1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
3 * Copyright (c) 2016 Facebook
4 * Copyright (c) 2018 Covalent IO, Inc. http://covalent.io
6 #include <uapi/linux/btf.h>
7 #include <linux/kernel.h>
8 #include <linux/types.h>
9 #include <linux/slab.h>
10 #include <linux/bpf.h>
11 #include <linux/btf.h>
12 #include <linux/bpf_verifier.h>
13 #include <linux/filter.h>
14 #include <net/netlink.h>
15 #include <linux/file.h>
16 #include <linux/vmalloc.h>
17 #include <linux/stringify.h>
18 #include <linux/bsearch.h>
19 #include <linux/sort.h>
20 #include <linux/perf_event.h>
21 #include <linux/ctype.h>
22 #include <linux/error-injection.h>
23 #include <linux/bpf_lsm.h>
24 #include <linux/btf_ids.h>
28 static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
29 #define BPF_PROG_TYPE(_id, _name, prog_ctx_type, kern_ctx_type) \
30 [_id] = & _name ## _verifier_ops,
31 #define BPF_MAP_TYPE(_id, _ops)
32 #define BPF_LINK_TYPE(_id, _name)
33 #include <linux/bpf_types.h>
39 /* bpf_check() is a static code analyzer that walks eBPF program
40 * instruction by instruction and updates register/stack state.
41 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
43 * The first pass is depth-first-search to check that the program is a DAG.
44 * It rejects the following programs:
45 * - larger than BPF_MAXINSNS insns
46 * - if loop is present (detected via back-edge)
47 * - unreachable insns exist (shouldn't be a forest. program = one function)
48 * - out of bounds or malformed jumps
49 * The second pass is all possible path descent from the 1st insn.
50 * Since it's analyzing all paths through the program, the length of the
51 * analysis is limited to 64k insn, which may be hit even if total number of
52 * insn is less then 4K, but there are too many branches that change stack/regs.
53 * Number of 'branches to be analyzed' is limited to 1k
55 * On entry to each instruction, each register has a type, and the instruction
56 * changes the types of the registers depending on instruction semantics.
57 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
60 * All registers are 64-bit.
61 * R0 - return register
62 * R1-R5 argument passing registers
63 * R6-R9 callee saved registers
64 * R10 - frame pointer read-only
66 * At the start of BPF program the register R1 contains a pointer to bpf_context
67 * and has type PTR_TO_CTX.
69 * Verifier tracks arithmetic operations on pointers in case:
70 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
71 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
72 * 1st insn copies R10 (which has FRAME_PTR) type into R1
73 * and 2nd arithmetic instruction is pattern matched to recognize
74 * that it wants to construct a pointer to some element within stack.
75 * So after 2nd insn, the register R1 has type PTR_TO_STACK
76 * (and -20 constant is saved for further stack bounds checking).
77 * Meaning that this reg is a pointer to stack plus known immediate constant.
79 * Most of the time the registers have SCALAR_VALUE type, which
80 * means the register has some value, but it's not a valid pointer.
81 * (like pointer plus pointer becomes SCALAR_VALUE type)
83 * When verifier sees load or store instructions the type of base register
84 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK, PTR_TO_SOCKET. These are
85 * four pointer types recognized by check_mem_access() function.
87 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
88 * and the range of [ptr, ptr + map's value_size) is accessible.
90 * registers used to pass values to function calls are checked against
91 * function argument constraints.
93 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
94 * It means that the register type passed to this function must be
95 * PTR_TO_STACK and it will be used inside the function as
96 * 'pointer to map element key'
98 * For example the argument constraints for bpf_map_lookup_elem():
99 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
100 * .arg1_type = ARG_CONST_MAP_PTR,
101 * .arg2_type = ARG_PTR_TO_MAP_KEY,
103 * ret_type says that this function returns 'pointer to map elem value or null'
104 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
105 * 2nd argument should be a pointer to stack, which will be used inside
106 * the helper function as a pointer to map element key.
108 * On the kernel side the helper function looks like:
109 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
111 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
112 * void *key = (void *) (unsigned long) r2;
115 * here kernel can access 'key' and 'map' pointers safely, knowing that
116 * [key, key + map->key_size) bytes are valid and were initialized on
117 * the stack of eBPF program.
120 * Corresponding eBPF program may look like:
121 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
122 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
123 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
124 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
125 * here verifier looks at prototype of map_lookup_elem() and sees:
126 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
127 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
129 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
130 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
131 * and were initialized prior to this call.
132 * If it's ok, then verifier allows this BPF_CALL insn and looks at
133 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
134 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
135 * returns either pointer to map value or NULL.
137 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
138 * insn, the register holding that pointer in the true branch changes state to
139 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
140 * branch. See check_cond_jmp_op().
142 * After the call R0 is set to return type of the function and registers R1-R5
143 * are set to NOT_INIT to indicate that they are no longer readable.
145 * The following reference types represent a potential reference to a kernel
146 * resource which, after first being allocated, must be checked and freed by
148 * - PTR_TO_SOCKET_OR_NULL, PTR_TO_SOCKET
150 * When the verifier sees a helper call return a reference type, it allocates a
151 * pointer id for the reference and stores it in the current function state.
152 * Similar to the way that PTR_TO_MAP_VALUE_OR_NULL is converted into
153 * PTR_TO_MAP_VALUE, PTR_TO_SOCKET_OR_NULL becomes PTR_TO_SOCKET when the type
154 * passes through a NULL-check conditional. For the branch wherein the state is
155 * changed to CONST_IMM, the verifier releases the reference.
157 * For each helper function that allocates a reference, such as
158 * bpf_sk_lookup_tcp(), there is a corresponding release function, such as
159 * bpf_sk_release(). When a reference type passes into the release function,
160 * the verifier also releases the reference. If any unchecked or unreleased
161 * reference remains at the end of the program, the verifier rejects it.
164 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
165 struct bpf_verifier_stack_elem {
166 /* verifer state is 'st'
167 * before processing instruction 'insn_idx'
168 * and after processing instruction 'prev_insn_idx'
170 struct bpf_verifier_state st;
173 struct bpf_verifier_stack_elem *next;
174 /* length of verifier log at the time this state was pushed on stack */
178 #define BPF_COMPLEXITY_LIMIT_JMP_SEQ 8192
179 #define BPF_COMPLEXITY_LIMIT_STATES 64
181 #define BPF_MAP_KEY_POISON (1ULL << 63)
182 #define BPF_MAP_KEY_SEEN (1ULL << 62)
184 #define BPF_MAP_PTR_UNPRIV 1UL
185 #define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \
186 POISON_POINTER_DELTA))
187 #define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
189 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
191 return BPF_MAP_PTR(aux->map_ptr_state) == BPF_MAP_PTR_POISON;
194 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
196 return aux->map_ptr_state & BPF_MAP_PTR_UNPRIV;
199 static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
200 const struct bpf_map *map, bool unpriv)
202 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON & BPF_MAP_PTR_UNPRIV);
203 unpriv |= bpf_map_ptr_unpriv(aux);
204 aux->map_ptr_state = (unsigned long)map |
205 (unpriv ? BPF_MAP_PTR_UNPRIV : 0UL);
208 static bool bpf_map_key_poisoned(const struct bpf_insn_aux_data *aux)
210 return aux->map_key_state & BPF_MAP_KEY_POISON;
213 static bool bpf_map_key_unseen(const struct bpf_insn_aux_data *aux)
215 return !(aux->map_key_state & BPF_MAP_KEY_SEEN);
218 static u64 bpf_map_key_immediate(const struct bpf_insn_aux_data *aux)
220 return aux->map_key_state & ~(BPF_MAP_KEY_SEEN | BPF_MAP_KEY_POISON);
223 static void bpf_map_key_store(struct bpf_insn_aux_data *aux, u64 state)
225 bool poisoned = bpf_map_key_poisoned(aux);
227 aux->map_key_state = state | BPF_MAP_KEY_SEEN |
228 (poisoned ? BPF_MAP_KEY_POISON : 0ULL);
231 static bool bpf_pseudo_call(const struct bpf_insn *insn)
233 return insn->code == (BPF_JMP | BPF_CALL) &&
234 insn->src_reg == BPF_PSEUDO_CALL;
237 static bool bpf_pseudo_kfunc_call(const struct bpf_insn *insn)
239 return insn->code == (BPF_JMP | BPF_CALL) &&
240 insn->src_reg == BPF_PSEUDO_KFUNC_CALL;
243 static bool bpf_pseudo_func(const struct bpf_insn *insn)
245 return insn->code == (BPF_LD | BPF_IMM | BPF_DW) &&
246 insn->src_reg == BPF_PSEUDO_FUNC;
249 struct bpf_call_arg_meta {
250 struct bpf_map *map_ptr;
267 struct btf *btf_vmlinux;
269 static DEFINE_MUTEX(bpf_verifier_lock);
271 static const struct bpf_line_info *
272 find_linfo(const struct bpf_verifier_env *env, u32 insn_off)
274 const struct bpf_line_info *linfo;
275 const struct bpf_prog *prog;
279 nr_linfo = prog->aux->nr_linfo;
281 if (!nr_linfo || insn_off >= prog->len)
284 linfo = prog->aux->linfo;
285 for (i = 1; i < nr_linfo; i++)
286 if (insn_off < linfo[i].insn_off)
289 return &linfo[i - 1];
292 void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
297 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
299 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
300 "verifier log line truncated - local buffer too short\n");
302 n = min(log->len_total - log->len_used - 1, n);
305 if (log->level == BPF_LOG_KERNEL) {
306 pr_err("BPF:%s\n", log->kbuf);
309 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
315 static void bpf_vlog_reset(struct bpf_verifier_log *log, u32 new_pos)
319 if (!bpf_verifier_log_needed(log))
322 log->len_used = new_pos;
323 if (put_user(zero, log->ubuf + new_pos))
327 /* log_level controls verbosity level of eBPF verifier.
328 * bpf_verifier_log_write() is used to dump the verification trace to the log,
329 * so the user can figure out what's wrong with the program
331 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
332 const char *fmt, ...)
336 if (!bpf_verifier_log_needed(&env->log))
340 bpf_verifier_vlog(&env->log, fmt, args);
343 EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
345 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
347 struct bpf_verifier_env *env = private_data;
350 if (!bpf_verifier_log_needed(&env->log))
354 bpf_verifier_vlog(&env->log, fmt, args);
358 __printf(2, 3) void bpf_log(struct bpf_verifier_log *log,
359 const char *fmt, ...)
363 if (!bpf_verifier_log_needed(log))
367 bpf_verifier_vlog(log, fmt, args);
371 static const char *ltrim(const char *s)
379 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env,
381 const char *prefix_fmt, ...)
383 const struct bpf_line_info *linfo;
385 if (!bpf_verifier_log_needed(&env->log))
388 linfo = find_linfo(env, insn_off);
389 if (!linfo || linfo == env->prev_linfo)
395 va_start(args, prefix_fmt);
396 bpf_verifier_vlog(&env->log, prefix_fmt, args);
401 ltrim(btf_name_by_offset(env->prog->aux->btf,
404 env->prev_linfo = linfo;
407 static void verbose_invalid_scalar(struct bpf_verifier_env *env,
408 struct bpf_reg_state *reg,
409 struct tnum *range, const char *ctx,
410 const char *reg_name)
414 verbose(env, "At %s the register %s ", ctx, reg_name);
415 if (!tnum_is_unknown(reg->var_off)) {
416 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
417 verbose(env, "has value %s", tn_buf);
419 verbose(env, "has unknown scalar value");
421 tnum_strn(tn_buf, sizeof(tn_buf), *range);
422 verbose(env, " should have been in %s\n", tn_buf);
425 static bool type_is_pkt_pointer(enum bpf_reg_type type)
427 return type == PTR_TO_PACKET ||
428 type == PTR_TO_PACKET_META;
431 static bool type_is_sk_pointer(enum bpf_reg_type type)
433 return type == PTR_TO_SOCKET ||
434 type == PTR_TO_SOCK_COMMON ||
435 type == PTR_TO_TCP_SOCK ||
436 type == PTR_TO_XDP_SOCK;
439 static bool reg_type_not_null(enum bpf_reg_type type)
441 return type == PTR_TO_SOCKET ||
442 type == PTR_TO_TCP_SOCK ||
443 type == PTR_TO_MAP_VALUE ||
444 type == PTR_TO_MAP_KEY ||
445 type == PTR_TO_SOCK_COMMON;
448 static bool reg_type_may_be_null(enum bpf_reg_type type)
450 return type == PTR_TO_MAP_VALUE_OR_NULL ||
451 type == PTR_TO_SOCKET_OR_NULL ||
452 type == PTR_TO_SOCK_COMMON_OR_NULL ||
453 type == PTR_TO_TCP_SOCK_OR_NULL ||
454 type == PTR_TO_BTF_ID_OR_NULL ||
455 type == PTR_TO_MEM_OR_NULL ||
456 type == PTR_TO_RDONLY_BUF_OR_NULL ||
457 type == PTR_TO_RDWR_BUF_OR_NULL;
460 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
462 return reg->type == PTR_TO_MAP_VALUE &&
463 map_value_has_spin_lock(reg->map_ptr);
466 static bool reg_type_may_be_refcounted_or_null(enum bpf_reg_type type)
468 return type == PTR_TO_SOCKET ||
469 type == PTR_TO_SOCKET_OR_NULL ||
470 type == PTR_TO_TCP_SOCK ||
471 type == PTR_TO_TCP_SOCK_OR_NULL ||
472 type == PTR_TO_MEM ||
473 type == PTR_TO_MEM_OR_NULL;
476 static bool arg_type_may_be_refcounted(enum bpf_arg_type type)
478 return type == ARG_PTR_TO_SOCK_COMMON;
481 static bool arg_type_may_be_null(enum bpf_arg_type type)
483 return type == ARG_PTR_TO_MAP_VALUE_OR_NULL ||
484 type == ARG_PTR_TO_MEM_OR_NULL ||
485 type == ARG_PTR_TO_CTX_OR_NULL ||
486 type == ARG_PTR_TO_SOCKET_OR_NULL ||
487 type == ARG_PTR_TO_ALLOC_MEM_OR_NULL ||
488 type == ARG_PTR_TO_STACK_OR_NULL;
491 /* Determine whether the function releases some resources allocated by another
492 * function call. The first reference type argument will be assumed to be
493 * released by release_reference().
495 static bool is_release_function(enum bpf_func_id func_id)
497 return func_id == BPF_FUNC_sk_release ||
498 func_id == BPF_FUNC_ringbuf_submit ||
499 func_id == BPF_FUNC_ringbuf_discard;
502 static bool may_be_acquire_function(enum bpf_func_id func_id)
504 return func_id == BPF_FUNC_sk_lookup_tcp ||
505 func_id == BPF_FUNC_sk_lookup_udp ||
506 func_id == BPF_FUNC_skc_lookup_tcp ||
507 func_id == BPF_FUNC_map_lookup_elem ||
508 func_id == BPF_FUNC_ringbuf_reserve;
511 static bool is_acquire_function(enum bpf_func_id func_id,
512 const struct bpf_map *map)
514 enum bpf_map_type map_type = map ? map->map_type : BPF_MAP_TYPE_UNSPEC;
516 if (func_id == BPF_FUNC_sk_lookup_tcp ||
517 func_id == BPF_FUNC_sk_lookup_udp ||
518 func_id == BPF_FUNC_skc_lookup_tcp ||
519 func_id == BPF_FUNC_ringbuf_reserve)
522 if (func_id == BPF_FUNC_map_lookup_elem &&
523 (map_type == BPF_MAP_TYPE_SOCKMAP ||
524 map_type == BPF_MAP_TYPE_SOCKHASH))
530 static bool is_ptr_cast_function(enum bpf_func_id func_id)
532 return func_id == BPF_FUNC_tcp_sock ||
533 func_id == BPF_FUNC_sk_fullsock ||
534 func_id == BPF_FUNC_skc_to_tcp_sock ||
535 func_id == BPF_FUNC_skc_to_tcp6_sock ||
536 func_id == BPF_FUNC_skc_to_udp6_sock ||
537 func_id == BPF_FUNC_skc_to_tcp_timewait_sock ||
538 func_id == BPF_FUNC_skc_to_tcp_request_sock;
541 static bool is_cmpxchg_insn(const struct bpf_insn *insn)
543 return BPF_CLASS(insn->code) == BPF_STX &&
544 BPF_MODE(insn->code) == BPF_ATOMIC &&
545 insn->imm == BPF_CMPXCHG;
548 /* string representation of 'enum bpf_reg_type' */
549 static const char * const reg_type_str[] = {
551 [SCALAR_VALUE] = "inv",
552 [PTR_TO_CTX] = "ctx",
553 [CONST_PTR_TO_MAP] = "map_ptr",
554 [PTR_TO_MAP_VALUE] = "map_value",
555 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
556 [PTR_TO_STACK] = "fp",
557 [PTR_TO_PACKET] = "pkt",
558 [PTR_TO_PACKET_META] = "pkt_meta",
559 [PTR_TO_PACKET_END] = "pkt_end",
560 [PTR_TO_FLOW_KEYS] = "flow_keys",
561 [PTR_TO_SOCKET] = "sock",
562 [PTR_TO_SOCKET_OR_NULL] = "sock_or_null",
563 [PTR_TO_SOCK_COMMON] = "sock_common",
564 [PTR_TO_SOCK_COMMON_OR_NULL] = "sock_common_or_null",
565 [PTR_TO_TCP_SOCK] = "tcp_sock",
566 [PTR_TO_TCP_SOCK_OR_NULL] = "tcp_sock_or_null",
567 [PTR_TO_TP_BUFFER] = "tp_buffer",
568 [PTR_TO_XDP_SOCK] = "xdp_sock",
569 [PTR_TO_BTF_ID] = "ptr_",
570 [PTR_TO_BTF_ID_OR_NULL] = "ptr_or_null_",
571 [PTR_TO_PERCPU_BTF_ID] = "percpu_ptr_",
572 [PTR_TO_MEM] = "mem",
573 [PTR_TO_MEM_OR_NULL] = "mem_or_null",
574 [PTR_TO_RDONLY_BUF] = "rdonly_buf",
575 [PTR_TO_RDONLY_BUF_OR_NULL] = "rdonly_buf_or_null",
576 [PTR_TO_RDWR_BUF] = "rdwr_buf",
577 [PTR_TO_RDWR_BUF_OR_NULL] = "rdwr_buf_or_null",
578 [PTR_TO_FUNC] = "func",
579 [PTR_TO_MAP_KEY] = "map_key",
582 static char slot_type_char[] = {
583 [STACK_INVALID] = '?',
589 static void print_liveness(struct bpf_verifier_env *env,
590 enum bpf_reg_liveness live)
592 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
594 if (live & REG_LIVE_READ)
596 if (live & REG_LIVE_WRITTEN)
598 if (live & REG_LIVE_DONE)
602 static struct bpf_func_state *func(struct bpf_verifier_env *env,
603 const struct bpf_reg_state *reg)
605 struct bpf_verifier_state *cur = env->cur_state;
607 return cur->frame[reg->frameno];
610 static const char *kernel_type_name(const struct btf* btf, u32 id)
612 return btf_name_by_offset(btf, btf_type_by_id(btf, id)->name_off);
615 static void print_verifier_state(struct bpf_verifier_env *env,
616 const struct bpf_func_state *state)
618 const struct bpf_reg_state *reg;
623 verbose(env, " frame%d:", state->frameno);
624 for (i = 0; i < MAX_BPF_REG; i++) {
625 reg = &state->regs[i];
629 verbose(env, " R%d", i);
630 print_liveness(env, reg->live);
631 verbose(env, "=%s", reg_type_str[t]);
632 if (t == SCALAR_VALUE && reg->precise)
634 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
635 tnum_is_const(reg->var_off)) {
636 /* reg->off should be 0 for SCALAR_VALUE */
637 verbose(env, "%lld", reg->var_off.value + reg->off);
639 if (t == PTR_TO_BTF_ID ||
640 t == PTR_TO_BTF_ID_OR_NULL ||
641 t == PTR_TO_PERCPU_BTF_ID)
642 verbose(env, "%s", kernel_type_name(reg->btf, reg->btf_id));
643 verbose(env, "(id=%d", reg->id);
644 if (reg_type_may_be_refcounted_or_null(t))
645 verbose(env, ",ref_obj_id=%d", reg->ref_obj_id);
646 if (t != SCALAR_VALUE)
647 verbose(env, ",off=%d", reg->off);
648 if (type_is_pkt_pointer(t))
649 verbose(env, ",r=%d", reg->range);
650 else if (t == CONST_PTR_TO_MAP ||
651 t == PTR_TO_MAP_KEY ||
652 t == PTR_TO_MAP_VALUE ||
653 t == PTR_TO_MAP_VALUE_OR_NULL)
654 verbose(env, ",ks=%d,vs=%d",
655 reg->map_ptr->key_size,
656 reg->map_ptr->value_size);
657 if (tnum_is_const(reg->var_off)) {
658 /* Typically an immediate SCALAR_VALUE, but
659 * could be a pointer whose offset is too big
662 verbose(env, ",imm=%llx", reg->var_off.value);
664 if (reg->smin_value != reg->umin_value &&
665 reg->smin_value != S64_MIN)
666 verbose(env, ",smin_value=%lld",
667 (long long)reg->smin_value);
668 if (reg->smax_value != reg->umax_value &&
669 reg->smax_value != S64_MAX)
670 verbose(env, ",smax_value=%lld",
671 (long long)reg->smax_value);
672 if (reg->umin_value != 0)
673 verbose(env, ",umin_value=%llu",
674 (unsigned long long)reg->umin_value);
675 if (reg->umax_value != U64_MAX)
676 verbose(env, ",umax_value=%llu",
677 (unsigned long long)reg->umax_value);
678 if (!tnum_is_unknown(reg->var_off)) {
681 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
682 verbose(env, ",var_off=%s", tn_buf);
684 if (reg->s32_min_value != reg->smin_value &&
685 reg->s32_min_value != S32_MIN)
686 verbose(env, ",s32_min_value=%d",
687 (int)(reg->s32_min_value));
688 if (reg->s32_max_value != reg->smax_value &&
689 reg->s32_max_value != S32_MAX)
690 verbose(env, ",s32_max_value=%d",
691 (int)(reg->s32_max_value));
692 if (reg->u32_min_value != reg->umin_value &&
693 reg->u32_min_value != U32_MIN)
694 verbose(env, ",u32_min_value=%d",
695 (int)(reg->u32_min_value));
696 if (reg->u32_max_value != reg->umax_value &&
697 reg->u32_max_value != U32_MAX)
698 verbose(env, ",u32_max_value=%d",
699 (int)(reg->u32_max_value));
704 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
705 char types_buf[BPF_REG_SIZE + 1];
709 for (j = 0; j < BPF_REG_SIZE; j++) {
710 if (state->stack[i].slot_type[j] != STACK_INVALID)
712 types_buf[j] = slot_type_char[
713 state->stack[i].slot_type[j]];
715 types_buf[BPF_REG_SIZE] = 0;
718 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
719 print_liveness(env, state->stack[i].spilled_ptr.live);
720 if (state->stack[i].slot_type[0] == STACK_SPILL) {
721 reg = &state->stack[i].spilled_ptr;
723 verbose(env, "=%s", reg_type_str[t]);
724 if (t == SCALAR_VALUE && reg->precise)
726 if (t == SCALAR_VALUE && tnum_is_const(reg->var_off))
727 verbose(env, "%lld", reg->var_off.value + reg->off);
729 verbose(env, "=%s", types_buf);
732 if (state->acquired_refs && state->refs[0].id) {
733 verbose(env, " refs=%d", state->refs[0].id);
734 for (i = 1; i < state->acquired_refs; i++)
735 if (state->refs[i].id)
736 verbose(env, ",%d", state->refs[i].id);
741 /* copy array src of length n * size bytes to dst. dst is reallocated if it's too
742 * small to hold src. This is different from krealloc since we don't want to preserve
743 * the contents of dst.
745 * Leaves dst untouched if src is NULL or length is zero. Returns NULL if memory could
748 static void *copy_array(void *dst, const void *src, size_t n, size_t size, gfp_t flags)
752 if (ZERO_OR_NULL_PTR(src))
755 if (unlikely(check_mul_overflow(n, size, &bytes)))
758 if (ksize(dst) < bytes) {
760 dst = kmalloc_track_caller(bytes, flags);
765 memcpy(dst, src, bytes);
767 return dst ? dst : ZERO_SIZE_PTR;
770 /* resize an array from old_n items to new_n items. the array is reallocated if it's too
771 * small to hold new_n items. new items are zeroed out if the array grows.
773 * Contrary to krealloc_array, does not free arr if new_n is zero.
775 static void *realloc_array(void *arr, size_t old_n, size_t new_n, size_t size)
777 if (!new_n || old_n == new_n)
780 arr = krealloc_array(arr, new_n, size, GFP_KERNEL);
785 memset(arr + old_n * size, 0, (new_n - old_n) * size);
788 return arr ? arr : ZERO_SIZE_PTR;
791 static int copy_reference_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
793 dst->refs = copy_array(dst->refs, src->refs, src->acquired_refs,
794 sizeof(struct bpf_reference_state), GFP_KERNEL);
798 dst->acquired_refs = src->acquired_refs;
802 static int copy_stack_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
804 size_t n = src->allocated_stack / BPF_REG_SIZE;
806 dst->stack = copy_array(dst->stack, src->stack, n, sizeof(struct bpf_stack_state),
811 dst->allocated_stack = src->allocated_stack;
815 static int resize_reference_state(struct bpf_func_state *state, size_t n)
817 state->refs = realloc_array(state->refs, state->acquired_refs, n,
818 sizeof(struct bpf_reference_state));
822 state->acquired_refs = n;
826 static int grow_stack_state(struct bpf_func_state *state, int size)
828 size_t old_n = state->allocated_stack / BPF_REG_SIZE, n = size / BPF_REG_SIZE;
833 state->stack = realloc_array(state->stack, old_n, n, sizeof(struct bpf_stack_state));
837 state->allocated_stack = size;
841 /* Acquire a pointer id from the env and update the state->refs to include
842 * this new pointer reference.
843 * On success, returns a valid pointer id to associate with the register
844 * On failure, returns a negative errno.
846 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
848 struct bpf_func_state *state = cur_func(env);
849 int new_ofs = state->acquired_refs;
852 err = resize_reference_state(state, state->acquired_refs + 1);
856 state->refs[new_ofs].id = id;
857 state->refs[new_ofs].insn_idx = insn_idx;
862 /* release function corresponding to acquire_reference_state(). Idempotent. */
863 static int release_reference_state(struct bpf_func_state *state, int ptr_id)
867 last_idx = state->acquired_refs - 1;
868 for (i = 0; i < state->acquired_refs; i++) {
869 if (state->refs[i].id == ptr_id) {
870 if (last_idx && i != last_idx)
871 memcpy(&state->refs[i], &state->refs[last_idx],
872 sizeof(*state->refs));
873 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
874 state->acquired_refs--;
881 static void free_func_state(struct bpf_func_state *state)
890 static void clear_jmp_history(struct bpf_verifier_state *state)
892 kfree(state->jmp_history);
893 state->jmp_history = NULL;
894 state->jmp_history_cnt = 0;
897 static void free_verifier_state(struct bpf_verifier_state *state,
902 for (i = 0; i <= state->curframe; i++) {
903 free_func_state(state->frame[i]);
904 state->frame[i] = NULL;
906 clear_jmp_history(state);
911 /* copy verifier state from src to dst growing dst stack space
912 * when necessary to accommodate larger src stack
914 static int copy_func_state(struct bpf_func_state *dst,
915 const struct bpf_func_state *src)
919 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
920 err = copy_reference_state(dst, src);
923 return copy_stack_state(dst, src);
926 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
927 const struct bpf_verifier_state *src)
929 struct bpf_func_state *dst;
932 dst_state->jmp_history = copy_array(dst_state->jmp_history, src->jmp_history,
933 src->jmp_history_cnt, sizeof(struct bpf_idx_pair),
935 if (!dst_state->jmp_history)
937 dst_state->jmp_history_cnt = src->jmp_history_cnt;
939 /* if dst has more stack frames then src frame, free them */
940 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
941 free_func_state(dst_state->frame[i]);
942 dst_state->frame[i] = NULL;
944 dst_state->speculative = src->speculative;
945 dst_state->curframe = src->curframe;
946 dst_state->active_spin_lock = src->active_spin_lock;
947 dst_state->branches = src->branches;
948 dst_state->parent = src->parent;
949 dst_state->first_insn_idx = src->first_insn_idx;
950 dst_state->last_insn_idx = src->last_insn_idx;
951 for (i = 0; i <= src->curframe; i++) {
952 dst = dst_state->frame[i];
954 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
957 dst_state->frame[i] = dst;
959 err = copy_func_state(dst, src->frame[i]);
966 static void update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
969 u32 br = --st->branches;
971 /* WARN_ON(br > 1) technically makes sense here,
972 * but see comment in push_stack(), hence:
974 WARN_ONCE((int)br < 0,
975 "BUG update_branch_counts:branches_to_explore=%d\n",
983 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
984 int *insn_idx, bool pop_log)
986 struct bpf_verifier_state *cur = env->cur_state;
987 struct bpf_verifier_stack_elem *elem, *head = env->head;
990 if (env->head == NULL)
994 err = copy_verifier_state(cur, &head->st);
999 bpf_vlog_reset(&env->log, head->log_pos);
1001 *insn_idx = head->insn_idx;
1003 *prev_insn_idx = head->prev_insn_idx;
1005 free_verifier_state(&head->st, false);
1012 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
1013 int insn_idx, int prev_insn_idx,
1016 struct bpf_verifier_state *cur = env->cur_state;
1017 struct bpf_verifier_stack_elem *elem;
1020 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
1024 elem->insn_idx = insn_idx;
1025 elem->prev_insn_idx = prev_insn_idx;
1026 elem->next = env->head;
1027 elem->log_pos = env->log.len_used;
1030 err = copy_verifier_state(&elem->st, cur);
1033 elem->st.speculative |= speculative;
1034 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
1035 verbose(env, "The sequence of %d jumps is too complex.\n",
1039 if (elem->st.parent) {
1040 ++elem->st.parent->branches;
1041 /* WARN_ON(branches > 2) technically makes sense here,
1043 * 1. speculative states will bump 'branches' for non-branch
1045 * 2. is_state_visited() heuristics may decide not to create
1046 * a new state for a sequence of branches and all such current
1047 * and cloned states will be pointing to a single parent state
1048 * which might have large 'branches' count.
1053 free_verifier_state(env->cur_state, true);
1054 env->cur_state = NULL;
1055 /* pop all elements and return */
1056 while (!pop_stack(env, NULL, NULL, false));
1060 #define CALLER_SAVED_REGS 6
1061 static const int caller_saved[CALLER_SAVED_REGS] = {
1062 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
1065 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1066 struct bpf_reg_state *reg);
1068 /* This helper doesn't clear reg->id */
1069 static void ___mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1071 reg->var_off = tnum_const(imm);
1072 reg->smin_value = (s64)imm;
1073 reg->smax_value = (s64)imm;
1074 reg->umin_value = imm;
1075 reg->umax_value = imm;
1077 reg->s32_min_value = (s32)imm;
1078 reg->s32_max_value = (s32)imm;
1079 reg->u32_min_value = (u32)imm;
1080 reg->u32_max_value = (u32)imm;
1083 /* Mark the unknown part of a register (variable offset or scalar value) as
1084 * known to have the value @imm.
1086 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1088 /* Clear id, off, and union(map_ptr, range) */
1089 memset(((u8 *)reg) + sizeof(reg->type), 0,
1090 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
1091 ___mark_reg_known(reg, imm);
1094 static void __mark_reg32_known(struct bpf_reg_state *reg, u64 imm)
1096 reg->var_off = tnum_const_subreg(reg->var_off, imm);
1097 reg->s32_min_value = (s32)imm;
1098 reg->s32_max_value = (s32)imm;
1099 reg->u32_min_value = (u32)imm;
1100 reg->u32_max_value = (u32)imm;
1103 /* Mark the 'variable offset' part of a register as zero. This should be
1104 * used only on registers holding a pointer type.
1106 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
1108 __mark_reg_known(reg, 0);
1111 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
1113 __mark_reg_known(reg, 0);
1114 reg->type = SCALAR_VALUE;
1117 static void mark_reg_known_zero(struct bpf_verifier_env *env,
1118 struct bpf_reg_state *regs, u32 regno)
1120 if (WARN_ON(regno >= MAX_BPF_REG)) {
1121 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
1122 /* Something bad happened, let's kill all regs */
1123 for (regno = 0; regno < MAX_BPF_REG; regno++)
1124 __mark_reg_not_init(env, regs + regno);
1127 __mark_reg_known_zero(regs + regno);
1130 static void mark_ptr_not_null_reg(struct bpf_reg_state *reg)
1132 switch (reg->type) {
1133 case PTR_TO_MAP_VALUE_OR_NULL: {
1134 const struct bpf_map *map = reg->map_ptr;
1136 if (map->inner_map_meta) {
1137 reg->type = CONST_PTR_TO_MAP;
1138 reg->map_ptr = map->inner_map_meta;
1139 /* transfer reg's id which is unique for every map_lookup_elem
1140 * as UID of the inner map.
1142 reg->map_uid = reg->id;
1143 } else if (map->map_type == BPF_MAP_TYPE_XSKMAP) {
1144 reg->type = PTR_TO_XDP_SOCK;
1145 } else if (map->map_type == BPF_MAP_TYPE_SOCKMAP ||
1146 map->map_type == BPF_MAP_TYPE_SOCKHASH) {
1147 reg->type = PTR_TO_SOCKET;
1149 reg->type = PTR_TO_MAP_VALUE;
1153 case PTR_TO_SOCKET_OR_NULL:
1154 reg->type = PTR_TO_SOCKET;
1156 case PTR_TO_SOCK_COMMON_OR_NULL:
1157 reg->type = PTR_TO_SOCK_COMMON;
1159 case PTR_TO_TCP_SOCK_OR_NULL:
1160 reg->type = PTR_TO_TCP_SOCK;
1162 case PTR_TO_BTF_ID_OR_NULL:
1163 reg->type = PTR_TO_BTF_ID;
1165 case PTR_TO_MEM_OR_NULL:
1166 reg->type = PTR_TO_MEM;
1168 case PTR_TO_RDONLY_BUF_OR_NULL:
1169 reg->type = PTR_TO_RDONLY_BUF;
1171 case PTR_TO_RDWR_BUF_OR_NULL:
1172 reg->type = PTR_TO_RDWR_BUF;
1175 WARN_ONCE(1, "unknown nullable register type");
1179 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
1181 return type_is_pkt_pointer(reg->type);
1184 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
1186 return reg_is_pkt_pointer(reg) ||
1187 reg->type == PTR_TO_PACKET_END;
1190 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
1191 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
1192 enum bpf_reg_type which)
1194 /* The register can already have a range from prior markings.
1195 * This is fine as long as it hasn't been advanced from its
1198 return reg->type == which &&
1201 tnum_equals_const(reg->var_off, 0);
1204 /* Reset the min/max bounds of a register */
1205 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
1207 reg->smin_value = S64_MIN;
1208 reg->smax_value = S64_MAX;
1209 reg->umin_value = 0;
1210 reg->umax_value = U64_MAX;
1212 reg->s32_min_value = S32_MIN;
1213 reg->s32_max_value = S32_MAX;
1214 reg->u32_min_value = 0;
1215 reg->u32_max_value = U32_MAX;
1218 static void __mark_reg64_unbounded(struct bpf_reg_state *reg)
1220 reg->smin_value = S64_MIN;
1221 reg->smax_value = S64_MAX;
1222 reg->umin_value = 0;
1223 reg->umax_value = U64_MAX;
1226 static void __mark_reg32_unbounded(struct bpf_reg_state *reg)
1228 reg->s32_min_value = S32_MIN;
1229 reg->s32_max_value = S32_MAX;
1230 reg->u32_min_value = 0;
1231 reg->u32_max_value = U32_MAX;
1234 static void __update_reg32_bounds(struct bpf_reg_state *reg)
1236 struct tnum var32_off = tnum_subreg(reg->var_off);
1238 /* min signed is max(sign bit) | min(other bits) */
1239 reg->s32_min_value = max_t(s32, reg->s32_min_value,
1240 var32_off.value | (var32_off.mask & S32_MIN));
1241 /* max signed is min(sign bit) | max(other bits) */
1242 reg->s32_max_value = min_t(s32, reg->s32_max_value,
1243 var32_off.value | (var32_off.mask & S32_MAX));
1244 reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)var32_off.value);
1245 reg->u32_max_value = min(reg->u32_max_value,
1246 (u32)(var32_off.value | var32_off.mask));
1249 static void __update_reg64_bounds(struct bpf_reg_state *reg)
1251 /* min signed is max(sign bit) | min(other bits) */
1252 reg->smin_value = max_t(s64, reg->smin_value,
1253 reg->var_off.value | (reg->var_off.mask & S64_MIN));
1254 /* max signed is min(sign bit) | max(other bits) */
1255 reg->smax_value = min_t(s64, reg->smax_value,
1256 reg->var_off.value | (reg->var_off.mask & S64_MAX));
1257 reg->umin_value = max(reg->umin_value, reg->var_off.value);
1258 reg->umax_value = min(reg->umax_value,
1259 reg->var_off.value | reg->var_off.mask);
1262 static void __update_reg_bounds(struct bpf_reg_state *reg)
1264 __update_reg32_bounds(reg);
1265 __update_reg64_bounds(reg);
1268 /* Uses signed min/max values to inform unsigned, and vice-versa */
1269 static void __reg32_deduce_bounds(struct bpf_reg_state *reg)
1271 /* Learn sign from signed bounds.
1272 * If we cannot cross the sign boundary, then signed and unsigned bounds
1273 * are the same, so combine. This works even in the negative case, e.g.
1274 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1276 if (reg->s32_min_value >= 0 || reg->s32_max_value < 0) {
1277 reg->s32_min_value = reg->u32_min_value =
1278 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1279 reg->s32_max_value = reg->u32_max_value =
1280 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1283 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1284 * boundary, so we must be careful.
1286 if ((s32)reg->u32_max_value >= 0) {
1287 /* Positive. We can't learn anything from the smin, but smax
1288 * is positive, hence safe.
1290 reg->s32_min_value = reg->u32_min_value;
1291 reg->s32_max_value = reg->u32_max_value =
1292 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1293 } else if ((s32)reg->u32_min_value < 0) {
1294 /* Negative. We can't learn anything from the smax, but smin
1295 * is negative, hence safe.
1297 reg->s32_min_value = reg->u32_min_value =
1298 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1299 reg->s32_max_value = reg->u32_max_value;
1303 static void __reg64_deduce_bounds(struct bpf_reg_state *reg)
1305 /* Learn sign from signed bounds.
1306 * If we cannot cross the sign boundary, then signed and unsigned bounds
1307 * are the same, so combine. This works even in the negative case, e.g.
1308 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1310 if (reg->smin_value >= 0 || reg->smax_value < 0) {
1311 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1313 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1317 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1318 * boundary, so we must be careful.
1320 if ((s64)reg->umax_value >= 0) {
1321 /* Positive. We can't learn anything from the smin, but smax
1322 * is positive, hence safe.
1324 reg->smin_value = reg->umin_value;
1325 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1327 } else if ((s64)reg->umin_value < 0) {
1328 /* Negative. We can't learn anything from the smax, but smin
1329 * is negative, hence safe.
1331 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1333 reg->smax_value = reg->umax_value;
1337 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
1339 __reg32_deduce_bounds(reg);
1340 __reg64_deduce_bounds(reg);
1343 /* Attempts to improve var_off based on unsigned min/max information */
1344 static void __reg_bound_offset(struct bpf_reg_state *reg)
1346 struct tnum var64_off = tnum_intersect(reg->var_off,
1347 tnum_range(reg->umin_value,
1349 struct tnum var32_off = tnum_intersect(tnum_subreg(reg->var_off),
1350 tnum_range(reg->u32_min_value,
1351 reg->u32_max_value));
1353 reg->var_off = tnum_or(tnum_clear_subreg(var64_off), var32_off);
1356 static void __reg_assign_32_into_64(struct bpf_reg_state *reg)
1358 reg->umin_value = reg->u32_min_value;
1359 reg->umax_value = reg->u32_max_value;
1360 /* Attempt to pull 32-bit signed bounds into 64-bit bounds
1361 * but must be positive otherwise set to worse case bounds
1362 * and refine later from tnum.
1364 if (reg->s32_min_value >= 0 && reg->s32_max_value >= 0)
1365 reg->smax_value = reg->s32_max_value;
1367 reg->smax_value = U32_MAX;
1368 if (reg->s32_min_value >= 0)
1369 reg->smin_value = reg->s32_min_value;
1371 reg->smin_value = 0;
1374 static void __reg_combine_32_into_64(struct bpf_reg_state *reg)
1376 /* special case when 64-bit register has upper 32-bit register
1377 * zeroed. Typically happens after zext or <<32, >>32 sequence
1378 * allowing us to use 32-bit bounds directly,
1380 if (tnum_equals_const(tnum_clear_subreg(reg->var_off), 0)) {
1381 __reg_assign_32_into_64(reg);
1383 /* Otherwise the best we can do is push lower 32bit known and
1384 * unknown bits into register (var_off set from jmp logic)
1385 * then learn as much as possible from the 64-bit tnum
1386 * known and unknown bits. The previous smin/smax bounds are
1387 * invalid here because of jmp32 compare so mark them unknown
1388 * so they do not impact tnum bounds calculation.
1390 __mark_reg64_unbounded(reg);
1391 __update_reg_bounds(reg);
1394 /* Intersecting with the old var_off might have improved our bounds
1395 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1396 * then new var_off is (0; 0x7f...fc) which improves our umax.
1398 __reg_deduce_bounds(reg);
1399 __reg_bound_offset(reg);
1400 __update_reg_bounds(reg);
1403 static bool __reg64_bound_s32(s64 a)
1405 return a > S32_MIN && a < S32_MAX;
1408 static bool __reg64_bound_u32(u64 a)
1410 return a > U32_MIN && a < U32_MAX;
1413 static void __reg_combine_64_into_32(struct bpf_reg_state *reg)
1415 __mark_reg32_unbounded(reg);
1417 if (__reg64_bound_s32(reg->smin_value) && __reg64_bound_s32(reg->smax_value)) {
1418 reg->s32_min_value = (s32)reg->smin_value;
1419 reg->s32_max_value = (s32)reg->smax_value;
1421 if (__reg64_bound_u32(reg->umin_value) && __reg64_bound_u32(reg->umax_value)) {
1422 reg->u32_min_value = (u32)reg->umin_value;
1423 reg->u32_max_value = (u32)reg->umax_value;
1426 /* Intersecting with the old var_off might have improved our bounds
1427 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1428 * then new var_off is (0; 0x7f...fc) which improves our umax.
1430 __reg_deduce_bounds(reg);
1431 __reg_bound_offset(reg);
1432 __update_reg_bounds(reg);
1435 /* Mark a register as having a completely unknown (scalar) value. */
1436 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
1437 struct bpf_reg_state *reg)
1440 * Clear type, id, off, and union(map_ptr, range) and
1441 * padding between 'type' and union
1443 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
1444 reg->type = SCALAR_VALUE;
1445 reg->var_off = tnum_unknown;
1447 reg->precise = env->subprog_cnt > 1 || !env->bpf_capable;
1448 __mark_reg_unbounded(reg);
1451 static void mark_reg_unknown(struct bpf_verifier_env *env,
1452 struct bpf_reg_state *regs, u32 regno)
1454 if (WARN_ON(regno >= MAX_BPF_REG)) {
1455 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
1456 /* Something bad happened, let's kill all regs except FP */
1457 for (regno = 0; regno < BPF_REG_FP; regno++)
1458 __mark_reg_not_init(env, regs + regno);
1461 __mark_reg_unknown(env, regs + regno);
1464 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1465 struct bpf_reg_state *reg)
1467 __mark_reg_unknown(env, reg);
1468 reg->type = NOT_INIT;
1471 static void mark_reg_not_init(struct bpf_verifier_env *env,
1472 struct bpf_reg_state *regs, u32 regno)
1474 if (WARN_ON(regno >= MAX_BPF_REG)) {
1475 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
1476 /* Something bad happened, let's kill all regs except FP */
1477 for (regno = 0; regno < BPF_REG_FP; regno++)
1478 __mark_reg_not_init(env, regs + regno);
1481 __mark_reg_not_init(env, regs + regno);
1484 static void mark_btf_ld_reg(struct bpf_verifier_env *env,
1485 struct bpf_reg_state *regs, u32 regno,
1486 enum bpf_reg_type reg_type,
1487 struct btf *btf, u32 btf_id)
1489 if (reg_type == SCALAR_VALUE) {
1490 mark_reg_unknown(env, regs, regno);
1493 mark_reg_known_zero(env, regs, regno);
1494 regs[regno].type = PTR_TO_BTF_ID;
1495 regs[regno].btf = btf;
1496 regs[regno].btf_id = btf_id;
1499 #define DEF_NOT_SUBREG (0)
1500 static void init_reg_state(struct bpf_verifier_env *env,
1501 struct bpf_func_state *state)
1503 struct bpf_reg_state *regs = state->regs;
1506 for (i = 0; i < MAX_BPF_REG; i++) {
1507 mark_reg_not_init(env, regs, i);
1508 regs[i].live = REG_LIVE_NONE;
1509 regs[i].parent = NULL;
1510 regs[i].subreg_def = DEF_NOT_SUBREG;
1514 regs[BPF_REG_FP].type = PTR_TO_STACK;
1515 mark_reg_known_zero(env, regs, BPF_REG_FP);
1516 regs[BPF_REG_FP].frameno = state->frameno;
1519 #define BPF_MAIN_FUNC (-1)
1520 static void init_func_state(struct bpf_verifier_env *env,
1521 struct bpf_func_state *state,
1522 int callsite, int frameno, int subprogno)
1524 state->callsite = callsite;
1525 state->frameno = frameno;
1526 state->subprogno = subprogno;
1527 init_reg_state(env, state);
1531 SRC_OP, /* register is used as source operand */
1532 DST_OP, /* register is used as destination operand */
1533 DST_OP_NO_MARK /* same as above, check only, don't mark */
1536 static int cmp_subprogs(const void *a, const void *b)
1538 return ((struct bpf_subprog_info *)a)->start -
1539 ((struct bpf_subprog_info *)b)->start;
1542 static int find_subprog(struct bpf_verifier_env *env, int off)
1544 struct bpf_subprog_info *p;
1546 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
1547 sizeof(env->subprog_info[0]), cmp_subprogs);
1550 return p - env->subprog_info;
1554 static int add_subprog(struct bpf_verifier_env *env, int off)
1556 int insn_cnt = env->prog->len;
1559 if (off >= insn_cnt || off < 0) {
1560 verbose(env, "call to invalid destination\n");
1563 ret = find_subprog(env, off);
1566 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
1567 verbose(env, "too many subprograms\n");
1570 /* determine subprog starts. The end is one before the next starts */
1571 env->subprog_info[env->subprog_cnt++].start = off;
1572 sort(env->subprog_info, env->subprog_cnt,
1573 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
1574 return env->subprog_cnt - 1;
1577 struct bpf_kfunc_desc {
1578 struct btf_func_model func_model;
1583 #define MAX_KFUNC_DESCS 256
1584 struct bpf_kfunc_desc_tab {
1585 struct bpf_kfunc_desc descs[MAX_KFUNC_DESCS];
1589 static int kfunc_desc_cmp_by_id(const void *a, const void *b)
1591 const struct bpf_kfunc_desc *d0 = a;
1592 const struct bpf_kfunc_desc *d1 = b;
1594 /* func_id is not greater than BTF_MAX_TYPE */
1595 return d0->func_id - d1->func_id;
1598 static const struct bpf_kfunc_desc *
1599 find_kfunc_desc(const struct bpf_prog *prog, u32 func_id)
1601 struct bpf_kfunc_desc desc = {
1604 struct bpf_kfunc_desc_tab *tab;
1606 tab = prog->aux->kfunc_tab;
1607 return bsearch(&desc, tab->descs, tab->nr_descs,
1608 sizeof(tab->descs[0]), kfunc_desc_cmp_by_id);
1611 static int add_kfunc_call(struct bpf_verifier_env *env, u32 func_id)
1613 const struct btf_type *func, *func_proto;
1614 struct bpf_kfunc_desc_tab *tab;
1615 struct bpf_prog_aux *prog_aux;
1616 struct bpf_kfunc_desc *desc;
1617 const char *func_name;
1621 prog_aux = env->prog->aux;
1622 tab = prog_aux->kfunc_tab;
1625 verbose(env, "calling kernel function is not supported without CONFIG_DEBUG_INFO_BTF\n");
1629 if (!env->prog->jit_requested) {
1630 verbose(env, "JIT is required for calling kernel function\n");
1634 if (!bpf_jit_supports_kfunc_call()) {
1635 verbose(env, "JIT does not support calling kernel function\n");
1639 if (!env->prog->gpl_compatible) {
1640 verbose(env, "cannot call kernel function from non-GPL compatible program\n");
1644 tab = kzalloc(sizeof(*tab), GFP_KERNEL);
1647 prog_aux->kfunc_tab = tab;
1650 if (find_kfunc_desc(env->prog, func_id))
1653 if (tab->nr_descs == MAX_KFUNC_DESCS) {
1654 verbose(env, "too many different kernel function calls\n");
1658 func = btf_type_by_id(btf_vmlinux, func_id);
1659 if (!func || !btf_type_is_func(func)) {
1660 verbose(env, "kernel btf_id %u is not a function\n",
1664 func_proto = btf_type_by_id(btf_vmlinux, func->type);
1665 if (!func_proto || !btf_type_is_func_proto(func_proto)) {
1666 verbose(env, "kernel function btf_id %u does not have a valid func_proto\n",
1671 func_name = btf_name_by_offset(btf_vmlinux, func->name_off);
1672 addr = kallsyms_lookup_name(func_name);
1674 verbose(env, "cannot find address for kernel function %s\n",
1679 desc = &tab->descs[tab->nr_descs++];
1680 desc->func_id = func_id;
1681 desc->imm = BPF_CAST_CALL(addr) - __bpf_call_base;
1682 err = btf_distill_func_proto(&env->log, btf_vmlinux,
1683 func_proto, func_name,
1686 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
1687 kfunc_desc_cmp_by_id, NULL);
1691 static int kfunc_desc_cmp_by_imm(const void *a, const void *b)
1693 const struct bpf_kfunc_desc *d0 = a;
1694 const struct bpf_kfunc_desc *d1 = b;
1696 if (d0->imm > d1->imm)
1698 else if (d0->imm < d1->imm)
1703 static void sort_kfunc_descs_by_imm(struct bpf_prog *prog)
1705 struct bpf_kfunc_desc_tab *tab;
1707 tab = prog->aux->kfunc_tab;
1711 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
1712 kfunc_desc_cmp_by_imm, NULL);
1715 bool bpf_prog_has_kfunc_call(const struct bpf_prog *prog)
1717 return !!prog->aux->kfunc_tab;
1720 const struct btf_func_model *
1721 bpf_jit_find_kfunc_model(const struct bpf_prog *prog,
1722 const struct bpf_insn *insn)
1724 const struct bpf_kfunc_desc desc = {
1727 const struct bpf_kfunc_desc *res;
1728 struct bpf_kfunc_desc_tab *tab;
1730 tab = prog->aux->kfunc_tab;
1731 res = bsearch(&desc, tab->descs, tab->nr_descs,
1732 sizeof(tab->descs[0]), kfunc_desc_cmp_by_imm);
1734 return res ? &res->func_model : NULL;
1737 static int add_subprog_and_kfunc(struct bpf_verifier_env *env)
1739 struct bpf_subprog_info *subprog = env->subprog_info;
1740 struct bpf_insn *insn = env->prog->insnsi;
1741 int i, ret, insn_cnt = env->prog->len;
1743 /* Add entry function. */
1744 ret = add_subprog(env, 0);
1748 for (i = 0; i < insn_cnt; i++, insn++) {
1749 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn) &&
1750 !bpf_pseudo_kfunc_call(insn))
1753 if (!env->bpf_capable) {
1754 verbose(env, "loading/calling other bpf or kernel functions are allowed for CAP_BPF and CAP_SYS_ADMIN\n");
1758 if (bpf_pseudo_func(insn)) {
1759 ret = add_subprog(env, i + insn->imm + 1);
1761 /* remember subprog */
1763 } else if (bpf_pseudo_call(insn)) {
1764 ret = add_subprog(env, i + insn->imm + 1);
1766 ret = add_kfunc_call(env, insn->imm);
1773 /* Add a fake 'exit' subprog which could simplify subprog iteration
1774 * logic. 'subprog_cnt' should not be increased.
