1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
3 * Copyright (c) 2016 Facebook
4 * Copyright (c) 2018 Covalent IO, Inc. http://covalent.io
6 #include <uapi/linux/btf.h>
7 #include <linux/kernel.h>
8 #include <linux/types.h>
9 #include <linux/slab.h>
10 #include <linux/bpf.h>
11 #include <linux/btf.h>
12 #include <linux/bpf_verifier.h>
13 #include <linux/filter.h>
14 #include <net/netlink.h>
15 #include <linux/file.h>
16 #include <linux/vmalloc.h>
17 #include <linux/stringify.h>
18 #include <linux/bsearch.h>
19 #include <linux/sort.h>
20 #include <linux/perf_event.h>
21 #include <linux/ctype.h>
22 #include <linux/error-injection.h>
23 #include <linux/bpf_lsm.h>
24 #include <linux/btf_ids.h>
28 static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
29 #define BPF_PROG_TYPE(_id, _name, prog_ctx_type, kern_ctx_type) \
30 [_id] = & _name ## _verifier_ops,
31 #define BPF_MAP_TYPE(_id, _ops)
32 #define BPF_LINK_TYPE(_id, _name)
33 #include <linux/bpf_types.h>
39 /* bpf_check() is a static code analyzer that walks eBPF program
40 * instruction by instruction and updates register/stack state.
41 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
43 * The first pass is depth-first-search to check that the program is a DAG.
44 * It rejects the following programs:
45 * - larger than BPF_MAXINSNS insns
46 * - if loop is present (detected via back-edge)
47 * - unreachable insns exist (shouldn't be a forest. program = one function)
48 * - out of bounds or malformed jumps
49 * The second pass is all possible path descent from the 1st insn.
50 * Since it's analyzing all paths through the program, the length of the
51 * analysis is limited to 64k insn, which may be hit even if total number of
52 * insn is less then 4K, but there are too many branches that change stack/regs.
53 * Number of 'branches to be analyzed' is limited to 1k
55 * On entry to each instruction, each register has a type, and the instruction
56 * changes the types of the registers depending on instruction semantics.
57 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
60 * All registers are 64-bit.
61 * R0 - return register
62 * R1-R5 argument passing registers
63 * R6-R9 callee saved registers
64 * R10 - frame pointer read-only
66 * At the start of BPF program the register R1 contains a pointer to bpf_context
67 * and has type PTR_TO_CTX.
69 * Verifier tracks arithmetic operations on pointers in case:
70 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
71 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
72 * 1st insn copies R10 (which has FRAME_PTR) type into R1
73 * and 2nd arithmetic instruction is pattern matched to recognize
74 * that it wants to construct a pointer to some element within stack.
75 * So after 2nd insn, the register R1 has type PTR_TO_STACK
76 * (and -20 constant is saved for further stack bounds checking).
77 * Meaning that this reg is a pointer to stack plus known immediate constant.
79 * Most of the time the registers have SCALAR_VALUE type, which
80 * means the register has some value, but it's not a valid pointer.
81 * (like pointer plus pointer becomes SCALAR_VALUE type)
83 * When verifier sees load or store instructions the type of base register
84 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK, PTR_TO_SOCKET. These are
85 * four pointer types recognized by check_mem_access() function.
87 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
88 * and the range of [ptr, ptr + map's value_size) is accessible.
90 * registers used to pass values to function calls are checked against
91 * function argument constraints.
93 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
94 * It means that the register type passed to this function must be
95 * PTR_TO_STACK and it will be used inside the function as
96 * 'pointer to map element key'
98 * For example the argument constraints for bpf_map_lookup_elem():
99 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
100 * .arg1_type = ARG_CONST_MAP_PTR,
101 * .arg2_type = ARG_PTR_TO_MAP_KEY,
103 * ret_type says that this function returns 'pointer to map elem value or null'
104 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
105 * 2nd argument should be a pointer to stack, which will be used inside
106 * the helper function as a pointer to map element key.
108 * On the kernel side the helper function looks like:
109 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
111 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
112 * void *key = (void *) (unsigned long) r2;
115 * here kernel can access 'key' and 'map' pointers safely, knowing that
116 * [key, key + map->key_size) bytes are valid and were initialized on
117 * the stack of eBPF program.
120 * Corresponding eBPF program may look like:
121 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
122 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
123 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
124 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
125 * here verifier looks at prototype of map_lookup_elem() and sees:
126 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
127 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
129 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
130 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
131 * and were initialized prior to this call.
132 * If it's ok, then verifier allows this BPF_CALL insn and looks at
133 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
134 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
135 * returns either pointer to map value or NULL.
137 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
138 * insn, the register holding that pointer in the true branch changes state to
139 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
140 * branch. See check_cond_jmp_op().
142 * After the call R0 is set to return type of the function and registers R1-R5
143 * are set to NOT_INIT to indicate that they are no longer readable.
145 * The following reference types represent a potential reference to a kernel
146 * resource which, after first being allocated, must be checked and freed by
148 * - PTR_TO_SOCKET_OR_NULL, PTR_TO_SOCKET
150 * When the verifier sees a helper call return a reference type, it allocates a
151 * pointer id for the reference and stores it in the current function state.
152 * Similar to the way that PTR_TO_MAP_VALUE_OR_NULL is converted into
153 * PTR_TO_MAP_VALUE, PTR_TO_SOCKET_OR_NULL becomes PTR_TO_SOCKET when the type
154 * passes through a NULL-check conditional. For the branch wherein the state is
155 * changed to CONST_IMM, the verifier releases the reference.
157 * For each helper function that allocates a reference, such as
158 * bpf_sk_lookup_tcp(), there is a corresponding release function, such as
159 * bpf_sk_release(). When a reference type passes into the release function,
160 * the verifier also releases the reference. If any unchecked or unreleased
161 * reference remains at the end of the program, the verifier rejects it.
164 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
165 struct bpf_verifier_stack_elem {
166 /* verifer state is 'st'
167 * before processing instruction 'insn_idx'
168 * and after processing instruction 'prev_insn_idx'
170 struct bpf_verifier_state st;
173 struct bpf_verifier_stack_elem *next;
174 /* length of verifier log at the time this state was pushed on stack */
178 #define BPF_COMPLEXITY_LIMIT_JMP_SEQ 8192
179 #define BPF_COMPLEXITY_LIMIT_STATES 64
181 #define BPF_MAP_KEY_POISON (1ULL << 63)
182 #define BPF_MAP_KEY_SEEN (1ULL << 62)
184 #define BPF_MAP_PTR_UNPRIV 1UL
185 #define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \
186 POISON_POINTER_DELTA))
187 #define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
189 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
191 return BPF_MAP_PTR(aux->map_ptr_state) == BPF_MAP_PTR_POISON;
194 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
196 return aux->map_ptr_state & BPF_MAP_PTR_UNPRIV;
199 static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
200 const struct bpf_map *map, bool unpriv)
202 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON & BPF_MAP_PTR_UNPRIV);
203 unpriv |= bpf_map_ptr_unpriv(aux);
204 aux->map_ptr_state = (unsigned long)map |
205 (unpriv ? BPF_MAP_PTR_UNPRIV : 0UL);
208 static bool bpf_map_key_poisoned(const struct bpf_insn_aux_data *aux)
210 return aux->map_key_state & BPF_MAP_KEY_POISON;
213 static bool bpf_map_key_unseen(const struct bpf_insn_aux_data *aux)
215 return !(aux->map_key_state & BPF_MAP_KEY_SEEN);
218 static u64 bpf_map_key_immediate(const struct bpf_insn_aux_data *aux)
220 return aux->map_key_state & ~(BPF_MAP_KEY_SEEN | BPF_MAP_KEY_POISON);
223 static void bpf_map_key_store(struct bpf_insn_aux_data *aux, u64 state)
225 bool poisoned = bpf_map_key_poisoned(aux);
227 aux->map_key_state = state | BPF_MAP_KEY_SEEN |
228 (poisoned ? BPF_MAP_KEY_POISON : 0ULL);
231 static bool bpf_pseudo_call(const struct bpf_insn *insn)
233 return insn->code == (BPF_JMP | BPF_CALL) &&
234 insn->src_reg == BPF_PSEUDO_CALL;
237 static bool bpf_pseudo_kfunc_call(const struct bpf_insn *insn)
239 return insn->code == (BPF_JMP | BPF_CALL) &&
240 insn->src_reg == BPF_PSEUDO_KFUNC_CALL;
243 static bool bpf_pseudo_func(const struct bpf_insn *insn)
245 return insn->code == (BPF_LD | BPF_IMM | BPF_DW) &&
246 insn->src_reg == BPF_PSEUDO_FUNC;
249 struct bpf_call_arg_meta {
250 struct bpf_map *map_ptr;
266 struct btf *btf_vmlinux;
268 static DEFINE_MUTEX(bpf_verifier_lock);
270 static const struct bpf_line_info *
271 find_linfo(const struct bpf_verifier_env *env, u32 insn_off)
273 const struct bpf_line_info *linfo;
274 const struct bpf_prog *prog;
278 nr_linfo = prog->aux->nr_linfo;
280 if (!nr_linfo || insn_off >= prog->len)
283 linfo = prog->aux->linfo;
284 for (i = 1; i < nr_linfo; i++)
285 if (insn_off < linfo[i].insn_off)
288 return &linfo[i - 1];
291 void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
296 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
298 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
299 "verifier log line truncated - local buffer too short\n");
301 n = min(log->len_total - log->len_used - 1, n);
304 if (log->level == BPF_LOG_KERNEL) {
305 pr_err("BPF:%s\n", log->kbuf);
308 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
314 static void bpf_vlog_reset(struct bpf_verifier_log *log, u32 new_pos)
318 if (!bpf_verifier_log_needed(log))
321 log->len_used = new_pos;
322 if (put_user(zero, log->ubuf + new_pos))
326 /* log_level controls verbosity level of eBPF verifier.
327 * bpf_verifier_log_write() is used to dump the verification trace to the log,
328 * so the user can figure out what's wrong with the program
330 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
331 const char *fmt, ...)
335 if (!bpf_verifier_log_needed(&env->log))
339 bpf_verifier_vlog(&env->log, fmt, args);
342 EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
344 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
346 struct bpf_verifier_env *env = private_data;
349 if (!bpf_verifier_log_needed(&env->log))
353 bpf_verifier_vlog(&env->log, fmt, args);
357 __printf(2, 3) void bpf_log(struct bpf_verifier_log *log,
358 const char *fmt, ...)
362 if (!bpf_verifier_log_needed(log))
366 bpf_verifier_vlog(log, fmt, args);
370 static const char *ltrim(const char *s)
378 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env,
380 const char *prefix_fmt, ...)
382 const struct bpf_line_info *linfo;
384 if (!bpf_verifier_log_needed(&env->log))
387 linfo = find_linfo(env, insn_off);
388 if (!linfo || linfo == env->prev_linfo)
394 va_start(args, prefix_fmt);
395 bpf_verifier_vlog(&env->log, prefix_fmt, args);
400 ltrim(btf_name_by_offset(env->prog->aux->btf,
403 env->prev_linfo = linfo;
406 static void verbose_invalid_scalar(struct bpf_verifier_env *env,
407 struct bpf_reg_state *reg,
408 struct tnum *range, const char *ctx,
409 const char *reg_name)
413 verbose(env, "At %s the register %s ", ctx, reg_name);
414 if (!tnum_is_unknown(reg->var_off)) {
415 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
416 verbose(env, "has value %s", tn_buf);
418 verbose(env, "has unknown scalar value");
420 tnum_strn(tn_buf, sizeof(tn_buf), *range);
421 verbose(env, " should have been in %s\n", tn_buf);
424 static bool type_is_pkt_pointer(enum bpf_reg_type type)
426 return type == PTR_TO_PACKET ||
427 type == PTR_TO_PACKET_META;
430 static bool type_is_sk_pointer(enum bpf_reg_type type)
432 return type == PTR_TO_SOCKET ||
433 type == PTR_TO_SOCK_COMMON ||
434 type == PTR_TO_TCP_SOCK ||
435 type == PTR_TO_XDP_SOCK;
438 static bool reg_type_not_null(enum bpf_reg_type type)
440 return type == PTR_TO_SOCKET ||
441 type == PTR_TO_TCP_SOCK ||
442 type == PTR_TO_MAP_VALUE ||
443 type == PTR_TO_MAP_KEY ||
444 type == PTR_TO_SOCK_COMMON;
447 static bool reg_type_may_be_null(enum bpf_reg_type type)
449 return type == PTR_TO_MAP_VALUE_OR_NULL ||
450 type == PTR_TO_SOCKET_OR_NULL ||
451 type == PTR_TO_SOCK_COMMON_OR_NULL ||
452 type == PTR_TO_TCP_SOCK_OR_NULL ||
453 type == PTR_TO_BTF_ID_OR_NULL ||
454 type == PTR_TO_MEM_OR_NULL ||
455 type == PTR_TO_RDONLY_BUF_OR_NULL ||
456 type == PTR_TO_RDWR_BUF_OR_NULL;
459 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
461 return reg->type == PTR_TO_MAP_VALUE &&
462 map_value_has_spin_lock(reg->map_ptr);
465 static bool reg_type_may_be_refcounted_or_null(enum bpf_reg_type type)
467 return type == PTR_TO_SOCKET ||
468 type == PTR_TO_SOCKET_OR_NULL ||
469 type == PTR_TO_TCP_SOCK ||
470 type == PTR_TO_TCP_SOCK_OR_NULL ||
471 type == PTR_TO_MEM ||
472 type == PTR_TO_MEM_OR_NULL;
475 static bool arg_type_may_be_refcounted(enum bpf_arg_type type)
477 return type == ARG_PTR_TO_SOCK_COMMON;
480 static bool arg_type_may_be_null(enum bpf_arg_type type)
482 return type == ARG_PTR_TO_MAP_VALUE_OR_NULL ||
483 type == ARG_PTR_TO_MEM_OR_NULL ||
484 type == ARG_PTR_TO_CTX_OR_NULL ||
485 type == ARG_PTR_TO_SOCKET_OR_NULL ||
486 type == ARG_PTR_TO_ALLOC_MEM_OR_NULL ||
487 type == ARG_PTR_TO_STACK_OR_NULL;
490 /* Determine whether the function releases some resources allocated by another
491 * function call. The first reference type argument will be assumed to be
492 * released by release_reference().
494 static bool is_release_function(enum bpf_func_id func_id)
496 return func_id == BPF_FUNC_sk_release ||
497 func_id == BPF_FUNC_ringbuf_submit ||
498 func_id == BPF_FUNC_ringbuf_discard;
501 static bool may_be_acquire_function(enum bpf_func_id func_id)
503 return func_id == BPF_FUNC_sk_lookup_tcp ||
504 func_id == BPF_FUNC_sk_lookup_udp ||
505 func_id == BPF_FUNC_skc_lookup_tcp ||
506 func_id == BPF_FUNC_map_lookup_elem ||
507 func_id == BPF_FUNC_ringbuf_reserve;
510 static bool is_acquire_function(enum bpf_func_id func_id,
511 const struct bpf_map *map)
513 enum bpf_map_type map_type = map ? map->map_type : BPF_MAP_TYPE_UNSPEC;
515 if (func_id == BPF_FUNC_sk_lookup_tcp ||
516 func_id == BPF_FUNC_sk_lookup_udp ||
517 func_id == BPF_FUNC_skc_lookup_tcp ||
518 func_id == BPF_FUNC_ringbuf_reserve)
521 if (func_id == BPF_FUNC_map_lookup_elem &&
522 (map_type == BPF_MAP_TYPE_SOCKMAP ||
523 map_type == BPF_MAP_TYPE_SOCKHASH))
529 static bool is_ptr_cast_function(enum bpf_func_id func_id)
531 return func_id == BPF_FUNC_tcp_sock ||
532 func_id == BPF_FUNC_sk_fullsock ||
533 func_id == BPF_FUNC_skc_to_tcp_sock ||
534 func_id == BPF_FUNC_skc_to_tcp6_sock ||
535 func_id == BPF_FUNC_skc_to_udp6_sock ||
536 func_id == BPF_FUNC_skc_to_tcp_timewait_sock ||
537 func_id == BPF_FUNC_skc_to_tcp_request_sock;
540 static bool is_cmpxchg_insn(const struct bpf_insn *insn)
542 return BPF_CLASS(insn->code) == BPF_STX &&
543 BPF_MODE(insn->code) == BPF_ATOMIC &&
544 insn->imm == BPF_CMPXCHG;
547 /* string representation of 'enum bpf_reg_type' */
548 static const char * const reg_type_str[] = {
550 [SCALAR_VALUE] = "inv",
551 [PTR_TO_CTX] = "ctx",
552 [CONST_PTR_TO_MAP] = "map_ptr",
553 [PTR_TO_MAP_VALUE] = "map_value",
554 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
555 [PTR_TO_STACK] = "fp",
556 [PTR_TO_PACKET] = "pkt",
557 [PTR_TO_PACKET_META] = "pkt_meta",
558 [PTR_TO_PACKET_END] = "pkt_end",
559 [PTR_TO_FLOW_KEYS] = "flow_keys",
560 [PTR_TO_SOCKET] = "sock",
561 [PTR_TO_SOCKET_OR_NULL] = "sock_or_null",
562 [PTR_TO_SOCK_COMMON] = "sock_common",
563 [PTR_TO_SOCK_COMMON_OR_NULL] = "sock_common_or_null",
564 [PTR_TO_TCP_SOCK] = "tcp_sock",
565 [PTR_TO_TCP_SOCK_OR_NULL] = "tcp_sock_or_null",
566 [PTR_TO_TP_BUFFER] = "tp_buffer",
567 [PTR_TO_XDP_SOCK] = "xdp_sock",
568 [PTR_TO_BTF_ID] = "ptr_",
569 [PTR_TO_BTF_ID_OR_NULL] = "ptr_or_null_",
570 [PTR_TO_PERCPU_BTF_ID] = "percpu_ptr_",
571 [PTR_TO_MEM] = "mem",
572 [PTR_TO_MEM_OR_NULL] = "mem_or_null",
573 [PTR_TO_RDONLY_BUF] = "rdonly_buf",
574 [PTR_TO_RDONLY_BUF_OR_NULL] = "rdonly_buf_or_null",
575 [PTR_TO_RDWR_BUF] = "rdwr_buf",
576 [PTR_TO_RDWR_BUF_OR_NULL] = "rdwr_buf_or_null",
577 [PTR_TO_FUNC] = "func",
578 [PTR_TO_MAP_KEY] = "map_key",
581 static char slot_type_char[] = {
582 [STACK_INVALID] = '?',
588 static void print_liveness(struct bpf_verifier_env *env,
589 enum bpf_reg_liveness live)
591 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
593 if (live & REG_LIVE_READ)
595 if (live & REG_LIVE_WRITTEN)
597 if (live & REG_LIVE_DONE)
601 static struct bpf_func_state *func(struct bpf_verifier_env *env,
602 const struct bpf_reg_state *reg)
604 struct bpf_verifier_state *cur = env->cur_state;
606 return cur->frame[reg->frameno];
609 static const char *kernel_type_name(const struct btf* btf, u32 id)
611 return btf_name_by_offset(btf, btf_type_by_id(btf, id)->name_off);
614 static void print_verifier_state(struct bpf_verifier_env *env,
615 const struct bpf_func_state *state)
617 const struct bpf_reg_state *reg;
622 verbose(env, " frame%d:", state->frameno);
623 for (i = 0; i < MAX_BPF_REG; i++) {
624 reg = &state->regs[i];
628 verbose(env, " R%d", i);
629 print_liveness(env, reg->live);
630 verbose(env, "=%s", reg_type_str[t]);
631 if (t == SCALAR_VALUE && reg->precise)
633 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
634 tnum_is_const(reg->var_off)) {
635 /* reg->off should be 0 for SCALAR_VALUE */
636 verbose(env, "%lld", reg->var_off.value + reg->off);
638 if (t == PTR_TO_BTF_ID ||
639 t == PTR_TO_BTF_ID_OR_NULL ||
640 t == PTR_TO_PERCPU_BTF_ID)
641 verbose(env, "%s", kernel_type_name(reg->btf, reg->btf_id));
642 verbose(env, "(id=%d", reg->id);
643 if (reg_type_may_be_refcounted_or_null(t))
644 verbose(env, ",ref_obj_id=%d", reg->ref_obj_id);
645 if (t != SCALAR_VALUE)
646 verbose(env, ",off=%d", reg->off);
647 if (type_is_pkt_pointer(t))
648 verbose(env, ",r=%d", reg->range);
649 else if (t == CONST_PTR_TO_MAP ||
650 t == PTR_TO_MAP_KEY ||
651 t == PTR_TO_MAP_VALUE ||
652 t == PTR_TO_MAP_VALUE_OR_NULL)
653 verbose(env, ",ks=%d,vs=%d",
654 reg->map_ptr->key_size,
655 reg->map_ptr->value_size);
656 if (tnum_is_const(reg->var_off)) {
657 /* Typically an immediate SCALAR_VALUE, but
658 * could be a pointer whose offset is too big
661 verbose(env, ",imm=%llx", reg->var_off.value);
663 if (reg->smin_value != reg->umin_value &&
664 reg->smin_value != S64_MIN)
665 verbose(env, ",smin_value=%lld",
666 (long long)reg->smin_value);
667 if (reg->smax_value != reg->umax_value &&
668 reg->smax_value != S64_MAX)
669 verbose(env, ",smax_value=%lld",
670 (long long)reg->smax_value);
671 if (reg->umin_value != 0)
672 verbose(env, ",umin_value=%llu",
673 (unsigned long long)reg->umin_value);
674 if (reg->umax_value != U64_MAX)
675 verbose(env, ",umax_value=%llu",
676 (unsigned long long)reg->umax_value);
677 if (!tnum_is_unknown(reg->var_off)) {
680 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
681 verbose(env, ",var_off=%s", tn_buf);
683 if (reg->s32_min_value != reg->smin_value &&
684 reg->s32_min_value != S32_MIN)
685 verbose(env, ",s32_min_value=%d",
686 (int)(reg->s32_min_value));
687 if (reg->s32_max_value != reg->smax_value &&
688 reg->s32_max_value != S32_MAX)
689 verbose(env, ",s32_max_value=%d",
690 (int)(reg->s32_max_value));
691 if (reg->u32_min_value != reg->umin_value &&
692 reg->u32_min_value != U32_MIN)
693 verbose(env, ",u32_min_value=%d",
694 (int)(reg->u32_min_value));
695 if (reg->u32_max_value != reg->umax_value &&
696 reg->u32_max_value != U32_MAX)
697 verbose(env, ",u32_max_value=%d",
698 (int)(reg->u32_max_value));
703 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
704 char types_buf[BPF_REG_SIZE + 1];
708 for (j = 0; j < BPF_REG_SIZE; j++) {
709 if (state->stack[i].slot_type[j] != STACK_INVALID)
711 types_buf[j] = slot_type_char[
712 state->stack[i].slot_type[j]];
714 types_buf[BPF_REG_SIZE] = 0;
717 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
718 print_liveness(env, state->stack[i].spilled_ptr.live);
719 if (state->stack[i].slot_type[0] == STACK_SPILL) {
720 reg = &state->stack[i].spilled_ptr;
722 verbose(env, "=%s", reg_type_str[t]);
723 if (t == SCALAR_VALUE && reg->precise)
725 if (t == SCALAR_VALUE && tnum_is_const(reg->var_off))
726 verbose(env, "%lld", reg->var_off.value + reg->off);
728 verbose(env, "=%s", types_buf);
731 if (state->acquired_refs && state->refs[0].id) {
732 verbose(env, " refs=%d", state->refs[0].id);
733 for (i = 1; i < state->acquired_refs; i++)
734 if (state->refs[i].id)
735 verbose(env, ",%d", state->refs[i].id);
740 /* copy array src of length n * size bytes to dst. dst is reallocated if it's too
741 * small to hold src. This is different from krealloc since we don't want to preserve
742 * the contents of dst.
744 * Leaves dst untouched if src is NULL or length is zero. Returns NULL if memory could
747 static void *copy_array(void *dst, const void *src, size_t n, size_t size, gfp_t flags)
751 if (ZERO_OR_NULL_PTR(src))
754 if (unlikely(check_mul_overflow(n, size, &bytes)))
757 if (ksize(dst) < bytes) {
759 dst = kmalloc_track_caller(bytes, flags);
764 memcpy(dst, src, bytes);
766 return dst ? dst : ZERO_SIZE_PTR;
769 /* resize an array from old_n items to new_n items. the array is reallocated if it's too
770 * small to hold new_n items. new items are zeroed out if the array grows.
772 * Contrary to krealloc_array, does not free arr if new_n is zero.
774 static void *realloc_array(void *arr, size_t old_n, size_t new_n, size_t size)
776 if (!new_n || old_n == new_n)
779 arr = krealloc_array(arr, new_n, size, GFP_KERNEL);
784 memset(arr + old_n * size, 0, (new_n - old_n) * size);
787 return arr ? arr : ZERO_SIZE_PTR;
790 static int copy_reference_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
792 dst->refs = copy_array(dst->refs, src->refs, src->acquired_refs,
793 sizeof(struct bpf_reference_state), GFP_KERNEL);
797 dst->acquired_refs = src->acquired_refs;
801 static int copy_stack_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
803 size_t n = src->allocated_stack / BPF_REG_SIZE;
805 dst->stack = copy_array(dst->stack, src->stack, n, sizeof(struct bpf_stack_state),
810 dst->allocated_stack = src->allocated_stack;
814 static int resize_reference_state(struct bpf_func_state *state, size_t n)
816 state->refs = realloc_array(state->refs, state->acquired_refs, n,
817 sizeof(struct bpf_reference_state));
821 state->acquired_refs = n;
825 static int grow_stack_state(struct bpf_func_state *state, int size)
827 size_t old_n = state->allocated_stack / BPF_REG_SIZE, n = size / BPF_REG_SIZE;
832 state->stack = realloc_array(state->stack, old_n, n, sizeof(struct bpf_stack_state));
836 state->allocated_stack = size;
840 /* Acquire a pointer id from the env and update the state->refs to include
841 * this new pointer reference.
842 * On success, returns a valid pointer id to associate with the register
843 * On failure, returns a negative errno.
845 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
847 struct bpf_func_state *state = cur_func(env);
848 int new_ofs = state->acquired_refs;
851 err = resize_reference_state(state, state->acquired_refs + 1);
855 state->refs[new_ofs].id = id;
856 state->refs[new_ofs].insn_idx = insn_idx;
861 /* release function corresponding to acquire_reference_state(). Idempotent. */
862 static int release_reference_state(struct bpf_func_state *state, int ptr_id)
866 last_idx = state->acquired_refs - 1;
867 for (i = 0; i < state->acquired_refs; i++) {
868 if (state->refs[i].id == ptr_id) {
869 if (last_idx && i != last_idx)
870 memcpy(&state->refs[i], &state->refs[last_idx],
871 sizeof(*state->refs));
872 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
873 state->acquired_refs--;
880 static void free_func_state(struct bpf_func_state *state)
889 static void clear_jmp_history(struct bpf_verifier_state *state)
891 kfree(state->jmp_history);
892 state->jmp_history = NULL;
893 state->jmp_history_cnt = 0;
896 static void free_verifier_state(struct bpf_verifier_state *state,
901 for (i = 0; i <= state->curframe; i++) {
902 free_func_state(state->frame[i]);
903 state->frame[i] = NULL;
905 clear_jmp_history(state);
910 /* copy verifier state from src to dst growing dst stack space
911 * when necessary to accommodate larger src stack
913 static int copy_func_state(struct bpf_func_state *dst,
914 const struct bpf_func_state *src)
918 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
919 err = copy_reference_state(dst, src);
922 return copy_stack_state(dst, src);
925 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
926 const struct bpf_verifier_state *src)
928 struct bpf_func_state *dst;
931 dst_state->jmp_history = copy_array(dst_state->jmp_history, src->jmp_history,
932 src->jmp_history_cnt, sizeof(struct bpf_idx_pair),
934 if (!dst_state->jmp_history)
936 dst_state->jmp_history_cnt = src->jmp_history_cnt;
938 /* if dst has more stack frames then src frame, free them */
939 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
940 free_func_state(dst_state->frame[i]);
941 dst_state->frame[i] = NULL;
943 dst_state->speculative = src->speculative;
944 dst_state->curframe = src->curframe;
945 dst_state->active_spin_lock = src->active_spin_lock;
946 dst_state->branches = src->branches;
947 dst_state->parent = src->parent;
948 dst_state->first_insn_idx = src->first_insn_idx;
949 dst_state->last_insn_idx = src->last_insn_idx;
950 for (i = 0; i <= src->curframe; i++) {
951 dst = dst_state->frame[i];
953 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
956 dst_state->frame[i] = dst;
958 err = copy_func_state(dst, src->frame[i]);
965 static void update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
968 u32 br = --st->branches;
970 /* WARN_ON(br > 1) technically makes sense here,
971 * but see comment in push_stack(), hence:
973 WARN_ONCE((int)br < 0,
974 "BUG update_branch_counts:branches_to_explore=%d\n",
982 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
983 int *insn_idx, bool pop_log)
985 struct bpf_verifier_state *cur = env->cur_state;
986 struct bpf_verifier_stack_elem *elem, *head = env->head;
989 if (env->head == NULL)
993 err = copy_verifier_state(cur, &head->st);
998 bpf_vlog_reset(&env->log, head->log_pos);
1000 *insn_idx = head->insn_idx;
1002 *prev_insn_idx = head->prev_insn_idx;
1004 free_verifier_state(&head->st, false);
1011 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
1012 int insn_idx, int prev_insn_idx,
1015 struct bpf_verifier_state *cur = env->cur_state;
1016 struct bpf_verifier_stack_elem *elem;
1019 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
1023 elem->insn_idx = insn_idx;
1024 elem->prev_insn_idx = prev_insn_idx;
1025 elem->next = env->head;
1026 elem->log_pos = env->log.len_used;
1029 err = copy_verifier_state(&elem->st, cur);
1032 elem->st.speculative |= speculative;
1033 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
1034 verbose(env, "The sequence of %d jumps is too complex.\n",
1038 if (elem->st.parent) {
1039 ++elem->st.parent->branches;
1040 /* WARN_ON(branches > 2) technically makes sense here,
1042 * 1. speculative states will bump 'branches' for non-branch
1044 * 2. is_state_visited() heuristics may decide not to create
1045 * a new state for a sequence of branches and all such current
1046 * and cloned states will be pointing to a single parent state
1047 * which might have large 'branches' count.
