s32_max_value = reg->u32_max_value;
1318 static void __reg64_deduce_bounds(struct bpf_reg_state *reg)
1320 /* Learn sign from signed bounds.
1321 * If we cannot cross the sign boundary, then signed and unsigned bounds
1322 * are the same, so combine. This works even in the negative case, e.g.
1323 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1325 if (reg->smin_value >= 0 || reg->smax_value < 0) {
1326 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1328 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1332 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1333 * boundary, so we must be careful.
1335 if ((s64)reg->umax_value >= 0) {
1336 /* Positive. We can't learn anything from the smin, but smax
1337 * is positive, hence safe.
1339 reg->smin_value = reg->umin_value;
1340 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1342 } else if ((s64)reg->umin_value < 0) {
1343 /* Negative. We can't learn anything from the smax, but smin
1344 * is negative, hence safe.
1346 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1348 reg->smax_value = reg->umax_value;
1352 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
1354 __reg32_deduce_bounds(reg);
1355 __reg64_deduce_bounds(reg);
1358 /* Attempts to improve var_off based on unsigned min/max information */
1359 static void __reg_bound_offset(struct bpf_reg_state *reg)
1361 struct tnum var64_off = tnum_intersect(reg->var_off,
1362 tnum_range(reg->umin_value,
1364 struct tnum var32_off = tnum_intersect(tnum_subreg(reg->var_off),
1365 tnum_range(reg->u32_min_value,
1366 reg->u32_max_value));
1368 reg->var_off = tnum_or(tnum_clear_subreg(var64_off), var32_off);
1371 static void __reg_assign_32_into_64(struct bpf_reg_state *reg)
1373 reg->umin_value = reg->u32_min_value;
1374 reg->umax_value = reg->u32_max_value;
1375 /* Attempt to pull 32-bit signed bounds into 64-bit bounds
1376 * but must be positive otherwise set to worse case bounds
1377 * and refine later from tnum.
1379 if (reg->s32_min_value >= 0 && reg->s32_max_value >= 0)
1380 reg->smax_value = reg->s32_max_value;
1382 reg->smax_value = U32_MAX;
1383 if (reg->s32_min_value >= 0)
1384 reg->smin_value = reg->s32_min_value;
1386 reg->smin_value = 0;
1389 static void __reg_combine_32_into_64(struct bpf_reg_state *reg)
1391 /* special case when 64-bit register has upper 32-bit register
1392 * zeroed. Typically happens after zext or <<32, >>32 sequence
1393 * allowing us to use 32-bit bounds directly,
1395 if (tnum_equals_const(tnum_clear_subreg(reg->var_off), 0)) {
1396 __reg_assign_32_into_64(reg);
1398 /* Otherwise the best we can do is push lower 32bit known and
1399 * unknown bits into register (var_off set from jmp logic)
1400 * then learn as much as possible from the 64-bit tnum
1401 * known and unknown bits. The previous smin/smax bounds are
1402 * invalid here because of jmp32 compare so mark them unknown
1403 * so they do not impact tnum bounds calculation.
1405 __mark_reg64_unbounded(reg);
1406 __update_reg_bounds(reg);
1409 /* Intersecting with the old var_off might have improved our bounds
1410 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1411 * then new var_off is (0; 0x7f...fc) which improves our umax.
1413 __reg_deduce_bounds(reg);
1414 __reg_bound_offset(reg);
1415 __update_reg_bounds(reg);
1418 static bool __reg64_bound_s32(s64 a)
1420 return a >= S32_MIN && a <= S32_MAX;
1423 static bool __reg64_bound_u32(u64 a)
1425 return a >= U32_MIN && a <= U32_MAX;
1428 static void __reg_combine_64_into_32(struct bpf_reg_state *reg)
1430 __mark_reg32_unbounded(reg);
1432 if (__reg64_bound_s32(reg->smin_value) && __reg64_bound_s32(reg->smax_value)) {
1433 reg->s32_min_value = (s32)reg->smin_value;
1434 reg->s32_max_value = (s32)reg->smax_value;
1436 if (__reg64_bound_u32(reg->umin_value) && __reg64_bound_u32(reg->umax_value)) {
1437 reg->u32_min_value = (u32)reg->umin_value;
1438 reg->u32_max_value = (u32)reg->umax_value;
1441 /* Intersecting with the old var_off might have improved our bounds
1442 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1443 * then new var_off is (0; 0x7f...fc) which improves our umax.
1445 __reg_deduce_bounds(reg);
1446 __reg_bound_offset(reg);
1447 __update_reg_bounds(reg);
1450 /* Mark a register as having a completely unknown (scalar) value. */
1451 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
1452 struct bpf_reg_state *reg)
1455 * Clear type, id, off, and union(map_ptr, range) and
1456 * padding between 'type' and union
1458 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
1459 reg->type = SCALAR_VALUE;
1460 reg->var_off = tnum_unknown;
1462 reg->precise = env->subprog_cnt > 1 || !env->bpf_capable;
1463 __mark_reg_unbounded(reg);
1466 static void mark_reg_unknown(struct bpf_verifier_env *env,
1467 struct bpf_reg_state *regs, u32 regno)
1469 if (WARN_ON(regno >= MAX_BPF_REG)) {
1470 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
1471 /* Something bad happened, let's kill all regs except FP */
1472 for (regno = 0; regno < BPF_REG_FP; regno++)
1473 __mark_reg_not_init(env, regs + regno);
1476 __mark_reg_unknown(env, regs + regno);
1479 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1480 struct bpf_reg_state *reg)
1482 __mark_reg_unknown(env, reg);
1483 reg->type = NOT_INIT;
1486 static void mark_reg_not_init(struct bpf_verifier_env *env,
1487 struct bpf_reg_state *regs, u32 regno)
1489 if (WARN_ON(regno >= MAX_BPF_REG)) {
1490 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
1491 /* Something bad happened, let's kill all regs except FP */
1492 for (regno = 0; regno < BPF_REG_FP; regno++)
1493 __mark_reg_not_init(env, regs + regno);
1496 __mark_reg_not_init(env, regs + regno);
1499 static void mark_btf_ld_reg(struct bpf_verifier_env *env,
1500 struct bpf_reg_state *regs, u32 regno,
1501 enum bpf_reg_type reg_type,
1502 struct btf *btf, u32 btf_id)
1504 if (reg_type == SCALAR_VALUE) {
1505 mark_reg_unknown(env, regs, regno);
1508 mark_reg_known_zero(env, regs, regno);
1509 regs[regno].type = PTR_TO_BTF_ID;
1510 regs[regno].btf = btf;
1511 regs[regno].btf_id = btf_id;
1514 #define DEF_NOT_SUBREG (0)
1515 static void init_reg_state(struct bpf_verifier_env *env,
1516 struct bpf_func_state *state)
1518 struct bpf_reg_state *regs = state->regs;
1521 for (i = 0; i < MAX_BPF_REG; i++) {
1522 mark_reg_not_init(env, regs, i);
1523 regs[i].live = REG_LIVE_NONE;
1524 regs[i].parent = NULL;
1525 regs[i].subreg_def = DEF_NOT_SUBREG;
1529 regs[BPF_REG_FP].type = PTR_TO_STACK;
1530 mark_reg_known_zero(env, regs, BPF_REG_FP);
1531 regs[BPF_REG_FP].frameno = state->frameno;
1534 #define BPF_MAIN_FUNC (-1)
1535 static void init_func_state(struct bpf_verifier_env *env,
1536 struct bpf_func_state *state,
1537 int callsite, int frameno, int subprogno)
1539 state->callsite = callsite;
1540 state->frameno = frameno;
1541 state->subprogno = subprogno;
1542 init_reg_state(env, state);
1545 /* Similar to push_stack(), but for async callbacks */
1546 static struct bpf_verifier_state *push_async_cb(struct bpf_verifier_env *env,
1547 int insn_idx, int prev_insn_idx,
1550 struct bpf_verifier_stack_elem *elem;
1551 struct bpf_func_state *frame;
1553 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
1557 elem->insn_idx = insn_idx;
1558 elem->prev_insn_idx = prev_insn_idx;
1559 elem->next = env->head;
1560 elem->log_pos = env->log.len_used;
1563 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
1565 "The sequence of %d jumps is too complex for async cb.\n",
1569 /* Unlike push_stack() do not copy_verifier_state().
1570 * The caller state doesn't matter.
1571 * This is async callback. It starts in a fresh stack.
1572 * Initialize it similar to do_check_common().
1574 elem->st.branches = 1;
1575 frame = kzalloc(sizeof(*frame), GFP_KERNEL);
1578 init_func_state(env, frame,
1579 BPF_MAIN_FUNC /* callsite */,
1580 0 /* frameno within this callchain */,
1581 subprog /* subprog number within this prog */);
1582 elem->st.frame[0] = frame;
1585 free_verifier_state(env->cur_state, true);
1586 env->cur_state = NULL;
1587 /* pop all elements and return */
1588 while (!pop_stack(env, NULL, NULL, false));
1594 SRC_OP, /* register is used as source operand */
1595 DST_OP, /* register is used as destination operand */
1596 DST_OP_NO_MARK /* same as above, check only, don't mark */
1599 static int cmp_subprogs(const void *a, const void *b)
1601 return ((struct bpf_subprog_info *)a)->start -
1602 ((struct bpf_subprog_info *)b)->start;
1605 static int find_subprog(struct bpf_verifier_env *env, int off)
1607 struct bpf_subprog_info *p;
1609 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
1610 sizeof(env->subprog_info[0]), cmp_subprogs);
1613 return p - env->subprog_info;
1617 static int add_subprog(struct bpf_verifier_env *env, int off)
1619 int insn_cnt = env->prog->len;
1622 if (off >= insn_cnt || off < 0) {
1623 verbose(env, "call to invalid destination\n");
1626 ret = find_subprog(env, off);
1629 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
1630 verbose(env, "too many subprograms\n");
1633 /* determine subprog starts. The end is one before the next starts */
1634 env->subprog_info[env->subprog_cnt++].start = off;
1635 sort(env->subprog_info, env->subprog_cnt,
1636 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
1637 return env->subprog_cnt - 1;
1640 #define MAX_KFUNC_DESCS 256
1641 #define MAX_KFUNC_BTFS 256
1643 struct bpf_kfunc_desc {
1644 struct btf_func_model func_model;
1650 struct bpf_kfunc_btf {
1652 struct module *module;
1656 struct bpf_kfunc_desc_tab {
1657 struct bpf_kfunc_desc descs[MAX_KFUNC_DESCS];
1661 struct bpf_kfunc_btf_tab {
1662 struct bpf_kfunc_btf descs[MAX_KFUNC_BTFS];
1666 static int kfunc_desc_cmp_by_id_off(const void *a, const void *b)
1668 const struct bpf_kfunc_desc *d0 = a;
1669 const struct bpf_kfunc_desc *d1 = b;
1671 /* func_id is not greater than BTF_MAX_TYPE */
1672 return d0->func_id - d1->func_id ?: d0->offset - d1->offset;
1675 static int kfunc_btf_cmp_by_off(const void *a, const void *b)
1677 const struct bpf_kfunc_btf *d0 = a;
1678 const struct bpf_kfunc_btf *d1 = b;
1680 return d0->offset - d1->offset;
1683 static const struct bpf_kfunc_desc *
1684 find_kfunc_desc(const struct bpf_prog *prog, u32 func_id, u16 offset)
1686 struct bpf_kfunc_desc desc = {
1690 struct bpf_kfunc_desc_tab *tab;
1692 tab = prog->aux->kfunc_tab;
1693 return bsearch(&desc, tab->descs, tab->nr_descs,
1694 sizeof(tab->descs[0]), kfunc_desc_cmp_by_id_off);
1697 static struct btf *__find_kfunc_desc_btf(struct bpf_verifier_env *env,
1698 s16 offset, struct module **btf_modp)
1700 struct bpf_kfunc_btf kf_btf = { .offset = offset };
1701 struct bpf_kfunc_btf_tab *tab;
1702 struct bpf_kfunc_btf *b;
1707 tab = env->prog->aux->kfunc_btf_tab;
1708 b = bsearch(&kf_btf, tab->descs, tab->nr_descs,
1709 sizeof(tab->descs[0]), kfunc_btf_cmp_by_off);
1711 if (tab->nr_descs == MAX_KFUNC_BTFS) {
1712 verbose(env, "too many different module BTFs\n");
1713 return ERR_PTR(-E2BIG);
1716 if (bpfptr_is_null(env->fd_array)) {
1717 verbose(env, "kfunc offset > 0 without fd_array is invalid\n");
1718 return ERR_PTR(-EPROTO);
1721 if (copy_from_bpfptr_offset(&btf_fd, env->fd_array,
1722 offset * sizeof(btf_fd),
1724 return ERR_PTR(-EFAULT);
1726 btf = btf_get_by_fd(btf_fd);
1728 verbose(env, "invalid module BTF fd specified\n");
1732 if (!btf_is_module(btf)) {
1733 verbose(env, "BTF fd for kfunc is not a module BTF\n");
1735 return ERR_PTR(-EINVAL);
1738 mod = btf_try_get_module(btf);
1741 return ERR_PTR(-ENXIO);
1744 b = &tab->descs[tab->nr_descs++];
1749 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
1750 kfunc_btf_cmp_by_off, NULL);
1753 *btf_modp = b->module;
1757 void bpf_free_kfunc_btf_tab(struct bpf_kfunc_btf_tab *tab)
1762 while (tab->nr_descs--) {
1763 module_put(tab->descs[tab->nr_descs].module);
1764 btf_put(tab->descs[tab->nr_descs].btf);
1769 static struct btf *find_kfunc_desc_btf(struct bpf_verifier_env *env,
1770 u32 func_id, s16 offset,
1771 struct module **btf_modp)
1775 /* In the future, this can be allowed to increase limit
1776 * of fd index into fd_array, interpreted as u16.
1778 verbose(env, "negative offset disallowed for kernel module function call\n");
1779 return ERR_PTR(-EINVAL);
1782 return __find_kfunc_desc_btf(env, offset, btf_modp);
1784 return btf_vmlinux ?: ERR_PTR(-ENOENT);
1787 static int add_kfunc_call(struct bpf_verifier_env *env, u32 func_id, s16 offset)
1789 const struct btf_type *func, *func_proto;
1790 struct bpf_kfunc_btf_tab *btf_tab;
1791 struct bpf_kfunc_desc_tab *tab;
1792 struct bpf_prog_aux *prog_aux;
1793 struct bpf_kfunc_desc *desc;
1794 const char *func_name;
1795 struct btf *desc_btf;
1799 prog_aux = env->prog->aux;
1800 tab = prog_aux->kfunc_tab;
1801 btf_tab = prog_aux->kfunc_btf_tab;
1804 verbose(env, "calling kernel function is not supported without CONFIG_DEBUG_INFO_BTF\n");
1808 if (!env->prog->jit_requested) {
1809 verbose(env, "JIT is required for calling kernel function\n");
1813 if (!bpf_jit_supports_kfunc_call()) {
1814 verbose(env, "JIT does not support calling kernel function\n");
1818 if (!env->prog->gpl_compatible) {
1819 verbose(env, "cannot call kernel function from non-GPL compatible program\n");
1823 tab = kzalloc(sizeof(*tab), GFP_KERNEL);
1826 prog_aux->kfunc_tab = tab;
1829 /* func_id == 0 is always invalid, but instead of returning an error, be
1830 * conservative and wait until the code elimination pass before returning
1831 * error, so that invalid calls that get pruned out can be in BPF programs
1832 * loaded from userspace. It is also required that offset be untouched
1835 if (!func_id && !offset)
1838 if (!btf_tab && offset) {
1839 btf_tab = kzalloc(sizeof(*btf_tab), GFP_KERNEL);
1842 prog_aux->kfunc_btf_tab = btf_tab;
1845 desc_btf = find_kfunc_desc_btf(env, func_id, offset, NULL);
1846 if (IS_ERR(desc_btf)) {
1847 verbose(env, "failed to find BTF for kernel function\n");
1848 return PTR_ERR(desc_btf);
1851 if (find_kfunc_desc(env->prog, func_id, offset))
1854 if (tab->nr_descs == MAX_KFUNC_DESCS) {
1855 verbose(env, "too many different kernel function calls\n");
1859 func = btf_type_by_id(desc_btf, func_id);
1860 if (!func || !btf_type_is_func(func)) {
1861 verbose(env, "kernel btf_id %u is not a function\n",
1865 func_proto = btf_type_by_id(desc_btf, func->type);
1866 if (!func_proto || !btf_type_is_func_proto(func_proto)) {
1867 verbose(env, "kernel function btf_id %u does not have a valid func_proto\n",
1872 func_name = btf_name_by_offset(desc_btf, func->name_off);
1873 addr = kallsyms_lookup_name(func_name);
1875 verbose(env, "cannot find address for kernel function %s\n",
1880 desc = &tab->descs[tab->nr_descs++];
1881 desc->func_id = func_id;
1882 desc->imm = BPF_CALL_IMM(addr);
1883 desc->offset = offset;
1884 err = btf_distill_func_proto(&env->log, desc_btf,
1885 func_proto, func_name,
1888 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
1889 kfunc_desc_cmp_by_id_off, NULL);
1893 static int kfunc_desc_cmp_by_imm(const void *a, const void *b)
1895 const struct bpf_kfunc_desc *d0 = a;
1896 const struct bpf_kfunc_desc *d1 = b;
1898 if (d0->imm > d1->imm)
1900 else if (d0->imm < d1->imm)
1905 static void sort_kfunc_descs_by_imm(struct bpf_prog *prog)
1907 struct bpf_kfunc_desc_tab *tab;
1909 tab = prog->aux->kfunc_tab;
1913 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
1914 kfunc_desc_cmp_by_imm, NULL);
1917 bool bpf_prog_has_kfunc_call(const struct bpf_prog *prog)
1919 return !!prog->aux->kfunc_tab;
1922 const struct btf_func_model *
1923 bpf_jit_find_kfunc_model(const struct bpf_prog *prog,
1924 const struct bpf_insn *insn)
1926 const struct bpf_kfunc_desc desc = {
1929 const struct bpf_kfunc_desc *res;
1930 struct bpf_kfunc_desc_tab *tab;
1932 tab = prog->aux->kfunc_tab;
1933 res = bsearch(&desc, tab->descs, tab->nr_descs,
1934 sizeof(tab->descs[0]), kfunc_desc_cmp_by_imm);
1936 return res ? &res->func_model : NULL;
1939 static int add_subprog_and_kfunc(struct bpf_verifier_env *env)
1941 struct bpf_subprog_info *subprog = env->subprog_info;
1942 struct bpf_insn *insn = env->prog->insnsi;
1943 int i, ret, insn_cnt = env->prog->len;
1945 /* Add entry function. */
1946 ret = add_subprog(env, 0);
1950 for (i = 0; i < insn_cnt; i++, insn++) {
1951 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn) &&
1952 !bpf_pseudo_kfunc_call(insn))
1955 if (!env->bpf_capable) {
1956 verbose(env, "loading/calling other bpf or kernel functions are allowed for CAP_BPF and CAP_SYS_ADMIN\n");
1960 if (bpf_pseudo_func(insn) || bpf_pseudo_call(insn))
1961 ret = add_subprog(env, i + insn->imm + 1);
1963 ret = add_kfunc_call(env, insn->imm, insn->off);
1969 /* Add a fake 'exit' subprog which could simplify subprog iteration
1970 * logic. 'subprog_cnt' should not be increased.
1972 subprog[env->subprog_cnt].start = insn_cnt;
1974 if (env->log.level & BPF_LOG_LEVEL2)
1975 for (i = 0; i < env->subprog_cnt; i++)
1976 verbose(env, "func#%d @%d\n", i, subprog[i].start);
1981 static int check_subprogs(struct bpf_verifier_env *env)
1983 int i, subprog_start, subprog_end, off, cur_subprog = 0;
1984 struct bpf_subprog_info *subprog = env->subprog_info;
1985 struct bpf_insn *insn = env->prog->insnsi;
1986 int insn_cnt = env->prog->len;
1988 /* now check that all jumps are within the same subprog */
1989 subprog_start = subprog[cur_subprog].start;
1990 subprog_end = subprog[cur_subprog + 1].start;
1991 for (i = 0; i < insn_cnt; i++) {
1992 u8 code = insn[i].code;
1994 if (code == (BPF_JMP | BPF_CALL) &&
1995 insn[i].imm == BPF_FUNC_tail_call &&
1996 insn[i].src_reg != BPF_PSEUDO_CALL)
1997 subprog[cur_subprog].has_tail_call = true;
1998 if (BPF_CLASS(code) == BPF_LD &&
1999 (BPF_MODE(code) == BPF_ABS || BPF_MODE(code) == BPF_IND))
2000 subprog[cur_subprog].has_ld_abs = true;
2001 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
2003 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
2005 off = i + insn[i].off + 1;
2006 if (off < subprog_start || off >= subprog_end) {
2007 verbose(env, "jump out of range from insn %d to %d\n", i, off);
2011 if (i == subprog_end - 1) {
2012 /* to avoid fall-through from one subprog into another
2013 * the last insn of the subprog should be either exit
2014 * or unconditional jump back
2016 if (code != (BPF_JMP | BPF_EXIT) &&
2017 code != (BPF_JMP | BPF_JA)) {
2018 verbose(env, "last insn is not an exit or jmp\n");
2021 subprog_start = subprog_end;
2023 if (cur_subprog < env->subprog_cnt)
2024 subprog_end = subprog[cur_subprog + 1].start;
2030 /* Parentage chain of this register (or stack slot) should take care of all
2031 * issues like callee-saved registers, stack slot allocation time, etc.
2033 static int mark_reg_read(struct bpf_verifier_env *env,
2034 const struct bpf_reg_state *state,
2035 struct bpf_reg_state *parent, u8 flag)
2037 bool writes = parent == state->parent; /* Observe write marks */
2041 /* if read wasn't screened by an earlier write ... */
2042 if (writes && state->live & REG_LIVE_WRITTEN)
2044 if (parent->live & REG_LIVE_DONE) {
2045 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
2046 reg_type_str[parent->type],
2047 parent->var_off.value, parent->off);
2050 /* The first condition is more likely to be true than the
2051 * second, checked it first.
2053 if ((parent->live & REG_LIVE_READ) == flag ||
2054 parent->live & REG_LIVE_READ64)
2055 /* The parentage chain never changes and
2056 * this parent was already marked as LIVE_READ.
2057 * There is no need to keep walking the chain again and
2058 * keep re-marking all parents as LIVE_READ.
2059 * This case happens when the same register is read
2060 * multiple times without writes into it in-between.
2061 * Also, if parent has the stronger REG_LIVE_READ64 set,
2062 * then no need to set the weak REG_LIVE_READ32.
2065 /* ... then we depend on parent's value */
2066 parent->live |= flag;
2067 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
2068 if (flag == REG_LIVE_READ64)
2069 parent->live &= ~REG_LIVE_READ32;
2071 parent = state->parent;
2076 if (env->longest_mark_read_walk < cnt)
2077 env->longest_mark_read_walk = cnt;
2081 /* This function is supposed to be used by the following 32-bit optimization
2082 * code only. It returns TRUE if the source or destination register operates
2083 * on 64-bit, otherwise return FALSE.
2085 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
2086 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
2091 class = BPF_CLASS(code);
2093 if (class == BPF_JMP) {
2094 /* BPF_EXIT for "main" will reach here. Return TRUE
2099 if (op == BPF_CALL) {
2100 /* BPF to BPF call will reach here because of marking
2101 * caller saved clobber with DST_OP_NO_MARK for which we
2102 * don't care the register def because they are anyway
2103 * marked as NOT_INIT already.
2105 if (insn->src_reg == BPF_PSEUDO_CALL)
2107 /* Helper call will reach here because of arg type
2108 * check, conservatively return TRUE.
2117 if (class == BPF_ALU64 || class == BPF_JMP ||
2118 /* BPF_END always use BPF_ALU class. */
2119 (class == BPF_ALU && op == BPF_END && insn->imm == 64))
2122 if (class == BPF_ALU || class == BPF_JMP32)
2125 if (class == BPF_LDX) {
2127 return BPF_SIZE(code) == BPF_DW;
2128 /* LDX source must be ptr. */
2132 if (class == BPF_STX) {
2133 /* BPF_STX (including atomic variants) has multiple source
2134 * operands, one of which is a ptr. Check whether the caller is
2137 if (t == SRC_OP && reg->type != SCALAR_VALUE)
2139 return BPF_SIZE(code) == BPF_DW;
2142 if (class == BPF_LD) {
2143 u8 mode = BPF_MODE(code);
2146 if (mode == BPF_IMM)
2149 /* Both LD_IND and LD_ABS return 32-bit data. */
2153 /* Implicit ctx ptr. */
2154 if (regno == BPF_REG_6)
2157 /* Explicit source could be any width. */
2161 if (class == BPF_ST)
2162 /* The only source register for BPF_ST is a ptr. */
2165 /* Conservatively return true at default. */
2169 /* Return the regno defined by the insn, or -1. */
2170 static int insn_def_regno(const struct bpf_insn *insn)
2172 switch (BPF_CLASS(insn->code)) {
2178 if (BPF_MODE(insn->code) == BPF_ATOMIC &&
2179 (insn->imm & BPF_FETCH)) {
2180 if (insn->imm == BPF_CMPXCHG)
2183 return insn->src_reg;
2188 return insn->dst_reg;
2192 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
2193 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
2195 int dst_reg = insn_def_regno(insn);
2200 return !is_reg64(env, insn, dst_reg, NULL, DST_OP);
2203 static void mark_insn_zext(struct bpf_verifier_env *env,
2204 struct bpf_reg_state *reg)
2206 s32 def_idx = reg->subreg_def;
2208 if (def_idx == DEF_NOT_SUBREG)
2211 env->insn_aux_data[def_idx - 1].zext_dst = true;
2212 /* The dst will be zero extended, so won't be sub-register anymore. */
2213 reg->subreg_def = DEF_NOT_SUBREG;
2216 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
2217 enum reg_arg_type t)
2219 struct bpf_verifier_state *vstate = env->cur_state;
2220 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2221 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
2222 struct bpf_reg_state *reg, *regs = state->regs;
2225 if (regno >= MAX_BPF_REG) {
2226 verbose(env, "R%d is invalid\n", regno);
2231 rw64 = is_reg64(env, insn, regno, reg, t);
2233 /* check whether register used as source operand can be read */
2234 if (reg->type == NOT_INIT) {
2235 verbose(env, "R%d !read_ok\n", regno);
2238 /* We don't need to worry about FP liveness because it's read-only */
2239 if (regno == BPF_REG_FP)
2243 mark_insn_zext(env, reg);
2245 return mark_reg_read(env, reg, reg->parent,
2246 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
2248 /* check whether register used as dest operand can be written to */
2249 if (regno == BPF_REG_FP) {
2250 verbose(env, "frame pointer is read only\n");
2253 reg->live |= REG_LIVE_WRITTEN;
2254 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
2256 mark_reg_unknown(env, regs, regno);
2261 /* for any branch, call, exit record the history of jmps in the given state */
2262 static int push_jmp_history(struct bpf_verifier_env *env,
2263 struct bpf_verifier_state *cur)
2265 u32 cnt = cur->jmp_history_cnt;
2266 struct bpf_idx_pair *p;
2269 p = krealloc(cur->jmp_history, cnt * sizeof(*p), GFP_USER);
2272 p[cnt - 1].idx = env->insn_idx;
2273 p[cnt - 1].prev_idx = env->prev_insn_idx;
2274 cur->jmp_history = p;
2275 cur->jmp_history_cnt = cnt;
2279 /* Backtrack one insn at a time. If idx is not at the top of recorded
2280 * history then previous instruction came from straight line execution.
2282 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
2287 if (cnt && st->jmp_history[cnt - 1].idx == i) {
2288 i = st->jmp_history[cnt - 1].prev_idx;
2296 static const char *disasm_kfunc_name(void *data, const struct bpf_insn *insn)
2298 const struct btf_type *func;
2299 struct btf *desc_btf;
2301 if (insn->src_reg != BPF_PSEUDO_KFUNC_CALL)
2304 desc_btf = find_kfunc_desc_btf(data, insn->imm, insn->off, NULL);
2305 if (IS_ERR(desc_btf))
2308 func = btf_type_by_id(desc_btf, insn->imm);
2309 return btf_name_by_offset(desc_btf, func->name_off);
2312 /* For given verifier state backtrack_insn() is called from the last insn to
2313 * the first insn. Its purpose is to compute a bitmask of registers and
2314 * stack slots that needs precision in the parent verifier state.
2316 static int backtrack_insn(struct bpf_verifier_env *env, int idx,
2317 u32 *reg_mask, u64 *stack_mask)
2319 const struct bpf_insn_cbs cbs = {
2320 .cb_call = disasm_kfunc_name,
2321 .cb_print = verbose,
2322 .private_data = env,
2324 struct bpf_insn *insn = env->prog->insnsi + idx;
2325 u8 class = BPF_CLASS(insn->code);
2326 u8 opcode = BPF_OP(insn->code);
2327 u8 mode = BPF_MODE(insn->code);
2328 u32 dreg = 1u << insn->dst_reg;
2329 u32 sreg = 1u << insn->src_reg;
2332 if (insn->code == 0)
2334 if (env->log.level & BPF_LOG_LEVEL) {
2335 verbose(env, "regs=%x stack=%llx before ", *reg_mask, *stack_mask);
2336 verbose(env, "%d: ", idx);
2337 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
2340 if (class == BPF_ALU || class == BPF_ALU64) {
2341 if (!(*reg_mask & dreg))
2343 if (opcode == BPF_MOV) {
2344 if (BPF_SRC(insn->code) == BPF_X) {
2346 * dreg needs precision after this insn
2347 * sreg needs precision before this insn
2353 * dreg needs precision after this insn.
2354 * Corresponding register is already marked
2355 * as precise=true in this verifier state.
2356 * No further markings in parent are necessary
2361 if (BPF_SRC(insn->code) == BPF_X) {
2363 * both dreg and sreg need precision
2368 * dreg still needs precision before this insn
2371 } else if (class == BPF_LDX) {
2372 if (!(*reg_mask & dreg))
2376 /* scalars can only be spilled into stack w/o losing precision.
2377 * Load from any other memory can be zero extended.
2378 * The desire to keep that precision is already indicated
2379 * by 'precise' mark in corresponding register of this state.
2380 * No further tracking necessary.
2382 if (insn->src_reg != BPF_REG_FP)
2384 if (BPF_SIZE(insn->code) != BPF_DW)
2387 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
2388 * that [fp - off] slot contains scalar that needs to be
2389 * tracked with precision
2391 spi = (-insn->off - 1) / BPF_REG_SIZE;
2393 verbose(env, "BUG spi %d\n", spi);
2394 WARN_ONCE(1, "verifier backtracking bug");
2397 *stack_mask |= 1ull << spi;
2398 } else if (class == BPF_STX || class == BPF_ST) {
2399 if (*reg_mask & dreg)
2400 /* stx & st shouldn't be using _scalar_ dst_reg
2401 * to access memory. It means backtracking
2402 * encountered a case of pointer subtraction.
2405 /* scalars can only be spilled into stack */
2406 if (insn->dst_reg != BPF_REG_FP)
2408 if (BPF_SIZE(insn->code) != BPF_DW)
2410 spi = (-insn->off - 1) / BPF_REG_SIZE;
2412 verbose(env, "BUG spi %d\n", spi);
2413 WARN_ONCE(1, "verifier backtracking bug");
2416 if (!(*stack_mask & (1ull << spi)))
2418 *stack_mask &= ~(1ull << spi);
2419 if (class == BPF_STX)
2421 } else if (class == BPF_JMP || class == BPF_JMP32) {
2422 if (opcode == BPF_CALL) {
2423 if (insn->src_reg == BPF_PSEUDO_CALL)
2425 /* regular helper call sets R0 */
2427 if (*reg_mask & 0x3f) {
2428 /* if backtracing was looking for registers R1-R5
2429 * they should have been found already.
2431 verbose(env, "BUG regs %x\n", *reg_mask);
2432 WARN_ONCE(1, "verifier backtracking bug");
2435 } else if (opcode == BPF_EXIT) {
2438 } else if (class == BPF_LD) {
2439 if (!(*reg_mask & dreg))
2442 /* It's ld_imm64 or ld_abs or ld_ind.
2443 * For ld_imm64 no further tracking of precision
2444 * into parent is necessary
2446 if (mode == BPF_IND || mode == BPF_ABS)
2447 /* to be analyzed */
2453 /* the scalar precision tracking algorithm:
2454 * . at the start all registers have precise=false.
2455 * . scalar ranges are tracked as normal through alu and jmp insns.
2456 * . once precise value of the scalar register is used in:
2457 * . ptr + scalar alu
2458 * . if (scalar cond K|scalar)
2459 * . helper_call(.., scalar, ...) where ARG_CONST is expected
2460 * backtrack through the verifier states and mark all registers and
2461 * stack slots with spilled constants that these scalar regisers
2462 * should be precise.
2463 * . during state pruning two registers (or spilled stack slots)
2464 * are equivalent if both are not precise.
2466 * Note the verifier cannot simply walk register parentage chain,
2467 * since many different registers and stack slots could have been
2468 * used to compute single precise scalar.
2470 * The approach of starting with precise=true for all registers and then
2471 * backtrack to mark a register as not precise when the verifier detects
2472 * that program doesn't care about specific value (e.g., when helper
2473 * takes register as ARG_ANYTHING parameter) is not safe.
2475 * It's ok to walk single parentage chain of the verifier states.
2476 * It's possible that this backtracking will go all the way till 1st insn.
2477 * All other branches will be explored for needing precision later.
2479 * The backtracking needs to deal with cases like:
2480 * 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)
2483 * if r5 > 0x79f goto pc+7
2484 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
2487 * call bpf_perf_event_output#25
2488 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
2492 * call foo // uses callee's r6 inside to compute r0
2496 * to track above reg_mask/stack_mask needs to be independent for each frame.
2498 * Also if parent's curframe > frame where backtracking started,
2499 * the verifier need to mark registers in both frames, otherwise callees
2500 * may incorrectly prune callers. This is similar to
2501 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
2503 * For now backtracking falls back into conservative marking.
2505 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
2506 struct bpf_verifier_state *st)
2508 struct bpf_func_state *func;
2509 struct bpf_reg_state *reg;
2512 /* big hammer: mark all scalars precise in this path.
2513 * pop_stack may still get !precise scalars.
2515 for (; st; st = st->parent)
2516 for (i = 0; i <= st->curframe; i++) {
2517 func = st->frame[i];
2518 for (j = 0; j < BPF_REG_FP; j++) {
2519 reg = &func->regs[j];
2520 if (reg->type != SCALAR_VALUE)
2522 reg->precise = true;
2524 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
2525 if (!is_spilled_reg(&func->stack[j]))
2527 reg = &func->stack[j].spilled_ptr;
2528 if (reg->type != SCALAR_VALUE)
2530 reg->precise = true;
2535 static int __mark_chain_precision(struct bpf_verifier_env *env, int regno,
2538 struct bpf_verifier_state *st = env->cur_state;
2539 int first_idx = st->first_insn_idx;
2540 int last_idx = env->insn_idx;
2541 struct bpf_func_state *func;
2542 struct bpf_reg_state *reg;
2543 u32 reg_mask = regno >= 0 ? 1u << regno : 0;
2544 u64 stack_mask = spi >= 0 ? 1ull << spi : 0;
2545 bool skip_first = true;
2546 bool new_marks = false;
2549 if (!env->bpf_capable)
2552 func = st->frame[st->curframe];
2554 reg = &func->regs[regno];
2555 if (reg->type != SCALAR_VALUE) {
2556 WARN_ONCE(1, "backtracing misuse");
2563 reg->precise = true;
2567 if (!is_spilled_reg(&func->stack[spi])) {
2571 reg = &func->stack[spi].spilled_ptr;
2572 if (reg->type != SCALAR_VALUE) {
2580 reg->precise = true;
2586 if (!reg_mask && !stack_mask)
2589 DECLARE_BITMAP(mask, 64);
2590 u32 history = st->jmp_history_cnt;
2592 if (env->log.level & BPF_LOG_LEVEL)
2593 verbose(env, "last_idx %d first_idx %d\n", last_idx, first_idx);
2594 for (i = last_idx;;) {
2599 err = backtrack_insn(env, i, ®_mask, &stack_mask);
2601 if (err == -ENOTSUPP) {
2602 mark_all_scalars_precise(env, st);
2607 if (!reg_mask && !stack_mask)
2608 /* Found assignment(s) into tracked register in this state.
2609 * Since this state is already marked, just return.
2610 * Nothing to be tracked further in the parent state.
2615 i = get_prev_insn_idx(st, i, &history);
2616 if (i >= env->prog->len) {
2617 /* This can happen if backtracking reached insn 0
2618 * and there are still reg_mask or stack_mask
2620 * It means the backtracking missed the spot where
2621 * particular register was initialized with a constant.
2623 verbose(env, "BUG backtracking idx %d\n", i);
2624 WARN_ONCE(1, "verifier backtracking bug");
2633 func = st->frame[st->curframe];
2634 bitmap_from_u64(mask, reg_mask);
2635 for_each_set_bit(i, mask, 32) {
2636 reg = &func->regs[i];
2637 if (reg->type != SCALAR_VALUE) {
2638 reg_mask &= ~(1u << i);
2643 reg->precise = true;
2646 bitmap_from_u64(mask, stack_mask);
2647 for_each_set_bit(i, mask, 64) {
2648 if (i >= func->allocated_stack / BPF_REG_SIZE) {
2649 /* the sequence of instructions:
2651 * 3: (7b) *(u64 *)(r3 -8) = r0
2652 * 4: (79) r4 = *(u64 *)(r10 -8)
2653 * doesn't contain jmps. It's backtracked
2654 * as a single block.
2655 * During backtracking insn 3 is not recognized as
2656 * stack access, so at the end of backtracking
2657 * stack slot fp-8 is still marked in stack_mask.
2658 * However the parent state may not have accessed
2659 * fp-8 and it's "unallocated" stack space.
2660 * In such case fallback to conservative.
2662 mark_all_scalars_precise(env, st);
2666 if (!is_spilled_reg(&func->stack[i])) {
2667 stack_mask &= ~(1ull << i);
2670 reg = &func->stack[i].spilled_ptr;
2671 if (reg->type != SCALAR_VALUE) {
2672 stack_mask &= ~(1ull << i);
2677 reg->precise = true;
2679 if (env->log.level & BPF_LOG_LEVEL) {
2680 print_verifier_state(env, func);
2681 verbose(env, "parent %s regs=%x stack=%llx marks\n",
2682 new_marks ? "didn't have" : "already had",
2683 reg_mask, stack_mask);
2686 if (!reg_mask && !stack_mask)
2691 last_idx = st->last_insn_idx;
2692 first_idx = st->first_insn_idx;
2697 static int mark_chain_precision(struct bpf_verifier_env *env, int regno)
2699 return __mark_chain_precision(env, regno, -1);
2702 static int mark_chain_precision_stack(struct bpf_verifier_env *env, int spi)
2704 return __mark_chain_precision(env, -1, spi);
2707 static bool is_spillable_regtype(enum bpf_reg_type type)
2710 case PTR_TO_MAP_VALUE:
2711 case PTR_TO_MAP_VALUE_OR_NULL:
2715 case PTR_TO_PACKET_META:
2716 case PTR_TO_PACKET_END:
2717 case PTR_TO_FLOW_KEYS:
2718 case CONST_PTR_TO_MAP:
2720 case PTR_TO_SOCKET_OR_NULL:
2721 case PTR_TO_SOCK_COMMON:
2722 case PTR_TO_SOCK_COMMON_OR_NULL:
2723 case PTR_TO_TCP_SOCK:
2724 case PTR_TO_TCP_SOCK_OR_NULL:
2725 case PTR_TO_XDP_SOCK:
2727 case PTR_TO_BTF_ID_OR_NULL:
2728 case PTR_TO_RDONLY_BUF:
2729 case PTR_TO_RDONLY_BUF_OR_NULL:
2730 case PTR_TO_RDWR_BUF:
2731 case PTR_TO_RDWR_BUF_OR_NULL:
2732 case PTR_TO_PERCPU_BTF_ID:
2734 case PTR_TO_MEM_OR_NULL:
2736 case PTR_TO_MAP_KEY:
2743 /* Does this register contain a constant zero? */
2744 static bool register_is_null(struct bpf_reg_state *reg)
2746 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
2749 static bool register_is_const(struct bpf_reg_state *reg)
2751 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
2754 static bool __is_scalar_unbounded(struct bpf_reg_state *reg)
2756 return tnum_is_unknown(reg->var_off) &&
2757 reg->smin_value == S64_MIN && reg->smax_value == S64_MAX &&
2758 reg->umin_value == 0 && reg->umax_value == U64_MAX &&
2759 reg->s32_min_value == S32_MIN && reg->s32_max_value == S32_MAX &&
2760 reg->u32_min_value == 0 && reg->u32_max_value == U32_MAX;
2763 static bool register_is_bounded(struct bpf_reg_state *reg)
2765 return reg->type == SCALAR_VALUE && !__is_scalar_unbounded(reg);
2768 static bool __is_pointer_value(bool allow_ptr_leaks,
2769 const struct bpf_reg_state *reg)
2771 if (allow_ptr_leaks)
2774 return reg->type != SCALAR_VALUE;
2777 static void save_register_state(struct bpf_func_state *state,
2778 int spi, struct bpf_reg_state *reg,
2783 state->stack[spi].spilled_ptr = *reg;
2784 if (size == BPF_REG_SIZE)
2785 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2787 for (i = BPF_REG_SIZE; i > BPF_REG_SIZE - size; i--)
2788 state->stack[spi].slot_type[i - 1] = STACK_SPILL;
2790 /* size < 8 bytes spill */
2792 scrub_spilled_slot(&state->stack[spi].slot_type[i - 1]);
2795 /* check_stack_{read,write}_fixed_off functions track spill/fill of registers,
2796 * stack boundary and alignment are checked in check_mem_access()
2798 static int check_stack_write_fixed_off(struct bpf_verifier_env *env,
2799 /* stack frame we're writing to */
2800 struct bpf_func_state *state,
2801 int off, int size, int value_regno,
2804 struct bpf_func_state *cur; /* state of the current function */
2805 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
2806 u32 dst_reg = env->prog->insnsi[insn_idx].dst_reg;
2807 struct bpf_reg_state *reg = NULL;
2809 err = grow_stack_state(state, round_up(slot + 1, BPF_REG_SIZE));
2812 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
2813 * so it's aligned access and [off, off + size) are within stack limits
2815 if (!env->allow_ptr_leaks &&
2816 state->stack[spi].slot_type[0] == STACK_SPILL &&
2817 size != BPF_REG_SIZE) {
2818 verbose(env, "attempt to corrupt spilled pointer on stack\n");
2822 cur = env->cur_state->frame[env->cur_state->curframe];
2823 if (value_regno >= 0)
2824 reg = &cur->regs[value_regno];
2825 if (!env->bypass_spec_v4) {
2826 bool sanitize = reg && is_spillable_regtype(reg->type);
2828 for (i = 0; i < size; i++) {
2829 if (state->stack[spi].slot_type[i] == STACK_INVALID) {
2836 env->insn_aux_data[insn_idx].sanitize_stack_spill = true;
2839 if (reg && !(off % BPF_REG_SIZE) && register_is_bounded(reg) &&
2840 !register_is_null(reg) && env->bpf_capable) {
2841 if (dst_reg != BPF_REG_FP) {
2842 /* The backtracking logic can only recognize explicit
2843 * stack slot address like [fp - 8]. Other spill of
2844 * scalar via different register has to be conservative.
2845 * Backtrack from here and mark all registers as precise
2846 * that contributed into 'reg' being a constant.
2848 err = mark_chain_precision(env, value_regno);
2852 save_register_state(state, spi, reg, size);
2853 } else if (reg && is_spillable_regtype(reg->type)) {
2854 /* register containing pointer is being spilled into stack */
2855 if (size != BPF_REG_SIZE) {
2856 verbose_linfo(env, insn_idx, "; ");
2857 verbose(env, "invalid size of register spill\n");
2860 if (state != cur && reg->type == PTR_TO_STACK) {
2861 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
2864 save_register_state(state, spi, reg, size);
2866 u8 type = STACK_MISC;
2868 /* regular write of data into stack destroys any spilled ptr */
2869 state->stack[spi].spilled_ptr.type = NOT_INIT;
2870 /* Mark slots as STACK_MISC if they belonged to spilled ptr. */
2871 if (is_spilled_reg(&state->stack[spi]))
2872 for (i = 0; i < BPF_REG_SIZE; i++)
2873 scrub_spilled_slot(&state->stack[spi].slot_type[i]);
2875 /* only mark the slot as written if all 8 bytes were written
2876 * otherwise read propagation may incorrectly stop too soon
2877 * when stack slots are partially written.
2878 * This heuristic means that read propagation will be
2879 * conservative, since it will add reg_live_read marks
2880 * to stack slots all the way to first state when programs
2881 * writes+reads less than 8 bytes
2883 if (size == BPF_REG_SIZE)
2884 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2886 /* when we zero initialize stack slots mark them as such */
2887 if (reg && register_is_null(reg)) {
2888 /* backtracking doesn't work for STACK_ZERO yet. */
2889 err = mark_chain_precision(env, value_regno);
2895 /* Mark slots affected by this stack write. */
2896 for (i = 0; i < size; i++)
2897 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
2903 /* Write the stack: 'stack[ptr_regno + off] = value_regno'. 'ptr_regno' is
2904 * known to contain a variable offset.