1776 subprog[env->subprog_cnt].start = insn_cnt;
1778 if (env->log.level & BPF_LOG_LEVEL2)
1779 for (i = 0; i < env->subprog_cnt; i++)
1780 verbose(env, "func#%d @%d\n", i, subprog[i].start);
1785 static int check_subprogs(struct bpf_verifier_env *env)
1787 int i, subprog_start, subprog_end, off, cur_subprog = 0;
1788 struct bpf_subprog_info *subprog = env->subprog_info;
1789 struct bpf_insn *insn = env->prog->insnsi;
1790 int insn_cnt = env->prog->len;
1792 /* now check that all jumps are within the same subprog */
1793 subprog_start = subprog[cur_subprog].start;
1794 subprog_end = subprog[cur_subprog + 1].start;
1795 for (i = 0; i < insn_cnt; i++) {
1796 u8 code = insn[i].code;
1798 if (code == (BPF_JMP | BPF_CALL) &&
1799 insn[i].imm == BPF_FUNC_tail_call &&
1800 insn[i].src_reg != BPF_PSEUDO_CALL)
1801 subprog[cur_subprog].has_tail_call = true;
1802 if (BPF_CLASS(code) == BPF_LD &&
1803 (BPF_MODE(code) == BPF_ABS || BPF_MODE(code) == BPF_IND))
1804 subprog[cur_subprog].has_ld_abs = true;
1805 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
1807 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
1809 off = i + insn[i].off + 1;
1810 if (off < subprog_start || off >= subprog_end) {
1811 verbose(env, "jump out of range from insn %d to %d\n", i, off);
1815 if (i == subprog_end - 1) {
1816 /* to avoid fall-through from one subprog into another
1817 * the last insn of the subprog should be either exit
1818 * or unconditional jump back
1820 if (code != (BPF_JMP | BPF_EXIT) &&
1821 code != (BPF_JMP | BPF_JA)) {
1822 verbose(env, "last insn is not an exit or jmp\n");
1825 subprog_start = subprog_end;
1827 if (cur_subprog < env->subprog_cnt)
1828 subprog_end = subprog[cur_subprog + 1].start;
1834 /* Parentage chain of this register (or stack slot) should take care of all
1835 * issues like callee-saved registers, stack slot allocation time, etc.
1837 static int mark_reg_read(struct bpf_verifier_env *env,
1838 const struct bpf_reg_state *state,
1839 struct bpf_reg_state *parent, u8 flag)
1841 bool writes = parent == state->parent; /* Observe write marks */
1845 /* if read wasn't screened by an earlier write ... */
1846 if (writes && state->live & REG_LIVE_WRITTEN)
1848 if (parent->live & REG_LIVE_DONE) {
1849 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
1850 reg_type_str[parent->type],
1851 parent->var_off.value, parent->off);
1854 /* The first condition is more likely to be true than the
1855 * second, checked it first.
1857 if ((parent->live & REG_LIVE_READ) == flag ||
1858 parent->live & REG_LIVE_READ64)
1859 /* The parentage chain never changes and
1860 * this parent was already marked as LIVE_READ.
1861 * There is no need to keep walking the chain again and
1862 * keep re-marking all parents as LIVE_READ.
1863 * This case happens when the same register is read
1864 * multiple times without writes into it in-between.
1865 * Also, if parent has the stronger REG_LIVE_READ64 set,
1866 * then no need to set the weak REG_LIVE_READ32.
1869 /* ... then we depend on parent's value */
1870 parent->live |= flag;
1871 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
1872 if (flag == REG_LIVE_READ64)
1873 parent->live &= ~REG_LIVE_READ32;
1875 parent = state->parent;
1880 if (env->longest_mark_read_walk < cnt)
1881 env->longest_mark_read_walk = cnt;
1885 /* This function is supposed to be used by the following 32-bit optimization
1886 * code only. It returns TRUE if the source or destination register operates
1887 * on 64-bit, otherwise return FALSE.
1889 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
1890 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
1895 class = BPF_CLASS(code);
1897 if (class == BPF_JMP) {
1898 /* BPF_EXIT for "main" will reach here. Return TRUE
1903 if (op == BPF_CALL) {
1904 /* BPF to BPF call will reach here because of marking
1905 * caller saved clobber with DST_OP_NO_MARK for which we
1906 * don't care the register def because they are anyway
1907 * marked as NOT_INIT already.
1909 if (insn->src_reg == BPF_PSEUDO_CALL)
1911 /* Helper call will reach here because of arg type
1912 * check, conservatively return TRUE.
1921 if (class == BPF_ALU64 || class == BPF_JMP ||
1922 /* BPF_END always use BPF_ALU class. */
1923 (class == BPF_ALU && op == BPF_END && insn->imm == 64))
1926 if (class == BPF_ALU || class == BPF_JMP32)
1929 if (class == BPF_LDX) {
1931 return BPF_SIZE(code) == BPF_DW;
1932 /* LDX source must be ptr. */
1936 if (class == BPF_STX) {
1937 /* BPF_STX (including atomic variants) has multiple source
1938 * operands, one of which is a ptr. Check whether the caller is
1941 if (t == SRC_OP && reg->type != SCALAR_VALUE)
1943 return BPF_SIZE(code) == BPF_DW;
1946 if (class == BPF_LD) {
1947 u8 mode = BPF_MODE(code);
1950 if (mode == BPF_IMM)
1953 /* Both LD_IND and LD_ABS return 32-bit data. */
1957 /* Implicit ctx ptr. */
1958 if (regno == BPF_REG_6)
1961 /* Explicit source could be any width. */
1965 if (class == BPF_ST)
1966 /* The only source register for BPF_ST is a ptr. */
1969 /* Conservatively return true at default. */
1973 /* Return the regno defined by the insn, or -1. */
1974 static int insn_def_regno(const struct bpf_insn *insn)
1976 switch (BPF_CLASS(insn->code)) {
1982 if (BPF_MODE(insn->code) == BPF_ATOMIC &&
1983 (insn->imm & BPF_FETCH)) {
1984 if (insn->imm == BPF_CMPXCHG)
1987 return insn->src_reg;
1992 return insn->dst_reg;
1996 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
1997 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
1999 int dst_reg = insn_def_regno(insn);
2004 return !is_reg64(env, insn, dst_reg, NULL, DST_OP);
2007 static void mark_insn_zext(struct bpf_verifier_env *env,
2008 struct bpf_reg_state *reg)
2010 s32 def_idx = reg->subreg_def;
2012 if (def_idx == DEF_NOT_SUBREG)
2015 env->insn_aux_data[def_idx - 1].zext_dst = true;
2016 /* The dst will be zero extended, so won't be sub-register anymore. */
2017 reg->subreg_def = DEF_NOT_SUBREG;
2020 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
2021 enum reg_arg_type t)
2023 struct bpf_verifier_state *vstate = env->cur_state;
2024 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2025 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
2026 struct bpf_reg_state *reg, *regs = state->regs;
2029 if (regno >= MAX_BPF_REG) {
2030 verbose(env, "R%d is invalid\n", regno);
2035 rw64 = is_reg64(env, insn, regno, reg, t);
2037 /* check whether register used as source operand can be read */
2038 if (reg->type == NOT_INIT) {
2039 verbose(env, "R%d !read_ok\n", regno);
2042 /* We don't need to worry about FP liveness because it's read-only */
2043 if (regno == BPF_REG_FP)
2047 mark_insn_zext(env, reg);
2049 return mark_reg_read(env, reg, reg->parent,
2050 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
2052 /* check whether register used as dest operand can be written to */
2053 if (regno == BPF_REG_FP) {
2054 verbose(env, "frame pointer is read only\n");
2057 reg->live |= REG_LIVE_WRITTEN;
2058 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
2060 mark_reg_unknown(env, regs, regno);
2065 /* for any branch, call, exit record the history of jmps in the given state */
2066 static int push_jmp_history(struct bpf_verifier_env *env,
2067 struct bpf_verifier_state *cur)
2069 u32 cnt = cur->jmp_history_cnt;
2070 struct bpf_idx_pair *p;
2073 p = krealloc(cur->jmp_history, cnt * sizeof(*p), GFP_USER);
2076 p[cnt - 1].idx = env->insn_idx;
2077 p[cnt - 1].prev_idx = env->prev_insn_idx;
2078 cur->jmp_history = p;
2079 cur->jmp_history_cnt = cnt;
2083 /* Backtrack one insn at a time. If idx is not at the top of recorded
2084 * history then previous instruction came from straight line execution.
2086 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
2091 if (cnt && st->jmp_history[cnt - 1].idx == i) {
2092 i = st->jmp_history[cnt - 1].prev_idx;
2100 static const char *disasm_kfunc_name(void *data, const struct bpf_insn *insn)
2102 const struct btf_type *func;
2104 if (insn->src_reg != BPF_PSEUDO_KFUNC_CALL)
2107 func = btf_type_by_id(btf_vmlinux, insn->imm);
2108 return btf_name_by_offset(btf_vmlinux, func->name_off);
2111 /* For given verifier state backtrack_insn() is called from the last insn to
2112 * the first insn. Its purpose is to compute a bitmask of registers and
2113 * stack slots that needs precision in the parent verifier state.
2115 static int backtrack_insn(struct bpf_verifier_env *env, int idx,
2116 u32 *reg_mask, u64 *stack_mask)
2118 const struct bpf_insn_cbs cbs = {
2119 .cb_call = disasm_kfunc_name,
2120 .cb_print = verbose,
2121 .private_data = env,
2123 struct bpf_insn *insn = env->prog->insnsi + idx;
2124 u8 class = BPF_CLASS(insn->code);
2125 u8 opcode = BPF_OP(insn->code);
2126 u8 mode = BPF_MODE(insn->code);
2127 u32 dreg = 1u << insn->dst_reg;
2128 u32 sreg = 1u << insn->src_reg;
2131 if (insn->code == 0)
2133 if (env->log.level & BPF_LOG_LEVEL) {
2134 verbose(env, "regs=%x stack=%llx before ", *reg_mask, *stack_mask);
2135 verbose(env, "%d: ", idx);
2136 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
2139 if (class == BPF_ALU || class == BPF_ALU64) {
2140 if (!(*reg_mask & dreg))
2142 if (opcode == BPF_MOV) {
2143 if (BPF_SRC(insn->code) == BPF_X) {
2145 * dreg needs precision after this insn
2146 * sreg needs precision before this insn
2152 * dreg needs precision after this insn.
2153 * Corresponding register is already marked
2154 * as precise=true in this verifier state.
2155 * No further markings in parent are necessary
2160 if (BPF_SRC(insn->code) == BPF_X) {
2162 * both dreg and sreg need precision
2167 * dreg still needs precision before this insn
2170 } else if (class == BPF_LDX) {
2171 if (!(*reg_mask & dreg))
2175 /* scalars can only be spilled into stack w/o losing precision.
2176 * Load from any other memory can be zero extended.
2177 * The desire to keep that precision is already indicated
2178 * by 'precise' mark in corresponding register of this state.
2179 * No further tracking necessary.
2181 if (insn->src_reg != BPF_REG_FP)
2183 if (BPF_SIZE(insn->code) != BPF_DW)
2186 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
2187 * that [fp - off] slot contains scalar that needs to be
2188 * tracked with precision
2190 spi = (-insn->off - 1) / BPF_REG_SIZE;
2192 verbose(env, "BUG spi %d\n", spi);
2193 WARN_ONCE(1, "verifier backtracking bug");
2196 *stack_mask |= 1ull << spi;
2197 } else if (class == BPF_STX || class == BPF_ST) {
2198 if (*reg_mask & dreg)
2199 /* stx & st shouldn't be using _scalar_ dst_reg
2200 * to access memory. It means backtracking
2201 * encountered a case of pointer subtraction.
2204 /* scalars can only be spilled into stack */
2205 if (insn->dst_reg != BPF_REG_FP)
2207 if (BPF_SIZE(insn->code) != BPF_DW)
2209 spi = (-insn->off - 1) / BPF_REG_SIZE;
2211 verbose(env, "BUG spi %d\n", spi);
2212 WARN_ONCE(1, "verifier backtracking bug");
2215 if (!(*stack_mask & (1ull << spi)))
2217 *stack_mask &= ~(1ull << spi);
2218 if (class == BPF_STX)
2220 } else if (class == BPF_JMP || class == BPF_JMP32) {
2221 if (opcode == BPF_CALL) {
2222 if (insn->src_reg == BPF_PSEUDO_CALL)
2224 /* regular helper call sets R0 */
2226 if (*reg_mask & 0x3f) {
2227 /* if backtracing was looking for registers R1-R5
2228 * they should have been found already.
2230 verbose(env, "BUG regs %x\n", *reg_mask);
2231 WARN_ONCE(1, "verifier backtracking bug");
2234 } else if (opcode == BPF_EXIT) {
2237 } else if (class == BPF_LD) {
2238 if (!(*reg_mask & dreg))
2241 /* It's ld_imm64 or ld_abs or ld_ind.
2242 * For ld_imm64 no further tracking of precision
2243 * into parent is necessary
2245 if (mode == BPF_IND || mode == BPF_ABS)
2246 /* to be analyzed */
2252 /* the scalar precision tracking algorithm:
2253 * . at the start all registers have precise=false.
2254 * . scalar ranges are tracked as normal through alu and jmp insns.
2255 * . once precise value of the scalar register is used in:
2256 * . ptr + scalar alu
2257 * . if (scalar cond K|scalar)
2258 * . helper_call(.., scalar, ...) where ARG_CONST is expected
2259 * backtrack through the verifier states and mark all registers and
2260 * stack slots with spilled constants that these scalar regisers
2261 * should be precise.
2262 * . during state pruning two registers (or spilled stack slots)
2263 * are equivalent if both are not precise.
2265 * Note the verifier cannot simply walk register parentage chain,
2266 * since many different registers and stack slots could have been
2267 * used to compute single precise scalar.
2269 * The approach of starting with precise=true for all registers and then
2270 * backtrack to mark a register as not precise when the verifier detects
2271 * that program doesn't care about specific value (e.g., when helper
2272 * takes register as ARG_ANYTHING parameter) is not safe.
2274 * It's ok to walk single parentage chain of the verifier states.
2275 * It's possible that this backtracking will go all the way till 1st insn.
2276 * All other branches will be explored for needing precision later.
2278 * The backtracking needs to deal with cases like:
2279 * 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)
2282 * if r5 > 0x79f goto pc+7
2283 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
2286 * call bpf_perf_event_output#25
2287 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
2291 * call foo // uses callee's r6 inside to compute r0
2295 * to track above reg_mask/stack_mask needs to be independent for each frame.
2297 * Also if parent's curframe > frame where backtracking started,
2298 * the verifier need to mark registers in both frames, otherwise callees
2299 * may incorrectly prune callers. This is similar to
2300 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
2302 * For now backtracking falls back into conservative marking.
2304 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
2305 struct bpf_verifier_state *st)
2307 struct bpf_func_state *func;
2308 struct bpf_reg_state *reg;
2311 /* big hammer: mark all scalars precise in this path.
2312 * pop_stack may still get !precise scalars.
2314 for (; st; st = st->parent)
2315 for (i = 0; i <= st->curframe; i++) {
2316 func = st->frame[i];
2317 for (j = 0; j < BPF_REG_FP; j++) {
2318 reg = &func->regs[j];
2319 if (reg->type != SCALAR_VALUE)
2321 reg->precise = true;
2323 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
2324 if (func->stack[j].slot_type[0] != STACK_SPILL)
2326 reg = &func->stack[j].spilled_ptr;
2327 if (reg->type != SCALAR_VALUE)
2329 reg->precise = true;
2334 static int __mark_chain_precision(struct bpf_verifier_env *env, int regno,
2337 struct bpf_verifier_state *st = env->cur_state;
2338 int first_idx = st->first_insn_idx;
2339 int last_idx = env->insn_idx;
2340 struct bpf_func_state *func;
2341 struct bpf_reg_state *reg;
2342 u32 reg_mask = regno >= 0 ? 1u << regno : 0;
2343 u64 stack_mask = spi >= 0 ? 1ull << spi : 0;
2344 bool skip_first = true;
2345 bool new_marks = false;
2348 if (!env->bpf_capable)
2351 func = st->frame[st->curframe];
2353 reg = &func->regs[regno];
2354 if (reg->type != SCALAR_VALUE) {
2355 WARN_ONCE(1, "backtracing misuse");
2362 reg->precise = true;
2366 if (func->stack[spi].slot_type[0] != STACK_SPILL) {
2370 reg = &func->stack[spi].spilled_ptr;
2371 if (reg->type != SCALAR_VALUE) {
2379 reg->precise = true;
2385 if (!reg_mask && !stack_mask)
2388 DECLARE_BITMAP(mask, 64);
2389 u32 history = st->jmp_history_cnt;
2391 if (env->log.level & BPF_LOG_LEVEL)
2392 verbose(env, "last_idx %d first_idx %d\n", last_idx, first_idx);
2393 for (i = last_idx;;) {
2398 err = backtrack_insn(env, i, ®_mask, &stack_mask);
2400 if (err == -ENOTSUPP) {
2401 mark_all_scalars_precise(env, st);
2406 if (!reg_mask && !stack_mask)
2407 /* Found assignment(s) into tracked register in this state.
2408 * Since this state is already marked, just return.
2409 * Nothing to be tracked further in the parent state.
2414 i = get_prev_insn_idx(st, i, &history);
2415 if (i >= env->prog->len) {
2416 /* This can happen if backtracking reached insn 0
2417 * and there are still reg_mask or stack_mask
2419 * It means the backtracking missed the spot where
2420 * particular register was initialized with a constant.
2422 verbose(env, "BUG backtracking idx %d\n", i);
2423 WARN_ONCE(1, "verifier backtracking bug");
2432 func = st->frame[st->curframe];
2433 bitmap_from_u64(mask, reg_mask);
2434 for_each_set_bit(i, mask, 32) {
2435 reg = &func->regs[i];
2436 if (reg->type != SCALAR_VALUE) {
2437 reg_mask &= ~(1u << i);
2442 reg->precise = true;
2445 bitmap_from_u64(mask, stack_mask);
2446 for_each_set_bit(i, mask, 64) {
2447 if (i >= func->allocated_stack / BPF_REG_SIZE) {
2448 /* the sequence of instructions:
2450 * 3: (7b) *(u64 *)(r3 -8) = r0
2451 * 4: (79) r4 = *(u64 *)(r10 -8)
2452 * doesn't contain jmps. It's backtracked
2453 * as a single block.
2454 * During backtracking insn 3 is not recognized as
2455 * stack access, so at the end of backtracking
2456 * stack slot fp-8 is still marked in stack_mask.
2457 * However the parent state may not have accessed
2458 * fp-8 and it's "unallocated" stack space.
2459 * In such case fallback to conservative.
2461 mark_all_scalars_precise(env, st);
2465 if (func->stack[i].slot_type[0] != STACK_SPILL) {
2466 stack_mask &= ~(1ull << i);
2469 reg = &func->stack[i].spilled_ptr;
2470 if (reg->type != SCALAR_VALUE) {
2471 stack_mask &= ~(1ull << i);
2476 reg->precise = true;
2478 if (env->log.level & BPF_LOG_LEVEL) {
2479 print_verifier_state(env, func);
2480 verbose(env, "parent %s regs=%x stack=%llx marks\n",
2481 new_marks ? "didn't have" : "already had",
2482 reg_mask, stack_mask);
2485 if (!reg_mask && !stack_mask)
2490 last_idx = st->last_insn_idx;
2491 first_idx = st->first_insn_idx;
2496 static int mark_chain_precision(struct bpf_verifier_env *env, int regno)
2498 return __mark_chain_precision(env, regno, -1);
2501 static int mark_chain_precision_stack(struct bpf_verifier_env *env, int spi)
2503 return __mark_chain_precision(env, -1, spi);
2506 static bool is_spillable_regtype(enum bpf_reg_type type)
2509 case PTR_TO_MAP_VALUE:
2510 case PTR_TO_MAP_VALUE_OR_NULL:
2514 case PTR_TO_PACKET_META:
2515 case PTR_TO_PACKET_END:
2516 case PTR_TO_FLOW_KEYS:
2517 case CONST_PTR_TO_MAP:
2519 case PTR_TO_SOCKET_OR_NULL:
2520 case PTR_TO_SOCK_COMMON:
2521 case PTR_TO_SOCK_COMMON_OR_NULL:
2522 case PTR_TO_TCP_SOCK:
2523 case PTR_TO_TCP_SOCK_OR_NULL:
2524 case PTR_TO_XDP_SOCK:
2526 case PTR_TO_BTF_ID_OR_NULL:
2527 case PTR_TO_RDONLY_BUF:
2528 case PTR_TO_RDONLY_BUF_OR_NULL:
2529 case PTR_TO_RDWR_BUF:
2530 case PTR_TO_RDWR_BUF_OR_NULL:
2531 case PTR_TO_PERCPU_BTF_ID:
2533 case PTR_TO_MEM_OR_NULL:
2535 case PTR_TO_MAP_KEY:
2542 /* Does this register contain a constant zero? */
2543 static bool register_is_null(struct bpf_reg_state *reg)
2545 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
2548 static bool register_is_const(struct bpf_reg_state *reg)
2550 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
2553 static bool __is_scalar_unbounded(struct bpf_reg_state *reg)
2555 return tnum_is_unknown(reg->var_off) &&
2556 reg->smin_value == S64_MIN && reg->smax_value == S64_MAX &&
2557 reg->umin_value == 0 && reg->umax_value == U64_MAX &&
2558 reg->s32_min_value == S32_MIN && reg->s32_max_value == S32_MAX &&
2559 reg->u32_min_value == 0 && reg->u32_max_value == U32_MAX;
2562 static bool register_is_bounded(struct bpf_reg_state *reg)
2564 return reg->type == SCALAR_VALUE && !__is_scalar_unbounded(reg);
2567 static bool __is_pointer_value(bool allow_ptr_leaks,
2568 const struct bpf_reg_state *reg)
2570 if (allow_ptr_leaks)
2573 return reg->type != SCALAR_VALUE;
2576 static void save_register_state(struct bpf_func_state *state,
2577 int spi, struct bpf_reg_state *reg)
2581 state->stack[spi].spilled_ptr = *reg;
2582 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2584 for (i = 0; i < BPF_REG_SIZE; i++)
2585 state->stack[spi].slot_type[i] = STACK_SPILL;
2588 /* check_stack_{read,write}_fixed_off functions track spill/fill of registers,
2589 * stack boundary and alignment are checked in check_mem_access()
2591 static int check_stack_write_fixed_off(struct bpf_verifier_env *env,
2592 /* stack frame we're writing to */
2593 struct bpf_func_state *state,
2594 int off, int size, int value_regno,
2597 struct bpf_func_state *cur; /* state of the current function */
2598 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
2599 u32 dst_reg = env->prog->insnsi[insn_idx].dst_reg;
2600 struct bpf_reg_state *reg = NULL;
2602 err = grow_stack_state(state, round_up(slot + 1, BPF_REG_SIZE));
2605 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
2606 * so it's aligned access and [off, off + size) are within stack limits
2608 if (!env->allow_ptr_leaks &&
2609 state->stack[spi].slot_type[0] == STACK_SPILL &&
2610 size != BPF_REG_SIZE) {
2611 verbose(env, "attempt to corrupt spilled pointer on stack\n");
2615 cur = env->cur_state->frame[env->cur_state->curframe];
2616 if (value_regno >= 0)
2617 reg = &cur->regs[value_regno];
2619 if (reg && size == BPF_REG_SIZE && register_is_bounded(reg) &&
2620 !register_is_null(reg) && env->bpf_capable) {
2621 if (dst_reg != BPF_REG_FP) {
2622 /* The backtracking logic can only recognize explicit
2623 * stack slot address like [fp - 8]. Other spill of
2624 * scalar via different register has to be conservative.
2625 * Backtrack from here and mark all registers as precise
2626 * that contributed into 'reg' being a constant.
2628 err = mark_chain_precision(env, value_regno);
2632 save_register_state(state, spi, reg);
2633 } else if (reg && is_spillable_regtype(reg->type)) {
2634 /* register containing pointer is being spilled into stack */
2635 if (size != BPF_REG_SIZE) {
2636 verbose_linfo(env, insn_idx, "; ");
2637 verbose(env, "invalid size of register spill\n");
2641 if (state != cur && reg->type == PTR_TO_STACK) {
2642 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
2646 if (!env->bypass_spec_v4) {
2647 bool sanitize = false;
2649 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
2650 register_is_const(&state->stack[spi].spilled_ptr))
2652 for (i = 0; i < BPF_REG_SIZE; i++)
2653 if (state->stack[spi].slot_type[i] == STACK_MISC) {
2658 int *poff = &env->insn_aux_data[insn_idx].sanitize_stack_off;
2659 int soff = (-spi - 1) * BPF_REG_SIZE;
2661 /* detected reuse of integer stack slot with a pointer
2662 * which means either llvm is reusing stack slot or
2663 * an attacker is trying to exploit CVE-2018-3639
2664 * (speculative store bypass)
2665 * Have to sanitize that slot with preemptive
2668 if (*poff && *poff != soff) {
2669 /* disallow programs where single insn stores
2670 * into two different stack slots, since verifier
2671 * cannot sanitize them
2674 "insn %d cannot access two stack slots fp%d and fp%d",
2675 insn_idx, *poff, soff);
2681 save_register_state(state, spi, reg);
2683 u8 type = STACK_MISC;
2685 /* regular write of data into stack destroys any spilled ptr */
2686 state->stack[spi].spilled_ptr.type = NOT_INIT;
2687 /* Mark slots as STACK_MISC if they belonged to spilled ptr. */
2688 if (state->stack[spi].slot_type[0] == STACK_SPILL)
2689 for (i = 0; i < BPF_REG_SIZE; i++)
2690 state->stack[spi].slot_type[i] = STACK_MISC;
2692 /* only mark the slot as written if all 8 bytes were written
2693 * otherwise read propagation may incorrectly stop too soon
2694 * when stack slots are partially written.
2695 * This heuristic means that read propagation will be
2696 * conservative, since it will add reg_live_read marks
2697 * to stack slots all the way to first state when programs
2698 * writes+reads less than 8 bytes
2700 if (size == BPF_REG_SIZE)
2701 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2703 /* when we zero initialize stack slots mark them as such */
2704 if (reg && register_is_null(reg)) {
2705 /* backtracking doesn't work for STACK_ZERO yet. */
2706 err = mark_chain_precision(env, value_regno);
2712 /* Mark slots affected by this stack write. */
2713 for (i = 0; i < size; i++)
2714 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
2720 /* Write the stack: 'stack[ptr_regno + off] = value_regno'. 'ptr_regno' is
2721 * known to contain a variable offset.
2722 * This function checks whether the write is permitted and conservatively
2723 * tracks the effects of the write, considering that each stack slot in the
2724 * dynamic range is potentially written to.
2726 * 'off' includes 'regno->off'.
2727 * 'value_regno' can be -1, meaning that an unknown value is being written to
2730 * Spilled pointers in range are not marked as written because we don't know
2731 * what's going to be actually written. This means that read propagation for
2732 * future reads cannot be terminated by this write.
2734 * For privileged programs, uninitialized stack slots are considered
2735 * initialized by this write (even though we don't know exactly what offsets
2736 * are going to be written to). The idea is that we don't want the verifier to
2737 * reject future reads that access slots written to through variable offsets.
2739 static int check_stack_write_var_off(struct bpf_verifier_env *env,
2740 /* func where register points to */
2741 struct bpf_func_state *state,
2742 int ptr_regno, int off, int size,
2743 int value_regno, int insn_idx)
2745 struct bpf_func_state *cur; /* state of the current function */
2746 int min_off, max_off;
2748 struct bpf_reg_state *ptr_reg = NULL, *value_reg = NULL;
2749 bool writing_zero = false;
2750 /* set if the fact that we're writing a zero is used to let any
2751 * stack slots remain STACK_ZERO
2753 bool zero_used = false;
2755 cur = env->cur_state->frame[env->cur_state->curframe];
2756 ptr_reg = &cur->regs[ptr_regno];
2757 min_off = ptr_reg->smin_value + off;
2758 max_off = ptr_reg->smax_value + off + size;
2759 if (value_regno >= 0)
2760 value_reg = &cur->regs[value_regno];
2761 if (value_reg && register_is_null(value_reg))
2762 writing_zero = true;
2764 err = grow_stack_state(state, round_up(-min_off, BPF_REG_SIZE));
2769 /* Variable offset writes destroy any spilled pointers in range. */
2770 for (i = min_off; i < max_off; i++) {
2771 u8 new_type, *stype;
2775 spi = slot / BPF_REG_SIZE;
2776 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
2778 if (!env->allow_ptr_leaks
2779 && *stype != NOT_INIT
2780 && *stype != SCALAR_VALUE) {
2781 /* Reject the write if there's are spilled pointers in
2782 * range. If we didn't reject here, the ptr status
2783 * would be erased below (even though not all slots are
2784 * actually overwritten), possibly opening the door to
2787 verbose(env, "spilled ptr in range of var-offset stack write; insn %d, ptr off: %d",
2792 /* Erase all spilled pointers. */
2793 state->stack[spi].spilled_ptr.type = NOT_INIT;
2795 /* Update the slot type. */
2796 new_type = STACK_MISC;
2797 if (writing_zero && *stype == STACK_ZERO) {
2798 new_type = STACK_ZERO;
2801 /* If the slot is STACK_INVALID, we check whether it's OK to
2802 * pretend that it will be initialized by this write. The slot
2803 * might not actually be written to, and so if we mark it as
2804 * initialized future reads might leak uninitialized memory.
2805 * For privileged programs, we will accept such reads to slots
2806 * that may or may not be written because, if we're reject
2807 * them, the error would be too confusing.
2809 if (*stype == STACK_INVALID && !env->allow_uninit_stack) {
2810 verbose(env, "uninit stack in range of var-offset write prohibited for !root; insn %d, off: %d",
2817 /* backtracking doesn't work for STACK_ZERO yet. */
2818 err = mark_chain_precision(env, value_regno);
2825 /* When register 'dst_regno' is assigned some values from stack[min_off,
2826 * max_off), we set the register's type according to the types of the
2827 * respective stack slots. If all the stack values are known to be zeros, then
2828 * so is the destination reg. Otherwise, the register is considered to be
2829 * SCALAR. This function does not deal with register filling; the caller must
2830 * ensure that all spilled registers in the stack range have been marked as
2833 static void mark_reg_stack_read(struct bpf_verifier_env *env,
2834 /* func where src register points to */
2835 struct bpf_func_state *ptr_state,
2836 int min_off, int max_off, int dst_regno)
2838 struct bpf_verifier_state *vstate = env->cur_state;
2839 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2844 for (i = min_off; i < max_off; i++) {
2846 spi = slot / BPF_REG_SIZE;
2847 stype = ptr_state->stack[spi].slot_type;
2848 if (stype[slot % BPF_REG_SIZE] != STACK_ZERO)
2852 if (zeros == max_off - min_off) {
2853 /* any access_size read into register is zero extended,
2854 * so the whole register == const_zero
2856 __mark_reg_const_zero(&state->regs[dst_regno]);
2857 /* backtracking doesn't support STACK_ZERO yet,
2858 * so mark it precise here, so that later
2859 * backtracking can stop here.
2860 * Backtracking may not need this if this register
2861 * doesn't participate in pointer adjustment.
2862 * Forward propagation of precise flag is not
2863 * necessary either. This mark is only to stop
2864 * backtracking. Any register that contributed
2865 * to const 0 was marked precise before spill.
2867 state->regs[dst_regno].precise = true;
2869 /* have read misc data from the stack */
2870 mark_reg_unknown(env, state->regs, dst_regno);
2872 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
2875 /* Read the stack at 'off' and put the results into the register indicated by
2876 * 'dst_regno'. It handles reg filling if the addressed stack slot is a
2879 * 'dst_regno' can be -1, meaning that the read value is not going to a
2882 * The access is assumed to be within the current stack bounds.
2884 static int check_stack_read_fixed_off(struct bpf_verifier_env *env,
2885 /* func where src register points to */
2886 struct bpf_func_state *reg_state,
2887 int off, int size, int dst_regno)
2889 struct bpf_verifier_state *vstate = env->cur_state;
2890 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2891 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
2892 struct bpf_reg_state *reg;
2895 stype = reg_state->stack[spi].slot_type;
2896 reg = ®_state->stack[spi].spilled_ptr;
2898 if (stype[0] == STACK_SPILL) {
2899 if (size != BPF_REG_SIZE) {
2900 if (reg->type != SCALAR_VALUE) {
2901 verbose_linfo(env, env->insn_idx, "; ");
2902 verbose(env, "invalid size of register fill\n");
2905 if (dst_regno >= 0) {
2906 mark_reg_unknown(env, state->regs, dst_regno);
2907 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
2909 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2912 for (i = 1; i < BPF_REG_SIZE; i++) {
2913 if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) {
2914 verbose(env, "corrupted spill memory\n");
2919 if (dst_regno >= 0) {
2920 /* restore register state from stack */
2921 state->regs[dst_regno] = *reg;
2922 /* mark reg as written since spilled pointer state likely
2923 * has its liveness marks cleared by is_state_visited()
2924 * which resets stack/reg liveness for state transitions
2926 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
2927 } else if (__is_pointer_value(env->allow_ptr_leaks, reg)) {
2928 /* If dst_regno==-1, the caller is asking us whether
2929 * it is acceptable to use this value as a SCALAR_VALUE
2931 * We must not allow unprivileged callers to do that
2932 * with spilled pointers.
2934 verbose(env, "leaking pointer from stack off %d\n",
2938 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2942 for (i = 0; i < size; i++) {
2943 type = stype[(slot - i) % BPF_REG_SIZE];
2944 if (type == STACK_MISC)
2946 if (type == STACK_ZERO)
2948 verbose(env, "invalid read from stack off %d+%d size %d\n",
2952 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2954 mark_reg_stack_read(env, reg_state, off, off + size, dst_regno);
2959 enum stack_access_src {
2960 ACCESS_DIRECT = 1, /* the access is performed by an instruction */
2961 ACCESS_HELPER = 2, /* the access is performed by a helper */
2964 static int check_stack_range_initialized(struct bpf_verifier_env *env,
2965 int regno, int off, int access_size,
2966 bool zero_size_allowed,
2967 enum stack_access_src type,
2968 struct bpf_call_arg_meta *meta);
2970 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
2972 return cur_regs(env) + regno;
2975 /* Read the stack at 'ptr_regno + off' and put the result into the register
2977 * 'off' includes the pointer register's fixed offset(i.e. 'ptr_regno.off'),
2978 * but not its variable offset.
2979 * 'size' is assumed to be <= reg size and the access is assumed to be aligned.
2981 * As opposed to check_stack_read_fixed_off, this function doesn't deal with
2982 * filling registers (i.e. reads of spilled register cannot be detected when
2983 * the offset is not fixed). We conservatively mark 'dst_regno' as containing
2984 * SCALAR_VALUE. That's why we assert that the 'ptr_regno' has a variable
2985 * offset; for a fixed offset check_stack_read_fixed_off should be used
2988 static int check_stack_read_var_off(struct bpf_verifier_env *env,
2989 int ptr_regno, int off, int size, int dst_regno)
2991 /* The state of the source register. */
2992 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
2993 struct bpf_func_state *ptr_state = func(env, reg);
2995 int min_off, max_off;
2997 /* Note that we pass a NULL meta, so raw access will not be permitted.
2999 err = check_stack_range_initialized(env, ptr_regno, off, size,
3000 false, ACCESS_DIRECT, NULL);
3004 min_off = reg->smin_value + off;
3005 max_off = reg->smax_value + off;
3006 mark_reg_stack_read(env, ptr_state, min_off, max_off + size, dst_regno);
3010 /* check_stack_read dispatches to check_stack_read_fixed_off or
3011 * check_stack_read_var_off.
3013 * The caller must ensure that the offset falls within the allocated stack
3016 * 'dst_regno' is a register which will receive the value from the stack. It
3017 * can be -1, meaning that the read value is not going to a register.
3019 static int check_stack_read(struct bpf_verifier_env *env,
3020 int ptr_regno, int off, int size,
3023 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
3024 struct bpf_func_state *state = func(env, reg);
3026 /* Some accesses are only permitted with a static offset. */
3027 bool var_off = !tnum_is_const(reg->var_off);
3029 /* The offset is required to be static when reads don't go to a
3030 * register, in order to not leak pointers (see
3031 * check_stack_read_fixed_off).
3033 if (dst_regno < 0 && var_off) {
3036 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3037 verbose(env, "variable offset stack pointer cannot be passed into helper function; var_off=%s off=%d size=%d\n",
3041 /* Variable offset is prohibited for unprivileged mode for simplicity
3042 * since it requires corresponding support in Spectre masking for stack
3043 * ALU. See also retrieve_ptr_limit().
3045 if (!env->bypass_spec_v1 && var_off) {
3048 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3049 verbose(env, "R%d variable offset stack access prohibited for !root, var_off=%s\n",
3055 off += reg->var_off.value;
3056 err = check_stack_read_fixed_off(env, state, off, size,
3059 /* Variable offset stack reads need more conservative handling
3060 * than fixed offset ones. Note that dst_regno >= 0 on this
3063 err = check_stack_read_var_off(env, ptr_regno, off, size,
3070 /* check_stack_write dispatches to check_stack_write_fixed_off or
3071 * check_stack_write_var_off.
3073 * 'ptr_regno' is the register used as a pointer into the stack.
3074 * 'off' includes 'ptr_regno->off', but not its variable offset (if any).
3075 * 'value_regno' is the register whose value we're writing to the stack. It can
3076 * be -1, meaning that we're not writing from a register.
3078 * The caller must ensure that the offset falls within the maximum stack size.
3080 static int check_stack_write(struct bpf_verifier_env *env,
3081 int ptr_regno, int off, int size,
3082 int value_regno, int insn_idx)
3084 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
3085 struct bpf_func_state *state = func(env, reg);
3088 if (tnum_is_const(reg->var_off)) {
3089 off += reg->var_off.value;
3090 err = check_stack_write_fixed_off(env, state, off, size,
3091 value_regno, insn_idx);
3093 /* Variable offset stack reads need more conservative handling
3094 * than fixed offset ones.
3096 err = check_stack_write_var_off(env, state,
3097 ptr_regno, off, size,
3098 value_regno, insn_idx);
3103 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
3104 int off, int size, enum bpf_access_type type)
3106 struct bpf_reg_state *regs = cur_regs(env);
3107 struct bpf_map *map = regs[regno].map_ptr;
3108 u32 cap = bpf_map_flags_to_cap(map);
3110 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
3111 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
3112 map->value_size, off, size);
3116 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
3117 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
3118 map->value_size, off, size);
3125 /* check read/write into memory region (e.g., map value, ringbuf sample, etc) */
3126 static int __check_mem_access(struct bpf_verifier_env *env, int regno,
3127 int off, int size, u32 mem_size,
3128 bool zero_size_allowed)
3130 bool size_ok = size > 0 || (size == 0 && zero_size_allowed);
3131 struct bpf_reg_state *reg;
3133 if (off >= 0 && size_ok && (u64)off + size <= mem_size)
3136 reg = &cur_regs(env)[regno];
3137 switch (reg->type) {
3138 case PTR_TO_MAP_KEY:
3139 verbose(env, "invalid access to map key, key_size=%d off=%d size=%d\n",
3140 mem_size, off, size);
3142 case PTR_TO_MAP_VALUE:
3143 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
3144 mem_size, off, size);
3147 case PTR_TO_PACKET_META:
3148 case PTR_TO_PACKET_END:
3149 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
3150 off, size, regno, reg->id, off, mem_size);
3154 verbose(env, "invalid access to memory, mem_size=%u off=%d size=%d\n",
3155 mem_size, off, size);
3161 /* check read/write into a memory region with possible variable offset */
3162 static int check_mem_region_access(struct bpf_verifier_env *env, u32 regno,
3163 int off, int size, u32 mem_size,
3164 bool zero_size_allowed)
3166 struct bpf_verifier_state *vstate = env->cur_state;
3167 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3168 struct bpf_reg_state *reg = &state->regs[regno];
3171 /* We may have adjusted the register pointing to memory region, so we
3172 * need to try adding each of min_value and max_value to off
3173 * to make sure our theoretical access will be safe.
3175 if (env->log.level & BPF_LOG_LEVEL)
3176 print_verifier_state(env, state);
3178 /* The minimum value is only important with signed
3179 * comparisons where we can't assume the floor of a
3180 * value is 0. If we are using signed variables for our
3181 * index'es we need to make sure that whatever we use
3182 * will have a set floor within our range.
3184 if (reg->smin_value < 0 &&
3185 (reg->smin_value == S64_MIN ||
3186 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
3187 reg->smin_value + off < 0)) {
3188 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3192 err = __check_mem_access(env, regno, reg->smin_value + off, size,
3193 mem_size, zero_size_allowed);
3195 verbose(env, "R%d min value is outside of the allowed memory range\n",
3200 /* If we haven't set a max value then we need to bail since we can't be
3201 * sure we won't do bad things.
3202 * If reg->umax_value + off could overflow, treat that as unbounded too.
3204 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
3205 verbose(env, "R%d unbounded memory access, make sure to bounds check any such access\n",
3209 err = __check_mem_access(env, regno, reg->umax_value + off, size,
3210 mem_size, zero_size_allowed);
3212 verbose(env, "R%d max value is outside of the allowed memory range\n",
3220 /* check read/write into a map element with possible variable offset */
3221 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
3222 int off, int size, bool zero_size_allowed)
3224 struct bpf_verifier_state *vstate = env->cur_state;
3225 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3226 struct bpf_reg_state *reg = &state->regs[regno];
3227 struct bpf_map *map = reg->map_ptr;
3230 err = check_mem_region_access(env, regno, off, size, map->value_size,
3235 if (map_value_has_spin_lock(map)) {
3236 u32 lock = map->spin_lock_off;
3238 /* if any part of struct bpf_spin_lock can be touched by
3239 * load/store reject this program.
3240 * To check that [x1, x2) overlaps with [y1, y2)
3241 * it is sufficient to check x1 < y2 && y1 < x2.
3243 if (reg->smin_value + off < lock + sizeof(struct bpf_spin_lock) &&
3244 lock < reg->umax_value + off + size) {
3245 verbose(env, "bpf_spin_lock cannot be accessed directly by load/store\n");
3249 if (map_value_has_timer(map)) {
3250 u32 t = map->timer_off;
3252 if (reg->smin_value + off < t + sizeof(struct bpf_timer) &&
3253 t < reg->umax_value + off + size) {
3254 verbose(env, "bpf_timer cannot be accessed directly by load/store\n");
3261 #define MAX_PACKET_OFF 0xffff
3263 static enum bpf_prog_type resolve_prog_type(struct bpf_prog *prog)
3265 return prog->aux->dst_prog ? prog->aux->dst_prog->type : prog->type;
3268 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
3269 const struct bpf_call_arg_meta *meta,
3270 enum bpf_access_type t)
3272 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
3274 switch (prog_type) {
3275 /* Program types only with direct read access go here! */
3276 case BPF_PROG_TYPE_LWT_IN:
3277 case BPF_PROG_TYPE_LWT_OUT:
3278 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
3279 case BPF_PROG_TYPE_SK_REUSEPORT:
3280 case BPF_PROG_TYPE_FLOW_DISSECTOR:
3281 case BPF_PROG_TYPE_CGROUP_SKB:
3286 /* Program types with direct read + write access go here! */
3287 case BPF_PROG_TYPE_SCHED_CLS:
3288 case BPF_PROG_TYPE_SCHED_ACT:
3289 case BPF_PROG_TYPE_XDP:
3290 case BPF_PROG_TYPE_LWT_XMIT:
3291 case BPF_PROG_TYPE_SK_SKB:
3292 case BPF_PROG_TYPE_SK_MSG:
3294 return meta->pkt_access;
3296 env->seen_direct_write = true;
3299 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
3301 env->seen_direct_write = true;
3310 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
3311 int size, bool zero_size_allowed)
3313 struct bpf_reg_state *regs = cur_regs(env);
3314 struct bpf_reg_state *reg = ®s[regno];
3317 /* We may have added a variable offset to the packet pointer; but any
3318 * reg->range we have comes after that. We are only checking the fixed
3322 /* We don't allow negative numbers, because we aren't tracking enough
3323 * detail to prove they're safe.
