1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
3 * Copyright (c) 2016 Facebook
4 * Copyright (c) 2018 Covalent IO, Inc. http://covalent.io
6 #include <uapi/linux/btf.h>
7 #include <linux/kernel.h>
8 #include <linux/types.h>
9 #include <linux/slab.h>
10 #include <linux/bpf.h>
11 #include <linux/btf.h>
12 #include <linux/bpf_verifier.h>
13 #include <linux/filter.h>
14 #include <net/netlink.h>
15 #include <linux/file.h>
16 #include <linux/vmalloc.h>
17 #include <linux/stringify.h>
18 #include <linux/bsearch.h>
19 #include <linux/sort.h>
20 #include <linux/perf_event.h>
21 #include <linux/ctype.h>
22 #include <linux/error-injection.h>
23 #include <linux/bpf_lsm.h>
24 #include <linux/btf_ids.h>
28 static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
29 #define BPF_PROG_TYPE(_id, _name, prog_ctx_type, kern_ctx_type) \
30 [_id] = & _name ## _verifier_ops,
31 #define BPF_MAP_TYPE(_id, _ops)
32 #define BPF_LINK_TYPE(_id, _name)
33 #include <linux/bpf_types.h>
39 /* bpf_check() is a static code analyzer that walks eBPF program
40 * instruction by instruction and updates register/stack state.
41 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
43 * The first pass is depth-first-search to check that the program is a DAG.
44 * It rejects the following programs:
45 * - larger than BPF_MAXINSNS insns
46 * - if loop is present (detected via back-edge)
47 * - unreachable insns exist (shouldn't be a forest. program = one function)
48 * - out of bounds or malformed jumps
49 * The second pass is all possible path descent from the 1st insn.
50 * Since it's analyzing all paths through the program, the length of the
51 * analysis is limited to 64k insn, which may be hit even if total number of
52 * insn is less then 4K, but there are too many branches that change stack/regs.
53 * Number of 'branches to be analyzed' is limited to 1k
55 * On entry to each instruction, each register has a type, and the instruction
56 * changes the types of the registers depending on instruction semantics.
57 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
60 * All registers are 64-bit.
61 * R0 - return register
62 * R1-R5 argument passing registers
63 * R6-R9 callee saved registers
64 * R10 - frame pointer read-only
66 * At the start of BPF program the register R1 contains a pointer to bpf_context
67 * and has type PTR_TO_CTX.
69 * Verifier tracks arithmetic operations on pointers in case:
70 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
71 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
72 * 1st insn copies R10 (which has FRAME_PTR) type into R1
73 * and 2nd arithmetic instruction is pattern matched to recognize
74 * that it wants to construct a pointer to some element within stack.
75 * So after 2nd insn, the register R1 has type PTR_TO_STACK
76 * (and -20 constant is saved for further stack bounds checking).
77 * Meaning that this reg is a pointer to stack plus known immediate constant.
79 * Most of the time the registers have SCALAR_VALUE type, which
80 * means the register has some value, but it's not a valid pointer.
81 * (like pointer plus pointer becomes SCALAR_VALUE type)
83 * When verifier sees load or store instructions the type of base register
84 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK, PTR_TO_SOCKET. These are
85 * four pointer types recognized by check_mem_access() function.
87 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
88 * and the range of [ptr, ptr + map's value_size) is accessible.
90 * registers used to pass values to function calls are checked against
91 * function argument constraints.
93 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
94 * It means that the register type passed to this function must be
95 * PTR_TO_STACK and it will be used inside the function as
96 * 'pointer to map element key'
98 * For example the argument constraints for bpf_map_lookup_elem():
99 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
100 * .arg1_type = ARG_CONST_MAP_PTR,
101 * .arg2_type = ARG_PTR_TO_MAP_KEY,
103 * ret_type says that this function returns 'pointer to map elem value or null'
104 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
105 * 2nd argument should be a pointer to stack, which will be used inside
106 * the helper function as a pointer to map element key.
108 * On the kernel side the helper function looks like:
109 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
111 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
112 * void *key = (void *) (unsigned long) r2;
115 * here kernel can access 'key' and 'map' pointers safely, knowing that
116 * [key, key + map->key_size) bytes are valid and were initialized on
117 * the stack of eBPF program.
120 * Corresponding eBPF program may look like:
121 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
122 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
123 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
124 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
125 * here verifier looks at prototype of map_lookup_elem() and sees:
126 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
127 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
129 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
130 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
131 * and were initialized prior to this call.
132 * If it's ok, then verifier allows this BPF_CALL insn and looks at
133 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
134 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
135 * returns either pointer to map value or NULL.
137 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
138 * insn, the register holding that pointer in the true branch changes state to
139 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
140 * branch. See check_cond_jmp_op().
142 * After the call R0 is set to return type of the function and registers R1-R5
143 * are set to NOT_INIT to indicate that they are no longer readable.
145 * The following reference types represent a potential reference to a kernel
146 * resource which, after first being allocated, must be checked and freed by
148 * - PTR_TO_SOCKET_OR_NULL, PTR_TO_SOCKET
150 * When the verifier sees a helper call return a reference type, it allocates a
151 * pointer id for the reference and stores it in the current function state.
152 * Similar to the way that PTR_TO_MAP_VALUE_OR_NULL is converted into
153 * PTR_TO_MAP_VALUE, PTR_TO_SOCKET_OR_NULL becomes PTR_TO_SOCKET when the type
154 * passes through a NULL-check conditional. For the branch wherein the state is
155 * changed to CONST_IMM, the verifier releases the reference.
157 * For each helper function that allocates a reference, such as
158 * bpf_sk_lookup_tcp(), there is a corresponding release function, such as
159 * bpf_sk_release(). When a reference type passes into the release function,
160 * the verifier also releases the reference. If any unchecked or unreleased
161 * reference remains at the end of the program, the verifier rejects it.
164 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
165 struct bpf_verifier_stack_elem {
166 /* verifer state is 'st'
167 * before processing instruction 'insn_idx'
168 * and after processing instruction 'prev_insn_idx'
170 struct bpf_verifier_state st;
173 struct bpf_verifier_stack_elem *next;
174 /* length of verifier log at the time this state was pushed on stack */
178 #define BPF_COMPLEXITY_LIMIT_JMP_SEQ 8192
179 #define BPF_COMPLEXITY_LIMIT_STATES 64
181 #define BPF_MAP_KEY_POISON (1ULL << 63)
182 #define BPF_MAP_KEY_SEEN (1ULL << 62)
184 #define BPF_MAP_PTR_UNPRIV 1UL
185 #define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \
186 POISON_POINTER_DELTA))
187 #define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
189 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
191 return BPF_MAP_PTR(aux->map_ptr_state) == BPF_MAP_PTR_POISON;
194 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
196 return aux->map_ptr_state & BPF_MAP_PTR_UNPRIV;
199 static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
200 const struct bpf_map *map, bool unpriv)
202 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON & BPF_MAP_PTR_UNPRIV);
203 unpriv |= bpf_map_ptr_unpriv(aux);
204 aux->map_ptr_state = (unsigned long)map |
205 (unpriv ? BPF_MAP_PTR_UNPRIV : 0UL);
208 static bool bpf_map_key_poisoned(const struct bpf_insn_aux_data *aux)
210 return aux->map_key_state & BPF_MAP_KEY_POISON;
213 static bool bpf_map_key_unseen(const struct bpf_insn_aux_data *aux)
215 return !(aux->map_key_state & BPF_MAP_KEY_SEEN);
218 static u64 bpf_map_key_immediate(const struct bpf_insn_aux_data *aux)
220 return aux->map_key_state & ~(BPF_MAP_KEY_SEEN | BPF_MAP_KEY_POISON);
223 static void bpf_map_key_store(struct bpf_insn_aux_data *aux, u64 state)
225 bool poisoned = bpf_map_key_poisoned(aux);
227 aux->map_key_state = state | BPF_MAP_KEY_SEEN |
228 (poisoned ? BPF_MAP_KEY_POISON : 0ULL);
231 static bool bpf_pseudo_call(const struct bpf_insn *insn)
233 return insn->code == (BPF_JMP | BPF_CALL) &&
234 insn->src_reg == BPF_PSEUDO_CALL;
237 static bool bpf_pseudo_kfunc_call(const struct bpf_insn *insn)
239 return insn->code == (BPF_JMP | BPF_CALL) &&
240 insn->src_reg == BPF_PSEUDO_KFUNC_CALL;
243 struct bpf_call_arg_meta {
244 struct bpf_map *map_ptr;
261 struct btf *btf_vmlinux;
263 static DEFINE_MUTEX(bpf_verifier_lock);
265 static const struct bpf_line_info *
266 find_linfo(const struct bpf_verifier_env *env, u32 insn_off)
268 const struct bpf_line_info *linfo;
269 const struct bpf_prog *prog;
273 nr_linfo = prog->aux->nr_linfo;
275 if (!nr_linfo || insn_off >= prog->len)
278 linfo = prog->aux->linfo;
279 for (i = 1; i < nr_linfo; i++)
280 if (insn_off < linfo[i].insn_off)
283 return &linfo[i - 1];
286 void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
291 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
293 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
294 "verifier log line truncated - local buffer too short\n");
296 n = min(log->len_total - log->len_used - 1, n);
299 if (log->level == BPF_LOG_KERNEL) {
300 pr_err("BPF:%s\n", log->kbuf);
303 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
309 static void bpf_vlog_reset(struct bpf_verifier_log *log, u32 new_pos)
313 if (!bpf_verifier_log_needed(log))
316 log->len_used = new_pos;
317 if (put_user(zero, log->ubuf + new_pos))
321 /* log_level controls verbosity level of eBPF verifier.
322 * bpf_verifier_log_write() is used to dump the verification trace to the log,
323 * so the user can figure out what's wrong with the program
325 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
326 const char *fmt, ...)
330 if (!bpf_verifier_log_needed(&env->log))
334 bpf_verifier_vlog(&env->log, fmt, args);
337 EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
339 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
341 struct bpf_verifier_env *env = private_data;
344 if (!bpf_verifier_log_needed(&env->log))
348 bpf_verifier_vlog(&env->log, fmt, args);
352 __printf(2, 3) void bpf_log(struct bpf_verifier_log *log,
353 const char *fmt, ...)
357 if (!bpf_verifier_log_needed(log))
361 bpf_verifier_vlog(log, fmt, args);
365 static const char *ltrim(const char *s)
373 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env,
375 const char *prefix_fmt, ...)
377 const struct bpf_line_info *linfo;
379 if (!bpf_verifier_log_needed(&env->log))
382 linfo = find_linfo(env, insn_off);
383 if (!linfo || linfo == env->prev_linfo)
389 va_start(args, prefix_fmt);
390 bpf_verifier_vlog(&env->log, prefix_fmt, args);
395 ltrim(btf_name_by_offset(env->prog->aux->btf,
398 env->prev_linfo = linfo;
401 static void verbose_invalid_scalar(struct bpf_verifier_env *env,
402 struct bpf_reg_state *reg,
403 struct tnum *range, const char *ctx,
404 const char *reg_name)
408 verbose(env, "At %s the register %s ", ctx, reg_name);
409 if (!tnum_is_unknown(reg->var_off)) {
410 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
411 verbose(env, "has value %s", tn_buf);
413 verbose(env, "has unknown scalar value");
415 tnum_strn(tn_buf, sizeof(tn_buf), *range);
416 verbose(env, " should have been in %s\n", tn_buf);
419 static bool type_is_pkt_pointer(enum bpf_reg_type type)
421 return type == PTR_TO_PACKET ||
422 type == PTR_TO_PACKET_META;
425 static bool type_is_sk_pointer(enum bpf_reg_type type)
427 return type == PTR_TO_SOCKET ||
428 type == PTR_TO_SOCK_COMMON ||
429 type == PTR_TO_TCP_SOCK ||
430 type == PTR_TO_XDP_SOCK;
433 static bool reg_type_not_null(enum bpf_reg_type type)
435 return type == PTR_TO_SOCKET ||
436 type == PTR_TO_TCP_SOCK ||
437 type == PTR_TO_MAP_VALUE ||
438 type == PTR_TO_MAP_KEY ||
439 type == PTR_TO_SOCK_COMMON;
442 static bool reg_type_may_be_null(enum bpf_reg_type type)
444 return type == PTR_TO_MAP_VALUE_OR_NULL ||
445 type == PTR_TO_SOCKET_OR_NULL ||
446 type == PTR_TO_SOCK_COMMON_OR_NULL ||
447 type == PTR_TO_TCP_SOCK_OR_NULL ||
448 type == PTR_TO_BTF_ID_OR_NULL ||
449 type == PTR_TO_MEM_OR_NULL ||
450 type == PTR_TO_RDONLY_BUF_OR_NULL ||
451 type == PTR_TO_RDWR_BUF_OR_NULL;
454 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
456 return reg->type == PTR_TO_MAP_VALUE &&
457 map_value_has_spin_lock(reg->map_ptr);
460 static bool reg_type_may_be_refcounted_or_null(enum bpf_reg_type type)
462 return type == PTR_TO_SOCKET ||
463 type == PTR_TO_SOCKET_OR_NULL ||
464 type == PTR_TO_TCP_SOCK ||
465 type == PTR_TO_TCP_SOCK_OR_NULL ||
466 type == PTR_TO_MEM ||
467 type == PTR_TO_MEM_OR_NULL;
470 static bool arg_type_may_be_refcounted(enum bpf_arg_type type)
472 return type == ARG_PTR_TO_SOCK_COMMON;
475 static bool arg_type_may_be_null(enum bpf_arg_type type)
477 return type == ARG_PTR_TO_MAP_VALUE_OR_NULL ||
478 type == ARG_PTR_TO_MEM_OR_NULL ||
479 type == ARG_PTR_TO_CTX_OR_NULL ||
480 type == ARG_PTR_TO_SOCKET_OR_NULL ||
481 type == ARG_PTR_TO_ALLOC_MEM_OR_NULL ||
482 type == ARG_PTR_TO_STACK_OR_NULL;
485 /* Determine whether the function releases some resources allocated by another
486 * function call. The first reference type argument will be assumed to be
487 * released by release_reference().
489 static bool is_release_function(enum bpf_func_id func_id)
491 return func_id == BPF_FUNC_sk_release ||
492 func_id == BPF_FUNC_ringbuf_submit ||
493 func_id == BPF_FUNC_ringbuf_discard;
496 static bool may_be_acquire_function(enum bpf_func_id func_id)
498 return func_id == BPF_FUNC_sk_lookup_tcp ||
499 func_id == BPF_FUNC_sk_lookup_udp ||
500 func_id == BPF_FUNC_skc_lookup_tcp ||
501 func_id == BPF_FUNC_map_lookup_elem ||
502 func_id == BPF_FUNC_ringbuf_reserve;
505 static bool is_acquire_function(enum bpf_func_id func_id,
506 const struct bpf_map *map)
508 enum bpf_map_type map_type = map ? map->map_type : BPF_MAP_TYPE_UNSPEC;
510 if (func_id == BPF_FUNC_sk_lookup_tcp ||
511 func_id == BPF_FUNC_sk_lookup_udp ||
512 func_id == BPF_FUNC_skc_lookup_tcp ||
513 func_id == BPF_FUNC_ringbuf_reserve)
516 if (func_id == BPF_FUNC_map_lookup_elem &&
517 (map_type == BPF_MAP_TYPE_SOCKMAP ||
518 map_type == BPF_MAP_TYPE_SOCKHASH))
524 static bool is_ptr_cast_function(enum bpf_func_id func_id)
526 return func_id == BPF_FUNC_tcp_sock ||
527 func_id == BPF_FUNC_sk_fullsock ||
528 func_id == BPF_FUNC_skc_to_tcp_sock ||
529 func_id == BPF_FUNC_skc_to_tcp6_sock ||
530 func_id == BPF_FUNC_skc_to_udp6_sock ||
531 func_id == BPF_FUNC_skc_to_tcp_timewait_sock ||
532 func_id == BPF_FUNC_skc_to_tcp_request_sock;
535 static bool is_cmpxchg_insn(const struct bpf_insn *insn)
537 return BPF_CLASS(insn->code) == BPF_STX &&
538 BPF_MODE(insn->code) == BPF_ATOMIC &&
539 insn->imm == BPF_CMPXCHG;
542 /* string representation of 'enum bpf_reg_type' */
543 static const char * const reg_type_str[] = {
545 [SCALAR_VALUE] = "inv",
546 [PTR_TO_CTX] = "ctx",
547 [CONST_PTR_TO_MAP] = "map_ptr",
548 [PTR_TO_MAP_VALUE] = "map_value",
549 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
550 [PTR_TO_STACK] = "fp",
551 [PTR_TO_PACKET] = "pkt",
552 [PTR_TO_PACKET_META] = "pkt_meta",
553 [PTR_TO_PACKET_END] = "pkt_end",
554 [PTR_TO_FLOW_KEYS] = "flow_keys",
555 [PTR_TO_SOCKET] = "sock",
556 [PTR_TO_SOCKET_OR_NULL] = "sock_or_null",
557 [PTR_TO_SOCK_COMMON] = "sock_common",
558 [PTR_TO_SOCK_COMMON_OR_NULL] = "sock_common_or_null",
559 [PTR_TO_TCP_SOCK] = "tcp_sock",
560 [PTR_TO_TCP_SOCK_OR_NULL] = "tcp_sock_or_null",
561 [PTR_TO_TP_BUFFER] = "tp_buffer",
562 [PTR_TO_XDP_SOCK] = "xdp_sock",
563 [PTR_TO_BTF_ID] = "ptr_",
564 [PTR_TO_BTF_ID_OR_NULL] = "ptr_or_null_",
565 [PTR_TO_PERCPU_BTF_ID] = "percpu_ptr_",
566 [PTR_TO_MEM] = "mem",
567 [PTR_TO_MEM_OR_NULL] = "mem_or_null",
568 [PTR_TO_RDONLY_BUF] = "rdonly_buf",
569 [PTR_TO_RDONLY_BUF_OR_NULL] = "rdonly_buf_or_null",
570 [PTR_TO_RDWR_BUF] = "rdwr_buf",
571 [PTR_TO_RDWR_BUF_OR_NULL] = "rdwr_buf_or_null",
572 [PTR_TO_FUNC] = "func",
573 [PTR_TO_MAP_KEY] = "map_key",
576 static char slot_type_char[] = {
577 [STACK_INVALID] = '?',
583 static void print_liveness(struct bpf_verifier_env *env,
584 enum bpf_reg_liveness live)
586 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
588 if (live & REG_LIVE_READ)
590 if (live & REG_LIVE_WRITTEN)
592 if (live & REG_LIVE_DONE)
596 static struct bpf_func_state *func(struct bpf_verifier_env *env,
597 const struct bpf_reg_state *reg)
599 struct bpf_verifier_state *cur = env->cur_state;
601 return cur->frame[reg->frameno];
604 static const char *kernel_type_name(const struct btf* btf, u32 id)
606 return btf_name_by_offset(btf, btf_type_by_id(btf, id)->name_off);
609 /* The reg state of a pointer or a bounded scalar was saved when
610 * it was spilled to the stack.
612 static bool is_spilled_reg(const struct bpf_stack_state *stack)
614 return stack->slot_type[BPF_REG_SIZE - 1] == STACK_SPILL;
617 static void scrub_spilled_slot(u8 *stype)
619 if (*stype != STACK_INVALID)
623 static void print_verifier_state(struct bpf_verifier_env *env,
624 const struct bpf_func_state *state)
626 const struct bpf_reg_state *reg;
631 verbose(env, " frame%d:", state->frameno);
632 for (i = 0; i < MAX_BPF_REG; i++) {
633 reg = &state->regs[i];
637 verbose(env, " R%d", i);
638 print_liveness(env, reg->live);
639 verbose(env, "=%s", reg_type_str[t]);
640 if (t == SCALAR_VALUE && reg->precise)
642 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
643 tnum_is_const(reg->var_off)) {
644 /* reg->off should be 0 for SCALAR_VALUE */
645 verbose(env, "%lld", reg->var_off.value + reg->off);
647 if (t == PTR_TO_BTF_ID ||
648 t == PTR_TO_BTF_ID_OR_NULL ||
649 t == PTR_TO_PERCPU_BTF_ID)
650 verbose(env, "%s", kernel_type_name(reg->btf, reg->btf_id));
651 verbose(env, "(id=%d", reg->id);
652 if (reg_type_may_be_refcounted_or_null(t))
653 verbose(env, ",ref_obj_id=%d", reg->ref_obj_id);
654 if (t != SCALAR_VALUE)
655 verbose(env, ",off=%d", reg->off);
656 if (type_is_pkt_pointer(t))
657 verbose(env, ",r=%d", reg->range);
658 else if (t == CONST_PTR_TO_MAP ||
659 t == PTR_TO_MAP_KEY ||
660 t == PTR_TO_MAP_VALUE ||
661 t == PTR_TO_MAP_VALUE_OR_NULL)
662 verbose(env, ",ks=%d,vs=%d",
663 reg->map_ptr->key_size,
664 reg->map_ptr->value_size);
665 if (tnum_is_const(reg->var_off)) {
666 /* Typically an immediate SCALAR_VALUE, but
667 * could be a pointer whose offset is too big
670 verbose(env, ",imm=%llx", reg->var_off.value);
672 if (reg->smin_value != reg->umin_value &&
673 reg->smin_value != S64_MIN)
674 verbose(env, ",smin_value=%lld",
675 (long long)reg->smin_value);
676 if (reg->smax_value != reg->umax_value &&
677 reg->smax_value != S64_MAX)
678 verbose(env, ",smax_value=%lld",
679 (long long)reg->smax_value);
680 if (reg->umin_value != 0)
681 verbose(env, ",umin_value=%llu",
682 (unsigned long long)reg->umin_value);
683 if (reg->umax_value != U64_MAX)
684 verbose(env, ",umax_value=%llu",
685 (unsigned long long)reg->umax_value);
686 if (!tnum_is_unknown(reg->var_off)) {
689 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
690 verbose(env, ",var_off=%s", tn_buf);
692 if (reg->s32_min_value != reg->smin_value &&
693 reg->s32_min_value != S32_MIN)
694 verbose(env, ",s32_min_value=%d",
695 (int)(reg->s32_min_value));
696 if (reg->s32_max_value != reg->smax_value &&
697 reg->s32_max_value != S32_MAX)
698 verbose(env, ",s32_max_value=%d",
699 (int)(reg->s32_max_value));
700 if (reg->u32_min_value != reg->umin_value &&
701 reg->u32_min_value != U32_MIN)
702 verbose(env, ",u32_min_value=%d",
703 (int)(reg->u32_min_value));
704 if (reg->u32_max_value != reg->umax_value &&
705 reg->u32_max_value != U32_MAX)
706 verbose(env, ",u32_max_value=%d",
707 (int)(reg->u32_max_value));
712 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
713 char types_buf[BPF_REG_SIZE + 1];
717 for (j = 0; j < BPF_REG_SIZE; j++) {
718 if (state->stack[i].slot_type[j] != STACK_INVALID)
720 types_buf[j] = slot_type_char[
721 state->stack[i].slot_type[j]];
723 types_buf[BPF_REG_SIZE] = 0;
726 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
727 print_liveness(env, state->stack[i].spilled_ptr.live);
728 if (is_spilled_reg(&state->stack[i])) {
729 reg = &state->stack[i].spilled_ptr;
731 verbose(env, "=%s", reg_type_str[t]);
732 if (t == SCALAR_VALUE && reg->precise)
734 if (t == SCALAR_VALUE && tnum_is_const(reg->var_off))
735 verbose(env, "%lld", reg->var_off.value + reg->off);
737 verbose(env, "=%s", types_buf);
740 if (state->acquired_refs && state->refs[0].id) {
741 verbose(env, " refs=%d", state->refs[0].id);
742 for (i = 1; i < state->acquired_refs; i++)
743 if (state->refs[i].id)
744 verbose(env, ",%d", state->refs[i].id);
746 if (state->in_callback_fn)
748 if (state->in_async_callback_fn)
749 verbose(env, " async_cb");
753 /* copy array src of length n * size bytes to dst. dst is reallocated if it's too
754 * small to hold src. This is different from krealloc since we don't want to preserve
755 * the contents of dst.
757 * Leaves dst untouched if src is NULL or length is zero. Returns NULL if memory could
760 static void *copy_array(void *dst, const void *src, size_t n, size_t size, gfp_t flags)
764 if (ZERO_OR_NULL_PTR(src))
767 if (unlikely(check_mul_overflow(n, size, &bytes)))
770 if (ksize(dst) < bytes) {
772 dst = kmalloc_track_caller(bytes, flags);
777 memcpy(dst, src, bytes);
779 return dst ? dst : ZERO_SIZE_PTR;
782 /* resize an array from old_n items to new_n items. the array is reallocated if it's too
783 * small to hold new_n items. new items are zeroed out if the array grows.
785 * Contrary to krealloc_array, does not free arr if new_n is zero.
787 static void *realloc_array(void *arr, size_t old_n, size_t new_n, size_t size)
789 if (!new_n || old_n == new_n)
792 arr = krealloc_array(arr, new_n, size, GFP_KERNEL);
797 memset(arr + old_n * size, 0, (new_n - old_n) * size);
800 return arr ? arr : ZERO_SIZE_PTR;
803 static int copy_reference_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
805 dst->refs = copy_array(dst->refs, src->refs, src->acquired_refs,
806 sizeof(struct bpf_reference_state), GFP_KERNEL);
810 dst->acquired_refs = src->acquired_refs;
814 static int copy_stack_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
816 size_t n = src->allocated_stack / BPF_REG_SIZE;
818 dst->stack = copy_array(dst->stack, src->stack, n, sizeof(struct bpf_stack_state),
823 dst->allocated_stack = src->allocated_stack;
827 static int resize_reference_state(struct bpf_func_state *state, size_t n)
829 state->refs = realloc_array(state->refs, state->acquired_refs, n,
830 sizeof(struct bpf_reference_state));
834 state->acquired_refs = n;
838 static int grow_stack_state(struct bpf_func_state *state, int size)
840 size_t old_n = state->allocated_stack / BPF_REG_SIZE, n = size / BPF_REG_SIZE;
845 state->stack = realloc_array(state->stack, old_n, n, sizeof(struct bpf_stack_state));
849 state->allocated_stack = size;
853 /* Acquire a pointer id from the env and update the state->refs to include
854 * this new pointer reference.
855 * On success, returns a valid pointer id to associate with the register
856 * On failure, returns a negative errno.
858 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
860 struct bpf_func_state *state = cur_func(env);
861 int new_ofs = state->acquired_refs;
864 err = resize_reference_state(state, state->acquired_refs + 1);
868 state->refs[new_ofs].id = id;
869 state->refs[new_ofs].insn_idx = insn_idx;
874 /* release function corresponding to acquire_reference_state(). Idempotent. */
875 static int release_reference_state(struct bpf_func_state *state, int ptr_id)
879 last_idx = state->acquired_refs - 1;
880 for (i = 0; i < state->acquired_refs; i++) {
881 if (state->refs[i].id == ptr_id) {
882 if (last_idx && i != last_idx)
883 memcpy(&state->refs[i], &state->refs[last_idx],
884 sizeof(*state->refs));
885 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
886 state->acquired_refs--;
893 static void free_func_state(struct bpf_func_state *state)
902 static void clear_jmp_history(struct bpf_verifier_state *state)
904 kfree(state->jmp_history);
905 state->jmp_history = NULL;
906 state->jmp_history_cnt = 0;
909 static void free_verifier_state(struct bpf_verifier_state *state,
914 for (i = 0; i <= state->curframe; i++) {
915 free_func_state(state->frame[i]);
916 state->frame[i] = NULL;
918 clear_jmp_history(state);
923 /* copy verifier state from src to dst growing dst stack space
924 * when necessary to accommodate larger src stack
926 static int copy_func_state(struct bpf_func_state *dst,
927 const struct bpf_func_state *src)
931 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
932 err = copy_reference_state(dst, src);
935 return copy_stack_state(dst, src);
938 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
939 const struct bpf_verifier_state *src)
941 struct bpf_func_state *dst;
944 dst_state->jmp_history = copy_array(dst_state->jmp_history, src->jmp_history,
945 src->jmp_history_cnt, sizeof(struct bpf_idx_pair),
947 if (!dst_state->jmp_history)
949 dst_state->jmp_history_cnt = src->jmp_history_cnt;
951 /* if dst has more stack frames then src frame, free them */
952 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
953 free_func_state(dst_state->frame[i]);
954 dst_state->frame[i] = NULL;
956 dst_state->speculative = src->speculative;
957 dst_state->curframe = src->curframe;
958 dst_state->active_spin_lock = src->active_spin_lock;
959 dst_state->branches = src->branches;
960 dst_state->parent = src->parent;
961 dst_state->first_insn_idx = src->first_insn_idx;
962 dst_state->last_insn_idx = src->last_insn_idx;
963 for (i = 0; i <= src->curframe; i++) {
964 dst = dst_state->frame[i];
966 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
969 dst_state->frame[i] = dst;
971 err = copy_func_state(dst, src->frame[i]);
978 static void update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
981 u32 br = --st->branches;
983 /* WARN_ON(br > 1) technically makes sense here,
984 * but see comment in push_stack(), hence:
986 WARN_ONCE((int)br < 0,
987 "BUG update_branch_counts:branches_to_explore=%d\n",
995 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
996 int *insn_idx, bool pop_log)
998 struct bpf_verifier_state *cur = env->cur_state;
999 struct bpf_verifier_stack_elem *elem, *head = env->head;
1002 if (env->head == NULL)
1006 err = copy_verifier_state(cur, &head->st);
1011 bpf_vlog_reset(&env->log, head->log_pos);
1013 *insn_idx = head->insn_idx;
1015 *prev_insn_idx = head->prev_insn_idx;
1017 free_verifier_state(&head->st, false);
1024 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
1025 int insn_idx, int prev_insn_idx,
1028 struct bpf_verifier_state *cur = env->cur_state;
1029 struct bpf_verifier_stack_elem *elem;
1032 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
1036 elem->insn_idx = insn_idx;
1037 elem->prev_insn_idx = prev_insn_idx;
1038 elem->next = env->head;
1039 elem->log_pos = env->log.len_used;
1042 err = copy_verifier_state(&elem->st, cur);
1045 elem->st.speculative |= speculative;
1046 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
1047 verbose(env, "The sequence of %d jumps is too complex.\n",
1051 if (elem->st.parent) {
1052 ++elem->st.parent->branches;
1053 /* WARN_ON(branches > 2) technically makes sense here,
1055 * 1. speculative states will bump 'branches' for non-branch
1057 * 2. is_state_visited() heuristics may decide not to create
1058 * a new state for a sequence of branches and all such current
1059 * and cloned states will be pointing to a single parent state
1060 * which might have large 'branches' count.
1065 free_verifier_state(env->cur_state, true);
1066 env->cur_state = NULL;
1067 /* pop all elements and return */
1068 while (!pop_stack(env, NULL, NULL, false));
1072 #define CALLER_SAVED_REGS 6
1073 static const int caller_saved[CALLER_SAVED_REGS] = {
1074 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
1077 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1078 struct bpf_reg_state *reg);
1080 /* This helper doesn't clear reg->id */
1081 static void ___mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1083 reg->var_off = tnum_const(imm);
1084 reg->smin_value = (s64)imm;
1085 reg->smax_value = (s64)imm;
1086 reg->umin_value = imm;
1087 reg->umax_value = imm;
1089 reg->s32_min_value = (s32)imm;
1090 reg->s32_max_value = (s32)imm;
1091 reg->u32_min_value = (u32)imm;
1092 reg->u32_max_value = (u32)imm;
1095 /* Mark the unknown part of a register (variable offset or scalar value) as
1096 * known to have the value @imm.
1098 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1100 /* Clear id, off, and union(map_ptr, range) */
1101 memset(((u8 *)reg) + sizeof(reg->type), 0,
1102 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
1103 ___mark_reg_known(reg, imm);
1106 static void __mark_reg32_known(struct bpf_reg_state *reg, u64 imm)
1108 reg->var_off = tnum_const_subreg(reg->var_off, imm);
1109 reg->s32_min_value = (s32)imm;
1110 reg->s32_max_value = (s32)imm;
1111 reg->u32_min_value = (u32)imm;
1112 reg->u32_max_value = (u32)imm;
1115 /* Mark the 'variable offset' part of a register as zero. This should be
1116 * used only on registers holding a pointer type.
1118 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
1120 __mark_reg_known(reg, 0);
1123 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
1125 __mark_reg_known(reg, 0);
1126 reg->type = SCALAR_VALUE;
1129 static void mark_reg_known_zero(struct bpf_verifier_env *env,
1130 struct bpf_reg_state *regs, u32 regno)
1132 if (WARN_ON(regno >= MAX_BPF_REG)) {
1133 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
1134 /* Something bad happened, let's kill all regs */
1135 for (regno = 0; regno < MAX_BPF_REG; regno++)
1136 __mark_reg_not_init(env, regs + regno);
1139 __mark_reg_known_zero(regs + regno);
1142 static void mark_ptr_not_null_reg(struct bpf_reg_state *reg)
1144 switch (reg->type) {
1145 case PTR_TO_MAP_VALUE_OR_NULL: {
1146 const struct bpf_map *map = reg->map_ptr;
1148 if (map->inner_map_meta) {
1149 reg->type = CONST_PTR_TO_MAP;
1150 reg->map_ptr = map->inner_map_meta;
1151 /* transfer reg's id which is unique for every map_lookup_elem
1152 * as UID of the inner map.
1154 if (map_value_has_timer(map->inner_map_meta))
1155 reg->map_uid = reg->id;
1156 } else if (map->map_type == BPF_MAP_TYPE_XSKMAP) {
1157 reg->type = PTR_TO_XDP_SOCK;
1158 } else if (map->map_type == BPF_MAP_TYPE_SOCKMAP ||
1159 map->map_type == BPF_MAP_TYPE_SOCKHASH) {
1160 reg->type = PTR_TO_SOCKET;
1162 reg->type = PTR_TO_MAP_VALUE;
1166 case PTR_TO_SOCKET_OR_NULL:
1167 reg->type = PTR_TO_SOCKET;
1169 case PTR_TO_SOCK_COMMON_OR_NULL:
1170 reg->type = PTR_TO_SOCK_COMMON;
1172 case PTR_TO_TCP_SOCK_OR_NULL:
1173 reg->type = PTR_TO_TCP_SOCK;
1175 case PTR_TO_BTF_ID_OR_NULL:
1176 reg->type = PTR_TO_BTF_ID;
1178 case PTR_TO_MEM_OR_NULL:
1179 reg->type = PTR_TO_MEM;
1181 case PTR_TO_RDONLY_BUF_OR_NULL:
1182 reg->type = PTR_TO_RDONLY_BUF;
1184 case PTR_TO_RDWR_BUF_OR_NULL:
1185 reg->type = PTR_TO_RDWR_BUF;
1188 WARN_ONCE(1, "unknown nullable register type");
1192 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
1194 return type_is_pkt_pointer(reg->type);
1197 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
1199 return reg_is_pkt_pointer(reg) ||
1200 reg->type == PTR_TO_PACKET_END;
1203 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
1204 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
1205 enum bpf_reg_type which)
1207 /* The register can already have a range from prior markings.
1208 * This is fine as long as it hasn't been advanced from its
1211 return reg->type == which &&
1214 tnum_equals_const(reg->var_off, 0);
1217 /* Reset the min/max bounds of a register */
1218 static void __mark_reg_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;
1225 reg->s32_min_value = S32_MIN;
1226 reg->s32_max_value = S32_MAX;
1227 reg->u32_min_value = 0;
1228 reg->u32_max_value = U32_MAX;
1231 static void __mark_reg64_unbounded(struct bpf_reg_state *reg)
1233 reg->smin_value = S64_MIN;
1234 reg->smax_value = S64_MAX;
1235 reg->umin_value = 0;
1236 reg->umax_value = U64_MAX;
1239 static void __mark_reg32_unbounded(struct bpf_reg_state *reg)
1241 reg->s32_min_value = S32_MIN;
1242 reg->s32_max_value = S32_MAX;
1243 reg->u32_min_value = 0;
1244 reg->u32_max_value = U32_MAX;
1247 static void __update_reg32_bounds(struct bpf_reg_state *reg)
1249 struct tnum var32_off = tnum_subreg(reg->var_off);
1251 /* min signed is max(sign bit) | min(other bits) */
1252 reg->s32_min_value = max_t(s32, reg->s32_min_value,
1253 var32_off.value | (var32_off.mask & S32_MIN));
1254 /* max signed is min(sign bit) | max(other bits) */
1255 reg->s32_max_value = min_t(s32, reg->s32_max_value,
1256 var32_off.value | (var32_off.mask & S32_MAX));
1257 reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)var32_off.value);
1258 reg->u32_max_value = min(reg->u32_max_value,
1259 (u32)(var32_off.value | var32_off.mask));
1262 static void __update_reg64_bounds(struct bpf_reg_state *reg)
1264 /* min signed is max(sign bit) | min(other bits) */
1265 reg->smin_value = max_t(s64, reg->smin_value,
1266 reg->var_off.value | (reg->var_off.mask & S64_MIN));
1267 /* max signed is min(sign bit) | max(other bits) */
1268 reg->smax_value = min_t(s64, reg->smax_value,
1269 reg->var_off.value | (reg->var_off.mask & S64_MAX));
1270 reg->umin_value = max(reg->umin_value, reg->var_off.value);
1271 reg->umax_value = min(reg->umax_value,
1272 reg->var_off.value | reg->var_off.mask);
1275 static void __update_reg_bounds(struct bpf_reg_state *reg)
1277 __update_reg32_bounds(reg);
1278 __update_reg64_bounds(reg);
1281 /* Uses signed min/max values to inform unsigned, and vice-versa */
1282 static void __reg32_deduce_bounds(struct bpf_reg_state *reg)
1284 /* Learn sign from signed bounds.
1285 * If we cannot cross the sign boundary, then signed and unsigned bounds
1286 * are the same, so combine. This works even in the negative case, e.g.
1287 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1289 if (reg->s32_min_value >= 0 || reg->s32_max_value < 0) {
1290 reg->s32_min_value = reg->u32_min_value =
1291 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1292 reg->s32_max_value = reg->u32_max_value =
1293 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1296 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1297 * boundary, so we must be careful.
1299 if ((s32)reg->u32_max_value >= 0) {
1300 /* Positive. We can't learn anything from the smin, but smax
1301 * is positive, hence safe.
1303 reg->s32_min_value = reg->u32_min_value;
1304 reg->s32_max_value = reg->u32_max_value =
1305 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1306 } else if ((s32)reg->u32_min_value < 0) {
1307 /* Negative. We can't learn anything from the smax, but smin
1308 * is negative, hence safe.
1310 reg->s32_min_value = reg->u32_min_value =
1311 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1312 reg->s32_max_value = reg->u32_max_value;
1316 static void __reg64_deduce_bounds(struct bpf_reg_state *reg)
1318 /* Learn sign from signed bounds.
1319 * If we cannot cross the sign boundary, then signed and unsigned bounds
1320 * are the same, so combine. This works even in the negative case, e.g.
1321 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1323 if (reg->smin_value >= 0 || reg->smax_value < 0) {
1324 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1326 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1330 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1331 * boundary, so we must be careful.
1333 if ((s64)reg->umax_value >= 0) {
1334 /* Positive. We can't learn anything from the smin, but smax
1335 * is positive, hence safe.
1337 reg->smin_value = reg->umin_value;
1338 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1340 } else if ((s64)reg->umin_value < 0) {
1341 /* Negative. We can't learn anything from the smax, but smin
1342 * is negative, hence safe.
1344 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1346 reg->smax_value = reg->umax_value;
1350 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
1352 __reg32_deduce_bounds(reg);
1353 __reg64_deduce_bounds(reg);
1356 /* Attempts to improve var_off based on unsigned min/max information */
1357 static void __reg_bound_offset(struct bpf_reg_state *reg)
1359 struct tnum var64_off = tnum_intersect(reg->var_off,
1360 tnum_range(reg->umin_value,
1362 struct tnum var32_off = tnum_intersect(tnum_subreg(reg->var_off),
1363 tnum_range(reg->u32_min_value,
1364 reg->u32_max_value));
1366 reg->var_off = tnum_or(tnum_clear_subreg(var64_off), var32_off);
1369 static void __reg_assign_32_into_64(struct bpf_reg_state *reg)
1371 reg->umin_value = reg->u32_min_value;
1372 reg->umax_value = reg->u32_max_value;
1373 /* Attempt to pull 32-bit signed bounds into 64-bit bounds
1374 * but must be positive otherwise set to worse case bounds
1375 * and refine later from tnum.
1377 if (reg->s32_min_value >= 0 && reg->s32_max_value >= 0)
1378 reg->smax_value = reg->s32_max_value;
1380 reg->smax_value = U32_MAX;
1381 if (reg->s32_min_value >= 0)
1382 reg->smin_value = reg->s32_min_value;
1384 reg->smin_value = 0;
1387 static void __reg_combine_32_into_64(struct bpf_reg_state *reg)
1389 /* special case when 64-bit register has upper 32-bit register
1390 * zeroed. Typically happens after zext or <<32, >>32 sequence
1391 * allowing us to use 32-bit bounds directly,
1393 if (tnum_equals_const(tnum_clear_subreg(reg->var_off), 0)) {
1394 __reg_assign_32_into_64(reg);
1396 /* Otherwise the best we can do is push lower 32bit known and
1397 * unknown bits into register (var_off set from jmp logic)
1398 * then learn as much as possible from the 64-bit tnum
1399 * known and unknown bits. The previous smin/smax bounds are
1400 * invalid here because of jmp32 compare so mark them unknown
1401 * so they do not impact tnum bounds calculation.
1403 __mark_reg64_unbounded(reg);
1404 __update_reg_bounds(reg);
1407 /* Intersecting with the old var_off might have improved our bounds
1408 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1409 * then new var_off is (0; 0x7f...fc) which improves our umax.
1411 __reg_deduce_bounds(reg);
1412 __reg_bound_offset(reg);
1413 __update_reg_bounds(reg);
1416 static bool __reg64_bound_s32(s64 a)
1418 return a >= S32_MIN && a <= S32_MAX;
1421 static bool __reg64_bound_u32(u64 a)
1423 return a >= U32_MIN && a <= U32_MAX;
1426 static void __reg_combine_64_into_32(struct bpf_reg_state *reg)
1428 __mark_reg32_unbounded(reg);
1430 if (__reg64_bound_s32(reg->smin_value) && __reg64_bound_s32(reg->smax_value)) {
1431 reg->s32_min_value = (s32)reg->smin_value;
1432 reg->s32_max_value = (s32)reg->smax_value;
1434 if (__reg64_bound_u32(reg->umin_value) && __reg64_bound_u32(reg->umax_value)) {
1435 reg->u32_min_value = (u32)reg->umin_value;
1436 reg->u32_max_value = (u32)reg->umax_value;
1439 /* Intersecting with the old var_off might have improved our bounds
1440 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1441 * then new var_off is (0; 0x7f...fc) which improves our umax.
1443 __reg_deduce_bounds(reg);
1444 __reg_bound_offset(reg);
1445 __update_reg_bounds(reg);
1448 /* Mark a register as having a completely unknown (scalar) value. */
1449 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
1450 struct bpf_reg_state *reg)
1453 * Clear type, id, off, and union(map_ptr, range) and
1454 * padding between 'type' and union
1456 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
1457 reg->type = SCALAR_VALUE;
1458 reg->var_off = tnum_unknown;
1460 reg->precise = env->subprog_cnt > 1 || !env->bpf_capable;
1461 __mark_reg_unbounded(reg);
1464 static void mark_reg_unknown(struct bpf_verifier_env *env,
1465 struct bpf_reg_state *regs, u32 regno)
1467 if (WARN_ON(regno >= MAX_BPF_REG)) {
1468 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
1469 /* Something bad happened, let's kill all regs except FP */
1470 for (regno = 0; regno < BPF_REG_FP; regno++)
1471 __mark_reg_not_init(env, regs + regno);
1474 __mark_reg_unknown(env, regs + regno);
1477 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1478 struct bpf_reg_state *reg)
1480 __mark_reg_unknown(env, reg);
1481 reg->type = NOT_INIT;
1484 static void mark_reg_not_init(struct bpf_verifier_env *env,
1485 struct bpf_reg_state *regs, u32 regno)
1487 if (WARN_ON(regno >= MAX_BPF_REG)) {
1488 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
1489 /* Something bad happened, let's kill all regs except FP */
1490 for (regno = 0; regno < BPF_REG_FP; regno++)
1491 __mark_reg_not_init(env, regs + regno);
1494 __mark_reg_not_init(env, regs + regno);
1497 static void mark_btf_ld_reg(struct bpf_verifier_env *env,
1498 struct bpf_reg_state *regs, u32 regno,
1499 enum bpf_reg_type reg_type,
1500 struct btf *btf, u32 btf_id)
1502 if (reg_type == SCALAR_VALUE) {
1503 mark_reg_unknown(env, regs, regno);
1506 mark_reg_known_zero(env, regs, regno);
1507 regs[regno].type = PTR_TO_BTF_ID;
1508 regs[regno].btf = btf;
1509 regs[regno].btf_id = btf_id;
1512 #define DEF_NOT_SUBREG (0)
1513 static void init_reg_state(struct bpf_verifier_env *env,
1514 struct bpf_func_state *state)
1516 struct bpf_reg_state *regs = state->regs;
1519 for (i = 0; i < MAX_BPF_REG; i++) {
1520 mark_reg_not_init(env, regs, i);
1521 regs[i].live = REG_LIVE_NONE;
1522 regs[i].parent = NULL;
1523 regs[i].subreg_def = DEF_NOT_SUBREG;
1527 regs[BPF_REG_FP].type = PTR_TO_STACK;
1528 mark_reg_known_zero(env, regs, BPF_REG_FP);
1529 regs[BPF_REG_FP].frameno = state->frameno;
1532 #define BPF_MAIN_FUNC (-1)
1533 static void init_func_state(struct bpf_verifier_env *env,
1534 struct bpf_func_state *state,
1535 int callsite, int frameno, int subprogno)
1537 state->callsite = callsite;
1538 state->frameno = frameno;
1539 state->subprogno = subprogno;
1540 init_reg_state(env, state);
1543 /* Similar to push_stack(), but for async callbacks */
1544 static struct bpf_verifier_state *push_async_cb(struct bpf_verifier_env *env,
1545 int insn_idx, int prev_insn_idx,
1548 struct bpf_verifier_stack_elem *elem;
1549 struct bpf_func_state *frame;
1551 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
1555 elem->insn_idx = insn_idx;
1556 elem->prev_insn_idx = prev_insn_idx;
1557 elem->next = env->head;
1558 elem->log_pos = env->log.len_used;
1561 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
1563 "The sequence of %d jumps is too complex for async cb.\n",
1567 /* Unlike push_stack() do not copy_verifier_state().
1568 * The caller state doesn't matter.
1569 * This is async callback. It starts in a fresh stack.
1570 * Initialize it similar to do_check_common().