2905 * This function checks whether the write is permitted and conservatively
2906 * tracks the effects of the write, considering that each stack slot in the
2907 * dynamic range is potentially written to.
2909 * 'off' includes 'regno->off'.
2910 * 'value_regno' can be -1, meaning that an unknown value is being written to
2913 * Spilled pointers in range are not marked as written because we don't know
2914 * what's going to be actually written. This means that read propagation for
2915 * future reads cannot be terminated by this write.
2917 * For privileged programs, uninitialized stack slots are considered
2918 * initialized by this write (even though we don't know exactly what offsets
2919 * are going to be written to). The idea is that we don't want the verifier to
2920 * reject future reads that access slots written to through variable offsets.
2922 static int check_stack_write_var_off(struct bpf_verifier_env *env,
2923 /* func where register points to */
2924 struct bpf_func_state *state,
2925 int ptr_regno, int off, int size,
2926 int value_regno, int insn_idx)
2928 struct bpf_func_state *cur; /* state of the current function */
2929 int min_off, max_off;
2931 struct bpf_reg_state *ptr_reg = NULL, *value_reg = NULL;
2932 bool writing_zero = false;
2933 /* set if the fact that we're writing a zero is used to let any
2934 * stack slots remain STACK_ZERO
2936 bool zero_used = false;
2938 cur = env->cur_state->frame[env->cur_state->curframe];
2939 ptr_reg = &cur->regs[ptr_regno];
2940 min_off = ptr_reg->smin_value + off;
2941 max_off = ptr_reg->smax_value + off + size;
2942 if (value_regno >= 0)
2943 value_reg = &cur->regs[value_regno];
2944 if (value_reg && register_is_null(value_reg))
2945 writing_zero = true;
2947 err = grow_stack_state(state, round_up(-min_off, BPF_REG_SIZE));
2952 /* Variable offset writes destroy any spilled pointers in range. */
2953 for (i = min_off; i < max_off; i++) {
2954 u8 new_type, *stype;
2958 spi = slot / BPF_REG_SIZE;
2959 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
2961 if (!env->allow_ptr_leaks
2962 && *stype != NOT_INIT
2963 && *stype != SCALAR_VALUE) {
2964 /* Reject the write if there's are spilled pointers in
2965 * range. If we didn't reject here, the ptr status
2966 * would be erased below (even though not all slots are
2967 * actually overwritten), possibly opening the door to
2970 verbose(env, "spilled ptr in range of var-offset stack write; insn %d, ptr off: %d",
2975 /* Erase all spilled pointers. */
2976 state->stack[spi].spilled_ptr.type = NOT_INIT;
2978 /* Update the slot type. */
2979 new_type = STACK_MISC;
2980 if (writing_zero && *stype == STACK_ZERO) {
2981 new_type = STACK_ZERO;
2984 /* If the slot is STACK_INVALID, we check whether it's OK to
2985 * pretend that it will be initialized by this write. The slot
2986 * might not actually be written to, and so if we mark it as
2987 * initialized future reads might leak uninitialized memory.
2988 * For privileged programs, we will accept such reads to slots
2989 * that may or may not be written because, if we're reject
2990 * them, the error would be too confusing.
2992 if (*stype == STACK_INVALID && !env->allow_uninit_stack) {
2993 verbose(env, "uninit stack in range of var-offset write prohibited for !root; insn %d, off: %d",
3000 /* backtracking doesn't work for STACK_ZERO yet. */
3001 err = mark_chain_precision(env, value_regno);
3008 /* When register 'dst_regno' is assigned some values from stack[min_off,
3009 * max_off), we set the register's type according to the types of the
3010 * respective stack slots. If all the stack values are known to be zeros, then
3011 * so is the destination reg. Otherwise, the register is considered to be
3012 * SCALAR. This function does not deal with register filling; the caller must
3013 * ensure that all spilled registers in the stack range have been marked as
3016 static void mark_reg_stack_read(struct bpf_verifier_env *env,
3017 /* func where src register points to */
3018 struct bpf_func_state *ptr_state,
3019 int min_off, int max_off, int dst_regno)
3021 struct bpf_verifier_state *vstate = env->cur_state;
3022 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3027 for (i = min_off; i < max_off; i++) {
3029 spi = slot / BPF_REG_SIZE;
3030 stype = ptr_state->stack[spi].slot_type;
3031 if (stype[slot % BPF_REG_SIZE] != STACK_ZERO)
3035 if (zeros == max_off - min_off) {
3036 /* any access_size read into register is zero extended,
3037 * so the whole register == const_zero
3039 __mark_reg_const_zero(&state->regs[dst_regno]);
3040 /* backtracking doesn't support STACK_ZERO yet,
3041 * so mark it precise here, so that later
3042 * backtracking can stop here.
3043 * Backtracking may not need this if this register
3044 * doesn't participate in pointer adjustment.
3045 * Forward propagation of precise flag is not
3046 * necessary either. This mark is only to stop
3047 * backtracking. Any register that contributed
3048 * to const 0 was marked precise before spill.
3050 state->regs[dst_regno].precise = true;
3052 /* have read misc data from the stack */
3053 mark_reg_unknown(env, state->regs, dst_regno);
3055 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
3058 /* Read the stack at 'off' and put the results into the register indicated by
3059 * 'dst_regno'. It handles reg filling if the addressed stack slot is a
3062 * 'dst_regno' can be -1, meaning that the read value is not going to a
3065 * The access is assumed to be within the current stack bounds.
3067 static int check_stack_read_fixed_off(struct bpf_verifier_env *env,
3068 /* func where src register points to */
3069 struct bpf_func_state *reg_state,
3070 int off, int size, int dst_regno)
3072 struct bpf_verifier_state *vstate = env->cur_state;
3073 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3074 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
3075 struct bpf_reg_state *reg;
3078 stype = reg_state->stack[spi].slot_type;
3079 reg = ®_state->stack[spi].spilled_ptr;
3081 if (is_spilled_reg(®_state->stack[spi])) {
3084 for (i = BPF_REG_SIZE - 1; i > 0 && stype[i - 1] == STACK_SPILL; i--)
3087 if (size != BPF_REG_SIZE || spill_size != BPF_REG_SIZE) {
3088 if (reg->type != SCALAR_VALUE) {
3089 verbose_linfo(env, env->insn_idx, "; ");
3090 verbose(env, "invalid size of register fill\n");
3094 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
3098 if (!(off % BPF_REG_SIZE) && size == spill_size) {
3099 /* The earlier check_reg_arg() has decided the
3100 * subreg_def for this insn. Save it first.
3102 s32 subreg_def = state->regs[dst_regno].subreg_def;
3104 state->regs[dst_regno] = *reg;
3105 state->regs[dst_regno].subreg_def = subreg_def;
3107 for (i = 0; i < size; i++) {
3108 type = stype[(slot - i) % BPF_REG_SIZE];
3109 if (type == STACK_SPILL)
3111 if (type == STACK_MISC)
3113 verbose(env, "invalid read from stack off %d+%d size %d\n",
3117 mark_reg_unknown(env, state->regs, dst_regno);
3119 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
3123 if (dst_regno >= 0) {
3124 /* restore register state from stack */
3125 state->regs[dst_regno] = *reg;
3126 /* mark reg as written since spilled pointer state likely
3127 * has its liveness marks cleared by is_state_visited()
3128 * which resets stack/reg liveness for state transitions
3130 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
3131 } else if (__is_pointer_value(env->allow_ptr_leaks, reg)) {
3132 /* If dst_regno==-1, the caller is asking us whether
3133 * it is acceptable to use this value as a SCALAR_VALUE
3135 * We must not allow unprivileged callers to do that
3136 * with spilled pointers.
3138 verbose(env, "leaking pointer from stack off %d\n",
3142 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
3144 for (i = 0; i < size; i++) {
3145 type = stype[(slot - i) % BPF_REG_SIZE];
3146 if (type == STACK_MISC)
3148 if (type == STACK_ZERO)
3150 verbose(env, "invalid read from stack off %d+%d size %d\n",
3154 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
3156 mark_reg_stack_read(env, reg_state, off, off + size, dst_regno);
3161 enum stack_access_src {
3162 ACCESS_DIRECT = 1, /* the access is performed by an instruction */
3163 ACCESS_HELPER = 2, /* the access is performed by a helper */
3166 static int check_stack_range_initialized(struct bpf_verifier_env *env,
3167 int regno, int off, int access_size,
3168 bool zero_size_allowed,
3169 enum stack_access_src type,
3170 struct bpf_call_arg_meta *meta);
3172 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
3174 return cur_regs(env) + regno;
3177 /* Read the stack at 'ptr_regno + off' and put the result into the register
3179 * 'off' includes the pointer register's fixed offset(i.e. 'ptr_regno.off'),
3180 * but not its variable offset.
3181 * 'size' is assumed to be <= reg size and the access is assumed to be aligned.
3183 * As opposed to check_stack_read_fixed_off, this function doesn't deal with
3184 * filling registers (i.e. reads of spilled register cannot be detected when
3185 * the offset is not fixed). We conservatively mark 'dst_regno' as containing
3186 * SCALAR_VALUE. That's why we assert that the 'ptr_regno' has a variable
3187 * offset; for a fixed offset check_stack_read_fixed_off should be used
3190 static int check_stack_read_var_off(struct bpf_verifier_env *env,
3191 int ptr_regno, int off, int size, int dst_regno)
3193 /* The state of the source register. */
3194 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
3195 struct bpf_func_state *ptr_state = func(env, reg);
3197 int min_off, max_off;
3199 /* Note that we pass a NULL meta, so raw access will not be permitted.
3201 err = check_stack_range_initialized(env, ptr_regno, off, size,
3202 false, ACCESS_DIRECT, NULL);
3206 min_off = reg->smin_value + off;
3207 max_off = reg->smax_value + off;
3208 mark_reg_stack_read(env, ptr_state, min_off, max_off + size, dst_regno);
3212 /* check_stack_read dispatches to check_stack_read_fixed_off or
3213 * check_stack_read_var_off.
3215 * The caller must ensure that the offset falls within the allocated stack
3218 * 'dst_regno' is a register which will receive the value from the stack. It
3219 * can be -1, meaning that the read value is not going to a register.
3221 static int check_stack_read(struct bpf_verifier_env *env,
3222 int ptr_regno, int off, int size,
3225 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
3226 struct bpf_func_state *state = func(env, reg);
3228 /* Some accesses are only permitted with a static offset. */
3229 bool var_off = !tnum_is_const(reg->var_off);
3231 /* The offset is required to be static when reads don't go to a
3232 * register, in order to not leak pointers (see
3233 * check_stack_read_fixed_off).
3235 if (dst_regno < 0 && var_off) {
3238 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3239 verbose(env, "variable offset stack pointer cannot be passed into helper function; var_off=%s off=%d size=%d\n",
3243 /* Variable offset is prohibited for unprivileged mode for simplicity
3244 * since it requires corresponding support in Spectre masking for stack
3245 * ALU. See also retrieve_ptr_limit().
3247 if (!env->bypass_spec_v1 && var_off) {
3250 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3251 verbose(env, "R%d variable offset stack access prohibited for !root, var_off=%s\n",
3257 off += reg->var_off.value;
3258 err = check_stack_read_fixed_off(env, state, off, size,
3261 /* Variable offset stack reads need more conservative handling
3262 * than fixed offset ones. Note that dst_regno >= 0 on this
3265 err = check_stack_read_var_off(env, ptr_regno, off, size,
3272 /* check_stack_write dispatches to check_stack_write_fixed_off or
3273 * check_stack_write_var_off.
3275 * 'ptr_regno' is the register used as a pointer into the stack.
3276 * 'off' includes 'ptr_regno->off', but not its variable offset (if any).
3277 * 'value_regno' is the register whose value we're writing to the stack. It can
3278 * be -1, meaning that we're not writing from a register.
3280 * The caller must ensure that the offset falls within the maximum stack size.
3282 static int check_stack_write(struct bpf_verifier_env *env,
3283 int ptr_regno, int off, int size,
3284 int value_regno, int insn_idx)
3286 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
3287 struct bpf_func_state *state = func(env, reg);
3290 if (tnum_is_const(reg->var_off)) {
3291 off += reg->var_off.value;
3292 err = check_stack_write_fixed_off(env, state, off, size,
3293 value_regno, insn_idx);
3295 /* Variable offset stack reads need more conservative handling
3296 * than fixed offset ones.
3298 err = check_stack_write_var_off(env, state,
3299 ptr_regno, off, size,
3300 value_regno, insn_idx);
3305 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
3306 int off, int size, enum bpf_access_type type)
3308 struct bpf_reg_state *regs = cur_regs(env);
3309 struct bpf_map *map = regs[regno].map_ptr;
3310 u32 cap = bpf_map_flags_to_cap(map);
3312 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
3313 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
3314 map->value_size, off, size);
3318 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
3319 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
3320 map->value_size, off, size);
3327 /* check read/write into memory region (e.g., map value, ringbuf sample, etc) */
3328 static int __check_mem_access(struct bpf_verifier_env *env, int regno,
3329 int off, int size, u32 mem_size,
3330 bool zero_size_allowed)
3332 bool size_ok = size > 0 || (size == 0 && zero_size_allowed);
3333 struct bpf_reg_state *reg;
3335 if (off >= 0 && size_ok && (u64)off + size <= mem_size)
3338 reg = &cur_regs(env)[regno];
3339 switch (reg->type) {
3340 case PTR_TO_MAP_KEY:
3341 verbose(env, "invalid access to map key, key_size=%d off=%d size=%d\n",
3342 mem_size, off, size);
3344 case PTR_TO_MAP_VALUE:
3345 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
3346 mem_size, off, size);
3349 case PTR_TO_PACKET_META:
3350 case PTR_TO_PACKET_END:
3351 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
3352 off, size, regno, reg->id, off, mem_size);
3356 verbose(env, "invalid access to memory, mem_size=%u off=%d size=%d\n",
3357 mem_size, off, size);
3363 /* check read/write into a memory region with possible variable offset */
3364 static int check_mem_region_access(struct bpf_verifier_env *env, u32 regno,
3365 int off, int size, u32 mem_size,
3366 bool zero_size_allowed)
3368 struct bpf_verifier_state *vstate = env->cur_state;
3369 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3370 struct bpf_reg_state *reg = &state->regs[regno];
3373 /* We may have adjusted the register pointing to memory region, so we
3374 * need to try adding each of min_value and max_value to off
3375 * to make sure our theoretical access will be safe.
3377 if (env->log.level & BPF_LOG_LEVEL)
3378 print_verifier_state(env, state);
3380 /* The minimum value is only important with signed
3381 * comparisons where we can't assume the floor of a
3382 * value is 0. If we are using signed variables for our
3383 * index'es we need to make sure that whatever we use
3384 * will have a set floor within our range.
3386 if (reg->smin_value < 0 &&
3387 (reg->smin_value == S64_MIN ||
3388 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
3389 reg->smin_value + off < 0)) {
3390 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3394 err = __check_mem_access(env, regno, reg->smin_value + off, size,
3395 mem_size, zero_size_allowed);
3397 verbose(env, "R%d min value is outside of the allowed memory range\n",
3402 /* If we haven't set a max value then we need to bail since we can't be
3403 * sure we won't do bad things.
3404 * If reg->umax_value + off could overflow, treat that as unbounded too.
3406 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
3407 verbose(env, "R%d unbounded memory access, make sure to bounds check any such access\n",
3411 err = __check_mem_access(env, regno, reg->umax_value + off, size,
3412 mem_size, zero_size_allowed);
3414 verbose(env, "R%d max value is outside of the allowed memory range\n",
3422 /* check read/write into a map element with possible variable offset */
3423 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
3424 int off, int size, bool zero_size_allowed)
3426 struct bpf_verifier_state *vstate = env->cur_state;
3427 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3428 struct bpf_reg_state *reg = &state->regs[regno];
3429 struct bpf_map *map = reg->map_ptr;
3432 err = check_mem_region_access(env, regno, off, size, map->value_size,
3437 if (map_value_has_spin_lock(map)) {
3438 u32 lock = map->spin_lock_off;
3440 /* if any part of struct bpf_spin_lock can be touched by
3441 * load/store reject this program.
3442 * To check that [x1, x2) overlaps with [y1, y2)
3443 * it is sufficient to check x1 < y2 && y1 < x2.
3445 if (reg->smin_value + off < lock + sizeof(struct bpf_spin_lock) &&
3446 lock < reg->umax_value + off + size) {
3447 verbose(env, "bpf_spin_lock cannot be accessed directly by load/store\n");
3451 if (map_value_has_timer(map)) {
3452 u32 t = map->timer_off;
3454 if (reg->smin_value + off < t + sizeof(struct bpf_timer) &&
3455 t < reg->umax_value + off + size) {
3456 verbose(env, "bpf_timer cannot be accessed directly by load/store\n");
3463 #define MAX_PACKET_OFF 0xffff
3465 static enum bpf_prog_type resolve_prog_type(struct bpf_prog *prog)
3467 return prog->aux->dst_prog ? prog->aux->dst_prog->type : prog->type;
3470 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
3471 const struct bpf_call_arg_meta *meta,
3472 enum bpf_access_type t)
3474 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
3476 switch (prog_type) {
3477 /* Program types only with direct read access go here! */
3478 case BPF_PROG_TYPE_LWT_IN:
3479 case BPF_PROG_TYPE_LWT_OUT:
3480 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
3481 case BPF_PROG_TYPE_SK_REUSEPORT:
3482 case BPF_PROG_TYPE_FLOW_DISSECTOR:
3483 case BPF_PROG_TYPE_CGROUP_SKB:
3488 /* Program types with direct read + write access go here! */
3489 case BPF_PROG_TYPE_SCHED_CLS:
3490 case BPF_PROG_TYPE_SCHED_ACT:
3491 case BPF_PROG_TYPE_XDP:
3492 case BPF_PROG_TYPE_LWT_XMIT:
3493 case BPF_PROG_TYPE_SK_SKB:
3494 case BPF_PROG_TYPE_SK_MSG:
3496 return meta->pkt_access;
3498 env->seen_direct_write = true;
3501 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
3503 env->seen_direct_write = true;
3512 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
3513 int size, bool zero_size_allowed)
3515 struct bpf_reg_state *regs = cur_regs(env);
3516 struct bpf_reg_state *reg = ®s[regno];
3519 /* We may have added a variable offset to the packet pointer; but any
3520 * reg->range we have comes after that. We are only checking the fixed
3524 /* We don't allow negative numbers, because we aren't tracking enough
3525 * detail to prove they're safe.
3527 if (reg->smin_value < 0) {
3528 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3533 err = reg->range < 0 ? -EINVAL :
3534 __check_mem_access(env, regno, off, size, reg->range,
3537 verbose(env, "R%d offset is outside of the packet\n", regno);
3541 /* __check_mem_access has made sure "off + size - 1" is within u16.
3542 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
3543 * otherwise find_good_pkt_pointers would have refused to set range info
3544 * that __check_mem_access would have rejected this pkt access.
3545 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
3547 env->prog->aux->max_pkt_offset =
3548 max_t(u32, env->prog->aux->max_pkt_offset,
3549 off + reg->umax_value + size - 1);
3554 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
3555 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
3556 enum bpf_access_type t, enum bpf_reg_type *reg_type,
3557 struct btf **btf, u32 *btf_id)
3559 struct bpf_insn_access_aux info = {
3560 .reg_type = *reg_type,
3564 if (env->ops->is_valid_access &&
3565 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
3566 /* A non zero info.ctx_field_size indicates that this field is a
3567 * candidate for later verifier transformation to load the whole
3568 * field and then apply a mask when accessed with a narrower
3569 * access than actual ctx access size. A zero info.ctx_field_size
3570 * will only allow for whole field access and rejects any other
3571 * type of narrower access.
3573 *reg_type = info.reg_type;
3575 if (*reg_type == PTR_TO_BTF_ID || *reg_type == PTR_TO_BTF_ID_OR_NULL) {
3577 *btf_id = info.btf_id;
3579 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
3581 /* remember the offset of last byte accessed in ctx */
3582 if (env->prog->aux->max_ctx_offset < off + size)
3583 env->prog->aux->max_ctx_offset = off + size;
3587 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
3591 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
3594 if (size < 0 || off < 0 ||
3595 (u64)off + size > sizeof(struct bpf_flow_keys)) {
3596 verbose(env, "invalid access to flow keys off=%d size=%d\n",
3603 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
3604 u32 regno, int off, int size,
3605 enum bpf_access_type t)
3607 struct bpf_reg_state *regs = cur_regs(env);
3608 struct bpf_reg_state *reg = ®s[regno];
3609 struct bpf_insn_access_aux info = {};
3612 if (reg->smin_value < 0) {
3613 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3618 switch (reg->type) {
3619 case PTR_TO_SOCK_COMMON:
3620 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
3623 valid = bpf_sock_is_valid_access(off, size, t, &info);
3625 case PTR_TO_TCP_SOCK:
3626 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
3628 case PTR_TO_XDP_SOCK:
3629 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
3637 env->insn_aux_data[insn_idx].ctx_field_size =
3638 info.ctx_field_size;
3642 verbose(env, "R%d invalid %s access off=%d size=%d\n",
3643 regno, reg_type_str[reg->type], off, size);
3648 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
3650 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
3653 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
3655 const struct bpf_reg_state *reg = reg_state(env, regno);
3657 return reg->type == PTR_TO_CTX;
3660 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
3662 const struct bpf_reg_state *reg = reg_state(env, regno);
3664 return type_is_sk_pointer(reg->type);
3667 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
3669 const struct bpf_reg_state *reg = reg_state(env, regno);
3671 return type_is_pkt_pointer(reg->type);
3674 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
3676 const struct bpf_reg_state *reg = reg_state(env, regno);
3678 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
3679 return reg->type == PTR_TO_FLOW_KEYS;
3682 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
3683 const struct bpf_reg_state *reg,
3684 int off, int size, bool strict)
3686 struct tnum reg_off;
3689 /* Byte size accesses are always allowed. */
3690 if (!strict || size == 1)
3693 /* For platforms that do not have a Kconfig enabling
3694 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
3695 * NET_IP_ALIGN is universally set to '2'. And on platforms
3696 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
3697 * to this code only in strict mode where we want to emulate
3698 * the NET_IP_ALIGN==2 checking. Therefore use an
3699 * unconditional IP align value of '2'.
3703 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
3704 if (!tnum_is_aligned(reg_off, size)) {
3707 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3709 "misaligned packet access off %d+%s+%d+%d size %d\n",
3710 ip_align, tn_buf, reg->off, off, size);
3717 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
3718 const struct bpf_reg_state *reg,
3719 const char *pointer_desc,
3720 int off, int size, bool strict)
3722 struct tnum reg_off;
3724 /* Byte size accesses are always allowed. */
3725 if (!strict || size == 1)
3728 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
3729 if (!tnum_is_aligned(reg_off, size)) {
3732 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3733 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
3734 pointer_desc, tn_buf, reg->off, off, size);
3741 static int check_ptr_alignment(struct bpf_verifier_env *env,
3742 const struct bpf_reg_state *reg, int off,
3743 int size, bool strict_alignment_once)
3745 bool strict = env->strict_alignment || strict_alignment_once;
3746 const char *pointer_desc = "";
3748 switch (reg->type) {
3750 case PTR_TO_PACKET_META:
3751 /* Special case, because of NET_IP_ALIGN. Given metadata sits
3752 * right in front, treat it the very same way.
3754 return check_pkt_ptr_alignment(env, reg, off, size, strict);
3755 case PTR_TO_FLOW_KEYS:
3756 pointer_desc = "flow keys ";
3758 case PTR_TO_MAP_KEY:
3759 pointer_desc = "key ";
3761 case PTR_TO_MAP_VALUE:
3762 pointer_desc = "value ";
3765 pointer_desc = "context ";
3768 pointer_desc = "stack ";
3769 /* The stack spill tracking logic in check_stack_write_fixed_off()
3770 * and check_stack_read_fixed_off() relies on stack accesses being
3776 pointer_desc = "sock ";
3778 case PTR_TO_SOCK_COMMON:
3779 pointer_desc = "sock_common ";
3781 case PTR_TO_TCP_SOCK:
3782 pointer_desc = "tcp_sock ";
3784 case PTR_TO_XDP_SOCK:
3785 pointer_desc = "xdp_sock ";
3790 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
3794 static int update_stack_depth(struct bpf_verifier_env *env,
3795 const struct bpf_func_state *func,
3798 u16 stack = env->subprog_info[func->subprogno].stack_depth;
3803 /* update known max for given subprogram */
3804 env->subprog_info[func->subprogno].stack_depth = -off;
3808 /* starting from main bpf function walk all instructions of the function
3809 * and recursively walk all callees that given function can call.
3810 * Ignore jump and exit insns.
3811 * Since recursion is prevented by check_cfg() this algorithm
3812 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
3814 static int check_max_stack_depth(struct bpf_verifier_env *env)
3816 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
3817 struct bpf_subprog_info *subprog = env->subprog_info;
3818 struct bpf_insn *insn = env->prog->insnsi;
3819 bool tail_call_reachable = false;
3820 int ret_insn[MAX_CALL_FRAMES];
3821 int ret_prog[MAX_CALL_FRAMES];
3825 /* protect against potential stack overflow that might happen when
3826 * bpf2bpf calls get combined with tailcalls. Limit the caller's stack
3827 * depth for such case down to 256 so that the worst case scenario
3828 * would result in 8k stack size (32 which is tailcall limit * 256 =
3831 * To get the idea what might happen, see an example:
3832 * func1 -> sub rsp, 128
3833 * subfunc1 -> sub rsp, 256
3834 * tailcall1 -> add rsp, 256
3835 * func2 -> sub rsp, 192 (total stack size = 128 + 192 = 320)
3836 * subfunc2 -> sub rsp, 64
3837 * subfunc22 -> sub rsp, 128
3838 * tailcall2 -> add rsp, 128
3839 * func3 -> sub rsp, 32 (total stack size 128 + 192 + 64 + 32 = 416)
3841 * tailcall will unwind the current stack frame but it will not get rid
3842 * of caller's stack as shown on the example above.
3844 if (idx && subprog[idx].has_tail_call && depth >= 256) {
3846 "tail_calls are not allowed when call stack of previous frames is %d bytes. Too large\n",
3850 /* round up to 32-bytes, since this is granularity
3851 * of interpreter stack size
3853 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
3854 if (depth > MAX_BPF_STACK) {
3855 verbose(env, "combined stack size of %d calls is %d. Too large\n",
3860 subprog_end = subprog[idx + 1].start;
3861 for (; i < subprog_end; i++) {
3864 if (!bpf_pseudo_call(insn + i) && !bpf_pseudo_func(insn + i))
3866 /* remember insn and function to return to */
3867 ret_insn[frame] = i + 1;
3868 ret_prog[frame] = idx;
3870 /* find the callee */
3871 next_insn = i + insn[i].imm + 1;
3872 idx = find_subprog(env, next_insn);
3874 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3878 if (subprog[idx].is_async_cb) {
3879 if (subprog[idx].has_tail_call) {
3880 verbose(env, "verifier bug. subprog has tail_call and async cb\n");
3883 /* async callbacks don't increase bpf prog stack size */
3888 if (subprog[idx].has_tail_call)
3889 tail_call_reachable = true;
3892 if (frame >= MAX_CALL_FRAMES) {
3893 verbose(env, "the call stack of %d frames is too deep !\n",
3899 /* if tail call got detected across bpf2bpf calls then mark each of the
3900 * currently present subprog frames as tail call reachable subprogs;
3901 * this info will be utilized by JIT so that we will be preserving the
3902 * tail call counter throughout bpf2bpf calls combined with tailcalls
3904 if (tail_call_reachable)
3905 for (j = 0; j < frame; j++)
3906 subprog[ret_prog[j]].tail_call_reachable = true;
3907 if (subprog[0].tail_call_reachable)
3908 env->prog->aux->tail_call_reachable = true;
3910 /* end of for() loop means the last insn of the 'subprog'
3911 * was reached. Doesn't matter whether it was JA or EXIT
3915 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
3917 i = ret_insn[frame];
3918 idx = ret_prog[frame];
3922 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
3923 static int get_callee_stack_depth(struct bpf_verifier_env *env,
3924 const struct bpf_insn *insn, int idx)
3926 int start = idx + insn->imm + 1, subprog;
3928 subprog = find_subprog(env, start);
3930 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3934 return env->subprog_info[subprog].stack_depth;
3938 int check_ctx_reg(struct bpf_verifier_env *env,
3939 const struct bpf_reg_state *reg, int regno)
3941 /* Access to ctx or passing it to a helper is only allowed in
3942 * its original, unmodified form.
3946 verbose(env, "dereference of modified ctx ptr R%d off=%d disallowed\n",
3951 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3954 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3955 verbose(env, "variable ctx access var_off=%s disallowed\n", tn_buf);
3962 static int __check_buffer_access(struct bpf_verifier_env *env,
3963 const char *buf_info,
3964 const struct bpf_reg_state *reg,
3965 int regno, int off, int size)
3969 "R%d invalid %s buffer access: off=%d, size=%d\n",
3970 regno, buf_info, off, size);
3973 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3976 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3978 "R%d invalid variable buffer offset: off=%d, var_off=%s\n",
3979 regno, off, tn_buf);
3986 static int check_tp_buffer_access(struct bpf_verifier_env *env,
3987 const struct bpf_reg_state *reg,
3988 int regno, int off, int size)
3992 err = __check_buffer_access(env, "tracepoint", reg, regno, off, size);
3996 if (off + size > env->prog->aux->max_tp_access)
3997 env->prog->aux->max_tp_access = off + size;
4002 static int check_buffer_access(struct bpf_verifier_env *env,
4003 const struct bpf_reg_state *reg,
4004 int regno, int off, int size,
4005 bool zero_size_allowed,
4006 const char *buf_info,
4011 err = __check_buffer_access(env, buf_info, reg, regno, off, size);
4015 if (off + size > *max_access)
4016 *max_access = off + size;
4021 /* BPF architecture zero extends alu32 ops into 64-bit registesr */
4022 static void zext_32_to_64(struct bpf_reg_state *reg)
4024 reg->var_off = tnum_subreg(reg->var_off);
4025 __reg_assign_32_into_64(reg);
4028 /* truncate register to smaller size (in bytes)
4029 * must be called with size < BPF_REG_SIZE
4031 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
4035 /* clear high bits in bit representation */
4036 reg->var_off = tnum_cast(reg->var_off, size);
4038 /* fix arithmetic bounds */
4039 mask = ((u64)1 << (size * 8)) - 1;
4040 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
4041 reg->umin_value &= mask;
4042 reg->umax_value &= mask;
4044 reg->umin_value = 0;
4045 reg->umax_value = mask;
4047 reg->smin_value = reg->umin_value;
4048 reg->smax_value = reg->umax_value;
4050 /* If size is smaller than 32bit register the 32bit register
4051 * values are also truncated so we push 64-bit bounds into
4052 * 32-bit bounds. Above were truncated < 32-bits already.
4056 __reg_combine_64_into_32(reg);
4059 static bool bpf_map_is_rdonly(const struct bpf_map *map)
4061 /* A map is considered read-only if the following condition are true:
4063 * 1) BPF program side cannot change any of the map content. The
4064 * BPF_F_RDONLY_PROG flag is throughout the lifetime of a map
4065 * and was set at map creation time.
4066 * 2) The map value(s) have been initialized from user space by a
4067 * loader and then "frozen", such that no new map update/delete
4068 * operations from syscall side are possible for the rest of
4069 * the map's lifetime from that point onwards.
4070 * 3) Any parallel/pending map update/delete operations from syscall
4071 * side have been completed. Only after that point, it's safe to
4072 * assume that map value(s) are immutable.
4074 return (map->map_flags & BPF_F_RDONLY_PROG) &&
4075 READ_ONCE(map->frozen) &&
4076 !bpf_map_write_active(map);
4079 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val)
4085 err = map->ops->map_direct_value_addr(map, &addr, off);
4088 ptr = (void *)(long)addr + off;
4092 *val = (u64)*(u8 *)ptr;
4095 *val = (u64)*(u16 *)ptr;
4098 *val = (u64)*(u32 *)ptr;
4109 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
4110 struct bpf_reg_state *regs,
4111 int regno, int off, int size,
4112 enum bpf_access_type atype,
4115 struct bpf_reg_state *reg = regs + regno;
4116 const struct btf_type *t = btf_type_by_id(reg->btf, reg->btf_id);
4117 const char *tname = btf_name_by_offset(reg->btf, t->name_off);
4123 "R%d is ptr_%s invalid negative access: off=%d\n",
4127 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
4130 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4132 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
4133 regno, tname, off, tn_buf);
4137 if (env->ops->btf_struct_access) {
4138 ret = env->ops->btf_struct_access(&env->log, reg->btf, t,
4139 off, size, atype, &btf_id);
4141 if (atype != BPF_READ) {
4142 verbose(env, "only read is supported\n");
4146 ret = btf_struct_access(&env->log, reg->btf, t, off, size,
4153 if (atype == BPF_READ && value_regno >= 0)
4154 mark_btf_ld_reg(env, regs, value_regno, ret, reg->btf, btf_id);
4159 static int check_ptr_to_map_access(struct bpf_verifier_env *env,
4160 struct bpf_reg_state *regs,
4161 int regno, int off, int size,
4162 enum bpf_access_type atype,
4165 struct bpf_reg_state *reg = regs + regno;
4166 struct bpf_map *map = reg->map_ptr;
4167 const struct btf_type *t;
4173 verbose(env, "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n");
4177 if (!map->ops->map_btf_id || !*map->ops->map_btf_id) {
4178 verbose(env, "map_ptr access not supported for map type %d\n",
4183 t = btf_type_by_id(btf_vmlinux, *map->ops->map_btf_id);
4184 tname = btf_name_by_offset(btf_vmlinux, t->name_off);
4186 if (!env->allow_ptr_to_map_access) {
4188 "%s access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
4194 verbose(env, "R%d is %s invalid negative access: off=%d\n",
4199 if (atype != BPF_READ) {
4200 verbose(env, "only read from %s is supported\n", tname);
4204 ret = btf_struct_access(&env->log, btf_vmlinux, t, off, size, atype, &btf_id);
4208 if (value_regno >= 0)
4209 mark_btf_ld_reg(env, regs, value_regno, ret, btf_vmlinux, btf_id);
4214 /* Check that the stack access at the given offset is within bounds. The
4215 * maximum valid offset is -1.
4217 * The minimum valid offset is -MAX_BPF_STACK for writes, and
4218 * -state->allocated_stack for reads.
4220 static int check_stack_slot_within_bounds(int off,
4221 struct bpf_func_state *state,
4222 enum bpf_access_type t)
4227 min_valid_off = -MAX_BPF_STACK;
4229 min_valid_off = -state->allocated_stack;
4231 if (off < min_valid_off || off > -1)
4236 /* Check that the stack access at 'regno + off' falls within the maximum stack
4239 * 'off' includes `regno->offset`, but not its dynamic part (if any).
4241 static int check_stack_access_within_bounds(
4242 struct bpf_verifier_env *env,
4243 int regno, int off, int access_size,
4244 enum stack_access_src src, enum bpf_access_type type)
4246 struct bpf_reg_state *regs = cur_regs(env);
4247 struct bpf_reg_state *reg = regs + regno;
4248 struct bpf_func_state *state = func(env, reg);
4249 int min_off, max_off;
4253 if (src == ACCESS_HELPER)
4254 /* We don't know if helpers are reading or writing (or both). */
4255 err_extra = " indirect access to";
4256 else if (type == BPF_READ)
4257 err_extra = " read from";
4259 err_extra = " write to";
4261 if (tnum_is_const(reg->var_off)) {
4262 min_off = reg->var_off.value + off;
4263 if (access_size > 0)
4264 max_off = min_off + access_size - 1;
4268 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
4269 reg->smin_value <= -BPF_MAX_VAR_OFF) {
4270 verbose(env, "invalid unbounded variable-offset%s stack R%d\n",
4274 min_off = reg->smin_value + off;
4275 if (access_size > 0)
4276 max_off = reg->smax_value + off + access_size - 1;
4281 err = check_stack_slot_within_bounds(min_off, state, type);
4283 err = check_stack_slot_within_bounds(max_off, state, type);
4286 if (tnum_is_const(reg->var_off)) {
4287 verbose(env, "invalid%s stack R%d off=%d size=%d\n",
4288 err_extra, regno, off, access_size);
4292 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4293 verbose(env, "invalid variable-offset%s stack R%d var_off=%s size=%d\n",
4294 err_extra, regno, tn_buf, access_size);
4300 /* check whether memory at (regno + off) is accessible for t = (read | write)
4301 * if t==write, value_regno is a register which value is stored into memory
4302 * if t==read, value_regno is a register which will receive the value from memory
4303 * if t==write && value_regno==-1, some unknown value is stored into memory
4304 * if t==read && value_regno==-1, don't care what we read from memory
4306 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
4307 int off, int bpf_size, enum bpf_access_type t,
4308 int value_regno, bool strict_alignment_once)
4310 struct bpf_reg_state *regs = cur_regs(env);
4311 struct bpf_reg_state *reg = regs + regno;
4312 struct bpf_func_state *state;
4315 size = bpf_size_to_bytes(bpf_size);
4319 /* alignment checks will add in reg->off themselves */
4320 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
4324 /* for access checks, reg->off is just part of off */
4327 if (reg->type == PTR_TO_MAP_KEY) {
4328 if (t == BPF_WRITE) {
4329 verbose(env, "write to change key R%d not allowed\n", regno);
4333 err = check_mem_region_access(env, regno, off, size,
4334 reg->map_ptr->key_size, false);
4337 if (value_regno >= 0)
4338 mark_reg_unknown(env, regs, value_regno);
4339 } else if (reg->type == PTR_TO_MAP_VALUE) {
4340 if (t == BPF_WRITE && value_regno >= 0 &&
4341 is_pointer_value(env, value_regno)) {
4342 verbose(env, "R%d leaks addr into map\n", value_regno);
4345 err = check_map_access_type(env, regno, off, size, t);
4348 err = check_map_access(env, regno, off, size, false);
4349 if (!err && t == BPF_READ && value_regno >= 0) {
4350 struct bpf_map *map = reg->map_ptr;
4352 /* if map is read-only, track its contents as scalars */
4353 if (tnum_is_const(reg->var_off) &&
4354 bpf_map_is_rdonly(map) &&
4355 map->ops->map_direct_value_addr) {
4356 int map_off = off + reg->var_off.value;
4359 err = bpf_map_direct_read(map, map_off, size,
4364 regs[value_regno].type = SCALAR_VALUE;
4365 __mark_reg_known(®s[value_regno], val);
4367 mark_reg_unknown(env, regs, value_regno);
4370 } else if (reg->type == PTR_TO_MEM) {
4371 if (t == BPF_WRITE && value_regno >= 0 &&
4372 is_pointer_value(env, value_regno)) {
4373 verbose(env, "R%d leaks addr into mem\n", value_regno);
4376 err = check_mem_region_access(env, regno, off, size,
4377 reg->mem_size, false);
4378 if (!err && t == BPF_READ && value_regno >= 0)
4379 mark_reg_unknown(env, regs, value_regno);
4380 } else if (reg->type == PTR_TO_CTX) {
4381 enum bpf_reg_type reg_type = SCALAR_VALUE;
4382 struct btf *btf = NULL;
4385 if (t == BPF_WRITE && value_regno >= 0 &&
4386 is_pointer_value(env, value_regno)) {
4387 verbose(env, "R%d leaks addr into ctx\n", value_regno);
4391 err = check_ctx_reg(env, reg, regno);
4395 err = check_ctx_access(env, insn_idx, off, size, t, ®_type, &btf, &btf_id);
4397 verbose_linfo(env, insn_idx, "; ");
4398 if (!err && t == BPF_READ && value_regno >= 0) {
4399 /* ctx access returns either a scalar, or a
4400 * PTR_TO_PACKET[_META,_END]. In the latter
4401 * case, we know the offset is zero.
4403 if (reg_type == SCALAR_VALUE) {
4404 mark_reg_unknown(env, regs, value_regno);
4406 mark_reg_known_zero(env, regs,
4408 if (reg_type_may_be_null(reg_type))
4409 regs[value_regno].id = ++env->id_gen;
4410 /* A load of ctx field could have different
4411 * actual load size with the one encoded in the
4412 * insn. When the dst is PTR, it is for sure not
4415 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
4416 if (reg_type == PTR_TO_BTF_ID ||
4417 reg_type == PTR_TO_BTF_ID_OR_NULL) {
4418 regs[value_regno].btf = btf;
4419 regs[value_regno].btf_id = btf_id;
4422 regs[value_regno].type = reg_type;
4425 } else if (reg->type == PTR_TO_STACK) {
4426 /* Basic bounds checks. */
4427 err = check_stack_access_within_bounds(env, regno, off, size, ACCESS_DIRECT, t);
4431 state = func(env, reg);
4432 err = update_stack_depth(env, state, off);
4437 err = check_stack_read(env, regno, off, size,
4440 err = check_stack_write(env, regno, off, size,
4441 value_regno, insn_idx);
4442 } else if (reg_is_pkt_pointer(reg)) {
4443 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
4444 verbose(env, "cannot write into packet\n");
4447 if (t == BPF_WRITE && value_regno >= 0 &&
4448 is_pointer_value(env, value_regno)) {
4449 verbose(env, "R%d leaks addr into packet\n",
4453 err = check_packet_access(env, regno, off, size, false);
4454 if (!err && t == BPF_READ && value_regno >= 0)
4455 mark_reg_unknown(env, regs, value_regno);
4456 } else if (reg->type == PTR_TO_FLOW_KEYS) {
4457 if (t == BPF_WRITE && value_regno >= 0 &&
4458 is_pointer_value(env, value_regno)) {
4459 verbose(env, "R%d leaks addr into flow keys\n",
4464 err = check_flow_keys_access(env, off, size);
4465 if (!err && t == BPF_READ && value_regno >= 0)
4466 mark_reg_unknown(env, regs, value_regno);
4467 } else if (type_is_sk_pointer(reg->type)) {
4468 if (t == BPF_WRITE) {
4469 verbose(env, "R%d cannot write into %s\n",
4470 regno, reg_type_str[reg->type]);
4473 err = check_sock_access(env, insn_idx, regno, off, size, t);
4474 if (!err && value_regno >= 0)
4475 mark_reg_unknown(env, regs, value_regno);
4476 } else if (reg->type == PTR_TO_TP_BUFFER) {
4477 err = check_tp_buffer_access(env, reg, regno, off, size);
4478 if (!err && t == BPF_READ && value_regno >= 0)
4479 mark_reg_unknown(env, regs, value_regno);
4480 } else if (reg->type == PTR_TO_BTF_ID) {
4481 err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
4483 } else if (reg->type == CONST_PTR_TO_MAP) {
4484 err = check_ptr_to_map_access(env, regs, regno, off, size, t,
4486 } else if (reg->type == PTR_TO_RDONLY_BUF) {
4487 if (t == BPF_WRITE) {
4488 verbose(env, "R%d cannot write into %s\n",
4489 regno, reg_type_str[reg->type]);
4492 err = check_buffer_access(env, reg, regno, off, size, false,
4494 &env->prog->aux->max_rdonly_access);
4495 if (!err && value_regno >= 0)
4496 mark_reg_unknown(env, regs, value_regno);
4497 } else if (reg->type == PTR_TO_RDWR_BUF) {
4498 err = check_buffer_access(env, reg, regno, off, size, false,
4500 &env->prog->aux->max_rdwr_access);
4501 if (!err && t == BPF_READ && value_regno >= 0)
4502 mark_reg_unknown(env, regs, value_regno);
4504 verbose(env, "R%d invalid mem access '%s'\n", regno,
4505 reg_type_str[reg->type]);
4509 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
4510 regs[value_regno].type == SCALAR_VALUE) {
4511 /* b/h/w load zero-extends, mark upper bits as known 0 */
4512 coerce_reg_to_size(®s[value_regno], size);
4517 static int check_atomic(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
4522 switch (insn->imm) {
4524 case BPF_ADD | BPF_FETCH:
4526 case BPF_AND | BPF_FETCH:
4528 case BPF_OR | BPF_FETCH:
4530 case BPF_XOR | BPF_FETCH:
4535 verbose(env, "BPF_ATOMIC uses invalid atomic opcode %02x\n", insn->imm);
4539 if (BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) {
4540 verbose(env, "invalid atomic operand size\n");
4544 /* check src1 operand */
4545 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4549 /* check src2 operand */
4550 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4554 if (insn->imm == BPF_CMPXCHG) {
4555 /* Check comparison of R0 with memory location */
4556 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
4561 if (is_pointer_value(env, insn->src_reg)) {
4562 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
4566 if (is_ctx_reg(env, insn->dst_reg) ||
4567 is_pkt_reg(env, insn->dst_reg) ||
4568 is_flow_key_reg(env, insn->dst_reg) ||
4569 is_sk_reg(env, insn->dst_reg)) {
4570 verbose(env, "BPF_ATOMIC stores into R%d %s is not allowed\n",
4572 reg_type_str[reg_state(env, insn->dst_reg)->type]);
4576 if (insn->imm & BPF_FETCH) {
4577 if (insn->imm == BPF_CMPXCHG)
4578 load_reg = BPF_REG_0;
4580 load_reg = insn->src_reg;
4582 /* check and record load of old value */
4583 err = check_reg_arg(env, load_reg, DST_OP);
4587 /* This instruction accesses a memory location but doesn't
4588 * actually load it into a register.
