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/bpf-cgroup.h>
8 #include <linux/kernel.h>
9 #include <linux/types.h>
10 #include <linux/slab.h>
11 #include <linux/bpf.h>
12 #include <linux/btf.h>
13 #include <linux/bpf_verifier.h>
14 #include <linux/filter.h>
15 #include <net/netlink.h>
16 #include <linux/file.h>
17 #include <linux/vmalloc.h>
18 #include <linux/stringify.h>
19 #include <linux/bsearch.h>
20 #include <linux/sort.h>
21 #include <linux/perf_event.h>
22 #include <linux/ctype.h>
23 #include <linux/error-injection.h>
24 #include <linux/bpf_lsm.h>
25 #include <linux/btf_ids.h>
29 static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
30 #define BPF_PROG_TYPE(_id, _name, prog_ctx_type, kern_ctx_type) \
31 [_id] = & _name ## _verifier_ops,
32 #define BPF_MAP_TYPE(_id, _ops)
33 #define BPF_LINK_TYPE(_id, _name)
34 #include <linux/bpf_types.h>
40 /* bpf_check() is a static code analyzer that walks eBPF program
41 * instruction by instruction and updates register/stack state.
42 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
44 * The first pass is depth-first-search to check that the program is a DAG.
45 * It rejects the following programs:
46 * - larger than BPF_MAXINSNS insns
47 * - if loop is present (detected via back-edge)
48 * - unreachable insns exist (shouldn't be a forest. program = one function)
49 * - out of bounds or malformed jumps
50 * The second pass is all possible path descent from the 1st insn.
51 * Since it's analyzing all paths through the program, the length of the
52 * analysis is limited to 64k insn, which may be hit even if total number of
53 * insn is less then 4K, but there are too many branches that change stack/regs.
54 * Number of 'branches to be analyzed' is limited to 1k
56 * On entry to each instruction, each register has a type, and the instruction
57 * changes the types of the registers depending on instruction semantics.
58 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
61 * All registers are 64-bit.
62 * R0 - return register
63 * R1-R5 argument passing registers
64 * R6-R9 callee saved registers
65 * R10 - frame pointer read-only
67 * At the start of BPF program the register R1 contains a pointer to bpf_context
68 * and has type PTR_TO_CTX.
70 * Verifier tracks arithmetic operations on pointers in case:
71 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
72 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
73 * 1st insn copies R10 (which has FRAME_PTR) type into R1
74 * and 2nd arithmetic instruction is pattern matched to recognize
75 * that it wants to construct a pointer to some element within stack.
76 * So after 2nd insn, the register R1 has type PTR_TO_STACK
77 * (and -20 constant is saved for further stack bounds checking).
78 * Meaning that this reg is a pointer to stack plus known immediate constant.
80 * Most of the time the registers have SCALAR_VALUE type, which
81 * means the register has some value, but it's not a valid pointer.
82 * (like pointer plus pointer becomes SCALAR_VALUE type)
84 * When verifier sees load or store instructions the type of base register
85 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK, PTR_TO_SOCKET. These are
86 * four pointer types recognized by check_mem_access() function.
88 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
89 * and the range of [ptr, ptr + map's value_size) is accessible.
91 * registers used to pass values to function calls are checked against
92 * function argument constraints.
94 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
95 * It means that the register type passed to this function must be
96 * PTR_TO_STACK and it will be used inside the function as
97 * 'pointer to map element key'
99 * For example the argument constraints for bpf_map_lookup_elem():
100 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
101 * .arg1_type = ARG_CONST_MAP_PTR,
102 * .arg2_type = ARG_PTR_TO_MAP_KEY,
104 * ret_type says that this function returns 'pointer to map elem value or null'
105 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
106 * 2nd argument should be a pointer to stack, which will be used inside
107 * the helper function as a pointer to map element key.
109 * On the kernel side the helper function looks like:
110 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
112 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
113 * void *key = (void *) (unsigned long) r2;
116 * here kernel can access 'key' and 'map' pointers safely, knowing that
117 * [key, key + map->key_size) bytes are valid and were initialized on
118 * the stack of eBPF program.
121 * Corresponding eBPF program may look like:
122 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
123 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
124 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
125 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
126 * here verifier looks at prototype of map_lookup_elem() and sees:
127 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
128 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
130 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
131 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
132 * and were initialized prior to this call.
133 * If it's ok, then verifier allows this BPF_CALL insn and looks at
134 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
135 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
136 * returns either pointer to map value or NULL.
138 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
139 * insn, the register holding that pointer in the true branch changes state to
140 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
141 * branch. See check_cond_jmp_op().
143 * After the call R0 is set to return type of the function and registers R1-R5
144 * are set to NOT_INIT to indicate that they are no longer readable.
146 * The following reference types represent a potential reference to a kernel
147 * resource which, after first being allocated, must be checked and freed by
149 * - PTR_TO_SOCKET_OR_NULL, PTR_TO_SOCKET
151 * When the verifier sees a helper call return a reference type, it allocates a
152 * pointer id for the reference and stores it in the current function state.
153 * Similar to the way that PTR_TO_MAP_VALUE_OR_NULL is converted into
154 * PTR_TO_MAP_VALUE, PTR_TO_SOCKET_OR_NULL becomes PTR_TO_SOCKET when the type
155 * passes through a NULL-check conditional. For the branch wherein the state is
156 * changed to CONST_IMM, the verifier releases the reference.
158 * For each helper function that allocates a reference, such as
159 * bpf_sk_lookup_tcp(), there is a corresponding release function, such as
160 * bpf_sk_release(). When a reference type passes into the release function,
161 * the verifier also releases the reference. If any unchecked or unreleased
162 * reference remains at the end of the program, the verifier rejects it.
165 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
166 struct bpf_verifier_stack_elem {
167 /* verifer state is 'st'
168 * before processing instruction 'insn_idx'
169 * and after processing instruction 'prev_insn_idx'
171 struct bpf_verifier_state st;
174 struct bpf_verifier_stack_elem *next;
175 /* length of verifier log at the time this state was pushed on stack */
179 #define BPF_COMPLEXITY_LIMIT_JMP_SEQ 8192
180 #define BPF_COMPLEXITY_LIMIT_STATES 64
182 #define BPF_MAP_KEY_POISON (1ULL << 63)
183 #define BPF_MAP_KEY_SEEN (1ULL << 62)
185 #define BPF_MAP_PTR_UNPRIV 1UL
186 #define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \
187 POISON_POINTER_DELTA))
188 #define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
190 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
192 return BPF_MAP_PTR(aux->map_ptr_state) == BPF_MAP_PTR_POISON;
195 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
197 return aux->map_ptr_state & BPF_MAP_PTR_UNPRIV;
200 static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
201 const struct bpf_map *map, bool unpriv)
203 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON & BPF_MAP_PTR_UNPRIV);
204 unpriv |= bpf_map_ptr_unpriv(aux);
205 aux->map_ptr_state = (unsigned long)map |
206 (unpriv ? BPF_MAP_PTR_UNPRIV : 0UL);
209 static bool bpf_map_key_poisoned(const struct bpf_insn_aux_data *aux)
211 return aux->map_key_state & BPF_MAP_KEY_POISON;
214 static bool bpf_map_key_unseen(const struct bpf_insn_aux_data *aux)
216 return !(aux->map_key_state & BPF_MAP_KEY_SEEN);
219 static u64 bpf_map_key_immediate(const struct bpf_insn_aux_data *aux)
221 return aux->map_key_state & ~(BPF_MAP_KEY_SEEN | BPF_MAP_KEY_POISON);
224 static void bpf_map_key_store(struct bpf_insn_aux_data *aux, u64 state)
226 bool poisoned = bpf_map_key_poisoned(aux);
228 aux->map_key_state = state | BPF_MAP_KEY_SEEN |
229 (poisoned ? BPF_MAP_KEY_POISON : 0ULL);
232 static bool bpf_pseudo_call(const struct bpf_insn *insn)
234 return insn->code == (BPF_JMP | BPF_CALL) &&
235 insn->src_reg == BPF_PSEUDO_CALL;
238 static bool bpf_pseudo_kfunc_call(const struct bpf_insn *insn)
240 return insn->code == (BPF_JMP | BPF_CALL) &&
241 insn->src_reg == BPF_PSEUDO_KFUNC_CALL;
244 struct bpf_call_arg_meta {
245 struct bpf_map *map_ptr;
262 struct btf *btf_vmlinux;
264 static DEFINE_MUTEX(bpf_verifier_lock);
266 static const struct bpf_line_info *
267 find_linfo(const struct bpf_verifier_env *env, u32 insn_off)
269 const struct bpf_line_info *linfo;
270 const struct bpf_prog *prog;
274 nr_linfo = prog->aux->nr_linfo;
276 if (!nr_linfo || insn_off >= prog->len)
279 linfo = prog->aux->linfo;
280 for (i = 1; i < nr_linfo; i++)
281 if (insn_off < linfo[i].insn_off)
284 return &linfo[i - 1];
287 void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
292 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
294 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
295 "verifier log line truncated - local buffer too short\n");
297 if (log->level == BPF_LOG_KERNEL) {
298 bool newline = n > 0 && log->kbuf[n - 1] == '\n';
300 pr_err("BPF: %s%s", log->kbuf, newline ? "" : "\n");
304 n = min(log->len_total - log->len_used - 1, n);
306 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
312 static void bpf_vlog_reset(struct bpf_verifier_log *log, u32 new_pos)
316 if (!bpf_verifier_log_needed(log))
319 log->len_used = new_pos;
320 if (put_user(zero, log->ubuf + new_pos))
324 /* log_level controls verbosity level of eBPF verifier.
325 * bpf_verifier_log_write() is used to dump the verification trace to the log,
326 * so the user can figure out what's wrong with the program
328 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
329 const char *fmt, ...)
333 if (!bpf_verifier_log_needed(&env->log))
337 bpf_verifier_vlog(&env->log, fmt, args);
340 EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
342 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
344 struct bpf_verifier_env *env = private_data;
347 if (!bpf_verifier_log_needed(&env->log))
351 bpf_verifier_vlog(&env->log, fmt, args);
355 __printf(2, 3) void bpf_log(struct bpf_verifier_log *log,
356 const char *fmt, ...)
360 if (!bpf_verifier_log_needed(log))
364 bpf_verifier_vlog(log, fmt, args);
368 static const char *ltrim(const char *s)
376 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env,
378 const char *prefix_fmt, ...)
380 const struct bpf_line_info *linfo;
382 if (!bpf_verifier_log_needed(&env->log))
385 linfo = find_linfo(env, insn_off);
386 if (!linfo || linfo == env->prev_linfo)
392 va_start(args, prefix_fmt);
393 bpf_verifier_vlog(&env->log, prefix_fmt, args);
398 ltrim(btf_name_by_offset(env->prog->aux->btf,
401 env->prev_linfo = linfo;
404 static void verbose_invalid_scalar(struct bpf_verifier_env *env,
405 struct bpf_reg_state *reg,
406 struct tnum *range, const char *ctx,
407 const char *reg_name)
411 verbose(env, "At %s the register %s ", ctx, reg_name);
412 if (!tnum_is_unknown(reg->var_off)) {
413 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
414 verbose(env, "has value %s", tn_buf);
416 verbose(env, "has unknown scalar value");
418 tnum_strn(tn_buf, sizeof(tn_buf), *range);
419 verbose(env, " should have been in %s\n", tn_buf);
422 static bool type_is_pkt_pointer(enum bpf_reg_type type)
424 return type == PTR_TO_PACKET ||
425 type == PTR_TO_PACKET_META;
428 static bool type_is_sk_pointer(enum bpf_reg_type type)
430 return type == PTR_TO_SOCKET ||
431 type == PTR_TO_SOCK_COMMON ||
432 type == PTR_TO_TCP_SOCK ||
433 type == PTR_TO_XDP_SOCK;
436 static bool reg_type_not_null(enum bpf_reg_type type)
438 return type == PTR_TO_SOCKET ||
439 type == PTR_TO_TCP_SOCK ||
440 type == PTR_TO_MAP_VALUE ||
441 type == PTR_TO_MAP_KEY ||
442 type == PTR_TO_SOCK_COMMON;
445 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
447 return reg->type == PTR_TO_MAP_VALUE &&
448 map_value_has_spin_lock(reg->map_ptr);
451 static bool reg_type_may_be_refcounted_or_null(enum bpf_reg_type type)
453 return base_type(type) == PTR_TO_SOCKET ||
454 base_type(type) == PTR_TO_TCP_SOCK ||
455 base_type(type) == PTR_TO_MEM;
458 static bool type_is_rdonly_mem(u32 type)
460 return type & MEM_RDONLY;
463 static bool arg_type_may_be_refcounted(enum bpf_arg_type type)
465 return type == ARG_PTR_TO_SOCK_COMMON;
468 static bool type_may_be_null(u32 type)
470 return type & PTR_MAYBE_NULL;
473 /* Determine whether the function releases some resources allocated by another
474 * function call. The first reference type argument will be assumed to be
475 * released by release_reference().
477 static bool is_release_function(enum bpf_func_id func_id)
479 return func_id == BPF_FUNC_sk_release ||
480 func_id == BPF_FUNC_ringbuf_submit ||
481 func_id == BPF_FUNC_ringbuf_discard;
484 static bool may_be_acquire_function(enum bpf_func_id func_id)
486 return func_id == BPF_FUNC_sk_lookup_tcp ||
487 func_id == BPF_FUNC_sk_lookup_udp ||
488 func_id == BPF_FUNC_skc_lookup_tcp ||
489 func_id == BPF_FUNC_map_lookup_elem ||
490 func_id == BPF_FUNC_ringbuf_reserve;
493 static bool is_acquire_function(enum bpf_func_id func_id,
494 const struct bpf_map *map)
496 enum bpf_map_type map_type = map ? map->map_type : BPF_MAP_TYPE_UNSPEC;
498 if (func_id == BPF_FUNC_sk_lookup_tcp ||
499 func_id == BPF_FUNC_sk_lookup_udp ||
500 func_id == BPF_FUNC_skc_lookup_tcp ||
501 func_id == BPF_FUNC_ringbuf_reserve)
504 if (func_id == BPF_FUNC_map_lookup_elem &&
505 (map_type == BPF_MAP_TYPE_SOCKMAP ||
506 map_type == BPF_MAP_TYPE_SOCKHASH))
512 static bool is_ptr_cast_function(enum bpf_func_id func_id)
514 return func_id == BPF_FUNC_tcp_sock ||
515 func_id == BPF_FUNC_sk_fullsock ||
516 func_id == BPF_FUNC_skc_to_tcp_sock ||
517 func_id == BPF_FUNC_skc_to_tcp6_sock ||
518 func_id == BPF_FUNC_skc_to_udp6_sock ||
519 func_id == BPF_FUNC_skc_to_tcp_timewait_sock ||
520 func_id == BPF_FUNC_skc_to_tcp_request_sock;
523 static bool is_cmpxchg_insn(const struct bpf_insn *insn)
525 return BPF_CLASS(insn->code) == BPF_STX &&
526 BPF_MODE(insn->code) == BPF_ATOMIC &&
527 insn->imm == BPF_CMPXCHG;
530 /* string representation of 'enum bpf_reg_type'
532 * Note that reg_type_str() can not appear more than once in a single verbose()
535 static const char *reg_type_str(struct bpf_verifier_env *env,
536 enum bpf_reg_type type)
538 char postfix[16] = {0}, prefix[16] = {0};
539 static const char * const str[] = {
541 [SCALAR_VALUE] = "inv",
542 [PTR_TO_CTX] = "ctx",
543 [CONST_PTR_TO_MAP] = "map_ptr",
544 [PTR_TO_MAP_VALUE] = "map_value",
545 [PTR_TO_STACK] = "fp",
546 [PTR_TO_PACKET] = "pkt",
547 [PTR_TO_PACKET_META] = "pkt_meta",
548 [PTR_TO_PACKET_END] = "pkt_end",
549 [PTR_TO_FLOW_KEYS] = "flow_keys",
550 [PTR_TO_SOCKET] = "sock",
551 [PTR_TO_SOCK_COMMON] = "sock_common",
552 [PTR_TO_TCP_SOCK] = "tcp_sock",
553 [PTR_TO_TP_BUFFER] = "tp_buffer",
554 [PTR_TO_XDP_SOCK] = "xdp_sock",
555 [PTR_TO_BTF_ID] = "ptr_",
556 [PTR_TO_PERCPU_BTF_ID] = "percpu_ptr_",
557 [PTR_TO_MEM] = "mem",
558 [PTR_TO_BUF] = "buf",
559 [PTR_TO_FUNC] = "func",
560 [PTR_TO_MAP_KEY] = "map_key",
563 if (type & PTR_MAYBE_NULL) {
564 if (base_type(type) == PTR_TO_BTF_ID ||
565 base_type(type) == PTR_TO_PERCPU_BTF_ID)
566 strncpy(postfix, "or_null_", 16);
568 strncpy(postfix, "_or_null", 16);
571 if (type & MEM_RDONLY)
572 strncpy(prefix, "rdonly_", 16);
573 if (type & MEM_ALLOC)
574 strncpy(prefix, "alloc_", 16);
576 snprintf(env->type_str_buf, TYPE_STR_BUF_LEN, "%s%s%s",
577 prefix, str[base_type(type)], postfix);
578 return env->type_str_buf;
581 static char slot_type_char[] = {
582 [STACK_INVALID] = '?',
588 static void print_liveness(struct bpf_verifier_env *env,
589 enum bpf_reg_liveness live)
591 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
593 if (live & REG_LIVE_READ)
595 if (live & REG_LIVE_WRITTEN)
597 if (live & REG_LIVE_DONE)
601 static struct bpf_func_state *func(struct bpf_verifier_env *env,
602 const struct bpf_reg_state *reg)
604 struct bpf_verifier_state *cur = env->cur_state;
606 return cur->frame[reg->frameno];
609 static const char *kernel_type_name(const struct btf* btf, u32 id)
611 return btf_name_by_offset(btf, btf_type_by_id(btf, id)->name_off);
614 static void mark_reg_scratched(struct bpf_verifier_env *env, u32 regno)
616 env->scratched_regs |= 1U << regno;
619 static void mark_stack_slot_scratched(struct bpf_verifier_env *env, u32 spi)
621 env->scratched_stack_slots |= 1ULL << spi;
624 static bool reg_scratched(const struct bpf_verifier_env *env, u32 regno)
626 return (env->scratched_regs >> regno) & 1;
629 static bool stack_slot_scratched(const struct bpf_verifier_env *env, u64 regno)
631 return (env->scratched_stack_slots >> regno) & 1;
634 static bool verifier_state_scratched(const struct bpf_verifier_env *env)
636 return env->scratched_regs || env->scratched_stack_slots;
639 static void mark_verifier_state_clean(struct bpf_verifier_env *env)
641 env->scratched_regs = 0U;
642 env->scratched_stack_slots = 0ULL;
645 /* Used for printing the entire verifier state. */
646 static void mark_verifier_state_scratched(struct bpf_verifier_env *env)
648 env->scratched_regs = ~0U;
649 env->scratched_stack_slots = ~0ULL;
652 /* The reg state of a pointer or a bounded scalar was saved when
653 * it was spilled to the stack.
655 static bool is_spilled_reg(const struct bpf_stack_state *stack)
657 return stack->slot_type[BPF_REG_SIZE - 1] == STACK_SPILL;
660 static void scrub_spilled_slot(u8 *stype)
662 if (*stype != STACK_INVALID)
666 static void print_verifier_state(struct bpf_verifier_env *env,
667 const struct bpf_func_state *state,
670 const struct bpf_reg_state *reg;
675 verbose(env, " frame%d:", state->frameno);
676 for (i = 0; i < MAX_BPF_REG; i++) {
677 reg = &state->regs[i];
681 if (!print_all && !reg_scratched(env, i))
683 verbose(env, " R%d", i);
684 print_liveness(env, reg->live);
685 verbose(env, "=%s", reg_type_str(env, t));
686 if (t == SCALAR_VALUE && reg->precise)
688 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
689 tnum_is_const(reg->var_off)) {
690 /* reg->off should be 0 for SCALAR_VALUE */
691 verbose(env, "%lld", reg->var_off.value + reg->off);
693 if (base_type(t) == PTR_TO_BTF_ID ||
694 base_type(t) == PTR_TO_PERCPU_BTF_ID)
695 verbose(env, "%s", kernel_type_name(reg->btf, reg->btf_id));
696 verbose(env, "(id=%d", reg->id);
697 if (reg_type_may_be_refcounted_or_null(t))
698 verbose(env, ",ref_obj_id=%d", reg->ref_obj_id);
699 if (t != SCALAR_VALUE)
700 verbose(env, ",off=%d", reg->off);
701 if (type_is_pkt_pointer(t))
702 verbose(env, ",r=%d", reg->range);
703 else if (base_type(t) == CONST_PTR_TO_MAP ||
704 base_type(t) == PTR_TO_MAP_KEY ||
705 base_type(t) == PTR_TO_MAP_VALUE)
706 verbose(env, ",ks=%d,vs=%d",
707 reg->map_ptr->key_size,
708 reg->map_ptr->value_size);
709 if (tnum_is_const(reg->var_off)) {
710 /* Typically an immediate SCALAR_VALUE, but
711 * could be a pointer whose offset is too big
714 verbose(env, ",imm=%llx", reg->var_off.value);
716 if (reg->smin_value != reg->umin_value &&
717 reg->smin_value != S64_MIN)
718 verbose(env, ",smin_value=%lld",
719 (long long)reg->smin_value);
720 if (reg->smax_value != reg->umax_value &&
721 reg->smax_value != S64_MAX)
722 verbose(env, ",smax_value=%lld",
723 (long long)reg->smax_value);
724 if (reg->umin_value != 0)
725 verbose(env, ",umin_value=%llu",
726 (unsigned long long)reg->umin_value);
727 if (reg->umax_value != U64_MAX)
728 verbose(env, ",umax_value=%llu",
729 (unsigned long long)reg->umax_value);
730 if (!tnum_is_unknown(reg->var_off)) {
733 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
734 verbose(env, ",var_off=%s", tn_buf);
736 if (reg->s32_min_value != reg->smin_value &&
737 reg->s32_min_value != S32_MIN)
738 verbose(env, ",s32_min_value=%d",
739 (int)(reg->s32_min_value));
740 if (reg->s32_max_value != reg->smax_value &&
741 reg->s32_max_value != S32_MAX)
742 verbose(env, ",s32_max_value=%d",
743 (int)(reg->s32_max_value));
744 if (reg->u32_min_value != reg->umin_value &&
745 reg->u32_min_value != U32_MIN)
746 verbose(env, ",u32_min_value=%d",
747 (int)(reg->u32_min_value));
748 if (reg->u32_max_value != reg->umax_value &&
749 reg->u32_max_value != U32_MAX)
750 verbose(env, ",u32_max_value=%d",
751 (int)(reg->u32_max_value));
756 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
757 char types_buf[BPF_REG_SIZE + 1];
761 for (j = 0; j < BPF_REG_SIZE; j++) {
762 if (state->stack[i].slot_type[j] != STACK_INVALID)
764 types_buf[j] = slot_type_char[
765 state->stack[i].slot_type[j]];
767 types_buf[BPF_REG_SIZE] = 0;
770 if (!print_all && !stack_slot_scratched(env, i))
772 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
773 print_liveness(env, state->stack[i].spilled_ptr.live);
774 if (is_spilled_reg(&state->stack[i])) {
775 reg = &state->stack[i].spilled_ptr;
777 verbose(env, "=%s", reg_type_str(env, t));
778 if (t == SCALAR_VALUE && reg->precise)
780 if (t == SCALAR_VALUE && tnum_is_const(reg->var_off))
781 verbose(env, "%lld", reg->var_off.value + reg->off);
783 verbose(env, "=%s", types_buf);
786 if (state->acquired_refs && state->refs[0].id) {
787 verbose(env, " refs=%d", state->refs[0].id);
788 for (i = 1; i < state->acquired_refs; i++)
789 if (state->refs[i].id)
790 verbose(env, ",%d", state->refs[i].id);
792 if (state->in_callback_fn)
794 if (state->in_async_callback_fn)
795 verbose(env, " async_cb");
797 mark_verifier_state_clean(env);
800 static inline u32 vlog_alignment(u32 pos)
802 return round_up(max(pos + BPF_LOG_MIN_ALIGNMENT / 2, BPF_LOG_ALIGNMENT),
803 BPF_LOG_MIN_ALIGNMENT) - pos - 1;
806 static void print_insn_state(struct bpf_verifier_env *env,
807 const struct bpf_func_state *state)
809 if (env->prev_log_len && env->prev_log_len == env->log.len_used) {
810 /* remove new line character */
811 bpf_vlog_reset(&env->log, env->prev_log_len - 1);
812 verbose(env, "%*c;", vlog_alignment(env->prev_insn_print_len), ' ');
814 verbose(env, "%d:", env->insn_idx);
816 print_verifier_state(env, state, false);
819 /* copy array src of length n * size bytes to dst. dst is reallocated if it's too
820 * small to hold src. This is different from krealloc since we don't want to preserve
821 * the contents of dst.
823 * Leaves dst untouched if src is NULL or length is zero. Returns NULL if memory could
826 static void *copy_array(void *dst, const void *src, size_t n, size_t size, gfp_t flags)
830 if (ZERO_OR_NULL_PTR(src))
833 if (unlikely(check_mul_overflow(n, size, &bytes)))
836 if (ksize(dst) < bytes) {
838 dst = kmalloc_track_caller(bytes, flags);
843 memcpy(dst, src, bytes);
845 return dst ? dst : ZERO_SIZE_PTR;
848 /* resize an array from old_n items to new_n items. the array is reallocated if it's too
849 * small to hold new_n items. new items are zeroed out if the array grows.
851 * Contrary to krealloc_array, does not free arr if new_n is zero.
853 static void *realloc_array(void *arr, size_t old_n, size_t new_n, size_t size)
855 if (!new_n || old_n == new_n)
858 arr = krealloc_array(arr, new_n, size, GFP_KERNEL);
863 memset(arr + old_n * size, 0, (new_n - old_n) * size);
866 return arr ? arr : ZERO_SIZE_PTR;
869 static int copy_reference_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
871 dst->refs = copy_array(dst->refs, src->refs, src->acquired_refs,
872 sizeof(struct bpf_reference_state), GFP_KERNEL);
876 dst->acquired_refs = src->acquired_refs;
880 static int copy_stack_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
882 size_t n = src->allocated_stack / BPF_REG_SIZE;
884 dst->stack = copy_array(dst->stack, src->stack, n, sizeof(struct bpf_stack_state),
889 dst->allocated_stack = src->allocated_stack;
893 static int resize_reference_state(struct bpf_func_state *state, size_t n)
895 state->refs = realloc_array(state->refs, state->acquired_refs, n,
896 sizeof(struct bpf_reference_state));
900 state->acquired_refs = n;
904 static int grow_stack_state(struct bpf_func_state *state, int size)
906 size_t old_n = state->allocated_stack / BPF_REG_SIZE, n = size / BPF_REG_SIZE;
911 state->stack = realloc_array(state->stack, old_n, n, sizeof(struct bpf_stack_state));
915 state->allocated_stack = size;
919 /* Acquire a pointer id from the env and update the state->refs to include
920 * this new pointer reference.
921 * On success, returns a valid pointer id to associate with the register
922 * On failure, returns a negative errno.
924 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
926 struct bpf_func_state *state = cur_func(env);
927 int new_ofs = state->acquired_refs;
930 err = resize_reference_state(state, state->acquired_refs + 1);
934 state->refs[new_ofs].id = id;
935 state->refs[new_ofs].insn_idx = insn_idx;
940 /* release function corresponding to acquire_reference_state(). Idempotent. */
941 static int release_reference_state(struct bpf_func_state *state, int ptr_id)
945 last_idx = state->acquired_refs - 1;
946 for (i = 0; i < state->acquired_refs; i++) {
947 if (state->refs[i].id == ptr_id) {
948 if (last_idx && i != last_idx)
949 memcpy(&state->refs[i], &state->refs[last_idx],
950 sizeof(*state->refs));
951 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
952 state->acquired_refs--;
959 static void free_func_state(struct bpf_func_state *state)
968 static void clear_jmp_history(struct bpf_verifier_state *state)
970 kfree(state->jmp_history);
971 state->jmp_history = NULL;
972 state->jmp_history_cnt = 0;
975 static void free_verifier_state(struct bpf_verifier_state *state,
980 for (i = 0; i <= state->curframe; i++) {
981 free_func_state(state->frame[i]);
982 state->frame[i] = NULL;
984 clear_jmp_history(state);
989 /* copy verifier state from src to dst growing dst stack space
990 * when necessary to accommodate larger src stack
992 static int copy_func_state(struct bpf_func_state *dst,
993 const struct bpf_func_state *src)
997 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
998 err = copy_reference_state(dst, src);
1001 return copy_stack_state(dst, src);
1004 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
1005 const struct bpf_verifier_state *src)
1007 struct bpf_func_state *dst;
1010 dst_state->jmp_history = copy_array(dst_state->jmp_history, src->jmp_history,
1011 src->jmp_history_cnt, sizeof(struct bpf_idx_pair),
1013 if (!dst_state->jmp_history)
1015 dst_state->jmp_history_cnt = src->jmp_history_cnt;
1017 /* if dst has more stack frames then src frame, free them */
1018 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
1019 free_func_state(dst_state->frame[i]);
1020 dst_state->frame[i] = NULL;
1022 dst_state->speculative = src->speculative;
1023 dst_state->curframe = src->curframe;
1024 dst_state->active_spin_lock = src->active_spin_lock;
1025 dst_state->branches = src->branches;
1026 dst_state->parent = src->parent;
1027 dst_state->first_insn_idx = src->first_insn_idx;
1028 dst_state->last_insn_idx = src->last_insn_idx;
1029 for (i = 0; i <= src->curframe; i++) {
1030 dst = dst_state->frame[i];
1032 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
1035 dst_state->frame[i] = dst;
1037 err = copy_func_state(dst, src->frame[i]);
1044 static void update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
1047 u32 br = --st->branches;
1049 /* WARN_ON(br > 1) technically makes sense here,
1050 * but see comment in push_stack(), hence:
1052 WARN_ONCE((int)br < 0,
1053 "BUG update_branch_counts:branches_to_explore=%d\n",
1061 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
1062 int *insn_idx, bool pop_log)
1064 struct bpf_verifier_state *cur = env->cur_state;
1065 struct bpf_verifier_stack_elem *elem, *head = env->head;
1068 if (env->head == NULL)
1072 err = copy_verifier_state(cur, &head->st);
1077 bpf_vlog_reset(&env->log, head->log_pos);
1079 *insn_idx = head->insn_idx;
1081 *prev_insn_idx = head->prev_insn_idx;
1083 free_verifier_state(&head->st, false);
1090 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
1091 int insn_idx, int prev_insn_idx,
1094 struct bpf_verifier_state *cur = env->cur_state;
1095 struct bpf_verifier_stack_elem *elem;
1098 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
1102 elem->insn_idx = insn_idx;
1103 elem->prev_insn_idx = prev_insn_idx;
1104 elem->next = env->head;
1105 elem->log_pos = env->log.len_used;
1108 err = copy_verifier_state(&elem->st, cur);
1111 elem->st.speculative |= speculative;
1112 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
1113 verbose(env, "The sequence of %d jumps is too complex.\n",
1117 if (elem->st.parent) {
1118 ++elem->st.parent->branches;
1119 /* WARN_ON(branches > 2) technically makes sense here,
1121 * 1. speculative states will bump 'branches' for non-branch
1123 * 2. is_state_visited() heuristics may decide not to create
1124 * a new state for a sequence of branches and all such current
1125 * and cloned states will be pointing to a single parent state
1126 * which might have large 'branches' count.
1131 free_verifier_state(env->cur_state, true);
1132 env->cur_state = NULL;
1133 /* pop all elements and return */
1134 while (!pop_stack(env, NULL, NULL, false));
1138 #define CALLER_SAVED_REGS 6
1139 static const int caller_saved[CALLER_SAVED_REGS] = {
1140 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
1143 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1144 struct bpf_reg_state *reg);
1146 /* This helper doesn't clear reg->id */
1147 static void ___mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1149 reg->var_off = tnum_const(imm);
1150 reg->smin_value = (s64)imm;
1151 reg->smax_value = (s64)imm;
1152 reg->umin_value = imm;
1153 reg->umax_value = imm;
1155 reg->s32_min_value = (s32)imm;
1156 reg->s32_max_value = (s32)imm;
1157 reg->u32_min_value = (u32)imm;
1158 reg->u32_max_value = (u32)imm;
1161 /* Mark the unknown part of a register (variable offset or scalar value) as
1162 * known to have the value @imm.
1164 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1166 /* Clear id, off, and union(map_ptr, range) */
1167 memset(((u8 *)reg) + sizeof(reg->type), 0,
1168 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
1169 ___mark_reg_known(reg, imm);
1172 static void __mark_reg32_known(struct bpf_reg_state *reg, u64 imm)
1174 reg->var_off = tnum_const_subreg(reg->var_off, imm);
1175 reg->s32_min_value = (s32)imm;
1176 reg->s32_max_value = (s32)imm;
1177 reg->u32_min_value = (u32)imm;
1178 reg->u32_max_value = (u32)imm;
1181 /* Mark the 'variable offset' part of a register as zero. This should be
1182 * used only on registers holding a pointer type.
1184 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
1186 __mark_reg_known(reg, 0);
1189 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
1191 __mark_reg_known(reg, 0);
1192 reg->type = SCALAR_VALUE;
1195 static void mark_reg_known_zero(struct bpf_verifier_env *env,
1196 struct bpf_reg_state *regs, u32 regno)
1198 if (WARN_ON(regno >= MAX_BPF_REG)) {
1199 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
1200 /* Something bad happened, let's kill all regs */
1201 for (regno = 0; regno < MAX_BPF_REG; regno++)
1202 __mark_reg_not_init(env, regs + regno);
1205 __mark_reg_known_zero(regs + regno);
1208 static void mark_ptr_not_null_reg(struct bpf_reg_state *reg)
1210 if (base_type(reg->type) == PTR_TO_MAP_VALUE) {
1211 const struct bpf_map *map = reg->map_ptr;
1213 if (map->inner_map_meta) {
1214 reg->type = CONST_PTR_TO_MAP;
1215 reg->map_ptr = map->inner_map_meta;
1216 /* transfer reg's id which is unique for every map_lookup_elem
1217 * as UID of the inner map.
1219 if (map_value_has_timer(map->inner_map_meta))
1220 reg->map_uid = reg->id;
1221 } else if (map->map_type == BPF_MAP_TYPE_XSKMAP) {
1222 reg->type = PTR_TO_XDP_SOCK;
1223 } else if (map->map_type == BPF_MAP_TYPE_SOCKMAP ||
1224 map->map_type == BPF_MAP_TYPE_SOCKHASH) {
1225 reg->type = PTR_TO_SOCKET;
1227 reg->type = PTR_TO_MAP_VALUE;
1232 reg->type &= ~PTR_MAYBE_NULL;
1235 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
1237 return type_is_pkt_pointer(reg->type);
1240 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
1242 return reg_is_pkt_pointer(reg) ||
1243 reg->type == PTR_TO_PACKET_END;
1246 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
1247 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
1248 enum bpf_reg_type which)
1250 /* The register can already have a range from prior markings.
1251 * This is fine as long as it hasn't been advanced from its
1254 return reg->type == which &&
1257 tnum_equals_const(reg->var_off, 0);
1260 /* Reset the min/max bounds of a register */
1261 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
1263 reg->smin_value = S64_MIN;
1264 reg->smax_value = S64_MAX;
1265 reg->umin_value = 0;
1266 reg->umax_value = U64_MAX;
1268 reg->s32_min_value = S32_MIN;
1269 reg->s32_max_value = S32_MAX;
1270 reg->u32_min_value = 0;
1271 reg->u32_max_value = U32_MAX;
1274 static void __mark_reg64_unbounded(struct bpf_reg_state *reg)
1276 reg->smin_value = S64_MIN;
1277 reg->smax_value = S64_MAX;
1278 reg->umin_value = 0;
1279 reg->umax_value = U64_MAX;
1282 static void __mark_reg32_unbounded(struct bpf_reg_state *reg)
1284 reg->s32_min_value = S32_MIN;
1285 reg->s32_max_value = S32_MAX;
1286 reg->u32_min_value = 0;
1287 reg->u32_max_value = U32_MAX;
1290 static void __update_reg32_bounds(struct bpf_reg_state *reg)
1292 struct tnum var32_off = tnum_subreg(reg->var_off);
1294 /* min signed is max(sign bit) | min(other bits) */
1295 reg->s32_min_value = max_t(s32, reg->s32_min_value,
1296 var32_off.value | (var32_off.mask & S32_MIN));
1297 /* max signed is min(sign bit) | max(other bits) */
1298 reg->s32_max_value = min_t(s32, reg->s32_max_value,
1299 var32_off.value | (var32_off.mask & S32_MAX));
1300 reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)var32_off.value);
1301 reg->u32_max_value = min(reg->u32_max_value,
1302 (u32)(var32_off.value | var32_off.mask));
1305 static void __update_reg64_bounds(struct bpf_reg_state *reg)
1307 /* min signed is max(sign bit) | min(other bits) */
1308 reg->smin_value = max_t(s64, reg->smin_value,
1309 reg->var_off.value | (reg->var_off.mask & S64_MIN));
1310 /* max signed is min(sign bit) | max(other bits) */
1311 reg->smax_value = min_t(s64, reg->smax_value,
1312 reg->var_off.value | (reg->var_off.mask & S64_MAX));
1313 reg->umin_value = max(reg->umin_value, reg->var_off.value);
1314 reg->umax_value = min(reg->umax_value,
1315 reg->var_off.value | reg->var_off.mask);
1318 static void __update_reg_bounds(struct bpf_reg_state *reg)
1320 __update_reg32_bounds(reg);
1321 __update_reg64_bounds(reg);
1324 /* Uses signed min/max values to inform unsigned, and vice-versa */
1325 static void __reg32_deduce_bounds(struct bpf_reg_state *reg)
1327 /* Learn sign from signed bounds.
1328 * If we cannot cross the sign boundary, then signed and unsigned bounds
1329 * are the same, so combine. This works even in the negative case, e.g.
1330 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1332 if (reg->s32_min_value >= 0 || reg->s32_max_value < 0) {
1333 reg->s32_min_value = reg->u32_min_value =
1334 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1335 reg->s32_max_value = reg->u32_max_value =
1336 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1339 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1340 * boundary, so we must be careful.
1342 if ((s32)reg->u32_max_value >= 0) {
1343 /* Positive. We can't learn anything from the smin, but smax
1344 * is positive, hence safe.
1346 reg->s32_min_value = reg->u32_min_value;
1347 reg->s32_max_value = reg->u32_max_value =
1348 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1349 } else if ((s32)reg->u32_min_value < 0) {
1350 /* Negative. We can't learn anything from the smax, but smin
1351 * is negative, hence safe.
1353 reg->s32_min_value = reg->u32_min_value =
1354 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1355 reg->s32_max_value = reg->u32_max_value;
1359 static void __reg64_deduce_bounds(struct bpf_reg_state *reg)
1361 /* Learn sign from signed bounds.
1362 * If we cannot cross the sign boundary, then signed and unsigned bounds
1363 * are the same, so combine. This works even in the negative case, e.g.
1364 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1366 if (reg->smin_value >= 0 || reg->smax_value < 0) {
1367 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1369 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1373 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1374 * boundary, so we must be careful.
1376 if ((s64)reg->umax_value >= 0) {
1377 /* Positive. We can't learn anything from the smin, but smax
1378 * is positive, hence safe.
1380 reg->smin_value = reg->umin_value;
1381 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1383 } else if ((s64)reg->umin_value < 0) {
1384 /* Negative. We can't learn anything from the smax, but smin
1385 * is negative, hence safe.
1387 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1389 reg->smax_value = reg->umax_value;
1393 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
1395 __reg32_deduce_bounds(reg);
1396 __reg64_deduce_bounds(reg);
1399 /* Attempts to improve var_off based on unsigned min/max information */
1400 static void __reg_bound_offset(struct bpf_reg_state *reg)
1402 struct tnum var64_off = tnum_intersect(reg->var_off,
1403 tnum_range(reg->umin_value,
1405 struct tnum var32_off = tnum_intersect(tnum_subreg(reg->var_off),
1406 tnum_range(reg->u32_min_value,
1407 reg->u32_max_value));
1409 reg->var_off = tnum_or(tnum_clear_subreg(var64_off), var32_off);
1412 static bool __reg32_bound_s64(s32 a)
1414 return a >= 0 && a <= S32_MAX;
1417 static void __reg_assign_32_into_64(struct bpf_reg_state *reg)
1419 reg->umin_value = reg->u32_min_value;
1420 reg->umax_value = reg->u32_max_value;
1422 /* Attempt to pull 32-bit signed bounds into 64-bit bounds but must
1423 * be positive otherwise set to worse case bounds and refine later
1426 if (__reg32_bound_s64(reg->s32_min_value) &&
1427 __reg32_bound_s64(reg->s32_max_value)) {
1428 reg->smin_value = reg->s32_min_value;
1429 reg->smax_value = reg->s32_max_value;
1431 reg->smin_value = 0;
1432 reg->smax_value = U32_MAX;
1436 static void __reg_combine_32_into_64(struct bpf_reg_state *reg)
1438 /* special case when 64-bit register has upper 32-bit register
1439 * zeroed. Typically happens after zext or <<32, >>32 sequence
1440 * allowing us to use 32-bit bounds directly,
1442 if (tnum_equals_const(tnum_clear_subreg(reg->var_off), 0)) {
1443 __reg_assign_32_into_64(reg);
1445 /* Otherwise the best we can do is push lower 32bit known and
1446 * unknown bits into register (var_off set from jmp logic)
1447 * then learn as much as possible from the 64-bit tnum
1448 * known and unknown bits. The previous smin/smax bounds are
1449 * invalid here because of jmp32 compare so mark them unknown
1450 * so they do not impact tnum bounds calculation.
1452 __mark_reg64_unbounded(reg);
1453 __update_reg_bounds(reg);
1456 /* Intersecting with the old var_off might have improved our bounds
1457 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1458 * then new var_off is (0; 0x7f...fc) which improves our umax.
1460 __reg_deduce_bounds(reg);
1461 __reg_bound_offset(reg);
1462 __update_reg_bounds(reg);
1465 static bool __reg64_bound_s32(s64 a)
1467 return a >= S32_MIN && a <= S32_MAX;
1470 static bool __reg64_bound_u32(u64 a)
1472 return a >= U32_MIN && a <= U32_MAX;
1475 static void __reg_combine_64_into_32(struct bpf_reg_state *reg)
1477 __mark_reg32_unbounded(reg);
1479 if (__reg64_bound_s32(reg->smin_value) && __reg64_bound_s32(reg->smax_value)) {
1480 reg->s32_min_value = (s32)reg->smin_value;
1481 reg->s32_max_value = (s32)reg->smax_value;
1483 if (__reg64_bound_u32(reg->umin_value) && __reg64_bound_u32(reg->umax_value)) {
1484 reg->u32_min_value = (u32)reg->umin_value;
1485 reg->u32_max_value = (u32)reg->umax_value;
1488 /* Intersecting with the old var_off might have improved our bounds
1489 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1490 * then new var_off is (0; 0x7f...fc) which improves our umax.
1492 __reg_deduce_bounds(reg);
1493 __reg_bound_offset(reg);
1494 __update_reg_bounds(reg);
1497 /* Mark a register as having a completely unknown (scalar) value. */
1498 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
1499 struct bpf_reg_state *reg)
1502 * Clear type, id, off, and union(map_ptr, range) and
1503 * padding between 'type' and union
1505 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
1506 reg->type = SCALAR_VALUE;
1507 reg->var_off = tnum_unknown;
1509 reg->precise = env->subprog_cnt > 1 || !env->bpf_capable;
1510 __mark_reg_unbounded(reg);
1513 static void mark_reg_unknown(struct bpf_verifier_env *env,
1514 struct bpf_reg_state *regs, u32 regno)
1516 if (WARN_ON(regno >= MAX_BPF_REG)) {
1517 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
1518 /* Something bad happened, let's kill all regs except FP */
1519 for (regno = 0; regno < BPF_REG_FP; regno++)
1520 __mark_reg_not_init(env, regs + regno);
1523 __mark_reg_unknown(env, regs + regno);
1526 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1527 struct bpf_reg_state *reg)
1529 __mark_reg_unknown(env, reg);
1530 reg->type = NOT_INIT;
1533 static void mark_reg_not_init(struct bpf_verifier_env *env,
1534 struct bpf_reg_state *regs, u32 regno)
1536 if (WARN_ON(regno >= MAX_BPF_REG)) {
1537 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
1538 /* Something bad happened, let's kill all regs except FP */
1539 for (regno = 0; regno < BPF_REG_FP; regno++)
1540 __mark_reg_not_init(env, regs + regno);
1543 __mark_reg_not_init(env, regs + regno);
1546 static void mark_btf_ld_reg(struct bpf_verifier_env *env,
1547 struct bpf_reg_state *regs, u32 regno,
1548 enum bpf_reg_type reg_type,
1549 struct btf *btf, u32 btf_id)
1551 if (reg_type == SCALAR_VALUE) {
1552 mark_reg_unknown(env, regs, regno);
1555 mark_reg_known_zero(env, regs, regno);
1556 regs[regno].type = PTR_TO_BTF_ID;
1557 regs[regno].btf = btf;
1558 regs[regno].btf_id = btf_id;
1561 #define DEF_NOT_SUBREG (0)
1562 static void init_reg_state(struct bpf_verifier_env *env,
1563 struct bpf_func_state *state)
1565 struct bpf_reg_state *regs = state->regs;
1568 for (i = 0; i < MAX_BPF_REG; i++) {
1569 mark_reg_not_init(env, regs, i);
1570 regs[i].live = REG_LIVE_NONE;
1571 regs[i].parent = NULL;
1572 regs[i].subreg_def = DEF_NOT_SUBREG;
1576 regs[BPF_REG_FP].type = PTR_TO_STACK;
1577 mark_reg_known_zero(env, regs, BPF_REG_FP);
1578 regs[BPF_REG_FP].frameno = state->frameno;
1581 #define BPF_MAIN_FUNC (-1)
1582 static void init_func_state(struct bpf_verifier_env *env,
1583 struct bpf_func_state *state,
1584 int callsite, int frameno, int subprogno)
1586 state->callsite = callsite;
1587 state->frameno = frameno;
1588 state->subprogno = subprogno;
1589 init_reg_state(env, state);
1590 mark_verifier_state_scratched(env);
1593 /* Similar to push_stack(), but for async callbacks */
1594 static struct bpf_verifier_state *push_async_cb(struct bpf_verifier_env *env,
1595 int insn_idx, int prev_insn_idx,
1598 struct bpf_verifier_stack_elem *elem;
1599 struct bpf_func_state *frame;
1601 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
1605 elem->insn_idx = insn_idx;
1606 elem->prev_insn_idx = prev_insn_idx;
1607 elem->next = env->head;
1608 elem->log_pos = env->log.len_used;
1611 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
1613 "The sequence of %d jumps is too complex for async cb.\n",
1617 /* Unlike push_stack() do not copy_verifier_state().
1618 * The caller state doesn't matter.
1619 * This is async callback. It starts in a fresh stack.
1620 * Initialize it similar to do_check_common().
1622 elem->st.branches = 1;
1623 frame = kzalloc(sizeof(*frame), GFP_KERNEL);
1626 init_func_state(env, frame,
1627 BPF_MAIN_FUNC /* callsite */,
1628 0 /* frameno within this callchain */,
1629 subprog /* subprog number within this prog */);
1630 elem->st.frame[0] = frame;
1633 free_verifier_state(env->cur_state, true);
1634 env->cur_state = NULL;
1635 /* pop all elements and return */
1636 while (!pop_stack(env, NULL, NULL, false));
1642 SRC_OP, /* register is used as source operand */
1643 DST_OP, /* register is used as destination operand */
1644 DST_OP_NO_MARK /* same as above, check only, don't mark */
1647 static int cmp_subprogs(const void *a, const void *b)
1649 return ((struct bpf_subprog_info *)a)->start -
1650 ((struct bpf_subprog_info *)b)->start;
1653 static int find_subprog(struct bpf_verifier_env *env, int off)
1655 struct bpf_subprog_info *p;
1657 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
1658 sizeof(env->subprog_info[0]), cmp_subprogs);
1661 return p - env->subprog_info;
1665 static int add_subprog(struct bpf_verifier_env *env, int off)
1667 int insn_cnt = env->prog->len;
1670 if (off >= insn_cnt || off < 0) {
1671 verbose(env, "call to invalid destination\n");
1674 ret = find_subprog(env, off);
1677 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
1678 verbose(env, "too many subprograms\n");
1681 /* determine subprog starts. The end is one before the next starts */
1682 env->subprog_info[env->subprog_cnt++].start = off;
1683 sort(env->subprog_info, env->subprog_cnt,
1684 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
1685 return env->subprog_cnt - 1;
1688 #define MAX_KFUNC_DESCS 256
1689 #define MAX_KFUNC_BTFS 256
1691 struct bpf_kfunc_desc {
1692 struct btf_func_model func_model;
1698 struct bpf_kfunc_btf {
1700 struct module *module;
1704 struct bpf_kfunc_desc_tab {
1705 struct bpf_kfunc_desc descs[MAX_KFUNC_DESCS];
1709 struct bpf_kfunc_btf_tab {
1710 struct bpf_kfunc_btf descs[MAX_KFUNC_BTFS];
1714 static int kfunc_desc_cmp_by_id_off(const void *a, const void *b)
1716 const struct bpf_kfunc_desc *d0 = a;
1717 const struct bpf_kfunc_desc *d1 = b;
1719 /* func_id is not greater than BTF_MAX_TYPE */
1720 return d0->func_id - d1->func_id ?: d0->offset - d1->offset;
1723 static int kfunc_btf_cmp_by_off(const void *a, const void *b)
1725 const struct bpf_kfunc_btf *d0 = a;
1726 const struct bpf_kfunc_btf *d1 = b;
1728 return d0->offset - d1->offset;
1731 static const struct bpf_kfunc_desc *
1732 find_kfunc_desc(const struct bpf_prog *prog, u32 func_id, u16 offset)
1734 struct bpf_kfunc_desc desc = {
1738 struct bpf_kfunc_desc_tab *tab;
1740 tab = prog->aux->kfunc_tab;
1741 return bsearch(&desc, tab->descs, tab->nr_descs,
1742 sizeof(tab->descs[0]), kfunc_desc_cmp_by_id_off);
1745 static struct btf *__find_kfunc_desc_btf(struct bpf_verifier_env *env,
1746 s16 offset, struct module **btf_modp)
1748 struct bpf_kfunc_btf kf_btf = { .offset = offset };
1749 struct bpf_kfunc_btf_tab *tab;
1750 struct bpf_kfunc_btf *b;
1755 tab = env->prog->aux->kfunc_btf_tab;
1756 b = bsearch(&kf_btf, tab->descs, tab->nr_descs,
1757 sizeof(tab->descs[0]), kfunc_btf_cmp_by_off);
1759 if (tab->nr_descs == MAX_KFUNC_BTFS) {
1760 verbose(env, "too many different module BTFs\n");
1761 return ERR_PTR(-E2BIG);
1764 if (bpfptr_is_null(env->fd_array)) {
1765 verbose(env, "kfunc offset > 0 without fd_array is invalid\n");
1766 return ERR_PTR(-EPROTO);
1769 if (copy_from_bpfptr_offset(&btf_fd, env->fd_array,
1770 offset * sizeof(btf_fd),
1772 return ERR_PTR(-EFAULT);
1774 btf = btf_get_by_fd(btf_fd);
1776 verbose(env, "invalid module BTF fd specified\n");
1780 if (!btf_is_module(btf)) {
1781 verbose(env, "BTF fd for kfunc is not a module BTF\n");
1783 return ERR_PTR(-EINVAL);
1786 mod = btf_try_get_module(btf);
1789 return ERR_PTR(-ENXIO);
1792 b = &tab->descs[tab->nr_descs++];
1797 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
1798 kfunc_btf_cmp_by_off, NULL);
1801 *btf_modp = b->module;
1805 void bpf_free_kfunc_btf_tab(struct bpf_kfunc_btf_tab *tab)
1810 while (tab->nr_descs--) {
1811 module_put(tab->descs[tab->nr_descs].module);
1812 btf_put(tab->descs[tab->nr_descs].btf);
1817 static struct btf *find_kfunc_desc_btf(struct bpf_verifier_env *env,
1818 u32 func_id, s16 offset,
1819 struct module **btf_modp)
1823 /* In the future, this can be allowed to increase limit
1824 * of fd index into fd_array, interpreted as u16.