1052 free_verifier_state(env->cur_state, true);
1053 env->cur_state = NULL;
1054 /* pop all elements and return */
1055 while (!pop_stack(env, NULL, NULL, false));
1059 #define CALLER_SAVED_REGS 6
1060 static const int caller_saved[CALLER_SAVED_REGS] = {
1061 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
1064 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1065 struct bpf_reg_state *reg);
1067 /* This helper doesn't clear reg->id */
1068 static void ___mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1070 reg->var_off = tnum_const(imm);
1071 reg->smin_value = (s64)imm;
1072 reg->smax_value = (s64)imm;
1073 reg->umin_value = imm;
1074 reg->umax_value = imm;
1076 reg->s32_min_value = (s32)imm;
1077 reg->s32_max_value = (s32)imm;
1078 reg->u32_min_value = (u32)imm;
1079 reg->u32_max_value = (u32)imm;
1082 /* Mark the unknown part of a register (variable offset or scalar value) as
1083 * known to have the value @imm.
1085 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1087 /* Clear id, off, and union(map_ptr, range) */
1088 memset(((u8 *)reg) + sizeof(reg->type), 0,
1089 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
1090 ___mark_reg_known(reg, imm);
1093 static void __mark_reg32_known(struct bpf_reg_state *reg, u64 imm)
1095 reg->var_off = tnum_const_subreg(reg->var_off, imm);
1096 reg->s32_min_value = (s32)imm;
1097 reg->s32_max_value = (s32)imm;
1098 reg->u32_min_value = (u32)imm;
1099 reg->u32_max_value = (u32)imm;
1102 /* Mark the 'variable offset' part of a register as zero. This should be
1103 * used only on registers holding a pointer type.
1105 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
1107 __mark_reg_known(reg, 0);
1110 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
1112 __mark_reg_known(reg, 0);
1113 reg->type = SCALAR_VALUE;
1116 static void mark_reg_known_zero(struct bpf_verifier_env *env,
1117 struct bpf_reg_state *regs, u32 regno)
1119 if (WARN_ON(regno >= MAX_BPF_REG)) {
1120 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
1121 /* Something bad happened, let's kill all regs */
1122 for (regno = 0; regno < MAX_BPF_REG; regno++)
1123 __mark_reg_not_init(env, regs + regno);
1126 __mark_reg_known_zero(regs + regno);
1129 static void mark_ptr_not_null_reg(struct bpf_reg_state *reg)
1131 switch (reg->type) {
1132 case PTR_TO_MAP_VALUE_OR_NULL: {
1133 const struct bpf_map *map = reg->map_ptr;
1135 if (map->inner_map_meta) {
1136 reg->type = CONST_PTR_TO_MAP;
1137 reg->map_ptr = map->inner_map_meta;
1138 } else if (map->map_type == BPF_MAP_TYPE_XSKMAP) {
1139 reg->type = PTR_TO_XDP_SOCK;
1140 } else if (map->map_type == BPF_MAP_TYPE_SOCKMAP ||
1141 map->map_type == BPF_MAP_TYPE_SOCKHASH) {
1142 reg->type = PTR_TO_SOCKET;
1144 reg->type = PTR_TO_MAP_VALUE;
1148 case PTR_TO_SOCKET_OR_NULL:
1149 reg->type = PTR_TO_SOCKET;
1151 case PTR_TO_SOCK_COMMON_OR_NULL:
1152 reg->type = PTR_TO_SOCK_COMMON;
1154 case PTR_TO_TCP_SOCK_OR_NULL:
1155 reg->type = PTR_TO_TCP_SOCK;
1157 case PTR_TO_BTF_ID_OR_NULL:
1158 reg->type = PTR_TO_BTF_ID;
1160 case PTR_TO_MEM_OR_NULL:
1161 reg->type = PTR_TO_MEM;
1163 case PTR_TO_RDONLY_BUF_OR_NULL:
1164 reg->type = PTR_TO_RDONLY_BUF;
1166 case PTR_TO_RDWR_BUF_OR_NULL:
1167 reg->type = PTR_TO_RDWR_BUF;
1170 WARN_ONCE(1, "unknown nullable register type");
1174 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
1176 return type_is_pkt_pointer(reg->type);
1179 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
1181 return reg_is_pkt_pointer(reg) ||
1182 reg->type == PTR_TO_PACKET_END;
1185 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
1186 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
1187 enum bpf_reg_type which)
1189 /* The register can already have a range from prior markings.
1190 * This is fine as long as it hasn't been advanced from its
1193 return reg->type == which &&
1196 tnum_equals_const(reg->var_off, 0);
1199 /* Reset the min/max bounds of a register */
1200 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
1202 reg->smin_value = S64_MIN;
1203 reg->smax_value = S64_MAX;
1204 reg->umin_value = 0;
1205 reg->umax_value = U64_MAX;
1207 reg->s32_min_value = S32_MIN;
1208 reg->s32_max_value = S32_MAX;
1209 reg->u32_min_value = 0;
1210 reg->u32_max_value = U32_MAX;
1213 static void __mark_reg64_unbounded(struct bpf_reg_state *reg)
1215 reg->smin_value = S64_MIN;
1216 reg->smax_value = S64_MAX;
1217 reg->umin_value = 0;
1218 reg->umax_value = U64_MAX;
1221 static void __mark_reg32_unbounded(struct bpf_reg_state *reg)
1223 reg->s32_min_value = S32_MIN;
1224 reg->s32_max_value = S32_MAX;
1225 reg->u32_min_value = 0;
1226 reg->u32_max_value = U32_MAX;
1229 static void __update_reg32_bounds(struct bpf_reg_state *reg)
1231 struct tnum var32_off = tnum_subreg(reg->var_off);
1233 /* min signed is max(sign bit) | min(other bits) */
1234 reg->s32_min_value = max_t(s32, reg->s32_min_value,
1235 var32_off.value | (var32_off.mask & S32_MIN));
1236 /* max signed is min(sign bit) | max(other bits) */
1237 reg->s32_max_value = min_t(s32, reg->s32_max_value,
1238 var32_off.value | (var32_off.mask & S32_MAX));
1239 reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)var32_off.value);
1240 reg->u32_max_value = min(reg->u32_max_value,
1241 (u32)(var32_off.value | var32_off.mask));
1244 static void __update_reg64_bounds(struct bpf_reg_state *reg)
1246 /* min signed is max(sign bit) | min(other bits) */
1247 reg->smin_value = max_t(s64, reg->smin_value,
1248 reg->var_off.value | (reg->var_off.mask & S64_MIN));
1249 /* max signed is min(sign bit) | max(other bits) */
1250 reg->smax_value = min_t(s64, reg->smax_value,
1251 reg->var_off.value | (reg->var_off.mask & S64_MAX));
1252 reg->umin_value = max(reg->umin_value, reg->var_off.value);
1253 reg->umax_value = min(reg->umax_value,
1254 reg->var_off.value | reg->var_off.mask);
1257 static void __update_reg_bounds(struct bpf_reg_state *reg)
1259 __update_reg32_bounds(reg);
1260 __update_reg64_bounds(reg);
1263 /* Uses signed min/max values to inform unsigned, and vice-versa */
1264 static void __reg32_deduce_bounds(struct bpf_reg_state *reg)
1266 /* Learn sign from signed bounds.
1267 * If we cannot cross the sign boundary, then signed and unsigned bounds
1268 * are the same, so combine. This works even in the negative case, e.g.
1269 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1271 if (reg->s32_min_value >= 0 || reg->s32_max_value < 0) {
1272 reg->s32_min_value = reg->u32_min_value =
1273 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1274 reg->s32_max_value = reg->u32_max_value =
1275 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1278 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1279 * boundary, so we must be careful.
1281 if ((s32)reg->u32_max_value >= 0) {
1282 /* Positive. We can't learn anything from the smin, but smax
1283 * is positive, hence safe.
1285 reg->s32_min_value = reg->u32_min_value;
1286 reg->s32_max_value = reg->u32_max_value =
1287 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1288 } else if ((s32)reg->u32_min_value < 0) {
1289 /* Negative. We can't learn anything from the smax, but smin
1290 * is negative, hence safe.
1292 reg->s32_min_value = reg->u32_min_value =
1293 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1294 reg->s32_max_value = reg->u32_max_value;
1298 static void __reg64_deduce_bounds(struct bpf_reg_state *reg)
1300 /* Learn sign from signed bounds.
1301 * If we cannot cross the sign boundary, then signed and unsigned bounds
1302 * are the same, so combine. This works even in the negative case, e.g.
1303 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1305 if (reg->smin_value >= 0 || reg->smax_value < 0) {
1306 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1308 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1312 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1313 * boundary, so we must be careful.
1315 if ((s64)reg->umax_value >= 0) {
1316 /* Positive. We can't learn anything from the smin, but smax
1317 * is positive, hence safe.
1319 reg->smin_value = reg->umin_value;
1320 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1322 } else if ((s64)reg->umin_value < 0) {
1323 /* Negative. We can't learn anything from the smax, but smin
1324 * is negative, hence safe.
1326 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1328 reg->smax_value = reg->umax_value;
1332 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
1334 __reg32_deduce_bounds(reg);
1335 __reg64_deduce_bounds(reg);
1338 /* Attempts to improve var_off based on unsigned min/max information */
1339 static void __reg_bound_offset(struct bpf_reg_state *reg)
1341 struct tnum var64_off = tnum_intersect(reg->var_off,
1342 tnum_range(reg->umin_value,
1344 struct tnum var32_off = tnum_intersect(tnum_subreg(reg->var_off),
1345 tnum_range(reg->u32_min_value,
1346 reg->u32_max_value));
1348 reg->var_off = tnum_or(tnum_clear_subreg(var64_off), var32_off);
1351 static void __reg_assign_32_into_64(struct bpf_reg_state *reg)
1353 reg->umin_value = reg->u32_min_value;
1354 reg->umax_value = reg->u32_max_value;
1355 /* Attempt to pull 32-bit signed bounds into 64-bit bounds
1356 * but must be positive otherwise set to worse case bounds
1357 * and refine later from tnum.
1359 if (reg->s32_min_value >= 0 && reg->s32_max_value >= 0)
1360 reg->smax_value = reg->s32_max_value;
1362 reg->smax_value = U32_MAX;
1363 if (reg->s32_min_value >= 0)
1364 reg->smin_value = reg->s32_min_value;
1366 reg->smin_value = 0;
1369 static void __reg_combine_32_into_64(struct bpf_reg_state *reg)
1371 /* special case when 64-bit register has upper 32-bit register
1372 * zeroed. Typically happens after zext or <<32, >>32 sequence
1373 * allowing us to use 32-bit bounds directly,
1375 if (tnum_equals_const(tnum_clear_subreg(reg->var_off), 0)) {
1376 __reg_assign_32_into_64(reg);
1378 /* Otherwise the best we can do is push lower 32bit known and
1379 * unknown bits into register (var_off set from jmp logic)
1380 * then learn as much as possible from the 64-bit tnum
1381 * known and unknown bits. The previous smin/smax bounds are
1382 * invalid here because of jmp32 compare so mark them unknown
1383 * so they do not impact tnum bounds calculation.
1385 __mark_reg64_unbounded(reg);
1386 __update_reg_bounds(reg);
1389 /* Intersecting with the old var_off might have improved our bounds
1390 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1391 * then new var_off is (0; 0x7f...fc) which improves our umax.
1393 __reg_deduce_bounds(reg);
1394 __reg_bound_offset(reg);
1395 __update_reg_bounds(reg);
1398 static bool __reg64_bound_s32(s64 a)
1400 return a > S32_MIN && a < S32_MAX;
1403 static bool __reg64_bound_u32(u64 a)
1405 return a > U32_MIN && a < U32_MAX;
1408 static void __reg_combine_64_into_32(struct bpf_reg_state *reg)
1410 __mark_reg32_unbounded(reg);
1412 if (__reg64_bound_s32(reg->smin_value) && __reg64_bound_s32(reg->smax_value)) {
1413 reg->s32_min_value = (s32)reg->smin_value;
1414 reg->s32_max_value = (s32)reg->smax_value;
1416 if (__reg64_bound_u32(reg->umin_value) && __reg64_bound_u32(reg->umax_value)) {
1417 reg->u32_min_value = (u32)reg->umin_value;
1418 reg->u32_max_value = (u32)reg->umax_value;
1421 /* Intersecting with the old var_off might have improved our bounds
1422 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1423 * then new var_off is (0; 0x7f...fc) which improves our umax.
1425 __reg_deduce_bounds(reg);
1426 __reg_bound_offset(reg);
1427 __update_reg_bounds(reg);
1430 /* Mark a register as having a completely unknown (scalar) value. */
1431 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
1432 struct bpf_reg_state *reg)
1435 * Clear type, id, off, and union(map_ptr, range) and
1436 * padding between 'type' and union
1438 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
1439 reg->type = SCALAR_VALUE;
1440 reg->var_off = tnum_unknown;
1442 reg->precise = env->subprog_cnt > 1 || !env->bpf_capable;
1443 __mark_reg_unbounded(reg);
1446 static void mark_reg_unknown(struct bpf_verifier_env *env,
1447 struct bpf_reg_state *regs, u32 regno)
1449 if (WARN_ON(regno >= MAX_BPF_REG)) {
1450 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
1451 /* Something bad happened, let's kill all regs except FP */
1452 for (regno = 0; regno < BPF_REG_FP; regno++)
1453 __mark_reg_not_init(env, regs + regno);
1456 __mark_reg_unknown(env, regs + regno);
1459 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1460 struct bpf_reg_state *reg)
1462 __mark_reg_unknown(env, reg);
1463 reg->type = NOT_INIT;
1466 static void mark_reg_not_init(struct bpf_verifier_env *env,
1467 struct bpf_reg_state *regs, u32 regno)
1469 if (WARN_ON(regno >= MAX_BPF_REG)) {
1470 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
1471 /* Something bad happened, let's kill all regs except FP */
1472 for (regno = 0; regno < BPF_REG_FP; regno++)
1473 __mark_reg_not_init(env, regs + regno);
1476 __mark_reg_not_init(env, regs + regno);
1479 static void mark_btf_ld_reg(struct bpf_verifier_env *env,
1480 struct bpf_reg_state *regs, u32 regno,
1481 enum bpf_reg_type reg_type,
1482 struct btf *btf, u32 btf_id)
1484 if (reg_type == SCALAR_VALUE) {
1485 mark_reg_unknown(env, regs, regno);
1488 mark_reg_known_zero(env, regs, regno);
1489 regs[regno].type = PTR_TO_BTF_ID;
1490 regs[regno].btf = btf;
1491 regs[regno].btf_id = btf_id;
1494 #define DEF_NOT_SUBREG (0)
1495 static void init_reg_state(struct bpf_verifier_env *env,
1496 struct bpf_func_state *state)
1498 struct bpf_reg_state *regs = state->regs;
1501 for (i = 0; i < MAX_BPF_REG; i++) {
1502 mark_reg_not_init(env, regs, i);
1503 regs[i].live = REG_LIVE_NONE;
1504 regs[i].parent = NULL;
1505 regs[i].subreg_def = DEF_NOT_SUBREG;
1509 regs[BPF_REG_FP].type = PTR_TO_STACK;
1510 mark_reg_known_zero(env, regs, BPF_REG_FP);
1511 regs[BPF_REG_FP].frameno = state->frameno;
1514 #define BPF_MAIN_FUNC (-1)
1515 static void init_func_state(struct bpf_verifier_env *env,
1516 struct bpf_func_state *state,
1517 int callsite, int frameno, int subprogno)
1519 state->callsite = callsite;
1520 state->frameno = frameno;
1521 state->subprogno = subprogno;
1522 init_reg_state(env, state);
1526 SRC_OP, /* register is used as source operand */
1527 DST_OP, /* register is used as destination operand */
1528 DST_OP_NO_MARK /* same as above, check only, don't mark */
1531 static int cmp_subprogs(const void *a, const void *b)
1533 return ((struct bpf_subprog_info *)a)->start -
1534 ((struct bpf_subprog_info *)b)->start;
1537 static int find_subprog(struct bpf_verifier_env *env, int off)
1539 struct bpf_subprog_info *p;
1541 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
1542 sizeof(env->subprog_info[0]), cmp_subprogs);
1545 return p - env->subprog_info;
1549 static int add_subprog(struct bpf_verifier_env *env, int off)
1551 int insn_cnt = env->prog->len;
1554 if (off >= insn_cnt || off < 0) {
1555 verbose(env, "call to invalid destination\n");
1558 ret = find_subprog(env, off);
1561 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
1562 verbose(env, "too many subprograms\n");
1565 /* determine subprog starts. The end is one before the next starts */
1566 env->subprog_info[env->subprog_cnt++].start = off;
1567 sort(env->subprog_info, env->subprog_cnt,
1568 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
1569 return env->subprog_cnt - 1;
1572 struct bpf_kfunc_desc {
1573 struct btf_func_model func_model;
1578 #define MAX_KFUNC_DESCS 256
1579 struct bpf_kfunc_desc_tab {
1580 struct bpf_kfunc_desc descs[MAX_KFUNC_DESCS];
1584 static int kfunc_desc_cmp_by_id(const void *a, const void *b)
1586 const struct bpf_kfunc_desc *d0 = a;
1587 const struct bpf_kfunc_desc *d1 = b;
1589 /* func_id is not greater than BTF_MAX_TYPE */
1590 return d0->func_id - d1->func_id;
1593 static const struct bpf_kfunc_desc *
1594 find_kfunc_desc(const struct bpf_prog *prog, u32 func_id)
1596 struct bpf_kfunc_desc desc = {
1599 struct bpf_kfunc_desc_tab *tab;
1601 tab = prog->aux->kfunc_tab;
1602 return bsearch(&desc, tab->descs, tab->nr_descs,
1603 sizeof(tab->descs[0]), kfunc_desc_cmp_by_id);
1606 static int add_kfunc_call(struct bpf_verifier_env *env, u32 func_id)
1608 const struct btf_type *func, *func_proto;
1609 struct bpf_kfunc_desc_tab *tab;
1610 struct bpf_prog_aux *prog_aux;
1611 struct bpf_kfunc_desc *desc;
1612 const char *func_name;
1616 prog_aux = env->prog->aux;
1617 tab = prog_aux->kfunc_tab;
1620 verbose(env, "calling kernel function is not supported without CONFIG_DEBUG_INFO_BTF\n");
1624 if (!env->prog->jit_requested) {
1625 verbose(env, "JIT is required for calling kernel function\n");
1629 if (!bpf_jit_supports_kfunc_call()) {
1630 verbose(env, "JIT does not support calling kernel function\n");
1634 if (!env->prog->gpl_compatible) {
1635 verbose(env, "cannot call kernel function from non-GPL compatible program\n");
1639 tab = kzalloc(sizeof(*tab), GFP_KERNEL);
1642 prog_aux->kfunc_tab = tab;
1645 if (find_kfunc_desc(env->prog, func_id))
1648 if (tab->nr_descs == MAX_KFUNC_DESCS) {
1649 verbose(env, "too many different kernel function calls\n");
1653 func = btf_type_by_id(btf_vmlinux, func_id);
1654 if (!func || !btf_type_is_func(func)) {
1655 verbose(env, "kernel btf_id %u is not a function\n",
1659 func_proto = btf_type_by_id(btf_vmlinux, func->type);
1660 if (!func_proto || !btf_type_is_func_proto(func_proto)) {
1661 verbose(env, "kernel function btf_id %u does not have a valid func_proto\n",
1666 func_name = btf_name_by_offset(btf_vmlinux, func->name_off);
1667 addr = kallsyms_lookup_name(func_name);
1669 verbose(env, "cannot find address for kernel function %s\n",
1674 desc = &tab->descs[tab->nr_descs++];
1675 desc->func_id = func_id;
1676 desc->imm = BPF_CAST_CALL(addr) - __bpf_call_base;
1677 err = btf_distill_func_proto(&env->log, btf_vmlinux,
1678 func_proto, func_name,
1681 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
1682 kfunc_desc_cmp_by_id, NULL);
1686 static int kfunc_desc_cmp_by_imm(const void *a, const void *b)
1688 const struct bpf_kfunc_desc *d0 = a;
1689 const struct bpf_kfunc_desc *d1 = b;
1691 if (d0->imm > d1->imm)
1693 else if (d0->imm < d1->imm)
1698 static void sort_kfunc_descs_by_imm(struct bpf_prog *prog)
1700 struct bpf_kfunc_desc_tab *tab;
1702 tab = prog->aux->kfunc_tab;
1706 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
1707 kfunc_desc_cmp_by_imm, NULL);
1710 bool bpf_prog_has_kfunc_call(const struct bpf_prog *prog)
1712 return !!prog->aux->kfunc_tab;
1715 const struct btf_func_model *
1716 bpf_jit_find_kfunc_model(const struct bpf_prog *prog,
1717 const struct bpf_insn *insn)
1719 const struct bpf_kfunc_desc desc = {
1722 const struct bpf_kfunc_desc *res;
1723 struct bpf_kfunc_desc_tab *tab;
1725 tab = prog->aux->kfunc_tab;
1726 res = bsearch(&desc, tab->descs, tab->nr_descs,
1727 sizeof(tab->descs[0]), kfunc_desc_cmp_by_imm);
1729 return res ? &res->func_model : NULL;
1732 static int add_subprog_and_kfunc(struct bpf_verifier_env *env)
1734 struct bpf_subprog_info *subprog = env->subprog_info;
1735 struct bpf_insn *insn = env->prog->insnsi;
1736 int i, ret, insn_cnt = env->prog->len;
1738 /* Add entry function. */
1739 ret = add_subprog(env, 0);
1743 for (i = 0; i < insn_cnt; i++, insn++) {
1744 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn) &&
1745 !bpf_pseudo_kfunc_call(insn))
1748 if (!env->bpf_capable) {
1749 verbose(env, "loading/calling other bpf or kernel functions are allowed for CAP_BPF and CAP_SYS_ADMIN\n");
1753 if (bpf_pseudo_func(insn)) {
1754 ret = add_subprog(env, i + insn->imm + 1);
1756 /* remember subprog */
1758 } else if (bpf_pseudo_call(insn)) {
1759 ret = add_subprog(env, i + insn->imm + 1);
1761 ret = add_kfunc_call(env, insn->imm);
1768 /* Add a fake 'exit' subprog which could simplify subprog iteration
1769 * logic. 'subprog_cnt' should not be increased.
1771 subprog[env->subprog_cnt].start = insn_cnt;
1773 if (env->log.level & BPF_LOG_LEVEL2)
1774 for (i = 0; i < env->subprog_cnt; i++)
1775 verbose(env, "func#%d @%d\n", i, subprog[i].start);
1780 static int check_subprogs(struct bpf_verifier_env *env)
1782 int i, subprog_start, subprog_end, off, cur_subprog = 0;
1783 struct bpf_subprog_info *subprog = env->subprog_info;
1784 struct bpf_insn *insn = env->prog->insnsi;
1785 int insn_cnt = env->prog->len;
1787 /* now check that all jumps are within the same subprog */
1788 subprog_start = subprog[cur_subprog].start;
1789 subprog_end = subprog[cur_subprog + 1].start;
1790 for (i = 0; i < insn_cnt; i++) {
1791 u8 code = insn[i].code;
1793 if (code == (BPF_JMP | BPF_CALL) &&
1794 insn[i].imm == BPF_FUNC_tail_call &&
1795 insn[i].src_reg != BPF_PSEUDO_CALL)
1796 subprog[cur_subprog].has_tail_call = true;
1797 if (BPF_CLASS(code) == BPF_LD &&
1798 (BPF_MODE(code) == BPF_ABS || BPF_MODE(code) == BPF_IND))
1799 subprog[cur_subprog].has_ld_abs = true;
1800 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
1802 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
1804 off = i + insn[i].off + 1;
1805 if (off < subprog_start || off >= subprog_end) {
1806 verbose(env, "jump out of range from insn %d to %d\n", i, off);
1810 if (i == subprog_end - 1) {
1811 /* to avoid fall-through from one subprog into another
1812 * the last insn of the subprog should be either exit
1813 * or unconditional jump back
1815 if (code != (BPF_JMP | BPF_EXIT) &&
1816 code != (BPF_JMP | BPF_JA)) {
1817 verbose(env, "last insn is not an exit or jmp\n");
1820 subprog_start = subprog_end;
1822 if (cur_subprog < env->subprog_cnt)
1823 subprog_end = subprog[cur_subprog + 1].start;
1829 /* Parentage chain of this register (or stack slot) should take care of all
1830 * issues like callee-saved registers, stack slot allocation time, etc.
1832 static int mark_reg_read(struct bpf_verifier_env *env,
1833 const struct bpf_reg_state *state,
1834 struct bpf_reg_state *parent, u8 flag)
1836 bool writes = parent == state->parent; /* Observe write marks */
1840 /* if read wasn't screened by an earlier write ... */
1841 if (writes && state->live & REG_LIVE_WRITTEN)
1843 if (parent->live & REG_LIVE_DONE) {
1844 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
1845 reg_type_str[parent->type],
1846 parent->var_off.value, parent->off);
1849 /* The first condition is more likely to be true than the
1850 * second, checked it first.
1852 if ((parent->live & REG_LIVE_READ) == flag ||
1853 parent->live & REG_LIVE_READ64)
1854 /* The parentage chain never changes and
1855 * this parent was already marked as LIVE_READ.
1856 * There is no need to keep walking the chain again and
1857 * keep re-marking all parents as LIVE_READ.
1858 * This case happens when the same register is read
1859 * multiple times without writes into it in-between.
1860 * Also, if parent has the stronger REG_LIVE_READ64 set,
1861 * then no need to set the weak REG_LIVE_READ32.
1864 /* ... then we depend on parent's value */
1865 parent->live |= flag;
1866 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
1867 if (flag == REG_LIVE_READ64)
1868 parent->live &= ~REG_LIVE_READ32;
1870 parent = state->parent;
1875 if (env->longest_mark_read_walk < cnt)
1876 env->longest_mark_read_walk = cnt;
1880 /* This function is supposed to be used by the following 32-bit optimization
1881 * code only. It returns TRUE if the source or destination register operates
1882 * on 64-bit, otherwise return FALSE.
1884 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
1885 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
1890 class = BPF_CLASS(code);
1892 if (class == BPF_JMP) {
1893 /* BPF_EXIT for "main" will reach here. Return TRUE
1898 if (op == BPF_CALL) {
1899 /* BPF to BPF call will reach here because of marking
1900 * caller saved clobber with DST_OP_NO_MARK for which we
1901 * don't care the register def because they are anyway
1902 * marked as NOT_INIT already.
1904 if (insn->src_reg == BPF_PSEUDO_CALL)
1906 /* Helper call will reach here because of arg type
1907 * check, conservatively return TRUE.