3325 if (reg->smin_value < 0) {
3326 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3331 err = reg->range < 0 ? -EINVAL :
3332 __check_mem_access(env, regno, off, size, reg->range,
3335 verbose(env, "R%d offset is outside of the packet\n", regno);
3339 /* __check_mem_access has made sure "off + size - 1" is within u16.
3340 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
3341 * otherwise find_good_pkt_pointers would have refused to set range info
3342 * that __check_mem_access would have rejected this pkt access.
3343 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
3345 env->prog->aux->max_pkt_offset =
3346 max_t(u32, env->prog->aux->max_pkt_offset,
3347 off + reg->umax_value + size - 1);
3352 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
3353 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
3354 enum bpf_access_type t, enum bpf_reg_type *reg_type,
3355 struct btf **btf, u32 *btf_id)
3357 struct bpf_insn_access_aux info = {
3358 .reg_type = *reg_type,
3362 if (env->ops->is_valid_access &&
3363 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
3364 /* A non zero info.ctx_field_size indicates that this field is a
3365 * candidate for later verifier transformation to load the whole
3366 * field and then apply a mask when accessed with a narrower
3367 * access than actual ctx access size. A zero info.ctx_field_size
3368 * will only allow for whole field access and rejects any other
3369 * type of narrower access.
3371 *reg_type = info.reg_type;
3373 if (*reg_type == PTR_TO_BTF_ID || *reg_type == PTR_TO_BTF_ID_OR_NULL) {
3375 *btf_id = info.btf_id;
3377 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
3379 /* remember the offset of last byte accessed in ctx */
3380 if (env->prog->aux->max_ctx_offset < off + size)
3381 env->prog->aux->max_ctx_offset = off + size;
3385 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
3389 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
3392 if (size < 0 || off < 0 ||
3393 (u64)off + size > sizeof(struct bpf_flow_keys)) {
3394 verbose(env, "invalid access to flow keys off=%d size=%d\n",
3401 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
3402 u32 regno, int off, int size,
3403 enum bpf_access_type t)
3405 struct bpf_reg_state *regs = cur_regs(env);
3406 struct bpf_reg_state *reg = ®s[regno];
3407 struct bpf_insn_access_aux info = {};
3410 if (reg->smin_value < 0) {
3411 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3416 switch (reg->type) {
3417 case PTR_TO_SOCK_COMMON:
3418 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
3421 valid = bpf_sock_is_valid_access(off, size, t, &info);
3423 case PTR_TO_TCP_SOCK:
3424 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
3426 case PTR_TO_XDP_SOCK:
3427 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
3435 env->insn_aux_data[insn_idx].ctx_field_size =
3436 info.ctx_field_size;
3440 verbose(env, "R%d invalid %s access off=%d size=%d\n",
3441 regno, reg_type_str[reg->type], off, size);
3446 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
3448 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
3451 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
3453 const struct bpf_reg_state *reg = reg_state(env, regno);
3455 return reg->type == PTR_TO_CTX;
3458 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
3460 const struct bpf_reg_state *reg = reg_state(env, regno);
3462 return type_is_sk_pointer(reg->type);
3465 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
3467 const struct bpf_reg_state *reg = reg_state(env, regno);
3469 return type_is_pkt_pointer(reg->type);
3472 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
3474 const struct bpf_reg_state *reg = reg_state(env, regno);
3476 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
3477 return reg->type == PTR_TO_FLOW_KEYS;
3480 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
3481 const struct bpf_reg_state *reg,
3482 int off, int size, bool strict)
3484 struct tnum reg_off;
3487 /* Byte size accesses are always allowed. */
3488 if (!strict || size == 1)
3491 /* For platforms that do not have a Kconfig enabling
3492 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
3493 * NET_IP_ALIGN is universally set to '2'. And on platforms
3494 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
3495 * to this code only in strict mode where we want to emulate
3496 * the NET_IP_ALIGN==2 checking. Therefore use an
3497 * unconditional IP align value of '2'.
3501 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
3502 if (!tnum_is_aligned(reg_off, size)) {
3505 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3507 "misaligned packet access off %d+%s+%d+%d size %d\n",
3508 ip_align, tn_buf, reg->off, off, size);
3515 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
3516 const struct bpf_reg_state *reg,
3517 const char *pointer_desc,
3518 int off, int size, bool strict)
3520 struct tnum reg_off;
3522 /* Byte size accesses are always allowed. */
3523 if (!strict || size == 1)
3526 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
3527 if (!tnum_is_aligned(reg_off, size)) {
3530 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3531 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
3532 pointer_desc, tn_buf, reg->off, off, size);
3539 static int check_ptr_alignment(struct bpf_verifier_env *env,
3540 const struct bpf_reg_state *reg, int off,
3541 int size, bool strict_alignment_once)
3543 bool strict = env->strict_alignment || strict_alignment_once;
3544 const char *pointer_desc = "";
3546 switch (reg->type) {
3548 case PTR_TO_PACKET_META:
3549 /* Special case, because of NET_IP_ALIGN. Given metadata sits
3550 * right in front, treat it the very same way.
3552 return check_pkt_ptr_alignment(env, reg, off, size, strict);
3553 case PTR_TO_FLOW_KEYS:
3554 pointer_desc = "flow keys ";
3556 case PTR_TO_MAP_KEY:
3557 pointer_desc = "key ";
3559 case PTR_TO_MAP_VALUE:
3560 pointer_desc = "value ";
3563 pointer_desc = "context ";
3566 pointer_desc = "stack ";
3567 /* The stack spill tracking logic in check_stack_write_fixed_off()
3568 * and check_stack_read_fixed_off() relies on stack accesses being
3574 pointer_desc = "sock ";
3576 case PTR_TO_SOCK_COMMON:
3577 pointer_desc = "sock_common ";
3579 case PTR_TO_TCP_SOCK:
3580 pointer_desc = "tcp_sock ";
3582 case PTR_TO_XDP_SOCK:
3583 pointer_desc = "xdp_sock ";
3588 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
3592 static int update_stack_depth(struct bpf_verifier_env *env,
3593 const struct bpf_func_state *func,
3596 u16 stack = env->subprog_info[func->subprogno].stack_depth;
3601 /* update known max for given subprogram */
3602 env->subprog_info[func->subprogno].stack_depth = -off;
3606 /* starting from main bpf function walk all instructions of the function
3607 * and recursively walk all callees that given function can call.
3608 * Ignore jump and exit insns.
3609 * Since recursion is prevented by check_cfg() this algorithm
3610 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
3612 static int check_max_stack_depth(struct bpf_verifier_env *env)
3614 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
3615 struct bpf_subprog_info *subprog = env->subprog_info;
3616 struct bpf_insn *insn = env->prog->insnsi;
3617 bool tail_call_reachable = false;
3618 int ret_insn[MAX_CALL_FRAMES];
3619 int ret_prog[MAX_CALL_FRAMES];
3623 /* protect against potential stack overflow that might happen when
3624 * bpf2bpf calls get combined with tailcalls. Limit the caller's stack
3625 * depth for such case down to 256 so that the worst case scenario
3626 * would result in 8k stack size (32 which is tailcall limit * 256 =
3629 * To get the idea what might happen, see an example:
3630 * func1 -> sub rsp, 128
3631 * subfunc1 -> sub rsp, 256
3632 * tailcall1 -> add rsp, 256
3633 * func2 -> sub rsp, 192 (total stack size = 128 + 192 = 320)
3634 * subfunc2 -> sub rsp, 64
3635 * subfunc22 -> sub rsp, 128
3636 * tailcall2 -> add rsp, 128
3637 * func3 -> sub rsp, 32 (total stack size 128 + 192 + 64 + 32 = 416)
3639 * tailcall will unwind the current stack frame but it will not get rid
3640 * of caller's stack as shown on the example above.
3642 if (idx && subprog[idx].has_tail_call && depth >= 256) {
3644 "tail_calls are not allowed when call stack of previous frames is %d bytes. Too large\n",
3648 /* round up to 32-bytes, since this is granularity
3649 * of interpreter stack size
3651 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
3652 if (depth > MAX_BPF_STACK) {
3653 verbose(env, "combined stack size of %d calls is %d. Too large\n",
3658 subprog_end = subprog[idx + 1].start;
3659 for (; i < subprog_end; i++) {
3660 if (!bpf_pseudo_call(insn + i) && !bpf_pseudo_func(insn + i))
3662 /* remember insn and function to return to */
3663 ret_insn[frame] = i + 1;
3664 ret_prog[frame] = idx;
3666 /* find the callee */
3667 i = i + insn[i].imm + 1;
3668 idx = find_subprog(env, i);
3670 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3675 if (subprog[idx].has_tail_call)
3676 tail_call_reachable = true;
3679 if (frame >= MAX_CALL_FRAMES) {
3680 verbose(env, "the call stack of %d frames is too deep !\n",
3686 /* if tail call got detected across bpf2bpf calls then mark each of the
3687 * currently present subprog frames as tail call reachable subprogs;
3688 * this info will be utilized by JIT so that we will be preserving the
3689 * tail call counter throughout bpf2bpf calls combined with tailcalls
3691 if (tail_call_reachable)
3692 for (j = 0; j < frame; j++)
3693 subprog[ret_prog[j]].tail_call_reachable = true;
3695 /* end of for() loop means the last insn of the 'subprog'
3696 * was reached. Doesn't matter whether it was JA or EXIT
3700 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
3702 i = ret_insn[frame];
3703 idx = ret_prog[frame];
3707 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
3708 static int get_callee_stack_depth(struct bpf_verifier_env *env,
3709 const struct bpf_insn *insn, int idx)
3711 int start = idx + insn->imm + 1, subprog;
3713 subprog = find_subprog(env, start);
3715 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3719 return env->subprog_info[subprog].stack_depth;
3723 int check_ctx_reg(struct bpf_verifier_env *env,
3724 const struct bpf_reg_state *reg, int regno)
3726 /* Access to ctx or passing it to a helper is only allowed in
3727 * its original, unmodified form.
3731 verbose(env, "dereference of modified ctx ptr R%d off=%d disallowed\n",
3736 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3739 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3740 verbose(env, "variable ctx access var_off=%s disallowed\n", tn_buf);
3747 static int __check_buffer_access(struct bpf_verifier_env *env,
3748 const char *buf_info,
3749 const struct bpf_reg_state *reg,
3750 int regno, int off, int size)
3754 "R%d invalid %s buffer access: off=%d, size=%d\n",
3755 regno, buf_info, off, size);
3758 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3761 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3763 "R%d invalid variable buffer offset: off=%d, var_off=%s\n",
3764 regno, off, tn_buf);
3771 static int check_tp_buffer_access(struct bpf_verifier_env *env,
3772 const struct bpf_reg_state *reg,
3773 int regno, int off, int size)
3777 err = __check_buffer_access(env, "tracepoint", reg, regno, off, size);
3781 if (off + size > env->prog->aux->max_tp_access)
3782 env->prog->aux->max_tp_access = off + size;
3787 static int check_buffer_access(struct bpf_verifier_env *env,
3788 const struct bpf_reg_state *reg,
3789 int regno, int off, int size,
3790 bool zero_size_allowed,
3791 const char *buf_info,
3796 err = __check_buffer_access(env, buf_info, reg, regno, off, size);
3800 if (off + size > *max_access)
3801 *max_access = off + size;
3806 /* BPF architecture zero extends alu32 ops into 64-bit registesr */
3807 static void zext_32_to_64(struct bpf_reg_state *reg)
3809 reg->var_off = tnum_subreg(reg->var_off);
3810 __reg_assign_32_into_64(reg);
3813 /* truncate register to smaller size (in bytes)
3814 * must be called with size < BPF_REG_SIZE
3816 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
3820 /* clear high bits in bit representation */
3821 reg->var_off = tnum_cast(reg->var_off, size);
3823 /* fix arithmetic bounds */
3824 mask = ((u64)1 << (size * 8)) - 1;
3825 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
3826 reg->umin_value &= mask;
3827 reg->umax_value &= mask;
3829 reg->umin_value = 0;
3830 reg->umax_value = mask;
3832 reg->smin_value = reg->umin_value;
3833 reg->smax_value = reg->umax_value;
3835 /* If size is smaller than 32bit register the 32bit register
3836 * values are also truncated so we push 64-bit bounds into
3837 * 32-bit bounds. Above were truncated < 32-bits already.
3841 __reg_combine_64_into_32(reg);
3844 static bool bpf_map_is_rdonly(const struct bpf_map *map)
3846 return (map->map_flags & BPF_F_RDONLY_PROG) && map->frozen;
3849 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val)
3855 err = map->ops->map_direct_value_addr(map, &addr, off);
3858 ptr = (void *)(long)addr + off;
3862 *val = (u64)*(u8 *)ptr;
3865 *val = (u64)*(u16 *)ptr;
3868 *val = (u64)*(u32 *)ptr;
3879 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
3880 struct bpf_reg_state *regs,
3881 int regno, int off, int size,
3882 enum bpf_access_type atype,
3885 struct bpf_reg_state *reg = regs + regno;
3886 const struct btf_type *t = btf_type_by_id(reg->btf, reg->btf_id);
3887 const char *tname = btf_name_by_offset(reg->btf, t->name_off);
3893 "R%d is ptr_%s invalid negative access: off=%d\n",
3897 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3900 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3902 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
3903 regno, tname, off, tn_buf);
3907 if (env->ops->btf_struct_access) {
3908 ret = env->ops->btf_struct_access(&env->log, reg->btf, t,
3909 off, size, atype, &btf_id);
3911 if (atype != BPF_READ) {
3912 verbose(env, "only read is supported\n");
3916 ret = btf_struct_access(&env->log, reg->btf, t, off, size,
3923 if (atype == BPF_READ && value_regno >= 0)
3924 mark_btf_ld_reg(env, regs, value_regno, ret, reg->btf, btf_id);
3929 static int check_ptr_to_map_access(struct bpf_verifier_env *env,
3930 struct bpf_reg_state *regs,
3931 int regno, int off, int size,
3932 enum bpf_access_type atype,
3935 struct bpf_reg_state *reg = regs + regno;
3936 struct bpf_map *map = reg->map_ptr;
3937 const struct btf_type *t;
3943 verbose(env, "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n");
3947 if (!map->ops->map_btf_id || !*map->ops->map_btf_id) {
3948 verbose(env, "map_ptr access not supported for map type %d\n",
3953 t = btf_type_by_id(btf_vmlinux, *map->ops->map_btf_id);
3954 tname = btf_name_by_offset(btf_vmlinux, t->name_off);
3956 if (!env->allow_ptr_to_map_access) {
3958 "%s access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
3964 verbose(env, "R%d is %s invalid negative access: off=%d\n",
3969 if (atype != BPF_READ) {
3970 verbose(env, "only read from %s is supported\n", tname);
3974 ret = btf_struct_access(&env->log, btf_vmlinux, t, off, size, atype, &btf_id);
3978 if (value_regno >= 0)
3979 mark_btf_ld_reg(env, regs, value_regno, ret, btf_vmlinux, btf_id);
3984 /* Check that the stack access at the given offset is within bounds. The
3985 * maximum valid offset is -1.
3987 * The minimum valid offset is -MAX_BPF_STACK for writes, and
3988 * -state->allocated_stack for reads.
3990 static int check_stack_slot_within_bounds(int off,
3991 struct bpf_func_state *state,
3992 enum bpf_access_type t)
3997 min_valid_off = -MAX_BPF_STACK;
3999 min_valid_off = -state->allocated_stack;
4001 if (off < min_valid_off || off > -1)
4006 /* Check that the stack access at 'regno + off' falls within the maximum stack
4009 * 'off' includes `regno->offset`, but not its dynamic part (if any).
4011 static int check_stack_access_within_bounds(
4012 struct bpf_verifier_env *env,
4013 int regno, int off, int access_size,
4014 enum stack_access_src src, enum bpf_access_type type)
4016 struct bpf_reg_state *regs = cur_regs(env);
4017 struct bpf_reg_state *reg = regs + regno;
4018 struct bpf_func_state *state = func(env, reg);
4019 int min_off, max_off;
4023 if (src == ACCESS_HELPER)
4024 /* We don't know if helpers are reading or writing (or both). */
4025 err_extra = " indirect access to";
4026 else if (type == BPF_READ)
4027 err_extra = " read from";
4029 err_extra = " write to";
4031 if (tnum_is_const(reg->var_off)) {
4032 min_off = reg->var_off.value + off;
4033 if (access_size > 0)
4034 max_off = min_off + access_size - 1;
4038 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
4039 reg->smin_value <= -BPF_MAX_VAR_OFF) {
4040 verbose(env, "invalid unbounded variable-offset%s stack R%d\n",
4044 min_off = reg->smin_value + off;
4045 if (access_size > 0)
4046 max_off = reg->smax_value + off + access_size - 1;
4051 err = check_stack_slot_within_bounds(min_off, state, type);
4053 err = check_stack_slot_within_bounds(max_off, state, type);
4056 if (tnum_is_const(reg->var_off)) {
4057 verbose(env, "invalid%s stack R%d off=%d size=%d\n",
4058 err_extra, regno, off, access_size);
4062 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4063 verbose(env, "invalid variable-offset%s stack R%d var_off=%s size=%d\n",
4064 err_extra, regno, tn_buf, access_size);
4070 /* check whether memory at (regno + off) is accessible for t = (read | write)
4071 * if t==write, value_regno is a register which value is stored into memory
4072 * if t==read, value_regno is a register which will receive the value from memory
4073 * if t==write && value_regno==-1, some unknown value is stored into memory
4074 * if t==read && value_regno==-1, don't care what we read from memory
4076 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
4077 int off, int bpf_size, enum bpf_access_type t,
4078 int value_regno, bool strict_alignment_once)
4080 struct bpf_reg_state *regs = cur_regs(env);
4081 struct bpf_reg_state *reg = regs + regno;
4082 struct bpf_func_state *state;
4085 size = bpf_size_to_bytes(bpf_size);
4089 /* alignment checks will add in reg->off themselves */
4090 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
4094 /* for access checks, reg->off is just part of off */
4097 if (reg->type == PTR_TO_MAP_KEY) {
4098 if (t == BPF_WRITE) {
4099 verbose(env, "write to change key R%d not allowed\n", regno);
4103 err = check_mem_region_access(env, regno, off, size,
4104 reg->map_ptr->key_size, false);
4107 if (value_regno >= 0)
4108 mark_reg_unknown(env, regs, value_regno);
4109 } else if (reg->type == PTR_TO_MAP_VALUE) {
4110 if (t == BPF_WRITE && value_regno >= 0 &&
4111 is_pointer_value(env, value_regno)) {
4112 verbose(env, "R%d leaks addr into map\n", value_regno);
4115 err = check_map_access_type(env, regno, off, size, t);
4118 err = check_map_access(env, regno, off, size, false);
4119 if (!err && t == BPF_READ && value_regno >= 0) {
4120 struct bpf_map *map = reg->map_ptr;
4122 /* if map is read-only, track its contents as scalars */
4123 if (tnum_is_const(reg->var_off) &&
4124 bpf_map_is_rdonly(map) &&
4125 map->ops->map_direct_value_addr) {
4126 int map_off = off + reg->var_off.value;
4129 err = bpf_map_direct_read(map, map_off, size,
4134 regs[value_regno].type = SCALAR_VALUE;
4135 __mark_reg_known(®s[value_regno], val);
4137 mark_reg_unknown(env, regs, value_regno);
4140 } else if (reg->type == PTR_TO_MEM) {
4141 if (t == BPF_WRITE && value_regno >= 0 &&
4142 is_pointer_value(env, value_regno)) {
4143 verbose(env, "R%d leaks addr into mem\n", value_regno);
4146 err = check_mem_region_access(env, regno, off, size,
4147 reg->mem_size, false);
4148 if (!err && t == BPF_READ && value_regno >= 0)
4149 mark_reg_unknown(env, regs, value_regno);
4150 } else if (reg->type == PTR_TO_CTX) {
4151 enum bpf_reg_type reg_type = SCALAR_VALUE;
4152 struct btf *btf = NULL;
4155 if (t == BPF_WRITE && value_regno >= 0 &&
4156 is_pointer_value(env, value_regno)) {
4157 verbose(env, "R%d leaks addr into ctx\n", value_regno);
4161 err = check_ctx_reg(env, reg, regno);
4165 err = check_ctx_access(env, insn_idx, off, size, t, ®_type, &btf, &btf_id);
4167 verbose_linfo(env, insn_idx, "; ");
4168 if (!err && t == BPF_READ && value_regno >= 0) {
4169 /* ctx access returns either a scalar, or a
4170 * PTR_TO_PACKET[_META,_END]. In the latter
4171 * case, we know the offset is zero.
4173 if (reg_type == SCALAR_VALUE) {
4174 mark_reg_unknown(env, regs, value_regno);
4176 mark_reg_known_zero(env, regs,
4178 if (reg_type_may_be_null(reg_type))
4179 regs[value_regno].id = ++env->id_gen;
4180 /* A load of ctx field could have different
4181 * actual load size with the one encoded in the
4182 * insn. When the dst is PTR, it is for sure not
4185 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
4186 if (reg_type == PTR_TO_BTF_ID ||
4187 reg_type == PTR_TO_BTF_ID_OR_NULL) {
4188 regs[value_regno].btf = btf;
4189 regs[value_regno].btf_id = btf_id;
4192 regs[value_regno].type = reg_type;
4195 } else if (reg->type == PTR_TO_STACK) {
4196 /* Basic bounds checks. */
4197 err = check_stack_access_within_bounds(env, regno, off, size, ACCESS_DIRECT, t);
4201 state = func(env, reg);
4202 err = update_stack_depth(env, state, off);
4207 err = check_stack_read(env, regno, off, size,
4210 err = check_stack_write(env, regno, off, size,
4211 value_regno, insn_idx);
4212 } else if (reg_is_pkt_pointer(reg)) {
4213 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
4214 verbose(env, "cannot write into packet\n");
4217 if (t == BPF_WRITE && value_regno >= 0 &&
4218 is_pointer_value(env, value_regno)) {
4219 verbose(env, "R%d leaks addr into packet\n",
4223 err = check_packet_access(env, regno, off, size, false);
4224 if (!err && t == BPF_READ && value_regno >= 0)
4225 mark_reg_unknown(env, regs, value_regno);
4226 } else if (reg->type == PTR_TO_FLOW_KEYS) {
4227 if (t == BPF_WRITE && value_regno >= 0 &&
4228 is_pointer_value(env, value_regno)) {
4229 verbose(env, "R%d leaks addr into flow keys\n",
4234 err = check_flow_keys_access(env, off, size);
4235 if (!err && t == BPF_READ && value_regno >= 0)
4236 mark_reg_unknown(env, regs, value_regno);
4237 } else if (type_is_sk_pointer(reg->type)) {
4238 if (t == BPF_WRITE) {
4239 verbose(env, "R%d cannot write into %s\n",
4240 regno, reg_type_str[reg->type]);
4243 err = check_sock_access(env, insn_idx, regno, off, size, t);
4244 if (!err && value_regno >= 0)
4245 mark_reg_unknown(env, regs, value_regno);
4246 } else if (reg->type == PTR_TO_TP_BUFFER) {
4247 err = check_tp_buffer_access(env, reg, regno, off, size);
4248 if (!err && t == BPF_READ && value_regno >= 0)
4249 mark_reg_unknown(env, regs, value_regno);
4250 } else if (reg->type == PTR_TO_BTF_ID) {
4251 err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
4253 } else if (reg->type == CONST_PTR_TO_MAP) {
4254 err = check_ptr_to_map_access(env, regs, regno, off, size, t,
4256 } else if (reg->type == PTR_TO_RDONLY_BUF) {
4257 if (t == BPF_WRITE) {
4258 verbose(env, "R%d cannot write into %s\n",
4259 regno, reg_type_str[reg->type]);
4262 err = check_buffer_access(env, reg, regno, off, size, false,
4264 &env->prog->aux->max_rdonly_access);
4265 if (!err && value_regno >= 0)
4266 mark_reg_unknown(env, regs, value_regno);
4267 } else if (reg->type == PTR_TO_RDWR_BUF) {
4268 err = check_buffer_access(env, reg, regno, off, size, false,
4270 &env->prog->aux->max_rdwr_access);
4271 if (!err && t == BPF_READ && value_regno >= 0)
4272 mark_reg_unknown(env, regs, value_regno);
4274 verbose(env, "R%d invalid mem access '%s'\n", regno,
4275 reg_type_str[reg->type]);
4279 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
4280 regs[value_regno].type == SCALAR_VALUE) {
4281 /* b/h/w load zero-extends, mark upper bits as known 0 */
4282 coerce_reg_to_size(®s[value_regno], size);
4287 static int check_atomic(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
4292 switch (insn->imm) {
4294 case BPF_ADD | BPF_FETCH:
4296 case BPF_AND | BPF_FETCH:
4298 case BPF_OR | BPF_FETCH:
4300 case BPF_XOR | BPF_FETCH:
4305 verbose(env, "BPF_ATOMIC uses invalid atomic opcode %02x\n", insn->imm);
4309 if (BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) {
4310 verbose(env, "invalid atomic operand size\n");
4314 /* check src1 operand */
4315 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4319 /* check src2 operand */
4320 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4324 if (insn->imm == BPF_CMPXCHG) {
4325 /* Check comparison of R0 with memory location */
4326 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
4331 if (is_pointer_value(env, insn->src_reg)) {
4332 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
4336 if (is_ctx_reg(env, insn->dst_reg) ||
4337 is_pkt_reg(env, insn->dst_reg) ||
4338 is_flow_key_reg(env, insn->dst_reg) ||
4339 is_sk_reg(env, insn->dst_reg)) {
4340 verbose(env, "BPF_ATOMIC stores into R%d %s is not allowed\n",
4342 reg_type_str[reg_state(env, insn->dst_reg)->type]);
4346 if (insn->imm & BPF_FETCH) {
4347 if (insn->imm == BPF_CMPXCHG)
4348 load_reg = BPF_REG_0;
4350 load_reg = insn->src_reg;
4352 /* check and record load of old value */
4353 err = check_reg_arg(env, load_reg, DST_OP);
4357 /* This instruction accesses a memory location but doesn't
4358 * actually load it into a register.
4363 /* check whether we can read the memory */
4364 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4365 BPF_SIZE(insn->code), BPF_READ, load_reg, true);
4369 /* check whether we can write into the same memory */
4370 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4371 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
4378 /* When register 'regno' is used to read the stack (either directly or through
4379 * a helper function) make sure that it's within stack boundary and, depending
4380 * on the access type, that all elements of the stack are initialized.
4382 * 'off' includes 'regno->off', but not its dynamic part (if any).
4384 * All registers that have been spilled on the stack in the slots within the
4385 * read offsets are marked as read.
4387 static int check_stack_range_initialized(
4388 struct bpf_verifier_env *env, int regno, int off,
4389 int access_size, bool zero_size_allowed,
4390 enum stack_access_src type, struct bpf_call_arg_meta *meta)
4392 struct bpf_reg_state *reg = reg_state(env, regno);
4393 struct bpf_func_state *state = func(env, reg);
4394 int err, min_off, max_off, i, j, slot, spi;
4395 char *err_extra = type == ACCESS_HELPER ? " indirect" : "";
4396 enum bpf_access_type bounds_check_type;
4397 /* Some accesses can write anything into the stack, others are
4400 bool clobber = false;
4402 if (access_size == 0 && !zero_size_allowed) {
4403 verbose(env, "invalid zero-sized read\n");
4407 if (type == ACCESS_HELPER) {
4408 /* The bounds checks for writes are more permissive than for
4409 * reads. However, if raw_mode is not set, we'll do extra
4412 bounds_check_type = BPF_WRITE;
4415 bounds_check_type = BPF_READ;
4417 err = check_stack_access_within_bounds(env, regno, off, access_size,
4418 type, bounds_check_type);
4423 if (tnum_is_const(reg->var_off)) {
4424 min_off = max_off = reg->var_off.value + off;
4426 /* Variable offset is prohibited for unprivileged mode for
4427 * simplicity since it requires corresponding support in
4428 * Spectre masking for stack ALU.
4429 * See also retrieve_ptr_limit().
4431 if (!env->bypass_spec_v1) {
4434 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4435 verbose(env, "R%d%s variable offset stack access prohibited for !root, var_off=%s\n",
4436 regno, err_extra, tn_buf);
4439 /* Only initialized buffer on stack is allowed to be accessed
4440 * with variable offset. With uninitialized buffer it's hard to
4441 * guarantee that whole memory is marked as initialized on
4442 * helper return since specific bounds are unknown what may
4443 * cause uninitialized stack leaking.
4445 if (meta && meta->raw_mode)
4448 min_off = reg->smin_value + off;
4449 max_off = reg->smax_value + off;
4452 if (meta && meta->raw_mode) {
4453 meta->access_size = access_size;
4454 meta->regno = regno;
4458 for (i = min_off; i < max_off + access_size; i++) {
4462 spi = slot / BPF_REG_SIZE;
4463 if (state->allocated_stack <= slot)
4465 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
4466 if (*stype == STACK_MISC)
4468 if (*stype == STACK_ZERO) {
4470 /* helper can write anything into the stack */
4471 *stype = STACK_MISC;
4476 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
4477 state->stack[spi].spilled_ptr.type == PTR_TO_BTF_ID)
4480 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
4481 (state->stack[spi].spilled_ptr.type == SCALAR_VALUE ||
4482 env->allow_ptr_leaks)) {
4484 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
4485 for (j = 0; j < BPF_REG_SIZE; j++)
4486 state->stack[spi].slot_type[j] = STACK_MISC;
4492 if (tnum_is_const(reg->var_off)) {
4493 verbose(env, "invalid%s read from stack R%d off %d+%d size %d\n",
4494 err_extra, regno, min_off, i - min_off, access_size);
4498 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4499 verbose(env, "invalid%s read from stack R%d var_off %s+%d size %d\n",
4500 err_extra, regno, tn_buf, i - min_off, access_size);
4504 /* reading any byte out of 8-byte 'spill_slot' will cause
4505 * the whole slot to be marked as 'read'
4507 mark_reg_read(env, &state->stack[spi].spilled_ptr,
4508 state->stack[spi].spilled_ptr.parent,
4511 return update_stack_depth(env, state, min_off);
4514 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
4515 int access_size, bool zero_size_allowed,
4516 struct bpf_call_arg_meta *meta)
4518 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4520 switch (reg->type) {
4522 case PTR_TO_PACKET_META:
4523 return check_packet_access(env, regno, reg->off, access_size,
4525 case PTR_TO_MAP_KEY:
4526 return check_mem_region_access(env, regno, reg->off, access_size,
4527 reg->map_ptr->key_size, false);
4528 case PTR_TO_MAP_VALUE:
4529 if (check_map_access_type(env, regno, reg->off, access_size,
4530 meta && meta->raw_mode ? BPF_WRITE :
4533 return check_map_access(env, regno, reg->off, access_size,
4536 return check_mem_region_access(env, regno, reg->off,
4537 access_size, reg->mem_size,
4539 case PTR_TO_RDONLY_BUF:
4540 if (meta && meta->raw_mode)
4542 return check_buffer_access(env, reg, regno, reg->off,
4543 access_size, zero_size_allowed,
4545 &env->prog->aux->max_rdonly_access);
4546 case PTR_TO_RDWR_BUF:
4547 return check_buffer_access(env, reg, regno, reg->off,
4548 access_size, zero_size_allowed,
4550 &env->prog->aux->max_rdwr_access);
4552 return check_stack_range_initialized(
4554 regno, reg->off, access_size,
4555 zero_size_allowed, ACCESS_HELPER, meta);
4556 default: /* scalar_value or invalid ptr */
4557 /* Allow zero-byte read from NULL, regardless of pointer type */
4558 if (zero_size_allowed && access_size == 0 &&
4559 register_is_null(reg))
4562 verbose(env, "R%d type=%s expected=%s\n", regno,
4563 reg_type_str[reg->type],
4564 reg_type_str[PTR_TO_STACK]);
4569 int check_mem_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
4570 u32 regno, u32 mem_size)
4572 if (register_is_null(reg))
4575 if (reg_type_may_be_null(reg->type)) {
4576 /* Assuming that the register contains a value check if the memory
4577 * access is safe. Temporarily save and restore the register's state as
4578 * the conversion shouldn't be visible to a caller.
4580 const struct bpf_reg_state saved_reg = *reg;
4583 mark_ptr_not_null_reg(reg);
4584 rv = check_helper_mem_access(env, regno, mem_size, true, NULL);
4589 return check_helper_mem_access(env, regno, mem_size, true, NULL);
4592 /* Implementation details:
4593 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL
4594 * Two bpf_map_lookups (even with the same key) will have different reg->id.
4595 * For traditional PTR_TO_MAP_VALUE the verifier clears reg->id after
4596 * value_or_null->value transition, since the verifier only cares about
4597 * the range of access to valid map value pointer and doesn't care about actual
4598 * address of the map element.
4599 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
4600 * reg->id > 0 after value_or_null->value transition. By doing so
4601 * two bpf_map_lookups will be considered two different pointers that
4602 * point to different bpf_spin_locks.
4603 * The verifier allows taking only one bpf_spin_lock at a time to avoid
4605 * Since only one bpf_spin_lock is allowed the checks are simpler than
4606 * reg_is_refcounted() logic. The verifier needs to remember only
4607 * one spin_lock instead of array of acquired_refs.
4608 * cur_state->active_spin_lock remembers which map value element got locked
4609 * and clears it after bpf_spin_unlock.
4611 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
4614 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4615 struct bpf_verifier_state *cur = env->cur_state;
4616 bool is_const = tnum_is_const(reg->var_off);
4617 struct bpf_map *map = reg->map_ptr;
4618 u64 val = reg->var_off.value;
4622 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
4628 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
4632 if (!map_value_has_spin_lock(map)) {
4633 if (map->spin_lock_off == -E2BIG)
4635 "map '%s' has more than one 'struct bpf_spin_lock'\n",
4637 else if (map->spin_lock_off == -ENOENT)
4639 "map '%s' doesn't have 'struct bpf_spin_lock'\n",
4643 "map '%s' is not a struct type or bpf_spin_lock is mangled\n",
4647 if (map->spin_lock_off != val + reg->off) {
4648 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock'\n",
4653 if (cur->active_spin_lock) {
4655 "Locking two bpf_spin_locks are not allowed\n");
4658 cur->active_spin_lock = reg->id;
4660 if (!cur->active_spin_lock) {
4661 verbose(env, "bpf_spin_unlock without taking a lock\n");
4664 if (cur->active_spin_lock != reg->id) {
4665 verbose(env, "bpf_spin_unlock of different lock\n");
4668 cur->active_spin_lock = 0;
4673 static int process_timer_func(struct bpf_verifier_env *env, int regno,
4674 struct bpf_call_arg_meta *meta)
4676 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4677 bool is_const = tnum_is_const(reg->var_off);
4678 struct bpf_map *map = reg->map_ptr;
4679 u64 val = reg->var_off.value;
4683 "R%d doesn't have constant offset. bpf_timer has to be at the constant offset\n",
4688 verbose(env, "map '%s' has to have BTF in order to use bpf_timer\n",
4692 if (!map_value_has_timer(map)) {
4693 if (map->timer_off == -E2BIG)
4695 "map '%s' has more than one 'struct bpf_timer'\n",
4697 else if (map->timer_off == -ENOENT)
4699 "map '%s' doesn't have 'struct bpf_timer'\n",
4703 "map '%s' is not a struct type or bpf_timer is mangled\n",
4707 if (map->timer_off != val + reg->off) {
4708 verbose(env, "off %lld doesn't point to 'struct bpf_timer' that is at %d\n",
4709 val + reg->off, map->timer_off);
4712 if (meta->map_ptr) {
4713 verbose(env, "verifier bug. Two map pointers in a timer helper\n");
4716 meta->map_uid = reg->map_uid;
4717 meta->map_ptr = map;
4721 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
4723 return type == ARG_PTR_TO_MEM ||
4724 type == ARG_PTR_TO_MEM_OR_NULL ||
4725 type == ARG_PTR_TO_UNINIT_MEM;
4728 static bool arg_type_is_mem_size(enum bpf_arg_type type)
4730 return type == ARG_CONST_SIZE ||
4731 type == ARG_CONST_SIZE_OR_ZERO;
4734 static bool arg_type_is_alloc_size(enum bpf_arg_type type)
4736 return type == ARG_CONST_ALLOC_SIZE_OR_ZERO;
4739 static bool arg_type_is_int_ptr(enum bpf_arg_type type)
4741 return type == ARG_PTR_TO_INT ||
4742 type == ARG_PTR_TO_LONG;
4745 static int int_ptr_type_to_size(enum bpf_arg_type type)
4747 if (type == ARG_PTR_TO_INT)
4749 else if (type == ARG_PTR_TO_LONG)
4755 static int resolve_map_arg_type(struct bpf_verifier_env *env,
4756 const struct bpf_call_arg_meta *meta,
4757 enum bpf_arg_type *arg_type)
4759 if (!meta->map_ptr) {
4760 /* kernel subsystem misconfigured verifier */
4761 verbose(env, "invalid map_ptr to access map->type\n");
4765 switch (meta->map_ptr->map_type) {
4766 case BPF_MAP_TYPE_SOCKMAP:
4767 case BPF_MAP_TYPE_SOCKHASH:
4768 if (*arg_type == ARG_PTR_TO_MAP_VALUE) {
4769 *arg_type = ARG_PTR_TO_BTF_ID_SOCK_COMMON;
4771 verbose(env, "invalid arg_type for sockmap/sockhash\n");
4782 struct bpf_reg_types {
4783 const enum bpf_reg_type types[10];
4787 static const struct bpf_reg_types map_key_value_types = {
4797 static const struct bpf_reg_types sock_types = {
4807 static const struct bpf_reg_types btf_id_sock_common_types = {
4815 .btf_id = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
4819 static const struct bpf_reg_types mem_types = {
4832 static const struct bpf_reg_types int_ptr_types = {
4842 static const struct bpf_reg_types fullsock_types = { .types = { PTR_TO_SOCKET } };
4843 static const struct bpf_reg_types scalar_types = { .types = { SCALAR_VALUE } };
4844 static const struct bpf_reg_types context_types = { .types = { PTR_TO_CTX } };
4845 static const struct bpf_reg_types alloc_mem_types = { .types = { PTR_TO_MEM } };
4846 static const struct bpf_reg_types const_map_ptr_types = { .types = { CONST_PTR_TO_MAP } };
4847 static const struct bpf_reg_types btf_ptr_types = { .types = { PTR_TO_BTF_ID } };
4848 static const struct bpf_reg_types spin_lock_types = { .types = { PTR_TO_MAP_VALUE } };
4849 static const struct bpf_reg_types percpu_btf_ptr_types = { .types = { PTR_TO_PERCPU_BTF_ID } };
4850 static const struct bpf_reg_types func_ptr_types = { .types = { PTR_TO_FUNC } };
4851 static const struct bpf_reg_types stack_ptr_types = { .types = { PTR_TO_STACK } };
4852 static const struct bpf_reg_types const_str_ptr_types = { .types = { PTR_TO_MAP_VALUE } };
4853 static const struct bpf_reg_types timer_types = { .types = { PTR_TO_MAP_VALUE } };
4855 static const struct bpf_reg_types *compatible_reg_types[__BPF_ARG_TYPE_MAX] = {
4856 [ARG_PTR_TO_MAP_KEY] = &map_key_value_types,
4857 [ARG_PTR_TO_MAP_VALUE] = &map_key_value_types,
4858 [ARG_PTR_TO_UNINIT_MAP_VALUE] = &map_key_value_types,
4859 [ARG_PTR_TO_MAP_VALUE_OR_NULL] = &map_key_value_types,
4860 [ARG_CONST_SIZE] = &scalar_types,
4861 [ARG_CONST_SIZE_OR_ZERO] = &scalar_types,
4862 [ARG_CONST_ALLOC_SIZE_OR_ZERO] = &scalar_types,
4863 [ARG_CONST_MAP_PTR] = &const_map_ptr_types,
4864 [ARG_PTR_TO_CTX] = &context_types,
4865 [ARG_PTR_TO_CTX_OR_NULL] = &context_types,
4866 [ARG_PTR_TO_SOCK_COMMON] = &sock_types,
4868 [ARG_PTR_TO_BTF_ID_SOCK_COMMON] = &btf_id_sock_common_types,
4870 [ARG_PTR_TO_SOCKET] = &fullsock_types,
4871 [ARG_PTR_TO_SOCKET_OR_NULL] = &fullsock_types,
4872 [ARG_PTR_TO_BTF_ID] = &btf_ptr_types,
4873 [ARG_PTR_TO_SPIN_LOCK] = &spin_lock_types,
4874 [ARG_PTR_TO_MEM] = &mem_types,
4875 [ARG_PTR_TO_MEM_OR_NULL] = &mem_types,
4876 [ARG_PTR_TO_UNINIT_MEM] = &mem_types,
4877 [ARG_PTR_TO_ALLOC_MEM] = &alloc_mem_types,
4878 [ARG_PTR_TO_ALLOC_MEM_OR_NULL] = &alloc_mem_types,
4879 [ARG_PTR_TO_INT] = &int_ptr_types,
4880 [ARG_PTR_TO_LONG] = &int_ptr_types,
4881 [ARG_PTR_TO_PERCPU_BTF_ID] = &percpu_btf_ptr_types,
4882 [ARG_PTR_TO_FUNC] = &func_ptr_types,
4883 [ARG_PTR_TO_STACK_OR_NULL] = &stack_ptr_types,
4884 [ARG_PTR_TO_CONST_STR] = &const_str_ptr_types,
4885 [ARG_PTR_TO_TIMER] = &timer_types,
4888 static int check_reg_type(struct bpf_verifier_env *env, u32 regno,
4889 enum bpf_arg_type arg_type,
4890 const u32 *arg_btf_id)
4892 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4893 enum bpf_reg_type expected, type = reg->type;
4894 const struct bpf_reg_types *compatible;
4897 compatible = compatible_reg_types[arg_type];
4899 verbose(env, "verifier internal error: unsupported arg type %d\n", arg_type);
4903 for (i = 0; i < ARRAY_SIZE(compatible->types); i++) {
4904 expected = compatible->types[i];
4905 if (expected == NOT_INIT)
4908 if (type == expected)
4912 verbose(env, "R%d type=%s expected=", regno, reg_type_str[type]);
4913 for (j = 0; j + 1 < i; j++)
4914 verbose(env, "%s, ", reg_type_str[compatible->types[j]]);
4915 verbose(env, "%s\n", reg_type_str[compatible->types[j]]);
4919 if (type == PTR_TO_BTF_ID) {
4921 if (!compatible->btf_id) {
4922 verbose(env, "verifier internal error: missing arg compatible BTF ID\n");
4925 arg_btf_id = compatible->btf_id;
4928 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
4929 btf_vmlinux, *arg_btf_id)) {
4930 verbose(env, "R%d is of type %s but %s is expected\n",
4931 regno, kernel_type_name(reg->btf, reg->btf_id),
4932 kernel_type_name(btf_vmlinux, *arg_btf_id));
4936 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
4937 verbose(env, "R%d is a pointer to in-kernel struct with non-zero offset\n",
4946 static int check_func_arg(struct bpf_verifier_env *env, u32 arg,
4947 struct bpf_call_arg_meta *meta,
4948 const struct bpf_func_proto *fn)
4950 u32 regno = BPF_REG_1 + arg;
4951 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4952 enum bpf_arg_type arg_type = fn->arg_type[arg];
4953 enum bpf_reg_type type = reg->type;
4956 if (arg_type == ARG_DONTCARE)
4959 err = check_reg_arg(env, regno, SRC_OP);
4963 if (arg_type == ARG_ANYTHING) {
4964 if (is_pointer_value(env, regno)) {
4965 verbose(env, "R%d leaks addr into helper function\n",
4972 if (type_is_pkt_pointer(type) &&
4973 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
4974 verbose(env, "helper access to the packet is not allowed\n");
4978 if (arg_type == ARG_PTR_TO_MAP_VALUE ||
4979 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE ||
4980 arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL) {
4981 err = resolve_map_arg_type(env, meta, &arg_type);
4986 if (register_is_null(reg) && arg_type_may_be_null(arg_type))
4987 /* A NULL register has a SCALAR_VALUE type, so skip
4990 goto skip_type_check;
4992 err = check_reg_type(env, regno, arg_type, fn->arg_btf_id[arg]);
4996 if (type == PTR_TO_CTX) {
4997 err = check_ctx_reg(env, reg, regno);
5003 if (reg->ref_obj_id) {
5004 if (meta->ref_obj_id) {
5005 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
5006 regno, reg->ref_obj_id,
5010 meta->ref_obj_id = reg->ref_obj_id;
5013 if (arg_type == ARG_CONST_MAP_PTR) {
5014 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
5015 if (meta->map_ptr) {
5016 /* Use map_uid (which is unique id of inner map) to reject:
5017 * inner_map1 = bpf_map_lookup_elem(outer_map, key1)
5018 * inner_map2 = bpf_map_lookup_elem(outer_map, key2)
5019 * if (inner_map1 && inner_map2) {
5020 * timer = bpf_map_lookup_elem(inner_map1);
5022 * // mismatch would have been allowed
5023 * bpf_timer_init(timer, inner_map2);
5026 * Comparing map_ptr is enough to distinguish normal and outer maps.
5028 if (meta->map_ptr != reg->map_ptr ||
5029 meta->map_uid != reg->map_uid) {
5031 "timer pointer in R1 map_uid=%d doesn't match map pointer in R2 map_uid=%d\n",
5032 meta->map_uid, reg->map_uid);
5036 meta->map_ptr = reg->map_ptr;
5037 meta->map_uid = reg->map_uid;
5038 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
5039 /* bpf_map_xxx(..., map_ptr, ..., key) call:
5040 * check that [key, key + map->key_size) are within
5041 * stack limits and initialized
5043 if (!meta->map_ptr) {
5044 /* in function declaration map_ptr must come before
5045 * map_key, so that it's verified and known before
5046 * we have to check map_key here. Otherwise it means
5047 * that kernel subsystem misconfigured verifier
5049 verbose(env, "invalid map_ptr to access map->key\n");
5052 err = check_helper_mem_access(env, regno,
5053 meta->map_ptr->key_size, false,
5055 } else if (arg_type == ARG_PTR_TO_MAP_VALUE ||
5056 (arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL &&
5057 !register_is_null(reg)) ||
5058 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
5059 /* bpf_map_xxx(..., map_ptr, ..., value) call:
5060 * check [value, value + map->value_size) validity
5062 if (!meta->map_ptr) {
5063 /* kernel subsystem misconfigured verifier */
5064 verbose(env, "invalid map_ptr to access map->value\n");
5067 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE);
5068 err = check_helper_mem_access(env, regno,
5069 meta->map_ptr->value_size, false,
5071 } else if (arg_type == ARG_PTR_TO_PERCPU_BTF_ID) {
5073 verbose(env, "Helper has invalid btf_id in R%d\n", regno);
5076 meta->ret_btf = reg->btf;
5077 meta->ret_btf_id = reg->btf_id;
5078 } else if (arg_type == ARG_PTR_TO_SPIN_LOCK) {
5079 if (meta->func_id == BPF_FUNC_spin_lock) {
5080 if (process_spin_lock(env, regno, true))
5082 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
5083 if (process_spin_lock(env, regno, false))
5086 verbose(env, "verifier internal error\n");
5089 } else if (arg_type == ARG_PTR_TO_TIMER) {
5090 if (process_timer_func(env, regno, meta))
5092 } else if (arg_type == ARG_PTR_TO_FUNC) {
5093 meta->subprogno = reg->subprogno;
5094 } else if (arg_type_is_mem_ptr(arg_type)) {
5095 /* The access to this pointer is only checked when we hit the
5096 * next is_mem_size argument below.