1826 verbose(env, "negative offset disallowed for kernel module function call\n");
1827 return ERR_PTR(-EINVAL);
1830 return __find_kfunc_desc_btf(env, offset, btf_modp);
1832 return btf_vmlinux ?: ERR_PTR(-ENOENT);
1835 static int add_kfunc_call(struct bpf_verifier_env *env, u32 func_id, s16 offset)
1837 const struct btf_type *func, *func_proto;
1838 struct bpf_kfunc_btf_tab *btf_tab;
1839 struct bpf_kfunc_desc_tab *tab;
1840 struct bpf_prog_aux *prog_aux;
1841 struct bpf_kfunc_desc *desc;
1842 const char *func_name;
1843 struct btf *desc_btf;
1847 prog_aux = env->prog->aux;
1848 tab = prog_aux->kfunc_tab;
1849 btf_tab = prog_aux->kfunc_btf_tab;
1852 verbose(env, "calling kernel function is not supported without CONFIG_DEBUG_INFO_BTF\n");
1856 if (!env->prog->jit_requested) {
1857 verbose(env, "JIT is required for calling kernel function\n");
1861 if (!bpf_jit_supports_kfunc_call()) {
1862 verbose(env, "JIT does not support calling kernel function\n");
1866 if (!env->prog->gpl_compatible) {
1867 verbose(env, "cannot call kernel function from non-GPL compatible program\n");
1871 tab = kzalloc(sizeof(*tab), GFP_KERNEL);
1874 prog_aux->kfunc_tab = tab;
1877 /* func_id == 0 is always invalid, but instead of returning an error, be
1878 * conservative and wait until the code elimination pass before returning
1879 * error, so that invalid calls that get pruned out can be in BPF programs
1880 * loaded from userspace. It is also required that offset be untouched
1883 if (!func_id && !offset)
1886 if (!btf_tab && offset) {
1887 btf_tab = kzalloc(sizeof(*btf_tab), GFP_KERNEL);
1890 prog_aux->kfunc_btf_tab = btf_tab;
1893 desc_btf = find_kfunc_desc_btf(env, func_id, offset, NULL);
1894 if (IS_ERR(desc_btf)) {
1895 verbose(env, "failed to find BTF for kernel function\n");
1896 return PTR_ERR(desc_btf);
1899 if (find_kfunc_desc(env->prog, func_id, offset))
1902 if (tab->nr_descs == MAX_KFUNC_DESCS) {
1903 verbose(env, "too many different kernel function calls\n");
1907 func = btf_type_by_id(desc_btf, func_id);
1908 if (!func || !btf_type_is_func(func)) {
1909 verbose(env, "kernel btf_id %u is not a function\n",
1913 func_proto = btf_type_by_id(desc_btf, func->type);
1914 if (!func_proto || !btf_type_is_func_proto(func_proto)) {
1915 verbose(env, "kernel function btf_id %u does not have a valid func_proto\n",
1920 func_name = btf_name_by_offset(desc_btf, func->name_off);
1921 addr = kallsyms_lookup_name(func_name);
1923 verbose(env, "cannot find address for kernel function %s\n",
1928 desc = &tab->descs[tab->nr_descs++];
1929 desc->func_id = func_id;
1930 desc->imm = BPF_CALL_IMM(addr);
1931 desc->offset = offset;
1932 err = btf_distill_func_proto(&env->log, desc_btf,
1933 func_proto, func_name,
1936 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
1937 kfunc_desc_cmp_by_id_off, NULL);
1941 static int kfunc_desc_cmp_by_imm(const void *a, const void *b)
1943 const struct bpf_kfunc_desc *d0 = a;
1944 const struct bpf_kfunc_desc *d1 = b;
1946 if (d0->imm > d1->imm)
1948 else if (d0->imm < d1->imm)
1953 static void sort_kfunc_descs_by_imm(struct bpf_prog *prog)
1955 struct bpf_kfunc_desc_tab *tab;
1957 tab = prog->aux->kfunc_tab;
1961 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
1962 kfunc_desc_cmp_by_imm, NULL);
1965 bool bpf_prog_has_kfunc_call(const struct bpf_prog *prog)
1967 return !!prog->aux->kfunc_tab;
1970 const struct btf_func_model *
1971 bpf_jit_find_kfunc_model(const struct bpf_prog *prog,
1972 const struct bpf_insn *insn)
1974 const struct bpf_kfunc_desc desc = {
1977 const struct bpf_kfunc_desc *res;
1978 struct bpf_kfunc_desc_tab *tab;
1980 tab = prog->aux->kfunc_tab;
1981 res = bsearch(&desc, tab->descs, tab->nr_descs,
1982 sizeof(tab->descs[0]), kfunc_desc_cmp_by_imm);
1984 return res ? &res->func_model : NULL;
1987 static int add_subprog_and_kfunc(struct bpf_verifier_env *env)
1989 struct bpf_subprog_info *subprog = env->subprog_info;
1990 struct bpf_insn *insn = env->prog->insnsi;
1991 int i, ret, insn_cnt = env->prog->len;
1993 /* Add entry function. */
1994 ret = add_subprog(env, 0);
1998 for (i = 0; i < insn_cnt; i++, insn++) {
1999 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn) &&
2000 !bpf_pseudo_kfunc_call(insn))
2003 if (!env->bpf_capable) {
2004 verbose(env, "loading/calling other bpf or kernel functions are allowed for CAP_BPF and CAP_SYS_ADMIN\n");
2008 if (bpf_pseudo_func(insn) || bpf_pseudo_call(insn))
2009 ret = add_subprog(env, i + insn->imm + 1);
2011 ret = add_kfunc_call(env, insn->imm, insn->off);
2017 /* Add a fake 'exit' subprog which could simplify subprog iteration
2018 * logic. 'subprog_cnt' should not be increased.
2020 subprog[env->subprog_cnt].start = insn_cnt;
2022 if (env->log.level & BPF_LOG_LEVEL2)
2023 for (i = 0; i < env->subprog_cnt; i++)
2024 verbose(env, "func#%d @%d\n", i, subprog[i].start);
2029 static int check_subprogs(struct bpf_verifier_env *env)
2031 int i, subprog_start, subprog_end, off, cur_subprog = 0;
2032 struct bpf_subprog_info *subprog = env->subprog_info;
2033 struct bpf_insn *insn = env->prog->insnsi;
2034 int insn_cnt = env->prog->len;
2036 /* now check that all jumps are within the same subprog */
2037 subprog_start = subprog[cur_subprog].start;
2038 subprog_end = subprog[cur_subprog + 1].start;
2039 for (i = 0; i < insn_cnt; i++) {
2040 u8 code = insn[i].code;
2042 if (code == (BPF_JMP | BPF_CALL) &&
2043 insn[i].imm == BPF_FUNC_tail_call &&
2044 insn[i].src_reg != BPF_PSEUDO_CALL)
2045 subprog[cur_subprog].has_tail_call = true;
2046 if (BPF_CLASS(code) == BPF_LD &&
2047 (BPF_MODE(code) == BPF_ABS || BPF_MODE(code) == BPF_IND))
2048 subprog[cur_subprog].has_ld_abs = true;
2049 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
2051 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
2053 off = i + insn[i].off + 1;
2054 if (off < subprog_start || off >= subprog_end) {
2055 verbose(env, "jump out of range from insn %d to %d\n", i, off);
2059 if (i == subprog_end - 1) {
2060 /* to avoid fall-through from one subprog into another
2061 * the last insn of the subprog should be either exit
2062 * or unconditional jump back
2064 if (code != (BPF_JMP | BPF_EXIT) &&
2065 code != (BPF_JMP | BPF_JA)) {
2066 verbose(env, "last insn is not an exit or jmp\n");
2069 subprog_start = subprog_end;
2071 if (cur_subprog < env->subprog_cnt)
2072 subprog_end = subprog[cur_subprog + 1].start;
2078 /* Parentage chain of this register (or stack slot) should take care of all
2079 * issues like callee-saved registers, stack slot allocation time, etc.
2081 static int mark_reg_read(struct bpf_verifier_env *env,
2082 const struct bpf_reg_state *state,
2083 struct bpf_reg_state *parent, u8 flag)
2085 bool writes = parent == state->parent; /* Observe write marks */
2089 /* if read wasn't screened by an earlier write ... */
2090 if (writes && state->live & REG_LIVE_WRITTEN)
2092 if (parent->live & REG_LIVE_DONE) {
2093 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
2094 reg_type_str(env, parent->type),
2095 parent->var_off.value, parent->off);
2098 /* The first condition is more likely to be true than the
2099 * second, checked it first.
2101 if ((parent->live & REG_LIVE_READ) == flag ||
2102 parent->live & REG_LIVE_READ64)
2103 /* The parentage chain never changes and
2104 * this parent was already marked as LIVE_READ.
2105 * There is no need to keep walking the chain again and
2106 * keep re-marking all parents as LIVE_READ.
2107 * This case happens when the same register is read
2108 * multiple times without writes into it in-between.
2109 * Also, if parent has the stronger REG_LIVE_READ64 set,
2110 * then no need to set the weak REG_LIVE_READ32.
2113 /* ... then we depend on parent's value */
2114 parent->live |= flag;
2115 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
2116 if (flag == REG_LIVE_READ64)
2117 parent->live &= ~REG_LIVE_READ32;
2119 parent = state->parent;
2124 if (env->longest_mark_read_walk < cnt)
2125 env->longest_mark_read_walk = cnt;
2129 /* This function is supposed to be used by the following 32-bit optimization
2130 * code only. It returns TRUE if the source or destination register operates
2131 * on 64-bit, otherwise return FALSE.
2133 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
2134 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
2139 class = BPF_CLASS(code);
2141 if (class == BPF_JMP) {
2142 /* BPF_EXIT for "main" will reach here. Return TRUE
2147 if (op == BPF_CALL) {
2148 /* BPF to BPF call will reach here because of marking
2149 * caller saved clobber with DST_OP_NO_MARK for which we
2150 * don't care the register def because they are anyway
2151 * marked as NOT_INIT already.
2153 if (insn->src_reg == BPF_PSEUDO_CALL)
2155 /* Helper call will reach here because of arg type
2156 * check, conservatively return TRUE.
2165 if (class == BPF_ALU64 || class == BPF_JMP ||
2166 /* BPF_END always use BPF_ALU class. */
2167 (class == BPF_ALU && op == BPF_END && insn->imm == 64))
2170 if (class == BPF_ALU || class == BPF_JMP32)
2173 if (class == BPF_LDX) {
2175 return BPF_SIZE(code) == BPF_DW;
2176 /* LDX source must be ptr. */
2180 if (class == BPF_STX) {
2181 /* BPF_STX (including atomic variants) has multiple source
2182 * operands, one of which is a ptr. Check whether the caller is
2185 if (t == SRC_OP && reg->type != SCALAR_VALUE)
2187 return BPF_SIZE(code) == BPF_DW;
2190 if (class == BPF_LD) {
2191 u8 mode = BPF_MODE(code);
2194 if (mode == BPF_IMM)
2197 /* Both LD_IND and LD_ABS return 32-bit data. */
2201 /* Implicit ctx ptr. */
2202 if (regno == BPF_REG_6)
2205 /* Explicit source could be any width. */
2209 if (class == BPF_ST)
2210 /* The only source register for BPF_ST is a ptr. */
2213 /* Conservatively return true at default. */
2217 /* Return the regno defined by the insn, or -1. */
2218 static int insn_def_regno(const struct bpf_insn *insn)
2220 switch (BPF_CLASS(insn->code)) {
2226 if (BPF_MODE(insn->code) == BPF_ATOMIC &&
2227 (insn->imm & BPF_FETCH)) {
2228 if (insn->imm == BPF_CMPXCHG)
2231 return insn->src_reg;
2236 return insn->dst_reg;
2240 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
2241 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
2243 int dst_reg = insn_def_regno(insn);
2248 return !is_reg64(env, insn, dst_reg, NULL, DST_OP);
2251 static void mark_insn_zext(struct bpf_verifier_env *env,
2252 struct bpf_reg_state *reg)
2254 s32 def_idx = reg->subreg_def;
2256 if (def_idx == DEF_NOT_SUBREG)
2259 env->insn_aux_data[def_idx - 1].zext_dst = true;
2260 /* The dst will be zero extended, so won't be sub-register anymore. */
2261 reg->subreg_def = DEF_NOT_SUBREG;
2264 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
2265 enum reg_arg_type t)
2267 struct bpf_verifier_state *vstate = env->cur_state;
2268 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2269 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
2270 struct bpf_reg_state *reg, *regs = state->regs;
2273 if (regno >= MAX_BPF_REG) {
2274 verbose(env, "R%d is invalid\n", regno);
2278 mark_reg_scratched(env, regno);
2281 rw64 = is_reg64(env, insn, regno, reg, t);
2283 /* check whether register used as source operand can be read */
2284 if (reg->type == NOT_INIT) {
2285 verbose(env, "R%d !read_ok\n", regno);
2288 /* We don't need to worry about FP liveness because it's read-only */
2289 if (regno == BPF_REG_FP)
2293 mark_insn_zext(env, reg);
2295 return mark_reg_read(env, reg, reg->parent,
2296 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
2298 /* check whether register used as dest operand can be written to */
2299 if (regno == BPF_REG_FP) {
2300 verbose(env, "frame pointer is read only\n");
2303 reg->live |= REG_LIVE_WRITTEN;
2304 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
2306 mark_reg_unknown(env, regs, regno);
2311 /* for any branch, call, exit record the history of jmps in the given state */
2312 static int push_jmp_history(struct bpf_verifier_env *env,
2313 struct bpf_verifier_state *cur)
2315 u32 cnt = cur->jmp_history_cnt;
2316 struct bpf_idx_pair *p;
2319 p = krealloc(cur->jmp_history, cnt * sizeof(*p), GFP_USER);
2322 p[cnt - 1].idx = env->insn_idx;
2323 p[cnt - 1].prev_idx = env->prev_insn_idx;
2324 cur->jmp_history = p;
2325 cur->jmp_history_cnt = cnt;
2329 /* Backtrack one insn at a time. If idx is not at the top of recorded
2330 * history then previous instruction came from straight line execution.
2332 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
2337 if (cnt && st->jmp_history[cnt - 1].idx == i) {
2338 i = st->jmp_history[cnt - 1].prev_idx;
2346 static const char *disasm_kfunc_name(void *data, const struct bpf_insn *insn)
2348 const struct btf_type *func;
2349 struct btf *desc_btf;
2351 if (insn->src_reg != BPF_PSEUDO_KFUNC_CALL)
2354 desc_btf = find_kfunc_desc_btf(data, insn->imm, insn->off, NULL);
2355 if (IS_ERR(desc_btf))
2358 func = btf_type_by_id(desc_btf, insn->imm);
2359 return btf_name_by_offset(desc_btf, func->name_off);
2362 /* For given verifier state backtrack_insn() is called from the last insn to
2363 * the first insn. Its purpose is to compute a bitmask of registers and
2364 * stack slots that needs precision in the parent verifier state.
2366 static int backtrack_insn(struct bpf_verifier_env *env, int idx,
2367 u32 *reg_mask, u64 *stack_mask)
2369 const struct bpf_insn_cbs cbs = {
2370 .cb_call = disasm_kfunc_name,
2371 .cb_print = verbose,
2372 .private_data = env,
2374 struct bpf_insn *insn = env->prog->insnsi + idx;
2375 u8 class = BPF_CLASS(insn->code);
2376 u8 opcode = BPF_OP(insn->code);
2377 u8 mode = BPF_MODE(insn->code);
2378 u32 dreg = 1u << insn->dst_reg;
2379 u32 sreg = 1u << insn->src_reg;
2382 if (insn->code == 0)
2384 if (env->log.level & BPF_LOG_LEVEL2) {
2385 verbose(env, "regs=%x stack=%llx before ", *reg_mask, *stack_mask);
2386 verbose(env, "%d: ", idx);
2387 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
2390 if (class == BPF_ALU || class == BPF_ALU64) {
2391 if (!(*reg_mask & dreg))
2393 if (opcode == BPF_MOV) {
2394 if (BPF_SRC(insn->code) == BPF_X) {
2396 * dreg needs precision after this insn
2397 * sreg needs precision before this insn
2403 * dreg needs precision after this insn.
2404 * Corresponding register is already marked
2405 * as precise=true in this verifier state.
2406 * No further markings in parent are necessary
2411 if (BPF_SRC(insn->code) == BPF_X) {
2413 * both dreg and sreg need precision
2418 * dreg still needs precision before this insn
2421 } else if (class == BPF_LDX) {
2422 if (!(*reg_mask & dreg))
2426 /* scalars can only be spilled into stack w/o losing precision.
2427 * Load from any other memory can be zero extended.
2428 * The desire to keep that precision is already indicated
2429 * by 'precise' mark in corresponding register of this state.
2430 * No further tracking necessary.
2432 if (insn->src_reg != BPF_REG_FP)
2435 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
2436 * that [fp - off] slot contains scalar that needs to be
2437 * tracked with precision
2439 spi = (-insn->off - 1) / BPF_REG_SIZE;
2441 verbose(env, "BUG spi %d\n", spi);
2442 WARN_ONCE(1, "verifier backtracking bug");
2445 *stack_mask |= 1ull << spi;
2446 } else if (class == BPF_STX || class == BPF_ST) {
2447 if (*reg_mask & dreg)
2448 /* stx & st shouldn't be using _scalar_ dst_reg
2449 * to access memory. It means backtracking
2450 * encountered a case of pointer subtraction.
2453 /* scalars can only be spilled into stack */
2454 if (insn->dst_reg != BPF_REG_FP)
2456 spi = (-insn->off - 1) / BPF_REG_SIZE;
2458 verbose(env, "BUG spi %d\n", spi);
2459 WARN_ONCE(1, "verifier backtracking bug");
2462 if (!(*stack_mask & (1ull << spi)))
2464 *stack_mask &= ~(1ull << spi);
2465 if (class == BPF_STX)
2467 } else if (class == BPF_JMP || class == BPF_JMP32) {
2468 if (opcode == BPF_CALL) {
2469 if (insn->src_reg == BPF_PSEUDO_CALL)
2471 /* regular helper call sets R0 */
2473 if (*reg_mask & 0x3f) {
2474 /* if backtracing was looking for registers R1-R5
2475 * they should have been found already.
2477 verbose(env, "BUG regs %x\n", *reg_mask);
2478 WARN_ONCE(1, "verifier backtracking bug");
2481 } else if (opcode == BPF_EXIT) {
2484 } else if (class == BPF_LD) {
2485 if (!(*reg_mask & dreg))
2488 /* It's ld_imm64 or ld_abs or ld_ind.
2489 * For ld_imm64 no further tracking of precision
2490 * into parent is necessary
2492 if (mode == BPF_IND || mode == BPF_ABS)
2493 /* to be analyzed */
2499 /* the scalar precision tracking algorithm:
2500 * . at the start all registers have precise=false.
2501 * . scalar ranges are tracked as normal through alu and jmp insns.
2502 * . once precise value of the scalar register is used in:
2503 * . ptr + scalar alu
2504 * . if (scalar cond K|scalar)
2505 * . helper_call(.., scalar, ...) where ARG_CONST is expected
2506 * backtrack through the verifier states and mark all registers and
2507 * stack slots with spilled constants that these scalar regisers
2508 * should be precise.
2509 * . during state pruning two registers (or spilled stack slots)
2510 * are equivalent if both are not precise.
2512 * Note the verifier cannot simply walk register parentage chain,
2513 * since many different registers and stack slots could have been
2514 * used to compute single precise scalar.
2516 * The approach of starting with precise=true for all registers and then
2517 * backtrack to mark a register as not precise when the verifier detects
2518 * that program doesn't care about specific value (e.g., when helper
2519 * takes register as ARG_ANYTHING parameter) is not safe.
2521 * It's ok to walk single parentage chain of the verifier states.
2522 * It's possible that this backtracking will go all the way till 1st insn.
2523 * All other branches will be explored for needing precision later.
2525 * The backtracking needs to deal with cases like:
2526 * 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)
2529 * if r5 > 0x79f goto pc+7
2530 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
2533 * call bpf_perf_event_output#25
2534 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
2538 * call foo // uses callee's r6 inside to compute r0
2542 * to track above reg_mask/stack_mask needs to be independent for each frame.
2544 * Also if parent's curframe > frame where backtracking started,
2545 * the verifier need to mark registers in both frames, otherwise callees
2546 * may incorrectly prune callers. This is similar to
2547 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
2549 * For now backtracking falls back into conservative marking.
2551 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
2552 struct bpf_verifier_state *st)
2554 struct bpf_func_state *func;
2555 struct bpf_reg_state *reg;
2558 /* big hammer: mark all scalars precise in this path.
2559 * pop_stack may still get !precise scalars.
2561 for (; st; st = st->parent)
2562 for (i = 0; i <= st->curframe; i++) {
2563 func = st->frame[i];
2564 for (j = 0; j < BPF_REG_FP; j++) {
2565 reg = &func->regs[j];
2566 if (reg->type != SCALAR_VALUE)
2568 reg->precise = true;
2570 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
2571 if (!is_spilled_reg(&func->stack[j]))
2573 reg = &func->stack[j].spilled_ptr;
2574 if (reg->type != SCALAR_VALUE)
2576 reg->precise = true;
2581 static int __mark_chain_precision(struct bpf_verifier_env *env, int regno,
2584 struct bpf_verifier_state *st = env->cur_state;
2585 int first_idx = st->first_insn_idx;
2586 int last_idx = env->insn_idx;
2587 struct bpf_func_state *func;
2588 struct bpf_reg_state *reg;
2589 u32 reg_mask = regno >= 0 ? 1u << regno : 0;
2590 u64 stack_mask = spi >= 0 ? 1ull << spi : 0;
2591 bool skip_first = true;
2592 bool new_marks = false;
2595 if (!env->bpf_capable)
2598 func = st->frame[st->curframe];
2600 reg = &func->regs[regno];
2601 if (reg->type != SCALAR_VALUE) {
2602 WARN_ONCE(1, "backtracing misuse");
2609 reg->precise = true;
2613 if (!is_spilled_reg(&func->stack[spi])) {
2617 reg = &func->stack[spi].spilled_ptr;
2618 if (reg->type != SCALAR_VALUE) {
2626 reg->precise = true;
2632 if (!reg_mask && !stack_mask)
2635 DECLARE_BITMAP(mask, 64);
2636 u32 history = st->jmp_history_cnt;
2638 if (env->log.level & BPF_LOG_LEVEL2)
2639 verbose(env, "last_idx %d first_idx %d\n", last_idx, first_idx);
2640 for (i = last_idx;;) {
2645 err = backtrack_insn(env, i, ®_mask, &stack_mask);
2647 if (err == -ENOTSUPP) {
2648 mark_all_scalars_precise(env, st);
2653 if (!reg_mask && !stack_mask)
2654 /* Found assignment(s) into tracked register in this state.
2655 * Since this state is already marked, just return.
2656 * Nothing to be tracked further in the parent state.
2661 i = get_prev_insn_idx(st, i, &history);
2662 if (i >= env->prog->len) {
2663 /* This can happen if backtracking reached insn 0
2664 * and there are still reg_mask or stack_mask
2666 * It means the backtracking missed the spot where
2667 * particular register was initialized with a constant.
2669 verbose(env, "BUG backtracking idx %d\n", i);
2670 WARN_ONCE(1, "verifier backtracking bug");
2679 func = st->frame[st->curframe];
2680 bitmap_from_u64(mask, reg_mask);
2681 for_each_set_bit(i, mask, 32) {
2682 reg = &func->regs[i];
2683 if (reg->type != SCALAR_VALUE) {
2684 reg_mask &= ~(1u << i);
2689 reg->precise = true;
2692 bitmap_from_u64(mask, stack_mask);
2693 for_each_set_bit(i, mask, 64) {
2694 if (i >= func->allocated_stack / BPF_REG_SIZE) {
2695 /* the sequence of instructions:
2697 * 3: (7b) *(u64 *)(r3 -8) = r0
2698 * 4: (79) r4 = *(u64 *)(r10 -8)
2699 * doesn't contain jmps. It's backtracked
2700 * as a single block.
2701 * During backtracking insn 3 is not recognized as
2702 * stack access, so at the end of backtracking
2703 * stack slot fp-8 is still marked in stack_mask.
2704 * However the parent state may not have accessed
2705 * fp-8 and it's "unallocated" stack space.
2706 * In such case fallback to conservative.
2708 mark_all_scalars_precise(env, st);
2712 if (!is_spilled_reg(&func->stack[i])) {
2713 stack_mask &= ~(1ull << i);
2716 reg = &func->stack[i].spilled_ptr;
2717 if (reg->type != SCALAR_VALUE) {
2718 stack_mask &= ~(1ull << i);
2723 reg->precise = true;
2725 if (env->log.level & BPF_LOG_LEVEL2) {
2726 verbose(env, "parent %s regs=%x stack=%llx marks:",
2727 new_marks ? "didn't have" : "already had",
2728 reg_mask, stack_mask);
2729 print_verifier_state(env, func, true);
2732 if (!reg_mask && !stack_mask)
2737 last_idx = st->last_insn_idx;
2738 first_idx = st->first_insn_idx;
2743 static int mark_chain_precision(struct bpf_verifier_env *env, int regno)
2745 return __mark_chain_precision(env, regno, -1);
2748 static int mark_chain_precision_stack(struct bpf_verifier_env *env, int spi)
2750 return __mark_chain_precision(env, -1, spi);
2753 static bool is_spillable_regtype(enum bpf_reg_type type)
2755 switch (base_type(type)) {
2756 case PTR_TO_MAP_VALUE:
2760 case PTR_TO_PACKET_META:
2761 case PTR_TO_PACKET_END:
2762 case PTR_TO_FLOW_KEYS:
2763 case CONST_PTR_TO_MAP:
2765 case PTR_TO_SOCK_COMMON:
2766 case PTR_TO_TCP_SOCK:
2767 case PTR_TO_XDP_SOCK:
2770 case PTR_TO_PERCPU_BTF_ID:
2773 case PTR_TO_MAP_KEY:
2780 /* Does this register contain a constant zero? */
2781 static bool register_is_null(struct bpf_reg_state *reg)
2783 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
2786 static bool register_is_const(struct bpf_reg_state *reg)
2788 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
2791 static bool __is_scalar_unbounded(struct bpf_reg_state *reg)
2793 return tnum_is_unknown(reg->var_off) &&
2794 reg->smin_value == S64_MIN && reg->smax_value == S64_MAX &&
2795 reg->umin_value == 0 && reg->umax_value == U64_MAX &&
2796 reg->s32_min_value == S32_MIN && reg->s32_max_value == S32_MAX &&
2797 reg->u32_min_value == 0 && reg->u32_max_value == U32_MAX;
2800 static bool register_is_bounded(struct bpf_reg_state *reg)
2802 return reg->type == SCALAR_VALUE && !__is_scalar_unbounded(reg);
2805 static bool __is_pointer_value(bool allow_ptr_leaks,
2806 const struct bpf_reg_state *reg)
2808 if (allow_ptr_leaks)
2811 return reg->type != SCALAR_VALUE;
2814 static void save_register_state(struct bpf_func_state *state,
2815 int spi, struct bpf_reg_state *reg,
2820 state->stack[spi].spilled_ptr = *reg;
2821 if (size == BPF_REG_SIZE)
2822 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2824 for (i = BPF_REG_SIZE; i > BPF_REG_SIZE - size; i--)
2825 state->stack[spi].slot_type[i - 1] = STACK_SPILL;
2827 /* size < 8 bytes spill */
2829 scrub_spilled_slot(&state->stack[spi].slot_type[i - 1]);
2832 /* check_stack_{read,write}_fixed_off functions track spill/fill of registers,
2833 * stack boundary and alignment are checked in check_mem_access()
2835 static int check_stack_write_fixed_off(struct bpf_verifier_env *env,
2836 /* stack frame we're writing to */
2837 struct bpf_func_state *state,
2838 int off, int size, int value_regno,
2841 struct bpf_func_state *cur; /* state of the current function */
2842 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
2843 u32 dst_reg = env->prog->insnsi[insn_idx].dst_reg;
2844 struct bpf_reg_state *reg = NULL;
2846 err = grow_stack_state(state, round_up(slot + 1, BPF_REG_SIZE));
2849 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
2850 * so it's aligned access and [off, off + size) are within stack limits
2852 if (!env->allow_ptr_leaks &&
2853 state->stack[spi].slot_type[0] == STACK_SPILL &&
2854 size != BPF_REG_SIZE) {
2855 verbose(env, "attempt to corrupt spilled pointer on stack\n");
2859 cur = env->cur_state->frame[env->cur_state->curframe];
2860 if (value_regno >= 0)
2861 reg = &cur->regs[value_regno];
2862 if (!env->bypass_spec_v4) {
2863 bool sanitize = reg && is_spillable_regtype(reg->type);
2865 for (i = 0; i < size; i++) {
2866 if (state->stack[spi].slot_type[i] == STACK_INVALID) {
2873 env->insn_aux_data[insn_idx].sanitize_stack_spill = true;
2876 mark_stack_slot_scratched(env, spi);
2877 if (reg && !(off % BPF_REG_SIZE) && register_is_bounded(reg) &&
2878 !register_is_null(reg) && env->bpf_capable) {
2879 if (dst_reg != BPF_REG_FP) {
2880 /* The backtracking logic can only recognize explicit
2881 * stack slot address like [fp - 8]. Other spill of
2882 * scalar via different register has to be conservative.
2883 * Backtrack from here and mark all registers as precise
2884 * that contributed into 'reg' being a constant.
2886 err = mark_chain_precision(env, value_regno);
2890 save_register_state(state, spi, reg, size);
2891 } else if (reg && is_spillable_regtype(reg->type)) {
2892 /* register containing pointer is being spilled into stack */
2893 if (size != BPF_REG_SIZE) {
2894 verbose_linfo(env, insn_idx, "; ");
2895 verbose(env, "invalid size of register spill\n");
2898 if (state != cur && reg->type == PTR_TO_STACK) {
2899 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
2902 save_register_state(state, spi, reg, size);
2904 u8 type = STACK_MISC;
2906 /* regular write of data into stack destroys any spilled ptr */
2907 state->stack[spi].spilled_ptr.type = NOT_INIT;
2908 /* Mark slots as STACK_MISC if they belonged to spilled ptr. */
2909 if (is_spilled_reg(&state->stack[spi]))
2910 for (i = 0; i < BPF_REG_SIZE; i++)
2911 scrub_spilled_slot(&state->stack[spi].slot_type[i]);
2913 /* only mark the slot as written if all 8 bytes were written
2914 * otherwise read propagation may incorrectly stop too soon
2915 * when stack slots are partially written.
2916 * This heuristic means that read propagation will be
2917 * conservative, since it will add reg_live_read marks
2918 * to stack slots all the way to first state when programs
2919 * writes+reads less than 8 bytes
2921 if (size == BPF_REG_SIZE)
2922 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2924 /* when we zero initialize stack slots mark them as such */
2925 if (reg && register_is_null(reg)) {
2926 /* backtracking doesn't work for STACK_ZERO yet. */
2927 err = mark_chain_precision(env, value_regno);
2933 /* Mark slots affected by this stack write. */
2934 for (i = 0; i < size; i++)
2935 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
2941 /* Write the stack: 'stack[ptr_regno + off] = value_regno'. 'ptr_regno' is
2942 * known to contain a variable offset.
2943 * This function checks whether the write is permitted and conservatively
2944 * tracks the effects of the write, considering that each stack slot in the
2945 * dynamic range is potentially written to.
2947 * 'off' includes 'regno->off'.
2948 * 'value_regno' can be -1, meaning that an unknown value is being written to
2951 * Spilled pointers in range are not marked as written because we don't know
2952 * what's going to be actually written. This means that read propagation for
2953 * future reads cannot be terminated by this write.
2955 * For privileged programs, uninitialized stack slots are considered
2956 * initialized by this write (even though we don't know exactly what offsets
2957 * are going to be written to). The idea is that we don't want the verifier to
2958 * reject future reads that access slots written to through variable offsets.
2960 static int check_stack_write_var_off(struct bpf_verifier_env *env,
2961 /* func where register points to */
2962 struct bpf_func_state *state,
2963 int ptr_regno, int off, int size,
2964 int value_regno, int insn_idx)
2966 struct bpf_func_state *cur; /* state of the current function */
2967 int min_off, max_off;
2969 struct bpf_reg_state *ptr_reg = NULL, *value_reg = NULL;
2970 bool writing_zero = false;
2971 /* set if the fact that we're writing a zero is used to let any
2972 * stack slots remain STACK_ZERO
2974 bool zero_used = false;
2976 cur = env->cur_state->frame[env->cur_state->curframe];
2977 ptr_reg = &cur->regs[ptr_regno];
2978 min_off = ptr_reg->smin_value + off;
2979 max_off = ptr_reg->smax_value + off + size;
2980 if (value_regno >= 0)
2981 value_reg = &cur->regs[value_regno];
2982 if (value_reg && register_is_null(value_reg))
2983 writing_zero = true;
2985 err = grow_stack_state(state, round_up(-min_off, BPF_REG_SIZE));
2990 /* Variable offset writes destroy any spilled pointers in range. */
2991 for (i = min_off; i < max_off; i++) {
2992 u8 new_type, *stype;
2996 spi = slot / BPF_REG_SIZE;
2997 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
2998 mark_stack_slot_scratched(env, spi);
3000 if (!env->allow_ptr_leaks
3001 && *stype != NOT_INIT
3002 && *stype != SCALAR_VALUE) {
3003 /* Reject the write if there's are spilled pointers in
3004 * range. If we didn't reject here, the ptr status
3005 * would be erased below (even though not all slots are
3006 * actually overwritten), possibly opening the door to
3009 verbose(env, "spilled ptr in range of var-offset stack write; insn %d, ptr off: %d",
3014 /* Erase all spilled pointers. */
3015 state->stack[spi].spilled_ptr.type = NOT_INIT;
3017 /* Update the slot type. */
3018 new_type = STACK_MISC;
3019 if (writing_zero && *stype == STACK_ZERO) {
3020 new_type = STACK_ZERO;
3023 /* If the slot is STACK_INVALID, we check whether it's OK to
3024 * pretend that it will be initialized by this write. The slot
3025 * might not actually be written to, and so if we mark it as
3026 * initialized future reads might leak uninitialized memory.
3027 * For privileged programs, we will accept such reads to slots
3028 * that may or may not be written because, if we're reject
3029 * them, the error would be too confusing.
3031 if (*stype == STACK_INVALID && !env->allow_uninit_stack) {
3032 verbose(env, "uninit stack in range of var-offset write prohibited for !root; insn %d, off: %d",
3039 /* backtracking doesn't work for STACK_ZERO yet. */
3040 err = mark_chain_precision(env, value_regno);
3047 /* When register 'dst_regno' is assigned some values from stack[min_off,
3048 * max_off), we set the register's type according to the types of the
3049 * respective stack slots. If all the stack values are known to be zeros, then
3050 * so is the destination reg. Otherwise, the register is considered to be
3051 * SCALAR. This function does not deal with register filling; the caller must
3052 * ensure that all spilled registers in the stack range have been marked as
3055 static void mark_reg_stack_read(struct bpf_verifier_env *env,
3056 /* func where src register points to */
3057 struct bpf_func_state *ptr_state,
3058 int min_off, int max_off, int dst_regno)
3060 struct bpf_verifier_state *vstate = env->cur_state;
3061 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3066 for (i = min_off; i < max_off; i++) {
3068 spi = slot / BPF_REG_SIZE;
3069 stype = ptr_state->stack[spi].slot_type;
3070 if (stype[slot % BPF_REG_SIZE] != STACK_ZERO)
3074 if (zeros == max_off - min_off) {
3075 /* any access_size read into register is zero extended,
3076 * so the whole register == const_zero
3078 __mark_reg_const_zero(&state->regs[dst_regno]);
3079 /* backtracking doesn't support STACK_ZERO yet,
3080 * so mark it precise here, so that later
3081 * backtracking can stop here.
3082 * Backtracking may not need this if this register
3083 * doesn't participate in pointer adjustment.
3084 * Forward propagation of precise flag is not
3085 * necessary either. This mark is only to stop
3086 * backtracking. Any register that contributed
3087 * to const 0 was marked precise before spill.
3089 state->regs[dst_regno].precise = true;
3091 /* have read misc data from the stack */
3092 mark_reg_unknown(env, state->regs, dst_regno);
3094 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
3097 /* Read the stack at 'off' and put the results into the register indicated by
3098 * 'dst_regno'. It handles reg filling if the addressed stack slot is a
3101 * 'dst_regno' can be -1, meaning that the read value is not going to a
3104 * The access is assumed to be within the current stack bounds.
3106 static int check_stack_read_fixed_off(struct bpf_verifier_env *env,
3107 /* func where src register points to */
3108 struct bpf_func_state *reg_state,
3109 int off, int size, int dst_regno)
3111 struct bpf_verifier_state *vstate = env->cur_state;
3112 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3113 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
3114 struct bpf_reg_state *reg;
3117 stype = reg_state->stack[spi].slot_type;
3118 reg = ®_state->stack[spi].spilled_ptr;
3120 if (is_spilled_reg(®_state->stack[spi])) {
3123 for (i = BPF_REG_SIZE - 1; i > 0 && stype[i - 1] == STACK_SPILL; i--)
3126 if (size != BPF_REG_SIZE || spill_size != BPF_REG_SIZE) {
3127 if (reg->type != SCALAR_VALUE) {
3128 verbose_linfo(env, env->insn_idx, "; ");
3129 verbose(env, "invalid size of register fill\n");
3133 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
3137 if (!(off % BPF_REG_SIZE) && size == spill_size) {
3138 /* The earlier check_reg_arg() has decided the
3139 * subreg_def for this insn. Save it first.
3141 s32 subreg_def = state->regs[dst_regno].subreg_def;
3143 state->regs[dst_regno] = *reg;
3144 state->regs[dst_regno].subreg_def = subreg_def;
3146 for (i = 0; i < size; i++) {
3147 type = stype[(slot - i) % BPF_REG_SIZE];
3148 if (type == STACK_SPILL)
3150 if (type == STACK_MISC)
3152 verbose(env, "invalid read from stack off %d+%d size %d\n",
3156 mark_reg_unknown(env, state->regs, dst_regno);
3158 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
3162 if (dst_regno >= 0) {
3163 /* restore register state from stack */
3164 state->regs[dst_regno] = *reg;
3165 /* mark reg as written since spilled pointer state likely
3166 * has its liveness marks cleared by is_state_visited()
3167 * which resets stack/reg liveness for state transitions
3169 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
3170 } else if (__is_pointer_value(env->allow_ptr_leaks, reg)) {
3171 /* If dst_regno==-1, the caller is asking us whether
3172 * it is acceptable to use this value as a SCALAR_VALUE
3174 * We must not allow unprivileged callers to do that
3175 * with spilled pointers.
3177 verbose(env, "leaking pointer from stack off %d\n",
3181 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
3183 for (i = 0; i < size; i++) {
3184 type = stype[(slot - i) % BPF_REG_SIZE];
3185 if (type == STACK_MISC)
3187 if (type == STACK_ZERO)
3189 verbose(env, "invalid read from stack off %d+%d size %d\n",
3193 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
3195 mark_reg_stack_read(env, reg_state, off, off + size, dst_regno);
3200 enum stack_access_src {
3201 ACCESS_DIRECT = 1, /* the access is performed by an instruction */
3202 ACCESS_HELPER = 2, /* the access is performed by a helper */
3205 static int check_stack_range_initialized(struct bpf_verifier_env *env,
3206 int regno, int off, int access_size,
3207 bool zero_size_allowed,
3208 enum stack_access_src type,
3209 struct bpf_call_arg_meta *meta);
3211 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
3213 return cur_regs(env) + regno;
3216 /* Read the stack at 'ptr_regno + off' and put the result into the register
3218 * 'off' includes the pointer register's fixed offset(i.e. 'ptr_regno.off'),
3219 * but not its variable offset.
3220 * 'size' is assumed to be <= reg size and the access is assumed to be aligned.
3222 * As opposed to check_stack_read_fixed_off, this function doesn't deal with
3223 * filling registers (i.e. reads of spilled register cannot be detected when
3224 * the offset is not fixed). We conservatively mark 'dst_regno' as containing
3225 * SCALAR_VALUE. That's why we assert that the 'ptr_regno' has a variable
3226 * offset; for a fixed offset check_stack_read_fixed_off should be used
3229 static int check_stack_read_var_off(struct bpf_verifier_env *env,
3230 int ptr_regno, int off, int size, int dst_regno)
3232 /* The state of the source register. */
3233 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
3234 struct bpf_func_state *ptr_state = func(env, reg);
3236 int min_off, max_off;
3238 /* Note that we pass a NULL meta, so raw access will not be permitted.
3240 err = check_stack_range_initialized(env, ptr_regno, off, size,
3241 false, ACCESS_DIRECT, NULL);
3245 min_off = reg->smin_value + off;
3246 max_off = reg->smax_value + off;
3247 mark_reg_stack_read(env, ptr_state, min_off, max_off + size, dst_regno);
3251 /* check_stack_read dispatches to check_stack_read_fixed_off or
3252 * check_stack_read_var_off.
3254 * The caller must ensure that the offset falls within the allocated stack
3257 * 'dst_regno' is a register which will receive the value from the stack. It
3258 * can be -1, meaning that the read value is not going to a register.
3260 static int check_stack_read(struct bpf_verifier_env *env,
3261 int ptr_regno, int off, int size,
3264 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
3265 struct bpf_func_state *state = func(env, reg);
3267 /* Some accesses are only permitted with a static offset. */
3268 bool var_off = !tnum_is_const(reg->var_off);
3270 /* The offset is required to be static when reads don't go to a
3271 * register, in order to not leak pointers (see
3272 * check_stack_read_fixed_off).
3274 if (dst_regno < 0 && var_off) {
3277 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3278 verbose(env, "variable offset stack pointer cannot be passed into helper function; var_off=%s off=%d size=%d\n",
3282 /* Variable offset is prohibited for unprivileged mode for simplicity
3283 * since it requires corresponding support in Spectre masking for stack
3284 * ALU. See also retrieve_ptr_limit().
3286 if (!env->bypass_spec_v1 && var_off) {
3289 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3290 verbose(env, "R%d variable offset stack access prohibited for !root, var_off=%s\n",
3296 off += reg->var_off.value;
3297 err = check_stack_read_fixed_off(env, state, off, size,
3300 /* Variable offset stack reads need more conservative handling
3301 * than fixed offset ones. Note that dst_regno >= 0 on this
3304 err = check_stack_read_var_off(env, ptr_regno, off, size,
3311 /* check_stack_write dispatches to check_stack_write_fixed_off or
3312 * check_stack_write_var_off.
3314 * 'ptr_regno' is the register used as a pointer into the stack.
3315 * 'off' includes 'ptr_regno->off', but not its variable offset (if any).
3316 * 'value_regno' is the register whose value we're writing to the stack. It can
3317 * be -1, meaning that we're not writing from a register.
3319 * The caller must ensure that the offset falls within the maximum stack size.
3321 static int check_stack_write(struct bpf_verifier_env *env,
3322 int ptr_regno, int off, int size,
3323 int value_regno, int insn_idx)
3325 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
3326 struct bpf_func_state *state = func(env, reg);
3329 if (tnum_is_const(reg->var_off)) {
3330 off += reg->var_off.value;
3331 err = check_stack_write_fixed_off(env, state, off, size,
3332 value_regno, insn_idx);
3334 /* Variable offset stack reads need more conservative handling
3335 * than fixed offset ones.
3337 err = check_stack_write_var_off(env, state,
3338 ptr_regno, off, size,
3339 value_regno, insn_idx);
3344 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
3345 int off, int size, enum bpf_access_type type)
3347 struct bpf_reg_state *regs = cur_regs(env);
3348 struct bpf_map *map = regs[regno].map_ptr;
3349 u32 cap = bpf_map_flags_to_cap(map);
3351 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
3352 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
3353 map->value_size, off, size);
3357 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
3358 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
3359 map->value_size, off, size);
3366 /* check read/write into memory region (e.g., map value, ringbuf sample, etc) */
3367 static int __check_mem_access(struct bpf_verifier_env *env, int regno,
3368 int off, int size, u32 mem_size,
3369 bool zero_size_allowed)
3371 bool size_ok = size > 0 || (size == 0 && zero_size_allowed);
3372 struct bpf_reg_state *reg;
3374 if (off >= 0 && size_ok && (u64)off + size <= mem_size)
3377 reg = &cur_regs(env)[regno];
3378 switch (reg->type) {
3379 case PTR_TO_MAP_KEY:
3380 verbose(env, "invalid access to map key, key_size=%d off=%d size=%d\n",
3381 mem_size, off, size);
3383 case PTR_TO_MAP_VALUE:
3384 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
3385 mem_size, off, size);
3388 case PTR_TO_PACKET_META:
3389 case PTR_TO_PACKET_END:
3390 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
3391 off, size, regno, reg->id, off, mem_size);
3395 verbose(env, "invalid access to memory, mem_size=%u off=%d size=%d\n",
3396 mem_size, off, size);
3402 /* check read/write into a memory region with possible variable offset */
3403 static int check_mem_region_access(struct bpf_verifier_env *env, u32 regno,
3404 int off, int size, u32 mem_size,
3405 bool zero_size_allowed)
3407 struct bpf_verifier_state *vstate = env->cur_state;
3408 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3409 struct bpf_reg_state *reg = &state->regs[regno];
3412 /* We may have adjusted the register pointing to memory region, so we
3413 * need to try adding each of min_value and max_value to off
3414 * to make sure our theoretical access will be safe.
3416 * The minimum value is only important with signed
3417 * comparisons where we can't assume the floor of a
3418 * value is 0. If we are using signed variables for our
3419 * index'es we need to make sure that whatever we use
3420 * will have a set floor within our range.
3422 if (reg->smin_value < 0 &&
3423 (reg->smin_value == S64_MIN ||
3424 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
3425 reg->smin_value + off < 0)) {
3426 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3430 err = __check_mem_access(env, regno, reg->smin_value + off, size,
3431 mem_size, zero_size_allowed);
3433 verbose(env, "R%d min value is outside of the allowed memory range\n",
3438 /* If we haven't set a max value then we need to bail since we can't be
3439 * sure we won't do bad things.
3440 * If reg->umax_value + off could overflow, treat that as unbounded too.
3442 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
3443 verbose(env, "R%d unbounded memory access, make sure to bounds check any such access\n",
3447 err = __check_mem_access(env, regno, reg->umax_value + off, size,
3448 mem_size, zero_size_allowed);
3450 verbose(env, "R%d max value is outside of the allowed memory range\n",
3458 /* check read/write into a map element with possible variable offset */
3459 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
3460 int off, int size, bool zero_size_allowed)
3462 struct bpf_verifier_state *vstate = env->cur_state;
3463 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3464 struct bpf_reg_state *reg = &state->regs[regno];
3465 struct bpf_map *map = reg->map_ptr;
3468 err = check_mem_region_access(env, regno, off, size, map->value_size,
3473 if (map_value_has_spin_lock(map)) {
3474 u32 lock = map->spin_lock_off;
3476 /* if any part of struct bpf_spin_lock can be touched by
3477 * load/store reject this program.
3478 * To check that [x1, x2) overlaps with [y1, y2)
3479 * it is sufficient to check x1 < y2 && y1 < x2.
3481 if (reg->smin_value + off < lock + sizeof(struct bpf_spin_lock) &&
3482 lock < reg->umax_value + off + size) {
3483 verbose(env, "bpf_spin_lock cannot be accessed directly by load/store\n");
3487 if (map_value_has_timer(map)) {
3488 u32 t = map->timer_off;
3490 if (reg->smin_value + off < t + sizeof(struct bpf_timer) &&
3491 t < reg->umax_value + off + size) {
3492 verbose(env, "bpf_timer cannot be accessed directly by load/store\n");
3499 #define MAX_PACKET_OFF 0xffff
3501 static enum bpf_prog_type resolve_prog_type(struct bpf_prog *prog)
3503 return prog->aux->dst_prog ? prog->aux->dst_prog->type : prog->type;
3506 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
3507 const struct bpf_call_arg_meta *meta,
3508 enum bpf_access_type t)
3510 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
3512 switch (prog_type) {
3513 /* Program types only with direct read access go here! */
3514 case BPF_PROG_TYPE_LWT_IN:
3515 case BPF_PROG_TYPE_LWT_OUT:
3516 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
3517 case BPF_PROG_TYPE_SK_REUSEPORT:
3518 case BPF_PROG_TYPE_FLOW_DISSECTOR:
3519 case BPF_PROG_TYPE_CGROUP_SKB:
3524 /* Program types with direct read + write access go here! */
3525 case BPF_PROG_TYPE_SCHED_CLS:
3526 case BPF_PROG_TYPE_SCHED_ACT:
3527 case BPF_PROG_TYPE_XDP:
3528 case BPF_PROG_TYPE_LWT_XMIT:
3529 case BPF_PROG_TYPE_SK_SKB:
3530 case BPF_PROG_TYPE_SK_MSG:
3532 return meta->pkt_access;
3534 env->seen_direct_write = true;
3537 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
3539 env->seen_direct_write = true;
3548 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
3549 int size, bool zero_size_allowed)
3551 struct bpf_reg_state *regs = cur_regs(env);
3552 struct bpf_reg_state *reg = ®s[regno];
3555 /* We may have added a variable offset to the packet pointer; but any
3556 * reg->range we have comes after that. We are only checking the fixed
3560 /* We don't allow negative numbers, because we aren't tracking enough
3561 * detail to prove they're safe.