1572 elem->st.branches = 1;
1573 frame = kzalloc(sizeof(*frame), GFP_KERNEL);
1576 init_func_state(env, frame,
1577 BPF_MAIN_FUNC /* callsite */,
1578 0 /* frameno within this callchain */,
1579 subprog /* subprog number within this prog */);
1580 elem->st.frame[0] = frame;
1583 free_verifier_state(env->cur_state, true);
1584 env->cur_state = NULL;
1585 /* pop all elements and return */
1586 while (!pop_stack(env, NULL, NULL, false));
1592 SRC_OP, /* register is used as source operand */
1593 DST_OP, /* register is used as destination operand */
1594 DST_OP_NO_MARK /* same as above, check only, don't mark */
1597 static int cmp_subprogs(const void *a, const void *b)
1599 return ((struct bpf_subprog_info *)a)->start -
1600 ((struct bpf_subprog_info *)b)->start;
1603 static int find_subprog(struct bpf_verifier_env *env, int off)
1605 struct bpf_subprog_info *p;
1607 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
1608 sizeof(env->subprog_info[0]), cmp_subprogs);
1611 return p - env->subprog_info;
1615 static int add_subprog(struct bpf_verifier_env *env, int off)
1617 int insn_cnt = env->prog->len;
1620 if (off >= insn_cnt || off < 0) {
1621 verbose(env, "call to invalid destination\n");
1624 ret = find_subprog(env, off);
1627 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
1628 verbose(env, "too many subprograms\n");
1631 /* determine subprog starts. The end is one before the next starts */
1632 env->subprog_info[env->subprog_cnt++].start = off;
1633 sort(env->subprog_info, env->subprog_cnt,
1634 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
1635 return env->subprog_cnt - 1;
1638 #define MAX_KFUNC_DESCS 256
1639 #define MAX_KFUNC_BTFS 256
1641 struct bpf_kfunc_desc {
1642 struct btf_func_model func_model;
1648 struct bpf_kfunc_btf {
1650 struct module *module;
1654 struct bpf_kfunc_desc_tab {
1655 struct bpf_kfunc_desc descs[MAX_KFUNC_DESCS];
1659 struct bpf_kfunc_btf_tab {
1660 struct bpf_kfunc_btf descs[MAX_KFUNC_BTFS];
1664 static int kfunc_desc_cmp_by_id_off(const void *a, const void *b)
1666 const struct bpf_kfunc_desc *d0 = a;
1667 const struct bpf_kfunc_desc *d1 = b;
1669 /* func_id is not greater than BTF_MAX_TYPE */
1670 return d0->func_id - d1->func_id ?: d0->offset - d1->offset;
1673 static int kfunc_btf_cmp_by_off(const void *a, const void *b)
1675 const struct bpf_kfunc_btf *d0 = a;
1676 const struct bpf_kfunc_btf *d1 = b;
1678 return d0->offset - d1->offset;
1681 static const struct bpf_kfunc_desc *
1682 find_kfunc_desc(const struct bpf_prog *prog, u32 func_id, u16 offset)
1684 struct bpf_kfunc_desc desc = {
1688 struct bpf_kfunc_desc_tab *tab;
1690 tab = prog->aux->kfunc_tab;
1691 return bsearch(&desc, tab->descs, tab->nr_descs,
1692 sizeof(tab->descs[0]), kfunc_desc_cmp_by_id_off);
1695 static struct btf *__find_kfunc_desc_btf(struct bpf_verifier_env *env,
1696 s16 offset, struct module **btf_modp)
1698 struct bpf_kfunc_btf kf_btf = { .offset = offset };
1699 struct bpf_kfunc_btf_tab *tab;
1700 struct bpf_kfunc_btf *b;
1705 tab = env->prog->aux->kfunc_btf_tab;
1706 b = bsearch(&kf_btf, tab->descs, tab->nr_descs,
1707 sizeof(tab->descs[0]), kfunc_btf_cmp_by_off);
1709 if (tab->nr_descs == MAX_KFUNC_BTFS) {
1710 verbose(env, "too many different module BTFs\n");
1711 return ERR_PTR(-E2BIG);
1714 if (bpfptr_is_null(env->fd_array)) {
1715 verbose(env, "kfunc offset > 0 without fd_array is invalid\n");
1716 return ERR_PTR(-EPROTO);
1719 if (copy_from_bpfptr_offset(&btf_fd, env->fd_array,
1720 offset * sizeof(btf_fd),
1722 return ERR_PTR(-EFAULT);
1724 btf = btf_get_by_fd(btf_fd);
1726 verbose(env, "invalid module BTF fd specified\n");
1730 if (!btf_is_module(btf)) {
1731 verbose(env, "BTF fd for kfunc is not a module BTF\n");
1733 return ERR_PTR(-EINVAL);
1736 mod = btf_try_get_module(btf);
1739 return ERR_PTR(-ENXIO);
1742 b = &tab->descs[tab->nr_descs++];
1747 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
1748 kfunc_btf_cmp_by_off, NULL);
1751 *btf_modp = b->module;
1755 void bpf_free_kfunc_btf_tab(struct bpf_kfunc_btf_tab *tab)
1760 while (tab->nr_descs--) {
1761 module_put(tab->descs[tab->nr_descs].module);
1762 btf_put(tab->descs[tab->nr_descs].btf);
1767 static struct btf *find_kfunc_desc_btf(struct bpf_verifier_env *env,
1768 u32 func_id, s16 offset,
1769 struct module **btf_modp)
1773 /* In the future, this can be allowed to increase limit
1774 * of fd index into fd_array, interpreted as u16.
1776 verbose(env, "negative offset disallowed for kernel module function call\n");
1777 return ERR_PTR(-EINVAL);
1780 return __find_kfunc_desc_btf(env, offset, btf_modp);
1782 return btf_vmlinux ?: ERR_PTR(-ENOENT);
1785 static int add_kfunc_call(struct bpf_verifier_env *env, u32 func_id, s16 offset)
1787 const struct btf_type *func, *func_proto;
1788 struct bpf_kfunc_btf_tab *btf_tab;
1789 struct bpf_kfunc_desc_tab *tab;
1790 struct bpf_prog_aux *prog_aux;
1791 struct bpf_kfunc_desc *desc;
1792 const char *func_name;
1793 struct btf *desc_btf;
1797 prog_aux = env->prog->aux;
1798 tab = prog_aux->kfunc_tab;
1799 btf_tab = prog_aux->kfunc_btf_tab;
1802 verbose(env, "calling kernel function is not supported without CONFIG_DEBUG_INFO_BTF\n");
1806 if (!env->prog->jit_requested) {
1807 verbose(env, "JIT is required for calling kernel function\n");
1811 if (!bpf_jit_supports_kfunc_call()) {
1812 verbose(env, "JIT does not support calling kernel function\n");
1816 if (!env->prog->gpl_compatible) {
1817 verbose(env, "cannot call kernel function from non-GPL compatible program\n");
1821 tab = kzalloc(sizeof(*tab), GFP_KERNEL);
1824 prog_aux->kfunc_tab = tab;
1827 /* func_id == 0 is always invalid, but instead of returning an error, be
1828 * conservative and wait until the code elimination pass before returning
1829 * error, so that invalid calls that get pruned out can be in BPF programs
1830 * loaded from userspace. It is also required that offset be untouched
1833 if (!func_id && !offset)
1836 if (!btf_tab && offset) {
1837 btf_tab = kzalloc(sizeof(*btf_tab), GFP_KERNEL);
1840 prog_aux->kfunc_btf_tab = btf_tab;
1843 desc_btf = find_kfunc_desc_btf(env, func_id, offset, NULL);
1844 if (IS_ERR(desc_btf)) {
1845 verbose(env, "failed to find BTF for kernel function\n");
1846 return PTR_ERR(desc_btf);
1849 if (find_kfunc_desc(env->prog, func_id, offset))
1852 if (tab->nr_descs == MAX_KFUNC_DESCS) {
1853 verbose(env, "too many different kernel function calls\n");
1857 func = btf_type_by_id(desc_btf, func_id);
1858 if (!func || !btf_type_is_func(func)) {
1859 verbose(env, "kernel btf_id %u is not a function\n",
1863 func_proto = btf_type_by_id(desc_btf, func->type);
1864 if (!func_proto || !btf_type_is_func_proto(func_proto)) {
1865 verbose(env, "kernel function btf_id %u does not have a valid func_proto\n",
1870 func_name = btf_name_by_offset(desc_btf, func->name_off);
1871 addr = kallsyms_lookup_name(func_name);
1873 verbose(env, "cannot find address for kernel function %s\n",
1878 desc = &tab->descs[tab->nr_descs++];
1879 desc->func_id = func_id;
1880 desc->imm = BPF_CALL_IMM(addr);
1881 desc->offset = offset;
1882 err = btf_distill_func_proto(&env->log, desc_btf,
1883 func_proto, func_name,
1886 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
1887 kfunc_desc_cmp_by_id_off, NULL);
1891 static int kfunc_desc_cmp_by_imm(const void *a, const void *b)
1893 const struct bpf_kfunc_desc *d0 = a;
1894 const struct bpf_kfunc_desc *d1 = b;
1896 if (d0->imm > d1->imm)
1898 else if (d0->imm < d1->imm)
1903 static void sort_kfunc_descs_by_imm(struct bpf_prog *prog)
1905 struct bpf_kfunc_desc_tab *tab;
1907 tab = prog->aux->kfunc_tab;
1911 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
1912 kfunc_desc_cmp_by_imm, NULL);
1915 bool bpf_prog_has_kfunc_call(const struct bpf_prog *prog)
1917 return !!prog->aux->kfunc_tab;
1920 const struct btf_func_model *
1921 bpf_jit_find_kfunc_model(const struct bpf_prog *prog,
1922 const struct bpf_insn *insn)
1924 const struct bpf_kfunc_desc desc = {
1927 const struct bpf_kfunc_desc *res;
1928 struct bpf_kfunc_desc_tab *tab;
1930 tab = prog->aux->kfunc_tab;
1931 res = bsearch(&desc, tab->descs, tab->nr_descs,
1932 sizeof(tab->descs[0]), kfunc_desc_cmp_by_imm);
1934 return res ? &res->func_model : NULL;
1937 static int add_subprog_and_kfunc(struct bpf_verifier_env *env)
1939 struct bpf_subprog_info *subprog = env->subprog_info;
1940 struct bpf_insn *insn = env->prog->insnsi;
1941 int i, ret, insn_cnt = env->prog->len;
1943 /* Add entry function. */
1944 ret = add_subprog(env, 0);
1948 for (i = 0; i < insn_cnt; i++, insn++) {
1949 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn) &&
1950 !bpf_pseudo_kfunc_call(insn))
1953 if (!env->bpf_capable) {
1954 verbose(env, "loading/calling other bpf or kernel functions are allowed for CAP_BPF and CAP_SYS_ADMIN\n");
1958 if (bpf_pseudo_func(insn) || bpf_pseudo_call(insn))
1959 ret = add_subprog(env, i + insn->imm + 1);
1961 ret = add_kfunc_call(env, insn->imm, insn->off);
1967 /* Add a fake 'exit' subprog which could simplify subprog iteration
1968 * logic. 'subprog_cnt' should not be increased.
1970 subprog[env->subprog_cnt].start = insn_cnt;
1972 if (env->log.level & BPF_LOG_LEVEL2)
1973 for (i = 0; i < env->subprog_cnt; i++)
1974 verbose(env, "func#%d @%d\n", i, subprog[i].start);
1979 static int check_subprogs(struct bpf_verifier_env *env)
1981 int i, subprog_start, subprog_end, off, cur_subprog = 0;
1982 struct bpf_subprog_info *subprog = env->subprog_info;
1983 struct bpf_insn *insn = env->prog->insnsi;
1984 int insn_cnt = env->prog->len;
1986 /* now check that all jumps are within the same subprog */
1987 subprog_start = subprog[cur_subprog].start;
1988 subprog_end = subprog[cur_subprog + 1].start;
1989 for (i = 0; i < insn_cnt; i++) {
1990 u8 code = insn[i].code;
1992 if (code == (BPF_JMP | BPF_CALL) &&
1993 insn[i].imm == BPF_FUNC_tail_call &&
1994 insn[i].src_reg != BPF_PSEUDO_CALL)
1995 subprog[cur_subprog].has_tail_call = true;
1996 if (BPF_CLASS(code) == BPF_LD &&
1997 (BPF_MODE(code) == BPF_ABS || BPF_MODE(code) == BPF_IND))
1998 subprog[cur_subprog].has_ld_abs = true;
1999 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
2001 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
2003 off = i + insn[i].off + 1;
2004 if (off < subprog_start || off >= subprog_end) {
2005 verbose(env, "jump out of range from insn %d to %d\n", i, off);
2009 if (i == subprog_end - 1) {
2010 /* to avoid fall-through from one subprog into another
2011 * the last insn of the subprog should be either exit
2012 * or unconditional jump back
2014 if (code != (BPF_JMP | BPF_EXIT) &&
2015 code != (BPF_JMP | BPF_JA)) {
2016 verbose(env, "last insn is not an exit or jmp\n");
2019 subprog_start = subprog_end;
2021 if (cur_subprog < env->subprog_cnt)
2022 subprog_end = subprog[cur_subprog + 1].start;
2028 /* Parentage chain of this register (or stack slot) should take care of all
2029 * issues like callee-saved registers, stack slot allocation time, etc.
2031 static int mark_reg_read(struct bpf_verifier_env *env,
2032 const struct bpf_reg_state *state,
2033 struct bpf_reg_state *parent, u8 flag)
2035 bool writes = parent == state->parent; /* Observe write marks */
2039 /* if read wasn't screened by an earlier write ... */
2040 if (writes && state->live & REG_LIVE_WRITTEN)
2042 if (parent->live & REG_LIVE_DONE) {
2043 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
2044 reg_type_str[parent->type],
2045 parent->var_off.value, parent->off);
2048 /* The first condition is more likely to be true than the
2049 * second, checked it first.
2051 if ((parent->live & REG_LIVE_READ) == flag ||
2052 parent->live & REG_LIVE_READ64)
2053 /* The parentage chain never changes and
2054 * this parent was already marked as LIVE_READ.
2055 * There is no need to keep walking the chain again and
2056 * keep re-marking all parents as LIVE_READ.
2057 * This case happens when the same register is read
2058 * multiple times without writes into it in-between.
2059 * Also, if parent has the stronger REG_LIVE_READ64 set,
2060 * then no need to set the weak REG_LIVE_READ32.
2063 /* ... then we depend on parent's value */
2064 parent->live |= flag;
2065 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
2066 if (flag == REG_LIVE_READ64)
2067 parent->live &= ~REG_LIVE_READ32;
2069 parent = state->parent;
2074 if (env->longest_mark_read_walk < cnt)
2075 env->longest_mark_read_walk = cnt;
2079 /* This function is supposed to be used by the following 32-bit optimization
2080 * code only. It returns TRUE if the source or destination register operates
2081 * on 64-bit, otherwise return FALSE.
2083 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
2084 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
2089 class = BPF_CLASS(code);
2091 if (class == BPF_JMP) {
2092 /* BPF_EXIT for "main" will reach here. Return TRUE
2097 if (op == BPF_CALL) {
2098 /* BPF to BPF call will reach here because of marking
2099 * caller saved clobber with DST_OP_NO_MARK for which we
2100 * don't care the register def because they are anyway
2101 * marked as NOT_INIT already.
2103 if (insn->src_reg == BPF_PSEUDO_CALL)
2105 /* Helper call will reach here because of arg type
2106 * check, conservatively return TRUE.
2115 if (class == BPF_ALU64 || class == BPF_JMP ||
2116 /* BPF_END always use BPF_ALU class. */
2117 (class == BPF_ALU && op == BPF_END && insn->imm == 64))
2120 if (class == BPF_ALU || class == BPF_JMP32)
2123 if (class == BPF_LDX) {
2125 return BPF_SIZE(code) == BPF_DW;
2126 /* LDX source must be ptr. */
2130 if (class == BPF_STX) {
2131 /* BPF_STX (including atomic variants) has multiple source
2132 * operands, one of which is a ptr. Check whether the caller is
2135 if (t == SRC_OP && reg->type != SCALAR_VALUE)
2137 return BPF_SIZE(code) == BPF_DW;
2140 if (class == BPF_LD) {
2141 u8 mode = BPF_MODE(code);
2144 if (mode == BPF_IMM)
2147 /* Both LD_IND and LD_ABS return 32-bit data. */
2151 /* Implicit ctx ptr. */
2152 if (regno == BPF_REG_6)
2155 /* Explicit source could be any width. */
2159 if (class == BPF_ST)
2160 /* The only source register for BPF_ST is a ptr. */
2163 /* Conservatively return true at default. */
2167 /* Return the regno defined by the insn, or -1. */
2168 static int insn_def_regno(const struct bpf_insn *insn)
2170 switch (BPF_CLASS(insn->code)) {
2176 if (BPF_MODE(insn->code) == BPF_ATOMIC &&
2177 (insn->imm & BPF_FETCH)) {
2178 if (insn->imm == BPF_CMPXCHG)
2181 return insn->src_reg;
2186 return insn->dst_reg;
2190 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
2191 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
2193 int dst_reg = insn_def_regno(insn);
2198 return !is_reg64(env, insn, dst_reg, NULL, DST_OP);
2201 static void mark_insn_zext(struct bpf_verifier_env *env,
2202 struct bpf_reg_state *reg)
2204 s32 def_idx = reg->subreg_def;
2206 if (def_idx == DEF_NOT_SUBREG)
2209 env->insn_aux_data[def_idx - 1].zext_dst = true;
2210 /* The dst will be zero extended, so won't be sub-register anymore. */
2211 reg->subreg_def = DEF_NOT_SUBREG;
2214 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
2215 enum reg_arg_type t)
2217 struct bpf_verifier_state *vstate = env->cur_state;
2218 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2219 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
2220 struct bpf_reg_state *reg, *regs = state->regs;
2223 if (regno >= MAX_BPF_REG) {
2224 verbose(env, "R%d is invalid\n", regno);
2229 rw64 = is_reg64(env, insn, regno, reg, t);
2231 /* check whether register used as source operand can be read */
2232 if (reg->type == NOT_INIT) {
2233 verbose(env, "R%d !read_ok\n", regno);
2236 /* We don't need to worry about FP liveness because it's read-only */
2237 if (regno == BPF_REG_FP)
2241 mark_insn_zext(env, reg);
2243 return mark_reg_read(env, reg, reg->parent,
2244 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
2246 /* check whether register used as dest operand can be written to */
2247 if (regno == BPF_REG_FP) {
2248 verbose(env, "frame pointer is read only\n");
2251 reg->live |= REG_LIVE_WRITTEN;
2252 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
2254 mark_reg_unknown(env, regs, regno);
2259 /* for any branch, call, exit record the history of jmps in the given state */
2260 static int push_jmp_history(struct bpf_verifier_env *env,
2261 struct bpf_verifier_state *cur)
2263 u32 cnt = cur->jmp_history_cnt;
2264 struct bpf_idx_pair *p;
2267 p = krealloc(cur->jmp_history, cnt * sizeof(*p), GFP_USER);
2270 p[cnt - 1].idx = env->insn_idx;
2271 p[cnt - 1].prev_idx = env->prev_insn_idx;
2272 cur->jmp_history = p;
2273 cur->jmp_history_cnt = cnt;
2277 /* Backtrack one insn at a time. If idx is not at the top of recorded
2278 * history then previous instruction came from straight line execution.
2280 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
2285 if (cnt && st->jmp_history[cnt - 1].idx == i) {
2286 i = st->jmp_history[cnt - 1].prev_idx;
2294 static const char *disasm_kfunc_name(void *data, const struct bpf_insn *insn)
2296 const struct btf_type *func;
2297 struct btf *desc_btf;
2299 if (insn->src_reg != BPF_PSEUDO_KFUNC_CALL)
2302 desc_btf = find_kfunc_desc_btf(data, insn->imm, insn->off, NULL);
2303 if (IS_ERR(desc_btf))
2306 func = btf_type_by_id(desc_btf, insn->imm);
2307 return btf_name_by_offset(desc_btf, func->name_off);
2310 /* For given verifier state backtrack_insn() is called from the last insn to
2311 * the first insn. Its purpose is to compute a bitmask of registers and
2312 * stack slots that needs precision in the parent verifier state.
2314 static int backtrack_insn(struct bpf_verifier_env *env, int idx,
2315 u32 *reg_mask, u64 *stack_mask)
2317 const struct bpf_insn_cbs cbs = {
2318 .cb_call = disasm_kfunc_name,
2319 .cb_print = verbose,
2320 .private_data = env,
2322 struct bpf_insn *insn = env->prog->insnsi + idx;
2323 u8 class = BPF_CLASS(insn->code);
2324 u8 opcode = BPF_OP(insn->code);
2325 u8 mode = BPF_MODE(insn->code);
2326 u32 dreg = 1u << insn->dst_reg;
2327 u32 sreg = 1u << insn->src_reg;
2330 if (insn->code == 0)
2332 if (env->log.level & BPF_LOG_LEVEL) {
2333 verbose(env, "regs=%x stack=%llx before ", *reg_mask, *stack_mask);
2334 verbose(env, "%d: ", idx);
2335 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
2338 if (class == BPF_ALU || class == BPF_ALU64) {
2339 if (!(*reg_mask & dreg))
2341 if (opcode == BPF_MOV) {
2342 if (BPF_SRC(insn->code) == BPF_X) {
2344 * dreg needs precision after this insn
2345 * sreg needs precision before this insn
2351 * dreg needs precision after this insn.
2352 * Corresponding register is already marked
2353 * as precise=true in this verifier state.
2354 * No further markings in parent are necessary
2359 if (BPF_SRC(insn->code) == BPF_X) {
2361 * both dreg and sreg need precision
2366 * dreg still needs precision before this insn
2369 } else if (class == BPF_LDX) {
2370 if (!(*reg_mask & dreg))
2374 /* scalars can only be spilled into stack w/o losing precision.
2375 * Load from any other memory can be zero extended.
2376 * The desire to keep that precision is already indicated
2377 * by 'precise' mark in corresponding register of this state.
2378 * No further tracking necessary.
2380 if (insn->src_reg != BPF_REG_FP)
2383 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
2384 * that [fp - off] slot contains scalar that needs to be
2385 * tracked with precision
2387 spi = (-insn->off - 1) / BPF_REG_SIZE;
2389 verbose(env, "BUG spi %d\n", spi);
2390 WARN_ONCE(1, "verifier backtracking bug");
2393 *stack_mask |= 1ull << spi;
2394 } else if (class == BPF_STX || class == BPF_ST) {
2395 if (*reg_mask & dreg)
2396 /* stx & st shouldn't be using _scalar_ dst_reg
2397 * to access memory. It means backtracking
2398 * encountered a case of pointer subtraction.
2401 /* scalars can only be spilled into stack */
2402 if (insn->dst_reg != BPF_REG_FP)
2404 spi = (-insn->off - 1) / BPF_REG_SIZE;
2406 verbose(env, "BUG spi %d\n", spi);
2407 WARN_ONCE(1, "verifier backtracking bug");
2410 if (!(*stack_mask & (1ull << spi)))
2412 *stack_mask &= ~(1ull << spi);
2413 if (class == BPF_STX)
2415 } else if (class == BPF_JMP || class == BPF_JMP32) {
2416 if (opcode == BPF_CALL) {
2417 if (insn->src_reg == BPF_PSEUDO_CALL)
2419 /* regular helper call sets R0 */
2421 if (*reg_mask & 0x3f) {
2422 /* if backtracing was looking for registers R1-R5
2423 * they should have been found already.
2425 verbose(env, "BUG regs %x\n", *reg_mask);
2426 WARN_ONCE(1, "verifier backtracking bug");
2429 } else if (opcode == BPF_EXIT) {
2432 } else if (class == BPF_LD) {
2433 if (!(*reg_mask & dreg))
2436 /* It's ld_imm64 or ld_abs or ld_ind.
2437 * For ld_imm64 no further tracking of precision
2438 * into parent is necessary
2440 if (mode == BPF_IND || mode == BPF_ABS)
2441 /* to be analyzed */
2447 /* the scalar precision tracking algorithm:
2448 * . at the start all registers have precise=false.
2449 * . scalar ranges are tracked as normal through alu and jmp insns.
2450 * . once precise value of the scalar register is used in:
2451 * . ptr + scalar alu
2452 * . if (scalar cond K|scalar)
2453 * . helper_call(.., scalar, ...) where ARG_CONST is expected
2454 * backtrack through the verifier states and mark all registers and
2455 * stack slots with spilled constants that these scalar regisers
2456 * should be precise.
2457 * . during state pruning two registers (or spilled stack slots)
2458 * are equivalent if both are not precise.
2460 * Note the verifier cannot simply walk register parentage chain,
2461 * since many different registers and stack slots could have been
2462 * used to compute single precise scalar.
2464 * The approach of starting with precise=true for all registers and then
2465 * backtrack to mark a register as not precise when the verifier detects
2466 * that program doesn't care about specific value (e.g., when helper
2467 * takes register as ARG_ANYTHING parameter) is not safe.
2469 * It's ok to walk single parentage chain of the verifier states.
2470 * It's possible that this backtracking will go all the way till 1st insn.
2471 * All other branches will be explored for needing precision later.
2473 * The backtracking needs to deal with cases like:
2474 * 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)
2477 * if r5 > 0x79f goto pc+7
2478 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
2481 * call bpf_perf_event_output#25
2482 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
2486 * call foo // uses callee's r6 inside to compute r0
2490 * to track above reg_mask/stack_mask needs to be independent for each frame.
2492 * Also if parent's curframe > frame where backtracking started,
2493 * the verifier need to mark registers in both frames, otherwise callees
2494 * may incorrectly prune callers. This is similar to
2495 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
2497 * For now backtracking falls back into conservative marking.
2499 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
2500 struct bpf_verifier_state *st)
2502 struct bpf_func_state *func;
2503 struct bpf_reg_state *reg;
2506 /* big hammer: mark all scalars precise in this path.
2507 * pop_stack may still get !precise scalars.
2509 for (; st; st = st->parent)
2510 for (i = 0; i <= st->curframe; i++) {
2511 func = st->frame[i];
2512 for (j = 0; j < BPF_REG_FP; j++) {
2513 reg = &func->regs[j];
2514 if (reg->type != SCALAR_VALUE)
2516 reg->precise = true;
2518 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
2519 if (!is_spilled_reg(&func->stack[j]))
2521 reg = &func->stack[j].spilled_ptr;
2522 if (reg->type != SCALAR_VALUE)
2524 reg->precise = true;
2529 static int __mark_chain_precision(struct bpf_verifier_env *env, int regno,
2532 struct bpf_verifier_state *st = env->cur_state;
2533 int first_idx = st->first_insn_idx;
2534 int last_idx = env->insn_idx;
2535 struct bpf_func_state *func;
2536 struct bpf_reg_state *reg;
2537 u32 reg_mask = regno >= 0 ? 1u << regno : 0;
2538 u64 stack_mask = spi >= 0 ? 1ull << spi : 0;
2539 bool skip_first = true;
2540 bool new_marks = false;
2543 if (!env->bpf_capable)
2546 func = st->frame[st->curframe];
2548 reg = &func->regs[regno];
2549 if (reg->type != SCALAR_VALUE) {
2550 WARN_ONCE(1, "backtracing misuse");
2557 reg->precise = true;
2561 if (!is_spilled_reg(&func->stack[spi])) {
2565 reg = &func->stack[spi].spilled_ptr;
2566 if (reg->type != SCALAR_VALUE) {
2574 reg->precise = true;
2580 if (!reg_mask && !stack_mask)
2583 DECLARE_BITMAP(mask, 64);
2584 u32 history = st->jmp_history_cnt;
2586 if (env->log.level & BPF_LOG_LEVEL)
2587 verbose(env, "last_idx %d first_idx %d\n", last_idx, first_idx);
2588 for (i = last_idx;;) {
2593 err = backtrack_insn(env, i, ®_mask, &stack_mask);
2595 if (err == -ENOTSUPP) {
2596 mark_all_scalars_precise(env, st);
2601 if (!reg_mask && !stack_mask)
2602 /* Found assignment(s) into tracked register in this state.
2603 * Since this state is already marked, just return.
2604 * Nothing to be tracked further in the parent state.
2609 i = get_prev_insn_idx(st, i, &history);
2610 if (i >= env->prog->len) {
2611 /* This can happen if backtracking reached insn 0
2612 * and there are still reg_mask or stack_mask
2614 * It means the backtracking missed the spot where
2615 * particular register was initialized with a constant.
2617 verbose(env, "BUG backtracking idx %d\n", i);
2618 WARN_ONCE(1, "verifier backtracking bug");
2627 func = st->frame[st->curframe];
2628 bitmap_from_u64(mask, reg_mask);
2629 for_each_set_bit(i, mask, 32) {
2630 reg = &func->regs[i];
2631 if (reg->type != SCALAR_VALUE) {
2632 reg_mask &= ~(1u << i);
2637 reg->precise = true;
2640 bitmap_from_u64(mask, stack_mask);
2641 for_each_set_bit(i, mask, 64) {
2642 if (i >= func->allocated_stack / BPF_REG_SIZE) {
2643 /* the sequence of instructions:
2645 * 3: (7b) *(u64 *)(r3 -8) = r0
2646 * 4: (79) r4 = *(u64 *)(r10 -8)
2647 * doesn't contain jmps. It's backtracked
2648 * as a single block.
2649 * During backtracking insn 3 is not recognized as
2650 * stack access, so at the end of backtracking
2651 * stack slot fp-8 is still marked in stack_mask.
2652 * However the parent state may not have accessed
2653 * fp-8 and it's "unallocated" stack space.
2654 * In such case fallback to conservative.
2656 mark_all_scalars_precise(env, st);
2660 if (!is_spilled_reg(&func->stack[i])) {
2661 stack_mask &= ~(1ull << i);
2664 reg = &func->stack[i].spilled_ptr;
2665 if (reg->type != SCALAR_VALUE) {
2666 stack_mask &= ~(1ull << i);
2671 reg->precise = true;
2673 if (env->log.level & BPF_LOG_LEVEL) {
2674 print_verifier_state(env, func);
2675 verbose(env, "parent %s regs=%x stack=%llx marks\n",
2676 new_marks ? "didn't have" : "already had",
2677 reg_mask, stack_mask);
2680 if (!reg_mask && !stack_mask)
2685 last_idx = st->last_insn_idx;
2686 first_idx = st->first_insn_idx;
2691 static int mark_chain_precision(struct bpf_verifier_env *env, int regno)
2693 return __mark_chain_precision(env, regno, -1);
2696 static int mark_chain_precision_stack(struct bpf_verifier_env *env, int spi)
2698 return __mark_chain_precision(env, -1, spi);
2701 static bool is_spillable_regtype(enum bpf_reg_type type)
2704 case PTR_TO_MAP_VALUE:
2705 case PTR_TO_MAP_VALUE_OR_NULL:
2709 case PTR_TO_PACKET_META:
2710 case PTR_TO_PACKET_END:
2711 case PTR_TO_FLOW_KEYS:
2712 case CONST_PTR_TO_MAP:
2714 case PTR_TO_SOCKET_OR_NULL:
2715 case PTR_TO_SOCK_COMMON:
2716 case PTR_TO_SOCK_COMMON_OR_NULL:
2717 case PTR_TO_TCP_SOCK:
2718 case PTR_TO_TCP_SOCK_OR_NULL:
2719 case PTR_TO_XDP_SOCK:
2721 case PTR_TO_BTF_ID_OR_NULL:
2722 case PTR_TO_RDONLY_BUF:
2723 case PTR_TO_RDONLY_BUF_OR_NULL:
2724 case PTR_TO_RDWR_BUF:
2725 case PTR_TO_RDWR_BUF_OR_NULL:
2726 case PTR_TO_PERCPU_BTF_ID:
2728 case PTR_TO_MEM_OR_NULL:
2730 case PTR_TO_MAP_KEY:
2737 /* Does this register contain a constant zero? */
2738 static bool register_is_null(struct bpf_reg_state *reg)
2740 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
2743 static bool register_is_const(struct bpf_reg_state *reg)
2745 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
2748 static bool __is_scalar_unbounded(struct bpf_reg_state *reg)
2750 return tnum_is_unknown(reg->var_off) &&
2751 reg->smin_value == S64_MIN && reg->smax_value == S64_MAX &&
2752 reg->umin_value == 0 && reg->umax_value == U64_MAX &&
2753 reg->s32_min_value == S32_MIN && reg->s32_max_value == S32_MAX &&
2754 reg->u32_min_value == 0 && reg->u32_max_value == U32_MAX;
2757 static bool register_is_bounded(struct bpf_reg_state *reg)
2759 return reg->type == SCALAR_VALUE && !__is_scalar_unbounded(reg);
2762 static bool __is_pointer_value(bool allow_ptr_leaks,
2763 const struct bpf_reg_state *reg)
2765 if (allow_ptr_leaks)
2768 return reg->type != SCALAR_VALUE;
2771 static void save_register_state(struct bpf_func_state *state,
2772 int spi, struct bpf_reg_state *reg,
2777 state->stack[spi].spilled_ptr = *reg;
2778 if (size == BPF_REG_SIZE)
2779 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2781 for (i = BPF_REG_SIZE; i > BPF_REG_SIZE - size; i--)
2782 state->stack[spi].slot_type[i - 1] = STACK_SPILL;
2784 /* size < 8 bytes spill */
2786 scrub_spilled_slot(&state->stack[spi].slot_type[i - 1]);
2789 /* check_stack_{read,write}_fixed_off functions track spill/fill of registers,
2790 * stack boundary and alignment are checked in check_mem_access()
2792 static int check_stack_write_fixed_off(struct bpf_verifier_env *env,
2793 /* stack frame we're writing to */
2794 struct bpf_func_state *state,
2795 int off, int size, int value_regno,
2798 struct bpf_func_state *cur; /* state of the current function */
2799 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
2800 u32 dst_reg = env->prog->insnsi[insn_idx].dst_reg;
2801 struct bpf_reg_state *reg = NULL;
2803 err = grow_stack_state(state, round_up(slot + 1, BPF_REG_SIZE));
2806 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
2807 * so it's aligned access and [off, off + size) are within stack limits
2809 if (!env->allow_ptr_leaks &&
2810 state->stack[spi].slot_type[0] == STACK_SPILL &&
2811 size != BPF_REG_SIZE) {
2812 verbose(env, "attempt to corrupt spilled pointer on stack\n");
2816 cur = env->cur_state->frame[env->cur_state->curframe];
2817 if (value_regno >= 0)
2818 reg = &cur->regs[value_regno];
2819 if (!env->bypass_spec_v4) {
2820 bool sanitize = reg && is_spillable_regtype(reg->type);
2822 for (i = 0; i < size; i++) {
2823 if (state->stack[spi].slot_type[i] == STACK_INVALID) {
2830 env->insn_aux_data[insn_idx].sanitize_stack_spill = true;
2833 if (reg && !(off % BPF_REG_SIZE) && register_is_bounded(reg) &&
2834 !register_is_null(reg) && env->bpf_capable) {
2835 if (dst_reg != BPF_REG_FP) {
2836 /* The backtracking logic can only recognize explicit
2837 * stack slot address like [fp - 8]. Other spill of
2838 * scalar via different register has to be conservative.
2839 * Backtrack from here and mark all registers as precise
2840 * that contributed into 'reg' being a constant.
2842 err = mark_chain_precision(env, value_regno);
2846 save_register_state(state, spi, reg, size);
2847 } else if (reg && is_spillable_regtype(reg->type)) {
2848 /* register containing pointer is being spilled into stack */
2849 if (size != BPF_REG_SIZE) {
2850 verbose_linfo(env, insn_idx, "; ");
2851 verbose(env, "invalid size of register spill\n");
2854 if (state != cur && reg->type == PTR_TO_STACK) {
2855 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
2858 save_register_state(state, spi, reg, size);
2860 u8 type = STACK_MISC;
2862 /* regular write of data into stack destroys any spilled ptr */
2863 state->stack[spi].spilled_ptr.type = NOT_INIT;
2864 /* Mark slots as STACK_MISC if they belonged to spilled ptr. */
2865 if (is_spilled_reg(&state->stack[spi]))
2866 for (i = 0; i < BPF_REG_SIZE; i++)
2867 scrub_spilled_slot(&state->stack[spi].slot_type[i]);
2869 /* only mark the slot as written if all 8 bytes were written
2870 * otherwise read propagation may incorrectly stop too soon
2871 * when stack slots are partially written.
2872 * This heuristic means that read propagation will be
2873 * conservative, since it will add reg_live_read marks
2874 * to stack slots all the way to first state when programs
2875 * writes+reads less than 8 bytes
2877 if (size == BPF_REG_SIZE)
2878 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2880 /* when we zero initialize stack slots mark them as such */
2881 if (reg && register_is_null(reg)) {
2882 /* backtracking doesn't work for STACK_ZERO yet. */
2883 err = mark_chain_precision(env, value_regno);
2889 /* Mark slots affected by this stack write. */
2890 for (i = 0; i < size; i++)
2891 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
2897 /* Write the stack: 'stack[ptr_regno + off] = value_regno'. 'ptr_regno' is
2898 * known to contain a variable offset.
2899 * This function checks whether the write is permitted and conservatively
2900 * tracks the effects of the write, considering that each stack slot in the
2901 * dynamic range is potentially written to.
2903 * 'off' includes 'regno->off'.
2904 * 'value_regno' can be -1, meaning that an unknown value is being written to
2907 * Spilled pointers in range are not marked as written because we don't know
2908 * what's going to be actually written. This means that read propagation for
2909 * future reads cannot be terminated by this write.
2911 * For privileged programs, uninitialized stack slots are considered
2912 * initialized by this write (even though we don't know exactly what offsets
2913 * are going to be written to). The idea is that we don't want the verifier to
2914 * reject future reads that access slots written to through variable offsets.
2916 static int check_stack_write_var_off(struct bpf_verifier_env *env,
2917 /* func where register points to */
2918 struct bpf_func_state *state,
2919 int ptr_regno, int off, int size,
2920 int value_regno, int insn_idx)
2922 struct bpf_func_state *cur; /* state of the current function */
2923 int min_off, max_off;
2925 struct bpf_reg_state *ptr_reg = NULL, *value_reg = NULL;
2926 bool writing_zero = false;
2927 /* set if the fact that we're writing a zero is used to let any
2928 * stack slots remain STACK_ZERO
2930 bool zero_used = false;
2932 cur = env->cur_state->frame[env->cur_state->curframe];
2933 ptr_reg = &cur->regs[ptr_regno];
2934 min_off = ptr_reg->smin_value + off;
2935 max_off = ptr_reg->smax_value + off + size;
2936 if (value_regno >= 0)
2937 value_reg = &cur->regs[value_regno];
2938 if (value_reg && register_is_null(value_reg))
2939 writing_zero = true;
2941 err = grow_stack_state(state, round_up(-min_off, BPF_REG_SIZE));
2946 /* Variable offset writes destroy any spilled pointers in range. */
2947 for (i = min_off; i < max_off; i++) {
2948 u8 new_type, *stype;
2952 spi = slot / BPF_REG_SIZE;
2953 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
2955 if (!env->allow_ptr_leaks
2956 && *stype != NOT_INIT
2957 && *stype != SCALAR_VALUE) {
2958 /* Reject the write if there's are spilled pointers in
2959 * range. If we didn't reject here, the ptr status
2960 * would be erased below (even though not all slots are
2961 * actually overwritten), possibly opening the door to
2964 verbose(env, "spilled ptr in range of var-offset stack write; insn %d, ptr off: %d",
2969 /* Erase all spilled pointers. */
2970 state->stack[spi].spilled_ptr.type = NOT_INIT;
2972 /* Update the slot type. */
2973 new_type = STACK_MISC;
2974 if (writing_zero && *stype == STACK_ZERO) {
2975 new_type = STACK_ZERO;
2978 /* If the slot is STACK_INVALID, we check whether it's OK to
2979 * pretend that it will be initialized by this write. The slot
2980 * might not actually be written to, and so if we mark it as
2981 * initialized future reads might leak uninitialized memory.
2982 * For privileged programs, we will accept such reads to slots
2983 * that may or may not be written because, if we're reject
2984 * them, the error would be too confusing.
2986 if (*stype == STACK_INVALID && !env->allow_uninit_stack) {
2987 verbose(env, "uninit stack in range of var-offset write prohibited for !root; insn %d, off: %d",
2994 /* backtracking doesn't work for STACK_ZERO yet. */
2995 err = mark_chain_precision(env, value_regno);
3002 /* When register 'dst_regno' is assigned some values from stack[min_off,
3003 * max_off), we set the register's type according to the types of the
3004 * respective stack slots. If all the stack values are known to be zeros, then
3005 * so is the destination reg. Otherwise, the register is considered to be
3006 * SCALAR. This function does not deal with register filling; the caller must
3007 * ensure that all spilled registers in the stack range have been marked as
3010 static void mark_reg_stack_read(struct bpf_verifier_env *env,
3011 /* func where src register points to */
3012 struct bpf_func_state *ptr_state,
3013 int min_off, int max_off, int dst_regno)
3015 struct bpf_verifier_state *vstate = env->cur_state;
3016 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3021 for (i = min_off; i < max_off; i++) {
3023 spi = slot / BPF_REG_SIZE;
3024 stype = ptr_state->stack[spi].slot_type;
3025 if (stype[slot % BPF_REG_SIZE] != STACK_ZERO)
3029 if (zeros == max_off - min_off) {
3030 /* any access_size read into register is zero extended,
3031 * so the whole register == const_zero
3033 __mark_reg_const_zero(&state->regs[dst_regno]);
3034 /* backtracking doesn't support STACK_ZERO yet,
3035 * so mark it precise here, so that later
3036 * backtracking can stop here.
3037 * Backtracking may not need this if this register
3038 * doesn't participate in pointer adjustment.
3039 * Forward propagation of precise flag is not
3040 * necessary either. This mark is only to stop
3041 * backtracking. Any register that contributed
3042 * to const 0 was marked precise before spill.
3044 state->regs[dst_regno].precise = true;
3046 /* have read misc data from the stack */
3047 mark_reg_unknown(env, state->regs, dst_regno);
3049 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
3052 /* Read the stack at 'off' and put the results into the register indicated by
3053 * 'dst_regno'. It handles reg filling if the addressed stack slot is a
3056 * 'dst_regno' can be -1, meaning that the read value is not going to a
3059 * The access is assumed to be within the current stack bounds.
3061 static int check_stack_read_fixed_off(struct bpf_verifier_env *env,
3062 /* func where src register points to */
3063 struct bpf_func_state *reg_state,
3064 int off, int size, int dst_regno)
3066 struct bpf_verifier_state *vstate = env->cur_state;
3067 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3068 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
3069 struct bpf_reg_state *reg;
3072 stype = reg_state->stack[spi].slot_type;
3073 reg = ®_state->stack[spi].spilled_ptr;
3075 if (is_spilled_reg(®_state->stack[spi])) {
3078 for (i = BPF_REG_SIZE - 1; i > 0 && stype[i - 1] == STACK_SPILL; i--)
3081 if (size != BPF_REG_SIZE || spill_size != BPF_REG_SIZE) {
3082 if (reg->type != SCALAR_VALUE) {
3083 verbose_linfo(env, env->insn_idx, "; ");
3084 verbose(env, "invalid size of register fill\n");
3088 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
3092 if (!(off % BPF_REG_SIZE) && size == spill_size) {
3093 /* The earlier check_reg_arg() has decided the
3094 * subreg_def for this insn. Save it first.
3096 s32 subreg_def = state->regs[dst_regno].subreg_def;
3098 state->regs[dst_regno] = *reg;
3099 state->regs[dst_regno].subreg_def = subreg_def;
3101 for (i = 0; i < size; i++) {
3102 type = stype[(slot - i) % BPF_REG_SIZE];
3103 if (type == STACK_SPILL)
3105 if (type == STACK_MISC)
3107 verbose(env, "invalid read from stack off %d+%d size %d\n",
3111 mark_reg_unknown(env, state->regs, dst_regno);
3113 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
3117 if (dst_regno >= 0) {
3118 /* restore register state from stack */
3119 state->regs[dst_regno] = *reg;
3120 /* mark reg as written since spilled pointer state likely
3121 * has its liveness marks cleared by is_state_visited()
3122 * which resets stack/reg liveness for state transitions
3124 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
3125 } else if (__is_pointer_value(env->allow_ptr_leaks, reg)) {
3126 /* If dst_regno==-1, the caller is asking us whether
3127 * it is acceptable to use this value as a SCALAR_VALUE
3129 * We must not allow unprivileged callers to do that
3130 * with spilled pointers.
3132 verbose(env, "leaking pointer from stack off %d\n",
3136 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
3138 for (i = 0; i < size; i++) {
3139 type = stype[(slot - i) % BPF_REG_SIZE];
3140 if (type == STACK_MISC)
3142 if (type == STACK_ZERO)
3144 verbose(env, "invalid read from stack off %d+%d size %d\n",
3148 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
3150 mark_reg_stack_read(env, reg_state, off, off + size, dst_regno);
3155 enum stack_access_src {
3156 ACCESS_DIRECT = 1, /* the access is performed by an instruction */
3157 ACCESS_HELPER = 2, /* the access is performed by a helper */
3160 static int check_stack_range_initialized(struct bpf_verifier_env *env,
3161 int regno, int off, int access_size,
3162 bool zero_size_allowed,
3163 enum stack_access_src type,
3164 struct bpf_call_arg_meta *meta);
3166 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
3168 return cur_regs(env) + regno;
3171 /* Read the stack at 'ptr_regno + off' and put the result into the register
3173 * 'off' includes the pointer register's fixed offset(i.e. 'ptr_regno.off'),
3174 * but not its variable offset.
3175 * 'size' is assumed to be <= reg size and the access is assumed to be aligned.
3177 * As opposed to check_stack_read_fixed_off, this function doesn't deal with
3178 * filling registers (i.e. reads of spilled register cannot be detected when
3179 * the offset is not fixed). We conservatively mark 'dst_regno' as containing
3180 * SCALAR_VALUE. That's why we assert that the 'ptr_regno' has a variable
3181 * offset; for a fixed offset check_stack_read_fixed_off should be used
3184 static int check_stack_read_var_off(struct bpf_verifier_env *env,
3185 int ptr_regno, int off, int size, int dst_regno)
3187 /* The state of the source register. */
3188 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
3189 struct bpf_func_state *ptr_state = func(env, reg);
3191 int min_off, max_off;
3193 /* Note that we pass a NULL meta, so raw access will not be permitted.
3195 err = check_stack_range_initialized(env, ptr_regno, off, size,
3196 false, ACCESS_DIRECT, NULL);
3200 min_off = reg->smin_value + off;
3201 max_off = reg->smax_value + off;
3202 mark_reg_stack_read(env, ptr_state, min_off, max_off + size, dst_regno);
3206 /* check_stack_read dispatches to check_stack_read_fixed_off or
3207 * check_stack_read_var_off.
3209 * The caller must ensure that the offset falls within the allocated stack
3212 * 'dst_regno' is a register which will receive the value from the stack. It
3213 * can be -1, meaning that the read value is not going to a register.
3215 static int check_stack_read(struct bpf_verifier_env *env,
3216 int ptr_regno, int off, int size,
3219 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
3220 struct bpf_func_state *state = func(env, reg);
3222 /* Some accesses are only permitted with a static offset. */
3223 bool var_off = !tnum_is_const(reg->var_off);
3225 /* The offset is required to be static when reads don't go to a
3226 * register, in order to not leak pointers (see
3227 * check_stack_read_fixed_off).
3229 if (dst_regno < 0 && var_off) {
3232 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3233 verbose(env, "variable offset stack pointer cannot be passed into helper function; var_off=%s off=%d size=%d\n",
3237 /* Variable offset is prohibited for unprivileged mode for simplicity
3238 * since it requires corresponding support in Spectre masking for stack
3239 * ALU. See also retrieve_ptr_limit().
3241 if (!env->bypass_spec_v1 && var_off) {
3244 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3245 verbose(env, "R%d variable offset stack access prohibited for !root, var_off=%s\n",
3251 off += reg->var_off.value;
3252 err = check_stack_read_fixed_off(env, state, off, size,
3255 /* Variable offset stack reads need more conservative handling
3256 * than fixed offset ones. Note that dst_regno >= 0 on this
3259 err = check_stack_read_var_off(env, ptr_regno, off, size,
3266 /* check_stack_write dispatches to check_stack_write_fixed_off or
3267 * check_stack_write_var_off.
3269 * 'ptr_regno' is the register used as a pointer into the stack.
3270 * 'off' includes 'ptr_regno->off', but not its variable offset (if any).
3271 * 'value_regno' is the register whose value we're writing to the stack. It can
3272 * be -1, meaning that we're not writing from a register.
3274 * The caller must ensure that the offset falls within the maximum stack size.
3276 static int check_stack_write(struct bpf_verifier_env *env,
3277 int ptr_regno, int off, int size,
3278 int value_regno, int insn_idx)
3280 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
3281 struct bpf_func_state *state = func(env, reg);
3284 if (tnum_is_const(reg->var_off)) {
3285 off += reg->var_off.value;
3286 err = check_stack_write_fixed_off(env, state, off, size,
3287 value_regno, insn_idx);
3289 /* Variable offset stack reads need more conservative handling
3290 * than fixed offset ones.
3292 err = check_stack_write_var_off(env, state,
3293 ptr_regno, off, size,
3294 value_regno, insn_idx);
3299 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
3300 int off, int size, enum bpf_access_type type)
3302 struct bpf_reg_state *regs = cur_regs(env);
3303 struct bpf_map *map = regs[regno].map_ptr;
3304 u32 cap = bpf_map_flags_to_cap(map);
3306 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
3307 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
3308 map->value_size, off, size);
3312 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
3313 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
3314 map->value_size, off, size);
3321 /* check read/write into memory region (e.g., map value, ringbuf sample, etc) */
3322 static int __check_mem_access(struct bpf_verifier_env *env, int regno,
3323 int off, int size, u32 mem_size,
3324 bool zero_size_allowed)
3326 bool size_ok = size > 0 || (size == 0 && zero_size_allowed);
3327 struct bpf_reg_state *reg;
3329 if (off >= 0 && size_ok && (u64)off + size <= mem_size)
3332 reg = &cur_regs(env)[regno];
3333 switch (reg->type) {
3334 case PTR_TO_MAP_KEY:
3335 verbose(env, "invalid access to map key, key_size=%d off=%d size=%d\n",
3336 mem_size, off, size);
3338 case PTR_TO_MAP_VALUE:
3339 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
3340 mem_size, off, size);
3343 case PTR_TO_PACKET_META:
3344 case PTR_TO_PACKET_END:
3345 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
3346 off, size, regno, reg->id, off, mem_size);
3350 verbose(env, "invalid access to memory, mem_size=%u off=%d size=%d\n",
3351 mem_size, off, size);
3357 /* check read/write into a memory region with possible variable offset */
3358 static int check_mem_region_access(struct bpf_verifier_env *env, u32 regno,
3359 int off, int size, u32 mem_size,
3360 bool zero_size_allowed)
3362 struct bpf_verifier_state *vstate = env->cur_state;
3363 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3364 struct bpf_reg_state *reg = &state->regs[regno];
3367 /* We may have adjusted the register pointing to memory region, so we
3368 * need to try adding each of min_value and max_value to off
3369 * to make sure our theoretical access will be safe.
3371 if (env->log.level & BPF_LOG_LEVEL)
3372 print_verifier_state(env, state);
3374 /* The minimum value is only important with signed
3375 * comparisons where we can't assume the floor of a
3376 * value is 0. If we are using signed variables for our
3377 * index'es we need to make sure that whatever we use
3378 * will have a set floor within our range.
3380 if (reg->smin_value < 0 &&
3381 (reg->smin_value == S64_MIN ||
3382 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
3383 reg->smin_value + off < 0)) {
3384 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3388 err = __check_mem_access(env, regno, reg->smin_value + off, size,
3389 mem_size, zero_size_allowed);
3391 verbose(env, "R%d min value is outside of the allowed memory range\n",
3396 /* If we haven't set a max value then we need to bail since we can't be
3397 * sure we won't do bad things.
3398 * If reg->umax_value + off could overflow, treat that as unbounded too.