5098 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MEM);
5099 } else if (arg_type_is_mem_size(arg_type)) {
5100 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
5102 /* This is used to refine r0 return value bounds for helpers
5103 * that enforce this value as an upper bound on return values.
5104 * See do_refine_retval_range() for helpers that can refine
5105 * the return value. C type of helper is u32 so we pull register
5106 * bound from umax_value however, if negative verifier errors
5107 * out. Only upper bounds can be learned because retval is an
5108 * int type and negative retvals are allowed.
5110 meta->msize_max_value = reg->umax_value;
5112 /* The register is SCALAR_VALUE; the access check
5113 * happens using its boundaries.
5115 if (!tnum_is_const(reg->var_off))
5116 /* For unprivileged variable accesses, disable raw
5117 * mode so that the program is required to
5118 * initialize all the memory that the helper could
5119 * just partially fill up.
5123 if (reg->smin_value < 0) {
5124 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
5129 if (reg->umin_value == 0) {
5130 err = check_helper_mem_access(env, regno - 1, 0,
5137 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
5138 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
5142 err = check_helper_mem_access(env, regno - 1,
5144 zero_size_allowed, meta);
5146 err = mark_chain_precision(env, regno);
5147 } else if (arg_type_is_alloc_size(arg_type)) {
5148 if (!tnum_is_const(reg->var_off)) {
5149 verbose(env, "R%d is not a known constant'\n",
5153 meta->mem_size = reg->var_off.value;
5154 } else if (arg_type_is_int_ptr(arg_type)) {
5155 int size = int_ptr_type_to_size(arg_type);
5157 err = check_helper_mem_access(env, regno, size, false, meta);
5160 err = check_ptr_alignment(env, reg, 0, size, true);
5161 } else if (arg_type == ARG_PTR_TO_CONST_STR) {
5162 struct bpf_map *map = reg->map_ptr;
5167 if (!bpf_map_is_rdonly(map)) {
5168 verbose(env, "R%d does not point to a readonly map'\n", regno);
5172 if (!tnum_is_const(reg->var_off)) {
5173 verbose(env, "R%d is not a constant address'\n", regno);
5177 if (!map->ops->map_direct_value_addr) {
5178 verbose(env, "no direct value access support for this map type\n");
5182 err = check_map_access(env, regno, reg->off,
5183 map->value_size - reg->off, false);
5187 map_off = reg->off + reg->var_off.value;
5188 err = map->ops->map_direct_value_addr(map, &map_addr, map_off);
5190 verbose(env, "direct value access on string failed\n");
5194 str_ptr = (char *)(long)(map_addr);
5195 if (!strnchr(str_ptr + map_off, map->value_size - map_off, 0)) {
5196 verbose(env, "string is not zero-terminated\n");
5204 static bool may_update_sockmap(struct bpf_verifier_env *env, int func_id)
5206 enum bpf_attach_type eatype = env->prog->expected_attach_type;
5207 enum bpf_prog_type type = resolve_prog_type(env->prog);
5209 if (func_id != BPF_FUNC_map_update_elem)
5212 /* It's not possible to get access to a locked struct sock in these
5213 * contexts, so updating is safe.
5216 case BPF_PROG_TYPE_TRACING:
5217 if (eatype == BPF_TRACE_ITER)
5220 case BPF_PROG_TYPE_SOCKET_FILTER:
5221 case BPF_PROG_TYPE_SCHED_CLS:
5222 case BPF_PROG_TYPE_SCHED_ACT:
5223 case BPF_PROG_TYPE_XDP:
5224 case BPF_PROG_TYPE_SK_REUSEPORT:
5225 case BPF_PROG_TYPE_FLOW_DISSECTOR:
5226 case BPF_PROG_TYPE_SK_LOOKUP:
5232 verbose(env, "cannot update sockmap in this context\n");
5236 static bool allow_tail_call_in_subprogs(struct bpf_verifier_env *env)
5238 return env->prog->jit_requested && IS_ENABLED(CONFIG_X86_64);
5241 static int check_map_func_compatibility(struct bpf_verifier_env *env,
5242 struct bpf_map *map, int func_id)
5247 /* We need a two way check, first is from map perspective ... */
5248 switch (map->map_type) {
5249 case BPF_MAP_TYPE_PROG_ARRAY:
5250 if (func_id != BPF_FUNC_tail_call)
5253 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
5254 if (func_id != BPF_FUNC_perf_event_read &&
5255 func_id != BPF_FUNC_perf_event_output &&
5256 func_id != BPF_FUNC_skb_output &&
5257 func_id != BPF_FUNC_perf_event_read_value &&
5258 func_id != BPF_FUNC_xdp_output)
5261 case BPF_MAP_TYPE_RINGBUF:
5262 if (func_id != BPF_FUNC_ringbuf_output &&
5263 func_id != BPF_FUNC_ringbuf_reserve &&
5264 func_id != BPF_FUNC_ringbuf_submit &&
5265 func_id != BPF_FUNC_ringbuf_discard &&
5266 func_id != BPF_FUNC_ringbuf_query)
5269 case BPF_MAP_TYPE_STACK_TRACE:
5270 if (func_id != BPF_FUNC_get_stackid)
5273 case BPF_MAP_TYPE_CGROUP_ARRAY:
5274 if (func_id != BPF_FUNC_skb_under_cgroup &&
5275 func_id != BPF_FUNC_current_task_under_cgroup)
5278 case BPF_MAP_TYPE_CGROUP_STORAGE:
5279 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
5280 if (func_id != BPF_FUNC_get_local_storage)
5283 case BPF_MAP_TYPE_DEVMAP:
5284 case BPF_MAP_TYPE_DEVMAP_HASH:
5285 if (func_id != BPF_FUNC_redirect_map &&
5286 func_id != BPF_FUNC_map_lookup_elem)
5289 /* Restrict bpf side of cpumap and xskmap, open when use-cases
5292 case BPF_MAP_TYPE_CPUMAP:
5293 if (func_id != BPF_FUNC_redirect_map)
5296 case BPF_MAP_TYPE_XSKMAP:
5297 if (func_id != BPF_FUNC_redirect_map &&
5298 func_id != BPF_FUNC_map_lookup_elem)
5301 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
5302 case BPF_MAP_TYPE_HASH_OF_MAPS:
5303 if (func_id != BPF_FUNC_map_lookup_elem)
5306 case BPF_MAP_TYPE_SOCKMAP:
5307 if (func_id != BPF_FUNC_sk_redirect_map &&
5308 func_id != BPF_FUNC_sock_map_update &&
5309 func_id != BPF_FUNC_map_delete_elem &&
5310 func_id != BPF_FUNC_msg_redirect_map &&
5311 func_id != BPF_FUNC_sk_select_reuseport &&
5312 func_id != BPF_FUNC_map_lookup_elem &&
5313 !may_update_sockmap(env, func_id))
5316 case BPF_MAP_TYPE_SOCKHASH:
5317 if (func_id != BPF_FUNC_sk_redirect_hash &&
5318 func_id != BPF_FUNC_sock_hash_update &&
5319 func_id != BPF_FUNC_map_delete_elem &&
5320 func_id != BPF_FUNC_msg_redirect_hash &&
5321 func_id != BPF_FUNC_sk_select_reuseport &&
5322 func_id != BPF_FUNC_map_lookup_elem &&
5323 !may_update_sockmap(env, func_id))
5326 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
5327 if (func_id != BPF_FUNC_sk_select_reuseport)
5330 case BPF_MAP_TYPE_QUEUE:
5331 case BPF_MAP_TYPE_STACK:
5332 if (func_id != BPF_FUNC_map_peek_elem &&
5333 func_id != BPF_FUNC_map_pop_elem &&
5334 func_id != BPF_FUNC_map_push_elem)
5337 case BPF_MAP_TYPE_SK_STORAGE:
5338 if (func_id != BPF_FUNC_sk_storage_get &&
5339 func_id != BPF_FUNC_sk_storage_delete)
5342 case BPF_MAP_TYPE_INODE_STORAGE:
5343 if (func_id != BPF_FUNC_inode_storage_get &&
5344 func_id != BPF_FUNC_inode_storage_delete)
5347 case BPF_MAP_TYPE_TASK_STORAGE:
5348 if (func_id != BPF_FUNC_task_storage_get &&
5349 func_id != BPF_FUNC_task_storage_delete)
5356 /* ... and second from the function itself. */
5358 case BPF_FUNC_tail_call:
5359 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
5361 if (env->subprog_cnt > 1 && !allow_tail_call_in_subprogs(env)) {
5362 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
5366 case BPF_FUNC_perf_event_read:
5367 case BPF_FUNC_perf_event_output:
5368 case BPF_FUNC_perf_event_read_value:
5369 case BPF_FUNC_skb_output:
5370 case BPF_FUNC_xdp_output:
5371 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
5374 case BPF_FUNC_get_stackid:
5375 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
5378 case BPF_FUNC_current_task_under_cgroup:
5379 case BPF_FUNC_skb_under_cgroup:
5380 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
5383 case BPF_FUNC_redirect_map:
5384 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
5385 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
5386 map->map_type != BPF_MAP_TYPE_CPUMAP &&
5387 map->map_type != BPF_MAP_TYPE_XSKMAP)
5390 case BPF_FUNC_sk_redirect_map:
5391 case BPF_FUNC_msg_redirect_map:
5392 case BPF_FUNC_sock_map_update:
5393 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
5396 case BPF_FUNC_sk_redirect_hash:
5397 case BPF_FUNC_msg_redirect_hash:
5398 case BPF_FUNC_sock_hash_update:
5399 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
5402 case BPF_FUNC_get_local_storage:
5403 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
5404 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
5407 case BPF_FUNC_sk_select_reuseport:
5408 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
5409 map->map_type != BPF_MAP_TYPE_SOCKMAP &&
5410 map->map_type != BPF_MAP_TYPE_SOCKHASH)
5413 case BPF_FUNC_map_peek_elem:
5414 case BPF_FUNC_map_pop_elem:
5415 case BPF_FUNC_map_push_elem:
5416 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
5417 map->map_type != BPF_MAP_TYPE_STACK)
5420 case BPF_FUNC_sk_storage_get:
5421 case BPF_FUNC_sk_storage_delete:
5422 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
5425 case BPF_FUNC_inode_storage_get:
5426 case BPF_FUNC_inode_storage_delete:
5427 if (map->map_type != BPF_MAP_TYPE_INODE_STORAGE)
5430 case BPF_FUNC_task_storage_get:
5431 case BPF_FUNC_task_storage_delete:
5432 if (map->map_type != BPF_MAP_TYPE_TASK_STORAGE)
5441 verbose(env, "cannot pass map_type %d into func %s#%d\n",
5442 map->map_type, func_id_name(func_id), func_id);
5446 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
5450 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
5452 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
5454 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
5456 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
5458 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
5461 /* We only support one arg being in raw mode at the moment,
5462 * which is sufficient for the helper functions we have
5468 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
5469 enum bpf_arg_type arg_next)
5471 return (arg_type_is_mem_ptr(arg_curr) &&
5472 !arg_type_is_mem_size(arg_next)) ||
5473 (!arg_type_is_mem_ptr(arg_curr) &&
5474 arg_type_is_mem_size(arg_next));
5477 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
5479 /* bpf_xxx(..., buf, len) call will access 'len'
5480 * bytes from memory 'buf'. Both arg types need
5481 * to be paired, so make sure there's no buggy
5482 * helper function specification.
5484 if (arg_type_is_mem_size(fn->arg1_type) ||
5485 arg_type_is_mem_ptr(fn->arg5_type) ||
5486 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
5487 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
5488 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
5489 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
5495 static bool check_refcount_ok(const struct bpf_func_proto *fn, int func_id)
5499 if (arg_type_may_be_refcounted(fn->arg1_type))
5501 if (arg_type_may_be_refcounted(fn->arg2_type))
5503 if (arg_type_may_be_refcounted(fn->arg3_type))
5505 if (arg_type_may_be_refcounted(fn->arg4_type))
5507 if (arg_type_may_be_refcounted(fn->arg5_type))
5510 /* A reference acquiring function cannot acquire
5511 * another refcounted ptr.
5513 if (may_be_acquire_function(func_id) && count)
5516 /* We only support one arg being unreferenced at the moment,
5517 * which is sufficient for the helper functions we have right now.
5522 static bool check_btf_id_ok(const struct bpf_func_proto *fn)
5526 for (i = 0; i < ARRAY_SIZE(fn->arg_type); i++) {
5527 if (fn->arg_type[i] == ARG_PTR_TO_BTF_ID && !fn->arg_btf_id[i])
5530 if (fn->arg_type[i] != ARG_PTR_TO_BTF_ID && fn->arg_btf_id[i])
5537 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
5539 return check_raw_mode_ok(fn) &&
5540 check_arg_pair_ok(fn) &&
5541 check_btf_id_ok(fn) &&
5542 check_refcount_ok(fn, func_id) ? 0 : -EINVAL;
5545 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
5546 * are now invalid, so turn them into unknown SCALAR_VALUE.
5548 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
5549 struct bpf_func_state *state)
5551 struct bpf_reg_state *regs = state->regs, *reg;
5554 for (i = 0; i < MAX_BPF_REG; i++)
5555 if (reg_is_pkt_pointer_any(®s[i]))
5556 mark_reg_unknown(env, regs, i);
5558 bpf_for_each_spilled_reg(i, state, reg) {
5561 if (reg_is_pkt_pointer_any(reg))
5562 __mark_reg_unknown(env, reg);
5566 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
5568 struct bpf_verifier_state *vstate = env->cur_state;
5571 for (i = 0; i <= vstate->curframe; i++)
5572 __clear_all_pkt_pointers(env, vstate->frame[i]);
5577 BEYOND_PKT_END = -2,
5580 static void mark_pkt_end(struct bpf_verifier_state *vstate, int regn, bool range_open)
5582 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5583 struct bpf_reg_state *reg = &state->regs[regn];
5585 if (reg->type != PTR_TO_PACKET)
5586 /* PTR_TO_PACKET_META is not supported yet */
5589 /* The 'reg' is pkt > pkt_end or pkt >= pkt_end.
5590 * How far beyond pkt_end it goes is unknown.
5591 * if (!range_open) it's the case of pkt >= pkt_end
5592 * if (range_open) it's the case of pkt > pkt_end
5593 * hence this pointer is at least 1 byte bigger than pkt_end
5596 reg->range = BEYOND_PKT_END;
5598 reg->range = AT_PKT_END;
5601 static void release_reg_references(struct bpf_verifier_env *env,
5602 struct bpf_func_state *state,
5605 struct bpf_reg_state *regs = state->regs, *reg;
5608 for (i = 0; i < MAX_BPF_REG; i++)
5609 if (regs[i].ref_obj_id == ref_obj_id)
5610 mark_reg_unknown(env, regs, i);
5612 bpf_for_each_spilled_reg(i, state, reg) {
5615 if (reg->ref_obj_id == ref_obj_id)
5616 __mark_reg_unknown(env, reg);
5620 /* The pointer with the specified id has released its reference to kernel
5621 * resources. Identify all copies of the same pointer and clear the reference.
5623 static int release_reference(struct bpf_verifier_env *env,
5626 struct bpf_verifier_state *vstate = env->cur_state;
5630 err = release_reference_state(cur_func(env), ref_obj_id);
5634 for (i = 0; i <= vstate->curframe; i++)
5635 release_reg_references(env, vstate->frame[i], ref_obj_id);
5640 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
5641 struct bpf_reg_state *regs)
5645 /* after the call registers r0 - r5 were scratched */
5646 for (i = 0; i < CALLER_SAVED_REGS; i++) {
5647 mark_reg_not_init(env, regs, caller_saved[i]);
5648 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
5652 typedef int (*set_callee_state_fn)(struct bpf_verifier_env *env,
5653 struct bpf_func_state *caller,
5654 struct bpf_func_state *callee,
5657 static int __check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
5658 int *insn_idx, int subprog,
5659 set_callee_state_fn set_callee_state_cb)
5661 struct bpf_verifier_state *state = env->cur_state;
5662 struct bpf_func_info_aux *func_info_aux;
5663 struct bpf_func_state *caller, *callee;
5665 bool is_global = false;
5667 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
5668 verbose(env, "the call stack of %d frames is too deep\n",
5669 state->curframe + 2);
5673 caller = state->frame[state->curframe];
5674 if (state->frame[state->curframe + 1]) {
5675 verbose(env, "verifier bug. Frame %d already allocated\n",
5676 state->curframe + 1);
5680 func_info_aux = env->prog->aux->func_info_aux;
5682 is_global = func_info_aux[subprog].linkage == BTF_FUNC_GLOBAL;
5683 err = btf_check_subprog_arg_match(env, subprog, caller->regs);
5688 verbose(env, "Caller passes invalid args into func#%d\n",
5692 if (env->log.level & BPF_LOG_LEVEL)
5694 "Func#%d is global and valid. Skipping.\n",
5696 clear_caller_saved_regs(env, caller->regs);
5698 /* All global functions return a 64-bit SCALAR_VALUE */
5699 mark_reg_unknown(env, caller->regs, BPF_REG_0);
5700 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
5702 /* continue with next insn after call */
5707 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
5710 state->frame[state->curframe + 1] = callee;
5712 /* callee cannot access r0, r6 - r9 for reading and has to write
5713 * into its own stack before reading from it.
5714 * callee can read/write into caller's stack
5716 init_func_state(env, callee,
5717 /* remember the callsite, it will be used by bpf_exit */
5718 *insn_idx /* callsite */,
5719 state->curframe + 1 /* frameno within this callchain */,
5720 subprog /* subprog number within this prog */);
5722 /* Transfer references to the callee */
5723 err = copy_reference_state(callee, caller);
5727 err = set_callee_state_cb(env, caller, callee, *insn_idx);
5731 clear_caller_saved_regs(env, caller->regs);
5733 /* only increment it after check_reg_arg() finished */
5736 /* and go analyze first insn of the callee */
5737 *insn_idx = env->subprog_info[subprog].start - 1;
5739 if (env->log.level & BPF_LOG_LEVEL) {
5740 verbose(env, "caller:\n");
5741 print_verifier_state(env, caller);
5742 verbose(env, "callee:\n");
5743 print_verifier_state(env, callee);
5748 int map_set_for_each_callback_args(struct bpf_verifier_env *env,
5749 struct bpf_func_state *caller,
5750 struct bpf_func_state *callee)
5752 /* bpf_for_each_map_elem(struct bpf_map *map, void *callback_fn,
5753 * void *callback_ctx, u64 flags);
5754 * callback_fn(struct bpf_map *map, void *key, void *value,
5755 * void *callback_ctx);
5757 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
5759 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
5760 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
5761 callee->regs[BPF_REG_2].map_ptr = caller->regs[BPF_REG_1].map_ptr;
5763 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
5764 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
5765 callee->regs[BPF_REG_3].map_ptr = caller->regs[BPF_REG_1].map_ptr;
5767 /* pointer to stack or null */
5768 callee->regs[BPF_REG_4] = caller->regs[BPF_REG_3];
5771 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
5775 static int set_callee_state(struct bpf_verifier_env *env,
5776 struct bpf_func_state *caller,
5777 struct bpf_func_state *callee, int insn_idx)
5781 /* copy r1 - r5 args that callee can access. The copy includes parent
5782 * pointers, which connects us up to the liveness chain
5784 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
5785 callee->regs[i] = caller->regs[i];
5789 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
5792 int subprog, target_insn;
5794 target_insn = *insn_idx + insn->imm + 1;
5795 subprog = find_subprog(env, target_insn);
5797 verbose(env, "verifier bug. No program starts at insn %d\n",
5802 return __check_func_call(env, insn, insn_idx, subprog, set_callee_state);
5805 static int set_map_elem_callback_state(struct bpf_verifier_env *env,
5806 struct bpf_func_state *caller,
5807 struct bpf_func_state *callee,
5810 struct bpf_insn_aux_data *insn_aux = &env->insn_aux_data[insn_idx];
5811 struct bpf_map *map;
5814 if (bpf_map_ptr_poisoned(insn_aux)) {
5815 verbose(env, "tail_call abusing map_ptr\n");
5819 map = BPF_MAP_PTR(insn_aux->map_ptr_state);
5820 if (!map->ops->map_set_for_each_callback_args ||
5821 !map->ops->map_for_each_callback) {
5822 verbose(env, "callback function not allowed for map\n");
5826 err = map->ops->map_set_for_each_callback_args(env, caller, callee);
5830 callee->in_callback_fn = true;
5834 static int set_timer_callback_state(struct bpf_verifier_env *env,
5835 struct bpf_func_state *caller,
5836 struct bpf_func_state *callee,
5839 struct bpf_map *map_ptr = caller->regs[BPF_REG_1].map_ptr;
5841 /* bpf_timer_set_callback(struct bpf_timer *timer, void *callback_fn);
5842 * callback_fn(struct bpf_map *map, void *key, void *value);
5844 callee->regs[BPF_REG_1].type = CONST_PTR_TO_MAP;
5845 __mark_reg_known_zero(&callee->regs[BPF_REG_1]);
5846 callee->regs[BPF_REG_1].map_ptr = map_ptr;
5848 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
5849 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
5850 callee->regs[BPF_REG_2].map_ptr = map_ptr;
5852 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
5853 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
5854 callee->regs[BPF_REG_3].map_ptr = map_ptr;
5857 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
5858 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
5862 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
5864 struct bpf_verifier_state *state = env->cur_state;
5865 struct bpf_func_state *caller, *callee;
5866 struct bpf_reg_state *r0;
5869 callee = state->frame[state->curframe];
5870 r0 = &callee->regs[BPF_REG_0];
5871 if (r0->type == PTR_TO_STACK) {
5872 /* technically it's ok to return caller's stack pointer
5873 * (or caller's caller's pointer) back to the caller,
5874 * since these pointers are valid. Only current stack
5875 * pointer will be invalid as soon as function exits,
5876 * but let's be conservative
5878 verbose(env, "cannot return stack pointer to the caller\n");
5883 caller = state->frame[state->curframe];
5884 if (callee->in_callback_fn) {
5885 /* enforce R0 return value range [0, 1]. */
5886 struct tnum range = tnum_range(0, 1);
5888 if (r0->type != SCALAR_VALUE) {
5889 verbose(env, "R0 not a scalar value\n");
5892 if (!tnum_in(range, r0->var_off)) {
5893 verbose_invalid_scalar(env, r0, &range, "callback return", "R0");
5897 /* return to the caller whatever r0 had in the callee */
5898 caller->regs[BPF_REG_0] = *r0;
5901 /* Transfer references to the caller */
5902 err = copy_reference_state(caller, callee);
5906 *insn_idx = callee->callsite + 1;
5907 if (env->log.level & BPF_LOG_LEVEL) {
5908 verbose(env, "returning from callee:\n");
5909 print_verifier_state(env, callee);
5910 verbose(env, "to caller at %d:\n", *insn_idx);
5911 print_verifier_state(env, caller);
5913 /* clear everything in the callee */
5914 free_func_state(callee);
5915 state->frame[state->curframe + 1] = NULL;
5919 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
5921 struct bpf_call_arg_meta *meta)
5923 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
5925 if (ret_type != RET_INTEGER ||
5926 (func_id != BPF_FUNC_get_stack &&
5927 func_id != BPF_FUNC_get_task_stack &&
5928 func_id != BPF_FUNC_probe_read_str &&
5929 func_id != BPF_FUNC_probe_read_kernel_str &&
5930 func_id != BPF_FUNC_probe_read_user_str))
5933 ret_reg->smax_value = meta->msize_max_value;
5934 ret_reg->s32_max_value = meta->msize_max_value;
5935 ret_reg->smin_value = -MAX_ERRNO;
5936 ret_reg->s32_min_value = -MAX_ERRNO;
5937 __reg_deduce_bounds(ret_reg);
5938 __reg_bound_offset(ret_reg);
5939 __update_reg_bounds(ret_reg);
5943 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
5944 int func_id, int insn_idx)
5946 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
5947 struct bpf_map *map = meta->map_ptr;
5949 if (func_id != BPF_FUNC_tail_call &&
5950 func_id != BPF_FUNC_map_lookup_elem &&
5951 func_id != BPF_FUNC_map_update_elem &&
5952 func_id != BPF_FUNC_map_delete_elem &&
5953 func_id != BPF_FUNC_map_push_elem &&
5954 func_id != BPF_FUNC_map_pop_elem &&
5955 func_id != BPF_FUNC_map_peek_elem &&
5956 func_id != BPF_FUNC_for_each_map_elem &&
5957 func_id != BPF_FUNC_redirect_map)
5961 verbose(env, "kernel subsystem misconfigured verifier\n");
5965 /* In case of read-only, some additional restrictions
5966 * need to be applied in order to prevent altering the
5967 * state of the map from program side.
5969 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
5970 (func_id == BPF_FUNC_map_delete_elem ||
5971 func_id == BPF_FUNC_map_update_elem ||
5972 func_id == BPF_FUNC_map_push_elem ||
5973 func_id == BPF_FUNC_map_pop_elem)) {
5974 verbose(env, "write into map forbidden\n");
5978 if (!BPF_MAP_PTR(aux->map_ptr_state))
5979 bpf_map_ptr_store(aux, meta->map_ptr,
5980 !meta->map_ptr->bypass_spec_v1);
5981 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
5982 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
5983 !meta->map_ptr->bypass_spec_v1);
5988 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
5989 int func_id, int insn_idx)
5991 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
5992 struct bpf_reg_state *regs = cur_regs(env), *reg;
5993 struct bpf_map *map = meta->map_ptr;
5998 if (func_id != BPF_FUNC_tail_call)
6000 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
6001 verbose(env, "kernel subsystem misconfigured verifier\n");
6005 range = tnum_range(0, map->max_entries - 1);
6006 reg = ®s[BPF_REG_3];
6008 if (!register_is_const(reg) || !tnum_in(range, reg->var_off)) {
6009 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
6013 err = mark_chain_precision(env, BPF_REG_3);
6017 val = reg->var_off.value;
6018 if (bpf_map_key_unseen(aux))
6019 bpf_map_key_store(aux, val);
6020 else if (!bpf_map_key_poisoned(aux) &&
6021 bpf_map_key_immediate(aux) != val)
6022 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
6026 static int check_reference_leak(struct bpf_verifier_env *env)
6028 struct bpf_func_state *state = cur_func(env);
6031 for (i = 0; i < state->acquired_refs; i++) {
6032 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
6033 state->refs[i].id, state->refs[i].insn_idx);
6035 return state->acquired_refs ? -EINVAL : 0;
6038 static int check_bpf_snprintf_call(struct bpf_verifier_env *env,
6039 struct bpf_reg_state *regs)
6041 struct bpf_reg_state *fmt_reg = ®s[BPF_REG_3];
6042 struct bpf_reg_state *data_len_reg = ®s[BPF_REG_5];
6043 struct bpf_map *fmt_map = fmt_reg->map_ptr;
6044 int err, fmt_map_off, num_args;
6048 /* data must be an array of u64 */
6049 if (data_len_reg->var_off.value % 8)
6051 num_args = data_len_reg->var_off.value / 8;
6053 /* fmt being ARG_PTR_TO_CONST_STR guarantees that var_off is const
6054 * and map_direct_value_addr is set.
6056 fmt_map_off = fmt_reg->off + fmt_reg->var_off.value;
6057 err = fmt_map->ops->map_direct_value_addr(fmt_map, &fmt_addr,
6060 verbose(env, "verifier bug\n");
6063 fmt = (char *)(long)fmt_addr + fmt_map_off;
6065 /* We are also guaranteed that fmt+fmt_map_off is NULL terminated, we
6066 * can focus on validating the format specifiers.
6068 err = bpf_bprintf_prepare(fmt, UINT_MAX, NULL, NULL, num_args);
6070 verbose(env, "Invalid format string\n");
6075 static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
6078 const struct bpf_func_proto *fn = NULL;
6079 struct bpf_reg_state *regs;
6080 struct bpf_call_arg_meta meta;
6081 int insn_idx = *insn_idx_p;
6083 int i, err, func_id;
6085 /* find function prototype */
6086 func_id = insn->imm;
6087 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
6088 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
6093 if (env->ops->get_func_proto)
6094 fn = env->ops->get_func_proto(func_id, env->prog);
6096 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
6101 /* eBPF programs must be GPL compatible to use GPL-ed functions */
6102 if (!env->prog->gpl_compatible && fn->gpl_only) {
6103 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
6107 if (fn->allowed && !fn->allowed(env->prog)) {
6108 verbose(env, "helper call is not allowed in probe\n");
6112 /* With LD_ABS/IND some JITs save/restore skb from r1. */
6113 changes_data = bpf_helper_changes_pkt_data(fn->func);
6114 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
6115 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
6116 func_id_name(func_id), func_id);
6120 memset(&meta, 0, sizeof(meta));
6121 meta.pkt_access = fn->pkt_access;
6123 err = check_func_proto(fn, func_id);
6125 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
6126 func_id_name(func_id), func_id);
6130 meta.func_id = func_id;
6132 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++) {
6133 err = check_func_arg(env, i, &meta, fn);
6138 err = record_func_map(env, &meta, func_id, insn_idx);
6142 err = record_func_key(env, &meta, func_id, insn_idx);
6146 /* Mark slots with STACK_MISC in case of raw mode, stack offset
6147 * is inferred from register state.
6149 for (i = 0; i < meta.access_size; i++) {
6150 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
6151 BPF_WRITE, -1, false);
6156 if (func_id == BPF_FUNC_tail_call) {
6157 err = check_reference_leak(env);
6159 verbose(env, "tail_call would lead to reference leak\n");
6162 } else if (is_release_function(func_id)) {
6163 err = release_reference(env, meta.ref_obj_id);
6165 verbose(env, "func %s#%d reference has not been acquired before\n",
6166 func_id_name(func_id), func_id);
6171 regs = cur_regs(env);
6173 /* check that flags argument in get_local_storage(map, flags) is 0,
6174 * this is required because get_local_storage() can't return an error.
6176 if (func_id == BPF_FUNC_get_local_storage &&
6177 !register_is_null(®s[BPF_REG_2])) {
6178 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
6182 if (func_id == BPF_FUNC_for_each_map_elem) {
6183 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
6184 set_map_elem_callback_state);
6189 if (func_id == BPF_FUNC_timer_set_callback) {
6190 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
6191 set_timer_callback_state);
6196 if (func_id == BPF_FUNC_snprintf) {
6197 err = check_bpf_snprintf_call(env, regs);
6202 /* reset caller saved regs */
6203 for (i = 0; i < CALLER_SAVED_REGS; i++) {
6204 mark_reg_not_init(env, regs, caller_saved[i]);
6205 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
6208 /* helper call returns 64-bit value. */
6209 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
6211 /* update return register (already marked as written above) */
6212 if (fn->ret_type == RET_INTEGER) {
6213 /* sets type to SCALAR_VALUE */
6214 mark_reg_unknown(env, regs, BPF_REG_0);
6215 } else if (fn->ret_type == RET_VOID) {
6216 regs[BPF_REG_0].type = NOT_INIT;
6217 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL ||
6218 fn->ret_type == RET_PTR_TO_MAP_VALUE) {
6219 /* There is no offset yet applied, variable or fixed */
6220 mark_reg_known_zero(env, regs, BPF_REG_0);
6221 /* remember map_ptr, so that check_map_access()
6222 * can check 'value_size' boundary of memory access
6223 * to map element returned from bpf_map_lookup_elem()
6225 if (meta.map_ptr == NULL) {
6227 "kernel subsystem misconfigured verifier\n");
6230 regs[BPF_REG_0].map_ptr = meta.map_ptr;
6231 regs[BPF_REG_0].map_uid = meta.map_uid;
6232 if (fn->ret_type == RET_PTR_TO_MAP_VALUE) {
6233 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE;
6234 if (map_value_has_spin_lock(meta.map_ptr))
6235 regs[BPF_REG_0].id = ++env->id_gen;
6237 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
6239 } else if (fn->ret_type == RET_PTR_TO_SOCKET_OR_NULL) {
6240 mark_reg_known_zero(env, regs, BPF_REG_0);
6241 regs[BPF_REG_0].type = PTR_TO_SOCKET_OR_NULL;
6242 } else if (fn->ret_type == RET_PTR_TO_SOCK_COMMON_OR_NULL) {
6243 mark_reg_known_zero(env, regs, BPF_REG_0);
6244 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON_OR_NULL;
6245 } else if (fn->ret_type == RET_PTR_TO_TCP_SOCK_OR_NULL) {
6246 mark_reg_known_zero(env, regs, BPF_REG_0);
6247 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK_OR_NULL;
6248 } else if (fn->ret_type == RET_PTR_TO_ALLOC_MEM_OR_NULL) {
6249 mark_reg_known_zero(env, regs, BPF_REG_0);
6250 regs[BPF_REG_0].type = PTR_TO_MEM_OR_NULL;
6251 regs[BPF_REG_0].mem_size = meta.mem_size;
6252 } else if (fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID_OR_NULL ||
6253 fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID) {
6254 const struct btf_type *t;
6256 mark_reg_known_zero(env, regs, BPF_REG_0);
6257 t = btf_type_skip_modifiers(meta.ret_btf, meta.ret_btf_id, NULL);
6258 if (!btf_type_is_struct(t)) {
6260 const struct btf_type *ret;
6263 /* resolve the type size of ksym. */
6264 ret = btf_resolve_size(meta.ret_btf, t, &tsize);
6266 tname = btf_name_by_offset(meta.ret_btf, t->name_off);
6267 verbose(env, "unable to resolve the size of type '%s': %ld\n",
6268 tname, PTR_ERR(ret));
6271 regs[BPF_REG_0].type =
6272 fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID ?
6273 PTR_TO_MEM : PTR_TO_MEM_OR_NULL;
6274 regs[BPF_REG_0].mem_size = tsize;
6276 regs[BPF_REG_0].type =
6277 fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID ?
6278 PTR_TO_BTF_ID : PTR_TO_BTF_ID_OR_NULL;
6279 regs[BPF_REG_0].btf = meta.ret_btf;
6280 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
6282 } else if (fn->ret_type == RET_PTR_TO_BTF_ID_OR_NULL ||
6283 fn->ret_type == RET_PTR_TO_BTF_ID) {
6286 mark_reg_known_zero(env, regs, BPF_REG_0);
6287 regs[BPF_REG_0].type = fn->ret_type == RET_PTR_TO_BTF_ID ?
6289 PTR_TO_BTF_ID_OR_NULL;
6290 ret_btf_id = *fn->ret_btf_id;
6291 if (ret_btf_id == 0) {
6292 verbose(env, "invalid return type %d of func %s#%d\n",
6293 fn->ret_type, func_id_name(func_id), func_id);
6296 /* current BPF helper definitions are only coming from
6297 * built-in code with type IDs from vmlinux BTF
6299 regs[BPF_REG_0].btf = btf_vmlinux;
6300 regs[BPF_REG_0].btf_id = ret_btf_id;
6302 verbose(env, "unknown return type %d of func %s#%d\n",
6303 fn->ret_type, func_id_name(func_id), func_id);
6307 if (reg_type_may_be_null(regs[BPF_REG_0].type))
6308 regs[BPF_REG_0].id = ++env->id_gen;
6310 if (is_ptr_cast_function(func_id)) {
6311 /* For release_reference() */
6312 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
6313 } else if (is_acquire_function(func_id, meta.map_ptr)) {
6314 int id = acquire_reference_state(env, insn_idx);
6318 /* For mark_ptr_or_null_reg() */
6319 regs[BPF_REG_0].id = id;
6320 /* For release_reference() */
6321 regs[BPF_REG_0].ref_obj_id = id;
6324 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
6326 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
6330 if ((func_id == BPF_FUNC_get_stack ||
6331 func_id == BPF_FUNC_get_task_stack) &&
6332 !env->prog->has_callchain_buf) {
6333 const char *err_str;
6335 #ifdef CONFIG_PERF_EVENTS
6336 err = get_callchain_buffers(sysctl_perf_event_max_stack);
6337 err_str = "cannot get callchain buffer for func %s#%d\n";
6340 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
6343 verbose(env, err_str, func_id_name(func_id), func_id);
6347 env->prog->has_callchain_buf = true;
6350 if (func_id == BPF_FUNC_get_stackid || func_id == BPF_FUNC_get_stack)
6351 env->prog->call_get_stack = true;
6354 clear_all_pkt_pointers(env);
6358 /* mark_btf_func_reg_size() is used when the reg size is determined by
6359 * the BTF func_proto's return value size and argument.
6361 static void mark_btf_func_reg_size(struct bpf_verifier_env *env, u32 regno,
6364 struct bpf_reg_state *reg = &cur_regs(env)[regno];
6366 if (regno == BPF_REG_0) {
6367 /* Function return value */
6368 reg->live |= REG_LIVE_WRITTEN;
6369 reg->subreg_def = reg_size == sizeof(u64) ?
6370 DEF_NOT_SUBREG : env->insn_idx + 1;
6372 /* Function argument */
6373 if (reg_size == sizeof(u64)) {
6374 mark_insn_zext(env, reg);
6375 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
6377 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ32);
6382 static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn)
6384 const struct btf_type *t, *func, *func_proto, *ptr_type;
6385 struct bpf_reg_state *regs = cur_regs(env);
6386 const char *func_name, *ptr_type_name;
6387 u32 i, nargs, func_id, ptr_type_id;
6388 const struct btf_param *args;
6391 func_id = insn->imm;
6392 func = btf_type_by_id(btf_vmlinux, func_id);
6393 func_name = btf_name_by_offset(btf_vmlinux, func->name_off);
6394 func_proto = btf_type_by_id(btf_vmlinux, func->type);
6396 if (!env->ops->check_kfunc_call ||
6397 !env->ops->check_kfunc_call(func_id)) {
6398 verbose(env, "calling kernel function %s is not allowed\n",
6403 /* Check the arguments */
6404 err = btf_check_kfunc_arg_match(env, btf_vmlinux, func_id, regs);
6408 for (i = 0; i < CALLER_SAVED_REGS; i++)
6409 mark_reg_not_init(env, regs, caller_saved[i]);
6411 /* Check return type */
6412 t = btf_type_skip_modifiers(btf_vmlinux, func_proto->type, NULL);
6413 if (btf_type_is_scalar(t)) {
6414 mark_reg_unknown(env, regs, BPF_REG_0);
6415 mark_btf_func_reg_size(env, BPF_REG_0, t->size);
6416 } else if (btf_type_is_ptr(t)) {
6417 ptr_type = btf_type_skip_modifiers(btf_vmlinux, t->type,
6419 if (!btf_type_is_struct(ptr_type)) {
6420 ptr_type_name = btf_name_by_offset(btf_vmlinux,
6421 ptr_type->name_off);
6422 verbose(env, "kernel function %s returns pointer type %s %s is not supported\n",
6423 func_name, btf_type_str(ptr_type),
6427 mark_reg_known_zero(env, regs, BPF_REG_0);
6428 regs[BPF_REG_0].btf = btf_vmlinux;
6429 regs[BPF_REG_0].type = PTR_TO_BTF_ID;
6430 regs[BPF_REG_0].btf_id = ptr_type_id;
6431 mark_btf_func_reg_size(env, BPF_REG_0, sizeof(void *));
6432 } /* else { add_kfunc_call() ensures it is btf_type_is_void(t) } */
6434 nargs = btf_type_vlen(func_proto);
6435 args = (const struct btf_param *)(func_proto + 1);
6436 for (i = 0; i < nargs; i++) {
6439 t = btf_type_skip_modifiers(btf_vmlinux, args[i].type, NULL);
6440 if (btf_type_is_ptr(t))
6441 mark_btf_func_reg_size(env, regno, sizeof(void *));
6443 /* scalar. ensured by btf_check_kfunc_arg_match() */
6444 mark_btf_func_reg_size(env, regno, t->size);
6450 static bool signed_add_overflows(s64 a, s64 b)
6452 /* Do the add in u64, where overflow is well-defined */
6453 s64 res = (s64)((u64)a + (u64)b);
6460 static bool signed_add32_overflows(s32 a, s32 b)
6462 /* Do the add in u32, where overflow is well-defined */
6463 s32 res = (s32)((u32)a + (u32)b);
6470 static bool signed_sub_overflows(s64 a, s64 b)
6472 /* Do the sub in u64, where overflow is well-defined */
6473 s64 res = (s64)((u64)a - (u64)b);
6480 static bool signed_sub32_overflows(s32 a, s32 b)
6482 /* Do the sub in u32, where overflow is well-defined */
6483 s32 res = (s32)((u32)a - (u32)b);
6490 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
6491 const struct bpf_reg_state *reg,
6492 enum bpf_reg_type type)
6494 bool known = tnum_is_const(reg->var_off);
6495 s64 val = reg->var_off.value;
6496 s64 smin = reg->smin_value;
6498 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
6499 verbose(env, "math between %s pointer and %lld is not allowed\n",
6500 reg_type_str[type], val);
6504 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
6505 verbose(env, "%s pointer offset %d is not allowed\n",
6506 reg_type_str[type], reg->off);
6510 if (smin == S64_MIN) {
6511 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
6512 reg_type_str[type]);
6516 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
6517 verbose(env, "value %lld makes %s pointer be out of bounds\n",
6518 smin, reg_type_str[type]);
6525 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
6527 return &env->insn_aux_data[env->insn_idx];
6538 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
6539 u32 *alu_limit, bool mask_to_left)
6541 u32 max = 0, ptr_limit = 0;
6543 switch (ptr_reg->type) {
6545 /* Offset 0 is out-of-bounds, but acceptable start for the
6546 * left direction, see BPF_REG_FP. Also, unknown scalar
6547 * offset where we would need to deal with min/max bounds is
6548 * currently prohibited for unprivileged.
6550 max = MAX_BPF_STACK + mask_to_left;
6551 ptr_limit = -(ptr_reg->var_off.value + ptr_reg->off);
6553 case PTR_TO_MAP_VALUE:
6554 max = ptr_reg->map_ptr->value_size;
6555 ptr_limit = (mask_to_left ?
6556 ptr_reg->smin_value :
6557 ptr_reg->umax_value) + ptr_reg->off;
6563 if (ptr_limit >= max)
6564 return REASON_LIMIT;
6565 *alu_limit = ptr_limit;
6569 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
6570 const struct bpf_insn *insn)
6572 return env->bypass_spec_v1 || BPF_SRC(insn->code) == BPF_K;
6575 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
6576 u32 alu_state, u32 alu_limit)
6578 /* If we arrived here from different branches with different
6579 * state or limits to sanitize, then this won't work.
6581 if (aux->alu_state &&
6582 (aux->alu_state != alu_state ||
6583 aux->alu_limit != alu_limit))
6584 return REASON_PATHS;
6586 /* Corresponding fixup done in do_misc_fixups(). */
6587 aux->alu_state = alu_state;
6588 aux->alu_limit = alu_limit;
6592 static int sanitize_val_alu(struct bpf_verifier_env *env,
6593 struct bpf_insn *insn)
6595 struct bpf_insn_aux_data *aux = cur_aux(env);
6597 if (can_skip_alu_sanitation(env, insn))
6600 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
6603 static bool sanitize_needed(u8 opcode)
6605 return opcode == BPF_ADD || opcode == BPF_SUB;
6608 struct bpf_sanitize_info {
6609 struct bpf_insn_aux_data aux;
6613 static struct bpf_verifier_state *
6614 sanitize_speculative_path(struct bpf_verifier_env *env,
6615 const struct bpf_insn *insn,
6616 u32 next_idx, u32 curr_idx)
6618 struct bpf_verifier_state *branch;
6619 struct bpf_reg_state *regs;
6621 branch = push_stack(env, next_idx, curr_idx, true);
6622 if (branch && insn) {
6623 regs = branch->frame[branch->curframe]->regs;
6624 if (BPF_SRC(insn->code) == BPF_K) {
6625 mark_reg_unknown(env, regs, insn->dst_reg);
6626 } else if (BPF_SRC(insn->code) == BPF_X) {
6627 mark_reg_unknown(env, regs, insn->dst_reg);
6628 mark_reg_unknown(env, regs, insn->src_reg);
6634 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
6635 struct bpf_insn *insn,
6636 const struct bpf_reg_state *ptr_reg,
6637 const struct bpf_reg_state *off_reg,
6638 struct bpf_reg_state *dst_reg,
6639 struct bpf_sanitize_info *info,
6640 const bool commit_window)
6642 struct bpf_insn_aux_data *aux = commit_window ? cur_aux(env) : &info->aux;
6643 struct bpf_verifier_state *vstate = env->cur_state;
6644 bool off_is_imm = tnum_is_const(off_reg->var_off);
6645 bool off_is_neg = off_reg->smin_value < 0;
6646 bool ptr_is_dst_reg = ptr_reg == dst_reg;
6647 u8 opcode = BPF_OP(insn->code);
6648 u32 alu_state, alu_limit;
6649 struct bpf_reg_state tmp;
6653 if (can_skip_alu_sanitation(env, insn))
6656 /* We already marked aux for masking from non-speculative
6657 * paths, thus we got here in the first place. We only care
6658 * to explore bad access from here.
6660 if (vstate->speculative)
6663 if (!commit_window) {
6664 if (!tnum_is_const(off_reg->var_off) &&
6665 (off_reg->smin_value < 0) != (off_reg->smax_value < 0))
6666 return REASON_BOUNDS;
6668 info->mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
6669 (opcode == BPF_SUB && !off_is_neg);
6672 err = retrieve_ptr_limit(ptr_reg, &alu_limit, info->mask_to_left);
6676 if (commit_window) {
6677 /* In commit phase we narrow the masking window based on
6678 * the observed pointer move after the simulated operation.
6680 alu_state = info->aux.alu_state;
6681 alu_limit = abs(info->aux.alu_limit - alu_limit);
6683 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
6684 alu_state |= off_is_imm ? BPF_ALU_IMMEDIATE : 0;
6685 alu_state |= ptr_is_dst_reg ?
6686 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
6689 err = update_alu_sanitation_state(aux, alu_state, alu_limit);
6693 /* If we're in commit phase, we're done here given we already
6694 * pushed the truncated dst_reg into the speculative verification
6697 * Also, when register is a known constant, we rewrite register-based
6698 * operation to immediate-based, and thus do not need masking (and as
6699 * a consequence, do not need to simulate the zero-truncation either).
6701 if (commit_window || off_is_imm)
6704 /* Simulate and find potential out-of-bounds access under
6705 * speculative execution from truncation as a result of
6706 * masking when off was not within expected range. If off
6707 * sits in dst, then we temporarily need to move ptr there
6708 * to simulate dst (== 0) +/-= ptr. Needed, for example,
6709 * for cases where we use K-based arithmetic in one direction
6710 * and truncated reg-based in the other in order to explore
6713 if (!ptr_is_dst_reg) {
6715 *dst_reg = *ptr_reg;
6717 ret = sanitize_speculative_path(env, NULL, env->insn_idx + 1,
6719 if (!ptr_is_dst_reg && ret)
6721 return !ret ? REASON_STACK : 0;
6724 static void sanitize_mark_insn_seen(struct bpf_verifier_env *env)
6726 struct bpf_verifier_state *vstate = env->cur_state;
6728 /* If we simulate paths under speculation, we don't update the
6729 * insn as 'seen' such that when we verify unreachable paths in
6730 * the non-speculative domain, sanitize_dead_code() can still
6731 * rewrite/sanitize them.