4593 /* check whether we can read the memory */
4594 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4595 BPF_SIZE(insn->code), BPF_READ, load_reg, true);
4599 /* check whether we can write into the same memory */
4600 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4601 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
4608 /* When register 'regno' is used to read the stack (either directly or through
4609 * a helper function) make sure that it's within stack boundary and, depending
4610 * on the access type, that all elements of the stack are initialized.
4612 * 'off' includes 'regno->off', but not its dynamic part (if any).
4614 * All registers that have been spilled on the stack in the slots within the
4615 * read offsets are marked as read.
4617 static int check_stack_range_initialized(
4618 struct bpf_verifier_env *env, int regno, int off,
4619 int access_size, bool zero_size_allowed,
4620 enum stack_access_src type, struct bpf_call_arg_meta *meta)
4622 struct bpf_reg_state *reg = reg_state(env, regno);
4623 struct bpf_func_state *state = func(env, reg);
4624 int err, min_off, max_off, i, j, slot, spi;
4625 char *err_extra = type == ACCESS_HELPER ? " indirect" : "";
4626 enum bpf_access_type bounds_check_type;
4627 /* Some accesses can write anything into the stack, others are
4630 bool clobber = false;
4632 if (access_size == 0 && !zero_size_allowed) {
4633 verbose(env, "invalid zero-sized read\n");
4637 if (type == ACCESS_HELPER) {
4638 /* The bounds checks for writes are more permissive than for
4639 * reads. However, if raw_mode is not set, we'll do extra
4642 bounds_check_type = BPF_WRITE;
4645 bounds_check_type = BPF_READ;
4647 err = check_stack_access_within_bounds(env, regno, off, access_size,
4648 type, bounds_check_type);
4653 if (tnum_is_const(reg->var_off)) {
4654 min_off = max_off = reg->var_off.value + off;
4656 /* Variable offset is prohibited for unprivileged mode for
4657 * simplicity since it requires corresponding support in
4658 * Spectre masking for stack ALU.
4659 * See also retrieve_ptr_limit().
4661 if (!env->bypass_spec_v1) {
4664 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4665 verbose(env, "R%d%s variable offset stack access prohibited for !root, var_off=%s\n",
4666 regno, err_extra, tn_buf);
4669 /* Only initialized buffer on stack is allowed to be accessed
4670 * with variable offset. With uninitialized buffer it's hard to
4671 * guarantee that whole memory is marked as initialized on
4672 * helper return since specific bounds are unknown what may
4673 * cause uninitialized stack leaking.
4675 if (meta && meta->raw_mode)
4678 min_off = reg->smin_value + off;
4679 max_off = reg->smax_value + off;
4682 if (meta && meta->raw_mode) {
4683 meta->access_size = access_size;
4684 meta->regno = regno;
4688 for (i = min_off; i < max_off + access_size; i++) {
4692 spi = slot / BPF_REG_SIZE;
4693 if (state->allocated_stack <= slot)
4695 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
4696 if (*stype == STACK_MISC)
4698 if (*stype == STACK_ZERO) {
4700 /* helper can write anything into the stack */
4701 *stype = STACK_MISC;
4706 if (is_spilled_reg(&state->stack[spi]) &&
4707 state->stack[spi].spilled_ptr.type == PTR_TO_BTF_ID)
4710 if (is_spilled_reg(&state->stack[spi]) &&
4711 (state->stack[spi].spilled_ptr.type == SCALAR_VALUE ||
4712 env->allow_ptr_leaks)) {
4714 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
4715 for (j = 0; j < BPF_REG_SIZE; j++)
4716 scrub_spilled_slot(&state->stack[spi].slot_type[j]);
4722 if (tnum_is_const(reg->var_off)) {
4723 verbose(env, "invalid%s read from stack R%d off %d+%d size %d\n",
4724 err_extra, regno, min_off, i - min_off, access_size);
4728 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4729 verbose(env, "invalid%s read from stack R%d var_off %s+%d size %d\n",
4730 err_extra, regno, tn_buf, i - min_off, access_size);
4734 /* reading any byte out of 8-byte 'spill_slot' will cause
4735 * the whole slot to be marked as 'read'
4737 mark_reg_read(env, &state->stack[spi].spilled_ptr,
4738 state->stack[spi].spilled_ptr.parent,
4741 return update_stack_depth(env, state, min_off);
4744 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
4745 int access_size, bool zero_size_allowed,
4746 struct bpf_call_arg_meta *meta)
4748 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4750 switch (reg->type) {
4752 case PTR_TO_PACKET_META:
4753 return check_packet_access(env, regno, reg->off, access_size,
4755 case PTR_TO_MAP_KEY:
4756 return check_mem_region_access(env, regno, reg->off, access_size,
4757 reg->map_ptr->key_size, false);
4758 case PTR_TO_MAP_VALUE:
4759 if (check_map_access_type(env, regno, reg->off, access_size,
4760 meta && meta->raw_mode ? BPF_WRITE :
4763 return check_map_access(env, regno, reg->off, access_size,
4766 return check_mem_region_access(env, regno, reg->off,
4767 access_size, reg->mem_size,
4769 case PTR_TO_RDONLY_BUF:
4770 if (meta && meta->raw_mode)
4772 return check_buffer_access(env, reg, regno, reg->off,
4773 access_size, zero_size_allowed,
4775 &env->prog->aux->max_rdonly_access);
4776 case PTR_TO_RDWR_BUF:
4777 return check_buffer_access(env, reg, regno, reg->off,
4778 access_size, zero_size_allowed,
4780 &env->prog->aux->max_rdwr_access);
4782 return check_stack_range_initialized(
4784 regno, reg->off, access_size,
4785 zero_size_allowed, ACCESS_HELPER, meta);
4786 default: /* scalar_value or invalid ptr */
4787 /* Allow zero-byte read from NULL, regardless of pointer type */
4788 if (zero_size_allowed && access_size == 0 &&
4789 register_is_null(reg))
4792 verbose(env, "R%d type=%s expected=%s\n", regno,
4793 reg_type_str[reg->type],
4794 reg_type_str[PTR_TO_STACK]);
4799 int check_mem_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
4800 u32 regno, u32 mem_size)
4802 if (register_is_null(reg))
4805 if (reg_type_may_be_null(reg->type)) {
4806 /* Assuming that the register contains a value check if the memory
4807 * access is safe. Temporarily save and restore the register's state as
4808 * the conversion shouldn't be visible to a caller.
4810 const struct bpf_reg_state saved_reg = *reg;
4813 mark_ptr_not_null_reg(reg);
4814 rv = check_helper_mem_access(env, regno, mem_size, true, NULL);
4819 return check_helper_mem_access(env, regno, mem_size, true, NULL);
4822 /* Implementation details:
4823 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL
4824 * Two bpf_map_lookups (even with the same key) will have different reg->id.
4825 * For traditional PTR_TO_MAP_VALUE the verifier clears reg->id after
4826 * value_or_null->value transition, since the verifier only cares about
4827 * the range of access to valid map value pointer and doesn't care about actual
4828 * address of the map element.
4829 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
4830 * reg->id > 0 after value_or_null->value transition. By doing so
4831 * two bpf_map_lookups will be considered two different pointers that
4832 * point to different bpf_spin_locks.
4833 * The verifier allows taking only one bpf_spin_lock at a time to avoid
4835 * Since only one bpf_spin_lock is allowed the checks are simpler than
4836 * reg_is_refcounted() logic. The verifier needs to remember only
4837 * one spin_lock instead of array of acquired_refs.
4838 * cur_state->active_spin_lock remembers which map value element got locked
4839 * and clears it after bpf_spin_unlock.
4841 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
4844 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4845 struct bpf_verifier_state *cur = env->cur_state;
4846 bool is_const = tnum_is_const(reg->var_off);
4847 struct bpf_map *map = reg->map_ptr;
4848 u64 val = reg->var_off.value;
4852 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
4858 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
4862 if (!map_value_has_spin_lock(map)) {
4863 if (map->spin_lock_off == -E2BIG)
4865 "map '%s' has more than one 'struct bpf_spin_lock'\n",
4867 else if (map->spin_lock_off == -ENOENT)
4869 "map '%s' doesn't have 'struct bpf_spin_lock'\n",
4873 "map '%s' is not a struct type or bpf_spin_lock is mangled\n",
4877 if (map->spin_lock_off != val + reg->off) {
4878 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock'\n",
4883 if (cur->active_spin_lock) {
4885 "Locking two bpf_spin_locks are not allowed\n");
4888 cur->active_spin_lock = reg->id;
4890 if (!cur->active_spin_lock) {
4891 verbose(env, "bpf_spin_unlock without taking a lock\n");
4894 if (cur->active_spin_lock != reg->id) {
4895 verbose(env, "bpf_spin_unlock of different lock\n");
4898 cur->active_spin_lock = 0;
4903 static int process_timer_func(struct bpf_verifier_env *env, int regno,
4904 struct bpf_call_arg_meta *meta)
4906 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4907 bool is_const = tnum_is_const(reg->var_off);
4908 struct bpf_map *map = reg->map_ptr;
4909 u64 val = reg->var_off.value;
4913 "R%d doesn't have constant offset. bpf_timer has to be at the constant offset\n",
4918 verbose(env, "map '%s' has to have BTF in order to use bpf_timer\n",
4922 if (!map_value_has_timer(map)) {
4923 if (map->timer_off == -E2BIG)
4925 "map '%s' has more than one 'struct bpf_timer'\n",
4927 else if (map->timer_off == -ENOENT)
4929 "map '%s' doesn't have 'struct bpf_timer'\n",
4933 "map '%s' is not a struct type or bpf_timer is mangled\n",
4937 if (map->timer_off != val + reg->off) {
4938 verbose(env, "off %lld doesn't point to 'struct bpf_timer' that is at %d\n",
4939 val + reg->off, map->timer_off);
4942 if (meta->map_ptr) {
4943 verbose(env, "verifier bug. Two map pointers in a timer helper\n");
4946 meta->map_uid = reg->map_uid;
4947 meta->map_ptr = map;
4951 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
4953 return type == ARG_PTR_TO_MEM ||
4954 type == ARG_PTR_TO_MEM_OR_NULL ||
4955 type == ARG_PTR_TO_UNINIT_MEM;
4958 static bool arg_type_is_mem_size(enum bpf_arg_type type)
4960 return type == ARG_CONST_SIZE ||
4961 type == ARG_CONST_SIZE_OR_ZERO;
4964 static bool arg_type_is_alloc_size(enum bpf_arg_type type)
4966 return type == ARG_CONST_ALLOC_SIZE_OR_ZERO;
4969 static bool arg_type_is_int_ptr(enum bpf_arg_type type)
4971 return type == ARG_PTR_TO_INT ||
4972 type == ARG_PTR_TO_LONG;
4975 static int int_ptr_type_to_size(enum bpf_arg_type type)
4977 if (type == ARG_PTR_TO_INT)
4979 else if (type == ARG_PTR_TO_LONG)
4985 static int resolve_map_arg_type(struct bpf_verifier_env *env,
4986 const struct bpf_call_arg_meta *meta,
4987 enum bpf_arg_type *arg_type)
4989 if (!meta->map_ptr) {
4990 /* kernel subsystem misconfigured verifier */
4991 verbose(env, "invalid map_ptr to access map->type\n");
4995 switch (meta->map_ptr->map_type) {
4996 case BPF_MAP_TYPE_SOCKMAP:
4997 case BPF_MAP_TYPE_SOCKHASH:
4998 if (*arg_type == ARG_PTR_TO_MAP_VALUE) {
4999 *arg_type = ARG_PTR_TO_BTF_ID_SOCK_COMMON;
5001 verbose(env, "invalid arg_type for sockmap/sockhash\n");
5005 case BPF_MAP_TYPE_BLOOM_FILTER:
5006 if (meta->func_id == BPF_FUNC_map_peek_elem)
5007 *arg_type = ARG_PTR_TO_MAP_VALUE;
5015 struct bpf_reg_types {
5016 const enum bpf_reg_type types[10];
5020 static const struct bpf_reg_types map_key_value_types = {
5030 static const struct bpf_reg_types sock_types = {
5040 static const struct bpf_reg_types btf_id_sock_common_types = {
5048 .btf_id = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
5052 static const struct bpf_reg_types mem_types = {
5065 static const struct bpf_reg_types int_ptr_types = {
5075 static const struct bpf_reg_types fullsock_types = { .types = { PTR_TO_SOCKET } };
5076 static const struct bpf_reg_types scalar_types = { .types = { SCALAR_VALUE } };
5077 static const struct bpf_reg_types context_types = { .types = { PTR_TO_CTX } };
5078 static const struct bpf_reg_types alloc_mem_types = { .types = { PTR_TO_MEM } };
5079 static const struct bpf_reg_types const_map_ptr_types = { .types = { CONST_PTR_TO_MAP } };
5080 static const struct bpf_reg_types btf_ptr_types = { .types = { PTR_TO_BTF_ID } };
5081 static const struct bpf_reg_types spin_lock_types = { .types = { PTR_TO_MAP_VALUE } };
5082 static const struct bpf_reg_types percpu_btf_ptr_types = { .types = { PTR_TO_PERCPU_BTF_ID } };
5083 static const struct bpf_reg_types func_ptr_types = { .types = { PTR_TO_FUNC } };
5084 static const struct bpf_reg_types stack_ptr_types = { .types = { PTR_TO_STACK } };
5085 static const struct bpf_reg_types const_str_ptr_types = { .types = { PTR_TO_MAP_VALUE } };
5086 static const struct bpf_reg_types timer_types = { .types = { PTR_TO_MAP_VALUE } };
5088 static const struct bpf_reg_types *compatible_reg_types[__BPF_ARG_TYPE_MAX] = {
5089 [ARG_PTR_TO_MAP_KEY] = &map_key_value_types,
5090 [ARG_PTR_TO_MAP_VALUE] = &map_key_value_types,
5091 [ARG_PTR_TO_UNINIT_MAP_VALUE] = &map_key_value_types,
5092 [ARG_PTR_TO_MAP_VALUE_OR_NULL] = &map_key_value_types,
5093 [ARG_CONST_SIZE] = &scalar_types,
5094 [ARG_CONST_SIZE_OR_ZERO] = &scalar_types,
5095 [ARG_CONST_ALLOC_SIZE_OR_ZERO] = &scalar_types,
5096 [ARG_CONST_MAP_PTR] = &const_map_ptr_types,
5097 [ARG_PTR_TO_CTX] = &context_types,
5098 [ARG_PTR_TO_CTX_OR_NULL] = &context_types,
5099 [ARG_PTR_TO_SOCK_COMMON] = &sock_types,
5101 [ARG_PTR_TO_BTF_ID_SOCK_COMMON] = &btf_id_sock_common_types,
5103 [ARG_PTR_TO_SOCKET] = &fullsock_types,
5104 [ARG_PTR_TO_SOCKET_OR_NULL] = &fullsock_types,
5105 [ARG_PTR_TO_BTF_ID] = &btf_ptr_types,
5106 [ARG_PTR_TO_SPIN_LOCK] = &spin_lock_types,
5107 [ARG_PTR_TO_MEM] = &mem_types,
5108 [ARG_PTR_TO_MEM_OR_NULL] = &mem_types,
5109 [ARG_PTR_TO_UNINIT_MEM] = &mem_types,
5110 [ARG_PTR_TO_ALLOC_MEM] = &alloc_mem_types,
5111 [ARG_PTR_TO_ALLOC_MEM_OR_NULL] = &alloc_mem_types,
5112 [ARG_PTR_TO_INT] = &int_ptr_types,
5113 [ARG_PTR_TO_LONG] = &int_ptr_types,
5114 [ARG_PTR_TO_PERCPU_BTF_ID] = &percpu_btf_ptr_types,
5115 [ARG_PTR_TO_FUNC] = &func_ptr_types,
5116 [ARG_PTR_TO_STACK_OR_NULL] = &stack_ptr_types,
5117 [ARG_PTR_TO_CONST_STR] = &const_str_ptr_types,
5118 [ARG_PTR_TO_TIMER] = &timer_types,
5121 static int check_reg_type(struct bpf_verifier_env *env, u32 regno,
5122 enum bpf_arg_type arg_type,
5123 const u32 *arg_btf_id)
5125 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
5126 enum bpf_reg_type expected, type = reg->type;
5127 const struct bpf_reg_types *compatible;
5130 compatible = compatible_reg_types[arg_type];
5132 verbose(env, "verifier internal error: unsupported arg type %d\n", arg_type);
5136 for (i = 0; i < ARRAY_SIZE(compatible->types); i++) {
5137 expected = compatible->types[i];
5138 if (expected == NOT_INIT)
5141 if (type == expected)
5145 verbose(env, "R%d type=%s expected=", regno, reg_type_str[type]);
5146 for (j = 0; j + 1 < i; j++)
5147 verbose(env, "%s, ", reg_type_str[compatible->types[j]]);
5148 verbose(env, "%s\n", reg_type_str[compatible->types[j]]);
5152 if (type == PTR_TO_BTF_ID) {
5154 if (!compatible->btf_id) {
5155 verbose(env, "verifier internal error: missing arg compatible BTF ID\n");
5158 arg_btf_id = compatible->btf_id;
5161 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
5162 btf_vmlinux, *arg_btf_id)) {
5163 verbose(env, "R%d is of type %s but %s is expected\n",
5164 regno, kernel_type_name(reg->btf, reg->btf_id),
5165 kernel_type_name(btf_vmlinux, *arg_btf_id));
5169 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
5170 verbose(env, "R%d is a pointer to in-kernel struct with non-zero offset\n",
5179 static int check_func_arg(struct bpf_verifier_env *env, u32 arg,
5180 struct bpf_call_arg_meta *meta,
5181 const struct bpf_func_proto *fn)
5183 u32 regno = BPF_REG_1 + arg;
5184 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
5185 enum bpf_arg_type arg_type = fn->arg_type[arg];
5186 enum bpf_reg_type type = reg->type;
5189 if (arg_type == ARG_DONTCARE)
5192 err = check_reg_arg(env, regno, SRC_OP);
5196 if (arg_type == ARG_ANYTHING) {
5197 if (is_pointer_value(env, regno)) {
5198 verbose(env, "R%d leaks addr into helper function\n",
5205 if (type_is_pkt_pointer(type) &&
5206 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
5207 verbose(env, "helper access to the packet is not allowed\n");
5211 if (arg_type == ARG_PTR_TO_MAP_VALUE ||
5212 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE ||
5213 arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL) {
5214 err = resolve_map_arg_type(env, meta, &arg_type);
5219 if (register_is_null(reg) && arg_type_may_be_null(arg_type))
5220 /* A NULL register has a SCALAR_VALUE type, so skip
5223 goto skip_type_check;
5225 err = check_reg_type(env, regno, arg_type, fn->arg_btf_id[arg]);
5229 if (type == PTR_TO_CTX) {
5230 err = check_ctx_reg(env, reg, regno);
5236 if (reg->ref_obj_id) {
5237 if (meta->ref_obj_id) {
5238 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
5239 regno, reg->ref_obj_id,
5243 meta->ref_obj_id = reg->ref_obj_id;
5246 if (arg_type == ARG_CONST_MAP_PTR) {
5247 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
5248 if (meta->map_ptr) {
5249 /* Use map_uid (which is unique id of inner map) to reject:
5250 * inner_map1 = bpf_map_lookup_elem(outer_map, key1)
5251 * inner_map2 = bpf_map_lookup_elem(outer_map, key2)
5252 * if (inner_map1 && inner_map2) {
5253 * timer = bpf_map_lookup_elem(inner_map1);
5255 * // mismatch would have been allowed
5256 * bpf_timer_init(timer, inner_map2);
5259 * Comparing map_ptr is enough to distinguish normal and outer maps.
5261 if (meta->map_ptr != reg->map_ptr ||
5262 meta->map_uid != reg->map_uid) {
5264 "timer pointer in R1 map_uid=%d doesn't match map pointer in R2 map_uid=%d\n",
5265 meta->map_uid, reg->map_uid);
5269 meta->map_ptr = reg->map_ptr;
5270 meta->map_uid = reg->map_uid;
5271 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
5272 /* bpf_map_xxx(..., map_ptr, ..., key) call:
5273 * check that [key, key + map->key_size) are within
5274 * stack limits and initialized
5276 if (!meta->map_ptr) {
5277 /* in function declaration map_ptr must come before
5278 * map_key, so that it's verified and known before
5279 * we have to check map_key here. Otherwise it means
5280 * that kernel subsystem misconfigured verifier
5282 verbose(env, "invalid map_ptr to access map->key\n");
5285 err = check_helper_mem_access(env, regno,
5286 meta->map_ptr->key_size, false,
5288 } else if (arg_type == ARG_PTR_TO_MAP_VALUE ||
5289 (arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL &&
5290 !register_is_null(reg)) ||
5291 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
5292 /* bpf_map_xxx(..., map_ptr, ..., value) call:
5293 * check [value, value + map->value_size) validity
5295 if (!meta->map_ptr) {
5296 /* kernel subsystem misconfigured verifier */
5297 verbose(env, "invalid map_ptr to access map->value\n");
5300 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE);
5301 err = check_helper_mem_access(env, regno,
5302 meta->map_ptr->value_size, false,
5304 } else if (arg_type == ARG_PTR_TO_PERCPU_BTF_ID) {
5306 verbose(env, "Helper has invalid btf_id in R%d\n", regno);
5309 meta->ret_btf = reg->btf;
5310 meta->ret_btf_id = reg->btf_id;
5311 } else if (arg_type == ARG_PTR_TO_SPIN_LOCK) {
5312 if (meta->func_id == BPF_FUNC_spin_lock) {
5313 if (process_spin_lock(env, regno, true))
5315 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
5316 if (process_spin_lock(env, regno, false))
5319 verbose(env, "verifier internal error\n");
5322 } else if (arg_type == ARG_PTR_TO_TIMER) {
5323 if (process_timer_func(env, regno, meta))
5325 } else if (arg_type == ARG_PTR_TO_FUNC) {
5326 meta->subprogno = reg->subprogno;
5327 } else if (arg_type_is_mem_ptr(arg_type)) {
5328 /* The access to this pointer is only checked when we hit the
5329 * next is_mem_size argument below.
5331 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MEM);
5332 } else if (arg_type_is_mem_size(arg_type)) {
5333 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
5335 /* This is used to refine r0 return value bounds for helpers
5336 * that enforce this value as an upper bound on return values.
5337 * See do_refine_retval_range() for helpers that can refine
5338 * the return value. C type of helper is u32 so we pull register
5339 * bound from umax_value however, if negative verifier errors
5340 * out. Only upper bounds can be learned because retval is an
5341 * int type and negative retvals are allowed.
5343 meta->msize_max_value = reg->umax_value;
5345 /* The register is SCALAR_VALUE; the access check
5346 * happens using its boundaries.
5348 if (!tnum_is_const(reg->var_off))
5349 /* For unprivileged variable accesses, disable raw
5350 * mode so that the program is required to
5351 * initialize all the memory that the helper could
5352 * just partially fill up.
5356 if (reg->smin_value < 0) {
5357 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
5362 if (reg->umin_value == 0) {
5363 err = check_helper_mem_access(env, regno - 1, 0,
5370 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
5371 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
5375 err = check_helper_mem_access(env, regno - 1,
5377 zero_size_allowed, meta);
5379 err = mark_chain_precision(env, regno);
5380 } else if (arg_type_is_alloc_size(arg_type)) {
5381 if (!tnum_is_const(reg->var_off)) {
5382 verbose(env, "R%d is not a known constant'\n",
5386 meta->mem_size = reg->var_off.value;
5387 } else if (arg_type_is_int_ptr(arg_type)) {
5388 int size = int_ptr_type_to_size(arg_type);
5390 err = check_helper_mem_access(env, regno, size, false, meta);
5393 err = check_ptr_alignment(env, reg, 0, size, true);
5394 } else if (arg_type == ARG_PTR_TO_CONST_STR) {
5395 struct bpf_map *map = reg->map_ptr;
5400 if (!bpf_map_is_rdonly(map)) {
5401 verbose(env, "R%d does not point to a readonly map'\n", regno);
5405 if (!tnum_is_const(reg->var_off)) {
5406 verbose(env, "R%d is not a constant address'\n", regno);
5410 if (!map->ops->map_direct_value_addr) {
5411 verbose(env, "no direct value access support for this map type\n");
5415 err = check_map_access(env, regno, reg->off,
5416 map->value_size - reg->off, false);
5420 map_off = reg->off + reg->var_off.value;
5421 err = map->ops->map_direct_value_addr(map, &map_addr, map_off);
5423 verbose(env, "direct value access on string failed\n");
5427 str_ptr = (char *)(long)(map_addr);
5428 if (!strnchr(str_ptr + map_off, map->value_size - map_off, 0)) {
5429 verbose(env, "string is not zero-terminated\n");
5437 static bool may_update_sockmap(struct bpf_verifier_env *env, int func_id)
5439 enum bpf_attach_type eatype = env->prog->expected_attach_type;
5440 enum bpf_prog_type type = resolve_prog_type(env->prog);
5442 if (func_id != BPF_FUNC_map_update_elem)
5445 /* It's not possible to get access to a locked struct sock in these
5446 * contexts, so updating is safe.
5449 case BPF_PROG_TYPE_TRACING:
5450 if (eatype == BPF_TRACE_ITER)
5453 case BPF_PROG_TYPE_SOCKET_FILTER:
5454 case BPF_PROG_TYPE_SCHED_CLS:
5455 case BPF_PROG_TYPE_SCHED_ACT:
5456 case BPF_PROG_TYPE_XDP:
5457 case BPF_PROG_TYPE_SK_REUSEPORT:
5458 case BPF_PROG_TYPE_FLOW_DISSECTOR:
5459 case BPF_PROG_TYPE_SK_LOOKUP:
5465 verbose(env, "cannot update sockmap in this context\n");
5469 static bool allow_tail_call_in_subprogs(struct bpf_verifier_env *env)
5471 return env->prog->jit_requested && IS_ENABLED(CONFIG_X86_64);
5474 static int check_map_func_compatibility(struct bpf_verifier_env *env,
5475 struct bpf_map *map, int func_id)
5480 /* We need a two way check, first is from map perspective ... */
5481 switch (map->map_type) {
5482 case BPF_MAP_TYPE_PROG_ARRAY:
5483 if (func_id != BPF_FUNC_tail_call)
5486 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
5487 if (func_id != BPF_FUNC_perf_event_read &&
5488 func_id != BPF_FUNC_perf_event_output &&
5489 func_id != BPF_FUNC_skb_output &&
5490 func_id != BPF_FUNC_perf_event_read_value &&
5491 func_id != BPF_FUNC_xdp_output)
5494 case BPF_MAP_TYPE_RINGBUF:
5495 if (func_id != BPF_FUNC_ringbuf_output &&
5496 func_id != BPF_FUNC_ringbuf_reserve &&
5497 func_id != BPF_FUNC_ringbuf_query)
5500 case BPF_MAP_TYPE_STACK_TRACE:
5501 if (func_id != BPF_FUNC_get_stackid)
5504 case BPF_MAP_TYPE_CGROUP_ARRAY:
5505 if (func_id != BPF_FUNC_skb_under_cgroup &&
5506 func_id != BPF_FUNC_current_task_under_cgroup)
5509 case BPF_MAP_TYPE_CGROUP_STORAGE:
5510 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
5511 if (func_id != BPF_FUNC_get_local_storage)
5514 case BPF_MAP_TYPE_DEVMAP:
5515 case BPF_MAP_TYPE_DEVMAP_HASH:
5516 if (func_id != BPF_FUNC_redirect_map &&
5517 func_id != BPF_FUNC_map_lookup_elem)
5520 /* Restrict bpf side of cpumap and xskmap, open when use-cases
5523 case BPF_MAP_TYPE_CPUMAP:
5524 if (func_id != BPF_FUNC_redirect_map)
5527 case BPF_MAP_TYPE_XSKMAP:
5528 if (func_id != BPF_FUNC_redirect_map &&
5529 func_id != BPF_FUNC_map_lookup_elem)
5532 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
5533 case BPF_MAP_TYPE_HASH_OF_MAPS:
5534 if (func_id != BPF_FUNC_map_lookup_elem)
5537 case BPF_MAP_TYPE_SOCKMAP:
5538 if (func_id != BPF_FUNC_sk_redirect_map &&
5539 func_id != BPF_FUNC_sock_map_update &&
5540 func_id != BPF_FUNC_map_delete_elem &&
5541 func_id != BPF_FUNC_msg_redirect_map &&
5542 func_id != BPF_FUNC_sk_select_reuseport &&
5543 func_id != BPF_FUNC_map_lookup_elem &&
5544 !may_update_sockmap(env, func_id))
5547 case BPF_MAP_TYPE_SOCKHASH:
5548 if (func_id != BPF_FUNC_sk_redirect_hash &&
5549 func_id != BPF_FUNC_sock_hash_update &&
5550 func_id != BPF_FUNC_map_delete_elem &&
5551 func_id != BPF_FUNC_msg_redirect_hash &&
5552 func_id != BPF_FUNC_sk_select_reuseport &&
5553 func_id != BPF_FUNC_map_lookup_elem &&
5554 !may_update_sockmap(env, func_id))
5557 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
5558 if (func_id != BPF_FUNC_sk_select_reuseport)
5561 case BPF_MAP_TYPE_QUEUE:
5562 case BPF_MAP_TYPE_STACK:
5563 if (func_id != BPF_FUNC_map_peek_elem &&
5564 func_id != BPF_FUNC_map_pop_elem &&
5565 func_id != BPF_FUNC_map_push_elem)
5568 case BPF_MAP_TYPE_SK_STORAGE:
5569 if (func_id != BPF_FUNC_sk_storage_get &&
5570 func_id != BPF_FUNC_sk_storage_delete)
5573 case BPF_MAP_TYPE_INODE_STORAGE:
5574 if (func_id != BPF_FUNC_inode_storage_get &&
5575 func_id != BPF_FUNC_inode_storage_delete)
5578 case BPF_MAP_TYPE_TASK_STORAGE:
5579 if (func_id != BPF_FUNC_task_storage_get &&
5580 func_id != BPF_FUNC_task_storage_delete)
5583 case BPF_MAP_TYPE_BLOOM_FILTER:
5584 if (func_id != BPF_FUNC_map_peek_elem &&
5585 func_id != BPF_FUNC_map_push_elem)
5592 /* ... and second from the function itself. */
5594 case BPF_FUNC_tail_call:
5595 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
5597 if (env->subprog_cnt > 1 && !allow_tail_call_in_subprogs(env)) {
5598 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
5602 case BPF_FUNC_perf_event_read:
5603 case BPF_FUNC_perf_event_output:
5604 case BPF_FUNC_perf_event_read_value:
5605 case BPF_FUNC_skb_output:
5606 case BPF_FUNC_xdp_output:
5607 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
5610 case BPF_FUNC_ringbuf_output:
5611 case BPF_FUNC_ringbuf_reserve:
5612 case BPF_FUNC_ringbuf_query:
5613 if (map->map_type != BPF_MAP_TYPE_RINGBUF)
5616 case BPF_FUNC_get_stackid:
5617 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
5620 case BPF_FUNC_current_task_under_cgroup:
5621 case BPF_FUNC_skb_under_cgroup:
5622 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
5625 case BPF_FUNC_redirect_map:
5626 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
5627 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
5628 map->map_type != BPF_MAP_TYPE_CPUMAP &&
5629 map->map_type != BPF_MAP_TYPE_XSKMAP)
5632 case BPF_FUNC_sk_redirect_map:
5633 case BPF_FUNC_msg_redirect_map:
5634 case BPF_FUNC_sock_map_update:
5635 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
5638 case BPF_FUNC_sk_redirect_hash:
5639 case BPF_FUNC_msg_redirect_hash:
5640 case BPF_FUNC_sock_hash_update:
5641 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
5644 case BPF_FUNC_get_local_storage:
5645 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
5646 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
5649 case BPF_FUNC_sk_select_reuseport:
5650 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
5651 map->map_type != BPF_MAP_TYPE_SOCKMAP &&
5652 map->map_type != BPF_MAP_TYPE_SOCKHASH)
5655 case BPF_FUNC_map_pop_elem:
5656 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
5657 map->map_type != BPF_MAP_TYPE_STACK)
5660 case BPF_FUNC_map_peek_elem:
5661 case BPF_FUNC_map_push_elem:
5662 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
5663 map->map_type != BPF_MAP_TYPE_STACK &&
5664 map->map_type != BPF_MAP_TYPE_BLOOM_FILTER)
5667 case BPF_FUNC_sk_storage_get:
5668 case BPF_FUNC_sk_storage_delete:
5669 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
5672 case BPF_FUNC_inode_storage_get:
5673 case BPF_FUNC_inode_storage_delete:
5674 if (map->map_type != BPF_MAP_TYPE_INODE_STORAGE)
5677 case BPF_FUNC_task_storage_get:
5678 case BPF_FUNC_task_storage_delete:
5679 if (map->map_type != BPF_MAP_TYPE_TASK_STORAGE)
5688 verbose(env, "cannot pass map_type %d into func %s#%d\n",
5689 map->map_type, func_id_name(func_id), func_id);
5693 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
5697 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
5699 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
5701 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
5703 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
5705 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
5708 /* We only support one arg being in raw mode at the moment,
5709 * which is sufficient for the helper functions we have
5715 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
5716 enum bpf_arg_type arg_next)
5718 return (arg_type_is_mem_ptr(arg_curr) &&
5719 !arg_type_is_mem_size(arg_next)) ||
5720 (!arg_type_is_mem_ptr(arg_curr) &&
5721 arg_type_is_mem_size(arg_next));
5724 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
5726 /* bpf_xxx(..., buf, len) call will access 'len'
5727 * bytes from memory 'buf'. Both arg types need
5728 * to be paired, so make sure there's no buggy
5729 * helper function specification.
5731 if (arg_type_is_mem_size(fn->arg1_type) ||
5732 arg_type_is_mem_ptr(fn->arg5_type) ||
5733 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
5734 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
5735 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
5736 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
5742 static bool check_refcount_ok(const struct bpf_func_proto *fn, int func_id)
5746 if (arg_type_may_be_refcounted(fn->arg1_type))
5748 if (arg_type_may_be_refcounted(fn->arg2_type))
5750 if (arg_type_may_be_refcounted(fn->arg3_type))
5752 if (arg_type_may_be_refcounted(fn->arg4_type))
5754 if (arg_type_may_be_refcounted(fn->arg5_type))
5757 /* A reference acquiring function cannot acquire
5758 * another refcounted ptr.
5760 if (may_be_acquire_function(func_id) && count)
5763 /* We only support one arg being unreferenced at the moment,
5764 * which is sufficient for the helper functions we have right now.
5769 static bool check_btf_id_ok(const struct bpf_func_proto *fn)
5773 for (i = 0; i < ARRAY_SIZE(fn->arg_type); i++) {
5774 if (fn->arg_type[i] == ARG_PTR_TO_BTF_ID && !fn->arg_btf_id[i])
5777 if (fn->arg_type[i] != ARG_PTR_TO_BTF_ID && fn->arg_btf_id[i])
5784 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
5786 return check_raw_mode_ok(fn) &&
5787 check_arg_pair_ok(fn) &&
5788 check_btf_id_ok(fn) &&
5789 check_refcount_ok(fn, func_id) ? 0 : -EINVAL;
5792 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
5793 * are now invalid, so turn them into unknown SCALAR_VALUE.
5795 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
5796 struct bpf_func_state *state)
5798 struct bpf_reg_state *regs = state->regs, *reg;
5801 for (i = 0; i < MAX_BPF_REG; i++)
5802 if (reg_is_pkt_pointer_any(®s[i]))
5803 mark_reg_unknown(env, regs, i);
5805 bpf_for_each_spilled_reg(i, state, reg) {
5808 if (reg_is_pkt_pointer_any(reg))
5809 __mark_reg_unknown(env, reg);
5813 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
5815 struct bpf_verifier_state *vstate = env->cur_state;
5818 for (i = 0; i <= vstate->curframe; i++)
5819 __clear_all_pkt_pointers(env, vstate->frame[i]);
5824 BEYOND_PKT_END = -2,
5827 static void mark_pkt_end(struct bpf_verifier_state *vstate, int regn, bool range_open)
5829 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5830 struct bpf_reg_state *reg = &state->regs[regn];
5832 if (reg->type != PTR_TO_PACKET)
5833 /* PTR_TO_PACKET_META is not supported yet */
5836 /* The 'reg' is pkt > pkt_end or pkt >= pkt_end.
5837 * How far beyond pkt_end it goes is unknown.
5838 * if (!range_open) it's the case of pkt >= pkt_end
5839 * if (range_open) it's the case of pkt > pkt_end
5840 * hence this pointer is at least 1 byte bigger than pkt_end
5843 reg->range = BEYOND_PKT_END;
5845 reg->range = AT_PKT_END;
5848 static void release_reg_references(struct bpf_verifier_env *env,
5849 struct bpf_func_state *state,
5852 struct bpf_reg_state *regs = state->regs, *reg;
5855 for (i = 0; i < MAX_BPF_REG; i++)
5856 if (regs[i].ref_obj_id == ref_obj_id)
5857 mark_reg_unknown(env, regs, i);
5859 bpf_for_each_spilled_reg(i, state, reg) {
5862 if (reg->ref_obj_id == ref_obj_id)
5863 __mark_reg_unknown(env, reg);
5867 /* The pointer with the specified id has released its reference to kernel
5868 * resources. Identify all copies of the same pointer and clear the reference.
5870 static int release_reference(struct bpf_verifier_env *env,
5873 struct bpf_verifier_state *vstate = env->cur_state;
5877 err = release_reference_state(cur_func(env), ref_obj_id);
5881 for (i = 0; i <= vstate->curframe; i++)
5882 release_reg_references(env, vstate->frame[i], ref_obj_id);
5887 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
5888 struct bpf_reg_state *regs)
5892 /* after the call registers r0 - r5 were scratched */
5893 for (i = 0; i < CALLER_SAVED_REGS; i++) {
5894 mark_reg_not_init(env, regs, caller_saved[i]);
5895 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
5899 typedef int (*set_callee_state_fn)(struct bpf_verifier_env *env,
5900 struct bpf_func_state *caller,
5901 struct bpf_func_state *callee,
5904 static int __check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
5905 int *insn_idx, int subprog,
5906 set_callee_state_fn set_callee_state_cb)
5908 struct bpf_verifier_state *state = env->cur_state;
5909 struct bpf_func_info_aux *func_info_aux;
5910 struct bpf_func_state *caller, *callee;
5912 bool is_global = false;
5914 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
5915 verbose(env, "the call stack of %d frames is too deep\n",
5916 state->curframe + 2);
5920 caller = state->frame[state->curframe];
5921 if (state->frame[state->curframe + 1]) {
5922 verbose(env, "verifier bug. Frame %d already allocated\n",
5923 state->curframe + 1);
5927 func_info_aux = env->prog->aux->func_info_aux;
5929 is_global = func_info_aux[subprog].linkage == BTF_FUNC_GLOBAL;
5930 err = btf_check_subprog_arg_match(env, subprog, caller->regs);
5935 verbose(env, "Caller passes invalid args into func#%d\n",
5939 if (env->log.level & BPF_LOG_LEVEL)
5941 "Func#%d is global and valid. Skipping.\n",
5943 clear_caller_saved_regs(env, caller->regs);
5945 /* All global functions return a 64-bit SCALAR_VALUE */
5946 mark_reg_unknown(env, caller->regs, BPF_REG_0);
5947 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
5949 /* continue with next insn after call */
5954 if (insn->code == (BPF_JMP | BPF_CALL) &&
5955 insn->imm == BPF_FUNC_timer_set_callback) {
5956 struct bpf_verifier_state *async_cb;
5958 /* there is no real recursion here. timer callbacks are async */
5959 env->subprog_info[subprog].is_async_cb = true;
5960 async_cb = push_async_cb(env, env->subprog_info[subprog].start,
5961 *insn_idx, subprog);
5964 callee = async_cb->frame[0];
5965 callee->async_entry_cnt = caller->async_entry_cnt + 1;
5967 /* Convert bpf_timer_set_callback() args into timer callback args */
5968 err = set_callee_state_cb(env, caller, callee, *insn_idx);
5972 clear_caller_saved_regs(env, caller->regs);
5973 mark_reg_unknown(env, caller->regs, BPF_REG_0);
5974 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
5975 /* continue with next insn after call */
5979 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
5982 state->frame[state->curframe + 1] = callee;
5984 /* callee cannot access r0, r6 - r9 for reading and has to write
5985 * into its own stack before reading from it.
5986 * callee can read/write into caller's stack
5988 init_func_state(env, callee,
5989 /* remember the callsite, it will be used by bpf_exit */
5990 *insn_idx /* callsite */,
5991 state->curframe + 1 /* frameno within this callchain */,
5992 subprog /* subprog number within this prog */);
5994 /* Transfer references to the callee */
5995 err = copy_reference_state(callee, caller);
5999 err = set_callee_state_cb(env, caller, callee, *insn_idx);
6003 clear_caller_saved_regs(env, caller->regs);
6005 /* only increment it after check_reg_arg() finished */
6008 /* and go analyze first insn of the callee */
6009 *insn_idx = env->subprog_info[subprog].start - 1;
6011 if (env->log.level & BPF_LOG_LEVEL) {
6012 verbose(env, "caller:\n");
6013 print_verifier_state(env, caller);
6014 verbose(env, "callee:\n");
6015 print_verifier_state(env, callee);
6020 int map_set_for_each_callback_args(struct bpf_verifier_env *env,
6021 struct bpf_func_state *caller,
6022 struct bpf_func_state *callee)
6024 /* bpf_for_each_map_elem(struct bpf_map *map, void *callback_fn,
6025 * void *callback_ctx, u64 flags);
6026 * callback_fn(struct bpf_map *map, void *key, void *value,
6027 * void *callback_ctx);
6029 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
6031 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
6032 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
6033 callee->regs[BPF_REG_2].map_ptr = caller->regs[BPF_REG_1].map_ptr;
6035 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
6036 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
6037 callee->regs[BPF_REG_3].map_ptr = caller->regs[BPF_REG_1].map_ptr;
6039 /* pointer to stack or null */
6040 callee->regs[BPF_REG_4] = caller->regs[BPF_REG_3];
6043 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
6047 static int set_callee_state(struct bpf_verifier_env *env,
6048 struct bpf_func_state *caller,
6049 struct bpf_func_state *callee, int insn_idx)
6053 /* copy r1 - r5 args that callee can access. The copy includes parent
6054 * pointers, which connects us up to the liveness chain
6056 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
6057 callee->regs[i] = caller->regs[i];
6061 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
6064 int subprog, target_insn;
6066 target_insn = *insn_idx + insn->imm + 1;
6067 subprog = find_subprog(env, target_insn);
6069 verbose(env, "verifier bug. No program starts at insn %d\n",
6074 return __check_func_call(env, insn, insn_idx, subprog, set_callee_state);
6077 static int set_map_elem_callback_state(struct bpf_verifier_env *env,
6078 struct bpf_func_state *caller,
6079 struct bpf_func_state *callee,
6082 struct bpf_insn_aux_data *insn_aux = &env->insn_aux_data[insn_idx];
6083 struct bpf_map *map;
6086 if (bpf_map_ptr_poisoned(insn_aux)) {
6087 verbose(env, "tail_call abusing map_ptr\n");
6091 map = BPF_MAP_PTR(insn_aux->map_ptr_state);
6092 if (!map->ops->map_set_for_each_callback_args ||
6093 !map->ops->map_for_each_callback) {
6094 verbose(env, "callback function not allowed for map\n");
6098 err = map->ops->map_set_for_each_callback_args(env, caller, callee);
6102 callee->in_callback_fn = true;
6106 static int set_loop_callback_state(struct bpf_verifier_env *env,
6107 struct bpf_func_state *caller,
6108 struct bpf_func_state *callee,
6111 /* bpf_loop(u32 nr_loops, void *callback_fn, void *callback_ctx,
6113 * callback_fn(u32 index, void *callback_ctx);
6115 callee->regs[BPF_REG_1].type = SCALAR_VALUE;
6116 callee->regs[BPF_REG_2] = caller->regs[BPF_REG_3];
6119 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
6120 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
6121 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
6123 callee->in_callback_fn = true;
6127 static int set_timer_callback_state(struct bpf_verifier_env *env,
6128 struct bpf_func_state *caller,
6129 struct bpf_func_state *callee,
6132 struct bpf_map *map_ptr = caller->regs[BPF_REG_1].map_ptr;
6134 /* bpf_timer_set_callback(struct bpf_timer *timer, void *callback_fn);
6135 * callback_fn(struct bpf_map *map, void *key, void *value);
6137 callee->regs[BPF_REG_1].type = CONST_PTR_TO_MAP;
6138 __mark_reg_known_zero(&callee->regs[BPF_REG_1]);
6139 callee->regs[BPF_REG_1].map_ptr = map_ptr;
6141 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
6142 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
6143 callee->regs[BPF_REG_2].map_ptr = map_ptr;
6145 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
6146 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
6147 callee->regs[BPF_REG_3].map_ptr = map_ptr;
6150 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
6151 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
6152 callee->in_async_callback_fn = true;
6156 static int set_find_vma_callback_state(struct bpf_verifier_env *env,
6157 struct bpf_func_state *caller,
6158 struct bpf_func_state *callee,
6161 /* bpf_find_vma(struct task_struct *task, u64 addr,
6162 * void *callback_fn, void *callback_ctx, u64 flags)
6163 * (callback_fn)(struct task_struct *task,
6164 * struct vm_area_struct *vma, void *callback_ctx);
6166 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
6168 callee->regs[BPF_REG_2].type = PTR_TO_BTF_ID;
6169 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
6170 callee->regs[BPF_REG_2].btf = btf_vmlinux;
6171 callee->regs[BPF_REG_2].btf_id = btf_tracing_ids[BTF_TRACING_TYPE_VMA],
6173 /* pointer to stack or null */
6174 callee->regs[BPF_REG_3] = caller->regs[BPF_REG_4];
6177 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
6178 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
6179 callee->in_callback_fn = true;
6183 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
6185 struct bpf_verifier_state *state = env->cur_state;
6186 struct bpf_func_state *caller, *callee;
6187 struct bpf_reg_state *r0;
6190 callee = state->frame[state->curframe];
6191 r0 = &callee->regs[BPF_REG_0];
6192 if (r0->type == PTR_TO_STACK) {
6193 /* technically it's ok to return caller's stack pointer
6194 * (or caller's caller's pointer) back to the caller,
6195 * since these pointers are valid. Only current stack
6196 * pointer will be invalid as soon as function exits,
6197 * but let's be conservative
6199 verbose(env, "cannot return stack pointer to the caller\n");
6204 caller = state->frame[state->curframe];
6205 if (callee->in_callback_fn) {
6206 /* enforce R0 return value range [0, 1]. */
6207 struct tnum range = tnum_range(0, 1);
6209 if (r0->type != SCALAR_VALUE) {
6210 verbose(env, "R0 not a scalar value\n");
6213 if (!tnum_in(range, r0->var_off)) {
6214 verbose_invalid_scalar(env, r0, &range, "callback return", "R0");
6218 /* return to the caller whatever r0 had in the callee */
6219 caller->regs[BPF_REG_0] = *r0;
6222 /* Transfer references to the caller */
6223 err = copy_reference_state(caller, callee);
6227 *insn_idx = callee->callsite + 1;
6228 if (env->log.level & BPF_LOG_LEVEL) {
6229 verbose(env, "returning from callee:\n");
6230 print_verifier_state(env, callee);
6231 verbose(env, "to caller at %d:\n", *insn_idx);
6232 print_verifier_state(env, caller);
6234 /* clear everything in the callee */
6235 free_func_state(callee);
6236 state->frame[state->curframe + 1] = NULL;
6240 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
6242 struct bpf_call_arg_meta *meta)
6244 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
6246 if (ret_type != RET_INTEGER ||
6247 (func_id != BPF_FUNC_get_stack &&
6248 func_id != BPF_FUNC_get_task_stack &&
6249 func_id != BPF_FUNC_probe_read_str &&
6250 func_id != BPF_FUNC_probe_read_kernel_str &&
6251 func_id != BPF_FUNC_probe_read_user_str))
6254 ret_reg->smax_value = meta->msize_max_value;
6255 ret_reg->s32_max_value = meta->msize_max_value;
6256 ret_reg->smin_value = -MAX_ERRNO;
6257 ret_reg->s32_min_value = -MAX_ERRNO;
6258 __reg_deduce_bounds(ret_reg);
6259 __reg_bound_offset(ret_reg);
6260 __update_reg_bounds(ret_reg);
6264 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
6265 int func_id, int insn_idx)
6267 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
6268 struct bpf_map *map = meta->map_ptr;
6270 if (func_id != BPF_FUNC_tail_call &&
6271 func_id != BPF_FUNC_map_lookup_elem &&
6272 func_id != BPF_FUNC_map_update_elem &&
6273 func_id != BPF_FUNC_map_delete_elem &&
6274 func_id != BPF_FUNC_map_push_elem &&
6275 func_id != BPF_FUNC_map_pop_elem &&
6276 func_id != BPF_FUNC_map_peek_elem &&
6277 func_id != BPF_FUNC_for_each_map_elem &&
6278 func_id != BPF_FUNC_redirect_map)
6282 verbose(env, "kernel subsystem misconfigured verifier\n");
6286 /* In case of read-only, some additional restrictions
6287 * need to be applied in order to prevent altering the
6288 * state of the map from program side.