3563 if (reg->smin_value < 0) {
3564 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3569 err = reg->range < 0 ? -EINVAL :
3570 __check_mem_access(env, regno, off, size, reg->range,
3573 verbose(env, "R%d offset is outside of the packet\n", regno);
3577 /* __check_mem_access has made sure "off + size - 1" is within u16.
3578 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
3579 * otherwise find_good_pkt_pointers would have refused to set range info
3580 * that __check_mem_access would have rejected this pkt access.
3581 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
3583 env->prog->aux->max_pkt_offset =
3584 max_t(u32, env->prog->aux->max_pkt_offset,
3585 off + reg->umax_value + size - 1);
3590 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
3591 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
3592 enum bpf_access_type t, enum bpf_reg_type *reg_type,
3593 struct btf **btf, u32 *btf_id)
3595 struct bpf_insn_access_aux info = {
3596 .reg_type = *reg_type,
3600 if (env->ops->is_valid_access &&
3601 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
3602 /* A non zero info.ctx_field_size indicates that this field is a
3603 * candidate for later verifier transformation to load the whole
3604 * field and then apply a mask when accessed with a narrower
3605 * access than actual ctx access size. A zero info.ctx_field_size
3606 * will only allow for whole field access and rejects any other
3607 * type of narrower access.
3609 *reg_type = info.reg_type;
3611 if (base_type(*reg_type) == PTR_TO_BTF_ID) {
3613 *btf_id = info.btf_id;
3615 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
3617 /* remember the offset of last byte accessed in ctx */
3618 if (env->prog->aux->max_ctx_offset < off + size)
3619 env->prog->aux->max_ctx_offset = off + size;
3623 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
3627 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
3630 if (size < 0 || off < 0 ||
3631 (u64)off + size > sizeof(struct bpf_flow_keys)) {
3632 verbose(env, "invalid access to flow keys off=%d size=%d\n",
3639 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
3640 u32 regno, int off, int size,
3641 enum bpf_access_type t)
3643 struct bpf_reg_state *regs = cur_regs(env);
3644 struct bpf_reg_state *reg = ®s[regno];
3645 struct bpf_insn_access_aux info = {};
3648 if (reg->smin_value < 0) {
3649 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3654 switch (reg->type) {
3655 case PTR_TO_SOCK_COMMON:
3656 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
3659 valid = bpf_sock_is_valid_access(off, size, t, &info);
3661 case PTR_TO_TCP_SOCK:
3662 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
3664 case PTR_TO_XDP_SOCK:
3665 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
3673 env->insn_aux_data[insn_idx].ctx_field_size =
3674 info.ctx_field_size;
3678 verbose(env, "R%d invalid %s access off=%d size=%d\n",
3679 regno, reg_type_str(env, reg->type), off, size);
3684 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
3686 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
3689 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
3691 const struct bpf_reg_state *reg = reg_state(env, regno);
3693 return reg->type == PTR_TO_CTX;
3696 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
3698 const struct bpf_reg_state *reg = reg_state(env, regno);
3700 return type_is_sk_pointer(reg->type);
3703 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
3705 const struct bpf_reg_state *reg = reg_state(env, regno);
3707 return type_is_pkt_pointer(reg->type);
3710 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
3712 const struct bpf_reg_state *reg = reg_state(env, regno);
3714 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
3715 return reg->type == PTR_TO_FLOW_KEYS;
3718 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
3719 const struct bpf_reg_state *reg,
3720 int off, int size, bool strict)
3722 struct tnum reg_off;
3725 /* Byte size accesses are always allowed. */
3726 if (!strict || size == 1)
3729 /* For platforms that do not have a Kconfig enabling
3730 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
3731 * NET_IP_ALIGN is universally set to '2'. And on platforms
3732 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
3733 * to this code only in strict mode where we want to emulate
3734 * the NET_IP_ALIGN==2 checking. Therefore use an
3735 * unconditional IP align value of '2'.
3739 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
3740 if (!tnum_is_aligned(reg_off, size)) {
3743 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3745 "misaligned packet access off %d+%s+%d+%d size %d\n",
3746 ip_align, tn_buf, reg->off, off, size);
3753 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
3754 const struct bpf_reg_state *reg,
3755 const char *pointer_desc,
3756 int off, int size, bool strict)
3758 struct tnum reg_off;
3760 /* Byte size accesses are always allowed. */
3761 if (!strict || size == 1)
3764 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
3765 if (!tnum_is_aligned(reg_off, size)) {
3768 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3769 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
3770 pointer_desc, tn_buf, reg->off, off, size);
3777 static int check_ptr_alignment(struct bpf_verifier_env *env,
3778 const struct bpf_reg_state *reg, int off,
3779 int size, bool strict_alignment_once)
3781 bool strict = env->strict_alignment || strict_alignment_once;
3782 const char *pointer_desc = "";
3784 switch (reg->type) {
3786 case PTR_TO_PACKET_META:
3787 /* Special case, because of NET_IP_ALIGN. Given metadata sits
3788 * right in front, treat it the very same way.
3790 return check_pkt_ptr_alignment(env, reg, off, size, strict);
3791 case PTR_TO_FLOW_KEYS:
3792 pointer_desc = "flow keys ";
3794 case PTR_TO_MAP_KEY:
3795 pointer_desc = "key ";
3797 case PTR_TO_MAP_VALUE:
3798 pointer_desc = "value ";
3801 pointer_desc = "context ";
3804 pointer_desc = "stack ";
3805 /* The stack spill tracking logic in check_stack_write_fixed_off()
3806 * and check_stack_read_fixed_off() relies on stack accesses being
3812 pointer_desc = "sock ";
3814 case PTR_TO_SOCK_COMMON:
3815 pointer_desc = "sock_common ";
3817 case PTR_TO_TCP_SOCK:
3818 pointer_desc = "tcp_sock ";
3820 case PTR_TO_XDP_SOCK:
3821 pointer_desc = "xdp_sock ";
3826 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
3830 static int update_stack_depth(struct bpf_verifier_env *env,
3831 const struct bpf_func_state *func,
3834 u16 stack = env->subprog_info[func->subprogno].stack_depth;
3839 /* update known max for given subprogram */
3840 env->subprog_info[func->subprogno].stack_depth = -off;
3844 /* starting from main bpf function walk all instructions of the function
3845 * and recursively walk all callees that given function can call.
3846 * Ignore jump and exit insns.
3847 * Since recursion is prevented by check_cfg() this algorithm
3848 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
3850 static int check_max_stack_depth(struct bpf_verifier_env *env)
3852 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
3853 struct bpf_subprog_info *subprog = env->subprog_info;
3854 struct bpf_insn *insn = env->prog->insnsi;
3855 bool tail_call_reachable = false;
3856 int ret_insn[MAX_CALL_FRAMES];
3857 int ret_prog[MAX_CALL_FRAMES];
3861 /* protect against potential stack overflow that might happen when
3862 * bpf2bpf calls get combined with tailcalls. Limit the caller's stack
3863 * depth for such case down to 256 so that the worst case scenario
3864 * would result in 8k stack size (32 which is tailcall limit * 256 =
3867 * To get the idea what might happen, see an example:
3868 * func1 -> sub rsp, 128
3869 * subfunc1 -> sub rsp, 256
3870 * tailcall1 -> add rsp, 256
3871 * func2 -> sub rsp, 192 (total stack size = 128 + 192 = 320)
3872 * subfunc2 -> sub rsp, 64
3873 * subfunc22 -> sub rsp, 128
3874 * tailcall2 -> add rsp, 128
3875 * func3 -> sub rsp, 32 (total stack size 128 + 192 + 64 + 32 = 416)
3877 * tailcall will unwind the current stack frame but it will not get rid
3878 * of caller's stack as shown on the example above.
3880 if (idx && subprog[idx].has_tail_call && depth >= 256) {
3882 "tail_calls are not allowed when call stack of previous frames is %d bytes. Too large\n",
3886 /* round up to 32-bytes, since this is granularity
3887 * of interpreter stack size
3889 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
3890 if (depth > MAX_BPF_STACK) {
3891 verbose(env, "combined stack size of %d calls is %d. Too large\n",
3896 subprog_end = subprog[idx + 1].start;
3897 for (; i < subprog_end; i++) {
3900 if (!bpf_pseudo_call(insn + i) && !bpf_pseudo_func(insn + i))
3902 /* remember insn and function to return to */
3903 ret_insn[frame] = i + 1;
3904 ret_prog[frame] = idx;
3906 /* find the callee */
3907 next_insn = i + insn[i].imm + 1;
3908 idx = find_subprog(env, next_insn);
3910 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3914 if (subprog[idx].is_async_cb) {
3915 if (subprog[idx].has_tail_call) {
3916 verbose(env, "verifier bug. subprog has tail_call and async cb\n");
3919 /* async callbacks don't increase bpf prog stack size */
3924 if (subprog[idx].has_tail_call)
3925 tail_call_reachable = true;
3928 if (frame >= MAX_CALL_FRAMES) {
3929 verbose(env, "the call stack of %d frames is too deep !\n",
3935 /* if tail call got detected across bpf2bpf calls then mark each of the
3936 * currently present subprog frames as tail call reachable subprogs;
3937 * this info will be utilized by JIT so that we will be preserving the
3938 * tail call counter throughout bpf2bpf calls combined with tailcalls
3940 if (tail_call_reachable)
3941 for (j = 0; j < frame; j++)
3942 subprog[ret_prog[j]].tail_call_reachable = true;
3943 if (subprog[0].tail_call_reachable)
3944 env->prog->aux->tail_call_reachable = true;
3946 /* end of for() loop means the last insn of the 'subprog'
3947 * was reached. Doesn't matter whether it was JA or EXIT
3951 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
3953 i = ret_insn[frame];
3954 idx = ret_prog[frame];
3958 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
3959 static int get_callee_stack_depth(struct bpf_verifier_env *env,
3960 const struct bpf_insn *insn, int idx)
3962 int start = idx + insn->imm + 1, subprog;
3964 subprog = find_subprog(env, start);
3966 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3970 return env->subprog_info[subprog].stack_depth;
3974 static int __check_ptr_off_reg(struct bpf_verifier_env *env,
3975 const struct bpf_reg_state *reg, int regno,
3978 /* Access to this pointer-typed register or passing it to a helper
3979 * is only allowed in its original, unmodified form.
3982 if (!fixed_off_ok && reg->off) {
3983 verbose(env, "dereference of modified %s ptr R%d off=%d disallowed\n",
3984 reg_type_str(env, reg->type), regno, reg->off);
3988 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3991 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3992 verbose(env, "variable %s access var_off=%s disallowed\n",
3993 reg_type_str(env, reg->type), tn_buf);
4000 int check_ptr_off_reg(struct bpf_verifier_env *env,
4001 const struct bpf_reg_state *reg, int regno)
4003 return __check_ptr_off_reg(env, reg, regno, false);
4006 static int __check_buffer_access(struct bpf_verifier_env *env,
4007 const char *buf_info,
4008 const struct bpf_reg_state *reg,
4009 int regno, int off, int size)
4013 "R%d invalid %s buffer access: off=%d, size=%d\n",
4014 regno, buf_info, off, size);
4017 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
4020 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4022 "R%d invalid variable buffer offset: off=%d, var_off=%s\n",
4023 regno, off, tn_buf);
4030 static int check_tp_buffer_access(struct bpf_verifier_env *env,
4031 const struct bpf_reg_state *reg,
4032 int regno, int off, int size)
4036 err = __check_buffer_access(env, "tracepoint", reg, regno, off, size);
4040 if (off + size > env->prog->aux->max_tp_access)
4041 env->prog->aux->max_tp_access = off + size;
4046 static int check_buffer_access(struct bpf_verifier_env *env,
4047 const struct bpf_reg_state *reg,
4048 int regno, int off, int size,
4049 bool zero_size_allowed,
4050 const char *buf_info,
4055 err = __check_buffer_access(env, buf_info, reg, regno, off, size);
4059 if (off + size > *max_access)
4060 *max_access = off + size;
4065 /* BPF architecture zero extends alu32 ops into 64-bit registesr */
4066 static void zext_32_to_64(struct bpf_reg_state *reg)
4068 reg->var_off = tnum_subreg(reg->var_off);
4069 __reg_assign_32_into_64(reg);
4072 /* truncate register to smaller size (in bytes)
4073 * must be called with size < BPF_REG_SIZE
4075 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
4079 /* clear high bits in bit representation */
4080 reg->var_off = tnum_cast(reg->var_off, size);
4082 /* fix arithmetic bounds */
4083 mask = ((u64)1 << (size * 8)) - 1;
4084 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
4085 reg->umin_value &= mask;
4086 reg->umax_value &= mask;
4088 reg->umin_value = 0;
4089 reg->umax_value = mask;
4091 reg->smin_value = reg->umin_value;
4092 reg->smax_value = reg->umax_value;
4094 /* If size is smaller than 32bit register the 32bit register
4095 * values are also truncated so we push 64-bit bounds into
4096 * 32-bit bounds. Above were truncated < 32-bits already.
4100 __reg_combine_64_into_32(reg);
4103 static bool bpf_map_is_rdonly(const struct bpf_map *map)
4105 /* A map is considered read-only if the following condition are true:
4107 * 1) BPF program side cannot change any of the map content. The
4108 * BPF_F_RDONLY_PROG flag is throughout the lifetime of a map
4109 * and was set at map creation time.
4110 * 2) The map value(s) have been initialized from user space by a
4111 * loader and then "frozen", such that no new map update/delete
4112 * operations from syscall side are possible for the rest of
4113 * the map's lifetime from that point onwards.
4114 * 3) Any parallel/pending map update/delete operations from syscall
4115 * side have been completed. Only after that point, it's safe to
4116 * assume that map value(s) are immutable.
4118 return (map->map_flags & BPF_F_RDONLY_PROG) &&
4119 READ_ONCE(map->frozen) &&
4120 !bpf_map_write_active(map);
4123 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val)
4129 err = map->ops->map_direct_value_addr(map, &addr, off);
4132 ptr = (void *)(long)addr + off;
4136 *val = (u64)*(u8 *)ptr;
4139 *val = (u64)*(u16 *)ptr;
4142 *val = (u64)*(u32 *)ptr;
4153 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
4154 struct bpf_reg_state *regs,
4155 int regno, int off, int size,
4156 enum bpf_access_type atype,
4159 struct bpf_reg_state *reg = regs + regno;
4160 const struct btf_type *t = btf_type_by_id(reg->btf, reg->btf_id);
4161 const char *tname = btf_name_by_offset(reg->btf, t->name_off);
4167 "R%d is ptr_%s invalid negative access: off=%d\n",
4171 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
4174 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4176 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
4177 regno, tname, off, tn_buf);
4181 if (env->ops->btf_struct_access) {
4182 ret = env->ops->btf_struct_access(&env->log, reg->btf, t,
4183 off, size, atype, &btf_id);
4185 if (atype != BPF_READ) {
4186 verbose(env, "only read is supported\n");
4190 ret = btf_struct_access(&env->log, reg->btf, t, off, size,
4197 if (atype == BPF_READ && value_regno >= 0)
4198 mark_btf_ld_reg(env, regs, value_regno, ret, reg->btf, btf_id);
4203 static int check_ptr_to_map_access(struct bpf_verifier_env *env,
4204 struct bpf_reg_state *regs,
4205 int regno, int off, int size,
4206 enum bpf_access_type atype,
4209 struct bpf_reg_state *reg = regs + regno;
4210 struct bpf_map *map = reg->map_ptr;
4211 const struct btf_type *t;
4217 verbose(env, "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n");
4221 if (!map->ops->map_btf_id || !*map->ops->map_btf_id) {
4222 verbose(env, "map_ptr access not supported for map type %d\n",
4227 t = btf_type_by_id(btf_vmlinux, *map->ops->map_btf_id);
4228 tname = btf_name_by_offset(btf_vmlinux, t->name_off);
4230 if (!env->allow_ptr_to_map_access) {
4232 "%s access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
4238 verbose(env, "R%d is %s invalid negative access: off=%d\n",
4243 if (atype != BPF_READ) {
4244 verbose(env, "only read from %s is supported\n", tname);
4248 ret = btf_struct_access(&env->log, btf_vmlinux, t, off, size, atype, &btf_id);
4252 if (value_regno >= 0)
4253 mark_btf_ld_reg(env, regs, value_regno, ret, btf_vmlinux, btf_id);
4258 /* Check that the stack access at the given offset is within bounds. The
4259 * maximum valid offset is -1.
4261 * The minimum valid offset is -MAX_BPF_STACK for writes, and
4262 * -state->allocated_stack for reads.
4264 static int check_stack_slot_within_bounds(int off,
4265 struct bpf_func_state *state,
4266 enum bpf_access_type t)
4271 min_valid_off = -MAX_BPF_STACK;
4273 min_valid_off = -state->allocated_stack;
4275 if (off < min_valid_off || off > -1)
4280 /* Check that the stack access at 'regno + off' falls within the maximum stack
4283 * 'off' includes `regno->offset`, but not its dynamic part (if any).
4285 static int check_stack_access_within_bounds(
4286 struct bpf_verifier_env *env,
4287 int regno, int off, int access_size,
4288 enum stack_access_src src, enum bpf_access_type type)
4290 struct bpf_reg_state *regs = cur_regs(env);
4291 struct bpf_reg_state *reg = regs + regno;
4292 struct bpf_func_state *state = func(env, reg);
4293 int min_off, max_off;
4297 if (src == ACCESS_HELPER)
4298 /* We don't know if helpers are reading or writing (or both). */
4299 err_extra = " indirect access to";
4300 else if (type == BPF_READ)
4301 err_extra = " read from";
4303 err_extra = " write to";
4305 if (tnum_is_const(reg->var_off)) {
4306 min_off = reg->var_off.value + off;
4307 if (access_size > 0)
4308 max_off = min_off + access_size - 1;
4312 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
4313 reg->smin_value <= -BPF_MAX_VAR_OFF) {
4314 verbose(env, "invalid unbounded variable-offset%s stack R%d\n",
4318 min_off = reg->smin_value + off;
4319 if (access_size > 0)
4320 max_off = reg->smax_value + off + access_size - 1;
4325 err = check_stack_slot_within_bounds(min_off, state, type);
4327 err = check_stack_slot_within_bounds(max_off, state, type);
4330 if (tnum_is_const(reg->var_off)) {
4331 verbose(env, "invalid%s stack R%d off=%d size=%d\n",
4332 err_extra, regno, off, access_size);
4336 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4337 verbose(env, "invalid variable-offset%s stack R%d var_off=%s size=%d\n",
4338 err_extra, regno, tn_buf, access_size);
4344 /* check whether memory at (regno + off) is accessible for t = (read | write)
4345 * if t==write, value_regno is a register which value is stored into memory
4346 * if t==read, value_regno is a register which will receive the value from memory
4347 * if t==write && value_regno==-1, some unknown value is stored into memory
4348 * if t==read && value_regno==-1, don't care what we read from memory
4350 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
4351 int off, int bpf_size, enum bpf_access_type t,
4352 int value_regno, bool strict_alignment_once)
4354 struct bpf_reg_state *regs = cur_regs(env);
4355 struct bpf_reg_state *reg = regs + regno;
4356 struct bpf_func_state *state;
4359 size = bpf_size_to_bytes(bpf_size);
4363 /* alignment checks will add in reg->off themselves */
4364 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
4368 /* for access checks, reg->off is just part of off */
4371 if (reg->type == PTR_TO_MAP_KEY) {
4372 if (t == BPF_WRITE) {
4373 verbose(env, "write to change key R%d not allowed\n", regno);
4377 err = check_mem_region_access(env, regno, off, size,
4378 reg->map_ptr->key_size, false);
4381 if (value_regno >= 0)
4382 mark_reg_unknown(env, regs, value_regno);
4383 } else if (reg->type == PTR_TO_MAP_VALUE) {
4384 if (t == BPF_WRITE && value_regno >= 0 &&
4385 is_pointer_value(env, value_regno)) {
4386 verbose(env, "R%d leaks addr into map\n", value_regno);
4389 err = check_map_access_type(env, regno, off, size, t);
4392 err = check_map_access(env, regno, off, size, false);
4393 if (!err && t == BPF_READ && value_regno >= 0) {
4394 struct bpf_map *map = reg->map_ptr;
4396 /* if map is read-only, track its contents as scalars */
4397 if (tnum_is_const(reg->var_off) &&
4398 bpf_map_is_rdonly(map) &&
4399 map->ops->map_direct_value_addr) {
4400 int map_off = off + reg->var_off.value;
4403 err = bpf_map_direct_read(map, map_off, size,
4408 regs[value_regno].type = SCALAR_VALUE;
4409 __mark_reg_known(®s[value_regno], val);
4411 mark_reg_unknown(env, regs, value_regno);
4414 } else if (base_type(reg->type) == PTR_TO_MEM) {
4415 bool rdonly_mem = type_is_rdonly_mem(reg->type);
4417 if (type_may_be_null(reg->type)) {
4418 verbose(env, "R%d invalid mem access '%s'\n", regno,
4419 reg_type_str(env, reg->type));
4423 if (t == BPF_WRITE && rdonly_mem) {
4424 verbose(env, "R%d cannot write into %s\n",
4425 regno, reg_type_str(env, reg->type));
4429 if (t == BPF_WRITE && value_regno >= 0 &&
4430 is_pointer_value(env, value_regno)) {
4431 verbose(env, "R%d leaks addr into mem\n", value_regno);
4435 err = check_mem_region_access(env, regno, off, size,
4436 reg->mem_size, false);
4437 if (!err && value_regno >= 0 && (t == BPF_READ || rdonly_mem))
4438 mark_reg_unknown(env, regs, value_regno);
4439 } else if (reg->type == PTR_TO_CTX) {
4440 enum bpf_reg_type reg_type = SCALAR_VALUE;
4441 struct btf *btf = NULL;
4444 if (t == BPF_WRITE && value_regno >= 0 &&
4445 is_pointer_value(env, value_regno)) {
4446 verbose(env, "R%d leaks addr into ctx\n", value_regno);
4450 err = check_ptr_off_reg(env, reg, regno);
4454 err = check_ctx_access(env, insn_idx, off, size, t, ®_type, &btf, &btf_id);
4456 verbose_linfo(env, insn_idx, "; ");
4457 if (!err && t == BPF_READ && value_regno >= 0) {
4458 /* ctx access returns either a scalar, or a
4459 * PTR_TO_PACKET[_META,_END]. In the latter
4460 * case, we know the offset is zero.
4462 if (reg_type == SCALAR_VALUE) {
4463 mark_reg_unknown(env, regs, value_regno);
4465 mark_reg_known_zero(env, regs,
4467 if (type_may_be_null(reg_type))
4468 regs[value_regno].id = ++env->id_gen;
4469 /* A load of ctx field could have different
4470 * actual load size with the one encoded in the
4471 * insn. When the dst is PTR, it is for sure not
4474 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
4475 if (base_type(reg_type) == PTR_TO_BTF_ID) {
4476 regs[value_regno].btf = btf;
4477 regs[value_regno].btf_id = btf_id;
4480 regs[value_regno].type = reg_type;
4483 } else if (reg->type == PTR_TO_STACK) {
4484 /* Basic bounds checks. */
4485 err = check_stack_access_within_bounds(env, regno, off, size, ACCESS_DIRECT, t);
4489 state = func(env, reg);
4490 err = update_stack_depth(env, state, off);
4495 err = check_stack_read(env, regno, off, size,
4498 err = check_stack_write(env, regno, off, size,
4499 value_regno, insn_idx);
4500 } else if (reg_is_pkt_pointer(reg)) {
4501 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
4502 verbose(env, "cannot write into packet\n");
4505 if (t == BPF_WRITE && value_regno >= 0 &&
4506 is_pointer_value(env, value_regno)) {
4507 verbose(env, "R%d leaks addr into packet\n",
4511 err = check_packet_access(env, regno, off, size, false);
4512 if (!err && t == BPF_READ && value_regno >= 0)
4513 mark_reg_unknown(env, regs, value_regno);
4514 } else if (reg->type == PTR_TO_FLOW_KEYS) {
4515 if (t == BPF_WRITE && value_regno >= 0 &&
4516 is_pointer_value(env, value_regno)) {
4517 verbose(env, "R%d leaks addr into flow keys\n",
4522 err = check_flow_keys_access(env, off, size);
4523 if (!err && t == BPF_READ && value_regno >= 0)
4524 mark_reg_unknown(env, regs, value_regno);
4525 } else if (type_is_sk_pointer(reg->type)) {
4526 if (t == BPF_WRITE) {
4527 verbose(env, "R%d cannot write into %s\n",
4528 regno, reg_type_str(env, reg->type));
4531 err = check_sock_access(env, insn_idx, regno, off, size, t);
4532 if (!err && value_regno >= 0)
4533 mark_reg_unknown(env, regs, value_regno);
4534 } else if (reg->type == PTR_TO_TP_BUFFER) {
4535 err = check_tp_buffer_access(env, reg, regno, off, size);
4536 if (!err && t == BPF_READ && value_regno >= 0)
4537 mark_reg_unknown(env, regs, value_regno);
4538 } else if (reg->type == PTR_TO_BTF_ID) {
4539 err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
4541 } else if (reg->type == CONST_PTR_TO_MAP) {
4542 err = check_ptr_to_map_access(env, regs, regno, off, size, t,
4544 } else if (base_type(reg->type) == PTR_TO_BUF) {
4545 bool rdonly_mem = type_is_rdonly_mem(reg->type);
4546 const char *buf_info;
4550 if (t == BPF_WRITE) {
4551 verbose(env, "R%d cannot write into %s\n",
4552 regno, reg_type_str(env, reg->type));
4555 buf_info = "rdonly";
4556 max_access = &env->prog->aux->max_rdonly_access;
4559 max_access = &env->prog->aux->max_rdwr_access;
4562 err = check_buffer_access(env, reg, regno, off, size, false,
4563 buf_info, max_access);
4565 if (!err && value_regno >= 0 && (rdonly_mem || t == BPF_READ))
4566 mark_reg_unknown(env, regs, value_regno);
4568 verbose(env, "R%d invalid mem access '%s'\n", regno,
4569 reg_type_str(env, reg->type));
4573 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
4574 regs[value_regno].type == SCALAR_VALUE) {
4575 /* b/h/w load zero-extends, mark upper bits as known 0 */
4576 coerce_reg_to_size(®s[value_regno], size);
4581 static int check_atomic(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
4586 switch (insn->imm) {
4588 case BPF_ADD | BPF_FETCH:
4590 case BPF_AND | BPF_FETCH:
4592 case BPF_OR | BPF_FETCH:
4594 case BPF_XOR | BPF_FETCH:
4599 verbose(env, "BPF_ATOMIC uses invalid atomic opcode %02x\n", insn->imm);
4603 if (BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) {
4604 verbose(env, "invalid atomic operand size\n");
4608 /* check src1 operand */
4609 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4613 /* check src2 operand */
4614 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4618 if (insn->imm == BPF_CMPXCHG) {
4619 /* Check comparison of R0 with memory location */
4620 const u32 aux_reg = BPF_REG_0;
4622 err = check_reg_arg(env, aux_reg, SRC_OP);
4626 if (is_pointer_value(env, aux_reg)) {
4627 verbose(env, "R%d leaks addr into mem\n", aux_reg);
4632 if (is_pointer_value(env, insn->src_reg)) {
4633 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
4637 if (is_ctx_reg(env, insn->dst_reg) ||
4638 is_pkt_reg(env, insn->dst_reg) ||
4639 is_flow_key_reg(env, insn->dst_reg) ||
4640 is_sk_reg(env, insn->dst_reg)) {
4641 verbose(env, "BPF_ATOMIC stores into R%d %s is not allowed\n",
4643 reg_type_str(env, reg_state(env, insn->dst_reg)->type));
4647 if (insn->imm & BPF_FETCH) {
4648 if (insn->imm == BPF_CMPXCHG)
4649 load_reg = BPF_REG_0;
4651 load_reg = insn->src_reg;
4653 /* check and record load of old value */
4654 err = check_reg_arg(env, load_reg, DST_OP);
4658 /* This instruction accesses a memory location but doesn't
4659 * actually load it into a register.
4664 /* Check whether we can read the memory, with second call for fetch
4665 * case to simulate the register fill.
4667 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4668 BPF_SIZE(insn->code), BPF_READ, -1, true);
4669 if (!err && load_reg >= 0)
4670 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4671 BPF_SIZE(insn->code), BPF_READ, load_reg,
4676 /* Check whether we can write into the same memory. */
4677 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4678 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
4685 /* When register 'regno' is used to read the stack (either directly or through
4686 * a helper function) make sure that it's within stack boundary and, depending
4687 * on the access type, that all elements of the stack are initialized.
4689 * 'off' includes 'regno->off', but not its dynamic part (if any).
4691 * All registers that have been spilled on the stack in the slots within the
4692 * read offsets are marked as read.
4694 static int check_stack_range_initialized(
4695 struct bpf_verifier_env *env, int regno, int off,
4696 int access_size, bool zero_size_allowed,
4697 enum stack_access_src type, struct bpf_call_arg_meta *meta)
4699 struct bpf_reg_state *reg = reg_state(env, regno);
4700 struct bpf_func_state *state = func(env, reg);
4701 int err, min_off, max_off, i, j, slot, spi;
4702 char *err_extra = type == ACCESS_HELPER ? " indirect" : "";
4703 enum bpf_access_type bounds_check_type;
4704 /* Some accesses can write anything into the stack, others are
4707 bool clobber = false;
4709 if (access_size == 0 && !zero_size_allowed) {
4710 verbose(env, "invalid zero-sized read\n");
4714 if (type == ACCESS_HELPER) {
4715 /* The bounds checks for writes are more permissive than for
4716 * reads. However, if raw_mode is not set, we'll do extra
4719 bounds_check_type = BPF_WRITE;
4722 bounds_check_type = BPF_READ;
4724 err = check_stack_access_within_bounds(env, regno, off, access_size,
4725 type, bounds_check_type);
4730 if (tnum_is_const(reg->var_off)) {
4731 min_off = max_off = reg->var_off.value + off;
4733 /* Variable offset is prohibited for unprivileged mode for
4734 * simplicity since it requires corresponding support in
4735 * Spectre masking for stack ALU.
4736 * See also retrieve_ptr_limit().
4738 if (!env->bypass_spec_v1) {
4741 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4742 verbose(env, "R%d%s variable offset stack access prohibited for !root, var_off=%s\n",
4743 regno, err_extra, tn_buf);
4746 /* Only initialized buffer on stack is allowed to be accessed
4747 * with variable offset. With uninitialized buffer it's hard to
4748 * guarantee that whole memory is marked as initialized on
4749 * helper return since specific bounds are unknown what may
4750 * cause uninitialized stack leaking.
4752 if (meta && meta->raw_mode)
4755 min_off = reg->smin_value + off;
4756 max_off = reg->smax_value + off;
4759 if (meta && meta->raw_mode) {
4760 meta->access_size = access_size;
4761 meta->regno = regno;
4765 for (i = min_off; i < max_off + access_size; i++) {
4769 spi = slot / BPF_REG_SIZE;
4770 if (state->allocated_stack <= slot)
4772 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
4773 if (*stype == STACK_MISC)
4775 if (*stype == STACK_ZERO) {
4777 /* helper can write anything into the stack */
4778 *stype = STACK_MISC;
4783 if (is_spilled_reg(&state->stack[spi]) &&
4784 state->stack[spi].spilled_ptr.type == PTR_TO_BTF_ID)
4787 if (is_spilled_reg(&state->stack[spi]) &&
4788 (state->stack[spi].spilled_ptr.type == SCALAR_VALUE ||
4789 env->allow_ptr_leaks)) {
4791 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
4792 for (j = 0; j < BPF_REG_SIZE; j++)
4793 scrub_spilled_slot(&state->stack[spi].slot_type[j]);
4799 if (tnum_is_const(reg->var_off)) {
4800 verbose(env, "invalid%s read from stack R%d off %d+%d size %d\n",
4801 err_extra, regno, min_off, i - min_off, access_size);
4805 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4806 verbose(env, "invalid%s read from stack R%d var_off %s+%d size %d\n",
4807 err_extra, regno, tn_buf, i - min_off, access_size);
4811 /* reading any byte out of 8-byte 'spill_slot' will cause
4812 * the whole slot to be marked as 'read'
4814 mark_reg_read(env, &state->stack[spi].spilled_ptr,
4815 state->stack[spi].spilled_ptr.parent,
4818 return update_stack_depth(env, state, min_off);
4821 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
4822 int access_size, bool zero_size_allowed,
4823 struct bpf_call_arg_meta *meta)
4825 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4826 const char *buf_info;
4829 switch (base_type(reg->type)) {
4831 case PTR_TO_PACKET_META:
4832 return check_packet_access(env, regno, reg->off, access_size,
4834 case PTR_TO_MAP_KEY:
4835 return check_mem_region_access(env, regno, reg->off, access_size,
4836 reg->map_ptr->key_size, false);
4837 case PTR_TO_MAP_VALUE:
4838 if (check_map_access_type(env, regno, reg->off, access_size,
4839 meta && meta->raw_mode ? BPF_WRITE :
4842 return check_map_access(env, regno, reg->off, access_size,
4845 return check_mem_region_access(env, regno, reg->off,
4846 access_size, reg->mem_size,
4849 if (type_is_rdonly_mem(reg->type)) {
4850 if (meta && meta->raw_mode)
4853 buf_info = "rdonly";
4854 max_access = &env->prog->aux->max_rdonly_access;
4857 max_access = &env->prog->aux->max_rdwr_access;
4859 return check_buffer_access(env, reg, regno, reg->off,
4860 access_size, zero_size_allowed,
4861 buf_info, max_access);
4863 return check_stack_range_initialized(
4865 regno, reg->off, access_size,
4866 zero_size_allowed, ACCESS_HELPER, meta);
4867 default: /* scalar_value or invalid ptr */
4868 /* Allow zero-byte read from NULL, regardless of pointer type */
4869 if (zero_size_allowed && access_size == 0 &&
4870 register_is_null(reg))
4873 verbose(env, "R%d type=%s ", regno,
4874 reg_type_str(env, reg->type));
4875 verbose(env, "expected=%s\n", reg_type_str(env, PTR_TO_STACK));
4880 int check_mem_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
4881 u32 regno, u32 mem_size)
4883 if (register_is_null(reg))
4886 if (type_may_be_null(reg->type)) {
4887 /* Assuming that the register contains a value check if the memory
4888 * access is safe. Temporarily save and restore the register's state as
4889 * the conversion shouldn't be visible to a caller.
4891 const struct bpf_reg_state saved_reg = *reg;
4894 mark_ptr_not_null_reg(reg);
4895 rv = check_helper_mem_access(env, regno, mem_size, true, NULL);
4900 return check_helper_mem_access(env, regno, mem_size, true, NULL);
4903 /* Implementation details:
4904 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL
4905 * Two bpf_map_lookups (even with the same key) will have different reg->id.
4906 * For traditional PTR_TO_MAP_VALUE the verifier clears reg->id after
4907 * value_or_null->value transition, since the verifier only cares about
4908 * the range of access to valid map value pointer and doesn't care about actual
4909 * address of the map element.
4910 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
4911 * reg->id > 0 after value_or_null->value transition. By doing so
4912 * two bpf_map_lookups will be considered two different pointers that
4913 * point to different bpf_spin_locks.
4914 * The verifier allows taking only one bpf_spin_lock at a time to avoid
4916 * Since only one bpf_spin_lock is allowed the checks are simpler than
4917 * reg_is_refcounted() logic. The verifier needs to remember only
4918 * one spin_lock instead of array of acquired_refs.
4919 * cur_state->active_spin_lock remembers which map value element got locked
4920 * and clears it after bpf_spin_unlock.
4922 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
4925 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4926 struct bpf_verifier_state *cur = env->cur_state;
4927 bool is_const = tnum_is_const(reg->var_off);
4928 struct bpf_map *map = reg->map_ptr;
4929 u64 val = reg->var_off.value;
4933 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
4939 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
4943 if (!map_value_has_spin_lock(map)) {
4944 if (map->spin_lock_off == -E2BIG)
4946 "map '%s' has more than one 'struct bpf_spin_lock'\n",
4948 else if (map->spin_lock_off == -ENOENT)
4950 "map '%s' doesn't have 'struct bpf_spin_lock'\n",
4954 "map '%s' is not a struct type or bpf_spin_lock is mangled\n",
4958 if (map->spin_lock_off != val + reg->off) {
4959 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock'\n",
4964 if (cur->active_spin_lock) {
4966 "Locking two bpf_spin_locks are not allowed\n");
4969 cur->active_spin_lock = reg->id;
4971 if (!cur->active_spin_lock) {
4972 verbose(env, "bpf_spin_unlock without taking a lock\n");
4975 if (cur->active_spin_lock != reg->id) {
4976 verbose(env, "bpf_spin_unlock of different lock\n");
4979 cur->active_spin_lock = 0;
4984 static int process_timer_func(struct bpf_verifier_env *env, int regno,
4985 struct bpf_call_arg_meta *meta)
4987 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4988 bool is_const = tnum_is_const(reg->var_off);
4989 struct bpf_map *map = reg->map_ptr;
4990 u64 val = reg->var_off.value;
4994 "R%d doesn't have constant offset. bpf_timer has to be at the constant offset\n",
4999 verbose(env, "map '%s' has to have BTF in order to use bpf_timer\n",
5003 if (!map_value_has_timer(map)) {
5004 if (map->timer_off == -E2BIG)
5006 "map '%s' has more than one 'struct bpf_timer'\n",
5008 else if (map->timer_off == -ENOENT)
5010 "map '%s' doesn't have 'struct bpf_timer'\n",
5014 "map '%s' is not a struct type or bpf_timer is mangled\n",
5018 if (map->timer_off != val + reg->off) {
5019 verbose(env, "off %lld doesn't point to 'struct bpf_timer' that is at %d\n",
5020 val + reg->off, map->timer_off);
5023 if (meta->map_ptr) {
5024 verbose(env, "verifier bug. Two map pointers in a timer helper\n");
5027 meta->map_uid = reg->map_uid;
5028 meta->map_ptr = map;
5032 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
5034 return base_type(type) == ARG_PTR_TO_MEM ||
5035 base_type(type) == ARG_PTR_TO_UNINIT_MEM;
5038 static bool arg_type_is_mem_size(enum bpf_arg_type type)
5040 return type == ARG_CONST_SIZE ||
5041 type == ARG_CONST_SIZE_OR_ZERO;
5044 static bool arg_type_is_alloc_size(enum bpf_arg_type type)
5046 return type == ARG_CONST_ALLOC_SIZE_OR_ZERO;
5049 static bool arg_type_is_int_ptr(enum bpf_arg_type type)
5051 return type == ARG_PTR_TO_INT ||
5052 type == ARG_PTR_TO_LONG;
5055 static int int_ptr_type_to_size(enum bpf_arg_type type)
5057 if (type == ARG_PTR_TO_INT)
5059 else if (type == ARG_PTR_TO_LONG)
5065 static int resolve_map_arg_type(struct bpf_verifier_env *env,
5066 const struct bpf_call_arg_meta *meta,
5067 enum bpf_arg_type *arg_type)
5069 if (!meta->map_ptr) {
5070 /* kernel subsystem misconfigured verifier */
5071 verbose(env, "invalid map_ptr to access map->type\n");
5075 switch (meta->map_ptr->map_type) {
5076 case BPF_MAP_TYPE_SOCKMAP:
5077 case BPF_MAP_TYPE_SOCKHASH:
5078 if (*arg_type == ARG_PTR_TO_MAP_VALUE) {
5079 *arg_type = ARG_PTR_TO_BTF_ID_SOCK_COMMON;
5081 verbose(env, "invalid arg_type for sockmap/sockhash\n");
5085 case BPF_MAP_TYPE_BLOOM_FILTER:
5086 if (meta->func_id == BPF_FUNC_map_peek_elem)
5087 *arg_type = ARG_PTR_TO_MAP_VALUE;
5095 struct bpf_reg_types {
5096 const enum bpf_reg_type types[10];
5100 static const struct bpf_reg_types map_key_value_types = {
5110 static const struct bpf_reg_types sock_types = {
5120 static const struct bpf_reg_types btf_id_sock_common_types = {
5128 .btf_id = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
5132 static const struct bpf_reg_types mem_types = {
5140 PTR_TO_MEM | MEM_ALLOC,
5145 static const struct bpf_reg_types int_ptr_types = {
5155 static const struct bpf_reg_types fullsock_types = { .types = { PTR_TO_SOCKET } };
5156 static const struct bpf_reg_types scalar_types = { .types = { SCALAR_VALUE } };
5157 static const struct bpf_reg_types context_types = { .types = { PTR_TO_CTX } };
5158 static const struct bpf_reg_types alloc_mem_types = { .types = { PTR_TO_MEM | MEM_ALLOC } };
5159 static const struct bpf_reg_types const_map_ptr_types = { .types = { CONST_PTR_TO_MAP } };
5160 static const struct bpf_reg_types btf_ptr_types = { .types = { PTR_TO_BTF_ID } };
5161 static const struct bpf_reg_types spin_lock_types = { .types = { PTR_TO_MAP_VALUE } };
5162 static const struct bpf_reg_types percpu_btf_ptr_types = { .types = { PTR_TO_PERCPU_BTF_ID } };
5163 static const struct bpf_reg_types func_ptr_types = { .types = { PTR_TO_FUNC } };
5164 static const struct bpf_reg_types stack_ptr_types = { .types = { PTR_TO_STACK } };
5165 static const struct bpf_reg_types const_str_ptr_types = { .types = { PTR_TO_MAP_VALUE } };
5166 static const struct bpf_reg_types timer_types = { .types = { PTR_TO_MAP_VALUE } };
5168 static const struct bpf_reg_types *compatible_reg_types[__BPF_ARG_TYPE_MAX] = {
5169 [ARG_PTR_TO_MAP_KEY] = &map_key_value_types,
5170 [ARG_PTR_TO_MAP_VALUE] = &map_key_value_types,
5171 [ARG_PTR_TO_UNINIT_MAP_VALUE] = &map_key_value_types,
5172 [ARG_CONST_SIZE] = &scalar_types,
5173 [ARG_CONST_SIZE_OR_ZERO] = &scalar_types,
5174 [ARG_CONST_ALLOC_SIZE_OR_ZERO] = &scalar_types,
5175 [ARG_CONST_MAP_PTR] = &const_map_ptr_types,
5176 [ARG_PTR_TO_CTX] = &context_types,
5177 [ARG_PTR_TO_SOCK_COMMON] = &sock_types,
5179 [ARG_PTR_TO_BTF_ID_SOCK_COMMON] = &btf_id_sock_common_types,
5181 [ARG_PTR_TO_SOCKET] = &fullsock_types,
5182 [ARG_PTR_TO_BTF_ID] = &btf_ptr_types,
5183 [ARG_PTR_TO_SPIN_LOCK] = &spin_lock_types,
5184 [ARG_PTR_TO_MEM] = &mem_types,
5185 [ARG_PTR_TO_UNINIT_MEM] = &mem_types,
5186 [ARG_PTR_TO_ALLOC_MEM] = &alloc_mem_types,
5187 [ARG_PTR_TO_INT] = &int_ptr_types,
5188 [ARG_PTR_TO_LONG] = &int_ptr_types,
5189 [ARG_PTR_TO_PERCPU_BTF_ID] = &percpu_btf_ptr_types,
5190 [ARG_PTR_TO_FUNC] = &func_ptr_types,
5191 [ARG_PTR_TO_STACK] = &stack_ptr_types,
5192 [ARG_PTR_TO_CONST_STR] = &const_str_ptr_types,
5193 [ARG_PTR_TO_TIMER] = &timer_types,
5196 static int check_reg_type(struct bpf_verifier_env *env, u32 regno,
5197 enum bpf_arg_type arg_type,
5198 const u32 *arg_btf_id)
5200 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
5201 enum bpf_reg_type expected, type = reg->type;
5202 const struct bpf_reg_types *compatible;
5205 compatible = compatible_reg_types[base_type(arg_type)];
5207 verbose(env, "verifier internal error: unsupported arg type %d\n", arg_type);
5211 /* ARG_PTR_TO_MEM + RDONLY is compatible with PTR_TO_MEM and PTR_TO_MEM + RDONLY,
5212 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM and NOT with PTR_TO_MEM + RDONLY
5214 * Same for MAYBE_NULL:
5216 * ARG_PTR_TO_MEM + MAYBE_NULL is compatible with PTR_TO_MEM and PTR_TO_MEM + MAYBE_NULL,
5217 * but ARG_PTR_TO_MEM is compatible only with PTR_TO_MEM but NOT with PTR_TO_MEM + MAYBE_NULL
5219 * Therefore we fold these flags depending on the arg_type before comparison.
5221 if (arg_type & MEM_RDONLY)
5222 type &= ~MEM_RDONLY;
5223 if (arg_type & PTR_MAYBE_NULL)
5224 type &= ~PTR_MAYBE_NULL;
5226 for (i = 0; i < ARRAY_SIZE(compatible->types); i++) {
5227 expected = compatible->types[i];
5228 if (expected == NOT_INIT)
5231 if (type == expected)
5235 verbose(env, "R%d type=%s expected=", regno, reg_type_str(env, reg->type));
5236 for (j = 0; j + 1 < i; j++)
5237 verbose(env, "%s, ", reg_type_str(env, compatible->types[j]));
5238 verbose(env, "%s\n", reg_type_str(env, compatible->types[j]));
5242 if (reg->type == PTR_TO_BTF_ID) {
5244 if (!compatible->btf_id) {
5245 verbose(env, "verifier internal error: missing arg compatible BTF ID\n");
5248 arg_btf_id = compatible->btf_id;
5251 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
5252 btf_vmlinux, *arg_btf_id)) {
5253 verbose(env, "R%d is of type %s but %s is expected\n",
5254 regno, kernel_type_name(reg->btf, reg->btf_id),
5255 kernel_type_name(btf_vmlinux, *arg_btf_id));
5263 static int check_func_arg(struct bpf_verifier_env *env, u32 arg,
5264 struct bpf_call_arg_meta *meta,
5265 const struct bpf_func_proto *fn)
5267 u32 regno = BPF_REG_1 + arg;
5268 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
5269 enum bpf_arg_type arg_type = fn->arg_type[arg];
5270 enum bpf_reg_type type = reg->type;
5273 if (arg_type == ARG_DONTCARE)
5276 err = check_reg_arg(env, regno, SRC_OP);
5280 if (arg_type == ARG_ANYTHING) {
5281 if (is_pointer_value(env, regno)) {
5282 verbose(env, "R%d leaks addr into helper function\n",
5289 if (type_is_pkt_pointer(type) &&
5290 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
5291 verbose(env, "helper access to the packet is not allowed\n");
5295 if (base_type(arg_type) == ARG_PTR_TO_MAP_VALUE ||
5296 base_type(arg_type) == ARG_PTR_TO_UNINIT_MAP_VALUE) {
5297 err = resolve_map_arg_type(env, meta, &arg_type);
5302 if (register_is_null(reg) && type_may_be_null(arg_type))
5303 /* A NULL register has a SCALAR_VALUE type, so skip
5306 goto skip_type_check;
5308 err = check_reg_type(env, regno, arg_type, fn->arg_btf_id[arg]);
5312 switch ((u32)type) {
5314 /* Pointer types where reg offset is explicitly allowed: */
5316 case PTR_TO_PACKET_META:
5317 case PTR_TO_MAP_KEY:
5318 case PTR_TO_MAP_VALUE:
5320 case PTR_TO_MEM | MEM_RDONLY:
5321 case PTR_TO_MEM | MEM_ALLOC:
5323 case PTR_TO_BUF | MEM_RDONLY:
5325 /* Some of the argument types nevertheless require a
5326 * zero register offset.
5328 if (arg_type == ARG_PTR_TO_ALLOC_MEM)
5329 goto force_off_check;
5331 /* All the rest must be rejected: */
5334 err = __check_ptr_off_reg(env, reg, regno,
5335 type == PTR_TO_BTF_ID);
5342 if (reg->ref_obj_id) {
5343 if (meta->ref_obj_id) {
5344 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
5345 regno, reg->ref_obj_id,
5349 meta->ref_obj_id = reg->ref_obj_id;
5352 if (arg_type == ARG_CONST_MAP_PTR) {
5353 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
5354 if (meta->map_ptr) {
5355 /* Use map_uid (which is unique id of inner map) to reject:
5356 * inner_map1 = bpf_map_lookup_elem(outer_map, key1)
5357 * inner_map2 = bpf_map_lookup_elem(outer_map, key2)
5358 * if (inner_map1 && inner_map2) {
5359 * timer = bpf_map_lookup_elem(inner_map1);
5361 * // mismatch would have been allowed
5362 * bpf_timer_init(timer, inner_map2);
5365 * Comparing map_ptr is enough to distinguish normal and outer maps.