6733 if (!vstate->speculative)
6734 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
6737 static int sanitize_err(struct bpf_verifier_env *env,
6738 const struct bpf_insn *insn, int reason,
6739 const struct bpf_reg_state *off_reg,
6740 const struct bpf_reg_state *dst_reg)
6742 static const char *err = "pointer arithmetic with it prohibited for !root";
6743 const char *op = BPF_OP(insn->code) == BPF_ADD ? "add" : "sub";
6744 u32 dst = insn->dst_reg, src = insn->src_reg;
6748 verbose(env, "R%d has unknown scalar with mixed signed bounds, %s\n",
6749 off_reg == dst_reg ? dst : src, err);
6752 verbose(env, "R%d has pointer with unsupported alu operation, %s\n",
6753 off_reg == dst_reg ? src : dst, err);
6756 verbose(env, "R%d tried to %s from different maps, paths or scalars, %s\n",
6760 verbose(env, "R%d tried to %s beyond pointer bounds, %s\n",
6764 verbose(env, "R%d could not be pushed for speculative verification, %s\n",
6768 verbose(env, "verifier internal error: unknown reason (%d)\n",
6776 /* check that stack access falls within stack limits and that 'reg' doesn't
6777 * have a variable offset.
6779 * Variable offset is prohibited for unprivileged mode for simplicity since it
6780 * requires corresponding support in Spectre masking for stack ALU. See also
6781 * retrieve_ptr_limit().
6784 * 'off' includes 'reg->off'.
6786 static int check_stack_access_for_ptr_arithmetic(
6787 struct bpf_verifier_env *env,
6789 const struct bpf_reg_state *reg,
6792 if (!tnum_is_const(reg->var_off)) {
6795 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6796 verbose(env, "R%d variable stack access prohibited for !root, var_off=%s off=%d\n",
6797 regno, tn_buf, off);
6801 if (off >= 0 || off < -MAX_BPF_STACK) {
6802 verbose(env, "R%d stack pointer arithmetic goes out of range, "
6803 "prohibited for !root; off=%d\n", regno, off);
6810 static int sanitize_check_bounds(struct bpf_verifier_env *env,
6811 const struct bpf_insn *insn,
6812 const struct bpf_reg_state *dst_reg)
6814 u32 dst = insn->dst_reg;
6816 /* For unprivileged we require that resulting offset must be in bounds
6817 * in order to be able to sanitize access later on.
6819 if (env->bypass_spec_v1)
6822 switch (dst_reg->type) {
6824 if (check_stack_access_for_ptr_arithmetic(env, dst, dst_reg,
6825 dst_reg->off + dst_reg->var_off.value))
6828 case PTR_TO_MAP_VALUE:
6829 if (check_map_access(env, dst, dst_reg->off, 1, false)) {
6830 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
6831 "prohibited for !root\n", dst);
6842 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
6843 * Caller should also handle BPF_MOV case separately.
6844 * If we return -EACCES, caller may want to try again treating pointer as a
6845 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
6847 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
6848 struct bpf_insn *insn,
6849 const struct bpf_reg_state *ptr_reg,
6850 const struct bpf_reg_state *off_reg)
6852 struct bpf_verifier_state *vstate = env->cur_state;
6853 struct bpf_func_state *state = vstate->frame[vstate->curframe];
6854 struct bpf_reg_state *regs = state->regs, *dst_reg;
6855 bool known = tnum_is_const(off_reg->var_off);
6856 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
6857 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
6858 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
6859 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
6860 struct bpf_sanitize_info info = {};
6861 u8 opcode = BPF_OP(insn->code);
6862 u32 dst = insn->dst_reg;
6865 dst_reg = ®s[dst];
6867 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
6868 smin_val > smax_val || umin_val > umax_val) {
6869 /* Taint dst register if offset had invalid bounds derived from
6870 * e.g. dead branches.
6872 __mark_reg_unknown(env, dst_reg);
6876 if (BPF_CLASS(insn->code) != BPF_ALU64) {
6877 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
6878 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
6879 __mark_reg_unknown(env, dst_reg);
6884 "R%d 32-bit pointer arithmetic prohibited\n",
6889 switch (ptr_reg->type) {
6890 case PTR_TO_MAP_VALUE_OR_NULL:
6891 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
6892 dst, reg_type_str[ptr_reg->type]);
6894 case CONST_PTR_TO_MAP:
6895 /* smin_val represents the known value */
6896 if (known && smin_val == 0 && opcode == BPF_ADD)
6899 case PTR_TO_PACKET_END:
6901 case PTR_TO_SOCKET_OR_NULL:
6902 case PTR_TO_SOCK_COMMON:
6903 case PTR_TO_SOCK_COMMON_OR_NULL:
6904 case PTR_TO_TCP_SOCK:
6905 case PTR_TO_TCP_SOCK_OR_NULL:
6906 case PTR_TO_XDP_SOCK:
6907 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
6908 dst, reg_type_str[ptr_reg->type]);
6914 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
6915 * The id may be overwritten later if we create a new variable offset.
6917 dst_reg->type = ptr_reg->type;
6918 dst_reg->id = ptr_reg->id;
6920 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
6921 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
6924 /* pointer types do not carry 32-bit bounds at the moment. */
6925 __mark_reg32_unbounded(dst_reg);
6927 if (sanitize_needed(opcode)) {
6928 ret = sanitize_ptr_alu(env, insn, ptr_reg, off_reg, dst_reg,
6931 return sanitize_err(env, insn, ret, off_reg, dst_reg);
6936 /* We can take a fixed offset as long as it doesn't overflow
6937 * the s32 'off' field
6939 if (known && (ptr_reg->off + smin_val ==
6940 (s64)(s32)(ptr_reg->off + smin_val))) {
6941 /* pointer += K. Accumulate it into fixed offset */
6942 dst_reg->smin_value = smin_ptr;
6943 dst_reg->smax_value = smax_ptr;
6944 dst_reg->umin_value = umin_ptr;
6945 dst_reg->umax_value = umax_ptr;
6946 dst_reg->var_off = ptr_reg->var_off;
6947 dst_reg->off = ptr_reg->off + smin_val;
6948 dst_reg->raw = ptr_reg->raw;
6951 /* A new variable offset is created. Note that off_reg->off
6952 * == 0, since it's a scalar.
6953 * dst_reg gets the pointer type and since some positive
6954 * integer value was added to the pointer, give it a new 'id'
6955 * if it's a PTR_TO_PACKET.
6956 * this creates a new 'base' pointer, off_reg (variable) gets
6957 * added into the variable offset, and we copy the fixed offset
6960 if (signed_add_overflows(smin_ptr, smin_val) ||
6961 signed_add_overflows(smax_ptr, smax_val)) {
6962 dst_reg->smin_value = S64_MIN;
6963 dst_reg->smax_value = S64_MAX;
6965 dst_reg->smin_value = smin_ptr + smin_val;
6966 dst_reg->smax_value = smax_ptr + smax_val;
6968 if (umin_ptr + umin_val < umin_ptr ||
6969 umax_ptr + umax_val < umax_ptr) {
6970 dst_reg->umin_value = 0;
6971 dst_reg->umax_value = U64_MAX;
6973 dst_reg->umin_value = umin_ptr + umin_val;
6974 dst_reg->umax_value = umax_ptr + umax_val;
6976 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
6977 dst_reg->off = ptr_reg->off;
6978 dst_reg->raw = ptr_reg->raw;
6979 if (reg_is_pkt_pointer(ptr_reg)) {
6980 dst_reg->id = ++env->id_gen;
6981 /* something was added to pkt_ptr, set range to zero */
6982 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
6986 if (dst_reg == off_reg) {
6987 /* scalar -= pointer. Creates an unknown scalar */
6988 verbose(env, "R%d tried to subtract pointer from scalar\n",
6992 /* We don't allow subtraction from FP, because (according to
6993 * test_verifier.c test "invalid fp arithmetic", JITs might not
6994 * be able to deal with it.
6996 if (ptr_reg->type == PTR_TO_STACK) {
6997 verbose(env, "R%d subtraction from stack pointer prohibited\n",
7001 if (known && (ptr_reg->off - smin_val ==
7002 (s64)(s32)(ptr_reg->off - smin_val))) {
7003 /* pointer -= K. Subtract it from fixed offset */
7004 dst_reg->smin_value = smin_ptr;
7005 dst_reg->smax_value = smax_ptr;
7006 dst_reg->umin_value = umin_ptr;
7007 dst_reg->umax_value = umax_ptr;
7008 dst_reg->var_off = ptr_reg->var_off;
7009 dst_reg->id = ptr_reg->id;
7010 dst_reg->off = ptr_reg->off - smin_val;
7011 dst_reg->raw = ptr_reg->raw;
7014 /* A new variable offset is created. If the subtrahend is known
7015 * nonnegative, then any reg->range we had before is still good.
7017 if (signed_sub_overflows(smin_ptr, smax_val) ||
7018 signed_sub_overflows(smax_ptr, smin_val)) {
7019 /* Overflow possible, we know nothing */
7020 dst_reg->smin_value = S64_MIN;
7021 dst_reg->smax_value = S64_MAX;
7023 dst_reg->smin_value = smin_ptr - smax_val;
7024 dst_reg->smax_value = smax_ptr - smin_val;
7026 if (umin_ptr < umax_val) {
7027 /* Overflow possible, we know nothing */
7028 dst_reg->umin_value = 0;
7029 dst_reg->umax_value = U64_MAX;
7031 /* Cannot overflow (as long as bounds are consistent) */
7032 dst_reg->umin_value = umin_ptr - umax_val;
7033 dst_reg->umax_value = umax_ptr - umin_val;
7035 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
7036 dst_reg->off = ptr_reg->off;
7037 dst_reg->raw = ptr_reg->raw;
7038 if (reg_is_pkt_pointer(ptr_reg)) {
7039 dst_reg->id = ++env->id_gen;
7040 /* something was added to pkt_ptr, set range to zero */
7042 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
7048 /* bitwise ops on pointers are troublesome, prohibit. */
7049 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
7050 dst, bpf_alu_string[opcode >> 4]);
7053 /* other operators (e.g. MUL,LSH) produce non-pointer results */
7054 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
7055 dst, bpf_alu_string[opcode >> 4]);
7059 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
7062 __update_reg_bounds(dst_reg);
7063 __reg_deduce_bounds(dst_reg);
7064 __reg_bound_offset(dst_reg);
7066 if (sanitize_check_bounds(env, insn, dst_reg) < 0)
7068 if (sanitize_needed(opcode)) {
7069 ret = sanitize_ptr_alu(env, insn, dst_reg, off_reg, dst_reg,
7072 return sanitize_err(env, insn, ret, off_reg, dst_reg);
7078 static void scalar32_min_max_add(struct bpf_reg_state *dst_reg,
7079 struct bpf_reg_state *src_reg)
7081 s32 smin_val = src_reg->s32_min_value;
7082 s32 smax_val = src_reg->s32_max_value;
7083 u32 umin_val = src_reg->u32_min_value;
7084 u32 umax_val = src_reg->u32_max_value;
7086 if (signed_add32_overflows(dst_reg->s32_min_value, smin_val) ||
7087 signed_add32_overflows(dst_reg->s32_max_value, smax_val)) {
7088 dst_reg->s32_min_value = S32_MIN;
7089 dst_reg->s32_max_value = S32_MAX;
7091 dst_reg->s32_min_value += smin_val;
7092 dst_reg->s32_max_value += smax_val;
7094 if (dst_reg->u32_min_value + umin_val < umin_val ||
7095 dst_reg->u32_max_value + umax_val < umax_val) {
7096 dst_reg->u32_min_value = 0;
7097 dst_reg->u32_max_value = U32_MAX;
7099 dst_reg->u32_min_value += umin_val;
7100 dst_reg->u32_max_value += umax_val;
7104 static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
7105 struct bpf_reg_state *src_reg)
7107 s64 smin_val = src_reg->smin_value;
7108 s64 smax_val = src_reg->smax_value;
7109 u64 umin_val = src_reg->umin_value;
7110 u64 umax_val = src_reg->umax_value;
7112 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
7113 signed_add_overflows(dst_reg->smax_value, smax_val)) {
7114 dst_reg->smin_value = S64_MIN;
7115 dst_reg->smax_value = S64_MAX;
7117 dst_reg->smin_value += smin_val;
7118 dst_reg->smax_value += smax_val;
7120 if (dst_reg->umin_value + umin_val < umin_val ||
7121 dst_reg->umax_value + umax_val < umax_val) {
7122 dst_reg->umin_value = 0;
7123 dst_reg->umax_value = U64_MAX;
7125 dst_reg->umin_value += umin_val;
7126 dst_reg->umax_value += umax_val;
7130 static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg,
7131 struct bpf_reg_state *src_reg)
7133 s32 smin_val = src_reg->s32_min_value;
7134 s32 smax_val = src_reg->s32_max_value;
7135 u32 umin_val = src_reg->u32_min_value;
7136 u32 umax_val = src_reg->u32_max_value;
7138 if (signed_sub32_overflows(dst_reg->s32_min_value, smax_val) ||
7139 signed_sub32_overflows(dst_reg->s32_max_value, smin_val)) {
7140 /* Overflow possible, we know nothing */
7141 dst_reg->s32_min_value = S32_MIN;
7142 dst_reg->s32_max_value = S32_MAX;
7144 dst_reg->s32_min_value -= smax_val;
7145 dst_reg->s32_max_value -= smin_val;
7147 if (dst_reg->u32_min_value < umax_val) {
7148 /* Overflow possible, we know nothing */
7149 dst_reg->u32_min_value = 0;
7150 dst_reg->u32_max_value = U32_MAX;
7152 /* Cannot overflow (as long as bounds are consistent) */
7153 dst_reg->u32_min_value -= umax_val;
7154 dst_reg->u32_max_value -= umin_val;
7158 static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
7159 struct bpf_reg_state *src_reg)
7161 s64 smin_val = src_reg->smin_value;
7162 s64 smax_val = src_reg->smax_value;
7163 u64 umin_val = src_reg->umin_value;
7164 u64 umax_val = src_reg->umax_value;
7166 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
7167 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
7168 /* Overflow possible, we know nothing */
7169 dst_reg->smin_value = S64_MIN;
7170 dst_reg->smax_value = S64_MAX;
7172 dst_reg->smin_value -= smax_val;
7173 dst_reg->smax_value -= smin_val;
7175 if (dst_reg->umin_value < umax_val) {
7176 /* Overflow possible, we know nothing */
7177 dst_reg->umin_value = 0;
7178 dst_reg->umax_value = U64_MAX;
7180 /* Cannot overflow (as long as bounds are consistent) */
7181 dst_reg->umin_value -= umax_val;
7182 dst_reg->umax_value -= umin_val;
7186 static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg,
7187 struct bpf_reg_state *src_reg)
7189 s32 smin_val = src_reg->s32_min_value;
7190 u32 umin_val = src_reg->u32_min_value;
7191 u32 umax_val = src_reg->u32_max_value;
7193 if (smin_val < 0 || dst_reg->s32_min_value < 0) {
7194 /* Ain't nobody got time to multiply that sign */
7195 __mark_reg32_unbounded(dst_reg);
7198 /* Both values are positive, so we can work with unsigned and
7199 * copy the result to signed (unless it exceeds S32_MAX).
7201 if (umax_val > U16_MAX || dst_reg->u32_max_value > U16_MAX) {
7202 /* Potential overflow, we know nothing */
7203 __mark_reg32_unbounded(dst_reg);
7206 dst_reg->u32_min_value *= umin_val;
7207 dst_reg->u32_max_value *= umax_val;
7208 if (dst_reg->u32_max_value > S32_MAX) {
7209 /* Overflow possible, we know nothing */
7210 dst_reg->s32_min_value = S32_MIN;
7211 dst_reg->s32_max_value = S32_MAX;
7213 dst_reg->s32_min_value = dst_reg->u32_min_value;
7214 dst_reg->s32_max_value = dst_reg->u32_max_value;
7218 static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
7219 struct bpf_reg_state *src_reg)
7221 s64 smin_val = src_reg->smin_value;
7222 u64 umin_val = src_reg->umin_value;
7223 u64 umax_val = src_reg->umax_value;
7225 if (smin_val < 0 || dst_reg->smin_value < 0) {
7226 /* Ain't nobody got time to multiply that sign */
7227 __mark_reg64_unbounded(dst_reg);
7230 /* Both values are positive, so we can work with unsigned and
7231 * copy the result to signed (unless it exceeds S64_MAX).
7233 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
7234 /* Potential overflow, we know nothing */
7235 __mark_reg64_unbounded(dst_reg);
7238 dst_reg->umin_value *= umin_val;
7239 dst_reg->umax_value *= umax_val;
7240 if (dst_reg->umax_value > S64_MAX) {
7241 /* Overflow possible, we know nothing */
7242 dst_reg->smin_value = S64_MIN;
7243 dst_reg->smax_value = S64_MAX;
7245 dst_reg->smin_value = dst_reg->umin_value;
7246 dst_reg->smax_value = dst_reg->umax_value;
7250 static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
7251 struct bpf_reg_state *src_reg)
7253 bool src_known = tnum_subreg_is_const(src_reg->var_off);
7254 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
7255 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
7256 s32 smin_val = src_reg->s32_min_value;
7257 u32 umax_val = src_reg->u32_max_value;
7259 if (src_known && dst_known) {
7260 __mark_reg32_known(dst_reg, var32_off.value);
7264 /* We get our minimum from the var_off, since that's inherently
7265 * bitwise. Our maximum is the minimum of the operands' maxima.
7267 dst_reg->u32_min_value = var32_off.value;
7268 dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val);
7269 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
7270 /* Lose signed bounds when ANDing negative numbers,
7271 * ain't nobody got time for that.
7273 dst_reg->s32_min_value = S32_MIN;
7274 dst_reg->s32_max_value = S32_MAX;
7276 /* ANDing two positives gives a positive, so safe to
7277 * cast result into s64.
7279 dst_reg->s32_min_value = dst_reg->u32_min_value;
7280 dst_reg->s32_max_value = dst_reg->u32_max_value;
7284 static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
7285 struct bpf_reg_state *src_reg)
7287 bool src_known = tnum_is_const(src_reg->var_off);
7288 bool dst_known = tnum_is_const(dst_reg->var_off);
7289 s64 smin_val = src_reg->smin_value;
7290 u64 umax_val = src_reg->umax_value;
7292 if (src_known && dst_known) {
7293 __mark_reg_known(dst_reg, dst_reg->var_off.value);
7297 /* We get our minimum from the var_off, since that's inherently
7298 * bitwise. Our maximum is the minimum of the operands' maxima.
7300 dst_reg->umin_value = dst_reg->var_off.value;
7301 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
7302 if (dst_reg->smin_value < 0 || smin_val < 0) {
7303 /* Lose signed bounds when ANDing negative numbers,
7304 * ain't nobody got time for that.
7306 dst_reg->smin_value = S64_MIN;
7307 dst_reg->smax_value = S64_MAX;
7309 /* ANDing two positives gives a positive, so safe to
7310 * cast result into s64.
7312 dst_reg->smin_value = dst_reg->umin_value;
7313 dst_reg->smax_value = dst_reg->umax_value;
7315 /* We may learn something more from the var_off */
7316 __update_reg_bounds(dst_reg);
7319 static void scalar32_min_max_or(struct bpf_reg_state *dst_reg,
7320 struct bpf_reg_state *src_reg)
7322 bool src_known = tnum_subreg_is_const(src_reg->var_off);
7323 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
7324 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
7325 s32 smin_val = src_reg->s32_min_value;
7326 u32 umin_val = src_reg->u32_min_value;
7328 if (src_known && dst_known) {
7329 __mark_reg32_known(dst_reg, var32_off.value);
7333 /* We get our maximum from the var_off, and our minimum is the
7334 * maximum of the operands' minima
7336 dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val);
7337 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
7338 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
7339 /* Lose signed bounds when ORing negative numbers,
7340 * ain't nobody got time for that.
7342 dst_reg->s32_min_value = S32_MIN;
7343 dst_reg->s32_max_value = S32_MAX;
7345 /* ORing two positives gives a positive, so safe to
7346 * cast result into s64.
7348 dst_reg->s32_min_value = dst_reg->u32_min_value;
7349 dst_reg->s32_max_value = dst_reg->u32_max_value;
7353 static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
7354 struct bpf_reg_state *src_reg)
7356 bool src_known = tnum_is_const(src_reg->var_off);
7357 bool dst_known = tnum_is_const(dst_reg->var_off);
7358 s64 smin_val = src_reg->smin_value;
7359 u64 umin_val = src_reg->umin_value;
7361 if (src_known && dst_known) {
7362 __mark_reg_known(dst_reg, dst_reg->var_off.value);
7366 /* We get our maximum from the var_off, and our minimum is the
7367 * maximum of the operands' minima
7369 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
7370 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
7371 if (dst_reg->smin_value < 0 || smin_val < 0) {
7372 /* Lose signed bounds when ORing negative numbers,
7373 * ain't nobody got time for that.
7375 dst_reg->smin_value = S64_MIN;
7376 dst_reg->smax_value = S64_MAX;
7378 /* ORing two positives gives a positive, so safe to
7379 * cast result into s64.
7381 dst_reg->smin_value = dst_reg->umin_value;
7382 dst_reg->smax_value = dst_reg->umax_value;
7384 /* We may learn something more from the var_off */
7385 __update_reg_bounds(dst_reg);
7388 static void scalar32_min_max_xor(struct bpf_reg_state *dst_reg,
7389 struct bpf_reg_state *src_reg)
7391 bool src_known = tnum_subreg_is_const(src_reg->var_off);
7392 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
7393 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
7394 s32 smin_val = src_reg->s32_min_value;
7396 if (src_known && dst_known) {
7397 __mark_reg32_known(dst_reg, var32_off.value);
7401 /* We get both minimum and maximum from the var32_off. */
7402 dst_reg->u32_min_value = var32_off.value;
7403 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
7405 if (dst_reg->s32_min_value >= 0 && smin_val >= 0) {
7406 /* XORing two positive sign numbers gives a positive,
7407 * so safe to cast u32 result into s32.
7409 dst_reg->s32_min_value = dst_reg->u32_min_value;
7410 dst_reg->s32_max_value = dst_reg->u32_max_value;
7412 dst_reg->s32_min_value = S32_MIN;
7413 dst_reg->s32_max_value = S32_MAX;
7417 static void scalar_min_max_xor(struct bpf_reg_state *dst_reg,
7418 struct bpf_reg_state *src_reg)
7420 bool src_known = tnum_is_const(src_reg->var_off);
7421 bool dst_known = tnum_is_const(dst_reg->var_off);
7422 s64 smin_val = src_reg->smin_value;
7424 if (src_known && dst_known) {
7425 /* dst_reg->var_off.value has been updated earlier */
7426 __mark_reg_known(dst_reg, dst_reg->var_off.value);
7430 /* We get both minimum and maximum from the var_off. */
7431 dst_reg->umin_value = dst_reg->var_off.value;
7432 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
7434 if (dst_reg->smin_value >= 0 && smin_val >= 0) {
7435 /* XORing two positive sign numbers gives a positive,
7436 * so safe to cast u64 result into s64.
7438 dst_reg->smin_value = dst_reg->umin_value;
7439 dst_reg->smax_value = dst_reg->umax_value;
7441 dst_reg->smin_value = S64_MIN;
7442 dst_reg->smax_value = S64_MAX;
7445 __update_reg_bounds(dst_reg);
7448 static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
7449 u64 umin_val, u64 umax_val)
7451 /* We lose all sign bit information (except what we can pick
7454 dst_reg->s32_min_value = S32_MIN;
7455 dst_reg->s32_max_value = S32_MAX;
7456 /* If we might shift our top bit out, then we know nothing */
7457 if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) {
7458 dst_reg->u32_min_value = 0;
7459 dst_reg->u32_max_value = U32_MAX;
7461 dst_reg->u32_min_value <<= umin_val;
7462 dst_reg->u32_max_value <<= umax_val;
7466 static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
7467 struct bpf_reg_state *src_reg)
7469 u32 umax_val = src_reg->u32_max_value;
7470 u32 umin_val = src_reg->u32_min_value;
7471 /* u32 alu operation will zext upper bits */
7472 struct tnum subreg = tnum_subreg(dst_reg->var_off);
7474 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
7475 dst_reg->var_off = tnum_subreg(tnum_lshift(subreg, umin_val));
7476 /* Not required but being careful mark reg64 bounds as unknown so
7477 * that we are forced to pick them up from tnum and zext later and
7478 * if some path skips this step we are still safe.
7480 __mark_reg64_unbounded(dst_reg);
7481 __update_reg32_bounds(dst_reg);
7484 static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg,
7485 u64 umin_val, u64 umax_val)
7487 /* Special case <<32 because it is a common compiler pattern to sign
7488 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
7489 * positive we know this shift will also be positive so we can track
7490 * bounds correctly. Otherwise we lose all sign bit information except
7491 * what we can pick up from var_off. Perhaps we can generalize this
7492 * later to shifts of any length.
7494 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0)
7495 dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32;
7497 dst_reg->smax_value = S64_MAX;
7499 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0)
7500 dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32;
7502 dst_reg->smin_value = S64_MIN;
7504 /* If we might shift our top bit out, then we know nothing */
7505 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
7506 dst_reg->umin_value = 0;
7507 dst_reg->umax_value = U64_MAX;
7509 dst_reg->umin_value <<= umin_val;
7510 dst_reg->umax_value <<= umax_val;
7514 static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
7515 struct bpf_reg_state *src_reg)
7517 u64 umax_val = src_reg->umax_value;
7518 u64 umin_val = src_reg->umin_value;
7520 /* scalar64 calc uses 32bit unshifted bounds so must be called first */
7521 __scalar64_min_max_lsh(dst_reg, umin_val, umax_val);
7522 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
7524 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
7525 /* We may learn something more from the var_off */
7526 __update_reg_bounds(dst_reg);
7529 static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg,
7530 struct bpf_reg_state *src_reg)
7532 struct tnum subreg = tnum_subreg(dst_reg->var_off);
7533 u32 umax_val = src_reg->u32_max_value;
7534 u32 umin_val = src_reg->u32_min_value;
7536 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
7537 * be negative, then either:
7538 * 1) src_reg might be zero, so the sign bit of the result is
7539 * unknown, so we lose our signed bounds
7540 * 2) it's known negative, thus the unsigned bounds capture the
7542 * 3) the signed bounds cross zero, so they tell us nothing
7544 * If the value in dst_reg is known nonnegative, then again the
7545 * unsigned bounds capture the signed bounds.
7546 * Thus, in all cases it suffices to blow away our signed bounds
7547 * and rely on inferring new ones from the unsigned bounds and
7548 * var_off of the result.
7550 dst_reg->s32_min_value = S32_MIN;
7551 dst_reg->s32_max_value = S32_MAX;
7553 dst_reg->var_off = tnum_rshift(subreg, umin_val);
7554 dst_reg->u32_min_value >>= umax_val;
7555 dst_reg->u32_max_value >>= umin_val;
7557 __mark_reg64_unbounded(dst_reg);
7558 __update_reg32_bounds(dst_reg);
7561 static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
7562 struct bpf_reg_state *src_reg)
7564 u64 umax_val = src_reg->umax_value;
7565 u64 umin_val = src_reg->umin_value;
7567 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
7568 * be negative, then either:
7569 * 1) src_reg might be zero, so the sign bit of the result is
7570 * unknown, so we lose our signed bounds
7571 * 2) it's known negative, thus the unsigned bounds capture the
7573 * 3) the signed bounds cross zero, so they tell us nothing
7575 * If the value in dst_reg is known nonnegative, then again the
7576 * unsigned bounds capture the signed bounds.
7577 * Thus, in all cases it suffices to blow away our signed bounds
7578 * and rely on inferring new ones from the unsigned bounds and
7579 * var_off of the result.
7581 dst_reg->smin_value = S64_MIN;
7582 dst_reg->smax_value = S64_MAX;
7583 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
7584 dst_reg->umin_value >>= umax_val;
7585 dst_reg->umax_value >>= umin_val;
7587 /* Its not easy to operate on alu32 bounds here because it depends
7588 * on bits being shifted in. Take easy way out and mark unbounded
7589 * so we can recalculate later from tnum.
7591 __mark_reg32_unbounded(dst_reg);
7592 __update_reg_bounds(dst_reg);
7595 static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg,
7596 struct bpf_reg_state *src_reg)
7598 u64 umin_val = src_reg->u32_min_value;
7600 /* Upon reaching here, src_known is true and
7601 * umax_val is equal to umin_val.
7603 dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val);
7604 dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val);
7606 dst_reg->var_off = tnum_arshift(tnum_subreg(dst_reg->var_off), umin_val, 32);
7608 /* blow away the dst_reg umin_value/umax_value and rely on
7609 * dst_reg var_off to refine the result.
7611 dst_reg->u32_min_value = 0;
7612 dst_reg->u32_max_value = U32_MAX;
7614 __mark_reg64_unbounded(dst_reg);
7615 __update_reg32_bounds(dst_reg);
7618 static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
7619 struct bpf_reg_state *src_reg)
7621 u64 umin_val = src_reg->umin_value;
7623 /* Upon reaching here, src_known is true and umax_val is equal
7626 dst_reg->smin_value >>= umin_val;
7627 dst_reg->smax_value >>= umin_val;
7629 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val, 64);
7631 /* blow away the dst_reg umin_value/umax_value and rely on
7632 * dst_reg var_off to refine the result.
7634 dst_reg->umin_value = 0;
7635 dst_reg->umax_value = U64_MAX;
7637 /* Its not easy to operate on alu32 bounds here because it depends
7638 * on bits being shifted in from upper 32-bits. Take easy way out
7639 * and mark unbounded so we can recalculate later from tnum.
7641 __mark_reg32_unbounded(dst_reg);
7642 __update_reg_bounds(dst_reg);
7645 /* WARNING: This function does calculations on 64-bit values, but the actual
7646 * execution may occur on 32-bit values. Therefore, things like bitshifts
7647 * need extra checks in the 32-bit case.
7649 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
7650 struct bpf_insn *insn,
7651 struct bpf_reg_state *dst_reg,
7652 struct bpf_reg_state src_reg)
7654 struct bpf_reg_state *regs = cur_regs(env);
7655 u8 opcode = BPF_OP(insn->code);
7657 s64 smin_val, smax_val;
7658 u64 umin_val, umax_val;
7659 s32 s32_min_val, s32_max_val;
7660 u32 u32_min_val, u32_max_val;
7661 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
7662 bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
7665 smin_val = src_reg.smin_value;
7666 smax_val = src_reg.smax_value;
7667 umin_val = src_reg.umin_value;
7668 umax_val = src_reg.umax_value;
7670 s32_min_val = src_reg.s32_min_value;
7671 s32_max_val = src_reg.s32_max_value;
7672 u32_min_val = src_reg.u32_min_value;
7673 u32_max_val = src_reg.u32_max_value;
7676 src_known = tnum_subreg_is_const(src_reg.var_off);
7678 (s32_min_val != s32_max_val || u32_min_val != u32_max_val)) ||
7679 s32_min_val > s32_max_val || u32_min_val > u32_max_val) {
7680 /* Taint dst register if offset had invalid bounds
7681 * derived from e.g. dead branches.
7683 __mark_reg_unknown(env, dst_reg);
7687 src_known = tnum_is_const(src_reg.var_off);
7689 (smin_val != smax_val || umin_val != umax_val)) ||
7690 smin_val > smax_val || umin_val > umax_val) {
7691 /* Taint dst register if offset had invalid bounds
7692 * derived from e.g. dead branches.
7694 __mark_reg_unknown(env, dst_reg);
7700 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
7701 __mark_reg_unknown(env, dst_reg);
7705 if (sanitize_needed(opcode)) {
7706 ret = sanitize_val_alu(env, insn);
7708 return sanitize_err(env, insn, ret, NULL, NULL);
7711 /* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
7712 * There are two classes of instructions: The first class we track both
7713 * alu32 and alu64 sign/unsigned bounds independently this provides the
7714 * greatest amount of precision when alu operations are mixed with jmp32
7715 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
7716 * and BPF_OR. This is possible because these ops have fairly easy to
7717 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
7718 * See alu32 verifier tests for examples. The second class of
7719 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
7720 * with regards to tracking sign/unsigned bounds because the bits may
7721 * cross subreg boundaries in the alu64 case. When this happens we mark
7722 * the reg unbounded in the subreg bound space and use the resulting
7723 * tnum to calculate an approximation of the sign/unsigned bounds.
7727 scalar32_min_max_add(dst_reg, &src_reg);
7728 scalar_min_max_add(dst_reg, &src_reg);
7729 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
7732 scalar32_min_max_sub(dst_reg, &src_reg);
7733 scalar_min_max_sub(dst_reg, &src_reg);
7734 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
7737 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
7738 scalar32_min_max_mul(dst_reg, &src_reg);
7739 scalar_min_max_mul(dst_reg, &src_reg);
7742 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
7743 scalar32_min_max_and(dst_reg, &src_reg);
7744 scalar_min_max_and(dst_reg, &src_reg);
7747 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
7748 scalar32_min_max_or(dst_reg, &src_reg);
7749 scalar_min_max_or(dst_reg, &src_reg);
7752 dst_reg->var_off = tnum_xor(dst_reg->var_off, src_reg.var_off);
7753 scalar32_min_max_xor(dst_reg, &src_reg);
7754 scalar_min_max_xor(dst_reg, &src_reg);
7757 if (umax_val >= insn_bitness) {
7758 /* Shifts greater than 31 or 63 are undefined.
7759 * This includes shifts by a negative number.
7761 mark_reg_unknown(env, regs, insn->dst_reg);
7765 scalar32_min_max_lsh(dst_reg, &src_reg);
7767 scalar_min_max_lsh(dst_reg, &src_reg);
7770 if (umax_val >= insn_bitness) {
7771 /* Shifts greater than 31 or 63 are undefined.
7772 * This includes shifts by a negative number.
7774 mark_reg_unknown(env, regs, insn->dst_reg);
7778 scalar32_min_max_rsh(dst_reg, &src_reg);
7780 scalar_min_max_rsh(dst_reg, &src_reg);
7783 if (umax_val >= insn_bitness) {
7784 /* Shifts greater than 31 or 63 are undefined.
7785 * This includes shifts by a negative number.
7787 mark_reg_unknown(env, regs, insn->dst_reg);
7791 scalar32_min_max_arsh(dst_reg, &src_reg);
7793 scalar_min_max_arsh(dst_reg, &src_reg);
7796 mark_reg_unknown(env, regs, insn->dst_reg);
7800 /* ALU32 ops are zero extended into 64bit register */
7802 zext_32_to_64(dst_reg);
7804 __update_reg_bounds(dst_reg);
7805 __reg_deduce_bounds(dst_reg);
7806 __reg_bound_offset(dst_reg);
7810 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
7813 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
7814 struct bpf_insn *insn)
7816 struct bpf_verifier_state *vstate = env->cur_state;
7817 struct bpf_func_state *state = vstate->frame[vstate->curframe];
7818 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
7819 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
7820 u8 opcode = BPF_OP(insn->code);
7823 dst_reg = ®s[insn->dst_reg];
7825 if (dst_reg->type != SCALAR_VALUE)
7828 /* Make sure ID is cleared otherwise dst_reg min/max could be
7829 * incorrectly propagated into other registers by find_equal_scalars()
7832 if (BPF_SRC(insn->code) == BPF_X) {
7833 src_reg = ®s[insn->src_reg];
7834 if (src_reg->type != SCALAR_VALUE) {
7835 if (dst_reg->type != SCALAR_VALUE) {
7836 /* Combining two pointers by any ALU op yields
7837 * an arbitrary scalar. Disallow all math except
7838 * pointer subtraction
7840 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
7841 mark_reg_unknown(env, regs, insn->dst_reg);
7844 verbose(env, "R%d pointer %s pointer prohibited\n",
7846 bpf_alu_string[opcode >> 4]);
7849 /* scalar += pointer
7850 * This is legal, but we have to reverse our
7851 * src/dest handling in computing the range
7853 err = mark_chain_precision(env, insn->dst_reg);
7856 return adjust_ptr_min_max_vals(env, insn,
7859 } else if (ptr_reg) {
7860 /* pointer += scalar */
7861 err = mark_chain_precision(env, insn->src_reg);
7864 return adjust_ptr_min_max_vals(env, insn,
7868 /* Pretend the src is a reg with a known value, since we only
7869 * need to be able to read from this state.
7871 off_reg.type = SCALAR_VALUE;
7872 __mark_reg_known(&off_reg, insn->imm);
7874 if (ptr_reg) /* pointer += K */
7875 return adjust_ptr_min_max_vals(env, insn,
7879 /* Got here implies adding two SCALAR_VALUEs */
7880 if (WARN_ON_ONCE(ptr_reg)) {
7881 print_verifier_state(env, state);
7882 verbose(env, "verifier internal error: unexpected ptr_reg\n");
7885 if (WARN_ON(!src_reg)) {
7886 print_verifier_state(env, state);
7887 verbose(env, "verifier internal error: no src_reg\n");
7890 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
7893 /* check validity of 32-bit and 64-bit arithmetic operations */
7894 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
7896 struct bpf_reg_state *regs = cur_regs(env);
7897 u8 opcode = BPF_OP(insn->code);
7900 if (opcode == BPF_END || opcode == BPF_NEG) {
7901 if (opcode == BPF_NEG) {
7902 if (BPF_SRC(insn->code) != 0 ||
7903 insn->src_reg != BPF_REG_0 ||
7904 insn->off != 0 || insn->imm != 0) {
7905 verbose(env, "BPF_NEG uses reserved fields\n");
7909 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
7910 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
7911 BPF_CLASS(insn->code) == BPF_ALU64) {
7912 verbose(env, "BPF_END uses reserved fields\n");
7917 /* check src operand */
7918 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
7922 if (is_pointer_value(env, insn->dst_reg)) {
7923 verbose(env, "R%d pointer arithmetic prohibited\n",
7928 /* check dest operand */
7929 err = check_reg_arg(env, insn->dst_reg, DST_OP);
7933 } else if (opcode == BPF_MOV) {
7935 if (BPF_SRC(insn->code) == BPF_X) {
7936 if (insn->imm != 0 || insn->off != 0) {
7937 verbose(env, "BPF_MOV uses reserved fields\n");
7941 /* check src operand */
7942 err = check_reg_arg(env, insn->src_reg, SRC_OP);
7946 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
7947 verbose(env, "BPF_MOV uses reserved fields\n");
7952 /* check dest operand, mark as required later */
7953 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
7957 if (BPF_SRC(insn->code) == BPF_X) {
7958 struct bpf_reg_state *src_reg = regs + insn->src_reg;
7959 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
7961 if (BPF_CLASS(insn->code) == BPF_ALU64) {
7963 * copy register state to dest reg
7965 if (src_reg->type == SCALAR_VALUE && !src_reg->id)
7966 /* Assign src and dst registers the same ID
7967 * that will be used by find_equal_scalars()
7968 * to propagate min/max range.
7970 src_reg->id = ++env->id_gen;
7971 *dst_reg = *src_reg;
7972 dst_reg->live |= REG_LIVE_WRITTEN;
7973 dst_reg->subreg_def = DEF_NOT_SUBREG;
7976 if (is_pointer_value(env, insn->src_reg)) {
7978 "R%d partial copy of pointer\n",
7981 } else if (src_reg->type == SCALAR_VALUE) {
7982 *dst_reg = *src_reg;
7983 /* Make sure ID is cleared otherwise
7984 * dst_reg min/max could be incorrectly
7985 * propagated into src_reg by find_equal_scalars()
7988 dst_reg->live |= REG_LIVE_WRITTEN;
7989 dst_reg->subreg_def = env->insn_idx + 1;
7991 mark_reg_unknown(env, regs,
7994 zext_32_to_64(dst_reg);
7998 * remember the value we stored into this reg
8000 /* clear any state __mark_reg_known doesn't set */
8001 mark_reg_unknown(env, regs, insn->dst_reg);
8002 regs[insn->dst_reg].type = SCALAR_VALUE;
8003 if (BPF_CLASS(insn->code) == BPF_ALU64) {
8004 __mark_reg_known(regs + insn->dst_reg,
8007 __mark_reg_known(regs + insn->dst_reg,
8012 } else if (opcode > BPF_END) {
8013 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
8016 } else { /* all other ALU ops: and, sub, xor, add, ... */
8018 if (BPF_SRC(insn->code) == BPF_X) {
8019 if (insn->imm != 0 || insn->off != 0) {
8020 verbose(env, "BPF_ALU uses reserved fields\n");
8023 /* check src1 operand */
8024 err = check_reg_arg(env, insn->src_reg, SRC_OP);
8028 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
8029 verbose(env, "BPF_ALU uses reserved fields\n");
8034 /* check src2 operand */
8035 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
8039 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
8040 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
8041 verbose(env, "div by zero\n");
8045 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
8046 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
8047 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
8049 if (insn->imm < 0 || insn->imm >= size) {
8050 verbose(env, "invalid shift %d\n", insn->imm);
8055 /* check dest operand */
8056 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
8060 return adjust_reg_min_max_vals(env, insn);
8066 static void __find_good_pkt_pointers(struct bpf_func_state *state,
8067 struct bpf_reg_state *dst_reg,
8068 enum bpf_reg_type type, int new_range)
8070 struct bpf_reg_state *reg;
8073 for (i = 0; i < MAX_BPF_REG; i++) {
8074 reg = &state->regs[i];
8075 if (reg->type == type && reg->id == dst_reg->id)
8076 /* keep the maximum range already checked */
8077 reg->range = max(reg->range, new_range);
8080 bpf_for_each_spilled_reg(i, state, reg) {
8083 if (reg->type == type && reg->id == dst_reg->id)
8084 reg->range = max(reg->range, new_range);
8088 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
8089 struct bpf_reg_state *dst_reg,
8090 enum bpf_reg_type type,
8091 bool range_right_open)
8095 if (dst_reg->off < 0 ||
8096 (dst_reg->off == 0 && range_right_open))
8097 /* This doesn't give us any range */
8100 if (dst_reg->umax_value > MAX_PACKET_OFF ||
8101 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
8102 /* Risk of overflow. For instance, ptr + (1<<63) may be less
8103 * than pkt_end, but that's because it's also less than pkt.
8107 new_range = dst_reg->off;
8108 if (range_right_open)
8111 /* Examples for register markings:
8113 * pkt_data in dst register:
8117 * if (r2 > pkt_end) goto <handle exception>
8122 * if (r2 < pkt_end) goto <access okay>
8123 * <handle exception>
8126 * r2 == dst_reg, pkt_end == src_reg
8127 * r2=pkt(id=n,off=8,r=0)
8128 * r3=pkt(id=n,off=0,r=0)
8130 * pkt_data in src register:
8134 * if (pkt_end >= r2) goto <access okay>
8135 * <handle exception>
8139 * if (pkt_end <= r2) goto <handle exception>
8143 * pkt_end == dst_reg, r2 == src_reg
8144 * r2=pkt(id=n,off=8,r=0)
8145 * r3=pkt(id=n,off=0,r=0)
8147 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
8148 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
8149 * and [r3, r3 + 8-1) respectively is safe to access depending on
8153 /* If our ids match, then we must have the same max_value. And we
8154 * don't care about the other reg's fixed offset, since if it's too big
8155 * the range won't allow anything.
8156 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
8158 for (i = 0; i <= vstate->curframe; i++)
8159 __find_good_pkt_pointers(vstate->frame[i], dst_reg, type,
8163 static int is_branch32_taken(struct bpf_reg_state *reg, u32 val, u8 opcode)
8165 struct tnum subreg = tnum_subreg(reg->var_off);
8166 s32 sval = (s32)val;
8170 if (tnum_is_const(subreg))
8171 return !!tnum_equals_const(subreg, val);
8174 if (tnum_is_const(subreg))
8175 return !tnum_equals_const(subreg, val);
8178 if ((~subreg.mask & subreg.value) & val)
8180 if (!((subreg.mask | subreg.value) & val))
8184 if (reg->u32_min_value > val)
8186 else if (reg->u32_max_value <= val)
8190 if (reg->s32_min_value > sval)
8192 else if (reg->s32_max_value <= sval)
8196 if (reg->u32_max_value < val)
8198 else if (reg->u32_min_value >= val)
8202 if (reg->s32_max_value < sval)
8204 else if (reg->s32_min_value >= sval)
8208 if (reg->u32_min_value >= val)
8210 else if (reg->u32_max_value < val)
8214 if (reg->s32_min_value >= sval)
8216 else if (reg->s32_max_value < sval)
8220 if (reg->u32_max_value <= val)
8222 else if (reg->u32_min_value > val)
8226 if (reg->s32_max_value <= sval)
8228 else if (reg->s32_min_value > sval)
8237 static int is_branch64_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
8239 s64 sval = (s64)val;
8243 if (tnum_is_const(reg->var_off))
8244 return !!tnum_equals_const(reg->var_off, val);
8247 if (tnum_is_const(reg->var_off))
8248 return !tnum_equals_const(reg->var_off, val);
8251 if ((~reg->var_off.mask & reg->var_off.value) & val)
8253 if (!((reg->var_off.mask | reg->var_off.value) & val))
8257 if (reg->umin_value > val)
8259 else if (reg->umax_value <= val)
8263 if (reg->smin_value > sval)
8265 else if (reg->smax_value <= sval)
8269 if (reg->umax_value < val)
8271 else if (reg->umin_value >= val)
8275 if (reg->smax_value < sval)
8277 else if (reg->smin_value >= sval)
8281 if (reg->umin_value >= val)
8283 else if (reg->umax_value < val)
8287 if (reg->smin_value >= sval)
8289 else if (reg->smax_value < sval)
8293 if (reg->umax_value <= val)
8295 else if (reg->umin_value > val)
8299 if (reg->smax_value <= sval)
8301 else if (reg->smin_value > sval)
8309 /* compute branch direction of the expression "if (reg opcode val) goto target;"
8311 * 1 - branch will be taken and "goto target" will be executed
8312 * 0 - branch will not be taken and fall-through to next insn
8313 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value
8316 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
8319 if (__is_pointer_value(false, reg)) {
8320 if (!reg_type_not_null(reg->type))
8323 /* If pointer is valid tests against zero will fail so we can
8324 * use this to direct branch taken.
8340 return is_branch32_taken(reg, val, opcode);
8341 return is_branch64_taken(reg, val, opcode);
8344 static int flip_opcode(u32 opcode)
8346 /* How can we transform "a <op> b" into "b <op> a"? */
8347 static const u8 opcode_flip[16] = {
8348 /* these stay the same */
8349 [BPF_JEQ >> 4] = BPF_JEQ,
8350 [BPF_JNE >> 4] = BPF_JNE,
8351 [BPF_JSET >> 4] = BPF_JSET,
8352 /* these swap "lesser" and "greater" (L and G in the opcodes) */
8353 [BPF_JGE >> 4] = BPF_JLE,
8354 [BPF_JGT >> 4] = BPF_JLT,
8355 [BPF_JLE >> 4] = BPF_JGE,
8356 [BPF_JLT >> 4] = BPF_JGT,
8357 [BPF_JSGE >> 4] = BPF_JSLE,
8358 [BPF_JSGT >> 4] = BPF_JSLT,
8359 [BPF_JSLE >> 4] = BPF_JSGE,
8360 [BPF_JSLT >> 4] = BPF_JSGT
8362 return opcode_flip[opcode >> 4];
8365 static int is_pkt_ptr_branch_taken(struct bpf_reg_state *dst_reg,
8366 struct bpf_reg_state *src_reg,
8369 struct bpf_reg_state *pkt;
8371 if (src_reg->type == PTR_TO_PACKET_END) {
8373 } else if (dst_reg->type == PTR_TO_PACKET_END) {
8375 opcode = flip_opcode(opcode);
8380 if (pkt->range >= 0)
8385 /* pkt <= pkt_end */
8389 if (pkt->range == BEYOND_PKT_END)
8390 /* pkt has at last one extra byte beyond pkt_end */
8391 return opcode == BPF_JGT;
8397 /* pkt >= pkt_end */
8398 if (pkt->range == BEYOND_PKT_END || pkt->range == AT_PKT_END)
8399 return opcode == BPF_JGE;
8405 /* Adjusts the register min/max values in the case that the dst_reg is the
8406 * variable register that we are working on, and src_reg is a constant or we're
8407 * simply doing a BPF_K check.
8408 * In JEQ/JNE cases we also adjust the var_off values.