3400 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
3401 verbose(env, "R%d unbounded memory access, make sure to bounds check any such access\n",
3405 err = __check_mem_access(env, regno, reg->umax_value + off, size,
3406 mem_size, zero_size_allowed);
3408 verbose(env, "R%d max value is outside of the allowed memory range\n",
3416 /* check read/write into a map element with possible variable offset */
3417 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
3418 int off, int size, bool zero_size_allowed)
3420 struct bpf_verifier_state *vstate = env->cur_state;
3421 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3422 struct bpf_reg_state *reg = &state->regs[regno];
3423 struct bpf_map *map = reg->map_ptr;
3426 err = check_mem_region_access(env, regno, off, size, map->value_size,
3431 if (map_value_has_spin_lock(map)) {
3432 u32 lock = map->spin_lock_off;
3434 /* if any part of struct bpf_spin_lock can be touched by
3435 * load/store reject this program.
3436 * To check that [x1, x2) overlaps with [y1, y2)
3437 * it is sufficient to check x1 < y2 && y1 < x2.
3439 if (reg->smin_value + off < lock + sizeof(struct bpf_spin_lock) &&
3440 lock < reg->umax_value + off + size) {
3441 verbose(env, "bpf_spin_lock cannot be accessed directly by load/store\n");
3445 if (map_value_has_timer(map)) {
3446 u32 t = map->timer_off;
3448 if (reg->smin_value + off < t + sizeof(struct bpf_timer) &&
3449 t < reg->umax_value + off + size) {
3450 verbose(env, "bpf_timer cannot be accessed directly by load/store\n");
3457 #define MAX_PACKET_OFF 0xffff
3459 static enum bpf_prog_type resolve_prog_type(struct bpf_prog *prog)
3461 return prog->aux->dst_prog ? prog->aux->dst_prog->type : prog->type;
3464 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
3465 const struct bpf_call_arg_meta *meta,
3466 enum bpf_access_type t)
3468 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
3470 switch (prog_type) {
3471 /* Program types only with direct read access go here! */
3472 case BPF_PROG_TYPE_LWT_IN:
3473 case BPF_PROG_TYPE_LWT_OUT:
3474 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
3475 case BPF_PROG_TYPE_SK_REUSEPORT:
3476 case BPF_PROG_TYPE_FLOW_DISSECTOR:
3477 case BPF_PROG_TYPE_CGROUP_SKB:
3482 /* Program types with direct read + write access go here! */
3483 case BPF_PROG_TYPE_SCHED_CLS:
3484 case BPF_PROG_TYPE_SCHED_ACT:
3485 case BPF_PROG_TYPE_XDP:
3486 case BPF_PROG_TYPE_LWT_XMIT:
3487 case BPF_PROG_TYPE_SK_SKB:
3488 case BPF_PROG_TYPE_SK_MSG:
3490 return meta->pkt_access;
3492 env->seen_direct_write = true;
3495 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
3497 env->seen_direct_write = true;
3506 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
3507 int size, bool zero_size_allowed)
3509 struct bpf_reg_state *regs = cur_regs(env);
3510 struct bpf_reg_state *reg = ®s[regno];
3513 /* We may have added a variable offset to the packet pointer; but any
3514 * reg->range we have comes after that. We are only checking the fixed
3518 /* We don't allow negative numbers, because we aren't tracking enough
3519 * detail to prove they're safe.
3521 if (reg->smin_value < 0) {
3522 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3527 err = reg->range < 0 ? -EINVAL :
3528 __check_mem_access(env, regno, off, size, reg->range,
3531 verbose(env, "R%d offset is outside of the packet\n", regno);
3535 /* __check_mem_access has made sure "off + size - 1" is within u16.
3536 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
3537 * otherwise find_good_pkt_pointers would have refused to set range info
3538 * that __check_mem_access would have rejected this pkt access.
3539 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
3541 env->prog->aux->max_pkt_offset =
3542 max_t(u32, env->prog->aux->max_pkt_offset,
3543 off + reg->umax_value + size - 1);
3548 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
3549 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
3550 enum bpf_access_type t, enum bpf_reg_type *reg_type,
3551 struct btf **btf, u32 *btf_id)
3553 struct bpf_insn_access_aux info = {
3554 .reg_type = *reg_type,
3558 if (env->ops->is_valid_access &&
3559 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
3560 /* A non zero info.ctx_field_size indicates that this field is a
3561 * candidate for later verifier transformation to load the whole
3562 * field and then apply a mask when accessed with a narrower
3563 * access than actual ctx access size. A zero info.ctx_field_size
3564 * will only allow for whole field access and rejects any other
3565 * type of narrower access.
3567 *reg_type = info.reg_type;
3569 if (*reg_type == PTR_TO_BTF_ID || *reg_type == PTR_TO_BTF_ID_OR_NULL) {
3571 *btf_id = info.btf_id;
3573 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
3575 /* remember the offset of last byte accessed in ctx */
3576 if (env->prog->aux->max_ctx_offset < off + size)
3577 env->prog->aux->max_ctx_offset = off + size;
3581 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
3585 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
3588 if (size < 0 || off < 0 ||
3589 (u64)off + size > sizeof(struct bpf_flow_keys)) {
3590 verbose(env, "invalid access to flow keys off=%d size=%d\n",
3597 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
3598 u32 regno, int off, int size,
3599 enum bpf_access_type t)
3601 struct bpf_reg_state *regs = cur_regs(env);
3602 struct bpf_reg_state *reg = ®s[regno];
3603 struct bpf_insn_access_aux info = {};
3606 if (reg->smin_value < 0) {
3607 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3612 switch (reg->type) {
3613 case PTR_TO_SOCK_COMMON:
3614 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
3617 valid = bpf_sock_is_valid_access(off, size, t, &info);
3619 case PTR_TO_TCP_SOCK:
3620 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
3622 case PTR_TO_XDP_SOCK:
3623 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
3631 env->insn_aux_data[insn_idx].ctx_field_size =
3632 info.ctx_field_size;
3636 verbose(env, "R%d invalid %s access off=%d size=%d\n",
3637 regno, reg_type_str[reg->type], off, size);
3642 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
3644 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
3647 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
3649 const struct bpf_reg_state *reg = reg_state(env, regno);
3651 return reg->type == PTR_TO_CTX;
3654 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
3656 const struct bpf_reg_state *reg = reg_state(env, regno);
3658 return type_is_sk_pointer(reg->type);
3661 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
3663 const struct bpf_reg_state *reg = reg_state(env, regno);
3665 return type_is_pkt_pointer(reg->type);
3668 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
3670 const struct bpf_reg_state *reg = reg_state(env, regno);
3672 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
3673 return reg->type == PTR_TO_FLOW_KEYS;
3676 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
3677 const struct bpf_reg_state *reg,
3678 int off, int size, bool strict)
3680 struct tnum reg_off;
3683 /* Byte size accesses are always allowed. */
3684 if (!strict || size == 1)
3687 /* For platforms that do not have a Kconfig enabling
3688 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
3689 * NET_IP_ALIGN is universally set to '2'. And on platforms
3690 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
3691 * to this code only in strict mode where we want to emulate
3692 * the NET_IP_ALIGN==2 checking. Therefore use an
3693 * unconditional IP align value of '2'.
3697 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
3698 if (!tnum_is_aligned(reg_off, size)) {
3701 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3703 "misaligned packet access off %d+%s+%d+%d size %d\n",
3704 ip_align, tn_buf, reg->off, off, size);
3711 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
3712 const struct bpf_reg_state *reg,
3713 const char *pointer_desc,
3714 int off, int size, bool strict)
3716 struct tnum reg_off;
3718 /* Byte size accesses are always allowed. */
3719 if (!strict || size == 1)
3722 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
3723 if (!tnum_is_aligned(reg_off, size)) {
3726 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3727 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
3728 pointer_desc, tn_buf, reg->off, off, size);
3735 static int check_ptr_alignment(struct bpf_verifier_env *env,
3736 const struct bpf_reg_state *reg, int off,
3737 int size, bool strict_alignment_once)
3739 bool strict = env->strict_alignment || strict_alignment_once;
3740 const char *pointer_desc = "";
3742 switch (reg->type) {
3744 case PTR_TO_PACKET_META:
3745 /* Special case, because of NET_IP_ALIGN. Given metadata sits
3746 * right in front, treat it the very same way.
3748 return check_pkt_ptr_alignment(env, reg, off, size, strict);
3749 case PTR_TO_FLOW_KEYS:
3750 pointer_desc = "flow keys ";
3752 case PTR_TO_MAP_KEY:
3753 pointer_desc = "key ";
3755 case PTR_TO_MAP_VALUE:
3756 pointer_desc = "value ";
3759 pointer_desc = "context ";
3762 pointer_desc = "stack ";
3763 /* The stack spill tracking logic in check_stack_write_fixed_off()
3764 * and check_stack_read_fixed_off() relies on stack accesses being
3770 pointer_desc = "sock ";
3772 case PTR_TO_SOCK_COMMON:
3773 pointer_desc = "sock_common ";
3775 case PTR_TO_TCP_SOCK:
3776 pointer_desc = "tcp_sock ";
3778 case PTR_TO_XDP_SOCK:
3779 pointer_desc = "xdp_sock ";
3784 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
3788 static int update_stack_depth(struct bpf_verifier_env *env,
3789 const struct bpf_func_state *func,
3792 u16 stack = env->subprog_info[func->subprogno].stack_depth;
3797 /* update known max for given subprogram */
3798 env->subprog_info[func->subprogno].stack_depth = -off;
3802 /* starting from main bpf function walk all instructions of the function
3803 * and recursively walk all callees that given function can call.
3804 * Ignore jump and exit insns.
3805 * Since recursion is prevented by check_cfg() this algorithm
3806 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
3808 static int check_max_stack_depth(struct bpf_verifier_env *env)
3810 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
3811 struct bpf_subprog_info *subprog = env->subprog_info;
3812 struct bpf_insn *insn = env->prog->insnsi;
3813 bool tail_call_reachable = false;
3814 int ret_insn[MAX_CALL_FRAMES];
3815 int ret_prog[MAX_CALL_FRAMES];
3819 /* protect against potential stack overflow that might happen when
3820 * bpf2bpf calls get combined with tailcalls. Limit the caller's stack
3821 * depth for such case down to 256 so that the worst case scenario
3822 * would result in 8k stack size (32 which is tailcall limit * 256 =
3825 * To get the idea what might happen, see an example:
3826 * func1 -> sub rsp, 128
3827 * subfunc1 -> sub rsp, 256
3828 * tailcall1 -> add rsp, 256
3829 * func2 -> sub rsp, 192 (total stack size = 128 + 192 = 320)
3830 * subfunc2 -> sub rsp, 64
3831 * subfunc22 -> sub rsp, 128
3832 * tailcall2 -> add rsp, 128
3833 * func3 -> sub rsp, 32 (total stack size 128 + 192 + 64 + 32 = 416)
3835 * tailcall will unwind the current stack frame but it will not get rid
3836 * of caller's stack as shown on the example above.
3838 if (idx && subprog[idx].has_tail_call && depth >= 256) {
3840 "tail_calls are not allowed when call stack of previous frames is %d bytes. Too large\n",
3844 /* round up to 32-bytes, since this is granularity
3845 * of interpreter stack size
3847 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
3848 if (depth > MAX_BPF_STACK) {
3849 verbose(env, "combined stack size of %d calls is %d. Too large\n",
3854 subprog_end = subprog[idx + 1].start;
3855 for (; i < subprog_end; i++) {
3858 if (!bpf_pseudo_call(insn + i) && !bpf_pseudo_func(insn + i))
3860 /* remember insn and function to return to */
3861 ret_insn[frame] = i + 1;
3862 ret_prog[frame] = idx;
3864 /* find the callee */
3865 next_insn = i + insn[i].imm + 1;
3866 idx = find_subprog(env, next_insn);
3868 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3872 if (subprog[idx].is_async_cb) {
3873 if (subprog[idx].has_tail_call) {
3874 verbose(env, "verifier bug. subprog has tail_call and async cb\n");
3877 /* async callbacks don't increase bpf prog stack size */
3882 if (subprog[idx].has_tail_call)
3883 tail_call_reachable = true;
3886 if (frame >= MAX_CALL_FRAMES) {
3887 verbose(env, "the call stack of %d frames is too deep !\n",
3893 /* if tail call got detected across bpf2bpf calls then mark each of the
3894 * currently present subprog frames as tail call reachable subprogs;
3895 * this info will be utilized by JIT so that we will be preserving the
3896 * tail call counter throughout bpf2bpf calls combined with tailcalls
3898 if (tail_call_reachable)
3899 for (j = 0; j < frame; j++)
3900 subprog[ret_prog[j]].tail_call_reachable = true;
3901 if (subprog[0].tail_call_reachable)
3902 env->prog->aux->tail_call_reachable = true;
3904 /* end of for() loop means the last insn of the 'subprog'
3905 * was reached. Doesn't matter whether it was JA or EXIT
3909 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
3911 i = ret_insn[frame];
3912 idx = ret_prog[frame];
3916 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
3917 static int get_callee_stack_depth(struct bpf_verifier_env *env,
3918 const struct bpf_insn *insn, int idx)
3920 int start = idx + insn->imm + 1, subprog;
3922 subprog = find_subprog(env, start);
3924 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3928 return env->subprog_info[subprog].stack_depth;
3932 int check_ctx_reg(struct bpf_verifier_env *env,
3933 const struct bpf_reg_state *reg, int regno)
3935 /* Access to ctx or passing it to a helper is only allowed in
3936 * its original, unmodified form.
3940 verbose(env, "dereference of modified ctx ptr R%d off=%d disallowed\n",
3945 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3948 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3949 verbose(env, "variable ctx access var_off=%s disallowed\n", tn_buf);
3956 static int __check_buffer_access(struct bpf_verifier_env *env,
3957 const char *buf_info,
3958 const struct bpf_reg_state *reg,
3959 int regno, int off, int size)
3963 "R%d invalid %s buffer access: off=%d, size=%d\n",
3964 regno, buf_info, off, size);
3967 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3970 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3972 "R%d invalid variable buffer offset: off=%d, var_off=%s\n",
3973 regno, off, tn_buf);
3980 static int check_tp_buffer_access(struct bpf_verifier_env *env,
3981 const struct bpf_reg_state *reg,
3982 int regno, int off, int size)
3986 err = __check_buffer_access(env, "tracepoint", reg, regno, off, size);
3990 if (off + size > env->prog->aux->max_tp_access)
3991 env->prog->aux->max_tp_access = off + size;
3996 static int check_buffer_access(struct bpf_verifier_env *env,
3997 const struct bpf_reg_state *reg,
3998 int regno, int off, int size,
3999 bool zero_size_allowed,
4000 const char *buf_info,
4005 err = __check_buffer_access(env, buf_info, reg, regno, off, size);
4009 if (off + size > *max_access)
4010 *max_access = off + size;
4015 /* BPF architecture zero extends alu32 ops into 64-bit registesr */
4016 static void zext_32_to_64(struct bpf_reg_state *reg)
4018 reg->var_off = tnum_subreg(reg->var_off);
4019 __reg_assign_32_into_64(reg);
4022 /* truncate register to smaller size (in bytes)
4023 * must be called with size < BPF_REG_SIZE
4025 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
4029 /* clear high bits in bit representation */
4030 reg->var_off = tnum_cast(reg->var_off, size);
4032 /* fix arithmetic bounds */
4033 mask = ((u64)1 << (size * 8)) - 1;
4034 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
4035 reg->umin_value &= mask;
4036 reg->umax_value &= mask;
4038 reg->umin_value = 0;
4039 reg->umax_value = mask;
4041 reg->smin_value = reg->umin_value;
4042 reg->smax_value = reg->umax_value;
4044 /* If size is smaller than 32bit register the 32bit register
4045 * values are also truncated so we push 64-bit bounds into
4046 * 32-bit bounds. Above were truncated < 32-bits already.
4050 __reg_combine_64_into_32(reg);
4053 static bool bpf_map_is_rdonly(const struct bpf_map *map)
4055 /* A map is considered read-only if the following condition are true:
4057 * 1) BPF program side cannot change any of the map content. The
4058 * BPF_F_RDONLY_PROG flag is throughout the lifetime of a map
4059 * and was set at map creation time.
4060 * 2) The map value(s) have been initialized from user space by a
4061 * loader and then "frozen", such that no new map update/delete
4062 * operations from syscall side are possible for the rest of
4063 * the map's lifetime from that point onwards.
4064 * 3) Any parallel/pending map update/delete operations from syscall
4065 * side have been completed. Only after that point, it's safe to
4066 * assume that map value(s) are immutable.
4068 return (map->map_flags & BPF_F_RDONLY_PROG) &&
4069 READ_ONCE(map->frozen) &&
4070 !bpf_map_write_active(map);
4073 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val)
4079 err = map->ops->map_direct_value_addr(map, &addr, off);
4082 ptr = (void *)(long)addr + off;
4086 *val = (u64)*(u8 *)ptr;
4089 *val = (u64)*(u16 *)ptr;
4092 *val = (u64)*(u32 *)ptr;
4103 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
4104 struct bpf_reg_state *regs,
4105 int regno, int off, int size,
4106 enum bpf_access_type atype,
4109 struct bpf_reg_state *reg = regs + regno;
4110 const struct btf_type *t = btf_type_by_id(reg->btf, reg->btf_id);
4111 const char *tname = btf_name_by_offset(reg->btf, t->name_off);
4117 "R%d is ptr_%s invalid negative access: off=%d\n",
4121 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
4124 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4126 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
4127 regno, tname, off, tn_buf);
4131 if (env->ops->btf_struct_access) {
4132 ret = env->ops->btf_struct_access(&env->log, reg->btf, t,
4133 off, size, atype, &btf_id);
4135 if (atype != BPF_READ) {
4136 verbose(env, "only read is supported\n");
4140 ret = btf_struct_access(&env->log, reg->btf, t, off, size,
4147 if (atype == BPF_READ && value_regno >= 0)
4148 mark_btf_ld_reg(env, regs, value_regno, ret, reg->btf, btf_id);
4153 static int check_ptr_to_map_access(struct bpf_verifier_env *env,
4154 struct bpf_reg_state *regs,
4155 int regno, int off, int size,
4156 enum bpf_access_type atype,
4159 struct bpf_reg_state *reg = regs + regno;
4160 struct bpf_map *map = reg->map_ptr;
4161 const struct btf_type *t;
4167 verbose(env, "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n");
4171 if (!map->ops->map_btf_id || !*map->ops->map_btf_id) {
4172 verbose(env, "map_ptr access not supported for map type %d\n",
4177 t = btf_type_by_id(btf_vmlinux, *map->ops->map_btf_id);
4178 tname = btf_name_by_offset(btf_vmlinux, t->name_off);
4180 if (!env->allow_ptr_to_map_access) {
4182 "%s access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
4188 verbose(env, "R%d is %s invalid negative access: off=%d\n",
4193 if (atype != BPF_READ) {
4194 verbose(env, "only read from %s is supported\n", tname);
4198 ret = btf_struct_access(&env->log, btf_vmlinux, t, off, size, atype, &btf_id);
4202 if (value_regno >= 0)
4203 mark_btf_ld_reg(env, regs, value_regno, ret, btf_vmlinux, btf_id);
4208 /* Check that the stack access at the given offset is within bounds. The
4209 * maximum valid offset is -1.
4211 * The minimum valid offset is -MAX_BPF_STACK for writes, and
4212 * -state->allocated_stack for reads.
4214 static int check_stack_slot_within_bounds(int off,
4215 struct bpf_func_state *state,
4216 enum bpf_access_type t)
4221 min_valid_off = -MAX_BPF_STACK;
4223 min_valid_off = -state->allocated_stack;
4225 if (off < min_valid_off || off > -1)
4230 /* Check that the stack access at 'regno + off' falls within the maximum stack
4233 * 'off' includes `regno->offset`, but not its dynamic part (if any).
4235 static int check_stack_access_within_bounds(
4236 struct bpf_verifier_env *env,
4237 int regno, int off, int access_size,
4238 enum stack_access_src src, enum bpf_access_type type)
4240 struct bpf_reg_state *regs = cur_regs(env);
4241 struct bpf_reg_state *reg = regs + regno;
4242 struct bpf_func_state *state = func(env, reg);
4243 int min_off, max_off;
4247 if (src == ACCESS_HELPER)
4248 /* We don't know if helpers are reading or writing (or both). */
4249 err_extra = " indirect access to";
4250 else if (type == BPF_READ)
4251 err_extra = " read from";
4253 err_extra = " write to";
4255 if (tnum_is_const(reg->var_off)) {
4256 min_off = reg->var_off.value + off;
4257 if (access_size > 0)
4258 max_off = min_off + access_size - 1;
4262 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
4263 reg->smin_value <= -BPF_MAX_VAR_OFF) {
4264 verbose(env, "invalid unbounded variable-offset%s stack R%d\n",
4268 min_off = reg->smin_value + off;
4269 if (access_size > 0)
4270 max_off = reg->smax_value + off + access_size - 1;
4275 err = check_stack_slot_within_bounds(min_off, state, type);
4277 err = check_stack_slot_within_bounds(max_off, state, type);
4280 if (tnum_is_const(reg->var_off)) {
4281 verbose(env, "invalid%s stack R%d off=%d size=%d\n",
4282 err_extra, regno, off, access_size);
4286 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4287 verbose(env, "invalid variable-offset%s stack R%d var_off=%s size=%d\n",
4288 err_extra, regno, tn_buf, access_size);
4294 /* check whether memory at (regno + off) is accessible for t = (read | write)
4295 * if t==write, value_regno is a register which value is stored into memory
4296 * if t==read, value_regno is a register which will receive the value from memory
4297 * if t==write && value_regno==-1, some unknown value is stored into memory
4298 * if t==read && value_regno==-1, don't care what we read from memory
4300 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
4301 int off, int bpf_size, enum bpf_access_type t,
4302 int value_regno, bool strict_alignment_once)
4304 struct bpf_reg_state *regs = cur_regs(env);
4305 struct bpf_reg_state *reg = regs + regno;
4306 struct bpf_func_state *state;
4309 size = bpf_size_to_bytes(bpf_size);
4313 /* alignment checks will add in reg->off themselves */
4314 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
4318 /* for access checks, reg->off is just part of off */
4321 if (reg->type == PTR_TO_MAP_KEY) {
4322 if (t == BPF_WRITE) {
4323 verbose(env, "write to change key R%d not allowed\n", regno);
4327 err = check_mem_region_access(env, regno, off, size,
4328 reg->map_ptr->key_size, false);
4331 if (value_regno >= 0)
4332 mark_reg_unknown(env, regs, value_regno);
4333 } else if (reg->type == PTR_TO_MAP_VALUE) {
4334 if (t == BPF_WRITE && value_regno >= 0 &&
4335 is_pointer_value(env, value_regno)) {
4336 verbose(env, "R%d leaks addr into map\n", value_regno);
4339 err = check_map_access_type(env, regno, off, size, t);
4342 err = check_map_access(env, regno, off, size, false);
4343 if (!err && t == BPF_READ && value_regno >= 0) {
4344 struct bpf_map *map = reg->map_ptr;
4346 /* if map is read-only, track its contents as scalars */
4347 if (tnum_is_const(reg->var_off) &&
4348 bpf_map_is_rdonly(map) &&
4349 map->ops->map_direct_value_addr) {
4350 int map_off = off + reg->var_off.value;
4353 err = bpf_map_direct_read(map, map_off, size,
4358 regs[value_regno].type = SCALAR_VALUE;
4359 __mark_reg_known(®s[value_regno], val);
4361 mark_reg_unknown(env, regs, value_regno);
4364 } else if (reg->type == PTR_TO_MEM) {
4365 if (t == BPF_WRITE && value_regno >= 0 &&
4366 is_pointer_value(env, value_regno)) {
4367 verbose(env, "R%d leaks addr into mem\n", value_regno);
4370 err = check_mem_region_access(env, regno, off, size,
4371 reg->mem_size, false);
4372 if (!err && t == BPF_READ && value_regno >= 0)
4373 mark_reg_unknown(env, regs, value_regno);
4374 } else if (reg->type == PTR_TO_CTX) {
4375 enum bpf_reg_type reg_type = SCALAR_VALUE;
4376 struct btf *btf = NULL;
4379 if (t == BPF_WRITE && value_regno >= 0 &&
4380 is_pointer_value(env, value_regno)) {
4381 verbose(env, "R%d leaks addr into ctx\n", value_regno);
4385 err = check_ctx_reg(env, reg, regno);
4389 err = check_ctx_access(env, insn_idx, off, size, t, ®_type, &btf, &btf_id);
4391 verbose_linfo(env, insn_idx, "; ");
4392 if (!err && t == BPF_READ && value_regno >= 0) {
4393 /* ctx access returns either a scalar, or a
4394 * PTR_TO_PACKET[_META,_END]. In the latter
4395 * case, we know the offset is zero.
4397 if (reg_type == SCALAR_VALUE) {
4398 mark_reg_unknown(env, regs, value_regno);
4400 mark_reg_known_zero(env, regs,
4402 if (reg_type_may_be_null(reg_type))
4403 regs[value_regno].id = ++env->id_gen;
4404 /* A load of ctx field could have different
4405 * actual load size with the one encoded in the
4406 * insn. When the dst is PTR, it is for sure not
4409 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
4410 if (reg_type == PTR_TO_BTF_ID ||
4411 reg_type == PTR_TO_BTF_ID_OR_NULL) {
4412 regs[value_regno].btf = btf;
4413 regs[value_regno].btf_id = btf_id;
4416 regs[value_regno].type = reg_type;
4419 } else if (reg->type == PTR_TO_STACK) {
4420 /* Basic bounds checks. */
4421 err = check_stack_access_within_bounds(env, regno, off, size, ACCESS_DIRECT, t);
4425 state = func(env, reg);
4426 err = update_stack_depth(env, state, off);
4431 err = check_stack_read(env, regno, off, size,
4434 err = check_stack_write(env, regno, off, size,
4435 value_regno, insn_idx);
4436 } else if (reg_is_pkt_pointer(reg)) {
4437 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
4438 verbose(env, "cannot write into packet\n");
4441 if (t == BPF_WRITE && value_regno >= 0 &&
4442 is_pointer_value(env, value_regno)) {
4443 verbose(env, "R%d leaks addr into packet\n",
4447 err = check_packet_access(env, regno, off, size, false);
4448 if (!err && t == BPF_READ && value_regno >= 0)
4449 mark_reg_unknown(env, regs, value_regno);
4450 } else if (reg->type == PTR_TO_FLOW_KEYS) {
4451 if (t == BPF_WRITE && value_regno >= 0 &&
4452 is_pointer_value(env, value_regno)) {
4453 verbose(env, "R%d leaks addr into flow keys\n",
4458 err = check_flow_keys_access(env, off, size);
4459 if (!err && t == BPF_READ && value_regno >= 0)
4460 mark_reg_unknown(env, regs, value_regno);
4461 } else if (type_is_sk_pointer(reg->type)) {
4462 if (t == BPF_WRITE) {
4463 verbose(env, "R%d cannot write into %s\n",
4464 regno, reg_type_str[reg->type]);
4467 err = check_sock_access(env, insn_idx, regno, off, size, t);
4468 if (!err && value_regno >= 0)
4469 mark_reg_unknown(env, regs, value_regno);
4470 } else if (reg->type == PTR_TO_TP_BUFFER) {
4471 err = check_tp_buffer_access(env, reg, regno, off, size);
4472 if (!err && t == BPF_READ && value_regno >= 0)
4473 mark_reg_unknown(env, regs, value_regno);
4474 } else if (reg->type == PTR_TO_BTF_ID) {
4475 err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
4477 } else if (reg->type == CONST_PTR_TO_MAP) {
4478 err = check_ptr_to_map_access(env, regs, regno, off, size, t,
4480 } else if (reg->type == PTR_TO_RDONLY_BUF) {
4481 if (t == BPF_WRITE) {
4482 verbose(env, "R%d cannot write into %s\n",
4483 regno, reg_type_str[reg->type]);
4486 err = check_buffer_access(env, reg, regno, off, size, false,
4488 &env->prog->aux->max_rdonly_access);
4489 if (!err && value_regno >= 0)
4490 mark_reg_unknown(env, regs, value_regno);
4491 } else if (reg->type == PTR_TO_RDWR_BUF) {
4492 err = check_buffer_access(env, reg, regno, off, size, false,
4494 &env->prog->aux->max_rdwr_access);
4495 if (!err && t == BPF_READ && value_regno >= 0)
4496 mark_reg_unknown(env, regs, value_regno);
4498 verbose(env, "R%d invalid mem access '%s'\n", regno,
4499 reg_type_str[reg->type]);
4503 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
4504 regs[value_regno].type == SCALAR_VALUE) {
4505 /* b/h/w load zero-extends, mark upper bits as known 0 */
4506 coerce_reg_to_size(®s[value_regno], size);
4511 static int check_atomic(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
4516 switch (insn->imm) {
4518 case BPF_ADD | BPF_FETCH:
4520 case BPF_AND | BPF_FETCH:
4522 case BPF_OR | BPF_FETCH:
4524 case BPF_XOR | BPF_FETCH:
4529 verbose(env, "BPF_ATOMIC uses invalid atomic opcode %02x\n", insn->imm);
4533 if (BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) {
4534 verbose(env, "invalid atomic operand size\n");
4538 /* check src1 operand */
4539 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4543 /* check src2 operand */
4544 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4548 if (insn->imm == BPF_CMPXCHG) {
4549 /* Check comparison of R0 with memory location */
4550 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
4555 if (is_pointer_value(env, insn->src_reg)) {
4556 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
4560 if (is_ctx_reg(env, insn->dst_reg) ||
4561 is_pkt_reg(env, insn->dst_reg) ||
4562 is_flow_key_reg(env, insn->dst_reg) ||
4563 is_sk_reg(env, insn->dst_reg)) {
4564 verbose(env, "BPF_ATOMIC stores into R%d %s is not allowed\n",
4566 reg_type_str[reg_state(env, insn->dst_reg)->type]);
4570 if (insn->imm & BPF_FETCH) {
4571 if (insn->imm == BPF_CMPXCHG)
4572 load_reg = BPF_REG_0;
4574 load_reg = insn->src_reg;
4576 /* check and record load of old value */
4577 err = check_reg_arg(env, load_reg, DST_OP);
4581 /* This instruction accesses a memory location but doesn't
4582 * actually load it into a register.
4587 /* Check whether we can read the memory, with second call for fetch
4588 * case to simulate the register fill.
4590 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4591 BPF_SIZE(insn->code), BPF_READ, -1, true);
4592 if (!err && load_reg >= 0)
4593 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4594 BPF_SIZE(insn->code), BPF_READ, load_reg,
4599 /* Check whether we can write into the same memory. */
4600 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4601 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
4608 /* When register 'regno' is used to read the stack (either directly or through
4609 * a helper function) make sure that it's within stack boundary and, depending
4610 * on the access type, that all elements of the stack are initialized.
4612 * 'off' includes 'regno->off', but not its dynamic part (if any).
4614 * All registers that have been spilled on the stack in the slots within the
4615 * read offsets are marked as read.
4617 static int check_stack_range_initialized(
4618 struct bpf_verifier_env *env, int regno, int off,
4619 int access_size, bool zero_size_allowed,
4620 enum stack_access_src type, struct bpf_call_arg_meta *meta)
4622 struct bpf_reg_state *reg = reg_state(env, regno);
4623 struct bpf_func_state *state = func(env, reg);
4624 int err, min_off, max_off, i, j, slot, spi;
4625 char *err_extra = type == ACCESS_HELPER ? " indirect" : "";
4626 enum bpf_access_type bounds_check_type;
4627 /* Some accesses can write anything into the stack, others are
4630 bool clobber = false;
4632 if (access_size == 0 && !zero_size_allowed) {
4633 verbose(env, "invalid zero-sized read\n");
4637 if (type == ACCESS_HELPER) {
4638 /* The bounds checks for writes are more permissive than for
4639 * reads. However, if raw_mode is not set, we'll do extra
4642 bounds_check_type = BPF_WRITE;
4645 bounds_check_type = BPF_READ;
4647 err = check_stack_access_within_bounds(env, regno, off, access_size,
4648 type, bounds_check_type);
4653 if (tnum_is_const(reg->var_off)) {
4654 min_off = max_off = reg->var_off.value + off;
4656 /* Variable offset is prohibited for unprivileged mode for
4657 * simplicity since it requires corresponding support in
4658 * Spectre masking for stack ALU.
4659 * See also retrieve_ptr_limit().
4661 if (!env->bypass_spec_v1) {
4664 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4665 verbose(env, "R%d%s variable offset stack access prohibited for !root, var_off=%s\n",
4666 regno, err_extra, tn_buf);
4669 /* Only initialized buffer on stack is allowed to be accessed
4670 * with variable offset. With uninitialized buffer it's hard to
4671 * guarantee that whole memory is marked as initialized on
4672 * helper return since specific bounds are unknown what may
4673 * cause uninitialized stack leaking.
4675 if (meta && meta->raw_mode)
4678 min_off = reg->smin_value + off;
4679 max_off = reg->smax_value + off;
4682 if (meta && meta->raw_mode) {
4683 meta->access_size = access_size;
4684 meta->regno = regno;
4688 for (i = min_off; i < max_off + access_size; i++) {
4692 spi = slot / BPF_REG_SIZE;
4693 if (state->allocated_stack <= slot)
4695 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
4696 if (*stype == STACK_MISC)
4698 if (*stype == STACK_ZERO) {
4700 /* helper can write anything into the stack */
4701 *stype = STACK_MISC;
4706 if (is_spilled_reg(&state->stack[spi]) &&
4707 state->stack[spi].spilled_ptr.type == PTR_TO_BTF_ID)
4710 if (is_spilled_reg(&state->stack[spi]) &&
4711 (state->stack[spi].spilled_ptr.type == SCALAR_VALUE ||
4712 env->allow_ptr_leaks)) {
4714 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
4715 for (j = 0; j < BPF_REG_SIZE; j++)
4716 scrub_spilled_slot(&state->stack[spi].slot_type[j]);
4722 if (tnum_is_const(reg->var_off)) {
4723 verbose(env, "invalid%s read from stack R%d off %d+%d size %d\n",
4724 err_extra, regno, min_off, i - min_off, access_size);
4728 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4729 verbose(env, "invalid%s read from stack R%d var_off %s+%d size %d\n",
4730 err_extra, regno, tn_buf, i - min_off, access_size);
4734 /* reading any byte out of 8-byte 'spill_slot' will cause
4735 * the whole slot to be marked as 'read'
4737 mark_reg_read(env, &state->stack[spi].spilled_ptr,
4738 state->stack[spi].spilled_ptr.parent,
4741 return update_stack_depth(env, state, min_off);
4744 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
4745 int access_size, bool zero_size_allowed,
4746 struct bpf_call_arg_meta *meta)
4748 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4750 switch (reg->type) {
4752 case PTR_TO_PACKET_META:
4753 return check_packet_access(env, regno, reg->off, access_size,
4755 case PTR_TO_MAP_KEY:
4756 return check_mem_region_access(env, regno, reg->off, access_size,
4757 reg->map_ptr->key_size, false);
4758 case PTR_TO_MAP_VALUE:
4759 if (check_map_access_type(env, regno, reg->off, access_size,
4760 meta && meta->raw_mode ? BPF_WRITE :
4763 return check_map_access(env, regno, reg->off, access_size,
4766 return check_mem_region_access(env, regno, reg->off,
4767 access_size, reg->mem_size,
4769 case PTR_TO_RDONLY_BUF:
4770 if (meta && meta->raw_mode)
4772 return check_buffer_access(env, reg, regno, reg->off,
4773 access_size, zero_size_allowed,
4775 &env->prog->aux->max_rdonly_access);
4776 case PTR_TO_RDWR_BUF:
4777 return check_buffer_access(env, reg, regno, reg->off,
4778 access_size, zero_size_allowed,
4780 &env->prog->aux->max_rdwr_access);
4782 return check_stack_range_initialized(
4784 regno, reg->off, access_size,
4785 zero_size_allowed, ACCESS_HELPER, meta);
4786 default: /* scalar_value or invalid ptr */
4787 /* Allow zero-byte read from NULL, regardless of pointer type */
4788 if (zero_size_allowed && access_size == 0 &&
4789 register_is_null(reg))
4792 verbose(env, "R%d type=%s expected=%s\n", regno,
4793 reg_type_str[reg->type],
4794 reg_type_str[PTR_TO_STACK]);
4799 int check_mem_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
4800 u32 regno, u32 mem_size)
4802 if (register_is_null(reg))
4805 if (reg_type_may_be_null(reg->type)) {
4806 /* Assuming that the register contains a value check if the memory
4807 * access is safe. Temporarily save and restore the register's state as
4808 * the conversion shouldn't be visible to a caller.
4810 const struct bpf_reg_state saved_reg = *reg;
4813 mark_ptr_not_null_reg(reg);
4814 rv = check_helper_mem_access(env, regno, mem_size, true, NULL);
4819 return check_helper_mem_access(env, regno, mem_size, true, NULL);
4822 /* Implementation details:
4823 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL
4824 * Two bpf_map_lookups (even with the same key) will have different reg->id.
4825 * For traditional PTR_TO_MAP_VALUE the verifier clears reg->id after
4826 * value_or_null->value transition, since the verifier only cares about
4827 * the range of access to valid map value pointer and doesn't care about actual
4828 * address of the map element.
4829 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
4830 * reg->id > 0 after value_or_null->value transition. By doing so
4831 * two bpf_map_lookups will be considered two different pointers that
4832 * point to different bpf_spin_locks.
4833 * The verifier allows taking only one bpf_spin_lock at a time to avoid
4835 * Since only one bpf_spin_lock is allowed the checks are simpler than
4836 * reg_is_refcounted() logic. The verifier needs to remember only
4837 * one spin_lock instead of array of acquired_refs.
4838 * cur_state->active_spin_lock remembers which map value element got locked
4839 * and clears it after bpf_spin_unlock.
4841 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
4844 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4845 struct bpf_verifier_state *cur = env->cur_state;
4846 bool is_const = tnum_is_const(reg->var_off);
4847 struct bpf_map *map = reg->map_ptr;
4848 u64 val = reg->var_off.value;
4852 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
4858 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
4862 if (!map_value_has_spin_lock(map)) {
4863 if (map->spin_lock_off == -E2BIG)
4865 "map '%s' has more than one 'struct bpf_spin_lock'\n",
4867 else if (map->spin_lock_off == -ENOENT)
4869 "map '%s' doesn't have 'struct bpf_spin_lock'\n",
4873 "map '%s' is not a struct type or bpf_spin_lock is mangled\n",
4877 if (map->spin_lock_off != val + reg->off) {
4878 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock'\n",
4883 if (cur->active_spin_lock) {
4885 "Locking two bpf_spin_locks are not allowed\n");
4888 cur->active_spin_lock = reg->id;
4890 if (!cur->active_spin_lock) {
4891 verbose(env, "bpf_spin_unlock without taking a lock\n");
4894 if (cur->active_spin_lock != reg->id) {
4895 verbose(env, "bpf_spin_unlock of different lock\n");
4898 cur->active_spin_lock = 0;
4903 static int process_timer_func(struct bpf_verifier_env *env, int regno,
4904 struct bpf_call_arg_meta *meta)
4906 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4907 bool is_const = tnum_is_const(reg->var_off);
4908 struct bpf_map *map = reg->map_ptr;
4909 u64 val = reg->var_off.value;
4913 "R%d doesn't have constant offset. bpf_timer has to be at the constant offset\n",
4918 verbose(env, "map '%s' has to have BTF in order to use bpf_timer\n",
4922 if (!map_value_has_timer(map)) {
4923 if (map->timer_off == -E2BIG)
4925 "map '%s' has more than one 'struct bpf_timer'\n",
4927 else if (map->timer_off == -ENOENT)
4929 "map '%s' doesn't have 'struct bpf_timer'\n",
4933 "map '%s' is not a struct type or bpf_timer is mangled\n",
4937 if (map->timer_off != val + reg->off) {
4938 verbose(env, "off %lld doesn't point to 'struct bpf_timer' that is at %d\n",
4939 val + reg->off, map->timer_off);
4942 if (meta->map_ptr) {
4943 verbose(env, "verifier bug. Two map pointers in a timer helper\n");
4946 meta->map_uid = reg->map_uid;
4947 meta->map_ptr = map;
4951 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
4953 return type == ARG_PTR_TO_MEM ||
4954 type == ARG_PTR_TO_MEM_OR_NULL ||
4955 type == ARG_PTR_TO_UNINIT_MEM;
4958 static bool arg_type_is_mem_size(enum bpf_arg_type type)
4960 return type == ARG_CONST_SIZE ||
4961 type == ARG_CONST_SIZE_OR_ZERO;
4964 static bool arg_type_is_alloc_size(enum bpf_arg_type type)
4966 return type == ARG_CONST_ALLOC_SIZE_OR_ZERO;
4969 static bool arg_type_is_int_ptr(enum bpf_arg_type type)
4971 return type == ARG_PTR_TO_INT ||
4972 type == ARG_PTR_TO_LONG;
4975 static int int_ptr_type_to_size(enum bpf_arg_type type)
4977 if (type == ARG_PTR_TO_INT)
4979 else if (type == ARG_PTR_TO_LONG)
4985 static int resolve_map_arg_type(struct bpf_verifier_env *env,
4986 const struct bpf_call_arg_meta *meta,
4987 enum bpf_arg_type *arg_type)
4989 if (!meta->map_ptr) {
4990 /* kernel subsystem misconfigured verifier */
4991 verbose(env, "invalid map_ptr to access map->type\n");
4995 switch (meta->map_ptr->map_type) {
4996 case BPF_MAP_TYPE_SOCKMAP:
4997 case BPF_MAP_TYPE_SOCKHASH:
4998 if (*arg_type == ARG_PTR_TO_MAP_VALUE) {
4999 *arg_type = ARG_PTR_TO_BTF_ID_SOCK_COMMON;
5001 verbose(env, "invalid arg_type for sockmap/sockhash\n");
5005 case BPF_MAP_TYPE_BLOOM_FILTER:
5006 if (meta->func_id == BPF_FUNC_map_peek_elem)
5007 *arg_type = ARG_PTR_TO_MAP_VALUE;
5015 struct bpf_reg_types {
5016 const enum bpf_reg_type types[10];
5020 static const struct bpf_reg_types map_key_value_types = {
5030 static const struct bpf_reg_types sock_types = {
5040 static const struct bpf_reg_types btf_id_sock_common_types = {
5048 .btf_id = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
5052 static const struct bpf_reg_types mem_types = {
5065 static const struct bpf_reg_types int_ptr_types = {
5075 static const struct bpf_reg_types fullsock_types = { .types = { PTR_TO_SOCKET } };
5076 static const struct bpf_reg_types scalar_types = { .types = { SCALAR_VALUE } };
5077 static const struct bpf_reg_types context_types = { .types = { PTR_TO_CTX } };
5078 static const struct bpf_reg_types alloc_mem_types = { .types = { PTR_TO_MEM } };
5079 static const struct bpf_reg_types const_map_ptr_types = { .types = { CONST_PTR_TO_MAP } };
5080 static const struct bpf_reg_types btf_ptr_types = { .types = { PTR_TO_BTF_ID } };
5081 static const struct bpf_reg_types spin_lock_types = { .types = { PTR_TO_MAP_VALUE } };
5082 static const struct bpf_reg_types percpu_btf_ptr_types = { .types = { PTR_TO_PERCPU_BTF_ID } };
5083 static const struct bpf_reg_types func_ptr_types = { .types = { PTR_TO_FUNC } };
5084 static const struct bpf_reg_types stack_ptr_types = { .types = { PTR_TO_STACK } };
5085 static const struct bpf_reg_types const_str_ptr_types = { .types = { PTR_TO_MAP_VALUE } };
5086 static const struct bpf_reg_types timer_types = { .types = { PTR_TO_MAP_VALUE } };
5088 static const struct bpf_reg_types *compatible_reg_types[__BPF_ARG_TYPE_MAX] = {
5089 [ARG_PTR_TO_MAP_KEY] = &map_key_value_types,
5090 [ARG_PTR_TO_MAP_VALUE] = &map_key_value_types,
5091 [ARG_PTR_TO_UNINIT_MAP_VALUE] = &map_key_value_types,
5092 [ARG_PTR_TO_MAP_VALUE_OR_NULL] = &map_key_value_types,
5093 [ARG_CONST_SIZE] = &scalar_types,
5094 [ARG_CONST_SIZE_OR_ZERO] = &scalar_types,
5095 [ARG_CONST_ALLOC_SIZE_OR_ZERO] = &scalar_types,
5096 [ARG_CONST_MAP_PTR] = &const_map_ptr_types,
5097 [ARG_PTR_TO_CTX] = &context_types,
5098 [ARG_PTR_TO_CTX_OR_NULL] = &context_types,
5099 [ARG_PTR_TO_SOCK_COMMON] = &sock_types,
5101 [ARG_PTR_TO_BTF_ID_SOCK_COMMON] = &btf_id_sock_common_types,
5103 [ARG_PTR_TO_SOCKET] = &fullsock_types,
5104 [ARG_PTR_TO_SOCKET_OR_NULL] = &fullsock_types,
5105 [ARG_PTR_TO_BTF_ID] = &btf_ptr_types,
5106 [ARG_PTR_TO_SPIN_LOCK] = &spin_lock_types,
5107 [ARG_PTR_TO_MEM] = &mem_types,
5108 [ARG_PTR_TO_MEM_OR_NULL] = &mem_types,
5109 [ARG_PTR_TO_UNINIT_MEM] = &mem_types,
5110 [ARG_PTR_TO_ALLOC_MEM] = &alloc_mem_types,
5111 [ARG_PTR_TO_ALLOC_MEM_OR_NULL] = &alloc_mem_types,
5112 [ARG_PTR_TO_INT] = &int_ptr_types,
5113 [ARG_PTR_TO_LONG] = &int_ptr_types,
5114 [ARG_PTR_TO_PERCPU_BTF_ID] = &percpu_btf_ptr_types,
5115 [ARG_PTR_TO_FUNC] = &func_ptr_types,
5116 [ARG_PTR_TO_STACK_OR_NULL] = &stack_ptr_types,
5117 [ARG_PTR_TO_CONST_STR] = &const_str_ptr_types,
5118 [ARG_PTR_TO_TIMER] = &timer_types,
5121 static int check_reg_type(struct bpf_verifier_env *env, u32 regno,
5122 enum bpf_arg_type arg_type,
5123 const u32 *arg_btf_id)
5125 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
5126 enum bpf_reg_type expected, type = reg->type;
5127 const struct bpf_reg_types *compatible;
5130 compatible = compatible_reg_types[arg_type];
5132 verbose(env, "verifier internal error: unsupported arg type %d\n", arg_type);
5136 for (i = 0; i < ARRAY_SIZE(compatible->types); i++) {
5137 expected = compatible->types[i];
5138 if (expected == NOT_INIT)
5141 if (type == expected)
5145 verbose(env, "R%d type=%s expected=", regno, reg_type_str[type]);
5146 for (j = 0; j + 1 < i; j++)
5147 verbose(env, "%s, ", reg_type_str[compatible->types[j]]);
5148 verbose(env, "%s\n", reg_type_str[compatible->types[j]]);
5152 if (type == PTR_TO_BTF_ID) {
5154 if (!compatible->btf_id) {
5155 verbose(env, "verifier internal error: missing arg compatible BTF ID\n");
5158 arg_btf_id = compatible->btf_id;
5161 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
5162 btf_vmlinux, *arg_btf_id)) {
5163 verbose(env, "R%d is of type %s but %s is expected\n",
5164 regno, kernel_type_name(reg->btf, reg->btf_id),
5165 kernel_type_name(btf_vmlinux, *arg_btf_id));
5169 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
5170 verbose(env, "R%d is a pointer to in-kernel struct with non-zero offset\n",
5179 static int check_func_arg(struct bpf_verifier_env *env, u32 arg,
5180 struct bpf_call_arg_meta *meta,
5181 const struct bpf_func_proto *fn)
5183 u32 regno = BPF_REG_1 + arg;
5184 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
5185 enum bpf_arg_type arg_type = fn->arg_type[arg];
5186 enum bpf_reg_type type = reg->type;
5189 if (arg_type == ARG_DONTCARE)
5192 err = check_reg_arg(env, regno, SRC_OP);
5196 if (arg_type == ARG_ANYTHING) {
5197 if (is_pointer_value(env, regno)) {
5198 verbose(env, "R%d leaks addr into helper function\n",
5205 if (type_is_pkt_pointer(type) &&
5206 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
5207 verbose(env, "helper access to the packet is not allowed\n");
5211 if (arg_type == ARG_PTR_TO_MAP_VALUE ||
5212 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE ||
5213 arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL) {
5214 err = resolve_map_arg_type(env, meta, &arg_type);
5219 if (register_is_null(reg) && arg_type_may_be_null(arg_type))
5220 /* A NULL register has a SCALAR_VALUE type, so skip
5223 goto skip_type_check;
5225 err = check_reg_type(env, regno, arg_type, fn->arg_btf_id[arg]);
5229 if (type == PTR_TO_CTX) {
5230 err = check_ctx_reg(env, reg, regno);
5236 if (reg->ref_obj_id) {
5237 if (meta->ref_obj_id) {
5238 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
5239 regno, reg->ref_obj_id,
5243 meta->ref_obj_id = reg->ref_obj_id;
5246 if (arg_type == ARG_CONST_MAP_PTR) {
5247 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
5248 if (meta->map_ptr) {
5249 /* Use map_uid (which is unique id of inner map) to reject:
5250 * inner_map1 = bpf_map_lookup_elem(outer_map, key1)
5251 * inner_map2 = bpf_map_lookup_elem(outer_map, key2)
5252 * if (inner_map1 && inner_map2) {
5253 * timer = bpf_map_lookup_elem(inner_map1);
5255 * // mismatch would have been allowed
5256 * bpf_timer_init(timer, inner_map2);
5259 * Comparing map_ptr is enough to distinguish normal and outer maps.