5367 if (meta->map_ptr != reg->map_ptr ||
5368 meta->map_uid != reg->map_uid) {
5370 "timer pointer in R1 map_uid=%d doesn't match map pointer in R2 map_uid=%d\n",
5371 meta->map_uid, reg->map_uid);
5375 meta->map_ptr = reg->map_ptr;
5376 meta->map_uid = reg->map_uid;
5377 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
5378 /* bpf_map_xxx(..., map_ptr, ..., key) call:
5379 * check that [key, key + map->key_size) are within
5380 * stack limits and initialized
5382 if (!meta->map_ptr) {
5383 /* in function declaration map_ptr must come before
5384 * map_key, so that it's verified and known before
5385 * we have to check map_key here. Otherwise it means
5386 * that kernel subsystem misconfigured verifier
5388 verbose(env, "invalid map_ptr to access map->key\n");
5391 err = check_helper_mem_access(env, regno,
5392 meta->map_ptr->key_size, false,
5394 } else if (base_type(arg_type) == ARG_PTR_TO_MAP_VALUE ||
5395 base_type(arg_type) == ARG_PTR_TO_UNINIT_MAP_VALUE) {
5396 if (type_may_be_null(arg_type) && register_is_null(reg))
5399 /* bpf_map_xxx(..., map_ptr, ..., value) call:
5400 * check [value, value + map->value_size) validity
5402 if (!meta->map_ptr) {
5403 /* kernel subsystem misconfigured verifier */
5404 verbose(env, "invalid map_ptr to access map->value\n");
5407 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE);
5408 err = check_helper_mem_access(env, regno,
5409 meta->map_ptr->value_size, false,
5411 } else if (arg_type == ARG_PTR_TO_PERCPU_BTF_ID) {
5413 verbose(env, "Helper has invalid btf_id in R%d\n", regno);
5416 meta->ret_btf = reg->btf;
5417 meta->ret_btf_id = reg->btf_id;
5418 } else if (arg_type == ARG_PTR_TO_SPIN_LOCK) {
5419 if (meta->func_id == BPF_FUNC_spin_lock) {
5420 if (process_spin_lock(env, regno, true))
5422 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
5423 if (process_spin_lock(env, regno, false))
5426 verbose(env, "verifier internal error\n");
5429 } else if (arg_type == ARG_PTR_TO_TIMER) {
5430 if (process_timer_func(env, regno, meta))
5432 } else if (arg_type == ARG_PTR_TO_FUNC) {
5433 meta->subprogno = reg->subprogno;
5434 } else if (arg_type_is_mem_ptr(arg_type)) {
5435 /* The access to this pointer is only checked when we hit the
5436 * next is_mem_size argument below.
5438 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MEM);
5439 } else if (arg_type_is_mem_size(arg_type)) {
5440 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
5442 /* This is used to refine r0 return value bounds for helpers
5443 * that enforce this value as an upper bound on return values.
5444 * See do_refine_retval_range() for helpers that can refine
5445 * the return value. C type of helper is u32 so we pull register
5446 * bound from umax_value however, if negative verifier errors
5447 * out. Only upper bounds can be learned because retval is an
5448 * int type and negative retvals are allowed.
5450 meta->msize_max_value = reg->umax_value;
5452 /* The register is SCALAR_VALUE; the access check
5453 * happens using its boundaries.
5455 if (!tnum_is_const(reg->var_off))
5456 /* For unprivileged variable accesses, disable raw
5457 * mode so that the program is required to
5458 * initialize all the memory that the helper could
5459 * just partially fill up.
5463 if (reg->smin_value < 0) {
5464 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
5469 if (reg->umin_value == 0) {
5470 err = check_helper_mem_access(env, regno - 1, 0,
5477 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
5478 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
5482 err = check_helper_mem_access(env, regno - 1,
5484 zero_size_allowed, meta);
5486 err = mark_chain_precision(env, regno);
5487 } else if (arg_type_is_alloc_size(arg_type)) {
5488 if (!tnum_is_const(reg->var_off)) {
5489 verbose(env, "R%d is not a known constant'\n",
5493 meta->mem_size = reg->var_off.value;
5494 } else if (arg_type_is_int_ptr(arg_type)) {
5495 int size = int_ptr_type_to_size(arg_type);
5497 err = check_helper_mem_access(env, regno, size, false, meta);
5500 err = check_ptr_alignment(env, reg, 0, size, true);
5501 } else if (arg_type == ARG_PTR_TO_CONST_STR) {
5502 struct bpf_map *map = reg->map_ptr;
5507 if (!bpf_map_is_rdonly(map)) {
5508 verbose(env, "R%d does not point to a readonly map'\n", regno);
5512 if (!tnum_is_const(reg->var_off)) {
5513 verbose(env, "R%d is not a constant address'\n", regno);
5517 if (!map->ops->map_direct_value_addr) {
5518 verbose(env, "no direct value access support for this map type\n");
5522 err = check_map_access(env, regno, reg->off,
5523 map->value_size - reg->off, false);
5527 map_off = reg->off + reg->var_off.value;
5528 err = map->ops->map_direct_value_addr(map, &map_addr, map_off);
5530 verbose(env, "direct value access on string failed\n");
5534 str_ptr = (char *)(long)(map_addr);
5535 if (!strnchr(str_ptr + map_off, map->value_size - map_off, 0)) {
5536 verbose(env, "string is not zero-terminated\n");
5544 static bool may_update_sockmap(struct bpf_verifier_env *env, int func_id)
5546 enum bpf_attach_type eatype = env->prog->expected_attach_type;
5547 enum bpf_prog_type type = resolve_prog_type(env->prog);
5549 if (func_id != BPF_FUNC_map_update_elem)
5552 /* It's not possible to get access to a locked struct sock in these
5553 * contexts, so updating is safe.
5556 case BPF_PROG_TYPE_TRACING:
5557 if (eatype == BPF_TRACE_ITER)
5560 case BPF_PROG_TYPE_SOCKET_FILTER:
5561 case BPF_PROG_TYPE_SCHED_CLS:
5562 case BPF_PROG_TYPE_SCHED_ACT:
5563 case BPF_PROG_TYPE_XDP:
5564 case BPF_PROG_TYPE_SK_REUSEPORT:
5565 case BPF_PROG_TYPE_FLOW_DISSECTOR:
5566 case BPF_PROG_TYPE_SK_LOOKUP:
5572 verbose(env, "cannot update sockmap in this context\n");
5576 static bool allow_tail_call_in_subprogs(struct bpf_verifier_env *env)
5578 return env->prog->jit_requested && IS_ENABLED(CONFIG_X86_64);
5581 static int check_map_func_compatibility(struct bpf_verifier_env *env,
5582 struct bpf_map *map, int func_id)
5587 /* We need a two way check, first is from map perspective ... */
5588 switch (map->map_type) {
5589 case BPF_MAP_TYPE_PROG_ARRAY:
5590 if (func_id != BPF_FUNC_tail_call)
5593 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
5594 if (func_id != BPF_FUNC_perf_event_read &&
5595 func_id != BPF_FUNC_perf_event_output &&
5596 func_id != BPF_FUNC_skb_output &&
5597 func_id != BPF_FUNC_perf_event_read_value &&
5598 func_id != BPF_FUNC_xdp_output)
5601 case BPF_MAP_TYPE_RINGBUF:
5602 if (func_id != BPF_FUNC_ringbuf_output &&
5603 func_id != BPF_FUNC_ringbuf_reserve &&
5604 func_id != BPF_FUNC_ringbuf_query)
5607 case BPF_MAP_TYPE_STACK_TRACE:
5608 if (func_id != BPF_FUNC_get_stackid)
5611 case BPF_MAP_TYPE_CGROUP_ARRAY:
5612 if (func_id != BPF_FUNC_skb_under_cgroup &&
5613 func_id != BPF_FUNC_current_task_under_cgroup)
5616 case BPF_MAP_TYPE_CGROUP_STORAGE:
5617 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
5618 if (func_id != BPF_FUNC_get_local_storage)
5621 case BPF_MAP_TYPE_DEVMAP:
5622 case BPF_MAP_TYPE_DEVMAP_HASH:
5623 if (func_id != BPF_FUNC_redirect_map &&
5624 func_id != BPF_FUNC_map_lookup_elem)
5627 /* Restrict bpf side of cpumap and xskmap, open when use-cases
5630 case BPF_MAP_TYPE_CPUMAP:
5631 if (func_id != BPF_FUNC_redirect_map)
5634 case BPF_MAP_TYPE_XSKMAP:
5635 if (func_id != BPF_FUNC_redirect_map &&
5636 func_id != BPF_FUNC_map_lookup_elem)
5639 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
5640 case BPF_MAP_TYPE_HASH_OF_MAPS:
5641 if (func_id != BPF_FUNC_map_lookup_elem)
5644 case BPF_MAP_TYPE_SOCKMAP:
5645 if (func_id != BPF_FUNC_sk_redirect_map &&
5646 func_id != BPF_FUNC_sock_map_update &&
5647 func_id != BPF_FUNC_map_delete_elem &&
5648 func_id != BPF_FUNC_msg_redirect_map &&
5649 func_id != BPF_FUNC_sk_select_reuseport &&
5650 func_id != BPF_FUNC_map_lookup_elem &&
5651 !may_update_sockmap(env, func_id))
5654 case BPF_MAP_TYPE_SOCKHASH:
5655 if (func_id != BPF_FUNC_sk_redirect_hash &&
5656 func_id != BPF_FUNC_sock_hash_update &&
5657 func_id != BPF_FUNC_map_delete_elem &&
5658 func_id != BPF_FUNC_msg_redirect_hash &&
5659 func_id != BPF_FUNC_sk_select_reuseport &&
5660 func_id != BPF_FUNC_map_lookup_elem &&
5661 !may_update_sockmap(env, func_id))
5664 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
5665 if (func_id != BPF_FUNC_sk_select_reuseport)
5668 case BPF_MAP_TYPE_QUEUE:
5669 case BPF_MAP_TYPE_STACK:
5670 if (func_id != BPF_FUNC_map_peek_elem &&
5671 func_id != BPF_FUNC_map_pop_elem &&
5672 func_id != BPF_FUNC_map_push_elem)
5675 case BPF_MAP_TYPE_SK_STORAGE:
5676 if (func_id != BPF_FUNC_sk_storage_get &&
5677 func_id != BPF_FUNC_sk_storage_delete)
5680 case BPF_MAP_TYPE_INODE_STORAGE:
5681 if (func_id != BPF_FUNC_inode_storage_get &&
5682 func_id != BPF_FUNC_inode_storage_delete)
5685 case BPF_MAP_TYPE_TASK_STORAGE:
5686 if (func_id != BPF_FUNC_task_storage_get &&
5687 func_id != BPF_FUNC_task_storage_delete)
5690 case BPF_MAP_TYPE_BLOOM_FILTER:
5691 if (func_id != BPF_FUNC_map_peek_elem &&
5692 func_id != BPF_FUNC_map_push_elem)
5699 /* ... and second from the function itself. */
5701 case BPF_FUNC_tail_call:
5702 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
5704 if (env->subprog_cnt > 1 && !allow_tail_call_in_subprogs(env)) {
5705 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
5709 case BPF_FUNC_perf_event_read:
5710 case BPF_FUNC_perf_event_output:
5711 case BPF_FUNC_perf_event_read_value:
5712 case BPF_FUNC_skb_output:
5713 case BPF_FUNC_xdp_output:
5714 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
5717 case BPF_FUNC_ringbuf_output:
5718 case BPF_FUNC_ringbuf_reserve:
5719 case BPF_FUNC_ringbuf_query:
5720 if (map->map_type != BPF_MAP_TYPE_RINGBUF)
5723 case BPF_FUNC_get_stackid:
5724 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
5727 case BPF_FUNC_current_task_under_cgroup:
5728 case BPF_FUNC_skb_under_cgroup:
5729 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
5732 case BPF_FUNC_redirect_map:
5733 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
5734 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
5735 map->map_type != BPF_MAP_TYPE_CPUMAP &&
5736 map->map_type != BPF_MAP_TYPE_XSKMAP)
5739 case BPF_FUNC_sk_redirect_map:
5740 case BPF_FUNC_msg_redirect_map:
5741 case BPF_FUNC_sock_map_update:
5742 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
5745 case BPF_FUNC_sk_redirect_hash:
5746 case BPF_FUNC_msg_redirect_hash:
5747 case BPF_FUNC_sock_hash_update:
5748 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
5751 case BPF_FUNC_get_local_storage:
5752 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
5753 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
5756 case BPF_FUNC_sk_select_reuseport:
5757 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
5758 map->map_type != BPF_MAP_TYPE_SOCKMAP &&
5759 map->map_type != BPF_MAP_TYPE_SOCKHASH)
5762 case BPF_FUNC_map_pop_elem:
5763 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
5764 map->map_type != BPF_MAP_TYPE_STACK)
5767 case BPF_FUNC_map_peek_elem:
5768 case BPF_FUNC_map_push_elem:
5769 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
5770 map->map_type != BPF_MAP_TYPE_STACK &&
5771 map->map_type != BPF_MAP_TYPE_BLOOM_FILTER)
5774 case BPF_FUNC_sk_storage_get:
5775 case BPF_FUNC_sk_storage_delete:
5776 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
5779 case BPF_FUNC_inode_storage_get:
5780 case BPF_FUNC_inode_storage_delete:
5781 if (map->map_type != BPF_MAP_TYPE_INODE_STORAGE)
5784 case BPF_FUNC_task_storage_get:
5785 case BPF_FUNC_task_storage_delete:
5786 if (map->map_type != BPF_MAP_TYPE_TASK_STORAGE)
5795 verbose(env, "cannot pass map_type %d into func %s#%d\n",
5796 map->map_type, func_id_name(func_id), func_id);
5800 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
5804 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
5806 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
5808 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
5810 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
5812 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
5815 /* We only support one arg being in raw mode at the moment,
5816 * which is sufficient for the helper functions we have
5822 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
5823 enum bpf_arg_type arg_next)
5825 return (arg_type_is_mem_ptr(arg_curr) &&
5826 !arg_type_is_mem_size(arg_next)) ||
5827 (!arg_type_is_mem_ptr(arg_curr) &&
5828 arg_type_is_mem_size(arg_next));
5831 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
5833 /* bpf_xxx(..., buf, len) call will access 'len'
5834 * bytes from memory 'buf'. Both arg types need
5835 * to be paired, so make sure there's no buggy
5836 * helper function specification.
5838 if (arg_type_is_mem_size(fn->arg1_type) ||
5839 arg_type_is_mem_ptr(fn->arg5_type) ||
5840 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
5841 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
5842 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
5843 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
5849 static bool check_refcount_ok(const struct bpf_func_proto *fn, int func_id)
5853 if (arg_type_may_be_refcounted(fn->arg1_type))
5855 if (arg_type_may_be_refcounted(fn->arg2_type))
5857 if (arg_type_may_be_refcounted(fn->arg3_type))
5859 if (arg_type_may_be_refcounted(fn->arg4_type))
5861 if (arg_type_may_be_refcounted(fn->arg5_type))
5864 /* A reference acquiring function cannot acquire
5865 * another refcounted ptr.
5867 if (may_be_acquire_function(func_id) && count)
5870 /* We only support one arg being unreferenced at the moment,
5871 * which is sufficient for the helper functions we have right now.
5876 static bool check_btf_id_ok(const struct bpf_func_proto *fn)
5880 for (i = 0; i < ARRAY_SIZE(fn->arg_type); i++) {
5881 if (fn->arg_type[i] == ARG_PTR_TO_BTF_ID && !fn->arg_btf_id[i])
5884 if (fn->arg_type[i] != ARG_PTR_TO_BTF_ID && fn->arg_btf_id[i])
5891 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
5893 return check_raw_mode_ok(fn) &&
5894 check_arg_pair_ok(fn) &&
5895 check_btf_id_ok(fn) &&
5896 check_refcount_ok(fn, func_id) ? 0 : -EINVAL;
5899 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
5900 * are now invalid, so turn them into unknown SCALAR_VALUE.
5902 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
5903 struct bpf_func_state *state)
5905 struct bpf_reg_state *regs = state->regs, *reg;
5908 for (i = 0; i < MAX_BPF_REG; i++)
5909 if (reg_is_pkt_pointer_any(®s[i]))
5910 mark_reg_unknown(env, regs, i);
5912 bpf_for_each_spilled_reg(i, state, reg) {
5915 if (reg_is_pkt_pointer_any(reg))
5916 __mark_reg_unknown(env, reg);
5920 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
5922 struct bpf_verifier_state *vstate = env->cur_state;
5925 for (i = 0; i <= vstate->curframe; i++)
5926 __clear_all_pkt_pointers(env, vstate->frame[i]);
5931 BEYOND_PKT_END = -2,
5934 static void mark_pkt_end(struct bpf_verifier_state *vstate, int regn, bool range_open)
5936 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5937 struct bpf_reg_state *reg = &state->regs[regn];
5939 if (reg->type != PTR_TO_PACKET)
5940 /* PTR_TO_PACKET_META is not supported yet */
5943 /* The 'reg' is pkt > pkt_end or pkt >= pkt_end.
5944 * How far beyond pkt_end it goes is unknown.
5945 * if (!range_open) it's the case of pkt >= pkt_end
5946 * if (range_open) it's the case of pkt > pkt_end
5947 * hence this pointer is at least 1 byte bigger than pkt_end
5950 reg->range = BEYOND_PKT_END;
5952 reg->range = AT_PKT_END;
5955 static void release_reg_references(struct bpf_verifier_env *env,
5956 struct bpf_func_state *state,
5959 struct bpf_reg_state *regs = state->regs, *reg;
5962 for (i = 0; i < MAX_BPF_REG; i++)
5963 if (regs[i].ref_obj_id == ref_obj_id)
5964 mark_reg_unknown(env, regs, i);
5966 bpf_for_each_spilled_reg(i, state, reg) {
5969 if (reg->ref_obj_id == ref_obj_id)
5970 __mark_reg_unknown(env, reg);
5974 /* The pointer with the specified id has released its reference to kernel
5975 * resources. Identify all copies of the same pointer and clear the reference.
5977 static int release_reference(struct bpf_verifier_env *env,
5980 struct bpf_verifier_state *vstate = env->cur_state;
5984 err = release_reference_state(cur_func(env), ref_obj_id);
5988 for (i = 0; i <= vstate->curframe; i++)
5989 release_reg_references(env, vstate->frame[i], ref_obj_id);
5994 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
5995 struct bpf_reg_state *regs)
5999 /* after the call registers r0 - r5 were scratched */
6000 for (i = 0; i < CALLER_SAVED_REGS; i++) {
6001 mark_reg_not_init(env, regs, caller_saved[i]);
6002 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
6006 typedef int (*set_callee_state_fn)(struct bpf_verifier_env *env,
6007 struct bpf_func_state *caller,
6008 struct bpf_func_state *callee,
6011 static int __check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
6012 int *insn_idx, int subprog,
6013 set_callee_state_fn set_callee_state_cb)
6015 struct bpf_verifier_state *state = env->cur_state;
6016 struct bpf_func_info_aux *func_info_aux;
6017 struct bpf_func_state *caller, *callee;
6019 bool is_global = false;
6021 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
6022 verbose(env, "the call stack of %d frames is too deep\n",
6023 state->curframe + 2);
6027 caller = state->frame[state->curframe];
6028 if (state->frame[state->curframe + 1]) {
6029 verbose(env, "verifier bug. Frame %d already allocated\n",
6030 state->curframe + 1);
6034 func_info_aux = env->prog->aux->func_info_aux;
6036 is_global = func_info_aux[subprog].linkage == BTF_FUNC_GLOBAL;
6037 err = btf_check_subprog_arg_match(env, subprog, caller->regs);
6042 verbose(env, "Caller passes invalid args into func#%d\n",
6046 if (env->log.level & BPF_LOG_LEVEL)
6048 "Func#%d is global and valid. Skipping.\n",
6050 clear_caller_saved_regs(env, caller->regs);
6052 /* All global functions return a 64-bit SCALAR_VALUE */
6053 mark_reg_unknown(env, caller->regs, BPF_REG_0);
6054 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
6056 /* continue with next insn after call */
6061 if (insn->code == (BPF_JMP | BPF_CALL) &&
6062 insn->src_reg == 0 &&
6063 insn->imm == BPF_FUNC_timer_set_callback) {
6064 struct bpf_verifier_state *async_cb;
6066 /* there is no real recursion here. timer callbacks are async */
6067 env->subprog_info[subprog].is_async_cb = true;
6068 async_cb = push_async_cb(env, env->subprog_info[subprog].start,
6069 *insn_idx, subprog);
6072 callee = async_cb->frame[0];
6073 callee->async_entry_cnt = caller->async_entry_cnt + 1;
6075 /* Convert bpf_timer_set_callback() args into timer callback args */
6076 err = set_callee_state_cb(env, caller, callee, *insn_idx);
6080 clear_caller_saved_regs(env, caller->regs);
6081 mark_reg_unknown(env, caller->regs, BPF_REG_0);
6082 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
6083 /* continue with next insn after call */
6087 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
6090 state->frame[state->curframe + 1] = callee;
6092 /* callee cannot access r0, r6 - r9 for reading and has to write
6093 * into its own stack before reading from it.
6094 * callee can read/write into caller's stack
6096 init_func_state(env, callee,
6097 /* remember the callsite, it will be used by bpf_exit */
6098 *insn_idx /* callsite */,
6099 state->curframe + 1 /* frameno within this callchain */,
6100 subprog /* subprog number within this prog */);
6102 /* Transfer references to the callee */
6103 err = copy_reference_state(callee, caller);
6107 err = set_callee_state_cb(env, caller, callee, *insn_idx);
6111 clear_caller_saved_regs(env, caller->regs);
6113 /* only increment it after check_reg_arg() finished */
6116 /* and go analyze first insn of the callee */
6117 *insn_idx = env->subprog_info[subprog].start - 1;
6119 if (env->log.level & BPF_LOG_LEVEL) {
6120 verbose(env, "caller:\n");
6121 print_verifier_state(env, caller, true);
6122 verbose(env, "callee:\n");
6123 print_verifier_state(env, callee, true);
6128 int map_set_for_each_callback_args(struct bpf_verifier_env *env,
6129 struct bpf_func_state *caller,
6130 struct bpf_func_state *callee)
6132 /* bpf_for_each_map_elem(struct bpf_map *map, void *callback_fn,
6133 * void *callback_ctx, u64 flags);
6134 * callback_fn(struct bpf_map *map, void *key, void *value,
6135 * void *callback_ctx);
6137 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
6139 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
6140 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
6141 callee->regs[BPF_REG_2].map_ptr = caller->regs[BPF_REG_1].map_ptr;
6143 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
6144 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
6145 callee->regs[BPF_REG_3].map_ptr = caller->regs[BPF_REG_1].map_ptr;
6147 /* pointer to stack or null */
6148 callee->regs[BPF_REG_4] = caller->regs[BPF_REG_3];
6151 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
6155 static int set_callee_state(struct bpf_verifier_env *env,
6156 struct bpf_func_state *caller,
6157 struct bpf_func_state *callee, int insn_idx)
6161 /* copy r1 - r5 args that callee can access. The copy includes parent
6162 * pointers, which connects us up to the liveness chain
6164 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
6165 callee->regs[i] = caller->regs[i];
6169 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
6172 int subprog, target_insn;
6174 target_insn = *insn_idx + insn->imm + 1;
6175 subprog = find_subprog(env, target_insn);
6177 verbose(env, "verifier bug. No program starts at insn %d\n",
6182 return __check_func_call(env, insn, insn_idx, subprog, set_callee_state);
6185 static int set_map_elem_callback_state(struct bpf_verifier_env *env,
6186 struct bpf_func_state *caller,
6187 struct bpf_func_state *callee,
6190 struct bpf_insn_aux_data *insn_aux = &env->insn_aux_data[insn_idx];
6191 struct bpf_map *map;
6194 if (bpf_map_ptr_poisoned(insn_aux)) {
6195 verbose(env, "tail_call abusing map_ptr\n");
6199 map = BPF_MAP_PTR(insn_aux->map_ptr_state);
6200 if (!map->ops->map_set_for_each_callback_args ||
6201 !map->ops->map_for_each_callback) {
6202 verbose(env, "callback function not allowed for map\n");
6206 err = map->ops->map_set_for_each_callback_args(env, caller, callee);
6210 callee->in_callback_fn = true;
6214 static int set_loop_callback_state(struct bpf_verifier_env *env,
6215 struct bpf_func_state *caller,
6216 struct bpf_func_state *callee,
6219 /* bpf_loop(u32 nr_loops, void *callback_fn, void *callback_ctx,
6221 * callback_fn(u32 index, void *callback_ctx);
6223 callee->regs[BPF_REG_1].type = SCALAR_VALUE;
6224 callee->regs[BPF_REG_2] = caller->regs[BPF_REG_3];
6227 __mark_reg_not_init(env, &callee->regs[BPF_REG_3]);
6228 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
6229 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
6231 callee->in_callback_fn = true;
6235 static int set_timer_callback_state(struct bpf_verifier_env *env,
6236 struct bpf_func_state *caller,
6237 struct bpf_func_state *callee,
6240 struct bpf_map *map_ptr = caller->regs[BPF_REG_1].map_ptr;
6242 /* bpf_timer_set_callback(struct bpf_timer *timer, void *callback_fn);
6243 * callback_fn(struct bpf_map *map, void *key, void *value);
6245 callee->regs[BPF_REG_1].type = CONST_PTR_TO_MAP;
6246 __mark_reg_known_zero(&callee->regs[BPF_REG_1]);
6247 callee->regs[BPF_REG_1].map_ptr = map_ptr;
6249 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
6250 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
6251 callee->regs[BPF_REG_2].map_ptr = map_ptr;
6253 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
6254 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
6255 callee->regs[BPF_REG_3].map_ptr = map_ptr;
6258 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
6259 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
6260 callee->in_async_callback_fn = true;
6264 static int set_find_vma_callback_state(struct bpf_verifier_env *env,
6265 struct bpf_func_state *caller,
6266 struct bpf_func_state *callee,
6269 /* bpf_find_vma(struct task_struct *task, u64 addr,
6270 * void *callback_fn, void *callback_ctx, u64 flags)
6271 * (callback_fn)(struct task_struct *task,
6272 * struct vm_area_struct *vma, void *callback_ctx);
6274 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
6276 callee->regs[BPF_REG_2].type = PTR_TO_BTF_ID;
6277 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
6278 callee->regs[BPF_REG_2].btf = btf_vmlinux;
6279 callee->regs[BPF_REG_2].btf_id = btf_tracing_ids[BTF_TRACING_TYPE_VMA],
6281 /* pointer to stack or null */
6282 callee->regs[BPF_REG_3] = caller->regs[BPF_REG_4];
6285 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
6286 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
6287 callee->in_callback_fn = true;
6291 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
6293 struct bpf_verifier_state *state = env->cur_state;
6294 struct bpf_func_state *caller, *callee;
6295 struct bpf_reg_state *r0;
6298 callee = state->frame[state->curframe];
6299 r0 = &callee->regs[BPF_REG_0];
6300 if (r0->type == PTR_TO_STACK) {
6301 /* technically it's ok to return caller's stack pointer
6302 * (or caller's caller's pointer) back to the caller,
6303 * since these pointers are valid. Only current stack
6304 * pointer will be invalid as soon as function exits,
6305 * but let's be conservative
6307 verbose(env, "cannot return stack pointer to the caller\n");
6312 caller = state->frame[state->curframe];
6313 if (callee->in_callback_fn) {
6314 /* enforce R0 return value range [0, 1]. */
6315 struct tnum range = tnum_range(0, 1);
6317 if (r0->type != SCALAR_VALUE) {
6318 verbose(env, "R0 not a scalar value\n");
6321 if (!tnum_in(range, r0->var_off)) {
6322 verbose_invalid_scalar(env, r0, &range, "callback return", "R0");
6326 /* return to the caller whatever r0 had in the callee */
6327 caller->regs[BPF_REG_0] = *r0;
6330 /* Transfer references to the caller */
6331 err = copy_reference_state(caller, callee);
6335 *insn_idx = callee->callsite + 1;
6336 if (env->log.level & BPF_LOG_LEVEL) {
6337 verbose(env, "returning from callee:\n");
6338 print_verifier_state(env, callee, true);
6339 verbose(env, "to caller at %d:\n", *insn_idx);
6340 print_verifier_state(env, caller, true);
6342 /* clear everything in the callee */
6343 free_func_state(callee);
6344 state->frame[state->curframe + 1] = NULL;
6348 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
6350 struct bpf_call_arg_meta *meta)
6352 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
6354 if (ret_type != RET_INTEGER ||
6355 (func_id != BPF_FUNC_get_stack &&
6356 func_id != BPF_FUNC_get_task_stack &&
6357 func_id != BPF_FUNC_probe_read_str &&
6358 func_id != BPF_FUNC_probe_read_kernel_str &&
6359 func_id != BPF_FUNC_probe_read_user_str))
6362 ret_reg->smax_value = meta->msize_max_value;
6363 ret_reg->s32_max_value = meta->msize_max_value;
6364 ret_reg->smin_value = -MAX_ERRNO;
6365 ret_reg->s32_min_value = -MAX_ERRNO;
6366 __reg_deduce_bounds(ret_reg);
6367 __reg_bound_offset(ret_reg);
6368 __update_reg_bounds(ret_reg);
6372 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
6373 int func_id, int insn_idx)
6375 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
6376 struct bpf_map *map = meta->map_ptr;
6378 if (func_id != BPF_FUNC_tail_call &&
6379 func_id != BPF_FUNC_map_lookup_elem &&
6380 func_id != BPF_FUNC_map_update_elem &&
6381 func_id != BPF_FUNC_map_delete_elem &&
6382 func_id != BPF_FUNC_map_push_elem &&
6383 func_id != BPF_FUNC_map_pop_elem &&
6384 func_id != BPF_FUNC_map_peek_elem &&
6385 func_id != BPF_FUNC_for_each_map_elem &&
6386 func_id != BPF_FUNC_redirect_map)
6390 verbose(env, "kernel subsystem misconfigured verifier\n");
6394 /* In case of read-only, some additional restrictions
6395 * need to be applied in order to prevent altering the
6396 * state of the map from program side.
6398 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
6399 (func_id == BPF_FUNC_map_delete_elem ||
6400 func_id == BPF_FUNC_map_update_elem ||
6401 func_id == BPF_FUNC_map_push_elem ||
6402 func_id == BPF_FUNC_map_pop_elem)) {
6403 verbose(env, "write into map forbidden\n");
6407 if (!BPF_MAP_PTR(aux->map_ptr_state))
6408 bpf_map_ptr_store(aux, meta->map_ptr,
6409 !meta->map_ptr->bypass_spec_v1);
6410 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
6411 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
6412 !meta->map_ptr->bypass_spec_v1);
6417 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
6418 int func_id, int insn_idx)
6420 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
6421 struct bpf_reg_state *regs = cur_regs(env), *reg;
6422 struct bpf_map *map = meta->map_ptr;
6427 if (func_id != BPF_FUNC_tail_call)
6429 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
6430 verbose(env, "kernel subsystem misconfigured verifier\n");
6434 range = tnum_range(0, map->max_entries - 1);
6435 reg = ®s[BPF_REG_3];
6437 if (!register_is_const(reg) || !tnum_in(range, reg->var_off)) {
6438 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
6442 err = mark_chain_precision(env, BPF_REG_3);
6446 val = reg->var_off.value;
6447 if (bpf_map_key_unseen(aux))
6448 bpf_map_key_store(aux, val);
6449 else if (!bpf_map_key_poisoned(aux) &&
6450 bpf_map_key_immediate(aux) != val)
6451 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
6455 static int check_reference_leak(struct bpf_verifier_env *env)
6457 struct bpf_func_state *state = cur_func(env);
6460 for (i = 0; i < state->acquired_refs; i++) {
6461 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
6462 state->refs[i].id, state->refs[i].insn_idx);
6464 return state->acquired_refs ? -EINVAL : 0;
6467 static int check_bpf_snprintf_call(struct bpf_verifier_env *env,
6468 struct bpf_reg_state *regs)
6470 struct bpf_reg_state *fmt_reg = ®s[BPF_REG_3];
6471 struct bpf_reg_state *data_len_reg = ®s[BPF_REG_5];
6472 struct bpf_map *fmt_map = fmt_reg->map_ptr;
6473 int err, fmt_map_off, num_args;
6477 /* data must be an array of u64 */
6478 if (data_len_reg->var_off.value % 8)
6480 num_args = data_len_reg->var_off.value / 8;
6482 /* fmt being ARG_PTR_TO_CONST_STR guarantees that var_off is const
6483 * and map_direct_value_addr is set.
6485 fmt_map_off = fmt_reg->off + fmt_reg->var_off.value;
6486 err = fmt_map->ops->map_direct_value_addr(fmt_map, &fmt_addr,
6489 verbose(env, "verifier bug\n");
6492 fmt = (char *)(long)fmt_addr + fmt_map_off;
6494 /* We are also guaranteed that fmt+fmt_map_off is NULL terminated, we
6495 * can focus on validating the format specifiers.
6497 err = bpf_bprintf_prepare(fmt, UINT_MAX, NULL, NULL, num_args);
6499 verbose(env, "Invalid format string\n");
6504 static int check_get_func_ip(struct bpf_verifier_env *env)
6506 enum bpf_prog_type type = resolve_prog_type(env->prog);
6507 int func_id = BPF_FUNC_get_func_ip;
6509 if (type == BPF_PROG_TYPE_TRACING) {
6510 if (!bpf_prog_has_trampoline(env->prog)) {
6511 verbose(env, "func %s#%d supported only for fentry/fexit/fmod_ret programs\n",
6512 func_id_name(func_id), func_id);
6516 } else if (type == BPF_PROG_TYPE_KPROBE) {
6520 verbose(env, "func %s#%d not supported for program type %d\n",
6521 func_id_name(func_id), func_id, type);
6525 static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
6528 const struct bpf_func_proto *fn = NULL;
6529 enum bpf_return_type ret_type;
6530 enum bpf_type_flag ret_flag;
6531 struct bpf_reg_state *regs;
6532 struct bpf_call_arg_meta meta;
6533 int insn_idx = *insn_idx_p;
6535 int i, err, func_id;
6537 /* find function prototype */
6538 func_id = insn->imm;
6539 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
6540 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
6545 if (env->ops->get_func_proto)
6546 fn = env->ops->get_func_proto(func_id, env->prog);
6548 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
6553 /* eBPF programs must be GPL compatible to use GPL-ed functions */
6554 if (!env->prog->gpl_compatible && fn->gpl_only) {
6555 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
6559 if (fn->allowed && !fn->allowed(env->prog)) {
6560 verbose(env, "helper call is not allowed in probe\n");
6564 /* With LD_ABS/IND some JITs save/restore skb from r1. */
6565 changes_data = bpf_helper_changes_pkt_data(fn->func);
6566 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
6567 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
6568 func_id_name(func_id), func_id);
6572 memset(&meta, 0, sizeof(meta));
6573 meta.pkt_access = fn->pkt_access;
6575 err = check_func_proto(fn, func_id);
6577 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
6578 func_id_name(func_id), func_id);
6582 meta.func_id = func_id;
6584 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++) {
6585 err = check_func_arg(env, i, &meta, fn);
6590 err = record_func_map(env, &meta, func_id, insn_idx);
6594 err = record_func_key(env, &meta, func_id, insn_idx);
6598 /* Mark slots with STACK_MISC in case of raw mode, stack offset
6599 * is inferred from register state.
6601 for (i = 0; i < meta.access_size; i++) {
6602 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
6603 BPF_WRITE, -1, false);
6608 if (is_release_function(func_id)) {
6609 err = release_reference(env, meta.ref_obj_id);
6611 verbose(env, "func %s#%d reference has not been acquired before\n",
6612 func_id_name(func_id), func_id);
6617 regs = cur_regs(env);
6620 case BPF_FUNC_tail_call:
6621 err = check_reference_leak(env);
6623 verbose(env, "tail_call would lead to reference leak\n");
6627 case BPF_FUNC_get_local_storage:
6628 /* check that flags argument in get_local_storage(map, flags) is 0,
6629 * this is required because get_local_storage() can't return an error.
6631 if (!register_is_null(®s[BPF_REG_2])) {
6632 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
6636 case BPF_FUNC_for_each_map_elem:
6637 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
6638 set_map_elem_callback_state);
6640 case BPF_FUNC_timer_set_callback:
6641 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
6642 set_timer_callback_state);
6644 case BPF_FUNC_find_vma:
6645 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
6646 set_find_vma_callback_state);
6648 case BPF_FUNC_snprintf:
6649 err = check_bpf_snprintf_call(env, regs);
6652 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
6653 set_loop_callback_state);
6660 /* reset caller saved regs */
6661 for (i = 0; i < CALLER_SAVED_REGS; i++) {
6662 mark_reg_not_init(env, regs, caller_saved[i]);
6663 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
6666 /* helper call returns 64-bit value. */
6667 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
6669 /* update return register (already marked as written above) */
6670 ret_type = fn->ret_type;
6671 ret_flag = type_flag(fn->ret_type);
6672 if (ret_type == RET_INTEGER) {
6673 /* sets type to SCALAR_VALUE */
6674 mark_reg_unknown(env, regs, BPF_REG_0);
6675 } else if (ret_type == RET_VOID) {
6676 regs[BPF_REG_0].type = NOT_INIT;
6677 } else if (base_type(ret_type) == RET_PTR_TO_MAP_VALUE) {
6678 /* There is no offset yet applied, variable or fixed */
6679 mark_reg_known_zero(env, regs, BPF_REG_0);
6680 /* remember map_ptr, so that check_map_access()
6681 * can check 'value_size' boundary of memory access
6682 * to map element returned from bpf_map_lookup_elem()
6684 if (meta.map_ptr == NULL) {
6686 "kernel subsystem misconfigured verifier\n");
6689 regs[BPF_REG_0].map_ptr = meta.map_ptr;
6690 regs[BPF_REG_0].map_uid = meta.map_uid;
6691 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE | ret_flag;
6692 if (!type_may_be_null(ret_type) &&
6693 map_value_has_spin_lock(meta.map_ptr)) {
6694 regs[BPF_REG_0].id = ++env->id_gen;
6696 } else if (base_type(ret_type) == RET_PTR_TO_SOCKET) {
6697 mark_reg_known_zero(env, regs, BPF_REG_0);
6698 regs[BPF_REG_0].type = PTR_TO_SOCKET | ret_flag;
6699 } else if (base_type(ret_type) == RET_PTR_TO_SOCK_COMMON) {
6700 mark_reg_known_zero(env, regs, BPF_REG_0);
6701 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON | ret_flag;
6702 } else if (base_type(ret_type) == RET_PTR_TO_TCP_SOCK) {
6703 mark_reg_known_zero(env, regs, BPF_REG_0);
6704 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK | ret_flag;
6705 } else if (base_type(ret_type) == RET_PTR_TO_ALLOC_MEM) {
6706 mark_reg_known_zero(env, regs, BPF_REG_0);
6707 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
6708 regs[BPF_REG_0].mem_size = meta.mem_size;
6709 } else if (base_type(ret_type) == RET_PTR_TO_MEM_OR_BTF_ID) {
6710 const struct btf_type *t;
6712 mark_reg_known_zero(env, regs, BPF_REG_0);
6713 t = btf_type_skip_modifiers(meta.ret_btf, meta.ret_btf_id, NULL);
6714 if (!btf_type_is_struct(t)) {
6716 const struct btf_type *ret;
6719 /* resolve the type size of ksym. */
6720 ret = btf_resolve_size(meta.ret_btf, t, &tsize);
6722 tname = btf_name_by_offset(meta.ret_btf, t->name_off);
6723 verbose(env, "unable to resolve the size of type '%s': %ld\n",
6724 tname, PTR_ERR(ret));
6727 regs[BPF_REG_0].type = PTR_TO_MEM | ret_flag;
6728 regs[BPF_REG_0].mem_size = tsize;
6730 /* MEM_RDONLY may be carried from ret_flag, but it
6731 * doesn't apply on PTR_TO_BTF_ID. Fold it, otherwise
6732 * it will confuse the check of PTR_TO_BTF_ID in
6733 * check_mem_access().
6735 ret_flag &= ~MEM_RDONLY;
6737 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
6738 regs[BPF_REG_0].btf = meta.ret_btf;
6739 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
6741 } else if (base_type(ret_type) == RET_PTR_TO_BTF_ID) {
6744 mark_reg_known_zero(env, regs, BPF_REG_0);
6745 regs[BPF_REG_0].type = PTR_TO_BTF_ID | ret_flag;
6746 ret_btf_id = *fn->ret_btf_id;
6747 if (ret_btf_id == 0) {
6748 verbose(env, "invalid return type %u of func %s#%d\n",
6749 base_type(ret_type), func_id_name(func_id),
6753 /* current BPF helper definitions are only coming from
6754 * built-in code with type IDs from vmlinux BTF
6756 regs[BPF_REG_0].btf = btf_vmlinux;
6757 regs[BPF_REG_0].btf_id = ret_btf_id;
6759 verbose(env, "unknown return type %u of func %s#%d\n",
6760 base_type(ret_type), func_id_name(func_id), func_id);
6764 if (type_may_be_null(regs[BPF_REG_0].type))
6765 regs[BPF_REG_0].id = ++env->id_gen;
6767 if (is_ptr_cast_function(func_id)) {
6768 /* For release_reference() */
6769 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
6770 } else if (is_acquire_function(func_id, meta.map_ptr)) {
6771 int id = acquire_reference_state(env, insn_idx);
6775 /* For mark_ptr_or_null_reg() */
6776 regs[BPF_REG_0].id = id;
6777 /* For release_reference() */
6778 regs[BPF_REG_0].ref_obj_id = id;
6781 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
6783 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
6787 if ((func_id == BPF_FUNC_get_stack ||
6788 func_id == BPF_FUNC_get_task_stack) &&
6789 !env->prog->has_callchain_buf) {
6790 const char *err_str;
6792 #ifdef CONFIG_PERF_EVENTS
6793 err = get_callchain_buffers(sysctl_perf_event_max_stack);
6794 err_str = "cannot get callchain buffer for func %s#%d\n";
6797 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
6800 verbose(env, err_str, func_id_name(func_id), func_id);
6804 env->prog->has_callchain_buf = true;
6807 if (func_id == BPF_FUNC_get_stackid || func_id == BPF_FUNC_get_stack)
6808 env->prog->call_get_stack = true;
6810 if (func_id == BPF_FUNC_get_func_ip) {
6811 if (check_get_func_ip(env))
6813 env->prog->call_get_func_ip = true;
6817 clear_all_pkt_pointers(env);
6821 /* mark_btf_func_reg_size() is used when the reg size is determined by
6822 * the BTF func_proto's return value size and argument.
6824 static void mark_btf_func_reg_size(struct bpf_verifier_env *env, u32 regno,
6827 struct bpf_reg_state *reg = &cur_regs(env)[regno];
6829 if (regno == BPF_REG_0) {
6830 /* Function return value */
6831 reg->live |= REG_LIVE_WRITTEN;
6832 reg->subreg_def = reg_size == sizeof(u64) ?
6833 DEF_NOT_SUBREG : env->insn_idx + 1;
6835 /* Function argument */
6836 if (reg_size == sizeof(u64)) {
6837 mark_insn_zext(env, reg);
6838 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
6840 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ32);
6845 static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn)
6847 const struct btf_type *t, *func, *func_proto, *ptr_type;
6848 struct bpf_reg_state *regs = cur_regs(env);
6849 const char *func_name, *ptr_type_name;
6850 u32 i, nargs, func_id, ptr_type_id;
6851 struct module *btf_mod = NULL;
6852 const struct btf_param *args;
6853 struct btf *desc_btf;
6856 /* skip for now, but return error when we find this in fixup_kfunc_call */
6860 desc_btf = find_kfunc_desc_btf(env, insn->imm, insn->off, &btf_mod);
6861 if (IS_ERR(desc_btf))
6862 return PTR_ERR(desc_btf);
6864 func_id = insn->imm;
6865 func = btf_type_by_id(desc_btf, func_id);
6866 func_name = btf_name_by_offset(desc_btf, func->name_off);
6867 func_proto = btf_type_by_id(desc_btf, func->type);
6869 if (!env->ops->check_kfunc_call ||
6870 !env->ops->check_kfunc_call(func_id, btf_mod)) {
6871 verbose(env, "calling kernel function %s is not allowed\n",
6876 /* Check the arguments */
6877 err = btf_check_kfunc_arg_match(env, desc_btf, func_id, regs);
6881 for (i = 0; i < CALLER_SAVED_REGS; i++)
6882 mark_reg_not_init(env, regs, caller_saved[i]);
6884 /* Check return type */
6885 t = btf_type_skip_modifiers(desc_btf, func_proto->type, NULL);
6886 if (btf_type_is_scalar(t)) {
6887 mark_reg_unknown(env, regs, BPF_REG_0);
6888 mark_btf_func_reg_size(env, BPF_REG_0, t->size);
6889 } else if (btf_type_is_ptr(t)) {
6890 ptr_type = btf_type_skip_modifiers(desc_btf, t->type,
6892 if (!btf_type_is_struct(ptr_type)) {
6893 ptr_type_name = btf_name_by_offset(desc_btf,
6894 ptr_type->name_off);
6895 verbose(env, "kernel function %s returns pointer type %s %s is not supported\n",
6896 func_name, btf_type_str(ptr_type),
6900 mark_reg_known_zero(env, regs, BPF_REG_0);
6901 regs[BPF_REG_0].btf = desc_btf;
6902 regs[BPF_REG_0].type = PTR_TO_BTF_ID;
6903 regs[BPF_REG_0].btf_id = ptr_type_id;
6904 mark_btf_func_reg_size(env, BPF_REG_0, sizeof(void *));
6905 } /* else { add_kfunc_call() ensures it is btf_type_is_void(t) } */
6907 nargs = btf_type_vlen(func_proto);
6908 args = (const struct btf_param *)(func_proto + 1);
6909 for (i = 0; i < nargs; i++) {
6912 t = btf_type_skip_modifiers(desc_btf, args[i].type, NULL);
6913 if (btf_type_is_ptr(t))
6914 mark_btf_func_reg_size(env, regno, sizeof(void *));
6916 /* scalar. ensured by btf_check_kfunc_arg_match() */
6917 mark_btf_func_reg_size(env, regno, t->size);
6923 static bool signed_add_overflows(s64 a, s64 b)
6925 /* Do the add in u64, where overflow is well-defined */
6926 s64 res = (s64)((u64)a + (u64)b);
6933 static bool signed_add32_overflows(s32 a, s32 b)
6935 /* Do the add in u32, where overflow is well-defined */
6936 s32 res = (s32)((u32)a + (u32)b);
6943 static bool signed_sub_overflows(s64 a, s64 b)
6945 /* Do the sub in u64, where overflow is well-defined */
6946 s64 res = (s64)((u64)a - (u64)b);
6953 static bool signed_sub32_overflows(s32 a, s32 b)
6955 /* Do the sub in u32, where overflow is well-defined */
6956 s32 res = (s32)((u32)a - (u32)b);
6963 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
6964 const struct bpf_reg_state *reg,
6965 enum bpf_reg_type type)
6967 bool known = tnum_is_const(reg->var_off);
6968 s64 val = reg->var_off.value;
6969 s64 smin = reg->smin_value;
6971 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
6972 verbose(env, "math between %s pointer and %lld is not allowed\n",
6973 reg_type_str(env, type), val);
6977 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
6978 verbose(env, "%s pointer offset %d is not allowed\n",
6979 reg_type_str(env, type), reg->off);
6983 if (smin == S64_MIN) {
6984 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
6985 reg_type_str(env, type));
6989 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
6990 verbose(env, "value %lld makes %s pointer be out of bounds\n",
6991 smin, reg_type_str(env, type));
6998 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
7000 return &env->insn_aux_data[env->insn_idx];
7011 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
7012 u32 *alu_limit, bool mask_to_left)
7014 u32 max = 0, ptr_limit = 0;
7016 switch (ptr_reg->type) {
7018 /* Offset 0 is out-of-bounds, but acceptable start for the
7019 * left direction, see BPF_REG_FP. Also, unknown scalar
7020 * offset where we would need to deal with min/max bounds is
7021 * currently prohibited for unprivileged.
7023 max = MAX_BPF_STACK + mask_to_left;
7024 ptr_limit = -(ptr_reg->var_off.value + ptr_reg->off);
7026 case PTR_TO_MAP_VALUE:
7027 max = ptr_reg->map_ptr->value_size;
7028 ptr_limit = (mask_to_left ?
7029 ptr_reg->smin_value :
7030 ptr_reg->umax_value) + ptr_reg->off;
7036 if (ptr_limit >= max)
7037 return REASON_LIMIT;
7038 *alu_limit = ptr_limit;
7042 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
7043 const struct bpf_insn *insn)
7045 return env->bypass_spec_v1 || BPF_SRC(insn->code) == BPF_K;
7048 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
7049 u32 alu_state, u32 alu_limit)
7051 /* If we arrived here from different branches with different
7052 * state or limits to sanitize, then this won't work.
7054 if (aux->alu_state &&
7055 (aux->alu_state != alu_state ||
7056 aux->alu_limit != alu_limit))
7057 return REASON_PATHS;
7059 /* Corresponding fixup done in do_misc_fixups(). */
7060 aux->alu_state = alu_state;
7061 aux->alu_limit = alu_limit;
7065 static int sanitize_val_alu(struct bpf_verifier_env *env,
7066 struct bpf_insn *insn)
7068 struct bpf_insn_aux_data *aux = cur_aux(env);
7070 if (can_skip_alu_sanitation(env, insn))
7073 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
7076 static bool sanitize_needed(u8 opcode)
7078 return opcode == BPF_ADD || opcode == BPF_SUB;
7081 struct bpf_sanitize_info {
7082 struct bpf_insn_aux_data aux;
7086 static struct bpf_verifier_state *
7087 sanitize_speculative_path(struct bpf_verifier_env *env,
7088 const struct bpf_insn *insn,
7089 u32 next_idx, u32 curr_idx)
7091 struct bpf_verifier_state *branch;
7092 struct bpf_reg_state *regs;
7094 branch = push_stack(env, next_idx, curr_idx, true);
7095 if (branch && insn) {
7096 regs = branch->frame[branch->curframe]->regs;
7097 if (BPF_SRC(insn->code) == BPF_K) {
7098 mark_reg_unknown(env, regs, insn->dst_reg);
7099 } else if (BPF_SRC(insn->code) == BPF_X) {
7100 mark_reg_unknown(env, regs, insn->dst_reg);
7101 mark_reg_unknown(env, regs, insn->src_reg);
7107 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
7108 struct bpf_insn *insn,
7109 const struct bpf_reg_state *ptr_reg,
7110 const struct bpf_reg_state *off_reg,
7111 struct bpf_reg_state *dst_reg,
7112 struct bpf_sanitize_info *info,
7113 const bool commit_window)
7115 struct bpf_insn_aux_data *aux = commit_window ? cur_aux(env) : &info->aux;
7116 struct bpf_verifier_state *vstate = env->cur_state;
7117 bool off_is_imm = tnum_is_const(off_reg->var_off);
7118 bool off_is_neg = off_reg->smin_value < 0;
7119 bool ptr_is_dst_reg = ptr_reg == dst_reg;
7120 u8 opcode = BPF_OP(insn->code);
7121 u32 alu_state, alu_limit;
7122 struct bpf_reg_state tmp;
7126 if (can_skip_alu_sanitation(env, insn))
7129 /* We already marked aux for masking from non-speculative
7130 * paths, thus we got here in the first place. We only care
7131 * to explore bad access from here.