8410 static void reg_set_min_max(struct bpf_reg_state *true_reg,
8411 struct bpf_reg_state *false_reg,
8413 u8 opcode, bool is_jmp32)
8415 struct tnum false_32off = tnum_subreg(false_reg->var_off);
8416 struct tnum false_64off = false_reg->var_off;
8417 struct tnum true_32off = tnum_subreg(true_reg->var_off);
8418 struct tnum true_64off = true_reg->var_off;
8419 s64 sval = (s64)val;
8420 s32 sval32 = (s32)val32;
8422 /* If the dst_reg is a pointer, we can't learn anything about its
8423 * variable offset from the compare (unless src_reg were a pointer into
8424 * the same object, but we don't bother with that.
8425 * Since false_reg and true_reg have the same type by construction, we
8426 * only need to check one of them for pointerness.
8428 if (__is_pointer_value(false, false_reg))
8435 struct bpf_reg_state *reg =
8436 opcode == BPF_JEQ ? true_reg : false_reg;
8438 /* JEQ/JNE comparison doesn't change the register equivalence.
8440 * if (r1 == 42) goto label;
8442 * label: // here both r1 and r2 are known to be 42.
8444 * Hence when marking register as known preserve it's ID.
8447 __mark_reg32_known(reg, val32);
8449 ___mark_reg_known(reg, val);
8454 false_32off = tnum_and(false_32off, tnum_const(~val32));
8455 if (is_power_of_2(val32))
8456 true_32off = tnum_or(true_32off,
8459 false_64off = tnum_and(false_64off, tnum_const(~val));
8460 if (is_power_of_2(val))
8461 true_64off = tnum_or(true_64off,
8469 u32 false_umax = opcode == BPF_JGT ? val32 : val32 - 1;
8470 u32 true_umin = opcode == BPF_JGT ? val32 + 1 : val32;
8472 false_reg->u32_max_value = min(false_reg->u32_max_value,
8474 true_reg->u32_min_value = max(true_reg->u32_min_value,
8477 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
8478 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
8480 false_reg->umax_value = min(false_reg->umax_value, false_umax);
8481 true_reg->umin_value = max(true_reg->umin_value, true_umin);
8489 s32 false_smax = opcode == BPF_JSGT ? sval32 : sval32 - 1;
8490 s32 true_smin = opcode == BPF_JSGT ? sval32 + 1 : sval32;
8492 false_reg->s32_max_value = min(false_reg->s32_max_value, false_smax);
8493 true_reg->s32_min_value = max(true_reg->s32_min_value, true_smin);
8495 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
8496 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
8498 false_reg->smax_value = min(false_reg->smax_value, false_smax);
8499 true_reg->smin_value = max(true_reg->smin_value, true_smin);
8507 u32 false_umin = opcode == BPF_JLT ? val32 : val32 + 1;
8508 u32 true_umax = opcode == BPF_JLT ? val32 - 1 : val32;
8510 false_reg->u32_min_value = max(false_reg->u32_min_value,
8512 true_reg->u32_max_value = min(true_reg->u32_max_value,
8515 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
8516 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
8518 false_reg->umin_value = max(false_reg->umin_value, false_umin);
8519 true_reg->umax_value = min(true_reg->umax_value, true_umax);
8527 s32 false_smin = opcode == BPF_JSLT ? sval32 : sval32 + 1;
8528 s32 true_smax = opcode == BPF_JSLT ? sval32 - 1 : sval32;
8530 false_reg->s32_min_value = max(false_reg->s32_min_value, false_smin);
8531 true_reg->s32_max_value = min(true_reg->s32_max_value, true_smax);
8533 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
8534 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
8536 false_reg->smin_value = max(false_reg->smin_value, false_smin);
8537 true_reg->smax_value = min(true_reg->smax_value, true_smax);
8546 false_reg->var_off = tnum_or(tnum_clear_subreg(false_64off),
8547 tnum_subreg(false_32off));
8548 true_reg->var_off = tnum_or(tnum_clear_subreg(true_64off),
8549 tnum_subreg(true_32off));
8550 __reg_combine_32_into_64(false_reg);
8551 __reg_combine_32_into_64(true_reg);
8553 false_reg->var_off = false_64off;
8554 true_reg->var_off = true_64off;
8555 __reg_combine_64_into_32(false_reg);
8556 __reg_combine_64_into_32(true_reg);
8560 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
8563 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
8564 struct bpf_reg_state *false_reg,
8566 u8 opcode, bool is_jmp32)
8568 opcode = flip_opcode(opcode);
8569 /* This uses zero as "not present in table"; luckily the zero opcode,
8570 * BPF_JA, can't get here.
8573 reg_set_min_max(true_reg, false_reg, val, val32, opcode, is_jmp32);
8576 /* Regs are known to be equal, so intersect their min/max/var_off */
8577 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
8578 struct bpf_reg_state *dst_reg)
8580 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
8581 dst_reg->umin_value);
8582 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
8583 dst_reg->umax_value);
8584 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
8585 dst_reg->smin_value);
8586 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
8587 dst_reg->smax_value);
8588 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
8590 /* We might have learned new bounds from the var_off. */
8591 __update_reg_bounds(src_reg);
8592 __update_reg_bounds(dst_reg);
8593 /* We might have learned something about the sign bit. */
8594 __reg_deduce_bounds(src_reg);
8595 __reg_deduce_bounds(dst_reg);
8596 /* We might have learned some bits from the bounds. */
8597 __reg_bound_offset(src_reg);
8598 __reg_bound_offset(dst_reg);
8599 /* Intersecting with the old var_off might have improved our bounds
8600 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
8601 * then new var_off is (0; 0x7f...fc) which improves our umax.
8603 __update_reg_bounds(src_reg);
8604 __update_reg_bounds(dst_reg);
8607 static void reg_combine_min_max(struct bpf_reg_state *true_src,
8608 struct bpf_reg_state *true_dst,
8609 struct bpf_reg_state *false_src,
8610 struct bpf_reg_state *false_dst,
8615 __reg_combine_min_max(true_src, true_dst);
8618 __reg_combine_min_max(false_src, false_dst);
8623 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
8624 struct bpf_reg_state *reg, u32 id,
8627 if (reg_type_may_be_null(reg->type) && reg->id == id &&
8628 !WARN_ON_ONCE(!reg->id)) {
8629 /* Old offset (both fixed and variable parts) should
8630 * have been known-zero, because we don't allow pointer
8631 * arithmetic on pointers that might be NULL.
8633 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
8634 !tnum_equals_const(reg->var_off, 0) ||
8636 __mark_reg_known_zero(reg);
8640 reg->type = SCALAR_VALUE;
8641 /* We don't need id and ref_obj_id from this point
8642 * onwards anymore, thus we should better reset it,
8643 * so that state pruning has chances to take effect.
8646 reg->ref_obj_id = 0;
8651 mark_ptr_not_null_reg(reg);
8653 if (!reg_may_point_to_spin_lock(reg)) {
8654 /* For not-NULL ptr, reg->ref_obj_id will be reset
8655 * in release_reg_references().
8657 * reg->id is still used by spin_lock ptr. Other
8658 * than spin_lock ptr type, reg->id can be reset.
8665 static void __mark_ptr_or_null_regs(struct bpf_func_state *state, u32 id,
8668 struct bpf_reg_state *reg;
8671 for (i = 0; i < MAX_BPF_REG; i++)
8672 mark_ptr_or_null_reg(state, &state->regs[i], id, is_null);
8674 bpf_for_each_spilled_reg(i, state, reg) {
8677 mark_ptr_or_null_reg(state, reg, id, is_null);
8681 /* The logic is similar to find_good_pkt_pointers(), both could eventually
8682 * be folded together at some point.
8684 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
8687 struct bpf_func_state *state = vstate->frame[vstate->curframe];
8688 struct bpf_reg_state *regs = state->regs;
8689 u32 ref_obj_id = regs[regno].ref_obj_id;
8690 u32 id = regs[regno].id;
8693 if (ref_obj_id && ref_obj_id == id && is_null)
8694 /* regs[regno] is in the " == NULL" branch.
8695 * No one could have freed the reference state before
8696 * doing the NULL check.
8698 WARN_ON_ONCE(release_reference_state(state, id));
8700 for (i = 0; i <= vstate->curframe; i++)
8701 __mark_ptr_or_null_regs(vstate->frame[i], id, is_null);
8704 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
8705 struct bpf_reg_state *dst_reg,
8706 struct bpf_reg_state *src_reg,
8707 struct bpf_verifier_state *this_branch,
8708 struct bpf_verifier_state *other_branch)
8710 if (BPF_SRC(insn->code) != BPF_X)
8713 /* Pointers are always 64-bit. */
8714 if (BPF_CLASS(insn->code) == BPF_JMP32)
8717 switch (BPF_OP(insn->code)) {
8719 if ((dst_reg->type == PTR_TO_PACKET &&
8720 src_reg->type == PTR_TO_PACKET_END) ||
8721 (dst_reg->type == PTR_TO_PACKET_META &&
8722 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
8723 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
8724 find_good_pkt_pointers(this_branch, dst_reg,
8725 dst_reg->type, false);
8726 mark_pkt_end(other_branch, insn->dst_reg, true);
8727 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
8728 src_reg->type == PTR_TO_PACKET) ||
8729 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
8730 src_reg->type == PTR_TO_PACKET_META)) {
8731 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
8732 find_good_pkt_pointers(other_branch, src_reg,
8733 src_reg->type, true);
8734 mark_pkt_end(this_branch, insn->src_reg, false);
8740 if ((dst_reg->type == PTR_TO_PACKET &&
8741 src_reg->type == PTR_TO_PACKET_END) ||
8742 (dst_reg->type == PTR_TO_PACKET_META &&
8743 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
8744 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
8745 find_good_pkt_pointers(other_branch, dst_reg,
8746 dst_reg->type, true);
8747 mark_pkt_end(this_branch, insn->dst_reg, false);
8748 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
8749 src_reg->type == PTR_TO_PACKET) ||
8750 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
8751 src_reg->type == PTR_TO_PACKET_META)) {
8752 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
8753 find_good_pkt_pointers(this_branch, src_reg,
8754 src_reg->type, false);
8755 mark_pkt_end(other_branch, insn->src_reg, true);
8761 if ((dst_reg->type == PTR_TO_PACKET &&
8762 src_reg->type == PTR_TO_PACKET_END) ||
8763 (dst_reg->type == PTR_TO_PACKET_META &&
8764 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
8765 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
8766 find_good_pkt_pointers(this_branch, dst_reg,
8767 dst_reg->type, true);
8768 mark_pkt_end(other_branch, insn->dst_reg, false);
8769 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
8770 src_reg->type == PTR_TO_PACKET) ||
8771 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
8772 src_reg->type == PTR_TO_PACKET_META)) {
8773 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
8774 find_good_pkt_pointers(other_branch, src_reg,
8775 src_reg->type, false);
8776 mark_pkt_end(this_branch, insn->src_reg, true);
8782 if ((dst_reg->type == PTR_TO_PACKET &&
8783 src_reg->type == PTR_TO_PACKET_END) ||
8784 (dst_reg->type == PTR_TO_PACKET_META &&
8785 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
8786 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
8787 find_good_pkt_pointers(other_branch, dst_reg,
8788 dst_reg->type, false);
8789 mark_pkt_end(this_branch, insn->dst_reg, true);
8790 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
8791 src_reg->type == PTR_TO_PACKET) ||
8792 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
8793 src_reg->type == PTR_TO_PACKET_META)) {
8794 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
8795 find_good_pkt_pointers(this_branch, src_reg,
8796 src_reg->type, true);
8797 mark_pkt_end(other_branch, insn->src_reg, false);
8809 static void find_equal_scalars(struct bpf_verifier_state *vstate,
8810 struct bpf_reg_state *known_reg)
8812 struct bpf_func_state *state;
8813 struct bpf_reg_state *reg;
8816 for (i = 0; i <= vstate->curframe; i++) {
8817 state = vstate->frame[i];
8818 for (j = 0; j < MAX_BPF_REG; j++) {
8819 reg = &state->regs[j];
8820 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
8824 bpf_for_each_spilled_reg(j, state, reg) {
8827 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
8833 static int check_cond_jmp_op(struct bpf_verifier_env *env,
8834 struct bpf_insn *insn, int *insn_idx)
8836 struct bpf_verifier_state *this_branch = env->cur_state;
8837 struct bpf_verifier_state *other_branch;
8838 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
8839 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
8840 u8 opcode = BPF_OP(insn->code);
8845 /* Only conditional jumps are expected to reach here. */
8846 if (opcode == BPF_JA || opcode > BPF_JSLE) {
8847 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
8851 if (BPF_SRC(insn->code) == BPF_X) {
8852 if (insn->imm != 0) {
8853 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
8857 /* check src1 operand */
8858 err = check_reg_arg(env, insn->src_reg, SRC_OP);
8862 if (is_pointer_value(env, insn->src_reg)) {
8863 verbose(env, "R%d pointer comparison prohibited\n",
8867 src_reg = ®s[insn->src_reg];
8869 if (insn->src_reg != BPF_REG_0) {
8870 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
8875 /* check src2 operand */
8876 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
8880 dst_reg = ®s[insn->dst_reg];
8881 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
8883 if (BPF_SRC(insn->code) == BPF_K) {
8884 pred = is_branch_taken(dst_reg, insn->imm, opcode, is_jmp32);
8885 } else if (src_reg->type == SCALAR_VALUE &&
8886 is_jmp32 && tnum_is_const(tnum_subreg(src_reg->var_off))) {
8887 pred = is_branch_taken(dst_reg,
8888 tnum_subreg(src_reg->var_off).value,
8891 } else if (src_reg->type == SCALAR_VALUE &&
8892 !is_jmp32 && tnum_is_const(src_reg->var_off)) {
8893 pred = is_branch_taken(dst_reg,
8894 src_reg->var_off.value,
8897 } else if (reg_is_pkt_pointer_any(dst_reg) &&
8898 reg_is_pkt_pointer_any(src_reg) &&
8900 pred = is_pkt_ptr_branch_taken(dst_reg, src_reg, opcode);
8904 /* If we get here with a dst_reg pointer type it is because
8905 * above is_branch_taken() special cased the 0 comparison.
8907 if (!__is_pointer_value(false, dst_reg))
8908 err = mark_chain_precision(env, insn->dst_reg);
8909 if (BPF_SRC(insn->code) == BPF_X && !err &&
8910 !__is_pointer_value(false, src_reg))
8911 err = mark_chain_precision(env, insn->src_reg);
8917 /* Only follow the goto, ignore fall-through. If needed, push
8918 * the fall-through branch for simulation under speculative
8921 if (!env->bypass_spec_v1 &&
8922 !sanitize_speculative_path(env, insn, *insn_idx + 1,
8925 *insn_idx += insn->off;
8927 } else if (pred == 0) {
8928 /* Only follow the fall-through branch, since that's where the
8929 * program will go. If needed, push the goto branch for
8930 * simulation under speculative execution.
8932 if (!env->bypass_spec_v1 &&
8933 !sanitize_speculative_path(env, insn,
8934 *insn_idx + insn->off + 1,
8940 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
8944 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
8946 /* detect if we are comparing against a constant value so we can adjust
8947 * our min/max values for our dst register.
8948 * this is only legit if both are scalars (or pointers to the same
8949 * object, I suppose, but we don't support that right now), because
8950 * otherwise the different base pointers mean the offsets aren't
8953 if (BPF_SRC(insn->code) == BPF_X) {
8954 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
8956 if (dst_reg->type == SCALAR_VALUE &&
8957 src_reg->type == SCALAR_VALUE) {
8958 if (tnum_is_const(src_reg->var_off) ||
8960 tnum_is_const(tnum_subreg(src_reg->var_off))))
8961 reg_set_min_max(&other_branch_regs[insn->dst_reg],
8963 src_reg->var_off.value,
8964 tnum_subreg(src_reg->var_off).value,
8966 else if (tnum_is_const(dst_reg->var_off) ||
8968 tnum_is_const(tnum_subreg(dst_reg->var_off))))
8969 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
8971 dst_reg->var_off.value,
8972 tnum_subreg(dst_reg->var_off).value,
8974 else if (!is_jmp32 &&
8975 (opcode == BPF_JEQ || opcode == BPF_JNE))
8976 /* Comparing for equality, we can combine knowledge */
8977 reg_combine_min_max(&other_branch_regs[insn->src_reg],
8978 &other_branch_regs[insn->dst_reg],
8979 src_reg, dst_reg, opcode);
8981 !WARN_ON_ONCE(src_reg->id != other_branch_regs[insn->src_reg].id)) {
8982 find_equal_scalars(this_branch, src_reg);
8983 find_equal_scalars(other_branch, &other_branch_regs[insn->src_reg]);
8987 } else if (dst_reg->type == SCALAR_VALUE) {
8988 reg_set_min_max(&other_branch_regs[insn->dst_reg],
8989 dst_reg, insn->imm, (u32)insn->imm,
8993 if (dst_reg->type == SCALAR_VALUE && dst_reg->id &&
8994 !WARN_ON_ONCE(dst_reg->id != other_branch_regs[insn->dst_reg].id)) {
8995 find_equal_scalars(this_branch, dst_reg);
8996 find_equal_scalars(other_branch, &other_branch_regs[insn->dst_reg]);
8999 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
9000 * NOTE: these optimizations below are related with pointer comparison
9001 * which will never be JMP32.
9003 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
9004 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
9005 reg_type_may_be_null(dst_reg->type)) {
9006 /* Mark all identical registers in each branch as either
9007 * safe or unknown depending R == 0 or R != 0 conditional.
9009 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
9011 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
9013 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
9014 this_branch, other_branch) &&
9015 is_pointer_value(env, insn->dst_reg)) {
9016 verbose(env, "R%d pointer comparison prohibited\n",
9020 if (env->log.level & BPF_LOG_LEVEL)
9021 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
9025 /* verify BPF_LD_IMM64 instruction */
9026 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
9028 struct bpf_insn_aux_data *aux = cur_aux(env);
9029 struct bpf_reg_state *regs = cur_regs(env);
9030 struct bpf_reg_state *dst_reg;
9031 struct bpf_map *map;
9034 if (BPF_SIZE(insn->code) != BPF_DW) {
9035 verbose(env, "invalid BPF_LD_IMM insn\n");
9038 if (insn->off != 0) {
9039 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
9043 err = check_reg_arg(env, insn->dst_reg, DST_OP);
9047 dst_reg = ®s[insn->dst_reg];
9048 if (insn->src_reg == 0) {
9049 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
9051 dst_reg->type = SCALAR_VALUE;
9052 __mark_reg_known(®s[insn->dst_reg], imm);
9056 if (insn->src_reg == BPF_PSEUDO_BTF_ID) {
9057 mark_reg_known_zero(env, regs, insn->dst_reg);
9059 dst_reg->type = aux->btf_var.reg_type;
9060 switch (dst_reg->type) {
9062 dst_reg->mem_size = aux->btf_var.mem_size;
9065 case PTR_TO_PERCPU_BTF_ID:
9066 dst_reg->btf = aux->btf_var.btf;
9067 dst_reg->btf_id = aux->btf_var.btf_id;
9070 verbose(env, "bpf verifier is misconfigured\n");
9076 if (insn->src_reg == BPF_PSEUDO_FUNC) {
9077 struct bpf_prog_aux *aux = env->prog->aux;
9078 u32 subprogno = insn[1].imm;
9080 if (!aux->func_info) {
9081 verbose(env, "missing btf func_info\n");
9084 if (aux->func_info_aux[subprogno].linkage != BTF_FUNC_STATIC) {
9085 verbose(env, "callback function not static\n");
9089 dst_reg->type = PTR_TO_FUNC;
9090 dst_reg->subprogno = subprogno;
9094 map = env->used_maps[aux->map_index];
9095 mark_reg_known_zero(env, regs, insn->dst_reg);
9096 dst_reg->map_ptr = map;
9098 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE ||
9099 insn->src_reg == BPF_PSEUDO_MAP_IDX_VALUE) {
9100 dst_reg->type = PTR_TO_MAP_VALUE;
9101 dst_reg->off = aux->map_off;
9102 if (map_value_has_spin_lock(map))
9103 dst_reg->id = ++env->id_gen;
9104 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD ||
9105 insn->src_reg == BPF_PSEUDO_MAP_IDX) {
9106 dst_reg->type = CONST_PTR_TO_MAP;
9108 verbose(env, "bpf verifier is misconfigured\n");
9115 static bool may_access_skb(enum bpf_prog_type type)
9118 case BPF_PROG_TYPE_SOCKET_FILTER:
9119 case BPF_PROG_TYPE_SCHED_CLS:
9120 case BPF_PROG_TYPE_SCHED_ACT:
9127 /* verify safety of LD_ABS|LD_IND instructions:
9128 * - they can only appear in the programs where ctx == skb
9129 * - since they are wrappers of function calls, they scratch R1-R5 registers,
9130 * preserve R6-R9, and store return value into R0
9133 * ctx == skb == R6 == CTX
9136 * SRC == any register
9137 * IMM == 32-bit immediate
9140 * R0 - 8/16/32-bit skb data converted to cpu endianness
9142 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
9144 struct bpf_reg_state *regs = cur_regs(env);
9145 static const int ctx_reg = BPF_REG_6;
9146 u8 mode = BPF_MODE(insn->code);
9149 if (!may_access_skb(resolve_prog_type(env->prog))) {
9150 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
9154 if (!env->ops->gen_ld_abs) {
9155 verbose(env, "bpf verifier is misconfigured\n");
9159 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
9160 BPF_SIZE(insn->code) == BPF_DW ||
9161 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
9162 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
9166 /* check whether implicit source operand (register R6) is readable */
9167 err = check_reg_arg(env, ctx_reg, SRC_OP);
9171 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
9172 * gen_ld_abs() may terminate the program at runtime, leading to
9175 err = check_reference_leak(env);
9177 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
9181 if (env->cur_state->active_spin_lock) {
9182 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
9186 if (regs[ctx_reg].type != PTR_TO_CTX) {
9188 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
9192 if (mode == BPF_IND) {
9193 /* check explicit source operand */
9194 err = check_reg_arg(env, insn->src_reg, SRC_OP);
9199 err = check_ctx_reg(env, ®s[ctx_reg], ctx_reg);
9203 /* reset caller saved regs to unreadable */
9204 for (i = 0; i < CALLER_SAVED_REGS; i++) {
9205 mark_reg_not_init(env, regs, caller_saved[i]);
9206 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
9209 /* mark destination R0 register as readable, since it contains
9210 * the value fetched from the packet.
9211 * Already marked as written above.
9213 mark_reg_unknown(env, regs, BPF_REG_0);
9214 /* ld_abs load up to 32-bit skb data. */
9215 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
9219 static int check_return_code(struct bpf_verifier_env *env)
9221 struct tnum enforce_attach_type_range = tnum_unknown;
9222 const struct bpf_prog *prog = env->prog;
9223 struct bpf_reg_state *reg;
9224 struct tnum range = tnum_range(0, 1);
9225 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
9227 const bool is_subprog = env->cur_state->frame[0]->subprogno;
9229 /* LSM and struct_ops func-ptr's return type could be "void" */
9231 (prog_type == BPF_PROG_TYPE_STRUCT_OPS ||
9232 prog_type == BPF_PROG_TYPE_LSM) &&
9233 !prog->aux->attach_func_proto->type)
9236 /* eBPF calling convention is such that R0 is used
9237 * to return the value from eBPF program.
9238 * Make sure that it's readable at this time
9239 * of bpf_exit, which means that program wrote
9240 * something into it earlier
9242 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
9246 if (is_pointer_value(env, BPF_REG_0)) {
9247 verbose(env, "R0 leaks addr as return value\n");
9251 reg = cur_regs(env) + BPF_REG_0;
9253 if (reg->type != SCALAR_VALUE) {
9254 verbose(env, "At subprogram exit the register R0 is not a scalar value (%s)\n",
9255 reg_type_str[reg->type]);
9261 switch (prog_type) {
9262 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
9263 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
9264 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG ||
9265 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETPEERNAME ||
9266 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETPEERNAME ||
9267 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETSOCKNAME ||
9268 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETSOCKNAME)
9269 range = tnum_range(1, 1);
9270 if (env->prog->expected_attach_type == BPF_CGROUP_INET4_BIND ||
9271 env->prog->expected_attach_type == BPF_CGROUP_INET6_BIND)
9272 range = tnum_range(0, 3);
9274 case BPF_PROG_TYPE_CGROUP_SKB:
9275 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
9276 range = tnum_range(0, 3);
9277 enforce_attach_type_range = tnum_range(2, 3);
9280 case BPF_PROG_TYPE_CGROUP_SOCK:
9281 case BPF_PROG_TYPE_SOCK_OPS:
9282 case BPF_PROG_TYPE_CGROUP_DEVICE:
9283 case BPF_PROG_TYPE_CGROUP_SYSCTL:
9284 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
9286 case BPF_PROG_TYPE_RAW_TRACEPOINT:
9287 if (!env->prog->aux->attach_btf_id)
9289 range = tnum_const(0);
9291 case BPF_PROG_TYPE_TRACING:
9292 switch (env->prog->expected_attach_type) {
9293 case BPF_TRACE_FENTRY:
9294 case BPF_TRACE_FEXIT:
9295 range = tnum_const(0);
9297 case BPF_TRACE_RAW_TP:
9298 case BPF_MODIFY_RETURN:
9300 case BPF_TRACE_ITER:
9306 case BPF_PROG_TYPE_SK_LOOKUP:
9307 range = tnum_range(SK_DROP, SK_PASS);
9309 case BPF_PROG_TYPE_EXT:
9310 /* freplace program can return anything as its return value
9311 * depends on the to-be-replaced kernel func or bpf program.
9317 if (reg->type != SCALAR_VALUE) {
9318 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
9319 reg_type_str[reg->type]);
9323 if (!tnum_in(range, reg->var_off)) {
9324 verbose_invalid_scalar(env, reg, &range, "program exit", "R0");
9328 if (!tnum_is_unknown(enforce_attach_type_range) &&
9329 tnum_in(enforce_attach_type_range, reg->var_off))
9330 env->prog->enforce_expected_attach_type = 1;
9334 /* non-recursive DFS pseudo code
9335 * 1 procedure DFS-iterative(G,v):
9336 * 2 label v as discovered
9337 * 3 let S be a stack
9339 * 5 while S is not empty
9341 * 7 if t is what we're looking for:
9343 * 9 for all edges e in G.adjacentEdges(t) do
9344 * 10 if edge e is already labelled
9345 * 11 continue with the next edge
9346 * 12 w <- G.adjacentVertex(t,e)
9347 * 13 if vertex w is not discovered and not explored
9348 * 14 label e as tree-edge
9349 * 15 label w as discovered
9352 * 18 else if vertex w is discovered
9353 * 19 label e as back-edge
9355 * 21 // vertex w is explored
9356 * 22 label e as forward- or cross-edge
9357 * 23 label t as explored
9362 * 0x11 - discovered and fall-through edge labelled
9363 * 0x12 - discovered and fall-through and branch edges labelled
9374 static u32 state_htab_size(struct bpf_verifier_env *env)
9376 return env->prog->len;
9379 static struct bpf_verifier_state_list **explored_state(
9380 struct bpf_verifier_env *env,
9383 struct bpf_verifier_state *cur = env->cur_state;
9384 struct bpf_func_state *state = cur->frame[cur->curframe];
9386 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
9389 static void init_explored_state(struct bpf_verifier_env *env, int idx)
9391 env->insn_aux_data[idx].prune_point = true;
9399 /* t, w, e - match pseudo-code above:
9400 * t - index of current instruction
9401 * w - next instruction
9404 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env,
9407 int *insn_stack = env->cfg.insn_stack;
9408 int *insn_state = env->cfg.insn_state;
9410 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
9411 return DONE_EXPLORING;
9413 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
9414 return DONE_EXPLORING;
9416 if (w < 0 || w >= env->prog->len) {
9417 verbose_linfo(env, t, "%d: ", t);
9418 verbose(env, "jump out of range from insn %d to %d\n", t, w);
9423 /* mark branch target for state pruning */
9424 init_explored_state(env, w);
9426 if (insn_state[w] == 0) {
9428 insn_state[t] = DISCOVERED | e;
9429 insn_state[w] = DISCOVERED;
9430 if (env->cfg.cur_stack >= env->prog->len)
9432 insn_stack[env->cfg.cur_stack++] = w;
9433 return KEEP_EXPLORING;
9434 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
9435 if (loop_ok && env->bpf_capable)
9436 return DONE_EXPLORING;
9437 verbose_linfo(env, t, "%d: ", t);
9438 verbose_linfo(env, w, "%d: ", w);
9439 verbose(env, "back-edge from insn %d to %d\n", t, w);
9441 } else if (insn_state[w] == EXPLORED) {
9442 /* forward- or cross-edge */
9443 insn_state[t] = DISCOVERED | e;
9445 verbose(env, "insn state internal bug\n");
9448 return DONE_EXPLORING;
9451 static int visit_func_call_insn(int t, int insn_cnt,
9452 struct bpf_insn *insns,
9453 struct bpf_verifier_env *env,
9458 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
9462 if (t + 1 < insn_cnt)
9463 init_explored_state(env, t + 1);
9465 init_explored_state(env, t);
9466 ret = push_insn(t, t + insns[t].imm + 1, BRANCH,
9472 /* Visits the instruction at index t and returns one of the following:
9473 * < 0 - an error occurred
9474 * DONE_EXPLORING - the instruction was fully explored
9475 * KEEP_EXPLORING - there is still work to be done before it is fully explored
9477 static int visit_insn(int t, int insn_cnt, struct bpf_verifier_env *env)
9479 struct bpf_insn *insns = env->prog->insnsi;
9482 if (bpf_pseudo_func(insns + t))
9483 return visit_func_call_insn(t, insn_cnt, insns, env, true);
9485 /* All non-branch instructions have a single fall-through edge. */
9486 if (BPF_CLASS(insns[t].code) != BPF_JMP &&
9487 BPF_CLASS(insns[t].code) != BPF_JMP32)
9488 return push_insn(t, t + 1, FALLTHROUGH, env, false);
9490 switch (BPF_OP(insns[t].code)) {
9492 return DONE_EXPLORING;
9495 return visit_func_call_insn(t, insn_cnt, insns, env,
9496 insns[t].src_reg == BPF_PSEUDO_CALL);
9499 if (BPF_SRC(insns[t].code) != BPF_K)
9502 /* unconditional jump with single edge */
9503 ret = push_insn(t, t + insns[t].off + 1, FALLTHROUGH, env,
9508 /* unconditional jmp is not a good pruning point,
9509 * but it's marked, since backtracking needs
9510 * to record jmp history in is_state_visited().
9512 init_explored_state(env, t + insns[t].off + 1);
9513 /* tell verifier to check for equivalent states
9514 * after every call and jump
9516 if (t + 1 < insn_cnt)
9517 init_explored_state(env, t + 1);
9522 /* conditional jump with two edges */
9523 init_explored_state(env, t);
9524 ret = push_insn(t, t + 1, FALLTHROUGH, env, true);
9528 return push_insn(t, t + insns[t].off + 1, BRANCH, env, true);
9532 /* non-recursive depth-first-search to detect loops in BPF program
9533 * loop == back-edge in directed graph
9535 static int check_cfg(struct bpf_verifier_env *env)
9537 int insn_cnt = env->prog->len;
9538 int *insn_stack, *insn_state;
9542 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
9546 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
9552 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
9553 insn_stack[0] = 0; /* 0 is the first instruction */
9554 env->cfg.cur_stack = 1;
9556 while (env->cfg.cur_stack > 0) {
9557 int t = insn_stack[env->cfg.cur_stack - 1];
9559 ret = visit_insn(t, insn_cnt, env);
9561 case DONE_EXPLORING:
9562 insn_state[t] = EXPLORED;
9563 env->cfg.cur_stack--;
9565 case KEEP_EXPLORING:
9569 verbose(env, "visit_insn internal bug\n");
9576 if (env->cfg.cur_stack < 0) {
9577 verbose(env, "pop stack internal bug\n");
9582 for (i = 0; i < insn_cnt; i++) {
9583 if (insn_state[i] != EXPLORED) {
9584 verbose(env, "unreachable insn %d\n", i);
9589 ret = 0; /* cfg looks good */
9594 env->cfg.insn_state = env->cfg.insn_stack = NULL;
9598 static int check_abnormal_return(struct bpf_verifier_env *env)
9602 for (i = 1; i < env->subprog_cnt; i++) {
9603 if (env->subprog_info[i].has_ld_abs) {
9604 verbose(env, "LD_ABS is not allowed in subprogs without BTF\n");
9607 if (env->subprog_info[i].has_tail_call) {
9608 verbose(env, "tail_call is not allowed in subprogs without BTF\n");
9615 /* The minimum supported BTF func info size */
9616 #define MIN_BPF_FUNCINFO_SIZE 8
9617 #define MAX_FUNCINFO_REC_SIZE 252
9619 static int check_btf_func(struct bpf_verifier_env *env,
9620 const union bpf_attr *attr,
9623 const struct btf_type *type, *func_proto, *ret_type;
9624 u32 i, nfuncs, urec_size, min_size;
9625 u32 krec_size = sizeof(struct bpf_func_info);
9626 struct bpf_func_info *krecord;
9627 struct bpf_func_info_aux *info_aux = NULL;
9628 struct bpf_prog *prog;
9629 const struct btf *btf;
9631 u32 prev_offset = 0;
9635 nfuncs = attr->func_info_cnt;
9637 if (check_abnormal_return(env))
9642 if (nfuncs != env->subprog_cnt) {
9643 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
9647 urec_size = attr->func_info_rec_size;
9648 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
9649 urec_size > MAX_FUNCINFO_REC_SIZE ||
9650 urec_size % sizeof(u32)) {
9651 verbose(env, "invalid func info rec size %u\n", urec_size);
9656 btf = prog->aux->btf;
9658 urecord = make_bpfptr(attr->func_info, uattr.is_kernel);
9659 min_size = min_t(u32, krec_size, urec_size);
9661 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
9664 info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN);
9668 for (i = 0; i < nfuncs; i++) {
9669 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
9671 if (ret == -E2BIG) {
9672 verbose(env, "nonzero tailing record in func info");
9673 /* set the size kernel expects so loader can zero
9674 * out the rest of the record.
9676 if (copy_to_bpfptr_offset(uattr,
9677 offsetof(union bpf_attr, func_info_rec_size),
9678 &min_size, sizeof(min_size)))
9684 if (copy_from_bpfptr(&krecord[i], urecord, min_size)) {
9689 /* check insn_off */
9692 if (krecord[i].insn_off) {
9694 "nonzero insn_off %u for the first func info record",
9695 krecord[i].insn_off);
9698 } else if (krecord[i].insn_off <= prev_offset) {
9700 "same or smaller insn offset (%u) than previous func info record (%u)",
9701 krecord[i].insn_off, prev_offset);
9705 if (env->subprog_info[i].start != krecord[i].insn_off) {
9706 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
9711 type = btf_type_by_id(btf, krecord[i].type_id);
9712 if (!type || !btf_type_is_func(type)) {
9713 verbose(env, "invalid type id %d in func info",
9714 krecord[i].type_id);
9717 info_aux[i].linkage = BTF_INFO_VLEN(type->info);
9719 func_proto = btf_type_by_id(btf, type->type);
9720 if (unlikely(!func_proto || !btf_type_is_func_proto(func_proto)))
9721 /* btf_func_check() already verified it during BTF load */
9723 ret_type = btf_type_skip_modifiers(btf, func_proto->type, NULL);
9725 btf_type_is_small_int(ret_type) || btf_type_is_enum(ret_type);
9726 if (i && !scalar_return && env->subprog_info[i].has_ld_abs) {
9727 verbose(env, "LD_ABS is only allowed in functions that return 'int'.\n");
9730 if (i && !scalar_return && env->subprog_info[i].has_tail_call) {
9731 verbose(env, "tail_call is only allowed in functions that return 'int'.\n");
9735 prev_offset = krecord[i].insn_off;
9736 bpfptr_add(&urecord, urec_size);
9739 prog->aux->func_info = krecord;
9740 prog->aux->func_info_cnt = nfuncs;
9741 prog->aux->func_info_aux = info_aux;
9750 static void adjust_btf_func(struct bpf_verifier_env *env)
9752 struct bpf_prog_aux *aux = env->prog->aux;
9755 if (!aux->func_info)
9758 for (i = 0; i < env->subprog_cnt; i++)
9759 aux->func_info[i].insn_off = env->subprog_info[i].start;
9762 #define MIN_BPF_LINEINFO_SIZE (offsetof(struct bpf_line_info, line_col) + \
9763 sizeof(((struct bpf_line_info *)(0))->line_col))
9764 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
9766 static int check_btf_line(struct bpf_verifier_env *env,
9767 const union bpf_attr *attr,
9770 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
9771 struct bpf_subprog_info *sub;
9772 struct bpf_line_info *linfo;
9773 struct bpf_prog *prog;
9774 const struct btf *btf;
9778 nr_linfo = attr->line_info_cnt;
9782 rec_size = attr->line_info_rec_size;
9783 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
9784 rec_size > MAX_LINEINFO_REC_SIZE ||
9785 rec_size & (sizeof(u32) - 1))
9788 /* Need to zero it in case the userspace may
9789 * pass in a smaller bpf_line_info object.
9791 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
9792 GFP_KERNEL | __GFP_NOWARN);
9797 btf = prog->aux->btf;
9800 sub = env->subprog_info;
9801 ulinfo = make_bpfptr(attr->line_info, uattr.is_kernel);
9802 expected_size = sizeof(struct bpf_line_info);
9803 ncopy = min_t(u32, expected_size, rec_size);
9804 for (i = 0; i < nr_linfo; i++) {
9805 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
9807 if (err == -E2BIG) {
9808 verbose(env, "nonzero tailing record in line_info");
9809 if (copy_to_bpfptr_offset(uattr,
9810 offsetof(union bpf_attr, line_info_rec_size),
9811 &expected_size, sizeof(expected_size)))
9817 if (copy_from_bpfptr(&linfo[i], ulinfo, ncopy)) {
9823 * Check insn_off to ensure
9824 * 1) strictly increasing AND
9825 * 2) bounded by prog->len
9827 * The linfo[0].insn_off == 0 check logically falls into
9828 * the later "missing bpf_line_info for func..." case
9829 * because the first linfo[0].insn_off must be the
9830 * first sub also and the first sub must have
9831 * subprog_info[0].start == 0.
9833 if ((i && linfo[i].insn_off <= prev_offset) ||
9834 linfo[i].insn_off >= prog->len) {
9835 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
9836 i, linfo[i].insn_off, prev_offset,
9842 if (!prog->insnsi[linfo[i].insn_off].code) {
9844 "Invalid insn code at line_info[%u].insn_off\n",
9850 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
9851 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
9852 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
9857 if (s != env->subprog_cnt) {
9858 if (linfo[i].insn_off == sub[s].start) {
9859 sub[s].linfo_idx = i;
9861 } else if (sub[s].start < linfo[i].insn_off) {
9862 verbose(env, "missing bpf_line_info for func#%u\n", s);
9868 prev_offset = linfo[i].insn_off;
9869 bpfptr_add(&ulinfo, rec_size);
9872 if (s != env->subprog_cnt) {
9873 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
9874 env->subprog_cnt - s, s);
9879 prog->aux->linfo = linfo;
9880 prog->aux->nr_linfo = nr_linfo;
9889 static int check_btf_info(struct bpf_verifier_env *env,
9890 const union bpf_attr *attr,
9896 if (!attr->func_info_cnt && !attr->line_info_cnt) {
9897 if (check_abnormal_return(env))
9902 btf = btf_get_by_fd(attr->prog_btf_fd);
9904 return PTR_ERR(btf);
9905 if (btf_is_kernel(btf)) {
9909 env->prog->aux->btf = btf;
9911 err = check_btf_func(env, attr, uattr);
9915 err = check_btf_line(env, attr, uattr);
9922 /* check %cur's range satisfies %old's */
9923 static bool range_within(struct bpf_reg_state *old,
9924 struct bpf_reg_state *cur)
9926 return old->umin_value <= cur->umin_value &&
9927 old->umax_value >= cur->umax_value &&
9928 old->smin_value <= cur->smin_value &&
9929 old->smax_value >= cur->smax_value &&
9930 old->u32_min_value <= cur->u32_min_value &&
9931 old->u32_max_value >= cur->u32_max_value &&
9932 old->s32_min_value <= cur->s32_min_value &&
9933 old->s32_max_value >= cur->s32_max_value;
9936 /* If in the old state two registers had the same id, then they need to have
9937 * the same id in the new state as well. But that id could be different from
9938 * the old state, so we need to track the mapping from old to new ids.
9939 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
9940 * regs with old id 5 must also have new id 9 for the new state to be safe. But
9941 * regs with a different old id could still have new id 9, we don't care about
9943 * So we look through our idmap to see if this old id has been seen before. If
9944 * so, we require the new id to match; otherwise, we add the id pair to the map.
9946 static bool check_ids(u32 old_id, u32 cur_id, struct bpf_id_pair *idmap)
9950 for (i = 0; i < BPF_ID_MAP_SIZE; i++) {
9951 if (!idmap[i].old) {
9952 /* Reached an empty slot; haven't seen this id before */
9953 idmap[i].old = old_id;
9954 idmap[i].cur = cur_id;
9957 if (idmap[i].old == old_id)
9958 return idmap[i].cur == cur_id;
9960 /* We ran out of idmap slots, which should be impossible */
9965 static void clean_func_state(struct bpf_verifier_env *env,
9966 struct bpf_func_state *st)
9968 enum bpf_reg_liveness live;
9971 for (i = 0; i < BPF_REG_FP; i++) {
9972 live = st->regs[i].live;
9973 /* liveness must not touch this register anymore */
9974 st->regs[i].live |= REG_LIVE_DONE;
9975 if (!(live & REG_LIVE_READ))
9976 /* since the register is unused, clear its state
9977 * to make further comparison simpler
9979 __mark_reg_not_init(env, &st->regs[i]);
9982 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
9983 live = st->stack[i].spilled_ptr.live;
9984 /* liveness must not touch this stack slot anymore */
9985 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
9986 if (!(live & REG_LIVE_READ)) {
9987 __mark_reg_not_init(env, &st->stack[i].spilled_ptr);
9988 for (j = 0; j < BPF_REG_SIZE; j++)
9989 st->stack[i].slot_type[j] = STACK_INVALID;
9994 static void clean_verifier_state(struct bpf_verifier_env *env,
9995 struct bpf_verifier_state *st)
9999 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
10000 /* all regs in this state in all frames were already marked */
10003 for (i = 0; i <= st->curframe; i++)
10004 clean_func_state(env, st->frame[i]);
10007 /* the parentage chains form a tree.
10008 * the verifier states are added to state lists at given insn and
10009 * pushed into state stack for future exploration.
10010 * when the verifier reaches bpf_exit insn some of the verifer states
10011 * stored in the state lists have their final liveness state already,
10012 * but a lot of states will get revised from liveness point of view when
10013 * the verifier explores other branches.
10016 * 2: if r1 == 100 goto pc+1
10019 * when the verifier reaches exit insn the register r0 in the state list of
10020 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
10021 * of insn 2 and goes exploring further. At the insn 4 it will walk the
10022 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
10024 * Since the verifier pushes the branch states as it sees them while exploring
10025 * the program the condition of walking the branch instruction for the second
10026 * time means that all states below this branch were already explored and
10027 * their final liveness marks are already propagated.
10028 * Hence when the verifier completes the search of state list in is_state_visited()
10029 * we can call this clean_live_states() function to mark all liveness states
10030 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
10031 * will not be used.
10032 * This function also clears the registers and stack for states that !READ
10033 * to simplify state merging.
10035 * Important note here that walking the same branch instruction in the callee
10036 * doesn't meant that the states are DONE. The verifier has to compare
10039 static void clean_live_states(struct bpf_verifier_env *env, int insn,
10040 struct bpf_verifier_state *cur)
10042 struct bpf_verifier_state_list *sl;
10045 sl = *explored_state(env, insn);
10047 if (sl->state.branches)
10049 if (sl->state.insn_idx != insn ||
10050 sl->state.curframe != cur->curframe)
10052 for (i = 0; i <= cur->curframe; i++)
10053 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
10055 clean_verifier_state(env, &sl->state);
10061 /* Returns true if (rold safe implies rcur safe) */
10062 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
10063 struct bpf_id_pair *idmap)
10067 if (!(rold->live & REG_LIVE_READ))
10068 /* explored state didn't use this */
10071 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
10073 if (rold->type == PTR_TO_STACK)
10074 /* two stack pointers are equal only if they're pointing to
10075 * the same stack frame, since fp-8 in foo != fp-8 in bar
10077 return equal && rold->frameno == rcur->frameno;
10082 if (rold->type == NOT_INIT)
10083 /* explored state can't have used this */
10085 if (rcur->type == NOT_INIT)
10087 switch (rold->type) {
10089 if (rcur->type == SCALAR_VALUE) {
10090 if (!rold->precise && !rcur->precise)
10092 /* new val must satisfy old val knowledge */
10093 return range_within(rold, rcur) &&
10094 tnum_in(rold->var_off, rcur->var_off);
10096 /* We're trying to use a pointer in place of a scalar.
10097 * Even if the scalar was unbounded, this could lead to
10098 * pointer leaks because scalars are allowed to leak
10099 * while pointers are not. We could make this safe in
10100 * special cases if root is calling us, but it's
10101 * probably not worth the hassle.
10105 case PTR_TO_MAP_KEY:
10106 case PTR_TO_MAP_VALUE:
10107 /* If the new min/max/var_off satisfy the old ones and
10108 * everything else matches, we are OK.
10109 * 'id' is not compared, since it's only used for maps with
10110 * bpf_spin_lock inside map element and in such cases if
10111 * the rest of the prog is valid for one map element then
10112 * it's valid for all map elements regardless of the key
10113 * used in bpf_map_lookup()
10115 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
10116 range_within(rold, rcur) &&
10117 tnum_in(rold->var_off, rcur->var_off);
10118 case PTR_TO_MAP_VALUE_OR_NULL:
10119 /* a PTR_TO_MAP_VALUE could be safe to use as a
10120 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
10121 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
10122 * checked, doing so could have affected others with the same
10123 * id, and we can't check for that because we lost the id when
10124 * we converted to a PTR_TO_MAP_VALUE.
10126 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
10128 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
10130 /* Check our ids match any regs they're supposed to */
10131 return check_ids(rold->id, rcur->id, idmap);
10132 case PTR_TO_PACKET_META:
10133 case PTR_TO_PACKET:
10134 if (rcur->type != rold->type)
10136 /* We must have at least as much range as the old ptr
10137 * did, so that any accesses which were safe before are
10138 * still safe. This is true even if old range < old off,
10139 * since someone could have accessed through (ptr - k), or
10140 * even done ptr -= k in a register, to get a safe access.
10142 if (rold->range > rcur->range)
10144 /* If the offsets don't match, we can't trust our alignment;
10145 * nor can we be sure that we won't fall out of range.
10147 if (rold->off != rcur->off)
10149 /* id relations must be preserved */
10150 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
10152 /* new val must satisfy old val knowledge */
10153 return range_within(rold, rcur) &&
10154 tnum_in(rold->var_off, rcur->var_off);
10156 case CONST_PTR_TO_MAP:
10157 case PTR_TO_PACKET_END:
10158 case PTR_TO_FLOW_KEYS:
10159 case PTR_TO_SOCKET:
10160 case PTR_TO_SOCKET_OR_NULL:
10161 case PTR_TO_SOCK_COMMON:
10162 case PTR_TO_SOCK_COMMON_OR_NULL:
10163 case PTR_TO_TCP_SOCK:
10164 case PTR_TO_TCP_SOCK_OR_NULL:
10165 case PTR_TO_XDP_SOCK:
10166 /* Only valid matches are exact, which memcmp() above
10167 * would have accepted
10170 /* Don't know what's going on, just say it's not safe */
10174 /* Shouldn't get here; if we do, say it's not safe */
10179 static bool stacksafe(struct bpf_func_state *old,
10180 struct bpf_func_state *cur,
10181 struct bpf_id_pair *idmap)
10185 /* walk slots of the explored stack and ignore any additional
10186 * slots in the current stack, since explored(safe) state
10189 for (i = 0; i < old->allocated_stack; i++) {
10190 spi = i / BPF_REG_SIZE;
10192 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
10193 i += BPF_REG_SIZE - 1;
10194 /* explored state didn't use this */
10198 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
10201 /* explored stack has more populated slots than current stack
10202 * and these slots were used
10204 if (i >= cur->allocated_stack)
10207 /* if old state was safe with misc data in the stack
10208 * it will be safe with zero-initialized stack.