5261 if (meta->map_ptr != reg->map_ptr ||
5262 meta->map_uid != reg->map_uid) {
5264 "timer pointer in R1 map_uid=%d doesn't match map pointer in R2 map_uid=%d\n",
5265 meta->map_uid, reg->map_uid);
5269 meta->map_ptr = reg->map_ptr;
5270 meta->map_uid = reg->map_uid;
5271 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
5272 /* bpf_map_xxx(..., map_ptr, ..., key) call:
5273 * check that [key, key + map->key_size) are within
5274 * stack limits and initialized
5276 if (!meta->map_ptr) {
5277 /* in function declaration map_ptr must come before
5278 * map_key, so that it's verified and known before
5279 * we have to check map_key here. Otherwise it means
5280 * that kernel subsystem misconfigured verifier
5282 verbose(env, "invalid map_ptr to access map->key\n");
5285 err = check_helper_mem_access(env, regno,
5286 meta->map_ptr->key_size, false,
5288 } else if (arg_type == ARG_PTR_TO_MAP_VALUE ||
5289 (arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL &&
5290 !register_is_null(reg)) ||
5291 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
5292 /* bpf_map_xxx(..., map_ptr, ..., value) call:
5293 * check [value, value + map->value_size) validity
5295 if (!meta->map_ptr) {
5296 /* kernel subsystem misconfigured verifier */
5297 verbose(env, "invalid map_ptr to access map->value\n");
5300 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE);
5301 err = check_helper_mem_access(env, regno,
5302 meta->map_ptr->value_size, false,
5304 } else if (arg_type == ARG_PTR_TO_PERCPU_BTF_ID) {
5306 verbose(env, "Helper has invalid btf_id in R%d\n", regno);
5309 meta->ret_btf = reg->btf;
5310 meta->ret_btf_id = reg->btf_id;
5311 } else if (arg_type == ARG_PTR_TO_SPIN_LOCK) {
5312 if (meta->func_id == BPF_FUNC_spin_lock) {
5313 if (process_spin_lock(env, regno, true))
5315 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
5316 if (process_spin_lock(env, regno, false))
5319 verbose(env, "verifier internal error\n");
5322 } else if (arg_type == ARG_PTR_TO_TIMER) {
5323 if (process_timer_func(env, regno, meta))
5325 } else if (arg_type == ARG_PTR_TO_FUNC) {
5326 meta->subprogno = reg->subprogno;
5327 } else if (arg_type_is_mem_ptr(arg_type)) {
5328 /* The access to this pointer is only checked when we hit the
5329 * next is_mem_size argument below.
5331 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MEM);
5332 } else if (arg_type_is_mem_size(arg_type)) {
5333 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
5335 /* This is used to refine r0 return value bounds for helpers
5336 * that enforce this value as an upper bound on return values.
5337 * See do_refine_retval_range() for helpers that can refine
5338 * the return value. C type of helper is u32 so we pull register
5339 * bound from umax_value however, if negative verifier errors
5340 * out. Only upper bounds can be learned because retval is an
5341 * int type and negative retvals are allowed.
5343 meta->msize_max_value = reg->umax_value;
5345 /* The register is SCALAR_VALUE; the access check
5346 * happens using its boundaries.
5348 if (!tnum_is_const(reg->var_off))
5349 /* For unprivileged variable accesses, disable raw
5350 * mode so that the program is required to
5351 * initialize all the memory that the helper could
5352 * just partially fill up.
5356 if (reg->smin_value < 0) {
5357 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
5362 if (reg->umin_value == 0) {
5363 err = check_helper_mem_access(env, regno - 1, 0,
5370 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
5371 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
5375 err = check_helper_mem_access(env, regno - 1,
5377 zero_size_allowed, meta);
5379 err = mark_chain_precision(env, regno);
5380 } else if (arg_type_is_alloc_size(arg_type)) {
5381 if (!tnum_is_const(reg->var_off)) {
5382 verbose(env, "R%d is not a known constant'\n",
5386 meta->mem_size = reg->var_off.value;
5387 } else if (arg_type_is_int_ptr(arg_type)) {
5388 int size = int_ptr_type_to_size(arg_type);
5390 err = check_helper_mem_access(env, regno, size, false, meta);
5393 err = check_ptr_alignment(env, reg, 0, size, true);
5394 } else if (arg_type == ARG_PTR_TO_CONST_STR) {
5395 struct bpf_map *map = reg->map_ptr;
5400 if (!bpf_map_is_rdonly(map)) {
5401 verbose(env, "R%d does not point to a readonly map'\n", regno);
5405 if (!tnum_is_const(reg->var_off)) {
5406 verbose(env, "R%d is not a constant address'\n", regno);
5410 if (!map->ops->map_direct_value_addr) {
5411 verbose(env, "no direct value access support for this map type\n");
5415 err = check_map_access(env, regno, reg->off,
5416 map->value_size - reg->off, false);
5420 map_off = reg->off + reg->var_off.value;
5421 err = map->ops->map_direct_value_addr(map, &map_addr, map_off);
5423 verbose(env, "direct value access on string failed\n");
5427 str_ptr = (char *)(long)(map_addr);
5428 if (!strnchr(str_ptr + map_off, map->value_size - map_off, 0)) {
5429 verbose(env, "string is not zero-terminated\n");
5437 static bool may_update_sockmap(struct bpf_verifier_env *env, int func_id)
5439 enum bpf_attach_type eatype = env->prog->expected_attach_type;
5440 enum bpf_prog_type type = resolve_prog_type(env->prog);
5442 if (func_id != BPF_FUNC_map_update_elem)
5445 /* It's not possible to get access to a locked struct sock in these
5446 * contexts, so updating is safe.
5449 case BPF_PROG_TYPE_TRACING:
5450 if (eatype == BPF_TRACE_ITER)
5453 case BPF_PROG_TYPE_SOCKET_FILTER:
5454 case BPF_PROG_TYPE_SCHED_CLS:
5455 case BPF_PROG_TYPE_SCHED_ACT:
5456 case BPF_PROG_TYPE_XDP:
5457 case BPF_PROG_TYPE_SK_REUSEPORT:
5458 case BPF_PROG_TYPE_FLOW_DISSECTOR:
5459 case BPF_PROG_TYPE_SK_LOOKUP:
5465 verbose(env, "cannot update sockmap in this context\n");
5469 static bool allow_tail_call_in_subprogs(struct bpf_verifier_env *env)
5471 return env->prog->jit_requested && IS_ENABLED(CONFIG_X86_64);
5474 static int check_map_func_compatibility(struct bpf_verifier_env *env,
5475 struct bpf_map *map, int func_id)
5480 /* We need a two way check, first is from map perspective ... */
5481 switch (map->map_type) {
5482 case BPF_MAP_TYPE_PROG_ARRAY:
5483 if (func_id != BPF_FUNC_tail_call)
5486 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
5487 if (func_id != BPF_FUNC_perf_event_read &&
5488 func_id != BPF_FUNC_perf_event_output &&
5489 func_id != BPF_FUNC_skb_output &&
5490 func_id != BPF_FUNC_perf_event_read_value &&
5491 func_id != BPF_FUNC_xdp_output)
5494 case BPF_MAP_TYPE_RINGBUF:
5495 if (func_id != BPF_FUNC_ringbuf_output &&
5496 func_id != BPF_FUNC_ringbuf_reserve &&
5497 func_id != BPF_FUNC_ringbuf_query)
5500 case BPF_MAP_TYPE_STACK_TRACE:
5501 if (func_id != BPF_FUNC_get_stackid)
5504 case BPF_MAP_TYPE_CGROUP_ARRAY:
5505 if (func_id != BPF_FUNC_skb_under_cgroup &&
5506 func_id != BPF_FUNC_current_task_under_cgroup)
5509 case BPF_MAP_TYPE_CGROUP_STORAGE:
5510 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
5511 if (func_id != BPF_FUNC_get_local_storage)
5514 case BPF_MAP_TYPE_DEVMAP:
5515 case BPF_MAP_TYPE_DEVMAP_HASH:
5516 if (func_id != BPF_FUNC_redirect_map &&
5517 func_id != BPF_FUNC_map_lookup_elem)
5520 /* Restrict bpf side of cpumap and xskmap, open when use-cases
5523 case BPF_MAP_TYPE_CPUMAP:
5524 if (func_id != BPF_FUNC_redirect_map)
5527 case BPF_MAP_TYPE_XSKMAP:
5528 if (func_id != BPF_FUNC_redirect_map &&
5529 func_id != BPF_FUNC_map_lookup_elem)
5532 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
5533 case BPF_MAP_TYPE_HASH_OF_MAPS:
5534 if (func_id != BPF_FUNC_map_lookup_elem)
5537 case BPF_MAP_TYPE_SOCKMAP:
5538 if (func_id != BPF_FUNC_sk_redirect_map &&
5539 func_id != BPF_FUNC_sock_map_update &&
5540 func_id != BPF_FUNC_map_delete_elem &&
5541 func_id != BPF_FUNC_msg_redirect_map &&
5542 func_id != BPF_FUNC_sk_select_reuseport &&
5543 func_id != BPF_FUNC_map_lookup_elem &&
5544 !may_update_sockmap(env, func_id))
5547 case BPF_MAP_TYPE_SOCKHASH:
5548 if (func_id != BPF_FUNC_sk_redirect_hash &&
5549 func_id != BPF_FUNC_sock_hash_update &&
5550 func_id != BPF_FUNC_map_delete_elem &&
5551 func_id != BPF_FUNC_msg_redirect_hash &&
5552 func_id != BPF_FUNC_sk_select_reuseport &&
5553 func_id != BPF_FUNC_map_lookup_elem &&
5554 !may_update_sockmap(env, func_id))
5557 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
5558 if (func_id != BPF_FUNC_sk_select_reuseport)
5561 case BPF_MAP_TYPE_QUEUE:
5562 case BPF_MAP_TYPE_STACK:
5563 if (func_id != BPF_FUNC_map_peek_elem &&
5564 func_id != BPF_FUNC_map_pop_elem &&
5565 func_id != BPF_FUNC_map_push_elem)
5568 case BPF_MAP_TYPE_SK_STORAGE:
5569 if (func_id != BPF_FUNC_sk_storage_get &&
5570 func_id != BPF_FUNC_sk_storage_delete)
5573 case BPF_MAP_TYPE_INODE_STORAGE:
5574 if (func_id != BPF_FUNC_inode_storage_get &&
5575 func_id != BPF_FUNC_inode_storage_delete)
5578 case BPF_MAP_TYPE_TASK_STORAGE:
5579 if (func_id != BPF_FUNC_task_storage_get &&
5580 func_id != BPF_FUNC_task_storage_delete)
5583 case BPF_MAP_TYPE_BLOOM_FILTER:
5584 if (func_id != BPF_FUNC_map_peek_elem &&
5585 func_id != BPF_FUNC_map_push_elem)
5592 /* ... and second from the function itself. */
5594 case BPF_FUNC_tail_call:
5595 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
5597 if (env->subprog_cnt > 1 && !allow_tail_call_in_subprogs(env)) {
5598 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
5602 case BPF_FUNC_perf_event_read:
5603 case BPF_FUNC_perf_event_output:
5604 case BPF_FUNC_perf_event_read_value:
5605 case BPF_FUNC_skb_output:
5606 case BPF_FUNC_xdp_output:
5607 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
5610 case BPF_FUNC_ringbuf_output:
5611 case BPF_FUNC_ringbuf_reserve:
5612 case BPF_FUNC_ringbuf_query:
5613 if (map->map_type != BPF_MAP_TYPE_RINGBUF)
5616 case BPF_FUNC_get_stackid:
5617 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
5620 case BPF_FUNC_current_task_under_cgroup:
5621 case BPF_FUNC_skb_under_cgroup:
5622 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
5625 case BPF_FUNC_redirect_map:
5626 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
5627 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
5628 map->map_type != BPF_MAP_TYPE_CPUMAP &&
5629 map->map_type != BPF_MAP_TYPE_XSKMAP)
5632 case BPF_FUNC_sk_redirect_map:
5633 case BPF_FUNC_msg_redirect_map:
5634 case BPF_FUNC_sock_map_update:
5635 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
5638 case BPF_FUNC_sk_redirect_hash:
5639 case BPF_FUNC_msg_redirect_hash:
5640 case BPF_FUNC_sock_hash_update:
5641 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
5644 case BPF_FUNC_get_local_storage:
5645 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
5646 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
5649 case BPF_FUNC_sk_select_reuseport:
5650 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
5651 map->map_type != BPF_MAP_TYPE_SOCKMAP &&
5652 map->map_type != BPF_MAP_TYPE_SOCKHASH)
5655 case BPF_FUNC_map_pop_elem:
5656 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
5657 map->map_type != BPF_MAP_TYPE_STACK)
5660 case BPF_FUNC_map_peek_elem:
5661 case BPF_FUNC_map_push_elem:
5662 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
5663 map->map_type != BPF_MAP_TYPE_STACK &&
5664 map->map_type != BPF_MAP_TYPE_BLOOM_FILTER)
5667 case BPF_FUNC_sk_storage_get:
5668 case BPF_FUNC_sk_storage_delete:
5669 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
5672 case BPF_FUNC_inode_storage_get:
5673 case BPF_FUNC_inode_storage_delete:
5674 if (map->map_type != BPF_MAP_TYPE_INODE_STORAGE)
5677 case BPF_FUNC_task_storage_get:
5678 case BPF_FUNC_task_storage_delete:
5679 if (map->map_type != BPF_MAP_TYPE_TASK_STORAGE)
5688 verbose(env, "cannot pass map_type %d into func %s#%d\n",
5689 map->map_type, func_id_name(func_id), func_id);
5693 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
5697 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
5699 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
5701 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
5703 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
5705 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
5708 /* We only support one arg being in raw mode at the moment,
5709 * which is sufficient for the helper functions we have
5715 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
5716 enum bpf_arg_type arg_next)
5718 return (arg_type_is_mem_ptr(arg_curr) &&
5719 !arg_type_is_mem_size(arg_next)) ||
5720 (!arg_type_is_mem_ptr(arg_curr) &&
5721 arg_type_is_mem_size(arg_next));
5724 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
5726 /* bpf_xxx(..., buf, len) call will access 'len'
5727 * bytes from memory 'buf'. Both arg types need
5728 * to be paired, so make sure there's no buggy
5729 * helper function specification.
5731 if (arg_type_is_mem_size(fn->arg1_type) ||
5732 arg_type_is_mem_ptr(fn->arg5_type) ||
5733 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
5734 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
5735 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
5736 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
5742 static bool check_refcount_ok(const struct bpf_func_proto *fn, int func_id)
5746 if (arg_type_may_be_refcounted(fn->arg1_type))
5748 if (arg_type_may_be_refcounted(fn->arg2_type))
5750 if (arg_type_may_be_refcounted(fn->arg3_type))
5752 if (arg_type_may_be_refcounted(fn->arg4_type))
5754 if (arg_type_may_be_refcounted(fn->arg5_type))
5757 /* A reference acquiring function cannot acquire
5758 * another refcounted ptr.
5760 if (may_be_acquire_function(func_id) && count)
5763 /* We only support one arg being unreferenced at the moment,
5764 * which is sufficient for the helper functions we have right now.
5769 static bool check_btf_id_ok(const struct bpf_func_proto *fn)
5773 for (i = 0; i < ARRAY_SIZE(fn->arg_type); i++) {
5774 if (fn->arg_type[i] == ARG_PTR_TO_BTF_ID && !fn->arg_btf_id[i])
5777 if (fn->arg_type[i] != ARG_PTR_TO_BTF_ID && fn->arg_btf_id[i])
5784 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
5786 return check_raw_mode_ok(fn) &&
5787 check_arg_pair_ok(fn) &&
5788 check_btf_id_ok(fn) &&
5789 check_refcount_ok(fn, func_id) ? 0 : -EINVAL;
5792 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
5793 * are now invalid, so turn them into unknown SCALAR_VALUE.
5795 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
5796 struct bpf_func_state *state)
5798 struct bpf_reg_state *regs = state->regs, *reg;
5801 for (i = 0; i < MAX_BPF_REG; i++)
5802 if (reg_is_pkt_pointer_any(®s[i]))
5803 mark_reg_unknown(env, regs, i);
5805 bpf_for_each_spilled_reg(i, state, reg) {
5808 if (reg_is_pkt_pointer_any(reg))
5809 __mark_reg_unknown(env, reg);
5813 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
5815 struct bpf_verifier_state *vstate = env->cur_state;
5818 for (i = 0; i <= vstate->curframe; i++)
5819 __clear_all_pkt_pointers(env, vstate->frame[i]);
5824 BEYOND_PKT_END = -2,
5827 static void mark_pkt_end(struct bpf_verifier_state *vstate, int regn, bool range_open)
5829 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5830 struct bpf_reg_state *reg = &state->regs[regn];
5832 if (reg->type != PTR_TO_PACKET)
5833 /* PTR_TO_PACKET_META is not supported yet */
5836 /* The 'reg' is pkt > pkt_end or pkt >= pkt_end.
5837 * How far beyond pkt_end it goes is unknown.
5838 * if (!range_open) it's the case of pkt >= pkt_end
5839 * if (range_open) it's the case of pkt > pkt_end
5840 * hence this pointer is at least 1 byte bigger than pkt_end
5843 reg->range = BEYOND_PKT_END;
5845 reg->range = AT_PKT_END;
5848 static void release_reg_references(struct bpf_verifier_env *env,
5849 struct bpf_func_state *state,
5852 struct bpf_reg_state *regs = state->regs, *reg;
5855 for (i = 0; i < MAX_BPF_REG; i++)
5856 if (regs[i].ref_obj_id == ref_obj_id)
5857 mark_reg_unknown(env, regs, i);
5859 bpf_for_each_spilled_reg(i, state, reg) {
5862 if (reg->ref_obj_id == ref_obj_id)
5863 __mark_reg_unknown(env, reg);
5867 /* The pointer with the specified id has released its reference to kernel
5868 * resources. Identify all copies of the same pointer and clear the reference.
5870 static int release_reference(struct bpf_verifier_env *env,
5873 struct bpf_verifier_state *vstate = env->cur_state;
5877 err = release_reference_state(cur_func(env), ref_obj_id);
5881 for (i = 0; i <= vstate->curframe; i++)
5882 release_reg_references(env, vstate->frame[i], ref_obj_id);
5887 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
5888 struct bpf_reg_state *regs)
5892 /* after the call registers r0 - r5 were scratched */
5893 for (i = 0; i < CALLER_SAVED_REGS; i++) {
5894 mark_reg_not_init(env, regs, caller_saved[i]);
5895 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
5899 typedef int (*set_callee_state_fn)(struct bpf_verifier_env *env,
5900 struct bpf_func_state *caller,
5901 struct bpf_func_state *callee,
5904 static int __check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
5905 int *insn_idx, int subprog,
5906 set_callee_state_fn set_callee_state_cb)
5908 struct bpf_verifier_state *state = env->cur_state;
5909 struct bpf_func_info_aux *func_info_aux;
5910 struct bpf_func_state *caller, *callee;
5912 bool is_global = false;
5914 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
5915 verbose(env, "the call stack of %d frames is too deep\n",
5916 state->curframe + 2);
5920 caller = state->frame[state->curframe];
5921 if (state->frame[state->curframe + 1]) {
5922 verbose(env, "verifier bug. Frame %d already allocated\n",
5923 state->curframe + 1);
5927 func_info_aux = env->prog->aux->func_info_aux;
5929 is_global = func_info_aux[subprog].linkage == BTF_FUNC_GLOBAL;
5930 err = btf_check_subprog_arg_match(env, subprog, caller->regs);
5935 verbose(env, "Caller passes invalid args into func#%d\n",
5939 if (env->log.level & BPF_LOG_LEVEL)
5941 "Func#%d is global and valid. Skipping.\n",
5943 clear_caller_saved_regs(env, caller->regs);
5945 /* All global functions return a 64-bit SCALAR_VALUE */
5946 mark_reg_unknown(env, caller->regs, BPF_REG_0);
5947 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
5949 /* continue with next insn after call */
5954 if (insn->code == (BPF_JMP | BPF_CALL) &&
5955 insn->imm == BPF_FUNC_timer_set_callback) {
5956 struct bpf_verifier_state *async_cb;
5958 /* there is no real recursion here. timer callbacks are async */
5959 env->subprog_info[subprog].is_async_cb = true;
5960 async_cb = push_async_cb(env, env->subprog_info[subprog].start,
5961 *insn_idx, subprog);
5964 callee = async_cb->frame[0];
5965 callee->async_entry_cnt = caller->async_entry_cnt + 1;
5967 /* Convert bpf_timer_set_callback() args into timer callback args */
5968 err = set_callee_state_cb(env, caller, callee, *insn_idx);
5972 clear_caller_saved_regs(env, caller->regs);
5973 mark_reg_unknown(env, caller->regs, BPF_REG_0);
5974 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
5975 /* continue with next insn after call */
5979 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
5982 state->frame[state->curframe + 1] = callee;
5984 /* callee cannot access r0, r6 - r9 for reading and has to write
5985 * into its own stack before reading from it.
5986 * callee can read/write into caller's stack
5988 init_func_state(env, callee,
5989 /* remember the callsite, it will be used by bpf_exit */
5990 *insn_idx /* callsite */,
5991 state->curframe + 1 /* frameno within this callchain */,
5992 subprog /* subprog number within this prog */);
5994 /* Transfer references to the callee */
5995 err = copy_reference_state(callee, caller);
5999 err = set_callee_state_cb(env, caller, callee, *insn_idx);
6003 clear_caller_saved_regs(env, caller->regs);
6005 /* only increment it after check_reg_arg() finished */
6008 /* and go analyze first insn of the callee */
6009 *insn_idx = env->subprog_info[subprog].start - 1;
6011 if (env->log.level & BPF_LOG_LEVEL) {
6012 verbose(env, "caller:\n");
6013 print_verifier_state(env, caller);
6014 verbose(env, "callee:\n");
6015 print_verifier_state(env, callee);
6020 int map_set_for_each_callback_args(struct bpf_verifier_env *env,
6021 struct bpf_func_state *caller,
6022 struct bpf_func_state *callee)
6024 /* bpf_for_each_map_elem(struct bpf_map *map, void *callback_fn,
6025 * void *callback_ctx, u64 flags);
6026 * callback_fn(struct bpf_map *map, void *key, void *value,
6027 * void *callback_ctx);
6029 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
6031 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
6032 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
6033 callee->regs[BPF_REG_2].map_ptr = caller->regs[BPF_REG_1].map_ptr;
6035 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
6036 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
6037 callee->regs[BPF_REG_3].map_ptr = caller->regs[BPF_REG_1].map_ptr;
6039 /* pointer to stack or null */
6040 callee->regs[BPF_REG_4] = caller->regs[BPF_REG_3];
6043 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
6047 static int set_callee_state(struct bpf_verifier_env *env,
6048 struct bpf_func_state *caller,
6049 struct bpf_func_state *callee, int insn_idx)
6053 /* copy r1 - r5 args that callee can access. The copy includes parent
6054 * pointers, which connects us up to the liveness chain
6056 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
6057 callee->regs[i] = caller->regs[i];
6061 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
6064 int subprog, target_insn;
6066 target_insn = *insn_idx + insn->imm + 1;
6067 subprog = find_subprog(env, target_insn);
6069 verbose(env, "verifier bug. No program starts at insn %d\n",
6074 return __check_func_call(env, insn, insn_idx, subprog, set_callee_state);
6077 static int set_map_elem_callback_state(struct bpf_verifier_env *env,
6078 struct bpf_func_state *caller,
6079 struct bpf_func_state *callee,
6082 struct bpf_insn_aux_data *insn_aux = &env->insn_aux_data[insn_idx];
6083 struct bpf_map *map;
6086 if (bpf_map_ptr_poisoned(insn_aux)) {
6087 verbose(env, "tail_call abusing map_ptr\n");
6091 map = BPF_MAP_PTR(insn_aux->map_ptr_state);
6092 if (!map->ops->map_set_for_each_callback_args ||
6093 !map->ops->map_for_each_callback) {
6094 verbose(env, "callback function not allowed for map\n");
6098 err = map->ops->map_set_for_each_callback_args(env, caller, callee);
6102 callee->in_callback_fn = true;
6106 static int set_timer_callback_state(struct bpf_verifier_env *env,
6107 struct bpf_func_state *caller,
6108 struct bpf_func_state *callee,
6111 struct bpf_map *map_ptr = caller->regs[BPF_REG_1].map_ptr;
6113 /* bpf_timer_set_callback(struct bpf_timer *timer, void *callback_fn);
6114 * callback_fn(struct bpf_map *map, void *key, void *value);
6116 callee->regs[BPF_REG_1].type = CONST_PTR_TO_MAP;
6117 __mark_reg_known_zero(&callee->regs[BPF_REG_1]);
6118 callee->regs[BPF_REG_1].map_ptr = map_ptr;
6120 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
6121 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
6122 callee->regs[BPF_REG_2].map_ptr = map_ptr;
6124 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
6125 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
6126 callee->regs[BPF_REG_3].map_ptr = map_ptr;
6129 __mark_reg_not_init(env, &callee->regs[BPF_REG_4]);
6130 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
6131 callee->in_async_callback_fn = true;
6135 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
6137 struct bpf_verifier_state *state = env->cur_state;
6138 struct bpf_func_state *caller, *callee;
6139 struct bpf_reg_state *r0;
6142 callee = state->frame[state->curframe];
6143 r0 = &callee->regs[BPF_REG_0];
6144 if (r0->type == PTR_TO_STACK) {
6145 /* technically it's ok to return caller's stack pointer
6146 * (or caller's caller's pointer) back to the caller,
6147 * since these pointers are valid. Only current stack
6148 * pointer will be invalid as soon as function exits,
6149 * but let's be conservative
6151 verbose(env, "cannot return stack pointer to the caller\n");
6156 caller = state->frame[state->curframe];
6157 if (callee->in_callback_fn) {
6158 /* enforce R0 return value range [0, 1]. */
6159 struct tnum range = tnum_range(0, 1);
6161 if (r0->type != SCALAR_VALUE) {
6162 verbose(env, "R0 not a scalar value\n");
6165 if (!tnum_in(range, r0->var_off)) {
6166 verbose_invalid_scalar(env, r0, &range, "callback return", "R0");
6170 /* return to the caller whatever r0 had in the callee */
6171 caller->regs[BPF_REG_0] = *r0;
6174 /* Transfer references to the caller */
6175 err = copy_reference_state(caller, callee);
6179 *insn_idx = callee->callsite + 1;
6180 if (env->log.level & BPF_LOG_LEVEL) {
6181 verbose(env, "returning from callee:\n");
6182 print_verifier_state(env, callee);
6183 verbose(env, "to caller at %d:\n", *insn_idx);
6184 print_verifier_state(env, caller);
6186 /* clear everything in the callee */
6187 free_func_state(callee);
6188 state->frame[state->curframe + 1] = NULL;
6192 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
6194 struct bpf_call_arg_meta *meta)
6196 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
6198 if (ret_type != RET_INTEGER ||
6199 (func_id != BPF_FUNC_get_stack &&
6200 func_id != BPF_FUNC_get_task_stack &&
6201 func_id != BPF_FUNC_probe_read_str &&
6202 func_id != BPF_FUNC_probe_read_kernel_str &&
6203 func_id != BPF_FUNC_probe_read_user_str))
6206 ret_reg->smax_value = meta->msize_max_value;
6207 ret_reg->s32_max_value = meta->msize_max_value;
6208 ret_reg->smin_value = -MAX_ERRNO;
6209 ret_reg->s32_min_value = -MAX_ERRNO;
6210 __reg_deduce_bounds(ret_reg);
6211 __reg_bound_offset(ret_reg);
6212 __update_reg_bounds(ret_reg);
6216 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
6217 int func_id, int insn_idx)
6219 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
6220 struct bpf_map *map = meta->map_ptr;
6222 if (func_id != BPF_FUNC_tail_call &&
6223 func_id != BPF_FUNC_map_lookup_elem &&
6224 func_id != BPF_FUNC_map_update_elem &&
6225 func_id != BPF_FUNC_map_delete_elem &&
6226 func_id != BPF_FUNC_map_push_elem &&
6227 func_id != BPF_FUNC_map_pop_elem &&
6228 func_id != BPF_FUNC_map_peek_elem &&
6229 func_id != BPF_FUNC_for_each_map_elem &&
6230 func_id != BPF_FUNC_redirect_map)
6234 verbose(env, "kernel subsystem misconfigured verifier\n");
6238 /* In case of read-only, some additional restrictions
6239 * need to be applied in order to prevent altering the
6240 * state of the map from program side.
6242 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
6243 (func_id == BPF_FUNC_map_delete_elem ||
6244 func_id == BPF_FUNC_map_update_elem ||
6245 func_id == BPF_FUNC_map_push_elem ||
6246 func_id == BPF_FUNC_map_pop_elem)) {
6247 verbose(env, "write into map forbidden\n");
6251 if (!BPF_MAP_PTR(aux->map_ptr_state))
6252 bpf_map_ptr_store(aux, meta->map_ptr,
6253 !meta->map_ptr->bypass_spec_v1);
6254 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
6255 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
6256 !meta->map_ptr->bypass_spec_v1);
6261 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
6262 int func_id, int insn_idx)
6264 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
6265 struct bpf_reg_state *regs = cur_regs(env), *reg;
6266 struct bpf_map *map = meta->map_ptr;
6271 if (func_id != BPF_FUNC_tail_call)
6273 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
6274 verbose(env, "kernel subsystem misconfigured verifier\n");
6278 range = tnum_range(0, map->max_entries - 1);
6279 reg = ®s[BPF_REG_3];
6281 if (!register_is_const(reg) || !tnum_in(range, reg->var_off)) {
6282 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
6286 err = mark_chain_precision(env, BPF_REG_3);
6290 val = reg->var_off.value;
6291 if (bpf_map_key_unseen(aux))
6292 bpf_map_key_store(aux, val);
6293 else if (!bpf_map_key_poisoned(aux) &&
6294 bpf_map_key_immediate(aux) != val)
6295 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
6299 static int check_reference_leak(struct bpf_verifier_env *env)
6301 struct bpf_func_state *state = cur_func(env);
6304 for (i = 0; i < state->acquired_refs; i++) {
6305 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
6306 state->refs[i].id, state->refs[i].insn_idx);
6308 return state->acquired_refs ? -EINVAL : 0;
6311 static int check_bpf_snprintf_call(struct bpf_verifier_env *env,
6312 struct bpf_reg_state *regs)
6314 struct bpf_reg_state *fmt_reg = ®s[BPF_REG_3];
6315 struct bpf_reg_state *data_len_reg = ®s[BPF_REG_5];
6316 struct bpf_map *fmt_map = fmt_reg->map_ptr;
6317 int err, fmt_map_off, num_args;
6321 /* data must be an array of u64 */
6322 if (data_len_reg->var_off.value % 8)
6324 num_args = data_len_reg->var_off.value / 8;
6326 /* fmt being ARG_PTR_TO_CONST_STR guarantees that var_off is const
6327 * and map_direct_value_addr is set.
6329 fmt_map_off = fmt_reg->off + fmt_reg->var_off.value;
6330 err = fmt_map->ops->map_direct_value_addr(fmt_map, &fmt_addr,
6333 verbose(env, "verifier bug\n");
6336 fmt = (char *)(long)fmt_addr + fmt_map_off;
6338 /* We are also guaranteed that fmt+fmt_map_off is NULL terminated, we
6339 * can focus on validating the format specifiers.
6341 err = bpf_bprintf_prepare(fmt, UINT_MAX, NULL, NULL, num_args);
6343 verbose(env, "Invalid format string\n");
6348 static int check_get_func_ip(struct bpf_verifier_env *env)
6350 enum bpf_attach_type eatype = env->prog->expected_attach_type;
6351 enum bpf_prog_type type = resolve_prog_type(env->prog);
6352 int func_id = BPF_FUNC_get_func_ip;
6354 if (type == BPF_PROG_TYPE_TRACING) {
6355 if (eatype != BPF_TRACE_FENTRY && eatype != BPF_TRACE_FEXIT &&
6356 eatype != BPF_MODIFY_RETURN) {
6357 verbose(env, "func %s#%d supported only for fentry/fexit/fmod_ret programs\n",
6358 func_id_name(func_id), func_id);
6362 } else if (type == BPF_PROG_TYPE_KPROBE) {
6366 verbose(env, "func %s#%d not supported for program type %d\n",
6367 func_id_name(func_id), func_id, type);
6371 static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
6374 const struct bpf_func_proto *fn = NULL;
6375 struct bpf_reg_state *regs;
6376 struct bpf_call_arg_meta meta;
6377 int insn_idx = *insn_idx_p;
6379 int i, err, func_id;
6381 /* find function prototype */
6382 func_id = insn->imm;
6383 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
6384 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
6389 if (env->ops->get_func_proto)
6390 fn = env->ops->get_func_proto(func_id, env->prog);
6392 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
6397 /* eBPF programs must be GPL compatible to use GPL-ed functions */
6398 if (!env->prog->gpl_compatible && fn->gpl_only) {
6399 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
6403 if (fn->allowed && !fn->allowed(env->prog)) {
6404 verbose(env, "helper call is not allowed in probe\n");
6408 /* With LD_ABS/IND some JITs save/restore skb from r1. */
6409 changes_data = bpf_helper_changes_pkt_data(fn->func);
6410 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
6411 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
6412 func_id_name(func_id), func_id);
6416 memset(&meta, 0, sizeof(meta));
6417 meta.pkt_access = fn->pkt_access;
6419 err = check_func_proto(fn, func_id);
6421 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
6422 func_id_name(func_id), func_id);
6426 meta.func_id = func_id;
6428 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++) {
6429 err = check_func_arg(env, i, &meta, fn);
6434 err = record_func_map(env, &meta, func_id, insn_idx);
6438 err = record_func_key(env, &meta, func_id, insn_idx);
6442 /* Mark slots with STACK_MISC in case of raw mode, stack offset
6443 * is inferred from register state.
6445 for (i = 0; i < meta.access_size; i++) {
6446 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
6447 BPF_WRITE, -1, false);
6452 if (func_id == BPF_FUNC_tail_call) {
6453 err = check_reference_leak(env);
6455 verbose(env, "tail_call would lead to reference leak\n");
6458 } else if (is_release_function(func_id)) {
6459 err = release_reference(env, meta.ref_obj_id);
6461 verbose(env, "func %s#%d reference has not been acquired before\n",
6462 func_id_name(func_id), func_id);
6467 regs = cur_regs(env);
6469 /* check that flags argument in get_local_storage(map, flags) is 0,
6470 * this is required because get_local_storage() can't return an error.
6472 if (func_id == BPF_FUNC_get_local_storage &&
6473 !register_is_null(®s[BPF_REG_2])) {
6474 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
6478 if (func_id == BPF_FUNC_for_each_map_elem) {
6479 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
6480 set_map_elem_callback_state);
6485 if (func_id == BPF_FUNC_timer_set_callback) {
6486 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
6487 set_timer_callback_state);
6492 if (func_id == BPF_FUNC_snprintf) {
6493 err = check_bpf_snprintf_call(env, regs);
6498 /* reset caller saved regs */
6499 for (i = 0; i < CALLER_SAVED_REGS; i++) {
6500 mark_reg_not_init(env, regs, caller_saved[i]);
6501 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
6504 /* helper call returns 64-bit value. */
6505 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
6507 /* update return register (already marked as written above) */
6508 if (fn->ret_type == RET_INTEGER) {
6509 /* sets type to SCALAR_VALUE */
6510 mark_reg_unknown(env, regs, BPF_REG_0);
6511 } else if (fn->ret_type == RET_VOID) {
6512 regs[BPF_REG_0].type = NOT_INIT;
6513 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL ||
6514 fn->ret_type == RET_PTR_TO_MAP_VALUE) {
6515 /* There is no offset yet applied, variable or fixed */
6516 mark_reg_known_zero(env, regs, BPF_REG_0);
6517 /* remember map_ptr, so that check_map_access()
6518 * can check 'value_size' boundary of memory access
6519 * to map element returned from bpf_map_lookup_elem()
6521 if (meta.map_ptr == NULL) {
6523 "kernel subsystem misconfigured verifier\n");
6526 regs[BPF_REG_0].map_ptr = meta.map_ptr;
6527 regs[BPF_REG_0].map_uid = meta.map_uid;
6528 if (fn->ret_type == RET_PTR_TO_MAP_VALUE) {
6529 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE;
6530 if (map_value_has_spin_lock(meta.map_ptr))
6531 regs[BPF_REG_0].id = ++env->id_gen;
6533 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
6535 } else if (fn->ret_type == RET_PTR_TO_SOCKET_OR_NULL) {
6536 mark_reg_known_zero(env, regs, BPF_REG_0);
6537 regs[BPF_REG_0].type = PTR_TO_SOCKET_OR_NULL;
6538 } else if (fn->ret_type == RET_PTR_TO_SOCK_COMMON_OR_NULL) {
6539 mark_reg_known_zero(env, regs, BPF_REG_0);
6540 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON_OR_NULL;
6541 } else if (fn->ret_type == RET_PTR_TO_TCP_SOCK_OR_NULL) {
6542 mark_reg_known_zero(env, regs, BPF_REG_0);
6543 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK_OR_NULL;
6544 } else if (fn->ret_type == RET_PTR_TO_ALLOC_MEM_OR_NULL) {
6545 mark_reg_known_zero(env, regs, BPF_REG_0);
6546 regs[BPF_REG_0].type = PTR_TO_MEM_OR_NULL;
6547 regs[BPF_REG_0].mem_size = meta.mem_size;
6548 } else if (fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID_OR_NULL ||
6549 fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID) {
6550 const struct btf_type *t;
6552 mark_reg_known_zero(env, regs, BPF_REG_0);
6553 t = btf_type_skip_modifiers(meta.ret_btf, meta.ret_btf_id, NULL);
6554 if (!btf_type_is_struct(t)) {
6556 const struct btf_type *ret;
6559 /* resolve the type size of ksym. */
6560 ret = btf_resolve_size(meta.ret_btf, t, &tsize);
6562 tname = btf_name_by_offset(meta.ret_btf, t->name_off);
6563 verbose(env, "unable to resolve the size of type '%s': %ld\n",
6564 tname, PTR_ERR(ret));
6567 regs[BPF_REG_0].type =
6568 fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID ?
6569 PTR_TO_MEM : PTR_TO_MEM_OR_NULL;
6570 regs[BPF_REG_0].mem_size = tsize;
6572 regs[BPF_REG_0].type =
6573 fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID ?
6574 PTR_TO_BTF_ID : PTR_TO_BTF_ID_OR_NULL;
6575 regs[BPF_REG_0].btf = meta.ret_btf;
6576 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
6578 } else if (fn->ret_type == RET_PTR_TO_BTF_ID_OR_NULL ||
6579 fn->ret_type == RET_PTR_TO_BTF_ID) {
6582 mark_reg_known_zero(env, regs, BPF_REG_0);
6583 regs[BPF_REG_0].type = fn->ret_type == RET_PTR_TO_BTF_ID ?
6585 PTR_TO_BTF_ID_OR_NULL;
6586 ret_btf_id = *fn->ret_btf_id;
6587 if (ret_btf_id == 0) {
6588 verbose(env, "invalid return type %d of func %s#%d\n",
6589 fn->ret_type, func_id_name(func_id), func_id);
6592 /* current BPF helper definitions are only coming from
6593 * built-in code with type IDs from vmlinux BTF
6595 regs[BPF_REG_0].btf = btf_vmlinux;
6596 regs[BPF_REG_0].btf_id = ret_btf_id;
6598 verbose(env, "unknown return type %d of func %s#%d\n",
6599 fn->ret_type, func_id_name(func_id), func_id);
6603 if (reg_type_may_be_null(regs[BPF_REG_0].type))
6604 regs[BPF_REG_0].id = ++env->id_gen;
6606 if (is_ptr_cast_function(func_id)) {
6607 /* For release_reference() */
6608 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
6609 } else if (is_acquire_function(func_id, meta.map_ptr)) {
6610 int id = acquire_reference_state(env, insn_idx);
6614 /* For mark_ptr_or_null_reg() */
6615 regs[BPF_REG_0].id = id;
6616 /* For release_reference() */
6617 regs[BPF_REG_0].ref_obj_id = id;
6620 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
6622 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
6626 if ((func_id == BPF_FUNC_get_stack ||
6627 func_id == BPF_FUNC_get_task_stack) &&
6628 !env->prog->has_callchain_buf) {
6629 const char *err_str;
6631 #ifdef CONFIG_PERF_EVENTS
6632 err = get_callchain_buffers(sysctl_perf_event_max_stack);
6633 err_str = "cannot get callchain buffer for func %s#%d\n";
6636 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
6639 verbose(env, err_str, func_id_name(func_id), func_id);
6643 env->prog->has_callchain_buf = true;
6646 if (func_id == BPF_FUNC_get_stackid || func_id == BPF_FUNC_get_stack)
6647 env->prog->call_get_stack = true;
6649 if (func_id == BPF_FUNC_get_func_ip) {
6650 if (check_get_func_ip(env))
6652 env->prog->call_get_func_ip = true;
6656 clear_all_pkt_pointers(env);
6660 /* mark_btf_func_reg_size() is used when the reg size is determined by
6661 * the BTF func_proto's return value size and argument.
6663 static void mark_btf_func_reg_size(struct bpf_verifier_env *env, u32 regno,
6666 struct bpf_reg_state *reg = &cur_regs(env)[regno];
6668 if (regno == BPF_REG_0) {
6669 /* Function return value */
6670 reg->live |= REG_LIVE_WRITTEN;
6671 reg->subreg_def = reg_size == sizeof(u64) ?
6672 DEF_NOT_SUBREG : env->insn_idx + 1;
6674 /* Function argument */
6675 if (reg_size == sizeof(u64)) {
6676 mark_insn_zext(env, reg);
6677 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
6679 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ32);
6684 static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn)
6686 const struct btf_type *t, *func, *func_proto, *ptr_type;
6687 struct bpf_reg_state *regs = cur_regs(env);
6688 const char *func_name, *ptr_type_name;
6689 u32 i, nargs, func_id, ptr_type_id;
6690 struct module *btf_mod = NULL;
6691 const struct btf_param *args;
6692 struct btf *desc_btf;
6695 /* skip for now, but return error when we find this in fixup_kfunc_call */
6699 desc_btf = find_kfunc_desc_btf(env, insn->imm, insn->off, &btf_mod);
6700 if (IS_ERR(desc_btf))
6701 return PTR_ERR(desc_btf);
6703 func_id = insn->imm;
6704 func = btf_type_by_id(desc_btf, func_id);
6705 func_name = btf_name_by_offset(desc_btf, func->name_off);
6706 func_proto = btf_type_by_id(desc_btf, func->type);
6708 if (!env->ops->check_kfunc_call ||
6709 !env->ops->check_kfunc_call(func_id, btf_mod)) {
6710 verbose(env, "calling kernel function %s is not allowed\n",
6715 /* Check the arguments */
6716 err = btf_check_kfunc_arg_match(env, desc_btf, func_id, regs);
6720 for (i = 0; i < CALLER_SAVED_REGS; i++)
6721 mark_reg_not_init(env, regs, caller_saved[i]);
6723 /* Check return type */
6724 t = btf_type_skip_modifiers(desc_btf, func_proto->type, NULL);
6725 if (btf_type_is_scalar(t)) {
6726 mark_reg_unknown(env, regs, BPF_REG_0);
6727 mark_btf_func_reg_size(env, BPF_REG_0, t->size);
6728 } else if (btf_type_is_ptr(t)) {
6729 ptr_type = btf_type_skip_modifiers(desc_btf, t->type,
6731 if (!btf_type_is_struct(ptr_type)) {
6732 ptr_type_name = btf_name_by_offset(desc_btf,
6733 ptr_type->name_off);
6734 verbose(env, "kernel function %s returns pointer type %s %s is not supported\n",
6735 func_name, btf_type_str(ptr_type),
6739 mark_reg_known_zero(env, regs, BPF_REG_0);
6740 regs[BPF_REG_0].btf = desc_btf;
6741 regs[BPF_REG_0].type = PTR_TO_BTF_ID;
6742 regs[BPF_REG_0].btf_id = ptr_type_id;
6743 mark_btf_func_reg_size(env, BPF_REG_0, sizeof(void *));
6744 } /* else { add_kfunc_call() ensures it is btf_type_is_void(t) } */
6746 nargs = btf_type_vlen(func_proto);
6747 args = (const struct btf_param *)(func_proto + 1);
6748 for (i = 0; i < nargs; i++) {
6751 t = btf_type_skip_modifiers(desc_btf, args[i].type, NULL);
6752 if (btf_type_is_ptr(t))
6753 mark_btf_func_reg_size(env, regno, sizeof(void *));
6755 /* scalar. ensured by btf_check_kfunc_arg_match() */
6756 mark_btf_func_reg_size(env, regno, t->size);
6762 static bool signed_add_overflows(s64 a, s64 b)
6764 /* Do the add in u64, where overflow is well-defined */
6765 s64 res = (s64)((u64)a + (u64)b);
6772 static bool signed_add32_overflows(s32 a, s32 b)
6774 /* Do the add in u32, where overflow is well-defined */
6775 s32 res = (s32)((u32)a + (u32)b);
6782 static bool signed_sub_overflows(s64 a, s64 b)
6784 /* Do the sub in u64, where overflow is well-defined */
6785 s64 res = (s64)((u64)a - (u64)b);
6792 static bool signed_sub32_overflows(s32 a, s32 b)
6794 /* Do the sub in u32, where overflow is well-defined */
6795 s32 res = (s32)((u32)a - (u32)b);
6802 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
6803 const struct bpf_reg_state *reg,
6804 enum bpf_reg_type type)
6806 bool known = tnum_is_const(reg->var_off);
6807 s64 val = reg->var_off.value;
6808 s64 smin = reg->smin_value;
6810 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
6811 verbose(env, "math between %s pointer and %lld is not allowed\n",
6812 reg_type_str[type], val);
6816 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
6817 verbose(env, "%s pointer offset %d is not allowed\n",
6818 reg_type_str[type], reg->off);
6822 if (smin == S64_MIN) {
6823 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
6824 reg_type_str[type]);
6828 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
6829 verbose(env, "value %lld makes %s pointer be out of bounds\n",
6830 smin, reg_type_str[type]);
6837 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
6839 return &env->insn_aux_data[env->insn_idx];
6850 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
6851 u32 *alu_limit, bool mask_to_left)
6853 u32 max = 0, ptr_limit = 0;
6855 switch (ptr_reg->type) {
6857 /* Offset 0 is out-of-bounds, but acceptable start for the
6858 * left direction, see BPF_REG_FP. Also, unknown scalar
6859 * offset where we would need to deal with min/max bounds is
6860 * currently prohibited for unprivileged.
6862 max = MAX_BPF_STACK + mask_to_left;
6863 ptr_limit = -(ptr_reg->var_off.value + ptr_reg->off);
6865 case PTR_TO_MAP_VALUE:
6866 max = ptr_reg->map_ptr->value_size;
6867 ptr_limit = (mask_to_left ?
6868 ptr_reg->smin_value :
6869 ptr_reg->umax_value) + ptr_reg->off;
6875 if (ptr_limit >= max)
6876 return REASON_LIMIT;
6877 *alu_limit = ptr_limit;
6881 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
6882 const struct bpf_insn *insn)
6884 return env->bypass_spec_v1 || BPF_SRC(insn->code) == BPF_K;
6887 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
6888 u32 alu_state, u32 alu_limit)
6890 /* If we arrived here from different branches with different
6891 * state or limits to sanitize, then this won't work.
6893 if (aux->alu_state &&
6894 (aux->alu_state != alu_state ||
6895 aux->alu_limit != alu_limit))
6896 return REASON_PATHS;
6898 /* Corresponding fixup done in do_misc_fixups(). */
6899 aux->alu_state = alu_state;
6900 aux->alu_limit = alu_limit;
6904 static int sanitize_val_alu(struct bpf_verifier_env *env,
6905 struct bpf_insn *insn)
6907 struct bpf_insn_aux_data *aux = cur_aux(env);
6909 if (can_skip_alu_sanitation(env, insn))
6912 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
6915 static bool sanitize_needed(u8 opcode)
6917 return opcode == BPF_ADD || opcode == BPF_SUB;
6920 struct bpf_sanitize_info {
6921 struct bpf_insn_aux_data aux;
6925 static struct bpf_verifier_state *
6926 sanitize_speculative_path(struct bpf_verifier_env *env,
6927 const struct bpf_insn *insn,
6928 u32 next_idx, u32 curr_idx)
6930 struct bpf_verifier_state *branch;
6931 struct bpf_reg_state *regs;
6933 branch = push_stack(env, next_idx, curr_idx, true);
6934 if (branch && insn) {
6935 regs = branch->frame[branch->curframe]->regs;
6936 if (BPF_SRC(insn->code) == BPF_K) {
6937 mark_reg_unknown(env, regs, insn->dst_reg);
6938 } else if (BPF_SRC(insn->code) == BPF_X) {
6939 mark_reg_unknown(env, regs, insn->dst_reg);
6940 mark_reg_unknown(env, regs, insn->src_reg);
6946 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
6947 struct bpf_insn *insn,
6948 const struct bpf_reg_state *ptr_reg,
6949 const struct bpf_reg_state *off_reg,
6950 struct bpf_reg_state *dst_reg,
6951 struct bpf_sanitize_info *info,
6952 const bool commit_window)
6954 struct bpf_insn_aux_data *aux = commit_window ? cur_aux(env) : &info->aux;
6955 struct bpf_verifier_state *vstate = env->cur_state;
6956 bool off_is_imm = tnum_is_const(off_reg->var_off);
6957 bool off_is_neg = off_reg->smin_value < 0;
6958 bool ptr_is_dst_reg = ptr_reg == dst_reg;
6959 u8 opcode = BPF_OP(insn->code);
6960 u32 alu_state, alu_limit;
6961 struct bpf_reg_state tmp;
6965 if (can_skip_alu_sanitation(env, insn))
6968 /* We already marked aux for masking from non-speculative
6969 * paths, thus we got here in the first place. We only care
6970 * to explore bad access from here.