6290 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
6291 (func_id == BPF_FUNC_map_delete_elem ||
6292 func_id == BPF_FUNC_map_update_elem ||
6293 func_id == BPF_FUNC_map_push_elem ||
6294 func_id == BPF_FUNC_map_pop_elem)) {
6295 verbose(env, "write into map forbidden\n");
6299 if (!BPF_MAP_PTR(aux->map_ptr_state))
6300 bpf_map_ptr_store(aux, meta->map_ptr,
6301 !meta->map_ptr->bypass_spec_v1);
6302 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
6303 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
6304 !meta->map_ptr->bypass_spec_v1);
6309 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
6310 int func_id, int insn_idx)
6312 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
6313 struct bpf_reg_state *regs = cur_regs(env), *reg;
6314 struct bpf_map *map = meta->map_ptr;
6319 if (func_id != BPF_FUNC_tail_call)
6321 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
6322 verbose(env, "kernel subsystem misconfigured verifier\n");
6326 range = tnum_range(0, map->max_entries - 1);
6327 reg = ®s[BPF_REG_3];
6329 if (!register_is_const(reg) || !tnum_in(range, reg->var_off)) {
6330 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
6334 err = mark_chain_precision(env, BPF_REG_3);
6338 val = reg->var_off.value;
6339 if (bpf_map_key_unseen(aux))
6340 bpf_map_key_store(aux, val);
6341 else if (!bpf_map_key_poisoned(aux) &&
6342 bpf_map_key_immediate(aux) != val)
6343 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
6347 static int check_reference_leak(struct bpf_verifier_env *env)
6349 struct bpf_func_state *state = cur_func(env);
6352 for (i = 0; i < state->acquired_refs; i++) {
6353 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
6354 state->refs[i].id, state->refs[i].insn_idx);
6356 return state->acquired_refs ? -EINVAL : 0;
6359 static int check_bpf_snprintf_call(struct bpf_verifier_env *env,
6360 struct bpf_reg_state *regs)
6362 struct bpf_reg_state *fmt_reg = ®s[BPF_REG_3];
6363 struct bpf_reg_state *data_len_reg = ®s[BPF_REG_5];
6364 struct bpf_map *fmt_map = fmt_reg->map_ptr;
6365 int err, fmt_map_off, num_args;
6369 /* data must be an array of u64 */
6370 if (data_len_reg->var_off.value % 8)
6372 num_args = data_len_reg->var_off.value / 8;
6374 /* fmt being ARG_PTR_TO_CONST_STR guarantees that var_off is const
6375 * and map_direct_value_addr is set.
6377 fmt_map_off = fmt_reg->off + fmt_reg->var_off.value;
6378 err = fmt_map->ops->map_direct_value_addr(fmt_map, &fmt_addr,
6381 verbose(env, "verifier bug\n");
6384 fmt = (char *)(long)fmt_addr + fmt_map_off;
6386 /* We are also guaranteed that fmt+fmt_map_off is NULL terminated, we
6387 * can focus on validating the format specifiers.
6389 err = bpf_bprintf_prepare(fmt, UINT_MAX, NULL, NULL, num_args);
6391 verbose(env, "Invalid format string\n");
6396 static int check_get_func_ip(struct bpf_verifier_env *env)
6398 enum bpf_prog_type type = resolve_prog_type(env->prog);
6399 int func_id = BPF_FUNC_get_func_ip;
6401 if (type == BPF_PROG_TYPE_TRACING) {
6402 if (!bpf_prog_has_trampoline(env->prog)) {
6403 verbose(env, "func %s#%d supported only for fentry/fexit/fmod_ret programs\n",
6404 func_id_name(func_id), func_id);
6408 } else if (type == BPF_PROG_TYPE_KPROBE) {
6412 verbose(env, "func %s#%d not supported for program type %d\n",
6413 func_id_name(func_id), func_id, type);
6417 static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
6420 const struct bpf_func_proto *fn = NULL;
6421 struct bpf_reg_state *regs;
6422 struct bpf_call_arg_meta meta;
6423 int insn_idx = *insn_idx_p;
6425 int i, err, func_id;
6427 /* find function prototype */
6428 func_id = insn->imm;
6429 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
6430 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
6435 if (env->ops->get_func_proto)
6436 fn = env->ops->get_func_proto(func_id, env->prog);
6438 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
6443 /* eBPF programs must be GPL compatible to use GPL-ed functions */
6444 if (!env->prog->gpl_compatible && fn->gpl_only) {
6445 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
6449 if (fn->allowed && !fn->allowed(env->prog)) {
6450 verbose(env, "helper call is not allowed in probe\n");
6454 /* With LD_ABS/IND some JITs save/restore skb from r1. */
6455 changes_data = bpf_helper_changes_pkt_data(fn->func);
6456 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
6457 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
6458 func_id_name(func_id), func_id);
6462 memset(&meta, 0, sizeof(meta));
6463 meta.pkt_access = fn->pkt_access;
6465 err = check_func_proto(fn, func_id);
6467 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
6468 func_id_name(func_id), func_id);
6472 meta.func_id = func_id;
6474 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++) {
6475 err = check_func_arg(env, i, &meta, fn);
6480 err = record_func_map(env, &meta, func_id, insn_idx);
6484 err = record_func_key(env, &meta, func_id, insn_idx);
6488 /* Mark slots with STACK_MISC in case of raw mode, stack offset
6489 * is inferred from register state.
6491 for (i = 0; i < meta.access_size; i++) {
6492 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
6493 BPF_WRITE, -1, false);
6498 if (is_release_function(func_id)) {
6499 err = release_reference(env, meta.ref_obj_id);
6501 verbose(env, "func %s#%d reference has not been acquired before\n",
6502 func_id_name(func_id), func_id);
6507 regs = cur_regs(env);
6510 case BPF_FUNC_tail_call:
6511 err = check_reference_leak(env);
6513 verbose(env, "tail_call would lead to reference leak\n");
6517 case BPF_FUNC_get_local_storage:
6518 /* check that flags argument in get_local_storage(map, flags) is 0,
6519 * this is required because get_local_storage() can't return an error.
6521 if (!register_is_null(®s[BPF_REG_2])) {
6522 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
6526 case BPF_FUNC_for_each_map_elem:
6527 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
6528 set_map_elem_callback_state);
6530 case BPF_FUNC_timer_set_callback:
6531 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
6532 set_timer_callback_state);
6534 case BPF_FUNC_find_vma:
6535 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
6536 set_find_vma_callback_state);
6538 case BPF_FUNC_snprintf:
6539 err = check_bpf_snprintf_call(env, regs);
6542 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
6543 set_loop_callback_state);
6550 /* reset caller saved regs */
6551 for (i = 0; i < CALLER_SAVED_REGS; i++) {
6552 mark_reg_not_init(env, regs, caller_saved[i]);
6553 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
6556 /* helper call returns 64-bit value. */
6557 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
6559 /* update return register (already marked as written above) */
6560 if (fn->ret_type == RET_INTEGER) {
6561 /* sets type to SCALAR_VALUE */
6562 mark_reg_unknown(env, regs, BPF_REG_0);
6563 } else if (fn->ret_type == RET_VOID) {
6564 regs[BPF_REG_0].type = NOT_INIT;
6565 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL ||
6566 fn->ret_type == RET_PTR_TO_MAP_VALUE) {
6567 /* There is no offset yet applied, variable or fixed */
6568 mark_reg_known_zero(env, regs, BPF_REG_0);
6569 /* remember map_ptr, so that check_map_access()
6570 * can check 'value_size' boundary of memory access
6571 * to map element returned from bpf_map_lookup_elem()
6573 if (meta.map_ptr == NULL) {
6575 "kernel subsystem misconfigured verifier\n");
6578 regs[BPF_REG_0].map_ptr = meta.map_ptr;
6579 regs[BPF_REG_0].map_uid = meta.map_uid;
6580 if (fn->ret_type == RET_PTR_TO_MAP_VALUE) {
6581 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE;
6582 if (map_value_has_spin_lock(meta.map_ptr))
6583 regs[BPF_REG_0].id = ++env->id_gen;
6585 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
6587 } else if (fn->ret_type == RET_PTR_TO_SOCKET_OR_NULL) {
6588 mark_reg_known_zero(env, regs, BPF_REG_0);
6589 regs[BPF_REG_0].type = PTR_TO_SOCKET_OR_NULL;
6590 } else if (fn->ret_type == RET_PTR_TO_SOCK_COMMON_OR_NULL) {
6591 mark_reg_known_zero(env, regs, BPF_REG_0);
6592 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON_OR_NULL;
6593 } else if (fn->ret_type == RET_PTR_TO_TCP_SOCK_OR_NULL) {
6594 mark_reg_known_zero(env, regs, BPF_REG_0);
6595 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK_OR_NULL;
6596 } else if (fn->ret_type == RET_PTR_TO_ALLOC_MEM_OR_NULL) {
6597 mark_reg_known_zero(env, regs, BPF_REG_0);
6598 regs[BPF_REG_0].type = PTR_TO_MEM_OR_NULL;
6599 regs[BPF_REG_0].mem_size = meta.mem_size;
6600 } else if (fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID_OR_NULL ||
6601 fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID) {
6602 const struct btf_type *t;
6604 mark_reg_known_zero(env, regs, BPF_REG_0);
6605 t = btf_type_skip_modifiers(meta.ret_btf, meta.ret_btf_id, NULL);
6606 if (!btf_type_is_struct(t)) {
6608 const struct btf_type *ret;
6611 /* resolve the type size of ksym. */
6612 ret = btf_resolve_size(meta.ret_btf, t, &tsize);
6614 tname = btf_name_by_offset(meta.ret_btf, t->name_off);
6615 verbose(env, "unable to resolve the size of type '%s': %ld\n",
6616 tname, PTR_ERR(ret));
6619 regs[BPF_REG_0].type =
6620 fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID ?
6621 PTR_TO_MEM : PTR_TO_MEM_OR_NULL;
6622 regs[BPF_REG_0].mem_size = tsize;
6624 regs[BPF_REG_0].type =
6625 fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID ?
6626 PTR_TO_BTF_ID : PTR_TO_BTF_ID_OR_NULL;
6627 regs[BPF_REG_0].btf = meta.ret_btf;
6628 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
6630 } else if (fn->ret_type == RET_PTR_TO_BTF_ID_OR_NULL ||
6631 fn->ret_type == RET_PTR_TO_BTF_ID) {
6634 mark_reg_known_zero(env, regs, BPF_REG_0);
6635 regs[BPF_REG_0].type = fn->ret_type == RET_PTR_TO_BTF_ID ?
6637 PTR_TO_BTF_ID_OR_NULL;
6638 ret_btf_id = *fn->ret_btf_id;
6639 if (ret_btf_id == 0) {
6640 verbose(env, "invalid return type %d of func %s#%d\n",
6641 fn->ret_type, func_id_name(func_id), func_id);
6644 /* current BPF helper definitions are only coming from
6645 * built-in code with type IDs from vmlinux BTF
6647 regs[BPF_REG_0].btf = btf_vmlinux;
6648 regs[BPF_REG_0].btf_id = ret_btf_id;
6650 verbose(env, "unknown return type %d of func %s#%d\n",
6651 fn->ret_type, func_id_name(func_id), func_id);
6655 if (reg_type_may_be_null(regs[BPF_REG_0].type))
6656 regs[BPF_REG_0].id = ++env->id_gen;
6658 if (is_ptr_cast_function(func_id)) {
6659 /* For release_reference() */
6660 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
6661 } else if (is_acquire_function(func_id, meta.map_ptr)) {
6662 int id = acquire_reference_state(env, insn_idx);
6666 /* For mark_ptr_or_null_reg() */
6667 regs[BPF_REG_0].id = id;
6668 /* For release_reference() */
6669 regs[BPF_REG_0].ref_obj_id = id;
6672 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
6674 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
6678 if ((func_id == BPF_FUNC_get_stack ||
6679 func_id == BPF_FUNC_get_task_stack) &&
6680 !env->prog->has_callchain_buf) {
6681 const char *err_str;
6683 #ifdef CONFIG_PERF_EVENTS
6684 err = get_callchain_buffers(sysctl_perf_event_max_stack);
6685 err_str = "cannot get callchain buffer for func %s#%d\n";
6688 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
6691 verbose(env, err_str, func_id_name(func_id), func_id);
6695 env->prog->has_callchain_buf = true;
6698 if (func_id == BPF_FUNC_get_stackid || func_id == BPF_FUNC_get_stack)
6699 env->prog->call_get_stack = true;
6701 if (func_id == BPF_FUNC_get_func_ip) {
6702 if (check_get_func_ip(env))
6704 env->prog->call_get_func_ip = true;
6708 clear_all_pkt_pointers(env);
6712 /* mark_btf_func_reg_size() is used when the reg size is determined by
6713 * the BTF func_proto's return value size and argument.
6715 static void mark_btf_func_reg_size(struct bpf_verifier_env *env, u32 regno,
6718 struct bpf_reg_state *reg = &cur_regs(env)[regno];
6720 if (regno == BPF_REG_0) {
6721 /* Function return value */
6722 reg->live |= REG_LIVE_WRITTEN;
6723 reg->subreg_def = reg_size == sizeof(u64) ?
6724 DEF_NOT_SUBREG : env->insn_idx + 1;
6726 /* Function argument */
6727 if (reg_size == sizeof(u64)) {
6728 mark_insn_zext(env, reg);
6729 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
6731 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ32);
6736 static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn)
6738 const struct btf_type *t, *func, *func_proto, *ptr_type;
6739 struct bpf_reg_state *regs = cur_regs(env);
6740 const char *func_name, *ptr_type_name;
6741 u32 i, nargs, func_id, ptr_type_id;
6742 struct module *btf_mod = NULL;
6743 const struct btf_param *args;
6744 struct btf *desc_btf;
6747 /* skip for now, but return error when we find this in fixup_kfunc_call */
6751 desc_btf = find_kfunc_desc_btf(env, insn->imm, insn->off, &btf_mod);
6752 if (IS_ERR(desc_btf))
6753 return PTR_ERR(desc_btf);
6755 func_id = insn->imm;
6756 func = btf_type_by_id(desc_btf, func_id);
6757 func_name = btf_name_by_offset(desc_btf, func->name_off);
6758 func_proto = btf_type_by_id(desc_btf, func->type);
6760 if (!env->ops->check_kfunc_call ||
6761 !env->ops->check_kfunc_call(func_id, btf_mod)) {
6762 verbose(env, "calling kernel function %s is not allowed\n",
6767 /* Check the arguments */
6768 err = btf_check_kfunc_arg_match(env, desc_btf, func_id, regs);
6772 for (i = 0; i < CALLER_SAVED_REGS; i++)
6773 mark_reg_not_init(env, regs, caller_saved[i]);
6775 /* Check return type */
6776 t = btf_type_skip_modifiers(desc_btf, func_proto->type, NULL);
6777 if (btf_type_is_scalar(t)) {
6778 mark_reg_unknown(env, regs, BPF_REG_0);
6779 mark_btf_func_reg_size(env, BPF_REG_0, t->size);
6780 } else if (btf_type_is_ptr(t)) {
6781 ptr_type = btf_type_skip_modifiers(desc_btf, t->type,
6783 if (!btf_type_is_struct(ptr_type)) {
6784 ptr_type_name = btf_name_by_offset(desc_btf,
6785 ptr_type->name_off);
6786 verbose(env, "kernel function %s returns pointer type %s %s is not supported\n",
6787 func_name, btf_type_str(ptr_type),
6791 mark_reg_known_zero(env, regs, BPF_REG_0);
6792 regs[BPF_REG_0].btf = desc_btf;
6793 regs[BPF_REG_0].type = PTR_TO_BTF_ID;
6794 regs[BPF_REG_0].btf_id = ptr_type_id;
6795 mark_btf_func_reg_size(env, BPF_REG_0, sizeof(void *));
6796 } /* else { add_kfunc_call() ensures it is btf_type_is_void(t) } */
6798 nargs = btf_type_vlen(func_proto);
6799 args = (const struct btf_param *)(func_proto + 1);
6800 for (i = 0; i < nargs; i++) {
6803 t = btf_type_skip_modifiers(desc_btf, args[i].type, NULL);
6804 if (btf_type_is_ptr(t))
6805 mark_btf_func_reg_size(env, regno, sizeof(void *));
6807 /* scalar. ensured by btf_check_kfunc_arg_match() */
6808 mark_btf_func_reg_size(env, regno, t->size);
6814 static bool signed_add_overflows(s64 a, s64 b)
6816 /* Do the add in u64, where overflow is well-defined */
6817 s64 res = (s64)((u64)a + (u64)b);
6824 static bool signed_add32_overflows(s32 a, s32 b)
6826 /* Do the add in u32, where overflow is well-defined */
6827 s32 res = (s32)((u32)a + (u32)b);
6834 static bool signed_sub_overflows(s64 a, s64 b)
6836 /* Do the sub in u64, where overflow is well-defined */
6837 s64 res = (s64)((u64)a - (u64)b);
6844 static bool signed_sub32_overflows(s32 a, s32 b)
6846 /* Do the sub in u32, where overflow is well-defined */
6847 s32 res = (s32)((u32)a - (u32)b);
6854 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
6855 const struct bpf_reg_state *reg,
6856 enum bpf_reg_type type)
6858 bool known = tnum_is_const(reg->var_off);
6859 s64 val = reg->var_off.value;
6860 s64 smin = reg->smin_value;
6862 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
6863 verbose(env, "math between %s pointer and %lld is not allowed\n",
6864 reg_type_str[type], val);
6868 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
6869 verbose(env, "%s pointer offset %d is not allowed\n",
6870 reg_type_str[type], reg->off);
6874 if (smin == S64_MIN) {
6875 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
6876 reg_type_str[type]);
6880 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
6881 verbose(env, "value %lld makes %s pointer be out of bounds\n",
6882 smin, reg_type_str[type]);
6889 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
6891 return &env->insn_aux_data[env->insn_idx];
6902 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
6903 u32 *alu_limit, bool mask_to_left)
6905 u32 max = 0, ptr_limit = 0;
6907 switch (ptr_reg->type) {
6909 /* Offset 0 is out-of-bounds, but acceptable start for the
6910 * left direction, see BPF_REG_FP. Also, unknown scalar
6911 * offset where we would need to deal with min/max bounds is
6912 * currently prohibited for unprivileged.
6914 max = MAX_BPF_STACK + mask_to_left;
6915 ptr_limit = -(ptr_reg->var_off.value + ptr_reg->off);
6917 case PTR_TO_MAP_VALUE:
6918 max = ptr_reg->map_ptr->value_size;
6919 ptr_limit = (mask_to_left ?
6920 ptr_reg->smin_value :
6921 ptr_reg->umax_value) + ptr_reg->off;
6927 if (ptr_limit >= max)
6928 return REASON_LIMIT;
6929 *alu_limit = ptr_limit;
6933 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
6934 const struct bpf_insn *insn)
6936 return env->bypass_spec_v1 || BPF_SRC(insn->code) == BPF_K;
6939 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
6940 u32 alu_state, u32 alu_limit)
6942 /* If we arrived here from different branches with different
6943 * state or limits to sanitize, then this won't work.
6945 if (aux->alu_state &&
6946 (aux->alu_state != alu_state ||
6947 aux->alu_limit != alu_limit))
6948 return REASON_PATHS;
6950 /* Corresponding fixup done in do_misc_fixups(). */
6951 aux->alu_state = alu_state;
6952 aux->alu_limit = alu_limit;
6956 static int sanitize_val_alu(struct bpf_verifier_env *env,
6957 struct bpf_insn *insn)
6959 struct bpf_insn_aux_data *aux = cur_aux(env);
6961 if (can_skip_alu_sanitation(env, insn))
6964 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
6967 static bool sanitize_needed(u8 opcode)
6969 return opcode == BPF_ADD || opcode == BPF_SUB;
6972 struct bpf_sanitize_info {
6973 struct bpf_insn_aux_data aux;
6977 static struct bpf_verifier_state *
6978 sanitize_speculative_path(struct bpf_verifier_env *env,
6979 const struct bpf_insn *insn,
6980 u32 next_idx, u32 curr_idx)
6982 struct bpf_verifier_state *branch;
6983 struct bpf_reg_state *regs;
6985 branch = push_stack(env, next_idx, curr_idx, true);
6986 if (branch && insn) {
6987 regs = branch->frame[branch->curframe]->regs;
6988 if (BPF_SRC(insn->code) == BPF_K) {
6989 mark_reg_unknown(env, regs, insn->dst_reg);
6990 } else if (BPF_SRC(insn->code) == BPF_X) {
6991 mark_reg_unknown(env, regs, insn->dst_reg);
6992 mark_reg_unknown(env, regs, insn->src_reg);
6998 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
6999 struct bpf_insn *insn,
7000 const struct bpf_reg_state *ptr_reg,
7001 const struct bpf_reg_state *off_reg,
7002 struct bpf_reg_state *dst_reg,
7003 struct bpf_sanitize_info *info,
7004 const bool commit_window)
7006 struct bpf_insn_aux_data *aux = commit_window ? cur_aux(env) : &info->aux;
7007 struct bpf_verifier_state *vstate = env->cur_state;
7008 bool off_is_imm = tnum_is_const(off_reg->var_off);
7009 bool off_is_neg = off_reg->smin_value < 0;
7010 bool ptr_is_dst_reg = ptr_reg == dst_reg;
7011 u8 opcode = BPF_OP(insn->code);
7012 u32 alu_state, alu_limit;
7013 struct bpf_reg_state tmp;
7017 if (can_skip_alu_sanitation(env, insn))
7020 /* We already marked aux for masking from non-speculative
7021 * paths, thus we got here in the first place. We only care
7022 * to explore bad access from here.
7024 if (vstate->speculative)
7027 if (!commit_window) {
7028 if (!tnum_is_const(off_reg->var_off) &&
7029 (off_reg->smin_value < 0) != (off_reg->smax_value < 0))
7030 return REASON_BOUNDS;
7032 info->mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
7033 (opcode == BPF_SUB && !off_is_neg);
7036 err = retrieve_ptr_limit(ptr_reg, &alu_limit, info->mask_to_left);
7040 if (commit_window) {
7041 /* In commit phase we narrow the masking window based on
7042 * the observed pointer move after the simulated operation.
7044 alu_state = info->aux.alu_state;
7045 alu_limit = abs(info->aux.alu_limit - alu_limit);
7047 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
7048 alu_state |= off_is_imm ? BPF_ALU_IMMEDIATE : 0;
7049 alu_state |= ptr_is_dst_reg ?
7050 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
7052 /* Limit pruning on unknown scalars to enable deep search for
7053 * potential masking differences from other program paths.
7056 env->explore_alu_limits = true;
7059 err = update_alu_sanitation_state(aux, alu_state, alu_limit);
7063 /* If we're in commit phase, we're done here given we already
7064 * pushed the truncated dst_reg into the speculative verification
7067 * Also, when register is a known constant, we rewrite register-based
7068 * operation to immediate-based, and thus do not need masking (and as
7069 * a consequence, do not need to simulate the zero-truncation either).
7071 if (commit_window || off_is_imm)
7074 /* Simulate and find potential out-of-bounds access under
7075 * speculative execution from truncation as a result of
7076 * masking when off was not within expected range. If off
7077 * sits in dst, then we temporarily need to move ptr there
7078 * to simulate dst (== 0) +/-= ptr. Needed, for example,
7079 * for cases where we use K-based arithmetic in one direction
7080 * and truncated reg-based in the other in order to explore
7083 if (!ptr_is_dst_reg) {
7085 *dst_reg = *ptr_reg;
7087 ret = sanitize_speculative_path(env, NULL, env->insn_idx + 1,
7089 if (!ptr_is_dst_reg && ret)
7091 return !ret ? REASON_STACK : 0;
7094 static void sanitize_mark_insn_seen(struct bpf_verifier_env *env)
7096 struct bpf_verifier_state *vstate = env->cur_state;
7098 /* If we simulate paths under speculation, we don't update the
7099 * insn as 'seen' such that when we verify unreachable paths in
7100 * the non-speculative domain, sanitize_dead_code() can still
7101 * rewrite/sanitize them.
7103 if (!vstate->speculative)
7104 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
7107 static int sanitize_err(struct bpf_verifier_env *env,
7108 const struct bpf_insn *insn, int reason,
7109 const struct bpf_reg_state *off_reg,
7110 const struct bpf_reg_state *dst_reg)
7112 static const char *err = "pointer arithmetic with it prohibited for !root";
7113 const char *op = BPF_OP(insn->code) == BPF_ADD ? "add" : "sub";
7114 u32 dst = insn->dst_reg, src = insn->src_reg;
7118 verbose(env, "R%d has unknown scalar with mixed signed bounds, %s\n",
7119 off_reg == dst_reg ? dst : src, err);
7122 verbose(env, "R%d has pointer with unsupported alu operation, %s\n",
7123 off_reg == dst_reg ? src : dst, err);
7126 verbose(env, "R%d tried to %s from different maps, paths or scalars, %s\n",
7130 verbose(env, "R%d tried to %s beyond pointer bounds, %s\n",
7134 verbose(env, "R%d could not be pushed for speculative verification, %s\n",
7138 verbose(env, "verifier internal error: unknown reason (%d)\n",
7146 /* check that stack access falls within stack limits and that 'reg' doesn't
7147 * have a variable offset.
7149 * Variable offset is prohibited for unprivileged mode for simplicity since it
7150 * requires corresponding support in Spectre masking for stack ALU. See also
7151 * retrieve_ptr_limit().
7154 * 'off' includes 'reg->off'.
7156 static int check_stack_access_for_ptr_arithmetic(
7157 struct bpf_verifier_env *env,
7159 const struct bpf_reg_state *reg,
7162 if (!tnum_is_const(reg->var_off)) {
7165 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
7166 verbose(env, "R%d variable stack access prohibited for !root, var_off=%s off=%d\n",
7167 regno, tn_buf, off);
7171 if (off >= 0 || off < -MAX_BPF_STACK) {
7172 verbose(env, "R%d stack pointer arithmetic goes out of range, "
7173 "prohibited for !root; off=%d\n", regno, off);
7180 static int sanitize_check_bounds(struct bpf_verifier_env *env,
7181 const struct bpf_insn *insn,
7182 const struct bpf_reg_state *dst_reg)
7184 u32 dst = insn->dst_reg;
7186 /* For unprivileged we require that resulting offset must be in bounds
7187 * in order to be able to sanitize access later on.
7189 if (env->bypass_spec_v1)
7192 switch (dst_reg->type) {
7194 if (check_stack_access_for_ptr_arithmetic(env, dst, dst_reg,
7195 dst_reg->off + dst_reg->var_off.value))
7198 case PTR_TO_MAP_VALUE:
7199 if (check_map_access(env, dst, dst_reg->off, 1, false)) {
7200 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
7201 "prohibited for !root\n", dst);
7212 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
7213 * Caller should also handle BPF_MOV case separately.
7214 * If we return -EACCES, caller may want to try again treating pointer as a
7215 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
7217 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
7218 struct bpf_insn *insn,
7219 const struct bpf_reg_state *ptr_reg,
7220 const struct bpf_reg_state *off_reg)
7222 struct bpf_verifier_state *vstate = env->cur_state;
7223 struct bpf_func_state *state = vstate->frame[vstate->curframe];
7224 struct bpf_reg_state *regs = state->regs, *dst_reg;
7225 bool known = tnum_is_const(off_reg->var_off);
7226 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
7227 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
7228 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
7229 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
7230 struct bpf_sanitize_info info = {};
7231 u8 opcode = BPF_OP(insn->code);
7232 u32 dst = insn->dst_reg;
7235 dst_reg = ®s[dst];
7237 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
7238 smin_val > smax_val || umin_val > umax_val) {
7239 /* Taint dst register if offset had invalid bounds derived from
7240 * e.g. dead branches.
7242 __mark_reg_unknown(env, dst_reg);
7246 if (BPF_CLASS(insn->code) != BPF_ALU64) {
7247 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
7248 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
7249 __mark_reg_unknown(env, dst_reg);
7254 "R%d 32-bit pointer arithmetic prohibited\n",
7259 switch (ptr_reg->type) {
7260 case PTR_TO_MAP_VALUE_OR_NULL:
7261 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
7262 dst, reg_type_str[ptr_reg->type]);
7264 case CONST_PTR_TO_MAP:
7265 /* smin_val represents the known value */
7266 if (known && smin_val == 0 && opcode == BPF_ADD)
7269 case PTR_TO_PACKET_END:
7271 case PTR_TO_SOCKET_OR_NULL:
7272 case PTR_TO_SOCK_COMMON:
7273 case PTR_TO_SOCK_COMMON_OR_NULL:
7274 case PTR_TO_TCP_SOCK:
7275 case PTR_TO_TCP_SOCK_OR_NULL:
7276 case PTR_TO_XDP_SOCK:
7277 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
7278 dst, reg_type_str[ptr_reg->type]);
7284 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
7285 * The id may be overwritten later if we create a new variable offset.
7287 dst_reg->type = ptr_reg->type;
7288 dst_reg->id = ptr_reg->id;
7290 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
7291 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
7294 /* pointer types do not carry 32-bit bounds at the moment. */
7295 __mark_reg32_unbounded(dst_reg);
7297 if (sanitize_needed(opcode)) {
7298 ret = sanitize_ptr_alu(env, insn, ptr_reg, off_reg, dst_reg,
7301 return sanitize_err(env, insn, ret, off_reg, dst_reg);
7306 /* We can take a fixed offset as long as it doesn't overflow
7307 * the s32 'off' field
7309 if (known && (ptr_reg->off + smin_val ==
7310 (s64)(s32)(ptr_reg->off + smin_val))) {
7311 /* pointer += K. Accumulate it into fixed offset */
7312 dst_reg->smin_value = smin_ptr;
7313 dst_reg->smax_value = smax_ptr;
7314 dst_reg->umin_value = umin_ptr;
7315 dst_reg->umax_value = umax_ptr;
7316 dst_reg->var_off = ptr_reg->var_off;
7317 dst_reg->off = ptr_reg->off + smin_val;
7318 dst_reg->raw = ptr_reg->raw;
7321 /* A new variable offset is created. Note that off_reg->off
7322 * == 0, since it's a scalar.
7323 * dst_reg gets the pointer type and since some positive
7324 * integer value was added to the pointer, give it a new 'id'
7325 * if it's a PTR_TO_PACKET.
7326 * this creates a new 'base' pointer, off_reg (variable) gets
7327 * added into the variable offset, and we copy the fixed offset
7330 if (signed_add_overflows(smin_ptr, smin_val) ||
7331 signed_add_overflows(smax_ptr, smax_val)) {
7332 dst_reg->smin_value = S64_MIN;
7333 dst_reg->smax_value = S64_MAX;
7335 dst_reg->smin_value = smin_ptr + smin_val;
7336 dst_reg->smax_value = smax_ptr + smax_val;
7338 if (umin_ptr + umin_val < umin_ptr ||
7339 umax_ptr + umax_val < umax_ptr) {
7340 dst_reg->umin_value = 0;
7341 dst_reg->umax_value = U64_MAX;
7343 dst_reg->umin_value = umin_ptr + umin_val;
7344 dst_reg->umax_value = umax_ptr + umax_val;
7346 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
7347 dst_reg->off = ptr_reg->off;
7348 dst_reg->raw = ptr_reg->raw;
7349 if (reg_is_pkt_pointer(ptr_reg)) {
7350 dst_reg->id = ++env->id_gen;
7351 /* something was added to pkt_ptr, set range to zero */
7352 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
7356 if (dst_reg == off_reg) {
7357 /* scalar -= pointer. Creates an unknown scalar */
7358 verbose(env, "R%d tried to subtract pointer from scalar\n",
7362 /* We don't allow subtraction from FP, because (according to
7363 * test_verifier.c test "invalid fp arithmetic", JITs might not
7364 * be able to deal with it.
7366 if (ptr_reg->type == PTR_TO_STACK) {
7367 verbose(env, "R%d subtraction from stack pointer prohibited\n",
7371 if (known && (ptr_reg->off - smin_val ==
7372 (s64)(s32)(ptr_reg->off - smin_val))) {
7373 /* pointer -= K. Subtract it from fixed offset */
7374 dst_reg->smin_value = smin_ptr;
7375 dst_reg->smax_value = smax_ptr;
7376 dst_reg->umin_value = umin_ptr;
7377 dst_reg->umax_value = umax_ptr;
7378 dst_reg->var_off = ptr_reg->var_off;
7379 dst_reg->id = ptr_reg->id;
7380 dst_reg->off = ptr_reg->off - smin_val;
7381 dst_reg->raw = ptr_reg->raw;
7384 /* A new variable offset is created. If the subtrahend is known
7385 * nonnegative, then any reg->range we had before is still good.
7387 if (signed_sub_overflows(smin_ptr, smax_val) ||
7388 signed_sub_overflows(smax_ptr, smin_val)) {
7389 /* Overflow possible, we know nothing */
7390 dst_reg->smin_value = S64_MIN;
7391 dst_reg->smax_value = S64_MAX;
7393 dst_reg->smin_value = smin_ptr - smax_val;
7394 dst_reg->smax_value = smax_ptr - smin_val;
7396 if (umin_ptr < umax_val) {
7397 /* Overflow possible, we know nothing */
7398 dst_reg->umin_value = 0;
7399 dst_reg->umax_value = U64_MAX;
7401 /* Cannot overflow (as long as bounds are consistent) */
7402 dst_reg->umin_value = umin_ptr - umax_val;
7403 dst_reg->umax_value = umax_ptr - umin_val;
7405 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
7406 dst_reg->off = ptr_reg->off;
7407 dst_reg->raw = ptr_reg->raw;
7408 if (reg_is_pkt_pointer(ptr_reg)) {
7409 dst_reg->id = ++env->id_gen;
7410 /* something was added to pkt_ptr, set range to zero */
7412 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
7418 /* bitwise ops on pointers are troublesome, prohibit. */
7419 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
7420 dst, bpf_alu_string[opcode >> 4]);
7423 /* other operators (e.g. MUL,LSH) produce non-pointer results */
7424 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
7425 dst, bpf_alu_string[opcode >> 4]);
7429 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
7432 __update_reg_bounds(dst_reg);
7433 __reg_deduce_bounds(dst_reg);
7434 __reg_bound_offset(dst_reg);
7436 if (sanitize_check_bounds(env, insn, dst_reg) < 0)
7438 if (sanitize_needed(opcode)) {
7439 ret = sanitize_ptr_alu(env, insn, dst_reg, off_reg, dst_reg,
7442 return sanitize_err(env, insn, ret, off_reg, dst_reg);
7448 static void scalar32_min_max_add(struct bpf_reg_state *dst_reg,
7449 struct bpf_reg_state *src_reg)
7451 s32 smin_val = src_reg->s32_min_value;
7452 s32 smax_val = src_reg->s32_max_value;
7453 u32 umin_val = src_reg->u32_min_value;
7454 u32 umax_val = src_reg->u32_max_value;
7456 if (signed_add32_overflows(dst_reg->s32_min_value, smin_val) ||
7457 signed_add32_overflows(dst_reg->s32_max_value, smax_val)) {
7458 dst_reg->s32_min_value = S32_MIN;
7459 dst_reg->s32_max_value = S32_MAX;
7461 dst_reg->s32_min_value += smin_val;
7462 dst_reg->s32_max_value += smax_val;
7464 if (dst_reg->u32_min_value + umin_val < umin_val ||
7465 dst_reg->u32_max_value + umax_val < umax_val) {
7466 dst_reg->u32_min_value = 0;
7467 dst_reg->u32_max_value = U32_MAX;
7469 dst_reg->u32_min_value += umin_val;
7470 dst_reg->u32_max_value += umax_val;
7474 static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
7475 struct bpf_reg_state *src_reg)
7477 s64 smin_val = src_reg->smin_value;
7478 s64 smax_val = src_reg->smax_value;
7479 u64 umin_val = src_reg->umin_value;
7480 u64 umax_val = src_reg->umax_value;
7482 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
7483 signed_add_overflows(dst_reg->smax_value, smax_val)) {
7484 dst_reg->smin_value = S64_MIN;
7485 dst_reg->smax_value = S64_MAX;
7487 dst_reg->smin_value += smin_val;
7488 dst_reg->smax_value += smax_val;
7490 if (dst_reg->umin_value + umin_val < umin_val ||
7491 dst_reg->umax_value + umax_val < umax_val) {
7492 dst_reg->umin_value = 0;
7493 dst_reg->umax_value = U64_MAX;
7495 dst_reg->umin_value += umin_val;
7496 dst_reg->umax_value += umax_val;
7500 static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg,
7501 struct bpf_reg_state *src_reg)
7503 s32 smin_val = src_reg->s32_min_value;
7504 s32 smax_val = src_reg->s32_max_value;
7505 u32 umin_val = src_reg->u32_min_value;
7506 u32 umax_val = src_reg->u32_max_value;
7508 if (signed_sub32_overflows(dst_reg->s32_min_value, smax_val) ||
7509 signed_sub32_overflows(dst_reg->s32_max_value, smin_val)) {
7510 /* Overflow possible, we know nothing */
7511 dst_reg->s32_min_value = S32_MIN;
7512 dst_reg->s32_max_value = S32_MAX;
7514 dst_reg->s32_min_value -= smax_val;
7515 dst_reg->s32_max_value -= smin_val;
7517 if (dst_reg->u32_min_value < umax_val) {
7518 /* Overflow possible, we know nothing */
7519 dst_reg->u32_min_value = 0;
7520 dst_reg->u32_max_value = U32_MAX;
7522 /* Cannot overflow (as long as bounds are consistent) */
7523 dst_reg->u32_min_value -= umax_val;
7524 dst_reg->u32_max_value -= umin_val;
7528 static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
7529 struct bpf_reg_state *src_reg)
7531 s64 smin_val = src_reg->smin_value;
7532 s64 smax_val = src_reg->smax_value;
7533 u64 umin_val = src_reg->umin_value;
7534 u64 umax_val = src_reg->umax_value;
7536 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
7537 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
7538 /* Overflow possible, we know nothing */
7539 dst_reg->smin_value = S64_MIN;
7540 dst_reg->smax_value = S64_MAX;
7542 dst_reg->smin_value -= smax_val;
7543 dst_reg->smax_value -= smin_val;
7545 if (dst_reg->umin_value < umax_val) {
7546 /* Overflow possible, we know nothing */
7547 dst_reg->umin_value = 0;
7548 dst_reg->umax_value = U64_MAX;
7550 /* Cannot overflow (as long as bounds are consistent) */
7551 dst_reg->umin_value -= umax_val;
7552 dst_reg->umax_value -= umin_val;
7556 static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg,
7557 struct bpf_reg_state *src_reg)
7559 s32 smin_val = src_reg->s32_min_value;
7560 u32 umin_val = src_reg->u32_min_value;
7561 u32 umax_val = src_reg->u32_max_value;
7563 if (smin_val < 0 || dst_reg->s32_min_value < 0) {
7564 /* Ain't nobody got time to multiply that sign */
7565 __mark_reg32_unbounded(dst_reg);
7568 /* Both values are positive, so we can work with unsigned and
7569 * copy the result to signed (unless it exceeds S32_MAX).
7571 if (umax_val > U16_MAX || dst_reg->u32_max_value > U16_MAX) {
7572 /* Potential overflow, we know nothing */
7573 __mark_reg32_unbounded(dst_reg);
7576 dst_reg->u32_min_value *= umin_val;
7577 dst_reg->u32_max_value *= umax_val;
7578 if (dst_reg->u32_max_value > S32_MAX) {
7579 /* Overflow possible, we know nothing */
7580 dst_reg->s32_min_value = S32_MIN;
7581 dst_reg->s32_max_value = S32_MAX;
7583 dst_reg->s32_min_value = dst_reg->u32_min_value;
7584 dst_reg->s32_max_value = dst_reg->u32_max_value;
7588 static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
7589 struct bpf_reg_state *src_reg)
7591 s64 smin_val = src_reg->smin_value;
7592 u64 umin_val = src_reg->umin_value;
7593 u64 umax_val = src_reg->umax_value;
7595 if (smin_val < 0 || dst_reg->smin_value < 0) {
7596 /* Ain't nobody got time to multiply that sign */
7597 __mark_reg64_unbounded(dst_reg);
7600 /* Both values are positive, so we can work with unsigned and
7601 * copy the result to signed (unless it exceeds S64_MAX).
7603 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
7604 /* Potential overflow, we know nothing */
7605 __mark_reg64_unbounded(dst_reg);
7608 dst_reg->umin_value *= umin_val;
7609 dst_reg->umax_value *= umax_val;
7610 if (dst_reg->umax_value > S64_MAX) {
7611 /* Overflow possible, we know nothing */
7612 dst_reg->smin_value = S64_MIN;
7613 dst_reg->smax_value = S64_MAX;
7615 dst_reg->smin_value = dst_reg->umin_value;
7616 dst_reg->smax_value = dst_reg->umax_value;
7620 static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
7621 struct bpf_reg_state *src_reg)
7623 bool src_known = tnum_subreg_is_const(src_reg->var_off);
7624 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
7625 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
7626 s32 smin_val = src_reg->s32_min_value;
7627 u32 umax_val = src_reg->u32_max_value;
7629 if (src_known && dst_known) {
7630 __mark_reg32_known(dst_reg, var32_off.value);
7634 /* We get our minimum from the var_off, since that's inherently
7635 * bitwise. Our maximum is the minimum of the operands' maxima.
7637 dst_reg->u32_min_value = var32_off.value;
7638 dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val);
7639 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
7640 /* Lose signed bounds when ANDing negative numbers,
7641 * ain't nobody got time for that.
7643 dst_reg->s32_min_value = S32_MIN;
7644 dst_reg->s32_max_value = S32_MAX;
7646 /* ANDing two positives gives a positive, so safe to
7647 * cast result into s64.