7133 if (vstate->speculative)
7136 if (!commit_window) {
7137 if (!tnum_is_const(off_reg->var_off) &&
7138 (off_reg->smin_value < 0) != (off_reg->smax_value < 0))
7139 return REASON_BOUNDS;
7141 info->mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
7142 (opcode == BPF_SUB && !off_is_neg);
7145 err = retrieve_ptr_limit(ptr_reg, &alu_limit, info->mask_to_left);
7149 if (commit_window) {
7150 /* In commit phase we narrow the masking window based on
7151 * the observed pointer move after the simulated operation.
7153 alu_state = info->aux.alu_state;
7154 alu_limit = abs(info->aux.alu_limit - alu_limit);
7156 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
7157 alu_state |= off_is_imm ? BPF_ALU_IMMEDIATE : 0;
7158 alu_state |= ptr_is_dst_reg ?
7159 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
7161 /* Limit pruning on unknown scalars to enable deep search for
7162 * potential masking differences from other program paths.
7165 env->explore_alu_limits = true;
7168 err = update_alu_sanitation_state(aux, alu_state, alu_limit);
7172 /* If we're in commit phase, we're done here given we already
7173 * pushed the truncated dst_reg into the speculative verification
7176 * Also, when register is a known constant, we rewrite register-based
7177 * operation to immediate-based, and thus do not need masking (and as
7178 * a consequence, do not need to simulate the zero-truncation either).
7180 if (commit_window || off_is_imm)
7183 /* Simulate and find potential out-of-bounds access under
7184 * speculative execution from truncation as a result of
7185 * masking when off was not within expected range. If off
7186 * sits in dst, then we temporarily need to move ptr there
7187 * to simulate dst (== 0) +/-= ptr. Needed, for example,
7188 * for cases where we use K-based arithmetic in one direction
7189 * and truncated reg-based in the other in order to explore
7192 if (!ptr_is_dst_reg) {
7194 *dst_reg = *ptr_reg;
7196 ret = sanitize_speculative_path(env, NULL, env->insn_idx + 1,
7198 if (!ptr_is_dst_reg && ret)
7200 return !ret ? REASON_STACK : 0;
7203 static void sanitize_mark_insn_seen(struct bpf_verifier_env *env)
7205 struct bpf_verifier_state *vstate = env->cur_state;
7207 /* If we simulate paths under speculation, we don't update the
7208 * insn as 'seen' such that when we verify unreachable paths in
7209 * the non-speculative domain, sanitize_dead_code() can still
7210 * rewrite/sanitize them.
7212 if (!vstate->speculative)
7213 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
7216 static int sanitize_err(struct bpf_verifier_env *env,
7217 const struct bpf_insn *insn, int reason,
7218 const struct bpf_reg_state *off_reg,
7219 const struct bpf_reg_state *dst_reg)
7221 static const char *err = "pointer arithmetic with it prohibited for !root";
7222 const char *op = BPF_OP(insn->code) == BPF_ADD ? "add" : "sub";
7223 u32 dst = insn->dst_reg, src = insn->src_reg;
7227 verbose(env, "R%d has unknown scalar with mixed signed bounds, %s\n",
7228 off_reg == dst_reg ? dst : src, err);
7231 verbose(env, "R%d has pointer with unsupported alu operation, %s\n",
7232 off_reg == dst_reg ? src : dst, err);
7235 verbose(env, "R%d tried to %s from different maps, paths or scalars, %s\n",
7239 verbose(env, "R%d tried to %s beyond pointer bounds, %s\n",
7243 verbose(env, "R%d could not be pushed for speculative verification, %s\n",
7247 verbose(env, "verifier internal error: unknown reason (%d)\n",
7255 /* check that stack access falls within stack limits and that 'reg' doesn't
7256 * have a variable offset.
7258 * Variable offset is prohibited for unprivileged mode for simplicity since it
7259 * requires corresponding support in Spectre masking for stack ALU. See also
7260 * retrieve_ptr_limit().
7263 * 'off' includes 'reg->off'.
7265 static int check_stack_access_for_ptr_arithmetic(
7266 struct bpf_verifier_env *env,
7268 const struct bpf_reg_state *reg,
7271 if (!tnum_is_const(reg->var_off)) {
7274 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
7275 verbose(env, "R%d variable stack access prohibited for !root, var_off=%s off=%d\n",
7276 regno, tn_buf, off);
7280 if (off >= 0 || off < -MAX_BPF_STACK) {
7281 verbose(env, "R%d stack pointer arithmetic goes out of range, "
7282 "prohibited for !root; off=%d\n", regno, off);
7289 static int sanitize_check_bounds(struct bpf_verifier_env *env,
7290 const struct bpf_insn *insn,
7291 const struct bpf_reg_state *dst_reg)
7293 u32 dst = insn->dst_reg;
7295 /* For unprivileged we require that resulting offset must be in bounds
7296 * in order to be able to sanitize access later on.
7298 if (env->bypass_spec_v1)
7301 switch (dst_reg->type) {
7303 if (check_stack_access_for_ptr_arithmetic(env, dst, dst_reg,
7304 dst_reg->off + dst_reg->var_off.value))
7307 case PTR_TO_MAP_VALUE:
7308 if (check_map_access(env, dst, dst_reg->off, 1, false)) {
7309 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
7310 "prohibited for !root\n", dst);
7321 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
7322 * Caller should also handle BPF_MOV case separately.
7323 * If we return -EACCES, caller may want to try again treating pointer as a
7324 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
7326 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
7327 struct bpf_insn *insn,
7328 const struct bpf_reg_state *ptr_reg,
7329 const struct bpf_reg_state *off_reg)
7331 struct bpf_verifier_state *vstate = env->cur_state;
7332 struct bpf_func_state *state = vstate->frame[vstate->curframe];
7333 struct bpf_reg_state *regs = state->regs, *dst_reg;
7334 bool known = tnum_is_const(off_reg->var_off);
7335 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
7336 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
7337 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
7338 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
7339 struct bpf_sanitize_info info = {};
7340 u8 opcode = BPF_OP(insn->code);
7341 u32 dst = insn->dst_reg;
7344 dst_reg = ®s[dst];
7346 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
7347 smin_val > smax_val || umin_val > umax_val) {
7348 /* Taint dst register if offset had invalid bounds derived from
7349 * e.g. dead branches.
7351 __mark_reg_unknown(env, dst_reg);
7355 if (BPF_CLASS(insn->code) != BPF_ALU64) {
7356 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
7357 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
7358 __mark_reg_unknown(env, dst_reg);
7363 "R%d 32-bit pointer arithmetic prohibited\n",
7368 if (ptr_reg->type & PTR_MAYBE_NULL) {
7369 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
7370 dst, reg_type_str(env, ptr_reg->type));
7374 switch (base_type(ptr_reg->type)) {
7375 case CONST_PTR_TO_MAP:
7376 /* smin_val represents the known value */
7377 if (known && smin_val == 0 && opcode == BPF_ADD)
7380 case PTR_TO_PACKET_END:
7382 case PTR_TO_SOCK_COMMON:
7383 case PTR_TO_TCP_SOCK:
7384 case PTR_TO_XDP_SOCK:
7385 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
7386 dst, reg_type_str(env, ptr_reg->type));
7392 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
7393 * The id may be overwritten later if we create a new variable offset.
7395 dst_reg->type = ptr_reg->type;
7396 dst_reg->id = ptr_reg->id;
7398 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
7399 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
7402 /* pointer types do not carry 32-bit bounds at the moment. */
7403 __mark_reg32_unbounded(dst_reg);
7405 if (sanitize_needed(opcode)) {
7406 ret = sanitize_ptr_alu(env, insn, ptr_reg, off_reg, dst_reg,
7409 return sanitize_err(env, insn, ret, off_reg, dst_reg);
7414 /* We can take a fixed offset as long as it doesn't overflow
7415 * the s32 'off' field
7417 if (known && (ptr_reg->off + smin_val ==
7418 (s64)(s32)(ptr_reg->off + smin_val))) {
7419 /* pointer += K. Accumulate it into fixed offset */
7420 dst_reg->smin_value = smin_ptr;
7421 dst_reg->smax_value = smax_ptr;
7422 dst_reg->umin_value = umin_ptr;
7423 dst_reg->umax_value = umax_ptr;
7424 dst_reg->var_off = ptr_reg->var_off;
7425 dst_reg->off = ptr_reg->off + smin_val;
7426 dst_reg->raw = ptr_reg->raw;
7429 /* A new variable offset is created. Note that off_reg->off
7430 * == 0, since it's a scalar.
7431 * dst_reg gets the pointer type and since some positive
7432 * integer value was added to the pointer, give it a new 'id'
7433 * if it's a PTR_TO_PACKET.
7434 * this creates a new 'base' pointer, off_reg (variable) gets
7435 * added into the variable offset, and we copy the fixed offset
7438 if (signed_add_overflows(smin_ptr, smin_val) ||
7439 signed_add_overflows(smax_ptr, smax_val)) {
7440 dst_reg->smin_value = S64_MIN;
7441 dst_reg->smax_value = S64_MAX;
7443 dst_reg->smin_value = smin_ptr + smin_val;
7444 dst_reg->smax_value = smax_ptr + smax_val;
7446 if (umin_ptr + umin_val < umin_ptr ||
7447 umax_ptr + umax_val < umax_ptr) {
7448 dst_reg->umin_value = 0;
7449 dst_reg->umax_value = U64_MAX;
7451 dst_reg->umin_value = umin_ptr + umin_val;
7452 dst_reg->umax_value = umax_ptr + umax_val;
7454 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
7455 dst_reg->off = ptr_reg->off;
7456 dst_reg->raw = ptr_reg->raw;
7457 if (reg_is_pkt_pointer(ptr_reg)) {
7458 dst_reg->id = ++env->id_gen;
7459 /* something was added to pkt_ptr, set range to zero */
7460 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
7464 if (dst_reg == off_reg) {
7465 /* scalar -= pointer. Creates an unknown scalar */
7466 verbose(env, "R%d tried to subtract pointer from scalar\n",
7470 /* We don't allow subtraction from FP, because (according to
7471 * test_verifier.c test "invalid fp arithmetic", JITs might not
7472 * be able to deal with it.
7474 if (ptr_reg->type == PTR_TO_STACK) {
7475 verbose(env, "R%d subtraction from stack pointer prohibited\n",
7479 if (known && (ptr_reg->off - smin_val ==
7480 (s64)(s32)(ptr_reg->off - smin_val))) {
7481 /* pointer -= K. Subtract it from fixed offset */
7482 dst_reg->smin_value = smin_ptr;
7483 dst_reg->smax_value = smax_ptr;
7484 dst_reg->umin_value = umin_ptr;
7485 dst_reg->umax_value = umax_ptr;
7486 dst_reg->var_off = ptr_reg->var_off;
7487 dst_reg->id = ptr_reg->id;
7488 dst_reg->off = ptr_reg->off - smin_val;
7489 dst_reg->raw = ptr_reg->raw;
7492 /* A new variable offset is created. If the subtrahend is known
7493 * nonnegative, then any reg->range we had before is still good.
7495 if (signed_sub_overflows(smin_ptr, smax_val) ||
7496 signed_sub_overflows(smax_ptr, smin_val)) {
7497 /* Overflow possible, we know nothing */
7498 dst_reg->smin_value = S64_MIN;
7499 dst_reg->smax_value = S64_MAX;
7501 dst_reg->smin_value = smin_ptr - smax_val;
7502 dst_reg->smax_value = smax_ptr - smin_val;
7504 if (umin_ptr < umax_val) {
7505 /* Overflow possible, we know nothing */
7506 dst_reg->umin_value = 0;
7507 dst_reg->umax_value = U64_MAX;
7509 /* Cannot overflow (as long as bounds are consistent) */
7510 dst_reg->umin_value = umin_ptr - umax_val;
7511 dst_reg->umax_value = umax_ptr - umin_val;
7513 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
7514 dst_reg->off = ptr_reg->off;
7515 dst_reg->raw = ptr_reg->raw;
7516 if (reg_is_pkt_pointer(ptr_reg)) {
7517 dst_reg->id = ++env->id_gen;
7518 /* something was added to pkt_ptr, set range to zero */
7520 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
7526 /* bitwise ops on pointers are troublesome, prohibit. */
7527 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
7528 dst, bpf_alu_string[opcode >> 4]);
7531 /* other operators (e.g. MUL,LSH) produce non-pointer results */
7532 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
7533 dst, bpf_alu_string[opcode >> 4]);
7537 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
7540 __update_reg_bounds(dst_reg);
7541 __reg_deduce_bounds(dst_reg);
7542 __reg_bound_offset(dst_reg);
7544 if (sanitize_check_bounds(env, insn, dst_reg) < 0)
7546 if (sanitize_needed(opcode)) {
7547 ret = sanitize_ptr_alu(env, insn, dst_reg, off_reg, dst_reg,
7550 return sanitize_err(env, insn, ret, off_reg, dst_reg);
7556 static void scalar32_min_max_add(struct bpf_reg_state *dst_reg,
7557 struct bpf_reg_state *src_reg)
7559 s32 smin_val = src_reg->s32_min_value;
7560 s32 smax_val = src_reg->s32_max_value;
7561 u32 umin_val = src_reg->u32_min_value;
7562 u32 umax_val = src_reg->u32_max_value;
7564 if (signed_add32_overflows(dst_reg->s32_min_value, smin_val) ||
7565 signed_add32_overflows(dst_reg->s32_max_value, smax_val)) {
7566 dst_reg->s32_min_value = S32_MIN;
7567 dst_reg->s32_max_value = S32_MAX;
7569 dst_reg->s32_min_value += smin_val;
7570 dst_reg->s32_max_value += smax_val;
7572 if (dst_reg->u32_min_value + umin_val < umin_val ||
7573 dst_reg->u32_max_value + umax_val < umax_val) {
7574 dst_reg->u32_min_value = 0;
7575 dst_reg->u32_max_value = U32_MAX;
7577 dst_reg->u32_min_value += umin_val;
7578 dst_reg->u32_max_value += umax_val;
7582 static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
7583 struct bpf_reg_state *src_reg)
7585 s64 smin_val = src_reg->smin_value;
7586 s64 smax_val = src_reg->smax_value;
7587 u64 umin_val = src_reg->umin_value;
7588 u64 umax_val = src_reg->umax_value;
7590 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
7591 signed_add_overflows(dst_reg->smax_value, smax_val)) {
7592 dst_reg->smin_value = S64_MIN;
7593 dst_reg->smax_value = S64_MAX;
7595 dst_reg->smin_value += smin_val;
7596 dst_reg->smax_value += smax_val;
7598 if (dst_reg->umin_value + umin_val < umin_val ||
7599 dst_reg->umax_value + umax_val < umax_val) {
7600 dst_reg->umin_value = 0;
7601 dst_reg->umax_value = U64_MAX;
7603 dst_reg->umin_value += umin_val;
7604 dst_reg->umax_value += umax_val;
7608 static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg,
7609 struct bpf_reg_state *src_reg)
7611 s32 smin_val = src_reg->s32_min_value;
7612 s32 smax_val = src_reg->s32_max_value;
7613 u32 umin_val = src_reg->u32_min_value;
7614 u32 umax_val = src_reg->u32_max_value;
7616 if (signed_sub32_overflows(dst_reg->s32_min_value, smax_val) ||
7617 signed_sub32_overflows(dst_reg->s32_max_value, smin_val)) {
7618 /* Overflow possible, we know nothing */
7619 dst_reg->s32_min_value = S32_MIN;
7620 dst_reg->s32_max_value = S32_MAX;
7622 dst_reg->s32_min_value -= smax_val;
7623 dst_reg->s32_max_value -= smin_val;
7625 if (dst_reg->u32_min_value < umax_val) {
7626 /* Overflow possible, we know nothing */
7627 dst_reg->u32_min_value = 0;
7628 dst_reg->u32_max_value = U32_MAX;
7630 /* Cannot overflow (as long as bounds are consistent) */
7631 dst_reg->u32_min_value -= umax_val;
7632 dst_reg->u32_max_value -= umin_val;
7636 static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
7637 struct bpf_reg_state *src_reg)
7639 s64 smin_val = src_reg->smin_value;
7640 s64 smax_val = src_reg->smax_value;
7641 u64 umin_val = src_reg->umin_value;
7642 u64 umax_val = src_reg->umax_value;
7644 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
7645 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
7646 /* Overflow possible, we know nothing */
7647 dst_reg->smin_value = S64_MIN;
7648 dst_reg->smax_value = S64_MAX;
7650 dst_reg->smin_value -= smax_val;
7651 dst_reg->smax_value -= smin_val;
7653 if (dst_reg->umin_value < umax_val) {
7654 /* Overflow possible, we know nothing */
7655 dst_reg->umin_value = 0;
7656 dst_reg->umax_value = U64_MAX;
7658 /* Cannot overflow (as long as bounds are consistent) */
7659 dst_reg->umin_value -= umax_val;
7660 dst_reg->umax_value -= umin_val;
7664 static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg,
7665 struct bpf_reg_state *src_reg)
7667 s32 smin_val = src_reg->s32_min_value;
7668 u32 umin_val = src_reg->u32_min_value;
7669 u32 umax_val = src_reg->u32_max_value;
7671 if (smin_val < 0 || dst_reg->s32_min_value < 0) {
7672 /* Ain't nobody got time to multiply that sign */
7673 __mark_reg32_unbounded(dst_reg);
7676 /* Both values are positive, so we can work with unsigned and
7677 * copy the result to signed (unless it exceeds S32_MAX).
7679 if (umax_val > U16_MAX || dst_reg->u32_max_value > U16_MAX) {
7680 /* Potential overflow, we know nothing */
7681 __mark_reg32_unbounded(dst_reg);
7684 dst_reg->u32_min_value *= umin_val;
7685 dst_reg->u32_max_value *= umax_val;
7686 if (dst_reg->u32_max_value > S32_MAX) {
7687 /* Overflow possible, we know nothing */
7688 dst_reg->s32_min_value = S32_MIN;
7689 dst_reg->s32_max_value = S32_MAX;
7691 dst_reg->s32_min_value = dst_reg->u32_min_value;
7692 dst_reg->s32_max_value = dst_reg->u32_max_value;
7696 static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
7697 struct bpf_reg_state *src_reg)
7699 s64 smin_val = src_reg->smin_value;
7700 u64 umin_val = src_reg->umin_value;
7701 u64 umax_val = src_reg->umax_value;
7703 if (smin_val < 0 || dst_reg->smin_value < 0) {
7704 /* Ain't nobody got time to multiply that sign */
7705 __mark_reg64_unbounded(dst_reg);
7708 /* Both values are positive, so we can work with unsigned and
7709 * copy the result to signed (unless it exceeds S64_MAX).
7711 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
7712 /* Potential overflow, we know nothing */
7713 __mark_reg64_unbounded(dst_reg);
7716 dst_reg->umin_value *= umin_val;
7717 dst_reg->umax_value *= umax_val;
7718 if (dst_reg->umax_value > S64_MAX) {
7719 /* Overflow possible, we know nothing */
7720 dst_reg->smin_value = S64_MIN;
7721 dst_reg->smax_value = S64_MAX;
7723 dst_reg->smin_value = dst_reg->umin_value;
7724 dst_reg->smax_value = dst_reg->umax_value;
7728 static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
7729 struct bpf_reg_state *src_reg)
7731 bool src_known = tnum_subreg_is_const(src_reg->var_off);
7732 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
7733 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
7734 s32 smin_val = src_reg->s32_min_value;
7735 u32 umax_val = src_reg->u32_max_value;
7737 if (src_known && dst_known) {
7738 __mark_reg32_known(dst_reg, var32_off.value);
7742 /* We get our minimum from the var_off, since that's inherently
7743 * bitwise. Our maximum is the minimum of the operands' maxima.
7745 dst_reg->u32_min_value = var32_off.value;
7746 dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val);
7747 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
7748 /* Lose signed bounds when ANDing negative numbers,
7749 * ain't nobody got time for that.
7751 dst_reg->s32_min_value = S32_MIN;
7752 dst_reg->s32_max_value = S32_MAX;
7754 /* ANDing two positives gives a positive, so safe to
7755 * cast result into s64.
7757 dst_reg->s32_min_value = dst_reg->u32_min_value;
7758 dst_reg->s32_max_value = dst_reg->u32_max_value;
7762 static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
7763 struct bpf_reg_state *src_reg)
7765 bool src_known = tnum_is_const(src_reg->var_off);
7766 bool dst_known = tnum_is_const(dst_reg->var_off);
7767 s64 smin_val = src_reg->smin_value;
7768 u64 umax_val = src_reg->umax_value;
7770 if (src_known && dst_known) {
7771 __mark_reg_known(dst_reg, dst_reg->var_off.value);
7775 /* We get our minimum from the var_off, since that's inherently
7776 * bitwise. Our maximum is the minimum of the operands' maxima.
7778 dst_reg->umin_value = dst_reg->var_off.value;
7779 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
7780 if (dst_reg->smin_value < 0 || smin_val < 0) {
7781 /* Lose signed bounds when ANDing negative numbers,
7782 * ain't nobody got time for that.
7784 dst_reg->smin_value = S64_MIN;
7785 dst_reg->smax_value = S64_MAX;
7787 /* ANDing two positives gives a positive, so safe to
7788 * cast result into s64.
7790 dst_reg->smin_value = dst_reg->umin_value;
7791 dst_reg->smax_value = dst_reg->umax_value;
7793 /* We may learn something more from the var_off */
7794 __update_reg_bounds(dst_reg);
7797 static void scalar32_min_max_or(struct bpf_reg_state *dst_reg,
7798 struct bpf_reg_state *src_reg)
7800 bool src_known = tnum_subreg_is_const(src_reg->var_off);
7801 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
7802 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
7803 s32 smin_val = src_reg->s32_min_value;
7804 u32 umin_val = src_reg->u32_min_value;
7806 if (src_known && dst_known) {
7807 __mark_reg32_known(dst_reg, var32_off.value);
7811 /* We get our maximum from the var_off, and our minimum is the
7812 * maximum of the operands' minima
7814 dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val);
7815 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
7816 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
7817 /* Lose signed bounds when ORing negative numbers,
7818 * ain't nobody got time for that.
7820 dst_reg->s32_min_value = S32_MIN;
7821 dst_reg->s32_max_value = S32_MAX;
7823 /* ORing two positives gives a positive, so safe to
7824 * cast result into s64.
7826 dst_reg->s32_min_value = dst_reg->u32_min_value;
7827 dst_reg->s32_max_value = dst_reg->u32_max_value;
7831 static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
7832 struct bpf_reg_state *src_reg)
7834 bool src_known = tnum_is_const(src_reg->var_off);
7835 bool dst_known = tnum_is_const(dst_reg->var_off);
7836 s64 smin_val = src_reg->smin_value;
7837 u64 umin_val = src_reg->umin_value;
7839 if (src_known && dst_known) {
7840 __mark_reg_known(dst_reg, dst_reg->var_off.value);
7844 /* We get our maximum from the var_off, and our minimum is the
7845 * maximum of the operands' minima
7847 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
7848 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
7849 if (dst_reg->smin_value < 0 || smin_val < 0) {
7850 /* Lose signed bounds when ORing negative numbers,
7851 * ain't nobody got time for that.
7853 dst_reg->smin_value = S64_MIN;
7854 dst_reg->smax_value = S64_MAX;
7856 /* ORing two positives gives a positive, so safe to
7857 * cast result into s64.
7859 dst_reg->smin_value = dst_reg->umin_value;
7860 dst_reg->smax_value = dst_reg->umax_value;
7862 /* We may learn something more from the var_off */
7863 __update_reg_bounds(dst_reg);
7866 static void scalar32_min_max_xor(struct bpf_reg_state *dst_reg,
7867 struct bpf_reg_state *src_reg)
7869 bool src_known = tnum_subreg_is_const(src_reg->var_off);
7870 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
7871 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
7872 s32 smin_val = src_reg->s32_min_value;
7874 if (src_known && dst_known) {
7875 __mark_reg32_known(dst_reg, var32_off.value);
7879 /* We get both minimum and maximum from the var32_off. */
7880 dst_reg->u32_min_value = var32_off.value;
7881 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
7883 if (dst_reg->s32_min_value >= 0 && smin_val >= 0) {
7884 /* XORing two positive sign numbers gives a positive,
7885 * so safe to cast u32 result into s32.
7887 dst_reg->s32_min_value = dst_reg->u32_min_value;
7888 dst_reg->s32_max_value = dst_reg->u32_max_value;
7890 dst_reg->s32_min_value = S32_MIN;
7891 dst_reg->s32_max_value = S32_MAX;
7895 static void scalar_min_max_xor(struct bpf_reg_state *dst_reg,
7896 struct bpf_reg_state *src_reg)
7898 bool src_known = tnum_is_const(src_reg->var_off);
7899 bool dst_known = tnum_is_const(dst_reg->var_off);
7900 s64 smin_val = src_reg->smin_value;
7902 if (src_known && dst_known) {
7903 /* dst_reg->var_off.value has been updated earlier */
7904 __mark_reg_known(dst_reg, dst_reg->var_off.value);
7908 /* We get both minimum and maximum from the var_off. */
7909 dst_reg->umin_value = dst_reg->var_off.value;
7910 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
7912 if (dst_reg->smin_value >= 0 && smin_val >= 0) {
7913 /* XORing two positive sign numbers gives a positive,
7914 * so safe to cast u64 result into s64.
7916 dst_reg->smin_value = dst_reg->umin_value;
7917 dst_reg->smax_value = dst_reg->umax_value;
7919 dst_reg->smin_value = S64_MIN;
7920 dst_reg->smax_value = S64_MAX;
7923 __update_reg_bounds(dst_reg);
7926 static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
7927 u64 umin_val, u64 umax_val)
7929 /* We lose all sign bit information (except what we can pick
7932 dst_reg->s32_min_value = S32_MIN;
7933 dst_reg->s32_max_value = S32_MAX;
7934 /* If we might shift our top bit out, then we know nothing */
7935 if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) {
7936 dst_reg->u32_min_value = 0;
7937 dst_reg->u32_max_value = U32_MAX;
7939 dst_reg->u32_min_value <<= umin_val;
7940 dst_reg->u32_max_value <<= umax_val;
7944 static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
7945 struct bpf_reg_state *src_reg)
7947 u32 umax_val = src_reg->u32_max_value;
7948 u32 umin_val = src_reg->u32_min_value;
7949 /* u32 alu operation will zext upper bits */
7950 struct tnum subreg = tnum_subreg(dst_reg->var_off);
7952 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
7953 dst_reg->var_off = tnum_subreg(tnum_lshift(subreg, umin_val));
7954 /* Not required but being careful mark reg64 bounds as unknown so
7955 * that we are forced to pick them up from tnum and zext later and
7956 * if some path skips this step we are still safe.
7958 __mark_reg64_unbounded(dst_reg);
7959 __update_reg32_bounds(dst_reg);
7962 static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg,
7963 u64 umin_val, u64 umax_val)
7965 /* Special case <<32 because it is a common compiler pattern to sign
7966 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
7967 * positive we know this shift will also be positive so we can track
7968 * bounds correctly. Otherwise we lose all sign bit information except
7969 * what we can pick up from var_off. Perhaps we can generalize this
7970 * later to shifts of any length.
7972 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0)
7973 dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32;
7975 dst_reg->smax_value = S64_MAX;
7977 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0)
7978 dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32;
7980 dst_reg->smin_value = S64_MIN;
7982 /* If we might shift our top bit out, then we know nothing */
7983 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
7984 dst_reg->umin_value = 0;
7985 dst_reg->umax_value = U64_MAX;
7987 dst_reg->umin_value <<= umin_val;
7988 dst_reg->umax_value <<= umax_val;
7992 static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
7993 struct bpf_reg_state *src_reg)
7995 u64 umax_val = src_reg->umax_value;
7996 u64 umin_val = src_reg->umin_value;
7998 /* scalar64 calc uses 32bit unshifted bounds so must be called first */
7999 __scalar64_min_max_lsh(dst_reg, umin_val, umax_val);
8000 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
8002 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
8003 /* We may learn something more from the var_off */
8004 __update_reg_bounds(dst_reg);
8007 static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg,
8008 struct bpf_reg_state *src_reg)
8010 struct tnum subreg = tnum_subreg(dst_reg->var_off);
8011 u32 umax_val = src_reg->u32_max_value;
8012 u32 umin_val = src_reg->u32_min_value;
8014 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
8015 * be negative, then either:
8016 * 1) src_reg might be zero, so the sign bit of the result is
8017 * unknown, so we lose our signed bounds
8018 * 2) it's known negative, thus the unsigned bounds capture the
8020 * 3) the signed bounds cross zero, so they tell us nothing
8022 * If the value in dst_reg is known nonnegative, then again the
8023 * unsigned bounds capture the signed bounds.
8024 * Thus, in all cases it suffices to blow away our signed bounds
8025 * and rely on inferring new ones from the unsigned bounds and
8026 * var_off of the result.
8028 dst_reg->s32_min_value = S32_MIN;
8029 dst_reg->s32_max_value = S32_MAX;
8031 dst_reg->var_off = tnum_rshift(subreg, umin_val);
8032 dst_reg->u32_min_value >>= umax_val;
8033 dst_reg->u32_max_value >>= umin_val;
8035 __mark_reg64_unbounded(dst_reg);
8036 __update_reg32_bounds(dst_reg);
8039 static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
8040 struct bpf_reg_state *src_reg)
8042 u64 umax_val = src_reg->umax_value;
8043 u64 umin_val = src_reg->umin_value;
8045 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
8046 * be negative, then either:
8047 * 1) src_reg might be zero, so the sign bit of the result is
8048 * unknown, so we lose our signed bounds
8049 * 2) it's known negative, thus the unsigned bounds capture the
8051 * 3) the signed bounds cross zero, so they tell us nothing
8053 * If the value in dst_reg is known nonnegative, then again the
8054 * unsigned bounds capture the signed bounds.
8055 * Thus, in all cases it suffices to blow away our signed bounds
8056 * and rely on inferring new ones from the unsigned bounds and
8057 * var_off of the result.
8059 dst_reg->smin_value = S64_MIN;
8060 dst_reg->smax_value = S64_MAX;
8061 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
8062 dst_reg->umin_value >>= umax_val;
8063 dst_reg->umax_value >>= umin_val;
8065 /* Its not easy to operate on alu32 bounds here because it depends
8066 * on bits being shifted in. Take easy way out and mark unbounded
8067 * so we can recalculate later from tnum.
8069 __mark_reg32_unbounded(dst_reg);
8070 __update_reg_bounds(dst_reg);
8073 static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg,
8074 struct bpf_reg_state *src_reg)
8076 u64 umin_val = src_reg->u32_min_value;
8078 /* Upon reaching here, src_known is true and
8079 * umax_val is equal to umin_val.
8081 dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val);
8082 dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val);
8084 dst_reg->var_off = tnum_arshift(tnum_subreg(dst_reg->var_off), umin_val, 32);
8086 /* blow away the dst_reg umin_value/umax_value and rely on
8087 * dst_reg var_off to refine the result.
8089 dst_reg->u32_min_value = 0;
8090 dst_reg->u32_max_value = U32_MAX;
8092 __mark_reg64_unbounded(dst_reg);
8093 __update_reg32_bounds(dst_reg);
8096 static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
8097 struct bpf_reg_state *src_reg)
8099 u64 umin_val = src_reg->umin_value;
8101 /* Upon reaching here, src_known is true and umax_val is equal
8104 dst_reg->smin_value >>= umin_val;
8105 dst_reg->smax_value >>= umin_val;
8107 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val, 64);
8109 /* blow away the dst_reg umin_value/umax_value and rely on
8110 * dst_reg var_off to refine the result.
8112 dst_reg->umin_value = 0;
8113 dst_reg->umax_value = U64_MAX;
8115 /* Its not easy to operate on alu32 bounds here because it depends
8116 * on bits being shifted in from upper 32-bits. Take easy way out
8117 * and mark unbounded so we can recalculate later from tnum.
8119 __mark_reg32_unbounded(dst_reg);
8120 __update_reg_bounds(dst_reg);
8123 /* WARNING: This function does calculations on 64-bit values, but the actual
8124 * execution may occur on 32-bit values. Therefore, things like bitshifts
8125 * need extra checks in the 32-bit case.
8127 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
8128 struct bpf_insn *insn,
8129 struct bpf_reg_state *dst_reg,
8130 struct bpf_reg_state src_reg)
8132 struct bpf_reg_state *regs = cur_regs(env);
8133 u8 opcode = BPF_OP(insn->code);
8135 s64 smin_val, smax_val;
8136 u64 umin_val, umax_val;
8137 s32 s32_min_val, s32_max_val;
8138 u32 u32_min_val, u32_max_val;
8139 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
8140 bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
8143 smin_val = src_reg.smin_value;
8144 smax_val = src_reg.smax_value;
8145 umin_val = src_reg.umin_value;
8146 umax_val = src_reg.umax_value;
8148 s32_min_val = src_reg.s32_min_value;
8149 s32_max_val = src_reg.s32_max_value;
8150 u32_min_val = src_reg.u32_min_value;
8151 u32_max_val = src_reg.u32_max_value;
8154 src_known = tnum_subreg_is_const(src_reg.var_off);
8156 (s32_min_val != s32_max_val || u32_min_val != u32_max_val)) ||
8157 s32_min_val > s32_max_val || u32_min_val > u32_max_val) {
8158 /* Taint dst register if offset had invalid bounds
8159 * derived from e.g. dead branches.
8161 __mark_reg_unknown(env, dst_reg);
8165 src_known = tnum_is_const(src_reg.var_off);
8167 (smin_val != smax_val || umin_val != umax_val)) ||
8168 smin_val > smax_val || umin_val > umax_val) {
8169 /* Taint dst register if offset had invalid bounds
8170 * derived from e.g. dead branches.
8172 __mark_reg_unknown(env, dst_reg);
8178 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
8179 __mark_reg_unknown(env, dst_reg);
8183 if (sanitize_needed(opcode)) {
8184 ret = sanitize_val_alu(env, insn);
8186 return sanitize_err(env, insn, ret, NULL, NULL);
8189 /* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
8190 * There are two classes of instructions: The first class we track both
8191 * alu32 and alu64 sign/unsigned bounds independently this provides the
8192 * greatest amount of precision when alu operations are mixed with jmp32
8193 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
8194 * and BPF_OR. This is possible because these ops have fairly easy to
8195 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
8196 * See alu32 verifier tests for examples. The second class of
8197 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
8198 * with regards to tracking sign/unsigned bounds because the bits may
8199 * cross subreg boundaries in the alu64 case. When this happens we mark
8200 * the reg unbounded in the subreg bound space and use the resulting
8201 * tnum to calculate an approximation of the sign/unsigned bounds.
8205 scalar32_min_max_add(dst_reg, &src_reg);
8206 scalar_min_max_add(dst_reg, &src_reg);
8207 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
8210 scalar32_min_max_sub(dst_reg, &src_reg);
8211 scalar_min_max_sub(dst_reg, &src_reg);
8212 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
8215 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
8216 scalar32_min_max_mul(dst_reg, &src_reg);
8217 scalar_min_max_mul(dst_reg, &src_reg);
8220 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
8221 scalar32_min_max_and(dst_reg, &src_reg);
8222 scalar_min_max_and(dst_reg, &src_reg);
8225 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
8226 scalar32_min_max_or(dst_reg, &src_reg);
8227 scalar_min_max_or(dst_reg, &src_reg);
8230 dst_reg->var_off = tnum_xor(dst_reg->var_off, src_reg.var_off);
8231 scalar32_min_max_xor(dst_reg, &src_reg);
8232 scalar_min_max_xor(dst_reg, &src_reg);
8235 if (umax_val >= insn_bitness) {
8236 /* Shifts greater than 31 or 63 are undefined.
8237 * This includes shifts by a negative number.
8239 mark_reg_unknown(env, regs, insn->dst_reg);
8243 scalar32_min_max_lsh(dst_reg, &src_reg);
8245 scalar_min_max_lsh(dst_reg, &src_reg);
8248 if (umax_val >= insn_bitness) {
8249 /* Shifts greater than 31 or 63 are undefined.
8250 * This includes shifts by a negative number.
8252 mark_reg_unknown(env, regs, insn->dst_reg);
8256 scalar32_min_max_rsh(dst_reg, &src_reg);
8258 scalar_min_max_rsh(dst_reg, &src_reg);
8261 if (umax_val >= insn_bitness) {
8262 /* Shifts greater than 31 or 63 are undefined.
8263 * This includes shifts by a negative number.
8265 mark_reg_unknown(env, regs, insn->dst_reg);
8269 scalar32_min_max_arsh(dst_reg, &src_reg);
8271 scalar_min_max_arsh(dst_reg, &src_reg);
8274 mark_reg_unknown(env, regs, insn->dst_reg);
8278 /* ALU32 ops are zero extended into 64bit register */
8280 zext_32_to_64(dst_reg);
8282 __update_reg_bounds(dst_reg);
8283 __reg_deduce_bounds(dst_reg);
8284 __reg_bound_offset(dst_reg);
8288 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
8291 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
8292 struct bpf_insn *insn)
8294 struct bpf_verifier_state *vstate = env->cur_state;
8295 struct bpf_func_state *state = vstate->frame[vstate->curframe];
8296 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
8297 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
8298 u8 opcode = BPF_OP(insn->code);
8301 dst_reg = ®s[insn->dst_reg];
8303 if (dst_reg->type != SCALAR_VALUE)
8306 /* Make sure ID is cleared otherwise dst_reg min/max could be
8307 * incorrectly propagated into other registers by find_equal_scalars()
8310 if (BPF_SRC(insn->code) == BPF_X) {
8311 src_reg = ®s[insn->src_reg];
8312 if (src_reg->type != SCALAR_VALUE) {
8313 if (dst_reg->type != SCALAR_VALUE) {
8314 /* Combining two pointers by any ALU op yields
8315 * an arbitrary scalar. Disallow all math except
8316 * pointer subtraction
8318 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
8319 mark_reg_unknown(env, regs, insn->dst_reg);
8322 verbose(env, "R%d pointer %s pointer prohibited\n",
8324 bpf_alu_string[opcode >> 4]);
8327 /* scalar += pointer
8328 * This is legal, but we have to reverse our
8329 * src/dest handling in computing the range
8331 err = mark_chain_precision(env, insn->dst_reg);
8334 return adjust_ptr_min_max_vals(env, insn,
8337 } else if (ptr_reg) {
8338 /* pointer += scalar */
8339 err = mark_chain_precision(env, insn->src_reg);
8342 return adjust_ptr_min_max_vals(env, insn,
8346 /* Pretend the src is a reg with a known value, since we only
8347 * need to be able to read from this state.
8349 off_reg.type = SCALAR_VALUE;
8350 __mark_reg_known(&off_reg, insn->imm);
8352 if (ptr_reg) /* pointer += K */
8353 return adjust_ptr_min_max_vals(env, insn,
8357 /* Got here implies adding two SCALAR_VALUEs */
8358 if (WARN_ON_ONCE(ptr_reg)) {
8359 print_verifier_state(env, state, true);
8360 verbose(env, "verifier internal error: unexpected ptr_reg\n");
8363 if (WARN_ON(!src_reg)) {
8364 print_verifier_state(env, state, true);
8365 verbose(env, "verifier internal error: no src_reg\n");
8368 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
8371 /* check validity of 32-bit and 64-bit arithmetic operations */
8372 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
8374 struct bpf_reg_state *regs = cur_regs(env);
8375 u8 opcode = BPF_OP(insn->code);
8378 if (opcode == BPF_END || opcode == BPF_NEG) {
8379 if (opcode == BPF_NEG) {
8380 if (BPF_SRC(insn->code) != 0 ||
8381 insn->src_reg != BPF_REG_0 ||
8382 insn->off != 0 || insn->imm != 0) {
8383 verbose(env, "BPF_NEG uses reserved fields\n");
8387 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
8388 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
8389 BPF_CLASS(insn->code) == BPF_ALU64) {
8390 verbose(env, "BPF_END uses reserved fields\n");
8395 /* check src operand */
8396 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
8400 if (is_pointer_value(env, insn->dst_reg)) {
8401 verbose(env, "R%d pointer arithmetic prohibited\n",
8406 /* check dest operand */
8407 err = check_reg_arg(env, insn->dst_reg, DST_OP);
8411 } else if (opcode == BPF_MOV) {
8413 if (BPF_SRC(insn->code) == BPF_X) {
8414 if (insn->imm != 0 || insn->off != 0) {
8415 verbose(env, "BPF_MOV uses reserved fields\n");
8419 /* check src operand */
8420 err = check_reg_arg(env, insn->src_reg, SRC_OP);
8424 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
8425 verbose(env, "BPF_MOV uses reserved fields\n");
8430 /* check dest operand, mark as required later */
8431 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
8435 if (BPF_SRC(insn->code) == BPF_X) {
8436 struct bpf_reg_state *src_reg = regs + insn->src_reg;
8437 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
8439 if (BPF_CLASS(insn->code) == BPF_ALU64) {
8441 * copy register state to dest reg
8443 if (src_reg->type == SCALAR_VALUE && !src_reg->id)
8444 /* Assign src and dst registers the same ID
8445 * that will be used by find_equal_scalars()
8446 * to propagate min/max range.
8448 src_reg->id = ++env->id_gen;
8449 *dst_reg = *src_reg;
8450 dst_reg->live |= REG_LIVE_WRITTEN;
8451 dst_reg->subreg_def = DEF_NOT_SUBREG;
8454 if (is_pointer_value(env, insn->src_reg)) {
8456 "R%d partial copy of pointer\n",
8459 } else if (src_reg->type == SCALAR_VALUE) {
8460 *dst_reg = *src_reg;
8461 /* Make sure ID is cleared otherwise
8462 * dst_reg min/max could be incorrectly
8463 * propagated into src_reg by find_equal_scalars()
8466 dst_reg->live |= REG_LIVE_WRITTEN;
8467 dst_reg->subreg_def = env->insn_idx + 1;
8469 mark_reg_unknown(env, regs,
8472 zext_32_to_64(dst_reg);
8474 __update_reg_bounds(dst_reg);
8475 __reg_deduce_bounds(dst_reg);
8476 __reg_bound_offset(dst_reg);
8480 * remember the value we stored into this reg
8482 /* clear any state __mark_reg_known doesn't set */
8483 mark_reg_unknown(env, regs, insn->dst_reg);
8484 regs[insn->dst_reg].type = SCALAR_VALUE;
8485 if (BPF_CLASS(insn->code) == BPF_ALU64) {
8486 __mark_reg_known(regs + insn->dst_reg,
8489 __mark_reg_known(regs + insn->dst_reg,
8494 } else if (opcode > BPF_END) {
8495 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
8498 } else { /* all other ALU ops: and, sub, xor, add, ... */
8500 if (BPF_SRC(insn->code) == BPF_X) {
8501 if (insn->imm != 0 || insn->off != 0) {
8502 verbose(env, "BPF_ALU uses reserved fields\n");
8505 /* check src1 operand */
8506 err = check_reg_arg(env, insn->src_reg, SRC_OP);
8510 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
8511 verbose(env, "BPF_ALU uses reserved fields\n");
8516 /* check src2 operand */
8517 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
8521 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
8522 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
8523 verbose(env, "div by zero\n");
8527 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
8528 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
8529 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
8531 if (insn->imm < 0 || insn->imm >= size) {
8532 verbose(env, "invalid shift %d\n", insn->imm);
8537 /* check dest operand */
8538 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
8542 return adjust_reg_min_max_vals(env, insn);
8548 static void __find_good_pkt_pointers(struct bpf_func_state *state,
8549 struct bpf_reg_state *dst_reg,
8550 enum bpf_reg_type type, int new_range)
8552 struct bpf_reg_state *reg;
8555 for (i = 0; i < MAX_BPF_REG; i++) {
8556 reg = &state->regs[i];
8557 if (reg->type == type && reg->id == dst_reg->id)
8558 /* keep the maximum range already checked */
8559 reg->range = max(reg->range, new_range);
8562 bpf_for_each_spilled_reg(i, state, reg) {
8565 if (reg->type == type && reg->id == dst_reg->id)
8566 reg->range = max(reg->range, new_range);
8570 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
8571 struct bpf_reg_state *dst_reg,
8572 enum bpf_reg_type type,
8573 bool range_right_open)
8577 if (dst_reg->off < 0 ||
8578 (dst_reg->off == 0 && range_right_open))
8579 /* This doesn't give us any range */
8582 if (dst_reg->umax_value > MAX_PACKET_OFF ||
8583 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
8584 /* Risk of overflow. For instance, ptr + (1<<63) may be less
8585 * than pkt_end, but that's because it's also less than pkt.
8589 new_range = dst_reg->off;
8590 if (range_right_open)
8593 /* Examples for register markings:
8595 * pkt_data in dst register:
8599 * if (r2 > pkt_end) goto <handle exception>
8604 * if (r2 < pkt_end) goto <access okay>
8605 * <handle exception>
8608 * r2 == dst_reg, pkt_end == src_reg
8609 * r2=pkt(id=n,off=8,r=0)
8610 * r3=pkt(id=n,off=0,r=0)
8612 * pkt_data in src register:
8616 * if (pkt_end >= r2) goto <access okay>
8617 * <handle exception>
8621 * if (pkt_end <= r2) goto <handle exception>
8625 * pkt_end == dst_reg, r2 == src_reg
8626 * r2=pkt(id=n,off=8,r=0)
8627 * r3=pkt(id=n,off=0,r=0)
8629 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
8630 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
8631 * and [r3, r3 + 8-1) respectively is safe to access depending on
8635 /* If our ids match, then we must have the same max_value. And we
8636 * don't care about the other reg's fixed offset, since if it's too big
8637 * the range won't allow anything.
8638 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
8640 for (i = 0; i <= vstate->curframe; i++)
8641 __find_good_pkt_pointers(vstate->frame[i], dst_reg, type,
8645 static int is_branch32_taken(struct bpf_reg_state *reg, u32 val, u8 opcode)
8647 struct tnum subreg = tnum_subreg(reg->var_off);
8648 s32 sval = (s32)val;
8652 if (tnum_is_const(subreg))
8653 return !!tnum_equals_const(subreg, val);
8656 if (tnum_is_const(subreg))
8657 return !tnum_equals_const(subreg, val);
8660 if ((~subreg.mask & subreg.value) & val)
8662 if (!((subreg.mask | subreg.value) & val))
8666 if (reg->u32_min_value > val)
8668 else if (reg->u32_max_value <= val)
8672 if (reg->s32_min_value > sval)
8674 else if (reg->s32_max_value <= sval)
8678 if (reg->u32_max_value < val)
8680 else if (reg->u32_min_value >= val)
8684 if (reg->s32_max_value < sval)
8686 else if (reg->s32_min_value >= sval)
8690 if (reg->u32_min_value >= val)
8692 else if (reg->u32_max_value < val)
8696 if (reg->s32_min_value >= sval)
8698 else if (reg->s32_max_value < sval)
8702 if (reg->u32_max_value <= val)
8704 else if (reg->u32_min_value > val)
8708 if (reg->s32_max_value <= sval)
8710 else if (reg->s32_min_value > sval)
8719 static int is_branch64_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
8721 s64 sval = (s64)val;
8725 if (tnum_is_const(reg->var_off))
8726 return !!tnum_equals_const(reg->var_off, val);
8729 if (tnum_is_const(reg->var_off))
8730 return !tnum_equals_const(reg->var_off, val);
8733 if ((~reg->var_off.mask & reg->var_off.value) & val)
8735 if (!((reg->var_off.mask | reg->var_off.value) & val))
8739 if (reg->umin_value > val)
8741 else if (reg->umax_value <= val)
8745 if (reg->smin_value > sval)
8747 else if (reg->smax_value <= sval)
8751 if (reg->umax_value < val)
8753 else if (reg->umin_value >= val)
8757 if (reg->smax_value < sval)
8759 else if (reg->smin_value >= sval)
8763 if (reg->umin_value >= val)
8765 else if (reg->umax_value < val)
8769 if (reg->smin_value >= sval)
8771 else if (reg->smax_value < sval)
8775 if (reg->umax_value <= val)
8777 else if (reg->umin_value > val)
8781 if (reg->smax_value <= sval)
8783 else if (reg->smin_value > sval)
8791 /* compute branch direction of the expression "if (reg opcode val) goto target;"
8793 * 1 - branch will be taken and "goto target" will be executed
8794 * 0 - branch will not be taken and fall-through to next insn
8795 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value
8798 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
8801 if (__is_pointer_value(false, reg)) {
8802 if (!reg_type_not_null(reg->type))
8805 /* If pointer is valid tests against zero will fail so we can
8806 * use this to direct branch taken.
8822 return is_branch32_taken(reg, val, opcode);
8823 return is_branch64_taken(reg, val, opcode);
8826 static int flip_opcode(u32 opcode)
8828 /* How can we transform "a <op> b" into "b <op> a"? */
8829 static const u8 opcode_flip[16] = {
8830 /* these stay the same */
8831 [BPF_JEQ >> 4] = BPF_JEQ,
8832 [BPF_JNE >> 4] = BPF_JNE,
8833 [BPF_JSET >> 4] = BPF_JSET,
8834 /* these swap "lesser" and "greater" (L and G in the opcodes) */
8835 [BPF_JGE >> 4] = BPF_JLE,
8836 [BPF_JGT >> 4] = BPF_JLT,
8837 [BPF_JLE >> 4] = BPF_JGE,
8838 [BPF_JLT >> 4] = BPF_JGT,
8839 [BPF_JSGE >> 4] = BPF_JSLE,
8840 [BPF_JSGT >> 4] = BPF_JSLT,
8841 [BPF_JSLE >> 4] = BPF_JSGE,
8842 [BPF_JSLT >> 4] = BPF_JSGT
8844 return opcode_flip[opcode >> 4];
8847 static int is_pkt_ptr_branch_taken(struct bpf_reg_state *dst_reg,
8848 struct bpf_reg_state *src_reg,
8851 struct bpf_reg_state *pkt;
8853 if (src_reg->type == PTR_TO_PACKET_END) {
8855 } else if (dst_reg->type == PTR_TO_PACKET_END) {
8857 opcode = flip_opcode(opcode);
8862 if (pkt->range >= 0)
8867 /* pkt <= pkt_end */
8871 if (pkt->range == BEYOND_PKT_END)
8872 /* pkt has at last one extra byte beyond pkt_end */
8873 return opcode == BPF_JGT;
8879 /* pkt >= pkt_end */
8880 if (pkt->range == BEYOND_PKT_END || pkt->range == AT_PKT_END)
8881 return opcode == BPF_JGE;
8887 /* Adjusts the register min/max values in the case that the dst_reg is the
8888 * variable register that we are working on, and src_reg is a constant or we're
8889 * simply doing a BPF_K check.