10209 * The opposite is not true
10211 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
10212 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
10214 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
10215 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
10216 /* Ex: old explored (safe) state has STACK_SPILL in
10217 * this stack slot, but current has STACK_MISC ->
10218 * this verifier states are not equivalent,
10219 * return false to continue verification of this path
10222 if (i % BPF_REG_SIZE)
10224 if (old->stack[spi].slot_type[0] != STACK_SPILL)
10226 if (!regsafe(&old->stack[spi].spilled_ptr,
10227 &cur->stack[spi].spilled_ptr,
10229 /* when explored and current stack slot are both storing
10230 * spilled registers, check that stored pointers types
10231 * are the same as well.
10232 * Ex: explored safe path could have stored
10233 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
10234 * but current path has stored:
10235 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
10236 * such verifier states are not equivalent.
10237 * return false to continue verification of this path
10244 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
10246 if (old->acquired_refs != cur->acquired_refs)
10248 return !memcmp(old->refs, cur->refs,
10249 sizeof(*old->refs) * old->acquired_refs);
10252 /* compare two verifier states
10254 * all states stored in state_list are known to be valid, since
10255 * verifier reached 'bpf_exit' instruction through them
10257 * this function is called when verifier exploring different branches of
10258 * execution popped from the state stack. If it sees an old state that has
10259 * more strict register state and more strict stack state then this execution
10260 * branch doesn't need to be explored further, since verifier already
10261 * concluded that more strict state leads to valid finish.
10263 * Therefore two states are equivalent if register state is more conservative
10264 * and explored stack state is more conservative than the current one.
10267 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
10268 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
10270 * In other words if current stack state (one being explored) has more
10271 * valid slots than old one that already passed validation, it means
10272 * the verifier can stop exploring and conclude that current state is valid too
10274 * Similarly with registers. If explored state has register type as invalid
10275 * whereas register type in current state is meaningful, it means that
10276 * the current state will reach 'bpf_exit' instruction safely
10278 static bool func_states_equal(struct bpf_verifier_env *env, struct bpf_func_state *old,
10279 struct bpf_func_state *cur)
10283 memset(env->idmap_scratch, 0, sizeof(env->idmap_scratch));
10284 for (i = 0; i < MAX_BPF_REG; i++)
10285 if (!regsafe(&old->regs[i], &cur->regs[i], env->idmap_scratch))
10288 if (!stacksafe(old, cur, env->idmap_scratch))
10291 if (!refsafe(old, cur))
10297 static bool states_equal(struct bpf_verifier_env *env,
10298 struct bpf_verifier_state *old,
10299 struct bpf_verifier_state *cur)
10303 if (old->curframe != cur->curframe)
10306 /* Verification state from speculative execution simulation
10307 * must never prune a non-speculative execution one.
10309 if (old->speculative && !cur->speculative)
10312 if (old->active_spin_lock != cur->active_spin_lock)
10315 /* for states to be equal callsites have to be the same
10316 * and all frame states need to be equivalent
10318 for (i = 0; i <= old->curframe; i++) {
10319 if (old->frame[i]->callsite != cur->frame[i]->callsite)
10321 if (!func_states_equal(env, old->frame[i], cur->frame[i]))
10327 /* Return 0 if no propagation happened. Return negative error code if error
10328 * happened. Otherwise, return the propagated bit.
10330 static int propagate_liveness_reg(struct bpf_verifier_env *env,
10331 struct bpf_reg_state *reg,
10332 struct bpf_reg_state *parent_reg)
10334 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
10335 u8 flag = reg->live & REG_LIVE_READ;
10338 /* When comes here, read flags of PARENT_REG or REG could be any of
10339 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
10340 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
10342 if (parent_flag == REG_LIVE_READ64 ||
10343 /* Or if there is no read flag from REG. */
10345 /* Or if the read flag from REG is the same as PARENT_REG. */
10346 parent_flag == flag)
10349 err = mark_reg_read(env, reg, parent_reg, flag);
10356 /* A write screens off any subsequent reads; but write marks come from the
10357 * straight-line code between a state and its parent. When we arrive at an
10358 * equivalent state (jump target or such) we didn't arrive by the straight-line
10359 * code, so read marks in the state must propagate to the parent regardless
10360 * of the state's write marks. That's what 'parent == state->parent' comparison
10361 * in mark_reg_read() is for.
10363 static int propagate_liveness(struct bpf_verifier_env *env,
10364 const struct bpf_verifier_state *vstate,
10365 struct bpf_verifier_state *vparent)
10367 struct bpf_reg_state *state_reg, *parent_reg;
10368 struct bpf_func_state *state, *parent;
10369 int i, frame, err = 0;
10371 if (vparent->curframe != vstate->curframe) {
10372 WARN(1, "propagate_live: parent frame %d current frame %d\n",
10373 vparent->curframe, vstate->curframe);
10376 /* Propagate read liveness of registers... */
10377 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
10378 for (frame = 0; frame <= vstate->curframe; frame++) {
10379 parent = vparent->frame[frame];
10380 state = vstate->frame[frame];
10381 parent_reg = parent->regs;
10382 state_reg = state->regs;
10383 /* We don't need to worry about FP liveness, it's read-only */
10384 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
10385 err = propagate_liveness_reg(env, &state_reg[i],
10389 if (err == REG_LIVE_READ64)
10390 mark_insn_zext(env, &parent_reg[i]);
10393 /* Propagate stack slots. */
10394 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
10395 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
10396 parent_reg = &parent->stack[i].spilled_ptr;
10397 state_reg = &state->stack[i].spilled_ptr;
10398 err = propagate_liveness_reg(env, state_reg,
10407 /* find precise scalars in the previous equivalent state and
10408 * propagate them into the current state
10410 static int propagate_precision(struct bpf_verifier_env *env,
10411 const struct bpf_verifier_state *old)
10413 struct bpf_reg_state *state_reg;
10414 struct bpf_func_state *state;
10417 state = old->frame[old->curframe];
10418 state_reg = state->regs;
10419 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
10420 if (state_reg->type != SCALAR_VALUE ||
10421 !state_reg->precise)
10423 if (env->log.level & BPF_LOG_LEVEL2)
10424 verbose(env, "propagating r%d\n", i);
10425 err = mark_chain_precision(env, i);
10430 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
10431 if (state->stack[i].slot_type[0] != STACK_SPILL)
10433 state_reg = &state->stack[i].spilled_ptr;
10434 if (state_reg->type != SCALAR_VALUE ||
10435 !state_reg->precise)
10437 if (env->log.level & BPF_LOG_LEVEL2)
10438 verbose(env, "propagating fp%d\n",
10439 (-i - 1) * BPF_REG_SIZE);
10440 err = mark_chain_precision_stack(env, i);
10447 static bool states_maybe_looping(struct bpf_verifier_state *old,
10448 struct bpf_verifier_state *cur)
10450 struct bpf_func_state *fold, *fcur;
10451 int i, fr = cur->curframe;
10453 if (old->curframe != fr)
10456 fold = old->frame[fr];
10457 fcur = cur->frame[fr];
10458 for (i = 0; i < MAX_BPF_REG; i++)
10459 if (memcmp(&fold->regs[i], &fcur->regs[i],
10460 offsetof(struct bpf_reg_state, parent)))
10466 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
10468 struct bpf_verifier_state_list *new_sl;
10469 struct bpf_verifier_state_list *sl, **pprev;
10470 struct bpf_verifier_state *cur = env->cur_state, *new;
10471 int i, j, err, states_cnt = 0;
10472 bool add_new_state = env->test_state_freq ? true : false;
10474 cur->last_insn_idx = env->prev_insn_idx;
10475 if (!env->insn_aux_data[insn_idx].prune_point)
10476 /* this 'insn_idx' instruction wasn't marked, so we will not
10477 * be doing state search here
10481 /* bpf progs typically have pruning point every 4 instructions
10482 * http://vger.kernel.org/bpfconf2019.html#session-1
10483 * Do not add new state for future pruning if the verifier hasn't seen
10484 * at least 2 jumps and at least 8 instructions.
10485 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
10486 * In tests that amounts to up to 50% reduction into total verifier
10487 * memory consumption and 20% verifier time speedup.
10489 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
10490 env->insn_processed - env->prev_insn_processed >= 8)
10491 add_new_state = true;
10493 pprev = explored_state(env, insn_idx);
10496 clean_live_states(env, insn_idx, cur);
10500 if (sl->state.insn_idx != insn_idx)
10502 if (sl->state.branches) {
10503 if (states_maybe_looping(&sl->state, cur) &&
10504 states_equal(env, &sl->state, cur)) {
10505 verbose_linfo(env, insn_idx, "; ");
10506 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
10509 /* if the verifier is processing a loop, avoid adding new state
10510 * too often, since different loop iterations have distinct
10511 * states and may not help future pruning.
10512 * This threshold shouldn't be too low to make sure that
10513 * a loop with large bound will be rejected quickly.
10514 * The most abusive loop will be:
10516 * if r1 < 1000000 goto pc-2
10517 * 1M insn_procssed limit / 100 == 10k peak states.
10518 * This threshold shouldn't be too high either, since states
10519 * at the end of the loop are likely to be useful in pruning.
10521 if (env->jmps_processed - env->prev_jmps_processed < 20 &&
10522 env->insn_processed - env->prev_insn_processed < 100)
10523 add_new_state = false;
10526 if (states_equal(env, &sl->state, cur)) {
10528 /* reached equivalent register/stack state,
10529 * prune the search.
10530 * Registers read by the continuation are read by us.
10531 * If we have any write marks in env->cur_state, they
10532 * will prevent corresponding reads in the continuation
10533 * from reaching our parent (an explored_state). Our
10534 * own state will get the read marks recorded, but
10535 * they'll be immediately forgotten as we're pruning
10536 * this state and will pop a new one.
10538 err = propagate_liveness(env, &sl->state, cur);
10540 /* if previous state reached the exit with precision and
10541 * current state is equivalent to it (except precsion marks)
10542 * the precision needs to be propagated back in
10543 * the current state.
10545 err = err ? : push_jmp_history(env, cur);
10546 err = err ? : propagate_precision(env, &sl->state);
10552 /* when new state is not going to be added do not increase miss count.
10553 * Otherwise several loop iterations will remove the state
10554 * recorded earlier. The goal of these heuristics is to have
10555 * states from some iterations of the loop (some in the beginning
10556 * and some at the end) to help pruning.
10560 /* heuristic to determine whether this state is beneficial
10561 * to keep checking from state equivalence point of view.
10562 * Higher numbers increase max_states_per_insn and verification time,
10563 * but do not meaningfully decrease insn_processed.
10565 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
10566 /* the state is unlikely to be useful. Remove it to
10567 * speed up verification
10570 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
10571 u32 br = sl->state.branches;
10574 "BUG live_done but branches_to_explore %d\n",
10576 free_verifier_state(&sl->state, false);
10578 env->peak_states--;
10580 /* cannot free this state, since parentage chain may
10581 * walk it later. Add it for free_list instead to
10582 * be freed at the end of verification
10584 sl->next = env->free_list;
10585 env->free_list = sl;
10595 if (env->max_states_per_insn < states_cnt)
10596 env->max_states_per_insn = states_cnt;
10598 if (!env->bpf_capable && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
10599 return push_jmp_history(env, cur);
10601 if (!add_new_state)
10602 return push_jmp_history(env, cur);
10604 /* There were no equivalent states, remember the current one.
10605 * Technically the current state is not proven to be safe yet,
10606 * but it will either reach outer most bpf_exit (which means it's safe)
10607 * or it will be rejected. When there are no loops the verifier won't be
10608 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
10609 * again on the way to bpf_exit.
10610 * When looping the sl->state.branches will be > 0 and this state
10611 * will not be considered for equivalence until branches == 0.
10613 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
10616 env->total_states++;
10617 env->peak_states++;
10618 env->prev_jmps_processed = env->jmps_processed;
10619 env->prev_insn_processed = env->insn_processed;
10621 /* add new state to the head of linked list */
10622 new = &new_sl->state;
10623 err = copy_verifier_state(new, cur);
10625 free_verifier_state(new, false);
10629 new->insn_idx = insn_idx;
10630 WARN_ONCE(new->branches != 1,
10631 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
10634 cur->first_insn_idx = insn_idx;
10635 clear_jmp_history(cur);
10636 new_sl->next = *explored_state(env, insn_idx);
10637 *explored_state(env, insn_idx) = new_sl;
10638 /* connect new state to parentage chain. Current frame needs all
10639 * registers connected. Only r6 - r9 of the callers are alive (pushed
10640 * to the stack implicitly by JITs) so in callers' frames connect just
10641 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
10642 * the state of the call instruction (with WRITTEN set), and r0 comes
10643 * from callee with its full parentage chain, anyway.
10645 /* clear write marks in current state: the writes we did are not writes
10646 * our child did, so they don't screen off its reads from us.
10647 * (There are no read marks in current state, because reads always mark
10648 * their parent and current state never has children yet. Only
10649 * explored_states can get read marks.)
10651 for (j = 0; j <= cur->curframe; j++) {
10652 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
10653 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
10654 for (i = 0; i < BPF_REG_FP; i++)
10655 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
10658 /* all stack frames are accessible from callee, clear them all */
10659 for (j = 0; j <= cur->curframe; j++) {
10660 struct bpf_func_state *frame = cur->frame[j];
10661 struct bpf_func_state *newframe = new->frame[j];
10663 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
10664 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
10665 frame->stack[i].spilled_ptr.parent =
10666 &newframe->stack[i].spilled_ptr;
10672 /* Return true if it's OK to have the same insn return a different type. */
10673 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
10677 case PTR_TO_SOCKET:
10678 case PTR_TO_SOCKET_OR_NULL:
10679 case PTR_TO_SOCK_COMMON:
10680 case PTR_TO_SOCK_COMMON_OR_NULL:
10681 case PTR_TO_TCP_SOCK:
10682 case PTR_TO_TCP_SOCK_OR_NULL:
10683 case PTR_TO_XDP_SOCK:
10684 case PTR_TO_BTF_ID:
10685 case PTR_TO_BTF_ID_OR_NULL:
10692 /* If an instruction was previously used with particular pointer types, then we
10693 * need to be careful to avoid cases such as the below, where it may be ok
10694 * for one branch accessing the pointer, but not ok for the other branch:
10699 * R1 = some_other_valid_ptr;
10702 * R2 = *(u32 *)(R1 + 0);
10704 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
10706 return src != prev && (!reg_type_mismatch_ok(src) ||
10707 !reg_type_mismatch_ok(prev));
10710 static int do_check(struct bpf_verifier_env *env)
10712 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
10713 struct bpf_verifier_state *state = env->cur_state;
10714 struct bpf_insn *insns = env->prog->insnsi;
10715 struct bpf_reg_state *regs;
10716 int insn_cnt = env->prog->len;
10717 bool do_print_state = false;
10718 int prev_insn_idx = -1;
10721 struct bpf_insn *insn;
10725 env->prev_insn_idx = prev_insn_idx;
10726 if (env->insn_idx >= insn_cnt) {
10727 verbose(env, "invalid insn idx %d insn_cnt %d\n",
10728 env->insn_idx, insn_cnt);
10732 insn = &insns[env->insn_idx];
10733 class = BPF_CLASS(insn->code);
10735 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
10737 "BPF program is too large. Processed %d insn\n",
10738 env->insn_processed);
10742 err = is_state_visited(env, env->insn_idx);
10746 /* found equivalent state, can prune the search */
10747 if (env->log.level & BPF_LOG_LEVEL) {
10748 if (do_print_state)
10749 verbose(env, "\nfrom %d to %d%s: safe\n",
10750 env->prev_insn_idx, env->insn_idx,
10751 env->cur_state->speculative ?
10752 " (speculative execution)" : "");
10754 verbose(env, "%d: safe\n", env->insn_idx);
10756 goto process_bpf_exit;
10759 if (signal_pending(current))
10762 if (need_resched())
10765 if (env->log.level & BPF_LOG_LEVEL2 ||
10766 (env->log.level & BPF_LOG_LEVEL && do_print_state)) {
10767 if (env->log.level & BPF_LOG_LEVEL2)
10768 verbose(env, "%d:", env->insn_idx);
10770 verbose(env, "\nfrom %d to %d%s:",
10771 env->prev_insn_idx, env->insn_idx,
10772 env->cur_state->speculative ?
10773 " (speculative execution)" : "");
10774 print_verifier_state(env, state->frame[state->curframe]);
10775 do_print_state = false;
10778 if (env->log.level & BPF_LOG_LEVEL) {
10779 const struct bpf_insn_cbs cbs = {
10780 .cb_call = disasm_kfunc_name,
10781 .cb_print = verbose,
10782 .private_data = env,
10785 verbose_linfo(env, env->insn_idx, "; ");
10786 verbose(env, "%d: ", env->insn_idx);
10787 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
10790 if (bpf_prog_is_dev_bound(env->prog->aux)) {
10791 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
10792 env->prev_insn_idx);
10797 regs = cur_regs(env);
10798 sanitize_mark_insn_seen(env);
10799 prev_insn_idx = env->insn_idx;
10801 if (class == BPF_ALU || class == BPF_ALU64) {
10802 err = check_alu_op(env, insn);
10806 } else if (class == BPF_LDX) {
10807 enum bpf_reg_type *prev_src_type, src_reg_type;
10809 /* check for reserved fields is already done */
10811 /* check src operand */
10812 err = check_reg_arg(env, insn->src_reg, SRC_OP);
10816 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
10820 src_reg_type = regs[insn->src_reg].type;
10822 /* check that memory (src_reg + off) is readable,
10823 * the state of dst_reg will be updated by this func
10825 err = check_mem_access(env, env->insn_idx, insn->src_reg,
10826 insn->off, BPF_SIZE(insn->code),
10827 BPF_READ, insn->dst_reg, false);
10831 prev_src_type = &env->insn_aux_data[env->insn_idx].ptr_type;
10833 if (*prev_src_type == NOT_INIT) {
10834 /* saw a valid insn
10835 * dst_reg = *(u32 *)(src_reg + off)
10836 * save type to validate intersecting paths
10838 *prev_src_type = src_reg_type;
10840 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
10841 /* ABuser program is trying to use the same insn
10842 * dst_reg = *(u32*) (src_reg + off)
10843 * with different pointer types:
10844 * src_reg == ctx in one branch and
10845 * src_reg == stack|map in some other branch.
10848 verbose(env, "same insn cannot be used with different pointers\n");
10852 } else if (class == BPF_STX) {
10853 enum bpf_reg_type *prev_dst_type, dst_reg_type;
10855 if (BPF_MODE(insn->code) == BPF_ATOMIC) {
10856 err = check_atomic(env, env->insn_idx, insn);
10863 if (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0) {
10864 verbose(env, "BPF_STX uses reserved fields\n");
10868 /* check src1 operand */
10869 err = check_reg_arg(env, insn->src_reg, SRC_OP);
10872 /* check src2 operand */
10873 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
10877 dst_reg_type = regs[insn->dst_reg].type;
10879 /* check that memory (dst_reg + off) is writeable */
10880 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
10881 insn->off, BPF_SIZE(insn->code),
10882 BPF_WRITE, insn->src_reg, false);
10886 prev_dst_type = &env->insn_aux_data[env->insn_idx].ptr_type;
10888 if (*prev_dst_type == NOT_INIT) {
10889 *prev_dst_type = dst_reg_type;
10890 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
10891 verbose(env, "same insn cannot be used with different pointers\n");
10895 } else if (class == BPF_ST) {
10896 if (BPF_MODE(insn->code) != BPF_MEM ||
10897 insn->src_reg != BPF_REG_0) {
10898 verbose(env, "BPF_ST uses reserved fields\n");
10901 /* check src operand */
10902 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
10906 if (is_ctx_reg(env, insn->dst_reg)) {
10907 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
10909 reg_type_str[reg_state(env, insn->dst_reg)->type]);
10913 /* check that memory (dst_reg + off) is writeable */
10914 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
10915 insn->off, BPF_SIZE(insn->code),
10916 BPF_WRITE, -1, false);
10920 } else if (class == BPF_JMP || class == BPF_JMP32) {
10921 u8 opcode = BPF_OP(insn->code);
10923 env->jmps_processed++;
10924 if (opcode == BPF_CALL) {
10925 if (BPF_SRC(insn->code) != BPF_K ||
10927 (insn->src_reg != BPF_REG_0 &&
10928 insn->src_reg != BPF_PSEUDO_CALL &&
10929 insn->src_reg != BPF_PSEUDO_KFUNC_CALL) ||
10930 insn->dst_reg != BPF_REG_0 ||
10931 class == BPF_JMP32) {
10932 verbose(env, "BPF_CALL uses reserved fields\n");
10936 if (env->cur_state->active_spin_lock &&
10937 (insn->src_reg == BPF_PSEUDO_CALL ||
10938 insn->imm != BPF_FUNC_spin_unlock)) {
10939 verbose(env, "function calls are not allowed while holding a lock\n");
10942 if (insn->src_reg == BPF_PSEUDO_CALL)
10943 err = check_func_call(env, insn, &env->insn_idx);
10944 else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL)
10945 err = check_kfunc_call(env, insn);
10947 err = check_helper_call(env, insn, &env->insn_idx);
10950 } else if (opcode == BPF_JA) {
10951 if (BPF_SRC(insn->code) != BPF_K ||
10953 insn->src_reg != BPF_REG_0 ||
10954 insn->dst_reg != BPF_REG_0 ||
10955 class == BPF_JMP32) {
10956 verbose(env, "BPF_JA uses reserved fields\n");
10960 env->insn_idx += insn->off + 1;
10963 } else if (opcode == BPF_EXIT) {
10964 if (BPF_SRC(insn->code) != BPF_K ||
10966 insn->src_reg != BPF_REG_0 ||
10967 insn->dst_reg != BPF_REG_0 ||
10968 class == BPF_JMP32) {
10969 verbose(env, "BPF_EXIT uses reserved fields\n");
10973 if (env->cur_state->active_spin_lock) {
10974 verbose(env, "bpf_spin_unlock is missing\n");
10978 if (state->curframe) {
10979 /* exit from nested function */
10980 err = prepare_func_exit(env, &env->insn_idx);
10983 do_print_state = true;
10987 err = check_reference_leak(env);
10991 err = check_return_code(env);
10995 update_branch_counts(env, env->cur_state);
10996 err = pop_stack(env, &prev_insn_idx,
10997 &env->insn_idx, pop_log);
10999 if (err != -ENOENT)
11003 do_print_state = true;
11007 err = check_cond_jmp_op(env, insn, &env->insn_idx);
11011 } else if (class == BPF_LD) {
11012 u8 mode = BPF_MODE(insn->code);
11014 if (mode == BPF_ABS || mode == BPF_IND) {
11015 err = check_ld_abs(env, insn);
11019 } else if (mode == BPF_IMM) {
11020 err = check_ld_imm(env, insn);
11025 sanitize_mark_insn_seen(env);
11027 verbose(env, "invalid BPF_LD mode\n");
11031 verbose(env, "unknown insn class %d\n", class);
11041 static int find_btf_percpu_datasec(struct btf *btf)
11043 const struct btf_type *t;
11048 * Both vmlinux and module each have their own ".data..percpu"
11049 * DATASECs in BTF. So for module's case, we need to skip vmlinux BTF
11050 * types to look at only module's own BTF types.
11052 n = btf_nr_types(btf);
11053 if (btf_is_module(btf))
11054 i = btf_nr_types(btf_vmlinux);
11058 for(; i < n; i++) {
11059 t = btf_type_by_id(btf, i);
11060 if (BTF_INFO_KIND(t->info) != BTF_KIND_DATASEC)
11063 tname = btf_name_by_offset(btf, t->name_off);
11064 if (!strcmp(tname, ".data..percpu"))
11071 /* replace pseudo btf_id with kernel symbol address */
11072 static int check_pseudo_btf_id(struct bpf_verifier_env *env,
11073 struct bpf_insn *insn,
11074 struct bpf_insn_aux_data *aux)
11076 const struct btf_var_secinfo *vsi;
11077 const struct btf_type *datasec;
11078 struct btf_mod_pair *btf_mod;
11079 const struct btf_type *t;
11080 const char *sym_name;
11081 bool percpu = false;
11082 u32 type, id = insn->imm;
11086 int i, btf_fd, err;
11088 btf_fd = insn[1].imm;
11090 btf = btf_get_by_fd(btf_fd);
11092 verbose(env, "invalid module BTF object FD specified.\n");
11096 if (!btf_vmlinux) {
11097 verbose(env, "kernel is missing BTF, make sure CONFIG_DEBUG_INFO_BTF=y is specified in Kconfig.\n");
11104 t = btf_type_by_id(btf, id);
11106 verbose(env, "ldimm64 insn specifies invalid btf_id %d.\n", id);
11111 if (!btf_type_is_var(t)) {
11112 verbose(env, "pseudo btf_id %d in ldimm64 isn't KIND_VAR.\n", id);
11117 sym_name = btf_name_by_offset(btf, t->name_off);
11118 addr = kallsyms_lookup_name(sym_name);
11120 verbose(env, "ldimm64 failed to find the address for kernel symbol '%s'.\n",
11126 datasec_id = find_btf_percpu_datasec(btf);
11127 if (datasec_id > 0) {
11128 datasec = btf_type_by_id(btf, datasec_id);
11129 for_each_vsi(i, datasec, vsi) {
11130 if (vsi->type == id) {
11137 insn[0].imm = (u32)addr;
11138 insn[1].imm = addr >> 32;
11141 t = btf_type_skip_modifiers(btf, type, NULL);
11143 aux->btf_var.reg_type = PTR_TO_PERCPU_BTF_ID;
11144 aux->btf_var.btf = btf;
11145 aux->btf_var.btf_id = type;
11146 } else if (!btf_type_is_struct(t)) {
11147 const struct btf_type *ret;
11151 /* resolve the type size of ksym. */
11152 ret = btf_resolve_size(btf, t, &tsize);
11154 tname = btf_name_by_offset(btf, t->name_off);
11155 verbose(env, "ldimm64 unable to resolve the size of type '%s': %ld\n",
11156 tname, PTR_ERR(ret));
11160 aux->btf_var.reg_type = PTR_TO_MEM;
11161 aux->btf_var.mem_size = tsize;
11163 aux->btf_var.reg_type = PTR_TO_BTF_ID;
11164 aux->btf_var.btf = btf;
11165 aux->btf_var.btf_id = type;
11168 /* check whether we recorded this BTF (and maybe module) already */
11169 for (i = 0; i < env->used_btf_cnt; i++) {
11170 if (env->used_btfs[i].btf == btf) {
11176 if (env->used_btf_cnt >= MAX_USED_BTFS) {
11181 btf_mod = &env->used_btfs[env->used_btf_cnt];
11182 btf_mod->btf = btf;
11183 btf_mod->module = NULL;
11185 /* if we reference variables from kernel module, bump its refcount */
11186 if (btf_is_module(btf)) {
11187 btf_mod->module = btf_try_get_module(btf);
11188 if (!btf_mod->module) {
11194 env->used_btf_cnt++;
11202 static int check_map_prealloc(struct bpf_map *map)
11204 return (map->map_type != BPF_MAP_TYPE_HASH &&
11205 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
11206 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
11207 !(map->map_flags & BPF_F_NO_PREALLOC);
11210 static bool is_tracing_prog_type(enum bpf_prog_type type)
11213 case BPF_PROG_TYPE_KPROBE:
11214 case BPF_PROG_TYPE_TRACEPOINT:
11215 case BPF_PROG_TYPE_PERF_EVENT:
11216 case BPF_PROG_TYPE_RAW_TRACEPOINT:
11223 static bool is_preallocated_map(struct bpf_map *map)
11225 if (!check_map_prealloc(map))
11227 if (map->inner_map_meta && !check_map_prealloc(map->inner_map_meta))
11232 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
11233 struct bpf_map *map,
11234 struct bpf_prog *prog)
11237 enum bpf_prog_type prog_type = resolve_prog_type(prog);
11239 * Validate that trace type programs use preallocated hash maps.
11241 * For programs attached to PERF events this is mandatory as the
11242 * perf NMI can hit any arbitrary code sequence.
11244 * All other trace types using preallocated hash maps are unsafe as
11245 * well because tracepoint or kprobes can be inside locked regions
11246 * of the memory allocator or at a place where a recursion into the
11247 * memory allocator would see inconsistent state.
11249 * On RT enabled kernels run-time allocation of all trace type
11250 * programs is strictly prohibited due to lock type constraints. On
11251 * !RT kernels it is allowed for backwards compatibility reasons for
11252 * now, but warnings are emitted so developers are made aware of
11253 * the unsafety and can fix their programs before this is enforced.
11255 if (is_tracing_prog_type(prog_type) && !is_preallocated_map(map)) {
11256 if (prog_type == BPF_PROG_TYPE_PERF_EVENT) {
11257 verbose(env, "perf_event programs can only use preallocated hash map\n");
11260 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
11261 verbose(env, "trace type programs can only use preallocated hash map\n");
11264 WARN_ONCE(1, "trace type BPF program uses run-time allocation\n");
11265 verbose(env, "trace type programs with run-time allocated hash maps are unsafe. Switch to preallocated hash maps.\n");
11268 if (map_value_has_spin_lock(map)) {
11269 if (prog_type == BPF_PROG_TYPE_SOCKET_FILTER) {
11270 verbose(env, "socket filter progs cannot use bpf_spin_lock yet\n");
11274 if (is_tracing_prog_type(prog_type)) {
11275 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
11279 if (prog->aux->sleepable) {
11280 verbose(env, "sleepable progs cannot use bpf_spin_lock yet\n");
11285 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
11286 !bpf_offload_prog_map_match(prog, map)) {
11287 verbose(env, "offload device mismatch between prog and map\n");
11291 if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
11292 verbose(env, "bpf_struct_ops map cannot be used in prog\n");
11296 if (prog->aux->sleepable)
11297 switch (map->map_type) {
11298 case BPF_MAP_TYPE_HASH:
11299 case BPF_MAP_TYPE_LRU_HASH:
11300 case BPF_MAP_TYPE_ARRAY:
11301 case BPF_MAP_TYPE_PERCPU_HASH:
11302 case BPF_MAP_TYPE_PERCPU_ARRAY:
11303 case BPF_MAP_TYPE_LRU_PERCPU_HASH:
11304 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
11305 case BPF_MAP_TYPE_HASH_OF_MAPS:
11306 if (!is_preallocated_map(map)) {
11308 "Sleepable programs can only use preallocated maps\n");
11312 case BPF_MAP_TYPE_RINGBUF:
11316 "Sleepable programs can only use array, hash, and ringbuf maps\n");
11323 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
11325 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
11326 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
11329 /* find and rewrite pseudo imm in ld_imm64 instructions:
11331 * 1. if it accesses map FD, replace it with actual map pointer.
11332 * 2. if it accesses btf_id of a VAR, replace it with pointer to the var.
11334 * NOTE: btf_vmlinux is required for converting pseudo btf_id.
11336 static int resolve_pseudo_ldimm64(struct bpf_verifier_env *env)
11338 struct bpf_insn *insn = env->prog->insnsi;
11339 int insn_cnt = env->prog->len;
11342 err = bpf_prog_calc_tag(env->prog);
11346 for (i = 0; i < insn_cnt; i++, insn++) {
11347 if (BPF_CLASS(insn->code) == BPF_LDX &&
11348 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
11349 verbose(env, "BPF_LDX uses reserved fields\n");
11353 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
11354 struct bpf_insn_aux_data *aux;
11355 struct bpf_map *map;
11360 if (i == insn_cnt - 1 || insn[1].code != 0 ||
11361 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
11362 insn[1].off != 0) {
11363 verbose(env, "invalid bpf_ld_imm64 insn\n");
11367 if (insn[0].src_reg == 0)
11368 /* valid generic load 64-bit imm */
11371 if (insn[0].src_reg == BPF_PSEUDO_BTF_ID) {
11372 aux = &env->insn_aux_data[i];
11373 err = check_pseudo_btf_id(env, insn, aux);
11379 if (insn[0].src_reg == BPF_PSEUDO_FUNC) {
11380 aux = &env->insn_aux_data[i];
11381 aux->ptr_type = PTR_TO_FUNC;
11385 /* In final convert_pseudo_ld_imm64() step, this is
11386 * converted into regular 64-bit imm load insn.
11388 switch (insn[0].src_reg) {
11389 case BPF_PSEUDO_MAP_VALUE:
11390 case BPF_PSEUDO_MAP_IDX_VALUE:
11392 case BPF_PSEUDO_MAP_FD:
11393 case BPF_PSEUDO_MAP_IDX:
11394 if (insn[1].imm == 0)
11398 verbose(env, "unrecognized bpf_ld_imm64 insn\n");
11402 switch (insn[0].src_reg) {
11403 case BPF_PSEUDO_MAP_IDX_VALUE:
11404 case BPF_PSEUDO_MAP_IDX:
11405 if (bpfptr_is_null(env->fd_array)) {
11406 verbose(env, "fd_idx without fd_array is invalid\n");
11409 if (copy_from_bpfptr_offset(&fd, env->fd_array,
11410 insn[0].imm * sizeof(fd),
11420 map = __bpf_map_get(f);
11422 verbose(env, "fd %d is not pointing to valid bpf_map\n",
11424 return PTR_ERR(map);
11427 err = check_map_prog_compatibility(env, map, env->prog);
11433 aux = &env->insn_aux_data[i];
11434 if (insn[0].src_reg == BPF_PSEUDO_MAP_FD ||
11435 insn[0].src_reg == BPF_PSEUDO_MAP_IDX) {
11436 addr = (unsigned long)map;
11438 u32 off = insn[1].imm;
11440 if (off >= BPF_MAX_VAR_OFF) {
11441 verbose(env, "direct value offset of %u is not allowed\n", off);
11446 if (!map->ops->map_direct_value_addr) {
11447 verbose(env, "no direct value access support for this map type\n");
11452 err = map->ops->map_direct_value_addr(map, &addr, off);
11454 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
11455 map->value_size, off);
11460 aux->map_off = off;
11464 insn[0].imm = (u32)addr;
11465 insn[1].imm = addr >> 32;
11467 /* check whether we recorded this map already */
11468 for (j = 0; j < env->used_map_cnt; j++) {
11469 if (env->used_maps[j] == map) {
11470 aux->map_index = j;
11476 if (env->used_map_cnt >= MAX_USED_MAPS) {
11481 /* hold the map. If the program is rejected by verifier,
11482 * the map will be released by release_maps() or it
11483 * will be used by the valid program until it's unloaded
11484 * and all maps are released in free_used_maps()
11488 aux->map_index = env->used_map_cnt;
11489 env->used_maps[env->used_map_cnt++] = map;
11491 if (bpf_map_is_cgroup_storage(map) &&
11492 bpf_cgroup_storage_assign(env->prog->aux, map)) {
11493 verbose(env, "only one cgroup storage of each type is allowed\n");
11505 /* Basic sanity check before we invest more work here. */
11506 if (!bpf_opcode_in_insntable(insn->code)) {
11507 verbose(env, "unknown opcode %02x\n", insn->code);
11512 /* now all pseudo BPF_LD_IMM64 instructions load valid
11513 * 'struct bpf_map *' into a register instead of user map_fd.
11514 * These pointers will be used later by verifier to validate map access.
11519 /* drop refcnt of maps used by the rejected program */
11520 static void release_maps(struct bpf_verifier_env *env)
11522 __bpf_free_used_maps(env->prog->aux, env->used_maps,
11523 env->used_map_cnt);
11526 /* drop refcnt of maps used by the rejected program */
11527 static void release_btfs(struct bpf_verifier_env *env)
11529 __bpf_free_used_btfs(env->prog->aux, env->used_btfs,
11530 env->used_btf_cnt);
11533 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
11534 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
11536 struct bpf_insn *insn = env->prog->insnsi;
11537 int insn_cnt = env->prog->len;
11540 for (i = 0; i < insn_cnt; i++, insn++) {
11541 if (insn->code != (BPF_LD | BPF_IMM | BPF_DW))
11543 if (insn->src_reg == BPF_PSEUDO_FUNC)
11549 /* single env->prog->insni[off] instruction was replaced with the range
11550 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
11551 * [0, off) and [off, end) to new locations, so the patched range stays zero
11553 static void adjust_insn_aux_data(struct bpf_verifier_env *env,
11554 struct bpf_insn_aux_data *new_data,
11555 struct bpf_prog *new_prog, u32 off, u32 cnt)
11557 struct bpf_insn_aux_data *old_data = env->insn_aux_data;
11558 struct bpf_insn *insn = new_prog->insnsi;
11559 u32 old_seen = old_data[off].seen;
11563 /* aux info at OFF always needs adjustment, no matter fast path
11564 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
11565 * original insn at old prog.
11567 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
11571 prog_len = new_prog->len;
11573 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
11574 memcpy(new_data + off + cnt - 1, old_data + off,
11575 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
11576 for (i = off; i < off + cnt - 1; i++) {
11577 /* Expand insni[off]'s seen count to the patched range. */
11578 new_data[i].seen = old_seen;
11579 new_data[i].zext_dst = insn_has_def32(env, insn + i);
11581 env->insn_aux_data = new_data;
11585 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
11591 /* NOTE: fake 'exit' subprog should be updated as well. */
11592 for (i = 0; i <= env->subprog_cnt; i++) {
11593 if (env->subprog_info[i].start <= off)
11595 env->subprog_info[i].start += len - 1;
11599 static void adjust_poke_descs(struct bpf_prog *prog, u32 off, u32 len)
11601 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
11602 int i, sz = prog->aux->size_poke_tab;
11603 struct bpf_jit_poke_descriptor *desc;
11605 for (i = 0; i < sz; i++) {
11607 if (desc->insn_idx <= off)
11609 desc->insn_idx += len - 1;
11613 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
11614 const struct bpf_insn *patch, u32 len)
11616 struct bpf_prog *new_prog;
11617 struct bpf_insn_aux_data *new_data = NULL;
11620 new_data = vzalloc(array_size(env->prog->len + len - 1,
11621 sizeof(struct bpf_insn_aux_data)));
11626 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
11627 if (IS_ERR(new_prog)) {
11628 if (PTR_ERR(new_prog) == -ERANGE)
11630 "insn %d cannot be patched due to 16-bit range\n",
11631 env->insn_aux_data[off].orig_idx);
11635 adjust_insn_aux_data(env, new_data, new_prog, off, len);
11636 adjust_subprog_starts(env, off, len);
11637 adjust_poke_descs(new_prog, off, len);
11641 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
11646 /* find first prog starting at or after off (first to remove) */
11647 for (i = 0; i < env->subprog_cnt; i++)
11648 if (env->subprog_info[i].start >= off)
11650 /* find first prog starting at or after off + cnt (first to stay) */
11651 for (j = i; j < env->subprog_cnt; j++)
11652 if (env->subprog_info[j].start >= off + cnt)
11654 /* if j doesn't start exactly at off + cnt, we are just removing
11655 * the front of previous prog
11657 if (env->subprog_info[j].start != off + cnt)
11661 struct bpf_prog_aux *aux = env->prog->aux;
11664 /* move fake 'exit' subprog as well */
11665 move = env->subprog_cnt + 1 - j;
11667 memmove(env->subprog_info + i,
11668 env->subprog_info + j,
11669 sizeof(*env->subprog_info) * move);
11670 env->subprog_cnt -= j - i;
11672 /* remove func_info */
11673 if (aux->func_info) {
11674 move = aux->func_info_cnt - j;
11676 memmove(aux->func_info + i,
11677 aux->func_info + j,
11678 sizeof(*aux->func_info) * move);
11679 aux->func_info_cnt -= j - i;
11680 /* func_info->insn_off is set after all code rewrites,
11681 * in adjust_btf_func() - no need to adjust
11685 /* convert i from "first prog to remove" to "first to adjust" */
11686 if (env->subprog_info[i].start == off)
11690 /* update fake 'exit' subprog as well */
11691 for (; i <= env->subprog_cnt; i++)
11692 env->subprog_info[i].start -= cnt;
11697 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
11700 struct bpf_prog *prog = env->prog;
11701 u32 i, l_off, l_cnt, nr_linfo;
11702 struct bpf_line_info *linfo;
11704 nr_linfo = prog->aux->nr_linfo;
11708 linfo = prog->aux->linfo;
11710 /* find first line info to remove, count lines to be removed */
11711 for (i = 0; i < nr_linfo; i++)
11712 if (linfo[i].insn_off >= off)
11717 for (; i < nr_linfo; i++)
11718 if (linfo[i].insn_off < off + cnt)
11723 /* First live insn doesn't match first live linfo, it needs to "inherit"
11724 * last removed linfo. prog is already modified, so prog->len == off
11725 * means no live instructions after (tail of the program was removed).
11727 if (prog->len != off && l_cnt &&
11728 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
11730 linfo[--i].insn_off = off + cnt;
11733 /* remove the line info which refer to the removed instructions */
11735 memmove(linfo + l_off, linfo + i,
11736 sizeof(*linfo) * (nr_linfo - i));
11738 prog->aux->nr_linfo -= l_cnt;
11739 nr_linfo = prog->aux->nr_linfo;
11742 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
11743 for (i = l_off; i < nr_linfo; i++)
11744 linfo[i].insn_off -= cnt;
11746 /* fix up all subprogs (incl. 'exit') which start >= off */
11747 for (i = 0; i <= env->subprog_cnt; i++)
11748 if (env->subprog_info[i].linfo_idx > l_off) {
11749 /* program may have started in the removed region but
11750 * may not be fully removed
11752 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
11753 env->subprog_info[i].linfo_idx -= l_cnt;
11755 env->subprog_info[i].linfo_idx = l_off;
11761 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
11763 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
11764 unsigned int orig_prog_len = env->prog->len;
11767 if (bpf_prog_is_dev_bound(env->prog->aux))
11768 bpf_prog_offload_remove_insns(env, off, cnt);
11770 err = bpf_remove_insns(env->prog, off, cnt);
11774 err = adjust_subprog_starts_after_remove(env, off, cnt);
11778 err = bpf_adj_linfo_after_remove(env, off, cnt);
11782 memmove(aux_data + off, aux_data + off + cnt,
11783 sizeof(*aux_data) * (orig_prog_len - off - cnt));
11788 /* The verifier does more data flow analysis than llvm and will not
11789 * explore branches that are dead at run time. Malicious programs can
11790 * have dead code too. Therefore replace all dead at-run-time code
11793 * Just nops are not optimal, e.g. if they would sit at the end of the
11794 * program and through another bug we would manage to jump there, then
11795 * we'd execute beyond program memory otherwise. Returning exception
11796 * code also wouldn't work since we can have subprogs where the dead
11797 * code could be located.
11799 static void sanitize_dead_code(struct bpf_verifier_env *env)
11801 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
11802 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
11803 struct bpf_insn *insn = env->prog->insnsi;
11804 const int insn_cnt = env->prog->len;
11807 for (i = 0; i < insn_cnt; i++) {
11808 if (aux_data[i].seen)
11810 memcpy(insn + i, &trap, sizeof(trap));
11814 static bool insn_is_cond_jump(u8 code)
11818 if (BPF_CLASS(code) == BPF_JMP32)
11821 if (BPF_CLASS(code) != BPF_JMP)
11825 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
11828 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
11830 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
11831 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
11832 struct bpf_insn *insn = env->prog->insnsi;
11833 const int insn_cnt = env->prog->len;
11836 for (i = 0; i < insn_cnt; i++, insn++) {
11837 if (!insn_is_cond_jump(insn->code))
11840 if (!aux_data[i + 1].seen)
11841 ja.off = insn->off;
11842 else if (!aux_data[i + 1 + insn->off].seen)
11847 if (bpf_prog_is_dev_bound(env->prog->aux))
11848 bpf_prog_offload_replace_insn(env, i, &ja);
11850 memcpy(insn, &ja, sizeof(ja));
11854 static int opt_remove_dead_code(struct bpf_verifier_env *env)
11856 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
11857 int insn_cnt = env->prog->len;
11860 for (i = 0; i < insn_cnt; i++) {
11864 while (i + j < insn_cnt && !aux_data[i + j].seen)
11869 err = verifier_remove_insns(env, i, j);
11872 insn_cnt = env->prog->len;
11878 static int opt_remove_nops(struct bpf_verifier_env *env)
11880 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
11881 struct bpf_insn *insn = env->prog->insnsi;
11882 int insn_cnt = env->prog->len;
11885 for (i = 0; i < insn_cnt; i++) {
11886 if (memcmp(&insn[i], &ja, sizeof(ja)))
11889 err = verifier_remove_insns(env, i, 1);
11899 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
11900 const union bpf_attr *attr)
11902 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
11903 struct bpf_insn_aux_data *aux = env->insn_aux_data;
11904 int i, patch_len, delta = 0, len = env->prog->len;
11905 struct bpf_insn *insns = env->prog->insnsi;
11906 struct bpf_prog *new_prog;
11909 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
11910 zext_patch[1] = BPF_ZEXT_REG(0);
11911 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
11912 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
11913 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
11914 for (i = 0; i < len; i++) {
11915 int adj_idx = i + delta;
11916 struct bpf_insn insn;
11919 insn = insns[adj_idx];
11920 load_reg = insn_def_regno(&insn);
11921 if (!aux[adj_idx].zext_dst) {
11929 class = BPF_CLASS(code);
11930 if (load_reg == -1)
11933 /* NOTE: arg "reg" (the fourth one) is only used for
11934 * BPF_STX + SRC_OP, so it is safe to pass NULL
11937 if (is_reg64(env, &insn, load_reg, NULL, DST_OP)) {
11938 if (class == BPF_LD &&
11939 BPF_MODE(code) == BPF_IMM)
11944 /* ctx load could be transformed into wider load. */
11945 if (class == BPF_LDX &&
11946 aux[adj_idx].ptr_type == PTR_TO_CTX)
11949 imm_rnd = get_random_int();
11950 rnd_hi32_patch[0] = insn;
11951 rnd_hi32_patch[1].imm = imm_rnd;
11952 rnd_hi32_patch[3].dst_reg = load_reg;
11953 patch = rnd_hi32_patch;
11955 goto apply_patch_buffer;
11958 /* Add in an zero-extend instruction if a) the JIT has requested
11959 * it or b) it's a CMPXCHG.
11961 * The latter is because: BPF_CMPXCHG always loads a value into
11962 * R0, therefore always zero-extends. However some archs'
11963 * equivalent instruction only does this load when the
11964 * comparison is successful. This detail of CMPXCHG is
11965 * orthogonal to the general zero-extension behaviour of the
11966 * CPU, so it's treated independently of bpf_jit_needs_zext.
11968 if (!bpf_jit_needs_zext() && !is_cmpxchg_insn(&insn))
11971 if (WARN_ON(load_reg == -1)) {
11972 verbose(env, "verifier bug. zext_dst is set, but no reg is defined\n");
11976 zext_patch[0] = insn;
11977 zext_patch[1].dst_reg = load_reg;
11978 zext_patch[1].src_reg = load_reg;
11979 patch = zext_patch;
11981 apply_patch_buffer:
11982 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
11985 env->prog = new_prog;
11986 insns = new_prog->insnsi;
11987 aux = env->insn_aux_data;
11988 delta += patch_len - 1;
11994 /* convert load instructions that access fields of a context type into a
11995 * sequence of instructions that access fields of the underlying structure:
11996 * struct __sk_buff -> struct sk_buff
11997 * struct bpf_sock_ops -> struct sock
11999 static int convert_ctx_accesses(struct bpf_verifier_env *env)
12001 const struct bpf_verifier_ops *ops = env->ops;
12002 int i, cnt, size, ctx_field_size, delta = 0;
12003 const int insn_cnt = env->prog->len;
12004 struct bpf_insn insn_buf[16], *insn;
12005 u32 target_size, size_default, off;
12006 struct bpf_prog *new_prog;
12007 enum bpf_access_type type;
12008 bool is_narrower_load;
12010 if (ops->gen_prologue || env->seen_direct_write) {
12011 if (!ops->gen_prologue) {
12012 verbose(env, "bpf verifier is misconfigured\n");
12015 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
12017 if (cnt >= ARRAY_SIZE(insn_buf)) {
12018 verbose(env, "bpf verifier is misconfigured\n");
12021 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
12025 env->prog = new_prog;
12030 if (bpf_prog_is_dev_bound(env->prog->aux))
12033 insn = env->prog->insnsi + delta;
12035 for (i = 0; i < insn_cnt; i++, insn++) {
12036 bpf_convert_ctx_access_t convert_ctx_access;
12038 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
12039 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
12040 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
12041 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
12043 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
12044 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
12045 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
12046 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
12051 if (type == BPF_WRITE &&
12052 env->insn_aux_data[i + delta].sanitize_stack_off) {
12053 struct bpf_insn patch[] = {
12054 /* Sanitize suspicious stack slot with zero.