7649 dst_reg->s32_min_value = dst_reg->u32_min_value;
7650 dst_reg->s32_max_value = dst_reg->u32_max_value;
7654 static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
7655 struct bpf_reg_state *src_reg)
7657 bool src_known = tnum_is_const(src_reg->var_off);
7658 bool dst_known = tnum_is_const(dst_reg->var_off);
7659 s64 smin_val = src_reg->smin_value;
7660 u64 umax_val = src_reg->umax_value;
7662 if (src_known && dst_known) {
7663 __mark_reg_known(dst_reg, dst_reg->var_off.value);
7667 /* We get our minimum from the var_off, since that's inherently
7668 * bitwise. Our maximum is the minimum of the operands' maxima.
7670 dst_reg->umin_value = dst_reg->var_off.value;
7671 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
7672 if (dst_reg->smin_value < 0 || smin_val < 0) {
7673 /* Lose signed bounds when ANDing negative numbers,
7674 * ain't nobody got time for that.
7676 dst_reg->smin_value = S64_MIN;
7677 dst_reg->smax_value = S64_MAX;
7679 /* ANDing two positives gives a positive, so safe to
7680 * cast result into s64.
7682 dst_reg->smin_value = dst_reg->umin_value;
7683 dst_reg->smax_value = dst_reg->umax_value;
7685 /* We may learn something more from the var_off */
7686 __update_reg_bounds(dst_reg);
7689 static void scalar32_min_max_or(struct bpf_reg_state *dst_reg,
7690 struct bpf_reg_state *src_reg)
7692 bool src_known = tnum_subreg_is_const(src_reg->var_off);
7693 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
7694 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
7695 s32 smin_val = src_reg->s32_min_value;
7696 u32 umin_val = src_reg->u32_min_value;
7698 if (src_known && dst_known) {
7699 __mark_reg32_known(dst_reg, var32_off.value);
7703 /* We get our maximum from the var_off, and our minimum is the
7704 * maximum of the operands' minima
7706 dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val);
7707 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
7708 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
7709 /* Lose signed bounds when ORing negative numbers,
7710 * ain't nobody got time for that.
7712 dst_reg->s32_min_value = S32_MIN;
7713 dst_reg->s32_max_value = S32_MAX;
7715 /* ORing two positives gives a positive, so safe to
7716 * cast result into s64.
7718 dst_reg->s32_min_value = dst_reg->u32_min_value;
7719 dst_reg->s32_max_value = dst_reg->u32_max_value;
7723 static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
7724 struct bpf_reg_state *src_reg)
7726 bool src_known = tnum_is_const(src_reg->var_off);
7727 bool dst_known = tnum_is_const(dst_reg->var_off);
7728 s64 smin_val = src_reg->smin_value;
7729 u64 umin_val = src_reg->umin_value;
7731 if (src_known && dst_known) {
7732 __mark_reg_known(dst_reg, dst_reg->var_off.value);
7736 /* We get our maximum from the var_off, and our minimum is the
7737 * maximum of the operands' minima
7739 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
7740 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
7741 if (dst_reg->smin_value < 0 || smin_val < 0) {
7742 /* Lose signed bounds when ORing negative numbers,
7743 * ain't nobody got time for that.
7745 dst_reg->smin_value = S64_MIN;
7746 dst_reg->smax_value = S64_MAX;
7748 /* ORing two positives gives a positive, so safe to
7749 * cast result into s64.
7751 dst_reg->smin_value = dst_reg->umin_value;
7752 dst_reg->smax_value = dst_reg->umax_value;
7754 /* We may learn something more from the var_off */
7755 __update_reg_bounds(dst_reg);
7758 static void scalar32_min_max_xor(struct bpf_reg_state *dst_reg,
7759 struct bpf_reg_state *src_reg)
7761 bool src_known = tnum_subreg_is_const(src_reg->var_off);
7762 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
7763 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
7764 s32 smin_val = src_reg->s32_min_value;
7766 if (src_known && dst_known) {
7767 __mark_reg32_known(dst_reg, var32_off.value);
7771 /* We get both minimum and maximum from the var32_off. */
7772 dst_reg->u32_min_value = var32_off.value;
7773 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
7775 if (dst_reg->s32_min_value >= 0 && smin_val >= 0) {
7776 /* XORing two positive sign numbers gives a positive,
7777 * so safe to cast u32 result into s32.
7779 dst_reg->s32_min_value = dst_reg->u32_min_value;
7780 dst_reg->s32_max_value = dst_reg->u32_max_value;
7782 dst_reg->s32_min_value = S32_MIN;
7783 dst_reg->s32_max_value = S32_MAX;
7787 static void scalar_min_max_xor(struct bpf_reg_state *dst_reg,
7788 struct bpf_reg_state *src_reg)
7790 bool src_known = tnum_is_const(src_reg->var_off);
7791 bool dst_known = tnum_is_const(dst_reg->var_off);
7792 s64 smin_val = src_reg->smin_value;
7794 if (src_known && dst_known) {
7795 /* dst_reg->var_off.value has been updated earlier */
7796 __mark_reg_known(dst_reg, dst_reg->var_off.value);
7800 /* We get both minimum and maximum from the var_off. */
7801 dst_reg->umin_value = dst_reg->var_off.value;
7802 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
7804 if (dst_reg->smin_value >= 0 && smin_val >= 0) {
7805 /* XORing two positive sign numbers gives a positive,
7806 * so safe to cast u64 result into s64.
7808 dst_reg->smin_value = dst_reg->umin_value;
7809 dst_reg->smax_value = dst_reg->umax_value;
7811 dst_reg->smin_value = S64_MIN;
7812 dst_reg->smax_value = S64_MAX;
7815 __update_reg_bounds(dst_reg);
7818 static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
7819 u64 umin_val, u64 umax_val)
7821 /* We lose all sign bit information (except what we can pick
7824 dst_reg->s32_min_value = S32_MIN;
7825 dst_reg->s32_max_value = S32_MAX;
7826 /* If we might shift our top bit out, then we know nothing */
7827 if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) {
7828 dst_reg->u32_min_value = 0;
7829 dst_reg->u32_max_value = U32_MAX;
7831 dst_reg->u32_min_value <<= umin_val;
7832 dst_reg->u32_max_value <<= umax_val;
7836 static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
7837 struct bpf_reg_state *src_reg)
7839 u32 umax_val = src_reg->u32_max_value;
7840 u32 umin_val = src_reg->u32_min_value;
7841 /* u32 alu operation will zext upper bits */
7842 struct tnum subreg = tnum_subreg(dst_reg->var_off);
7844 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
7845 dst_reg->var_off = tnum_subreg(tnum_lshift(subreg, umin_val));
7846 /* Not required but being careful mark reg64 bounds as unknown so
7847 * that we are forced to pick them up from tnum and zext later and
7848 * if some path skips this step we are still safe.
7850 __mark_reg64_unbounded(dst_reg);
7851 __update_reg32_bounds(dst_reg);
7854 static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg,
7855 u64 umin_val, u64 umax_val)
7857 /* Special case <<32 because it is a common compiler pattern to sign
7858 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
7859 * positive we know this shift will also be positive so we can track
7860 * bounds correctly. Otherwise we lose all sign bit information except
7861 * what we can pick up from var_off. Perhaps we can generalize this
7862 * later to shifts of any length.
7864 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0)
7865 dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32;
7867 dst_reg->smax_value = S64_MAX;
7869 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0)
7870 dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32;
7872 dst_reg->smin_value = S64_MIN;
7874 /* If we might shift our top bit out, then we know nothing */
7875 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
7876 dst_reg->umin_value = 0;
7877 dst_reg->umax_value = U64_MAX;
7879 dst_reg->umin_value <<= umin_val;
7880 dst_reg->umax_value <<= umax_val;
7884 static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
7885 struct bpf_reg_state *src_reg)
7887 u64 umax_val = src_reg->umax_value;
7888 u64 umin_val = src_reg->umin_value;
7890 /* scalar64 calc uses 32bit unshifted bounds so must be called first */
7891 __scalar64_min_max_lsh(dst_reg, umin_val, umax_val);
7892 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
7894 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
7895 /* We may learn something more from the var_off */
7896 __update_reg_bounds(dst_reg);
7899 static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg,
7900 struct bpf_reg_state *src_reg)
7902 struct tnum subreg = tnum_subreg(dst_reg->var_off);
7903 u32 umax_val = src_reg->u32_max_value;
7904 u32 umin_val = src_reg->u32_min_value;
7906 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
7907 * be negative, then either:
7908 * 1) src_reg might be zero, so the sign bit of the result is
7909 * unknown, so we lose our signed bounds
7910 * 2) it's known negative, thus the unsigned bounds capture the
7912 * 3) the signed bounds cross zero, so they tell us nothing
7914 * If the value in dst_reg is known nonnegative, then again the
7915 * unsigned bounds capture the signed bounds.
7916 * Thus, in all cases it suffices to blow away our signed bounds
7917 * and rely on inferring new ones from the unsigned bounds and
7918 * var_off of the result.
7920 dst_reg->s32_min_value = S32_MIN;
7921 dst_reg->s32_max_value = S32_MAX;
7923 dst_reg->var_off = tnum_rshift(subreg, umin_val);
7924 dst_reg->u32_min_value >>= umax_val;
7925 dst_reg->u32_max_value >>= umin_val;
7927 __mark_reg64_unbounded(dst_reg);
7928 __update_reg32_bounds(dst_reg);
7931 static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
7932 struct bpf_reg_state *src_reg)
7934 u64 umax_val = src_reg->umax_value;
7935 u64 umin_val = src_reg->umin_value;
7937 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
7938 * be negative, then either:
7939 * 1) src_reg might be zero, so the sign bit of the result is
7940 * unknown, so we lose our signed bounds
7941 * 2) it's known negative, thus the unsigned bounds capture the
7943 * 3) the signed bounds cross zero, so they tell us nothing
7945 * If the value in dst_reg is known nonnegative, then again the
7946 * unsigned bounds capture the signed bounds.
7947 * Thus, in all cases it suffices to blow away our signed bounds
7948 * and rely on inferring new ones from the unsigned bounds and
7949 * var_off of the result.
7951 dst_reg->smin_value = S64_MIN;
7952 dst_reg->smax_value = S64_MAX;
7953 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
7954 dst_reg->umin_value >>= umax_val;
7955 dst_reg->umax_value >>= umin_val;
7957 /* Its not easy to operate on alu32 bounds here because it depends
7958 * on bits being shifted in. Take easy way out and mark unbounded
7959 * so we can recalculate later from tnum.
7961 __mark_reg32_unbounded(dst_reg);
7962 __update_reg_bounds(dst_reg);
7965 static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg,
7966 struct bpf_reg_state *src_reg)
7968 u64 umin_val = src_reg->u32_min_value;
7970 /* Upon reaching here, src_known is true and
7971 * umax_val is equal to umin_val.
7973 dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val);
7974 dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val);
7976 dst_reg->var_off = tnum_arshift(tnum_subreg(dst_reg->var_off), umin_val, 32);
7978 /* blow away the dst_reg umin_value/umax_value and rely on
7979 * dst_reg var_off to refine the result.
7981 dst_reg->u32_min_value = 0;
7982 dst_reg->u32_max_value = U32_MAX;
7984 __mark_reg64_unbounded(dst_reg);
7985 __update_reg32_bounds(dst_reg);
7988 static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
7989 struct bpf_reg_state *src_reg)
7991 u64 umin_val = src_reg->umin_value;
7993 /* Upon reaching here, src_known is true and umax_val is equal
7996 dst_reg->smin_value >>= umin_val;
7997 dst_reg->smax_value >>= umin_val;
7999 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val, 64);
8001 /* blow away the dst_reg umin_value/umax_value and rely on
8002 * dst_reg var_off to refine the result.
8004 dst_reg->umin_value = 0;
8005 dst_reg->umax_value = U64_MAX;
8007 /* Its not easy to operate on alu32 bounds here because it depends
8008 * on bits being shifted in from upper 32-bits. Take easy way out
8009 * and mark unbounded so we can recalculate later from tnum.
8011 __mark_reg32_unbounded(dst_reg);
8012 __update_reg_bounds(dst_reg);
8015 /* WARNING: This function does calculations on 64-bit values, but the actual
8016 * execution may occur on 32-bit values. Therefore, things like bitshifts
8017 * need extra checks in the 32-bit case.
8019 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
8020 struct bpf_insn *insn,
8021 struct bpf_reg_state *dst_reg,
8022 struct bpf_reg_state src_reg)
8024 struct bpf_reg_state *regs = cur_regs(env);
8025 u8 opcode = BPF_OP(insn->code);
8027 s64 smin_val, smax_val;
8028 u64 umin_val, umax_val;
8029 s32 s32_min_val, s32_max_val;
8030 u32 u32_min_val, u32_max_val;
8031 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
8032 bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
8035 smin_val = src_reg.smin_value;
8036 smax_val = src_reg.smax_value;
8037 umin_val = src_reg.umin_value;
8038 umax_val = src_reg.umax_value;
8040 s32_min_val = src_reg.s32_min_value;
8041 s32_max_val = src_reg.s32_max_value;
8042 u32_min_val = src_reg.u32_min_value;
8043 u32_max_val = src_reg.u32_max_value;
8046 src_known = tnum_subreg_is_const(src_reg.var_off);
8048 (s32_min_val != s32_max_val || u32_min_val != u32_max_val)) ||
8049 s32_min_val > s32_max_val || u32_min_val > u32_max_val) {
8050 /* Taint dst register if offset had invalid bounds
8051 * derived from e.g. dead branches.
8053 __mark_reg_unknown(env, dst_reg);
8057 src_known = tnum_is_const(src_reg.var_off);
8059 (smin_val != smax_val || umin_val != umax_val)) ||
8060 smin_val > smax_val || umin_val > umax_val) {
8061 /* Taint dst register if offset had invalid bounds
8062 * derived from e.g. dead branches.
8064 __mark_reg_unknown(env, dst_reg);
8070 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
8071 __mark_reg_unknown(env, dst_reg);
8075 if (sanitize_needed(opcode)) {
8076 ret = sanitize_val_alu(env, insn);
8078 return sanitize_err(env, insn, ret, NULL, NULL);
8081 /* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
8082 * There are two classes of instructions: The first class we track both
8083 * alu32 and alu64 sign/unsigned bounds independently this provides the
8084 * greatest amount of precision when alu operations are mixed with jmp32
8085 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
8086 * and BPF_OR. This is possible because these ops have fairly easy to
8087 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
8088 * See alu32 verifier tests for examples. The second class of
8089 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
8090 * with regards to tracking sign/unsigned bounds because the bits may
8091 * cross subreg boundaries in the alu64 case. When this happens we mark
8092 * the reg unbounded in the subreg bound space and use the resulting
8093 * tnum to calculate an approximation of the sign/unsigned bounds.
8097 scalar32_min_max_add(dst_reg, &src_reg);
8098 scalar_min_max_add(dst_reg, &src_reg);
8099 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
8102 scalar32_min_max_sub(dst_reg, &src_reg);
8103 scalar_min_max_sub(dst_reg, &src_reg);
8104 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
8107 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
8108 scalar32_min_max_mul(dst_reg, &src_reg);
8109 scalar_min_max_mul(dst_reg, &src_reg);
8112 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
8113 scalar32_min_max_and(dst_reg, &src_reg);
8114 scalar_min_max_and(dst_reg, &src_reg);
8117 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
8118 scalar32_min_max_or(dst_reg, &src_reg);
8119 scalar_min_max_or(dst_reg, &src_reg);
8122 dst_reg->var_off = tnum_xor(dst_reg->var_off, src_reg.var_off);
8123 scalar32_min_max_xor(dst_reg, &src_reg);
8124 scalar_min_max_xor(dst_reg, &src_reg);
8127 if (umax_val >= insn_bitness) {
8128 /* Shifts greater than 31 or 63 are undefined.
8129 * This includes shifts by a negative number.
8131 mark_reg_unknown(env, regs, insn->dst_reg);
8135 scalar32_min_max_lsh(dst_reg, &src_reg);
8137 scalar_min_max_lsh(dst_reg, &src_reg);
8140 if (umax_val >= insn_bitness) {
8141 /* Shifts greater than 31 or 63 are undefined.
8142 * This includes shifts by a negative number.
8144 mark_reg_unknown(env, regs, insn->dst_reg);
8148 scalar32_min_max_rsh(dst_reg, &src_reg);
8150 scalar_min_max_rsh(dst_reg, &src_reg);
8153 if (umax_val >= insn_bitness) {
8154 /* Shifts greater than 31 or 63 are undefined.
8155 * This includes shifts by a negative number.
8157 mark_reg_unknown(env, regs, insn->dst_reg);
8161 scalar32_min_max_arsh(dst_reg, &src_reg);
8163 scalar_min_max_arsh(dst_reg, &src_reg);
8166 mark_reg_unknown(env, regs, insn->dst_reg);
8170 /* ALU32 ops are zero extended into 64bit register */
8172 zext_32_to_64(dst_reg);
8174 __update_reg_bounds(dst_reg);
8175 __reg_deduce_bounds(dst_reg);
8176 __reg_bound_offset(dst_reg);
8180 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
8183 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
8184 struct bpf_insn *insn)
8186 struct bpf_verifier_state *vstate = env->cur_state;
8187 struct bpf_func_state *state = vstate->frame[vstate->curframe];
8188 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
8189 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
8190 u8 opcode = BPF_OP(insn->code);
8193 dst_reg = ®s[insn->dst_reg];
8195 if (dst_reg->type != SCALAR_VALUE)
8198 /* Make sure ID is cleared otherwise dst_reg min/max could be
8199 * incorrectly propagated into other registers by find_equal_scalars()
8202 if (BPF_SRC(insn->code) == BPF_X) {
8203 src_reg = ®s[insn->src_reg];
8204 if (src_reg->type != SCALAR_VALUE) {
8205 if (dst_reg->type != SCALAR_VALUE) {
8206 /* Combining two pointers by any ALU op yields
8207 * an arbitrary scalar. Disallow all math except
8208 * pointer subtraction
8210 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
8211 mark_reg_unknown(env, regs, insn->dst_reg);
8214 verbose(env, "R%d pointer %s pointer prohibited\n",
8216 bpf_alu_string[opcode >> 4]);
8219 /* scalar += pointer
8220 * This is legal, but we have to reverse our
8221 * src/dest handling in computing the range
8223 err = mark_chain_precision(env, insn->dst_reg);
8226 return adjust_ptr_min_max_vals(env, insn,
8229 } else if (ptr_reg) {
8230 /* pointer += scalar */
8231 err = mark_chain_precision(env, insn->src_reg);
8234 return adjust_ptr_min_max_vals(env, insn,
8238 /* Pretend the src is a reg with a known value, since we only
8239 * need to be able to read from this state.
8241 off_reg.type = SCALAR_VALUE;
8242 __mark_reg_known(&off_reg, insn->imm);
8244 if (ptr_reg) /* pointer += K */
8245 return adjust_ptr_min_max_vals(env, insn,
8249 /* Got here implies adding two SCALAR_VALUEs */
8250 if (WARN_ON_ONCE(ptr_reg)) {
8251 print_verifier_state(env, state);
8252 verbose(env, "verifier internal error: unexpected ptr_reg\n");
8255 if (WARN_ON(!src_reg)) {
8256 print_verifier_state(env, state);
8257 verbose(env, "verifier internal error: no src_reg\n");
8260 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
8263 /* check validity of 32-bit and 64-bit arithmetic operations */
8264 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
8266 struct bpf_reg_state *regs = cur_regs(env);
8267 u8 opcode = BPF_OP(insn->code);
8270 if (opcode == BPF_END || opcode == BPF_NEG) {
8271 if (opcode == BPF_NEG) {
8272 if (BPF_SRC(insn->code) != 0 ||
8273 insn->src_reg != BPF_REG_0 ||
8274 insn->off != 0 || insn->imm != 0) {
8275 verbose(env, "BPF_NEG uses reserved fields\n");
8279 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
8280 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
8281 BPF_CLASS(insn->code) == BPF_ALU64) {
8282 verbose(env, "BPF_END uses reserved fields\n");
8287 /* check src operand */
8288 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
8292 if (is_pointer_value(env, insn->dst_reg)) {
8293 verbose(env, "R%d pointer arithmetic prohibited\n",
8298 /* check dest operand */
8299 err = check_reg_arg(env, insn->dst_reg, DST_OP);
8303 } else if (opcode == BPF_MOV) {
8305 if (BPF_SRC(insn->code) == BPF_X) {
8306 if (insn->imm != 0 || insn->off != 0) {
8307 verbose(env, "BPF_MOV uses reserved fields\n");
8311 /* check src operand */
8312 err = check_reg_arg(env, insn->src_reg, SRC_OP);
8316 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
8317 verbose(env, "BPF_MOV uses reserved fields\n");
8322 /* check dest operand, mark as required later */
8323 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
8327 if (BPF_SRC(insn->code) == BPF_X) {
8328 struct bpf_reg_state *src_reg = regs + insn->src_reg;
8329 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
8331 if (BPF_CLASS(insn->code) == BPF_ALU64) {
8333 * copy register state to dest reg
8335 if (src_reg->type == SCALAR_VALUE && !src_reg->id)
8336 /* Assign src and dst registers the same ID
8337 * that will be used by find_equal_scalars()
8338 * to propagate min/max range.
8340 src_reg->id = ++env->id_gen;
8341 *dst_reg = *src_reg;
8342 dst_reg->live |= REG_LIVE_WRITTEN;
8343 dst_reg->subreg_def = DEF_NOT_SUBREG;
8346 if (is_pointer_value(env, insn->src_reg)) {
8348 "R%d partial copy of pointer\n",
8351 } else if (src_reg->type == SCALAR_VALUE) {
8352 *dst_reg = *src_reg;
8353 /* Make sure ID is cleared otherwise
8354 * dst_reg min/max could be incorrectly
8355 * propagated into src_reg by find_equal_scalars()
8358 dst_reg->live |= REG_LIVE_WRITTEN;
8359 dst_reg->subreg_def = env->insn_idx + 1;
8361 mark_reg_unknown(env, regs,
8364 zext_32_to_64(dst_reg);
8368 * remember the value we stored into this reg
8370 /* clear any state __mark_reg_known doesn't set */
8371 mark_reg_unknown(env, regs, insn->dst_reg);
8372 regs[insn->dst_reg].type = SCALAR_VALUE;
8373 if (BPF_CLASS(insn->code) == BPF_ALU64) {
8374 __mark_reg_known(regs + insn->dst_reg,
8377 __mark_reg_known(regs + insn->dst_reg,
8382 } else if (opcode > BPF_END) {
8383 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
8386 } else { /* all other ALU ops: and, sub, xor, add, ... */
8388 if (BPF_SRC(insn->code) == BPF_X) {
8389 if (insn->imm != 0 || insn->off != 0) {
8390 verbose(env, "BPF_ALU uses reserved fields\n");
8393 /* check src1 operand */
8394 err = check_reg_arg(env, insn->src_reg, SRC_OP);
8398 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
8399 verbose(env, "BPF_ALU uses reserved fields\n");
8404 /* check src2 operand */
8405 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
8409 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
8410 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
8411 verbose(env, "div by zero\n");
8415 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
8416 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
8417 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
8419 if (insn->imm < 0 || insn->imm >= size) {
8420 verbose(env, "invalid shift %d\n", insn->imm);
8425 /* check dest operand */
8426 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
8430 return adjust_reg_min_max_vals(env, insn);
8436 static void __find_good_pkt_pointers(struct bpf_func_state *state,
8437 struct bpf_reg_state *dst_reg,
8438 enum bpf_reg_type type, int new_range)
8440 struct bpf_reg_state *reg;
8443 for (i = 0; i < MAX_BPF_REG; i++) {
8444 reg = &state->regs[i];
8445 if (reg->type == type && reg->id == dst_reg->id)
8446 /* keep the maximum range already checked */
8447 reg->range = max(reg->range, new_range);
8450 bpf_for_each_spilled_reg(i, state, reg) {
8453 if (reg->type == type && reg->id == dst_reg->id)
8454 reg->range = max(reg->range, new_range);
8458 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
8459 struct bpf_reg_state *dst_reg,
8460 enum bpf_reg_type type,
8461 bool range_right_open)
8465 if (dst_reg->off < 0 ||
8466 (dst_reg->off == 0 && range_right_open))
8467 /* This doesn't give us any range */
8470 if (dst_reg->umax_value > MAX_PACKET_OFF ||
8471 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
8472 /* Risk of overflow. For instance, ptr + (1<<63) may be less
8473 * than pkt_end, but that's because it's also less than pkt.
8477 new_range = dst_reg->off;
8478 if (range_right_open)
8481 /* Examples for register markings:
8483 * pkt_data in dst register:
8487 * if (r2 > pkt_end) goto <handle exception>
8492 * if (r2 < pkt_end) goto <access okay>
8493 * <handle exception>
8496 * r2 == dst_reg, pkt_end == src_reg
8497 * r2=pkt(id=n,off=8,r=0)
8498 * r3=pkt(id=n,off=0,r=0)
8500 * pkt_data in src register:
8504 * if (pkt_end >= r2) goto <access okay>
8505 * <handle exception>
8509 * if (pkt_end <= r2) goto <handle exception>
8513 * pkt_end == dst_reg, r2 == src_reg
8514 * r2=pkt(id=n,off=8,r=0)
8515 * r3=pkt(id=n,off=0,r=0)
8517 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
8518 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
8519 * and [r3, r3 + 8-1) respectively is safe to access depending on
8523 /* If our ids match, then we must have the same max_value. And we
8524 * don't care about the other reg's fixed offset, since if it's too big
8525 * the range won't allow anything.
8526 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
8528 for (i = 0; i <= vstate->curframe; i++)
8529 __find_good_pkt_pointers(vstate->frame[i], dst_reg, type,
8533 static int is_branch32_taken(struct bpf_reg_state *reg, u32 val, u8 opcode)
8535 struct tnum subreg = tnum_subreg(reg->var_off);
8536 s32 sval = (s32)val;
8540 if (tnum_is_const(subreg))
8541 return !!tnum_equals_const(subreg, val);
8544 if (tnum_is_const(subreg))
8545 return !tnum_equals_const(subreg, val);
8548 if ((~subreg.mask & subreg.value) & val)
8550 if (!((subreg.mask | subreg.value) & val))
8554 if (reg->u32_min_value > val)
8556 else if (reg->u32_max_value <= val)
8560 if (reg->s32_min_value > sval)
8562 else if (reg->s32_max_value <= sval)
8566 if (reg->u32_max_value < val)
8568 else if (reg->u32_min_value >= val)
8572 if (reg->s32_max_value < sval)
8574 else if (reg->s32_min_value >= sval)
8578 if (reg->u32_min_value >= val)
8580 else if (reg->u32_max_value < val)
8584 if (reg->s32_min_value >= sval)
8586 else if (reg->s32_max_value < sval)
8590 if (reg->u32_max_value <= val)
8592 else if (reg->u32_min_value > val)
8596 if (reg->s32_max_value <= sval)
8598 else if (reg->s32_min_value > sval)
8607 static int is_branch64_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
8609 s64 sval = (s64)val;
8613 if (tnum_is_const(reg->var_off))
8614 return !!tnum_equals_const(reg->var_off, val);
8617 if (tnum_is_const(reg->var_off))
8618 return !tnum_equals_const(reg->var_off, val);
8621 if ((~reg->var_off.mask & reg->var_off.value) & val)
8623 if (!((reg->var_off.mask | reg->var_off.value) & val))
8627 if (reg->umin_value > val)
8629 else if (reg->umax_value <= val)
8633 if (reg->smin_value > sval)
8635 else if (reg->smax_value <= sval)
8639 if (reg->umax_value < val)
8641 else if (reg->umin_value >= val)
8645 if (reg->smax_value < sval)
8647 else if (reg->smin_value >= sval)
8651 if (reg->umin_value >= val)
8653 else if (reg->umax_value < val)
8657 if (reg->smin_value >= sval)
8659 else if (reg->smax_value < sval)
8663 if (reg->umax_value <= val)
8665 else if (reg->umin_value > val)
8669 if (reg->smax_value <= sval)
8671 else if (reg->smin_value > sval)
8679 /* compute branch direction of the expression "if (reg opcode val) goto target;"
8681 * 1 - branch will be taken and "goto target" will be executed
8682 * 0 - branch will not be taken and fall-through to next insn
8683 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value
8686 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
8689 if (__is_pointer_value(false, reg)) {
8690 if (!reg_type_not_null(reg->type))
8693 /* If pointer is valid tests against zero will fail so we can
8694 * use this to direct branch taken.
8710 return is_branch32_taken(reg, val, opcode);
8711 return is_branch64_taken(reg, val, opcode);
8714 static int flip_opcode(u32 opcode)
8716 /* How can we transform "a <op> b" into "b <op> a"? */
8717 static const u8 opcode_flip[16] = {
8718 /* these stay the same */
8719 [BPF_JEQ >> 4] = BPF_JEQ,
8720 [BPF_JNE >> 4] = BPF_JNE,
8721 [BPF_JSET >> 4] = BPF_JSET,
8722 /* these swap "lesser" and "greater" (L and G in the opcodes) */
8723 [BPF_JGE >> 4] = BPF_JLE,
8724 [BPF_JGT >> 4] = BPF_JLT,
8725 [BPF_JLE >> 4] = BPF_JGE,
8726 [BPF_JLT >> 4] = BPF_JGT,
8727 [BPF_JSGE >> 4] = BPF_JSLE,
8728 [BPF_JSGT >> 4] = BPF_JSLT,
8729 [BPF_JSLE >> 4] = BPF_JSGE,
8730 [BPF_JSLT >> 4] = BPF_JSGT
8732 return opcode_flip[opcode >> 4];
8735 static int is_pkt_ptr_branch_taken(struct bpf_reg_state *dst_reg,
8736 struct bpf_reg_state *src_reg,
8739 struct bpf_reg_state *pkt;
8741 if (src_reg->type == PTR_TO_PACKET_END) {
8743 } else if (dst_reg->type == PTR_TO_PACKET_END) {
8745 opcode = flip_opcode(opcode);
8750 if (pkt->range >= 0)
8755 /* pkt <= pkt_end */
8759 if (pkt->range == BEYOND_PKT_END)
8760 /* pkt has at last one extra byte beyond pkt_end */
8761 return opcode == BPF_JGT;
8767 /* pkt >= pkt_end */
8768 if (pkt->range == BEYOND_PKT_END || pkt->range == AT_PKT_END)
8769 return opcode == BPF_JGE;
8775 /* Adjusts the register min/max values in the case that the dst_reg is the
8776 * variable register that we are working on, and src_reg is a constant or we're
8777 * simply doing a BPF_K check.
8778 * In JEQ/JNE cases we also adjust the var_off values.
8780 static void reg_set_min_max(struct bpf_reg_state *true_reg,
8781 struct bpf_reg_state *false_reg,
8783 u8 opcode, bool is_jmp32)
8785 struct tnum false_32off = tnum_subreg(false_reg->var_off);
8786 struct tnum false_64off = false_reg->var_off;
8787 struct tnum true_32off = tnum_subreg(true_reg->var_off);
8788 struct tnum true_64off = true_reg->var_off;
8789 s64 sval = (s64)val;
8790 s32 sval32 = (s32)val32;
8792 /* If the dst_reg is a pointer, we can't learn anything about its
8793 * variable offset from the compare (unless src_reg were a pointer into
8794 * the same object, but we don't bother with that.
8795 * Since false_reg and true_reg have the same type by construction, we
8796 * only need to check one of them for pointerness.
8798 if (__is_pointer_value(false, false_reg))
8805 struct bpf_reg_state *reg =
8806 opcode == BPF_JEQ ? true_reg : false_reg;
8808 /* JEQ/JNE comparison doesn't change the register equivalence.
8810 * if (r1 == 42) goto label;
8812 * label: // here both r1 and r2 are known to be 42.
8814 * Hence when marking register as known preserve it's ID.
8817 __mark_reg32_known(reg, val32);
8819 ___mark_reg_known(reg, val);
8824 false_32off = tnum_and(false_32off, tnum_const(~val32));
8825 if (is_power_of_2(val32))
8826 true_32off = tnum_or(true_32off,
8829 false_64off = tnum_and(false_64off, tnum_const(~val));
8830 if (is_power_of_2(val))
8831 true_64off = tnum_or(true_64off,
8839 u32 false_umax = opcode == BPF_JGT ? val32 : val32 - 1;
8840 u32 true_umin = opcode == BPF_JGT ? val32 + 1 : val32;
8842 false_reg->u32_max_value = min(false_reg->u32_max_value,
8844 true_reg->u32_min_value = max(true_reg->u32_min_value,
8847 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
8848 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
8850 false_reg->umax_value = min(false_reg->umax_value, false_umax);
8851 true_reg->umin_value = max(true_reg->umin_value, true_umin);
8859 s32 false_smax = opcode == BPF_JSGT ? sval32 : sval32 - 1;
8860 s32 true_smin = opcode == BPF_JSGT ? sval32 + 1 : sval32;
8862 false_reg->s32_max_value = min(false_reg->s32_max_value, false_smax);
8863 true_reg->s32_min_value = max(true_reg->s32_min_value, true_smin);
8865 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
8866 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
8868 false_reg->smax_value = min(false_reg->smax_value, false_smax);
8869 true_reg->smin_value = max(true_reg->smin_value, true_smin);
8877 u32 false_umin = opcode == BPF_JLT ? val32 : val32 + 1;
8878 u32 true_umax = opcode == BPF_JLT ? val32 - 1 : val32;
8880 false_reg->u32_min_value = max(false_reg->u32_min_value,
8882 true_reg->u32_max_value = min(true_reg->u32_max_value,
8885 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
8886 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
8888 false_reg->umin_value = max(false_reg->umin_value, false_umin);
8889 true_reg->umax_value = min(true_reg->umax_value, true_umax);
8897 s32 false_smin = opcode == BPF_JSLT ? sval32 : sval32 + 1;
8898 s32 true_smax = opcode == BPF_JSLT ? sval32 - 1 : sval32;
8900 false_reg->s32_min_value = max(false_reg->s32_min_value, false_smin);
8901 true_reg->s32_max_value = min(true_reg->s32_max_value, true_smax);
8903 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
8904 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
8906 false_reg->smin_value = max(false_reg->smin_value, false_smin);
8907 true_reg->smax_value = min(true_reg->smax_value, true_smax);
8916 false_reg->var_off = tnum_or(tnum_clear_subreg(false_64off),
8917 tnum_subreg(false_32off));
8918 true_reg->var_off = tnum_or(tnum_clear_subreg(true_64off),
8919 tnum_subreg(true_32off));
8920 __reg_combine_32_into_64(false_reg);
8921 __reg_combine_32_into_64(true_reg);
8923 false_reg->var_off = false_64off;
8924 true_reg->var_off = true_64off;
8925 __reg_combine_64_into_32(false_reg);
8926 __reg_combine_64_into_32(true_reg);
8930 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
8933 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
8934 struct bpf_reg_state *false_reg,
8936 u8 opcode, bool is_jmp32)
8938 opcode = flip_opcode(opcode);
8939 /* This uses zero as "not present in table"; luckily the zero opcode,
8940 * BPF_JA, can't get here.
8943 reg_set_min_max(true_reg, false_reg, val, val32, opcode, is_jmp32);
8946 /* Regs are known to be equal, so intersect their min/max/var_off */
8947 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
8948 struct bpf_reg_state *dst_reg)
8950 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
8951 dst_reg->umin_value);
8952 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
8953 dst_reg->umax_value);
8954 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
8955 dst_reg->smin_value);
8956 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
8957 dst_reg->smax_value);
8958 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
8960 /* We might have learned new bounds from the var_off. */
8961 __update_reg_bounds(src_reg);
8962 __update_reg_bounds(dst_reg);
8963 /* We might have learned something about the sign bit. */
8964 __reg_deduce_bounds(src_reg);
8965 __reg_deduce_bounds(dst_reg);
8966 /* We might have learned some bits from the bounds. */
8967 __reg_bound_offset(src_reg);
8968 __reg_bound_offset(dst_reg);
8969 /* Intersecting with the old var_off might have improved our bounds
8970 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
8971 * then new var_off is (0; 0x7f...fc) which improves our umax.
8973 __update_reg_bounds(src_reg);
8974 __update_reg_bounds(dst_reg);
8977 static void reg_combine_min_max(struct bpf_reg_state *true_src,
8978 struct bpf_reg_state *true_dst,
8979 struct bpf_reg_state *false_src,
8980 struct bpf_reg_state *false_dst,
8985 __reg_combine_min_max(true_src, true_dst);
8988 __reg_combine_min_max(false_src, false_dst);
8993 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
8994 struct bpf_reg_state *reg, u32 id,
8997 if (reg_type_may_be_null(reg->type) && reg->id == id &&
8998 !WARN_ON_ONCE(!reg->id)) {
8999 /* Old offset (both fixed and variable parts) should
9000 * have been known-zero, because we don't allow pointer
9001 * arithmetic on pointers that might be NULL.
9003 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
9004 !tnum_equals_const(reg->var_off, 0) ||
9006 __mark_reg_known_zero(reg);
9010 reg->type = SCALAR_VALUE;
9011 /* We don't need id and ref_obj_id from this point
9012 * onwards anymore, thus we should better reset it,
9013 * so that state pruning has chances to take effect.
9016 reg->ref_obj_id = 0;
9021 mark_ptr_not_null_reg(reg);
9023 if (!reg_may_point_to_spin_lock(reg)) {
9024 /* For not-NULL ptr, reg->ref_obj_id will be reset
9025 * in release_reg_references().
9027 * reg->id is still used by spin_lock ptr. Other
9028 * than spin_lock ptr type, reg->id can be reset.
9035 static void __mark_ptr_or_null_regs(struct bpf_func_state *state, u32 id,
9038 struct bpf_reg_state *reg;
9041 for (i = 0; i < MAX_BPF_REG; i++)
9042 mark_ptr_or_null_reg(state, &state->regs[i], id, is_null);
9044 bpf_for_each_spilled_reg(i, state, reg) {
9047 mark_ptr_or_null_reg(state, reg, id, is_null);
9051 /* The logic is similar to find_good_pkt_pointers(), both could eventually
9052 * be folded together at some point.
9054 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
9057 struct bpf_func_state *state = vstate->frame[vstate->curframe];
9058 struct bpf_reg_state *regs = state->regs;
9059 u32 ref_obj_id = regs[regno].ref_obj_id;
9060 u32 id = regs[regno].id;
9063 if (ref_obj_id && ref_obj_id == id && is_null)
9064 /* regs[regno] is in the " == NULL" branch.
9065 * No one could have freed the reference state before
9066 * doing the NULL check.
9068 WARN_ON_ONCE(release_reference_state(state, id));
9070 for (i = 0; i <= vstate->curframe; i++)
9071 __mark_ptr_or_null_regs(vstate->frame[i], id, is_null);
9074 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
9075 struct bpf_reg_state *dst_reg,
9076 struct bpf_reg_state *src_reg,
9077 struct bpf_verifier_state *this_branch,
9078 struct bpf_verifier_state *other_branch)
9080 if (BPF_SRC(insn->code) != BPF_X)
9083 /* Pointers are always 64-bit. */
9084 if (BPF_CLASS(insn->code) == BPF_JMP32)
9087 switch (BPF_OP(insn->code)) {
9089 if ((dst_reg->type == PTR_TO_PACKET &&
9090 src_reg->type == PTR_TO_PACKET_END) ||
9091 (dst_reg->type == PTR_TO_PACKET_META &&
9092 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
9093 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
9094 find_good_pkt_pointers(this_branch, dst_reg,
9095 dst_reg->type, false);
9096 mark_pkt_end(other_branch, insn->dst_reg, true);
9097 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
9098 src_reg->type == PTR_TO_PACKET) ||
9099 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
9100 src_reg->type == PTR_TO_PACKET_META)) {
9101 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
9102 find_good_pkt_pointers(other_branch, src_reg,
9103 src_reg->type, true);
9104 mark_pkt_end(this_branch, insn->src_reg, false);
9110 if ((dst_reg->type == PTR_TO_PACKET &&
9111 src_reg->type == PTR_TO_PACKET_END) ||
9112 (dst_reg->type == PTR_TO_PACKET_META &&
9113 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
9114 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
9115 find_good_pkt_pointers(other_branch, dst_reg,
9116 dst_reg->type, true);
9117 mark_pkt_end(this_branch, insn->dst_reg, false);
9118 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
9119 src_reg->type == PTR_TO_PACKET) ||
9120 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
9121 src_reg->type == PTR_TO_PACKET_META)) {
9122 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
9123 find_good_pkt_pointers(this_branch, src_reg,
9124 src_reg->type, false);
9125 mark_pkt_end(other_branch, insn->src_reg, true);
9131 if ((dst_reg->type == PTR_TO_PACKET &&
9132 src_reg->type == PTR_TO_PACKET_END) ||
9133 (dst_reg->type == PTR_TO_PACKET_META &&
9134 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
9135 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
9136 find_good_pkt_pointers(this_branch, dst_reg,
9137 dst_reg->type, true);
9138 mark_pkt_end(other_branch, insn->dst_reg, false);
9139 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
9140 src_reg->type == PTR_TO_PACKET) ||
9141 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
9142 src_reg->type == PTR_TO_PACKET_META)) {
9143 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
9144 find_good_pkt_pointers(other_branch, src_reg,
9145 src_reg->type, false);
9146 mark_pkt_end(this_branch, insn->src_reg, true);
9152 if ((dst_reg->type == PTR_TO_PACKET &&
9153 src_reg->type == PTR_TO_PACKET_END) ||
9154 (dst_reg->type == PTR_TO_PACKET_META &&
9155 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
9156 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
9157 find_good_pkt_pointers(other_branch, dst_reg,
9158 dst_reg->type, false);
9159 mark_pkt_end(this_branch, insn->dst_reg, true);
9160 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
9161 src_reg->type == PTR_TO_PACKET) ||
9162 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
9163 src_reg->type == PTR_TO_PACKET_META)) {
9164 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
9165 find_good_pkt_pointers(this_branch, src_reg,
9166 src_reg->type, true);
9167 mark_pkt_end(other_branch, insn->src_reg, false);
9179 static void find_equal_scalars(struct bpf_verifier_state *vstate,
9180 struct bpf_reg_state *known_reg)
9182 struct bpf_func_state *state;
9183 struct bpf_reg_state *reg;
9186 for (i = 0; i <= vstate->curframe; i++) {
9187 state = vstate->frame[i];
9188 for (j = 0; j < MAX_BPF_REG; j++) {
9189 reg = &state->regs[j];
9190 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
9194 bpf_for_each_spilled_reg(j, state, reg) {
9197 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
9203 static int check_cond_jmp_op(struct bpf_verifier_env *env,
9204 struct bpf_insn *insn, int *insn_idx)
9206 struct bpf_verifier_state *this_branch = env->cur_state;
9207 struct bpf_verifier_state *other_branch;
9208 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
9209 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
9210 u8 opcode = BPF_OP(insn->code);
9215 /* Only conditional jumps are expected to reach here. */
9216 if (opcode == BPF_JA || opcode > BPF_JSLE) {
9217 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
9221 if (BPF_SRC(insn->code) == BPF_X) {
9222 if (insn->imm != 0) {
9223 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
9227 /* check src1 operand */
9228 err = check_reg_arg(env, insn->src_reg, SRC_OP);
9232 if (is_pointer_value(env, insn->src_reg)) {
9233 verbose(env, "R%d pointer comparison prohibited\n",
9237 src_reg = ®s[insn->src_reg];
9239 if (insn->src_reg != BPF_REG_0) {
9240 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
9245 /* check src2 operand */
9246 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
9250 dst_reg = ®s[insn->dst_reg];
9251 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
9253 if (BPF_SRC(insn->code) == BPF_K) {
9254 pred = is_branch_taken(dst_reg, insn->imm, opcode, is_jmp32);
9255 } else if (src_reg->type == SCALAR_VALUE &&
9256 is_jmp32 && tnum_is_const(tnum_subreg(src_reg->var_off))) {
9257 pred = is_branch_taken(dst_reg,
9258 tnum_subreg(src_reg->var_off).value,
9261 } else if (src_reg->type == SCALAR_VALUE &&
9262 !is_jmp32 && tnum_is_const(src_reg->var_off)) {
9263 pred = is_branch_taken(dst_reg,
9264 src_reg->var_off.value,
9267 } else if (reg_is_pkt_pointer_any(dst_reg) &&
9268 reg_is_pkt_pointer_any(src_reg) &&
9270 pred = is_pkt_ptr_branch_taken(dst_reg, src_reg, opcode);
9274 /* If we get here with a dst_reg pointer type it is because
9275 * above is_branch_taken() special cased the 0 comparison.
9277 if (!__is_pointer_value(false, dst_reg))
9278 err = mark_chain_precision(env, insn->dst_reg);
9279 if (BPF_SRC(insn->code) == BPF_X && !err &&
9280 !__is_pointer_value(false, src_reg))
9281 err = mark_chain_precision(env, insn->src_reg);
9287 /* Only follow the goto, ignore fall-through. If needed, push
9288 * the fall-through branch for simulation under speculative
9291 if (!env->bypass_spec_v1 &&
9292 !sanitize_speculative_path(env, insn, *insn_idx + 1,
9295 *insn_idx += insn->off;
9297 } else if (pred == 0) {
9298 /* Only follow the fall-through branch, since that's where the
9299 * program will go. If needed, push the goto branch for
9300 * simulation under speculative execution.