8890 * In JEQ/JNE cases we also adjust the var_off values.
8892 static void reg_set_min_max(struct bpf_reg_state *true_reg,
8893 struct bpf_reg_state *false_reg,
8895 u8 opcode, bool is_jmp32)
8897 struct tnum false_32off = tnum_subreg(false_reg->var_off);
8898 struct tnum false_64off = false_reg->var_off;
8899 struct tnum true_32off = tnum_subreg(true_reg->var_off);
8900 struct tnum true_64off = true_reg->var_off;
8901 s64 sval = (s64)val;
8902 s32 sval32 = (s32)val32;
8904 /* If the dst_reg is a pointer, we can't learn anything about its
8905 * variable offset from the compare (unless src_reg were a pointer into
8906 * the same object, but we don't bother with that.
8907 * Since false_reg and true_reg have the same type by construction, we
8908 * only need to check one of them for pointerness.
8910 if (__is_pointer_value(false, false_reg))
8917 struct bpf_reg_state *reg =
8918 opcode == BPF_JEQ ? true_reg : false_reg;
8920 /* JEQ/JNE comparison doesn't change the register equivalence.
8922 * if (r1 == 42) goto label;
8924 * label: // here both r1 and r2 are known to be 42.
8926 * Hence when marking register as known preserve it's ID.
8929 __mark_reg32_known(reg, val32);
8931 ___mark_reg_known(reg, val);
8936 false_32off = tnum_and(false_32off, tnum_const(~val32));
8937 if (is_power_of_2(val32))
8938 true_32off = tnum_or(true_32off,
8941 false_64off = tnum_and(false_64off, tnum_const(~val));
8942 if (is_power_of_2(val))
8943 true_64off = tnum_or(true_64off,
8951 u32 false_umax = opcode == BPF_JGT ? val32 : val32 - 1;
8952 u32 true_umin = opcode == BPF_JGT ? val32 + 1 : val32;
8954 false_reg->u32_max_value = min(false_reg->u32_max_value,
8956 true_reg->u32_min_value = max(true_reg->u32_min_value,
8959 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
8960 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
8962 false_reg->umax_value = min(false_reg->umax_value, false_umax);
8963 true_reg->umin_value = max(true_reg->umin_value, true_umin);
8971 s32 false_smax = opcode == BPF_JSGT ? sval32 : sval32 - 1;
8972 s32 true_smin = opcode == BPF_JSGT ? sval32 + 1 : sval32;
8974 false_reg->s32_max_value = min(false_reg->s32_max_value, false_smax);
8975 true_reg->s32_min_value = max(true_reg->s32_min_value, true_smin);
8977 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
8978 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
8980 false_reg->smax_value = min(false_reg->smax_value, false_smax);
8981 true_reg->smin_value = max(true_reg->smin_value, true_smin);
8989 u32 false_umin = opcode == BPF_JLT ? val32 : val32 + 1;
8990 u32 true_umax = opcode == BPF_JLT ? val32 - 1 : val32;
8992 false_reg->u32_min_value = max(false_reg->u32_min_value,
8994 true_reg->u32_max_value = min(true_reg->u32_max_value,
8997 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
8998 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
9000 false_reg->umin_value = max(false_reg->umin_value, false_umin);
9001 true_reg->umax_value = min(true_reg->umax_value, true_umax);
9009 s32 false_smin = opcode == BPF_JSLT ? sval32 : sval32 + 1;
9010 s32 true_smax = opcode == BPF_JSLT ? sval32 - 1 : sval32;
9012 false_reg->s32_min_value = max(false_reg->s32_min_value, false_smin);
9013 true_reg->s32_max_value = min(true_reg->s32_max_value, true_smax);
9015 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
9016 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
9018 false_reg->smin_value = max(false_reg->smin_value, false_smin);
9019 true_reg->smax_value = min(true_reg->smax_value, true_smax);
9028 false_reg->var_off = tnum_or(tnum_clear_subreg(false_64off),
9029 tnum_subreg(false_32off));
9030 true_reg->var_off = tnum_or(tnum_clear_subreg(true_64off),
9031 tnum_subreg(true_32off));
9032 __reg_combine_32_into_64(false_reg);
9033 __reg_combine_32_into_64(true_reg);
9035 false_reg->var_off = false_64off;
9036 true_reg->var_off = true_64off;
9037 __reg_combine_64_into_32(false_reg);
9038 __reg_combine_64_into_32(true_reg);
9042 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
9045 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
9046 struct bpf_reg_state *false_reg,
9048 u8 opcode, bool is_jmp32)
9050 opcode = flip_opcode(opcode);
9051 /* This uses zero as "not present in table"; luckily the zero opcode,
9052 * BPF_JA, can't get here.
9055 reg_set_min_max(true_reg, false_reg, val, val32, opcode, is_jmp32);
9058 /* Regs are known to be equal, so intersect their min/max/var_off */
9059 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
9060 struct bpf_reg_state *dst_reg)
9062 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
9063 dst_reg->umin_value);
9064 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
9065 dst_reg->umax_value);
9066 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
9067 dst_reg->smin_value);
9068 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
9069 dst_reg->smax_value);
9070 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
9072 /* We might have learned new bounds from the var_off. */
9073 __update_reg_bounds(src_reg);
9074 __update_reg_bounds(dst_reg);
9075 /* We might have learned something about the sign bit. */
9076 __reg_deduce_bounds(src_reg);
9077 __reg_deduce_bounds(dst_reg);
9078 /* We might have learned some bits from the bounds. */
9079 __reg_bound_offset(src_reg);
9080 __reg_bound_offset(dst_reg);
9081 /* Intersecting with the old var_off might have improved our bounds
9082 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
9083 * then new var_off is (0; 0x7f...fc) which improves our umax.
9085 __update_reg_bounds(src_reg);
9086 __update_reg_bounds(dst_reg);
9089 static void reg_combine_min_max(struct bpf_reg_state *true_src,
9090 struct bpf_reg_state *true_dst,
9091 struct bpf_reg_state *false_src,
9092 struct bpf_reg_state *false_dst,
9097 __reg_combine_min_max(true_src, true_dst);
9100 __reg_combine_min_max(false_src, false_dst);
9105 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
9106 struct bpf_reg_state *reg, u32 id,
9109 if (type_may_be_null(reg->type) && reg->id == id &&
9110 !WARN_ON_ONCE(!reg->id)) {
9111 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
9112 !tnum_equals_const(reg->var_off, 0) ||
9114 /* Old offset (both fixed and variable parts) should
9115 * have been known-zero, because we don't allow pointer
9116 * arithmetic on pointers that might be NULL. If we
9117 * see this happening, don't convert the register.
9122 reg->type = SCALAR_VALUE;
9123 /* We don't need id and ref_obj_id from this point
9124 * onwards anymore, thus we should better reset it,
9125 * so that state pruning has chances to take effect.
9128 reg->ref_obj_id = 0;
9133 mark_ptr_not_null_reg(reg);
9135 if (!reg_may_point_to_spin_lock(reg)) {
9136 /* For not-NULL ptr, reg->ref_obj_id will be reset
9137 * in release_reg_references().
9139 * reg->id is still used by spin_lock ptr. Other
9140 * than spin_lock ptr type, reg->id can be reset.
9147 static void __mark_ptr_or_null_regs(struct bpf_func_state *state, u32 id,
9150 struct bpf_reg_state *reg;
9153 for (i = 0; i < MAX_BPF_REG; i++)
9154 mark_ptr_or_null_reg(state, &state->regs[i], id, is_null);
9156 bpf_for_each_spilled_reg(i, state, reg) {
9159 mark_ptr_or_null_reg(state, reg, id, is_null);
9163 /* The logic is similar to find_good_pkt_pointers(), both could eventually
9164 * be folded together at some point.
9166 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
9169 struct bpf_func_state *state = vstate->frame[vstate->curframe];
9170 struct bpf_reg_state *regs = state->regs;
9171 u32 ref_obj_id = regs[regno].ref_obj_id;
9172 u32 id = regs[regno].id;
9175 if (ref_obj_id && ref_obj_id == id && is_null)
9176 /* regs[regno] is in the " == NULL" branch.
9177 * No one could have freed the reference state before
9178 * doing the NULL check.
9180 WARN_ON_ONCE(release_reference_state(state, id));
9182 for (i = 0; i <= vstate->curframe; i++)
9183 __mark_ptr_or_null_regs(vstate->frame[i], id, is_null);
9186 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
9187 struct bpf_reg_state *dst_reg,
9188 struct bpf_reg_state *src_reg,
9189 struct bpf_verifier_state *this_branch,
9190 struct bpf_verifier_state *other_branch)
9192 if (BPF_SRC(insn->code) != BPF_X)
9195 /* Pointers are always 64-bit. */
9196 if (BPF_CLASS(insn->code) == BPF_JMP32)
9199 switch (BPF_OP(insn->code)) {
9201 if ((dst_reg->type == PTR_TO_PACKET &&
9202 src_reg->type == PTR_TO_PACKET_END) ||
9203 (dst_reg->type == PTR_TO_PACKET_META &&
9204 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
9205 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
9206 find_good_pkt_pointers(this_branch, dst_reg,
9207 dst_reg->type, false);
9208 mark_pkt_end(other_branch, insn->dst_reg, true);
9209 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
9210 src_reg->type == PTR_TO_PACKET) ||
9211 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
9212 src_reg->type == PTR_TO_PACKET_META)) {
9213 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
9214 find_good_pkt_pointers(other_branch, src_reg,
9215 src_reg->type, true);
9216 mark_pkt_end(this_branch, insn->src_reg, false);
9222 if ((dst_reg->type == PTR_TO_PACKET &&
9223 src_reg->type == PTR_TO_PACKET_END) ||
9224 (dst_reg->type == PTR_TO_PACKET_META &&
9225 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
9226 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
9227 find_good_pkt_pointers(other_branch, dst_reg,
9228 dst_reg->type, true);
9229 mark_pkt_end(this_branch, insn->dst_reg, false);
9230 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
9231 src_reg->type == PTR_TO_PACKET) ||
9232 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
9233 src_reg->type == PTR_TO_PACKET_META)) {
9234 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
9235 find_good_pkt_pointers(this_branch, src_reg,
9236 src_reg->type, false);
9237 mark_pkt_end(other_branch, insn->src_reg, true);
9243 if ((dst_reg->type == PTR_TO_PACKET &&
9244 src_reg->type == PTR_TO_PACKET_END) ||
9245 (dst_reg->type == PTR_TO_PACKET_META &&
9246 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
9247 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
9248 find_good_pkt_pointers(this_branch, dst_reg,
9249 dst_reg->type, true);
9250 mark_pkt_end(other_branch, insn->dst_reg, false);
9251 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
9252 src_reg->type == PTR_TO_PACKET) ||
9253 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
9254 src_reg->type == PTR_TO_PACKET_META)) {
9255 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
9256 find_good_pkt_pointers(other_branch, src_reg,
9257 src_reg->type, false);
9258 mark_pkt_end(this_branch, insn->src_reg, true);
9264 if ((dst_reg->type == PTR_TO_PACKET &&
9265 src_reg->type == PTR_TO_PACKET_END) ||
9266 (dst_reg->type == PTR_TO_PACKET_META &&
9267 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
9268 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
9269 find_good_pkt_pointers(other_branch, dst_reg,
9270 dst_reg->type, false);
9271 mark_pkt_end(this_branch, insn->dst_reg, true);
9272 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
9273 src_reg->type == PTR_TO_PACKET) ||
9274 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
9275 src_reg->type == PTR_TO_PACKET_META)) {
9276 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
9277 find_good_pkt_pointers(this_branch, src_reg,
9278 src_reg->type, true);
9279 mark_pkt_end(other_branch, insn->src_reg, false);
9291 static void find_equal_scalars(struct bpf_verifier_state *vstate,
9292 struct bpf_reg_state *known_reg)
9294 struct bpf_func_state *state;
9295 struct bpf_reg_state *reg;
9298 for (i = 0; i <= vstate->curframe; i++) {
9299 state = vstate->frame[i];
9300 for (j = 0; j < MAX_BPF_REG; j++) {
9301 reg = &state->regs[j];
9302 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
9306 bpf_for_each_spilled_reg(j, state, reg) {
9309 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
9315 static int check_cond_jmp_op(struct bpf_verifier_env *env,
9316 struct bpf_insn *insn, int *insn_idx)
9318 struct bpf_verifier_state *this_branch = env->cur_state;
9319 struct bpf_verifier_state *other_branch;
9320 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
9321 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
9322 u8 opcode = BPF_OP(insn->code);
9327 /* Only conditional jumps are expected to reach here. */
9328 if (opcode == BPF_JA || opcode > BPF_JSLE) {
9329 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
9333 if (BPF_SRC(insn->code) == BPF_X) {
9334 if (insn->imm != 0) {
9335 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
9339 /* check src1 operand */
9340 err = check_reg_arg(env, insn->src_reg, SRC_OP);
9344 if (is_pointer_value(env, insn->src_reg)) {
9345 verbose(env, "R%d pointer comparison prohibited\n",
9349 src_reg = ®s[insn->src_reg];
9351 if (insn->src_reg != BPF_REG_0) {
9352 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
9357 /* check src2 operand */
9358 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
9362 dst_reg = ®s[insn->dst_reg];
9363 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
9365 if (BPF_SRC(insn->code) == BPF_K) {
9366 pred = is_branch_taken(dst_reg, insn->imm, opcode, is_jmp32);
9367 } else if (src_reg->type == SCALAR_VALUE &&
9368 is_jmp32 && tnum_is_const(tnum_subreg(src_reg->var_off))) {
9369 pred = is_branch_taken(dst_reg,
9370 tnum_subreg(src_reg->var_off).value,
9373 } else if (src_reg->type == SCALAR_VALUE &&
9374 !is_jmp32 && tnum_is_const(src_reg->var_off)) {
9375 pred = is_branch_taken(dst_reg,
9376 src_reg->var_off.value,
9379 } else if (reg_is_pkt_pointer_any(dst_reg) &&
9380 reg_is_pkt_pointer_any(src_reg) &&
9382 pred = is_pkt_ptr_branch_taken(dst_reg, src_reg, opcode);
9386 /* If we get here with a dst_reg pointer type it is because
9387 * above is_branch_taken() special cased the 0 comparison.
9389 if (!__is_pointer_value(false, dst_reg))
9390 err = mark_chain_precision(env, insn->dst_reg);
9391 if (BPF_SRC(insn->code) == BPF_X && !err &&
9392 !__is_pointer_value(false, src_reg))
9393 err = mark_chain_precision(env, insn->src_reg);
9399 /* Only follow the goto, ignore fall-through. If needed, push
9400 * the fall-through branch for simulation under speculative
9403 if (!env->bypass_spec_v1 &&
9404 !sanitize_speculative_path(env, insn, *insn_idx + 1,
9407 *insn_idx += insn->off;
9409 } else if (pred == 0) {
9410 /* Only follow the fall-through branch, since that's where the
9411 * program will go. If needed, push the goto branch for
9412 * simulation under speculative execution.
9414 if (!env->bypass_spec_v1 &&
9415 !sanitize_speculative_path(env, insn,
9416 *insn_idx + insn->off + 1,
9422 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
9426 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
9428 /* detect if we are comparing against a constant value so we can adjust
9429 * our min/max values for our dst register.
9430 * this is only legit if both are scalars (or pointers to the same
9431 * object, I suppose, but we don't support that right now), because
9432 * otherwise the different base pointers mean the offsets aren't
9435 if (BPF_SRC(insn->code) == BPF_X) {
9436 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
9438 if (dst_reg->type == SCALAR_VALUE &&
9439 src_reg->type == SCALAR_VALUE) {
9440 if (tnum_is_const(src_reg->var_off) ||
9442 tnum_is_const(tnum_subreg(src_reg->var_off))))
9443 reg_set_min_max(&other_branch_regs[insn->dst_reg],
9445 src_reg->var_off.value,
9446 tnum_subreg(src_reg->var_off).value,
9448 else if (tnum_is_const(dst_reg->var_off) ||
9450 tnum_is_const(tnum_subreg(dst_reg->var_off))))
9451 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
9453 dst_reg->var_off.value,
9454 tnum_subreg(dst_reg->var_off).value,
9456 else if (!is_jmp32 &&
9457 (opcode == BPF_JEQ || opcode == BPF_JNE))
9458 /* Comparing for equality, we can combine knowledge */
9459 reg_combine_min_max(&other_branch_regs[insn->src_reg],
9460 &other_branch_regs[insn->dst_reg],
9461 src_reg, dst_reg, opcode);
9463 !WARN_ON_ONCE(src_reg->id != other_branch_regs[insn->src_reg].id)) {
9464 find_equal_scalars(this_branch, src_reg);
9465 find_equal_scalars(other_branch, &other_branch_regs[insn->src_reg]);
9469 } else if (dst_reg->type == SCALAR_VALUE) {
9470 reg_set_min_max(&other_branch_regs[insn->dst_reg],
9471 dst_reg, insn->imm, (u32)insn->imm,
9475 if (dst_reg->type == SCALAR_VALUE && dst_reg->id &&
9476 !WARN_ON_ONCE(dst_reg->id != other_branch_regs[insn->dst_reg].id)) {
9477 find_equal_scalars(this_branch, dst_reg);
9478 find_equal_scalars(other_branch, &other_branch_regs[insn->dst_reg]);
9481 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
9482 * NOTE: these optimizations below are related with pointer comparison
9483 * which will never be JMP32.
9485 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
9486 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
9487 type_may_be_null(dst_reg->type)) {
9488 /* Mark all identical registers in each branch as either
9489 * safe or unknown depending R == 0 or R != 0 conditional.
9491 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
9493 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
9495 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
9496 this_branch, other_branch) &&
9497 is_pointer_value(env, insn->dst_reg)) {
9498 verbose(env, "R%d pointer comparison prohibited\n",
9502 if (env->log.level & BPF_LOG_LEVEL)
9503 print_insn_state(env, this_branch->frame[this_branch->curframe]);
9507 /* verify BPF_LD_IMM64 instruction */
9508 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
9510 struct bpf_insn_aux_data *aux = cur_aux(env);
9511 struct bpf_reg_state *regs = cur_regs(env);
9512 struct bpf_reg_state *dst_reg;
9513 struct bpf_map *map;
9516 if (BPF_SIZE(insn->code) != BPF_DW) {
9517 verbose(env, "invalid BPF_LD_IMM insn\n");
9520 if (insn->off != 0) {
9521 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
9525 err = check_reg_arg(env, insn->dst_reg, DST_OP);
9529 dst_reg = ®s[insn->dst_reg];
9530 if (insn->src_reg == 0) {
9531 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
9533 dst_reg->type = SCALAR_VALUE;
9534 __mark_reg_known(®s[insn->dst_reg], imm);
9538 /* All special src_reg cases are listed below. From this point onwards
9539 * we either succeed and assign a corresponding dst_reg->type after
9540 * zeroing the offset, or fail and reject the program.
9542 mark_reg_known_zero(env, regs, insn->dst_reg);
9544 if (insn->src_reg == BPF_PSEUDO_BTF_ID) {
9545 dst_reg->type = aux->btf_var.reg_type;
9546 switch (base_type(dst_reg->type)) {
9548 dst_reg->mem_size = aux->btf_var.mem_size;
9551 case PTR_TO_PERCPU_BTF_ID:
9552 dst_reg->btf = aux->btf_var.btf;
9553 dst_reg->btf_id = aux->btf_var.btf_id;
9556 verbose(env, "bpf verifier is misconfigured\n");
9562 if (insn->src_reg == BPF_PSEUDO_FUNC) {
9563 struct bpf_prog_aux *aux = env->prog->aux;
9564 u32 subprogno = find_subprog(env,
9565 env->insn_idx + insn->imm + 1);
9567 if (!aux->func_info) {
9568 verbose(env, "missing btf func_info\n");
9571 if (aux->func_info_aux[subprogno].linkage != BTF_FUNC_STATIC) {
9572 verbose(env, "callback function not static\n");
9576 dst_reg->type = PTR_TO_FUNC;
9577 dst_reg->subprogno = subprogno;
9581 map = env->used_maps[aux->map_index];
9582 dst_reg->map_ptr = map;
9584 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE ||
9585 insn->src_reg == BPF_PSEUDO_MAP_IDX_VALUE) {
9586 dst_reg->type = PTR_TO_MAP_VALUE;
9587 dst_reg->off = aux->map_off;
9588 if (map_value_has_spin_lock(map))
9589 dst_reg->id = ++env->id_gen;
9590 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD ||
9591 insn->src_reg == BPF_PSEUDO_MAP_IDX) {
9592 dst_reg->type = CONST_PTR_TO_MAP;
9594 verbose(env, "bpf verifier is misconfigured\n");
9601 static bool may_access_skb(enum bpf_prog_type type)
9604 case BPF_PROG_TYPE_SOCKET_FILTER:
9605 case BPF_PROG_TYPE_SCHED_CLS:
9606 case BPF_PROG_TYPE_SCHED_ACT:
9613 /* verify safety of LD_ABS|LD_IND instructions:
9614 * - they can only appear in the programs where ctx == skb
9615 * - since they are wrappers of function calls, they scratch R1-R5 registers,
9616 * preserve R6-R9, and store return value into R0
9619 * ctx == skb == R6 == CTX
9622 * SRC == any register
9623 * IMM == 32-bit immediate
9626 * R0 - 8/16/32-bit skb data converted to cpu endianness
9628 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
9630 struct bpf_reg_state *regs = cur_regs(env);
9631 static const int ctx_reg = BPF_REG_6;
9632 u8 mode = BPF_MODE(insn->code);
9635 if (!may_access_skb(resolve_prog_type(env->prog))) {
9636 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
9640 if (!env->ops->gen_ld_abs) {
9641 verbose(env, "bpf verifier is misconfigured\n");
9645 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
9646 BPF_SIZE(insn->code) == BPF_DW ||
9647 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
9648 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
9652 /* check whether implicit source operand (register R6) is readable */
9653 err = check_reg_arg(env, ctx_reg, SRC_OP);
9657 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
9658 * gen_ld_abs() may terminate the program at runtime, leading to
9661 err = check_reference_leak(env);
9663 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
9667 if (env->cur_state->active_spin_lock) {
9668 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
9672 if (regs[ctx_reg].type != PTR_TO_CTX) {
9674 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
9678 if (mode == BPF_IND) {
9679 /* check explicit source operand */
9680 err = check_reg_arg(env, insn->src_reg, SRC_OP);
9685 err = check_ptr_off_reg(env, ®s[ctx_reg], ctx_reg);
9689 /* reset caller saved regs to unreadable */
9690 for (i = 0; i < CALLER_SAVED_REGS; i++) {
9691 mark_reg_not_init(env, regs, caller_saved[i]);
9692 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
9695 /* mark destination R0 register as readable, since it contains
9696 * the value fetched from the packet.
9697 * Already marked as written above.
9699 mark_reg_unknown(env, regs, BPF_REG_0);
9700 /* ld_abs load up to 32-bit skb data. */
9701 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
9705 static int check_return_code(struct bpf_verifier_env *env)
9707 struct tnum enforce_attach_type_range = tnum_unknown;
9708 const struct bpf_prog *prog = env->prog;
9709 struct bpf_reg_state *reg;
9710 struct tnum range = tnum_range(0, 1);
9711 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
9713 struct bpf_func_state *frame = env->cur_state->frame[0];
9714 const bool is_subprog = frame->subprogno;
9716 /* LSM and struct_ops func-ptr's return type could be "void" */
9718 (prog_type == BPF_PROG_TYPE_STRUCT_OPS ||
9719 prog_type == BPF_PROG_TYPE_LSM) &&
9720 !prog->aux->attach_func_proto->type)
9723 /* eBPF calling convention is such that R0 is used
9724 * to return the value from eBPF program.
9725 * Make sure that it's readable at this time
9726 * of bpf_exit, which means that program wrote
9727 * something into it earlier
9729 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
9733 if (is_pointer_value(env, BPF_REG_0)) {
9734 verbose(env, "R0 leaks addr as return value\n");
9738 reg = cur_regs(env) + BPF_REG_0;
9740 if (frame->in_async_callback_fn) {
9741 /* enforce return zero from async callbacks like timer */
9742 if (reg->type != SCALAR_VALUE) {
9743 verbose(env, "In async callback the register R0 is not a known value (%s)\n",
9744 reg_type_str(env, reg->type));
9748 if (!tnum_in(tnum_const(0), reg->var_off)) {
9749 verbose_invalid_scalar(env, reg, &range, "async callback", "R0");
9756 if (reg->type != SCALAR_VALUE) {
9757 verbose(env, "At subprogram exit the register R0 is not a scalar value (%s)\n",
9758 reg_type_str(env, reg->type));
9764 switch (prog_type) {
9765 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
9766 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
9767 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG ||
9768 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETPEERNAME ||
9769 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETPEERNAME ||
9770 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETSOCKNAME ||
9771 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETSOCKNAME)
9772 range = tnum_range(1, 1);
9773 if (env->prog->expected_attach_type == BPF_CGROUP_INET4_BIND ||
9774 env->prog->expected_attach_type == BPF_CGROUP_INET6_BIND)
9775 range = tnum_range(0, 3);
9777 case BPF_PROG_TYPE_CGROUP_SKB:
9778 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
9779 range = tnum_range(0, 3);
9780 enforce_attach_type_range = tnum_range(2, 3);
9783 case BPF_PROG_TYPE_CGROUP_SOCK:
9784 case BPF_PROG_TYPE_SOCK_OPS:
9785 case BPF_PROG_TYPE_CGROUP_DEVICE:
9786 case BPF_PROG_TYPE_CGROUP_SYSCTL:
9787 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
9789 case BPF_PROG_TYPE_RAW_TRACEPOINT:
9790 if (!env->prog->aux->attach_btf_id)
9792 range = tnum_const(0);
9794 case BPF_PROG_TYPE_TRACING:
9795 switch (env->prog->expected_attach_type) {
9796 case BPF_TRACE_FENTRY:
9797 case BPF_TRACE_FEXIT:
9798 range = tnum_const(0);
9800 case BPF_TRACE_RAW_TP:
9801 case BPF_MODIFY_RETURN:
9803 case BPF_TRACE_ITER:
9809 case BPF_PROG_TYPE_SK_LOOKUP:
9810 range = tnum_range(SK_DROP, SK_PASS);
9812 case BPF_PROG_TYPE_EXT:
9813 /* freplace program can return anything as its return value
9814 * depends on the to-be-replaced kernel func or bpf program.
9820 if (reg->type != SCALAR_VALUE) {
9821 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
9822 reg_type_str(env, reg->type));
9826 if (!tnum_in(range, reg->var_off)) {
9827 verbose_invalid_scalar(env, reg, &range, "program exit", "R0");
9831 if (!tnum_is_unknown(enforce_attach_type_range) &&
9832 tnum_in(enforce_attach_type_range, reg->var_off))
9833 env->prog->enforce_expected_attach_type = 1;
9837 /* non-recursive DFS pseudo code
9838 * 1 procedure DFS-iterative(G,v):
9839 * 2 label v as discovered
9840 * 3 let S be a stack
9842 * 5 while S is not empty
9844 * 7 if t is what we're looking for:
9846 * 9 for all edges e in G.adjacentEdges(t) do
9847 * 10 if edge e is already labelled
9848 * 11 continue with the next edge
9849 * 12 w <- G.adjacentVertex(t,e)
9850 * 13 if vertex w is not discovered and not explored
9851 * 14 label e as tree-edge
9852 * 15 label w as discovered
9855 * 18 else if vertex w is discovered
9856 * 19 label e as back-edge
9858 * 21 // vertex w is explored
9859 * 22 label e as forward- or cross-edge
9860 * 23 label t as explored
9865 * 0x11 - discovered and fall-through edge labelled
9866 * 0x12 - discovered and fall-through and branch edges labelled
9877 static u32 state_htab_size(struct bpf_verifier_env *env)
9879 return env->prog->len;
9882 static struct bpf_verifier_state_list **explored_state(
9883 struct bpf_verifier_env *env,
9886 struct bpf_verifier_state *cur = env->cur_state;
9887 struct bpf_func_state *state = cur->frame[cur->curframe];
9889 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
9892 static void init_explored_state(struct bpf_verifier_env *env, int idx)
9894 env->insn_aux_data[idx].prune_point = true;
9902 /* t, w, e - match pseudo-code above:
9903 * t - index of current instruction
9904 * w - next instruction
9907 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env,
9910 int *insn_stack = env->cfg.insn_stack;
9911 int *insn_state = env->cfg.insn_state;
9913 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
9914 return DONE_EXPLORING;
9916 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
9917 return DONE_EXPLORING;
9919 if (w < 0 || w >= env->prog->len) {
9920 verbose_linfo(env, t, "%d: ", t);
9921 verbose(env, "jump out of range from insn %d to %d\n", t, w);
9926 /* mark branch target for state pruning */
9927 init_explored_state(env, w);
9929 if (insn_state[w] == 0) {
9931 insn_state[t] = DISCOVERED | e;
9932 insn_state[w] = DISCOVERED;
9933 if (env->cfg.cur_stack >= env->prog->len)
9935 insn_stack[env->cfg.cur_stack++] = w;
9936 return KEEP_EXPLORING;
9937 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
9938 if (loop_ok && env->bpf_capable)
9939 return DONE_EXPLORING;
9940 verbose_linfo(env, t, "%d: ", t);
9941 verbose_linfo(env, w, "%d: ", w);
9942 verbose(env, "back-edge from insn %d to %d\n", t, w);
9944 } else if (insn_state[w] == EXPLORED) {
9945 /* forward- or cross-edge */
9946 insn_state[t] = DISCOVERED | e;
9948 verbose(env, "insn state internal bug\n");
9951 return DONE_EXPLORING;
9954 static int visit_func_call_insn(int t, int insn_cnt,
9955 struct bpf_insn *insns,
9956 struct bpf_verifier_env *env,
9961 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
9965 if (t + 1 < insn_cnt)
9966 init_explored_state(env, t + 1);
9968 init_explored_state(env, t);
9969 ret = push_insn(t, t + insns[t].imm + 1, BRANCH, env,
9970 /* It's ok to allow recursion from CFG point of
9971 * view. __check_func_call() will do the actual
9974 bpf_pseudo_func(insns + t));
9979 /* Visits the instruction at index t and returns one of the following:
9980 * < 0 - an error occurred
9981 * DONE_EXPLORING - the instruction was fully explored
9982 * KEEP_EXPLORING - there is still work to be done before it is fully explored
9984 static int visit_insn(int t, int insn_cnt, struct bpf_verifier_env *env)
9986 struct bpf_insn *insns = env->prog->insnsi;
9989 if (bpf_pseudo_func(insns + t))
9990 return visit_func_call_insn(t, insn_cnt, insns, env, true);
9992 /* All non-branch instructions have a single fall-through edge. */
9993 if (BPF_CLASS(insns[t].code) != BPF_JMP &&
9994 BPF_CLASS(insns[t].code) != BPF_JMP32)
9995 return push_insn(t, t + 1, FALLTHROUGH, env, false);
9997 switch (BPF_OP(insns[t].code)) {
9999 return DONE_EXPLORING;
10002 if (insns[t].imm == BPF_FUNC_timer_set_callback)
10003 /* Mark this call insn to trigger is_state_visited() check
10004 * before call itself is processed by __check_func_call().
10005 * Otherwise new async state will be pushed for further
10008 init_explored_state(env, t);
10009 return visit_func_call_insn(t, insn_cnt, insns, env,
10010 insns[t].src_reg == BPF_PSEUDO_CALL);
10013 if (BPF_SRC(insns[t].code) != BPF_K)
10016 /* unconditional jump with single edge */
10017 ret = push_insn(t, t + insns[t].off + 1, FALLTHROUGH, env,
10022 /* unconditional jmp is not a good pruning point,
10023 * but it's marked, since backtracking needs
10024 * to record jmp history in is_state_visited().
10026 init_explored_state(env, t + insns[t].off + 1);
10027 /* tell verifier to check for equivalent states
10028 * after every call and jump
10030 if (t + 1 < insn_cnt)
10031 init_explored_state(env, t + 1);
10036 /* conditional jump with two edges */
10037 init_explored_state(env, t);
10038 ret = push_insn(t, t + 1, FALLTHROUGH, env, true);
10042 return push_insn(t, t + insns[t].off + 1, BRANCH, env, true);
10046 /* non-recursive depth-first-search to detect loops in BPF program
10047 * loop == back-edge in directed graph
10049 static int check_cfg(struct bpf_verifier_env *env)
10051 int insn_cnt = env->prog->len;
10052 int *insn_stack, *insn_state;
10056 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
10060 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
10062 kvfree(insn_state);
10066 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
10067 insn_stack[0] = 0; /* 0 is the first instruction */
10068 env->cfg.cur_stack = 1;
10070 while (env->cfg.cur_stack > 0) {
10071 int t = insn_stack[env->cfg.cur_stack - 1];
10073 ret = visit_insn(t, insn_cnt, env);
10075 case DONE_EXPLORING:
10076 insn_state[t] = EXPLORED;
10077 env->cfg.cur_stack--;
10079 case KEEP_EXPLORING:
10083 verbose(env, "visit_insn internal bug\n");
10090 if (env->cfg.cur_stack < 0) {
10091 verbose(env, "pop stack internal bug\n");
10096 for (i = 0; i < insn_cnt; i++) {
10097 if (insn_state[i] != EXPLORED) {
10098 verbose(env, "unreachable insn %d\n", i);
10103 ret = 0; /* cfg looks good */
10106 kvfree(insn_state);
10107 kvfree(insn_stack);
10108 env->cfg.insn_state = env->cfg.insn_stack = NULL;
10112 static int check_abnormal_return(struct bpf_verifier_env *env)
10116 for (i = 1; i < env->subprog_cnt; i++) {
10117 if (env->subprog_info[i].has_ld_abs) {
10118 verbose(env, "LD_ABS is not allowed in subprogs without BTF\n");
10121 if (env->subprog_info[i].has_tail_call) {
10122 verbose(env, "tail_call is not allowed in subprogs without BTF\n");
10129 /* The minimum supported BTF func info size */
10130 #define MIN_BPF_FUNCINFO_SIZE 8
10131 #define MAX_FUNCINFO_REC_SIZE 252
10133 static int check_btf_func(struct bpf_verifier_env *env,
10134 const union bpf_attr *attr,
10137 const struct btf_type *type, *func_proto, *ret_type;
10138 u32 i, nfuncs, urec_size, min_size;
10139 u32 krec_size = sizeof(struct bpf_func_info);
10140 struct bpf_func_info *krecord;
10141 struct bpf_func_info_aux *info_aux = NULL;
10142 struct bpf_prog *prog;
10143 const struct btf *btf;
10145 u32 prev_offset = 0;
10146 bool scalar_return;
10149 nfuncs = attr->func_info_cnt;
10151 if (check_abnormal_return(env))
10156 if (nfuncs != env->subprog_cnt) {
10157 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
10161 urec_size = attr->func_info_rec_size;
10162 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
10163 urec_size > MAX_FUNCINFO_REC_SIZE ||
10164 urec_size % sizeof(u32)) {
10165 verbose(env, "invalid func info rec size %u\n", urec_size);
10170 btf = prog->aux->btf;
10172 urecord = make_bpfptr(attr->func_info, uattr.is_kernel);
10173 min_size = min_t(u32, krec_size, urec_size);
10175 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
10178 info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN);
10182 for (i = 0; i < nfuncs; i++) {
10183 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
10185 if (ret == -E2BIG) {
10186 verbose(env, "nonzero tailing record in func info");
10187 /* set the size kernel expects so loader can zero
10188 * out the rest of the record.
10190 if (copy_to_bpfptr_offset(uattr,
10191 offsetof(union bpf_attr, func_info_rec_size),
10192 &min_size, sizeof(min_size)))
10198 if (copy_from_bpfptr(&krecord[i], urecord, min_size)) {
10203 /* check insn_off */
10206 if (krecord[i].insn_off) {
10208 "nonzero insn_off %u for the first func info record",
10209 krecord[i].insn_off);
10212 } else if (krecord[i].insn_off <= prev_offset) {
10214 "same or smaller insn offset (%u) than previous func info record (%u)",
10215 krecord[i].insn_off, prev_offset);
10219 if (env->subprog_info[i].start != krecord[i].insn_off) {
10220 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
10224 /* check type_id */
10225 type = btf_type_by_id(btf, krecord[i].type_id);
10226 if (!type || !btf_type_is_func(type)) {
10227 verbose(env, "invalid type id %d in func info",
10228 krecord[i].type_id);
10231 info_aux[i].linkage = BTF_INFO_VLEN(type->info);
10233 func_proto = btf_type_by_id(btf, type->type);
10234 if (unlikely(!func_proto || !btf_type_is_func_proto(func_proto)))
10235 /* btf_func_check() already verified it during BTF load */
10237 ret_type = btf_type_skip_modifiers(btf, func_proto->type, NULL);
10239 btf_type_is_small_int(ret_type) || btf_type_is_enum(ret_type);
10240 if (i && !scalar_return && env->subprog_info[i].has_ld_abs) {
10241 verbose(env, "LD_ABS is only allowed in functions that return 'int'.\n");
10244 if (i && !scalar_return && env->subprog_info[i].has_tail_call) {
10245 verbose(env, "tail_call is only allowed in functions that return 'int'.\n");
10249 prev_offset = krecord[i].insn_off;
10250 bpfptr_add(&urecord, urec_size);
10253 prog->aux->func_info = krecord;
10254 prog->aux->func_info_cnt = nfuncs;
10255 prog->aux->func_info_aux = info_aux;
10264 static void adjust_btf_func(struct bpf_verifier_env *env)
10266 struct bpf_prog_aux *aux = env->prog->aux;
10269 if (!aux->func_info)
10272 for (i = 0; i < env->subprog_cnt; i++)
10273 aux->func_info[i].insn_off = env->subprog_info[i].start;
10276 #define MIN_BPF_LINEINFO_SIZE (offsetof(struct bpf_line_info, line_col) + \
10277 sizeof(((struct bpf_line_info *)(0))->line_col))
10278 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
10280 static int check_btf_line(struct bpf_verifier_env *env,
10281 const union bpf_attr *attr,
10284 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
10285 struct bpf_subprog_info *sub;
10286 struct bpf_line_info *linfo;
10287 struct bpf_prog *prog;
10288 const struct btf *btf;
10292 nr_linfo = attr->line_info_cnt;
10295 if (nr_linfo > INT_MAX / sizeof(struct bpf_line_info))
10298 rec_size = attr->line_info_rec_size;
10299 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
10300 rec_size > MAX_LINEINFO_REC_SIZE ||
10301 rec_size & (sizeof(u32) - 1))
10304 /* Need to zero it in case the userspace may
10305 * pass in a smaller bpf_line_info object.
10307 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
10308 GFP_KERNEL | __GFP_NOWARN);
10313 btf = prog->aux->btf;
10316 sub = env->subprog_info;
10317 ulinfo = make_bpfptr(attr->line_info, uattr.is_kernel);
10318 expected_size = sizeof(struct bpf_line_info);
10319 ncopy = min_t(u32, expected_size, rec_size);
10320 for (i = 0; i < nr_linfo; i++) {
10321 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
10323 if (err == -E2BIG) {
10324 verbose(env, "nonzero tailing record in line_info");
10325 if (copy_to_bpfptr_offset(uattr,
10326 offsetof(union bpf_attr, line_info_rec_size),
10327 &expected_size, sizeof(expected_size)))
10333 if (copy_from_bpfptr(&linfo[i], ulinfo, ncopy)) {
10339 * Check insn_off to ensure
10340 * 1) strictly increasing AND
10341 * 2) bounded by prog->len
10343 * The linfo[0].insn_off == 0 check logically falls into
10344 * the later "missing bpf_line_info for func..." case
10345 * because the first linfo[0].insn_off must be the
10346 * first sub also and the first sub must have
10347 * subprog_info[0].start == 0.
10349 if ((i && linfo[i].insn_off <= prev_offset) ||
10350 linfo[i].insn_off >= prog->len) {
10351 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
10352 i, linfo[i].insn_off, prev_offset,
10358 if (!prog->insnsi[linfo[i].insn_off].code) {
10360 "Invalid insn code at line_info[%u].insn_off\n",
10366 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
10367 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
10368 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
10373 if (s != env->subprog_cnt) {
10374 if (linfo[i].insn_off == sub[s].start) {
10375 sub[s].linfo_idx = i;
10377 } else if (sub[s].start < linfo[i].insn_off) {
10378 verbose(env, "missing bpf_line_info for func#%u\n", s);
10384 prev_offset = linfo[i].insn_off;
10385 bpfptr_add(&ulinfo, rec_size);
10388 if (s != env->subprog_cnt) {
10389 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
10390 env->subprog_cnt - s, s);
10395 prog->aux->linfo = linfo;
10396 prog->aux->nr_linfo = nr_linfo;
10405 #define MIN_CORE_RELO_SIZE sizeof(struct bpf_core_relo)
10406 #define MAX_CORE_RELO_SIZE MAX_FUNCINFO_REC_SIZE
10408 static int check_core_relo(struct bpf_verifier_env *env,
10409 const union bpf_attr *attr,
10412 u32 i, nr_core_relo, ncopy, expected_size, rec_size;
10413 struct bpf_core_relo core_relo = {};
10414 struct bpf_prog *prog = env->prog;
10415 const struct btf *btf = prog->aux->btf;
10416 struct bpf_core_ctx ctx = {
10420 bpfptr_t u_core_relo;
10423 nr_core_relo = attr->core_relo_cnt;
10426 if (nr_core_relo > INT_MAX / sizeof(struct bpf_core_relo))
10429 rec_size = attr->core_relo_rec_size;
10430 if (rec_size < MIN_CORE_RELO_SIZE ||
10431 rec_size > MAX_CORE_RELO_SIZE ||
10432 rec_size % sizeof(u32))
10435 u_core_relo = make_bpfptr(attr->core_relos, uattr.is_kernel);
10436 expected_size = sizeof(struct bpf_core_relo);
10437 ncopy = min_t(u32, expected_size, rec_size);
10439 /* Unlike func_info and line_info, copy and apply each CO-RE
10440 * relocation record one at a time.
10442 for (i = 0; i < nr_core_relo; i++) {
10443 /* future proofing when sizeof(bpf_core_relo) changes */
10444 err = bpf_check_uarg_tail_zero(u_core_relo, expected_size, rec_size);
10446 if (err == -E2BIG) {
10447 verbose(env, "nonzero tailing record in core_relo");
10448 if (copy_to_bpfptr_offset(uattr,
10449 offsetof(union bpf_attr, core_relo_rec_size),
10450 &expected_size, sizeof(expected_size)))
10456 if (copy_from_bpfptr(&core_relo, u_core_relo, ncopy)) {
10461 if (core_relo.insn_off % 8 || core_relo.insn_off / 8 >= prog->len) {
10462 verbose(env, "Invalid core_relo[%u].insn_off:%u prog->len:%u\n",
10463 i, core_relo.insn_off, prog->len);
10468 err = bpf_core_apply(&ctx, &core_relo, i,
10469 &prog->insnsi[core_relo.insn_off / 8]);
10472 bpfptr_add(&u_core_relo, rec_size);
10477 static int check_btf_info(struct bpf_verifier_env *env,
10478 const union bpf_attr *attr,
10484 if (!attr->func_info_cnt && !attr->line_info_cnt) {
10485 if (check_abnormal_return(env))
10490 btf = btf_get_by_fd(attr->prog_btf_fd);
10492 return PTR_ERR(btf);
10493 if (btf_is_kernel(btf)) {
10497 env->prog->aux->btf = btf;
10499 err = check_btf_func(env, attr, uattr);
10503 err = check_btf_line(env, attr, uattr);
10507 err = check_core_relo(env, attr, uattr);
10514 /* check %cur's range satisfies %old's */
10515 static bool range_within(struct bpf_reg_state *old,
10516 struct bpf_reg_state *cur)
10518 return old->umin_value <= cur->umin_value &&
10519 old->umax_value >= cur->umax_value &&
10520 old->smin_value <= cur->smin_value &&
10521 old->smax_value >= cur->smax_value &&
10522 old->u32_min_value <= cur->u32_min_value &&
10523 old->u32_max_value >= cur->u32_max_value &&
10524 old->s32_min_value <= cur->s32_min_value &&
10525 old->s32_max_value >= cur->s32_max_value;
10528 /* If in the old state two registers had the same id, then they need to have
10529 * the same id in the new state as well. But that id could be different from
10530 * the old state, so we need to track the mapping from old to new ids.
10531 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
10532 * regs with old id 5 must also have new id 9 for the new state to be safe. But
10533 * regs with a different old id could still have new id 9, we don't care about
10535 * So we look through our idmap to see if this old id has been seen before. If
10536 * so, we require the new id to match; otherwise, we add the id pair to the map.
10538 static bool check_ids(u32 old_id, u32 cur_id, struct bpf_id_pair *idmap)
10542 for (i = 0; i < BPF_ID_MAP_SIZE; i++) {
10543 if (!idmap[i].old) {
10544 /* Reached an empty slot; haven't seen this id before */
10545 idmap[i].old = old_id;
10546 idmap[i].cur = cur_id;
10549 if (idmap[i].old == old_id)
10550 return idmap[i].cur == cur_id;
10552 /* We ran out of idmap slots, which should be impossible */
10557 static void clean_func_state(struct bpf_verifier_env *env,
10558 struct bpf_func_state *st)
10560 enum bpf_reg_liveness live;
10563 for (i = 0; i < BPF_REG_FP; i++) {
10564 live = st->regs[i].live;
10565 /* liveness must not touch this register anymore */
10566 st->regs[i].live |= REG_LIVE_DONE;
10567 if (!(live & REG_LIVE_READ))
10568 /* since the register is unused, clear its state
10569 * to make further comparison simpler
10571 __mark_reg_not_init(env, &st->regs[i]);
10574 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
10575 live = st->stack[i].spilled_ptr.live;
10576 /* liveness must not touch this stack slot anymore */
10577 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
10578 if (!(live & REG_LIVE_READ)) {
10579 __mark_reg_not_init(env, &st->stack[i].spilled_ptr);
10580 for (j = 0; j < BPF_REG_SIZE; j++)
10581 st->stack[i].slot_type[j] = STACK_INVALID;
10586 static void clean_verifier_state(struct bpf_verifier_env *env,
10587 struct bpf_verifier_state *st)
10591 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
10592 /* all regs in this state in all frames were already marked */
10595 for (i = 0; i <= st->curframe; i++)
10596 clean_func_state(env, st->frame[i]);
10599 /* the parentage chains form a tree.
10600 * the verifier states are added to state lists at given insn and
10601 * pushed into state stack for future exploration.
10602 * when the verifier reaches bpf_exit insn some of the verifer states
10603 * stored in the state lists have their final liveness state already,
10604 * but a lot of states will get revised from liveness point of view when
10605 * the verifier explores other branches.
10608 * 2: if r1 == 100 goto pc+1
10611 * when the verifier reaches exit insn the register r0 in the state list of
10612 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
10613 * of insn 2 and goes exploring further. At the insn 4 it will walk the
10614 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
10616 * Since the verifier pushes the branch states as it sees them while exploring
10617 * the program the condition of walking the branch instruction for the second
10618 * time means that all states below this branch were already explored and
10619 * their final liveness marks are already propagated.
10620 * Hence when the verifier completes the search of state list in is_state_visited()
10621 * we can call this clean_live_states() function to mark all liveness states
10622 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
10623 * will not be used.
10624 * This function also clears the registers and stack for states that !READ
10625 * to simplify state merging.
10627 * Important note here that walking the same branch instruction in the callee
10628 * doesn't meant that the states are DONE. The verifier has to compare
10631 static void clean_live_states(struct bpf_verifier_env *env, int insn,
10632 struct bpf_verifier_state *cur)
10634 struct bpf_verifier_state_list *sl;
10637 sl = *explored_state(env, insn);
10639 if (sl->state.branches)
10641 if (sl->state.insn_idx != insn ||
10642 sl->state.curframe != cur->curframe)
10644 for (i = 0; i <= cur->curframe; i++)
10645 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
10647 clean_verifier_state(env, &sl->state);
10653 /* Returns true if (rold safe implies rcur safe) */
10654 static bool regsafe(struct bpf_verifier_env *env, struct bpf_reg_state *rold,
10655 struct bpf_reg_state *rcur, struct bpf_id_pair *idmap)
10659 if (!(rold->live & REG_LIVE_READ))
10660 /* explored state didn't use this */
10663 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
10665 if (rold->type == PTR_TO_STACK)
10666 /* two stack pointers are equal only if they're pointing to
10667 * the same stack frame, since fp-8 in foo != fp-8 in bar
10669 return equal && rold->frameno == rcur->frameno;
10674 if (rold->type == NOT_INIT)
10675 /* explored state can't have used this */
10677 if (rcur->type == NOT_INIT)
10679 switch (base_type(rold->type)) {
10681 if (env->explore_alu_limits)
10683 if (rcur->type == SCALAR_VALUE) {
10684 if (!rold->precise && !rcur->precise)
10686 /* new val must satisfy old val knowledge */
10687 return range_within(rold, rcur) &&
10688 tnum_in(rold->var_off, rcur->var_off);
10690 /* We're trying to use a pointer in place of a scalar.
10691 * Even if the scalar was unbounded, this could lead to
10692 * pointer leaks because scalars are allowed to leak
10693 * while pointers are not. We could make this safe in
10694 * special cases if root is calling us, but it's
10695 * probably not worth the hassle.
10699 case PTR_TO_MAP_KEY:
10700 case PTR_TO_MAP_VALUE:
10701 /* a PTR_TO_MAP_VALUE could be safe to use as a
10702 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
10703 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
10704 * checked, doing so could have affected others with the same
10705 * id, and we can't check for that because we lost the id when
10706 * we converted to a PTR_TO_MAP_VALUE.
10708 if (type_may_be_null(rold->type)) {
10709 if (!type_may_be_null(rcur->type))
10711 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
10713 /* Check our ids match any regs they're supposed to */
10714 return check_ids(rold->id, rcur->id, idmap);
10717 /* If the new min/max/var_off satisfy the old ones and
10718 * everything else matches, we are OK.
10719 * 'id' is not compared, since it's only used for maps with
10720 * bpf_spin_lock inside map element and in such cases if
10721 * the rest of the prog is valid for one map element then
10722 * it's valid for all map elements regardless of the key
10723 * used in bpf_map_lookup()
10725 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
10726 range_within(rold, rcur) &&
10727 tnum_in(rold->var_off, rcur->var_off);
10728 case PTR_TO_PACKET_META:
10729 case PTR_TO_PACKET:
10730 if (rcur->type != rold->type)
10732 /* We must have at least as much range as the old ptr
10733 * did, so that any accesses which were safe before are
10734 * still safe. This is true even if old range < old off,
10735 * since someone could have accessed through (ptr - k), or
10736 * even done ptr -= k in a register, to get a safe access.