6972 if (vstate->speculative)
6975 if (!commit_window) {
6976 if (!tnum_is_const(off_reg->var_off) &&
6977 (off_reg->smin_value < 0) != (off_reg->smax_value < 0))
6978 return REASON_BOUNDS;
6980 info->mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
6981 (opcode == BPF_SUB && !off_is_neg);
6984 err = retrieve_ptr_limit(ptr_reg, &alu_limit, info->mask_to_left);
6988 if (commit_window) {
6989 /* In commit phase we narrow the masking window based on
6990 * the observed pointer move after the simulated operation.
6992 alu_state = info->aux.alu_state;
6993 alu_limit = abs(info->aux.alu_limit - alu_limit);
6995 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
6996 alu_state |= off_is_imm ? BPF_ALU_IMMEDIATE : 0;
6997 alu_state |= ptr_is_dst_reg ?
6998 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
7000 /* Limit pruning on unknown scalars to enable deep search for
7001 * potential masking differences from other program paths.
7004 env->explore_alu_limits = true;
7007 err = update_alu_sanitation_state(aux, alu_state, alu_limit);
7011 /* If we're in commit phase, we're done here given we already
7012 * pushed the truncated dst_reg into the speculative verification
7015 * Also, when register is a known constant, we rewrite register-based
7016 * operation to immediate-based, and thus do not need masking (and as
7017 * a consequence, do not need to simulate the zero-truncation either).
7019 if (commit_window || off_is_imm)
7022 /* Simulate and find potential out-of-bounds access under
7023 * speculative execution from truncation as a result of
7024 * masking when off was not within expected range. If off
7025 * sits in dst, then we temporarily need to move ptr there
7026 * to simulate dst (== 0) +/-= ptr. Needed, for example,
7027 * for cases where we use K-based arithmetic in one direction
7028 * and truncated reg-based in the other in order to explore
7031 if (!ptr_is_dst_reg) {
7033 *dst_reg = *ptr_reg;
7035 ret = sanitize_speculative_path(env, NULL, env->insn_idx + 1,
7037 if (!ptr_is_dst_reg && ret)
7039 return !ret ? REASON_STACK : 0;
7042 static void sanitize_mark_insn_seen(struct bpf_verifier_env *env)
7044 struct bpf_verifier_state *vstate = env->cur_state;
7046 /* If we simulate paths under speculation, we don't update the
7047 * insn as 'seen' such that when we verify unreachable paths in
7048 * the non-speculative domain, sanitize_dead_code() can still
7049 * rewrite/sanitize them.
7051 if (!vstate->speculative)
7052 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
7055 static int sanitize_err(struct bpf_verifier_env *env,
7056 const struct bpf_insn *insn, int reason,
7057 const struct bpf_reg_state *off_reg,
7058 const struct bpf_reg_state *dst_reg)
7060 static const char *err = "pointer arithmetic with it prohibited for !root";
7061 const char *op = BPF_OP(insn->code) == BPF_ADD ? "add" : "sub";
7062 u32 dst = insn->dst_reg, src = insn->src_reg;
7066 verbose(env, "R%d has unknown scalar with mixed signed bounds, %s\n",
7067 off_reg == dst_reg ? dst : src, err);
7070 verbose(env, "R%d has pointer with unsupported alu operation, %s\n",
7071 off_reg == dst_reg ? src : dst, err);
7074 verbose(env, "R%d tried to %s from different maps, paths or scalars, %s\n",
7078 verbose(env, "R%d tried to %s beyond pointer bounds, %s\n",
7082 verbose(env, "R%d could not be pushed for speculative verification, %s\n",
7086 verbose(env, "verifier internal error: unknown reason (%d)\n",
7094 /* check that stack access falls within stack limits and that 'reg' doesn't
7095 * have a variable offset.
7097 * Variable offset is prohibited for unprivileged mode for simplicity since it
7098 * requires corresponding support in Spectre masking for stack ALU. See also
7099 * retrieve_ptr_limit().
7102 * 'off' includes 'reg->off'.
7104 static int check_stack_access_for_ptr_arithmetic(
7105 struct bpf_verifier_env *env,
7107 const struct bpf_reg_state *reg,
7110 if (!tnum_is_const(reg->var_off)) {
7113 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
7114 verbose(env, "R%d variable stack access prohibited for !root, var_off=%s off=%d\n",
7115 regno, tn_buf, off);
7119 if (off >= 0 || off < -MAX_BPF_STACK) {
7120 verbose(env, "R%d stack pointer arithmetic goes out of range, "
7121 "prohibited for !root; off=%d\n", regno, off);
7128 static int sanitize_check_bounds(struct bpf_verifier_env *env,
7129 const struct bpf_insn *insn,
7130 const struct bpf_reg_state *dst_reg)
7132 u32 dst = insn->dst_reg;
7134 /* For unprivileged we require that resulting offset must be in bounds
7135 * in order to be able to sanitize access later on.
7137 if (env->bypass_spec_v1)
7140 switch (dst_reg->type) {
7142 if (check_stack_access_for_ptr_arithmetic(env, dst, dst_reg,
7143 dst_reg->off + dst_reg->var_off.value))
7146 case PTR_TO_MAP_VALUE:
7147 if (check_map_access(env, dst, dst_reg->off, 1, false)) {
7148 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
7149 "prohibited for !root\n", dst);
7160 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
7161 * Caller should also handle BPF_MOV case separately.
7162 * If we return -EACCES, caller may want to try again treating pointer as a
7163 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
7165 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
7166 struct bpf_insn *insn,
7167 const struct bpf_reg_state *ptr_reg,
7168 const struct bpf_reg_state *off_reg)
7170 struct bpf_verifier_state *vstate = env->cur_state;
7171 struct bpf_func_state *state = vstate->frame[vstate->curframe];
7172 struct bpf_reg_state *regs = state->regs, *dst_reg;
7173 bool known = tnum_is_const(off_reg->var_off);
7174 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
7175 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
7176 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
7177 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
7178 struct bpf_sanitize_info info = {};
7179 u8 opcode = BPF_OP(insn->code);
7180 u32 dst = insn->dst_reg;
7183 dst_reg = ®s[dst];
7185 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
7186 smin_val > smax_val || umin_val > umax_val) {
7187 /* Taint dst register if offset had invalid bounds derived from
7188 * e.g. dead branches.
7190 __mark_reg_unknown(env, dst_reg);
7194 if (BPF_CLASS(insn->code) != BPF_ALU64) {
7195 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
7196 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
7197 __mark_reg_unknown(env, dst_reg);
7202 "R%d 32-bit pointer arithmetic prohibited\n",
7207 switch (ptr_reg->type) {
7208 case PTR_TO_MAP_VALUE_OR_NULL:
7209 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
7210 dst, reg_type_str[ptr_reg->type]);
7212 case CONST_PTR_TO_MAP:
7213 /* smin_val represents the known value */
7214 if (known && smin_val == 0 && opcode == BPF_ADD)
7217 case PTR_TO_PACKET_END:
7219 case PTR_TO_SOCKET_OR_NULL:
7220 case PTR_TO_SOCK_COMMON:
7221 case PTR_TO_SOCK_COMMON_OR_NULL:
7222 case PTR_TO_TCP_SOCK:
7223 case PTR_TO_TCP_SOCK_OR_NULL:
7224 case PTR_TO_XDP_SOCK:
7225 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
7226 dst, reg_type_str[ptr_reg->type]);
7232 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
7233 * The id may be overwritten later if we create a new variable offset.
7235 dst_reg->type = ptr_reg->type;
7236 dst_reg->id = ptr_reg->id;
7238 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
7239 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
7242 /* pointer types do not carry 32-bit bounds at the moment. */
7243 __mark_reg32_unbounded(dst_reg);
7245 if (sanitize_needed(opcode)) {
7246 ret = sanitize_ptr_alu(env, insn, ptr_reg, off_reg, dst_reg,
7249 return sanitize_err(env, insn, ret, off_reg, dst_reg);
7254 /* We can take a fixed offset as long as it doesn't overflow
7255 * the s32 'off' field
7257 if (known && (ptr_reg->off + smin_val ==
7258 (s64)(s32)(ptr_reg->off + smin_val))) {
7259 /* pointer += K. Accumulate it into fixed offset */
7260 dst_reg->smin_value = smin_ptr;
7261 dst_reg->smax_value = smax_ptr;
7262 dst_reg->umin_value = umin_ptr;
7263 dst_reg->umax_value = umax_ptr;
7264 dst_reg->var_off = ptr_reg->var_off;
7265 dst_reg->off = ptr_reg->off + smin_val;
7266 dst_reg->raw = ptr_reg->raw;
7269 /* A new variable offset is created. Note that off_reg->off
7270 * == 0, since it's a scalar.
7271 * dst_reg gets the pointer type and since some positive
7272 * integer value was added to the pointer, give it a new 'id'
7273 * if it's a PTR_TO_PACKET.
7274 * this creates a new 'base' pointer, off_reg (variable) gets
7275 * added into the variable offset, and we copy the fixed offset
7278 if (signed_add_overflows(smin_ptr, smin_val) ||
7279 signed_add_overflows(smax_ptr, smax_val)) {
7280 dst_reg->smin_value = S64_MIN;
7281 dst_reg->smax_value = S64_MAX;
7283 dst_reg->smin_value = smin_ptr + smin_val;
7284 dst_reg->smax_value = smax_ptr + smax_val;
7286 if (umin_ptr + umin_val < umin_ptr ||
7287 umax_ptr + umax_val < umax_ptr) {
7288 dst_reg->umin_value = 0;
7289 dst_reg->umax_value = U64_MAX;
7291 dst_reg->umin_value = umin_ptr + umin_val;
7292 dst_reg->umax_value = umax_ptr + umax_val;
7294 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
7295 dst_reg->off = ptr_reg->off;
7296 dst_reg->raw = ptr_reg->raw;
7297 if (reg_is_pkt_pointer(ptr_reg)) {
7298 dst_reg->id = ++env->id_gen;
7299 /* something was added to pkt_ptr, set range to zero */
7300 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
7304 if (dst_reg == off_reg) {
7305 /* scalar -= pointer. Creates an unknown scalar */
7306 verbose(env, "R%d tried to subtract pointer from scalar\n",
7310 /* We don't allow subtraction from FP, because (according to
7311 * test_verifier.c test "invalid fp arithmetic", JITs might not
7312 * be able to deal with it.
7314 if (ptr_reg->type == PTR_TO_STACK) {
7315 verbose(env, "R%d subtraction from stack pointer prohibited\n",
7319 if (known && (ptr_reg->off - smin_val ==
7320 (s64)(s32)(ptr_reg->off - smin_val))) {
7321 /* pointer -= K. Subtract it from fixed offset */
7322 dst_reg->smin_value = smin_ptr;
7323 dst_reg->smax_value = smax_ptr;
7324 dst_reg->umin_value = umin_ptr;
7325 dst_reg->umax_value = umax_ptr;
7326 dst_reg->var_off = ptr_reg->var_off;
7327 dst_reg->id = ptr_reg->id;
7328 dst_reg->off = ptr_reg->off - smin_val;
7329 dst_reg->raw = ptr_reg->raw;
7332 /* A new variable offset is created. If the subtrahend is known
7333 * nonnegative, then any reg->range we had before is still good.
7335 if (signed_sub_overflows(smin_ptr, smax_val) ||
7336 signed_sub_overflows(smax_ptr, smin_val)) {
7337 /* Overflow possible, we know nothing */
7338 dst_reg->smin_value = S64_MIN;
7339 dst_reg->smax_value = S64_MAX;
7341 dst_reg->smin_value = smin_ptr - smax_val;
7342 dst_reg->smax_value = smax_ptr - smin_val;
7344 if (umin_ptr < umax_val) {
7345 /* Overflow possible, we know nothing */
7346 dst_reg->umin_value = 0;
7347 dst_reg->umax_value = U64_MAX;
7349 /* Cannot overflow (as long as bounds are consistent) */
7350 dst_reg->umin_value = umin_ptr - umax_val;
7351 dst_reg->umax_value = umax_ptr - umin_val;
7353 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
7354 dst_reg->off = ptr_reg->off;
7355 dst_reg->raw = ptr_reg->raw;
7356 if (reg_is_pkt_pointer(ptr_reg)) {
7357 dst_reg->id = ++env->id_gen;
7358 /* something was added to pkt_ptr, set range to zero */
7360 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
7366 /* bitwise ops on pointers are troublesome, prohibit. */
7367 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
7368 dst, bpf_alu_string[opcode >> 4]);
7371 /* other operators (e.g. MUL,LSH) produce non-pointer results */
7372 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
7373 dst, bpf_alu_string[opcode >> 4]);
7377 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
7380 __update_reg_bounds(dst_reg);
7381 __reg_deduce_bounds(dst_reg);
7382 __reg_bound_offset(dst_reg);
7384 if (sanitize_check_bounds(env, insn, dst_reg) < 0)
7386 if (sanitize_needed(opcode)) {
7387 ret = sanitize_ptr_alu(env, insn, dst_reg, off_reg, dst_reg,
7390 return sanitize_err(env, insn, ret, off_reg, dst_reg);
7396 static void scalar32_min_max_add(struct bpf_reg_state *dst_reg,
7397 struct bpf_reg_state *src_reg)
7399 s32 smin_val = src_reg->s32_min_value;
7400 s32 smax_val = src_reg->s32_max_value;
7401 u32 umin_val = src_reg->u32_min_value;
7402 u32 umax_val = src_reg->u32_max_value;
7404 if (signed_add32_overflows(dst_reg->s32_min_value, smin_val) ||
7405 signed_add32_overflows(dst_reg->s32_max_value, smax_val)) {
7406 dst_reg->s32_min_value = S32_MIN;
7407 dst_reg->s32_max_value = S32_MAX;
7409 dst_reg->s32_min_value += smin_val;
7410 dst_reg->s32_max_value += smax_val;
7412 if (dst_reg->u32_min_value + umin_val < umin_val ||
7413 dst_reg->u32_max_value + umax_val < umax_val) {
7414 dst_reg->u32_min_value = 0;
7415 dst_reg->u32_max_value = U32_MAX;
7417 dst_reg->u32_min_value += umin_val;
7418 dst_reg->u32_max_value += umax_val;
7422 static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
7423 struct bpf_reg_state *src_reg)
7425 s64 smin_val = src_reg->smin_value;
7426 s64 smax_val = src_reg->smax_value;
7427 u64 umin_val = src_reg->umin_value;
7428 u64 umax_val = src_reg->umax_value;
7430 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
7431 signed_add_overflows(dst_reg->smax_value, smax_val)) {
7432 dst_reg->smin_value = S64_MIN;
7433 dst_reg->smax_value = S64_MAX;
7435 dst_reg->smin_value += smin_val;
7436 dst_reg->smax_value += smax_val;
7438 if (dst_reg->umin_value + umin_val < umin_val ||
7439 dst_reg->umax_value + umax_val < umax_val) {
7440 dst_reg->umin_value = 0;
7441 dst_reg->umax_value = U64_MAX;
7443 dst_reg->umin_value += umin_val;
7444 dst_reg->umax_value += umax_val;
7448 static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg,
7449 struct bpf_reg_state *src_reg)
7451 s32 smin_val = src_reg->s32_min_value;
7452 s32 smax_val = src_reg->s32_max_value;
7453 u32 umin_val = src_reg->u32_min_value;
7454 u32 umax_val = src_reg->u32_max_value;
7456 if (signed_sub32_overflows(dst_reg->s32_min_value, smax_val) ||
7457 signed_sub32_overflows(dst_reg->s32_max_value, smin_val)) {
7458 /* Overflow possible, we know nothing */
7459 dst_reg->s32_min_value = S32_MIN;
7460 dst_reg->s32_max_value = S32_MAX;
7462 dst_reg->s32_min_value -= smax_val;
7463 dst_reg->s32_max_value -= smin_val;
7465 if (dst_reg->u32_min_value < umax_val) {
7466 /* Overflow possible, we know nothing */
7467 dst_reg->u32_min_value = 0;
7468 dst_reg->u32_max_value = U32_MAX;
7470 /* Cannot overflow (as long as bounds are consistent) */
7471 dst_reg->u32_min_value -= umax_val;
7472 dst_reg->u32_max_value -= umin_val;
7476 static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
7477 struct bpf_reg_state *src_reg)
7479 s64 smin_val = src_reg->smin_value;
7480 s64 smax_val = src_reg->smax_value;
7481 u64 umin_val = src_reg->umin_value;
7482 u64 umax_val = src_reg->umax_value;
7484 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
7485 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
7486 /* Overflow possible, we know nothing */
7487 dst_reg->smin_value = S64_MIN;
7488 dst_reg->smax_value = S64_MAX;
7490 dst_reg->smin_value -= smax_val;
7491 dst_reg->smax_value -= smin_val;
7493 if (dst_reg->umin_value < umax_val) {
7494 /* Overflow possible, we know nothing */
7495 dst_reg->umin_value = 0;
7496 dst_reg->umax_value = U64_MAX;
7498 /* Cannot overflow (as long as bounds are consistent) */
7499 dst_reg->umin_value -= umax_val;
7500 dst_reg->umax_value -= umin_val;
7504 static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg,
7505 struct bpf_reg_state *src_reg)
7507 s32 smin_val = src_reg->s32_min_value;
7508 u32 umin_val = src_reg->u32_min_value;
7509 u32 umax_val = src_reg->u32_max_value;
7511 if (smin_val < 0 || dst_reg->s32_min_value < 0) {
7512 /* Ain't nobody got time to multiply that sign */
7513 __mark_reg32_unbounded(dst_reg);
7516 /* Both values are positive, so we can work with unsigned and
7517 * copy the result to signed (unless it exceeds S32_MAX).
7519 if (umax_val > U16_MAX || dst_reg->u32_max_value > U16_MAX) {
7520 /* Potential overflow, we know nothing */
7521 __mark_reg32_unbounded(dst_reg);
7524 dst_reg->u32_min_value *= umin_val;
7525 dst_reg->u32_max_value *= umax_val;
7526 if (dst_reg->u32_max_value > S32_MAX) {
7527 /* Overflow possible, we know nothing */
7528 dst_reg->s32_min_value = S32_MIN;
7529 dst_reg->s32_max_value = S32_MAX;
7531 dst_reg->s32_min_value = dst_reg->u32_min_value;
7532 dst_reg->s32_max_value = dst_reg->u32_max_value;
7536 static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
7537 struct bpf_reg_state *src_reg)
7539 s64 smin_val = src_reg->smin_value;
7540 u64 umin_val = src_reg->umin_value;
7541 u64 umax_val = src_reg->umax_value;
7543 if (smin_val < 0 || dst_reg->smin_value < 0) {
7544 /* Ain't nobody got time to multiply that sign */
7545 __mark_reg64_unbounded(dst_reg);
7548 /* Both values are positive, so we can work with unsigned and
7549 * copy the result to signed (unless it exceeds S64_MAX).
7551 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
7552 /* Potential overflow, we know nothing */
7553 __mark_reg64_unbounded(dst_reg);
7556 dst_reg->umin_value *= umin_val;
7557 dst_reg->umax_value *= umax_val;
7558 if (dst_reg->umax_value > S64_MAX) {
7559 /* Overflow possible, we know nothing */
7560 dst_reg->smin_value = S64_MIN;
7561 dst_reg->smax_value = S64_MAX;
7563 dst_reg->smin_value = dst_reg->umin_value;
7564 dst_reg->smax_value = dst_reg->umax_value;
7568 static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
7569 struct bpf_reg_state *src_reg)
7571 bool src_known = tnum_subreg_is_const(src_reg->var_off);
7572 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
7573 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
7574 s32 smin_val = src_reg->s32_min_value;
7575 u32 umax_val = src_reg->u32_max_value;
7577 if (src_known && dst_known) {
7578 __mark_reg32_known(dst_reg, var32_off.value);
7582 /* We get our minimum from the var_off, since that's inherently
7583 * bitwise. Our maximum is the minimum of the operands' maxima.
7585 dst_reg->u32_min_value = var32_off.value;
7586 dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val);
7587 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
7588 /* Lose signed bounds when ANDing negative numbers,
7589 * ain't nobody got time for that.
7591 dst_reg->s32_min_value = S32_MIN;
7592 dst_reg->s32_max_value = S32_MAX;
7594 /* ANDing two positives gives a positive, so safe to
7595 * cast result into s64.
7597 dst_reg->s32_min_value = dst_reg->u32_min_value;
7598 dst_reg->s32_max_value = dst_reg->u32_max_value;
7602 static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
7603 struct bpf_reg_state *src_reg)
7605 bool src_known = tnum_is_const(src_reg->var_off);
7606 bool dst_known = tnum_is_const(dst_reg->var_off);
7607 s64 smin_val = src_reg->smin_value;
7608 u64 umax_val = src_reg->umax_value;
7610 if (src_known && dst_known) {
7611 __mark_reg_known(dst_reg, dst_reg->var_off.value);
7615 /* We get our minimum from the var_off, since that's inherently
7616 * bitwise. Our maximum is the minimum of the operands' maxima.
7618 dst_reg->umin_value = dst_reg->var_off.value;
7619 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
7620 if (dst_reg->smin_value < 0 || smin_val < 0) {
7621 /* Lose signed bounds when ANDing negative numbers,
7622 * ain't nobody got time for that.
7624 dst_reg->smin_value = S64_MIN;
7625 dst_reg->smax_value = S64_MAX;
7627 /* ANDing two positives gives a positive, so safe to
7628 * cast result into s64.
7630 dst_reg->smin_value = dst_reg->umin_value;
7631 dst_reg->smax_value = dst_reg->umax_value;
7633 /* We may learn something more from the var_off */
7634 __update_reg_bounds(dst_reg);
7637 static void scalar32_min_max_or(struct bpf_reg_state *dst_reg,
7638 struct bpf_reg_state *src_reg)
7640 bool src_known = tnum_subreg_is_const(src_reg->var_off);
7641 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
7642 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
7643 s32 smin_val = src_reg->s32_min_value;
7644 u32 umin_val = src_reg->u32_min_value;
7646 if (src_known && dst_known) {
7647 __mark_reg32_known(dst_reg, var32_off.value);
7651 /* We get our maximum from the var_off, and our minimum is the
7652 * maximum of the operands' minima
7654 dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val);
7655 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
7656 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
7657 /* Lose signed bounds when ORing negative numbers,
7658 * ain't nobody got time for that.
7660 dst_reg->s32_min_value = S32_MIN;
7661 dst_reg->s32_max_value = S32_MAX;
7663 /* ORing two positives gives a positive, so safe to
7664 * cast result into s64.
7666 dst_reg->s32_min_value = dst_reg->u32_min_value;
7667 dst_reg->s32_max_value = dst_reg->u32_max_value;
7671 static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
7672 struct bpf_reg_state *src_reg)
7674 bool src_known = tnum_is_const(src_reg->var_off);
7675 bool dst_known = tnum_is_const(dst_reg->var_off);
7676 s64 smin_val = src_reg->smin_value;
7677 u64 umin_val = src_reg->umin_value;
7679 if (src_known && dst_known) {
7680 __mark_reg_known(dst_reg, dst_reg->var_off.value);
7684 /* We get our maximum from the var_off, and our minimum is the
7685 * maximum of the operands' minima
7687 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
7688 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
7689 if (dst_reg->smin_value < 0 || smin_val < 0) {
7690 /* Lose signed bounds when ORing negative numbers,
7691 * ain't nobody got time for that.
7693 dst_reg->smin_value = S64_MIN;
7694 dst_reg->smax_value = S64_MAX;
7696 /* ORing two positives gives a positive, so safe to
7697 * cast result into s64.
7699 dst_reg->smin_value = dst_reg->umin_value;
7700 dst_reg->smax_value = dst_reg->umax_value;
7702 /* We may learn something more from the var_off */
7703 __update_reg_bounds(dst_reg);
7706 static void scalar32_min_max_xor(struct bpf_reg_state *dst_reg,
7707 struct bpf_reg_state *src_reg)
7709 bool src_known = tnum_subreg_is_const(src_reg->var_off);
7710 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
7711 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
7712 s32 smin_val = src_reg->s32_min_value;
7714 if (src_known && dst_known) {
7715 __mark_reg32_known(dst_reg, var32_off.value);
7719 /* We get both minimum and maximum from the var32_off. */
7720 dst_reg->u32_min_value = var32_off.value;
7721 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
7723 if (dst_reg->s32_min_value >= 0 && smin_val >= 0) {
7724 /* XORing two positive sign numbers gives a positive,
7725 * so safe to cast u32 result into s32.
7727 dst_reg->s32_min_value = dst_reg->u32_min_value;
7728 dst_reg->s32_max_value = dst_reg->u32_max_value;
7730 dst_reg->s32_min_value = S32_MIN;
7731 dst_reg->s32_max_value = S32_MAX;
7735 static void scalar_min_max_xor(struct bpf_reg_state *dst_reg,
7736 struct bpf_reg_state *src_reg)
7738 bool src_known = tnum_is_const(src_reg->var_off);
7739 bool dst_known = tnum_is_const(dst_reg->var_off);
7740 s64 smin_val = src_reg->smin_value;
7742 if (src_known && dst_known) {
7743 /* dst_reg->var_off.value has been updated earlier */
7744 __mark_reg_known(dst_reg, dst_reg->var_off.value);
7748 /* We get both minimum and maximum from the var_off. */
7749 dst_reg->umin_value = dst_reg->var_off.value;
7750 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
7752 if (dst_reg->smin_value >= 0 && smin_val >= 0) {
7753 /* XORing two positive sign numbers gives a positive,
7754 * so safe to cast u64 result into s64.
7756 dst_reg->smin_value = dst_reg->umin_value;
7757 dst_reg->smax_value = dst_reg->umax_value;
7759 dst_reg->smin_value = S64_MIN;
7760 dst_reg->smax_value = S64_MAX;
7763 __update_reg_bounds(dst_reg);
7766 static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
7767 u64 umin_val, u64 umax_val)
7769 /* We lose all sign bit information (except what we can pick
7772 dst_reg->s32_min_value = S32_MIN;
7773 dst_reg->s32_max_value = S32_MAX;
7774 /* If we might shift our top bit out, then we know nothing */
7775 if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) {
7776 dst_reg->u32_min_value = 0;
7777 dst_reg->u32_max_value = U32_MAX;
7779 dst_reg->u32_min_value <<= umin_val;
7780 dst_reg->u32_max_value <<= umax_val;
7784 static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
7785 struct bpf_reg_state *src_reg)
7787 u32 umax_val = src_reg->u32_max_value;
7788 u32 umin_val = src_reg->u32_min_value;
7789 /* u32 alu operation will zext upper bits */
7790 struct tnum subreg = tnum_subreg(dst_reg->var_off);
7792 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
7793 dst_reg->var_off = tnum_subreg(tnum_lshift(subreg, umin_val));
7794 /* Not required but being careful mark reg64 bounds as unknown so
7795 * that we are forced to pick them up from tnum and zext later and
7796 * if some path skips this step we are still safe.
7798 __mark_reg64_unbounded(dst_reg);
7799 __update_reg32_bounds(dst_reg);
7802 static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg,
7803 u64 umin_val, u64 umax_val)
7805 /* Special case <<32 because it is a common compiler pattern to sign
7806 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
7807 * positive we know this shift will also be positive so we can track
7808 * bounds correctly. Otherwise we lose all sign bit information except
7809 * what we can pick up from var_off. Perhaps we can generalize this
7810 * later to shifts of any length.
7812 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0)
7813 dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32;
7815 dst_reg->smax_value = S64_MAX;
7817 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0)
7818 dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32;
7820 dst_reg->smin_value = S64_MIN;
7822 /* If we might shift our top bit out, then we know nothing */
7823 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
7824 dst_reg->umin_value = 0;
7825 dst_reg->umax_value = U64_MAX;
7827 dst_reg->umin_value <<= umin_val;
7828 dst_reg->umax_value <<= umax_val;
7832 static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
7833 struct bpf_reg_state *src_reg)
7835 u64 umax_val = src_reg->umax_value;
7836 u64 umin_val = src_reg->umin_value;
7838 /* scalar64 calc uses 32bit unshifted bounds so must be called first */
7839 __scalar64_min_max_lsh(dst_reg, umin_val, umax_val);
7840 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
7842 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
7843 /* We may learn something more from the var_off */
7844 __update_reg_bounds(dst_reg);
7847 static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg,
7848 struct bpf_reg_state *src_reg)
7850 struct tnum subreg = tnum_subreg(dst_reg->var_off);
7851 u32 umax_val = src_reg->u32_max_value;
7852 u32 umin_val = src_reg->u32_min_value;
7854 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
7855 * be negative, then either:
7856 * 1) src_reg might be zero, so the sign bit of the result is
7857 * unknown, so we lose our signed bounds
7858 * 2) it's known negative, thus the unsigned bounds capture the
7860 * 3) the signed bounds cross zero, so they tell us nothing
7862 * If the value in dst_reg is known nonnegative, then again the
7863 * unsigned bounds capture the signed bounds.
7864 * Thus, in all cases it suffices to blow away our signed bounds
7865 * and rely on inferring new ones from the unsigned bounds and
7866 * var_off of the result.
7868 dst_reg->s32_min_value = S32_MIN;
7869 dst_reg->s32_max_value = S32_MAX;
7871 dst_reg->var_off = tnum_rshift(subreg, umin_val);
7872 dst_reg->u32_min_value >>= umax_val;
7873 dst_reg->u32_max_value >>= umin_val;
7875 __mark_reg64_unbounded(dst_reg);
7876 __update_reg32_bounds(dst_reg);
7879 static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
7880 struct bpf_reg_state *src_reg)
7882 u64 umax_val = src_reg->umax_value;
7883 u64 umin_val = src_reg->umin_value;
7885 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
7886 * be negative, then either:
7887 * 1) src_reg might be zero, so the sign bit of the result is
7888 * unknown, so we lose our signed bounds
7889 * 2) it's known negative, thus the unsigned bounds capture the
7891 * 3) the signed bounds cross zero, so they tell us nothing
7893 * If the value in dst_reg is known nonnegative, then again the
7894 * unsigned bounds capture the signed bounds.
7895 * Thus, in all cases it suffices to blow away our signed bounds
7896 * and rely on inferring new ones from the unsigned bounds and
7897 * var_off of the result.
7899 dst_reg->smin_value = S64_MIN;
7900 dst_reg->smax_value = S64_MAX;
7901 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
7902 dst_reg->umin_value >>= umax_val;
7903 dst_reg->umax_value >>= umin_val;
7905 /* Its not easy to operate on alu32 bounds here because it depends
7906 * on bits being shifted in. Take easy way out and mark unbounded
7907 * so we can recalculate later from tnum.
7909 __mark_reg32_unbounded(dst_reg);
7910 __update_reg_bounds(dst_reg);
7913 static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg,
7914 struct bpf_reg_state *src_reg)
7916 u64 umin_val = src_reg->u32_min_value;
7918 /* Upon reaching here, src_known is true and
7919 * umax_val is equal to umin_val.
7921 dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val);
7922 dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val);
7924 dst_reg->var_off = tnum_arshift(tnum_subreg(dst_reg->var_off), umin_val, 32);
7926 /* blow away the dst_reg umin_value/umax_value and rely on
7927 * dst_reg var_off to refine the result.
7929 dst_reg->u32_min_value = 0;
7930 dst_reg->u32_max_value = U32_MAX;
7932 __mark_reg64_unbounded(dst_reg);
7933 __update_reg32_bounds(dst_reg);
7936 static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
7937 struct bpf_reg_state *src_reg)
7939 u64 umin_val = src_reg->umin_value;
7941 /* Upon reaching here, src_known is true and umax_val is equal
7944 dst_reg->smin_value >>= umin_val;
7945 dst_reg->smax_value >>= umin_val;
7947 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val, 64);
7949 /* blow away the dst_reg umin_value/umax_value and rely on
7950 * dst_reg var_off to refine the result.
7952 dst_reg->umin_value = 0;
7953 dst_reg->umax_value = U64_MAX;
7955 /* Its not easy to operate on alu32 bounds here because it depends
7956 * on bits being shifted in from upper 32-bits. Take easy way out
7957 * and mark unbounded so we can recalculate later from tnum.
7959 __mark_reg32_unbounded(dst_reg);
7960 __update_reg_bounds(dst_reg);
7963 /* WARNING: This function does calculations on 64-bit values, but the actual
7964 * execution may occur on 32-bit values. Therefore, things like bitshifts
7965 * need extra checks in the 32-bit case.
7967 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
7968 struct bpf_insn *insn,
7969 struct bpf_reg_state *dst_reg,
7970 struct bpf_reg_state src_reg)
7972 struct bpf_reg_state *regs = cur_regs(env);
7973 u8 opcode = BPF_OP(insn->code);
7975 s64 smin_val, smax_val;
7976 u64 umin_val, umax_val;
7977 s32 s32_min_val, s32_max_val;
7978 u32 u32_min_val, u32_max_val;
7979 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
7980 bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
7983 smin_val = src_reg.smin_value;
7984 smax_val = src_reg.smax_value;
7985 umin_val = src_reg.umin_value;
7986 umax_val = src_reg.umax_value;
7988 s32_min_val = src_reg.s32_min_value;
7989 s32_max_val = src_reg.s32_max_value;
7990 u32_min_val = src_reg.u32_min_value;
7991 u32_max_val = src_reg.u32_max_value;
7994 src_known = tnum_subreg_is_const(src_reg.var_off);
7996 (s32_min_val != s32_max_val || u32_min_val != u32_max_val)) ||
7997 s32_min_val > s32_max_val || u32_min_val > u32_max_val) {
7998 /* Taint dst register if offset had invalid bounds
7999 * derived from e.g. dead branches.
8001 __mark_reg_unknown(env, dst_reg);
8005 src_known = tnum_is_const(src_reg.var_off);
8007 (smin_val != smax_val || umin_val != umax_val)) ||
8008 smin_val > smax_val || umin_val > umax_val) {
8009 /* Taint dst register if offset had invalid bounds
8010 * derived from e.g. dead branches.
8012 __mark_reg_unknown(env, dst_reg);
8018 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
8019 __mark_reg_unknown(env, dst_reg);
8023 if (sanitize_needed(opcode)) {
8024 ret = sanitize_val_alu(env, insn);
8026 return sanitize_err(env, insn, ret, NULL, NULL);
8029 /* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
8030 * There are two classes of instructions: The first class we track both
8031 * alu32 and alu64 sign/unsigned bounds independently this provides the
8032 * greatest amount of precision when alu operations are mixed with jmp32
8033 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
8034 * and BPF_OR. This is possible because these ops have fairly easy to
8035 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
8036 * See alu32 verifier tests for examples. The second class of
8037 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
8038 * with regards to tracking sign/unsigned bounds because the bits may
8039 * cross subreg boundaries in the alu64 case. When this happens we mark
8040 * the reg unbounded in the subreg bound space and use the resulting
8041 * tnum to calculate an approximation of the sign/unsigned bounds.
8045 scalar32_min_max_add(dst_reg, &src_reg);
8046 scalar_min_max_add(dst_reg, &src_reg);
8047 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
8050 scalar32_min_max_sub(dst_reg, &src_reg);
8051 scalar_min_max_sub(dst_reg, &src_reg);
8052 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
8055 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
8056 scalar32_min_max_mul(dst_reg, &src_reg);
8057 scalar_min_max_mul(dst_reg, &src_reg);
8060 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
8061 scalar32_min_max_and(dst_reg, &src_reg);
8062 scalar_min_max_and(dst_reg, &src_reg);
8065 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
8066 scalar32_min_max_or(dst_reg, &src_reg);
8067 scalar_min_max_or(dst_reg, &src_reg);
8070 dst_reg->var_off = tnum_xor(dst_reg->var_off, src_reg.var_off);
8071 scalar32_min_max_xor(dst_reg, &src_reg);
8072 scalar_min_max_xor(dst_reg, &src_reg);
8075 if (umax_val >= insn_bitness) {
8076 /* Shifts greater than 31 or 63 are undefined.
8077 * This includes shifts by a negative number.
8079 mark_reg_unknown(env, regs, insn->dst_reg);
8083 scalar32_min_max_lsh(dst_reg, &src_reg);
8085 scalar_min_max_lsh(dst_reg, &src_reg);
8088 if (umax_val >= insn_bitness) {
8089 /* Shifts greater than 31 or 63 are undefined.
8090 * This includes shifts by a negative number.
8092 mark_reg_unknown(env, regs, insn->dst_reg);
8096 scalar32_min_max_rsh(dst_reg, &src_reg);
8098 scalar_min_max_rsh(dst_reg, &src_reg);
8101 if (umax_val >= insn_bitness) {
8102 /* Shifts greater than 31 or 63 are undefined.
8103 * This includes shifts by a negative number.
8105 mark_reg_unknown(env, regs, insn->dst_reg);
8109 scalar32_min_max_arsh(dst_reg, &src_reg);
8111 scalar_min_max_arsh(dst_reg, &src_reg);
8114 mark_reg_unknown(env, regs, insn->dst_reg);
8118 /* ALU32 ops are zero extended into 64bit register */
8120 zext_32_to_64(dst_reg);
8122 __update_reg_bounds(dst_reg);
8123 __reg_deduce_bounds(dst_reg);
8124 __reg_bound_offset(dst_reg);
8128 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
8131 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
8132 struct bpf_insn *insn)
8134 struct bpf_verifier_state *vstate = env->cur_state;
8135 struct bpf_func_state *state = vstate->frame[vstate->curframe];
8136 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
8137 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
8138 u8 opcode = BPF_OP(insn->code);
8141 dst_reg = ®s[insn->dst_reg];
8143 if (dst_reg->type != SCALAR_VALUE)
8146 /* Make sure ID is cleared otherwise dst_reg min/max could be
8147 * incorrectly propagated into other registers by find_equal_scalars()
8150 if (BPF_SRC(insn->code) == BPF_X) {
8151 src_reg = ®s[insn->src_reg];
8152 if (src_reg->type != SCALAR_VALUE) {
8153 if (dst_reg->type != SCALAR_VALUE) {
8154 /* Combining two pointers by any ALU op yields
8155 * an arbitrary scalar. Disallow all math except
8156 * pointer subtraction
8158 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
8159 mark_reg_unknown(env, regs, insn->dst_reg);
8162 verbose(env, "R%d pointer %s pointer prohibited\n",
8164 bpf_alu_string[opcode >> 4]);
8167 /* scalar += pointer
8168 * This is legal, but we have to reverse our
8169 * src/dest handling in computing the range
8171 err = mark_chain_precision(env, insn->dst_reg);
8174 return adjust_ptr_min_max_vals(env, insn,
8177 } else if (ptr_reg) {
8178 /* pointer += scalar */
8179 err = mark_chain_precision(env, insn->src_reg);
8182 return adjust_ptr_min_max_vals(env, insn,
8186 /* Pretend the src is a reg with a known value, since we only
8187 * need to be able to read from this state.
8189 off_reg.type = SCALAR_VALUE;
8190 __mark_reg_known(&off_reg, insn->imm);
8192 if (ptr_reg) /* pointer += K */
8193 return adjust_ptr_min_max_vals(env, insn,
8197 /* Got here implies adding two SCALAR_VALUEs */
8198 if (WARN_ON_ONCE(ptr_reg)) {
8199 print_verifier_state(env, state);
8200 verbose(env, "verifier internal error: unexpected ptr_reg\n");
8203 if (WARN_ON(!src_reg)) {
8204 print_verifier_state(env, state);
8205 verbose(env, "verifier internal error: no src_reg\n");
8208 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
8211 /* check validity of 32-bit and 64-bit arithmetic operations */
8212 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
8214 struct bpf_reg_state *regs = cur_regs(env);
8215 u8 opcode = BPF_OP(insn->code);
8218 if (opcode == BPF_END || opcode == BPF_NEG) {
8219 if (opcode == BPF_NEG) {
8220 if (BPF_SRC(insn->code) != 0 ||
8221 insn->src_reg != BPF_REG_0 ||
8222 insn->off != 0 || insn->imm != 0) {
8223 verbose(env, "BPF_NEG uses reserved fields\n");
8227 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
8228 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
8229 BPF_CLASS(insn->code) == BPF_ALU64) {
8230 verbose(env, "BPF_END uses reserved fields\n");
8235 /* check src operand */
8236 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
8240 if (is_pointer_value(env, insn->dst_reg)) {
8241 verbose(env, "R%d pointer arithmetic prohibited\n",
8246 /* check dest operand */
8247 err = check_reg_arg(env, insn->dst_reg, DST_OP);
8251 } else if (opcode == BPF_MOV) {
8253 if (BPF_SRC(insn->code) == BPF_X) {
8254 if (insn->imm != 0 || insn->off != 0) {
8255 verbose(env, "BPF_MOV uses reserved fields\n");
8259 /* check src operand */
8260 err = check_reg_arg(env, insn->src_reg, SRC_OP);
8264 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
8265 verbose(env, "BPF_MOV uses reserved fields\n");
8270 /* check dest operand, mark as required later */
8271 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
8275 if (BPF_SRC(insn->code) == BPF_X) {
8276 struct bpf_reg_state *src_reg = regs + insn->src_reg;
8277 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
8279 if (BPF_CLASS(insn->code) == BPF_ALU64) {
8281 * copy register state to dest reg
8283 if (src_reg->type == SCALAR_VALUE && !src_reg->id)
8284 /* Assign src and dst registers the same ID
8285 * that will be used by find_equal_scalars()
8286 * to propagate min/max range.
8288 src_reg->id = ++env->id_gen;
8289 *dst_reg = *src_reg;
8290 dst_reg->live |= REG_LIVE_WRITTEN;
8291 dst_reg->subreg_def = DEF_NOT_SUBREG;
8294 if (is_pointer_value(env, insn->src_reg)) {
8296 "R%d partial copy of pointer\n",
8299 } else if (src_reg->type == SCALAR_VALUE) {
8300 *dst_reg = *src_reg;
8301 /* Make sure ID is cleared otherwise
8302 * dst_reg min/max could be incorrectly
8303 * propagated into src_reg by find_equal_scalars()
8306 dst_reg->live |= REG_LIVE_WRITTEN;
8307 dst_reg->subreg_def = env->insn_idx + 1;
8309 mark_reg_unknown(env, regs,
8312 zext_32_to_64(dst_reg);
8316 * remember the value we stored into this reg
8318 /* clear any state __mark_reg_known doesn't set */
8319 mark_reg_unknown(env, regs, insn->dst_reg);
8320 regs[insn->dst_reg].type = SCALAR_VALUE;
8321 if (BPF_CLASS(insn->code) == BPF_ALU64) {
8322 __mark_reg_known(regs + insn->dst_reg,
8325 __mark_reg_known(regs + insn->dst_reg,
8330 } else if (opcode > BPF_END) {
8331 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
8334 } else { /* all other ALU ops: and, sub, xor, add, ... */
8336 if (BPF_SRC(insn->code) == BPF_X) {
8337 if (insn->imm != 0 || insn->off != 0) {
8338 verbose(env, "BPF_ALU uses reserved fields\n");
8341 /* check src1 operand */
8342 err = check_reg_arg(env, insn->src_reg, SRC_OP);
8346 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
8347 verbose(env, "BPF_ALU uses reserved fields\n");
8352 /* check src2 operand */
8353 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
8357 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
8358 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
8359 verbose(env, "div by zero\n");
8363 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
8364 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
8365 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
8367 if (insn->imm < 0 || insn->imm >= size) {
8368 verbose(env, "invalid shift %d\n", insn->imm);
8373 /* check dest operand */
8374 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
8378 return adjust_reg_min_max_vals(env, insn);
8384 static void __find_good_pkt_pointers(struct bpf_func_state *state,
8385 struct bpf_reg_state *dst_reg,
8386 enum bpf_reg_type type, int new_range)
8388 struct bpf_reg_state *reg;
8391 for (i = 0; i < MAX_BPF_REG; i++) {
8392 reg = &state->regs[i];
8393 if (reg->type == type && reg->id == dst_reg->id)
8394 /* keep the maximum range already checked */
8395 reg->range = max(reg->range, new_range);
8398 bpf_for_each_spilled_reg(i, state, reg) {
8401 if (reg->type == type && reg->id == dst_reg->id)
8402 reg->range = max(reg->range, new_range);
8406 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
8407 struct bpf_reg_state *dst_reg,
8408 enum bpf_reg_type type,
8409 bool range_right_open)
8413 if (dst_reg->off < 0 ||
8414 (dst_reg->off == 0 && range_right_open))
8415 /* This doesn't give us any range */
8418 if (dst_reg->umax_value > MAX_PACKET_OFF ||
8419 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
8420 /* Risk of overflow. For instance, ptr + (1<<63) may be less
8421 * than pkt_end, but that's because it's also less than pkt.
8425 new_range = dst_reg->off;
8426 if (range_right_open)
8429 /* Examples for register markings:
8431 * pkt_data in dst register:
8435 * if (r2 > pkt_end) goto <handle exception>
8440 * if (r2 < pkt_end) goto <access okay>
8441 * <handle exception>
8444 * r2 == dst_reg, pkt_end == src_reg
8445 * r2=pkt(id=n,off=8,r=0)
8446 * r3=pkt(id=n,off=0,r=0)
8448 * pkt_data in src register:
8452 * if (pkt_end >= r2) goto <access okay>
8453 * <handle exception>
8457 * if (pkt_end <= r2) goto <handle exception>
8461 * pkt_end == dst_reg, r2 == src_reg
8462 * r2=pkt(id=n,off=8,r=0)
8463 * r3=pkt(id=n,off=0,r=0)
8465 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
8466 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
8467 * and [r3, r3 + 8-1) respectively is safe to access depending on
8471 /* If our ids match, then we must have the same max_value. And we
8472 * don't care about the other reg's fixed offset, since if it's too big
8473 * the range won't allow anything.
8474 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
8476 for (i = 0; i <= vstate->curframe; i++)
8477 __find_good_pkt_pointers(vstate->frame[i], dst_reg, type,
8481 static int is_branch32_taken(struct bpf_reg_state *reg, u32 val, u8 opcode)
8483 struct tnum subreg = tnum_subreg(reg->var_off);
8484 s32 sval = (s32)val;
8488 if (tnum_is_const(subreg))
8489 return !!tnum_equals_const(subreg, val);
8492 if (tnum_is_const(subreg))
8493 return !tnum_equals_const(subreg, val);
8496 if ((~subreg.mask & subreg.value) & val)
8498 if (!((subreg.mask | subreg.value) & val))
8502 if (reg->u32_min_value > val)
8504 else if (reg->u32_max_value <= val)
8508 if (reg->s32_min_value > sval)
8510 else if (reg->s32_max_value <= sval)
8514 if (reg->u32_max_value < val)
8516 else if (reg->u32_min_value >= val)
8520 if (reg->s32_max_value < sval)
8522 else if (reg->s32_min_value >= sval)
8526 if (reg->u32_min_value >= val)
8528 else if (reg->u32_max_value < val)
8532 if (reg->s32_min_value >= sval)
8534 else if (reg->s32_max_value < sval)
8538 if (reg->u32_max_value <= val)
8540 else if (reg->u32_min_value > val)
8544 if (reg->s32_max_value <= sval)
8546 else if (reg->s32_min_value > sval)
8555 static int is_branch64_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
8557 s64 sval = (s64)val;
8561 if (tnum_is_const(reg->var_off))
8562 return !!tnum_equals_const(reg->var_off, val);
8565 if (tnum_is_const(reg->var_off))
8566 return !tnum_equals_const(reg->var_off, val);
8569 if ((~reg->var_off.mask & reg->var_off.value) & val)
8571 if (!((reg->var_off.mask | reg->var_off.value) & val))
8575 if (reg->umin_value > val)
8577 else if (reg->umax_value <= val)
8581 if (reg->smin_value > sval)
8583 else if (reg->smax_value <= sval)
8587 if (reg->umax_value < val)
8589 else if (reg->umin_value >= val)
8593 if (reg->smax_value < sval)
8595 else if (reg->smin_value >= sval)
8599 if (reg->umin_value >= val)
8601 else if (reg->umax_value < val)
8605 if (reg->smin_value >= sval)
8607 else if (reg->smax_value < sval)
8611 if (reg->umax_value <= val)
8613 else if (reg->umin_value > val)
8617 if (reg->smax_value <= sval)
8619 else if (reg->smin_value > sval)
8627 /* compute branch direction of the expression "if (reg opcode val) goto target;"
8629 * 1 - branch will be taken and "goto target" will be executed
8630 * 0 - branch will not be taken and fall-through to next insn
8631 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value
8634 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
8637 if (__is_pointer_value(false, reg)) {
8638 if (!reg_type_not_null(reg->type))
8641 /* If pointer is valid tests against zero will fail so we can
8642 * use this to direct branch taken.