9302 if (!env->bypass_spec_v1 &&
9303 !sanitize_speculative_path(env, insn,
9304 *insn_idx + insn->off + 1,
9310 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
9314 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
9316 /* detect if we are comparing against a constant value so we can adjust
9317 * our min/max values for our dst register.
9318 * this is only legit if both are scalars (or pointers to the same
9319 * object, I suppose, but we don't support that right now), because
9320 * otherwise the different base pointers mean the offsets aren't
9323 if (BPF_SRC(insn->code) == BPF_X) {
9324 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
9326 if (dst_reg->type == SCALAR_VALUE &&
9327 src_reg->type == SCALAR_VALUE) {
9328 if (tnum_is_const(src_reg->var_off) ||
9330 tnum_is_const(tnum_subreg(src_reg->var_off))))
9331 reg_set_min_max(&other_branch_regs[insn->dst_reg],
9333 src_reg->var_off.value,
9334 tnum_subreg(src_reg->var_off).value,
9336 else if (tnum_is_const(dst_reg->var_off) ||
9338 tnum_is_const(tnum_subreg(dst_reg->var_off))))
9339 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
9341 dst_reg->var_off.value,
9342 tnum_subreg(dst_reg->var_off).value,
9344 else if (!is_jmp32 &&
9345 (opcode == BPF_JEQ || opcode == BPF_JNE))
9346 /* Comparing for equality, we can combine knowledge */
9347 reg_combine_min_max(&other_branch_regs[insn->src_reg],
9348 &other_branch_regs[insn->dst_reg],
9349 src_reg, dst_reg, opcode);
9351 !WARN_ON_ONCE(src_reg->id != other_branch_regs[insn->src_reg].id)) {
9352 find_equal_scalars(this_branch, src_reg);
9353 find_equal_scalars(other_branch, &other_branch_regs[insn->src_reg]);
9357 } else if (dst_reg->type == SCALAR_VALUE) {
9358 reg_set_min_max(&other_branch_regs[insn->dst_reg],
9359 dst_reg, insn->imm, (u32)insn->imm,
9363 if (dst_reg->type == SCALAR_VALUE && dst_reg->id &&
9364 !WARN_ON_ONCE(dst_reg->id != other_branch_regs[insn->dst_reg].id)) {
9365 find_equal_scalars(this_branch, dst_reg);
9366 find_equal_scalars(other_branch, &other_branch_regs[insn->dst_reg]);
9369 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
9370 * NOTE: these optimizations below are related with pointer comparison
9371 * which will never be JMP32.
9373 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
9374 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
9375 reg_type_may_be_null(dst_reg->type)) {
9376 /* Mark all identical registers in each branch as either
9377 * safe or unknown depending R == 0 or R != 0 conditional.
9379 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
9381 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
9383 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
9384 this_branch, other_branch) &&
9385 is_pointer_value(env, insn->dst_reg)) {
9386 verbose(env, "R%d pointer comparison prohibited\n",
9390 if (env->log.level & BPF_LOG_LEVEL)
9391 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
9395 /* verify BPF_LD_IMM64 instruction */
9396 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
9398 struct bpf_insn_aux_data *aux = cur_aux(env);
9399 struct bpf_reg_state *regs = cur_regs(env);
9400 struct bpf_reg_state *dst_reg;
9401 struct bpf_map *map;
9404 if (BPF_SIZE(insn->code) != BPF_DW) {
9405 verbose(env, "invalid BPF_LD_IMM insn\n");
9408 if (insn->off != 0) {
9409 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
9413 err = check_reg_arg(env, insn->dst_reg, DST_OP);
9417 dst_reg = ®s[insn->dst_reg];
9418 if (insn->src_reg == 0) {
9419 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
9421 dst_reg->type = SCALAR_VALUE;
9422 __mark_reg_known(®s[insn->dst_reg], imm);
9426 if (insn->src_reg == BPF_PSEUDO_BTF_ID) {
9427 mark_reg_known_zero(env, regs, insn->dst_reg);
9429 dst_reg->type = aux->btf_var.reg_type;
9430 switch (dst_reg->type) {
9432 dst_reg->mem_size = aux->btf_var.mem_size;
9435 case PTR_TO_PERCPU_BTF_ID:
9436 dst_reg->btf = aux->btf_var.btf;
9437 dst_reg->btf_id = aux->btf_var.btf_id;
9440 verbose(env, "bpf verifier is misconfigured\n");
9446 if (insn->src_reg == BPF_PSEUDO_FUNC) {
9447 struct bpf_prog_aux *aux = env->prog->aux;
9448 u32 subprogno = find_subprog(env,
9449 env->insn_idx + insn->imm + 1);
9451 if (!aux->func_info) {
9452 verbose(env, "missing btf func_info\n");
9455 if (aux->func_info_aux[subprogno].linkage != BTF_FUNC_STATIC) {
9456 verbose(env, "callback function not static\n");
9460 dst_reg->type = PTR_TO_FUNC;
9461 dst_reg->subprogno = subprogno;
9465 map = env->used_maps[aux->map_index];
9466 mark_reg_known_zero(env, regs, insn->dst_reg);
9467 dst_reg->map_ptr = map;
9469 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE ||
9470 insn->src_reg == BPF_PSEUDO_MAP_IDX_VALUE) {
9471 dst_reg->type = PTR_TO_MAP_VALUE;
9472 dst_reg->off = aux->map_off;
9473 if (map_value_has_spin_lock(map))
9474 dst_reg->id = ++env->id_gen;
9475 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD ||
9476 insn->src_reg == BPF_PSEUDO_MAP_IDX) {
9477 dst_reg->type = CONST_PTR_TO_MAP;
9479 verbose(env, "bpf verifier is misconfigured\n");
9486 static bool may_access_skb(enum bpf_prog_type type)
9489 case BPF_PROG_TYPE_SOCKET_FILTER:
9490 case BPF_PROG_TYPE_SCHED_CLS:
9491 case BPF_PROG_TYPE_SCHED_ACT:
9498 /* verify safety of LD_ABS|LD_IND instructions:
9499 * - they can only appear in the programs where ctx == skb
9500 * - since they are wrappers of function calls, they scratch R1-R5 registers,
9501 * preserve R6-R9, and store return value into R0
9504 * ctx == skb == R6 == CTX
9507 * SRC == any register
9508 * IMM == 32-bit immediate
9511 * R0 - 8/16/32-bit skb data converted to cpu endianness
9513 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
9515 struct bpf_reg_state *regs = cur_regs(env);
9516 static const int ctx_reg = BPF_REG_6;
9517 u8 mode = BPF_MODE(insn->code);
9520 if (!may_access_skb(resolve_prog_type(env->prog))) {
9521 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
9525 if (!env->ops->gen_ld_abs) {
9526 verbose(env, "bpf verifier is misconfigured\n");
9530 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
9531 BPF_SIZE(insn->code) == BPF_DW ||
9532 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
9533 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
9537 /* check whether implicit source operand (register R6) is readable */
9538 err = check_reg_arg(env, ctx_reg, SRC_OP);
9542 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
9543 * gen_ld_abs() may terminate the program at runtime, leading to
9546 err = check_reference_leak(env);
9548 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
9552 if (env->cur_state->active_spin_lock) {
9553 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
9557 if (regs[ctx_reg].type != PTR_TO_CTX) {
9559 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
9563 if (mode == BPF_IND) {
9564 /* check explicit source operand */
9565 err = check_reg_arg(env, insn->src_reg, SRC_OP);
9570 err = check_ctx_reg(env, ®s[ctx_reg], ctx_reg);
9574 /* reset caller saved regs to unreadable */
9575 for (i = 0; i < CALLER_SAVED_REGS; i++) {
9576 mark_reg_not_init(env, regs, caller_saved[i]);
9577 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
9580 /* mark destination R0 register as readable, since it contains
9581 * the value fetched from the packet.
9582 * Already marked as written above.
9584 mark_reg_unknown(env, regs, BPF_REG_0);
9585 /* ld_abs load up to 32-bit skb data. */
9586 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
9590 static int check_return_code(struct bpf_verifier_env *env)
9592 struct tnum enforce_attach_type_range = tnum_unknown;
9593 const struct bpf_prog *prog = env->prog;
9594 struct bpf_reg_state *reg;
9595 struct tnum range = tnum_range(0, 1);
9596 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
9598 struct bpf_func_state *frame = env->cur_state->frame[0];
9599 const bool is_subprog = frame->subprogno;
9601 /* LSM and struct_ops func-ptr's return type could be "void" */
9603 (prog_type == BPF_PROG_TYPE_STRUCT_OPS ||
9604 prog_type == BPF_PROG_TYPE_LSM) &&
9605 !prog->aux->attach_func_proto->type)
9608 /* eBPF calling convention is such that R0 is used
9609 * to return the value from eBPF program.
9610 * Make sure that it's readable at this time
9611 * of bpf_exit, which means that program wrote
9612 * something into it earlier
9614 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
9618 if (is_pointer_value(env, BPF_REG_0)) {
9619 verbose(env, "R0 leaks addr as return value\n");
9623 reg = cur_regs(env) + BPF_REG_0;
9625 if (frame->in_async_callback_fn) {
9626 /* enforce return zero from async callbacks like timer */
9627 if (reg->type != SCALAR_VALUE) {
9628 verbose(env, "In async callback the register R0 is not a known value (%s)\n",
9629 reg_type_str[reg->type]);
9633 if (!tnum_in(tnum_const(0), reg->var_off)) {
9634 verbose_invalid_scalar(env, reg, &range, "async callback", "R0");
9641 if (reg->type != SCALAR_VALUE) {
9642 verbose(env, "At subprogram exit the register R0 is not a scalar value (%s)\n",
9643 reg_type_str[reg->type]);
9649 switch (prog_type) {
9650 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
9651 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
9652 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG ||
9653 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETPEERNAME ||
9654 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETPEERNAME ||
9655 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETSOCKNAME ||
9656 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETSOCKNAME)
9657 range = tnum_range(1, 1);
9658 if (env->prog->expected_attach_type == BPF_CGROUP_INET4_BIND ||
9659 env->prog->expected_attach_type == BPF_CGROUP_INET6_BIND)
9660 range = tnum_range(0, 3);
9662 case BPF_PROG_TYPE_CGROUP_SKB:
9663 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
9664 range = tnum_range(0, 3);
9665 enforce_attach_type_range = tnum_range(2, 3);
9668 case BPF_PROG_TYPE_CGROUP_SOCK:
9669 case BPF_PROG_TYPE_SOCK_OPS:
9670 case BPF_PROG_TYPE_CGROUP_DEVICE:
9671 case BPF_PROG_TYPE_CGROUP_SYSCTL:
9672 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
9674 case BPF_PROG_TYPE_RAW_TRACEPOINT:
9675 if (!env->prog->aux->attach_btf_id)
9677 range = tnum_const(0);
9679 case BPF_PROG_TYPE_TRACING:
9680 switch (env->prog->expected_attach_type) {
9681 case BPF_TRACE_FENTRY:
9682 case BPF_TRACE_FEXIT:
9683 range = tnum_const(0);
9685 case BPF_TRACE_RAW_TP:
9686 case BPF_MODIFY_RETURN:
9688 case BPF_TRACE_ITER:
9694 case BPF_PROG_TYPE_SK_LOOKUP:
9695 range = tnum_range(SK_DROP, SK_PASS);
9697 case BPF_PROG_TYPE_EXT:
9698 /* freplace program can return anything as its return value
9699 * depends on the to-be-replaced kernel func or bpf program.
9705 if (reg->type != SCALAR_VALUE) {
9706 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
9707 reg_type_str[reg->type]);
9711 if (!tnum_in(range, reg->var_off)) {
9712 verbose_invalid_scalar(env, reg, &range, "program exit", "R0");
9716 if (!tnum_is_unknown(enforce_attach_type_range) &&
9717 tnum_in(enforce_attach_type_range, reg->var_off))
9718 env->prog->enforce_expected_attach_type = 1;
9722 /* non-recursive DFS pseudo code
9723 * 1 procedure DFS-iterative(G,v):
9724 * 2 label v as discovered
9725 * 3 let S be a stack
9727 * 5 while S is not empty
9729 * 7 if t is what we're looking for:
9731 * 9 for all edges e in G.adjacentEdges(t) do
9732 * 10 if edge e is already labelled
9733 * 11 continue with the next edge
9734 * 12 w <- G.adjacentVertex(t,e)
9735 * 13 if vertex w is not discovered and not explored
9736 * 14 label e as tree-edge
9737 * 15 label w as discovered
9740 * 18 else if vertex w is discovered
9741 * 19 label e as back-edge
9743 * 21 // vertex w is explored
9744 * 22 label e as forward- or cross-edge
9745 * 23 label t as explored
9750 * 0x11 - discovered and fall-through edge labelled
9751 * 0x12 - discovered and fall-through and branch edges labelled
9762 static u32 state_htab_size(struct bpf_verifier_env *env)
9764 return env->prog->len;
9767 static struct bpf_verifier_state_list **explored_state(
9768 struct bpf_verifier_env *env,
9771 struct bpf_verifier_state *cur = env->cur_state;
9772 struct bpf_func_state *state = cur->frame[cur->curframe];
9774 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
9777 static void init_explored_state(struct bpf_verifier_env *env, int idx)
9779 env->insn_aux_data[idx].prune_point = true;
9787 /* t, w, e - match pseudo-code above:
9788 * t - index of current instruction
9789 * w - next instruction
9792 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env,
9795 int *insn_stack = env->cfg.insn_stack;
9796 int *insn_state = env->cfg.insn_state;
9798 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
9799 return DONE_EXPLORING;
9801 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
9802 return DONE_EXPLORING;
9804 if (w < 0 || w >= env->prog->len) {
9805 verbose_linfo(env, t, "%d: ", t);
9806 verbose(env, "jump out of range from insn %d to %d\n", t, w);
9811 /* mark branch target for state pruning */
9812 init_explored_state(env, w);
9814 if (insn_state[w] == 0) {
9816 insn_state[t] = DISCOVERED | e;
9817 insn_state[w] = DISCOVERED;
9818 if (env->cfg.cur_stack >= env->prog->len)
9820 insn_stack[env->cfg.cur_stack++] = w;
9821 return KEEP_EXPLORING;
9822 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
9823 if (loop_ok && env->bpf_capable)
9824 return DONE_EXPLORING;
9825 verbose_linfo(env, t, "%d: ", t);
9826 verbose_linfo(env, w, "%d: ", w);
9827 verbose(env, "back-edge from insn %d to %d\n", t, w);
9829 } else if (insn_state[w] == EXPLORED) {
9830 /* forward- or cross-edge */
9831 insn_state[t] = DISCOVERED | e;
9833 verbose(env, "insn state internal bug\n");
9836 return DONE_EXPLORING;
9839 static int visit_func_call_insn(int t, int insn_cnt,
9840 struct bpf_insn *insns,
9841 struct bpf_verifier_env *env,
9846 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
9850 if (t + 1 < insn_cnt)
9851 init_explored_state(env, t + 1);
9853 init_explored_state(env, t);
9854 ret = push_insn(t, t + insns[t].imm + 1, BRANCH, env,
9855 /* It's ok to allow recursion from CFG point of
9856 * view. __check_func_call() will do the actual
9859 bpf_pseudo_func(insns + t));
9864 /* Visits the instruction at index t and returns one of the following:
9865 * < 0 - an error occurred
9866 * DONE_EXPLORING - the instruction was fully explored
9867 * KEEP_EXPLORING - there is still work to be done before it is fully explored
9869 static int visit_insn(int t, int insn_cnt, struct bpf_verifier_env *env)
9871 struct bpf_insn *insns = env->prog->insnsi;
9874 if (bpf_pseudo_func(insns + t))
9875 return visit_func_call_insn(t, insn_cnt, insns, env, true);
9877 /* All non-branch instructions have a single fall-through edge. */
9878 if (BPF_CLASS(insns[t].code) != BPF_JMP &&
9879 BPF_CLASS(insns[t].code) != BPF_JMP32)
9880 return push_insn(t, t + 1, FALLTHROUGH, env, false);
9882 switch (BPF_OP(insns[t].code)) {
9884 return DONE_EXPLORING;
9887 if (insns[t].imm == BPF_FUNC_timer_set_callback)
9888 /* Mark this call insn to trigger is_state_visited() check
9889 * before call itself is processed by __check_func_call().
9890 * Otherwise new async state will be pushed for further
9893 init_explored_state(env, t);
9894 return visit_func_call_insn(t, insn_cnt, insns, env,
9895 insns[t].src_reg == BPF_PSEUDO_CALL);
9898 if (BPF_SRC(insns[t].code) != BPF_K)
9901 /* unconditional jump with single edge */
9902 ret = push_insn(t, t + insns[t].off + 1, FALLTHROUGH, env,
9907 /* unconditional jmp is not a good pruning point,
9908 * but it's marked, since backtracking needs
9909 * to record jmp history in is_state_visited().
9911 init_explored_state(env, t + insns[t].off + 1);
9912 /* tell verifier to check for equivalent states
9913 * after every call and jump
9915 if (t + 1 < insn_cnt)
9916 init_explored_state(env, t + 1);
9921 /* conditional jump with two edges */
9922 init_explored_state(env, t);
9923 ret = push_insn(t, t + 1, FALLTHROUGH, env, true);
9927 return push_insn(t, t + insns[t].off + 1, BRANCH, env, true);
9931 /* non-recursive depth-first-search to detect loops in BPF program
9932 * loop == back-edge in directed graph
9934 static int check_cfg(struct bpf_verifier_env *env)
9936 int insn_cnt = env->prog->len;
9937 int *insn_stack, *insn_state;
9941 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
9945 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
9951 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
9952 insn_stack[0] = 0; /* 0 is the first instruction */
9953 env->cfg.cur_stack = 1;
9955 while (env->cfg.cur_stack > 0) {
9956 int t = insn_stack[env->cfg.cur_stack - 1];
9958 ret = visit_insn(t, insn_cnt, env);
9960 case DONE_EXPLORING:
9961 insn_state[t] = EXPLORED;
9962 env->cfg.cur_stack--;
9964 case KEEP_EXPLORING:
9968 verbose(env, "visit_insn internal bug\n");
9975 if (env->cfg.cur_stack < 0) {
9976 verbose(env, "pop stack internal bug\n");
9981 for (i = 0; i < insn_cnt; i++) {
9982 if (insn_state[i] != EXPLORED) {
9983 verbose(env, "unreachable insn %d\n", i);
9988 ret = 0; /* cfg looks good */
9993 env->cfg.insn_state = env->cfg.insn_stack = NULL;
9997 static int check_abnormal_return(struct bpf_verifier_env *env)
10001 for (i = 1; i < env->subprog_cnt; i++) {
10002 if (env->subprog_info[i].has_ld_abs) {
10003 verbose(env, "LD_ABS is not allowed in subprogs without BTF\n");
10006 if (env->subprog_info[i].has_tail_call) {
10007 verbose(env, "tail_call is not allowed in subprogs without BTF\n");
10014 /* The minimum supported BTF func info size */
10015 #define MIN_BPF_FUNCINFO_SIZE 8
10016 #define MAX_FUNCINFO_REC_SIZE 252
10018 static int check_btf_func(struct bpf_verifier_env *env,
10019 const union bpf_attr *attr,
10022 const struct btf_type *type, *func_proto, *ret_type;
10023 u32 i, nfuncs, urec_size, min_size;
10024 u32 krec_size = sizeof(struct bpf_func_info);
10025 struct bpf_func_info *krecord;
10026 struct bpf_func_info_aux *info_aux = NULL;
10027 struct bpf_prog *prog;
10028 const struct btf *btf;
10030 u32 prev_offset = 0;
10031 bool scalar_return;
10034 nfuncs = attr->func_info_cnt;
10036 if (check_abnormal_return(env))
10041 if (nfuncs != env->subprog_cnt) {
10042 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
10046 urec_size = attr->func_info_rec_size;
10047 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
10048 urec_size > MAX_FUNCINFO_REC_SIZE ||
10049 urec_size % sizeof(u32)) {
10050 verbose(env, "invalid func info rec size %u\n", urec_size);
10055 btf = prog->aux->btf;
10057 urecord = make_bpfptr(attr->func_info, uattr.is_kernel);
10058 min_size = min_t(u32, krec_size, urec_size);
10060 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
10063 info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN);
10067 for (i = 0; i < nfuncs; i++) {
10068 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
10070 if (ret == -E2BIG) {
10071 verbose(env, "nonzero tailing record in func info");
10072 /* set the size kernel expects so loader can zero
10073 * out the rest of the record.
10075 if (copy_to_bpfptr_offset(uattr,
10076 offsetof(union bpf_attr, func_info_rec_size),
10077 &min_size, sizeof(min_size)))
10083 if (copy_from_bpfptr(&krecord[i], urecord, min_size)) {
10088 /* check insn_off */
10091 if (krecord[i].insn_off) {
10093 "nonzero insn_off %u for the first func info record",
10094 krecord[i].insn_off);
10097 } else if (krecord[i].insn_off <= prev_offset) {
10099 "same or smaller insn offset (%u) than previous func info record (%u)",
10100 krecord[i].insn_off, prev_offset);
10104 if (env->subprog_info[i].start != krecord[i].insn_off) {
10105 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
10109 /* check type_id */
10110 type = btf_type_by_id(btf, krecord[i].type_id);
10111 if (!type || !btf_type_is_func(type)) {
10112 verbose(env, "invalid type id %d in func info",
10113 krecord[i].type_id);
10116 info_aux[i].linkage = BTF_INFO_VLEN(type->info);
10118 func_proto = btf_type_by_id(btf, type->type);
10119 if (unlikely(!func_proto || !btf_type_is_func_proto(func_proto)))
10120 /* btf_func_check() already verified it during BTF load */
10122 ret_type = btf_type_skip_modifiers(btf, func_proto->type, NULL);
10124 btf_type_is_small_int(ret_type) || btf_type_is_enum(ret_type);
10125 if (i && !scalar_return && env->subprog_info[i].has_ld_abs) {
10126 verbose(env, "LD_ABS is only allowed in functions that return 'int'.\n");
10129 if (i && !scalar_return && env->subprog_info[i].has_tail_call) {
10130 verbose(env, "tail_call is only allowed in functions that return 'int'.\n");
10134 prev_offset = krecord[i].insn_off;
10135 bpfptr_add(&urecord, urec_size);
10138 prog->aux->func_info = krecord;
10139 prog->aux->func_info_cnt = nfuncs;
10140 prog->aux->func_info_aux = info_aux;
10149 static void adjust_btf_func(struct bpf_verifier_env *env)
10151 struct bpf_prog_aux *aux = env->prog->aux;
10154 if (!aux->func_info)
10157 for (i = 0; i < env->subprog_cnt; i++)
10158 aux->func_info[i].insn_off = env->subprog_info[i].start;
10161 #define MIN_BPF_LINEINFO_SIZE (offsetof(struct bpf_line_info, line_col) + \
10162 sizeof(((struct bpf_line_info *)(0))->line_col))
10163 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
10165 static int check_btf_line(struct bpf_verifier_env *env,
10166 const union bpf_attr *attr,
10169 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
10170 struct bpf_subprog_info *sub;
10171 struct bpf_line_info *linfo;
10172 struct bpf_prog *prog;
10173 const struct btf *btf;
10177 nr_linfo = attr->line_info_cnt;
10180 if (nr_linfo > INT_MAX / sizeof(struct bpf_line_info))
10183 rec_size = attr->line_info_rec_size;
10184 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
10185 rec_size > MAX_LINEINFO_REC_SIZE ||
10186 rec_size & (sizeof(u32) - 1))
10189 /* Need to zero it in case the userspace may
10190 * pass in a smaller bpf_line_info object.
10192 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
10193 GFP_KERNEL | __GFP_NOWARN);
10198 btf = prog->aux->btf;
10201 sub = env->subprog_info;
10202 ulinfo = make_bpfptr(attr->line_info, uattr.is_kernel);
10203 expected_size = sizeof(struct bpf_line_info);
10204 ncopy = min_t(u32, expected_size, rec_size);
10205 for (i = 0; i < nr_linfo; i++) {
10206 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
10208 if (err == -E2BIG) {
10209 verbose(env, "nonzero tailing record in line_info");
10210 if (copy_to_bpfptr_offset(uattr,
10211 offsetof(union bpf_attr, line_info_rec_size),
10212 &expected_size, sizeof(expected_size)))
10218 if (copy_from_bpfptr(&linfo[i], ulinfo, ncopy)) {
10224 * Check insn_off to ensure
10225 * 1) strictly increasing AND
10226 * 2) bounded by prog->len
10228 * The linfo[0].insn_off == 0 check logically falls into
10229 * the later "missing bpf_line_info for func..." case
10230 * because the first linfo[0].insn_off must be the
10231 * first sub also and the first sub must have
10232 * subprog_info[0].start == 0.
10234 if ((i && linfo[i].insn_off <= prev_offset) ||
10235 linfo[i].insn_off >= prog->len) {
10236 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
10237 i, linfo[i].insn_off, prev_offset,
10243 if (!prog->insnsi[linfo[i].insn_off].code) {
10245 "Invalid insn code at line_info[%u].insn_off\n",
10251 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
10252 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
10253 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
10258 if (s != env->subprog_cnt) {
10259 if (linfo[i].insn_off == sub[s].start) {
10260 sub[s].linfo_idx = i;
10262 } else if (sub[s].start < linfo[i].insn_off) {
10263 verbose(env, "missing bpf_line_info for func#%u\n", s);
10269 prev_offset = linfo[i].insn_off;
10270 bpfptr_add(&ulinfo, rec_size);
10273 if (s != env->subprog_cnt) {
10274 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
10275 env->subprog_cnt - s, s);
10280 prog->aux->linfo = linfo;
10281 prog->aux->nr_linfo = nr_linfo;
10290 #define MIN_CORE_RELO_SIZE sizeof(struct bpf_core_relo)
10291 #define MAX_CORE_RELO_SIZE MAX_FUNCINFO_REC_SIZE
10293 static int check_core_relo(struct bpf_verifier_env *env,
10294 const union bpf_attr *attr,
10297 u32 i, nr_core_relo, ncopy, expected_size, rec_size;
10298 struct bpf_core_relo core_relo = {};
10299 struct bpf_prog *prog = env->prog;
10300 const struct btf *btf = prog->aux->btf;
10301 struct bpf_core_ctx ctx = {
10305 bpfptr_t u_core_relo;
10308 nr_core_relo = attr->core_relo_cnt;
10311 if (nr_core_relo > INT_MAX / sizeof(struct bpf_core_relo))
10314 rec_size = attr->core_relo_rec_size;
10315 if (rec_size < MIN_CORE_RELO_SIZE ||
10316 rec_size > MAX_CORE_RELO_SIZE ||
10317 rec_size % sizeof(u32))
10320 u_core_relo = make_bpfptr(attr->core_relos, uattr.is_kernel);
10321 expected_size = sizeof(struct bpf_core_relo);
10322 ncopy = min_t(u32, expected_size, rec_size);
10324 /* Unlike func_info and line_info, copy and apply each CO-RE
10325 * relocation record one at a time.
10327 for (i = 0; i < nr_core_relo; i++) {
10328 /* future proofing when sizeof(bpf_core_relo) changes */
10329 err = bpf_check_uarg_tail_zero(u_core_relo, expected_size, rec_size);
10331 if (err == -E2BIG) {
10332 verbose(env, "nonzero tailing record in core_relo");
10333 if (copy_to_bpfptr_offset(uattr,
10334 offsetof(union bpf_attr, core_relo_rec_size),
10335 &expected_size, sizeof(expected_size)))
10341 if (copy_from_bpfptr(&core_relo, u_core_relo, ncopy)) {
10346 if (core_relo.insn_off % 8 || core_relo.insn_off / 8 >= prog->len) {
10347 verbose(env, "Invalid core_relo[%u].insn_off:%u prog->len:%u\n",
10348 i, core_relo.insn_off, prog->len);
10353 err = bpf_core_apply(&ctx, &core_relo, i,
10354 &prog->insnsi[core_relo.insn_off / 8]);
10357 bpfptr_add(&u_core_relo, rec_size);
10362 static int check_btf_info(struct bpf_verifier_env *env,
10363 const union bpf_attr *attr,
10369 if (!attr->func_info_cnt && !attr->line_info_cnt) {
10370 if (check_abnormal_return(env))
10375 btf = btf_get_by_fd(attr->prog_btf_fd);
10377 return PTR_ERR(btf);
10378 if (btf_is_kernel(btf)) {
10382 env->prog->aux->btf = btf;
10384 err = check_btf_func(env, attr, uattr);
10388 err = check_btf_line(env, attr, uattr);
10392 err = check_core_relo(env, attr, uattr);
10399 /* check %cur's range satisfies %old's */
10400 static bool range_within(struct bpf_reg_state *old,
10401 struct bpf_reg_state *cur)
10403 return old->umin_value <= cur->umin_value &&
10404 old->umax_value >= cur->umax_value &&
10405 old->smin_value <= cur->smin_value &&
10406 old->smax_value >= cur->smax_value &&
10407 old->u32_min_value <= cur->u32_min_value &&
10408 old->u32_max_value >= cur->u32_max_value &&
10409 old->s32_min_value <= cur->s32_min_value &&
10410 old->s32_max_value >= cur->s32_max_value;
10413 /* If in the old state two registers had the same id, then they need to have
10414 * the same id in the new state as well. But that id could be different from
10415 * the old state, so we need to track the mapping from old to new ids.
10416 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
10417 * regs with old id 5 must also have new id 9 for the new state to be safe. But
10418 * regs with a different old id could still have new id 9, we don't care about
10420 * So we look through our idmap to see if this old id has been seen before. If
10421 * so, we require the new id to match; otherwise, we add the id pair to the map.
10423 static bool check_ids(u32 old_id, u32 cur_id, struct bpf_id_pair *idmap)
10427 for (i = 0; i < BPF_ID_MAP_SIZE; i++) {
10428 if (!idmap[i].old) {
10429 /* Reached an empty slot; haven't seen this id before */
10430 idmap[i].old = old_id;
10431 idmap[i].cur = cur_id;
10434 if (idmap[i].old == old_id)
10435 return idmap[i].cur == cur_id;
10437 /* We ran out of idmap slots, which should be impossible */
10442 static void clean_func_state(struct bpf_verifier_env *env,
10443 struct bpf_func_state *st)
10445 enum bpf_reg_liveness live;
10448 for (i = 0; i < BPF_REG_FP; i++) {
10449 live = st->regs[i].live;
10450 /* liveness must not touch this register anymore */
10451 st->regs[i].live |= REG_LIVE_DONE;
10452 if (!(live & REG_LIVE_READ))
10453 /* since the register is unused, clear its state
10454 * to make further comparison simpler
10456 __mark_reg_not_init(env, &st->regs[i]);
10459 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
10460 live = st->stack[i].spilled_ptr.live;
10461 /* liveness must not touch this stack slot anymore */
10462 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
10463 if (!(live & REG_LIVE_READ)) {
10464 __mark_reg_not_init(env, &st->stack[i].spilled_ptr);
10465 for (j = 0; j < BPF_REG_SIZE; j++)
10466 st->stack[i].slot_type[j] = STACK_INVALID;
10471 static void clean_verifier_state(struct bpf_verifier_env *env,
10472 struct bpf_verifier_state *st)
10476 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
10477 /* all regs in this state in all frames were already marked */
10480 for (i = 0; i <= st->curframe; i++)
10481 clean_func_state(env, st->frame[i]);
10484 /* the parentage chains form a tree.
10485 * the verifier states are added to state lists at given insn and
10486 * pushed into state stack for future exploration.
10487 * when the verifier reaches bpf_exit insn some of the verifer states
10488 * stored in the state lists have their final liveness state already,
10489 * but a lot of states will get revised from liveness point of view when
10490 * the verifier explores other branches.
10493 * 2: if r1 == 100 goto pc+1
10496 * when the verifier reaches exit insn the register r0 in the state list of
10497 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
10498 * of insn 2 and goes exploring further. At the insn 4 it will walk the
10499 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
10501 * Since the verifier pushes the branch states as it sees them while exploring
10502 * the program the condition of walking the branch instruction for the second
10503 * time means that all states below this branch were already explored and
10504 * their final liveness marks are already propagated.
10505 * Hence when the verifier completes the search of state list in is_state_visited()
10506 * we can call this clean_live_states() function to mark all liveness states
10507 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
10508 * will not be used.
10509 * This function also clears the registers and stack for states that !READ
10510 * to simplify state merging.
10512 * Important note here that walking the same branch instruction in the callee
10513 * doesn't meant that the states are DONE. The verifier has to compare
10516 static void clean_live_states(struct bpf_verifier_env *env, int insn,
10517 struct bpf_verifier_state *cur)
10519 struct bpf_verifier_state_list *sl;
10522 sl = *explored_state(env, insn);
10524 if (sl->state.branches)
10526 if (sl->state.insn_idx != insn ||
10527 sl->state.curframe != cur->curframe)
10529 for (i = 0; i <= cur->curframe; i++)
10530 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
10532 clean_verifier_state(env, &sl->state);
10538 /* Returns true if (rold safe implies rcur safe) */
10539 static bool regsafe(struct bpf_verifier_env *env, struct bpf_reg_state *rold,
10540 struct bpf_reg_state *rcur, struct bpf_id_pair *idmap)
10544 if (!(rold->live & REG_LIVE_READ))
10545 /* explored state didn't use this */
10548 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
10550 if (rold->type == PTR_TO_STACK)
10551 /* two stack pointers are equal only if they're pointing to
10552 * the same stack frame, since fp-8 in foo != fp-8 in bar
10554 return equal && rold->frameno == rcur->frameno;
10559 if (rold->type == NOT_INIT)
10560 /* explored state can't have used this */
10562 if (rcur->type == NOT_INIT)
10564 switch (rold->type) {
10566 if (env->explore_alu_limits)
10568 if (rcur->type == SCALAR_VALUE) {
10569 if (!rold->precise && !rcur->precise)
10571 /* new val must satisfy old val knowledge */
10572 return range_within(rold, rcur) &&
10573 tnum_in(rold->var_off, rcur->var_off);
10575 /* We're trying to use a pointer in place of a scalar.
10576 * Even if the scalar was unbounded, this could lead to
10577 * pointer leaks because scalars are allowed to leak
10578 * while pointers are not. We could make this safe in
10579 * special cases if root is calling us, but it's
10580 * probably not worth the hassle.
10584 case PTR_TO_MAP_KEY:
10585 case PTR_TO_MAP_VALUE:
10586 /* If the new min/max/var_off satisfy the old ones and
10587 * everything else matches, we are OK.
10588 * 'id' is not compared, since it's only used for maps with
10589 * bpf_spin_lock inside map element and in such cases if
10590 * the rest of the prog is valid for one map element then
10591 * it's valid for all map elements regardless of the key
10592 * used in bpf_map_lookup()
10594 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
10595 range_within(rold, rcur) &&
10596 tnum_in(rold->var_off, rcur->var_off);
10597 case PTR_TO_MAP_VALUE_OR_NULL:
10598 /* a PTR_TO_MAP_VALUE could be safe to use as a
10599 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
10600 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
10601 * checked, doing so could have affected others with the same
10602 * id, and we can't check for that because we lost the id when
10603 * we converted to a PTR_TO_MAP_VALUE.
10605 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
10607 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
10609 /* Check our ids match any regs they're supposed to */
10610 return check_ids(rold->id, rcur->id, idmap);
10611 case PTR_TO_PACKET_META:
10612 case PTR_TO_PACKET:
10613 if (rcur->type != rold->type)
10615 /* We must have at least as much range as the old ptr
10616 * did, so that any accesses which were safe before are
10617 * still safe. This is true even if old range < old off,
10618 * since someone could have accessed through (ptr - k), or
10619 * even done ptr -= k in a register, to get a safe access.
10621 if (rold->range > rcur->range)
10623 /* If the offsets don't match, we can't trust our alignment;
10624 * nor can we be sure that we won't fall out of range.
10626 if (rold->off != rcur->off)
10628 /* id relations must be preserved */
10629 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
10631 /* new val must satisfy old val knowledge */
10632 return range_within(rold, rcur) &&
10633 tnum_in(rold->var_off, rcur->var_off);
10635 case CONST_PTR_TO_MAP:
10636 case PTR_TO_PACKET_END:
10637 case PTR_TO_FLOW_KEYS:
10638 case PTR_TO_SOCKET:
10639 case PTR_TO_SOCKET_OR_NULL:
10640 case PTR_TO_SOCK_COMMON:
10641 case PTR_TO_SOCK_COMMON_OR_NULL:
10642 case PTR_TO_TCP_SOCK:
10643 case PTR_TO_TCP_SOCK_OR_NULL:
10644 case PTR_TO_XDP_SOCK:
10645 /* Only valid matches are exact, which memcmp() above
10646 * would have accepted
10649 /* Don't know what's going on, just say it's not safe */
10653 /* Shouldn't get here; if we do, say it's not safe */
10658 static bool stacksafe(struct bpf_verifier_env *env, struct bpf_func_state *old,
10659 struct bpf_func_state *cur, struct bpf_id_pair *idmap)
10663 /* walk slots of the explored stack and ignore any additional
10664 * slots in the current stack, since explored(safe) state
10667 for (i = 0; i < old->allocated_stack; i++) {
10668 spi = i / BPF_REG_SIZE;
10670 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
10671 i += BPF_REG_SIZE - 1;
10672 /* explored state didn't use this */
10676 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
10679 /* explored stack has more populated slots than current stack
10680 * and these slots were used
10682 if (i >= cur->allocated_stack)
10685 /* if old state was safe with misc data in the stack
10686 * it will be safe with zero-initialized stack.
10687 * The opposite is not true
10689 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
10690 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
10692 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
10693 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
10694 /* Ex: old explored (safe) state has STACK_SPILL in
10695 * this stack slot, but current has STACK_MISC ->
10696 * this verifier states are not equivalent,
10697 * return false to continue verification of this path
10700 if (i % BPF_REG_SIZE != BPF_REG_SIZE - 1)
10702 if (!is_spilled_reg(&old->stack[spi]))
10704 if (!regsafe(env, &old->stack[spi].spilled_ptr,
10705 &cur->stack[spi].spilled_ptr, idmap))
10706 /* when explored and current stack slot are both storing
10707 * spilled registers, check that stored pointers types
10708 * are the same as well.
10709 * Ex: explored safe path could have stored
10710 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
10711 * but current path has stored:
10712 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
10713 * such verifier states are not equivalent.
10714 * return false to continue verification of this path
10721 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
10723 if (old->acquired_refs != cur->acquired_refs)
10725 return !memcmp(old->refs, cur->refs,
10726 sizeof(*old->refs) * old->acquired_refs);
10729 /* compare two verifier states
10731 * all states stored in state_list are known to be valid, since
10732 * verifier reached 'bpf_exit' instruction through them
10734 * this function is called when verifier exploring different branches of
10735 * execution popped from the state stack. If it sees an old state that has
10736 * more strict register state and more strict stack state then this execution
10737 * branch doesn't need to be explored further, since verifier already
10738 * concluded that more strict state leads to valid finish.
10740 * Therefore two states are equivalent if register state is more conservative
10741 * and explored stack state is more conservative than the current one.
10744 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
10745 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
10747 * In other words if current stack state (one being explored) has more
10748 * valid slots than old one that already passed validation, it means
10749 * the verifier can stop exploring and conclude that current state is valid too
10751 * Similarly with registers. If explored state has register type as invalid
10752 * whereas register type in current state is meaningful, it means that
10753 * the current state will reach 'bpf_exit' instruction safely
10755 static bool func_states_equal(struct bpf_verifier_env *env, struct bpf_func_state *old,
10756 struct bpf_func_state *cur)
10760 memset(env->idmap_scratch, 0, sizeof(env->idmap_scratch));
10761 for (i = 0; i < MAX_BPF_REG; i++)
10762 if (!regsafe(env, &old->regs[i], &cur->regs[i],
10763 env->idmap_scratch))
10766 if (!stacksafe(env, old, cur, env->idmap_scratch))
10769 if (!refsafe(old, cur))
10775 static bool states_equal(struct bpf_verifier_env *env,
10776 struct bpf_verifier_state *old,
10777 struct bpf_verifier_state *cur)
10781 if (old->curframe != cur->curframe)
10784 /* Verification state from speculative execution simulation
10785 * must never prune a non-speculative execution one.
10787 if (old->speculative && !cur->speculative)
10790 if (old->active_spin_lock != cur->active_spin_lock)
10793 /* for states to be equal callsites have to be the same
10794 * and all frame states need to be equivalent
10796 for (i = 0; i <= old->curframe; i++) {
10797 if (old->frame[i]->callsite != cur->frame[i]->callsite)
10799 if (!func_states_equal(env, old->frame[i], cur->frame[i]))
10805 /* Return 0 if no propagation happened. Return negative error code if error
10806 * happened. Otherwise, return the propagated bit.
10808 static int propagate_liveness_reg(struct bpf_verifier_env *env,
10809 struct bpf_reg_state *reg,
10810 struct bpf_reg_state *parent_reg)
10812 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
10813 u8 flag = reg->live & REG_LIVE_READ;
10816 /* When comes here, read flags of PARENT_REG or REG could be any of
10817 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
10818 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
10820 if (parent_flag == REG_LIVE_READ64 ||
10821 /* Or if there is no read flag from REG. */
10823 /* Or if the read flag from REG is the same as PARENT_REG. */
10824 parent_flag == flag)
10827 err = mark_reg_read(env, reg, parent_reg, flag);
10834 /* A write screens off any subsequent reads; but write marks come from the
10835 * straight-line code between a state and its parent. When we arrive at an
10836 * equivalent state (jump target or such) we didn't arrive by the straight-line
10837 * code, so read marks in the state must propagate to the parent regardless
10838 * of the state's write marks. That's what 'parent == state->parent' comparison
10839 * in mark_reg_read() is for.
10841 static int propagate_liveness(struct bpf_verifier_env *env,
10842 const struct bpf_verifier_state *vstate,
10843 struct bpf_verifier_state *vparent)
10845 struct bpf_reg_state *state_reg, *parent_reg;
10846 struct bpf_func_state *state, *parent;
10847 int i, frame, err = 0;
10849 if (vparent->curframe != vstate->curframe) {
10850 WARN(1, "propagate_live: parent frame %d current frame %d\n",
10851 vparent->curframe, vstate->curframe);
10854 /* Propagate read liveness of registers... */
10855 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
10856 for (frame = 0; frame <= vstate->curframe; frame++) {
10857 parent = vparent->frame[frame];
10858 state = vstate->frame[frame];
10859 parent_reg = parent->regs;
10860 state_reg = state->regs;
10861 /* We don't need to worry about FP liveness, it's read-only */
10862 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
10863 err = propagate_liveness_reg(env, &state_reg[i],
10867 if (err == REG_LIVE_READ64)
10868 mark_insn_zext(env, &parent_reg[i]);
10871 /* Propagate stack slots. */
10872 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
10873 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
10874 parent_reg = &parent->stack[i].spilled_ptr;
10875 state_reg = &state->stack[i].spilled_ptr;
10876 err = propagate_liveness_reg(env, state_reg,
10885 /* find precise scalars in the previous equivalent state and
10886 * propagate them into the current state
10888 static int propagate_precision(struct bpf_verifier_env *env,
10889 const struct bpf_verifier_state *old)
10891 struct bpf_reg_state *state_reg;
10892 struct bpf_func_state *state;
10895 state = old->frame[old->curframe];
10896 state_reg = state->regs;
10897 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
10898 if (state_reg->type != SCALAR_VALUE ||
10899 !state_reg->precise)
10901 if (env->log.level & BPF_LOG_LEVEL2)
10902 verbose(env, "propagating r%d\n", i);
10903 err = mark_chain_precision(env, i);
10908 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
10909 if (!is_spilled_reg(&state->stack[i]))
10911 state_reg = &state->stack[i].spilled_ptr;
10912 if (state_reg->type != SCALAR_VALUE ||
10913 !state_reg->precise)
10915 if (env->log.level & BPF_LOG_LEVEL2)
10916 verbose(env, "propagating fp%d\n",
10917 (-i - 1) * BPF_REG_SIZE);
10918 err = mark_chain_precision_stack(env, i);
10925 static bool states_maybe_looping(struct bpf_verifier_state *old,
10926 struct bpf_verifier_state *cur)
10928 struct bpf_func_state *fold, *fcur;
10929 int i, fr = cur->curframe;
10931 if (old->curframe != fr)
10934 fold = old->frame[fr];
10935 fcur = cur->frame[fr];
10936 for (i = 0; i < MAX_BPF_REG; i++)
10937 if (memcmp(&fold->regs[i], &fcur->regs[i],
10938 offsetof(struct bpf_reg_state, parent)))
10944 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
10946 struct bpf_verifier_state_list *new_sl;
10947 struct bpf_verifier_state_list *sl, **pprev;
10948 struct bpf_verifier_state *cur = env->cur_state, *new;
10949 int i, j, err, states_cnt = 0;
10950 bool add_new_state = env->test_state_freq ? true : false;
10952 cur->last_insn_idx = env->prev_insn_idx;
10953 if (!env->insn_aux_data[insn_idx].prune_point)
10954 /* this 'insn_idx' instruction wasn't marked, so we will not
10955 * be doing state search here
10959 /* bpf progs typically have pruning point every 4 instructions
10960 * http://vger.kernel.org/bpfconf2019.html#session-1
10961 * Do not add new state for future pruning if the verifier hasn't seen
10962 * at least 2 jumps and at least 8 instructions.
10963 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
10964 * In tests that amounts to up to 50% reduction into total verifier
10965 * memory consumption and 20% verifier time speedup.