10738 if (rold->range > rcur->range)
10740 /* If the offsets don't match, we can't trust our alignment;
10741 * nor can we be sure that we won't fall out of range.
10743 if (rold->off != rcur->off)
10745 /* id relations must be preserved */
10746 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
10748 /* new val must satisfy old val knowledge */
10749 return range_within(rold, rcur) &&
10750 tnum_in(rold->var_off, rcur->var_off);
10752 case CONST_PTR_TO_MAP:
10753 case PTR_TO_PACKET_END:
10754 case PTR_TO_FLOW_KEYS:
10755 case PTR_TO_SOCKET:
10756 case PTR_TO_SOCK_COMMON:
10757 case PTR_TO_TCP_SOCK:
10758 case PTR_TO_XDP_SOCK:
10759 /* Only valid matches are exact, which memcmp() above
10760 * would have accepted
10763 /* Don't know what's going on, just say it's not safe */
10767 /* Shouldn't get here; if we do, say it's not safe */
10772 static bool stacksafe(struct bpf_verifier_env *env, struct bpf_func_state *old,
10773 struct bpf_func_state *cur, struct bpf_id_pair *idmap)
10777 /* walk slots of the explored stack and ignore any additional
10778 * slots in the current stack, since explored(safe) state
10781 for (i = 0; i < old->allocated_stack; i++) {
10782 spi = i / BPF_REG_SIZE;
10784 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
10785 i += BPF_REG_SIZE - 1;
10786 /* explored state didn't use this */
10790 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
10793 /* explored stack has more populated slots than current stack
10794 * and these slots were used
10796 if (i >= cur->allocated_stack)
10799 /* if old state was safe with misc data in the stack
10800 * it will be safe with zero-initialized stack.
10801 * The opposite is not true
10803 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
10804 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
10806 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
10807 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
10808 /* Ex: old explored (safe) state has STACK_SPILL in
10809 * this stack slot, but current has STACK_MISC ->
10810 * this verifier states are not equivalent,
10811 * return false to continue verification of this path
10814 if (i % BPF_REG_SIZE != BPF_REG_SIZE - 1)
10816 if (!is_spilled_reg(&old->stack[spi]))
10818 if (!regsafe(env, &old->stack[spi].spilled_ptr,
10819 &cur->stack[spi].spilled_ptr, idmap))
10820 /* when explored and current stack slot are both storing
10821 * spilled registers, check that stored pointers types
10822 * are the same as well.
10823 * Ex: explored safe path could have stored
10824 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
10825 * but current path has stored:
10826 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
10827 * such verifier states are not equivalent.
10828 * return false to continue verification of this path
10835 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
10837 if (old->acquired_refs != cur->acquired_refs)
10839 return !memcmp(old->refs, cur->refs,
10840 sizeof(*old->refs) * old->acquired_refs);
10843 /* compare two verifier states
10845 * all states stored in state_list are known to be valid, since
10846 * verifier reached 'bpf_exit' instruction through them
10848 * this function is called when verifier exploring different branches of
10849 * execution popped from the state stack. If it sees an old state that has
10850 * more strict register state and more strict stack state then this execution
10851 * branch doesn't need to be explored further, since verifier already
10852 * concluded that more strict state leads to valid finish.
10854 * Therefore two states are equivalent if register state is more conservative
10855 * and explored stack state is more conservative than the current one.
10858 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
10859 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
10861 * In other words if current stack state (one being explored) has more
10862 * valid slots than old one that already passed validation, it means
10863 * the verifier can stop exploring and conclude that current state is valid too
10865 * Similarly with registers. If explored state has register type as invalid
10866 * whereas register type in current state is meaningful, it means that
10867 * the current state will reach 'bpf_exit' instruction safely
10869 static bool func_states_equal(struct bpf_verifier_env *env, struct bpf_func_state *old,
10870 struct bpf_func_state *cur)
10874 memset(env->idmap_scratch, 0, sizeof(env->idmap_scratch));
10875 for (i = 0; i < MAX_BPF_REG; i++)
10876 if (!regsafe(env, &old->regs[i], &cur->regs[i],
10877 env->idmap_scratch))
10880 if (!stacksafe(env, old, cur, env->idmap_scratch))
10883 if (!refsafe(old, cur))
10889 static bool states_equal(struct bpf_verifier_env *env,
10890 struct bpf_verifier_state *old,
10891 struct bpf_verifier_state *cur)
10895 if (old->curframe != cur->curframe)
10898 /* Verification state from speculative execution simulation
10899 * must never prune a non-speculative execution one.
10901 if (old->speculative && !cur->speculative)
10904 if (old->active_spin_lock != cur->active_spin_lock)
10907 /* for states to be equal callsites have to be the same
10908 * and all frame states need to be equivalent
10910 for (i = 0; i <= old->curframe; i++) {
10911 if (old->frame[i]->callsite != cur->frame[i]->callsite)
10913 if (!func_states_equal(env, old->frame[i], cur->frame[i]))
10919 /* Return 0 if no propagation happened. Return negative error code if error
10920 * happened. Otherwise, return the propagated bit.
10922 static int propagate_liveness_reg(struct bpf_verifier_env *env,
10923 struct bpf_reg_state *reg,
10924 struct bpf_reg_state *parent_reg)
10926 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
10927 u8 flag = reg->live & REG_LIVE_READ;
10930 /* When comes here, read flags of PARENT_REG or REG could be any of
10931 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
10932 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
10934 if (parent_flag == REG_LIVE_READ64 ||
10935 /* Or if there is no read flag from REG. */
10937 /* Or if the read flag from REG is the same as PARENT_REG. */
10938 parent_flag == flag)
10941 err = mark_reg_read(env, reg, parent_reg, flag);
10948 /* A write screens off any subsequent reads; but write marks come from the
10949 * straight-line code between a state and its parent. When we arrive at an
10950 * equivalent state (jump target or such) we didn't arrive by the straight-line
10951 * code, so read marks in the state must propagate to the parent regardless
10952 * of the state's write marks. That's what 'parent == state->parent' comparison
10953 * in mark_reg_read() is for.
10955 static int propagate_liveness(struct bpf_verifier_env *env,
10956 const struct bpf_verifier_state *vstate,
10957 struct bpf_verifier_state *vparent)
10959 struct bpf_reg_state *state_reg, *parent_reg;
10960 struct bpf_func_state *state, *parent;
10961 int i, frame, err = 0;
10963 if (vparent->curframe != vstate->curframe) {
10964 WARN(1, "propagate_live: parent frame %d current frame %d\n",
10965 vparent->curframe, vstate->curframe);
10968 /* Propagate read liveness of registers... */
10969 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
10970 for (frame = 0; frame <= vstate->curframe; frame++) {
10971 parent = vparent->frame[frame];
10972 state = vstate->frame[frame];
10973 parent_reg = parent->regs;
10974 state_reg = state->regs;
10975 /* We don't need to worry about FP liveness, it's read-only */
10976 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
10977 err = propagate_liveness_reg(env, &state_reg[i],
10981 if (err == REG_LIVE_READ64)
10982 mark_insn_zext(env, &parent_reg[i]);
10985 /* Propagate stack slots. */
10986 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
10987 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
10988 parent_reg = &parent->stack[i].spilled_ptr;
10989 state_reg = &state->stack[i].spilled_ptr;
10990 err = propagate_liveness_reg(env, state_reg,
10999 /* find precise scalars in the previous equivalent state and
11000 * propagate them into the current state
11002 static int propagate_precision(struct bpf_verifier_env *env,
11003 const struct bpf_verifier_state *old)
11005 struct bpf_reg_state *state_reg;
11006 struct bpf_func_state *state;
11009 state = old->frame[old->curframe];
11010 state_reg = state->regs;
11011 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
11012 if (state_reg->type != SCALAR_VALUE ||
11013 !state_reg->precise)
11015 if (env->log.level & BPF_LOG_LEVEL2)
11016 verbose(env, "propagating r%d\n", i);
11017 err = mark_chain_precision(env, i);
11022 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
11023 if (!is_spilled_reg(&state->stack[i]))
11025 state_reg = &state->stack[i].spilled_ptr;
11026 if (state_reg->type != SCALAR_VALUE ||
11027 !state_reg->precise)
11029 if (env->log.level & BPF_LOG_LEVEL2)
11030 verbose(env, "propagating fp%d\n",
11031 (-i - 1) * BPF_REG_SIZE);
11032 err = mark_chain_precision_stack(env, i);
11039 static bool states_maybe_looping(struct bpf_verifier_state *old,
11040 struct bpf_verifier_state *cur)
11042 struct bpf_func_state *fold, *fcur;
11043 int i, fr = cur->curframe;
11045 if (old->curframe != fr)
11048 fold = old->frame[fr];
11049 fcur = cur->frame[fr];
11050 for (i = 0; i < MAX_BPF_REG; i++)
11051 if (memcmp(&fold->regs[i], &fcur->regs[i],
11052 offsetof(struct bpf_reg_state, parent)))
11058 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
11060 struct bpf_verifier_state_list *new_sl;
11061 struct bpf_verifier_state_list *sl, **pprev;
11062 struct bpf_verifier_state *cur = env->cur_state, *new;
11063 int i, j, err, states_cnt = 0;
11064 bool add_new_state = env->test_state_freq ? true : false;
11066 cur->last_insn_idx = env->prev_insn_idx;
11067 if (!env->insn_aux_data[insn_idx].prune_point)
11068 /* this 'insn_idx' instruction wasn't marked, so we will not
11069 * be doing state search here
11073 /* bpf progs typically have pruning point every 4 instructions
11074 * http://vger.kernel.org/bpfconf2019.html#session-1
11075 * Do not add new state for future pruning if the verifier hasn't seen
11076 * at least 2 jumps and at least 8 instructions.
11077 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
11078 * In tests that amounts to up to 50% reduction into total verifier
11079 * memory consumption and 20% verifier time speedup.
11081 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
11082 env->insn_processed - env->prev_insn_processed >= 8)
11083 add_new_state = true;
11085 pprev = explored_state(env, insn_idx);
11088 clean_live_states(env, insn_idx, cur);
11092 if (sl->state.insn_idx != insn_idx)
11095 if (sl->state.branches) {
11096 struct bpf_func_state *frame = sl->state.frame[sl->state.curframe];
11098 if (frame->in_async_callback_fn &&
11099 frame->async_entry_cnt != cur->frame[cur->curframe]->async_entry_cnt) {
11100 /* Different async_entry_cnt means that the verifier is
11101 * processing another entry into async callback.
11102 * Seeing the same state is not an indication of infinite
11103 * loop or infinite recursion.
11104 * But finding the same state doesn't mean that it's safe
11105 * to stop processing the current state. The previous state
11106 * hasn't yet reached bpf_exit, since state.branches > 0.
11107 * Checking in_async_callback_fn alone is not enough either.
11108 * Since the verifier still needs to catch infinite loops
11109 * inside async callbacks.
11111 } else if (states_maybe_looping(&sl->state, cur) &&
11112 states_equal(env, &sl->state, cur)) {
11113 verbose_linfo(env, insn_idx, "; ");
11114 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
11117 /* if the verifier is processing a loop, avoid adding new state
11118 * too often, since different loop iterations have distinct
11119 * states and may not help future pruning.
11120 * This threshold shouldn't be too low to make sure that
11121 * a loop with large bound will be rejected quickly.
11122 * The most abusive loop will be:
11124 * if r1 < 1000000 goto pc-2
11125 * 1M insn_procssed limit / 100 == 10k peak states.
11126 * This threshold shouldn't be too high either, since states
11127 * at the end of the loop are likely to be useful in pruning.
11129 if (env->jmps_processed - env->prev_jmps_processed < 20 &&
11130 env->insn_processed - env->prev_insn_processed < 100)
11131 add_new_state = false;
11134 if (states_equal(env, &sl->state, cur)) {
11136 /* reached equivalent register/stack state,
11137 * prune the search.
11138 * Registers read by the continuation are read by us.
11139 * If we have any write marks in env->cur_state, they
11140 * will prevent corresponding reads in the continuation
11141 * from reaching our parent (an explored_state). Our
11142 * own state will get the read marks recorded, but
11143 * they'll be immediately forgotten as we're pruning
11144 * this state and will pop a new one.
11146 err = propagate_liveness(env, &sl->state, cur);
11148 /* if previous state reached the exit with precision and
11149 * current state is equivalent to it (except precsion marks)
11150 * the precision needs to be propagated back in
11151 * the current state.
11153 err = err ? : push_jmp_history(env, cur);
11154 err = err ? : propagate_precision(env, &sl->state);
11160 /* when new state is not going to be added do not increase miss count.
11161 * Otherwise several loop iterations will remove the state
11162 * recorded earlier. The goal of these heuristics is to have
11163 * states from some iterations of the loop (some in the beginning
11164 * and some at the end) to help pruning.
11168 /* heuristic to determine whether this state is beneficial
11169 * to keep checking from state equivalence point of view.
11170 * Higher numbers increase max_states_per_insn and verification time,
11171 * but do not meaningfully decrease insn_processed.
11173 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
11174 /* the state is unlikely to be useful. Remove it to
11175 * speed up verification
11178 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
11179 u32 br = sl->state.branches;
11182 "BUG live_done but branches_to_explore %d\n",
11184 free_verifier_state(&sl->state, false);
11186 env->peak_states--;
11188 /* cannot free this state, since parentage chain may
11189 * walk it later. Add it for free_list instead to
11190 * be freed at the end of verification
11192 sl->next = env->free_list;
11193 env->free_list = sl;
11203 if (env->max_states_per_insn < states_cnt)
11204 env->max_states_per_insn = states_cnt;
11206 if (!env->bpf_capable && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
11207 return push_jmp_history(env, cur);
11209 if (!add_new_state)
11210 return push_jmp_history(env, cur);
11212 /* There were no equivalent states, remember the current one.
11213 * Technically the current state is not proven to be safe yet,
11214 * but it will either reach outer most bpf_exit (which means it's safe)
11215 * or it will be rejected. When there are no loops the verifier won't be
11216 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
11217 * again on the way to bpf_exit.
11218 * When looping the sl->state.branches will be > 0 and this state
11219 * will not be considered for equivalence until branches == 0.
11221 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
11224 env->total_states++;
11225 env->peak_states++;
11226 env->prev_jmps_processed = env->jmps_processed;
11227 env->prev_insn_processed = env->insn_processed;
11229 /* add new state to the head of linked list */
11230 new = &new_sl->state;
11231 err = copy_verifier_state(new, cur);
11233 free_verifier_state(new, false);
11237 new->insn_idx = insn_idx;
11238 WARN_ONCE(new->branches != 1,
11239 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
11242 cur->first_insn_idx = insn_idx;
11243 clear_jmp_history(cur);
11244 new_sl->next = *explored_state(env, insn_idx);
11245 *explored_state(env, insn_idx) = new_sl;
11246 /* connect new state to parentage chain. Current frame needs all
11247 * registers connected. Only r6 - r9 of the callers are alive (pushed
11248 * to the stack implicitly by JITs) so in callers' frames connect just
11249 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
11250 * the state of the call instruction (with WRITTEN set), and r0 comes
11251 * from callee with its full parentage chain, anyway.
11253 /* clear write marks in current state: the writes we did are not writes
11254 * our child did, so they don't screen off its reads from us.
11255 * (There are no read marks in current state, because reads always mark
11256 * their parent and current state never has children yet. Only
11257 * explored_states can get read marks.)
11259 for (j = 0; j <= cur->curframe; j++) {
11260 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
11261 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
11262 for (i = 0; i < BPF_REG_FP; i++)
11263 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
11266 /* all stack frames are accessible from callee, clear them all */
11267 for (j = 0; j <= cur->curframe; j++) {
11268 struct bpf_func_state *frame = cur->frame[j];
11269 struct bpf_func_state *newframe = new->frame[j];
11271 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
11272 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
11273 frame->stack[i].spilled_ptr.parent =
11274 &newframe->stack[i].spilled_ptr;
11280 /* Return true if it's OK to have the same insn return a different type. */
11281 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
11283 switch (base_type(type)) {
11285 case PTR_TO_SOCKET:
11286 case PTR_TO_SOCK_COMMON:
11287 case PTR_TO_TCP_SOCK:
11288 case PTR_TO_XDP_SOCK:
11289 case PTR_TO_BTF_ID:
11296 /* If an instruction was previously used with particular pointer types, then we
11297 * need to be careful to avoid cases such as the below, where it may be ok
11298 * for one branch accessing the pointer, but not ok for the other branch:
11303 * R1 = some_other_valid_ptr;
11306 * R2 = *(u32 *)(R1 + 0);
11308 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
11310 return src != prev && (!reg_type_mismatch_ok(src) ||
11311 !reg_type_mismatch_ok(prev));
11314 static int do_check(struct bpf_verifier_env *env)
11316 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
11317 struct bpf_verifier_state *state = env->cur_state;
11318 struct bpf_insn *insns = env->prog->insnsi;
11319 struct bpf_reg_state *regs;
11320 int insn_cnt = env->prog->len;
11321 bool do_print_state = false;
11322 int prev_insn_idx = -1;
11325 struct bpf_insn *insn;
11329 env->prev_insn_idx = prev_insn_idx;
11330 if (env->insn_idx >= insn_cnt) {
11331 verbose(env, "invalid insn idx %d insn_cnt %d\n",
11332 env->insn_idx, insn_cnt);
11336 insn = &insns[env->insn_idx];
11337 class = BPF_CLASS(insn->code);
11339 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
11341 "BPF program is too large. Processed %d insn\n",
11342 env->insn_processed);
11346 err = is_state_visited(env, env->insn_idx);
11350 /* found equivalent state, can prune the search */
11351 if (env->log.level & BPF_LOG_LEVEL) {
11352 if (do_print_state)
11353 verbose(env, "\nfrom %d to %d%s: safe\n",
11354 env->prev_insn_idx, env->insn_idx,
11355 env->cur_state->speculative ?
11356 " (speculative execution)" : "");
11358 verbose(env, "%d: safe\n", env->insn_idx);
11360 goto process_bpf_exit;
11363 if (signal_pending(current))
11366 if (need_resched())
11369 if (env->log.level & BPF_LOG_LEVEL2 && do_print_state) {
11370 verbose(env, "\nfrom %d to %d%s:",
11371 env->prev_insn_idx, env->insn_idx,
11372 env->cur_state->speculative ?
11373 " (speculative execution)" : "");
11374 print_verifier_state(env, state->frame[state->curframe], true);
11375 do_print_state = false;
11378 if (env->log.level & BPF_LOG_LEVEL) {
11379 const struct bpf_insn_cbs cbs = {
11380 .cb_call = disasm_kfunc_name,
11381 .cb_print = verbose,
11382 .private_data = env,
11385 if (verifier_state_scratched(env))
11386 print_insn_state(env, state->frame[state->curframe]);
11388 verbose_linfo(env, env->insn_idx, "; ");
11389 env->prev_log_len = env->log.len_used;
11390 verbose(env, "%d: ", env->insn_idx);
11391 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
11392 env->prev_insn_print_len = env->log.len_used - env->prev_log_len;
11393 env->prev_log_len = env->log.len_used;
11396 if (bpf_prog_is_dev_bound(env->prog->aux)) {
11397 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
11398 env->prev_insn_idx);
11403 regs = cur_regs(env);
11404 sanitize_mark_insn_seen(env);
11405 prev_insn_idx = env->insn_idx;
11407 if (class == BPF_ALU || class == BPF_ALU64) {
11408 err = check_alu_op(env, insn);
11412 } else if (class == BPF_LDX) {
11413 enum bpf_reg_type *prev_src_type, src_reg_type;
11415 /* check for reserved fields is already done */
11417 /* check src operand */
11418 err = check_reg_arg(env, insn->src_reg, SRC_OP);
11422 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
11426 src_reg_type = regs[insn->src_reg].type;
11428 /* check that memory (src_reg + off) is readable,
11429 * the state of dst_reg will be updated by this func
11431 err = check_mem_access(env, env->insn_idx, insn->src_reg,
11432 insn->off, BPF_SIZE(insn->code),
11433 BPF_READ, insn->dst_reg, false);
11437 prev_src_type = &env->insn_aux_data[env->insn_idx].ptr_type;
11439 if (*prev_src_type == NOT_INIT) {
11440 /* saw a valid insn
11441 * dst_reg = *(u32 *)(src_reg + off)
11442 * save type to validate intersecting paths
11444 *prev_src_type = src_reg_type;
11446 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
11447 /* ABuser program is trying to use the same insn
11448 * dst_reg = *(u32*) (src_reg + off)
11449 * with different pointer types:
11450 * src_reg == ctx in one branch and
11451 * src_reg == stack|map in some other branch.
11454 verbose(env, "same insn cannot be used with different pointers\n");
11458 } else if (class == BPF_STX) {
11459 enum bpf_reg_type *prev_dst_type, dst_reg_type;
11461 if (BPF_MODE(insn->code) == BPF_ATOMIC) {
11462 err = check_atomic(env, env->insn_idx, insn);
11469 if (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0) {
11470 verbose(env, "BPF_STX uses reserved fields\n");
11474 /* check src1 operand */
11475 err = check_reg_arg(env, insn->src_reg, SRC_OP);
11478 /* check src2 operand */
11479 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
11483 dst_reg_type = regs[insn->dst_reg].type;
11485 /* check that memory (dst_reg + off) is writeable */
11486 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
11487 insn->off, BPF_SIZE(insn->code),
11488 BPF_WRITE, insn->src_reg, false);
11492 prev_dst_type = &env->insn_aux_data[env->insn_idx].ptr_type;
11494 if (*prev_dst_type == NOT_INIT) {
11495 *prev_dst_type = dst_reg_type;
11496 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
11497 verbose(env, "same insn cannot be used with different pointers\n");
11501 } else if (class == BPF_ST) {
11502 if (BPF_MODE(insn->code) != BPF_MEM ||
11503 insn->src_reg != BPF_REG_0) {
11504 verbose(env, "BPF_ST uses reserved fields\n");
11507 /* check src operand */
11508 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
11512 if (is_ctx_reg(env, insn->dst_reg)) {
11513 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
11515 reg_type_str(env, reg_state(env, insn->dst_reg)->type));
11519 /* check that memory (dst_reg + off) is writeable */
11520 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
11521 insn->off, BPF_SIZE(insn->code),
11522 BPF_WRITE, -1, false);
11526 } else if (class == BPF_JMP || class == BPF_JMP32) {
11527 u8 opcode = BPF_OP(insn->code);
11529 env->jmps_processed++;
11530 if (opcode == BPF_CALL) {
11531 if (BPF_SRC(insn->code) != BPF_K ||
11532 (insn->src_reg != BPF_PSEUDO_KFUNC_CALL
11533 && insn->off != 0) ||
11534 (insn->src_reg != BPF_REG_0 &&
11535 insn->src_reg != BPF_PSEUDO_CALL &&
11536 insn->src_reg != BPF_PSEUDO_KFUNC_CALL) ||
11537 insn->dst_reg != BPF_REG_0 ||
11538 class == BPF_JMP32) {
11539 verbose(env, "BPF_CALL uses reserved fields\n");
11543 if (env->cur_state->active_spin_lock &&
11544 (insn->src_reg == BPF_PSEUDO_CALL ||
11545 insn->imm != BPF_FUNC_spin_unlock)) {
11546 verbose(env, "function calls are not allowed while holding a lock\n");
11549 if (insn->src_reg == BPF_PSEUDO_CALL)
11550 err = check_func_call(env, insn, &env->insn_idx);
11551 else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL)
11552 err = check_kfunc_call(env, insn);
11554 err = check_helper_call(env, insn, &env->insn_idx);
11557 } else if (opcode == BPF_JA) {
11558 if (BPF_SRC(insn->code) != BPF_K ||
11560 insn->src_reg != BPF_REG_0 ||
11561 insn->dst_reg != BPF_REG_0 ||
11562 class == BPF_JMP32) {
11563 verbose(env, "BPF_JA uses reserved fields\n");
11567 env->insn_idx += insn->off + 1;
11570 } else if (opcode == BPF_EXIT) {
11571 if (BPF_SRC(insn->code) != BPF_K ||
11573 insn->src_reg != BPF_REG_0 ||
11574 insn->dst_reg != BPF_REG_0 ||
11575 class == BPF_JMP32) {
11576 verbose(env, "BPF_EXIT uses reserved fields\n");
11580 if (env->cur_state->active_spin_lock) {
11581 verbose(env, "bpf_spin_unlock is missing\n");
11585 if (state->curframe) {
11586 /* exit from nested function */
11587 err = prepare_func_exit(env, &env->insn_idx);
11590 do_print_state = true;
11594 err = check_reference_leak(env);
11598 err = check_return_code(env);
11602 mark_verifier_state_scratched(env);
11603 update_branch_counts(env, env->cur_state);
11604 err = pop_stack(env, &prev_insn_idx,
11605 &env->insn_idx, pop_log);
11607 if (err != -ENOENT)
11611 do_print_state = true;
11615 err = check_cond_jmp_op(env, insn, &env->insn_idx);
11619 } else if (class == BPF_LD) {
11620 u8 mode = BPF_MODE(insn->code);
11622 if (mode == BPF_ABS || mode == BPF_IND) {
11623 err = check_ld_abs(env, insn);
11627 } else if (mode == BPF_IMM) {
11628 err = check_ld_imm(env, insn);
11633 sanitize_mark_insn_seen(env);
11635 verbose(env, "invalid BPF_LD mode\n");
11639 verbose(env, "unknown insn class %d\n", class);
11649 static int find_btf_percpu_datasec(struct btf *btf)
11651 const struct btf_type *t;
11656 * Both vmlinux and module each have their own ".data..percpu"
11657 * DATASECs in BTF. So for module's case, we need to skip vmlinux BTF
11658 * types to look at only module's own BTF types.
11660 n = btf_nr_types(btf);
11661 if (btf_is_module(btf))
11662 i = btf_nr_types(btf_vmlinux);
11666 for(; i < n; i++) {
11667 t = btf_type_by_id(btf, i);
11668 if (BTF_INFO_KIND(t->info) != BTF_KIND_DATASEC)
11671 tname = btf_name_by_offset(btf, t->name_off);
11672 if (!strcmp(tname, ".data..percpu"))
11679 /* replace pseudo btf_id with kernel symbol address */
11680 static int check_pseudo_btf_id(struct bpf_verifier_env *env,
11681 struct bpf_insn *insn,
11682 struct bpf_insn_aux_data *aux)
11684 const struct btf_var_secinfo *vsi;
11685 const struct btf_type *datasec;
11686 struct btf_mod_pair *btf_mod;
11687 const struct btf_type *t;
11688 const char *sym_name;
11689 bool percpu = false;
11690 u32 type, id = insn->imm;
11694 int i, btf_fd, err;
11696 btf_fd = insn[1].imm;
11698 btf = btf_get_by_fd(btf_fd);
11700 verbose(env, "invalid module BTF object FD specified.\n");
11704 if (!btf_vmlinux) {
11705 verbose(env, "kernel is missing BTF, make sure CONFIG_DEBUG_INFO_BTF=y is specified in Kconfig.\n");
11712 t = btf_type_by_id(btf, id);
11714 verbose(env, "ldimm64 insn specifies invalid btf_id %d.\n", id);
11719 if (!btf_type_is_var(t)) {
11720 verbose(env, "pseudo btf_id %d in ldimm64 isn't KIND_VAR.\n", id);
11725 sym_name = btf_name_by_offset(btf, t->name_off);
11726 addr = kallsyms_lookup_name(sym_name);
11728 verbose(env, "ldimm64 failed to find the address for kernel symbol '%s'.\n",
11734 datasec_id = find_btf_percpu_datasec(btf);
11735 if (datasec_id > 0) {
11736 datasec = btf_type_by_id(btf, datasec_id);
11737 for_each_vsi(i, datasec, vsi) {
11738 if (vsi->type == id) {
11745 insn[0].imm = (u32)addr;
11746 insn[1].imm = addr >> 32;
11749 t = btf_type_skip_modifiers(btf, type, NULL);
11751 aux->btf_var.reg_type = PTR_TO_PERCPU_BTF_ID;
11752 aux->btf_var.btf = btf;
11753 aux->btf_var.btf_id = type;
11754 } else if (!btf_type_is_struct(t)) {
11755 const struct btf_type *ret;
11759 /* resolve the type size of ksym. */
11760 ret = btf_resolve_size(btf, t, &tsize);
11762 tname = btf_name_by_offset(btf, t->name_off);
11763 verbose(env, "ldimm64 unable to resolve the size of type '%s': %ld\n",
11764 tname, PTR_ERR(ret));
11768 aux->btf_var.reg_type = PTR_TO_MEM | MEM_RDONLY;
11769 aux->btf_var.mem_size = tsize;
11771 aux->btf_var.reg_type = PTR_TO_BTF_ID;
11772 aux->btf_var.btf = btf;
11773 aux->btf_var.btf_id = type;
11776 /* check whether we recorded this BTF (and maybe module) already */
11777 for (i = 0; i < env->used_btf_cnt; i++) {
11778 if (env->used_btfs[i].btf == btf) {
11784 if (env->used_btf_cnt >= MAX_USED_BTFS) {
11789 btf_mod = &env->used_btfs[env->used_btf_cnt];
11790 btf_mod->btf = btf;
11791 btf_mod->module = NULL;
11793 /* if we reference variables from kernel module, bump its refcount */
11794 if (btf_is_module(btf)) {
11795 btf_mod->module = btf_try_get_module(btf);
11796 if (!btf_mod->module) {
11802 env->used_btf_cnt++;
11810 static int check_map_prealloc(struct bpf_map *map)
11812 return (map->map_type != BPF_MAP_TYPE_HASH &&
11813 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
11814 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
11815 !(map->map_flags & BPF_F_NO_PREALLOC);
11818 static bool is_tracing_prog_type(enum bpf_prog_type type)
11821 case BPF_PROG_TYPE_KPROBE:
11822 case BPF_PROG_TYPE_TRACEPOINT:
11823 case BPF_PROG_TYPE_PERF_EVENT:
11824 case BPF_PROG_TYPE_RAW_TRACEPOINT:
11831 static bool is_preallocated_map(struct bpf_map *map)
11833 if (!check_map_prealloc(map))
11835 if (map->inner_map_meta && !check_map_prealloc(map->inner_map_meta))
11840 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
11841 struct bpf_map *map,
11842 struct bpf_prog *prog)
11845 enum bpf_prog_type prog_type = resolve_prog_type(prog);
11847 * Validate that trace type programs use preallocated hash maps.
11849 * For programs attached to PERF events this is mandatory as the
11850 * perf NMI can hit any arbitrary code sequence.
11852 * All other trace types using preallocated hash maps are unsafe as
11853 * well because tracepoint or kprobes can be inside locked regions
11854 * of the memory allocator or at a place where a recursion into the
11855 * memory allocator would see inconsistent state.
11857 * On RT enabled kernels run-time allocation of all trace type
11858 * programs is strictly prohibited due to lock type constraints. On
11859 * !RT kernels it is allowed for backwards compatibility reasons for
11860 * now, but warnings are emitted so developers are made aware of
11861 * the unsafety and can fix their programs before this is enforced.
11863 if (is_tracing_prog_type(prog_type) && !is_preallocated_map(map)) {
11864 if (prog_type == BPF_PROG_TYPE_PERF_EVENT) {
11865 verbose(env, "perf_event programs can only use preallocated hash map\n");
11868 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
11869 verbose(env, "trace type programs can only use preallocated hash map\n");
11872 WARN_ONCE(1, "trace type BPF program uses run-time allocation\n");
11873 verbose(env, "trace type programs with run-time allocated hash maps are unsafe. Switch to preallocated hash maps.\n");
11876 if (map_value_has_spin_lock(map)) {
11877 if (prog_type == BPF_PROG_TYPE_SOCKET_FILTER) {
11878 verbose(env, "socket filter progs cannot use bpf_spin_lock yet\n");
11882 if (is_tracing_prog_type(prog_type)) {
11883 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
11887 if (prog->aux->sleepable) {
11888 verbose(env, "sleepable progs cannot use bpf_spin_lock yet\n");
11893 if (map_value_has_timer(map)) {
11894 if (is_tracing_prog_type(prog_type)) {
11895 verbose(env, "tracing progs cannot use bpf_timer yet\n");
11900 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
11901 !bpf_offload_prog_map_match(prog, map)) {
11902 verbose(env, "offload device mismatch between prog and map\n");
11906 if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
11907 verbose(env, "bpf_struct_ops map cannot be used in prog\n");
11911 if (prog->aux->sleepable)
11912 switch (map->map_type) {
11913 case BPF_MAP_TYPE_HASH:
11914 case BPF_MAP_TYPE_LRU_HASH:
11915 case BPF_MAP_TYPE_ARRAY:
11916 case BPF_MAP_TYPE_PERCPU_HASH:
11917 case BPF_MAP_TYPE_PERCPU_ARRAY:
11918 case BPF_MAP_TYPE_LRU_PERCPU_HASH:
11919 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
11920 case BPF_MAP_TYPE_HASH_OF_MAPS:
11921 if (!is_preallocated_map(map)) {
11923 "Sleepable programs can only use preallocated maps\n");
11927 case BPF_MAP_TYPE_RINGBUF:
11928 case BPF_MAP_TYPE_INODE_STORAGE:
11929 case BPF_MAP_TYPE_SK_STORAGE:
11930 case BPF_MAP_TYPE_TASK_STORAGE:
11934 "Sleepable programs can only use array, hash, and ringbuf maps\n");
11941 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
11943 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
11944 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
11947 /* find and rewrite pseudo imm in ld_imm64 instructions:
11949 * 1. if it accesses map FD, replace it with actual map pointer.
11950 * 2. if it accesses btf_id of a VAR, replace it with pointer to the var.
11952 * NOTE: btf_vmlinux is required for converting pseudo btf_id.
11954 static int resolve_pseudo_ldimm64(struct bpf_verifier_env *env)
11956 struct bpf_insn *insn = env->prog->insnsi;
11957 int insn_cnt = env->prog->len;
11960 err = bpf_prog_calc_tag(env->prog);
11964 for (i = 0; i < insn_cnt; i++, insn++) {
11965 if (BPF_CLASS(insn->code) == BPF_LDX &&
11966 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
11967 verbose(env, "BPF_LDX uses reserved fields\n");
11971 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
11972 struct bpf_insn_aux_data *aux;
11973 struct bpf_map *map;
11978 if (i == insn_cnt - 1 || insn[1].code != 0 ||
11979 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
11980 insn[1].off != 0) {
11981 verbose(env, "invalid bpf_ld_imm64 insn\n");
11985 if (insn[0].src_reg == 0)
11986 /* valid generic load 64-bit imm */
11989 if (insn[0].src_reg == BPF_PSEUDO_BTF_ID) {
11990 aux = &env->insn_aux_data[i];
11991 err = check_pseudo_btf_id(env, insn, aux);
11997 if (insn[0].src_reg == BPF_PSEUDO_FUNC) {
11998 aux = &env->insn_aux_data[i];
11999 aux->ptr_type = PTR_TO_FUNC;
12003 /* In final convert_pseudo_ld_imm64() step, this is
12004 * converted into regular 64-bit imm load insn.
12006 switch (insn[0].src_reg) {
12007 case BPF_PSEUDO_MAP_VALUE:
12008 case BPF_PSEUDO_MAP_IDX_VALUE:
12010 case BPF_PSEUDO_MAP_FD:
12011 case BPF_PSEUDO_MAP_IDX:
12012 if (insn[1].imm == 0)
12016 verbose(env, "unrecognized bpf_ld_imm64 insn\n");
12020 switch (insn[0].src_reg) {
12021 case BPF_PSEUDO_MAP_IDX_VALUE:
12022 case BPF_PSEUDO_MAP_IDX:
12023 if (bpfptr_is_null(env->fd_array)) {
12024 verbose(env, "fd_idx without fd_array is invalid\n");
12027 if (copy_from_bpfptr_offset(&fd, env->fd_array,
12028 insn[0].imm * sizeof(fd),
12038 map = __bpf_map_get(f);
12040 verbose(env, "fd %d is not pointing to valid bpf_map\n",
12042 return PTR_ERR(map);
12045 err = check_map_prog_compatibility(env, map, env->prog);
12051 aux = &env->insn_aux_data[i];
12052 if (insn[0].src_reg == BPF_PSEUDO_MAP_FD ||
12053 insn[0].src_reg == BPF_PSEUDO_MAP_IDX) {
12054 addr = (unsigned long)map;
12056 u32 off = insn[1].imm;
12058 if (off >= BPF_MAX_VAR_OFF) {
12059 verbose(env, "direct value offset of %u is not allowed\n", off);
12064 if (!map->ops->map_direct_value_addr) {
12065 verbose(env, "no direct value access support for this map type\n");
12070 err = map->ops->map_direct_value_addr(map, &addr, off);
12072 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
12073 map->value_size, off);
12078 aux->map_off = off;
12082 insn[0].imm = (u32)addr;
12083 insn[1].imm = addr >> 32;
12085 /* check whether we recorded this map already */
12086 for (j = 0; j < env->used_map_cnt; j++) {
12087 if (env->used_maps[j] == map) {
12088 aux->map_index = j;
12094 if (env->used_map_cnt >= MAX_USED_MAPS) {
12099 /* hold the map. If the program is rejected by verifier,
12100 * the map will be released by release_maps() or it
12101 * will be used by the valid program until it's unloaded
12102 * and all maps are released in free_used_maps()
12106 aux->map_index = env->used_map_cnt;
12107 env->used_maps[env->used_map_cnt++] = map;
12109 if (bpf_map_is_cgroup_storage(map) &&
12110 bpf_cgroup_storage_assign(env->prog->aux, map)) {
12111 verbose(env, "only one cgroup storage of each type is allowed\n");
12123 /* Basic sanity check before we invest more work here. */
12124 if (!bpf_opcode_in_insntable(insn->code)) {
12125 verbose(env, "unknown opcode %02x\n", insn->code);
12130 /* now all pseudo BPF_LD_IMM64 instructions load valid
12131 * 'struct bpf_map *' into a register instead of user map_fd.
12132 * These pointers will be used later by verifier to validate map access.
12137 /* drop refcnt of maps used by the rejected program */
12138 static void release_maps(struct bpf_verifier_env *env)
12140 __bpf_free_used_maps(env->prog->aux, env->used_maps,
12141 env->used_map_cnt);
12144 /* drop refcnt of maps used by the rejected program */
12145 static void release_btfs(struct bpf_verifier_env *env)
12147 __bpf_free_used_btfs(env->prog->aux, env->used_btfs,
12148 env->used_btf_cnt);
12151 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
12152 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
12154 struct bpf_insn *insn = env->prog->insnsi;
12155 int insn_cnt = env->prog->len;
12158 for (i = 0; i < insn_cnt; i++, insn++) {
12159 if (insn->code != (BPF_LD | BPF_IMM | BPF_DW))
12161 if (insn->src_reg == BPF_PSEUDO_FUNC)
12167 /* single env->prog->insni[off] instruction was replaced with the range
12168 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
12169 * [0, off) and [off, end) to new locations, so the patched range stays zero
12171 static void adjust_insn_aux_data(struct bpf_verifier_env *env,
12172 struct bpf_insn_aux_data *new_data,
12173 struct bpf_prog *new_prog, u32 off, u32 cnt)
12175 struct bpf_insn_aux_data *old_data = env->insn_aux_data;
12176 struct bpf_insn *insn = new_prog->insnsi;
12177 u32 old_seen = old_data[off].seen;
12181 /* aux info at OFF always needs adjustment, no matter fast path
12182 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
12183 * original insn at old prog.
12185 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
12189 prog_len = new_prog->len;
12191 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
12192 memcpy(new_data + off + cnt - 1, old_data + off,
12193 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
12194 for (i = off; i < off + cnt - 1; i++) {
12195 /* Expand insni[off]'s seen count to the patched range. */
12196 new_data[i].seen = old_seen;
12197 new_data[i].zext_dst = insn_has_def32(env, insn + i);
12199 env->insn_aux_data = new_data;
12203 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
12209 /* NOTE: fake 'exit' subprog should be updated as well. */
12210 for (i = 0; i <= env->subprog_cnt; i++) {
12211 if (env->subprog_info[i].start <= off)
12213 env->subprog_info[i].start += len - 1;
12217 static void adjust_poke_descs(struct bpf_prog *prog, u32 off, u32 len)
12219 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
12220 int i, sz = prog->aux->size_poke_tab;
12221 struct bpf_jit_poke_descriptor *desc;
12223 for (i = 0; i < sz; i++) {
12225 if (desc->insn_idx <= off)
12227 desc->insn_idx += len - 1;
12231 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
12232 const struct bpf_insn *patch, u32 len)
12234 struct bpf_prog *new_prog;
12235 struct bpf_insn_aux_data *new_data = NULL;
12238 new_data = vzalloc(array_size(env->prog->len + len - 1,
12239 sizeof(struct bpf_insn_aux_data)));
12244 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
12245 if (IS_ERR(new_prog)) {
12246 if (PTR_ERR(new_prog) == -ERANGE)
12248 "insn %d cannot be patched due to 16-bit range\n",
12249 env->insn_aux_data[off].orig_idx);
12253 adjust_insn_aux_data(env, new_data, new_prog, off, len);
12254 adjust_subprog_starts(env, off, len);
12255 adjust_poke_descs(new_prog, off, len);
12259 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
12264 /* find first prog starting at or after off (first to remove) */
12265 for (i = 0; i < env->subprog_cnt; i++)
12266 if (env->subprog_info[i].start >= off)
12268 /* find first prog starting at or after off + cnt (first to stay) */
12269 for (j = i; j < env->subprog_cnt; j++)
12270 if (env->subprog_info[j].start >= off + cnt)
12272 /* if j doesn't start exactly at off + cnt, we are just removing
12273 * the front of previous prog
12275 if (env->subprog_info[j].start != off + cnt)
12279 struct bpf_prog_aux *aux = env->prog->aux;
12282 /* move fake 'exit' subprog as well */
12283 move = env->subprog_cnt + 1 - j;
12285 memmove(env->subprog_info + i,
12286 env->subprog_info + j,
12287 sizeof(*env->subprog_info) * move);
12288 env->subprog_cnt -= j - i;
12290 /* remove func_info */
12291 if (aux->func_info) {
12292 move = aux->func_info_cnt - j;
12294 memmove(aux->func_info + i,
12295 aux->func_info + j,
12296 sizeof(*aux->func_info) * move);
12297 aux->func_info_cnt -= j - i;
12298 /* func_info->insn_off is set after all code rewrites,
12299 * in adjust_btf_func() - no need to adjust
12303 /* convert i from "first prog to remove" to "first to adjust" */
12304 if (env->subprog_info[i].start == off)
12308 /* update fake 'exit' subprog as well */
12309 for (; i <= env->subprog_cnt; i++)
12310 env->subprog_info[i].start -= cnt;
12315 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
12318 struct bpf_prog *prog = env->prog;
12319 u32 i, l_off, l_cnt, nr_linfo;
12320 struct bpf_line_info *linfo;
12322 nr_linfo = prog->aux->nr_linfo;
12326 linfo = prog->aux->linfo;
12328 /* find first line info to remove, count lines to be removed */
12329 for (i = 0; i < nr_linfo; i++)
12330 if (linfo[i].insn_off >= off)
12335 for (; i < nr_linfo; i++)
12336 if (linfo[i].insn_off < off + cnt)
12341 /* First live insn doesn't match first live linfo, it needs to "inherit"
12342 * last removed linfo. prog is already modified, so prog->len == off
12343 * means no live instructions after (tail of the program was removed).
12345 if (prog->len != off && l_cnt &&
12346 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
12348 linfo[--i].insn_off = off + cnt;
12351 /* remove the line info which refer to the removed instructions */
12353 memmove(linfo + l_off, linfo + i,
12354 sizeof(*linfo) * (nr_linfo - i));
12356 prog->aux->nr_linfo -= l_cnt;
12357 nr_linfo = prog->aux->nr_linfo;
12360 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
12361 for (i = l_off; i < nr_linfo; i++)
12362 linfo[i].insn_off -= cnt;
12364 /* fix up all subprogs (incl. 'exit') which start >= off */
12365 for (i = 0; i <= env->subprog_cnt; i++)
12366 if (env->subprog_info[i].linfo_idx > l_off) {
12367 /* program may have started in the removed region but
12368 * may not be fully removed
12370 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
12371 env->subprog_info[i].linfo_idx -= l_cnt;
12373 env->subprog_info[i].linfo_idx = l_off;
12379 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
12381 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
12382 unsigned int orig_prog_len = env->prog->len;
12385 if (bpf_prog_is_dev_bound(env->prog->aux))
12386 bpf_prog_offload_remove_insns(env, off, cnt);
12388 err = bpf_remove_insns(env->prog, off, cnt);
12392 err = adjust_subprog_starts_after_remove(env, off, cnt);
12396 err = bpf_adj_linfo_after_remove(env, off, cnt);
12400 memmove(aux_data + off, aux_data + off + cnt,
12401 sizeof(*aux_data) * (orig_prog_len - off - cnt));
12406 /* The verifier does more data flow analysis than llvm and will not
12407 * explore branches that are dead at run time. Malicious programs can
12408 * have dead code too. Therefore replace all dead at-run-time code
12411 * Just nops are not optimal, e.g. if they would sit at the end of the
12412 * program and through another bug we would manage to jump there, then
12413 * we'd execute beyond program memory otherwise. Returning exception
12414 * code also wouldn't work since we can have subprogs where the dead
12415 * code could be located.
12417 static void sanitize_dead_code(struct bpf_verifier_env *env)
12419 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
12420 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
12421 struct bpf_insn *insn = env->prog->insnsi;
12422 const int insn_cnt = env->prog->len;
12425 for (i = 0; i < insn_cnt; i++) {
12426 if (aux_data[i].seen)
12428 memcpy(insn + i, &trap, sizeof(trap));
12429 aux_data[i].zext_dst = false;
12433 static bool insn_is_cond_jump(u8 code)
12437 if (BPF_CLASS(code) == BPF_JMP32)
12440 if (BPF_CLASS(code) != BPF_JMP)
12444 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
12447 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
12449 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
12450 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
12451 struct bpf_insn *insn = env->prog->insnsi;
12452 const int insn_cnt = env->prog->len;
12455 for (i = 0; i < insn_cnt; i++, insn++) {
12456 if (!insn_is_cond_jump(insn->code))
12459 if (!aux_data[i + 1].seen)
12460 ja.off = insn->off;
12461 else if (!aux_data[i + 1 + insn->off].seen)
12466 if (bpf_prog_is_dev_bound(env->prog->aux))
12467 bpf_prog_offload_replace_insn(env, i, &ja);
12469 memcpy(insn, &ja, sizeof(ja));
12473 static int opt_remove_dead_code(struct bpf_verifier_env *env)
12475 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
12476 int insn_cnt = env->prog->len;
12479 for (i = 0; i < insn_cnt; i++) {
12483 while (i + j < insn_cnt && !aux_data[i + j].seen)
12488 err = verifier_remove_insns(env, i, j);
12491 insn_cnt = env->prog->len;
12497 static int opt_remove_nops(struct bpf_verifier_env *env)
12499 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
12500 struct bpf_insn *insn = env->prog->insnsi;
12501 int insn_cnt = env->prog->len;
12504 for (i = 0; i < insn_cnt; i++) {
12505 if (memcmp(&insn[i], &ja, sizeof(ja)))
12508 err = verifier_remove_insns(env, i, 1);
12518 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
12519 const union bpf_attr *attr)
12521 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
12522 struct bpf_insn_aux_data *aux = env->insn_aux_data;
12523 int i, patch_len, delta = 0, len = env->prog->len;
12524 struct bpf_insn *insns = env->prog->insnsi;
12525 struct bpf_prog *new_prog;
12528 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
12529 zext_patch[1] = BPF_ZEXT_REG(0);
12530 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
12531 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
12532 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
12533 for (i = 0; i < len; i++) {
12534 int adj_idx = i + delta;
12535 struct bpf_insn insn;
12538 insn = insns[adj_idx];
12539 load_reg = insn_def_regno(&insn);
12540 if (!aux[adj_idx].zext_dst) {
12548 class = BPF_CLASS(code);
12549 if (load_reg == -1)
12552 /* NOTE: arg "reg" (the fourth one) is only used for
12553 * BPF_STX + SRC_OP, so it is safe to pass NULL
12556 if (is_reg64(env, &insn, load_reg, NULL, DST_OP)) {
12557 if (class == BPF_LD &&
12558 BPF_MODE(code) == BPF_IMM)
12563 /* ctx load could be transformed into wider load. */
12564 if (class == BPF_LDX &&
12565 aux[adj_idx].ptr_type == PTR_TO_CTX)
12568 imm_rnd = get_random_int();
12569 rnd_hi32_patch[0] = insn;
12570 rnd_hi32_patch[1].imm = imm_rnd;
12571 rnd_hi32_patch[3].dst_reg = load_reg;
12572 patch = rnd_hi32_patch;
12574 goto apply_patch_buffer;
12577 /* Add in an zero-extend instruction if a) the JIT has requested
12578 * it or b) it's a CMPXCHG.
12580 * The latter is because: BPF_CMPXCHG always loads a value into
12581 * R0, therefore always zero-extends. However some archs'
12582 * equivalent instruction only does this load when the
12583 * comparison is successful. This detail of CMPXCHG is
12584 * orthogonal to the general zero-extension behaviour of the
12585 * CPU, so it's treated independently of bpf_jit_needs_zext.