1916 if (class == BPF_ALU64 || class == BPF_JMP ||
1917 /* BPF_END always use BPF_ALU class. */
1918 (class == BPF_ALU && op == BPF_END && insn->imm == 64))
1921 if (class == BPF_ALU || class == BPF_JMP32)
1924 if (class == BPF_LDX) {
1926 return BPF_SIZE(code) == BPF_DW;
1927 /* LDX source must be ptr. */
1931 if (class == BPF_STX) {
1932 /* BPF_STX (including atomic variants) has multiple source
1933 * operands, one of which is a ptr. Check whether the caller is
1936 if (t == SRC_OP && reg->type != SCALAR_VALUE)
1938 return BPF_SIZE(code) == BPF_DW;
1941 if (class == BPF_LD) {
1942 u8 mode = BPF_MODE(code);
1945 if (mode == BPF_IMM)
1948 /* Both LD_IND and LD_ABS return 32-bit data. */
1952 /* Implicit ctx ptr. */
1953 if (regno == BPF_REG_6)
1956 /* Explicit source could be any width. */
1960 if (class == BPF_ST)
1961 /* The only source register for BPF_ST is a ptr. */
1964 /* Conservatively return true at default. */
1968 /* Return the regno defined by the insn, or -1. */
1969 static int insn_def_regno(const struct bpf_insn *insn)
1971 switch (BPF_CLASS(insn->code)) {
1977 if (BPF_MODE(insn->code) == BPF_ATOMIC &&
1978 (insn->imm & BPF_FETCH)) {
1979 if (insn->imm == BPF_CMPXCHG)
1982 return insn->src_reg;
1987 return insn->dst_reg;
1991 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
1992 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
1994 int dst_reg = insn_def_regno(insn);
1999 return !is_reg64(env, insn, dst_reg, NULL, DST_OP);
2002 static void mark_insn_zext(struct bpf_verifier_env *env,
2003 struct bpf_reg_state *reg)
2005 s32 def_idx = reg->subreg_def;
2007 if (def_idx == DEF_NOT_SUBREG)
2010 env->insn_aux_data[def_idx - 1].zext_dst = true;
2011 /* The dst will be zero extended, so won't be sub-register anymore. */
2012 reg->subreg_def = DEF_NOT_SUBREG;
2015 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
2016 enum reg_arg_type t)
2018 struct bpf_verifier_state *vstate = env->cur_state;
2019 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2020 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
2021 struct bpf_reg_state *reg, *regs = state->regs;
2024 if (regno >= MAX_BPF_REG) {
2025 verbose(env, "R%d is invalid\n", regno);
2030 rw64 = is_reg64(env, insn, regno, reg, t);
2032 /* check whether register used as source operand can be read */
2033 if (reg->type == NOT_INIT) {
2034 verbose(env, "R%d !read_ok\n", regno);
2037 /* We don't need to worry about FP liveness because it's read-only */
2038 if (regno == BPF_REG_FP)
2042 mark_insn_zext(env, reg);
2044 return mark_reg_read(env, reg, reg->parent,
2045 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
2047 /* check whether register used as dest operand can be written to */
2048 if (regno == BPF_REG_FP) {
2049 verbose(env, "frame pointer is read only\n");
2052 reg->live |= REG_LIVE_WRITTEN;
2053 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
2055 mark_reg_unknown(env, regs, regno);
2060 /* for any branch, call, exit record the history of jmps in the given state */
2061 static int push_jmp_history(struct bpf_verifier_env *env,
2062 struct bpf_verifier_state *cur)
2064 u32 cnt = cur->jmp_history_cnt;
2065 struct bpf_idx_pair *p;
2068 p = krealloc(cur->jmp_history, cnt * sizeof(*p), GFP_USER);
2071 p[cnt - 1].idx = env->insn_idx;
2072 p[cnt - 1].prev_idx = env->prev_insn_idx;
2073 cur->jmp_history = p;
2074 cur->jmp_history_cnt = cnt;
2078 /* Backtrack one insn at a time. If idx is not at the top of recorded
2079 * history then previous instruction came from straight line execution.
2081 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
2086 if (cnt && st->jmp_history[cnt - 1].idx == i) {
2087 i = st->jmp_history[cnt - 1].prev_idx;
2095 static const char *disasm_kfunc_name(void *data, const struct bpf_insn *insn)
2097 const struct btf_type *func;
2099 if (insn->src_reg != BPF_PSEUDO_KFUNC_CALL)
2102 func = btf_type_by_id(btf_vmlinux, insn->imm);
2103 return btf_name_by_offset(btf_vmlinux, func->name_off);
2106 /* For given verifier state backtrack_insn() is called from the last insn to
2107 * the first insn. Its purpose is to compute a bitmask of registers and
2108 * stack slots that needs precision in the parent verifier state.
2110 static int backtrack_insn(struct bpf_verifier_env *env, int idx,
2111 u32 *reg_mask, u64 *stack_mask)
2113 const struct bpf_insn_cbs cbs = {
2114 .cb_call = disasm_kfunc_name,
2115 .cb_print = verbose,
2116 .private_data = env,
2118 struct bpf_insn *insn = env->prog->insnsi + idx;
2119 u8 class = BPF_CLASS(insn->code);
2120 u8 opcode = BPF_OP(insn->code);
2121 u8 mode = BPF_MODE(insn->code);
2122 u32 dreg = 1u << insn->dst_reg;
2123 u32 sreg = 1u << insn->src_reg;
2126 if (insn->code == 0)
2128 if (env->log.level & BPF_LOG_LEVEL) {
2129 verbose(env, "regs=%x stack=%llx before ", *reg_mask, *stack_mask);
2130 verbose(env, "%d: ", idx);
2131 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
2134 if (class == BPF_ALU || class == BPF_ALU64) {
2135 if (!(*reg_mask & dreg))
2137 if (opcode == BPF_MOV) {
2138 if (BPF_SRC(insn->code) == BPF_X) {
2140 * dreg needs precision after this insn
2141 * sreg needs precision before this insn
2147 * dreg needs precision after this insn.
2148 * Corresponding register is already marked
2149 * as precise=true in this verifier state.
2150 * No further markings in parent are necessary
2155 if (BPF_SRC(insn->code) == BPF_X) {
2157 * both dreg and sreg need precision
2162 * dreg still needs precision before this insn
2165 } else if (class == BPF_LDX) {
2166 if (!(*reg_mask & dreg))
2170 /* scalars can only be spilled into stack w/o losing precision.
2171 * Load from any other memory can be zero extended.
2172 * The desire to keep that precision is already indicated
2173 * by 'precise' mark in corresponding register of this state.
2174 * No further tracking necessary.
2176 if (insn->src_reg != BPF_REG_FP)
2178 if (BPF_SIZE(insn->code) != BPF_DW)
2181 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
2182 * that [fp - off] slot contains scalar that needs to be
2183 * tracked with precision
2185 spi = (-insn->off - 1) / BPF_REG_SIZE;
2187 verbose(env, "BUG spi %d\n", spi);
2188 WARN_ONCE(1, "verifier backtracking bug");
2191 *stack_mask |= 1ull << spi;
2192 } else if (class == BPF_STX || class == BPF_ST) {
2193 if (*reg_mask & dreg)
2194 /* stx & st shouldn't be using _scalar_ dst_reg
2195 * to access memory. It means backtracking
2196 * encountered a case of pointer subtraction.
2199 /* scalars can only be spilled into stack */
2200 if (insn->dst_reg != BPF_REG_FP)
2202 if (BPF_SIZE(insn->code) != BPF_DW)
2204 spi = (-insn->off - 1) / BPF_REG_SIZE;
2206 verbose(env, "BUG spi %d\n", spi);
2207 WARN_ONCE(1, "verifier backtracking bug");
2210 if (!(*stack_mask & (1ull << spi)))
2212 *stack_mask &= ~(1ull << spi);
2213 if (class == BPF_STX)
2215 } else if (class == BPF_JMP || class == BPF_JMP32) {
2216 if (opcode == BPF_CALL) {
2217 if (insn->src_reg == BPF_PSEUDO_CALL)
2219 /* regular helper call sets R0 */
2221 if (*reg_mask & 0x3f) {
2222 /* if backtracing was looking for registers R1-R5
2223 * they should have been found already.
2225 verbose(env, "BUG regs %x\n", *reg_mask);
2226 WARN_ONCE(1, "verifier backtracking bug");
2229 } else if (opcode == BPF_EXIT) {
2232 } else if (class == BPF_LD) {
2233 if (!(*reg_mask & dreg))
2236 /* It's ld_imm64 or ld_abs or ld_ind.
2237 * For ld_imm64 no further tracking of precision
2238 * into parent is necessary
2240 if (mode == BPF_IND || mode == BPF_ABS)
2241 /* to be analyzed */
2247 /* the scalar precision tracking algorithm:
2248 * . at the start all registers have precise=false.
2249 * . scalar ranges are tracked as normal through alu and jmp insns.
2250 * . once precise value of the scalar register is used in:
2251 * . ptr + scalar alu
2252 * . if (scalar cond K|scalar)
2253 * . helper_call(.., scalar, ...) where ARG_CONST is expected
2254 * backtrack through the verifier states and mark all registers and
2255 * stack slots with spilled constants that these scalar regisers
2256 * should be precise.
2257 * . during state pruning two registers (or spilled stack slots)
2258 * are equivalent if both are not precise.
2260 * Note the verifier cannot simply walk register parentage chain,
2261 * since many different registers and stack slots could have been
2262 * used to compute single precise scalar.
2264 * The approach of starting with precise=true for all registers and then
2265 * backtrack to mark a register as not precise when the verifier detects
2266 * that program doesn't care about specific value (e.g., when helper
2267 * takes register as ARG_ANYTHING parameter) is not safe.
2269 * It's ok to walk single parentage chain of the verifier states.
2270 * It's possible that this backtracking will go all the way till 1st insn.
2271 * All other branches will be explored for needing precision later.
2273 * The backtracking needs to deal with cases like:
2274 * 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)
2277 * if r5 > 0x79f goto pc+7
2278 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
2281 * call bpf_perf_event_output#25
2282 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
2286 * call foo // uses callee's r6 inside to compute r0
2290 * to track above reg_mask/stack_mask needs to be independent for each frame.
2292 * Also if parent's curframe > frame where backtracking started,
2293 * the verifier need to mark registers in both frames, otherwise callees
2294 * may incorrectly prune callers. This is similar to
2295 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
2297 * For now backtracking falls back into conservative marking.
2299 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
2300 struct bpf_verifier_state *st)
2302 struct bpf_func_state *func;
2303 struct bpf_reg_state *reg;
2306 /* big hammer: mark all scalars precise in this path.
2307 * pop_stack may still get !precise scalars.
2309 for (; st; st = st->parent)
2310 for (i = 0; i <= st->curframe; i++) {
2311 func = st->frame[i];
2312 for (j = 0; j < BPF_REG_FP; j++) {
2313 reg = &func->regs[j];
2314 if (reg->type != SCALAR_VALUE)
2316 reg->precise = true;
2318 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
2319 if (func->stack[j].slot_type[0] != STACK_SPILL)
2321 reg = &func->stack[j].spilled_ptr;
2322 if (reg->type != SCALAR_VALUE)
2324 reg->precise = true;
2329 static int __mark_chain_precision(struct bpf_verifier_env *env, int regno,
2332 struct bpf_verifier_state *st = env->cur_state;
2333 int first_idx = st->first_insn_idx;
2334 int last_idx = env->insn_idx;
2335 struct bpf_func_state *func;
2336 struct bpf_reg_state *reg;
2337 u32 reg_mask = regno >= 0 ? 1u << regno : 0;
2338 u64 stack_mask = spi >= 0 ? 1ull << spi : 0;
2339 bool skip_first = true;
2340 bool new_marks = false;
2343 if (!env->bpf_capable)
2346 func = st->frame[st->curframe];
2348 reg = &func->regs[regno];
2349 if (reg->type != SCALAR_VALUE) {
2350 WARN_ONCE(1, "backtracing misuse");
2357 reg->precise = true;
2361 if (func->stack[spi].slot_type[0] != STACK_SPILL) {
2365 reg = &func->stack[spi].spilled_ptr;
2366 if (reg->type != SCALAR_VALUE) {
2374 reg->precise = true;
2380 if (!reg_mask && !stack_mask)
2383 DECLARE_BITMAP(mask, 64);
2384 u32 history = st->jmp_history_cnt;
2386 if (env->log.level & BPF_LOG_LEVEL)
2387 verbose(env, "last_idx %d first_idx %d\n", last_idx, first_idx);
2388 for (i = last_idx;;) {
2393 err = backtrack_insn(env, i, ®_mask, &stack_mask);
2395 if (err == -ENOTSUPP) {
2396 mark_all_scalars_precise(env, st);
2401 if (!reg_mask && !stack_mask)
2402 /* Found assignment(s) into tracked register in this state.
2403 * Since this state is already marked, just return.
2404 * Nothing to be tracked further in the parent state.
2409 i = get_prev_insn_idx(st, i, &history);
2410 if (i >= env->prog->len) {
2411 /* This can happen if backtracking reached insn 0
2412 * and there are still reg_mask or stack_mask
2414 * It means the backtracking missed the spot where
2415 * particular register was initialized with a constant.
2417 verbose(env, "BUG backtracking idx %d\n", i);
2418 WARN_ONCE(1, "verifier backtracking bug");
2427 func = st->frame[st->curframe];
2428 bitmap_from_u64(mask, reg_mask);
2429 for_each_set_bit(i, mask, 32) {
2430 reg = &func->regs[i];
2431 if (reg->type != SCALAR_VALUE) {
2432 reg_mask &= ~(1u << i);
2437 reg->precise = true;
2440 bitmap_from_u64(mask, stack_mask);
2441 for_each_set_bit(i, mask, 64) {
2442 if (i >= func->allocated_stack / BPF_REG_SIZE) {
2443 /* the sequence of instructions:
2445 * 3: (7b) *(u64 *)(r3 -8) = r0
2446 * 4: (79) r4 = *(u64 *)(r10 -8)
2447 * doesn't contain jmps. It's backtracked
2448 * as a single block.
2449 * During backtracking insn 3 is not recognized as
2450 * stack access, so at the end of backtracking
2451 * stack slot fp-8 is still marked in stack_mask.
2452 * However the parent state may not have accessed
2453 * fp-8 and it's "unallocated" stack space.
2454 * In such case fallback to conservative.
2456 mark_all_scalars_precise(env, st);
2460 if (func->stack[i].slot_type[0] != STACK_SPILL) {
2461 stack_mask &= ~(1ull << i);
2464 reg = &func->stack[i].spilled_ptr;
2465 if (reg->type != SCALAR_VALUE) {
2466 stack_mask &= ~(1ull << i);
2471 reg->precise = true;
2473 if (env->log.level & BPF_LOG_LEVEL) {
2474 print_verifier_state(env, func);
2475 verbose(env, "parent %s regs=%x stack=%llx marks\n",
2476 new_marks ? "didn't have" : "already had",
2477 reg_mask, stack_mask);
2480 if (!reg_mask && !stack_mask)
2485 last_idx = st->last_insn_idx;
2486 first_idx = st->first_insn_idx;
2491 static int mark_chain_precision(struct bpf_verifier_env *env, int regno)
2493 return __mark_chain_precision(env, regno, -1);
2496 static int mark_chain_precision_stack(struct bpf_verifier_env *env, int spi)
2498 return __mark_chain_precision(env, -1, spi);
2501 static bool is_spillable_regtype(enum bpf_reg_type type)
2504 case PTR_TO_MAP_VALUE:
2505 case PTR_TO_MAP_VALUE_OR_NULL:
2509 case PTR_TO_PACKET_META:
2510 case PTR_TO_PACKET_END:
2511 case PTR_TO_FLOW_KEYS:
2512 case CONST_PTR_TO_MAP:
2514 case PTR_TO_SOCKET_OR_NULL:
2515 case PTR_TO_SOCK_COMMON:
2516 case PTR_TO_SOCK_COMMON_OR_NULL:
2517 case PTR_TO_TCP_SOCK:
2518 case PTR_TO_TCP_SOCK_OR_NULL:
2519 case PTR_TO_XDP_SOCK:
2521 case PTR_TO_BTF_ID_OR_NULL:
2522 case PTR_TO_RDONLY_BUF:
2523 case PTR_TO_RDONLY_BUF_OR_NULL:
2524 case PTR_TO_RDWR_BUF:
2525 case PTR_TO_RDWR_BUF_OR_NULL:
2526 case PTR_TO_PERCPU_BTF_ID:
2528 case PTR_TO_MEM_OR_NULL:
2530 case PTR_TO_MAP_KEY:
2537 /* Does this register contain a constant zero? */
2538 static bool register_is_null(struct bpf_reg_state *reg)
2540 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
2543 static bool register_is_const(struct bpf_reg_state *reg)
2545 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
2548 static bool __is_scalar_unbounded(struct bpf_reg_state *reg)
2550 return tnum_is_unknown(reg->var_off) &&
2551 reg->smin_value == S64_MIN && reg->smax_value == S64_MAX &&
2552 reg->umin_value == 0 && reg->umax_value == U64_MAX &&
2553 reg->s32_min_value == S32_MIN && reg->s32_max_value == S32_MAX &&
2554 reg->u32_min_value == 0 && reg->u32_max_value == U32_MAX;
2557 static bool register_is_bounded(struct bpf_reg_state *reg)
2559 return reg->type == SCALAR_VALUE && !__is_scalar_unbounded(reg);
2562 static bool __is_pointer_value(bool allow_ptr_leaks,
2563 const struct bpf_reg_state *reg)
2565 if (allow_ptr_leaks)
2568 return reg->type != SCALAR_VALUE;
2571 static void save_register_state(struct bpf_func_state *state,
2572 int spi, struct bpf_reg_state *reg)
2576 state->stack[spi].spilled_ptr = *reg;
2577 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2579 for (i = 0; i < BPF_REG_SIZE; i++)
2580 state->stack[spi].slot_type[i] = STACK_SPILL;
2583 /* check_stack_{read,write}_fixed_off functions track spill/fill of registers,
2584 * stack boundary and alignment are checked in check_mem_access()
2586 static int check_stack_write_fixed_off(struct bpf_verifier_env *env,
2587 /* stack frame we're writing to */
2588 struct bpf_func_state *state,
2589 int off, int size, int value_regno,
2592 struct bpf_func_state *cur; /* state of the current function */
2593 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
2594 u32 dst_reg = env->prog->insnsi[insn_idx].dst_reg;
2595 struct bpf_reg_state *reg = NULL;
2597 err = grow_stack_state(state, round_up(slot + 1, BPF_REG_SIZE));
2600 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
2601 * so it's aligned access and [off, off + size) are within stack limits
2603 if (!env->allow_ptr_leaks &&
2604 state->stack[spi].slot_type[0] == STACK_SPILL &&
2605 size != BPF_REG_SIZE) {
2606 verbose(env, "attempt to corrupt spilled pointer on stack\n");
2610 cur = env->cur_state->frame[env->cur_state->curframe];
2611 if (value_regno >= 0)
2612 reg = &cur->regs[value_regno];
2613 if (!env->bypass_spec_v4) {
2614 bool sanitize = reg && is_spillable_regtype(reg->type);
2616 for (i = 0; i < size; i++) {
2617 if (state->stack[spi].slot_type[i] == STACK_INVALID) {
2624 env->insn_aux_data[insn_idx].sanitize_stack_spill = true;
2627 if (reg && size == BPF_REG_SIZE && register_is_bounded(reg) &&
2628 !register_is_null(reg) && env->bpf_capable) {
2629 if (dst_reg != BPF_REG_FP) {
2630 /* The backtracking logic can only recognize explicit
2631 * stack slot address like [fp - 8]. Other spill of
2632 * scalar via different register has to be conservative.
2633 * Backtrack from here and mark all registers as precise
2634 * that contributed into 'reg' being a constant.
2636 err = mark_chain_precision(env, value_regno);
2640 save_register_state(state, spi, reg);
2641 } else if (reg && is_spillable_regtype(reg->type)) {
2642 /* register containing pointer is being spilled into stack */
2643 if (size != BPF_REG_SIZE) {
2644 verbose_linfo(env, insn_idx, "; ");
2645 verbose(env, "invalid size of register spill\n");
2648 if (state != cur && reg->type == PTR_TO_STACK) {
2649 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
2652 save_register_state(state, spi, reg);
2654 u8 type = STACK_MISC;
2656 /* regular write of data into stack destroys any spilled ptr */
2657 state->stack[spi].spilled_ptr.type = NOT_INIT;
2658 /* Mark slots as STACK_MISC if they belonged to spilled ptr. */
2659 if (state->stack[spi].slot_type[0] == STACK_SPILL)
2660 for (i = 0; i < BPF_REG_SIZE; i++)
2661 state->stack[spi].slot_type[i] = STACK_MISC;
2663 /* only mark the slot as written if all 8 bytes were written
2664 * otherwise read propagation may incorrectly stop too soon
2665 * when stack slots are partially written.
2666 * This heuristic means that read propagation will be
2667 * conservative, since it will add reg_live_read marks
2668 * to stack slots all the way to first state when programs
2669 * writes+reads less than 8 bytes
2671 if (size == BPF_REG_SIZE)
2672 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2674 /* when we zero initialize stack slots mark them as such */
2675 if (reg && register_is_null(reg)) {
2676 /* backtracking doesn't work for STACK_ZERO yet. */
2677 err = mark_chain_precision(env, value_regno);
2683 /* Mark slots affected by this stack write. */
2684 for (i = 0; i < size; i++)
2685 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
2691 /* Write the stack: 'stack[ptr_regno + off] = value_regno'. 'ptr_regno' is
2692 * known to contain a variable offset.
2693 * This function checks whether the write is permitted and conservatively
2694 * tracks the effects of the write, considering that each stack slot in the
2695 * dynamic range is potentially written to.
2697 * 'off' includes 'regno->off'.
2698 * 'value_regno' can be -1, meaning that an unknown value is being written to
2701 * Spilled pointers in range are not marked as written because we don't know
2702 * what's going to be actually written. This means that read propagation for
2703 * future reads cannot be terminated by this write.
2705 * For privileged programs, uninitialized stack slots are considered
2706 * initialized by this write (even though we don't know exactly what offsets
2707 * are going to be written to). The idea is that we don't want the verifier to
2708 * reject future reads that access slots written to through variable offsets.
2710 static int check_stack_write_var_off(struct bpf_verifier_env *env,
2711 /* func where register points to */
2712 struct bpf_func_state *state,
2713 int ptr_regno, int off, int size,
2714 int value_regno, int insn_idx)
2716 struct bpf_func_state *cur; /* state of the current function */
2717 int min_off, max_off;
2719 struct bpf_reg_state *ptr_reg = NULL, *value_reg = NULL;
2720 bool writing_zero = false;
2721 /* set if the fact that we're writing a zero is used to let any
2722 * stack slots remain STACK_ZERO
2724 bool zero_used = false;
2726 cur = env->cur_state->frame[env->cur_state->curframe];
2727 ptr_reg = &cur->regs[ptr_regno];
2728 min_off = ptr_reg->smin_value + off;
2729 max_off = ptr_reg->smax_value + off + size;
2730 if (value_regno >= 0)
2731 value_reg = &cur->regs[value_regno];
2732 if (value_reg && register_is_null(value_reg))
2733 writing_zero = true;
2735 err = grow_stack_state(state, round_up(-min_off, BPF_REG_SIZE));
2740 /* Variable offset writes destroy any spilled pointers in range. */
2741 for (i = min_off; i < max_off; i++) {
2742 u8 new_type, *stype;
2746 spi = slot / BPF_REG_SIZE;
2747 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
2749 if (!env->allow_ptr_leaks
2750 && *stype != NOT_INIT
2751 && *stype != SCALAR_VALUE) {
2752 /* Reject the write if there's are spilled pointers in
2753 * range. If we didn't reject here, the ptr status
2754 * would be erased below (even though not all slots are
2755 * actually overwritten), possibly opening the door to
2758 verbose(env, "spilled ptr in range of var-offset stack write; insn %d, ptr off: %d",
2763 /* Erase all spilled pointers. */
2764 state->stack[spi].spilled_ptr.type = NOT_INIT;
2766 /* Update the slot type. */
2767 new_type = STACK_MISC;
2768 if (writing_zero && *stype == STACK_ZERO) {
2769 new_type = STACK_ZERO;
2772 /* If the slot is STACK_INVALID, we check whether it's OK to
2773 * pretend that it will be initialized by this write. The slot
2774 * might not actually be written to, and so if we mark it as
2775 * initialized future reads might leak uninitialized memory.
2776 * For privileged programs, we will accept such reads to slots
2777 * that may or may not be written because, if we're reject
2778 * them, the error would be too confusing.
2780 if (*stype == STACK_INVALID && !env->allow_uninit_stack) {
2781 verbose(env, "uninit stack in range of var-offset write prohibited for !root; insn %d, off: %d",
2788 /* backtracking doesn't work for STACK_ZERO yet. */
2789 err = mark_chain_precision(env, value_regno);
2796 /* When register 'dst_regno' is assigned some values from stack[min_off,
2797 * max_off), we set the register's type according to the types of the
2798 * respective stack slots. If all the stack values are known to be zeros, then
2799 * so is the destination reg. Otherwise, the register is considered to be
2800 * SCALAR. This function does not deal with register filling; the caller must
2801 * ensure that all spilled registers in the stack range have been marked as
2804 static void mark_reg_stack_read(struct bpf_verifier_env *env,
2805 /* func where src register points to */
2806 struct bpf_func_state *ptr_state,
2807 int min_off, int max_off, int dst_regno)
2809 struct bpf_verifier_state *vstate = env->cur_state;
2810 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2815 for (i = min_off; i < max_off; i++) {
2817 spi = slot / BPF_REG_SIZE;
2818 stype = ptr_state->stack[spi].slot_type;
2819 if (stype[slot % BPF_REG_SIZE] != STACK_ZERO)
2823 if (zeros == max_off - min_off) {
2824 /* any access_size read into register is zero extended,
2825 * so the whole register == const_zero
2827 __mark_reg_const_zero(&state->regs[dst_regno]);
2828 /* backtracking doesn't support STACK_ZERO yet,
2829 * so mark it precise here, so that later
2830 * backtracking can stop here.
2831 * Backtracking may not need this if this register
2832 * doesn't participate in pointer adjustment.
2833 * Forward propagation of precise flag is not
2834 * necessary either. This mark is only to stop
2835 * backtracking. Any register that contributed
2836 * to const 0 was marked precise before spill.
2838 state->regs[dst_regno].precise = true;
2840 /* have read misc data from the stack */
2841 mark_reg_unknown(env, state->regs, dst_regno);
2843 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
2846 /* Read the stack at 'off' and put the results into the register indicated by
2847 * 'dst_regno'. It handles reg filling if the addressed stack slot is a
2850 * 'dst_regno' can be -1, meaning that the read value is not going to a
2853 * The access is assumed to be within the current stack bounds.
2855 static int check_stack_read_fixed_off(struct bpf_verifier_env *env,
2856 /* func where src register points to */
2857 struct bpf_func_state *reg_state,
2858 int off, int size, int dst_regno)
2860 struct bpf_verifier_state *vstate = env->cur_state;
2861 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2862 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
2863 struct bpf_reg_state *reg;
2866 stype = reg_state->stack[spi].slot_type;
2867 reg = ®_state->stack[spi].spilled_ptr;
2869 if (stype[0] == STACK_SPILL) {
2870 if (size != BPF_REG_SIZE) {
2871 if (reg->type != SCALAR_VALUE) {
2872 verbose_linfo(env, env->insn_idx, "; ");
2873 verbose(env, "invalid size of register fill\n");
2876 if (dst_regno >= 0) {
2877 mark_reg_unknown(env, state->regs, dst_regno);
2878 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
2880 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2883 for (i = 1; i < BPF_REG_SIZE; i++) {
2884 if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) {
2885 verbose(env, "corrupted spill memory\n");
2890 if (dst_regno >= 0) {
2891 /* restore register state from stack */
2892 state->regs[dst_regno] = *reg;
2893 /* mark reg as written since spilled pointer state likely
2894 * has its liveness marks cleared by is_state_visited()
2895 * which resets stack/reg liveness for state transitions
2897 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
2898 } else if (__is_pointer_value(env->allow_ptr_leaks, reg)) {
2899 /* If dst_regno==-1, the caller is asking us whether
2900 * it is acceptable to use this value as a SCALAR_VALUE
2902 * We must not allow unprivileged callers to do that
2903 * with spilled pointers.
2905 verbose(env, "leaking pointer from stack off %d\n",
2909 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2913 for (i = 0; i < size; i++) {
2914 type = stype[(slot - i) % BPF_REG_SIZE];
2915 if (type == STACK_MISC)
2917 if (type == STACK_ZERO)
2919 verbose(env, "invalid read from stack off %d+%d size %d\n",
2923 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2925 mark_reg_stack_read(env, reg_state, off, off + size, dst_regno);
2930 enum stack_access_src {
2931 ACCESS_DIRECT = 1, /* the access is performed by an instruction */
2932 ACCESS_HELPER = 2, /* the access is performed by a helper */
2935 static int check_stack_range_initialized(struct bpf_verifier_env *env,
2936 int regno, int off, int access_size,
2937 bool zero_size_allowed,
2938 enum stack_access_src type,
2939 struct bpf_call_arg_meta *meta);
2941 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
2943 return cur_regs(env) + regno;
2946 /* Read the stack at 'ptr_regno + off' and put the result into the register
2948 * 'off' includes the pointer register's fixed offset(i.e. 'ptr_regno.off'),
2949 * but not its variable offset.
2950 * 'size' is assumed to be <= reg size and the access is assumed to be aligned.
2952 * As opposed to check_stack_read_fixed_off, this function doesn't deal with
2953 * filling registers (i.e. reads of spilled register cannot be detected when
2954 * the offset is not fixed). We conservatively mark 'dst_regno' as containing
2955 * SCALAR_VALUE. That's why we assert that the 'ptr_regno' has a variable
2956 * offset; for a fixed offset check_stack_read_fixed_off should be used
2959 static int check_stack_read_var_off(struct bpf_verifier_env *env,
2960 int ptr_regno, int off, int size, int dst_regno)
2962 /* The state of the source register. */
2963 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
2964 struct bpf_func_state *ptr_state = func(env, reg);
2966 int min_off, max_off;
2968 /* Note that we pass a NULL meta, so raw access will not be permitted.