8658 return is_branch32_taken(reg, val, opcode);
8659 return is_branch64_taken(reg, val, opcode);
8662 static int flip_opcode(u32 opcode)
8664 /* How can we transform "a <op> b" into "b <op> a"? */
8665 static const u8 opcode_flip[16] = {
8666 /* these stay the same */
8667 [BPF_JEQ >> 4] = BPF_JEQ,
8668 [BPF_JNE >> 4] = BPF_JNE,
8669 [BPF_JSET >> 4] = BPF_JSET,
8670 /* these swap "lesser" and "greater" (L and G in the opcodes) */
8671 [BPF_JGE >> 4] = BPF_JLE,
8672 [BPF_JGT >> 4] = BPF_JLT,
8673 [BPF_JLE >> 4] = BPF_JGE,
8674 [BPF_JLT >> 4] = BPF_JGT,
8675 [BPF_JSGE >> 4] = BPF_JSLE,
8676 [BPF_JSGT >> 4] = BPF_JSLT,
8677 [BPF_JSLE >> 4] = BPF_JSGE,
8678 [BPF_JSLT >> 4] = BPF_JSGT
8680 return opcode_flip[opcode >> 4];
8683 static int is_pkt_ptr_branch_taken(struct bpf_reg_state *dst_reg,
8684 struct bpf_reg_state *src_reg,
8687 struct bpf_reg_state *pkt;
8689 if (src_reg->type == PTR_TO_PACKET_END) {
8691 } else if (dst_reg->type == PTR_TO_PACKET_END) {
8693 opcode = flip_opcode(opcode);
8698 if (pkt->range >= 0)
8703 /* pkt <= pkt_end */
8707 if (pkt->range == BEYOND_PKT_END)
8708 /* pkt has at last one extra byte beyond pkt_end */
8709 return opcode == BPF_JGT;
8715 /* pkt >= pkt_end */
8716 if (pkt->range == BEYOND_PKT_END || pkt->range == AT_PKT_END)
8717 return opcode == BPF_JGE;
8723 /* Adjusts the register min/max values in the case that the dst_reg is the
8724 * variable register that we are working on, and src_reg is a constant or we're
8725 * simply doing a BPF_K check.
8726 * In JEQ/JNE cases we also adjust the var_off values.
8728 static void reg_set_min_max(struct bpf_reg_state *true_reg,
8729 struct bpf_reg_state *false_reg,
8731 u8 opcode, bool is_jmp32)
8733 struct tnum false_32off = tnum_subreg(false_reg->var_off);
8734 struct tnum false_64off = false_reg->var_off;
8735 struct tnum true_32off = tnum_subreg(true_reg->var_off);
8736 struct tnum true_64off = true_reg->var_off;
8737 s64 sval = (s64)val;
8738 s32 sval32 = (s32)val32;
8740 /* If the dst_reg is a pointer, we can't learn anything about its
8741 * variable offset from the compare (unless src_reg were a pointer into
8742 * the same object, but we don't bother with that.
8743 * Since false_reg and true_reg have the same type by construction, we
8744 * only need to check one of them for pointerness.
8746 if (__is_pointer_value(false, false_reg))
8753 struct bpf_reg_state *reg =
8754 opcode == BPF_JEQ ? true_reg : false_reg;
8756 /* JEQ/JNE comparison doesn't change the register equivalence.
8758 * if (r1 == 42) goto label;
8760 * label: // here both r1 and r2 are known to be 42.
8762 * Hence when marking register as known preserve it's ID.
8765 __mark_reg32_known(reg, val32);
8767 ___mark_reg_known(reg, val);
8772 false_32off = tnum_and(false_32off, tnum_const(~val32));
8773 if (is_power_of_2(val32))
8774 true_32off = tnum_or(true_32off,
8777 false_64off = tnum_and(false_64off, tnum_const(~val));
8778 if (is_power_of_2(val))
8779 true_64off = tnum_or(true_64off,
8787 u32 false_umax = opcode == BPF_JGT ? val32 : val32 - 1;
8788 u32 true_umin = opcode == BPF_JGT ? val32 + 1 : val32;
8790 false_reg->u32_max_value = min(false_reg->u32_max_value,
8792 true_reg->u32_min_value = max(true_reg->u32_min_value,
8795 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
8796 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
8798 false_reg->umax_value = min(false_reg->umax_value, false_umax);
8799 true_reg->umin_value = max(true_reg->umin_value, true_umin);
8807 s32 false_smax = opcode == BPF_JSGT ? sval32 : sval32 - 1;
8808 s32 true_smin = opcode == BPF_JSGT ? sval32 + 1 : sval32;
8810 false_reg->s32_max_value = min(false_reg->s32_max_value, false_smax);
8811 true_reg->s32_min_value = max(true_reg->s32_min_value, true_smin);
8813 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
8814 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
8816 false_reg->smax_value = min(false_reg->smax_value, false_smax);
8817 true_reg->smin_value = max(true_reg->smin_value, true_smin);
8825 u32 false_umin = opcode == BPF_JLT ? val32 : val32 + 1;
8826 u32 true_umax = opcode == BPF_JLT ? val32 - 1 : val32;
8828 false_reg->u32_min_value = max(false_reg->u32_min_value,
8830 true_reg->u32_max_value = min(true_reg->u32_max_value,
8833 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
8834 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
8836 false_reg->umin_value = max(false_reg->umin_value, false_umin);
8837 true_reg->umax_value = min(true_reg->umax_value, true_umax);
8845 s32 false_smin = opcode == BPF_JSLT ? sval32 : sval32 + 1;
8846 s32 true_smax = opcode == BPF_JSLT ? sval32 - 1 : sval32;
8848 false_reg->s32_min_value = max(false_reg->s32_min_value, false_smin);
8849 true_reg->s32_max_value = min(true_reg->s32_max_value, true_smax);
8851 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
8852 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
8854 false_reg->smin_value = max(false_reg->smin_value, false_smin);
8855 true_reg->smax_value = min(true_reg->smax_value, true_smax);
8864 false_reg->var_off = tnum_or(tnum_clear_subreg(false_64off),
8865 tnum_subreg(false_32off));
8866 true_reg->var_off = tnum_or(tnum_clear_subreg(true_64off),
8867 tnum_subreg(true_32off));
8868 __reg_combine_32_into_64(false_reg);
8869 __reg_combine_32_into_64(true_reg);
8871 false_reg->var_off = false_64off;
8872 true_reg->var_off = true_64off;
8873 __reg_combine_64_into_32(false_reg);
8874 __reg_combine_64_into_32(true_reg);
8878 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
8881 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
8882 struct bpf_reg_state *false_reg,
8884 u8 opcode, bool is_jmp32)
8886 opcode = flip_opcode(opcode);
8887 /* This uses zero as "not present in table"; luckily the zero opcode,
8888 * BPF_JA, can't get here.
8891 reg_set_min_max(true_reg, false_reg, val, val32, opcode, is_jmp32);
8894 /* Regs are known to be equal, so intersect their min/max/var_off */
8895 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
8896 struct bpf_reg_state *dst_reg)
8898 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
8899 dst_reg->umin_value);
8900 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
8901 dst_reg->umax_value);
8902 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
8903 dst_reg->smin_value);
8904 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
8905 dst_reg->smax_value);
8906 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
8908 /* We might have learned new bounds from the var_off. */
8909 __update_reg_bounds(src_reg);
8910 __update_reg_bounds(dst_reg);
8911 /* We might have learned something about the sign bit. */
8912 __reg_deduce_bounds(src_reg);
8913 __reg_deduce_bounds(dst_reg);
8914 /* We might have learned some bits from the bounds. */
8915 __reg_bound_offset(src_reg);
8916 __reg_bound_offset(dst_reg);
8917 /* Intersecting with the old var_off might have improved our bounds
8918 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
8919 * then new var_off is (0; 0x7f...fc) which improves our umax.
8921 __update_reg_bounds(src_reg);
8922 __update_reg_bounds(dst_reg);
8925 static void reg_combine_min_max(struct bpf_reg_state *true_src,
8926 struct bpf_reg_state *true_dst,
8927 struct bpf_reg_state *false_src,
8928 struct bpf_reg_state *false_dst,
8933 __reg_combine_min_max(true_src, true_dst);
8936 __reg_combine_min_max(false_src, false_dst);
8941 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
8942 struct bpf_reg_state *reg, u32 id,
8945 if (reg_type_may_be_null(reg->type) && reg->id == id &&
8946 !WARN_ON_ONCE(!reg->id)) {
8947 /* Old offset (both fixed and variable parts) should
8948 * have been known-zero, because we don't allow pointer
8949 * arithmetic on pointers that might be NULL.
8951 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
8952 !tnum_equals_const(reg->var_off, 0) ||
8954 __mark_reg_known_zero(reg);
8958 reg->type = SCALAR_VALUE;
8959 /* We don't need id and ref_obj_id from this point
8960 * onwards anymore, thus we should better reset it,
8961 * so that state pruning has chances to take effect.
8964 reg->ref_obj_id = 0;
8969 mark_ptr_not_null_reg(reg);
8971 if (!reg_may_point_to_spin_lock(reg)) {
8972 /* For not-NULL ptr, reg->ref_obj_id will be reset
8973 * in release_reg_references().
8975 * reg->id is still used by spin_lock ptr. Other
8976 * than spin_lock ptr type, reg->id can be reset.
8983 static void __mark_ptr_or_null_regs(struct bpf_func_state *state, u32 id,
8986 struct bpf_reg_state *reg;
8989 for (i = 0; i < MAX_BPF_REG; i++)
8990 mark_ptr_or_null_reg(state, &state->regs[i], id, is_null);
8992 bpf_for_each_spilled_reg(i, state, reg) {
8995 mark_ptr_or_null_reg(state, reg, id, is_null);
8999 /* The logic is similar to find_good_pkt_pointers(), both could eventually
9000 * be folded together at some point.
9002 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
9005 struct bpf_func_state *state = vstate->frame[vstate->curframe];
9006 struct bpf_reg_state *regs = state->regs;
9007 u32 ref_obj_id = regs[regno].ref_obj_id;
9008 u32 id = regs[regno].id;
9011 if (ref_obj_id && ref_obj_id == id && is_null)
9012 /* regs[regno] is in the " == NULL" branch.
9013 * No one could have freed the reference state before
9014 * doing the NULL check.
9016 WARN_ON_ONCE(release_reference_state(state, id));
9018 for (i = 0; i <= vstate->curframe; i++)
9019 __mark_ptr_or_null_regs(vstate->frame[i], id, is_null);
9022 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
9023 struct bpf_reg_state *dst_reg,
9024 struct bpf_reg_state *src_reg,
9025 struct bpf_verifier_state *this_branch,
9026 struct bpf_verifier_state *other_branch)
9028 if (BPF_SRC(insn->code) != BPF_X)
9031 /* Pointers are always 64-bit. */
9032 if (BPF_CLASS(insn->code) == BPF_JMP32)
9035 switch (BPF_OP(insn->code)) {
9037 if ((dst_reg->type == PTR_TO_PACKET &&
9038 src_reg->type == PTR_TO_PACKET_END) ||
9039 (dst_reg->type == PTR_TO_PACKET_META &&
9040 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
9041 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
9042 find_good_pkt_pointers(this_branch, dst_reg,
9043 dst_reg->type, false);
9044 mark_pkt_end(other_branch, insn->dst_reg, true);
9045 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
9046 src_reg->type == PTR_TO_PACKET) ||
9047 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
9048 src_reg->type == PTR_TO_PACKET_META)) {
9049 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
9050 find_good_pkt_pointers(other_branch, src_reg,
9051 src_reg->type, true);
9052 mark_pkt_end(this_branch, insn->src_reg, false);
9058 if ((dst_reg->type == PTR_TO_PACKET &&
9059 src_reg->type == PTR_TO_PACKET_END) ||
9060 (dst_reg->type == PTR_TO_PACKET_META &&
9061 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
9062 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
9063 find_good_pkt_pointers(other_branch, dst_reg,
9064 dst_reg->type, true);
9065 mark_pkt_end(this_branch, insn->dst_reg, false);
9066 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
9067 src_reg->type == PTR_TO_PACKET) ||
9068 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
9069 src_reg->type == PTR_TO_PACKET_META)) {
9070 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
9071 find_good_pkt_pointers(this_branch, src_reg,
9072 src_reg->type, false);
9073 mark_pkt_end(other_branch, insn->src_reg, true);
9079 if ((dst_reg->type == PTR_TO_PACKET &&
9080 src_reg->type == PTR_TO_PACKET_END) ||
9081 (dst_reg->type == PTR_TO_PACKET_META &&
9082 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
9083 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
9084 find_good_pkt_pointers(this_branch, dst_reg,
9085 dst_reg->type, true);
9086 mark_pkt_end(other_branch, insn->dst_reg, false);
9087 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
9088 src_reg->type == PTR_TO_PACKET) ||
9089 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
9090 src_reg->type == PTR_TO_PACKET_META)) {
9091 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
9092 find_good_pkt_pointers(other_branch, src_reg,
9093 src_reg->type, false);
9094 mark_pkt_end(this_branch, insn->src_reg, true);
9100 if ((dst_reg->type == PTR_TO_PACKET &&
9101 src_reg->type == PTR_TO_PACKET_END) ||
9102 (dst_reg->type == PTR_TO_PACKET_META &&
9103 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
9104 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
9105 find_good_pkt_pointers(other_branch, dst_reg,
9106 dst_reg->type, false);
9107 mark_pkt_end(this_branch, insn->dst_reg, true);
9108 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
9109 src_reg->type == PTR_TO_PACKET) ||
9110 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
9111 src_reg->type == PTR_TO_PACKET_META)) {
9112 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
9113 find_good_pkt_pointers(this_branch, src_reg,
9114 src_reg->type, true);
9115 mark_pkt_end(other_branch, insn->src_reg, false);
9127 static void find_equal_scalars(struct bpf_verifier_state *vstate,
9128 struct bpf_reg_state *known_reg)
9130 struct bpf_func_state *state;
9131 struct bpf_reg_state *reg;
9134 for (i = 0; i <= vstate->curframe; i++) {
9135 state = vstate->frame[i];
9136 for (j = 0; j < MAX_BPF_REG; j++) {
9137 reg = &state->regs[j];
9138 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
9142 bpf_for_each_spilled_reg(j, state, reg) {
9145 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
9151 static int check_cond_jmp_op(struct bpf_verifier_env *env,
9152 struct bpf_insn *insn, int *insn_idx)
9154 struct bpf_verifier_state *this_branch = env->cur_state;
9155 struct bpf_verifier_state *other_branch;
9156 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
9157 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
9158 u8 opcode = BPF_OP(insn->code);
9163 /* Only conditional jumps are expected to reach here. */
9164 if (opcode == BPF_JA || opcode > BPF_JSLE) {
9165 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
9169 if (BPF_SRC(insn->code) == BPF_X) {
9170 if (insn->imm != 0) {
9171 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
9175 /* check src1 operand */
9176 err = check_reg_arg(env, insn->src_reg, SRC_OP);
9180 if (is_pointer_value(env, insn->src_reg)) {
9181 verbose(env, "R%d pointer comparison prohibited\n",
9185 src_reg = ®s[insn->src_reg];
9187 if (insn->src_reg != BPF_REG_0) {
9188 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
9193 /* check src2 operand */
9194 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
9198 dst_reg = ®s[insn->dst_reg];
9199 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
9201 if (BPF_SRC(insn->code) == BPF_K) {
9202 pred = is_branch_taken(dst_reg, insn->imm, opcode, is_jmp32);
9203 } else if (src_reg->type == SCALAR_VALUE &&
9204 is_jmp32 && tnum_is_const(tnum_subreg(src_reg->var_off))) {
9205 pred = is_branch_taken(dst_reg,
9206 tnum_subreg(src_reg->var_off).value,
9209 } else if (src_reg->type == SCALAR_VALUE &&
9210 !is_jmp32 && tnum_is_const(src_reg->var_off)) {
9211 pred = is_branch_taken(dst_reg,
9212 src_reg->var_off.value,
9215 } else if (reg_is_pkt_pointer_any(dst_reg) &&
9216 reg_is_pkt_pointer_any(src_reg) &&
9218 pred = is_pkt_ptr_branch_taken(dst_reg, src_reg, opcode);
9222 /* If we get here with a dst_reg pointer type it is because
9223 * above is_branch_taken() special cased the 0 comparison.
9225 if (!__is_pointer_value(false, dst_reg))
9226 err = mark_chain_precision(env, insn->dst_reg);
9227 if (BPF_SRC(insn->code) == BPF_X && !err &&
9228 !__is_pointer_value(false, src_reg))
9229 err = mark_chain_precision(env, insn->src_reg);
9235 /* Only follow the goto, ignore fall-through. If needed, push
9236 * the fall-through branch for simulation under speculative
9239 if (!env->bypass_spec_v1 &&
9240 !sanitize_speculative_path(env, insn, *insn_idx + 1,
9243 *insn_idx += insn->off;
9245 } else if (pred == 0) {
9246 /* Only follow the fall-through branch, since that's where the
9247 * program will go. If needed, push the goto branch for
9248 * simulation under speculative execution.
9250 if (!env->bypass_spec_v1 &&
9251 !sanitize_speculative_path(env, insn,
9252 *insn_idx + insn->off + 1,
9258 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
9262 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
9264 /* detect if we are comparing against a constant value so we can adjust
9265 * our min/max values for our dst register.
9266 * this is only legit if both are scalars (or pointers to the same
9267 * object, I suppose, but we don't support that right now), because
9268 * otherwise the different base pointers mean the offsets aren't
9271 if (BPF_SRC(insn->code) == BPF_X) {
9272 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
9274 if (dst_reg->type == SCALAR_VALUE &&
9275 src_reg->type == SCALAR_VALUE) {
9276 if (tnum_is_const(src_reg->var_off) ||
9278 tnum_is_const(tnum_subreg(src_reg->var_off))))
9279 reg_set_min_max(&other_branch_regs[insn->dst_reg],
9281 src_reg->var_off.value,
9282 tnum_subreg(src_reg->var_off).value,
9284 else if (tnum_is_const(dst_reg->var_off) ||
9286 tnum_is_const(tnum_subreg(dst_reg->var_off))))
9287 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
9289 dst_reg->var_off.value,
9290 tnum_subreg(dst_reg->var_off).value,
9292 else if (!is_jmp32 &&
9293 (opcode == BPF_JEQ || opcode == BPF_JNE))
9294 /* Comparing for equality, we can combine knowledge */
9295 reg_combine_min_max(&other_branch_regs[insn->src_reg],
9296 &other_branch_regs[insn->dst_reg],
9297 src_reg, dst_reg, opcode);
9299 !WARN_ON_ONCE(src_reg->id != other_branch_regs[insn->src_reg].id)) {
9300 find_equal_scalars(this_branch, src_reg);
9301 find_equal_scalars(other_branch, &other_branch_regs[insn->src_reg]);
9305 } else if (dst_reg->type == SCALAR_VALUE) {
9306 reg_set_min_max(&other_branch_regs[insn->dst_reg],
9307 dst_reg, insn->imm, (u32)insn->imm,
9311 if (dst_reg->type == SCALAR_VALUE && dst_reg->id &&
9312 !WARN_ON_ONCE(dst_reg->id != other_branch_regs[insn->dst_reg].id)) {
9313 find_equal_scalars(this_branch, dst_reg);
9314 find_equal_scalars(other_branch, &other_branch_regs[insn->dst_reg]);
9317 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
9318 * NOTE: these optimizations below are related with pointer comparison
9319 * which will never be JMP32.
9321 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
9322 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
9323 reg_type_may_be_null(dst_reg->type)) {
9324 /* Mark all identical registers in each branch as either
9325 * safe or unknown depending R == 0 or R != 0 conditional.
9327 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
9329 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
9331 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
9332 this_branch, other_branch) &&
9333 is_pointer_value(env, insn->dst_reg)) {
9334 verbose(env, "R%d pointer comparison prohibited\n",
9338 if (env->log.level & BPF_LOG_LEVEL)
9339 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
9343 /* verify BPF_LD_IMM64 instruction */
9344 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
9346 struct bpf_insn_aux_data *aux = cur_aux(env);
9347 struct bpf_reg_state *regs = cur_regs(env);
9348 struct bpf_reg_state *dst_reg;
9349 struct bpf_map *map;
9352 if (BPF_SIZE(insn->code) != BPF_DW) {
9353 verbose(env, "invalid BPF_LD_IMM insn\n");
9356 if (insn->off != 0) {
9357 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
9361 err = check_reg_arg(env, insn->dst_reg, DST_OP);
9365 dst_reg = ®s[insn->dst_reg];
9366 if (insn->src_reg == 0) {
9367 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
9369 dst_reg->type = SCALAR_VALUE;
9370 __mark_reg_known(®s[insn->dst_reg], imm);
9374 if (insn->src_reg == BPF_PSEUDO_BTF_ID) {
9375 mark_reg_known_zero(env, regs, insn->dst_reg);
9377 dst_reg->type = aux->btf_var.reg_type;
9378 switch (dst_reg->type) {
9380 dst_reg->mem_size = aux->btf_var.mem_size;
9383 case PTR_TO_PERCPU_BTF_ID:
9384 dst_reg->btf = aux->btf_var.btf;
9385 dst_reg->btf_id = aux->btf_var.btf_id;
9388 verbose(env, "bpf verifier is misconfigured\n");
9394 if (insn->src_reg == BPF_PSEUDO_FUNC) {
9395 struct bpf_prog_aux *aux = env->prog->aux;
9396 u32 subprogno = find_subprog(env,
9397 env->insn_idx + insn->imm + 1);
9399 if (!aux->func_info) {
9400 verbose(env, "missing btf func_info\n");
9403 if (aux->func_info_aux[subprogno].linkage != BTF_FUNC_STATIC) {
9404 verbose(env, "callback function not static\n");
9408 dst_reg->type = PTR_TO_FUNC;
9409 dst_reg->subprogno = subprogno;
9413 map = env->used_maps[aux->map_index];
9414 mark_reg_known_zero(env, regs, insn->dst_reg);
9415 dst_reg->map_ptr = map;
9417 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE ||
9418 insn->src_reg == BPF_PSEUDO_MAP_IDX_VALUE) {
9419 dst_reg->type = PTR_TO_MAP_VALUE;
9420 dst_reg->off = aux->map_off;
9421 if (map_value_has_spin_lock(map))
9422 dst_reg->id = ++env->id_gen;
9423 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD ||
9424 insn->src_reg == BPF_PSEUDO_MAP_IDX) {
9425 dst_reg->type = CONST_PTR_TO_MAP;
9427 verbose(env, "bpf verifier is misconfigured\n");
9434 static bool may_access_skb(enum bpf_prog_type type)
9437 case BPF_PROG_TYPE_SOCKET_FILTER:
9438 case BPF_PROG_TYPE_SCHED_CLS:
9439 case BPF_PROG_TYPE_SCHED_ACT:
9446 /* verify safety of LD_ABS|LD_IND instructions:
9447 * - they can only appear in the programs where ctx == skb
9448 * - since they are wrappers of function calls, they scratch R1-R5 registers,
9449 * preserve R6-R9, and store return value into R0
9452 * ctx == skb == R6 == CTX
9455 * SRC == any register
9456 * IMM == 32-bit immediate
9459 * R0 - 8/16/32-bit skb data converted to cpu endianness
9461 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
9463 struct bpf_reg_state *regs = cur_regs(env);
9464 static const int ctx_reg = BPF_REG_6;
9465 u8 mode = BPF_MODE(insn->code);
9468 if (!may_access_skb(resolve_prog_type(env->prog))) {
9469 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
9473 if (!env->ops->gen_ld_abs) {
9474 verbose(env, "bpf verifier is misconfigured\n");
9478 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
9479 BPF_SIZE(insn->code) == BPF_DW ||
9480 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
9481 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
9485 /* check whether implicit source operand (register R6) is readable */
9486 err = check_reg_arg(env, ctx_reg, SRC_OP);
9490 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
9491 * gen_ld_abs() may terminate the program at runtime, leading to
9494 err = check_reference_leak(env);
9496 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
9500 if (env->cur_state->active_spin_lock) {
9501 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
9505 if (regs[ctx_reg].type != PTR_TO_CTX) {
9507 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
9511 if (mode == BPF_IND) {
9512 /* check explicit source operand */
9513 err = check_reg_arg(env, insn->src_reg, SRC_OP);
9518 err = check_ctx_reg(env, ®s[ctx_reg], ctx_reg);
9522 /* reset caller saved regs to unreadable */
9523 for (i = 0; i < CALLER_SAVED_REGS; i++) {
9524 mark_reg_not_init(env, regs, caller_saved[i]);
9525 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
9528 /* mark destination R0 register as readable, since it contains
9529 * the value fetched from the packet.
9530 * Already marked as written above.
9532 mark_reg_unknown(env, regs, BPF_REG_0);
9533 /* ld_abs load up to 32-bit skb data. */
9534 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
9538 static int check_return_code(struct bpf_verifier_env *env)
9540 struct tnum enforce_attach_type_range = tnum_unknown;
9541 const struct bpf_prog *prog = env->prog;
9542 struct bpf_reg_state *reg;
9543 struct tnum range = tnum_range(0, 1);
9544 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
9546 struct bpf_func_state *frame = env->cur_state->frame[0];
9547 const bool is_subprog = frame->subprogno;
9549 /* LSM and struct_ops func-ptr's return type could be "void" */
9551 (prog_type == BPF_PROG_TYPE_STRUCT_OPS ||
9552 prog_type == BPF_PROG_TYPE_LSM) &&
9553 !prog->aux->attach_func_proto->type)
9556 /* eBPF calling convention is such that R0 is used
9557 * to return the value from eBPF program.
9558 * Make sure that it's readable at this time
9559 * of bpf_exit, which means that program wrote
9560 * something into it earlier
9562 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
9566 if (is_pointer_value(env, BPF_REG_0)) {
9567 verbose(env, "R0 leaks addr as return value\n");
9571 reg = cur_regs(env) + BPF_REG_0;
9573 if (frame->in_async_callback_fn) {
9574 /* enforce return zero from async callbacks like timer */
9575 if (reg->type != SCALAR_VALUE) {
9576 verbose(env, "In async callback the register R0 is not a known value (%s)\n",
9577 reg_type_str[reg->type]);
9581 if (!tnum_in(tnum_const(0), reg->var_off)) {
9582 verbose_invalid_scalar(env, reg, &range, "async callback", "R0");
9589 if (reg->type != SCALAR_VALUE) {
9590 verbose(env, "At subprogram exit the register R0 is not a scalar value (%s)\n",
9591 reg_type_str[reg->type]);
9597 switch (prog_type) {
9598 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
9599 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
9600 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG ||
9601 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETPEERNAME ||
9602 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETPEERNAME ||
9603 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETSOCKNAME ||
9604 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETSOCKNAME)
9605 range = tnum_range(1, 1);
9606 if (env->prog->expected_attach_type == BPF_CGROUP_INET4_BIND ||
9607 env->prog->expected_attach_type == BPF_CGROUP_INET6_BIND)
9608 range = tnum_range(0, 3);
9610 case BPF_PROG_TYPE_CGROUP_SKB:
9611 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
9612 range = tnum_range(0, 3);
9613 enforce_attach_type_range = tnum_range(2, 3);
9616 case BPF_PROG_TYPE_CGROUP_SOCK:
9617 case BPF_PROG_TYPE_SOCK_OPS:
9618 case BPF_PROG_TYPE_CGROUP_DEVICE:
9619 case BPF_PROG_TYPE_CGROUP_SYSCTL:
9620 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
9622 case BPF_PROG_TYPE_RAW_TRACEPOINT:
9623 if (!env->prog->aux->attach_btf_id)
9625 range = tnum_const(0);
9627 case BPF_PROG_TYPE_TRACING:
9628 switch (env->prog->expected_attach_type) {
9629 case BPF_TRACE_FENTRY:
9630 case BPF_TRACE_FEXIT:
9631 range = tnum_const(0);
9633 case BPF_TRACE_RAW_TP:
9634 case BPF_MODIFY_RETURN:
9636 case BPF_TRACE_ITER:
9642 case BPF_PROG_TYPE_SK_LOOKUP:
9643 range = tnum_range(SK_DROP, SK_PASS);
9645 case BPF_PROG_TYPE_EXT:
9646 /* freplace program can return anything as its return value
9647 * depends on the to-be-replaced kernel func or bpf program.
9653 if (reg->type != SCALAR_VALUE) {
9654 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
9655 reg_type_str[reg->type]);
9659 if (!tnum_in(range, reg->var_off)) {
9660 verbose_invalid_scalar(env, reg, &range, "program exit", "R0");
9664 if (!tnum_is_unknown(enforce_attach_type_range) &&
9665 tnum_in(enforce_attach_type_range, reg->var_off))
9666 env->prog->enforce_expected_attach_type = 1;
9670 /* non-recursive DFS pseudo code
9671 * 1 procedure DFS-iterative(G,v):
9672 * 2 label v as discovered
9673 * 3 let S be a stack
9675 * 5 while S is not empty
9677 * 7 if t is what we're looking for:
9679 * 9 for all edges e in G.adjacentEdges(t) do
9680 * 10 if edge e is already labelled
9681 * 11 continue with the next edge
9682 * 12 w <- G.adjacentVertex(t,e)
9683 * 13 if vertex w is not discovered and not explored
9684 * 14 label e as tree-edge
9685 * 15 label w as discovered
9688 * 18 else if vertex w is discovered
9689 * 19 label e as back-edge
9691 * 21 // vertex w is explored
9692 * 22 label e as forward- or cross-edge
9693 * 23 label t as explored
9698 * 0x11 - discovered and fall-through edge labelled
9699 * 0x12 - discovered and fall-through and branch edges labelled
9710 static u32 state_htab_size(struct bpf_verifier_env *env)
9712 return env->prog->len;
9715 static struct bpf_verifier_state_list **explored_state(
9716 struct bpf_verifier_env *env,
9719 struct bpf_verifier_state *cur = env->cur_state;
9720 struct bpf_func_state *state = cur->frame[cur->curframe];
9722 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
9725 static void init_explored_state(struct bpf_verifier_env *env, int idx)
9727 env->insn_aux_data[idx].prune_point = true;
9735 /* t, w, e - match pseudo-code above:
9736 * t - index of current instruction
9737 * w - next instruction
9740 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env,
9743 int *insn_stack = env->cfg.insn_stack;
9744 int *insn_state = env->cfg.insn_state;
9746 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
9747 return DONE_EXPLORING;
9749 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
9750 return DONE_EXPLORING;
9752 if (w < 0 || w >= env->prog->len) {
9753 verbose_linfo(env, t, "%d: ", t);
9754 verbose(env, "jump out of range from insn %d to %d\n", t, w);
9759 /* mark branch target for state pruning */
9760 init_explored_state(env, w);
9762 if (insn_state[w] == 0) {
9764 insn_state[t] = DISCOVERED | e;
9765 insn_state[w] = DISCOVERED;
9766 if (env->cfg.cur_stack >= env->prog->len)
9768 insn_stack[env->cfg.cur_stack++] = w;
9769 return KEEP_EXPLORING;
9770 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
9771 if (loop_ok && env->bpf_capable)
9772 return DONE_EXPLORING;
9773 verbose_linfo(env, t, "%d: ", t);
9774 verbose_linfo(env, w, "%d: ", w);
9775 verbose(env, "back-edge from insn %d to %d\n", t, w);
9777 } else if (insn_state[w] == EXPLORED) {
9778 /* forward- or cross-edge */
9779 insn_state[t] = DISCOVERED | e;
9781 verbose(env, "insn state internal bug\n");
9784 return DONE_EXPLORING;
9787 static int visit_func_call_insn(int t, int insn_cnt,
9788 struct bpf_insn *insns,
9789 struct bpf_verifier_env *env,
9794 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
9798 if (t + 1 < insn_cnt)
9799 init_explored_state(env, t + 1);
9801 init_explored_state(env, t);
9802 ret = push_insn(t, t + insns[t].imm + 1, BRANCH, env,
9803 /* It's ok to allow recursion from CFG point of
9804 * view. __check_func_call() will do the actual
9807 bpf_pseudo_func(insns + t));
9812 /* Visits the instruction at index t and returns one of the following:
9813 * < 0 - an error occurred
9814 * DONE_EXPLORING - the instruction was fully explored
9815 * KEEP_EXPLORING - there is still work to be done before it is fully explored
9817 static int visit_insn(int t, int insn_cnt, struct bpf_verifier_env *env)
9819 struct bpf_insn *insns = env->prog->insnsi;
9822 if (bpf_pseudo_func(insns + t))
9823 return visit_func_call_insn(t, insn_cnt, insns, env, true);
9825 /* All non-branch instructions have a single fall-through edge. */
9826 if (BPF_CLASS(insns[t].code) != BPF_JMP &&
9827 BPF_CLASS(insns[t].code) != BPF_JMP32)
9828 return push_insn(t, t + 1, FALLTHROUGH, env, false);
9830 switch (BPF_OP(insns[t].code)) {
9832 return DONE_EXPLORING;
9835 if (insns[t].imm == BPF_FUNC_timer_set_callback)
9836 /* Mark this call insn to trigger is_state_visited() check
9837 * before call itself is processed by __check_func_call().
9838 * Otherwise new async state will be pushed for further
9841 init_explored_state(env, t);
9842 return visit_func_call_insn(t, insn_cnt, insns, env,
9843 insns[t].src_reg == BPF_PSEUDO_CALL);
9846 if (BPF_SRC(insns[t].code) != BPF_K)
9849 /* unconditional jump with single edge */
9850 ret = push_insn(t, t + insns[t].off + 1, FALLTHROUGH, env,
9855 /* unconditional jmp is not a good pruning point,
9856 * but it's marked, since backtracking needs
9857 * to record jmp history in is_state_visited().
9859 init_explored_state(env, t + insns[t].off + 1);
9860 /* tell verifier to check for equivalent states
9861 * after every call and jump
9863 if (t + 1 < insn_cnt)
9864 init_explored_state(env, t + 1);
9869 /* conditional jump with two edges */
9870 init_explored_state(env, t);
9871 ret = push_insn(t, t + 1, FALLTHROUGH, env, true);
9875 return push_insn(t, t + insns[t].off + 1, BRANCH, env, true);
9879 /* non-recursive depth-first-search to detect loops in BPF program
9880 * loop == back-edge in directed graph
9882 static int check_cfg(struct bpf_verifier_env *env)
9884 int insn_cnt = env->prog->len;
9885 int *insn_stack, *insn_state;
9889 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
9893 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
9899 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
9900 insn_stack[0] = 0; /* 0 is the first instruction */
9901 env->cfg.cur_stack = 1;
9903 while (env->cfg.cur_stack > 0) {
9904 int t = insn_stack[env->cfg.cur_stack - 1];
9906 ret = visit_insn(t, insn_cnt, env);
9908 case DONE_EXPLORING:
9909 insn_state[t] = EXPLORED;
9910 env->cfg.cur_stack--;
9912 case KEEP_EXPLORING:
9916 verbose(env, "visit_insn internal bug\n");
9923 if (env->cfg.cur_stack < 0) {
9924 verbose(env, "pop stack internal bug\n");
9929 for (i = 0; i < insn_cnt; i++) {
9930 if (insn_state[i] != EXPLORED) {
9931 verbose(env, "unreachable insn %d\n", i);
9936 ret = 0; /* cfg looks good */
9941 env->cfg.insn_state = env->cfg.insn_stack = NULL;
9945 static int check_abnormal_return(struct bpf_verifier_env *env)
9949 for (i = 1; i < env->subprog_cnt; i++) {
9950 if (env->subprog_info[i].has_ld_abs) {
9951 verbose(env, "LD_ABS is not allowed in subprogs without BTF\n");
9954 if (env->subprog_info[i].has_tail_call) {
9955 verbose(env, "tail_call is not allowed in subprogs without BTF\n");
9962 /* The minimum supported BTF func info size */
9963 #define MIN_BPF_FUNCINFO_SIZE 8
9964 #define MAX_FUNCINFO_REC_SIZE 252
9966 static int check_btf_func(struct bpf_verifier_env *env,
9967 const union bpf_attr *attr,
9970 const struct btf_type *type, *func_proto, *ret_type;
9971 u32 i, nfuncs, urec_size, min_size;
9972 u32 krec_size = sizeof(struct bpf_func_info);
9973 struct bpf_func_info *krecord;
9974 struct bpf_func_info_aux *info_aux = NULL;
9975 struct bpf_prog *prog;
9976 const struct btf *btf;
9978 u32 prev_offset = 0;
9982 nfuncs = attr->func_info_cnt;
9984 if (check_abnormal_return(env))
9989 if (nfuncs != env->subprog_cnt) {
9990 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
9994 urec_size = attr->func_info_rec_size;
9995 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
9996 urec_size > MAX_FUNCINFO_REC_SIZE ||
9997 urec_size % sizeof(u32)) {
9998 verbose(env, "invalid func info rec size %u\n", urec_size);
10003 btf = prog->aux->btf;
10005 urecord = make_bpfptr(attr->func_info, uattr.is_kernel);
10006 min_size = min_t(u32, krec_size, urec_size);
10008 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
10011 info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN);
10015 for (i = 0; i < nfuncs; i++) {
10016 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
10018 if (ret == -E2BIG) {
10019 verbose(env, "nonzero tailing record in func info");
10020 /* set the size kernel expects so loader can zero
10021 * out the rest of the record.
10023 if (copy_to_bpfptr_offset(uattr,
10024 offsetof(union bpf_attr, func_info_rec_size),
10025 &min_size, sizeof(min_size)))
10031 if (copy_from_bpfptr(&krecord[i], urecord, min_size)) {
10036 /* check insn_off */
10039 if (krecord[i].insn_off) {
10041 "nonzero insn_off %u for the first func info record",
10042 krecord[i].insn_off);
10045 } else if (krecord[i].insn_off <= prev_offset) {
10047 "same or smaller insn offset (%u) than previous func info record (%u)",
10048 krecord[i].insn_off, prev_offset);
10052 if (env->subprog_info[i].start != krecord[i].insn_off) {
10053 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
10057 /* check type_id */
10058 type = btf_type_by_id(btf, krecord[i].type_id);
10059 if (!type || !btf_type_is_func(type)) {
10060 verbose(env, "invalid type id %d in func info",
10061 krecord[i].type_id);
10064 info_aux[i].linkage = BTF_INFO_VLEN(type->info);
10066 func_proto = btf_type_by_id(btf, type->type);
10067 if (unlikely(!func_proto || !btf_type_is_func_proto(func_proto)))
10068 /* btf_func_check() already verified it during BTF load */
10070 ret_type = btf_type_skip_modifiers(btf, func_proto->type, NULL);
10072 btf_type_is_small_int(ret_type) || btf_type_is_enum(ret_type);
10073 if (i && !scalar_return && env->subprog_info[i].has_ld_abs) {
10074 verbose(env, "LD_ABS is only allowed in functions that return 'int'.\n");
10077 if (i && !scalar_return && env->subprog_info[i].has_tail_call) {
10078 verbose(env, "tail_call is only allowed in functions that return 'int'.\n");
10082 prev_offset = krecord[i].insn_off;
10083 bpfptr_add(&urecord, urec_size);
10086 prog->aux->func_info = krecord;
10087 prog->aux->func_info_cnt = nfuncs;
10088 prog->aux->func_info_aux = info_aux;
10097 static void adjust_btf_func(struct bpf_verifier_env *env)
10099 struct bpf_prog_aux *aux = env->prog->aux;
10102 if (!aux->func_info)
10105 for (i = 0; i < env->subprog_cnt; i++)
10106 aux->func_info[i].insn_off = env->subprog_info[i].start;
10109 #define MIN_BPF_LINEINFO_SIZE (offsetof(struct bpf_line_info, line_col) + \
10110 sizeof(((struct bpf_line_info *)(0))->line_col))
10111 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
10113 static int check_btf_line(struct bpf_verifier_env *env,
10114 const union bpf_attr *attr,
10117 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
10118 struct bpf_subprog_info *sub;
10119 struct bpf_line_info *linfo;
10120 struct bpf_prog *prog;
10121 const struct btf *btf;
10125 nr_linfo = attr->line_info_cnt;
10128 if (nr_linfo > INT_MAX / sizeof(struct bpf_line_info))
10131 rec_size = attr->line_info_rec_size;
10132 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
10133 rec_size > MAX_LINEINFO_REC_SIZE ||
10134 rec_size & (sizeof(u32) - 1))
10137 /* Need to zero it in case the userspace may
10138 * pass in a smaller bpf_line_info object.
10140 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
10141 GFP_KERNEL | __GFP_NOWARN);
10146 btf = prog->aux->btf;
10149 sub = env->subprog_info;
10150 ulinfo = make_bpfptr(attr->line_info, uattr.is_kernel);
10151 expected_size = sizeof(struct bpf_line_info);
10152 ncopy = min_t(u32, expected_size, rec_size);
10153 for (i = 0; i < nr_linfo; i++) {
10154 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
10156 if (err == -E2BIG) {
10157 verbose(env, "nonzero tailing record in line_info");
10158 if (copy_to_bpfptr_offset(uattr,
10159 offsetof(union bpf_attr, line_info_rec_size),
10160 &expected_size, sizeof(expected_size)))
10166 if (copy_from_bpfptr(&linfo[i], ulinfo, ncopy)) {
10172 * Check insn_off to ensure
10173 * 1) strictly increasing AND
10174 * 2) bounded by prog->len
10176 * The linfo[0].insn_off == 0 check logically falls into
10177 * the later "missing bpf_line_info for func..." case
10178 * because the first linfo[0].insn_off must be the
10179 * first sub also and the first sub must have
10180 * subprog_info[0].start == 0.
10182 if ((i && linfo[i].insn_off <= prev_offset) ||
10183 linfo[i].insn_off >= prog->len) {
10184 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
10185 i, linfo[i].insn_off, prev_offset,
10191 if (!prog->insnsi[linfo[i].insn_off].code) {
10193 "Invalid insn code at line_info[%u].insn_off\n",
10199 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
10200 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
10201 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
10206 if (s != env->subprog_cnt) {
10207 if (linfo[i].insn_off == sub[s].start) {
10208 sub[s].linfo_idx = i;
10210 } else if (sub[s].start < linfo[i].insn_off) {
10211 verbose(env, "missing bpf_line_info for func#%u\n", s);
10217 prev_offset = linfo[i].insn_off;
10218 bpfptr_add(&ulinfo, rec_size);
10221 if (s != env->subprog_cnt) {
10222 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
10223 env->subprog_cnt - s, s);
10228 prog->aux->linfo = linfo;
10229 prog->aux->nr_linfo = nr_linfo;
10238 static int check_btf_info(struct bpf_verifier_env *env,
10239 const union bpf_attr *attr,
10245 if (!attr->func_info_cnt && !attr->line_info_cnt) {
10246 if (check_abnormal_return(env))
10251 btf = btf_get_by_fd(attr->prog_btf_fd);
10253 return PTR_ERR(btf);
10254 if (btf_is_kernel(btf)) {
10258 env->prog->aux->btf = btf;
10260 err = check_btf_func(env, attr, uattr);
10264 err = check_btf_line(env, attr, uattr);
10271 /* check %cur's range satisfies %old's */
10272 static bool range_within(struct bpf_reg_state *old,
10273 struct bpf_reg_state *cur)
10275 return old->umin_value <= cur->umin_value &&
10276 old->umax_value >= cur->umax_value &&
10277 old->smin_value <= cur->smin_value &&
10278 old->smax_value >= cur->smax_value &&
10279 old->u32_min_value <= cur->u32_min_value &&
10280 old->u32_max_value >= cur->u32_max_value &&
10281 old->s32_min_value <= cur->s32_min_value &&
10282 old->s32_max_value >= cur->s32_max_value;
10285 /* If in the old state two registers had the same id, then they need to have
10286 * the same id in the new state as well. But that id could be different from
10287 * the old state, so we need to track the mapping from old to new ids.
10288 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
10289 * regs with old id 5 must also have new id 9 for the new state to be safe. But
10290 * regs with a different old id could still have new id 9, we don't care about
10292 * So we look through our idmap to see if this old id has been seen before. If
10293 * so, we require the new id to match; otherwise, we add the id pair to the map.
10295 static bool check_ids(u32 old_id, u32 cur_id, struct bpf_id_pair *idmap)
10299 for (i = 0; i < BPF_ID_MAP_SIZE; i++) {
10300 if (!idmap[i].old) {
10301 /* Reached an empty slot; haven't seen this id before */
10302 idmap[i].old = old_id;
10303 idmap[i].cur = cur_id;
10306 if (idmap[i].old == old_id)
10307 return idmap[i].cur == cur_id;
10309 /* We ran out of idmap slots, which should be impossible */
10314 static void clean_func_state(struct bpf_verifier_env *env,
10315 struct bpf_func_state *st)
10317 enum bpf_reg_liveness live;
10320 for (i = 0; i < BPF_REG_FP; i++) {
10321 live = st->regs[i].live;
10322 /* liveness must not touch this register anymore */
10323 st->regs[i].live |= REG_LIVE_DONE;
10324 if (!(live & REG_LIVE_READ))
10325 /* since the register is unused, clear its state
10326 * to make further comparison simpler
10328 __mark_reg_not_init(env, &st->regs[i]);
10331 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
10332 live = st->stack[i].spilled_ptr.live;
10333 /* liveness must not touch this stack slot anymore */
10334 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
10335 if (!(live & REG_LIVE_READ)) {
10336 __mark_reg_not_init(env, &st->stack[i].spilled_ptr);
10337 for (j = 0; j < BPF_REG_SIZE; j++)
10338 st->stack[i].slot_type[j] = STACK_INVALID;
10343 static void clean_verifier_state(struct bpf_verifier_env *env,
10344 struct bpf_verifier_state *st)
10348 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
10349 /* all regs in this state in all frames were already marked */
10352 for (i = 0; i <= st->curframe; i++)
10353 clean_func_state(env, st->frame[i]);
10356 /* the parentage chains form a tree.
10357 * the verifier states are added to state lists at given insn and
10358 * pushed into state stack for future exploration.
10359 * when the verifier reaches bpf_exit insn some of the verifer states
10360 * stored in the state lists have their final liveness state already,
10361 * but a lot of states will get revised from liveness point of view when
10362 * the verifier explores other branches.
10365 * 2: if r1 == 100 goto pc+1
10368 * when the verifier reaches exit insn the register r0 in the state list of
10369 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
10370 * of insn 2 and goes exploring further. At the insn 4 it will walk the
10371 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
10373 * Since the verifier pushes the branch states as it sees them while exploring
10374 * the program the condition of walking the branch instruction for the second
10375 * time means that all states below this branch were already explored and
10376 * their final liveness marks are already propagated.
10377 * Hence when the verifier completes the search of state list in is_state_visited()
10378 * we can call this clean_live_states() function to mark all liveness states
10379 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
10380 * will not be used.
10381 * This function also clears the registers and stack for states that !READ
10382 * to simplify state merging.
10384 * Important note here that walking the same branch instruction in the callee
10385 * doesn't meant that the states are DONE. The verifier has to compare
10388 static void clean_live_states(struct bpf_verifier_env *env, int insn,
10389 struct bpf_verifier_state *cur)
10391 struct bpf_verifier_state_list *sl;
10394 sl = *explored_state(env, insn);
10396 if (sl->state.branches)
10398 if (sl->state.insn_idx != insn ||
10399 sl->state.curframe != cur->curframe)
10401 for (i = 0; i <= cur->curframe; i++)
10402 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
10404 clean_verifier_state(env, &sl->state);
10410 /* Returns true if (rold safe implies rcur safe) */
10411 static bool regsafe(struct bpf_verifier_env *env, struct bpf_reg_state *rold,
10412 struct bpf_reg_state *rcur, struct bpf_id_pair *idmap)
10416 if (!(rold->live & REG_LIVE_READ))
10417 /* explored state didn't use this */
10420 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
10422 if (rold->type == PTR_TO_STACK)
10423 /* two stack pointers are equal only if they're pointing to
10424 * the same stack frame, since fp-8 in foo != fp-8 in bar
10426 return equal && rold->frameno == rcur->frameno;
10431 if (rold->type == NOT_INIT)
10432 /* explored state can't have used this */
10434 if (rcur->type == NOT_INIT)
10436 switch (rold->type) {
10438 if (env->explore_alu_limits)
10440 if (rcur->type == SCALAR_VALUE) {
10441 if (!rold->precise && !rcur->precise)
10443 /* new val must satisfy old val knowledge */
10444 return range_within(rold, rcur) &&
10445 tnum_in(rold->var_off, rcur->var_off);
10447 /* We're trying to use a pointer in place of a scalar.
10448 * Even if the scalar was unbounded, this could lead to
10449 * pointer leaks because scalars are allowed to leak
10450 * while pointers are not. We could make this safe in
10451 * special cases if root is calling us, but it's
10452 * probably not worth the hassle.
10456 case PTR_TO_MAP_KEY:
10457 case PTR_TO_MAP_VALUE:
10458 /* If the new min/max/var_off satisfy the old ones and
10459 * everything else matches, we are OK.
10460 * 'id' is not compared, since it's only used for maps with
10461 * bpf_spin_lock inside map element and in such cases if
10462 * the rest of the prog is valid for one map element then
10463 * it's valid for all map elements regardless of the key
10464 * used in bpf_map_lookup()
10466 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
10467 range_within(rold, rcur) &&
10468 tnum_in(rold->var_off, rcur->var_off);
10469 case PTR_TO_MAP_VALUE_OR_NULL:
10470 /* a PTR_TO_MAP_VALUE could be safe to use as a
10471 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
10472 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
10473 * checked, doing so could have affected others with the same
10474 * id, and we can't check for that because we lost the id when
10475 * we converted to a PTR_TO_MAP_VALUE.