10967 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
10968 env->insn_processed - env->prev_insn_processed >= 8)
10969 add_new_state = true;
10971 pprev = explored_state(env, insn_idx);
10974 clean_live_states(env, insn_idx, cur);
10978 if (sl->state.insn_idx != insn_idx)
10981 if (sl->state.branches) {
10982 struct bpf_func_state *frame = sl->state.frame[sl->state.curframe];
10984 if (frame->in_async_callback_fn &&
10985 frame->async_entry_cnt != cur->frame[cur->curframe]->async_entry_cnt) {
10986 /* Different async_entry_cnt means that the verifier is
10987 * processing another entry into async callback.
10988 * Seeing the same state is not an indication of infinite
10989 * loop or infinite recursion.
10990 * But finding the same state doesn't mean that it's safe
10991 * to stop processing the current state. The previous state
10992 * hasn't yet reached bpf_exit, since state.branches > 0.
10993 * Checking in_async_callback_fn alone is not enough either.
10994 * Since the verifier still needs to catch infinite loops
10995 * inside async callbacks.
10997 } else if (states_maybe_looping(&sl->state, cur) &&
10998 states_equal(env, &sl->state, cur)) {
10999 verbose_linfo(env, insn_idx, "; ");
11000 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
11003 /* if the verifier is processing a loop, avoid adding new state
11004 * too often, since different loop iterations have distinct
11005 * states and may not help future pruning.
11006 * This threshold shouldn't be too low to make sure that
11007 * a loop with large bound will be rejected quickly.
11008 * The most abusive loop will be:
11010 * if r1 < 1000000 goto pc-2
11011 * 1M insn_procssed limit / 100 == 10k peak states.
11012 * This threshold shouldn't be too high either, since states
11013 * at the end of the loop are likely to be useful in pruning.
11015 if (env->jmps_processed - env->prev_jmps_processed < 20 &&
11016 env->insn_processed - env->prev_insn_processed < 100)
11017 add_new_state = false;
11020 if (states_equal(env, &sl->state, cur)) {
11022 /* reached equivalent register/stack state,
11023 * prune the search.
11024 * Registers read by the continuation are read by us.
11025 * If we have any write marks in env->cur_state, they
11026 * will prevent corresponding reads in the continuation
11027 * from reaching our parent (an explored_state). Our
11028 * own state will get the read marks recorded, but
11029 * they'll be immediately forgotten as we're pruning
11030 * this state and will pop a new one.
11032 err = propagate_liveness(env, &sl->state, cur);
11034 /* if previous state reached the exit with precision and
11035 * current state is equivalent to it (except precsion marks)
11036 * the precision needs to be propagated back in
11037 * the current state.
11039 err = err ? : push_jmp_history(env, cur);
11040 err = err ? : propagate_precision(env, &sl->state);
11046 /* when new state is not going to be added do not increase miss count.
11047 * Otherwise several loop iterations will remove the state
11048 * recorded earlier. The goal of these heuristics is to have
11049 * states from some iterations of the loop (some in the beginning
11050 * and some at the end) to help pruning.
11054 /* heuristic to determine whether this state is beneficial
11055 * to keep checking from state equivalence point of view.
11056 * Higher numbers increase max_states_per_insn and verification time,
11057 * but do not meaningfully decrease insn_processed.
11059 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
11060 /* the state is unlikely to be useful. Remove it to
11061 * speed up verification
11064 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
11065 u32 br = sl->state.branches;
11068 "BUG live_done but branches_to_explore %d\n",
11070 free_verifier_state(&sl->state, false);
11072 env->peak_states--;
11074 /* cannot free this state, since parentage chain may
11075 * walk it later. Add it for free_list instead to
11076 * be freed at the end of verification
11078 sl->next = env->free_list;
11079 env->free_list = sl;
11089 if (env->max_states_per_insn < states_cnt)
11090 env->max_states_per_insn = states_cnt;
11092 if (!env->bpf_capable && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
11093 return push_jmp_history(env, cur);
11095 if (!add_new_state)
11096 return push_jmp_history(env, cur);
11098 /* There were no equivalent states, remember the current one.
11099 * Technically the current state is not proven to be safe yet,
11100 * but it will either reach outer most bpf_exit (which means it's safe)
11101 * or it will be rejected. When there are no loops the verifier won't be
11102 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
11103 * again on the way to bpf_exit.
11104 * When looping the sl->state.branches will be > 0 and this state
11105 * will not be considered for equivalence until branches == 0.
11107 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
11110 env->total_states++;
11111 env->peak_states++;
11112 env->prev_jmps_processed = env->jmps_processed;
11113 env->prev_insn_processed = env->insn_processed;
11115 /* add new state to the head of linked list */
11116 new = &new_sl->state;
11117 err = copy_verifier_state(new, cur);
11119 free_verifier_state(new, false);
11123 new->insn_idx = insn_idx;
11124 WARN_ONCE(new->branches != 1,
11125 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
11128 cur->first_insn_idx = insn_idx;
11129 clear_jmp_history(cur);
11130 new_sl->next = *explored_state(env, insn_idx);
11131 *explored_state(env, insn_idx) = new_sl;
11132 /* connect new state to parentage chain. Current frame needs all
11133 * registers connected. Only r6 - r9 of the callers are alive (pushed
11134 * to the stack implicitly by JITs) so in callers' frames connect just
11135 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
11136 * the state of the call instruction (with WRITTEN set), and r0 comes
11137 * from callee with its full parentage chain, anyway.
11139 /* clear write marks in current state: the writes we did are not writes
11140 * our child did, so they don't screen off its reads from us.
11141 * (There are no read marks in current state, because reads always mark
11142 * their parent and current state never has children yet. Only
11143 * explored_states can get read marks.)
11145 for (j = 0; j <= cur->curframe; j++) {
11146 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
11147 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
11148 for (i = 0; i < BPF_REG_FP; i++)
11149 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
11152 /* all stack frames are accessible from callee, clear them all */
11153 for (j = 0; j <= cur->curframe; j++) {
11154 struct bpf_func_state *frame = cur->frame[j];
11155 struct bpf_func_state *newframe = new->frame[j];
11157 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
11158 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
11159 frame->stack[i].spilled_ptr.parent =
11160 &newframe->stack[i].spilled_ptr;
11166 /* Return true if it's OK to have the same insn return a different type. */
11167 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
11171 case PTR_TO_SOCKET:
11172 case PTR_TO_SOCKET_OR_NULL:
11173 case PTR_TO_SOCK_COMMON:
11174 case PTR_TO_SOCK_COMMON_OR_NULL:
11175 case PTR_TO_TCP_SOCK:
11176 case PTR_TO_TCP_SOCK_OR_NULL:
11177 case PTR_TO_XDP_SOCK:
11178 case PTR_TO_BTF_ID:
11179 case PTR_TO_BTF_ID_OR_NULL:
11186 /* If an instruction was previously used with particular pointer types, then we
11187 * need to be careful to avoid cases such as the below, where it may be ok
11188 * for one branch accessing the pointer, but not ok for the other branch:
11193 * R1 = some_other_valid_ptr;
11196 * R2 = *(u32 *)(R1 + 0);
11198 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
11200 return src != prev && (!reg_type_mismatch_ok(src) ||
11201 !reg_type_mismatch_ok(prev));
11204 static int do_check(struct bpf_verifier_env *env)
11206 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
11207 struct bpf_verifier_state *state = env->cur_state;
11208 struct bpf_insn *insns = env->prog->insnsi;
11209 struct bpf_reg_state *regs;
11210 int insn_cnt = env->prog->len;
11211 bool do_print_state = false;
11212 int prev_insn_idx = -1;
11215 struct bpf_insn *insn;
11219 env->prev_insn_idx = prev_insn_idx;
11220 if (env->insn_idx >= insn_cnt) {
11221 verbose(env, "invalid insn idx %d insn_cnt %d\n",
11222 env->insn_idx, insn_cnt);
11226 insn = &insns[env->insn_idx];
11227 class = BPF_CLASS(insn->code);
11229 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
11231 "BPF program is too large. Processed %d insn\n",
11232 env->insn_processed);
11236 err = is_state_visited(env, env->insn_idx);
11240 /* found equivalent state, can prune the search */
11241 if (env->log.level & BPF_LOG_LEVEL) {
11242 if (do_print_state)
11243 verbose(env, "\nfrom %d to %d%s: safe\n",
11244 env->prev_insn_idx, env->insn_idx,
11245 env->cur_state->speculative ?
11246 " (speculative execution)" : "");
11248 verbose(env, "%d: safe\n", env->insn_idx);
11250 goto process_bpf_exit;
11253 if (signal_pending(current))
11256 if (need_resched())
11259 if (env->log.level & BPF_LOG_LEVEL2 ||
11260 (env->log.level & BPF_LOG_LEVEL && do_print_state)) {
11261 if (env->log.level & BPF_LOG_LEVEL2)
11262 verbose(env, "%d:", env->insn_idx);
11264 verbose(env, "\nfrom %d to %d%s:",
11265 env->prev_insn_idx, env->insn_idx,
11266 env->cur_state->speculative ?
11267 " (speculative execution)" : "");
11268 print_verifier_state(env, state->frame[state->curframe]);
11269 do_print_state = false;
11272 if (env->log.level & BPF_LOG_LEVEL) {
11273 const struct bpf_insn_cbs cbs = {
11274 .cb_call = disasm_kfunc_name,
11275 .cb_print = verbose,
11276 .private_data = env,
11279 verbose_linfo(env, env->insn_idx, "; ");
11280 verbose(env, "%d: ", env->insn_idx);
11281 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
11284 if (bpf_prog_is_dev_bound(env->prog->aux)) {
11285 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
11286 env->prev_insn_idx);
11291 regs = cur_regs(env);
11292 sanitize_mark_insn_seen(env);
11293 prev_insn_idx = env->insn_idx;
11295 if (class == BPF_ALU || class == BPF_ALU64) {
11296 err = check_alu_op(env, insn);
11300 } else if (class == BPF_LDX) {
11301 enum bpf_reg_type *prev_src_type, src_reg_type;
11303 /* check for reserved fields is already done */
11305 /* check src operand */
11306 err = check_reg_arg(env, insn->src_reg, SRC_OP);
11310 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
11314 src_reg_type = regs[insn->src_reg].type;
11316 /* check that memory (src_reg + off) is readable,
11317 * the state of dst_reg will be updated by this func
11319 err = check_mem_access(env, env->insn_idx, insn->src_reg,
11320 insn->off, BPF_SIZE(insn->code),
11321 BPF_READ, insn->dst_reg, false);
11325 prev_src_type = &env->insn_aux_data[env->insn_idx].ptr_type;
11327 if (*prev_src_type == NOT_INIT) {
11328 /* saw a valid insn
11329 * dst_reg = *(u32 *)(src_reg + off)
11330 * save type to validate intersecting paths
11332 *prev_src_type = src_reg_type;
11334 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
11335 /* ABuser program is trying to use the same insn
11336 * dst_reg = *(u32*) (src_reg + off)
11337 * with different pointer types:
11338 * src_reg == ctx in one branch and
11339 * src_reg == stack|map in some other branch.
11342 verbose(env, "same insn cannot be used with different pointers\n");
11346 } else if (class == BPF_STX) {
11347 enum bpf_reg_type *prev_dst_type, dst_reg_type;
11349 if (BPF_MODE(insn->code) == BPF_ATOMIC) {
11350 err = check_atomic(env, env->insn_idx, insn);
11357 if (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0) {
11358 verbose(env, "BPF_STX uses reserved fields\n");
11362 /* check src1 operand */
11363 err = check_reg_arg(env, insn->src_reg, SRC_OP);
11366 /* check src2 operand */
11367 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
11371 dst_reg_type = regs[insn->dst_reg].type;
11373 /* check that memory (dst_reg + off) is writeable */
11374 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
11375 insn->off, BPF_SIZE(insn->code),
11376 BPF_WRITE, insn->src_reg, false);
11380 prev_dst_type = &env->insn_aux_data[env->insn_idx].ptr_type;
11382 if (*prev_dst_type == NOT_INIT) {
11383 *prev_dst_type = dst_reg_type;
11384 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
11385 verbose(env, "same insn cannot be used with different pointers\n");
11389 } else if (class == BPF_ST) {
11390 if (BPF_MODE(insn->code) != BPF_MEM ||
11391 insn->src_reg != BPF_REG_0) {
11392 verbose(env, "BPF_ST uses reserved fields\n");
11395 /* check src operand */
11396 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
11400 if (is_ctx_reg(env, insn->dst_reg)) {
11401 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
11403 reg_type_str[reg_state(env, insn->dst_reg)->type]);
11407 /* check that memory (dst_reg + off) is writeable */
11408 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
11409 insn->off, BPF_SIZE(insn->code),
11410 BPF_WRITE, -1, false);
11414 } else if (class == BPF_JMP || class == BPF_JMP32) {
11415 u8 opcode = BPF_OP(insn->code);
11417 env->jmps_processed++;
11418 if (opcode == BPF_CALL) {
11419 if (BPF_SRC(insn->code) != BPF_K ||
11420 (insn->src_reg != BPF_PSEUDO_KFUNC_CALL
11421 && insn->off != 0) ||
11422 (insn->src_reg != BPF_REG_0 &&
11423 insn->src_reg != BPF_PSEUDO_CALL &&
11424 insn->src_reg != BPF_PSEUDO_KFUNC_CALL) ||
11425 insn->dst_reg != BPF_REG_0 ||
11426 class == BPF_JMP32) {
11427 verbose(env, "BPF_CALL uses reserved fields\n");
11431 if (env->cur_state->active_spin_lock &&
11432 (insn->src_reg == BPF_PSEUDO_CALL ||
11433 insn->imm != BPF_FUNC_spin_unlock)) {
11434 verbose(env, "function calls are not allowed while holding a lock\n");
11437 if (insn->src_reg == BPF_PSEUDO_CALL)
11438 err = check_func_call(env, insn, &env->insn_idx);
11439 else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL)
11440 err = check_kfunc_call(env, insn);
11442 err = check_helper_call(env, insn, &env->insn_idx);
11445 } else if (opcode == BPF_JA) {
11446 if (BPF_SRC(insn->code) != BPF_K ||
11448 insn->src_reg != BPF_REG_0 ||
11449 insn->dst_reg != BPF_REG_0 ||
11450 class == BPF_JMP32) {
11451 verbose(env, "BPF_JA uses reserved fields\n");
11455 env->insn_idx += insn->off + 1;
11458 } else if (opcode == BPF_EXIT) {
11459 if (BPF_SRC(insn->code) != BPF_K ||
11461 insn->src_reg != BPF_REG_0 ||
11462 insn->dst_reg != BPF_REG_0 ||
11463 class == BPF_JMP32) {
11464 verbose(env, "BPF_EXIT uses reserved fields\n");
11468 if (env->cur_state->active_spin_lock) {
11469 verbose(env, "bpf_spin_unlock is missing\n");
11473 if (state->curframe) {
11474 /* exit from nested function */
11475 err = prepare_func_exit(env, &env->insn_idx);
11478 do_print_state = true;
11482 err = check_reference_leak(env);
11486 err = check_return_code(env);
11490 update_branch_counts(env, env->cur_state);
11491 err = pop_stack(env, &prev_insn_idx,
11492 &env->insn_idx, pop_log);
11494 if (err != -ENOENT)
11498 do_print_state = true;
11502 err = check_cond_jmp_op(env, insn, &env->insn_idx);
11506 } else if (class == BPF_LD) {
11507 u8 mode = BPF_MODE(insn->code);
11509 if (mode == BPF_ABS || mode == BPF_IND) {
11510 err = check_ld_abs(env, insn);
11514 } else if (mode == BPF_IMM) {
11515 err = check_ld_imm(env, insn);
11520 sanitize_mark_insn_seen(env);
11522 verbose(env, "invalid BPF_LD mode\n");
11526 verbose(env, "unknown insn class %d\n", class);
11536 static int find_btf_percpu_datasec(struct btf *btf)
11538 const struct btf_type *t;
11543 * Both vmlinux and module each have their own ".data..percpu"
11544 * DATASECs in BTF. So for module's case, we need to skip vmlinux BTF
11545 * types to look at only module's own BTF types.
11547 n = btf_nr_types(btf);
11548 if (btf_is_module(btf))
11549 i = btf_nr_types(btf_vmlinux);
11553 for(; i < n; i++) {
11554 t = btf_type_by_id(btf, i);
11555 if (BTF_INFO_KIND(t->info) != BTF_KIND_DATASEC)
11558 tname = btf_name_by_offset(btf, t->name_off);
11559 if (!strcmp(tname, ".data..percpu"))
11566 /* replace pseudo btf_id with kernel symbol address */
11567 static int check_pseudo_btf_id(struct bpf_verifier_env *env,
11568 struct bpf_insn *insn,
11569 struct bpf_insn_aux_data *aux)
11571 const struct btf_var_secinfo *vsi;
11572 const struct btf_type *datasec;
11573 struct btf_mod_pair *btf_mod;
11574 const struct btf_type *t;
11575 const char *sym_name;
11576 bool percpu = false;
11577 u32 type, id = insn->imm;
11581 int i, btf_fd, err;
11583 btf_fd = insn[1].imm;
11585 btf = btf_get_by_fd(btf_fd);
11587 verbose(env, "invalid module BTF object FD specified.\n");
11591 if (!btf_vmlinux) {
11592 verbose(env, "kernel is missing BTF, make sure CONFIG_DEBUG_INFO_BTF=y is specified in Kconfig.\n");
11599 t = btf_type_by_id(btf, id);
11601 verbose(env, "ldimm64 insn specifies invalid btf_id %d.\n", id);
11606 if (!btf_type_is_var(t)) {
11607 verbose(env, "pseudo btf_id %d in ldimm64 isn't KIND_VAR.\n", id);
11612 sym_name = btf_name_by_offset(btf, t->name_off);
11613 addr = kallsyms_lookup_name(sym_name);
11615 verbose(env, "ldimm64 failed to find the address for kernel symbol '%s'.\n",
11621 datasec_id = find_btf_percpu_datasec(btf);
11622 if (datasec_id > 0) {
11623 datasec = btf_type_by_id(btf, datasec_id);
11624 for_each_vsi(i, datasec, vsi) {
11625 if (vsi->type == id) {
11632 insn[0].imm = (u32)addr;
11633 insn[1].imm = addr >> 32;
11636 t = btf_type_skip_modifiers(btf, type, NULL);
11638 aux->btf_var.reg_type = PTR_TO_PERCPU_BTF_ID;
11639 aux->btf_var.btf = btf;
11640 aux->btf_var.btf_id = type;
11641 } else if (!btf_type_is_struct(t)) {
11642 const struct btf_type *ret;
11646 /* resolve the type size of ksym. */
11647 ret = btf_resolve_size(btf, t, &tsize);
11649 tname = btf_name_by_offset(btf, t->name_off);
11650 verbose(env, "ldimm64 unable to resolve the size of type '%s': %ld\n",
11651 tname, PTR_ERR(ret));
11655 aux->btf_var.reg_type = PTR_TO_MEM;
11656 aux->btf_var.mem_size = tsize;
11658 aux->btf_var.reg_type = PTR_TO_BTF_ID;
11659 aux->btf_var.btf = btf;
11660 aux->btf_var.btf_id = type;
11663 /* check whether we recorded this BTF (and maybe module) already */
11664 for (i = 0; i < env->used_btf_cnt; i++) {
11665 if (env->used_btfs[i].btf == btf) {
11671 if (env->used_btf_cnt >= MAX_USED_BTFS) {
11676 btf_mod = &env->used_btfs[env->used_btf_cnt];
11677 btf_mod->btf = btf;
11678 btf_mod->module = NULL;
11680 /* if we reference variables from kernel module, bump its refcount */
11681 if (btf_is_module(btf)) {
11682 btf_mod->module = btf_try_get_module(btf);
11683 if (!btf_mod->module) {
11689 env->used_btf_cnt++;
11697 static int check_map_prealloc(struct bpf_map *map)
11699 return (map->map_type != BPF_MAP_TYPE_HASH &&
11700 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
11701 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
11702 !(map->map_flags & BPF_F_NO_PREALLOC);
11705 static bool is_tracing_prog_type(enum bpf_prog_type type)
11708 case BPF_PROG_TYPE_KPROBE:
11709 case BPF_PROG_TYPE_TRACEPOINT:
11710 case BPF_PROG_TYPE_PERF_EVENT:
11711 case BPF_PROG_TYPE_RAW_TRACEPOINT:
11718 static bool is_preallocated_map(struct bpf_map *map)
11720 if (!check_map_prealloc(map))
11722 if (map->inner_map_meta && !check_map_prealloc(map->inner_map_meta))
11727 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
11728 struct bpf_map *map,
11729 struct bpf_prog *prog)
11732 enum bpf_prog_type prog_type = resolve_prog_type(prog);
11734 * Validate that trace type programs use preallocated hash maps.
11736 * For programs attached to PERF events this is mandatory as the
11737 * perf NMI can hit any arbitrary code sequence.
11739 * All other trace types using preallocated hash maps are unsafe as
11740 * well because tracepoint or kprobes can be inside locked regions
11741 * of the memory allocator or at a place where a recursion into the
11742 * memory allocator would see inconsistent state.
11744 * On RT enabled kernels run-time allocation of all trace type
11745 * programs is strictly prohibited due to lock type constraints. On
11746 * !RT kernels it is allowed for backwards compatibility reasons for
11747 * now, but warnings are emitted so developers are made aware of
11748 * the unsafety and can fix their programs before this is enforced.
11750 if (is_tracing_prog_type(prog_type) && !is_preallocated_map(map)) {
11751 if (prog_type == BPF_PROG_TYPE_PERF_EVENT) {
11752 verbose(env, "perf_event programs can only use preallocated hash map\n");
11755 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
11756 verbose(env, "trace type programs can only use preallocated hash map\n");
11759 WARN_ONCE(1, "trace type BPF program uses run-time allocation\n");
11760 verbose(env, "trace type programs with run-time allocated hash maps are unsafe. Switch to preallocated hash maps.\n");
11763 if (map_value_has_spin_lock(map)) {
11764 if (prog_type == BPF_PROG_TYPE_SOCKET_FILTER) {
11765 verbose(env, "socket filter progs cannot use bpf_spin_lock yet\n");
11769 if (is_tracing_prog_type(prog_type)) {
11770 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
11774 if (prog->aux->sleepable) {
11775 verbose(env, "sleepable progs cannot use bpf_spin_lock yet\n");
11780 if (map_value_has_timer(map)) {
11781 if (is_tracing_prog_type(prog_type)) {
11782 verbose(env, "tracing progs cannot use bpf_timer yet\n");
11787 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
11788 !bpf_offload_prog_map_match(prog, map)) {
11789 verbose(env, "offload device mismatch between prog and map\n");
11793 if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
11794 verbose(env, "bpf_struct_ops map cannot be used in prog\n");
11798 if (prog->aux->sleepable)
11799 switch (map->map_type) {
11800 case BPF_MAP_TYPE_HASH:
11801 case BPF_MAP_TYPE_LRU_HASH:
11802 case BPF_MAP_TYPE_ARRAY:
11803 case BPF_MAP_TYPE_PERCPU_HASH:
11804 case BPF_MAP_TYPE_PERCPU_ARRAY:
11805 case BPF_MAP_TYPE_LRU_PERCPU_HASH:
11806 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
11807 case BPF_MAP_TYPE_HASH_OF_MAPS:
11808 if (!is_preallocated_map(map)) {
11810 "Sleepable programs can only use preallocated maps\n");
11814 case BPF_MAP_TYPE_RINGBUF:
11818 "Sleepable programs can only use array, hash, and ringbuf maps\n");
11825 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
11827 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
11828 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
11831 /* find and rewrite pseudo imm in ld_imm64 instructions:
11833 * 1. if it accesses map FD, replace it with actual map pointer.
11834 * 2. if it accesses btf_id of a VAR, replace it with pointer to the var.
11836 * NOTE: btf_vmlinux is required for converting pseudo btf_id.
11838 static int resolve_pseudo_ldimm64(struct bpf_verifier_env *env)
11840 struct bpf_insn *insn = env->prog->insnsi;
11841 int insn_cnt = env->prog->len;
11844 err = bpf_prog_calc_tag(env->prog);
11848 for (i = 0; i < insn_cnt; i++, insn++) {
11849 if (BPF_CLASS(insn->code) == BPF_LDX &&
11850 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
11851 verbose(env, "BPF_LDX uses reserved fields\n");
11855 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
11856 struct bpf_insn_aux_data *aux;
11857 struct bpf_map *map;
11862 if (i == insn_cnt - 1 || insn[1].code != 0 ||
11863 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
11864 insn[1].off != 0) {
11865 verbose(env, "invalid bpf_ld_imm64 insn\n");
11869 if (insn[0].src_reg == 0)
11870 /* valid generic load 64-bit imm */
11873 if (insn[0].src_reg == BPF_PSEUDO_BTF_ID) {
11874 aux = &env->insn_aux_data[i];
11875 err = check_pseudo_btf_id(env, insn, aux);
11881 if (insn[0].src_reg == BPF_PSEUDO_FUNC) {
11882 aux = &env->insn_aux_data[i];
11883 aux->ptr_type = PTR_TO_FUNC;
11887 /* In final convert_pseudo_ld_imm64() step, this is
11888 * converted into regular 64-bit imm load insn.
11890 switch (insn[0].src_reg) {
11891 case BPF_PSEUDO_MAP_VALUE:
11892 case BPF_PSEUDO_MAP_IDX_VALUE:
11894 case BPF_PSEUDO_MAP_FD:
11895 case BPF_PSEUDO_MAP_IDX:
11896 if (insn[1].imm == 0)
11900 verbose(env, "unrecognized bpf_ld_imm64 insn\n");
11904 switch (insn[0].src_reg) {
11905 case BPF_PSEUDO_MAP_IDX_VALUE:
11906 case BPF_PSEUDO_MAP_IDX:
11907 if (bpfptr_is_null(env->fd_array)) {
11908 verbose(env, "fd_idx without fd_array is invalid\n");
11911 if (copy_from_bpfptr_offset(&fd, env->fd_array,
11912 insn[0].imm * sizeof(fd),
11922 map = __bpf_map_get(f);
11924 verbose(env, "fd %d is not pointing to valid bpf_map\n",
11926 return PTR_ERR(map);
11929 err = check_map_prog_compatibility(env, map, env->prog);
11935 aux = &env->insn_aux_data[i];
11936 if (insn[0].src_reg == BPF_PSEUDO_MAP_FD ||
11937 insn[0].src_reg == BPF_PSEUDO_MAP_IDX) {
11938 addr = (unsigned long)map;
11940 u32 off = insn[1].imm;
11942 if (off >= BPF_MAX_VAR_OFF) {
11943 verbose(env, "direct value offset of %u is not allowed\n", off);
11948 if (!map->ops->map_direct_value_addr) {
11949 verbose(env, "no direct value access support for this map type\n");
11954 err = map->ops->map_direct_value_addr(map, &addr, off);
11956 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
11957 map->value_size, off);
11962 aux->map_off = off;
11966 insn[0].imm = (u32)addr;
11967 insn[1].imm = addr >> 32;
11969 /* check whether we recorded this map already */
11970 for (j = 0; j < env->used_map_cnt; j++) {
11971 if (env->used_maps[j] == map) {
11972 aux->map_index = j;
11978 if (env->used_map_cnt >= MAX_USED_MAPS) {
11983 /* hold the map. If the program is rejected by verifier,
11984 * the map will be released by release_maps() or it
11985 * will be used by the valid program until it's unloaded
11986 * and all maps are released in free_used_maps()
11990 aux->map_index = env->used_map_cnt;
11991 env->used_maps[env->used_map_cnt++] = map;
11993 if (bpf_map_is_cgroup_storage(map) &&
11994 bpf_cgroup_storage_assign(env->prog->aux, map)) {
11995 verbose(env, "only one cgroup storage of each type is allowed\n");
12007 /* Basic sanity check before we invest more work here. */
12008 if (!bpf_opcode_in_insntable(insn->code)) {
12009 verbose(env, "unknown opcode %02x\n", insn->code);
12014 /* now all pseudo BPF_LD_IMM64 instructions load valid
12015 * 'struct bpf_map *' into a register instead of user map_fd.
12016 * These pointers will be used later by verifier to validate map access.
12021 /* drop refcnt of maps used by the rejected program */
12022 static void release_maps(struct bpf_verifier_env *env)
12024 __bpf_free_used_maps(env->prog->aux, env->used_maps,
12025 env->used_map_cnt);
12028 /* drop refcnt of maps used by the rejected program */
12029 static void release_btfs(struct bpf_verifier_env *env)
12031 __bpf_free_used_btfs(env->prog->aux, env->used_btfs,
12032 env->used_btf_cnt);
12035 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
12036 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
12038 struct bpf_insn *insn = env->prog->insnsi;
12039 int insn_cnt = env->prog->len;
12042 for (i = 0; i < insn_cnt; i++, insn++) {
12043 if (insn->code != (BPF_LD | BPF_IMM | BPF_DW))
12045 if (insn->src_reg == BPF_PSEUDO_FUNC)
12051 /* single env->prog->insni[off] instruction was replaced with the range
12052 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
12053 * [0, off) and [off, end) to new locations, so the patched range stays zero
12055 static void adjust_insn_aux_data(struct bpf_verifier_env *env,
12056 struct bpf_insn_aux_data *new_data,
12057 struct bpf_prog *new_prog, u32 off, u32 cnt)
12059 struct bpf_insn_aux_data *old_data = env->insn_aux_data;
12060 struct bpf_insn *insn = new_prog->insnsi;
12061 u32 old_seen = old_data[off].seen;
12065 /* aux info at OFF always needs adjustment, no matter fast path
12066 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
12067 * original insn at old prog.
12069 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
12073 prog_len = new_prog->len;
12075 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
12076 memcpy(new_data + off + cnt - 1, old_data + off,
12077 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
12078 for (i = off; i < off + cnt - 1; i++) {
12079 /* Expand insni[off]'s seen count to the patched range. */
12080 new_data[i].seen = old_seen;
12081 new_data[i].zext_dst = insn_has_def32(env, insn + i);
12083 env->insn_aux_data = new_data;
12087 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
12093 /* NOTE: fake 'exit' subprog should be updated as well. */
12094 for (i = 0; i <= env->subprog_cnt; i++) {
12095 if (env->subprog_info[i].start <= off)
12097 env->subprog_info[i].start += len - 1;
12101 static void adjust_poke_descs(struct bpf_prog *prog, u32 off, u32 len)
12103 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
12104 int i, sz = prog->aux->size_poke_tab;
12105 struct bpf_jit_poke_descriptor *desc;
12107 for (i = 0; i < sz; i++) {
12109 if (desc->insn_idx <= off)
12111 desc->insn_idx += len - 1;
12115 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
12116 const struct bpf_insn *patch, u32 len)
12118 struct bpf_prog *new_prog;
12119 struct bpf_insn_aux_data *new_data = NULL;
12122 new_data = vzalloc(array_size(env->prog->len + len - 1,
12123 sizeof(struct bpf_insn_aux_data)));
12128 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
12129 if (IS_ERR(new_prog)) {
12130 if (PTR_ERR(new_prog) == -ERANGE)
12132 "insn %d cannot be patched due to 16-bit range\n",
12133 env->insn_aux_data[off].orig_idx);
12137 adjust_insn_aux_data(env, new_data, new_prog, off, len);
12138 adjust_subprog_starts(env, off, len);
12139 adjust_poke_descs(new_prog, off, len);
12143 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
12148 /* find first prog starting at or after off (first to remove) */
12149 for (i = 0; i < env->subprog_cnt; i++)
12150 if (env->subprog_info[i].start >= off)
12152 /* find first prog starting at or after off + cnt (first to stay) */
12153 for (j = i; j < env->subprog_cnt; j++)
12154 if (env->subprog_info[j].start >= off + cnt)
12156 /* if j doesn't start exactly at off + cnt, we are just removing
12157 * the front of previous prog
12159 if (env->subprog_info[j].start != off + cnt)
12163 struct bpf_prog_aux *aux = env->prog->aux;
12166 /* move fake 'exit' subprog as well */
12167 move = env->subprog_cnt + 1 - j;
12169 memmove(env->subprog_info + i,
12170 env->subprog_info + j,
12171 sizeof(*env->subprog_info) * move);
12172 env->subprog_cnt -= j - i;
12174 /* remove func_info */
12175 if (aux->func_info) {
12176 move = aux->func_info_cnt - j;
12178 memmove(aux->func_info + i,
12179 aux->func_info + j,
12180 sizeof(*aux->func_info) * move);
12181 aux->func_info_cnt -= j - i;
12182 /* func_info->insn_off is set after all code rewrites,
12183 * in adjust_btf_func() - no need to adjust
12187 /* convert i from "first prog to remove" to "first to adjust" */
12188 if (env->subprog_info[i].start == off)
12192 /* update fake 'exit' subprog as well */
12193 for (; i <= env->subprog_cnt; i++)
12194 env->subprog_info[i].start -= cnt;
12199 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
12202 struct bpf_prog *prog = env->prog;
12203 u32 i, l_off, l_cnt, nr_linfo;
12204 struct bpf_line_info *linfo;
12206 nr_linfo = prog->aux->nr_linfo;
12210 linfo = prog->aux->linfo;
12212 /* find first line info to remove, count lines to be removed */
12213 for (i = 0; i < nr_linfo; i++)
12214 if (linfo[i].insn_off >= off)
12219 for (; i < nr_linfo; i++)
12220 if (linfo[i].insn_off < off + cnt)
12225 /* First live insn doesn't match first live linfo, it needs to "inherit"
12226 * last removed linfo. prog is already modified, so prog->len == off
12227 * means no live instructions after (tail of the program was removed).
12229 if (prog->len != off && l_cnt &&
12230 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
12232 linfo[--i].insn_off = off + cnt;
12235 /* remove the line info which refer to the removed instructions */
12237 memmove(linfo + l_off, linfo + i,
12238 sizeof(*linfo) * (nr_linfo - i));
12240 prog->aux->nr_linfo -= l_cnt;
12241 nr_linfo = prog->aux->nr_linfo;
12244 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
12245 for (i = l_off; i < nr_linfo; i++)
12246 linfo[i].insn_off -= cnt;
12248 /* fix up all subprogs (incl. 'exit') which start >= off */
12249 for (i = 0; i <= env->subprog_cnt; i++)
12250 if (env->subprog_info[i].linfo_idx > l_off) {
12251 /* program may have started in the removed region but
12252 * may not be fully removed
12254 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
12255 env->subprog_info[i].linfo_idx -= l_cnt;
12257 env->subprog_info[i].linfo_idx = l_off;
12263 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
12265 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
12266 unsigned int orig_prog_len = env->prog->len;
12269 if (bpf_prog_is_dev_bound(env->prog->aux))
12270 bpf_prog_offload_remove_insns(env, off, cnt);
12272 err = bpf_remove_insns(env->prog, off, cnt);
12276 err = adjust_subprog_starts_after_remove(env, off, cnt);
12280 err = bpf_adj_linfo_after_remove(env, off, cnt);
12284 memmove(aux_data + off, aux_data + off + cnt,
12285 sizeof(*aux_data) * (orig_prog_len - off - cnt));
12290 /* The verifier does more data flow analysis than llvm and will not
12291 * explore branches that are dead at run time. Malicious programs can
12292 * have dead code too. Therefore replace all dead at-run-time code
12295 * Just nops are not optimal, e.g. if they would sit at the end of the
12296 * program and through another bug we would manage to jump there, then
12297 * we'd execute beyond program memory otherwise. Returning exception
12298 * code also wouldn't work since we can have subprogs where the dead
12299 * code could be located.
12301 static void sanitize_dead_code(struct bpf_verifier_env *env)
12303 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
12304 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
12305 struct bpf_insn *insn = env->prog->insnsi;
12306 const int insn_cnt = env->prog->len;
12309 for (i = 0; i < insn_cnt; i++) {
12310 if (aux_data[i].seen)
12312 memcpy(insn + i, &trap, sizeof(trap));
12313 aux_data[i].zext_dst = false;
12317 static bool insn_is_cond_jump(u8 code)
12321 if (BPF_CLASS(code) == BPF_JMP32)
12324 if (BPF_CLASS(code) != BPF_JMP)
12328 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
12331 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
12333 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
12334 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
12335 struct bpf_insn *insn = env->prog->insnsi;
12336 const int insn_cnt = env->prog->len;
12339 for (i = 0; i < insn_cnt; i++, insn++) {
12340 if (!insn_is_cond_jump(insn->code))
12343 if (!aux_data[i + 1].seen)
12344 ja.off = insn->off;
12345 else if (!aux_data[i + 1 + insn->off].seen)
12350 if (bpf_prog_is_dev_bound(env->prog->aux))
12351 bpf_prog_offload_replace_insn(env, i, &ja);
12353 memcpy(insn, &ja, sizeof(ja));
12357 static int opt_remove_dead_code(struct bpf_verifier_env *env)
12359 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
12360 int insn_cnt = env->prog->len;
12363 for (i = 0; i < insn_cnt; i++) {
12367 while (i + j < insn_cnt && !aux_data[i + j].seen)
12372 err = verifier_remove_insns(env, i, j);
12375 insn_cnt = env->prog->len;
12381 static int opt_remove_nops(struct bpf_verifier_env *env)
12383 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
12384 struct bpf_insn *insn = env->prog->insnsi;
12385 int insn_cnt = env->prog->len;
12388 for (i = 0; i < insn_cnt; i++) {
12389 if (memcmp(&insn[i], &ja, sizeof(ja)))
12392 err = verifier_remove_insns(env, i, 1);
12402 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
12403 const union bpf_attr *attr)
12405 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
12406 struct bpf_insn_aux_data *aux = env->insn_aux_data;
12407 int i, patch_len, delta = 0, len = env->prog->len;
12408 struct bpf_insn *insns = env->prog->insnsi;
12409 struct bpf_prog *new_prog;
12412 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
12413 zext_patch[1] = BPF_ZEXT_REG(0);
12414 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
12415 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
12416 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
12417 for (i = 0; i < len; i++) {
12418 int adj_idx = i + delta;
12419 struct bpf_insn insn;
12422 insn = insns[adj_idx];
12423 load_reg = insn_def_regno(&insn);
12424 if (!aux[adj_idx].zext_dst) {
12432 class = BPF_CLASS(code);
12433 if (load_reg == -1)
12436 /* NOTE: arg "reg" (the fourth one) is only used for
12437 * BPF_STX + SRC_OP, so it is safe to pass NULL
12440 if (is_reg64(env, &insn, load_reg, NULL, DST_OP)) {
12441 if (class == BPF_LD &&
12442 BPF_MODE(code) == BPF_IMM)
12447 /* ctx load could be transformed into wider load. */
12448 if (class == BPF_LDX &&
12449 aux[adj_idx].ptr_type == PTR_TO_CTX)
12452 imm_rnd = get_random_int();
12453 rnd_hi32_patch[0] = insn;
12454 rnd_hi32_patch[1].imm = imm_rnd;
12455 rnd_hi32_patch[3].dst_reg = load_reg;
12456 patch = rnd_hi32_patch;
12458 goto apply_patch_buffer;
12461 /* Add in an zero-extend instruction if a) the JIT has requested
12462 * it or b) it's a CMPXCHG.
12464 * The latter is because: BPF_CMPXCHG always loads a value into
12465 * R0, therefore always zero-extends. However some archs'
12466 * equivalent instruction only does this load when the
12467 * comparison is successful. This detail of CMPXCHG is
12468 * orthogonal to the general zero-extension behaviour of the
12469 * CPU, so it's treated independently of bpf_jit_needs_zext.
12471 if (!bpf_jit_needs_zext() && !is_cmpxchg_insn(&insn))
12474 if (WARN_ON(load_reg == -1)) {
12475 verbose(env, "verifier bug. zext_dst is set, but no reg is defined\n");
12479 zext_patch[0] = insn;
12480 zext_patch[1].dst_reg = load_reg;
12481 zext_patch[1].src_reg = load_reg;
12482 patch = zext_patch;
12484 apply_patch_buffer:
12485 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
12488 env->prog = new_prog;
12489 insns = new_prog->insnsi;
12490 aux = env->insn_aux_data;
12491 delta += patch_len - 1;
12497 /* convert load instructions that access fields of a context type into a
12498 * sequence of instructions that access fields of the underlying structure:
12499 * struct __sk_buff -> struct sk_buff
12500 * struct bpf_sock_ops -> struct sock
12502 static int convert_ctx_accesses(struct bpf_verifier_env *env)
12504 const struct bpf_verifier_ops *ops = env->ops;
12505 int i, cnt, size, ctx_field_size, delta = 0;
12506 const int insn_cnt = env->prog->len;
12507 struct bpf_insn insn_buf[16], *insn;
12508 u32 target_size, size_default, off;
12509 struct bpf_prog *new_prog;
12510 enum bpf_access_type type;
12511 bool is_narrower_load;
12513 if (ops->gen_prologue || env->seen_direct_write) {
12514 if (!ops->gen_prologue) {
12515 verbose(env, "bpf verifier is misconfigured\n");
12518 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
12520 if (cnt >= ARRAY_SIZE(insn_buf)) {
12521 verbose(env, "bpf verifier is misconfigured\n");
12524 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
12528 env->prog = new_prog;
12533 if (bpf_prog_is_dev_bound(env->prog->aux))
12536 insn = env->prog->insnsi + delta;
12538 for (i = 0; i < insn_cnt; i++, insn++) {
12539 bpf_convert_ctx_access_t convert_ctx_access;
12542 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
12543 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
12544 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
12545 insn->code == (BPF_LDX | BPF_MEM | BPF_DW)) {
12548 } else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
12549 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
12550 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
12551 insn->code == (BPF_STX | BPF_MEM | BPF_DW) ||
12552 insn->code == (BPF_ST | BPF_MEM | BPF_B) ||
12553 insn->code == (BPF_ST | BPF_MEM | BPF_H) ||
12554 insn->code == (BPF_ST | BPF_MEM | BPF_W) ||
12555 insn->code == (BPF_ST | BPF_MEM | BPF_DW)) {
12557 ctx_access = BPF_CLASS(insn->code) == BPF_STX;
12562 if (type == BPF_WRITE &&
12563 env->insn_aux_data[i + delta].sanitize_stack_spill) {
12564 struct bpf_insn patch[] = {
12569 cnt = ARRAY_SIZE(patch);
12570 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
12575 env->prog = new_prog;
12576 insn = new_prog->insnsi + i + delta;
12583 switch (env->insn_aux_data[i + delta].ptr_type) {
12585 if (!ops->convert_ctx_access)
12587 convert_ctx_access = ops->convert_ctx_access;
12589 case PTR_TO_SOCKET:
12590 case PTR_TO_SOCK_COMMON:
12591 convert_ctx_access = bpf_sock_convert_ctx_access;
12593 case PTR_TO_TCP_SOCK:
12594 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
12596 case PTR_TO_XDP_SOCK:
12597 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
12599 case PTR_TO_BTF_ID:
12600 if (type == BPF_READ) {
12601 insn->code = BPF_LDX | BPF_PROBE_MEM |
12602 BPF_SIZE((insn)->code);
12603 env->prog->aux->num_exentries++;
12604 } else if (resolve_prog_type(env->prog) != BPF_PROG_TYPE_STRUCT_OPS) {
12605 verbose(env, "Writes through BTF pointers are not allowed\n");
12613 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
12614 size = BPF_LDST_BYTES(insn);
12616 /* If the read access is a narrower load of the field,
12617 * convert to a 4/8-byte load, to minimum program type specific
12618 * convert_ctx_access changes. If conversion is successful,
12619 * we will apply proper mask to the result.
12621 is_narrower_load = size < ctx_field_size;
12622 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
12624 if (is_narrower_load) {
12627 if (type == BPF_WRITE) {
12628 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
12633 if (ctx_field_size == 4)
12635 else if (ctx_field_size == 8)
12636 size_code = BPF_DW;
12638 insn->off = off & ~(size_default - 1);
12639 insn->code = BPF_LDX | BPF_MEM | size_code;
12643 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
12645 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
12646 (ctx_field_size && !target_size)) {
12647 verbose(env, "bpf verifier is misconfigured\n");
12651 if (is_narrower_load && size < target_size) {
12652 u8 shift = bpf_ctx_narrow_access_offset(
12653 off, size, size_default) * 8;
12654 if (shift && cnt + 1 >= ARRAY_SIZE(insn_buf)) {
12655 verbose(env, "bpf verifier narrow ctx load misconfigured\n");
12658 if (ctx_field_size <= 4) {
12660 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
12663 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
12664 (1 << size * 8) - 1);
12667 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
12670 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
12671 (1ULL << size * 8) - 1);
12675 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12681 /* keep walking new program and skip insns we just inserted */
12682 env->prog = new_prog;
12683 insn = new_prog->insnsi + i + delta;
12689 static int jit_subprogs(struct bpf_verifier_env *env)
12691 struct bpf_prog *prog = env->prog, **func, *tmp;
12692 int i, j, subprog_start, subprog_end = 0, len, subprog;
12693 struct bpf_map *map_ptr;
12694 struct bpf_insn *insn;
12695 void *old_bpf_func;
12696 int err, num_exentries;
12698 if (env->subprog_cnt <= 1)
12701 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
12702 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn))
12705 /* Upon error here we cannot fall back to interpreter but
12706 * need a hard reject of the program. Thus -EFAULT is
12707 * propagated in any case.