2970 err = check_stack_range_initialized(env, ptr_regno, off, size,
2971 false, ACCESS_DIRECT, NULL);
2975 min_off = reg->smin_value + off;
2976 max_off = reg->smax_value + off;
2977 mark_reg_stack_read(env, ptr_state, min_off, max_off + size, dst_regno);
2981 /* check_stack_read dispatches to check_stack_read_fixed_off or
2982 * check_stack_read_var_off.
2984 * The caller must ensure that the offset falls within the allocated stack
2987 * 'dst_regno' is a register which will receive the value from the stack. It
2988 * can be -1, meaning that the read value is not going to a register.
2990 static int check_stack_read(struct bpf_verifier_env *env,
2991 int ptr_regno, int off, int size,
2994 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
2995 struct bpf_func_state *state = func(env, reg);
2997 /* Some accesses are only permitted with a static offset. */
2998 bool var_off = !tnum_is_const(reg->var_off);
3000 /* The offset is required to be static when reads don't go to a
3001 * register, in order to not leak pointers (see
3002 * check_stack_read_fixed_off).
3004 if (dst_regno < 0 && var_off) {
3007 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3008 verbose(env, "variable offset stack pointer cannot be passed into helper function; var_off=%s off=%d size=%d\n",
3012 /* Variable offset is prohibited for unprivileged mode for simplicity
3013 * since it requires corresponding support in Spectre masking for stack
3014 * ALU. See also retrieve_ptr_limit().
3016 if (!env->bypass_spec_v1 && var_off) {
3019 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3020 verbose(env, "R%d variable offset stack access prohibited for !root, var_off=%s\n",
3026 off += reg->var_off.value;
3027 err = check_stack_read_fixed_off(env, state, off, size,
3030 /* Variable offset stack reads need more conservative handling
3031 * than fixed offset ones. Note that dst_regno >= 0 on this
3034 err = check_stack_read_var_off(env, ptr_regno, off, size,
3041 /* check_stack_write dispatches to check_stack_write_fixed_off or
3042 * check_stack_write_var_off.
3044 * 'ptr_regno' is the register used as a pointer into the stack.
3045 * 'off' includes 'ptr_regno->off', but not its variable offset (if any).
3046 * 'value_regno' is the register whose value we're writing to the stack. It can
3047 * be -1, meaning that we're not writing from a register.
3049 * The caller must ensure that the offset falls within the maximum stack size.
3051 static int check_stack_write(struct bpf_verifier_env *env,
3052 int ptr_regno, int off, int size,
3053 int value_regno, int insn_idx)
3055 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
3056 struct bpf_func_state *state = func(env, reg);
3059 if (tnum_is_const(reg->var_off)) {
3060 off += reg->var_off.value;
3061 err = check_stack_write_fixed_off(env, state, off, size,
3062 value_regno, insn_idx);
3064 /* Variable offset stack reads need more conservative handling
3065 * than fixed offset ones.
3067 err = check_stack_write_var_off(env, state,
3068 ptr_regno, off, size,
3069 value_regno, insn_idx);
3074 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
3075 int off, int size, enum bpf_access_type type)
3077 struct bpf_reg_state *regs = cur_regs(env);
3078 struct bpf_map *map = regs[regno].map_ptr;
3079 u32 cap = bpf_map_flags_to_cap(map);
3081 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
3082 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
3083 map->value_size, off, size);
3087 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
3088 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
3089 map->value_size, off, size);
3096 /* check read/write into memory region (e.g., map value, ringbuf sample, etc) */
3097 static int __check_mem_access(struct bpf_verifier_env *env, int regno,
3098 int off, int size, u32 mem_size,
3099 bool zero_size_allowed)
3101 bool size_ok = size > 0 || (size == 0 && zero_size_allowed);
3102 struct bpf_reg_state *reg;
3104 if (off >= 0 && size_ok && (u64)off + size <= mem_size)
3107 reg = &cur_regs(env)[regno];
3108 switch (reg->type) {
3109 case PTR_TO_MAP_KEY:
3110 verbose(env, "invalid access to map key, key_size=%d off=%d size=%d\n",
3111 mem_size, off, size);
3113 case PTR_TO_MAP_VALUE:
3114 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
3115 mem_size, off, size);
3118 case PTR_TO_PACKET_META:
3119 case PTR_TO_PACKET_END:
3120 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
3121 off, size, regno, reg->id, off, mem_size);
3125 verbose(env, "invalid access to memory, mem_size=%u off=%d size=%d\n",
3126 mem_size, off, size);
3132 /* check read/write into a memory region with possible variable offset */
3133 static int check_mem_region_access(struct bpf_verifier_env *env, u32 regno,
3134 int off, int size, u32 mem_size,
3135 bool zero_size_allowed)
3137 struct bpf_verifier_state *vstate = env->cur_state;
3138 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3139 struct bpf_reg_state *reg = &state->regs[regno];
3142 /* We may have adjusted the register pointing to memory region, so we
3143 * need to try adding each of min_value and max_value to off
3144 * to make sure our theoretical access will be safe.
3146 if (env->log.level & BPF_LOG_LEVEL)
3147 print_verifier_state(env, state);
3149 /* The minimum value is only important with signed
3150 * comparisons where we can't assume the floor of a
3151 * value is 0. If we are using signed variables for our
3152 * index'es we need to make sure that whatever we use
3153 * will have a set floor within our range.
3155 if (reg->smin_value < 0 &&
3156 (reg->smin_value == S64_MIN ||
3157 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
3158 reg->smin_value + off < 0)) {
3159 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3163 err = __check_mem_access(env, regno, reg->smin_value + off, size,
3164 mem_size, zero_size_allowed);
3166 verbose(env, "R%d min value is outside of the allowed memory range\n",
3171 /* If we haven't set a max value then we need to bail since we can't be
3172 * sure we won't do bad things.
3173 * If reg->umax_value + off could overflow, treat that as unbounded too.
3175 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
3176 verbose(env, "R%d unbounded memory access, make sure to bounds check any such access\n",
3180 err = __check_mem_access(env, regno, reg->umax_value + off, size,
3181 mem_size, zero_size_allowed);
3183 verbose(env, "R%d max value is outside of the allowed memory range\n",
3191 /* check read/write into a map element with possible variable offset */
3192 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
3193 int off, int size, bool zero_size_allowed)
3195 struct bpf_verifier_state *vstate = env->cur_state;
3196 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3197 struct bpf_reg_state *reg = &state->regs[regno];
3198 struct bpf_map *map = reg->map_ptr;
3201 err = check_mem_region_access(env, regno, off, size, map->value_size,
3206 if (map_value_has_spin_lock(map)) {
3207 u32 lock = map->spin_lock_off;
3209 /* if any part of struct bpf_spin_lock can be touched by
3210 * load/store reject this program.
3211 * To check that [x1, x2) overlaps with [y1, y2)
3212 * it is sufficient to check x1 < y2 && y1 < x2.
3214 if (reg->smin_value + off < lock + sizeof(struct bpf_spin_lock) &&
3215 lock < reg->umax_value + off + size) {
3216 verbose(env, "bpf_spin_lock cannot be accessed directly by load/store\n");
3223 #define MAX_PACKET_OFF 0xffff
3225 static enum bpf_prog_type resolve_prog_type(struct bpf_prog *prog)
3227 return prog->aux->dst_prog ? prog->aux->dst_prog->type : prog->type;
3230 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
3231 const struct bpf_call_arg_meta *meta,
3232 enum bpf_access_type t)
3234 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
3236 switch (prog_type) {
3237 /* Program types only with direct read access go here! */
3238 case BPF_PROG_TYPE_LWT_IN:
3239 case BPF_PROG_TYPE_LWT_OUT:
3240 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
3241 case BPF_PROG_TYPE_SK_REUSEPORT:
3242 case BPF_PROG_TYPE_FLOW_DISSECTOR:
3243 case BPF_PROG_TYPE_CGROUP_SKB:
3248 /* Program types with direct read + write access go here! */
3249 case BPF_PROG_TYPE_SCHED_CLS:
3250 case BPF_PROG_TYPE_SCHED_ACT:
3251 case BPF_PROG_TYPE_XDP:
3252 case BPF_PROG_TYPE_LWT_XMIT:
3253 case BPF_PROG_TYPE_SK_SKB:
3254 case BPF_PROG_TYPE_SK_MSG:
3256 return meta->pkt_access;
3258 env->seen_direct_write = true;
3261 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
3263 env->seen_direct_write = true;
3272 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
3273 int size, bool zero_size_allowed)
3275 struct bpf_reg_state *regs = cur_regs(env);
3276 struct bpf_reg_state *reg = ®s[regno];
3279 /* We may have added a variable offset to the packet pointer; but any
3280 * reg->range we have comes after that. We are only checking the fixed
3284 /* We don't allow negative numbers, because we aren't tracking enough
3285 * detail to prove they're safe.
3287 if (reg->smin_value < 0) {
3288 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3293 err = reg->range < 0 ? -EINVAL :
3294 __check_mem_access(env, regno, off, size, reg->range,
3297 verbose(env, "R%d offset is outside of the packet\n", regno);
3301 /* __check_mem_access has made sure "off + size - 1" is within u16.
3302 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
3303 * otherwise find_good_pkt_pointers would have refused to set range info
3304 * that __check_mem_access would have rejected this pkt access.
3305 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
3307 env->prog->aux->max_pkt_offset =
3308 max_t(u32, env->prog->aux->max_pkt_offset,
3309 off + reg->umax_value + size - 1);
3314 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
3315 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
3316 enum bpf_access_type t, enum bpf_reg_type *reg_type,
3317 struct btf **btf, u32 *btf_id)
3319 struct bpf_insn_access_aux info = {
3320 .reg_type = *reg_type,
3324 if (env->ops->is_valid_access &&
3325 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
3326 /* A non zero info.ctx_field_size indicates that this field is a
3327 * candidate for later verifier transformation to load the whole
3328 * field and then apply a mask when accessed with a narrower
3329 * access than actual ctx access size. A zero info.ctx_field_size
3330 * will only allow for whole field access and rejects any other
3331 * type of narrower access.
3333 *reg_type = info.reg_type;
3335 if (*reg_type == PTR_TO_BTF_ID || *reg_type == PTR_TO_BTF_ID_OR_NULL) {
3337 *btf_id = info.btf_id;
3339 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
3341 /* remember the offset of last byte accessed in ctx */
3342 if (env->prog->aux->max_ctx_offset < off + size)
3343 env->prog->aux->max_ctx_offset = off + size;
3347 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
3351 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
3354 if (size < 0 || off < 0 ||
3355 (u64)off + size > sizeof(struct bpf_flow_keys)) {
3356 verbose(env, "invalid access to flow keys off=%d size=%d\n",
3363 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
3364 u32 regno, int off, int size,
3365 enum bpf_access_type t)
3367 struct bpf_reg_state *regs = cur_regs(env);
3368 struct bpf_reg_state *reg = ®s[regno];
3369 struct bpf_insn_access_aux info = {};
3372 if (reg->smin_value < 0) {
3373 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3378 switch (reg->type) {
3379 case PTR_TO_SOCK_COMMON:
3380 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
3383 valid = bpf_sock_is_valid_access(off, size, t, &info);
3385 case PTR_TO_TCP_SOCK:
3386 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
3388 case PTR_TO_XDP_SOCK:
3389 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
3397 env->insn_aux_data[insn_idx].ctx_field_size =
3398 info.ctx_field_size;
3402 verbose(env, "R%d invalid %s access off=%d size=%d\n",
3403 regno, reg_type_str[reg->type], off, size);
3408 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
3410 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
3413 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
3415 const struct bpf_reg_state *reg = reg_state(env, regno);
3417 return reg->type == PTR_TO_CTX;
3420 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
3422 const struct bpf_reg_state *reg = reg_state(env, regno);
3424 return type_is_sk_pointer(reg->type);
3427 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
3429 const struct bpf_reg_state *reg = reg_state(env, regno);
3431 return type_is_pkt_pointer(reg->type);
3434 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
3436 const struct bpf_reg_state *reg = reg_state(env, regno);
3438 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
3439 return reg->type == PTR_TO_FLOW_KEYS;
3442 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
3443 const struct bpf_reg_state *reg,
3444 int off, int size, bool strict)
3446 struct tnum reg_off;
3449 /* Byte size accesses are always allowed. */
3450 if (!strict || size == 1)
3453 /* For platforms that do not have a Kconfig enabling
3454 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
3455 * NET_IP_ALIGN is universally set to '2'. And on platforms
3456 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
3457 * to this code only in strict mode where we want to emulate
3458 * the NET_IP_ALIGN==2 checking. Therefore use an
3459 * unconditional IP align value of '2'.
3463 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
3464 if (!tnum_is_aligned(reg_off, size)) {
3467 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3469 "misaligned packet access off %d+%s+%d+%d size %d\n",
3470 ip_align, tn_buf, reg->off, off, size);
3477 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
3478 const struct bpf_reg_state *reg,
3479 const char *pointer_desc,
3480 int off, int size, bool strict)
3482 struct tnum reg_off;
3484 /* Byte size accesses are always allowed. */
3485 if (!strict || size == 1)
3488 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
3489 if (!tnum_is_aligned(reg_off, size)) {
3492 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3493 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
3494 pointer_desc, tn_buf, reg->off, off, size);
3501 static int check_ptr_alignment(struct bpf_verifier_env *env,
3502 const struct bpf_reg_state *reg, int off,
3503 int size, bool strict_alignment_once)
3505 bool strict = env->strict_alignment || strict_alignment_once;
3506 const char *pointer_desc = "";
3508 switch (reg->type) {
3510 case PTR_TO_PACKET_META:
3511 /* Special case, because of NET_IP_ALIGN. Given metadata sits
3512 * right in front, treat it the very same way.
3514 return check_pkt_ptr_alignment(env, reg, off, size, strict);
3515 case PTR_TO_FLOW_KEYS:
3516 pointer_desc = "flow keys ";
3518 case PTR_TO_MAP_KEY:
3519 pointer_desc = "key ";
3521 case PTR_TO_MAP_VALUE:
3522 pointer_desc = "value ";
3525 pointer_desc = "context ";
3528 pointer_desc = "stack ";
3529 /* The stack spill tracking logic in check_stack_write_fixed_off()
3530 * and check_stack_read_fixed_off() relies on stack accesses being
3536 pointer_desc = "sock ";
3538 case PTR_TO_SOCK_COMMON:
3539 pointer_desc = "sock_common ";
3541 case PTR_TO_TCP_SOCK:
3542 pointer_desc = "tcp_sock ";
3544 case PTR_TO_XDP_SOCK:
3545 pointer_desc = "xdp_sock ";
3550 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
3554 static int update_stack_depth(struct bpf_verifier_env *env,
3555 const struct bpf_func_state *func,
3558 u16 stack = env->subprog_info[func->subprogno].stack_depth;
3563 /* update known max for given subprogram */
3564 env->subprog_info[func->subprogno].stack_depth = -off;
3568 /* starting from main bpf function walk all instructions of the function
3569 * and recursively walk all callees that given function can call.
3570 * Ignore jump and exit insns.
3571 * Since recursion is prevented by check_cfg() this algorithm
3572 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
3574 static int check_max_stack_depth(struct bpf_verifier_env *env)
3576 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
3577 struct bpf_subprog_info *subprog = env->subprog_info;
3578 struct bpf_insn *insn = env->prog->insnsi;
3579 bool tail_call_reachable = false;
3580 int ret_insn[MAX_CALL_FRAMES];
3581 int ret_prog[MAX_CALL_FRAMES];
3585 /* protect against potential stack overflow that might happen when
3586 * bpf2bpf calls get combined with tailcalls. Limit the caller's stack
3587 * depth for such case down to 256 so that the worst case scenario
3588 * would result in 8k stack size (32 which is tailcall limit * 256 =
3591 * To get the idea what might happen, see an example:
3592 * func1 -> sub rsp, 128
3593 * subfunc1 -> sub rsp, 256
3594 * tailcall1 -> add rsp, 256
3595 * func2 -> sub rsp, 192 (total stack size = 128 + 192 = 320)
3596 * subfunc2 -> sub rsp, 64
3597 * subfunc22 -> sub rsp, 128
3598 * tailcall2 -> add rsp, 128
3599 * func3 -> sub rsp, 32 (total stack size 128 + 192 + 64 + 32 = 416)
3601 * tailcall will unwind the current stack frame but it will not get rid
3602 * of caller's stack as shown on the example above.
3604 if (idx && subprog[idx].has_tail_call && depth >= 256) {
3606 "tail_calls are not allowed when call stack of previous frames is %d bytes. Too large\n",
3610 /* round up to 32-bytes, since this is granularity
3611 * of interpreter stack size
3613 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
3614 if (depth > MAX_BPF_STACK) {
3615 verbose(env, "combined stack size of %d calls is %d. Too large\n",
3620 subprog_end = subprog[idx + 1].start;
3621 for (; i < subprog_end; i++) {
3622 if (!bpf_pseudo_call(insn + i) && !bpf_pseudo_func(insn + i))
3624 /* remember insn and function to return to */
3625 ret_insn[frame] = i + 1;
3626 ret_prog[frame] = idx;
3628 /* find the callee */
3629 i = i + insn[i].imm + 1;
3630 idx = find_subprog(env, i);
3632 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3637 if (subprog[idx].has_tail_call)
3638 tail_call_reachable = true;
3641 if (frame >= MAX_CALL_FRAMES) {
3642 verbose(env, "the call stack of %d frames is too deep !\n",
3648 /* if tail call got detected across bpf2bpf calls then mark each of the
3649 * currently present subprog frames as tail call reachable subprogs;
3650 * this info will be utilized by JIT so that we will be preserving the
3651 * tail call counter throughout bpf2bpf calls combined with tailcalls
3653 if (tail_call_reachable)
3654 for (j = 0; j < frame; j++)
3655 subprog[ret_prog[j]].tail_call_reachable = true;
3656 if (subprog[0].tail_call_reachable)
3657 env->prog->aux->tail_call_reachable = true;
3659 /* end of for() loop means the last insn of the 'subprog'
3660 * was reached. Doesn't matter whether it was JA or EXIT
3664 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
3666 i = ret_insn[frame];
3667 idx = ret_prog[frame];
3671 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
3672 static int get_callee_stack_depth(struct bpf_verifier_env *env,
3673 const struct bpf_insn *insn, int idx)
3675 int start = idx + insn->imm + 1, subprog;
3677 subprog = find_subprog(env, start);
3679 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3683 return env->subprog_info[subprog].stack_depth;
3687 int check_ctx_reg(struct bpf_verifier_env *env,
3688 const struct bpf_reg_state *reg, int regno)
3690 /* Access to ctx or passing it to a helper is only allowed in
3691 * its original, unmodified form.
3695 verbose(env, "dereference of modified ctx ptr R%d off=%d disallowed\n",
3700 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3703 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3704 verbose(env, "variable ctx access var_off=%s disallowed\n", tn_buf);
3711 static int __check_buffer_access(struct bpf_verifier_env *env,
3712 const char *buf_info,
3713 const struct bpf_reg_state *reg,
3714 int regno, int off, int size)
3718 "R%d invalid %s buffer access: off=%d, size=%d\n",
3719 regno, buf_info, off, size);
3722 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3725 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3727 "R%d invalid variable buffer offset: off=%d, var_off=%s\n",
3728 regno, off, tn_buf);
3735 static int check_tp_buffer_access(struct bpf_verifier_env *env,
3736 const struct bpf_reg_state *reg,
3737 int regno, int off, int size)
3741 err = __check_buffer_access(env, "tracepoint", reg, regno, off, size);
3745 if (off + size > env->prog->aux->max_tp_access)
3746 env->prog->aux->max_tp_access = off + size;
3751 static int check_buffer_access(struct bpf_verifier_env *env,
3752 const struct bpf_reg_state *reg,
3753 int regno, int off, int size,
3754 bool zero_size_allowed,
3755 const char *buf_info,
3760 err = __check_buffer_access(env, buf_info, reg, regno, off, size);
3764 if (off + size > *max_access)
3765 *max_access = off + size;
3770 /* BPF architecture zero extends alu32 ops into 64-bit registesr */
3771 static void zext_32_to_64(struct bpf_reg_state *reg)
3773 reg->var_off = tnum_subreg(reg->var_off);
3774 __reg_assign_32_into_64(reg);
3777 /* truncate register to smaller size (in bytes)
3778 * must be called with size < BPF_REG_SIZE
3780 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
3784 /* clear high bits in bit representation */
3785 reg->var_off = tnum_cast(reg->var_off, size);
3787 /* fix arithmetic bounds */
3788 mask = ((u64)1 << (size * 8)) - 1;
3789 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
3790 reg->umin_value &= mask;
3791 reg->umax_value &= mask;
3793 reg->umin_value = 0;
3794 reg->umax_value = mask;
3796 reg->smin_value = reg->umin_value;
3797 reg->smax_value = reg->umax_value;
3799 /* If size is smaller than 32bit register the 32bit register
3800 * values are also truncated so we push 64-bit bounds into
3801 * 32-bit bounds. Above were truncated < 32-bits already.
3805 __reg_combine_64_into_32(reg);
3808 static bool bpf_map_is_rdonly(const struct bpf_map *map)
3810 return (map->map_flags & BPF_F_RDONLY_PROG) && map->frozen;
3813 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val)
3819 err = map->ops->map_direct_value_addr(map, &addr, off);
3822 ptr = (void *)(long)addr + off;
3826 *val = (u64)*(u8 *)ptr;
3829 *val = (u64)*(u16 *)ptr;
3832 *val = (u64)*(u32 *)ptr;
3843 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
3844 struct bpf_reg_state *regs,
3845 int regno, int off, int size,
3846 enum bpf_access_type atype,
3849 struct bpf_reg_state *reg = regs + regno;
3850 const struct btf_type *t = btf_type_by_id(reg->btf, reg->btf_id);
3851 const char *tname = btf_name_by_offset(reg->btf, t->name_off);
3857 "R%d is ptr_%s invalid negative access: off=%d\n",
3861 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3864 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3866 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
3867 regno, tname, off, tn_buf);
3871 if (env->ops->btf_struct_access) {
3872 ret = env->ops->btf_struct_access(&env->log, reg->btf, t,
3873 off, size, atype, &btf_id);
3875 if (atype != BPF_READ) {
3876 verbose(env, "only read is supported\n");
3880 ret = btf_struct_access(&env->log, reg->btf, t, off, size,
3887 if (atype == BPF_READ && value_regno >= 0)
3888 mark_btf_ld_reg(env, regs, value_regno, ret, reg->btf, btf_id);
3893 static int check_ptr_to_map_access(struct bpf_verifier_env *env,
3894 struct bpf_reg_state *regs,
3895 int regno, int off, int size,
3896 enum bpf_access_type atype,
3899 struct bpf_reg_state *reg = regs + regno;
3900 struct bpf_map *map = reg->map_ptr;
3901 const struct btf_type *t;
3907 verbose(env, "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n");
3911 if (!map->ops->map_btf_id || !*map->ops->map_btf_id) {
3912 verbose(env, "map_ptr access not supported for map type %d\n",
3917 t = btf_type_by_id(btf_vmlinux, *map->ops->map_btf_id);
3918 tname = btf_name_by_offset(btf_vmlinux, t->name_off);
3920 if (!env->allow_ptr_to_map_access) {
3922 "%s access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
3928 verbose(env, "R%d is %s invalid negative access: off=%d\n",
3933 if (atype != BPF_READ) {
3934 verbose(env, "only read from %s is supported\n", tname);
3938 ret = btf_struct_access(&env->log, btf_vmlinux, t, off, size, atype, &btf_id);
3942 if (value_regno >= 0)
3943 mark_btf_ld_reg(env, regs, value_regno, ret, btf_vmlinux, btf_id);
3948 /* Check that the stack access at the given offset is within bounds. The
3949 * maximum valid offset is -1.
3951 * The minimum valid offset is -MAX_BPF_STACK for writes, and
3952 * -state->allocated_stack for reads.
3954 static int check_stack_slot_within_bounds(int off,
3955 struct bpf_func_state *state,
3956 enum bpf_access_type t)
3961 min_valid_off = -MAX_BPF_STACK;
3963 min_valid_off = -state->allocated_stack;
3965 if (off < min_valid_off || off > -1)
3970 /* Check that the stack access at 'regno + off' falls within the maximum stack
3973 * 'off' includes `regno->offset`, but not its dynamic part (if any).
3975 static int check_stack_access_within_bounds(
3976 struct bpf_verifier_env *env,
3977 int regno, int off, int access_size,
3978 enum stack_access_src src, enum bpf_access_type type)
3980 struct bpf_reg_state *regs = cur_regs(env);
3981 struct bpf_reg_state *reg = regs + regno;
3982 struct bpf_func_state *state = func(env, reg);
3983 int min_off, max_off;
3987 if (src == ACCESS_HELPER)
3988 /* We don't know if helpers are reading or writing (or both). */
3989 err_extra = " indirect access to";
3990 else if (type == BPF_READ)
3991 err_extra = " read from";
3993 err_extra = " write to";
3995 if (tnum_is_const(reg->var_off)) {
3996 min_off = reg->var_off.value + off;
3997 if (access_size > 0)
3998 max_off = min_off + access_size - 1;
4002 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
4003 reg->smin_value <= -BPF_MAX_VAR_OFF) {
4004 verbose(env, "invalid unbounded variable-offset%s stack R%d\n",
4008 min_off = reg->smin_value + off;
4009 if (access_size > 0)
4010 max_off = reg->smax_value + off + access_size - 1;
4015 err = check_stack_slot_within_bounds(min_off, state, type);
4017 err = check_stack_slot_within_bounds(max_off, state, type);
4020 if (tnum_is_const(reg->var_off)) {
4021 verbose(env, "invalid%s stack R%d off=%d size=%d\n",
4022 err_extra, regno, off, access_size);
4026 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4027 verbose(env, "invalid variable-offset%s stack R%d var_off=%s size=%d\n",
4028 err_extra, regno, tn_buf, access_size);
4034 /* check whether memory at (regno + off) is accessible for t = (read | write)
4035 * if t==write, value_regno is a register which value is stored into memory
4036 * if t==read, value_regno is a register which will receive the value from memory
4037 * if t==write && value_regno==-1, some unknown value is stored into memory
4038 * if t==read && value_regno==-1, don't care what we read from memory
4040 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
4041 int off, int bpf_size, enum bpf_access_type t,
4042 int value_regno, bool strict_alignment_once)
4044 struct bpf_reg_state *regs = cur_regs(env);
4045 struct bpf_reg_state *reg = regs + regno;
4046 struct bpf_func_state *state;
4049 size = bpf_size_to_bytes(bpf_size);
4053 /* alignment checks will add in reg->off themselves */
4054 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
4058 /* for access checks, reg->off is just part of off */
4061 if (reg->type == PTR_TO_MAP_KEY) {
4062 if (t == BPF_WRITE) {
4063 verbose(env, "write to change key R%d not allowed\n", regno);
4067 err = check_mem_region_access(env, regno, off, size,
4068 reg->map_ptr->key_size, false);
4071 if (value_regno >= 0)
4072 mark_reg_unknown(env, regs, value_regno);
4073 } else if (reg->type == PTR_TO_MAP_VALUE) {
4074 if (t == BPF_WRITE && value_regno >= 0 &&
4075 is_pointer_value(env, value_regno)) {
4076 verbose(env, "R%d leaks addr into map\n", value_regno);
4079 err = check_map_access_type(env, regno, off, size, t);
4082 err = check_map_access(env, regno, off, size, false);
4083 if (!err && t == BPF_READ && value_regno >= 0) {
4084 struct bpf_map *map = reg->map_ptr;
4086 /* if map is read-only, track its contents as scalars */
4087 if (tnum_is_const(reg->var_off) &&
4088 bpf_map_is_rdonly(map) &&
4089 map->ops->map_direct_value_addr) {
4090 int map_off = off + reg->var_off.value;
4093 err = bpf_map_direct_read(map, map_off, size,
4098 regs[value_regno].type = SCALAR_VALUE;
4099 __mark_reg_known(®s[value_regno], val);
4101 mark_reg_unknown(env, regs, value_regno);
4104 } else if (reg->type == PTR_TO_MEM) {
4105 if (t == BPF_WRITE && value_regno >= 0 &&
4106 is_pointer_value(env, value_regno)) {
4107 verbose(env, "R%d leaks addr into mem\n", value_regno);
4110 err = check_mem_region_access(env, regno, off, size,
4111 reg->mem_size, false);
4112 if (!err && t == BPF_READ && value_regno >= 0)
4113 mark_reg_unknown(env, regs, value_regno);
4114 } else if (reg->type == PTR_TO_CTX) {
4115 enum bpf_reg_type reg_type = SCALAR_VALUE;
4116 struct btf *btf = NULL;
4119 if (t == BPF_WRITE && value_regno >= 0 &&
4120 is_pointer_value(env, value_regno)) {
4121 verbose(env, "R%d leaks addr into ctx\n", value_regno);
4125 err = check_ctx_reg(env, reg, regno);
4129 err = check_ctx_access(env, insn_idx, off, size, t, ®_type, &btf, &btf_id);
4131 verbose_linfo(env, insn_idx, "; ");
4132 if (!err && t == BPF_READ && value_regno >= 0) {
4133 /* ctx access returns either a scalar, or a
4134 * PTR_TO_PACKET[_META,_END]. In the latter
4135 * case, we know the offset is zero.