10477 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
10479 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
10481 /* Check our ids match any regs they're supposed to */
10482 return check_ids(rold->id, rcur->id, idmap);
10483 case PTR_TO_PACKET_META:
10484 case PTR_TO_PACKET:
10485 if (rcur->type != rold->type)
10487 /* We must have at least as much range as the old ptr
10488 * did, so that any accesses which were safe before are
10489 * still safe. This is true even if old range < old off,
10490 * since someone could have accessed through (ptr - k), or
10491 * even done ptr -= k in a register, to get a safe access.
10493 if (rold->range > rcur->range)
10495 /* If the offsets don't match, we can't trust our alignment;
10496 * nor can we be sure that we won't fall out of range.
10498 if (rold->off != rcur->off)
10500 /* id relations must be preserved */
10501 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
10503 /* new val must satisfy old val knowledge */
10504 return range_within(rold, rcur) &&
10505 tnum_in(rold->var_off, rcur->var_off);
10507 case CONST_PTR_TO_MAP:
10508 case PTR_TO_PACKET_END:
10509 case PTR_TO_FLOW_KEYS:
10510 case PTR_TO_SOCKET:
10511 case PTR_TO_SOCKET_OR_NULL:
10512 case PTR_TO_SOCK_COMMON:
10513 case PTR_TO_SOCK_COMMON_OR_NULL:
10514 case PTR_TO_TCP_SOCK:
10515 case PTR_TO_TCP_SOCK_OR_NULL:
10516 case PTR_TO_XDP_SOCK:
10517 /* Only valid matches are exact, which memcmp() above
10518 * would have accepted
10521 /* Don't know what's going on, just say it's not safe */
10525 /* Shouldn't get here; if we do, say it's not safe */
10530 static bool stacksafe(struct bpf_verifier_env *env, struct bpf_func_state *old,
10531 struct bpf_func_state *cur, struct bpf_id_pair *idmap)
10535 /* walk slots of the explored stack and ignore any additional
10536 * slots in the current stack, since explored(safe) state
10539 for (i = 0; i < old->allocated_stack; i++) {
10540 spi = i / BPF_REG_SIZE;
10542 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
10543 i += BPF_REG_SIZE - 1;
10544 /* explored state didn't use this */
10548 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
10551 /* explored stack has more populated slots than current stack
10552 * and these slots were used
10554 if (i >= cur->allocated_stack)
10557 /* if old state was safe with misc data in the stack
10558 * it will be safe with zero-initialized stack.
10559 * The opposite is not true
10561 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
10562 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
10564 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
10565 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
10566 /* Ex: old explored (safe) state has STACK_SPILL in
10567 * this stack slot, but current has STACK_MISC ->
10568 * this verifier states are not equivalent,
10569 * return false to continue verification of this path
10572 if (i % BPF_REG_SIZE != BPF_REG_SIZE - 1)
10574 if (!is_spilled_reg(&old->stack[spi]))
10576 if (!regsafe(env, &old->stack[spi].spilled_ptr,
10577 &cur->stack[spi].spilled_ptr, idmap))
10578 /* when explored and current stack slot are both storing
10579 * spilled registers, check that stored pointers types
10580 * are the same as well.
10581 * Ex: explored safe path could have stored
10582 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
10583 * but current path has stored:
10584 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
10585 * such verifier states are not equivalent.
10586 * return false to continue verification of this path
10593 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
10595 if (old->acquired_refs != cur->acquired_refs)
10597 return !memcmp(old->refs, cur->refs,
10598 sizeof(*old->refs) * old->acquired_refs);
10601 /* compare two verifier states
10603 * all states stored in state_list are known to be valid, since
10604 * verifier reached 'bpf_exit' instruction through them
10606 * this function is called when verifier exploring different branches of
10607 * execution popped from the state stack. If it sees an old state that has
10608 * more strict register state and more strict stack state then this execution
10609 * branch doesn't need to be explored further, since verifier already
10610 * concluded that more strict state leads to valid finish.
10612 * Therefore two states are equivalent if register state is more conservative
10613 * and explored stack state is more conservative than the current one.
10616 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
10617 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
10619 * In other words if current stack state (one being explored) has more
10620 * valid slots than old one that already passed validation, it means
10621 * the verifier can stop exploring and conclude that current state is valid too
10623 * Similarly with registers. If explored state has register type as invalid
10624 * whereas register type in current state is meaningful, it means that
10625 * the current state will reach 'bpf_exit' instruction safely
10627 static bool func_states_equal(struct bpf_verifier_env *env, struct bpf_func_state *old,
10628 struct bpf_func_state *cur)
10632 memset(env->idmap_scratch, 0, sizeof(env->idmap_scratch));
10633 for (i = 0; i < MAX_BPF_REG; i++)
10634 if (!regsafe(env, &old->regs[i], &cur->regs[i],
10635 env->idmap_scratch))
10638 if (!stacksafe(env, old, cur, env->idmap_scratch))
10641 if (!refsafe(old, cur))
10647 static bool states_equal(struct bpf_verifier_env *env,
10648 struct bpf_verifier_state *old,
10649 struct bpf_verifier_state *cur)
10653 if (old->curframe != cur->curframe)
10656 /* Verification state from speculative execution simulation
10657 * must never prune a non-speculative execution one.
10659 if (old->speculative && !cur->speculative)
10662 if (old->active_spin_lock != cur->active_spin_lock)
10665 /* for states to be equal callsites have to be the same
10666 * and all frame states need to be equivalent
10668 for (i = 0; i <= old->curframe; i++) {
10669 if (old->frame[i]->callsite != cur->frame[i]->callsite)
10671 if (!func_states_equal(env, old->frame[i], cur->frame[i]))
10677 /* Return 0 if no propagation happened. Return negative error code if error
10678 * happened. Otherwise, return the propagated bit.
10680 static int propagate_liveness_reg(struct bpf_verifier_env *env,
10681 struct bpf_reg_state *reg,
10682 struct bpf_reg_state *parent_reg)
10684 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
10685 u8 flag = reg->live & REG_LIVE_READ;
10688 /* When comes here, read flags of PARENT_REG or REG could be any of
10689 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
10690 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
10692 if (parent_flag == REG_LIVE_READ64 ||
10693 /* Or if there is no read flag from REG. */
10695 /* Or if the read flag from REG is the same as PARENT_REG. */
10696 parent_flag == flag)
10699 err = mark_reg_read(env, reg, parent_reg, flag);
10706 /* A write screens off any subsequent reads; but write marks come from the
10707 * straight-line code between a state and its parent. When we arrive at an
10708 * equivalent state (jump target or such) we didn't arrive by the straight-line
10709 * code, so read marks in the state must propagate to the parent regardless
10710 * of the state's write marks. That's what 'parent == state->parent' comparison
10711 * in mark_reg_read() is for.
10713 static int propagate_liveness(struct bpf_verifier_env *env,
10714 const struct bpf_verifier_state *vstate,
10715 struct bpf_verifier_state *vparent)
10717 struct bpf_reg_state *state_reg, *parent_reg;
10718 struct bpf_func_state *state, *parent;
10719 int i, frame, err = 0;
10721 if (vparent->curframe != vstate->curframe) {
10722 WARN(1, "propagate_live: parent frame %d current frame %d\n",
10723 vparent->curframe, vstate->curframe);
10726 /* Propagate read liveness of registers... */
10727 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
10728 for (frame = 0; frame <= vstate->curframe; frame++) {
10729 parent = vparent->frame[frame];
10730 state = vstate->frame[frame];
10731 parent_reg = parent->regs;
10732 state_reg = state->regs;
10733 /* We don't need to worry about FP liveness, it's read-only */
10734 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
10735 err = propagate_liveness_reg(env, &state_reg[i],
10739 if (err == REG_LIVE_READ64)
10740 mark_insn_zext(env, &parent_reg[i]);
10743 /* Propagate stack slots. */
10744 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
10745 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
10746 parent_reg = &parent->stack[i].spilled_ptr;
10747 state_reg = &state->stack[i].spilled_ptr;
10748 err = propagate_liveness_reg(env, state_reg,
10757 /* find precise scalars in the previous equivalent state and
10758 * propagate them into the current state
10760 static int propagate_precision(struct bpf_verifier_env *env,
10761 const struct bpf_verifier_state *old)
10763 struct bpf_reg_state *state_reg;
10764 struct bpf_func_state *state;
10767 state = old->frame[old->curframe];
10768 state_reg = state->regs;
10769 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
10770 if (state_reg->type != SCALAR_VALUE ||
10771 !state_reg->precise)
10773 if (env->log.level & BPF_LOG_LEVEL2)
10774 verbose(env, "propagating r%d\n", i);
10775 err = mark_chain_precision(env, i);
10780 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
10781 if (!is_spilled_reg(&state->stack[i]))
10783 state_reg = &state->stack[i].spilled_ptr;
10784 if (state_reg->type != SCALAR_VALUE ||
10785 !state_reg->precise)
10787 if (env->log.level & BPF_LOG_LEVEL2)
10788 verbose(env, "propagating fp%d\n",
10789 (-i - 1) * BPF_REG_SIZE);
10790 err = mark_chain_precision_stack(env, i);
10797 static bool states_maybe_looping(struct bpf_verifier_state *old,
10798 struct bpf_verifier_state *cur)
10800 struct bpf_func_state *fold, *fcur;
10801 int i, fr = cur->curframe;
10803 if (old->curframe != fr)
10806 fold = old->frame[fr];
10807 fcur = cur->frame[fr];
10808 for (i = 0; i < MAX_BPF_REG; i++)
10809 if (memcmp(&fold->regs[i], &fcur->regs[i],
10810 offsetof(struct bpf_reg_state, parent)))
10816 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
10818 struct bpf_verifier_state_list *new_sl;
10819 struct bpf_verifier_state_list *sl, **pprev;
10820 struct bpf_verifier_state *cur = env->cur_state, *new;
10821 int i, j, err, states_cnt = 0;
10822 bool add_new_state = env->test_state_freq ? true : false;
10824 cur->last_insn_idx = env->prev_insn_idx;
10825 if (!env->insn_aux_data[insn_idx].prune_point)
10826 /* this 'insn_idx' instruction wasn't marked, so we will not
10827 * be doing state search here
10831 /* bpf progs typically have pruning point every 4 instructions
10832 * http://vger.kernel.org/bpfconf2019.html#session-1
10833 * Do not add new state for future pruning if the verifier hasn't seen
10834 * at least 2 jumps and at least 8 instructions.
10835 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
10836 * In tests that amounts to up to 50% reduction into total verifier
10837 * memory consumption and 20% verifier time speedup.
10839 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
10840 env->insn_processed - env->prev_insn_processed >= 8)
10841 add_new_state = true;
10843 pprev = explored_state(env, insn_idx);
10846 clean_live_states(env, insn_idx, cur);
10850 if (sl->state.insn_idx != insn_idx)
10853 if (sl->state.branches) {
10854 struct bpf_func_state *frame = sl->state.frame[sl->state.curframe];
10856 if (frame->in_async_callback_fn &&
10857 frame->async_entry_cnt != cur->frame[cur->curframe]->async_entry_cnt) {
10858 /* Different async_entry_cnt means that the verifier is
10859 * processing another entry into async callback.
10860 * Seeing the same state is not an indication of infinite
10861 * loop or infinite recursion.
10862 * But finding the same state doesn't mean that it's safe
10863 * to stop processing the current state. The previous state
10864 * hasn't yet reached bpf_exit, since state.branches > 0.
10865 * Checking in_async_callback_fn alone is not enough either.
10866 * Since the verifier still needs to catch infinite loops
10867 * inside async callbacks.
10869 } else if (states_maybe_looping(&sl->state, cur) &&
10870 states_equal(env, &sl->state, cur)) {
10871 verbose_linfo(env, insn_idx, "; ");
10872 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
10875 /* if the verifier is processing a loop, avoid adding new state
10876 * too often, since different loop iterations have distinct
10877 * states and may not help future pruning.
10878 * This threshold shouldn't be too low to make sure that
10879 * a loop with large bound will be rejected quickly.
10880 * The most abusive loop will be:
10882 * if r1 < 1000000 goto pc-2
10883 * 1M insn_procssed limit / 100 == 10k peak states.
10884 * This threshold shouldn't be too high either, since states
10885 * at the end of the loop are likely to be useful in pruning.
10887 if (env->jmps_processed - env->prev_jmps_processed < 20 &&
10888 env->insn_processed - env->prev_insn_processed < 100)
10889 add_new_state = false;
10892 if (states_equal(env, &sl->state, cur)) {
10894 /* reached equivalent register/stack state,
10895 * prune the search.
10896 * Registers read by the continuation are read by us.
10897 * If we have any write marks in env->cur_state, they
10898 * will prevent corresponding reads in the continuation
10899 * from reaching our parent (an explored_state). Our
10900 * own state will get the read marks recorded, but
10901 * they'll be immediately forgotten as we're pruning
10902 * this state and will pop a new one.
10904 err = propagate_liveness(env, &sl->state, cur);
10906 /* if previous state reached the exit with precision and
10907 * current state is equivalent to it (except precsion marks)
10908 * the precision needs to be propagated back in
10909 * the current state.
10911 err = err ? : push_jmp_history(env, cur);
10912 err = err ? : propagate_precision(env, &sl->state);
10918 /* when new state is not going to be added do not increase miss count.
10919 * Otherwise several loop iterations will remove the state
10920 * recorded earlier. The goal of these heuristics is to have
10921 * states from some iterations of the loop (some in the beginning
10922 * and some at the end) to help pruning.
10926 /* heuristic to determine whether this state is beneficial
10927 * to keep checking from state equivalence point of view.
10928 * Higher numbers increase max_states_per_insn and verification time,
10929 * but do not meaningfully decrease insn_processed.
10931 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
10932 /* the state is unlikely to be useful. Remove it to
10933 * speed up verification
10936 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
10937 u32 br = sl->state.branches;
10940 "BUG live_done but branches_to_explore %d\n",
10942 free_verifier_state(&sl->state, false);
10944 env->peak_states--;
10946 /* cannot free this state, since parentage chain may
10947 * walk it later. Add it for free_list instead to
10948 * be freed at the end of verification
10950 sl->next = env->free_list;
10951 env->free_list = sl;
10961 if (env->max_states_per_insn < states_cnt)
10962 env->max_states_per_insn = states_cnt;
10964 if (!env->bpf_capable && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
10965 return push_jmp_history(env, cur);
10967 if (!add_new_state)
10968 return push_jmp_history(env, cur);
10970 /* There were no equivalent states, remember the current one.
10971 * Technically the current state is not proven to be safe yet,
10972 * but it will either reach outer most bpf_exit (which means it's safe)
10973 * or it will be rejected. When there are no loops the verifier won't be
10974 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
10975 * again on the way to bpf_exit.
10976 * When looping the sl->state.branches will be > 0 and this state
10977 * will not be considered for equivalence until branches == 0.
10979 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
10982 env->total_states++;
10983 env->peak_states++;
10984 env->prev_jmps_processed = env->jmps_processed;
10985 env->prev_insn_processed = env->insn_processed;
10987 /* add new state to the head of linked list */
10988 new = &new_sl->state;
10989 err = copy_verifier_state(new, cur);
10991 free_verifier_state(new, false);
10995 new->insn_idx = insn_idx;
10996 WARN_ONCE(new->branches != 1,
10997 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
11000 cur->first_insn_idx = insn_idx;
11001 clear_jmp_history(cur);
11002 new_sl->next = *explored_state(env, insn_idx);
11003 *explored_state(env, insn_idx) = new_sl;
11004 /* connect new state to parentage chain. Current frame needs all
11005 * registers connected. Only r6 - r9 of the callers are alive (pushed
11006 * to the stack implicitly by JITs) so in callers' frames connect just
11007 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
11008 * the state of the call instruction (with WRITTEN set), and r0 comes
11009 * from callee with its full parentage chain, anyway.
11011 /* clear write marks in current state: the writes we did are not writes
11012 * our child did, so they don't screen off its reads from us.
11013 * (There are no read marks in current state, because reads always mark
11014 * their parent and current state never has children yet. Only
11015 * explored_states can get read marks.)
11017 for (j = 0; j <= cur->curframe; j++) {
11018 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
11019 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
11020 for (i = 0; i < BPF_REG_FP; i++)
11021 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
11024 /* all stack frames are accessible from callee, clear them all */
11025 for (j = 0; j <= cur->curframe; j++) {
11026 struct bpf_func_state *frame = cur->frame[j];
11027 struct bpf_func_state *newframe = new->frame[j];
11029 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
11030 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
11031 frame->stack[i].spilled_ptr.parent =
11032 &newframe->stack[i].spilled_ptr;
11038 /* Return true if it's OK to have the same insn return a different type. */
11039 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
11043 case PTR_TO_SOCKET:
11044 case PTR_TO_SOCKET_OR_NULL:
11045 case PTR_TO_SOCK_COMMON:
11046 case PTR_TO_SOCK_COMMON_OR_NULL:
11047 case PTR_TO_TCP_SOCK:
11048 case PTR_TO_TCP_SOCK_OR_NULL:
11049 case PTR_TO_XDP_SOCK:
11050 case PTR_TO_BTF_ID:
11051 case PTR_TO_BTF_ID_OR_NULL:
11058 /* If an instruction was previously used with particular pointer types, then we
11059 * need to be careful to avoid cases such as the below, where it may be ok
11060 * for one branch accessing the pointer, but not ok for the other branch:
11065 * R1 = some_other_valid_ptr;
11068 * R2 = *(u32 *)(R1 + 0);
11070 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
11072 return src != prev && (!reg_type_mismatch_ok(src) ||
11073 !reg_type_mismatch_ok(prev));
11076 static int do_check(struct bpf_verifier_env *env)
11078 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
11079 struct bpf_verifier_state *state = env->cur_state;
11080 struct bpf_insn *insns = env->prog->insnsi;
11081 struct bpf_reg_state *regs;
11082 int insn_cnt = env->prog->len;
11083 bool do_print_state = false;
11084 int prev_insn_idx = -1;
11087 struct bpf_insn *insn;
11091 env->prev_insn_idx = prev_insn_idx;
11092 if (env->insn_idx >= insn_cnt) {
11093 verbose(env, "invalid insn idx %d insn_cnt %d\n",
11094 env->insn_idx, insn_cnt);
11098 insn = &insns[env->insn_idx];
11099 class = BPF_CLASS(insn->code);
11101 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
11103 "BPF program is too large. Processed %d insn\n",
11104 env->insn_processed);
11108 err = is_state_visited(env, env->insn_idx);
11112 /* found equivalent state, can prune the search */
11113 if (env->log.level & BPF_LOG_LEVEL) {
11114 if (do_print_state)
11115 verbose(env, "\nfrom %d to %d%s: safe\n",
11116 env->prev_insn_idx, env->insn_idx,
11117 env->cur_state->speculative ?
11118 " (speculative execution)" : "");
11120 verbose(env, "%d: safe\n", env->insn_idx);
11122 goto process_bpf_exit;
11125 if (signal_pending(current))
11128 if (need_resched())
11131 if (env->log.level & BPF_LOG_LEVEL2 ||
11132 (env->log.level & BPF_LOG_LEVEL && do_print_state)) {
11133 if (env->log.level & BPF_LOG_LEVEL2)
11134 verbose(env, "%d:", env->insn_idx);
11136 verbose(env, "\nfrom %d to %d%s:",
11137 env->prev_insn_idx, env->insn_idx,
11138 env->cur_state->speculative ?
11139 " (speculative execution)" : "");
11140 print_verifier_state(env, state->frame[state->curframe]);
11141 do_print_state = false;
11144 if (env->log.level & BPF_LOG_LEVEL) {
11145 const struct bpf_insn_cbs cbs = {
11146 .cb_call = disasm_kfunc_name,
11147 .cb_print = verbose,
11148 .private_data = env,
11151 verbose_linfo(env, env->insn_idx, "; ");
11152 verbose(env, "%d: ", env->insn_idx);
11153 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
11156 if (bpf_prog_is_dev_bound(env->prog->aux)) {
11157 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
11158 env->prev_insn_idx);
11163 regs = cur_regs(env);
11164 sanitize_mark_insn_seen(env);
11165 prev_insn_idx = env->insn_idx;
11167 if (class == BPF_ALU || class == BPF_ALU64) {
11168 err = check_alu_op(env, insn);
11172 } else if (class == BPF_LDX) {
11173 enum bpf_reg_type *prev_src_type, src_reg_type;
11175 /* check for reserved fields is already done */
11177 /* check src operand */
11178 err = check_reg_arg(env, insn->src_reg, SRC_OP);
11182 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
11186 src_reg_type = regs[insn->src_reg].type;
11188 /* check that memory (src_reg + off) is readable,
11189 * the state of dst_reg will be updated by this func
11191 err = check_mem_access(env, env->insn_idx, insn->src_reg,
11192 insn->off, BPF_SIZE(insn->code),
11193 BPF_READ, insn->dst_reg, false);
11197 prev_src_type = &env->insn_aux_data[env->insn_idx].ptr_type;
11199 if (*prev_src_type == NOT_INIT) {
11200 /* saw a valid insn
11201 * dst_reg = *(u32 *)(src_reg + off)
11202 * save type to validate intersecting paths
11204 *prev_src_type = src_reg_type;
11206 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
11207 /* ABuser program is trying to use the same insn
11208 * dst_reg = *(u32*) (src_reg + off)
11209 * with different pointer types:
11210 * src_reg == ctx in one branch and
11211 * src_reg == stack|map in some other branch.
11214 verbose(env, "same insn cannot be used with different pointers\n");
11218 } else if (class == BPF_STX) {
11219 enum bpf_reg_type *prev_dst_type, dst_reg_type;
11221 if (BPF_MODE(insn->code) == BPF_ATOMIC) {
11222 err = check_atomic(env, env->insn_idx, insn);
11229 if (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0) {
11230 verbose(env, "BPF_STX uses reserved fields\n");
11234 /* check src1 operand */
11235 err = check_reg_arg(env, insn->src_reg, SRC_OP);
11238 /* check src2 operand */
11239 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
11243 dst_reg_type = regs[insn->dst_reg].type;
11245 /* check that memory (dst_reg + off) is writeable */
11246 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
11247 insn->off, BPF_SIZE(insn->code),
11248 BPF_WRITE, insn->src_reg, false);
11252 prev_dst_type = &env->insn_aux_data[env->insn_idx].ptr_type;
11254 if (*prev_dst_type == NOT_INIT) {
11255 *prev_dst_type = dst_reg_type;
11256 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
11257 verbose(env, "same insn cannot be used with different pointers\n");
11261 } else if (class == BPF_ST) {
11262 if (BPF_MODE(insn->code) != BPF_MEM ||
11263 insn->src_reg != BPF_REG_0) {
11264 verbose(env, "BPF_ST uses reserved fields\n");
11267 /* check src operand */
11268 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
11272 if (is_ctx_reg(env, insn->dst_reg)) {
11273 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
11275 reg_type_str[reg_state(env, insn->dst_reg)->type]);
11279 /* check that memory (dst_reg + off) is writeable */
11280 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
11281 insn->off, BPF_SIZE(insn->code),
11282 BPF_WRITE, -1, false);
11286 } else if (class == BPF_JMP || class == BPF_JMP32) {
11287 u8 opcode = BPF_OP(insn->code);
11289 env->jmps_processed++;
11290 if (opcode == BPF_CALL) {
11291 if (BPF_SRC(insn->code) != BPF_K ||
11292 (insn->src_reg != BPF_PSEUDO_KFUNC_CALL
11293 && insn->off != 0) ||
11294 (insn->src_reg != BPF_REG_0 &&
11295 insn->src_reg != BPF_PSEUDO_CALL &&
11296 insn->src_reg != BPF_PSEUDO_KFUNC_CALL) ||
11297 insn->dst_reg != BPF_REG_0 ||
11298 class == BPF_JMP32) {
11299 verbose(env, "BPF_CALL uses reserved fields\n");
11303 if (env->cur_state->active_spin_lock &&
11304 (insn->src_reg == BPF_PSEUDO_CALL ||
11305 insn->imm != BPF_FUNC_spin_unlock)) {
11306 verbose(env, "function calls are not allowed while holding a lock\n");
11309 if (insn->src_reg == BPF_PSEUDO_CALL)
11310 err = check_func_call(env, insn, &env->insn_idx);
11311 else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL)
11312 err = check_kfunc_call(env, insn);
11314 err = check_helper_call(env, insn, &env->insn_idx);
11317 } else if (opcode == BPF_JA) {
11318 if (BPF_SRC(insn->code) != BPF_K ||
11320 insn->src_reg != BPF_REG_0 ||
11321 insn->dst_reg != BPF_REG_0 ||
11322 class == BPF_JMP32) {
11323 verbose(env, "BPF_JA uses reserved fields\n");
11327 env->insn_idx += insn->off + 1;
11330 } else if (opcode == BPF_EXIT) {
11331 if (BPF_SRC(insn->code) != BPF_K ||
11333 insn->src_reg != BPF_REG_0 ||
11334 insn->dst_reg != BPF_REG_0 ||
11335 class == BPF_JMP32) {
11336 verbose(env, "BPF_EXIT uses reserved fields\n");
11340 if (env->cur_state->active_spin_lock) {
11341 verbose(env, "bpf_spin_unlock is missing\n");
11345 if (state->curframe) {
11346 /* exit from nested function */
11347 err = prepare_func_exit(env, &env->insn_idx);
11350 do_print_state = true;
11354 err = check_reference_leak(env);
11358 err = check_return_code(env);
11362 update_branch_counts(env, env->cur_state);
11363 err = pop_stack(env, &prev_insn_idx,
11364 &env->insn_idx, pop_log);
11366 if (err != -ENOENT)
11370 do_print_state = true;
11374 err = check_cond_jmp_op(env, insn, &env->insn_idx);
11378 } else if (class == BPF_LD) {
11379 u8 mode = BPF_MODE(insn->code);
11381 if (mode == BPF_ABS || mode == BPF_IND) {
11382 err = check_ld_abs(env, insn);
11386 } else if (mode == BPF_IMM) {
11387 err = check_ld_imm(env, insn);
11392 sanitize_mark_insn_seen(env);
11394 verbose(env, "invalid BPF_LD mode\n");
11398 verbose(env, "unknown insn class %d\n", class);
11408 static int find_btf_percpu_datasec(struct btf *btf)
11410 const struct btf_type *t;
11415 * Both vmlinux and module each have their own ".data..percpu"
11416 * DATASECs in BTF. So for module's case, we need to skip vmlinux BTF
11417 * types to look at only module's own BTF types.
11419 n = btf_nr_types(btf);
11420 if (btf_is_module(btf))
11421 i = btf_nr_types(btf_vmlinux);
11425 for(; i < n; i++) {
11426 t = btf_type_by_id(btf, i);
11427 if (BTF_INFO_KIND(t->info) != BTF_KIND_DATASEC)
11430 tname = btf_name_by_offset(btf, t->name_off);
11431 if (!strcmp(tname, ".data..percpu"))
11438 /* replace pseudo btf_id with kernel symbol address */
11439 static int check_pseudo_btf_id(struct bpf_verifier_env *env,
11440 struct bpf_insn *insn,
11441 struct bpf_insn_aux_data *aux)
11443 const struct btf_var_secinfo *vsi;
11444 const struct btf_type *datasec;
11445 struct btf_mod_pair *btf_mod;
11446 const struct btf_type *t;
11447 const char *sym_name;
11448 bool percpu = false;
11449 u32 type, id = insn->imm;
11453 int i, btf_fd, err;
11455 btf_fd = insn[1].imm;
11457 btf = btf_get_by_fd(btf_fd);
11459 verbose(env, "invalid module BTF object FD specified.\n");
11463 if (!btf_vmlinux) {
11464 verbose(env, "kernel is missing BTF, make sure CONFIG_DEBUG_INFO_BTF=y is specified in Kconfig.\n");
11471 t = btf_type_by_id(btf, id);
11473 verbose(env, "ldimm64 insn specifies invalid btf_id %d.\n", id);
11478 if (!btf_type_is_var(t)) {
11479 verbose(env, "pseudo btf_id %d in ldimm64 isn't KIND_VAR.\n", id);
11484 sym_name = btf_name_by_offset(btf, t->name_off);
11485 addr = kallsyms_lookup_name(sym_name);
11487 verbose(env, "ldimm64 failed to find the address for kernel symbol '%s'.\n",
11493 datasec_id = find_btf_percpu_datasec(btf);
11494 if (datasec_id > 0) {
11495 datasec = btf_type_by_id(btf, datasec_id);
11496 for_each_vsi(i, datasec, vsi) {
11497 if (vsi->type == id) {
11504 insn[0].imm = (u32)addr;
11505 insn[1].imm = addr >> 32;
11508 t = btf_type_skip_modifiers(btf, type, NULL);
11510 aux->btf_var.reg_type = PTR_TO_PERCPU_BTF_ID;
11511 aux->btf_var.btf = btf;
11512 aux->btf_var.btf_id = type;
11513 } else if (!btf_type_is_struct(t)) {
11514 const struct btf_type *ret;
11518 /* resolve the type size of ksym. */
11519 ret = btf_resolve_size(btf, t, &tsize);
11521 tname = btf_name_by_offset(btf, t->name_off);
11522 verbose(env, "ldimm64 unable to resolve the size of type '%s': %ld\n",
11523 tname, PTR_ERR(ret));
11527 aux->btf_var.reg_type = PTR_TO_MEM;
11528 aux->btf_var.mem_size = tsize;
11530 aux->btf_var.reg_type = PTR_TO_BTF_ID;
11531 aux->btf_var.btf = btf;
11532 aux->btf_var.btf_id = type;
11535 /* check whether we recorded this BTF (and maybe module) already */
11536 for (i = 0; i < env->used_btf_cnt; i++) {
11537 if (env->used_btfs[i].btf == btf) {
11543 if (env->used_btf_cnt >= MAX_USED_BTFS) {
11548 btf_mod = &env->used_btfs[env->used_btf_cnt];
11549 btf_mod->btf = btf;
11550 btf_mod->module = NULL;
11552 /* if we reference variables from kernel module, bump its refcount */
11553 if (btf_is_module(btf)) {
11554 btf_mod->module = btf_try_get_module(btf);
11555 if (!btf_mod->module) {
11561 env->used_btf_cnt++;
11569 static int check_map_prealloc(struct bpf_map *map)
11571 return (map->map_type != BPF_MAP_TYPE_HASH &&
11572 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
11573 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
11574 !(map->map_flags & BPF_F_NO_PREALLOC);
11577 static bool is_tracing_prog_type(enum bpf_prog_type type)
11580 case BPF_PROG_TYPE_KPROBE:
11581 case BPF_PROG_TYPE_TRACEPOINT:
11582 case BPF_PROG_TYPE_PERF_EVENT:
11583 case BPF_PROG_TYPE_RAW_TRACEPOINT:
11590 static bool is_preallocated_map(struct bpf_map *map)
11592 if (!check_map_prealloc(map))
11594 if (map->inner_map_meta && !check_map_prealloc(map->inner_map_meta))
11599 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
11600 struct bpf_map *map,
11601 struct bpf_prog *prog)
11604 enum bpf_prog_type prog_type = resolve_prog_type(prog);
11606 * Validate that trace type programs use preallocated hash maps.
11608 * For programs attached to PERF events this is mandatory as the
11609 * perf NMI can hit any arbitrary code sequence.
11611 * All other trace types using preallocated hash maps are unsafe as
11612 * well because tracepoint or kprobes can be inside locked regions
11613 * of the memory allocator or at a place where a recursion into the
11614 * memory allocator would see inconsistent state.
11616 * On RT enabled kernels run-time allocation of all trace type
11617 * programs is strictly prohibited due to lock type constraints. On
11618 * !RT kernels it is allowed for backwards compatibility reasons for
11619 * now, but warnings are emitted so developers are made aware of
11620 * the unsafety and can fix their programs before this is enforced.
11622 if (is_tracing_prog_type(prog_type) && !is_preallocated_map(map)) {
11623 if (prog_type == BPF_PROG_TYPE_PERF_EVENT) {
11624 verbose(env, "perf_event programs can only use preallocated hash map\n");
11627 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
11628 verbose(env, "trace type programs can only use preallocated hash map\n");
11631 WARN_ONCE(1, "trace type BPF program uses run-time allocation\n");
11632 verbose(env, "trace type programs with run-time allocated hash maps are unsafe. Switch to preallocated hash maps.\n");
11635 if (map_value_has_spin_lock(map)) {
11636 if (prog_type == BPF_PROG_TYPE_SOCKET_FILTER) {
11637 verbose(env, "socket filter progs cannot use bpf_spin_lock yet\n");
11641 if (is_tracing_prog_type(prog_type)) {
11642 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
11646 if (prog->aux->sleepable) {
11647 verbose(env, "sleepable progs cannot use bpf_spin_lock yet\n");
11652 if (map_value_has_timer(map)) {
11653 if (is_tracing_prog_type(prog_type)) {
11654 verbose(env, "tracing progs cannot use bpf_timer yet\n");
11659 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
11660 !bpf_offload_prog_map_match(prog, map)) {
11661 verbose(env, "offload device mismatch between prog and map\n");
11665 if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
11666 verbose(env, "bpf_struct_ops map cannot be used in prog\n");
11670 if (prog->aux->sleepable)
11671 switch (map->map_type) {
11672 case BPF_MAP_TYPE_HASH:
11673 case BPF_MAP_TYPE_LRU_HASH:
11674 case BPF_MAP_TYPE_ARRAY:
11675 case BPF_MAP_TYPE_PERCPU_HASH:
11676 case BPF_MAP_TYPE_PERCPU_ARRAY:
11677 case BPF_MAP_TYPE_LRU_PERCPU_HASH:
11678 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
11679 case BPF_MAP_TYPE_HASH_OF_MAPS:
11680 if (!is_preallocated_map(map)) {
11682 "Sleepable programs can only use preallocated maps\n");
11686 case BPF_MAP_TYPE_RINGBUF:
11690 "Sleepable programs can only use array, hash, and ringbuf maps\n");
11697 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
11699 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
11700 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
11703 /* find and rewrite pseudo imm in ld_imm64 instructions:
11705 * 1. if it accesses map FD, replace it with actual map pointer.
11706 * 2. if it accesses btf_id of a VAR, replace it with pointer to the var.
11708 * NOTE: btf_vmlinux is required for converting pseudo btf_id.
11710 static int resolve_pseudo_ldimm64(struct bpf_verifier_env *env)
11712 struct bpf_insn *insn = env->prog->insnsi;
11713 int insn_cnt = env->prog->len;
11716 err = bpf_prog_calc_tag(env->prog);
11720 for (i = 0; i < insn_cnt; i++, insn++) {
11721 if (BPF_CLASS(insn->code) == BPF_LDX &&
11722 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
11723 verbose(env, "BPF_LDX uses reserved fields\n");
11727 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
11728 struct bpf_insn_aux_data *aux;
11729 struct bpf_map *map;
11734 if (i == insn_cnt - 1 || insn[1].code != 0 ||
11735 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
11736 insn[1].off != 0) {
11737 verbose(env, "invalid bpf_ld_imm64 insn\n");
11741 if (insn[0].src_reg == 0)
11742 /* valid generic load 64-bit imm */
11745 if (insn[0].src_reg == BPF_PSEUDO_BTF_ID) {
11746 aux = &env->insn_aux_data[i];
11747 err = check_pseudo_btf_id(env, insn, aux);
11753 if (insn[0].src_reg == BPF_PSEUDO_FUNC) {
11754 aux = &env->insn_aux_data[i];
11755 aux->ptr_type = PTR_TO_FUNC;
11759 /* In final convert_pseudo_ld_imm64() step, this is
11760 * converted into regular 64-bit imm load insn.
11762 switch (insn[0].src_reg) {
11763 case BPF_PSEUDO_MAP_VALUE:
11764 case BPF_PSEUDO_MAP_IDX_VALUE:
11766 case BPF_PSEUDO_MAP_FD:
11767 case BPF_PSEUDO_MAP_IDX:
11768 if (insn[1].imm == 0)
11772 verbose(env, "unrecognized bpf_ld_imm64 insn\n");
11776 switch (insn[0].src_reg) {
11777 case BPF_PSEUDO_MAP_IDX_VALUE:
11778 case BPF_PSEUDO_MAP_IDX:
11779 if (bpfptr_is_null(env->fd_array)) {
11780 verbose(env, "fd_idx without fd_array is invalid\n");
11783 if (copy_from_bpfptr_offset(&fd, env->fd_array,
11784 insn[0].imm * sizeof(fd),
11794 map = __bpf_map_get(f);
11796 verbose(env, "fd %d is not pointing to valid bpf_map\n",
11798 return PTR_ERR(map);
11801 err = check_map_prog_compatibility(env, map, env->prog);
11807 aux = &env->insn_aux_data[i];
11808 if (insn[0].src_reg == BPF_PSEUDO_MAP_FD ||
11809 insn[0].src_reg == BPF_PSEUDO_MAP_IDX) {
11810 addr = (unsigned long)map;
11812 u32 off = insn[1].imm;
11814 if (off >= BPF_MAX_VAR_OFF) {
11815 verbose(env, "direct value offset of %u is not allowed\n", off);
11820 if (!map->ops->map_direct_value_addr) {
11821 verbose(env, "no direct value access support for this map type\n");
11826 err = map->ops->map_direct_value_addr(map, &addr, off);
11828 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
11829 map->value_size, off);
11834 aux->map_off = off;
11838 insn[0].imm = (u32)addr;
11839 insn[1].imm = addr >> 32;
11841 /* check whether we recorded this map already */
11842 for (j = 0; j < env->used_map_cnt; j++) {
11843 if (env->used_maps[j] == map) {
11844 aux->map_index = j;
11850 if (env->used_map_cnt >= MAX_USED_MAPS) {
11855 /* hold the map. If the program is rejected by verifier,
11856 * the map will be released by release_maps() or it
11857 * will be used by the valid program until it's unloaded
11858 * and all maps are released in free_used_maps()
11862 aux->map_index = env->used_map_cnt;
11863 env->used_maps[env->used_map_cnt++] = map;
11865 if (bpf_map_is_cgroup_storage(map) &&
11866 bpf_cgroup_storage_assign(env->prog->aux, map)) {
11867 verbose(env, "only one cgroup storage of each type is allowed\n");
11879 /* Basic sanity check before we invest more work here. */
11880 if (!bpf_opcode_in_insntable(insn->code)) {
11881 verbose(env, "unknown opcode %02x\n", insn->code);
11886 /* now all pseudo BPF_LD_IMM64 instructions load valid
11887 * 'struct bpf_map *' into a register instead of user map_fd.
11888 * These pointers will be used later by verifier to validate map access.
11893 /* drop refcnt of maps used by the rejected program */
11894 static void release_maps(struct bpf_verifier_env *env)
11896 __bpf_free_used_maps(env->prog->aux, env->used_maps,
11897 env->used_map_cnt);
11900 /* drop refcnt of maps used by the rejected program */
11901 static void release_btfs(struct bpf_verifier_env *env)
11903 __bpf_free_used_btfs(env->prog->aux, env->used_btfs,
11904 env->used_btf_cnt);
11907 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
11908 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
11910 struct bpf_insn *insn = env->prog->insnsi;
11911 int insn_cnt = env->prog->len;
11914 for (i = 0; i < insn_cnt; i++, insn++) {
11915 if (insn->code != (BPF_LD | BPF_IMM | BPF_DW))
11917 if (insn->src_reg == BPF_PSEUDO_FUNC)
11923 /* single env->prog->insni[off] instruction was replaced with the range
11924 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
11925 * [0, off) and [off, end) to new locations, so the patched range stays zero
11927 static void adjust_insn_aux_data(struct bpf_verifier_env *env,
11928 struct bpf_insn_aux_data *new_data,
11929 struct bpf_prog *new_prog, u32 off, u32 cnt)
11931 struct bpf_insn_aux_data *old_data = env->insn_aux_data;
11932 struct bpf_insn *insn = new_prog->insnsi;
11933 u32 old_seen = old_data[off].seen;
11937 /* aux info at OFF always needs adjustment, no matter fast path
11938 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
11939 * original insn at old prog.
11941 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
11945 prog_len = new_prog->len;
11947 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
11948 memcpy(new_data + off + cnt - 1, old_data + off,
11949 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
11950 for (i = off; i < off + cnt - 1; i++) {
11951 /* Expand insni[off]'s seen count to the patched range. */
11952 new_data[i].seen = old_seen;
11953 new_data[i].zext_dst = insn_has_def32(env, insn + i);
11955 env->insn_aux_data = new_data;
11959 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
11965 /* NOTE: fake 'exit' subprog should be updated as well. */
11966 for (i = 0; i <= env->subprog_cnt; i++) {
11967 if (env->subprog_info[i].start <= off)
11969 env->subprog_info[i].start += len - 1;
11973 static void adjust_poke_descs(struct bpf_prog *prog, u32 off, u32 len)
11975 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
11976 int i, sz = prog->aux->size_poke_tab;
11977 struct bpf_jit_poke_descriptor *desc;
11979 for (i = 0; i < sz; i++) {
11981 if (desc->insn_idx <= off)
11983 desc->insn_idx += len - 1;
11987 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
11988 const struct bpf_insn *patch, u32 len)
11990 struct bpf_prog *new_prog;
11991 struct bpf_insn_aux_data *new_data = NULL;
11994 new_data = vzalloc(array_size(env->prog->len + len - 1,
11995 sizeof(struct bpf_insn_aux_data)));
12000 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
12001 if (IS_ERR(new_prog)) {
12002 if (PTR_ERR(new_prog) == -ERANGE)
12004 "insn %d cannot be patched due to 16-bit range\n",
12005 env->insn_aux_data[off].orig_idx);
12009 adjust_insn_aux_data(env, new_data, new_prog, off, len);
12010 adjust_subprog_starts(env, off, len);
12011 adjust_poke_descs(new_prog, off, len);
12015 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
12020 /* find first prog starting at or after off (first to remove) */
12021 for (i = 0; i < env->subprog_cnt; i++)
12022 if (env->subprog_info[i].start >= off)
12024 /* find first prog starting at or after off + cnt (first to stay) */
12025 for (j = i; j < env->subprog_cnt; j++)
12026 if (env->subprog_info[j].start >= off + cnt)
12028 /* if j doesn't start exactly at off + cnt, we are just removing
12029 * the front of previous prog
12031 if (env->subprog_info[j].start != off + cnt)
12035 struct bpf_prog_aux *aux = env->prog->aux;
12038 /* move fake 'exit' subprog as well */
12039 move = env->subprog_cnt + 1 - j;
12041 memmove(env->subprog_info + i,
12042 env->subprog_info + j,
12043 sizeof(*env->subprog_info) * move);
12044 env->subprog_cnt -= j - i;
12046 /* remove func_info */
12047 if (aux->func_info) {
12048 move = aux->func_info_cnt - j;
12050 memmove(aux->func_info + i,
12051 aux->func_info + j,
12052 sizeof(*aux->func_info) * move);
12053 aux->func_info_cnt -= j - i;
12054 /* func_info->insn_off is set after all code rewrites,
12055 * in adjust_btf_func() - no need to adjust
12059 /* convert i from "first prog to remove" to "first to adjust" */
12060 if (env->subprog_info[i].start == off)
12064 /* update fake 'exit' subprog as well */
12065 for (; i <= env->subprog_cnt; i++)
12066 env->subprog_info[i].start -= cnt;
12071 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
12074 struct bpf_prog *prog = env->prog;
12075 u32 i, l_off, l_cnt, nr_linfo;
12076 struct bpf_line_info *linfo;
12078 nr_linfo = prog->aux->nr_linfo;
12082 linfo = prog->aux->linfo;
12084 /* find first line info to remove, count lines to be removed */
12085 for (i = 0; i < nr_linfo; i++)
12086 if (linfo[i].insn_off >= off)
12091 for (; i < nr_linfo; i++)
12092 if (linfo[i].insn_off < off + cnt)
12097 /* First live insn doesn't match first live linfo, it needs to "inherit"
12098 * last removed linfo. prog is already modified, so prog->len == off
12099 * means no live instructions after (tail of the program was removed).
12101 if (prog->len != off && l_cnt &&
12102 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
12104 linfo[--i].insn_off = off + cnt;
12107 /* remove the line info which refer to the removed instructions */
12109 memmove(linfo + l_off, linfo + i,
12110 sizeof(*linfo) * (nr_linfo - i));
12112 prog->aux->nr_linfo -= l_cnt;
12113 nr_linfo = prog->aux->nr_linfo;
12116 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
12117 for (i = l_off; i < nr_linfo; i++)
12118 linfo[i].insn_off -= cnt;
12120 /* fix up all subprogs (incl. 'exit') which start >= off */
12121 for (i = 0; i <= env->subprog_cnt; i++)
12122 if (env->subprog_info[i].linfo_idx > l_off) {
12123 /* program may have started in the removed region but
12124 * may not be fully removed
12126 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
12127 env->subprog_info[i].linfo_idx -= l_cnt;
12129 env->subprog_info[i].linfo_idx = l_off;
12135 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
12137 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
12138 unsigned int orig_prog_len = env->prog->len;
12141 if (bpf_prog_is_dev_bound(env->prog->aux))
12142 bpf_prog_offload_remove_insns(env, off, cnt);
12144 err = bpf_remove_insns(env->prog, off, cnt);
12148 err = adjust_subprog_starts_after_remove(env, off, cnt);
12152 err = bpf_adj_linfo_after_remove(env, off, cnt);
12156 memmove(aux_data + off, aux_data + off + cnt,
12157 sizeof(*aux_data) * (orig_prog_len - off - cnt));
12162 /* The verifier does more data flow analysis than llvm and will not
12163 * explore branches that are dead at run time. Malicious programs can
12164 * have dead code too. Therefore replace all dead at-run-time code
12167 * Just nops are not optimal, e.g. if they would sit at the end of the
12168 * program and through another bug we would manage to jump there, then
12169 * we'd execute beyond program memory otherwise. Returning exception
12170 * code also wouldn't work since we can have subprogs where the dead
12171 * code could be located.
12173 static void sanitize_dead_code(struct bpf_verifier_env *env)
12175 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
12176 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
12177 struct bpf_insn *insn = env->prog->insnsi;
12178 const int insn_cnt = env->prog->len;
12181 for (i = 0; i < insn_cnt; i++) {
12182 if (aux_data[i].seen)
12184 memcpy(insn + i, &trap, sizeof(trap));
12185 aux_data[i].zext_dst = false;
12189 static bool insn_is_cond_jump(u8 code)
12193 if (BPF_CLASS(code) == BPF_JMP32)
12196 if (BPF_CLASS(code) != BPF_JMP)
12200 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
12203 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
12205 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
12206 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
12207 struct bpf_insn *insn = env->prog->insnsi;
12208 const int insn_cnt = env->prog->len;
12211 for (i = 0; i < insn_cnt; i++, insn++) {
12212 if (!insn_is_cond_jump(insn->code))
12215 if (!aux_data[i + 1].seen)
12216 ja.off = insn->off;
12217 else if (!aux_data[i + 1 + insn->off].seen)
12222 if (bpf_prog_is_dev_bound(env->prog->aux))
12223 bpf_prog_offload_replace_insn(env, i, &ja);
12225 memcpy(insn, &ja, sizeof(ja));
12229 static int opt_remove_dead_code(struct bpf_verifier_env *env)
12231 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
12232 int insn_cnt = env->prog->len;
12235 for (i = 0; i < insn_cnt; i++) {
12239 while (i + j < insn_cnt && !aux_data[i + j].seen)
12244 err = verifier_remove_insns(env, i, j);
12247 insn_cnt = env->prog->len;
12253 static int opt_remove_nops(struct bpf_verifier_env *env)
12255 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
12256 struct bpf_insn *insn = env->prog->insnsi;
12257 int insn_cnt = env->prog->len;
12260 for (i = 0; i < insn_cnt; i++) {
12261 if (memcmp(&insn[i], &ja, sizeof(ja)))
12264 err = verifier_remove_insns(env, i, 1);
12274 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
12275 const union bpf_attr *attr)
12277 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
12278 struct bpf_insn_aux_data *aux = env->insn_aux_data;
12279 int i, patch_len, delta = 0, len = env->prog->len;
12280 struct bpf_insn *insns = env->prog->insnsi;
12281 struct bpf_prog *new_prog;
12284 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
12285 zext_patch[1] = BPF_ZEXT_REG(0);
12286 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
12287 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
12288 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
12289 for (i = 0; i < len; i++) {
12290 int adj_idx = i + delta;
12291 struct bpf_insn insn;
12294 insn = insns[adj_idx];
12295 load_reg = insn_def_regno(&insn);
12296 if (!aux[adj_idx].zext_dst) {
12304 class = BPF_CLASS(code);
12305 if (load_reg == -1)
12308 /* NOTE: arg "reg" (the fourth one) is only used for
12309 * BPF_STX + SRC_OP, so it is safe to pass NULL
12312 if (is_reg64(env, &insn, load_reg, NULL, DST_OP)) {
12313 if (class == BPF_LD &&
12314 BPF_MODE(code) == BPF_IMM)
12319 /* ctx load could be transformed into wider load. */
12320 if (class == BPF_LDX &&
12321 aux[adj_idx].ptr_type == PTR_TO_CTX)
12324 imm_rnd = get_random_int();
12325 rnd_hi32_patch[0] = insn;
12326 rnd_hi32_patch[1].imm = imm_rnd;
12327 rnd_hi32_patch[3].dst_reg = load_reg;
12328 patch = rnd_hi32_patch;
12330 goto apply_patch_buffer;
12333 /* Add in an zero-extend instruction if a) the JIT has requested
12334 * it or b) it's a CMPXCHG.