12055 * There are no memory dependencies for this store,
12056 * since it's only using frame pointer and immediate
12059 BPF_ST_MEM(BPF_DW, BPF_REG_FP,
12060 env->insn_aux_data[i + delta].sanitize_stack_off,
12062 /* the original STX instruction will immediately
12063 * overwrite the same stack slot with appropriate value
12068 cnt = ARRAY_SIZE(patch);
12069 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
12074 env->prog = new_prog;
12075 insn = new_prog->insnsi + i + delta;
12079 switch (env->insn_aux_data[i + delta].ptr_type) {
12081 if (!ops->convert_ctx_access)
12083 convert_ctx_access = ops->convert_ctx_access;
12085 case PTR_TO_SOCKET:
12086 case PTR_TO_SOCK_COMMON:
12087 convert_ctx_access = bpf_sock_convert_ctx_access;
12089 case PTR_TO_TCP_SOCK:
12090 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
12092 case PTR_TO_XDP_SOCK:
12093 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
12095 case PTR_TO_BTF_ID:
12096 if (type == BPF_READ) {
12097 insn->code = BPF_LDX | BPF_PROBE_MEM |
12098 BPF_SIZE((insn)->code);
12099 env->prog->aux->num_exentries++;
12100 } else if (resolve_prog_type(env->prog) != BPF_PROG_TYPE_STRUCT_OPS) {
12101 verbose(env, "Writes through BTF pointers are not allowed\n");
12109 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
12110 size = BPF_LDST_BYTES(insn);
12112 /* If the read access is a narrower load of the field,
12113 * convert to a 4/8-byte load, to minimum program type specific
12114 * convert_ctx_access changes. If conversion is successful,
12115 * we will apply proper mask to the result.
12117 is_narrower_load = size < ctx_field_size;
12118 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
12120 if (is_narrower_load) {
12123 if (type == BPF_WRITE) {
12124 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
12129 if (ctx_field_size == 4)
12131 else if (ctx_field_size == 8)
12132 size_code = BPF_DW;
12134 insn->off = off & ~(size_default - 1);
12135 insn->code = BPF_LDX | BPF_MEM | size_code;
12139 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
12141 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
12142 (ctx_field_size && !target_size)) {
12143 verbose(env, "bpf verifier is misconfigured\n");
12147 if (is_narrower_load && size < target_size) {
12148 u8 shift = bpf_ctx_narrow_access_offset(
12149 off, size, size_default) * 8;
12150 if (ctx_field_size <= 4) {
12152 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
12155 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
12156 (1 << size * 8) - 1);
12159 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
12162 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
12163 (1ULL << size * 8) - 1);
12167 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12173 /* keep walking new program and skip insns we just inserted */
12174 env->prog = new_prog;
12175 insn = new_prog->insnsi + i + delta;
12181 static int jit_subprogs(struct bpf_verifier_env *env)
12183 struct bpf_prog *prog = env->prog, **func, *tmp;
12184 int i, j, subprog_start, subprog_end = 0, len, subprog;
12185 struct bpf_map *map_ptr;
12186 struct bpf_insn *insn;
12187 void *old_bpf_func;
12188 int err, num_exentries;
12190 if (env->subprog_cnt <= 1)
12193 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
12194 if (bpf_pseudo_func(insn)) {
12195 env->insn_aux_data[i].call_imm = insn->imm;
12196 /* subprog is encoded in insn[1].imm */
12200 if (!bpf_pseudo_call(insn))
12202 /* Upon error here we cannot fall back to interpreter but
12203 * need a hard reject of the program. Thus -EFAULT is
12204 * propagated in any case.
12206 subprog = find_subprog(env, i + insn->imm + 1);
12208 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
12209 i + insn->imm + 1);
12212 /* temporarily remember subprog id inside insn instead of
12213 * aux_data, since next loop will split up all insns into funcs
12215 insn->off = subprog;
12216 /* remember original imm in case JIT fails and fallback
12217 * to interpreter will be needed
12219 env->insn_aux_data[i].call_imm = insn->imm;
12220 /* point imm to __bpf_call_base+1 from JITs point of view */
12224 err = bpf_prog_alloc_jited_linfo(prog);
12226 goto out_undo_insn;
12229 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
12231 goto out_undo_insn;
12233 for (i = 0; i < env->subprog_cnt; i++) {
12234 subprog_start = subprog_end;
12235 subprog_end = env->subprog_info[i + 1].start;
12237 len = subprog_end - subprog_start;
12238 /* BPF_PROG_RUN doesn't call subprogs directly,
12239 * hence main prog stats include the runtime of subprogs.
12240 * subprogs don't have IDs and not reachable via prog_get_next_id
12241 * func[i]->stats will never be accessed and stays NULL
12243 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
12246 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
12247 len * sizeof(struct bpf_insn));
12248 func[i]->type = prog->type;
12249 func[i]->len = len;
12250 if (bpf_prog_calc_tag(func[i]))
12252 func[i]->is_func = 1;
12253 func[i]->aux->func_idx = i;
12254 /* the btf and func_info will be freed only at prog->aux */
12255 func[i]->aux->btf = prog->aux->btf;
12256 func[i]->aux->func_info = prog->aux->func_info;
12258 for (j = 0; j < prog->aux->size_poke_tab; j++) {
12259 u32 insn_idx = prog->aux->poke_tab[j].insn_idx;
12262 if (!(insn_idx >= subprog_start &&
12263 insn_idx <= subprog_end))
12266 ret = bpf_jit_add_poke_descriptor(func[i],
12267 &prog->aux->poke_tab[j]);
12269 verbose(env, "adding tail call poke descriptor failed\n");
12273 func[i]->insnsi[insn_idx - subprog_start].imm = ret + 1;
12275 map_ptr = func[i]->aux->poke_tab[ret].tail_call.map;
12276 ret = map_ptr->ops->map_poke_track(map_ptr, func[i]->aux);
12278 verbose(env, "tracking tail call prog failed\n");
12283 /* Use bpf_prog_F_tag to indicate functions in stack traces.
12284 * Long term would need debug info to populate names
12286 func[i]->aux->name[0] = 'F';
12287 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
12288 func[i]->jit_requested = 1;
12289 func[i]->aux->kfunc_tab = prog->aux->kfunc_tab;
12290 func[i]->aux->linfo = prog->aux->linfo;
12291 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
12292 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
12293 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
12295 insn = func[i]->insnsi;
12296 for (j = 0; j < func[i]->len; j++, insn++) {
12297 if (BPF_CLASS(insn->code) == BPF_LDX &&
12298 BPF_MODE(insn->code) == BPF_PROBE_MEM)
12301 func[i]->aux->num_exentries = num_exentries;
12302 func[i]->aux->tail_call_reachable = env->subprog_info[i].tail_call_reachable;
12303 func[i] = bpf_int_jit_compile(func[i]);
12304 if (!func[i]->jited) {
12311 /* Untrack main program's aux structs so that during map_poke_run()
12312 * we will not stumble upon the unfilled poke descriptors; each
12313 * of the main program's poke descs got distributed across subprogs
12314 * and got tracked onto map, so we are sure that none of them will
12315 * be missed after the operation below
12317 for (i = 0; i < prog->aux->size_poke_tab; i++) {
12318 map_ptr = prog->aux->poke_tab[i].tail_call.map;
12320 map_ptr->ops->map_poke_untrack(map_ptr, prog->aux);
12323 /* at this point all bpf functions were successfully JITed
12324 * now populate all bpf_calls with correct addresses and
12325 * run last pass of JIT
12327 for (i = 0; i < env->subprog_cnt; i++) {
12328 insn = func[i]->insnsi;
12329 for (j = 0; j < func[i]->len; j++, insn++) {
12330 if (bpf_pseudo_func(insn)) {
12331 subprog = insn[1].imm;
12332 insn[0].imm = (u32)(long)func[subprog]->bpf_func;
12333 insn[1].imm = ((u64)(long)func[subprog]->bpf_func) >> 32;
12336 if (!bpf_pseudo_call(insn))
12338 subprog = insn->off;
12339 insn->imm = BPF_CAST_CALL(func[subprog]->bpf_func) -
12343 /* we use the aux data to keep a list of the start addresses
12344 * of the JITed images for each function in the program
12346 * for some architectures, such as powerpc64, the imm field
12347 * might not be large enough to hold the offset of the start
12348 * address of the callee's JITed image from __bpf_call_base
12350 * in such cases, we can lookup the start address of a callee
12351 * by using its subprog id, available from the off field of
12352 * the call instruction, as an index for this list
12354 func[i]->aux->func = func;
12355 func[i]->aux->func_cnt = env->subprog_cnt;
12357 for (i = 0; i < env->subprog_cnt; i++) {
12358 old_bpf_func = func[i]->bpf_func;
12359 tmp = bpf_int_jit_compile(func[i]);
12360 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
12361 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
12368 /* finally lock prog and jit images for all functions and
12369 * populate kallsysm
12371 for (i = 0; i < env->subprog_cnt; i++) {
12372 bpf_prog_lock_ro(func[i]);
12373 bpf_prog_kallsyms_add(func[i]);
12376 /* Last step: make now unused interpreter insns from main
12377 * prog consistent for later dump requests, so they can
12378 * later look the same as if they were interpreted only.
12380 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
12381 if (bpf_pseudo_func(insn)) {
12382 insn[0].imm = env->insn_aux_data[i].call_imm;
12383 insn[1].imm = find_subprog(env, i + insn[0].imm + 1);
12386 if (!bpf_pseudo_call(insn))
12388 insn->off = env->insn_aux_data[i].call_imm;
12389 subprog = find_subprog(env, i + insn->off + 1);
12390 insn->imm = subprog;
12394 prog->bpf_func = func[0]->bpf_func;
12395 prog->aux->func = func;
12396 prog->aux->func_cnt = env->subprog_cnt;
12397 bpf_prog_jit_attempt_done(prog);
12400 for (i = 0; i < env->subprog_cnt; i++) {
12404 for (j = 0; j < func[i]->aux->size_poke_tab; j++) {
12405 map_ptr = func[i]->aux->poke_tab[j].tail_call.map;
12406 map_ptr->ops->map_poke_untrack(map_ptr, func[i]->aux);
12408 bpf_jit_free(func[i]);
12412 /* cleanup main prog to be interpreted */
12413 prog->jit_requested = 0;
12414 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
12415 if (!bpf_pseudo_call(insn))
12418 insn->imm = env->insn_aux_data[i].call_imm;
12420 bpf_prog_jit_attempt_done(prog);
12424 static int fixup_call_args(struct bpf_verifier_env *env)
12426 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
12427 struct bpf_prog *prog = env->prog;
12428 struct bpf_insn *insn = prog->insnsi;
12429 bool has_kfunc_call = bpf_prog_has_kfunc_call(prog);
12434 if (env->prog->jit_requested &&
12435 !bpf_prog_is_dev_bound(env->prog->aux)) {
12436 err = jit_subprogs(env);
12439 if (err == -EFAULT)
12442 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
12443 if (has_kfunc_call) {
12444 verbose(env, "calling kernel functions are not allowed in non-JITed programs\n");
12447 if (env->subprog_cnt > 1 && env->prog->aux->tail_call_reachable) {
12448 /* When JIT fails the progs with bpf2bpf calls and tail_calls
12449 * have to be rejected, since interpreter doesn't support them yet.
12451 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
12454 for (i = 0; i < prog->len; i++, insn++) {
12455 if (bpf_pseudo_func(insn)) {
12456 /* When JIT fails the progs with callback calls
12457 * have to be rejected, since interpreter doesn't support them yet.
12459 verbose(env, "callbacks are not allowed in non-JITed programs\n");
12463 if (!bpf_pseudo_call(insn))
12465 depth = get_callee_stack_depth(env, insn, i);
12468 bpf_patch_call_args(insn, depth);
12475 static int fixup_kfunc_call(struct bpf_verifier_env *env,
12476 struct bpf_insn *insn)
12478 const struct bpf_kfunc_desc *desc;
12480 /* insn->imm has the btf func_id. Replace it with
12481 * an address (relative to __bpf_base_call).
12483 desc = find_kfunc_desc(env->prog, insn->imm);
12485 verbose(env, "verifier internal error: kernel function descriptor not found for func_id %u\n",
12490 insn->imm = desc->imm;
12495 /* Do various post-verification rewrites in a single program pass.
12496 * These rewrites simplify JIT and interpreter implementations.
12498 static int do_misc_fixups(struct bpf_verifier_env *env)
12500 struct bpf_prog *prog = env->prog;
12501 bool expect_blinding = bpf_jit_blinding_enabled(prog);
12502 struct bpf_insn *insn = prog->insnsi;
12503 const struct bpf_func_proto *fn;
12504 const int insn_cnt = prog->len;
12505 const struct bpf_map_ops *ops;
12506 struct bpf_insn_aux_data *aux;
12507 struct bpf_insn insn_buf[16];
12508 struct bpf_prog *new_prog;
12509 struct bpf_map *map_ptr;
12510 int i, ret, cnt, delta = 0;
12512 for (i = 0; i < insn_cnt; i++, insn++) {
12513 /* Make divide-by-zero exceptions impossible. */
12514 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
12515 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
12516 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
12517 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
12518 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
12519 bool isdiv = BPF_OP(insn->code) == BPF_DIV;
12520 struct bpf_insn *patchlet;
12521 struct bpf_insn chk_and_div[] = {
12522 /* [R,W]x div 0 -> 0 */
12523 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
12524 BPF_JNE | BPF_K, insn->src_reg,
12526 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
12527 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
12530 struct bpf_insn chk_and_mod[] = {
12531 /* [R,W]x mod 0 -> [R,W]x */
12532 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
12533 BPF_JEQ | BPF_K, insn->src_reg,
12534 0, 1 + (is64 ? 0 : 1), 0),
12536 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
12537 BPF_MOV32_REG(insn->dst_reg, insn->dst_reg),
12540 patchlet = isdiv ? chk_and_div : chk_and_mod;
12541 cnt = isdiv ? ARRAY_SIZE(chk_and_div) :
12542 ARRAY_SIZE(chk_and_mod) - (is64 ? 2 : 0);
12544 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
12549 env->prog = prog = new_prog;
12550 insn = new_prog->insnsi + i + delta;
12554 /* Implement LD_ABS and LD_IND with a rewrite, if supported by the program type. */
12555 if (BPF_CLASS(insn->code) == BPF_LD &&
12556 (BPF_MODE(insn->code) == BPF_ABS ||
12557 BPF_MODE(insn->code) == BPF_IND)) {
12558 cnt = env->ops->gen_ld_abs(insn, insn_buf);
12559 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
12560 verbose(env, "bpf verifier is misconfigured\n");
12564 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12569 env->prog = prog = new_prog;
12570 insn = new_prog->insnsi + i + delta;
12574 /* Rewrite pointer arithmetic to mitigate speculation attacks. */
12575 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
12576 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
12577 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
12578 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
12579 struct bpf_insn *patch = &insn_buf[0];
12580 bool issrc, isneg, isimm;
12583 aux = &env->insn_aux_data[i + delta];
12584 if (!aux->alu_state ||
12585 aux->alu_state == BPF_ALU_NON_POINTER)
12588 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
12589 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
12590 BPF_ALU_SANITIZE_SRC;
12591 isimm = aux->alu_state & BPF_ALU_IMMEDIATE;
12593 off_reg = issrc ? insn->src_reg : insn->dst_reg;
12595 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
12598 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
12599 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
12600 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
12601 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
12602 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
12603 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
12604 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX, off_reg);
12607 *patch++ = BPF_MOV64_REG(insn->dst_reg, insn->src_reg);
12608 insn->src_reg = BPF_REG_AX;
12610 insn->code = insn->code == code_add ?
12611 code_sub : code_add;
12613 if (issrc && isneg && !isimm)
12614 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
12615 cnt = patch - insn_buf;
12617 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12622 env->prog = prog = new_prog;
12623 insn = new_prog->insnsi + i + delta;
12627 if (insn->code != (BPF_JMP | BPF_CALL))
12629 if (insn->src_reg == BPF_PSEUDO_CALL)
12631 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
12632 ret = fixup_kfunc_call(env, insn);
12638 if (insn->imm == BPF_FUNC_get_route_realm)
12639 prog->dst_needed = 1;
12640 if (insn->imm == BPF_FUNC_get_prandom_u32)
12641 bpf_user_rnd_init_once();
12642 if (insn->imm == BPF_FUNC_override_return)
12643 prog->kprobe_override = 1;
12644 if (insn->imm == BPF_FUNC_tail_call) {
12645 /* If we tail call into other programs, we
12646 * cannot make any assumptions since they can
12647 * be replaced dynamically during runtime in
12648 * the program array.
12650 prog->cb_access = 1;
12651 if (!allow_tail_call_in_subprogs(env))
12652 prog->aux->stack_depth = MAX_BPF_STACK;
12653 prog->aux->max_pkt_offset = MAX_PACKET_OFF;
12655 /* mark bpf_tail_call as different opcode to avoid
12656 * conditional branch in the interpreter for every normal
12657 * call and to prevent accidental JITing by JIT compiler
12658 * that doesn't support bpf_tail_call yet
12661 insn->code = BPF_JMP | BPF_TAIL_CALL;
12663 aux = &env->insn_aux_data[i + delta];
12664 if (env->bpf_capable && !expect_blinding &&
12665 prog->jit_requested &&
12666 !bpf_map_key_poisoned(aux) &&
12667 !bpf_map_ptr_poisoned(aux) &&
12668 !bpf_map_ptr_unpriv(aux)) {
12669 struct bpf_jit_poke_descriptor desc = {
12670 .reason = BPF_POKE_REASON_TAIL_CALL,
12671 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
12672 .tail_call.key = bpf_map_key_immediate(aux),
12673 .insn_idx = i + delta,
12676 ret = bpf_jit_add_poke_descriptor(prog, &desc);
12678 verbose(env, "adding tail call poke descriptor failed\n");
12682 insn->imm = ret + 1;
12686 if (!bpf_map_ptr_unpriv(aux))
12689 /* instead of changing every JIT dealing with tail_call
12690 * emit two extra insns:
12691 * if (index >= max_entries) goto out;
12692 * index &= array->index_mask;
12693 * to avoid out-of-bounds cpu speculation
12695 if (bpf_map_ptr_poisoned(aux)) {
12696 verbose(env, "tail_call abusing map_ptr\n");
12700 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
12701 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
12702 map_ptr->max_entries, 2);
12703 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
12704 container_of(map_ptr,
12707 insn_buf[2] = *insn;
12709 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12714 env->prog = prog = new_prog;
12715 insn = new_prog->insnsi + i + delta;
12719 if (insn->imm == BPF_FUNC_timer_set_callback) {
12720 /* The verifier will process callback_fn as many times as necessary
12721 * with different maps and the register states prepared by
12722 * set_timer_callback_state will be accurate.
12724 * The following use case is valid:
12725 * map1 is shared by prog1, prog2, prog3.
12726 * prog1 calls bpf_timer_init for some map1 elements
12727 * prog2 calls bpf_timer_set_callback for some map1 elements.
12728 * Those that were not bpf_timer_init-ed will return -EINVAL.
12729 * prog3 calls bpf_timer_start for some map1 elements.
12730 * Those that were not both bpf_timer_init-ed and
12731 * bpf_timer_set_callback-ed will return -EINVAL.
12733 struct bpf_insn ld_addrs[2] = {
12734 BPF_LD_IMM64(BPF_REG_3, (long)prog->aux),
12737 insn_buf[0] = ld_addrs[0];
12738 insn_buf[1] = ld_addrs[1];
12739 insn_buf[2] = *insn;
12742 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12747 env->prog = prog = new_prog;
12748 insn = new_prog->insnsi + i + delta;
12749 goto patch_call_imm;
12752 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
12753 * and other inlining handlers are currently limited to 64 bit
12756 if (prog->jit_requested && BITS_PER_LONG == 64 &&
12757 (insn->imm == BPF_FUNC_map_lookup_elem ||
12758 insn->imm == BPF_FUNC_map_update_elem ||
12759 insn->imm == BPF_FUNC_map_delete_elem ||
12760 insn->imm == BPF_FUNC_map_push_elem ||
12761 insn->imm == BPF_FUNC_map_pop_elem ||
12762 insn->imm == BPF_FUNC_map_peek_elem ||
12763 insn->imm == BPF_FUNC_redirect_map)) {
12764 aux = &env->insn_aux_data[i + delta];
12765 if (bpf_map_ptr_poisoned(aux))
12766 goto patch_call_imm;
12768 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
12769 ops = map_ptr->ops;
12770 if (insn->imm == BPF_FUNC_map_lookup_elem &&
12771 ops->map_gen_lookup) {
12772 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
12773 if (cnt == -EOPNOTSUPP)
12774 goto patch_map_ops_generic;
12775 if (cnt <= 0 || cnt >= ARRAY_SIZE(insn_buf)) {
12776 verbose(env, "bpf verifier is misconfigured\n");
12780 new_prog = bpf_patch_insn_data(env, i + delta,
12786 env->prog = prog = new_prog;
12787 insn = new_prog->insnsi + i + delta;
12791 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
12792 (void *(*)(struct bpf_map *map, void *key))NULL));
12793 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
12794 (int (*)(struct bpf_map *map, void *key))NULL));
12795 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
12796 (int (*)(struct bpf_map *map, void *key, void *value,
12798 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
12799 (int (*)(struct bpf_map *map, void *value,
12801 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
12802 (int (*)(struct bpf_map *map, void *value))NULL));
12803 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
12804 (int (*)(struct bpf_map *map, void *value))NULL));
12805 BUILD_BUG_ON(!__same_type(ops->map_redirect,
12806 (int (*)(struct bpf_map *map, u32 ifindex, u64 flags))NULL));
12808 patch_map_ops_generic:
12809 switch (insn->imm) {
12810 case BPF_FUNC_map_lookup_elem:
12811 insn->imm = BPF_CAST_CALL(ops->map_lookup_elem) -
12814 case BPF_FUNC_map_update_elem:
12815 insn->imm = BPF_CAST_CALL(ops->map_update_elem) -
12818 case BPF_FUNC_map_delete_elem:
12819 insn->imm = BPF_CAST_CALL(ops->map_delete_elem) -
12822 case BPF_FUNC_map_push_elem:
12823 insn->imm = BPF_CAST_CALL(ops->map_push_elem) -
12826 case BPF_FUNC_map_pop_elem:
12827 insn->imm = BPF_CAST_CALL(ops->map_pop_elem) -
12830 case BPF_FUNC_map_peek_elem:
12831 insn->imm = BPF_CAST_CALL(ops->map_peek_elem) -
12834 case BPF_FUNC_redirect_map:
12835 insn->imm = BPF_CAST_CALL(ops->map_redirect) -
12840 goto patch_call_imm;
12843 /* Implement bpf_jiffies64 inline. */
12844 if (prog->jit_requested && BITS_PER_LONG == 64 &&
12845 insn->imm == BPF_FUNC_jiffies64) {
12846 struct bpf_insn ld_jiffies_addr[2] = {
12847 BPF_LD_IMM64(BPF_REG_0,
12848 (unsigned long)&jiffies),
12851 insn_buf[0] = ld_jiffies_addr[0];
12852 insn_buf[1] = ld_jiffies_addr[1];
12853 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
12857 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
12863 env->prog = prog = new_prog;
12864 insn = new_prog->insnsi + i + delta;
12869 fn = env->ops->get_func_proto(insn->imm, env->prog);
12870 /* all functions that have prototype and verifier allowed
12871 * programs to call them, must be real in-kernel functions
12875 "kernel subsystem misconfigured func %s#%d\n",
12876 func_id_name(insn->imm), insn->imm);
12879 insn->imm = fn->func - __bpf_call_base;
12882 /* Since poke tab is now finalized, publish aux to tracker. */
12883 for (i = 0; i < prog->aux->size_poke_tab; i++) {
12884 map_ptr = prog->aux->poke_tab[i].tail_call.map;
12885 if (!map_ptr->ops->map_poke_track ||
12886 !map_ptr->ops->map_poke_untrack ||
12887 !map_ptr->ops->map_poke_run) {
12888 verbose(env, "bpf verifier is misconfigured\n");
12892 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
12894 verbose(env, "tracking tail call prog failed\n");
12899 sort_kfunc_descs_by_imm(env->prog);
12904 static void free_states(struct bpf_verifier_env *env)
12906 struct bpf_verifier_state_list *sl, *sln;
12909 sl = env->free_list;
12912 free_verifier_state(&sl->state, false);
12916 env->free_list = NULL;
12918 if (!env->explored_states)
12921 for (i = 0; i < state_htab_size(env); i++) {
12922 sl = env->explored_states[i];
12926 free_verifier_state(&sl->state, false);
12930 env->explored_states[i] = NULL;
12934 /* The verifier is using insn_aux_data[] to store temporary data during
12935 * verification and to store information for passes that run after the
12936 * verification like dead code sanitization. do_check_common() for subprogram N
12937 * may analyze many other subprograms. sanitize_insn_aux_data() clears all
12938 * temporary data after do_check_common() finds that subprogram N cannot be
12939 * verified independently. pass_cnt counts the number of times
12940 * do_check_common() was run and insn->aux->seen tells the pass number
12941 * insn_aux_data was touched. These variables are compared to clear temporary
12942 * data from failed pass. For testing and experiments do_check_common() can be
12943 * run multiple times even when prior attempt to verify is unsuccessful.
12945 * Note that special handling is needed on !env->bypass_spec_v1 if this is
12946 * ever called outside of error path with subsequent program rejection.
12948 static void sanitize_insn_aux_data(struct bpf_verifier_env *env)
12950 struct bpf_insn *insn = env->prog->insnsi;
12951 struct bpf_insn_aux_data *aux;
12954 for (i = 0; i < env->prog->len; i++) {
12955 class = BPF_CLASS(insn[i].code);
12956 if (class != BPF_LDX && class != BPF_STX)
12958 aux = &env->insn_aux_data[i];
12959 if (aux->seen != env->pass_cnt)
12961 memset(aux, 0, offsetof(typeof(*aux), orig_idx));
12965 static int do_check_common(struct bpf_verifier_env *env, int subprog)
12967 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
12968 struct bpf_verifier_state *state;
12969 struct bpf_reg_state *regs;
12972 env->prev_linfo = NULL;
12975 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
12978 state->curframe = 0;
12979 state->speculative = false;
12980 state->branches = 1;
12981 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
12982 if (!state->frame[0]) {
12986 env->cur_state = state;
12987 init_func_state(env, state->frame[0],
12988 BPF_MAIN_FUNC /* callsite */,
12992 regs = state->frame[state->curframe]->regs;
12993 if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
12994 ret = btf_prepare_func_args(env, subprog, regs);
12997 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
12998 if (regs[i].type == PTR_TO_CTX)
12999 mark_reg_known_zero(env, regs, i);
13000 else if (regs[i].type == SCALAR_VALUE)
13001 mark_reg_unknown(env, regs, i);
13002 else if (regs[i].type == PTR_TO_MEM_OR_NULL) {
13003 const u32 mem_size = regs[i].mem_size;
13005 mark_reg_known_zero(env, regs, i);
13006 regs[i].mem_size = mem_size;
13007 regs[i].id = ++env->id_gen;
13011 /* 1st arg to a function */
13012 regs[BPF_REG_1].type = PTR_TO_CTX;
13013 mark_reg_known_zero(env, regs, BPF_REG_1);
13014 ret = btf_check_subprog_arg_match(env, subprog, regs);
13015 if (ret == -EFAULT)
13016 /* unlikely verifier bug. abort.
13017 * ret == 0 and ret < 0 are sadly acceptable for
13018 * main() function due to backward compatibility.
13019 * Like socket filter program may be written as:
13020 * int bpf_prog(struct pt_regs *ctx)
13021 * and never dereference that ctx in the program.
13022 * 'struct pt_regs' is a type mismatch for socket
13023 * filter that should be using 'struct __sk_buff'.
13028 ret = do_check(env);
13030 /* check for NULL is necessary, since cur_state can be freed inside
13031 * do_check() under memory pressure.
13033 if (env->cur_state) {
13034 free_verifier_state(env->cur_state, true);
13035 env->cur_state = NULL;
13037 while (!pop_stack(env, NULL, NULL, false));
13038 if (!ret && pop_log)
13039 bpf_vlog_reset(&env->log, 0);
13042 /* clean aux data in case subprog was rejected */
13043 sanitize_insn_aux_data(env);
13047 /* Verify all global functions in a BPF program one by one based on their BTF.
13048 * All global functions must pass verification. Otherwise the whole program is rejected.
13059 * foo() will be verified first for R1=any_scalar_value. During verification it
13060 * will be assumed that bar() already verified successfully and call to bar()
13061 * from foo() will be checked for type match only. Later bar() will be verified
13062 * independently to check that it's safe for R1=any_scalar_value.
13064 static int do_check_subprogs(struct bpf_verifier_env *env)
13066 struct bpf_prog_aux *aux = env->prog->aux;
13069 if (!aux->func_info)
13072 for (i = 1; i < env->subprog_cnt; i++) {
13073 if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL)
13075 env->insn_idx = env->subprog_info[i].start;
13076 WARN_ON_ONCE(env->insn_idx == 0);
13077 ret = do_check_common(env, i);
13080 } else if (env->log.level & BPF_LOG_LEVEL) {
13082 "Func#%d is safe for any args that match its prototype\n",
13089 static int do_check_main(struct bpf_verifier_env *env)
13094 ret = do_check_common(env, 0);
13096 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
13101 static void print_verification_stats(struct bpf_verifier_env *env)
13105 if (env->log.level & BPF_LOG_STATS) {
13106 verbose(env, "verification time %lld usec\n",
13107 div_u64(env->verification_time, 1000));
13108 verbose(env, "stack depth ");
13109 for (i = 0; i < env->subprog_cnt; i++) {
13110 u32 depth = env->subprog_info[i].stack_depth;
13112 verbose(env, "%d", depth);
13113 if (i + 1 < env->subprog_cnt)
13116 verbose(env, "\n");
13118 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
13119 "total_states %d peak_states %d mark_read %d\n",
13120 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
13121 env->max_states_per_insn, env->total_states,
13122 env->peak_states, env->longest_mark_read_walk);
13125 static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
13127 const struct btf_type *t, *func_proto;
13128 const struct bpf_struct_ops *st_ops;
13129 const struct btf_member *member;
13130 struct bpf_prog *prog = env->prog;
13131 u32 btf_id, member_idx;
13134 if (!prog->gpl_compatible) {
13135 verbose(env, "struct ops programs must have a GPL compatible license\n");
13139 btf_id = prog->aux->attach_btf_id;
13140 st_ops = bpf_struct_ops_find(btf_id);
13142 verbose(env, "attach_btf_id %u is not a supported struct\n",
13148 member_idx = prog->expected_attach_type;
13149 if (member_idx >= btf_type_vlen(t)) {
13150 verbose(env, "attach to invalid member idx %u of struct %s\n",
13151 member_idx, st_ops->name);
13155 member = &btf_type_member(t)[member_idx];
13156 mname = btf_name_by_offset(btf_vmlinux, member->name_off);
13157 func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type,
13160 verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
13161 mname, member_idx, st_ops->name);
13165 if (st_ops->check_member) {
13166 int err = st_ops->check_member(t, member);
13169 verbose(env, "attach to unsupported member %s of struct %s\n",
13170 mname, st_ops->name);
13175 prog->aux->attach_func_proto = func_proto;
13176 prog->aux->attach_func_name = mname;
13177 env->ops = st_ops->verifier_ops;
13181 #define SECURITY_PREFIX "security_"
13183 static int check_attach_modify_return(unsigned long addr, const char *func_name)
13185 if (within_error_injection_list(addr) ||
13186 !strncmp(SECURITY_PREFIX, func_name, sizeof(SECURITY_PREFIX) - 1))
13192 /* list of non-sleepable functions that are otherwise on
13193 * ALLOW_ERROR_INJECTION list
13195 BTF_SET_START(btf_non_sleepable_error_inject)
13196 /* Three functions below can be called from sleepable and non-sleepable context.
13197 * Assume non-sleepable from bpf safety point of view.
13199 BTF_ID(func, __add_to_page_cache_locked)
13200 BTF_ID(func, should_fail_alloc_page)
13201 BTF_ID(func, should_failslab)
13202 BTF_SET_END(btf_non_sleepable_error_inject)
13204 static int check_non_sleepable_error_inject(u32 btf_id)
13206 return btf_id_set_contains(&btf_non_sleepable_error_inject, btf_id);
13209 int bpf_check_attach_target(struct bpf_verifier_log *log,
13210 const struct bpf_prog *prog,
13211 const struct bpf_prog *tgt_prog,
13213 struct bpf_attach_target_info *tgt_info)
13215 bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
13216 const char prefix[] = "btf_trace_";
13217 int ret = 0, subprog = -1, i;
13218 const struct btf_type *t;
13219 bool conservative = true;
13225 bpf_log(log, "Tracing programs must provide btf_id\n");
13228 btf = tgt_prog ? tgt_prog->aux->btf : prog->aux->attach_btf;
13231 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
13234 t = btf_type_by_id(btf, btf_id);
13236 bpf_log(log, "attach_btf_id %u is invalid\n", btf_id);
13239 tname = btf_name_by_offset(btf, t->name_off);
13241 bpf_log(log, "attach_btf_id %u doesn't have a name\n", btf_id);
13245 struct bpf_prog_aux *aux = tgt_prog->aux;
13247 for (i = 0; i < aux->func_info_cnt; i++)
13248 if (aux->func_info[i].type_id == btf_id) {
13252 if (subprog == -1) {
13253 bpf_log(log, "Subprog %s doesn't exist\n", tname);
13256 conservative = aux->func_info_aux[subprog].unreliable;
13257 if (prog_extension) {
13258 if (conservative) {
13260 "Cannot replace static functions\n");
13263 if (!prog->jit_requested) {
13265 "Extension programs should be JITed\n");
13269 if (!tgt_prog->jited) {
13270 bpf_log(log, "Can attach to only JITed progs\n");
13273 if (tgt_prog->type == prog->type) {
13274 /* Cannot fentry/fexit another fentry/fexit program.
13275 * Cannot attach program extension to another extension.
13276 * It's ok to attach fentry/fexit to extension program.
13278 bpf_log(log, "Cannot recursively attach\n");
13281 if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
13283 (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
13284 tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
13285 /* Program extensions can extend all program types
13286 * except fentry/fexit. The reason is the following.
13287 * The fentry/fexit programs are used for performance
13288 * analysis, stats and can be attached to any program
13289 * type except themselves. When extension program is
13290 * replacing XDP function it is necessary to allow
13291 * performance analysis of all functions. Both original
13292 * XDP program and its program extension. Hence
13293 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
13294 * allowed. If extending of fentry/fexit was allowed it
13295 * would be possible to create long call chain
13296 * fentry->extension->fentry->extension beyond
13297 * reasonable stack size. Hence extending fentry is not
13300 bpf_log(log, "Cannot extend fentry/fexit\n");
13304 if (prog_extension) {
13305 bpf_log(log, "Cannot replace kernel functions\n");
13310 switch (prog->expected_attach_type) {
13311 case BPF_TRACE_RAW_TP:
13314 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
13317 if (!btf_type_is_typedef(t)) {
13318 bpf_log(log, "attach_btf_id %u is not a typedef\n",
13322 if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
13323 bpf_log(log, "attach_btf_id %u points to wrong type name %s\n",
13327 tname += sizeof(prefix) - 1;
13328 t = btf_type_by_id(btf, t->type);
13329 if (!btf_type_is_ptr(t))
13330 /* should never happen in valid vmlinux build */
13332 t = btf_type_by_id(btf, t->type);
13333 if (!btf_type_is_func_proto(t))
13334 /* should never happen in valid vmlinux build */
13338 case BPF_TRACE_ITER:
13339 if (!btf_type_is_func(t)) {
13340 bpf_log(log, "attach_btf_id %u is not a function\n",
13344 t = btf_type_by_id(btf, t->type);
13345 if (!btf_type_is_func_proto(t))
13347 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
13352 if (!prog_extension)
13355 case BPF_MODIFY_RETURN:
13357 case BPF_TRACE_FENTRY:
13358 case BPF_TRACE_FEXIT:
13359 if (!btf_type_is_func(t)) {
13360 bpf_log(log, "attach_btf_id %u is not a function\n",
13364 if (prog_extension &&
13365 btf_check_type_match(log, prog, btf, t))
13367 t = btf_type_by_id(btf, t->type);
13368 if (!btf_type_is_func_proto(t))
13371 if ((prog->aux->saved_dst_prog_type || prog->aux->saved_dst_attach_type) &&
13372 (!tgt_prog || prog->aux->saved_dst_prog_type != tgt_prog->type ||
13373 prog->aux->saved_dst_attach_type != tgt_prog->expected_attach_type))
13376 if (tgt_prog && conservative)
13379 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
13385 addr = (long) tgt_prog->bpf_func;
13387 addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
13389 addr = kallsyms_lookup_name(tname);
13392 "The address of function %s cannot be found\n",
13398 if (prog->aux->sleepable) {
13400 switch (prog->type) {
13401 case BPF_PROG_TYPE_TRACING:
13402 /* fentry/fexit/fmod_ret progs can be sleepable only if they are
13403 * attached to ALLOW_ERROR_INJECTION and are not in denylist.
13405 if (!check_non_sleepable_error_inject(btf_id) &&
13406 within_error_injection_list(addr))
13409 case BPF_PROG_TYPE_LSM:
13410 /* LSM progs check that they are attached to bpf_lsm_*() funcs.
13411 * Only some of them are sleepable.
13413 if (bpf_lsm_is_sleepable_hook(btf_id))
13420 bpf_log(log, "%s is not sleepable\n", tname);
13423 } else if (prog->expected_attach_type == BPF_MODIFY_RETURN) {
13425 bpf_log(log, "can't modify return codes of BPF programs\n");
13428 ret = check_attach_modify_return(addr, tname);
13430 bpf_log(log, "%s() is not modifiable\n", tname);
13437 tgt_info->tgt_addr = addr;
13438 tgt_info->tgt_name = tname;
13439 tgt_info->tgt_type = t;
13443 BTF_SET_START(btf_id_deny)
13446 BTF_ID(func, migrate_disable)
13447 BTF_ID(func, migrate_enable)
13449 #if !defined CONFIG_PREEMPT_RCU && !defined CONFIG_TINY_RCU
13450 BTF_ID(func, rcu_read_unlock_strict)
13452 BTF_SET_END(btf_id_deny)
13454 static int check_attach_btf_id(struct bpf_verifier_env *env)
13456 struct bpf_prog *prog = env->prog;
13457 struct bpf_prog *tgt_prog = prog->aux->dst_prog;
13458 struct bpf_attach_target_info tgt_info = {};
13459 u32 btf_id = prog->aux->attach_btf_id;
13460 struct bpf_trampoline *tr;
13464 if (prog->type == BPF_PROG_TYPE_SYSCALL) {
13465 if (prog->aux->sleepable)
13466 /* attach_btf_id checked to be zero already */
13468 verbose(env, "Syscall programs can only be sleepable\n");
13472 if (prog->aux->sleepable && prog->type != BPF_PROG_TYPE_TRACING &&
13473 prog->type != BPF_PROG_TYPE_LSM) {
13474 verbose(env, "Only fentry/fexit/fmod_ret and lsm programs can be sleepable\n");
13478 if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
13479 return check_struct_ops_btf_id(env);
13481 if (prog->type != BPF_PROG_TYPE_TRACING &&
13482 prog->type != BPF_PROG_TYPE_LSM &&
13483 prog->type != BPF_PROG_TYPE_EXT)
13486 ret = bpf_check_attach_target(&env->log, prog, tgt_prog, btf_id, &tgt_info);
13490 if (tgt_prog && prog->type == BPF_PROG_TYPE_EXT) {
13491 /* to make freplace equivalent to their targets, they need to
13492 * inherit env->ops and expected_attach_type for the rest of the
13495 env->ops = bpf_verifier_ops[tgt_prog->type];
13496 prog->expected_attach_type = tgt_prog->expected_attach_type;
13499 /* store info about the attachment target that will be used later */
13500 prog->aux->attach_func_proto = tgt_info.tgt_type;
13501 prog->aux->attach_func_name = tgt_info.tgt_name;
13504 prog->aux->saved_dst_prog_type = tgt_prog->type;
13505 prog->aux->saved_dst_attach_type = tgt_prog->expected_attach_type;
13508 if (prog->expected_attach_type == BPF_TRACE_RAW_TP) {
13509 prog->aux->attach_btf_trace = true;
13511 } else if (prog->expected_attach_type == BPF_TRACE_ITER) {
13512 if (!bpf_iter_prog_supported(prog))
13517 if (prog->type == BPF_PROG_TYPE_LSM) {
13518 ret = bpf_lsm_verify_prog(&env->log, prog);
13521 } else if (prog->type == BPF_PROG_TYPE_TRACING &&
13522 btf_id_set_contains(&btf_id_deny, btf_id)) {
13526 key = bpf_trampoline_compute_key(tgt_prog, prog->aux->attach_btf, btf_id);
13527 tr = bpf_trampoline_get(key, &tgt_info);
13531 prog->aux->dst_trampoline = tr;
13535 struct btf *bpf_get_btf_vmlinux(void)
13537 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
13538 mutex_lock(&bpf_verifier_lock);
13540 btf_vmlinux = btf_parse_vmlinux();
13541 mutex_unlock(&bpf_verifier_lock);
13543 return btf_vmlinux;
13546 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr, bpfptr_t uattr)
13548 u64 start_time = ktime_get_ns();
13549 struct bpf_verifier_env *env;
13550 struct bpf_verifier_log *log;
13551 int i, len, ret = -EINVAL;
13554 /* no program is valid */
13555 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
13558 /* 'struct bpf_verifier_env' can be global, but since it's not small,
13559 * allocate/free it every time bpf_check() is called
13561 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
13566 len = (*prog)->len;
13567 env->insn_aux_data =
13568 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
13570 if (!env->insn_aux_data)
13572 for (i = 0; i < len; i++)
13573 env->insn_aux_data[i].orig_idx = i;
13575 env->ops = bpf_verifier_ops[env->prog->type];
13576 env->fd_array = make_bpfptr(attr->fd_array, uattr.is_kernel);
13577 is_priv = bpf_capable();
13579 bpf_get_btf_vmlinux();
13581 /* grab the mutex to protect few globals used by verifier */
13583 mutex_lock(&bpf_verifier_lock);
13585 if (attr->log_level || attr->log_buf || attr->log_size) {
13586 /* user requested verbose verifier output
13587 * and supplied buffer to store the verification trace
13589 log->level = attr->log_level;
13590 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
13591 log->len_total = attr->log_size;
13594 /* log attributes have to be sane */
13595 if (log->len_total < 128 || log->len_total > UINT_MAX >> 2 ||
13596 !log->level || !log->ubuf || log->level & ~BPF_LOG_MASK)
13600 if (IS_ERR(btf_vmlinux)) {
13601 /* Either gcc or pahole or kernel are broken. */
13602 verbose(env, "in-kernel BTF is malformed\n");
13603 ret = PTR_ERR(btf_vmlinux);
13604 goto skip_full_check;
13607 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
13608 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
13609 env->strict_alignment = true;
13610 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
13611 env->strict_alignment = false;
13613 env->allow_ptr_leaks = bpf_allow_ptr_leaks();
13614 env->allow_uninit_stack = bpf_allow_uninit_stack();
13615 env->allow_ptr_to_map_access = bpf_allow_ptr_to_map_access();
13616 env->bypass_spec_v1 = bpf_bypass_spec_v1();
13617 env->bypass_spec_v4 = bpf_bypass_spec_v4();
13618 env->bpf_capable = bpf_capable();
13621 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
13623 env->explored_states = kvcalloc(state_htab_size(env),
13624 sizeof(struct bpf_verifier_state_list *),
13627 if (!env->explored_states)
13628 goto skip_full_check;
13630 ret = add_subprog_and_kfunc(env);
13632 goto skip_full_check;
13634 ret = check_subprogs(env);
13636 goto skip_full_check;
13638 ret = check_btf_info(env, attr, uattr);
13640 goto skip_full_check;
13642 ret = check_attach_btf_id(env);
13644 goto skip_full_check;
13646 ret = resolve_pseudo_ldimm64(env);
13648 goto skip_full_check;
13650 if (bpf_prog_is_dev_bound(env->prog->aux)) {
13651 ret = bpf_prog_offload_verifier_prep(env->prog);
13653 goto skip_full_check;
13656 ret = check_cfg(env);
13658 goto skip_full_check;
13660 ret = do_check_subprogs(env);
13661 ret = ret ?: do_check_main(env);
13663 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
13664 ret = bpf_prog_offload_finalize(env);
13667 kvfree(env->explored_states);
13670 ret = check_max_stack_depth(env);
13672 /* instruction rewrites happen after this point */
13675 opt_hard_wire_dead_code_branches(env);
13677 ret = opt_remove_dead_code(env);
13679 ret = opt_remove_nops(env);
13682 sanitize_dead_code(env);
13686 /* program is valid, convert *(u32*)(ctx + off) accesses */
13687 ret = convert_ctx_accesses(env);
13690 ret = do_misc_fixups(env);
13692 /* do 32-bit optimization after insn patching has done so those patched
13693 * insns could be handled correctly.
13695 if (ret == 0 && !bpf_prog_is_dev_bound(env->prog->aux)) {
13696 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
13697 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
13702 ret = fixup_call_args(env);
13704 env->verification_time = ktime_get_ns() - start_time;
13705 print_verification_stats(env);
13707 if (log->level && bpf_verifier_log_full(log))
13709 if (log->level && !log->ubuf) {
13711 goto err_release_maps;
13715 goto err_release_maps;
13717 if (env->used_map_cnt) {
13718 /* if program passed verifier, update used_maps in bpf_prog_info */
13719 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
13720 sizeof(env->used_maps[0]),
13723 if (!env->prog->aux->used_maps) {
13725 goto err_release_maps;
13728 memcpy(env->prog->aux->used_maps, env->used_maps,
13729 sizeof(env->used_maps[0]) * env->used_map_cnt);
13730 env->prog->aux->used_map_cnt = env->used_map_cnt;
13732 if (env->used_btf_cnt) {
13733 /* if program passed verifier, update used_btfs in bpf_prog_aux */
13734 env->prog->aux->used_btfs = kmalloc_array(env->used_btf_cnt,
13735 sizeof(env->used_btfs[0]),
13737 if (!env->prog->aux->used_btfs) {
13739 goto err_release_maps;
13742 memcpy(env->prog->aux->used_btfs, env->used_btfs,
13743 sizeof(env->used_btfs[0]) * env->used_btf_cnt);
13744 env->prog->aux->used_btf_cnt = env->used_btf_cnt;
13746 if (env->used_map_cnt || env->used_btf_cnt) {
13747 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
13748 * bpf_ld_imm64 instructions
13750 convert_pseudo_ld_imm64(env);
13753 adjust_btf_func(env);
13756 if (!env->prog->aux->used_maps)
13757 /* if we didn't copy map pointers into bpf_prog_info, release
13758 * them now. Otherwise free_used_maps() will release them.
13761 if (!env->prog->aux->used_btfs)
13764 /* extension progs temporarily inherit the attach_type of their targets
13765 for verification purposes, so set it back to zero before returning
13767 if (env->prog->type == BPF_PROG_TYPE_EXT)
13768 env->prog->expected_attach_type = 0;
13773 mutex_unlock(&bpf_verifier_lock);
13774 vfree(env->insn_aux_data);