4137 if (reg_type == SCALAR_VALUE) {
4138 mark_reg_unknown(env, regs, value_regno);
4140 mark_reg_known_zero(env, regs,
4142 if (reg_type_may_be_null(reg_type))
4143 regs[value_regno].id = ++env->id_gen;
4144 /* A load of ctx field could have different
4145 * actual load size with the one encoded in the
4146 * insn. When the dst is PTR, it is for sure not
4149 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
4150 if (reg_type == PTR_TO_BTF_ID ||
4151 reg_type == PTR_TO_BTF_ID_OR_NULL) {
4152 regs[value_regno].btf = btf;
4153 regs[value_regno].btf_id = btf_id;
4156 regs[value_regno].type = reg_type;
4159 } else if (reg->type == PTR_TO_STACK) {
4160 /* Basic bounds checks. */
4161 err = check_stack_access_within_bounds(env, regno, off, size, ACCESS_DIRECT, t);
4165 state = func(env, reg);
4166 err = update_stack_depth(env, state, off);
4171 err = check_stack_read(env, regno, off, size,
4174 err = check_stack_write(env, regno, off, size,
4175 value_regno, insn_idx);
4176 } else if (reg_is_pkt_pointer(reg)) {
4177 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
4178 verbose(env, "cannot write into packet\n");
4181 if (t == BPF_WRITE && value_regno >= 0 &&
4182 is_pointer_value(env, value_regno)) {
4183 verbose(env, "R%d leaks addr into packet\n",
4187 err = check_packet_access(env, regno, off, size, false);
4188 if (!err && t == BPF_READ && value_regno >= 0)
4189 mark_reg_unknown(env, regs, value_regno);
4190 } else if (reg->type == PTR_TO_FLOW_KEYS) {
4191 if (t == BPF_WRITE && value_regno >= 0 &&
4192 is_pointer_value(env, value_regno)) {
4193 verbose(env, "R%d leaks addr into flow keys\n",
4198 err = check_flow_keys_access(env, off, size);
4199 if (!err && t == BPF_READ && value_regno >= 0)
4200 mark_reg_unknown(env, regs, value_regno);
4201 } else if (type_is_sk_pointer(reg->type)) {
4202 if (t == BPF_WRITE) {
4203 verbose(env, "R%d cannot write into %s\n",
4204 regno, reg_type_str[reg->type]);
4207 err = check_sock_access(env, insn_idx, regno, off, size, t);
4208 if (!err && value_regno >= 0)
4209 mark_reg_unknown(env, regs, value_regno);
4210 } else if (reg->type == PTR_TO_TP_BUFFER) {
4211 err = check_tp_buffer_access(env, reg, regno, off, size);
4212 if (!err && t == BPF_READ && value_regno >= 0)
4213 mark_reg_unknown(env, regs, value_regno);
4214 } else if (reg->type == PTR_TO_BTF_ID) {
4215 err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
4217 } else if (reg->type == CONST_PTR_TO_MAP) {
4218 err = check_ptr_to_map_access(env, regs, regno, off, size, t,
4220 } else if (reg->type == PTR_TO_RDONLY_BUF) {
4221 if (t == BPF_WRITE) {
4222 verbose(env, "R%d cannot write into %s\n",
4223 regno, reg_type_str[reg->type]);
4226 err = check_buffer_access(env, reg, regno, off, size, false,
4228 &env->prog->aux->max_rdonly_access);
4229 if (!err && value_regno >= 0)
4230 mark_reg_unknown(env, regs, value_regno);
4231 } else if (reg->type == PTR_TO_RDWR_BUF) {
4232 err = check_buffer_access(env, reg, regno, off, size, false,
4234 &env->prog->aux->max_rdwr_access);
4235 if (!err && t == BPF_READ && value_regno >= 0)
4236 mark_reg_unknown(env, regs, value_regno);
4238 verbose(env, "R%d invalid mem access '%s'\n", regno,
4239 reg_type_str[reg->type]);
4243 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
4244 regs[value_regno].type == SCALAR_VALUE) {
4245 /* b/h/w load zero-extends, mark upper bits as known 0 */
4246 coerce_reg_to_size(®s[value_regno], size);
4251 static int check_atomic(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
4256 switch (insn->imm) {
4258 case BPF_ADD | BPF_FETCH:
4260 case BPF_AND | BPF_FETCH:
4262 case BPF_OR | BPF_FETCH:
4264 case BPF_XOR | BPF_FETCH:
4269 verbose(env, "BPF_ATOMIC uses invalid atomic opcode %02x\n", insn->imm);
4273 if (BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) {
4274 verbose(env, "invalid atomic operand size\n");
4278 /* check src1 operand */
4279 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4283 /* check src2 operand */
4284 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4288 if (insn->imm == BPF_CMPXCHG) {
4289 /* Check comparison of R0 with memory location */
4290 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
4295 if (is_pointer_value(env, insn->src_reg)) {
4296 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
4300 if (is_ctx_reg(env, insn->dst_reg) ||
4301 is_pkt_reg(env, insn->dst_reg) ||
4302 is_flow_key_reg(env, insn->dst_reg) ||
4303 is_sk_reg(env, insn->dst_reg)) {
4304 verbose(env, "BPF_ATOMIC stores into R%d %s is not allowed\n",
4306 reg_type_str[reg_state(env, insn->dst_reg)->type]);
4310 if (insn->imm & BPF_FETCH) {
4311 if (insn->imm == BPF_CMPXCHG)
4312 load_reg = BPF_REG_0;
4314 load_reg = insn->src_reg;
4316 /* check and record load of old value */
4317 err = check_reg_arg(env, load_reg, DST_OP);
4321 /* This instruction accesses a memory location but doesn't
4322 * actually load it into a register.
4327 /* check whether we can read the memory */
4328 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4329 BPF_SIZE(insn->code), BPF_READ, load_reg, true);
4333 /* check whether we can write into the same memory */
4334 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4335 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
4342 /* When register 'regno' is used to read the stack (either directly or through
4343 * a helper function) make sure that it's within stack boundary and, depending
4344 * on the access type, that all elements of the stack are initialized.
4346 * 'off' includes 'regno->off', but not its dynamic part (if any).
4348 * All registers that have been spilled on the stack in the slots within the
4349 * read offsets are marked as read.
4351 static int check_stack_range_initialized(
4352 struct bpf_verifier_env *env, int regno, int off,
4353 int access_size, bool zero_size_allowed,
4354 enum stack_access_src type, struct bpf_call_arg_meta *meta)
4356 struct bpf_reg_state *reg = reg_state(env, regno);
4357 struct bpf_func_state *state = func(env, reg);
4358 int err, min_off, max_off, i, j, slot, spi;
4359 char *err_extra = type == ACCESS_HELPER ? " indirect" : "";
4360 enum bpf_access_type bounds_check_type;
4361 /* Some accesses can write anything into the stack, others are
4364 bool clobber = false;
4366 if (access_size == 0 && !zero_size_allowed) {
4367 verbose(env, "invalid zero-sized read\n");
4371 if (type == ACCESS_HELPER) {
4372 /* The bounds checks for writes are more permissive than for
4373 * reads. However, if raw_mode is not set, we'll do extra
4376 bounds_check_type = BPF_WRITE;
4379 bounds_check_type = BPF_READ;
4381 err = check_stack_access_within_bounds(env, regno, off, access_size,
4382 type, bounds_check_type);
4387 if (tnum_is_const(reg->var_off)) {
4388 min_off = max_off = reg->var_off.value + off;
4390 /* Variable offset is prohibited for unprivileged mode for
4391 * simplicity since it requires corresponding support in
4392 * Spectre masking for stack ALU.
4393 * See also retrieve_ptr_limit().
4395 if (!env->bypass_spec_v1) {
4398 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4399 verbose(env, "R%d%s variable offset stack access prohibited for !root, var_off=%s\n",
4400 regno, err_extra, tn_buf);
4403 /* Only initialized buffer on stack is allowed to be accessed
4404 * with variable offset. With uninitialized buffer it's hard to
4405 * guarantee that whole memory is marked as initialized on
4406 * helper return since specific bounds are unknown what may
4407 * cause uninitialized stack leaking.
4409 if (meta && meta->raw_mode)
4412 min_off = reg->smin_value + off;
4413 max_off = reg->smax_value + off;
4416 if (meta && meta->raw_mode) {
4417 meta->access_size = access_size;
4418 meta->regno = regno;
4422 for (i = min_off; i < max_off + access_size; i++) {
4426 spi = slot / BPF_REG_SIZE;
4427 if (state->allocated_stack <= slot)
4429 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
4430 if (*stype == STACK_MISC)
4432 if (*stype == STACK_ZERO) {
4434 /* helper can write anything into the stack */
4435 *stype = STACK_MISC;
4440 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
4441 state->stack[spi].spilled_ptr.type == PTR_TO_BTF_ID)
4444 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
4445 (state->stack[spi].spilled_ptr.type == SCALAR_VALUE ||
4446 env->allow_ptr_leaks)) {
4448 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
4449 for (j = 0; j < BPF_REG_SIZE; j++)
4450 state->stack[spi].slot_type[j] = STACK_MISC;
4456 if (tnum_is_const(reg->var_off)) {
4457 verbose(env, "invalid%s read from stack R%d off %d+%d size %d\n",
4458 err_extra, regno, min_off, i - min_off, access_size);
4462 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4463 verbose(env, "invalid%s read from stack R%d var_off %s+%d size %d\n",
4464 err_extra, regno, tn_buf, i - min_off, access_size);
4468 /* reading any byte out of 8-byte 'spill_slot' will cause
4469 * the whole slot to be marked as 'read'
4471 mark_reg_read(env, &state->stack[spi].spilled_ptr,
4472 state->stack[spi].spilled_ptr.parent,
4475 return update_stack_depth(env, state, min_off);
4478 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
4479 int access_size, bool zero_size_allowed,
4480 struct bpf_call_arg_meta *meta)
4482 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4484 switch (reg->type) {
4486 case PTR_TO_PACKET_META:
4487 return check_packet_access(env, regno, reg->off, access_size,
4489 case PTR_TO_MAP_KEY:
4490 return check_mem_region_access(env, regno, reg->off, access_size,
4491 reg->map_ptr->key_size, false);
4492 case PTR_TO_MAP_VALUE:
4493 if (check_map_access_type(env, regno, reg->off, access_size,
4494 meta && meta->raw_mode ? BPF_WRITE :
4497 return check_map_access(env, regno, reg->off, access_size,
4500 return check_mem_region_access(env, regno, reg->off,
4501 access_size, reg->mem_size,
4503 case PTR_TO_RDONLY_BUF:
4504 if (meta && meta->raw_mode)
4506 return check_buffer_access(env, reg, regno, reg->off,
4507 access_size, zero_size_allowed,
4509 &env->prog->aux->max_rdonly_access);
4510 case PTR_TO_RDWR_BUF:
4511 return check_buffer_access(env, reg, regno, reg->off,
4512 access_size, zero_size_allowed,
4514 &env->prog->aux->max_rdwr_access);
4516 return check_stack_range_initialized(
4518 regno, reg->off, access_size,
4519 zero_size_allowed, ACCESS_HELPER, meta);
4520 default: /* scalar_value or invalid ptr */
4521 /* Allow zero-byte read from NULL, regardless of pointer type */
4522 if (zero_size_allowed && access_size == 0 &&
4523 register_is_null(reg))
4526 verbose(env, "R%d type=%s expected=%s\n", regno,
4527 reg_type_str[reg->type],
4528 reg_type_str[PTR_TO_STACK]);
4533 int check_mem_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
4534 u32 regno, u32 mem_size)
4536 if (register_is_null(reg))
4539 if (reg_type_may_be_null(reg->type)) {
4540 /* Assuming that the register contains a value check if the memory
4541 * access is safe. Temporarily save and restore the register's state as
4542 * the conversion shouldn't be visible to a caller.
4544 const struct bpf_reg_state saved_reg = *reg;
4547 mark_ptr_not_null_reg(reg);
4548 rv = check_helper_mem_access(env, regno, mem_size, true, NULL);
4553 return check_helper_mem_access(env, regno, mem_size, true, NULL);
4556 /* Implementation details:
4557 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL
4558 * Two bpf_map_lookups (even with the same key) will have different reg->id.
4559 * For traditional PTR_TO_MAP_VALUE the verifier clears reg->id after
4560 * value_or_null->value transition, since the verifier only cares about
4561 * the range of access to valid map value pointer and doesn't care about actual
4562 * address of the map element.
4563 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
4564 * reg->id > 0 after value_or_null->value transition. By doing so
4565 * two bpf_map_lookups will be considered two different pointers that
4566 * point to different bpf_spin_locks.
4567 * The verifier allows taking only one bpf_spin_lock at a time to avoid
4569 * Since only one bpf_spin_lock is allowed the checks are simpler than
4570 * reg_is_refcounted() logic. The verifier needs to remember only
4571 * one spin_lock instead of array of acquired_refs.
4572 * cur_state->active_spin_lock remembers which map value element got locked
4573 * and clears it after bpf_spin_unlock.
4575 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
4578 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4579 struct bpf_verifier_state *cur = env->cur_state;
4580 bool is_const = tnum_is_const(reg->var_off);
4581 struct bpf_map *map = reg->map_ptr;
4582 u64 val = reg->var_off.value;
4586 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
4592 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
4596 if (!map_value_has_spin_lock(map)) {
4597 if (map->spin_lock_off == -E2BIG)
4599 "map '%s' has more than one 'struct bpf_spin_lock'\n",
4601 else if (map->spin_lock_off == -ENOENT)
4603 "map '%s' doesn't have 'struct bpf_spin_lock'\n",
4607 "map '%s' is not a struct type or bpf_spin_lock is mangled\n",
4611 if (map->spin_lock_off != val + reg->off) {
4612 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock'\n",
4617 if (cur->active_spin_lock) {
4619 "Locking two bpf_spin_locks are not allowed\n");
4622 cur->active_spin_lock = reg->id;
4624 if (!cur->active_spin_lock) {
4625 verbose(env, "bpf_spin_unlock without taking a lock\n");
4628 if (cur->active_spin_lock != reg->id) {
4629 verbose(env, "bpf_spin_unlock of different lock\n");
4632 cur->active_spin_lock = 0;
4637 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
4639 return type == ARG_PTR_TO_MEM ||
4640 type == ARG_PTR_TO_MEM_OR_NULL ||
4641 type == ARG_PTR_TO_UNINIT_MEM;
4644 static bool arg_type_is_mem_size(enum bpf_arg_type type)
4646 return type == ARG_CONST_SIZE ||
4647 type == ARG_CONST_SIZE_OR_ZERO;
4650 static bool arg_type_is_alloc_size(enum bpf_arg_type type)
4652 return type == ARG_CONST_ALLOC_SIZE_OR_ZERO;
4655 static bool arg_type_is_int_ptr(enum bpf_arg_type type)
4657 return type == ARG_PTR_TO_INT ||
4658 type == ARG_PTR_TO_LONG;
4661 static int int_ptr_type_to_size(enum bpf_arg_type type)
4663 if (type == ARG_PTR_TO_INT)
4665 else if (type == ARG_PTR_TO_LONG)
4671 static int resolve_map_arg_type(struct bpf_verifier_env *env,
4672 const struct bpf_call_arg_meta *meta,
4673 enum bpf_arg_type *arg_type)
4675 if (!meta->map_ptr) {
4676 /* kernel subsystem misconfigured verifier */
4677 verbose(env, "invalid map_ptr to access map->type\n");
4681 switch (meta->map_ptr->map_type) {
4682 case BPF_MAP_TYPE_SOCKMAP:
4683 case BPF_MAP_TYPE_SOCKHASH:
4684 if (*arg_type == ARG_PTR_TO_MAP_VALUE) {
4685 *arg_type = ARG_PTR_TO_BTF_ID_SOCK_COMMON;
4687 verbose(env, "invalid arg_type for sockmap/sockhash\n");
4698 struct bpf_reg_types {
4699 const enum bpf_reg_type types[10];
4703 static const struct bpf_reg_types map_key_value_types = {
4713 static const struct bpf_reg_types sock_types = {
4723 static const struct bpf_reg_types btf_id_sock_common_types = {
4731 .btf_id = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
4735 static const struct bpf_reg_types mem_types = {
4748 static const struct bpf_reg_types int_ptr_types = {
4758 static const struct bpf_reg_types fullsock_types = { .types = { PTR_TO_SOCKET } };
4759 static const struct bpf_reg_types scalar_types = { .types = { SCALAR_VALUE } };
4760 static const struct bpf_reg_types context_types = { .types = { PTR_TO_CTX } };
4761 static const struct bpf_reg_types alloc_mem_types = { .types = { PTR_TO_MEM } };
4762 static const struct bpf_reg_types const_map_ptr_types = { .types = { CONST_PTR_TO_MAP } };
4763 static const struct bpf_reg_types btf_ptr_types = { .types = { PTR_TO_BTF_ID } };
4764 static const struct bpf_reg_types spin_lock_types = { .types = { PTR_TO_MAP_VALUE } };
4765 static const struct bpf_reg_types percpu_btf_ptr_types = { .types = { PTR_TO_PERCPU_BTF_ID } };
4766 static const struct bpf_reg_types func_ptr_types = { .types = { PTR_TO_FUNC } };
4767 static const struct bpf_reg_types stack_ptr_types = { .types = { PTR_TO_STACK } };
4768 static const struct bpf_reg_types const_str_ptr_types = { .types = { PTR_TO_MAP_VALUE } };
4770 static const struct bpf_reg_types *compatible_reg_types[__BPF_ARG_TYPE_MAX] = {
4771 [ARG_PTR_TO_MAP_KEY] = &map_key_value_types,
4772 [ARG_PTR_TO_MAP_VALUE] = &map_key_value_types,
4773 [ARG_PTR_TO_UNINIT_MAP_VALUE] = &map_key_value_types,
4774 [ARG_PTR_TO_MAP_VALUE_OR_NULL] = &map_key_value_types,
4775 [ARG_CONST_SIZE] = &scalar_types,
4776 [ARG_CONST_SIZE_OR_ZERO] = &scalar_types,
4777 [ARG_CONST_ALLOC_SIZE_OR_ZERO] = &scalar_types,
4778 [ARG_CONST_MAP_PTR] = &const_map_ptr_types,
4779 [ARG_PTR_TO_CTX] = &context_types,
4780 [ARG_PTR_TO_CTX_OR_NULL] = &context_types,
4781 [ARG_PTR_TO_SOCK_COMMON] = &sock_types,
4783 [ARG_PTR_TO_BTF_ID_SOCK_COMMON] = &btf_id_sock_common_types,
4785 [ARG_PTR_TO_SOCKET] = &fullsock_types,
4786 [ARG_PTR_TO_SOCKET_OR_NULL] = &fullsock_types,
4787 [ARG_PTR_TO_BTF_ID] = &btf_ptr_types,
4788 [ARG_PTR_TO_SPIN_LOCK] = &spin_lock_types,
4789 [ARG_PTR_TO_MEM] = &mem_types,
4790 [ARG_PTR_TO_MEM_OR_NULL] = &mem_types,
4791 [ARG_PTR_TO_UNINIT_MEM] = &mem_types,
4792 [ARG_PTR_TO_ALLOC_MEM] = &alloc_mem_types,
4793 [ARG_PTR_TO_ALLOC_MEM_OR_NULL] = &alloc_mem_types,
4794 [ARG_PTR_TO_INT] = &int_ptr_types,
4795 [ARG_PTR_TO_LONG] = &int_ptr_types,
4796 [ARG_PTR_TO_PERCPU_BTF_ID] = &percpu_btf_ptr_types,
4797 [ARG_PTR_TO_FUNC] = &func_ptr_types,
4798 [ARG_PTR_TO_STACK_OR_NULL] = &stack_ptr_types,
4799 [ARG_PTR_TO_CONST_STR] = &const_str_ptr_types,
4802 static int check_reg_type(struct bpf_verifier_env *env, u32 regno,
4803 enum bpf_arg_type arg_type,
4804 const u32 *arg_btf_id)
4806 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4807 enum bpf_reg_type expected, type = reg->type;
4808 const struct bpf_reg_types *compatible;
4811 compatible = compatible_reg_types[arg_type];
4813 verbose(env, "verifier internal error: unsupported arg type %d\n", arg_type);
4817 for (i = 0; i < ARRAY_SIZE(compatible->types); i++) {
4818 expected = compatible->types[i];
4819 if (expected == NOT_INIT)
4822 if (type == expected)
4826 verbose(env, "R%d type=%s expected=", regno, reg_type_str[type]);
4827 for (j = 0; j + 1 < i; j++)
4828 verbose(env, "%s, ", reg_type_str[compatible->types[j]]);
4829 verbose(env, "%s\n", reg_type_str[compatible->types[j]]);
4833 if (type == PTR_TO_BTF_ID) {
4835 if (!compatible->btf_id) {
4836 verbose(env, "verifier internal error: missing arg compatible BTF ID\n");
4839 arg_btf_id = compatible->btf_id;
4842 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
4843 btf_vmlinux, *arg_btf_id)) {
4844 verbose(env, "R%d is of type %s but %s is expected\n",
4845 regno, kernel_type_name(reg->btf, reg->btf_id),
4846 kernel_type_name(btf_vmlinux, *arg_btf_id));
4850 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
4851 verbose(env, "R%d is a pointer to in-kernel struct with non-zero offset\n",
4860 static int check_func_arg(struct bpf_verifier_env *env, u32 arg,
4861 struct bpf_call_arg_meta *meta,
4862 const struct bpf_func_proto *fn)
4864 u32 regno = BPF_REG_1 + arg;
4865 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4866 enum bpf_arg_type arg_type = fn->arg_type[arg];
4867 enum bpf_reg_type type = reg->type;
4870 if (arg_type == ARG_DONTCARE)
4873 err = check_reg_arg(env, regno, SRC_OP);
4877 if (arg_type == ARG_ANYTHING) {
4878 if (is_pointer_value(env, regno)) {
4879 verbose(env, "R%d leaks addr into helper function\n",
4886 if (type_is_pkt_pointer(type) &&
4887 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
4888 verbose(env, "helper access to the packet is not allowed\n");
4892 if (arg_type == ARG_PTR_TO_MAP_VALUE ||
4893 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE ||
4894 arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL) {
4895 err = resolve_map_arg_type(env, meta, &arg_type);
4900 if (register_is_null(reg) && arg_type_may_be_null(arg_type))
4901 /* A NULL register has a SCALAR_VALUE type, so skip
4904 goto skip_type_check;
4906 err = check_reg_type(env, regno, arg_type, fn->arg_btf_id[arg]);
4910 if (type == PTR_TO_CTX) {
4911 err = check_ctx_reg(env, reg, regno);
4917 if (reg->ref_obj_id) {
4918 if (meta->ref_obj_id) {
4919 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
4920 regno, reg->ref_obj_id,
4924 meta->ref_obj_id = reg->ref_obj_id;
4927 if (arg_type == ARG_CONST_MAP_PTR) {
4928 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
4929 meta->map_ptr = reg->map_ptr;
4930 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
4931 /* bpf_map_xxx(..., map_ptr, ..., key) call:
4932 * check that [key, key + map->key_size) are within
4933 * stack limits and initialized
4935 if (!meta->map_ptr) {
4936 /* in function declaration map_ptr must come before
4937 * map_key, so that it's verified and known before
4938 * we have to check map_key here. Otherwise it means
4939 * that kernel subsystem misconfigured verifier
4941 verbose(env, "invalid map_ptr to access map->key\n");
4944 err = check_helper_mem_access(env, regno,
4945 meta->map_ptr->key_size, false,
4947 } else if (arg_type == ARG_PTR_TO_MAP_VALUE ||
4948 (arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL &&
4949 !register_is_null(reg)) ||
4950 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
4951 /* bpf_map_xxx(..., map_ptr, ..., value) call:
4952 * check [value, value + map->value_size) validity
4954 if (!meta->map_ptr) {
4955 /* kernel subsystem misconfigured verifier */
4956 verbose(env, "invalid map_ptr to access map->value\n");
4959 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE);
4960 err = check_helper_mem_access(env, regno,
4961 meta->map_ptr->value_size, false,
4963 } else if (arg_type == ARG_PTR_TO_PERCPU_BTF_ID) {
4965 verbose(env, "Helper has invalid btf_id in R%d\n", regno);
4968 meta->ret_btf = reg->btf;
4969 meta->ret_btf_id = reg->btf_id;
4970 } else if (arg_type == ARG_PTR_TO_SPIN_LOCK) {
4971 if (meta->func_id == BPF_FUNC_spin_lock) {
4972 if (process_spin_lock(env, regno, true))
4974 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
4975 if (process_spin_lock(env, regno, false))
4978 verbose(env, "verifier internal error\n");
4981 } else if (arg_type == ARG_PTR_TO_FUNC) {
4982 meta->subprogno = reg->subprogno;
4983 } else if (arg_type_is_mem_ptr(arg_type)) {
4984 /* The access to this pointer is only checked when we hit the
4985 * next is_mem_size argument below.
4987 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MEM);
4988 } else if (arg_type_is_mem_size(arg_type)) {
4989 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
4991 /* This is used to refine r0 return value bounds for helpers
4992 * that enforce this value as an upper bound on return values.
4993 * See do_refine_retval_range() for helpers that can refine
4994 * the return value. C type of helper is u32 so we pull register
4995 * bound from umax_value however, if negative verifier errors
4996 * out. Only upper bounds can be learned because retval is an
4997 * int type and negative retvals are allowed.
4999 meta->msize_max_value = reg->umax_value;
5001 /* The register is SCALAR_VALUE; the access check
5002 * happens using its boundaries.
5004 if (!tnum_is_const(reg->var_off))
5005 /* For unprivileged variable accesses, disable raw
5006 * mode so that the program is required to
5007 * initialize all the memory that the helper could
5008 * just partially fill up.
5012 if (reg->smin_value < 0) {
5013 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
5018 if (reg->umin_value == 0) {
5019 err = check_helper_mem_access(env, regno - 1, 0,
5026 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
5027 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
5031 err = check_helper_mem_access(env, regno - 1,
5033 zero_size_allowed, meta);
5035 err = mark_chain_precision(env, regno);
5036 } else if (arg_type_is_alloc_size(arg_type)) {
5037 if (!tnum_is_const(reg->var_off)) {
5038 verbose(env, "R%d is not a known constant'\n",
5042 meta->mem_size = reg->var_off.value;
5043 } else if (arg_type_is_int_ptr(arg_type)) {
5044 int size = int_ptr_type_to_size(arg_type);
5046 err = check_helper_mem_access(env, regno, size, false, meta);
5049 err = check_ptr_alignment(env, reg, 0, size, true);
5050 } else if (arg_type == ARG_PTR_TO_CONST_STR) {
5051 struct bpf_map *map = reg->map_ptr;
5056 if (!bpf_map_is_rdonly(map)) {
5057 verbose(env, "R%d does not point to a readonly map'\n", regno);
5061 if (!tnum_is_const(reg->var_off)) {
5062 verbose(env, "R%d is not a constant address'\n", regno);
5066 if (!map->ops->map_direct_value_addr) {
5067 verbose(env, "no direct value access support for this map type\n");
5071 err = check_map_access(env, regno, reg->off,
5072 map->value_size - reg->off, false);
5076 map_off = reg->off + reg->var_off.value;
5077 err = map->ops->map_direct_value_addr(map, &map_addr, map_off);
5079 verbose(env, "direct value access on string failed\n");
5083 str_ptr = (char *)(long)(map_addr);
5084 if (!strnchr(str_ptr + map_off, map->value_size - map_off, 0)) {
5085 verbose(env, "string is not zero-terminated\n");
5093 static bool may_update_sockmap(struct bpf_verifier_env *env, int func_id)
5095 enum bpf_attach_type eatype = env->prog->expected_attach_type;
5096 enum bpf_prog_type type = resolve_prog_type(env->prog);
5098 if (func_id != BPF_FUNC_map_update_elem)
5101 /* It's not possible to get access to a locked struct sock in these
5102 * contexts, so updating is safe.