12336 * The latter is because: BPF_CMPXCHG always loads a value into
12337 * R0, therefore always zero-extends. However some archs'
12338 * equivalent instruction only does this load when the
12339 * comparison is successful. This detail of CMPXCHG is
12340 * orthogonal to the general zero-extension behaviour of the
12341 * CPU, so it's treated independently of bpf_jit_needs_zext.
12343 if (!bpf_jit_needs_zext() && !is_cmpxchg_insn(&insn))
12346 if (WARN_ON(load_reg == -1)) {
12347 verbose(env, "verifier bug. zext_dst is set, but no reg is defined\n");
12351 zext_patch[0] = insn;
12352 zext_patch[1].dst_reg = load_reg;
12353 zext_patch[1].src_reg = load_reg;
12354 patch = zext_patch;
12356 apply_patch_buffer:
12357 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
12360 env->prog = new_prog;
12361 insns = new_prog->insnsi;
12362 aux = env->insn_aux_data;
12363 delta += patch_len - 1;
12369 /* convert load instructions that access fields of a context type into a
12370 * sequence of instructions that access fields of the underlying structure:
12371 * struct __sk_buff -> struct sk_buff
12372 * struct bpf_sock_ops -> struct sock
12374 static int convert_ctx_accesses(struct bpf_verifier_env *env)
12376 const struct bpf_verifier_ops *ops = env->ops;
12377 int i, cnt, size, ctx_field_size, delta = 0;
12378 const int insn_cnt = env->prog->len;
12379 struct bpf_insn insn_buf[16], *insn;
12380 u32 target_size, size_default, off;
12381 struct bpf_prog *new_prog;
12382 enum bpf_access_type type;
12383 bool is_narrower_load;
12385 if (ops->gen_prologue || env->seen_direct_write) {
12386 if (!ops->gen_prologue) {
12387 verbose(env, "bpf verifier is misconfigured\n");
12390 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
12392 if (cnt >= ARRAY_SIZE(insn_buf)) {
12393 verbose(env, "bpf verifier is misconfigured\n");
12396 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
12400 env->prog = new_prog;
12405 if (bpf_prog_is_dev_bound(env->prog->aux))
12408 insn = env->prog->insnsi + delta;
12410 for (i = 0; i < insn_cnt; i++, insn++) {
12411 bpf_convert_ctx_access_t convert_ctx_access;
12414 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
12415 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
12416 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
12417 insn->code == (BPF_LDX | BPF_MEM | BPF_DW)) {
12420 } else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
12421 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
12422 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
12423 insn->code == (BPF_STX | BPF_MEM | BPF_DW) ||
12424 insn->code == (BPF_ST | BPF_MEM | BPF_B) ||
12425 insn->code == (BPF_ST | BPF_MEM | BPF_H) ||
12426 insn->code == (BPF_ST | BPF_MEM | BPF_W) ||
12427 insn->code == (BPF_ST | BPF_MEM | BPF_DW)) {
12429 ctx_access = BPF_CLASS(insn->code) == BPF_STX;
12434 if (type == BPF_WRITE &&
12435 env->insn_aux_data[i + delta].sanitize_stack_spill) {
12436 struct bpf_insn patch[] = {
12441 cnt = ARRAY_SIZE(patch);
12442 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
12447 env->prog = new_prog;
12448 insn = new_prog->insnsi + i + delta;
12455 switch (env->insn_aux_data[i + delta].ptr_type) {
12457 if (!ops->convert_ctx_access)
12459 convert_ctx_access = ops->convert_ctx_access;
12461 case PTR_TO_SOCKET:
12462 case PTR_TO_SOCK_COMMON:
12463 convert_ctx_access = bpf_sock_convert_ctx_access;
12465 case PTR_TO_TCP_SOCK:
12466 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
12468 case PTR_TO_XDP_SOCK:
12469 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
12471 case PTR_TO_BTF_ID:
12472 if (type == BPF_READ) {
12473 insn->code = BPF_LDX | BPF_PROBE_MEM |
12474 BPF_SIZE((insn)->code);
12475 env->prog->aux->num_exentries++;
12476 } else if (resolve_prog_type(env->prog) != BPF_PROG_TYPE_STRUCT_OPS) {
12477 verbose(env, "Writes through BTF pointers are not allowed\n");
12485 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
12486 size = BPF_LDST_BYTES(insn);
12488 /* If the read access is a narrower load of the field,
12489 * convert to a 4/8-byte load, to minimum program type specific
12490 * convert_ctx_access changes. If conversion is successful,
12491 * we will apply proper mask to the result.
12493 is_narrower_load = size < ctx_field_size;
12494 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
12496 if (is_narrower_load) {
12499 if (type == BPF_WRITE) {
12500 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
12505 if (ctx_field_size == 4)
12507 else if (ctx_field_size == 8)
12508 size_code = BPF_DW;
12510 insn->off = off & ~(size_default - 1);
12511 insn->code = BPF_LDX | BPF_MEM | size_code;
12515 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
12517 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
12518 (ctx_field_size && !target_size)) {
12519 verbose(env, "bpf verifier is misconfigured\n");
12523 if (is_narrower_load && size < target_size) {
12524 u8 shift = bpf_ctx_narrow_access_offset(
12525 off, size, size_default) * 8;
12526 if (shift && cnt + 1 >= ARRAY_SIZE(insn_buf)) {
12527 verbose(env, "bpf verifier narrow ctx load misconfigured\n");
12530 if (ctx_field_size <= 4) {
12532 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
12535 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
12536 (1 << size * 8) - 1);
12539 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
12542 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
12543 (1ULL << size * 8) - 1);
12547 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12553 /* keep walking new program and skip insns we just inserted */
12554 env->prog = new_prog;
12555 insn = new_prog->insnsi + i + delta;
12561 static int jit_subprogs(struct bpf_verifier_env *env)
12563 struct bpf_prog *prog = env->prog, **func, *tmp;
12564 int i, j, subprog_start, subprog_end = 0, len, subprog;
12565 struct bpf_map *map_ptr;
12566 struct bpf_insn *insn;
12567 void *old_bpf_func;
12568 int err, num_exentries;
12570 if (env->subprog_cnt <= 1)
12573 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
12574 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn))
12577 /* Upon error here we cannot fall back to interpreter but
12578 * need a hard reject of the program. Thus -EFAULT is
12579 * propagated in any case.
12581 subprog = find_subprog(env, i + insn->imm + 1);
12583 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
12584 i + insn->imm + 1);
12587 /* temporarily remember subprog id inside insn instead of
12588 * aux_data, since next loop will split up all insns into funcs
12590 insn->off = subprog;
12591 /* remember original imm in case JIT fails and fallback
12592 * to interpreter will be needed
12594 env->insn_aux_data[i].call_imm = insn->imm;
12595 /* point imm to __bpf_call_base+1 from JITs point of view */
12597 if (bpf_pseudo_func(insn))
12598 /* jit (e.g. x86_64) may emit fewer instructions
12599 * if it learns a u32 imm is the same as a u64 imm.
12600 * Force a non zero here.
12605 err = bpf_prog_alloc_jited_linfo(prog);
12607 goto out_undo_insn;
12610 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
12612 goto out_undo_insn;
12614 for (i = 0; i < env->subprog_cnt; i++) {
12615 subprog_start = subprog_end;
12616 subprog_end = env->subprog_info[i + 1].start;
12618 len = subprog_end - subprog_start;
12619 /* bpf_prog_run() doesn't call subprogs directly,
12620 * hence main prog stats include the runtime of subprogs.
12621 * subprogs don't have IDs and not reachable via prog_get_next_id
12622 * func[i]->stats will never be accessed and stays NULL
12624 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
12627 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
12628 len * sizeof(struct bpf_insn));
12629 func[i]->type = prog->type;
12630 func[i]->len = len;
12631 if (bpf_prog_calc_tag(func[i]))
12633 func[i]->is_func = 1;
12634 func[i]->aux->func_idx = i;
12635 /* Below members will be freed only at prog->aux */
12636 func[i]->aux->btf = prog->aux->btf;
12637 func[i]->aux->func_info = prog->aux->func_info;
12638 func[i]->aux->poke_tab = prog->aux->poke_tab;
12639 func[i]->aux->size_poke_tab = prog->aux->size_poke_tab;
12641 for (j = 0; j < prog->aux->size_poke_tab; j++) {
12642 struct bpf_jit_poke_descriptor *poke;
12644 poke = &prog->aux->poke_tab[j];
12645 if (poke->insn_idx < subprog_end &&
12646 poke->insn_idx >= subprog_start)
12647 poke->aux = func[i]->aux;
12650 /* Use bpf_prog_F_tag to indicate functions in stack traces.
12651 * Long term would need debug info to populate names
12653 func[i]->aux->name[0] = 'F';
12654 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
12655 func[i]->jit_requested = 1;
12656 func[i]->aux->kfunc_tab = prog->aux->kfunc_tab;
12657 func[i]->aux->kfunc_btf_tab = prog->aux->kfunc_btf_tab;
12658 func[i]->aux->linfo = prog->aux->linfo;
12659 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
12660 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
12661 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
12663 insn = func[i]->insnsi;
12664 for (j = 0; j < func[i]->len; j++, insn++) {
12665 if (BPF_CLASS(insn->code) == BPF_LDX &&
12666 BPF_MODE(insn->code) == BPF_PROBE_MEM)
12669 func[i]->aux->num_exentries = num_exentries;
12670 func[i]->aux->tail_call_reachable = env->subprog_info[i].tail_call_reachable;
12671 func[i] = bpf_int_jit_compile(func[i]);
12672 if (!func[i]->jited) {
12679 /* at this point all bpf functions were successfully JITed
12680 * now populate all bpf_calls with correct addresses and
12681 * run last pass of JIT
12683 for (i = 0; i < env->subprog_cnt; i++) {
12684 insn = func[i]->insnsi;
12685 for (j = 0; j < func[i]->len; j++, insn++) {
12686 if (bpf_pseudo_func(insn)) {
12687 subprog = insn->off;
12688 insn[0].imm = (u32)(long)func[subprog]->bpf_func;
12689 insn[1].imm = ((u64)(long)func[subprog]->bpf_func) >> 32;
12692 if (!bpf_pseudo_call(insn))
12694 subprog = insn->off;
12695 insn->imm = BPF_CALL_IMM(func[subprog]->bpf_func);
12698 /* we use the aux data to keep a list of the start addresses
12699 * of the JITed images for each function in the program
12701 * for some architectures, such as powerpc64, the imm field
12702 * might not be large enough to hold the offset of the start
12703 * address of the callee's JITed image from __bpf_call_base
12705 * in such cases, we can lookup the start address of a callee
12706 * by using its subprog id, available from the off field of
12707 * the call instruction, as an index for this list
12709 func[i]->aux->func = func;
12710 func[i]->aux->func_cnt = env->subprog_cnt;
12712 for (i = 0; i < env->subprog_cnt; i++) {
12713 old_bpf_func = func[i]->bpf_func;
12714 tmp = bpf_int_jit_compile(func[i]);
12715 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
12716 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
12723 /* finally lock prog and jit images for all functions and
12724 * populate kallsysm
12726 for (i = 0; i < env->subprog_cnt; i++) {
12727 bpf_prog_lock_ro(func[i]);
12728 bpf_prog_kallsyms_add(func[i]);
12731 /* Last step: make now unused interpreter insns from main
12732 * prog consistent for later dump requests, so they can
12733 * later look the same as if they were interpreted only.
12735 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
12736 if (bpf_pseudo_func(insn)) {
12737 insn[0].imm = env->insn_aux_data[i].call_imm;
12738 insn[1].imm = insn->off;
12742 if (!bpf_pseudo_call(insn))
12744 insn->off = env->insn_aux_data[i].call_imm;
12745 subprog = find_subprog(env, i + insn->off + 1);
12746 insn->imm = subprog;
12750 prog->bpf_func = func[0]->bpf_func;
12751 prog->aux->func = func;
12752 prog->aux->func_cnt = env->subprog_cnt;
12753 bpf_prog_jit_attempt_done(prog);
12756 /* We failed JIT'ing, so at this point we need to unregister poke
12757 * descriptors from subprogs, so that kernel is not attempting to
12758 * patch it anymore as we're freeing the subprog JIT memory.
12760 for (i = 0; i < prog->aux->size_poke_tab; i++) {
12761 map_ptr = prog->aux->poke_tab[i].tail_call.map;
12762 map_ptr->ops->map_poke_untrack(map_ptr, prog->aux);
12764 /* At this point we're guaranteed that poke descriptors are not
12765 * live anymore. We can just unlink its descriptor table as it's
12766 * released with the main prog.
12768 for (i = 0; i < env->subprog_cnt; i++) {
12771 func[i]->aux->poke_tab = NULL;
12772 bpf_jit_free(func[i]);
12776 /* cleanup main prog to be interpreted */
12777 prog->jit_requested = 0;
12778 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
12779 if (!bpf_pseudo_call(insn))
12782 insn->imm = env->insn_aux_data[i].call_imm;
12784 bpf_prog_jit_attempt_done(prog);
12788 static int fixup_call_args(struct bpf_verifier_env *env)
12790 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
12791 struct bpf_prog *prog = env->prog;
12792 struct bpf_insn *insn = prog->insnsi;
12793 bool has_kfunc_call = bpf_prog_has_kfunc_call(prog);
12798 if (env->prog->jit_requested &&
12799 !bpf_prog_is_dev_bound(env->prog->aux)) {
12800 err = jit_subprogs(env);
12803 if (err == -EFAULT)
12806 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
12807 if (has_kfunc_call) {
12808 verbose(env, "calling kernel functions are not allowed in non-JITed programs\n");
12811 if (env->subprog_cnt > 1 && env->prog->aux->tail_call_reachable) {
12812 /* When JIT fails the progs with bpf2bpf calls and tail_calls
12813 * have to be rejected, since interpreter doesn't support them yet.
12815 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
12818 for (i = 0; i < prog->len; i++, insn++) {
12819 if (bpf_pseudo_func(insn)) {
12820 /* When JIT fails the progs with callback calls
12821 * have to be rejected, since interpreter doesn't support them yet.
12823 verbose(env, "callbacks are not allowed in non-JITed programs\n");
12827 if (!bpf_pseudo_call(insn))
12829 depth = get_callee_stack_depth(env, insn, i);
12832 bpf_patch_call_args(insn, depth);
12839 static int fixup_kfunc_call(struct bpf_verifier_env *env,
12840 struct bpf_insn *insn)
12842 const struct bpf_kfunc_desc *desc;
12845 verbose(env, "invalid kernel function call not eliminated in verifier pass\n");
12849 /* insn->imm has the btf func_id. Replace it with
12850 * an address (relative to __bpf_base_call).
12852 desc = find_kfunc_desc(env->prog, insn->imm, insn->off);
12854 verbose(env, "verifier internal error: kernel function descriptor not found for func_id %u\n",
12859 insn->imm = desc->imm;
12864 /* Do various post-verification rewrites in a single program pass.
12865 * These rewrites simplify JIT and interpreter implementations.
12867 static int do_misc_fixups(struct bpf_verifier_env *env)
12869 struct bpf_prog *prog = env->prog;
12870 bool expect_blinding = bpf_jit_blinding_enabled(prog);
12871 enum bpf_prog_type prog_type = resolve_prog_type(prog);
12872 struct bpf_insn *insn = prog->insnsi;
12873 const struct bpf_func_proto *fn;
12874 const int insn_cnt = prog->len;
12875 const struct bpf_map_ops *ops;
12876 struct bpf_insn_aux_data *aux;
12877 struct bpf_insn insn_buf[16];
12878 struct bpf_prog *new_prog;
12879 struct bpf_map *map_ptr;
12880 int i, ret, cnt, delta = 0;
12882 for (i = 0; i < insn_cnt; i++, insn++) {
12883 /* Make divide-by-zero exceptions impossible. */
12884 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
12885 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
12886 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
12887 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
12888 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
12889 bool isdiv = BPF_OP(insn->code) == BPF_DIV;
12890 struct bpf_insn *patchlet;
12891 struct bpf_insn chk_and_div[] = {
12892 /* [R,W]x div 0 -> 0 */
12893 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
12894 BPF_JNE | BPF_K, insn->src_reg,
12896 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
12897 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
12900 struct bpf_insn chk_and_mod[] = {
12901 /* [R,W]x mod 0 -> [R,W]x */
12902 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
12903 BPF_JEQ | BPF_K, insn->src_reg,
12904 0, 1 + (is64 ? 0 : 1), 0),
12906 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
12907 BPF_MOV32_REG(insn->dst_reg, insn->dst_reg),
12910 patchlet = isdiv ? chk_and_div : chk_and_mod;
12911 cnt = isdiv ? ARRAY_SIZE(chk_and_div) :
12912 ARRAY_SIZE(chk_and_mod) - (is64 ? 2 : 0);
12914 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
12919 env->prog = prog = new_prog;
12920 insn = new_prog->insnsi + i + delta;
12924 /* Implement LD_ABS and LD_IND with a rewrite, if supported by the program type. */
12925 if (BPF_CLASS(insn->code) == BPF_LD &&
12926 (BPF_MODE(insn->code) == BPF_ABS ||
12927 BPF_MODE(insn->code) == BPF_IND)) {
12928 cnt = env->ops->gen_ld_abs(insn, insn_buf);
12929 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
12930 verbose(env, "bpf verifier is misconfigured\n");
12934 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12939 env->prog = prog = new_prog;
12940 insn = new_prog->insnsi + i + delta;
12944 /* Rewrite pointer arithmetic to mitigate speculation attacks. */
12945 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
12946 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
12947 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
12948 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
12949 struct bpf_insn *patch = &insn_buf[0];
12950 bool issrc, isneg, isimm;
12953 aux = &env->insn_aux_data[i + delta];
12954 if (!aux->alu_state ||
12955 aux->alu_state == BPF_ALU_NON_POINTER)
12958 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
12959 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
12960 BPF_ALU_SANITIZE_SRC;
12961 isimm = aux->alu_state & BPF_ALU_IMMEDIATE;
12963 off_reg = issrc ? insn->src_reg : insn->dst_reg;
12965 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
12968 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
12969 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
12970 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
12971 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
12972 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
12973 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
12974 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX, off_reg);
12977 *patch++ = BPF_MOV64_REG(insn->dst_reg, insn->src_reg);
12978 insn->src_reg = BPF_REG_AX;
12980 insn->code = insn->code == code_add ?
12981 code_sub : code_add;
12983 if (issrc && isneg && !isimm)
12984 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
12985 cnt = patch - insn_buf;
12987 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12992 env->prog = prog = new_prog;
12993 insn = new_prog->insnsi + i + delta;
12997 if (insn->code != (BPF_JMP | BPF_CALL))
12999 if (insn->src_reg == BPF_PSEUDO_CALL)
13001 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
13002 ret = fixup_kfunc_call(env, insn);
13008 if (insn->imm == BPF_FUNC_get_route_realm)
13009 prog->dst_needed = 1;
13010 if (insn->imm == BPF_FUNC_get_prandom_u32)
13011 bpf_user_rnd_init_once();
13012 if (insn->imm == BPF_FUNC_override_return)
13013 prog->kprobe_override = 1;
13014 if (insn->imm == BPF_FUNC_tail_call) {
13015 /* If we tail call into other programs, we
13016 * cannot make any assumptions since they can
13017 * be replaced dynamically during runtime in
13018 * the program array.
13020 prog->cb_access = 1;
13021 if (!allow_tail_call_in_subprogs(env))
13022 prog->aux->stack_depth = MAX_BPF_STACK;
13023 prog->aux->max_pkt_offset = MAX_PACKET_OFF;
13025 /* mark bpf_tail_call as different opcode to avoid
13026 * conditional branch in the interpreter for every normal
13027 * call and to prevent accidental JITing by JIT compiler
13028 * that doesn't support bpf_tail_call yet
13031 insn->code = BPF_JMP | BPF_TAIL_CALL;
13033 aux = &env->insn_aux_data[i + delta];
13034 if (env->bpf_capable && !expect_blinding &&
13035 prog->jit_requested &&
13036 !bpf_map_key_poisoned(aux) &&
13037 !bpf_map_ptr_poisoned(aux) &&
13038 !bpf_map_ptr_unpriv(aux)) {
13039 struct bpf_jit_poke_descriptor desc = {
13040 .reason = BPF_POKE_REASON_TAIL_CALL,
13041 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
13042 .tail_call.key = bpf_map_key_immediate(aux),
13043 .insn_idx = i + delta,
13046 ret = bpf_jit_add_poke_descriptor(prog, &desc);
13048 verbose(env, "adding tail call poke descriptor failed\n");
13052 insn->imm = ret + 1;
13056 if (!bpf_map_ptr_unpriv(aux))
13059 /* instead of changing every JIT dealing with tail_call
13060 * emit two extra insns:
13061 * if (index >= max_entries) goto out;
13062 * index &= array->index_mask;
13063 * to avoid out-of-bounds cpu speculation
13065 if (bpf_map_ptr_poisoned(aux)) {
13066 verbose(env, "tail_call abusing map_ptr\n");
13070 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
13071 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
13072 map_ptr->max_entries, 2);
13073 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
13074 container_of(map_ptr,
13077 insn_buf[2] = *insn;
13079 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
13084 env->prog = prog = new_prog;
13085 insn = new_prog->insnsi + i + delta;
13089 if (insn->imm == BPF_FUNC_timer_set_callback) {
13090 /* The verifier will process callback_fn as many times as necessary
13091 * with different maps and the register states prepared by
13092 * set_timer_callback_state will be accurate.
13094 * The following use case is valid:
13095 * map1 is shared by prog1, prog2, prog3.
13096 * prog1 calls bpf_timer_init for some map1 elements
13097 * prog2 calls bpf_timer_set_callback for some map1 elements.
13098 * Those that were not bpf_timer_init-ed will return -EINVAL.
13099 * prog3 calls bpf_timer_start for some map1 elements.
13100 * Those that were not both bpf_timer_init-ed and
13101 * bpf_timer_set_callback-ed will return -EINVAL.
13103 struct bpf_insn ld_addrs[2] = {
13104 BPF_LD_IMM64(BPF_REG_3, (long)prog->aux),
13107 insn_buf[0] = ld_addrs[0];
13108 insn_buf[1] = ld_addrs[1];
13109 insn_buf[2] = *insn;
13112 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
13117 env->prog = prog = new_prog;
13118 insn = new_prog->insnsi + i + delta;
13119 goto patch_call_imm;
13122 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
13123 * and other inlining handlers are currently limited to 64 bit
13126 if (prog->jit_requested && BITS_PER_LONG == 64 &&
13127 (insn->imm == BPF_FUNC_map_lookup_elem ||
13128 insn->imm == BPF_FUNC_map_update_elem ||
13129 insn->imm == BPF_FUNC_map_delete_elem ||
13130 insn->imm == BPF_FUNC_map_push_elem ||
13131 insn->imm == BPF_FUNC_map_pop_elem ||
13132 insn->imm == BPF_FUNC_map_peek_elem ||
13133 insn->imm == BPF_FUNC_redirect_map ||
13134 insn->imm == BPF_FUNC_for_each_map_elem)) {
13135 aux = &env->insn_aux_data[i + delta];
13136 if (bpf_map_ptr_poisoned(aux))
13137 goto patch_call_imm;
13139 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
13140 ops = map_ptr->ops;
13141 if (insn->imm == BPF_FUNC_map_lookup_elem &&
13142 ops->map_gen_lookup) {
13143 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
13144 if (cnt == -EOPNOTSUPP)
13145 goto patch_map_ops_generic;
13146 if (cnt <= 0 || cnt >= ARRAY_SIZE(insn_buf)) {
13147 verbose(env, "bpf verifier is misconfigured\n");
13151 new_prog = bpf_patch_insn_data(env, i + delta,
13157 env->prog = prog = new_prog;
13158 insn = new_prog->insnsi + i + delta;
13162 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
13163 (void *(*)(struct bpf_map *map, void *key))NULL));
13164 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
13165 (int (*)(struct bpf_map *map, void *key))NULL));
13166 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
13167 (int (*)(struct bpf_map *map, void *key, void *value,
13169 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
13170 (int (*)(struct bpf_map *map, void *value,
13172 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
13173 (int (*)(struct bpf_map *map, void *value))NULL));
13174 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
13175 (int (*)(struct bpf_map *map, void *value))NULL));
13176 BUILD_BUG_ON(!__same_type(ops->map_redirect,
13177 (int (*)(struct bpf_map *map, u32 ifindex, u64 flags))NULL));
13178 BUILD_BUG_ON(!__same_type(ops->map_for_each_callback,
13179 (int (*)(struct bpf_map *map,
13180 bpf_callback_t callback_fn,
13181 void *callback_ctx,
13184 patch_map_ops_generic:
13185 switch (insn->imm) {
13186 case BPF_FUNC_map_lookup_elem:
13187 insn->imm = BPF_CALL_IMM(ops->map_lookup_elem);
13189 case BPF_FUNC_map_update_elem:
13190 insn->imm = BPF_CALL_IMM(ops->map_update_elem);
13192 case BPF_FUNC_map_delete_elem:
13193 insn->imm = BPF_CALL_IMM(ops->map_delete_elem);
13195 case BPF_FUNC_map_push_elem:
13196 insn->imm = BPF_CALL_IMM(ops->map_push_elem);
13198 case BPF_FUNC_map_pop_elem:
13199 insn->imm = BPF_CALL_IMM(ops->map_pop_elem);
13201 case BPF_FUNC_map_peek_elem:
13202 insn->imm = BPF_CALL_IMM(ops->map_peek_elem);
13204 case BPF_FUNC_redirect_map:
13205 insn->imm = BPF_CALL_IMM(ops->map_redirect);
13207 case BPF_FUNC_for_each_map_elem:
13208 insn->imm = BPF_CALL_IMM(ops->map_for_each_callback);
13212 goto patch_call_imm;
13215 /* Implement bpf_jiffies64 inline. */
13216 if (prog->jit_requested && BITS_PER_LONG == 64 &&
13217 insn->imm == BPF_FUNC_jiffies64) {
13218 struct bpf_insn ld_jiffies_addr[2] = {
13219 BPF_LD_IMM64(BPF_REG_0,
13220 (unsigned long)&jiffies),
13223 insn_buf[0] = ld_jiffies_addr[0];
13224 insn_buf[1] = ld_jiffies_addr[1];
13225 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
13229 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
13235 env->prog = prog = new_prog;
13236 insn = new_prog->insnsi + i + delta;
13240 /* Implement bpf_get_func_ip inline. */
13241 if (prog_type == BPF_PROG_TYPE_TRACING &&
13242 insn->imm == BPF_FUNC_get_func_ip) {
13243 /* Load IP address from ctx - 8 */
13244 insn_buf[0] = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_1, -8);
13246 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, 1);
13250 env->prog = prog = new_prog;
13251 insn = new_prog->insnsi + i + delta;
13256 fn = env->ops->get_func_proto(insn->imm, env->prog);
13257 /* all functions that have prototype and verifier allowed
13258 * programs to call them, must be real in-kernel functions
13262 "kernel subsystem misconfigured func %s#%d\n",
13263 func_id_name(insn->imm), insn->imm);
13266 insn->imm = fn->func - __bpf_call_base;
13269 /* Since poke tab is now finalized, publish aux to tracker. */
13270 for (i = 0; i < prog->aux->size_poke_tab; i++) {
13271 map_ptr = prog->aux->poke_tab[i].tail_call.map;
13272 if (!map_ptr->ops->map_poke_track ||
13273 !map_ptr->ops->map_poke_untrack ||
13274 !map_ptr->ops->map_poke_run) {
13275 verbose(env, "bpf verifier is misconfigured\n");
13279 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
13281 verbose(env, "tracking tail call prog failed\n");
13286 sort_kfunc_descs_by_imm(env->prog);
13291 static void free_states(struct bpf_verifier_env *env)
13293 struct bpf_verifier_state_list *sl, *sln;
13296 sl = env->free_list;
13299 free_verifier_state(&sl->state, false);
13303 env->free_list = NULL;
13305 if (!env->explored_states)
13308 for (i = 0; i < state_htab_size(env); i++) {
13309 sl = env->explored_states[i];
13313 free_verifier_state(&sl->state, false);
13317 env->explored_states[i] = NULL;
13321 static int do_check_common(struct bpf_verifier_env *env, int subprog)
13323 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
13324 struct bpf_verifier_state *state;
13325 struct bpf_reg_state *regs;
13328 env->prev_linfo = NULL;
13331 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
13334 state->curframe = 0;
13335 state->speculative = false;
13336 state->branches = 1;
13337 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
13338 if (!state->frame[0]) {
13342 env->cur_state = state;
13343 init_func_state(env, state->frame[0],
13344 BPF_MAIN_FUNC /* callsite */,
13348 regs = state->frame[state->curframe]->regs;
13349 if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
13350 ret = btf_prepare_func_args(env, subprog, regs);
13353 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
13354 if (regs[i].type == PTR_TO_CTX)
13355 mark_reg_known_zero(env, regs, i);
13356 else if (regs[i].type == SCALAR_VALUE)
13357 mark_reg_unknown(env, regs, i);
13358 else if (regs[i].type == PTR_TO_MEM_OR_NULL) {
13359 const u32 mem_size = regs[i].mem_size;
13361 mark_reg_known_zero(env, regs, i);
13362 regs[i].mem_size = mem_size;
13363 regs[i].id = ++env->id_gen;
13367 /* 1st arg to a function */
13368 regs[BPF_REG_1].type = PTR_TO_CTX;
13369 mark_reg_known_zero(env, regs, BPF_REG_1);
13370 ret = btf_check_subprog_arg_match(env, subprog, regs);
13371 if (ret == -EFAULT)
13372 /* unlikely verifier bug. abort.
13373 * ret == 0 and ret < 0 are sadly acceptable for
13374 * main() function due to backward compatibility.
13375 * Like socket filter program may be written as:
13376 * int bpf_prog(struct pt_regs *ctx)
13377 * and never dereference that ctx in the program.
13378 * 'struct pt_regs' is a type mismatch for socket
13379 * filter that should be using 'struct __sk_buff'.
13384 ret = do_check(env);
13386 /* check for NULL is necessary, since cur_state can be freed inside
13387 * do_check() under memory pressure.
13389 if (env->cur_state) {
13390 free_verifier_state(env->cur_state, true);
13391 env->cur_state = NULL;
13393 while (!pop_stack(env, NULL, NULL, false));
13394 if (!ret && pop_log)
13395 bpf_vlog_reset(&env->log, 0);
13400 /* Verify all global functions in a BPF program one by one based on their BTF.
13401 * All global functions must pass verification. Otherwise the whole program is rejected.
13412 * foo() will be verified first for R1=any_scalar_value. During verification it
13413 * will be assumed that bar() already verified successfully and call to bar()
13414 * from foo() will be checked for type match only. Later bar() will be verified
13415 * independently to check that it's safe for R1=any_scalar_value.
13417 static int do_check_subprogs(struct bpf_verifier_env *env)
13419 struct bpf_prog_aux *aux = env->prog->aux;
13422 if (!aux->func_info)
13425 for (i = 1; i < env->subprog_cnt; i++) {
13426 if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL)
13428 env->insn_idx = env->subprog_info[i].start;
13429 WARN_ON_ONCE(env->insn_idx == 0);
13430 ret = do_check_common(env, i);
13433 } else if (env->log.level & BPF_LOG_LEVEL) {
13435 "Func#%d is safe for any args that match its prototype\n",
13442 static int do_check_main(struct bpf_verifier_env *env)
13447 ret = do_check_common(env, 0);
13449 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
13454 static void print_verification_stats(struct bpf_verifier_env *env)
13458 if (env->log.level & BPF_LOG_STATS) {
13459 verbose(env, "verification time %lld usec\n",
13460 div_u64(env->verification_time, 1000));
13461 verbose(env, "stack depth ");
13462 for (i = 0; i < env->subprog_cnt; i++) {
13463 u32 depth = env->subprog_info[i].stack_depth;
13465 verbose(env, "%d", depth);
13466 if (i + 1 < env->subprog_cnt)
13469 verbose(env, "\n");
13471 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
13472 "total_states %d peak_states %d mark_read %d\n",
13473 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
13474 env->max_states_per_insn, env->total_states,
13475 env->peak_states, env->longest_mark_read_walk);
13478 static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
13480 const struct btf_type *t, *func_proto;
13481 const struct bpf_struct_ops *st_ops;
13482 const struct btf_member *member;
13483 struct bpf_prog *prog = env->prog;
13484 u32 btf_id, member_idx;
13487 if (!prog->gpl_compatible) {
13488 verbose(env, "struct ops programs must have a GPL compatible license\n");
13492 btf_id = prog->aux->attach_btf_id;
13493 st_ops = bpf_struct_ops_find(btf_id);
13495 verbose(env, "attach_btf_id %u is not a supported struct\n",
13501 member_idx = prog->expected_attach_type;
13502 if (member_idx >= btf_type_vlen(t)) {
13503 verbose(env, "attach to invalid member idx %u of struct %s\n",
13504 member_idx, st_ops->name);
13508 member = &btf_type_member(t)[member_idx];
13509 mname = btf_name_by_offset(btf_vmlinux, member->name_off);
13510 func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type,
13513 verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
13514 mname, member_idx, st_ops->name);
13518 if (st_ops->check_member) {
13519 int err = st_ops->check_member(t, member);
13522 verbose(env, "attach to unsupported member %s of struct %s\n",
13523 mname, st_ops->name);
13528 prog->aux->attach_func_proto = func_proto;
13529 prog->aux->attach_func_name = mname;
13530 env->ops = st_ops->verifier_ops;
13534 #define SECURITY_PREFIX "security_"
13536 static int check_attach_modify_return(unsigned long addr, const char *func_name)
13538 if (within_error_injection_list(addr) ||
13539 !strncmp(SECURITY_PREFIX, func_name, sizeof(SECURITY_PREFIX) - 1))
13545 /* list of non-sleepable functions that are otherwise on
13546 * ALLOW_ERROR_INJECTION list
13548 BTF_SET_START(btf_non_sleepable_error_inject)
13549 /* Three functions below can be called from sleepable and non-sleepable context.
13550 * Assume non-sleepable from bpf safety point of view.
13552 BTF_ID(func, __filemap_add_folio)
13553 BTF_ID(func, should_fail_alloc_page)
13554 BTF_ID(func, should_failslab)
13555 BTF_SET_END(btf_non_sleepable_error_inject)
13557 static int check_non_sleepable_error_inject(u32 btf_id)
13559 return btf_id_set_contains(&btf_non_sleepable_error_inject, btf_id);
13562 int bpf_check_attach_target(struct bpf_verifier_log *log,
13563 const struct bpf_prog *prog,
13564 const struct bpf_prog *tgt_prog,
13566 struct bpf_attach_target_info *tgt_info)
13568 bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
13569 const char prefix[] = "btf_trace_";
13570 int ret = 0, subprog = -1, i;
13571 const struct btf_type *t;
13572 bool conservative = true;
13578 bpf_log(log, "Tracing programs must provide btf_id\n");
13581 btf = tgt_prog ? tgt_prog->aux->btf : prog->aux->attach_btf;
13584 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
13587 t = btf_type_by_id(btf, btf_id);
13589 bpf_log(log, "attach_btf_id %u is invalid\n", btf_id);
13592 tname = btf_name_by_offset(btf, t->name_off);
13594 bpf_log(log, "attach_btf_id %u doesn't have a name\n", btf_id);
13598 struct bpf_prog_aux *aux = tgt_prog->aux;
13600 for (i = 0; i < aux->func_info_cnt; i++)
13601 if (aux->func_info[i].type_id == btf_id) {
13605 if (subprog == -1) {
13606 bpf_log(log, "Subprog %s doesn't exist\n", tname);
13609 conservative = aux->func_info_aux[subprog].unreliable;
13610 if (prog_extension) {
13611 if (conservative) {
13613 "Cannot replace static functions\n");
13616 if (!prog->jit_requested) {
13618 "Extension programs should be JITed\n");
13622 if (!tgt_prog->jited) {
13623 bpf_log(log, "Can attach to only JITed progs\n");
13626 if (tgt_prog->type == prog->type) {
13627 /* Cannot fentry/fexit another fentry/fexit program.
13628 * Cannot attach program extension to another extension.
13629 * It's ok to attach fentry/fexit to extension program.
13631 bpf_log(log, "Cannot recursively attach\n");
13634 if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
13636 (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
13637 tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
13638 /* Program extensions can extend all program types
13639 * except fentry/fexit. The reason is the following.
13640 * The fentry/fexit programs are used for performance
13641 * analysis, stats and can be attached to any program
13642 * type except themselves. When extension program is
13643 * replacing XDP function it is necessary to allow
13644 * performance analysis of all functions. Both original
13645 * XDP program and its program extension. Hence
13646 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
13647 * allowed. If extending of fentry/fexit was allowed it
13648 * would be possible to create long call chain
13649 * fentry->extension->fentry->extension beyond
13650 * reasonable stack size. Hence extending fentry is not
13653 bpf_log(log, "Cannot extend fentry/fexit\n");
13657 if (prog_extension) {
13658 bpf_log(log, "Cannot replace kernel functions\n");
13663 switch (prog->expected_attach_type) {
13664 case BPF_TRACE_RAW_TP:
13667 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
13670 if (!btf_type_is_typedef(t)) {
13671 bpf_log(log, "attach_btf_id %u is not a typedef\n",
13675 if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
13676 bpf_log(log, "attach_btf_id %u points to wrong type name %s\n",
13680 tname += sizeof(prefix) - 1;
13681 t = btf_type_by_id(btf, t->type);
13682 if (!btf_type_is_ptr(t))
13683 /* should never happen in valid vmlinux build */
13685 t = btf_type_by_id(btf, t->type);
13686 if (!btf_type_is_func_proto(t))
13687 /* should never happen in valid vmlinux build */
13691 case BPF_TRACE_ITER:
13692 if (!btf_type_is_func(t)) {
13693 bpf_log(log, "attach_btf_id %u is not a function\n",
13697 t = btf_type_by_id(btf, t->type);
13698 if (!btf_type_is_func_proto(t))
13700 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
13705 if (!prog_extension)
13708 case BPF_MODIFY_RETURN:
13710 case BPF_TRACE_FENTRY:
13711 case BPF_TRACE_FEXIT:
13712 if (!btf_type_is_func(t)) {
13713 bpf_log(log, "attach_btf_id %u is not a function\n",
13717 if (prog_extension &&
13718 btf_check_type_match(log, prog, btf, t))
13720 t = btf_type_by_id(btf, t->type);
13721 if (!btf_type_is_func_proto(t))
13724 if ((prog->aux->saved_dst_prog_type || prog->aux->saved_dst_attach_type) &&
13725 (!tgt_prog || prog->aux->saved_dst_prog_type != tgt_prog->type ||
13726 prog->aux->saved_dst_attach_type != tgt_prog->expected_attach_type))
13729 if (tgt_prog && conservative)
13732 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
13738 addr = (long) tgt_prog->bpf_func;
13740 addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
13742 addr = kallsyms_lookup_name(tname);
13745 "The address of function %s cannot be found\n",
13751 if (prog->aux->sleepable) {
13753 switch (prog->type) {
13754 case BPF_PROG_TYPE_TRACING:
13755 /* fentry/fexit/fmod_ret progs can be sleepable only if they are
13756 * attached to ALLOW_ERROR_INJECTION and are not in denylist.
13758 if (!check_non_sleepable_error_inject(btf_id) &&
13759 within_error_injection_list(addr))
13762 case BPF_PROG_TYPE_LSM:
13763 /* LSM progs check that they are attached to bpf_lsm_*() funcs.
13764 * Only some of them are sleepable.
13766 if (bpf_lsm_is_sleepable_hook(btf_id))
13773 bpf_log(log, "%s is not sleepable\n", tname);
13776 } else if (prog->expected_attach_type == BPF_MODIFY_RETURN) {
13778 bpf_log(log, "can't modify return codes of BPF programs\n");
13781 ret = check_attach_modify_return(addr, tname);
13783 bpf_log(log, "%s() is not modifiable\n", tname);
13790 tgt_info->tgt_addr = addr;
13791 tgt_info->tgt_name = tname;
13792 tgt_info->tgt_type = t;
13796 BTF_SET_START(btf_id_deny)
13799 BTF_ID(func, migrate_disable)
13800 BTF_ID(func, migrate_enable)
13802 #if !defined CONFIG_PREEMPT_RCU && !defined CONFIG_TINY_RCU
13803 BTF_ID(func, rcu_read_unlock_strict)
13805 BTF_SET_END(btf_id_deny)
13807 static int check_attach_btf_id(struct bpf_verifier_env *env)
13809 struct bpf_prog *prog = env->prog;
13810 struct bpf_prog *tgt_prog = prog->aux->dst_prog;
13811 struct bpf_attach_target_info tgt_info = {};
13812 u32 btf_id = prog->aux->attach_btf_id;
13813 struct bpf_trampoline *tr;
13817 if (prog->type == BPF_PROG_TYPE_SYSCALL) {
13818 if (prog->aux->sleepable)
13819 /* attach_btf_id checked to be zero already */
13821 verbose(env, "Syscall programs can only be sleepable\n");
13825 if (prog->aux->sleepable && prog->type != BPF_PROG_TYPE_TRACING &&
13826 prog->type != BPF_PROG_TYPE_LSM) {
13827 verbose(env, "Only fentry/fexit/fmod_ret and lsm programs can be sleepable\n");
13831 if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
13832 return check_struct_ops_btf_id(env);
13834 if (prog->type != BPF_PROG_TYPE_TRACING &&
13835 prog->type != BPF_PROG_TYPE_LSM &&
13836 prog->type != BPF_PROG_TYPE_EXT)
13839 ret = bpf_check_attach_target(&env->log, prog, tgt_prog, btf_id, &tgt_info);
13843 if (tgt_prog && prog->type == BPF_PROG_TYPE_EXT) {
13844 /* to make freplace equivalent to their targets, they need to
13845 * inherit env->ops and expected_attach_type for the rest of the
13848 env->ops = bpf_verifier_ops[tgt_prog->type];
13849 prog->expected_attach_type = tgt_prog->expected_attach_type;
13852 /* store info about the attachment target that will be used later */
13853 prog->aux->attach_func_proto = tgt_info.tgt_type;
13854 prog->aux->attach_func_name = tgt_info.tgt_name;
13857 prog->aux->saved_dst_prog_type = tgt_prog->type;
13858 prog->aux->saved_dst_attach_type = tgt_prog->expected_attach_type;
13861 if (prog->expected_attach_type == BPF_TRACE_RAW_TP) {
13862 prog->aux->attach_btf_trace = true;
13864 } else if (prog->expected_attach_type == BPF_TRACE_ITER) {
13865 if (!bpf_iter_prog_supported(prog))
13870 if (prog->type == BPF_PROG_TYPE_LSM) {
13871 ret = bpf_lsm_verify_prog(&env->log, prog);
13874 } else if (prog->type == BPF_PROG_TYPE_TRACING &&
13875 btf_id_set_contains(&btf_id_deny, btf_id)) {
13879 key = bpf_trampoline_compute_key(tgt_prog, prog->aux->attach_btf, btf_id);
13880 tr = bpf_trampoline_get(key, &tgt_info);
13884 prog->aux->dst_trampoline = tr;
13888 struct btf *bpf_get_btf_vmlinux(void)
13890 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
13891 mutex_lock(&bpf_verifier_lock);
13893 btf_vmlinux = btf_parse_vmlinux();
13894 mutex_unlock(&bpf_verifier_lock);
13896 return btf_vmlinux;
13899 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr, bpfptr_t uattr)
13901 u64 start_time = ktime_get_ns();
13902 struct bpf_verifier_env *env;
13903 struct bpf_verifier_log *log;
13904 int i, len, ret = -EINVAL;
13907 /* no program is valid */
13908 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
13911 /* 'struct bpf_verifier_env' can be global, but since it's not small,
13912 * allocate/free it every time bpf_check() is called
13914 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
13919 len = (*prog)->len;
13920 env->insn_aux_data =
13921 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
13923 if (!env->insn_aux_data)
13925 for (i = 0; i < len; i++)
13926 env->insn_aux_data[i].orig_idx = i;
13928 env->ops = bpf_verifier_ops[env->prog->type];
13929 env->fd_array = make_bpfptr(attr->fd_array, uattr.is_kernel);
13930 is_priv = bpf_capable();
13932 bpf_get_btf_vmlinux();
13934 /* grab the mutex to protect few globals used by verifier */
13936 mutex_lock(&bpf_verifier_lock);
13938 if (attr->log_level || attr->log_buf || attr->log_size) {
13939 /* user requested verbose verifier output
13940 * and supplied buffer to store the verification trace
13942 log->level = attr->log_level;
13943 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
13944 log->len_total = attr->log_size;
13947 /* log attributes have to be sane */
13948 if (log->len_total < 128 || log->len_total > UINT_MAX >> 2 ||
13949 !log->level || !log->ubuf || log->level & ~BPF_LOG_MASK)
13953 if (IS_ERR(btf_vmlinux)) {
13954 /* Either gcc or pahole or kernel are broken. */
13955 verbose(env, "in-kernel BTF is malformed\n");
13956 ret = PTR_ERR(btf_vmlinux);
13957 goto skip_full_check;
13960 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
13961 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
13962 env->strict_alignment = true;
13963 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
13964 env->strict_alignment = false;
13966 env->allow_ptr_leaks = bpf_allow_ptr_leaks();
13967 env->allow_uninit_stack = bpf_allow_uninit_stack();
13968 env->allow_ptr_to_map_access = bpf_allow_ptr_to_map_access();
13969 env->bypass_spec_v1 = bpf_bypass_spec_v1();
13970 env->bypass_spec_v4 = bpf_bypass_spec_v4();
13971 env->bpf_capable = bpf_capable();
13974 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
13976 env->explored_states = kvcalloc(state_htab_size(env),
13977 sizeof(struct bpf_verifier_state_list *),
13980 if (!env->explored_states)
13981 goto skip_full_check;
13983 ret = add_subprog_and_kfunc(env);
13985 goto skip_full_check;
13987 ret = check_subprogs(env);
13989 goto skip_full_check;
13991 ret = check_btf_info(env, attr, uattr);
13993 goto skip_full_check;
13995 ret = check_attach_btf_id(env);
13997 goto skip_full_check;
13999 ret = resolve_pseudo_ldimm64(env);
14001 goto skip_full_check;
14003 if (bpf_prog_is_dev_bound(env->prog->aux)) {
14004 ret = bpf_prog_offload_verifier_prep(env->prog);
14006 goto skip_full_check;
14009 ret = check_cfg(env);
14011 goto skip_full_check;
14013 ret = do_check_subprogs(env);
14014 ret = ret ?: do_check_main(env);
14016 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
14017 ret = bpf_prog_offload_finalize(env);
14020 kvfree(env->explored_states);
14023 ret = check_max_stack_depth(env);
14025 /* instruction rewrites happen after this point */
14028 opt_hard_wire_dead_code_branches(env);
14030 ret = opt_remove_dead_code(env);
14032 ret = opt_remove_nops(env);
14035 sanitize_dead_code(env);
14039 /* program is valid, convert *(u32*)(ctx + off) accesses */
14040 ret = convert_ctx_accesses(env);
14043 ret = do_misc_fixups(env);
14045 /* do 32-bit optimization after insn patching has done so those patched
14046 * insns could be handled correctly.
14048 if (ret == 0 && !bpf_prog_is_dev_bound(env->prog->aux)) {
14049 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
14050 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
14055 ret = fixup_call_args(env);
14057 env->verification_time = ktime_get_ns() - start_time;
14058 print_verification_stats(env);
14059 env->prog->aux->verified_insns = env->insn_processed;
14061 if (log->level && bpf_verifier_log_full(log))
14063 if (log->level && !log->ubuf) {
14065 goto err_release_maps;
14069 goto err_release_maps;
14071 if (env->used_map_cnt) {
14072 /* if program passed verifier, update used_maps in bpf_prog_info */
14073 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
14074 sizeof(env->used_maps[0]),
14077 if (!env->prog->aux->used_maps) {
14079 goto err_release_maps;
14082 memcpy(env->prog->aux->used_maps, env->used_maps,
14083 sizeof(env->used_maps[0]) * env->used_map_cnt);
14084 env->prog->aux->used_map_cnt = env->used_map_cnt;
14086 if (env->used_btf_cnt) {
14087 /* if program passed verifier, update used_btfs in bpf_prog_aux */
14088 env->prog->aux->used_btfs = kmalloc_array(env->used_btf_cnt,
14089 sizeof(env->used_btfs[0]),
14091 if (!env->prog->aux->used_btfs) {
14093 goto err_release_maps;
14096 memcpy(env->prog->aux->used_btfs, env->used_btfs,
14097 sizeof(env->used_btfs[0]) * env->used_btf_cnt);
14098 env->prog->aux->used_btf_cnt = env->used_btf_cnt;
14100 if (env->used_map_cnt || env->used_btf_cnt) {
14101 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
14102 * bpf_ld_imm64 instructions
14104 convert_pseudo_ld_imm64(env);
14107 adjust_btf_func(env);
14110 if (!env->prog->aux->used_maps)
14111 /* if we didn't copy map pointers into bpf_prog_info, release
14112 * them now. Otherwise free_used_maps() will release them.
14115 if (!env->prog->aux->used_btfs)
14118 /* extension progs temporarily inherit the attach_type of their targets
14119 for verification purposes, so set it back to zero before returning
14121 if (env->prog->type == BPF_PROG_TYPE_EXT)
14122 env->prog->expected_attach_type = 0;
14127 mutex_unlock(&bpf_verifier_lock);
14128 vfree(env->insn_aux_data);