12587 if (!bpf_jit_needs_zext() && !is_cmpxchg_insn(&insn))
12590 if (WARN_ON(load_reg == -1)) {
12591 verbose(env, "verifier bug. zext_dst is set, but no reg is defined\n");
12595 zext_patch[0] = insn;
12596 zext_patch[1].dst_reg = load_reg;
12597 zext_patch[1].src_reg = load_reg;
12598 patch = zext_patch;
12600 apply_patch_buffer:
12601 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
12604 env->prog = new_prog;
12605 insns = new_prog->insnsi;
12606 aux = env->insn_aux_data;
12607 delta += patch_len - 1;
12613 /* convert load instructions that access fields of a context type into a
12614 * sequence of instructions that access fields of the underlying structure:
12615 * struct __sk_buff -> struct sk_buff
12616 * struct bpf_sock_ops -> struct sock
12618 static int convert_ctx_accesses(struct bpf_verifier_env *env)
12620 const struct bpf_verifier_ops *ops = env->ops;
12621 int i, cnt, size, ctx_field_size, delta = 0;
12622 const int insn_cnt = env->prog->len;
12623 struct bpf_insn insn_buf[16], *insn;
12624 u32 target_size, size_default, off;
12625 struct bpf_prog *new_prog;
12626 enum bpf_access_type type;
12627 bool is_narrower_load;
12629 if (ops->gen_prologue || env->seen_direct_write) {
12630 if (!ops->gen_prologue) {
12631 verbose(env, "bpf verifier is misconfigured\n");
12634 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
12636 if (cnt >= ARRAY_SIZE(insn_buf)) {
12637 verbose(env, "bpf verifier is misconfigured\n");
12640 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
12644 env->prog = new_prog;
12649 if (bpf_prog_is_dev_bound(env->prog->aux))
12652 insn = env->prog->insnsi + delta;
12654 for (i = 0; i < insn_cnt; i++, insn++) {
12655 bpf_convert_ctx_access_t convert_ctx_access;
12658 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
12659 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
12660 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
12661 insn->code == (BPF_LDX | BPF_MEM | BPF_DW)) {
12664 } else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
12665 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
12666 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
12667 insn->code == (BPF_STX | BPF_MEM | BPF_DW) ||
12668 insn->code == (BPF_ST | BPF_MEM | BPF_B) ||
12669 insn->code == (BPF_ST | BPF_MEM | BPF_H) ||
12670 insn->code == (BPF_ST | BPF_MEM | BPF_W) ||
12671 insn->code == (BPF_ST | BPF_MEM | BPF_DW)) {
12673 ctx_access = BPF_CLASS(insn->code) == BPF_STX;
12678 if (type == BPF_WRITE &&
12679 env->insn_aux_data[i + delta].sanitize_stack_spill) {
12680 struct bpf_insn patch[] = {
12685 cnt = ARRAY_SIZE(patch);
12686 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
12691 env->prog = new_prog;
12692 insn = new_prog->insnsi + i + delta;
12699 switch (env->insn_aux_data[i + delta].ptr_type) {
12701 if (!ops->convert_ctx_access)
12703 convert_ctx_access = ops->convert_ctx_access;
12705 case PTR_TO_SOCKET:
12706 case PTR_TO_SOCK_COMMON:
12707 convert_ctx_access = bpf_sock_convert_ctx_access;
12709 case PTR_TO_TCP_SOCK:
12710 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
12712 case PTR_TO_XDP_SOCK:
12713 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
12715 case PTR_TO_BTF_ID:
12716 if (type == BPF_READ) {
12717 insn->code = BPF_LDX | BPF_PROBE_MEM |
12718 BPF_SIZE((insn)->code);
12719 env->prog->aux->num_exentries++;
12720 } else if (resolve_prog_type(env->prog) != BPF_PROG_TYPE_STRUCT_OPS) {
12721 verbose(env, "Writes through BTF pointers are not allowed\n");
12729 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
12730 size = BPF_LDST_BYTES(insn);
12732 /* If the read access is a narrower load of the field,
12733 * convert to a 4/8-byte load, to minimum program type specific
12734 * convert_ctx_access changes. If conversion is successful,
12735 * we will apply proper mask to the result.
12737 is_narrower_load = size < ctx_field_size;
12738 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
12740 if (is_narrower_load) {
12743 if (type == BPF_WRITE) {
12744 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
12749 if (ctx_field_size == 4)
12751 else if (ctx_field_size == 8)
12752 size_code = BPF_DW;
12754 insn->off = off & ~(size_default - 1);
12755 insn->code = BPF_LDX | BPF_MEM | size_code;
12759 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
12761 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
12762 (ctx_field_size && !target_size)) {
12763 verbose(env, "bpf verifier is misconfigured\n");
12767 if (is_narrower_load && size < target_size) {
12768 u8 shift = bpf_ctx_narrow_access_offset(
12769 off, size, size_default) * 8;
12770 if (shift && cnt + 1 >= ARRAY_SIZE(insn_buf)) {
12771 verbose(env, "bpf verifier narrow ctx load misconfigured\n");
12774 if (ctx_field_size <= 4) {
12776 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
12779 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
12780 (1 << size * 8) - 1);
12783 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
12786 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
12787 (1ULL << size * 8) - 1);
12791 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12797 /* keep walking new program and skip insns we just inserted */
12798 env->prog = new_prog;
12799 insn = new_prog->insnsi + i + delta;
12805 static int jit_subprogs(struct bpf_verifier_env *env)
12807 struct bpf_prog *prog = env->prog, **func, *tmp;
12808 int i, j, subprog_start, subprog_end = 0, len, subprog;
12809 struct bpf_map *map_ptr;
12810 struct bpf_insn *insn;
12811 void *old_bpf_func;
12812 int err, num_exentries;
12814 if (env->subprog_cnt <= 1)
12817 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
12818 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn))
12821 /* Upon error here we cannot fall back to interpreter but
12822 * need a hard reject of the program. Thus -EFAULT is
12823 * propagated in any case.
12825 subprog = find_subprog(env, i + insn->imm + 1);
12827 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
12828 i + insn->imm + 1);
12831 /* temporarily remember subprog id inside insn instead of
12832 * aux_data, since next loop will split up all insns into funcs
12834 insn->off = subprog;
12835 /* remember original imm in case JIT fails and fallback
12836 * to interpreter will be needed
12838 env->insn_aux_data[i].call_imm = insn->imm;
12839 /* point imm to __bpf_call_base+1 from JITs point of view */
12841 if (bpf_pseudo_func(insn))
12842 /* jit (e.g. x86_64) may emit fewer instructions
12843 * if it learns a u32 imm is the same as a u64 imm.
12844 * Force a non zero here.
12849 err = bpf_prog_alloc_jited_linfo(prog);
12851 goto out_undo_insn;
12854 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
12856 goto out_undo_insn;
12858 for (i = 0; i < env->subprog_cnt; i++) {
12859 subprog_start = subprog_end;
12860 subprog_end = env->subprog_info[i + 1].start;
12862 len = subprog_end - subprog_start;
12863 /* bpf_prog_run() doesn't call subprogs directly,
12864 * hence main prog stats include the runtime of subprogs.
12865 * subprogs don't have IDs and not reachable via prog_get_next_id
12866 * func[i]->stats will never be accessed and stays NULL
12868 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
12871 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
12872 len * sizeof(struct bpf_insn));
12873 func[i]->type = prog->type;
12874 func[i]->len = len;
12875 if (bpf_prog_calc_tag(func[i]))
12877 func[i]->is_func = 1;
12878 func[i]->aux->func_idx = i;
12879 /* Below members will be freed only at prog->aux */
12880 func[i]->aux->btf = prog->aux->btf;
12881 func[i]->aux->func_info = prog->aux->func_info;
12882 func[i]->aux->poke_tab = prog->aux->poke_tab;
12883 func[i]->aux->size_poke_tab = prog->aux->size_poke_tab;
12885 for (j = 0; j < prog->aux->size_poke_tab; j++) {
12886 struct bpf_jit_poke_descriptor *poke;
12888 poke = &prog->aux->poke_tab[j];
12889 if (poke->insn_idx < subprog_end &&
12890 poke->insn_idx >= subprog_start)
12891 poke->aux = func[i]->aux;
12894 /* Use bpf_prog_F_tag to indicate functions in stack traces.
12895 * Long term would need debug info to populate names
12897 func[i]->aux->name[0] = 'F';
12898 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
12899 func[i]->jit_requested = 1;
12900 func[i]->aux->kfunc_tab = prog->aux->kfunc_tab;
12901 func[i]->aux->kfunc_btf_tab = prog->aux->kfunc_btf_tab;
12902 func[i]->aux->linfo = prog->aux->linfo;
12903 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
12904 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
12905 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
12907 insn = func[i]->insnsi;
12908 for (j = 0; j < func[i]->len; j++, insn++) {
12909 if (BPF_CLASS(insn->code) == BPF_LDX &&
12910 BPF_MODE(insn->code) == BPF_PROBE_MEM)
12913 func[i]->aux->num_exentries = num_exentries;
12914 func[i]->aux->tail_call_reachable = env->subprog_info[i].tail_call_reachable;
12915 func[i] = bpf_int_jit_compile(func[i]);
12916 if (!func[i]->jited) {
12923 /* at this point all bpf functions were successfully JITed
12924 * now populate all bpf_calls with correct addresses and
12925 * run last pass of JIT
12927 for (i = 0; i < env->subprog_cnt; i++) {
12928 insn = func[i]->insnsi;
12929 for (j = 0; j < func[i]->len; j++, insn++) {
12930 if (bpf_pseudo_func(insn)) {
12931 subprog = insn->off;
12932 insn[0].imm = (u32)(long)func[subprog]->bpf_func;
12933 insn[1].imm = ((u64)(long)func[subprog]->bpf_func) >> 32;
12936 if (!bpf_pseudo_call(insn))
12938 subprog = insn->off;
12939 insn->imm = BPF_CALL_IMM(func[subprog]->bpf_func);
12942 /* we use the aux data to keep a list of the start addresses
12943 * of the JITed images for each function in the program
12945 * for some architectures, such as powerpc64, the imm field
12946 * might not be large enough to hold the offset of the start
12947 * address of the callee's JITed image from __bpf_call_base
12949 * in such cases, we can lookup the start address of a callee
12950 * by using its subprog id, available from the off field of
12951 * the call instruction, as an index for this list
12953 func[i]->aux->func = func;
12954 func[i]->aux->func_cnt = env->subprog_cnt;
12956 for (i = 0; i < env->subprog_cnt; i++) {
12957 old_bpf_func = func[i]->bpf_func;
12958 tmp = bpf_int_jit_compile(func[i]);
12959 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
12960 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
12967 /* finally lock prog and jit images for all functions and
12968 * populate kallsysm
12970 for (i = 0; i < env->subprog_cnt; i++) {
12971 bpf_prog_lock_ro(func[i]);
12972 bpf_prog_kallsyms_add(func[i]);
12975 /* Last step: make now unused interpreter insns from main
12976 * prog consistent for later dump requests, so they can
12977 * later look the same as if they were interpreted only.
12979 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
12980 if (bpf_pseudo_func(insn)) {
12981 insn[0].imm = env->insn_aux_data[i].call_imm;
12982 insn[1].imm = insn->off;
12986 if (!bpf_pseudo_call(insn))
12988 insn->off = env->insn_aux_data[i].call_imm;
12989 subprog = find_subprog(env, i + insn->off + 1);
12990 insn->imm = subprog;
12994 prog->bpf_func = func[0]->bpf_func;
12995 prog->aux->func = func;
12996 prog->aux->func_cnt = env->subprog_cnt;
12997 bpf_prog_jit_attempt_done(prog);
13000 /* We failed JIT'ing, so at this point we need to unregister poke
13001 * descriptors from subprogs, so that kernel is not attempting to
13002 * patch it anymore as we're freeing the subprog JIT memory.
13004 for (i = 0; i < prog->aux->size_poke_tab; i++) {
13005 map_ptr = prog->aux->poke_tab[i].tail_call.map;
13006 map_ptr->ops->map_poke_untrack(map_ptr, prog->aux);
13008 /* At this point we're guaranteed that poke descriptors are not
13009 * live anymore. We can just unlink its descriptor table as it's
13010 * released with the main prog.
13012 for (i = 0; i < env->subprog_cnt; i++) {
13015 func[i]->aux->poke_tab = NULL;
13016 bpf_jit_free(func[i]);
13020 /* cleanup main prog to be interpreted */
13021 prog->jit_requested = 0;
13022 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
13023 if (!bpf_pseudo_call(insn))
13026 insn->imm = env->insn_aux_data[i].call_imm;
13028 bpf_prog_jit_attempt_done(prog);
13032 static int fixup_call_args(struct bpf_verifier_env *env)
13034 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
13035 struct bpf_prog *prog = env->prog;
13036 struct bpf_insn *insn = prog->insnsi;
13037 bool has_kfunc_call = bpf_prog_has_kfunc_call(prog);
13042 if (env->prog->jit_requested &&
13043 !bpf_prog_is_dev_bound(env->prog->aux)) {
13044 err = jit_subprogs(env);
13047 if (err == -EFAULT)
13050 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
13051 if (has_kfunc_call) {
13052 verbose(env, "calling kernel functions are not allowed in non-JITed programs\n");
13055 if (env->subprog_cnt > 1 && env->prog->aux->tail_call_reachable) {
13056 /* When JIT fails the progs with bpf2bpf calls and tail_calls
13057 * have to be rejected, since interpreter doesn't support them yet.
13059 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
13062 for (i = 0; i < prog->len; i++, insn++) {
13063 if (bpf_pseudo_func(insn)) {
13064 /* When JIT fails the progs with callback calls
13065 * have to be rejected, since interpreter doesn't support them yet.
13067 verbose(env, "callbacks are not allowed in non-JITed programs\n");
13071 if (!bpf_pseudo_call(insn))
13073 depth = get_callee_stack_depth(env, insn, i);
13076 bpf_patch_call_args(insn, depth);
13083 static int fixup_kfunc_call(struct bpf_verifier_env *env,
13084 struct bpf_insn *insn)
13086 const struct bpf_kfunc_desc *desc;
13089 verbose(env, "invalid kernel function call not eliminated in verifier pass\n");
13093 /* insn->imm has the btf func_id. Replace it with
13094 * an address (relative to __bpf_base_call).
13096 desc = find_kfunc_desc(env->prog, insn->imm, insn->off);
13098 verbose(env, "verifier internal error: kernel function descriptor not found for func_id %u\n",
13103 insn->imm = desc->imm;
13108 /* Do various post-verification rewrites in a single program pass.
13109 * These rewrites simplify JIT and interpreter implementations.
13111 static int do_misc_fixups(struct bpf_verifier_env *env)
13113 struct bpf_prog *prog = env->prog;
13114 enum bpf_attach_type eatype = prog->expected_attach_type;
13115 bool expect_blinding = bpf_jit_blinding_enabled(prog);
13116 enum bpf_prog_type prog_type = resolve_prog_type(prog);
13117 struct bpf_insn *insn = prog->insnsi;
13118 const struct bpf_func_proto *fn;
13119 const int insn_cnt = prog->len;
13120 const struct bpf_map_ops *ops;
13121 struct bpf_insn_aux_data *aux;
13122 struct bpf_insn insn_buf[16];
13123 struct bpf_prog *new_prog;
13124 struct bpf_map *map_ptr;
13125 int i, ret, cnt, delta = 0;
13127 for (i = 0; i < insn_cnt; i++, insn++) {
13128 /* Make divide-by-zero exceptions impossible. */
13129 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
13130 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
13131 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
13132 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
13133 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
13134 bool isdiv = BPF_OP(insn->code) == BPF_DIV;
13135 struct bpf_insn *patchlet;
13136 struct bpf_insn chk_and_div[] = {
13137 /* [R,W]x div 0 -> 0 */
13138 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
13139 BPF_JNE | BPF_K, insn->src_reg,
13141 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
13142 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
13145 struct bpf_insn chk_and_mod[] = {
13146 /* [R,W]x mod 0 -> [R,W]x */
13147 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
13148 BPF_JEQ | BPF_K, insn->src_reg,
13149 0, 1 + (is64 ? 0 : 1), 0),
13151 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
13152 BPF_MOV32_REG(insn->dst_reg, insn->dst_reg),
13155 patchlet = isdiv ? chk_and_div : chk_and_mod;
13156 cnt = isdiv ? ARRAY_SIZE(chk_and_div) :
13157 ARRAY_SIZE(chk_and_mod) - (is64 ? 2 : 0);
13159 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
13164 env->prog = prog = new_prog;
13165 insn = new_prog->insnsi + i + delta;
13169 /* Implement LD_ABS and LD_IND with a rewrite, if supported by the program type. */
13170 if (BPF_CLASS(insn->code) == BPF_LD &&
13171 (BPF_MODE(insn->code) == BPF_ABS ||
13172 BPF_MODE(insn->code) == BPF_IND)) {
13173 cnt = env->ops->gen_ld_abs(insn, insn_buf);
13174 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
13175 verbose(env, "bpf verifier is misconfigured\n");
13179 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
13184 env->prog = prog = new_prog;
13185 insn = new_prog->insnsi + i + delta;
13189 /* Rewrite pointer arithmetic to mitigate speculation attacks. */
13190 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
13191 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
13192 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
13193 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
13194 struct bpf_insn *patch = &insn_buf[0];
13195 bool issrc, isneg, isimm;
13198 aux = &env->insn_aux_data[i + delta];
13199 if (!aux->alu_state ||
13200 aux->alu_state == BPF_ALU_NON_POINTER)
13203 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
13204 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
13205 BPF_ALU_SANITIZE_SRC;
13206 isimm = aux->alu_state & BPF_ALU_IMMEDIATE;
13208 off_reg = issrc ? insn->src_reg : insn->dst_reg;
13210 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
13213 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
13214 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
13215 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
13216 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
13217 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
13218 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
13219 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX, off_reg);
13222 *patch++ = BPF_MOV64_REG(insn->dst_reg, insn->src_reg);
13223 insn->src_reg = BPF_REG_AX;
13225 insn->code = insn->code == code_add ?
13226 code_sub : code_add;
13228 if (issrc && isneg && !isimm)
13229 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
13230 cnt = patch - insn_buf;
13232 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
13237 env->prog = prog = new_prog;
13238 insn = new_prog->insnsi + i + delta;
13242 if (insn->code != (BPF_JMP | BPF_CALL))
13244 if (insn->src_reg == BPF_PSEUDO_CALL)
13246 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
13247 ret = fixup_kfunc_call(env, insn);
13253 if (insn->imm == BPF_FUNC_get_route_realm)
13254 prog->dst_needed = 1;
13255 if (insn->imm == BPF_FUNC_get_prandom_u32)
13256 bpf_user_rnd_init_once();
13257 if (insn->imm == BPF_FUNC_override_return)
13258 prog->kprobe_override = 1;
13259 if (insn->imm == BPF_FUNC_tail_call) {
13260 /* If we tail call into other programs, we
13261 * cannot make any assumptions since they can
13262 * be replaced dynamically during runtime in
13263 * the program array.
13265 prog->cb_access = 1;
13266 if (!allow_tail_call_in_subprogs(env))
13267 prog->aux->stack_depth = MAX_BPF_STACK;
13268 prog->aux->max_pkt_offset = MAX_PACKET_OFF;
13270 /* mark bpf_tail_call as different opcode to avoid
13271 * conditional branch in the interpreter for every normal
13272 * call and to prevent accidental JITing by JIT compiler
13273 * that doesn't support bpf_tail_call yet
13276 insn->code = BPF_JMP | BPF_TAIL_CALL;
13278 aux = &env->insn_aux_data[i + delta];
13279 if (env->bpf_capable && !expect_blinding &&
13280 prog->jit_requested &&
13281 !bpf_map_key_poisoned(aux) &&
13282 !bpf_map_ptr_poisoned(aux) &&
13283 !bpf_map_ptr_unpriv(aux)) {
13284 struct bpf_jit_poke_descriptor desc = {
13285 .reason = BPF_POKE_REASON_TAIL_CALL,
13286 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
13287 .tail_call.key = bpf_map_key_immediate(aux),
13288 .insn_idx = i + delta,
13291 ret = bpf_jit_add_poke_descriptor(prog, &desc);
13293 verbose(env, "adding tail call poke descriptor failed\n");
13297 insn->imm = ret + 1;
13301 if (!bpf_map_ptr_unpriv(aux))
13304 /* instead of changing every JIT dealing with tail_call
13305 * emit two extra insns:
13306 * if (index >= max_entries) goto out;
13307 * index &= array->index_mask;
13308 * to avoid out-of-bounds cpu speculation
13310 if (bpf_map_ptr_poisoned(aux)) {
13311 verbose(env, "tail_call abusing map_ptr\n");
13315 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
13316 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
13317 map_ptr->max_entries, 2);
13318 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
13319 container_of(map_ptr,
13322 insn_buf[2] = *insn;
13324 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
13329 env->prog = prog = new_prog;
13330 insn = new_prog->insnsi + i + delta;
13334 if (insn->imm == BPF_FUNC_timer_set_callback) {
13335 /* The verifier will process callback_fn as many times as necessary
13336 * with different maps and the register states prepared by
13337 * set_timer_callback_state will be accurate.
13339 * The following use case is valid:
13340 * map1 is shared by prog1, prog2, prog3.
13341 * prog1 calls bpf_timer_init for some map1 elements
13342 * prog2 calls bpf_timer_set_callback for some map1 elements.
13343 * Those that were not bpf_timer_init-ed will return -EINVAL.
13344 * prog3 calls bpf_timer_start for some map1 elements.
13345 * Those that were not both bpf_timer_init-ed and
13346 * bpf_timer_set_callback-ed will return -EINVAL.
13348 struct bpf_insn ld_addrs[2] = {
13349 BPF_LD_IMM64(BPF_REG_3, (long)prog->aux),
13352 insn_buf[0] = ld_addrs[0];
13353 insn_buf[1] = ld_addrs[1];
13354 insn_buf[2] = *insn;
13357 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
13362 env->prog = prog = new_prog;
13363 insn = new_prog->insnsi + i + delta;
13364 goto patch_call_imm;
13367 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
13368 * and other inlining handlers are currently limited to 64 bit
13371 if (prog->jit_requested && BITS_PER_LONG == 64 &&
13372 (insn->imm == BPF_FUNC_map_lookup_elem ||
13373 insn->imm == BPF_FUNC_map_update_elem ||
13374 insn->imm == BPF_FUNC_map_delete_elem ||
13375 insn->imm == BPF_FUNC_map_push_elem ||
13376 insn->imm == BPF_FUNC_map_pop_elem ||
13377 insn->imm == BPF_FUNC_map_peek_elem ||
13378 insn->imm == BPF_FUNC_redirect_map ||
13379 insn->imm == BPF_FUNC_for_each_map_elem)) {
13380 aux = &env->insn_aux_data[i + delta];
13381 if (bpf_map_ptr_poisoned(aux))
13382 goto patch_call_imm;
13384 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
13385 ops = map_ptr->ops;
13386 if (insn->imm == BPF_FUNC_map_lookup_elem &&
13387 ops->map_gen_lookup) {
13388 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
13389 if (cnt == -EOPNOTSUPP)
13390 goto patch_map_ops_generic;
13391 if (cnt <= 0 || cnt >= ARRAY_SIZE(insn_buf)) {
13392 verbose(env, "bpf verifier is misconfigured\n");
13396 new_prog = bpf_patch_insn_data(env, i + delta,
13402 env->prog = prog = new_prog;
13403 insn = new_prog->insnsi + i + delta;
13407 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
13408 (void *(*)(struct bpf_map *map, void *key))NULL));
13409 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
13410 (int (*)(struct bpf_map *map, void *key))NULL));
13411 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
13412 (int (*)(struct bpf_map *map, void *key, void *value,
13414 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
13415 (int (*)(struct bpf_map *map, void *value,
13417 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
13418 (int (*)(struct bpf_map *map, void *value))NULL));
13419 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
13420 (int (*)(struct bpf_map *map, void *value))NULL));
13421 BUILD_BUG_ON(!__same_type(ops->map_redirect,
13422 (int (*)(struct bpf_map *map, u32 ifindex, u64 flags))NULL));
13423 BUILD_BUG_ON(!__same_type(ops->map_for_each_callback,
13424 (int (*)(struct bpf_map *map,
13425 bpf_callback_t callback_fn,
13426 void *callback_ctx,
13429 patch_map_ops_generic:
13430 switch (insn->imm) {
13431 case BPF_FUNC_map_lookup_elem:
13432 insn->imm = BPF_CALL_IMM(ops->map_lookup_elem);
13434 case BPF_FUNC_map_update_elem:
13435 insn->imm = BPF_CALL_IMM(ops->map_update_elem);
13437 case BPF_FUNC_map_delete_elem:
13438 insn->imm = BPF_CALL_IMM(ops->map_delete_elem);
13440 case BPF_FUNC_map_push_elem:
13441 insn->imm = BPF_CALL_IMM(ops->map_push_elem);
13443 case BPF_FUNC_map_pop_elem:
13444 insn->imm = BPF_CALL_IMM(ops->map_pop_elem);
13446 case BPF_FUNC_map_peek_elem:
13447 insn->imm = BPF_CALL_IMM(ops->map_peek_elem);
13449 case BPF_FUNC_redirect_map:
13450 insn->imm = BPF_CALL_IMM(ops->map_redirect);
13452 case BPF_FUNC_for_each_map_elem:
13453 insn->imm = BPF_CALL_IMM(ops->map_for_each_callback);
13457 goto patch_call_imm;
13460 /* Implement bpf_jiffies64 inline. */
13461 if (prog->jit_requested && BITS_PER_LONG == 64 &&
13462 insn->imm == BPF_FUNC_jiffies64) {
13463 struct bpf_insn ld_jiffies_addr[2] = {
13464 BPF_LD_IMM64(BPF_REG_0,
13465 (unsigned long)&jiffies),
13468 insn_buf[0] = ld_jiffies_addr[0];
13469 insn_buf[1] = ld_jiffies_addr[1];
13470 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
13474 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
13480 env->prog = prog = new_prog;
13481 insn = new_prog->insnsi + i + delta;
13485 /* Implement bpf_get_func_arg inline. */
13486 if (prog_type == BPF_PROG_TYPE_TRACING &&
13487 insn->imm == BPF_FUNC_get_func_arg) {
13488 /* Load nr_args from ctx - 8 */
13489 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
13490 insn_buf[1] = BPF_JMP32_REG(BPF_JGE, BPF_REG_2, BPF_REG_0, 6);
13491 insn_buf[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_2, 3);
13492 insn_buf[3] = BPF_ALU64_REG(BPF_ADD, BPF_REG_2, BPF_REG_1);
13493 insn_buf[4] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_2, 0);
13494 insn_buf[5] = BPF_STX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
13495 insn_buf[6] = BPF_MOV64_IMM(BPF_REG_0, 0);
13496 insn_buf[7] = BPF_JMP_A(1);
13497 insn_buf[8] = BPF_MOV64_IMM(BPF_REG_0, -EINVAL);
13500 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
13505 env->prog = prog = new_prog;
13506 insn = new_prog->insnsi + i + delta;
13510 /* Implement bpf_get_func_ret inline. */
13511 if (prog_type == BPF_PROG_TYPE_TRACING &&
13512 insn->imm == BPF_FUNC_get_func_ret) {
13513 if (eatype == BPF_TRACE_FEXIT ||
13514 eatype == BPF_MODIFY_RETURN) {
13515 /* Load nr_args from ctx - 8 */
13516 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
13517 insn_buf[1] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_0, 3);
13518 insn_buf[2] = BPF_ALU64_REG(BPF_ADD, BPF_REG_0, BPF_REG_1);
13519 insn_buf[3] = BPF_LDX_MEM(BPF_DW, BPF_REG_3, BPF_REG_0, 0);
13520 insn_buf[4] = BPF_STX_MEM(BPF_DW, BPF_REG_2, BPF_REG_3, 0);
13521 insn_buf[5] = BPF_MOV64_IMM(BPF_REG_0, 0);
13524 insn_buf[0] = BPF_MOV64_IMM(BPF_REG_0, -EOPNOTSUPP);
13528 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
13533 env->prog = prog = new_prog;
13534 insn = new_prog->insnsi + i + delta;
13538 /* Implement get_func_arg_cnt inline. */
13539 if (prog_type == BPF_PROG_TYPE_TRACING &&
13540 insn->imm == BPF_FUNC_get_func_arg_cnt) {
13541 /* Load nr_args from ctx - 8 */
13542 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
13544 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
13548 env->prog = prog = new_prog;
13549 insn = new_prog->insnsi + i + delta;
13553 /* Implement bpf_get_func_ip inline. */
13554 if (prog_type == BPF_PROG_TYPE_TRACING &&
13555 insn->imm == BPF_FUNC_get_func_ip) {
13556 /* Load IP address from ctx - 16 */
13557 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -16);
13559 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
13563 env->prog = prog = new_prog;
13564 insn = new_prog->insnsi + i + delta;
13569 fn = env->ops->get_func_proto(insn->imm, env->prog);
13570 /* all functions that have prototype and verifier allowed
13571 * programs to call them, must be real in-kernel functions
13575 "kernel subsystem misconfigured func %s#%d\n",
13576 func_id_name(insn->imm), insn->imm);
13579 insn->imm = fn->func - __bpf_call_base;
13582 /* Since poke tab is now finalized, publish aux to tracker. */
13583 for (i = 0; i < prog->aux->size_poke_tab; i++) {
13584 map_ptr = prog->aux->poke_tab[i].tail_call.map;
13585 if (!map_ptr->ops->map_poke_track ||
13586 !map_ptr->ops->map_poke_untrack ||
13587 !map_ptr->ops->map_poke_run) {
13588 verbose(env, "bpf verifier is misconfigured\n");
13592 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
13594 verbose(env, "tracking tail call prog failed\n");
13599 sort_kfunc_descs_by_imm(env->prog);
13604 static void free_states(struct bpf_verifier_env *env)
13606 struct bpf_verifier_state_list *sl, *sln;
13609 sl = env->free_list;
13612 free_verifier_state(&sl->state, false);
13616 env->free_list = NULL;
13618 if (!env->explored_states)
13621 for (i = 0; i < state_htab_size(env); i++) {
13622 sl = env->explored_states[i];
13626 free_verifier_state(&sl->state, false);
13630 env->explored_states[i] = NULL;
13634 static int do_check_common(struct bpf_verifier_env *env, int subprog)
13636 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
13637 struct bpf_verifier_state *state;
13638 struct bpf_reg_state *regs;
13641 env->prev_linfo = NULL;
13644 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
13647 state->curframe = 0;
13648 state->speculative = false;
13649 state->branches = 1;
13650 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
13651 if (!state->frame[0]) {
13655 env->cur_state = state;
13656 init_func_state(env, state->frame[0],
13657 BPF_MAIN_FUNC /* callsite */,
13661 regs = state->frame[state->curframe]->regs;
13662 if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
13663 ret = btf_prepare_func_args(env, subprog, regs);
13666 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
13667 if (regs[i].type == PTR_TO_CTX)
13668 mark_reg_known_zero(env, regs, i);
13669 else if (regs[i].type == SCALAR_VALUE)
13670 mark_reg_unknown(env, regs, i);
13671 else if (base_type(regs[i].type) == PTR_TO_MEM) {
13672 const u32 mem_size = regs[i].mem_size;
13674 mark_reg_known_zero(env, regs, i);
13675 regs[i].mem_size = mem_size;
13676 regs[i].id = ++env->id_gen;
13680 /* 1st arg to a function */
13681 regs[BPF_REG_1].type = PTR_TO_CTX;
13682 mark_reg_known_zero(env, regs, BPF_REG_1);
13683 ret = btf_check_subprog_arg_match(env, subprog, regs);
13684 if (ret == -EFAULT)
13685 /* unlikely verifier bug. abort.
13686 * ret == 0 and ret < 0 are sadly acceptable for
13687 * main() function due to backward compatibility.
13688 * Like socket filter program may be written as:
13689 * int bpf_prog(struct pt_regs *ctx)
13690 * and never dereference that ctx in the program.
13691 * 'struct pt_regs' is a type mismatch for socket
13692 * filter that should be using 'struct __sk_buff'.
13697 ret = do_check(env);
13699 /* check for NULL is necessary, since cur_state can be freed inside
13700 * do_check() under memory pressure.
13702 if (env->cur_state) {
13703 free_verifier_state(env->cur_state, true);
13704 env->cur_state = NULL;
13706 while (!pop_stack(env, NULL, NULL, false));
13707 if (!ret && pop_log)
13708 bpf_vlog_reset(&env->log, 0);
13713 /* Verify all global functions in a BPF program one by one based on their BTF.
13714 * All global functions must pass verification. Otherwise the whole program is rejected.
13725 * foo() will be verified first for R1=any_scalar_value. During verification it
13726 * will be assumed that bar() already verified successfully and call to bar()
13727 * from foo() will be checked for type match only. Later bar() will be verified
13728 * independently to check that it's safe for R1=any_scalar_value.
13730 static int do_check_subprogs(struct bpf_verifier_env *env)
13732 struct bpf_prog_aux *aux = env->prog->aux;
13735 if (!aux->func_info)
13738 for (i = 1; i < env->subprog_cnt; i++) {
13739 if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL)
13741 env->insn_idx = env->subprog_info[i].start;
13742 WARN_ON_ONCE(env->insn_idx == 0);
13743 ret = do_check_common(env, i);
13746 } else if (env->log.level & BPF_LOG_LEVEL) {
13748 "Func#%d is safe for any args that match its prototype\n",
13755 static int do_check_main(struct bpf_verifier_env *env)
13760 ret = do_check_common(env, 0);
13762 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
13767 static void print_verification_stats(struct bpf_verifier_env *env)
13771 if (env->log.level & BPF_LOG_STATS) {
13772 verbose(env, "verification time %lld usec\n",
13773 div_u64(env->verification_time, 1000));
13774 verbose(env, "stack depth ");
13775 for (i = 0; i < env->subprog_cnt; i++) {
13776 u32 depth = env->subprog_info[i].stack_depth;
13778 verbose(env, "%d", depth);
13779 if (i + 1 < env->subprog_cnt)
13782 verbose(env, "\n");
13784 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
13785 "total_states %d peak_states %d mark_read %d\n",
13786 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
13787 env->max_states_per_insn, env->total_states,
13788 env->peak_states, env->longest_mark_read_walk);
13791 static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
13793 const struct btf_type *t, *func_proto;
13794 const struct bpf_struct_ops *st_ops;
13795 const struct btf_member *member;
13796 struct bpf_prog *prog = env->prog;
13797 u32 btf_id, member_idx;
13800 if (!prog->gpl_compatible) {
13801 verbose(env, "struct ops programs must have a GPL compatible license\n");
13805 btf_id = prog->aux->attach_btf_id;
13806 st_ops = bpf_struct_ops_find(btf_id);
13808 verbose(env, "attach_btf_id %u is not a supported struct\n",
13814 member_idx = prog->expected_attach_type;
13815 if (member_idx >= btf_type_vlen(t)) {
13816 verbose(env, "attach to invalid member idx %u of struct %s\n",
13817 member_idx, st_ops->name);
13821 member = &btf_type_member(t)[member_idx];
13822 mname = btf_name_by_offset(btf_vmlinux, member->name_off);
13823 func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type,
13826 verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
13827 mname, member_idx, st_ops->name);
13831 if (st_ops->check_member) {
13832 int err = st_ops->check_member(t, member);
13835 verbose(env, "attach to unsupported member %s of struct %s\n",
13836 mname, st_ops->name);
13841 prog->aux->attach_func_proto = func_proto;
13842 prog->aux->attach_func_name = mname;
13843 env->ops = st_ops->verifier_ops;
13847 #define SECURITY_PREFIX "security_"
13849 static int check_attach_modify_return(unsigned long addr, const char *func_name)
13851 if (within_error_injection_list(addr) ||
13852 !strncmp(SECURITY_PREFIX, func_name, sizeof(SECURITY_PREFIX) - 1))
13858 /* list of non-sleepable functions that are otherwise on
13859 * ALLOW_ERROR_INJECTION list
13861 BTF_SET_START(btf_non_sleepable_error_inject)
13862 /* Three functions below can be called from sleepable and non-sleepable context.
13863 * Assume non-sleepable from bpf safety point of view.
13865 BTF_ID(func, __filemap_add_folio)
13866 BTF_ID(func, should_fail_alloc_page)
13867 BTF_ID(func, should_failslab)
13868 BTF_SET_END(btf_non_sleepable_error_inject)
13870 static int check_non_sleepable_error_inject(u32 btf_id)
13872 return btf_id_set_contains(&btf_non_sleepable_error_inject, btf_id);
13875 int bpf_check_attach_target(struct bpf_verifier_log *log,
13876 const struct bpf_prog *prog,
13877 const struct bpf_prog *tgt_prog,
13879 struct bpf_attach_target_info *tgt_info)
13881 bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
13882 const char prefix[] = "btf_trace_";
13883 int ret = 0, subprog = -1, i;
13884 const struct btf_type *t;
13885 bool conservative = true;
13891 bpf_log(log, "Tracing programs must provide btf_id\n");
13894 btf = tgt_prog ? tgt_prog->aux->btf : prog->aux->attach_btf;
13897 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
13900 t = btf_type_by_id(btf, btf_id);
13902 bpf_log(log, "attach_btf_id %u is invalid\n", btf_id);
13905 tname = btf_name_by_offset(btf, t->name_off);
13907 bpf_log(log, "attach_btf_id %u doesn't have a name\n", btf_id);
13911 struct bpf_prog_aux *aux = tgt_prog->aux;
13913 for (i = 0; i < aux->func_info_cnt; i++)
13914 if (aux->func_info[i].type_id == btf_id) {
13918 if (subprog == -1) {
13919 bpf_log(log, "Subprog %s doesn't exist\n", tname);
13922 conservative = aux->func_info_aux[subprog].unreliable;
13923 if (prog_extension) {
13924 if (conservative) {
13926 "Cannot replace static functions\n");
13929 if (!prog->jit_requested) {
13931 "Extension programs should be JITed\n");
13935 if (!tgt_prog->jited) {
13936 bpf_log(log, "Can attach to only JITed progs\n");
13939 if (tgt_prog->type == prog->type) {
13940 /* Cannot fentry/fexit another fentry/fexit program.
13941 * Cannot attach program extension to another extension.
13942 * It's ok to attach fentry/fexit to extension program.
13944 bpf_log(log, "Cannot recursively attach\n");
13947 if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
13949 (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
13950 tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
13951 /* Program extensions can extend all program types
13952 * except fentry/fexit. The reason is the following.
13953 * The fentry/fexit programs are used for performance
13954 * analysis, stats and can be attached to any program
13955 * type except themselves. When extension program is
13956 * replacing XDP function it is necessary to allow
13957 * performance analysis of all functions. Both original
13958 * XDP program and its program extension. Hence
13959 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
13960 * allowed. If extending of fentry/fexit was allowed it
13961 * would be possible to create long call chain
13962 * fentry->extension->fentry->extension beyond
13963 * reasonable stack size. Hence extending fentry is not
13966 bpf_log(log, "Cannot extend fentry/fexit\n");
13970 if (prog_extension) {
13971 bpf_log(log, "Cannot replace kernel functions\n");
13976 switch (prog->expected_attach_type) {
13977 case BPF_TRACE_RAW_TP:
13980 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
13983 if (!btf_type_is_typedef(t)) {
13984 bpf_log(log, "attach_btf_id %u is not a typedef\n",
13988 if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
13989 bpf_log(log, "attach_btf_id %u points to wrong type name %s\n",
13993 tname += sizeof(prefix) - 1;
13994 t = btf_type_by_id(btf, t->type);
13995 if (!btf_type_is_ptr(t))
13996 /* should never happen in valid vmlinux build */
13998 t = btf_type_by_id(btf, t->type);
13999 if (!btf_type_is_func_proto(t))
14000 /* should never happen in valid vmlinux build */
14004 case BPF_TRACE_ITER:
14005 if (!btf_type_is_func(t)) {
14006 bpf_log(log, "attach_btf_id %u is not a function\n",
14010 t = btf_type_by_id(btf, t->type);
14011 if (!btf_type_is_func_proto(t))
14013 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
14018 if (!prog_extension)
14021 case BPF_MODIFY_RETURN:
14023 case BPF_TRACE_FENTRY:
14024 case BPF_TRACE_FEXIT:
14025 if (!btf_type_is_func(t)) {
14026 bpf_log(log, "attach_btf_id %u is not a function\n",
14030 if (prog_extension &&
14031 btf_check_type_match(log, prog, btf, t))
14033 t = btf_type_by_id(btf, t->type);
14034 if (!btf_type_is_func_proto(t))
14037 if ((prog->aux->saved_dst_prog_type || prog->aux->saved_dst_attach_type) &&
14038 (!tgt_prog || prog->aux->saved_dst_prog_type != tgt_prog->type ||
14039 prog->aux->saved_dst_attach_type != tgt_prog->expected_attach_type))
14042 if (tgt_prog && conservative)
14045 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
14051 addr = (long) tgt_prog->bpf_func;
14053 addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
14055 addr = kallsyms_lookup_name(tname);
14058 "The address of function %s cannot be found\n",
14064 if (prog->aux->sleepable) {
14066 switch (prog->type) {
14067 case BPF_PROG_TYPE_TRACING:
14068 /* fentry/fexit/fmod_ret progs can be sleepable only if they are
14069 * attached to ALLOW_ERROR_INJECTION and are not in denylist.
14071 if (!check_non_sleepable_error_inject(btf_id) &&
14072 within_error_injection_list(addr))
14075 case BPF_PROG_TYPE_LSM:
14076 /* LSM progs check that they are attached to bpf_lsm_*() funcs.
14077 * Only some of them are sleepable.
14079 if (bpf_lsm_is_sleepable_hook(btf_id))
14086 bpf_log(log, "%s is not sleepable\n", tname);
14089 } else if (prog->expected_attach_type == BPF_MODIFY_RETURN) {
14091 bpf_log(log, "can't modify return codes of BPF programs\n");
14094 ret = check_attach_modify_return(addr, tname);
14096 bpf_log(log, "%s() is not modifiable\n", tname);
14103 tgt_info->tgt_addr = addr;
14104 tgt_info->tgt_name = tname;
14105 tgt_info->tgt_type = t;
14109 BTF_SET_START(btf_id_deny)
14112 BTF_ID(func, migrate_disable)
14113 BTF_ID(func, migrate_enable)
14115 #if !defined CONFIG_PREEMPT_RCU && !defined CONFIG_TINY_RCU
14116 BTF_ID(func, rcu_read_unlock_strict)
14118 BTF_SET_END(btf_id_deny)
14120 static int check_attach_btf_id(struct bpf_verifier_env *env)
14122 struct bpf_prog *prog = env->prog;
14123 struct bpf_prog *tgt_prog = prog->aux->dst_prog;
14124 struct bpf_attach_target_info tgt_info = {};
14125 u32 btf_id = prog->aux->attach_btf_id;
14126 struct bpf_trampoline *tr;
14130 if (prog->type == BPF_PROG_TYPE_SYSCALL) {
14131 if (prog->aux->sleepable)
14132 /* attach_btf_id checked to be zero already */
14134 verbose(env, "Syscall programs can only be sleepable\n");
14138 if (prog->aux->sleepable && prog->type != BPF_PROG_TYPE_TRACING &&
14139 prog->type != BPF_PROG_TYPE_LSM) {
14140 verbose(env, "Only fentry/fexit/fmod_ret and lsm programs can be sleepable\n");
14144 if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
14145 return check_struct_ops_btf_id(env);
14147 if (prog->type != BPF_PROG_TYPE_TRACING &&
14148 prog->type != BPF_PROG_TYPE_LSM &&
14149 prog->type != BPF_PROG_TYPE_EXT)
14152 ret = bpf_check_attach_target(&env->log, prog, tgt_prog, btf_id, &tgt_info);
14156 if (tgt_prog && prog->type == BPF_PROG_TYPE_EXT) {
14157 /* to make freplace equivalent to their targets, they need to
14158 * inherit env->ops and expected_attach_type for the rest of the
14161 env->ops = bpf_verifier_ops[tgt_prog->type];
14162 prog->expected_attach_type = tgt_prog->expected_attach_type;
14165 /* store info about the attachment target that will be used later */
14166 prog->aux->attach_func_proto = tgt_info.tgt_type;
14167 prog->aux->attach_func_name = tgt_info.tgt_name;
14170 prog->aux->saved_dst_prog_type = tgt_prog->type;
14171 prog->aux->saved_dst_attach_type = tgt_prog->expected_attach_type;
14174 if (prog->expected_attach_type == BPF_TRACE_RAW_TP) {
14175 prog->aux->attach_btf_trace = true;
14177 } else if (prog->expected_attach_type == BPF_TRACE_ITER) {
14178 if (!bpf_iter_prog_supported(prog))
14183 if (prog->type == BPF_PROG_TYPE_LSM) {
14184 ret = bpf_lsm_verify_prog(&env->log, prog);
14187 } else if (prog->type == BPF_PROG_TYPE_TRACING &&
14188 btf_id_set_contains(&btf_id_deny, btf_id)) {
14192 key = bpf_trampoline_compute_key(tgt_prog, prog->aux->attach_btf, btf_id);
14193 tr = bpf_trampoline_get(key, &tgt_info);
14197 prog->aux->dst_trampoline = tr;
14201 struct btf *bpf_get_btf_vmlinux(void)
14203 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
14204 mutex_lock(&bpf_verifier_lock);
14206 btf_vmlinux = btf_parse_vmlinux();
14207 mutex_unlock(&bpf_verifier_lock);
14209 return btf_vmlinux;
14212 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr, bpfptr_t uattr)
14214 u64 start_time = ktime_get_ns();
14215 struct bpf_verifier_env *env;
14216 struct bpf_verifier_log *log;
14217 int i, len, ret = -EINVAL;
14220 /* no program is valid */
14221 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
14224 /* 'struct bpf_verifier_env' can be global, but since it's not small,
14225 * allocate/free it every time bpf_check() is called
14227 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
14232 len = (*prog)->len;
14233 env->insn_aux_data =
14234 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
14236 if (!env->insn_aux_data)
14238 for (i = 0; i < len; i++)
14239 env->insn_aux_data[i].orig_idx = i;
14241 env->ops = bpf_verifier_ops[env->prog->type];
14242 env->fd_array = make_bpfptr(attr->fd_array, uattr.is_kernel);
14243 is_priv = bpf_capable();
14245 bpf_get_btf_vmlinux();
14247 /* grab the mutex to protect few globals used by verifier */
14249 mutex_lock(&bpf_verifier_lock);
14251 if (attr->log_level || attr->log_buf || attr->log_size) {
14252 /* user requested verbose verifier output
14253 * and supplied buffer to store the verification trace
14255 log->level = attr->log_level;
14256 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
14257 log->len_total = attr->log_size;
14259 /* log attributes have to be sane */
14260 if (!bpf_verifier_log_attr_valid(log)) {
14266 mark_verifier_state_clean(env);
14268 if (IS_ERR(btf_vmlinux)) {
14269 /* Either gcc or pahole or kernel are broken. */
14270 verbose(env, "in-kernel BTF is malformed\n");
14271 ret = PTR_ERR(btf_vmlinux);
14272 goto skip_full_check;
14275 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
14276 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
14277 env->strict_alignment = true;
14278 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
14279 env->strict_alignment = false;
14281 env->allow_ptr_leaks = bpf_allow_ptr_leaks();
14282 env->allow_uninit_stack = bpf_allow_uninit_stack();
14283 env->allow_ptr_to_map_access = bpf_allow_ptr_to_map_access();
14284 env->bypass_spec_v1 = bpf_bypass_spec_v1();
14285 env->bypass_spec_v4 = bpf_bypass_spec_v4();
14286 env->bpf_capable = bpf_capable();
14289 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
14291 env->explored_states = kvcalloc(state_htab_size(env),
14292 sizeof(struct bpf_verifier_state_list *),
14295 if (!env->explored_states)
14296 goto skip_full_check;
14298 ret = add_subprog_and_kfunc(env);
14300 goto skip_full_check;
14302 ret = check_subprogs(env);
14304 goto skip_full_check;
14306 ret = check_btf_info(env, attr, uattr);
14308 goto skip_full_check;
14310 ret = check_attach_btf_id(env);
14312 goto skip_full_check;
14314 ret = resolve_pseudo_ldimm64(env);
14316 goto skip_full_check;
14318 if (bpf_prog_is_dev_bound(env->prog->aux)) {
14319 ret = bpf_prog_offload_verifier_prep(env->prog);
14321 goto skip_full_check;
14324 ret = check_cfg(env);
14326 goto skip_full_check;
14328 ret = do_check_subprogs(env);
14329 ret = ret ?: do_check_main(env);
14331 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
14332 ret = bpf_prog_offload_finalize(env);
14335 kvfree(env->explored_states);
14338 ret = check_max_stack_depth(env);
14340 /* instruction rewrites happen after this point */
14343 opt_hard_wire_dead_code_branches(env);
14345 ret = opt_remove_dead_code(env);
14347 ret = opt_remove_nops(env);
14350 sanitize_dead_code(env);
14354 /* program is valid, convert *(u32*)(ctx + off) accesses */
14355 ret = convert_ctx_accesses(env);
14358 ret = do_misc_fixups(env);
14360 /* do 32-bit optimization after insn patching has done so those patched
14361 * insns could be handled correctly.
14363 if (ret == 0 && !bpf_prog_is_dev_bound(env->prog->aux)) {
14364 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
14365 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
14370 ret = fixup_call_args(env);
14372 env->verification_time = ktime_get_ns() - start_time;
14373 print_verification_stats(env);
14374 env->prog->aux->verified_insns = env->insn_processed;
14376 if (log->level && bpf_verifier_log_full(log))
14378 if (log->level && !log->ubuf) {
14380 goto err_release_maps;
14384 goto err_release_maps;
14386 if (env->used_map_cnt) {
14387 /* if program passed verifier, update used_maps in bpf_prog_info */
14388 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
14389 sizeof(env->used_maps[0]),
14392 if (!env->prog->aux->used_maps) {
14394 goto err_release_maps;
14397 memcpy(env->prog->aux->used_maps, env->used_maps,
14398 sizeof(env->used_maps[0]) * env->used_map_cnt);
14399 env->prog->aux->used_map_cnt = env->used_map_cnt;
14401 if (env->used_btf_cnt) {
14402 /* if program passed verifier, update used_btfs in bpf_prog_aux */
14403 env->prog->aux->used_btfs = kmalloc_array(env->used_btf_cnt,
14404 sizeof(env->used_btfs[0]),
14406 if (!env->prog->aux->used_btfs) {
14408 goto err_release_maps;
14411 memcpy(env->prog->aux->used_btfs, env->used_btfs,
14412 sizeof(env->used_btfs[0]) * env->used_btf_cnt);
14413 env->prog->aux->used_btf_cnt = env->used_btf_cnt;
14415 if (env->used_map_cnt || env->used_btf_cnt) {
14416 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
14417 * bpf_ld_imm64 instructions
14419 convert_pseudo_ld_imm64(env);
14422 adjust_btf_func(env);
14425 if (!env->prog->aux->used_maps)
14426 /* if we didn't copy map pointers into bpf_prog_info, release
14427 * them now. Otherwise free_used_maps() will release them.
14430 if (!env->prog->aux->used_btfs)
14433 /* extension progs temporarily inherit the attach_type of their targets
14434 for verification purposes, so set it back to zero before returning
14436 if (env->prog->type == BPF_PROG_TYPE_EXT)
14437 env->prog->expected_attach_type = 0;
14442 mutex_unlock(&bpf_verifier_lock);
14443 vfree(env->insn_aux_data);