5105 case BPF_PROG_TYPE_TRACING:
5106 if (eatype == BPF_TRACE_ITER)
5109 case BPF_PROG_TYPE_SOCKET_FILTER:
5110 case BPF_PROG_TYPE_SCHED_CLS:
5111 case BPF_PROG_TYPE_SCHED_ACT:
5112 case BPF_PROG_TYPE_XDP:
5113 case BPF_PROG_TYPE_SK_REUSEPORT:
5114 case BPF_PROG_TYPE_FLOW_DISSECTOR:
5115 case BPF_PROG_TYPE_SK_LOOKUP:
5121 verbose(env, "cannot update sockmap in this context\n");
5125 static bool allow_tail_call_in_subprogs(struct bpf_verifier_env *env)
5127 return env->prog->jit_requested && IS_ENABLED(CONFIG_X86_64);
5130 static int check_map_func_compatibility(struct bpf_verifier_env *env,
5131 struct bpf_map *map, int func_id)
5136 /* We need a two way check, first is from map perspective ... */
5137 switch (map->map_type) {
5138 case BPF_MAP_TYPE_PROG_ARRAY:
5139 if (func_id != BPF_FUNC_tail_call)
5142 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
5143 if (func_id != BPF_FUNC_perf_event_read &&
5144 func_id != BPF_FUNC_perf_event_output &&
5145 func_id != BPF_FUNC_skb_output &&
5146 func_id != BPF_FUNC_perf_event_read_value &&
5147 func_id != BPF_FUNC_xdp_output)
5150 case BPF_MAP_TYPE_RINGBUF:
5151 if (func_id != BPF_FUNC_ringbuf_output &&
5152 func_id != BPF_FUNC_ringbuf_reserve &&
5153 func_id != BPF_FUNC_ringbuf_submit &&
5154 func_id != BPF_FUNC_ringbuf_discard &&
5155 func_id != BPF_FUNC_ringbuf_query)
5158 case BPF_MAP_TYPE_STACK_TRACE:
5159 if (func_id != BPF_FUNC_get_stackid)
5162 case BPF_MAP_TYPE_CGROUP_ARRAY:
5163 if (func_id != BPF_FUNC_skb_under_cgroup &&
5164 func_id != BPF_FUNC_current_task_under_cgroup)
5167 case BPF_MAP_TYPE_CGROUP_STORAGE:
5168 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
5169 if (func_id != BPF_FUNC_get_local_storage)
5172 case BPF_MAP_TYPE_DEVMAP:
5173 case BPF_MAP_TYPE_DEVMAP_HASH:
5174 if (func_id != BPF_FUNC_redirect_map &&
5175 func_id != BPF_FUNC_map_lookup_elem)
5178 /* Restrict bpf side of cpumap and xskmap, open when use-cases
5181 case BPF_MAP_TYPE_CPUMAP:
5182 if (func_id != BPF_FUNC_redirect_map)
5185 case BPF_MAP_TYPE_XSKMAP:
5186 if (func_id != BPF_FUNC_redirect_map &&
5187 func_id != BPF_FUNC_map_lookup_elem)
5190 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
5191 case BPF_MAP_TYPE_HASH_OF_MAPS:
5192 if (func_id != BPF_FUNC_map_lookup_elem)
5195 case BPF_MAP_TYPE_SOCKMAP:
5196 if (func_id != BPF_FUNC_sk_redirect_map &&
5197 func_id != BPF_FUNC_sock_map_update &&
5198 func_id != BPF_FUNC_map_delete_elem &&
5199 func_id != BPF_FUNC_msg_redirect_map &&
5200 func_id != BPF_FUNC_sk_select_reuseport &&
5201 func_id != BPF_FUNC_map_lookup_elem &&
5202 !may_update_sockmap(env, func_id))
5205 case BPF_MAP_TYPE_SOCKHASH:
5206 if (func_id != BPF_FUNC_sk_redirect_hash &&
5207 func_id != BPF_FUNC_sock_hash_update &&
5208 func_id != BPF_FUNC_map_delete_elem &&
5209 func_id != BPF_FUNC_msg_redirect_hash &&
5210 func_id != BPF_FUNC_sk_select_reuseport &&
5211 func_id != BPF_FUNC_map_lookup_elem &&
5212 !may_update_sockmap(env, func_id))
5215 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
5216 if (func_id != BPF_FUNC_sk_select_reuseport)
5219 case BPF_MAP_TYPE_QUEUE:
5220 case BPF_MAP_TYPE_STACK:
5221 if (func_id != BPF_FUNC_map_peek_elem &&
5222 func_id != BPF_FUNC_map_pop_elem &&
5223 func_id != BPF_FUNC_map_push_elem)
5226 case BPF_MAP_TYPE_SK_STORAGE:
5227 if (func_id != BPF_FUNC_sk_storage_get &&
5228 func_id != BPF_FUNC_sk_storage_delete)
5231 case BPF_MAP_TYPE_INODE_STORAGE:
5232 if (func_id != BPF_FUNC_inode_storage_get &&
5233 func_id != BPF_FUNC_inode_storage_delete)
5236 case BPF_MAP_TYPE_TASK_STORAGE:
5237 if (func_id != BPF_FUNC_task_storage_get &&
5238 func_id != BPF_FUNC_task_storage_delete)
5245 /* ... and second from the function itself. */
5247 case BPF_FUNC_tail_call:
5248 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
5250 if (env->subprog_cnt > 1 && !allow_tail_call_in_subprogs(env)) {
5251 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
5255 case BPF_FUNC_perf_event_read:
5256 case BPF_FUNC_perf_event_output:
5257 case BPF_FUNC_perf_event_read_value:
5258 case BPF_FUNC_skb_output:
5259 case BPF_FUNC_xdp_output:
5260 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
5263 case BPF_FUNC_get_stackid:
5264 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
5267 case BPF_FUNC_current_task_under_cgroup:
5268 case BPF_FUNC_skb_under_cgroup:
5269 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
5272 case BPF_FUNC_redirect_map:
5273 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
5274 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
5275 map->map_type != BPF_MAP_TYPE_CPUMAP &&
5276 map->map_type != BPF_MAP_TYPE_XSKMAP)
5279 case BPF_FUNC_sk_redirect_map:
5280 case BPF_FUNC_msg_redirect_map:
5281 case BPF_FUNC_sock_map_update:
5282 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
5285 case BPF_FUNC_sk_redirect_hash:
5286 case BPF_FUNC_msg_redirect_hash:
5287 case BPF_FUNC_sock_hash_update:
5288 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
5291 case BPF_FUNC_get_local_storage:
5292 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
5293 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
5296 case BPF_FUNC_sk_select_reuseport:
5297 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
5298 map->map_type != BPF_MAP_TYPE_SOCKMAP &&
5299 map->map_type != BPF_MAP_TYPE_SOCKHASH)
5302 case BPF_FUNC_map_peek_elem:
5303 case BPF_FUNC_map_pop_elem:
5304 case BPF_FUNC_map_push_elem:
5305 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
5306 map->map_type != BPF_MAP_TYPE_STACK)
5309 case BPF_FUNC_sk_storage_get:
5310 case BPF_FUNC_sk_storage_delete:
5311 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
5314 case BPF_FUNC_inode_storage_get:
5315 case BPF_FUNC_inode_storage_delete:
5316 if (map->map_type != BPF_MAP_TYPE_INODE_STORAGE)
5319 case BPF_FUNC_task_storage_get:
5320 case BPF_FUNC_task_storage_delete:
5321 if (map->map_type != BPF_MAP_TYPE_TASK_STORAGE)
5330 verbose(env, "cannot pass map_type %d into func %s#%d\n",
5331 map->map_type, func_id_name(func_id), func_id);
5335 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
5339 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
5341 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
5343 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
5345 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
5347 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
5350 /* We only support one arg being in raw mode at the moment,
5351 * which is sufficient for the helper functions we have
5357 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
5358 enum bpf_arg_type arg_next)
5360 return (arg_type_is_mem_ptr(arg_curr) &&
5361 !arg_type_is_mem_size(arg_next)) ||
5362 (!arg_type_is_mem_ptr(arg_curr) &&
5363 arg_type_is_mem_size(arg_next));
5366 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
5368 /* bpf_xxx(..., buf, len) call will access 'len'
5369 * bytes from memory 'buf'. Both arg types need
5370 * to be paired, so make sure there's no buggy
5371 * helper function specification.
5373 if (arg_type_is_mem_size(fn->arg1_type) ||
5374 arg_type_is_mem_ptr(fn->arg5_type) ||
5375 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
5376 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
5377 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
5378 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
5384 static bool check_refcount_ok(const struct bpf_func_proto *fn, int func_id)
5388 if (arg_type_may_be_refcounted(fn->arg1_type))
5390 if (arg_type_may_be_refcounted(fn->arg2_type))
5392 if (arg_type_may_be_refcounted(fn->arg3_type))
5394 if (arg_type_may_be_refcounted(fn->arg4_type))
5396 if (arg_type_may_be_refcounted(fn->arg5_type))
5399 /* A reference acquiring function cannot acquire
5400 * another refcounted ptr.
5402 if (may_be_acquire_function(func_id) && count)
5405 /* We only support one arg being unreferenced at the moment,
5406 * which is sufficient for the helper functions we have right now.
5411 static bool check_btf_id_ok(const struct bpf_func_proto *fn)
5415 for (i = 0; i < ARRAY_SIZE(fn->arg_type); i++) {
5416 if (fn->arg_type[i] == ARG_PTR_TO_BTF_ID && !fn->arg_btf_id[i])
5419 if (fn->arg_type[i] != ARG_PTR_TO_BTF_ID && fn->arg_btf_id[i])
5426 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
5428 return check_raw_mode_ok(fn) &&
5429 check_arg_pair_ok(fn) &&
5430 check_btf_id_ok(fn) &&
5431 check_refcount_ok(fn, func_id) ? 0 : -EINVAL;
5434 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
5435 * are now invalid, so turn them into unknown SCALAR_VALUE.
5437 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
5438 struct bpf_func_state *state)
5440 struct bpf_reg_state *regs = state->regs, *reg;
5443 for (i = 0; i < MAX_BPF_REG; i++)
5444 if (reg_is_pkt_pointer_any(®s[i]))
5445 mark_reg_unknown(env, regs, i);
5447 bpf_for_each_spilled_reg(i, state, reg) {
5450 if (reg_is_pkt_pointer_any(reg))
5451 __mark_reg_unknown(env, reg);
5455 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
5457 struct bpf_verifier_state *vstate = env->cur_state;
5460 for (i = 0; i <= vstate->curframe; i++)
5461 __clear_all_pkt_pointers(env, vstate->frame[i]);
5466 BEYOND_PKT_END = -2,
5469 static void mark_pkt_end(struct bpf_verifier_state *vstate, int regn, bool range_open)
5471 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5472 struct bpf_reg_state *reg = &state->regs[regn];
5474 if (reg->type != PTR_TO_PACKET)
5475 /* PTR_TO_PACKET_META is not supported yet */
5478 /* The 'reg' is pkt > pkt_end or pkt >= pkt_end.
5479 * How far beyond pkt_end it goes is unknown.
5480 * if (!range_open) it's the case of pkt >= pkt_end
5481 * if (range_open) it's the case of pkt > pkt_end
5482 * hence this pointer is at least 1 byte bigger than pkt_end
5485 reg->range = BEYOND_PKT_END;
5487 reg->range = AT_PKT_END;
5490 static void release_reg_references(struct bpf_verifier_env *env,
5491 struct bpf_func_state *state,
5494 struct bpf_reg_state *regs = state->regs, *reg;
5497 for (i = 0; i < MAX_BPF_REG; i++)
5498 if (regs[i].ref_obj_id == ref_obj_id)
5499 mark_reg_unknown(env, regs, i);
5501 bpf_for_each_spilled_reg(i, state, reg) {
5504 if (reg->ref_obj_id == ref_obj_id)
5505 __mark_reg_unknown(env, reg);
5509 /* The pointer with the specified id has released its reference to kernel
5510 * resources. Identify all copies of the same pointer and clear the reference.
5512 static int release_reference(struct bpf_verifier_env *env,
5515 struct bpf_verifier_state *vstate = env->cur_state;
5519 err = release_reference_state(cur_func(env), ref_obj_id);
5523 for (i = 0; i <= vstate->curframe; i++)
5524 release_reg_references(env, vstate->frame[i], ref_obj_id);
5529 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
5530 struct bpf_reg_state *regs)
5534 /* after the call registers r0 - r5 were scratched */
5535 for (i = 0; i < CALLER_SAVED_REGS; i++) {
5536 mark_reg_not_init(env, regs, caller_saved[i]);
5537 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
5541 typedef int (*set_callee_state_fn)(struct bpf_verifier_env *env,
5542 struct bpf_func_state *caller,
5543 struct bpf_func_state *callee,
5546 static int __check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
5547 int *insn_idx, int subprog,
5548 set_callee_state_fn set_callee_state_cb)
5550 struct bpf_verifier_state *state = env->cur_state;
5551 struct bpf_func_info_aux *func_info_aux;
5552 struct bpf_func_state *caller, *callee;
5554 bool is_global = false;
5556 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
5557 verbose(env, "the call stack of %d frames is too deep\n",
5558 state->curframe + 2);
5562 caller = state->frame[state->curframe];
5563 if (state->frame[state->curframe + 1]) {
5564 verbose(env, "verifier bug. Frame %d already allocated\n",
5565 state->curframe + 1);
5569 func_info_aux = env->prog->aux->func_info_aux;
5571 is_global = func_info_aux[subprog].linkage == BTF_FUNC_GLOBAL;
5572 err = btf_check_subprog_arg_match(env, subprog, caller->regs);
5577 verbose(env, "Caller passes invalid args into func#%d\n",
5581 if (env->log.level & BPF_LOG_LEVEL)
5583 "Func#%d is global and valid. Skipping.\n",
5585 clear_caller_saved_regs(env, caller->regs);
5587 /* All global functions return a 64-bit SCALAR_VALUE */
5588 mark_reg_unknown(env, caller->regs, BPF_REG_0);
5589 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
5591 /* continue with next insn after call */
5596 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
5599 state->frame[state->curframe + 1] = callee;
5601 /* callee cannot access r0, r6 - r9 for reading and has to write
5602 * into its own stack before reading from it.
5603 * callee can read/write into caller's stack
5605 init_func_state(env, callee,
5606 /* remember the callsite, it will be used by bpf_exit */
5607 *insn_idx /* callsite */,
5608 state->curframe + 1 /* frameno within this callchain */,
5609 subprog /* subprog number within this prog */);
5611 /* Transfer references to the callee */
5612 err = copy_reference_state(callee, caller);
5616 err = set_callee_state_cb(env, caller, callee, *insn_idx);
5620 clear_caller_saved_regs(env, caller->regs);
5622 /* only increment it after check_reg_arg() finished */
5625 /* and go analyze first insn of the callee */
5626 *insn_idx = env->subprog_info[subprog].start - 1;
5628 if (env->log.level & BPF_LOG_LEVEL) {
5629 verbose(env, "caller:\n");
5630 print_verifier_state(env, caller);
5631 verbose(env, "callee:\n");
5632 print_verifier_state(env, callee);
5637 int map_set_for_each_callback_args(struct bpf_verifier_env *env,
5638 struct bpf_func_state *caller,
5639 struct bpf_func_state *callee)
5641 /* bpf_for_each_map_elem(struct bpf_map *map, void *callback_fn,
5642 * void *callback_ctx, u64 flags);
5643 * callback_fn(struct bpf_map *map, void *key, void *value,
5644 * void *callback_ctx);
5646 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
5648 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
5649 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
5650 callee->regs[BPF_REG_2].map_ptr = caller->regs[BPF_REG_1].map_ptr;
5652 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
5653 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
5654 callee->regs[BPF_REG_3].map_ptr = caller->regs[BPF_REG_1].map_ptr;
5656 /* pointer to stack or null */
5657 callee->regs[BPF_REG_4] = caller->regs[BPF_REG_3];
5660 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
5664 static int set_callee_state(struct bpf_verifier_env *env,
5665 struct bpf_func_state *caller,
5666 struct bpf_func_state *callee, int insn_idx)
5670 /* copy r1 - r5 args that callee can access. The copy includes parent
5671 * pointers, which connects us up to the liveness chain
5673 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
5674 callee->regs[i] = caller->regs[i];
5678 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
5681 int subprog, target_insn;
5683 target_insn = *insn_idx + insn->imm + 1;
5684 subprog = find_subprog(env, target_insn);
5686 verbose(env, "verifier bug. No program starts at insn %d\n",
5691 return __check_func_call(env, insn, insn_idx, subprog, set_callee_state);
5694 static int set_map_elem_callback_state(struct bpf_verifier_env *env,
5695 struct bpf_func_state *caller,
5696 struct bpf_func_state *callee,
5699 struct bpf_insn_aux_data *insn_aux = &env->insn_aux_data[insn_idx];
5700 struct bpf_map *map;
5703 if (bpf_map_ptr_poisoned(insn_aux)) {
5704 verbose(env, "tail_call abusing map_ptr\n");
5708 map = BPF_MAP_PTR(insn_aux->map_ptr_state);
5709 if (!map->ops->map_set_for_each_callback_args ||
5710 !map->ops->map_for_each_callback) {
5711 verbose(env, "callback function not allowed for map\n");
5715 err = map->ops->map_set_for_each_callback_args(env, caller, callee);
5719 callee->in_callback_fn = true;
5723 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
5725 struct bpf_verifier_state *state = env->cur_state;
5726 struct bpf_func_state *caller, *callee;
5727 struct bpf_reg_state *r0;
5730 callee = state->frame[state->curframe];
5731 r0 = &callee->regs[BPF_REG_0];
5732 if (r0->type == PTR_TO_STACK) {
5733 /* technically it's ok to return caller's stack pointer
5734 * (or caller's caller's pointer) back to the caller,
5735 * since these pointers are valid. Only current stack
5736 * pointer will be invalid as soon as function exits,
5737 * but let's be conservative
5739 verbose(env, "cannot return stack pointer to the caller\n");
5744 caller = state->frame[state->curframe];
5745 if (callee->in_callback_fn) {
5746 /* enforce R0 return value range [0, 1]. */
5747 struct tnum range = tnum_range(0, 1);
5749 if (r0->type != SCALAR_VALUE) {
5750 verbose(env, "R0 not a scalar value\n");
5753 if (!tnum_in(range, r0->var_off)) {
5754 verbose_invalid_scalar(env, r0, &range, "callback return", "R0");
5758 /* return to the caller whatever r0 had in the callee */
5759 caller->regs[BPF_REG_0] = *r0;
5762 /* Transfer references to the caller */
5763 err = copy_reference_state(caller, callee);
5767 *insn_idx = callee->callsite + 1;
5768 if (env->log.level & BPF_LOG_LEVEL) {
5769 verbose(env, "returning from callee:\n");
5770 print_verifier_state(env, callee);
5771 verbose(env, "to caller at %d:\n", *insn_idx);
5772 print_verifier_state(env, caller);
5774 /* clear everything in the callee */
5775 free_func_state(callee);
5776 state->frame[state->curframe + 1] = NULL;
5780 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
5782 struct bpf_call_arg_meta *meta)
5784 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
5786 if (ret_type != RET_INTEGER ||
5787 (func_id != BPF_FUNC_get_stack &&
5788 func_id != BPF_FUNC_get_task_stack &&
5789 func_id != BPF_FUNC_probe_read_str &&
5790 func_id != BPF_FUNC_probe_read_kernel_str &&
5791 func_id != BPF_FUNC_probe_read_user_str))
5794 ret_reg->smax_value = meta->msize_max_value;
5795 ret_reg->s32_max_value = meta->msize_max_value;
5796 ret_reg->smin_value = -MAX_ERRNO;
5797 ret_reg->s32_min_value = -MAX_ERRNO;
5798 __reg_deduce_bounds(ret_reg);
5799 __reg_bound_offset(ret_reg);
5800 __update_reg_bounds(ret_reg);
5804 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
5805 int func_id, int insn_idx)
5807 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
5808 struct bpf_map *map = meta->map_ptr;
5810 if (func_id != BPF_FUNC_tail_call &&
5811 func_id != BPF_FUNC_map_lookup_elem &&
5812 func_id != BPF_FUNC_map_update_elem &&
5813 func_id != BPF_FUNC_map_delete_elem &&
5814 func_id != BPF_FUNC_map_push_elem &&
5815 func_id != BPF_FUNC_map_pop_elem &&
5816 func_id != BPF_FUNC_map_peek_elem &&
5817 func_id != BPF_FUNC_for_each_map_elem &&
5818 func_id != BPF_FUNC_redirect_map)
5822 verbose(env, "kernel subsystem misconfigured verifier\n");
5826 /* In case of read-only, some additional restrictions
5827 * need to be applied in order to prevent altering the
5828 * state of the map from program side.
5830 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
5831 (func_id == BPF_FUNC_map_delete_elem ||
5832 func_id == BPF_FUNC_map_update_elem ||
5833 func_id == BPF_FUNC_map_push_elem ||
5834 func_id == BPF_FUNC_map_pop_elem)) {
5835 verbose(env, "write into map forbidden\n");
5839 if (!BPF_MAP_PTR(aux->map_ptr_state))
5840 bpf_map_ptr_store(aux, meta->map_ptr,
5841 !meta->map_ptr->bypass_spec_v1);
5842 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
5843 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
5844 !meta->map_ptr->bypass_spec_v1);
5849 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
5850 int func_id, int insn_idx)
5852 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
5853 struct bpf_reg_state *regs = cur_regs(env), *reg;
5854 struct bpf_map *map = meta->map_ptr;
5859 if (func_id != BPF_FUNC_tail_call)
5861 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
5862 verbose(env, "kernel subsystem misconfigured verifier\n");
5866 range = tnum_range(0, map->max_entries - 1);
5867 reg = ®s[BPF_REG_3];
5869 if (!register_is_const(reg) || !tnum_in(range, reg->var_off)) {
5870 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
5874 err = mark_chain_precision(env, BPF_REG_3);
5878 val = reg->var_off.value;
5879 if (bpf_map_key_unseen(aux))
5880 bpf_map_key_store(aux, val);
5881 else if (!bpf_map_key_poisoned(aux) &&
5882 bpf_map_key_immediate(aux) != val)
5883 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
5887 static int check_reference_leak(struct bpf_verifier_env *env)
5889 struct bpf_func_state *state = cur_func(env);
5892 for (i = 0; i < state->acquired_refs; i++) {
5893 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
5894 state->refs[i].id, state->refs[i].insn_idx);
5896 return state->acquired_refs ? -EINVAL : 0;
5899 static int check_bpf_snprintf_call(struct bpf_verifier_env *env,
5900 struct bpf_reg_state *regs)
5902 struct bpf_reg_state *fmt_reg = ®s[BPF_REG_3];
5903 struct bpf_reg_state *data_len_reg = ®s[BPF_REG_5];
5904 struct bpf_map *fmt_map = fmt_reg->map_ptr;
5905 int err, fmt_map_off, num_args;
5909 /* data must be an array of u64 */
5910 if (data_len_reg->var_off.value % 8)
5912 num_args = data_len_reg->var_off.value / 8;
5914 /* fmt being ARG_PTR_TO_CONST_STR guarantees that var_off is const
5915 * and map_direct_value_addr is set.
5917 fmt_map_off = fmt_reg->off + fmt_reg->var_off.value;
5918 err = fmt_map->ops->map_direct_value_addr(fmt_map, &fmt_addr,
5921 verbose(env, "verifier bug\n");
5924 fmt = (char *)(long)fmt_addr + fmt_map_off;
5926 /* We are also guaranteed that fmt+fmt_map_off is NULL terminated, we
5927 * can focus on validating the format specifiers.
5929 err = bpf_bprintf_prepare(fmt, UINT_MAX, NULL, NULL, num_args);
5931 verbose(env, "Invalid format string\n");
5936 static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
5939 const struct bpf_func_proto *fn = NULL;
5940 struct bpf_reg_state *regs;
5941 struct bpf_call_arg_meta meta;
5942 int insn_idx = *insn_idx_p;
5944 int i, err, func_id;
5946 /* find function prototype */
5947 func_id = insn->imm;
5948 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
5949 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
5954 if (env->ops->get_func_proto)
5955 fn = env->ops->get_func_proto(func_id, env->prog);
5957 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
5962 /* eBPF programs must be GPL compatible to use GPL-ed functions */
5963 if (!env->prog->gpl_compatible && fn->gpl_only) {
5964 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
5968 if (fn->allowed && !fn->allowed(env->prog)) {
5969 verbose(env, "helper call is not allowed in probe\n");
5973 /* With LD_ABS/IND some JITs save/restore skb from r1. */
5974 changes_data = bpf_helper_changes_pkt_data(fn->func);
5975 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
5976 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
5977 func_id_name(func_id), func_id);
5981 memset(&meta, 0, sizeof(meta));
5982 meta.pkt_access = fn->pkt_access;
5984 err = check_func_proto(fn, func_id);
5986 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
5987 func_id_name(func_id), func_id);
5991 meta.func_id = func_id;
5993 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++) {
5994 err = check_func_arg(env, i, &meta, fn);
5999 err = record_func_map(env, &meta, func_id, insn_idx);
6003 err = record_func_key(env, &meta, func_id, insn_idx);
6007 /* Mark slots with STACK_MISC in case of raw mode, stack offset
6008 * is inferred from register state.
6010 for (i = 0; i < meta.access_size; i++) {
6011 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
6012 BPF_WRITE, -1, false);
6017 if (func_id == BPF_FUNC_tail_call) {
6018 err = check_reference_leak(env);
6020 verbose(env, "tail_call would lead to reference leak\n");
6023 } else if (is_release_function(func_id)) {
6024 err = release_reference(env, meta.ref_obj_id);
6026 verbose(env, "func %s#%d reference has not been acquired before\n",
6027 func_id_name(func_id), func_id);
6032 regs = cur_regs(env);
6034 /* check that flags argument in get_local_storage(map, flags) is 0,
6035 * this is required because get_local_storage() can't return an error.
6037 if (func_id == BPF_FUNC_get_local_storage &&
6038 !register_is_null(®s[BPF_REG_2])) {
6039 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
6043 if (func_id == BPF_FUNC_for_each_map_elem) {
6044 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
6045 set_map_elem_callback_state);
6050 if (func_id == BPF_FUNC_snprintf) {
6051 err = check_bpf_snprintf_call(env, regs);
6056 /* reset caller saved regs */
6057 for (i = 0; i < CALLER_SAVED_REGS; i++) {
6058 mark_reg_not_init(env, regs, caller_saved[i]);
6059 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
6062 /* helper call returns 64-bit value. */
6063 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
6065 /* update return register (already marked as written above) */
6066 if (fn->ret_type == RET_INTEGER) {
6067 /* sets type to SCALAR_VALUE */
6068 mark_reg_unknown(env, regs, BPF_REG_0);
6069 } else if (fn->ret_type == RET_VOID) {
6070 regs[BPF_REG_0].type = NOT_INIT;
6071 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL ||
6072 fn->ret_type == RET_PTR_TO_MAP_VALUE) {
6073 /* There is no offset yet applied, variable or fixed */
6074 mark_reg_known_zero(env, regs, BPF_REG_0);
6075 /* remember map_ptr, so that check_map_access()
6076 * can check 'value_size' boundary of memory access
6077 * to map element returned from bpf_map_lookup_elem()
6079 if (meta.map_ptr == NULL) {
6081 "kernel subsystem misconfigured verifier\n");
6084 regs[BPF_REG_0].map_ptr = meta.map_ptr;
6085 if (fn->ret_type == RET_PTR_TO_MAP_VALUE) {
6086 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE;
6087 if (map_value_has_spin_lock(meta.map_ptr))
6088 regs[BPF_REG_0].id = ++env->id_gen;
6090 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
6092 } else if (fn->ret_type == RET_PTR_TO_SOCKET_OR_NULL) {
6093 mark_reg_known_zero(env, regs, BPF_REG_0);
6094 regs[BPF_REG_0].type = PTR_TO_SOCKET_OR_NULL;
6095 } else if (fn->ret_type == RET_PTR_TO_SOCK_COMMON_OR_NULL) {
6096 mark_reg_known_zero(env, regs, BPF_REG_0);
6097 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON_OR_NULL;
6098 } else if (fn->ret_type == RET_PTR_TO_TCP_SOCK_OR_NULL) {
6099 mark_reg_known_zero(env, regs, BPF_REG_0);
6100 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK_OR_NULL;
6101 } else if (fn->ret_type == RET_PTR_TO_ALLOC_MEM_OR_NULL) {
6102 mark_reg_known_zero(env, regs, BPF_REG_0);
6103 regs[BPF_REG_0].type = PTR_TO_MEM_OR_NULL;
6104 regs[BPF_REG_0].mem_size = meta.mem_size;
6105 } else if (fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID_OR_NULL ||
6106 fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID) {
6107 const struct btf_type *t;
6109 mark_reg_known_zero(env, regs, BPF_REG_0);
6110 t = btf_type_skip_modifiers(meta.ret_btf, meta.ret_btf_id, NULL);
6111 if (!btf_type_is_struct(t)) {
6113 const struct btf_type *ret;
6116 /* resolve the type size of ksym. */
6117 ret = btf_resolve_size(meta.ret_btf, t, &tsize);
6119 tname = btf_name_by_offset(meta.ret_btf, t->name_off);
6120 verbose(env, "unable to resolve the size of type '%s': %ld\n",
6121 tname, PTR_ERR(ret));
6124 regs[BPF_REG_0].type =
6125 fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID ?
6126 PTR_TO_MEM : PTR_TO_MEM_OR_NULL;
6127 regs[BPF_REG_0].mem_size = tsize;
6129 regs[BPF_REG_0].type =
6130 fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID ?
6131 PTR_TO_BTF_ID : PTR_TO_BTF_ID_OR_NULL;
6132 regs[BPF_REG_0].btf = meta.ret_btf;
6133 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
6135 } else if (fn->ret_type == RET_PTR_TO_BTF_ID_OR_NULL ||
6136 fn->ret_type == RET_PTR_TO_BTF_ID) {
6139 mark_reg_known_zero(env, regs, BPF_REG_0);
6140 regs[BPF_REG_0].type = fn->ret_type == RET_PTR_TO_BTF_ID ?