12709 subprog = find_subprog(env, i + insn->imm + 1);
12711 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
12712 i + insn->imm + 1);
12715 /* temporarily remember subprog id inside insn instead of
12716 * aux_data, since next loop will split up all insns into funcs
12718 insn->off = subprog;
12719 /* remember original imm in case JIT fails and fallback
12720 * to interpreter will be needed
12722 env->insn_aux_data[i].call_imm = insn->imm;
12723 /* point imm to __bpf_call_base+1 from JITs point of view */
12725 if (bpf_pseudo_func(insn))
12726 /* jit (e.g. x86_64) may emit fewer instructions
12727 * if it learns a u32 imm is the same as a u64 imm.
12728 * Force a non zero here.
12733 err = bpf_prog_alloc_jited_linfo(prog);
12735 goto out_undo_insn;
12738 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
12740 goto out_undo_insn;
12742 for (i = 0; i < env->subprog_cnt; i++) {
12743 subprog_start = subprog_end;
12744 subprog_end = env->subprog_info[i + 1].start;
12746 len = subprog_end - subprog_start;
12747 /* bpf_prog_run() doesn't call subprogs directly,
12748 * hence main prog stats include the runtime of subprogs.
12749 * subprogs don't have IDs and not reachable via prog_get_next_id
12750 * func[i]->stats will never be accessed and stays NULL
12752 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
12755 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
12756 len * sizeof(struct bpf_insn));
12757 func[i]->type = prog->type;
12758 func[i]->len = len;
12759 if (bpf_prog_calc_tag(func[i]))
12761 func[i]->is_func = 1;
12762 func[i]->aux->func_idx = i;
12763 /* Below members will be freed only at prog->aux */
12764 func[i]->aux->btf = prog->aux->btf;
12765 func[i]->aux->func_info = prog->aux->func_info;
12766 func[i]->aux->poke_tab = prog->aux->poke_tab;
12767 func[i]->aux->size_poke_tab = prog->aux->size_poke_tab;
12769 for (j = 0; j < prog->aux->size_poke_tab; j++) {
12770 struct bpf_jit_poke_descriptor *poke;
12772 poke = &prog->aux->poke_tab[j];
12773 if (poke->insn_idx < subprog_end &&
12774 poke->insn_idx >= subprog_start)
12775 poke->aux = func[i]->aux;
12778 /* Use bpf_prog_F_tag to indicate functions in stack traces.
12779 * Long term would need debug info to populate names
12781 func[i]->aux->name[0] = 'F';
12782 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
12783 func[i]->jit_requested = 1;
12784 func[i]->aux->kfunc_tab = prog->aux->kfunc_tab;
12785 func[i]->aux->kfunc_btf_tab = prog->aux->kfunc_btf_tab;
12786 func[i]->aux->linfo = prog->aux->linfo;
12787 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
12788 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
12789 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
12791 insn = func[i]->insnsi;
12792 for (j = 0; j < func[i]->len; j++, insn++) {
12793 if (BPF_CLASS(insn->code) == BPF_LDX &&
12794 BPF_MODE(insn->code) == BPF_PROBE_MEM)
12797 func[i]->aux->num_exentries = num_exentries;
12798 func[i]->aux->tail_call_reachable = env->subprog_info[i].tail_call_reachable;
12799 func[i] = bpf_int_jit_compile(func[i]);
12800 if (!func[i]->jited) {
12807 /* at this point all bpf functions were successfully JITed
12808 * now populate all bpf_calls with correct addresses and
12809 * run last pass of JIT
12811 for (i = 0; i < env->subprog_cnt; i++) {
12812 insn = func[i]->insnsi;
12813 for (j = 0; j < func[i]->len; j++, insn++) {
12814 if (bpf_pseudo_func(insn)) {
12815 subprog = insn->off;
12816 insn[0].imm = (u32)(long)func[subprog]->bpf_func;
12817 insn[1].imm = ((u64)(long)func[subprog]->bpf_func) >> 32;
12820 if (!bpf_pseudo_call(insn))
12822 subprog = insn->off;
12823 insn->imm = BPF_CALL_IMM(func[subprog]->bpf_func);
12826 /* we use the aux data to keep a list of the start addresses
12827 * of the JITed images for each function in the program
12829 * for some architectures, such as powerpc64, the imm field
12830 * might not be large enough to hold the offset of the start
12831 * address of the callee's JITed image from __bpf_call_base
12833 * in such cases, we can lookup the start address of a callee
12834 * by using its subprog id, available from the off field of
12835 * the call instruction, as an index for this list
12837 func[i]->aux->func = func;
12838 func[i]->aux->func_cnt = env->subprog_cnt;
12840 for (i = 0; i < env->subprog_cnt; i++) {
12841 old_bpf_func = func[i]->bpf_func;
12842 tmp = bpf_int_jit_compile(func[i]);
12843 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
12844 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
12851 /* finally lock prog and jit images for all functions and
12852 * populate kallsysm
12854 for (i = 0; i < env->subprog_cnt; i++) {
12855 bpf_prog_lock_ro(func[i]);
12856 bpf_prog_kallsyms_add(func[i]);
12859 /* Last step: make now unused interpreter insns from main
12860 * prog consistent for later dump requests, so they can
12861 * later look the same as if they were interpreted only.
12863 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
12864 if (bpf_pseudo_func(insn)) {
12865 insn[0].imm = env->insn_aux_data[i].call_imm;
12866 insn[1].imm = insn->off;
12870 if (!bpf_pseudo_call(insn))
12872 insn->off = env->insn_aux_data[i].call_imm;
12873 subprog = find_subprog(env, i + insn->off + 1);
12874 insn->imm = subprog;
12878 prog->bpf_func = func[0]->bpf_func;
12879 prog->aux->func = func;
12880 prog->aux->func_cnt = env->subprog_cnt;
12881 bpf_prog_jit_attempt_done(prog);
12884 /* We failed JIT'ing, so at this point we need to unregister poke
12885 * descriptors from subprogs, so that kernel is not attempting to
12886 * patch it anymore as we're freeing the subprog JIT memory.
12888 for (i = 0; i < prog->aux->size_poke_tab; i++) {
12889 map_ptr = prog->aux->poke_tab[i].tail_call.map;
12890 map_ptr->ops->map_poke_untrack(map_ptr, prog->aux);
12892 /* At this point we're guaranteed that poke descriptors are not
12893 * live anymore. We can just unlink its descriptor table as it's
12894 * released with the main prog.
12896 for (i = 0; i < env->subprog_cnt; i++) {
12899 func[i]->aux->poke_tab = NULL;
12900 bpf_jit_free(func[i]);
12904 /* cleanup main prog to be interpreted */
12905 prog->jit_requested = 0;
12906 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
12907 if (!bpf_pseudo_call(insn))
12910 insn->imm = env->insn_aux_data[i].call_imm;
12912 bpf_prog_jit_attempt_done(prog);
12916 static int fixup_call_args(struct bpf_verifier_env *env)
12918 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
12919 struct bpf_prog *prog = env->prog;
12920 struct bpf_insn *insn = prog->insnsi;
12921 bool has_kfunc_call = bpf_prog_has_kfunc_call(prog);
12926 if (env->prog->jit_requested &&
12927 !bpf_prog_is_dev_bound(env->prog->aux)) {
12928 err = jit_subprogs(env);
12931 if (err == -EFAULT)
12934 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
12935 if (has_kfunc_call) {
12936 verbose(env, "calling kernel functions are not allowed in non-JITed programs\n");
12939 if (env->subprog_cnt > 1 && env->prog->aux->tail_call_reachable) {
12940 /* When JIT fails the progs with bpf2bpf calls and tail_calls
12941 * have to be rejected, since interpreter doesn't support them yet.
12943 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
12946 for (i = 0; i < prog->len; i++, insn++) {
12947 if (bpf_pseudo_func(insn)) {
12948 /* When JIT fails the progs with callback calls
12949 * have to be rejected, since interpreter doesn't support them yet.
12951 verbose(env, "callbacks are not allowed in non-JITed programs\n");
12955 if (!bpf_pseudo_call(insn))
12957 depth = get_callee_stack_depth(env, insn, i);
12960 bpf_patch_call_args(insn, depth);
12967 static int fixup_kfunc_call(struct bpf_verifier_env *env,
12968 struct bpf_insn *insn)
12970 const struct bpf_kfunc_desc *desc;
12973 verbose(env, "invalid kernel function call not eliminated in verifier pass\n");
12977 /* insn->imm has the btf func_id. Replace it with
12978 * an address (relative to __bpf_base_call).
12980 desc = find_kfunc_desc(env->prog, insn->imm, insn->off);
12982 verbose(env, "verifier internal error: kernel function descriptor not found for func_id %u\n",
12987 insn->imm = desc->imm;
12992 /* Do various post-verification rewrites in a single program pass.
12993 * These rewrites simplify JIT and interpreter implementations.
12995 static int do_misc_fixups(struct bpf_verifier_env *env)
12997 struct bpf_prog *prog = env->prog;
12998 enum bpf_attach_type eatype = prog->expected_attach_type;
12999 bool expect_blinding = bpf_jit_blinding_enabled(prog);
13000 enum bpf_prog_type prog_type = resolve_prog_type(prog);
13001 struct bpf_insn *insn = prog->insnsi;
13002 const struct bpf_func_proto *fn;
13003 const int insn_cnt = prog->len;
13004 const struct bpf_map_ops *ops;
13005 struct bpf_insn_aux_data *aux;
13006 struct bpf_insn insn_buf[16];
13007 struct bpf_prog *new_prog;
13008 struct bpf_map *map_ptr;
13009 int i, ret, cnt, delta = 0;
13011 for (i = 0; i < insn_cnt; i++, insn++) {
13012 /* Make divide-by-zero exceptions impossible. */
13013 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
13014 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
13015 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
13016 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
13017 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
13018 bool isdiv = BPF_OP(insn->code) == BPF_DIV;
13019 struct bpf_insn *patchlet;
13020 struct bpf_insn chk_and_div[] = {
13021 /* [R,W]x div 0 -> 0 */
13022 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
13023 BPF_JNE | BPF_K, insn->src_reg,
13025 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
13026 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
13029 struct bpf_insn chk_and_mod[] = {
13030 /* [R,W]x mod 0 -> [R,W]x */
13031 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
13032 BPF_JEQ | BPF_K, insn->src_reg,
13033 0, 1 + (is64 ? 0 : 1), 0),
13035 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
13036 BPF_MOV32_REG(insn->dst_reg, insn->dst_reg),
13039 patchlet = isdiv ? chk_and_div : chk_and_mod;
13040 cnt = isdiv ? ARRAY_SIZE(chk_and_div) :
13041 ARRAY_SIZE(chk_and_mod) - (is64 ? 2 : 0);
13043 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
13048 env->prog = prog = new_prog;
13049 insn = new_prog->insnsi + i + delta;
13053 /* Implement LD_ABS and LD_IND with a rewrite, if supported by the program type. */
13054 if (BPF_CLASS(insn->code) == BPF_LD &&
13055 (BPF_MODE(insn->code) == BPF_ABS ||
13056 BPF_MODE(insn->code) == BPF_IND)) {
13057 cnt = env->ops->gen_ld_abs(insn, insn_buf);
13058 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
13059 verbose(env, "bpf verifier is misconfigured\n");
13063 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
13068 env->prog = prog = new_prog;
13069 insn = new_prog->insnsi + i + delta;
13073 /* Rewrite pointer arithmetic to mitigate speculation attacks. */
13074 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
13075 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
13076 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
13077 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
13078 struct bpf_insn *patch = &insn_buf[0];
13079 bool issrc, isneg, isimm;
13082 aux = &env->insn_aux_data[i + delta];
13083 if (!aux->alu_state ||
13084 aux->alu_state == BPF_ALU_NON_POINTER)
13087 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
13088 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
13089 BPF_ALU_SANITIZE_SRC;
13090 isimm = aux->alu_state & BPF_ALU_IMMEDIATE;
13092 off_reg = issrc ? insn->src_reg : insn->dst_reg;
13094 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
13097 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
13098 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
13099 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
13100 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
13101 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
13102 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
13103 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX, off_reg);
13106 *patch++ = BPF_MOV64_REG(insn->dst_reg, insn->src_reg);
13107 insn->src_reg = BPF_REG_AX;
13109 insn->code = insn->code == code_add ?
13110 code_sub : code_add;
13112 if (issrc && isneg && !isimm)
13113 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
13114 cnt = patch - insn_buf;
13116 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
13121 env->prog = prog = new_prog;
13122 insn = new_prog->insnsi + i + delta;
13126 if (insn->code != (BPF_JMP | BPF_CALL))
13128 if (insn->src_reg == BPF_PSEUDO_CALL)
13130 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
13131 ret = fixup_kfunc_call(env, insn);
13137 if (insn->imm == BPF_FUNC_get_route_realm)
13138 prog->dst_needed = 1;
13139 if (insn->imm == BPF_FUNC_get_prandom_u32)
13140 bpf_user_rnd_init_once();
13141 if (insn->imm == BPF_FUNC_override_return)
13142 prog->kprobe_override = 1;
13143 if (insn->imm == BPF_FUNC_tail_call) {
13144 /* If we tail call into other programs, we
13145 * cannot make any assumptions since they can
13146 * be replaced dynamically during runtime in
13147 * the program array.
13149 prog->cb_access = 1;
13150 if (!allow_tail_call_in_subprogs(env))
13151 prog->aux->stack_depth = MAX_BPF_STACK;
13152 prog->aux->max_pkt_offset = MAX_PACKET_OFF;
13154 /* mark bpf_tail_call as different opcode to avoid
13155 * conditional branch in the interpreter for every normal
13156 * call and to prevent accidental JITing by JIT compiler
13157 * that doesn't support bpf_tail_call yet
13160 insn->code = BPF_JMP | BPF_TAIL_CALL;
13162 aux = &env->insn_aux_data[i + delta];
13163 if (env->bpf_capable && !expect_blinding &&
13164 prog->jit_requested &&
13165 !bpf_map_key_poisoned(aux) &&
13166 !bpf_map_ptr_poisoned(aux) &&
13167 !bpf_map_ptr_unpriv(aux)) {
13168 struct bpf_jit_poke_descriptor desc = {
13169 .reason = BPF_POKE_REASON_TAIL_CALL,
13170 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
13171 .tail_call.key = bpf_map_key_immediate(aux),
13172 .insn_idx = i + delta,
13175 ret = bpf_jit_add_poke_descriptor(prog, &desc);
13177 verbose(env, "adding tail call poke descriptor failed\n");
13181 insn->imm = ret + 1;
13185 if (!bpf_map_ptr_unpriv(aux))
13188 /* instead of changing every JIT dealing with tail_call
13189 * emit two extra insns:
13190 * if (index >= max_entries) goto out;
13191 * index &= array->index_mask;
13192 * to avoid out-of-bounds cpu speculation
13194 if (bpf_map_ptr_poisoned(aux)) {
13195 verbose(env, "tail_call abusing map_ptr\n");
13199 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
13200 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
13201 map_ptr->max_entries, 2);
13202 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
13203 container_of(map_ptr,
13206 insn_buf[2] = *insn;
13208 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
13213 env->prog = prog = new_prog;
13214 insn = new_prog->insnsi + i + delta;
13218 if (insn->imm == BPF_FUNC_timer_set_callback) {
13219 /* The verifier will process callback_fn as many times as necessary
13220 * with different maps and the register states prepared by
13221 * set_timer_callback_state will be accurate.
13223 * The following use case is valid:
13224 * map1 is shared by prog1, prog2, prog3.
13225 * prog1 calls bpf_timer_init for some map1 elements
13226 * prog2 calls bpf_timer_set_callback for some map1 elements.
13227 * Those that were not bpf_timer_init-ed will return -EINVAL.
13228 * prog3 calls bpf_timer_start for some map1 elements.
13229 * Those that were not both bpf_timer_init-ed and
13230 * bpf_timer_set_callback-ed will return -EINVAL.
13232 struct bpf_insn ld_addrs[2] = {
13233 BPF_LD_IMM64(BPF_REG_3, (long)prog->aux),
13236 insn_buf[0] = ld_addrs[0];
13237 insn_buf[1] = ld_addrs[1];
13238 insn_buf[2] = *insn;
13241 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
13246 env->prog = prog = new_prog;
13247 insn = new_prog->insnsi + i + delta;
13248 goto patch_call_imm;
13251 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
13252 * and other inlining handlers are currently limited to 64 bit
13255 if (prog->jit_requested && BITS_PER_LONG == 64 &&
13256 (insn->imm == BPF_FUNC_map_lookup_elem ||
13257 insn->imm == BPF_FUNC_map_update_elem ||
13258 insn->imm == BPF_FUNC_map_delete_elem ||
13259 insn->imm == BPF_FUNC_map_push_elem ||
13260 insn->imm == BPF_FUNC_map_pop_elem ||
13261 insn->imm == BPF_FUNC_map_peek_elem ||
13262 insn->imm == BPF_FUNC_redirect_map ||
13263 insn->imm == BPF_FUNC_for_each_map_elem)) {
13264 aux = &env->insn_aux_data[i + delta];
13265 if (bpf_map_ptr_poisoned(aux))
13266 goto patch_call_imm;
13268 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
13269 ops = map_ptr->ops;
13270 if (insn->imm == BPF_FUNC_map_lookup_elem &&
13271 ops->map_gen_lookup) {
13272 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
13273 if (cnt == -EOPNOTSUPP)
13274 goto patch_map_ops_generic;
13275 if (cnt <= 0 || cnt >= ARRAY_SIZE(insn_buf)) {
13276 verbose(env, "bpf verifier is misconfigured\n");
13280 new_prog = bpf_patch_insn_data(env, i + delta,
13286 env->prog = prog = new_prog;
13287 insn = new_prog->insnsi + i + delta;
13291 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
13292 (void *(*)(struct bpf_map *map, void *key))NULL));
13293 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
13294 (int (*)(struct bpf_map *map, void *key))NULL));
13295 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
13296 (int (*)(struct bpf_map *map, void *key, void *value,
13298 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
13299 (int (*)(struct bpf_map *map, void *value,
13301 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
13302 (int (*)(struct bpf_map *map, void *value))NULL));
13303 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
13304 (int (*)(struct bpf_map *map, void *value))NULL));
13305 BUILD_BUG_ON(!__same_type(ops->map_redirect,
13306 (int (*)(struct bpf_map *map, u32 ifindex, u64 flags))NULL));
13307 BUILD_BUG_ON(!__same_type(ops->map_for_each_callback,
13308 (int (*)(struct bpf_map *map,
13309 bpf_callback_t callback_fn,
13310 void *callback_ctx,
13313 patch_map_ops_generic:
13314 switch (insn->imm) {
13315 case BPF_FUNC_map_lookup_elem:
13316 insn->imm = BPF_CALL_IMM(ops->map_lookup_elem);
13318 case BPF_FUNC_map_update_elem:
13319 insn->imm = BPF_CALL_IMM(ops->map_update_elem);
13321 case BPF_FUNC_map_delete_elem:
13322 insn->imm = BPF_CALL_IMM(ops->map_delete_elem);
13324 case BPF_FUNC_map_push_elem:
13325 insn->imm = BPF_CALL_IMM(ops->map_push_elem);
13327 case BPF_FUNC_map_pop_elem:
13328 insn->imm = BPF_CALL_IMM(ops->map_pop_elem);
13330 case BPF_FUNC_map_peek_elem:
13331 insn->imm = BPF_CALL_IMM(ops->map_peek_elem);
13333 case BPF_FUNC_redirect_map:
13334 insn->imm = BPF_CALL_IMM(ops->map_redirect);
13336 case BPF_FUNC_for_each_map_elem:
13337 insn->imm = BPF_CALL_IMM(ops->map_for_each_callback);
13341 goto patch_call_imm;
13344 /* Implement bpf_jiffies64 inline. */
13345 if (prog->jit_requested && BITS_PER_LONG == 64 &&
13346 insn->imm == BPF_FUNC_jiffies64) {
13347 struct bpf_insn ld_jiffies_addr[2] = {
13348 BPF_LD_IMM64(BPF_REG_0,
13349 (unsigned long)&jiffies),
13352 insn_buf[0] = ld_jiffies_addr[0];
13353 insn_buf[1] = ld_jiffies_addr[1];
13354 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
13358 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
13364 env->prog = prog = new_prog;
13365 insn = new_prog->insnsi + i + delta;
13369 /* Implement bpf_get_func_arg inline. */
13370 if (prog_type == BPF_PROG_TYPE_TRACING &&
13371 insn->imm == BPF_FUNC_get_func_arg) {
13372 /* Load nr_args from ctx - 8 */
13373 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
13374 insn_buf[1] = BPF_JMP32_REG(BPF_JGE, BPF_REG_2, BPF_REG_0, 6);
13375 insn_buf[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_2, 3);
13376 insn_buf[3] = BPF_ALU64_REG(BPF_ADD, BPF_REG_2, BPF_REG_1);
13377 insn_buf[4] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_2, 0);
13378 insn_buf[5] = BPF_STX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
13379 insn_buf[6] = BPF_MOV64_IMM(BPF_REG_0, 0);
13380 insn_buf[7] = BPF_JMP_A(1);
13381 insn_buf[8] = BPF_MOV64_IMM(BPF_REG_0, -EINVAL);
13384 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
13389 env->prog = prog = new_prog;
13390 insn = new_prog->insnsi + i + delta;
13394 /* Implement bpf_get_func_ret inline. */
13395 if (prog_type == BPF_PROG_TYPE_TRACING &&
13396 insn->imm == BPF_FUNC_get_func_ret) {
13397 if (eatype == BPF_TRACE_FEXIT ||
13398 eatype == BPF_MODIFY_RETURN) {
13399 /* Load nr_args from ctx - 8 */
13400 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
13401 insn_buf[1] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_0, 3);
13402 insn_buf[2] = BPF_ALU64_REG(BPF_ADD, BPF_REG_0, BPF_REG_1);
13403 insn_buf[3] = BPF_LDX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
13404 insn_buf[4] = BPF_STX_MEM(BPF_DW, BPF_REG_2, BPF_REG_3, 0);
13405 insn_buf[5] = BPF_MOV64_IMM(BPF_REG_0, 0);
13408 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_0, -EOPNOTSUPP);
13412 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
13417 env->prog = prog = new_prog;
13418 insn = new_prog->insnsi + i + delta;
13422 /* Implement get_func_arg_cnt inline. */
13423 if (prog_type == BPF_PROG_TYPE_TRACING &&
13424 insn->imm == BPF_FUNC_get_func_arg_cnt) {
13425 /* Load nr_args from ctx - 8 */
13426 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
13428 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
13432 env->prog = prog = new_prog;
13433 insn = new_prog->insnsi + i + delta;
13437 /* Implement bpf_get_func_ip inline. */
13438 if (prog_type == BPF_PROG_TYPE_TRACING &&
13439 insn->imm == BPF_FUNC_get_func_ip) {
13440 /* Load IP address from ctx - 16 */
13441 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -16);
13443 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
13447 env->prog = prog = new_prog;
13448 insn = new_prog->insnsi + i + delta;
13453 fn = env->ops->get_func_proto(insn->imm, env->prog);
13454 /* all functions that have prototype and verifier allowed
13455 * programs to call them, must be real in-kernel functions
13459 "kernel subsystem misconfigured func %s#%d\n",
13460 func_id_name(insn->imm), insn->imm);
13463 insn->imm = fn->func - __bpf_call_base;
13466 /* Since poke tab is now finalized, publish aux to tracker. */
13467 for (i = 0; i < prog->aux->size_poke_tab; i++) {
13468 map_ptr = prog->aux->poke_tab[i].tail_call.map;
13469 if (!map_ptr->ops->map_poke_track ||
13470 !map_ptr->ops->map_poke_untrack ||
13471 !map_ptr->ops->map_poke_run) {
13472 verbose(env, "bpf verifier is misconfigured\n");
13476 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
13478 verbose(env, "tracking tail call prog failed\n");
13483 sort_kfunc_descs_by_imm(env->prog);
13488 static void free_states(struct bpf_verifier_env *env)
13490 struct bpf_verifier_state_list *sl, *sln;
13493 sl = env->free_list;
13496 free_verifier_state(&sl->state, false);
13500 env->free_list = NULL;
13502 if (!env->explored_states)
13505 for (i = 0; i < state_htab_size(env); i++) {
13506 sl = env->explored_states[i];
13510 free_verifier_state(&sl->state, false);
13514 env->explored_states[i] = NULL;
13518 static int do_check_common(struct bpf_verifier_env *env, int subprog)
13520 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
13521 struct bpf_verifier_state *state;
13522 struct bpf_reg_state *regs;
13525 env->prev_linfo = NULL;
13528 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
13531 state->curframe = 0;
13532 state->speculative = false;
13533 state->branches = 1;
13534 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
13535 if (!state->frame[0]) {
13539 env->cur_state = state;
13540 init_func_state(env, state->frame[0],
13541 BPF_MAIN_FUNC /* callsite */,
13545 regs = state->frame[state->curframe]->regs;
13546 if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
13547 ret = btf_prepare_func_args(env, subprog, regs);
13550 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
13551 if (regs[i].type == PTR_TO_CTX)
13552 mark_reg_known_zero(env, regs, i);
13553 else if (regs[i].type == SCALAR_VALUE)
13554 mark_reg_unknown(env, regs, i);
13555 else if (regs[i].type == PTR_TO_MEM_OR_NULL) {
13556 const u32 mem_size = regs[i].mem_size;
13558 mark_reg_known_zero(env, regs, i);
13559 regs[i].mem_size = mem_size;
13560 regs[i].id = ++env->id_gen;
13564 /* 1st arg to a function */
13565 regs[BPF_REG_1].type = PTR_TO_CTX;
13566 mark_reg_known_zero(env, regs, BPF_REG_1);
13567 ret = btf_check_subprog_arg_match(env, subprog, regs);
13568 if (ret == -EFAULT)
13569 /* unlikely verifier bug. abort.
13570 * ret == 0 and ret < 0 are sadly acceptable for
13571 * main() function due to backward compatibility.
13572 * Like socket filter program may be written as:
13573 * int bpf_prog(struct pt_regs *ctx)
13574 * and never dereference that ctx in the program.
13575 * 'struct pt_regs' is a type mismatch for socket
13576 * filter that should be using 'struct __sk_buff'.
13581 ret = do_check(env);
13583 /* check for NULL is necessary, since cur_state can be freed inside
13584 * do_check() under memory pressure.
13586 if (env->cur_state) {
13587 free_verifier_state(env->cur_state, true);
13588 env->cur_state = NULL;
13590 while (!pop_stack(env, NULL, NULL, false));
13591 if (!ret && pop_log)
13592 bpf_vlog_reset(&env->log, 0);
13597 /* Verify all global functions in a BPF program one by one based on their BTF.
13598 * All global functions must pass verification. Otherwise the whole program is rejected.
13609 * foo() will be verified first for R1=any_scalar_value. During verification it
13610 * will be assumed that bar() already verified successfully and call to bar()
13611 * from foo() will be checked for type match only. Later bar() will be verified
13612 * independently to check that it's safe for R1=any_scalar_value.
13614 static int do_check_subprogs(struct bpf_verifier_env *env)
13616 struct bpf_prog_aux *aux = env->prog->aux;
13619 if (!aux->func_info)
13622 for (i = 1; i < env->subprog_cnt; i++) {
13623 if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL)
13625 env->insn_idx = env->subprog_info[i].start;
13626 WARN_ON_ONCE(env->insn_idx == 0);
13627 ret = do_check_common(env, i);
13630 } else if (env->log.level & BPF_LOG_LEVEL) {
13632 "Func#%d is safe for any args that match its prototype\n",
13639 static int do_check_main(struct bpf_verifier_env *env)
13644 ret = do_check_common(env, 0);
13646 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
13651 static void print_verification_stats(struct bpf_verifier_env *env)
13655 if (env->log.level & BPF_LOG_STATS) {
13656 verbose(env, "verification time %lld usec\n",
13657 div_u64(env->verification_time, 1000));
13658 verbose(env, "stack depth ");
13659 for (i = 0; i < env->subprog_cnt; i++) {
13660 u32 depth = env->subprog_info[i].stack_depth;
13662 verbose(env, "%d", depth);
13663 if (i + 1 < env->subprog_cnt)
13666 verbose(env, "\n");
13668 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
13669 "total_states %d peak_states %d mark_read %d\n",
13670 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
13671 env->max_states_per_insn, env->total_states,
13672 env->peak_states, env->longest_mark_read_walk);
13675 static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
13677 const struct btf_type *t, *func_proto;
13678 const struct bpf_struct_ops *st_ops;
13679 const struct btf_member *member;
13680 struct bpf_prog *prog = env->prog;
13681 u32 btf_id, member_idx;
13684 if (!prog->gpl_compatible) {
13685 verbose(env, "struct ops programs must have a GPL compatible license\n");
13689 btf_id = prog->aux->attach_btf_id;
13690 st_ops = bpf_struct_ops_find(btf_id);
13692 verbose(env, "attach_btf_id %u is not a supported struct\n",
13698 member_idx = prog->expected_attach_type;
13699 if (member_idx >= btf_type_vlen(t)) {
13700 verbose(env, "attach to invalid member idx %u of struct %s\n",
13701 member_idx, st_ops->name);
13705 member = &btf_type_member(t)[member_idx];
13706 mname = btf_name_by_offset(btf_vmlinux, member->name_off);
13707 func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type,
13710 verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
13711 mname, member_idx, st_ops->name);
13715 if (st_ops->check_member) {
13716 int err = st_ops->check_member(t, member);
13719 verbose(env, "attach to unsupported member %s of struct %s\n",
13720 mname, st_ops->name);
13725 prog->aux->attach_func_proto = func_proto;
13726 prog->aux->attach_func_name = mname;
13727 env->ops = st_ops->verifier_ops;
13731 #define SECURITY_PREFIX "security_"
13733 static int check_attach_modify_return(unsigned long addr, const char *func_name)
13735 if (within_error_injection_list(addr) ||
13736 !strncmp(SECURITY_PREFIX, func_name, sizeof(SECURITY_PREFIX) - 1))
13742 /* list of non-sleepable functions that are otherwise on
13743 * ALLOW_ERROR_INJECTION list
13745 BTF_SET_START(btf_non_sleepable_error_inject)
13746 /* Three functions below can be called from sleepable and non-sleepable context.
13747 * Assume non-sleepable from bpf safety point of view.
13749 BTF_ID(func, __filemap_add_folio)
13750 BTF_ID(func, should_fail_alloc_page)
13751 BTF_ID(func, should_failslab)
13752 BTF_SET_END(btf_non_sleepable_error_inject)
13754 static int check_non_sleepable_error_inject(u32 btf_id)
13756 return btf_id_set_contains(&btf_non_sleepable_error_inject, btf_id);
13759 int bpf_check_attach_target(struct bpf_verifier_log *log,
13760 const struct bpf_prog *prog,
13761 const struct bpf_prog *tgt_prog,
13763 struct bpf_attach_target_info *tgt_info)
13765 bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
13766 const char prefix[] = "btf_trace_";
13767 int ret = 0, subprog = -1, i;
13768 const struct btf_type *t;
13769 bool conservative = true;
13775 bpf_log(log, "Tracing programs must provide btf_id\n");
13778 btf = tgt_prog ? tgt_prog->aux->btf : prog->aux->attach_btf;
13781 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
13784 t = btf_type_by_id(btf, btf_id);
13786 bpf_log(log, "attach_btf_id %u is invalid\n", btf_id);
13789 tname = btf_name_by_offset(btf, t->name_off);
13791 bpf_log(log, "attach_btf_id %u doesn't have a name\n", btf_id);
13795 struct bpf_prog_aux *aux = tgt_prog->aux;
13797 for (i = 0; i < aux->func_info_cnt; i++)
13798 if (aux->func_info[i].type_id == btf_id) {
13802 if (subprog == -1) {
13803 bpf_log(log, "Subprog %s doesn't exist\n", tname);
13806 conservative = aux->func_info_aux[subprog].unreliable;
13807 if (prog_extension) {
13808 if (conservative) {
13810 "Cannot replace static functions\n");
13813 if (!prog->jit_requested) {
13815 "Extension programs should be JITed\n");
13819 if (!tgt_prog->jited) {
13820 bpf_log(log, "Can attach to only JITed progs\n");
13823 if (tgt_prog->type == prog->type) {
13824 /* Cannot fentry/fexit another fentry/fexit program.
13825 * Cannot attach program extension to another extension.
13826 * It's ok to attach fentry/fexit to extension program.
13828 bpf_log(log, "Cannot recursively attach\n");
13831 if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
13833 (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
13834 tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
13835 /* Program extensions can extend all program types
13836 * except fentry/fexit. The reason is the following.
13837 * The fentry/fexit programs are used for performance
13838 * analysis, stats and can be attached to any program
13839 * type except themselves. When extension program is
13840 * replacing XDP function it is necessary to allow
13841 * performance analysis of all functions. Both original
13842 * XDP program and its program extension. Hence
13843 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
13844 * allowed. If extending of fentry/fexit was allowed it
13845 * would be possible to create long call chain
13846 * fentry->extension->fentry->extension beyond
13847 * reasonable stack size. Hence extending fentry is not
13850 bpf_log(log, "Cannot extend fentry/fexit\n");
13854 if (prog_extension) {
13855 bpf_log(log, "Cannot replace kernel functions\n");
13860 switch (prog->expected_attach_type) {
13861 case BPF_TRACE_RAW_TP:
13864 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
13867 if (!btf_type_is_typedef(t)) {
13868 bpf_log(log, "attach_btf_id %u is not a typedef\n",
13872 if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
13873 bpf_log(log, "attach_btf_id %u points to wrong type name %s\n",
13877 tname += sizeof(prefix) - 1;
13878 t = btf_type_by_id(btf, t->type);
13879 if (!btf_type_is_ptr(t))
13880 /* should never happen in valid vmlinux build */
13882 t = btf_type_by_id(btf, t->type);
13883 if (!btf_type_is_func_proto(t))
13884 /* should never happen in valid vmlinux build */
13888 case BPF_TRACE_ITER:
13889 if (!btf_type_is_func(t)) {
13890 bpf_log(log, "attach_btf_id %u is not a function\n",
13894 t = btf_type_by_id(btf, t->type);
13895 if (!btf_type_is_func_proto(t))
13897 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
13902 if (!prog_extension)
13905 case BPF_MODIFY_RETURN:
13907 case BPF_TRACE_FENTRY:
13908 case BPF_TRACE_FEXIT:
13909 if (!btf_type_is_func(t)) {
13910 bpf_log(log, "attach_btf_id %u is not a function\n",
13914 if (prog_extension &&
13915 btf_check_type_match(log, prog, btf, t))
13917 t = btf_type_by_id(btf, t->type);
13918 if (!btf_type_is_func_proto(t))
13921 if ((prog->aux->saved_dst_prog_type || prog->aux->saved_dst_attach_type) &&
13922 (!tgt_prog || prog->aux->saved_dst_prog_type != tgt_prog->type ||
13923 prog->aux->saved_dst_attach_type != tgt_prog->expected_attach_type))
13926 if (tgt_prog && conservative)
13929 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
13935 addr = (long) tgt_prog->bpf_func;
13937 addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
13939 addr = kallsyms_lookup_name(tname);
13942 "The address of function %s cannot be found\n",
13948 if (prog->aux->sleepable) {
13950 switch (prog->type) {
13951 case BPF_PROG_TYPE_TRACING:
13952 /* fentry/fexit/fmod_ret progs can be sleepable only if they are
13953 * attached to ALLOW_ERROR_INJECTION and are not in denylist.
13955 if (!check_non_sleepable_error_inject(btf_id) &&
13956 within_error_injection_list(addr))
13959 case BPF_PROG_TYPE_LSM:
13960 /* LSM progs check that they are attached to bpf_lsm_*() funcs.
13961 * Only some of them are sleepable.
13963 if (bpf_lsm_is_sleepable_hook(btf_id))
13970 bpf_log(log, "%s is not sleepable\n", tname);
13973 } else if (prog->expected_attach_type == BPF_MODIFY_RETURN) {
13975 bpf_log(log, "can't modify return codes of BPF programs\n");
13978 ret = check_attach_modify_return(addr, tname);
13980 bpf_log(log, "%s() is not modifiable\n", tname);
13987 tgt_info->tgt_addr = addr;
13988 tgt_info->tgt_name = tname;
13989 tgt_info->tgt_type = t;
13993 BTF_SET_START(btf_id_deny)
13996 BTF_ID(func, migrate_disable)
13997 BTF_ID(func, migrate_enable)
13999 #if !defined CONFIG_PREEMPT_RCU && !defined CONFIG_TINY_RCU
14000 BTF_ID(func, rcu_read_unlock_strict)
14002 BTF_SET_END(btf_id_deny)
14004 static int check_attach_btf_id(struct bpf_verifier_env *env)
14006 struct bpf_prog *prog = env->prog;
14007 struct bpf_prog *tgt_prog = prog->aux->dst_prog;
14008 struct bpf_attach_target_info tgt_info = {};
14009 u32 btf_id = prog->aux->attach_btf_id;
14010 struct bpf_trampoline *tr;
14014 if (prog->type == BPF_PROG_TYPE_SYSCALL) {
14015 if (prog->aux->sleepable)
14016 /* attach_btf_id checked to be zero already */
14018 verbose(env, "Syscall programs can only be sleepable\n");
14022 if (prog->aux->sleepable && prog->type != BPF_PROG_TYPE_TRACING &&
14023 prog->type != BPF_PROG_TYPE_LSM) {
14024 verbose(env, "Only fentry/fexit/fmod_ret and lsm programs can be sleepable\n");
14028 if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
14029 return check_struct_ops_btf_id(env);
14031 if (prog->type != BPF_PROG_TYPE_TRACING &&
14032 prog->type != BPF_PROG_TYPE_LSM &&
14033 prog->type != BPF_PROG_TYPE_EXT)
14036 ret = bpf_check_attach_target(&env->log, prog, tgt_prog, btf_id, &tgt_info);
14040 if (tgt_prog && prog->type == BPF_PROG_TYPE_EXT) {
14041 /* to make freplace equivalent to their targets, they need to
14042 * inherit env->ops and expected_attach_type for the rest of the
14045 env->ops = bpf_verifier_ops[tgt_prog->type];
14046 prog->expected_attach_type = tgt_prog->expected_attach_type;
14049 /* store info about the attachment target that will be used later */
14050 prog->aux->attach_func_proto = tgt_info.tgt_type;
14051 prog->aux->attach_func_name = tgt_info.tgt_name;
14054 prog->aux->saved_dst_prog_type = tgt_prog->type;
14055 prog->aux->saved_dst_attach_type = tgt_prog->expected_attach_type;
14058 if (prog->expected_attach_type == BPF_TRACE_RAW_TP) {
14059 prog->aux->attach_btf_trace = true;
14061 } else if (prog->expected_attach_type == BPF_TRACE_ITER) {
14062 if (!bpf_iter_prog_supported(prog))
14067 if (prog->type == BPF_PROG_TYPE_LSM) {
14068 ret = bpf_lsm_verify_prog(&env->log, prog);
14071 } else if (prog->type == BPF_PROG_TYPE_TRACING &&
14072 btf_id_set_contains(&btf_id_deny, btf_id)) {
14076 key = bpf_trampoline_compute_key(tgt_prog, prog->aux->attach_btf, btf_id);
14077 tr = bpf_trampoline_get(key, &tgt_info);
14081 prog->aux->dst_trampoline = tr;
14085 struct btf *bpf_get_btf_vmlinux(void)
14087 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
14088 mutex_lock(&bpf_verifier_lock);
14090 btf_vmlinux = btf_parse_vmlinux();
14091 mutex_unlock(&bpf_verifier_lock);
14093 return btf_vmlinux;
14096 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr, bpfptr_t uattr)
14098 u64 start_time = ktime_get_ns();
14099 struct bpf_verifier_env *env;
14100 struct bpf_verifier_log *log;
14101 int i, len, ret = -EINVAL;
14104 /* no program is valid */
14105 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
14108 /* 'struct bpf_verifier_env' can be global, but since it's not small,
14109 * allocate/free it every time bpf_check() is called
14111 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
14116 len = (*prog)->len;
14117 env->insn_aux_data =
14118 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
14120 if (!env->insn_aux_data)
14122 for (i = 0; i < len; i++)
14123 env->insn_aux_data[i].orig_idx = i;
14125 env->ops = bpf_verifier_ops[env->prog->type];
14126 env->fd_array = make_bpfptr(attr->fd_array, uattr.is_kernel);
14127 is_priv = bpf_capable();
14129 bpf_get_btf_vmlinux();
14131 /* grab the mutex to protect few globals used by verifier */
14133 mutex_lock(&bpf_verifier_lock);
14135 if (attr->log_level || attr->log_buf || attr->log_size) {
14136 /* user requested verbose verifier output
14137 * and supplied buffer to store the verification trace
14139 log->level = attr->log_level;
14140 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
14141 log->len_total = attr->log_size;
14143 /* log attributes have to be sane */
14144 if (!bpf_verifier_log_attr_valid(log)) {
14150 if (IS_ERR(btf_vmlinux)) {
14151 /* Either gcc or pahole or kernel are broken. */
14152 verbose(env, "in-kernel BTF is malformed\n");
14153 ret = PTR_ERR(btf_vmlinux);
14154 goto skip_full_check;
14157 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
14158 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
14159 env->strict_alignment = true;
14160 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
14161 env->strict_alignment = false;
14163 env->allow_ptr_leaks = bpf_allow_ptr_leaks();
14164 env->allow_uninit_stack = bpf_allow_uninit_stack();
14165 env->allow_ptr_to_map_access = bpf_allow_ptr_to_map_access();
14166 env->bypass_spec_v1 = bpf_bypass_spec_v1();
14167 env->bypass_spec_v4 = bpf_bypass_spec_v4();
14168 env->bpf_capable = bpf_capable();
14171 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
14173 env->explored_states = kvcalloc(state_htab_size(env),
14174 sizeof(struct bpf_verifier_state_list *),
14177 if (!env->explored_states)
14178 goto skip_full_check;
14180 ret = add_subprog_and_kfunc(env);
14182 goto skip_full_check;
14184 ret = check_subprogs(env);
14186 goto skip_full_check;
14188 ret = check_btf_info(env, attr, uattr);
14190 goto skip_full_check;
14192 ret = check_attach_btf_id(env);
14194 goto skip_full_check;
14196 ret = resolve_pseudo_ldimm64(env);
14198 goto skip_full_check;
14200 if (bpf_prog_is_dev_bound(env->prog->aux)) {
14201 ret = bpf_prog_offload_verifier_prep(env->prog);
14203 goto skip_full_check;
14206 ret = check_cfg(env);
14208 goto skip_full_check;
14210 ret = do_check_subprogs(env);
14211 ret = ret ?: do_check_main(env);
14213 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
14214 ret = bpf_prog_offload_finalize(env);
14217 kvfree(env->explored_states);
14220 ret = check_max_stack_depth(env);
14222 /* instruction rewrites happen after this point */
14225 opt_hard_wire_dead_code_branches(env);
14227 ret = opt_remove_dead_code(env);
14229 ret = opt_remove_nops(env);
14232 sanitize_dead_code(env);
14236 /* program is valid, convert *(u32*)(ctx + off) accesses */
14237 ret = convert_ctx_accesses(env);
14240 ret = do_misc_fixups(env);
14242 /* do 32-bit optimization after insn patching has done so those patched
14243 * insns could be handled correctly.
14245 if (ret == 0 && !bpf_prog_is_dev_bound(env->prog->aux)) {
14246 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
14247 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
14252 ret = fixup_call_args(env);
14254 env->verification_time = ktime_get_ns() - start_time;
14255 print_verification_stats(env);
14256 env->prog->aux->verified_insns = env->insn_processed;
14258 if (log->level && bpf_verifier_log_full(log))
14260 if (log->level && !log->ubuf) {
14262 goto err_release_maps;
14266 goto err_release_maps;
14268 if (env->used_map_cnt) {
14269 /* if program passed verifier, update used_maps in bpf_prog_info */
14270 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
14271 sizeof(env->used_maps[0]),
14274 if (!env->prog->aux->used_maps) {
14276 goto err_release_maps;
14279 memcpy(env->prog->aux->used_maps, env->used_maps,
14280 sizeof(env->used_maps[0]) * env->used_map_cnt);
14281 env->prog->aux->used_map_cnt = env->used_map_cnt;
14283 if (env->used_btf_cnt) {
14284 /* if program passed verifier, update used_btfs in bpf_prog_aux */
14285 env->prog->aux->used_btfs = kmalloc_array(env->used_btf_cnt,
14286 sizeof(env->used_btfs[0]),
14288 if (!env->prog->aux->used_btfs) {
14290 goto err_release_maps;
14293 memcpy(env->prog->aux->used_btfs, env->used_btfs,
14294 sizeof(env->used_btfs[0]) * env->used_btf_cnt);
14295 env->prog->aux->used_btf_cnt = env->used_btf_cnt;
14297 if (env->used_map_cnt || env->used_btf_cnt) {
14298 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
14299 * bpf_ld_imm64 instructions
14301 convert_pseudo_ld_imm64(env);
14304 adjust_btf_func(env);
14307 if (!env->prog->aux->used_maps)
14308 /* if we didn't copy map pointers into bpf_prog_info, release
14309 * them now. Otherwise free_used_maps() will release them.
14312 if (!env->prog->aux->used_btfs)
14315 /* extension progs temporarily inherit the attach_type of their targets
14316 for verification purposes, so set it back to zero before returning
14318 if (env->prog->type == BPF_PROG_TYPE_EXT)
14319 env->prog->expected_attach_type = 0;
14324 mutex_unlock(&bpf_verifier_lock);
14325 vfree(env->insn_aux_data);