6142 PTR_TO_BTF_ID_OR_NULL;
6143 ret_btf_id = *fn->ret_btf_id;
6144 if (ret_btf_id == 0) {
6145 verbose(env, "invalid return type %d of func %s#%d\n",
6146 fn->ret_type, func_id_name(func_id), func_id);
6149 /* current BPF helper definitions are only coming from
6150 * built-in code with type IDs from vmlinux BTF
6152 regs[BPF_REG_0].btf = btf_vmlinux;
6153 regs[BPF_REG_0].btf_id = ret_btf_id;
6155 verbose(env, "unknown return type %d of func %s#%d\n",
6156 fn->ret_type, func_id_name(func_id), func_id);
6160 if (reg_type_may_be_null(regs[BPF_REG_0].type))
6161 regs[BPF_REG_0].id = ++env->id_gen;
6163 if (is_ptr_cast_function(func_id)) {
6164 /* For release_reference() */
6165 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
6166 } else if (is_acquire_function(func_id, meta.map_ptr)) {
6167 int id = acquire_reference_state(env, insn_idx);
6171 /* For mark_ptr_or_null_reg() */
6172 regs[BPF_REG_0].id = id;
6173 /* For release_reference() */
6174 regs[BPF_REG_0].ref_obj_id = id;
6177 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
6179 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
6183 if ((func_id == BPF_FUNC_get_stack ||
6184 func_id == BPF_FUNC_get_task_stack) &&
6185 !env->prog->has_callchain_buf) {
6186 const char *err_str;
6188 #ifdef CONFIG_PERF_EVENTS
6189 err = get_callchain_buffers(sysctl_perf_event_max_stack);
6190 err_str = "cannot get callchain buffer for func %s#%d\n";
6193 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
6196 verbose(env, err_str, func_id_name(func_id), func_id);
6200 env->prog->has_callchain_buf = true;
6203 if (func_id == BPF_FUNC_get_stackid || func_id == BPF_FUNC_get_stack)
6204 env->prog->call_get_stack = true;
6207 clear_all_pkt_pointers(env);
6211 /* mark_btf_func_reg_size() is used when the reg size is determined by
6212 * the BTF func_proto's return value size and argument.
6214 static void mark_btf_func_reg_size(struct bpf_verifier_env *env, u32 regno,
6217 struct bpf_reg_state *reg = &cur_regs(env)[regno];
6219 if (regno == BPF_REG_0) {
6220 /* Function return value */
6221 reg->live |= REG_LIVE_WRITTEN;
6222 reg->subreg_def = reg_size == sizeof(u64) ?
6223 DEF_NOT_SUBREG : env->insn_idx + 1;
6225 /* Function argument */
6226 if (reg_size == sizeof(u64)) {
6227 mark_insn_zext(env, reg);
6228 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
6230 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ32);
6235 static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn)
6237 const struct btf_type *t, *func, *func_proto, *ptr_type;
6238 struct bpf_reg_state *regs = cur_regs(env);
6239 const char *func_name, *ptr_type_name;
6240 u32 i, nargs, func_id, ptr_type_id;
6241 const struct btf_param *args;
6244 func_id = insn->imm;
6245 func = btf_type_by_id(btf_vmlinux, func_id);
6246 func_name = btf_name_by_offset(btf_vmlinux, func->name_off);
6247 func_proto = btf_type_by_id(btf_vmlinux, func->type);
6249 if (!env->ops->check_kfunc_call ||
6250 !env->ops->check_kfunc_call(func_id)) {
6251 verbose(env, "calling kernel function %s is not allowed\n",
6256 /* Check the arguments */
6257 err = btf_check_kfunc_arg_match(env, btf_vmlinux, func_id, regs);
6261 for (i = 0; i < CALLER_SAVED_REGS; i++)
6262 mark_reg_not_init(env, regs, caller_saved[i]);
6264 /* Check return type */
6265 t = btf_type_skip_modifiers(btf_vmlinux, func_proto->type, NULL);
6266 if (btf_type_is_scalar(t)) {
6267 mark_reg_unknown(env, regs, BPF_REG_0);
6268 mark_btf_func_reg_size(env, BPF_REG_0, t->size);
6269 } else if (btf_type_is_ptr(t)) {
6270 ptr_type = btf_type_skip_modifiers(btf_vmlinux, t->type,
6272 if (!btf_type_is_struct(ptr_type)) {
6273 ptr_type_name = btf_name_by_offset(btf_vmlinux,
6274 ptr_type->name_off);
6275 verbose(env, "kernel function %s returns pointer type %s %s is not supported\n",
6276 func_name, btf_type_str(ptr_type),
6280 mark_reg_known_zero(env, regs, BPF_REG_0);
6281 regs[BPF_REG_0].btf = btf_vmlinux;
6282 regs[BPF_REG_0].type = PTR_TO_BTF_ID;
6283 regs[BPF_REG_0].btf_id = ptr_type_id;
6284 mark_btf_func_reg_size(env, BPF_REG_0, sizeof(void *));
6285 } /* else { add_kfunc_call() ensures it is btf_type_is_void(t) } */
6287 nargs = btf_type_vlen(func_proto);
6288 args = (const struct btf_param *)(func_proto + 1);
6289 for (i = 0; i < nargs; i++) {
6292 t = btf_type_skip_modifiers(btf_vmlinux, args[i].type, NULL);
6293 if (btf_type_is_ptr(t))
6294 mark_btf_func_reg_size(env, regno, sizeof(void *));
6296 /* scalar. ensured by btf_check_kfunc_arg_match() */
6297 mark_btf_func_reg_size(env, regno, t->size);
6303 static bool signed_add_overflows(s64 a, s64 b)
6305 /* Do the add in u64, where overflow is well-defined */
6306 s64 res = (s64)((u64)a + (u64)b);
6313 static bool signed_add32_overflows(s32 a, s32 b)
6315 /* Do the add in u32, where overflow is well-defined */
6316 s32 res = (s32)((u32)a + (u32)b);
6323 static bool signed_sub_overflows(s64 a, s64 b)
6325 /* Do the sub in u64, where overflow is well-defined */
6326 s64 res = (s64)((u64)a - (u64)b);
6333 static bool signed_sub32_overflows(s32 a, s32 b)
6335 /* Do the sub in u32, where overflow is well-defined */
6336 s32 res = (s32)((u32)a - (u32)b);
6343 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
6344 const struct bpf_reg_state *reg,
6345 enum bpf_reg_type type)
6347 bool known = tnum_is_const(reg->var_off);
6348 s64 val = reg->var_off.value;
6349 s64 smin = reg->smin_value;
6351 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
6352 verbose(env, "math between %s pointer and %lld is not allowed\n",
6353 reg_type_str[type], val);
6357 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
6358 verbose(env, "%s pointer offset %d is not allowed\n",
6359 reg_type_str[type], reg->off);
6363 if (smin == S64_MIN) {
6364 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
6365 reg_type_str[type]);
6369 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
6370 verbose(env, "value %lld makes %s pointer be out of bounds\n",
6371 smin, reg_type_str[type]);
6378 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
6380 return &env->insn_aux_data[env->insn_idx];
6391 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
6392 u32 *alu_limit, bool mask_to_left)
6394 u32 max = 0, ptr_limit = 0;
6396 switch (ptr_reg->type) {
6398 /* Offset 0 is out-of-bounds, but acceptable start for the
6399 * left direction, see BPF_REG_FP. Also, unknown scalar
6400 * offset where we would need to deal with min/max bounds is
6401 * currently prohibited for unprivileged.
6403 max = MAX_BPF_STACK + mask_to_left;
6404 ptr_limit = -(ptr_reg->var_off.value + ptr_reg->off);
6406 case PTR_TO_MAP_VALUE:
6407 max = ptr_reg->map_ptr->value_size;
6408 ptr_limit = (mask_to_left ?
6409 ptr_reg->smin_value :
6410 ptr_reg->umax_value) + ptr_reg->off;
6416 if (ptr_limit >= max)
6417 return REASON_LIMIT;
6418 *alu_limit = ptr_limit;
6422 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
6423 const struct bpf_insn *insn)
6425 return env->bypass_spec_v1 || BPF_SRC(insn->code) == BPF_K;
6428 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
6429 u32 alu_state, u32 alu_limit)
6431 /* If we arrived here from different branches with different
6432 * state or limits to sanitize, then this won't work.
6434 if (aux->alu_state &&
6435 (aux->alu_state != alu_state ||
6436 aux->alu_limit != alu_limit))
6437 return REASON_PATHS;
6439 /* Corresponding fixup done in do_misc_fixups(). */
6440 aux->alu_state = alu_state;
6441 aux->alu_limit = alu_limit;
6445 static int sanitize_val_alu(struct bpf_verifier_env *env,
6446 struct bpf_insn *insn)
6448 struct bpf_insn_aux_data *aux = cur_aux(env);
6450 if (can_skip_alu_sanitation(env, insn))
6453 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
6456 static bool sanitize_needed(u8 opcode)
6458 return opcode == BPF_ADD || opcode == BPF_SUB;
6461 struct bpf_sanitize_info {
6462 struct bpf_insn_aux_data aux;
6466 static struct bpf_verifier_state *
6467 sanitize_speculative_path(struct bpf_verifier_env *env,
6468 const struct bpf_insn *insn,
6469 u32 next_idx, u32 curr_idx)
6471 struct bpf_verifier_state *branch;
6472 struct bpf_reg_state *regs;
6474 branch = push_stack(env, next_idx, curr_idx, true);
6475 if (branch && insn) {
6476 regs = branch->frame[branch->curframe]->regs;
6477 if (BPF_SRC(insn->code) == BPF_K) {
6478 mark_reg_unknown(env, regs, insn->dst_reg);
6479 } else if (BPF_SRC(insn->code) == BPF_X) {
6480 mark_reg_unknown(env, regs, insn->dst_reg);
6481 mark_reg_unknown(env, regs, insn->src_reg);
6487 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
6488 struct bpf_insn *insn,
6489 const struct bpf_reg_state *ptr_reg,
6490 const struct bpf_reg_state *off_reg,
6491 struct bpf_reg_state *dst_reg,
6492 struct bpf_sanitize_info *info,
6493 const bool commit_window)
6495 struct bpf_insn_aux_data *aux = commit_window ? cur_aux(env) : &info->aux;
6496 struct bpf_verifier_state *vstate = env->cur_state;
6497 bool off_is_imm = tnum_is_const(off_reg->var_off);
6498 bool off_is_neg = off_reg->smin_value < 0;
6499 bool ptr_is_dst_reg = ptr_reg == dst_reg;
6500 u8 opcode = BPF_OP(insn->code);
6501 u32 alu_state, alu_limit;
6502 struct bpf_reg_state tmp;
6506 if (can_skip_alu_sanitation(env, insn))
6509 /* We already marked aux for masking from non-speculative
6510 * paths, thus we got here in the first place. We only care
6511 * to explore bad access from here.
6513 if (vstate->speculative)
6516 if (!commit_window) {
6517 if (!tnum_is_const(off_reg->var_off) &&
6518 (off_reg->smin_value < 0) != (off_reg->smax_value < 0))
6519 return REASON_BOUNDS;
6521 info->mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
6522 (opcode == BPF_SUB && !off_is_neg);
6525 err = retrieve_ptr_limit(ptr_reg, &alu_limit, info->mask_to_left);
6529 if (commit_window) {
6530 /* In commit phase we narrow the masking window based on
6531 * the observed pointer move after the simulated operation.
6533 alu_state = info->aux.alu_state;
6534 alu_limit = abs(info->aux.alu_limit - alu_limit);
6536 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
6537 alu_state |= off_is_imm ? BPF_ALU_IMMEDIATE : 0;
6538 alu_state |= ptr_is_dst_reg ?
6539 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
6541 /* Limit pruning on unknown scalars to enable deep search for
6542 * potential masking differences from other program paths.
6545 env->explore_alu_limits = true;
6548 err = update_alu_sanitation_state(aux, alu_state, alu_limit);
6552 /* If we're in commit phase, we're done here given we already
6553 * pushed the truncated dst_reg into the speculative verification
6556 * Also, when register is a known constant, we rewrite register-based
6557 * operation to immediate-based, and thus do not need masking (and as
6558 * a consequence, do not need to simulate the zero-truncation either).
6560 if (commit_window || off_is_imm)
6563 /* Simulate and find potential out-of-bounds access under
6564 * speculative execution from truncation as a result of
6565 * masking when off was not within expected range. If off
6566 * sits in dst, then we temporarily need to move ptr there
6567 * to simulate dst (== 0) +/-= ptr. Needed, for example,
6568 * for cases where we use K-based arithmetic in one direction
6569 * and truncated reg-based in the other in order to explore
6572 if (!ptr_is_dst_reg) {
6574 *dst_reg = *ptr_reg;
6576 ret = sanitize_speculative_path(env, NULL, env->insn_idx + 1,
6578 if (!ptr_is_dst_reg && ret)
6580 return !ret ? REASON_STACK : 0;
6583 static void sanitize_mark_insn_seen(struct bpf_verifier_env *env)
6585 struct bpf_verifier_state *vstate = env->cur_state;
6587 /* If we simulate paths under speculation, we don't update the
6588 * insn as 'seen' such that when we verify unreachable paths in
6589 * the non-speculative domain, sanitize_dead_code() can still
6590 * rewrite/sanitize them.
6592 if (!vstate->speculative)
6593 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
6596 static int sanitize_err(struct bpf_verifier_env *env,
6597 const struct bpf_insn *insn, int reason,
6598 const struct bpf_reg_state *off_reg,
6599 const struct bpf_reg_state *dst_reg)
6601 static const char *err = "pointer arithmetic with it prohibited for !root";
6602 const char *op = BPF_OP(insn->code) == BPF_ADD ? "add" : "sub";
6603 u32 dst = insn->dst_reg, src = insn->src_reg;
6607 verbose(env, "R%d has unknown scalar with mixed signed bounds, %s\n",
6608 off_reg == dst_reg ? dst : src, err);
6611 verbose(env, "R%d has pointer with unsupported alu operation, %s\n",
6612 off_reg == dst_reg ? src : dst, err);
6615 verbose(env, "R%d tried to %s from different maps, paths or scalars, %s\n",
6619 verbose(env, "R%d tried to %s beyond pointer bounds, %s\n",
6623 verbose(env, "R%d could not be pushed for speculative verification, %s\n",
6627 verbose(env, "verifier internal error: unknown reason (%d)\n",
6635 /* check that stack access falls within stack limits and that 'reg' doesn't
6636 * have a variable offset.
6638 * Variable offset is prohibited for unprivileged mode for simplicity since it
6639 * requires corresponding support in Spectre masking for stack ALU. See also
6640 * retrieve_ptr_limit().
6643 * 'off' includes 'reg->off'.
6645 static int check_stack_access_for_ptr_arithmetic(
6646 struct bpf_verifier_env *env,
6648 const struct bpf_reg_state *reg,
6651 if (!tnum_is_const(reg->var_off)) {
6654 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6655 verbose(env, "R%d variable stack access prohibited for !root, var_off=%s off=%d\n",
6656 regno, tn_buf, off);
6660 if (off >= 0 || off < -MAX_BPF_STACK) {
6661 verbose(env, "R%d stack pointer arithmetic goes out of range, "
6662 "prohibited for !root; off=%d\n", regno, off);
6669 static int sanitize_check_bounds(struct bpf_verifier_env *env,
6670 const struct bpf_insn *insn,
6671 const struct bpf_reg_state *dst_reg)
6673 u32 dst = insn->dst_reg;
6675 /* For unprivileged we require that resulting offset must be in bounds
6676 * in order to be able to sanitize access later on.
6678 if (env->bypass_spec_v1)
6681 switch (dst_reg->type) {
6683 if (check_stack_access_for_ptr_arithmetic(env, dst, dst_reg,
6684 dst_reg->off + dst_reg->var_off.value))
6687 case PTR_TO_MAP_VALUE:
6688 if (check_map_access(env, dst, dst_reg->off, 1, false)) {
6689 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
6690 "prohibited for !root\n", dst);
6701 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
6702 * Caller should also handle BPF_MOV case separately.
6703 * If we return -EACCES, caller may want to try again treating pointer as a
6704 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
6706 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
6707 struct bpf_insn *insn,
6708 const struct bpf_reg_state *ptr_reg,
6709 const struct bpf_reg_state *off_reg)
6711 struct bpf_verifier_state *vstate = env->cur_state;
6712 struct bpf_func_state *state = vstate->frame[vstate->curframe];
6713 struct bpf_reg_state *regs = state->regs, *dst_reg;
6714 bool known = tnum_is_const(off_reg->var_off);
6715 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
6716 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
6717 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
6718 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
6719 struct bpf_sanitize_info info = {};
6720 u8 opcode = BPF_OP(insn->code);
6721 u32 dst = insn->dst_reg;
6724 dst_reg = ®s[dst];
6726 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
6727 smin_val > smax_val || umin_val > umax_val) {
6728 /* Taint dst register if offset had invalid bounds derived from
6729 * e.g. dead branches.
6731 __mark_reg_unknown(env, dst_reg);
6735 if (BPF_CLASS(insn->code) != BPF_ALU64) {
6736 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
6737 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
6738 __mark_reg_unknown(env, dst_reg);
6743 "R%d 32-bit pointer arithmetic prohibited\n",
6748 switch (ptr_reg->type) {
6749 case PTR_TO_MAP_VALUE_OR_NULL:
6750 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
6751 dst, reg_type_str[ptr_reg->type]);
6753 case CONST_PTR_TO_MAP:
6754 /* smin_val represents the known value */
6755 if (known && smin_val == 0 && opcode == BPF_ADD)
6758 case PTR_TO_PACKET_END:
6760 case PTR_TO_SOCKET_OR_NULL:
6761 case PTR_TO_SOCK_COMMON:
6762 case PTR_TO_SOCK_COMMON_OR_NULL:
6763 case PTR_TO_TCP_SOCK:
6764 case PTR_TO_TCP_SOCK_OR_NULL:
6765 case PTR_TO_XDP_SOCK:
6766 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
6767 dst, reg_type_str[ptr_reg->type]);
6773 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
6774 * The id may be overwritten later if we create a new variable offset.
6776 dst_reg->type = ptr_reg->type;
6777 dst_reg->id = ptr_reg->id;
6779 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
6780 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
6783 /* pointer types do not carry 32-bit bounds at the moment. */
6784 __mark_reg32_unbounded(dst_reg);
6786 if (sanitize_needed(opcode)) {
6787 ret = sanitize_ptr_alu(env, insn, ptr_reg, off_reg, dst_reg,
6790 return sanitize_err(env, insn, ret, off_reg, dst_reg);
6795 /* We can take a fixed offset as long as it doesn't overflow
6796 * the s32 'off' field
6798 if (known && (ptr_reg->off + smin_val ==
6799 (s64)(s32)(ptr_reg->off + smin_val))) {
6800 /* pointer += K. Accumulate it into fixed offset */
6801 dst_reg->smin_value = smin_ptr;
6802 dst_reg->smax_value = smax_ptr;
6803 dst_reg->umin_value = umin_ptr;
6804 dst_reg->umax_value = umax_ptr;
6805 dst_reg->var_off = ptr_reg->var_off;
6806 dst_reg->off = ptr_reg->off + smin_val;
6807 dst_reg->raw = ptr_reg->raw;
6810 /* A new variable offset is created. Note that off_reg->off
6811 * == 0, since it's a scalar.
6812 * dst_reg gets the pointer type and since some positive
6813 * integer value was added to the pointer, give it a new 'id'
6814 * if it's a PTR_TO_PACKET.
6815 * this creates a new 'base' pointer, off_reg (variable) gets
6816 * added into the variable offset, and we copy the fixed offset
6819 if (signed_add_overflows(smin_ptr, smin_val) ||
6820 signed_add_overflows(smax_ptr, smax_val)) {
6821 dst_reg->smin_value = S64_MIN;
6822 dst_reg->smax_value = S64_MAX;
6824 dst_reg->smin_value = smin_ptr + smin_val;
6825 dst_reg->smax_value = smax_ptr + smax_val;
6827 if (umin_ptr + umin_val < umin_ptr ||
6828 umax_ptr + umax_val < umax_ptr) {
6829 dst_reg->umin_value = 0;
6830 dst_reg->umax_value = U64_MAX;
6832 dst_reg->umin_value = umin_ptr + umin_val;
6833 dst_reg->umax_value = umax_ptr + umax_val;
6835 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
6836 dst_reg->off = ptr_reg->off;
6837 dst_reg->raw = ptr_reg->raw;
6838 if (reg_is_pkt_pointer(ptr_reg)) {
6839 dst_reg->id = ++env->id_gen;
6840 /* something was added to pkt_ptr, set range to zero */
6841 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
6845 if (dst_reg == off_reg) {
6846 /* scalar -= pointer. Creates an unknown scalar */
6847 verbose(env, "R%d tried to subtract pointer from scalar\n",
6851 /* We don't allow subtraction from FP, because (according to
6852 * test_verifier.c test "invalid fp arithmetic", JITs might not
6853 * be able to deal with it.
6855 if (ptr_reg->type == PTR_TO_STACK) {
6856 verbose(env, "R%d subtraction from stack pointer prohibited\n",
6860 if (known && (ptr_reg->off - smin_val ==
6861 (s64)(s32)(ptr_reg->off - smin_val))) {
6862 /* pointer -= K. Subtract it from fixed offset */
6863 dst_reg->smin_value = smin_ptr;
6864 dst_reg->smax_value = smax_ptr;
6865 dst_reg->umin_value = umin_ptr;
6866 dst_reg->umax_value = umax_ptr;
6867 dst_reg->var_off = ptr_reg->var_off;
6868 dst_reg->id = ptr_reg->id;
6869 dst_reg->off = ptr_reg->off - smin_val;
6870 dst_reg->raw = ptr_reg->raw;
6873 /* A new variable offset is created. If the subtrahend is known
6874 * nonnegative, then any reg->range we had before is still good.
6876 if (signed_sub_overflows(smin_ptr, smax_val) ||
6877 signed_sub_overflows(smax_ptr, smin_val)) {
6878 /* Overflow possible, we know nothing */
6879 dst_reg->smin_value = S64_MIN;
6880 dst_reg->smax_value = S64_MAX;
6882 dst_reg->smin_value = smin_ptr - smax_val;
6883 dst_reg->smax_value = smax_ptr - smin_val;
6885 if (umin_ptr < umax_val) {
6886 /* Overflow possible, we know nothing */
6887 dst_reg->umin_value = 0;
6888 dst_reg->umax_value = U64_MAX;
6890 /* Cannot overflow (as long as bounds are consistent) */
6891 dst_reg->umin_value = umin_ptr - umax_val;
6892 dst_reg->umax_value = umax_ptr - umin_val;
6894 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
6895 dst_reg->off = ptr_reg->off;
6896 dst_reg->raw = ptr_reg->raw;
6897 if (reg_is_pkt_pointer(ptr_reg)) {
6898 dst_reg->id = ++env->id_gen;
6899 /* something was added to pkt_ptr, set range to zero */
6901 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
6907 /* bitwise ops on pointers are troublesome, prohibit. */
6908 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
6909 dst, bpf_alu_string[opcode >> 4]);
6912 /* other operators (e.g. MUL,LSH) produce non-pointer results */
6913 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
6914 dst, bpf_alu_string[opcode >> 4]);
6918 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
6921 __update_reg_bounds(dst_reg);
6922 __reg_deduce_bounds(dst_reg);
6923 __reg_bound_offset(dst_reg);
6925 if (sanitize_check_bounds(env, insn, dst_reg) < 0)
6927 if (sanitize_needed(opcode)) {
6928 ret = sanitize_ptr_alu(env, insn, dst_reg, off_reg, dst_reg,
6931 return sanitize_err(env, insn, ret, off_reg, dst_reg);
6937 static void scalar32_min_max_add(struct bpf_reg_state *dst_reg,
6938 struct bpf_reg_state *src_reg)
6940 s32 smin_val = src_reg->s32_min_value;
6941 s32 smax_val = src_reg->s32_max_value;
6942 u32 umin_val = src_reg->u32_min_value;
6943 u32 umax_val = src_reg->u32_max_value;
6945 if (signed_add32_overflows(dst_reg->s32_min_value, smin_val) ||
6946 signed_add32_overflows(dst_reg->s32_max_value, smax_val)) {
6947 dst_reg->s32_min_value = S32_MIN;
6948 dst_reg->s32_max_value = S32_MAX;
6950 dst_reg->s32_min_value += smin_val;
6951 dst_reg->s32_max_value += smax_val;
6953 if (dst_reg->u32_min_value + umin_val < umin_val ||
6954 dst_reg->u32_max_value + umax_val < umax_val) {
6955 dst_reg->u32_min_value = 0;
6956 dst_reg->u32_max_value = U32_MAX;
6958 dst_reg->u32_min_value += umin_val;
6959 dst_reg->u32_max_value += umax_val;
6963 static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
6964 struct bpf_reg_state *src_reg)
6966 s64 smin_val = src_reg->smin_value;
6967 s64 smax_val = src_reg->smax_value;
6968 u64 umin_val = src_reg->umin_value;
6969 u64 umax_val = src_reg->umax_value;
6971 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
6972 signed_add_overflows(dst_reg->smax_value, smax_val)) {
6973 dst_reg->smin_value = S64_MIN;
6974 dst_reg->smax_value = S64_MAX;
6976 dst_reg->smin_value += smin_val;
6977 dst_reg->smax_value += smax_val;
6979 if (dst_reg->umin_value + umin_val < umin_val ||
6980 dst_reg->umax_value + umax_val < umax_val) {
6981 dst_reg->umin_value = 0;
6982 dst_reg->umax_value = U64_MAX;
6984 dst_reg->umin_value += umin_val;
6985 dst_reg->umax_value += umax_val;
6989 static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg,
6990 struct bpf_reg_state *src_reg)
6992 s32 smin_val = src_reg->s32_min_value;
6993 s32 smax_val = src_reg->s32_max_value;
6994 u32 umin_val = src_reg->u32_min_value;
6995 u32 umax_val = src_reg->u32_max_value;
6997 if (signed_sub32_overflows(dst_reg->s32_min_value, smax_val) ||
6998 signed_sub32_overflows(dst_reg->s32_max_value, smin_val)) {
6999 /* Overflow possible, we know nothing */
7000 dst_reg->s32_min_value = S32_MIN;
7001 dst_reg->s32_max_value = S32_MAX;
7003 dst_reg->s32_min_value -= smax_val;
7004 dst_reg->s32_max_value -= smin_val;
7006 if (dst_reg->u32_min_value < umax_val) {
7007 /* Overflow possible, we know nothing */
7008 dst_reg->u32_min_value = 0;
7009 dst_reg->u32_max_value = U32_MAX;
7011 /* Cannot overflow (as long as bounds are consistent) */
7012 dst_reg->u32_min_value -= umax_val;
7013 dst_reg->u32_max_value -= umin_val;
7017 static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
7018 struct bpf_reg_state *src_reg)
7020 s64 smin_val = src_reg->smin_value;
7021 s64 smax_val = src_reg->smax_value;
7022 u64 umin_val = src_reg->umin_value;
7023 u64 umax_val = src_reg->umax_value;
7025 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
7026 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
7027 /* Overflow possible, we know nothing */
7028 dst_reg->smin_value = S64_MIN;
7029 dst_reg->smax_value = S64_MAX;
7031 dst_reg->smin_value -= smax_val;
7032 dst_reg->smax_value -= smin_val;
7034 if (dst_reg->umin_value < umax_val) {
7035 /* Overflow possible, we know nothing */
7036 dst_reg->umin_value = 0;
7037 dst_reg->umax_value = U64_MAX;
7039 /* Cannot overflow (as long as bounds are consistent) */
7040 dst_reg->umin_value -= umax_val;
7041 dst_reg->umax_value -= umin_val;
7045 static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg,
7046 struct bpf_reg_state *src_reg)
7048 s32 smin_val = src_reg->s32_min_value;
7049 u32 umin_val = src_reg->u32_min_value;
7050 u32 umax_val = src_reg->u32_max_value;
7052 if (smin_val < 0 || dst_reg->s32_min_value < 0) {
7053 /* Ain't nobody got time to multiply that sign */
7054 __mark_reg32_unbounded(dst_reg);
7057 /* Both values are positive, so we can work with unsigned and
7058 * copy the result to signed (unless it exceeds S32_MAX).
7060 if (umax_val > U16_MAX || dst_reg->u32_max_value > U16_MAX) {
7061 /* Potential overflow, we know nothing */
7062 __mark_reg32_unbounded(dst_reg);
7065 dst_reg->u32_min_value *= umin_val;
7066 dst_reg->u32_max_value *= umax_val;
7067 if (dst_reg->u32_max_value > S32_MAX) {
7068 /* Overflow possible, we know nothing */
7069 dst_reg->s32_min_value = S32_MIN;
7070 dst_reg->s32_max_value = S32_MAX;
7072 dst_reg->s32_min_value = dst_reg->u32_min_value;
7073 dst_reg->s32_max_value = dst_reg->u32_max_value;
7077 static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
7078 struct bpf_reg_state *src_reg)
7080 s64 smin_val = src_reg->smin_value;
7081 u64 umin_val = src_reg->umin_value;
7082 u64 umax_val = src_reg->umax_value;
7084 if (smin_val < 0 || dst_reg->smin_value < 0) {
7085 /* Ain't nobody got time to multiply that sign */
7086 __mark_reg64_unbounded(dst_reg);
7089 /* Both values are positive, so we can work with unsigned and
7090 * copy the result to signed (unless it exceeds S64_MAX).
7092 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
7093 /* Potential overflow, we know nothing */
7094 __mark_reg64_unbounded(dst_reg);
7097 dst_reg->umin_value *= umin_val;
7098 dst_reg->umax_value *= umax_val;
7099 if (dst_reg->umax_value > S64_MAX) {
7100 /* Overflow possible, we know nothing */
7101 dst_reg->smin_value = S64_MIN;
7102 dst_reg->smax_value = S64_MAX;
7104 dst_reg->smin_value = dst_reg->umin_value;
7105 dst_reg->smax_value = dst_reg->umax_value;
7109 static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
7110 struct bpf_reg_state *src_reg)
7112 bool src_known = tnum_subreg_is_const(src_reg->var_off);
7113 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
7114 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
7115 s32 smin_val = src_reg->s32_min_value;
7116 u32 umax_val = src_reg->u32_max_value;
7118 if (src_known && dst_known) {
7119 __mark_reg32_known(dst_reg, var32_off.value);
7123 /* We get our minimum from the var_off, since that's inherently
7124 * bitwise. Our maximum is the minimum of the operands' maxima.
7126 dst_reg->u32_min_value = var32_off.value;
7127 dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val);
7128 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
7129 /* Lose signed bounds when ANDing negative numbers,
7130 * ain't nobody got time for that.
7132 dst_reg->s32_min_value = S32_MIN;
7133 dst_reg->s32_max_value = S32_MAX;
7135 /* ANDing two positives gives a positive, so safe to
7136 * cast result into s64.
7138 dst_reg->s32_min_value = dst_reg->u32_min_value;
7139 dst_reg->s32_max_value = dst_reg->u32_max_value;
7143 static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
7144 struct bpf_reg_state *src_reg)
7146 bool src_known = tnum_is_const(src_reg->var_off);
7147 bool dst_known = tnum_is_const(dst_reg->var_off);
7148 s64 smin_val = src_reg->smin_value;
7149 u64 umax_val = src_reg->umax_value;
7151 if (src_known && dst_known) {
7152 __mark_reg_known(dst_reg, dst_reg->var_off.value);
7156 /* We get our minimum from the var_off, since that's inherently
7157 * bitwise. Our maximum is the minimum of the operands' maxima.
7159 dst_reg->umin_value = dst_reg->var_off.value;
7160 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
7161 if (dst_reg->smin_value < 0 || smin_val < 0) {
7162 /* Lose signed bounds when ANDing negative numbers,
7163 * ain't nobody got time for that.
7165 dst_reg->smin_value = S64_MIN;
7166 dst_reg->smax_value = S64_MAX;
7168 /* ANDing two positives gives a positive, so safe to
7169 * cast result into s64.
7171 dst_reg->smin_value = dst_reg->umin_value;
7172 dst_reg->smax_value = dst_reg->umax_value;
7174 /* We may learn something more from the var_off */
7175 __update_reg_bounds(dst_reg);
7178 static void scalar32_min_max_or(struct bpf_reg_state *dst_reg,
7179 struct bpf_reg_state *src_reg)
7181 bool src_known = tnum_subreg_is_const(src_reg->var_off);
7182 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
7183 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
7184 s32 smin_val = src_reg->s32_min_value;
7185 u32 umin_val = src_reg->u32_min_value;
7187 if (src_known && dst_known) {
7188 __mark_reg32_known(dst_reg, var32_off.value);
7192 /* We get our maximum from the var_off, and our minimum is the
7193 * maximum of the operands' minima
7195 dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val);
7196 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
7197 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
7198 /* Lose signed bounds when ORing negative numbers,
7199 * ain't nobody got time for that.
7201 dst_reg->s32_min_value = S32_MIN;
7202 dst_reg->s32_max_value = S32_MAX;
7204 /* ORing two positives gives a positive, so safe to
7205 * cast result into s64.
7207 dst_reg->s32_min_value = dst_reg->u32_min_value;
7208 dst_reg->s32_max_value = dst_reg->u32_max_value;
7212 static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
7213 struct bpf_reg_state *src_reg)
7215 bool src_known = tnum_is_const(src_reg->var_off);
7216 bool dst_known = tnum_is_const(dst_reg->var_off);
7217 s64 smin_val = src_reg->smin_value;
7218 u64 umin_val = src_reg->umin_value;
7220 if (src_known && dst_known) {
7221 __mark_reg_known(dst_reg, dst_reg->var_off.value);
7225 /* We get our maximum from the var_off, and our minimum is the
7226 * maximum of the operands' minima
7228 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
7229 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
7230 if (dst_reg->smin_value < 0 || smin_val < 0) {
7231 /* Lose signed bounds when ORing negative numbers,
7232 * ain't nobody got time for that.
7234 dst_reg->smin_value = S64_MIN;
7235 dst_reg->smax_value = S64_MAX;
7237 /* ORing two positives gives a positive, so safe to
7238 * cast result into s64.
7240 dst_reg->smin_value = dst_reg->umin_value;
7241 dst_reg->smax_value = dst_reg->umax_value;
7243 /* We may learn something more from the var_off */
7244 __update_reg_bounds(dst_reg);
7247 static void scalar32_min_max_xor(struct bpf_reg_state *dst_reg,
7248 struct bpf_reg_state *src_reg)
7250 bool src_known = tnum_subreg_is_const(src_reg->var_off);
7251 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
7252 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
7253 s32 smin_val = src_reg->s32_min_value;
7255 if (src_known && dst_known) {
7256 __mark_reg32_known(dst_reg, var32_off.value);
7260 /* We get both minimum and maximum from the var32_off. */
7261 dst_reg->u32_min_value = var32_off.value;
7262 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
7264 if (dst_reg->s32_min_value >= 0 && smin_val >= 0) {
7265 /* XORing two positive sign numbers gives a positive,
7266 * so safe to cast u32 result into s32.
7268 dst_reg->s32_min_value = dst_reg->u32_min_value;
7269 dst_reg->s32_max_value = dst_reg->u32_max_value;
7271 dst_reg->s32_min_value = S32_MIN;
7272 dst_reg->s32_max_value = S32_MAX;
7276 static void scalar_min_max_xor(struct bpf_reg_state *dst_reg,
7277 struct bpf_reg_state *src_reg)
7279 bool src_known = tnum_is_const(src_reg->var_off);
7280 bool dst_known = tnum_is_const(dst_reg->var_off);
7281 s64 smin_val = src_reg->smin_value;
7283 if (src_known && dst_known) {
7284 /* dst_reg->var_off.value has been updated earlier */
7285 __mark_reg_known(dst_reg, dst_reg->var_off.value);
7289 /* We get both minimum and maximum from the var_off. */
7290 dst_reg->umin_value = dst_reg->var_off.value;
7291 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
7293 if (dst_reg->smin_value >= 0 && smin_val >= 0) {
7294 /* XORing two positive sign numbers gives a positive,
7295 * so safe to cast u64 result into s64.
7297 dst_reg->smin_value = dst_reg->umin_value;
7298 dst_reg->smax_value = dst_reg->umax_value;
7300 dst_reg->smin_value = S64_MIN;
7301 dst_reg->smax_value = S64_MAX;
7304 __update_reg_bounds(dst_reg);
7307 static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
7308 u64 umin_val, u64 umax_val)
7310 /* We lose all sign bit information (except what we can pick
7313 dst_reg->s32_min_value = S32_MIN;
7314 dst_reg->s32_max_value = S32_MAX;
7315 /* If we might shift our top bit out, then we know nothing */
7316 if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) {
7317 dst_reg->u32_min_value = 0;
7318 dst_reg->u32_max_value = U32_MAX;
7320 dst_reg->u32_min_value <<= umin_val;
7321 dst_reg->u32_max_value <<= umax_val;
7325 static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
7326 struct bpf_reg_state *src_reg)
7328 u32 umax_val = src_reg->u32_max_value;
7329 u32 umin_val = src_reg->u32_min_value;
7330 /* u32 alu operation will zext upper bits */
7331 struct tnum subreg = tnum_subreg(dst_reg->var_off);
7333 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
7334 dst_reg->var_off = tnum_subreg(tnum_lshift(subreg, umin_val));
7335 /* Not required but being careful mark reg64 bounds as unknown so
7336 * that we are forced to pick them up from tnum and zext later and
7337 * if some path skips this step we are still safe.
7339 __mark_reg64_unbounded(dst_reg);
7340 __update_reg32_bounds(dst_reg);
7343 static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg,
7344 u64 umin_val, u64 umax_val)
7346 /* Special case <<32 because it is a common compiler pattern to sign
7347 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
7348 * positive we know this shift will also be positive so we can track
7349 * bounds correctly. Otherwise we lose all sign bit information except
7350 * what we can pick up from var_off. Perhaps we can generalize this
7351 * later to shifts of any length.
7353 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0)
7354 dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32;
7356 dst_reg->smax_value = S64_MAX;
7358 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0)
7359 dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32;
7361 dst_reg->smin_value = S64_MIN;
7363 /* If we might shift our top bit out, then we know nothing */
7364 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
7365 dst_reg->umin_value = 0;
7366 dst_reg->umax_value = U64_MAX;
7368 dst_reg->umin_value <<= umin_val;
7369 dst_reg->umax_value <<= umax_val;
7373 static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
7374 struct bpf_reg_state *src_reg)
7376 u64 umax_val = src_reg->umax_value;
7377 u64 umin_val = src_reg->umin_value;
7379 /* scalar64 calc uses 32bit unshifted bounds so must be called first */
7380 __scalar64_min_max_lsh(dst_reg, umin_val, umax_val);
7381 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
7383 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
7384 /* We may learn something more from the var_off */
7385 __update_reg_bounds(dst_reg);
7388 static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg,
7389 struct bpf_reg_state *src_reg)
7391 struct tnum subreg = tnum_subreg(dst_reg->var_off);
7392 u32 umax_val = src_reg->u32_max_value;
7393 u32 umin_val = src_reg->u32_min_value;
7395 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
7396 * be negative, then either:
7397 * 1) src_reg might be zero, so the sign bit of the result is
7398 * unknown, so we lose our signed bounds
7399 * 2) it's known negative, thus the unsigned bounds capture the
7401 * 3) the signed bounds cross zero, so they tell us nothing
7403 * If the value in dst_reg is known nonnegative, then again the
7404 * unsigned bounds capture the signed bounds.
7405 * Thus, in all cases it suffices to blow away our signed bounds
7406 * and rely on inferring new ones from the unsigned bounds and
7407 * var_off of the result.
7409 dst_reg->s32_min_value = S32_MIN;
7410 dst_reg->s32_max_value = S32_MAX;
7412 dst_reg->var_off = tnum_rshift(subreg, umin_val);
7413 dst_reg->u32_min_value >>= umax_val;
7414 dst_reg->u32_max_value >>= umin_val;
7416 __mark_reg64_unbounded(dst_reg);
7417 __update_reg32_bounds(dst_reg);
7420 static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
7421 struct bpf_reg_state *src_reg)
7423 u64 umax_val = src_reg->umax_value;
7424 u64 umin_val = src_reg->umin_value;
7426 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
7427 * be negative, then either:
7428 * 1) src_reg might be zero, so the sign bit of the result is
7429 * unknown, so we lose our signed bounds
7430 * 2) it's known negative, thus the unsigned bounds capture the
7432 * 3) the signed bounds cross zero, so they tell us nothing
7434 * If the value in dst_reg is known nonnegative, then again the
7435 * unsigned bounds capture the signed bounds.
7436 * Thus, in all cases it suffices to blow away our signed bounds
7437 * and rely on inferring new ones from the unsigned bounds and
7438 * var_off of the result.
7440 dst_reg->smin_value = S64_MIN;
7441 dst_reg->smax_value = S64_MAX;
7442 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
7443 dst_reg->umin_value >>= umax_val;
7444 dst_reg->umax_value >>= umin_val;
7446 /* Its not easy to operate on alu32 bounds here because it depends
7447 * on bits being shifted in. Take easy way out and mark unbounded
7448 * so we can recalculate later from tnum.
7450 __mark_reg32_unbounded(dst_reg);
7451 __update_reg_bounds(dst_reg);
7454 static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg,
7455 struct bpf_reg_state *src_reg)
7457 u64 umin_val = src_reg->u32_min_value;
7459 /* Upon reaching here, src_known is true and
7460 * umax_val is equal to umin_val.
7462 dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val);
7463 dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val);
7465 dst_reg->var_off = tnum_arshift(tnum_subreg(dst_reg->var_off), umin_val, 32);
7467 /* blow away the dst_reg umin_value/umax_value and rely on
7468 * dst_reg var_off to refine the result.
7470 dst_reg->u32_min_value = 0;
7471 dst_reg->u32_max_value = U32_MAX;
7473 __mark_reg64_unbounded(dst_reg);
7474 __update_reg32_bounds(dst_reg);
7477 static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
7478 struct bpf_reg_state *src_reg)
7480 u64 umin_val = src_reg->umin_value;
7482 /* Upon reaching here, src_known is true and umax_val is equal
7485 dst_reg->smin_value >>= umin_val;
7486 dst_reg->smax_value >>= umin_val;
7488 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val, 64);
7490 /* blow away the dst_reg umin_value/umax_value and rely on
7491 * dst_reg var_off to refine the result.
7493 dst_reg->umin_value = 0;
7494 dst_reg->umax_value = U64_MAX;
7496 /* Its not easy to operate on alu32 bounds here because it depends
7497 * on bits being shifted in from upper 32-bits. Take easy way out
7498 * and mark unbounded so we can recalculate later from tnum.
7500 __mark_reg32_unbounded(dst_reg);
7501 __update_reg_bounds(dst_reg);
7504 /* WARNING: This function does calculations on 64-bit values, but the actual
7505 * execution may occur on 32-bit values. Therefore, things like bitshifts
7506 * need extra checks in the 32-bit case.
7508 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
7509 struct bpf_insn *insn,
7510 struct bpf_reg_state *dst_reg,
7511 struct bpf_reg_state src_reg)
7513 struct bpf_reg_state *regs = cur_regs(env);
7514 u8 opcode = BPF_OP(insn->code);
7516 s64 smin_val, smax_val;
7517 u64 umin_val, umax_val;
7518 s32 s32_min_val, s32_max_val;
7519 u32 u32_min_val, u32_max_val;
7520 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
7521 bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
7524 smin_val = src_reg.smin_value;
7525 smax_val = src_reg.smax_value;
7526 umin_val = src_reg.umin_value;
7527 umax_val = src_reg.umax_value;
7529 s32_min_val = src_reg.s32_min_value;
7530 s32_max_val = src_reg.s32_max_value;
7531 u32_min_val = src_reg.u32_min_value;
7532 u32_max_val = src_reg.u32_max_value;
7535 src_known = tnum_subreg_is_const(src_reg.var_off);
7537 (s32_min_val != s32_max_val || u32_min_val != u32_max_val)) ||
7538 s32_min_val > s32_max_val || u32_min_val > u32_max_val) {
7539 /* Taint dst register if offset had invalid bounds
7540 * derived from e.g. dead branches.
7542 __mark_reg_unknown(env, dst_reg);
7546 src_known = tnum_is_const(src_reg.var_off);
7548 (smin_val != smax_val || umin_val != umax_val)) ||
7549 smin_val > smax_val || umin_val > umax_val) {
7550 /* Taint dst register if offset had invalid bounds
7551 * derived from e.g. dead branches.
7553 __mark_reg_unknown(env, dst_reg);
7559 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
7560 __mark_reg_unknown(env, dst_reg);
7564 if (sanitize_needed(opcode)) {
7565 ret = sanitize_val_alu(env, insn);
7567 return sanitize_err(env, insn, ret, NULL, NULL);
7570 /* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
7571 * There are two classes of instructions: The first class we track both
7572 * alu32 and alu64 sign/unsigned bounds independently this provides the
7573 * greatest amount of precision when alu operations are mixed with jmp32
7574 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
7575 * and BPF_OR. This is possible because these ops have fairly easy to
7576 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
7577 * See alu32 verifier tests for examples. The second class of
7578 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
7579 * with regards to tracking sign/unsigned bounds because the bits may
7580 * cross subreg boundaries in the alu64 case. When this happens we mark
7581 * the reg unbounded in the subreg bound space and use the resulting
7582 * tnum to calculate an approximation of the sign/unsigned bounds.
7586 scalar32_min_max_add(dst_reg, &src_reg);
7587 scalar_min_max_add(dst_reg, &src_reg);
7588 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
7591 scalar32_min_max_sub(dst_reg, &src_reg);
7592 scalar_min_max_sub(dst_reg, &src_reg);
7593 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
7596 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
7597 scalar32_min_max_mul(dst_reg, &src_reg);
7598 scalar_min_max_mul(dst_reg, &src_reg);
7601 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
7602 scalar32_min_max_and(dst_reg, &src_reg);
7603 scalar_min_max_and(dst_reg, &src_reg);
7606 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
7607 scalar32_min_max_or(dst_reg, &src_reg);
7608 scalar_min_max_or(dst_reg, &src_reg);
7611 dst_reg->var_off = tnum_xor(dst_reg->var_off, src_reg.var_off);
7612 scalar32_min_max_xor(dst_reg, &src_reg);
7613 scalar_min_max_xor(dst_reg, &src_reg);
7616 if (umax_val >= insn_bitness) {
7617 /* Shifts greater than 31 or 63 are undefined.
7618 * This includes shifts by a negative number.
7620 mark_reg_unknown(env, regs, insn->dst_reg);
7624 scalar32_min_max_lsh(dst_reg, &src_reg);
7626 scalar_min_max_lsh(dst_reg, &src_reg);
7629 if (umax_val >= insn_bitness) {
7630 /* Shifts greater than 31 or 63 are undefined.
7631 * This includes shifts by a negative number.
7633 mark_reg_unknown(env, regs, insn->dst_reg);
7637 scalar32_min_max_rsh(dst_reg, &src_reg);
7639 scalar_min_max_rsh(dst_reg, &src_reg);
7642 if (umax_val >= insn_bitness) {
7643 /* Shifts greater than 31 or 63 are undefined.
7644 * This includes shifts by a negative number.
7646 mark_reg_unknown(env, regs, insn->dst_reg);
7650 scalar32_min_max_arsh(dst_reg, &src_reg);
7652 scalar_min_max_arsh(dst_reg, &src_reg);
7655 mark_reg_unknown(env, regs, insn->dst_reg);
7659 /* ALU32 ops are zero extended into 64bit register */
7661 zext_32_to_64(dst_reg);
7663 __update_reg_bounds(dst_reg);
7664 __reg_deduce_bounds(dst_reg);
7665 __reg_bound_offset(dst_reg);
7669 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
7672 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
7673 struct bpf_insn *insn)
7675 struct bpf_verifier_state *vstate = env->cur_state;
7676 struct bpf_func_state *state = vstate->frame[vstate->curframe];
7677 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
7678 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
7679 u8 opcode = BPF_OP(insn->code);
7682 dst_reg = ®s[insn->dst_reg];
7684 if (dst_reg->type != SCALAR_VALUE)
7687 /* Make sure ID is cleared otherwise dst_reg min/max could be
7688 * incorrectly propagated into other registers by find_equal_scalars()
7691 if (BPF_SRC(insn->code) == BPF_X) {
7692 src_reg = ®s[insn->src_reg];
7693 if (src_reg->type != SCALAR_VALUE) {
7694 if (dst_reg->type != SCALAR_VALUE) {
7695 /* Combining two pointers by any ALU op yields
7696 * an arbitrary scalar. Disallow all math except
7697 * pointer subtraction
7699 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
7700 mark_reg_unknown(env, regs, insn->dst_reg);
7703 verbose(env, "R%d pointer %s pointer prohibited\n",
7705 bpf_alu_string[opcode >> 4]);
7708 /* scalar += pointer
7709 * This is legal, but we have to reverse our
7710 * src/dest handling in computing the range
7712 err = mark_chain_precision(env, insn->dst_reg);
7715 return adjust_ptr_min_max_vals(env, insn,
7718 } else if (ptr_reg) {
7719 /* pointer += scalar */
7720 err = mark_chain_precision(env, insn->src_reg);
7723 return adjust_ptr_min_max_vals(env, insn,
7727 /* Pretend the src is a reg with a known value, since we only
7728 * need to be able to read from this state.
7730 off_reg.type = SCALAR_VALUE;
7731 __mark_reg_known(&off_reg, insn->imm);
7733 if (ptr_reg) /* pointer += K */
7734 return adjust_ptr_min_max_vals(env, insn,
7738 /* Got here implies adding two SCALAR_VALUEs */
7739 if (WARN_ON_ONCE(ptr_reg)) {
7740 print_verifier_state(env, state);
7741 verbose(env, "verifier internal error: unexpected ptr_reg\n");
7744 if (WARN_ON(!src_reg)) {
7745 print_verifier_state(env, state);
7746 verbose(env, "verifier internal error: no src_reg\n");
7749 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
7752 /* check validity of 32-bit and 64-bit arithmetic operations */
7753 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
7755 struct bpf_reg_state *regs = cur_regs(env);
7756 u8 opcode = BPF_OP(insn->code);
7759 if (opcode == BPF_END || opcode == BPF_NEG) {
7760 if (opcode == BPF_NEG) {
7761 if (BPF_SRC(insn->code) != 0 ||
7762 insn->src_reg != BPF_REG_0 ||
7763 insn->off != 0 || insn->imm != 0) {
7764 verbose(env, "BPF_NEG uses reserved fields\n");
7768 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
7769 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
7770 BPF_CLASS(insn->code) == BPF_ALU64) {
7771 verbose(env, "BPF_END uses reserved fields\n");
7776 /* check src operand */
7777 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
7781 if (is_pointer_value(env, insn->dst_reg)) {
7782 verbose(env, "R%d pointer arithmetic prohibited\n",
7787 /* check dest operand */
7788 err = check_reg_arg(env, insn->dst_reg, DST_OP);
7792 } else if (opcode == BPF_MOV) {
7794 if (BPF_SRC(insn->code) == BPF_X) {
7795 if (insn->imm != 0 || insn->off != 0) {
7796 verbose(env, "BPF_MOV uses reserved fields\n");
7800 /* check src operand */
7801 err = check_reg_arg(env, insn->src_reg, SRC_OP);
7805 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
7806 verbose(env, "BPF_MOV uses reserved fields\n");
7811 /* check dest operand, mark as required later */
7812 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
7816 if (BPF_SRC(insn->code) == BPF_X) {
7817 struct bpf_reg_state *src_reg = regs + insn->src_reg;
7818 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
7820 if (BPF_CLASS(insn->code) == BPF_ALU64) {
7822 * copy register state to dest reg
7824 if (src_reg->type == SCALAR_VALUE && !src_reg->id)
7825 /* Assign src and dst registers the same ID
7826 * that will be used by find_equal_scalars()
7827 * to propagate min/max range.
7829 src_reg->id = ++env->id_gen;
7830 *dst_reg = *src_reg;
7831 dst_reg->live |= REG_LIVE_WRITTEN;
7832 dst_reg->subreg_def = DEF_NOT_SUBREG;
7835 if (is_pointer_value(env, insn->src_reg)) {
7837 "R%d partial copy of pointer\n",
7840 } else if (src_reg->type == SCALAR_VALUE) {
7841 *dst_reg = *src_reg;
7842 /* Make sure ID is cleared otherwise
7843 * dst_reg min/max could be incorrectly
7844 * propagated into src_reg by find_equal_scalars()
7847 dst_reg->live |= REG_LIVE_WRITTEN;
7848 dst_reg->subreg_def = env->insn_idx + 1;
7850 mark_reg_unknown(env, regs,
7853 zext_32_to_64(dst_reg);
7857 * remember the value we stored into this reg
7859 /* clear any state __mark_reg_known doesn't set */
7860 mark_reg_unknown(env, regs, insn->dst_reg);
7861 regs[insn->dst_reg].type = SCALAR_VALUE;
7862 if (BPF_CLASS(insn->code) == BPF_ALU64) {
7863 __mark_reg_known(regs + insn->dst_reg,
7866 __mark_reg_known(regs + insn->dst_reg,
7871 } else if (opcode > BPF_END) {
7872 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
7875 } else { /* all other ALU ops: and, sub, xor, add, ... */
7877 if (BPF_SRC(insn->code) == BPF_X) {
7878 if (insn->imm != 0 || insn->off != 0) {
7879 verbose(env, "BPF_ALU uses reserved fields\n");
7882 /* check src1 operand */
7883 err = check_reg_arg(env, insn->src_reg, SRC_OP);
7887 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
7888 verbose(env, "BPF_ALU uses reserved fields\n");
7893 /* check src2 operand */
7894 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
7898 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
7899 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
7900 verbose(env, "div by zero\n");
7904 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
7905 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
7906 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
7908 if (insn->imm < 0 || insn->imm >= size) {
7909 verbose(env, "invalid shift %d\n", insn->imm);
7914 /* check dest operand */
7915 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
7919 return adjust_reg_min_max_vals(env, insn);
7925 static void __find_good_pkt_pointers(struct bpf_func_state *state,
7926 struct bpf_reg_state *dst_reg,
7927 enum bpf_reg_type type, int new_range)
7929 struct bpf_reg_state *reg;
7932 for (i = 0; i < MAX_BPF_REG; i++) {
7933 reg = &state->regs[i];
7934 if (reg->type == type && reg->id == dst_reg->id)
7935 /* keep the maximum range already checked */
7936 reg->range = max(reg->range, new_range);
7939 bpf_for_each_spilled_reg(i, state, reg) {
7942 if (reg->type == type && reg->id == dst_reg->id)
7943 reg->range = max(reg->range, new_range);
7947 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
7948 struct bpf_reg_state *dst_reg,
7949 enum bpf_reg_type type,
7950 bool range_right_open)
7954 if (dst_reg->off < 0 ||
7955 (dst_reg->off == 0 && range_right_open))
7956 /* This doesn't give us any range */
7959 if (dst_reg->umax_value > MAX_PACKET_OFF ||
7960 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
7961 /* Risk of overflow. For instance, ptr + (1<<63) may be less
7962 * than pkt_end, but that's because it's also less than pkt.
7966 new_range = dst_reg->off;
7967 if (range_right_open)
7970 /* Examples for register markings:
7972 * pkt_data in dst register:
7976 * if (r2 > pkt_end) goto <handle exception>
7981 * if (r2 < pkt_end) goto <access okay>
7982 * <handle exception>
7985 * r2 == dst_reg, pkt_end == src_reg
7986 * r2=pkt(id=n,off=8,r=0)
7987 * r3=pkt(id=n,off=0,r=0)
7989 * pkt_data in src register:
7993 * if (pkt_end >= r2) goto <access okay>
7994 * <handle exception>
7998 * if (pkt_end <= r2) goto <handle exception>
8002 * pkt_end == dst_reg, r2 == src_reg
8003 * r2=pkt(id=n,off=8,r=0)
8004 * r3=pkt(id=n,off=0,r=0)
8006 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
8007 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
8008 * and [r3, r3 + 8-1) respectively is safe to access depending on
8012 /* If our ids match, then we must have the same max_value. And we
8013 * don't care about the other reg's fixed offset, since if it's too big
8014 * the range won't allow anything.
8015 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
8017 for (i = 0; i <= vstate->curframe; i++)
8018 __find_good_pkt_pointers(vstate->frame[i], dst_reg, type,
8022 static int is_branch32_taken(struct bpf_reg_state *reg, u32 val, u8 opcode)
8024 struct tnum subreg = tnum_subreg(reg->var_off);
8025 s32 sval = (s32)val;
8029 if (tnum_is_const(subreg))
8030 return !!tnum_equals_const(subreg, val);
8033 if (tnum_is_const(subreg))
8034 return !tnum_equals_const(subreg, val);
8037 if ((~subreg.mask & subreg.value) & val)
8039 if (!((subreg.mask | subreg.value) & val))
8043 if (reg->u32_min_value > val)
8045 else if (reg->u32_max_value <= val)
8049 if (reg->s32_min_value > sval)
8051 else if (reg->s32_max_value <= sval)
8055 if (reg->u32_max_value < val)
8057 else if (reg->u32_min_value >= val)
8061 if (reg->s32_max_value < sval)
8063 else if (reg->s32_min_value >= sval)
8067 if (reg->u32_min_value >= val)
8069 else if (reg->u32_max_value < val)
8073 if (reg->s32_min_value >= sval)
8075 else if (reg->s32_max_value < sval)
8079 if (reg->u32_max_value <= val)
8081 else if (reg->u32_min_value > val)
8085 if (reg->s32_max_value <= sval)
8087 else if (reg->s32_min_value > sval)
8096 static int is_branch64_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
8098 s64 sval = (s64)val;
8102 if (tnum_is_const(reg->var_off))
8103 return !!tnum_equals_const(reg->var_off, val);
8106 if (tnum_is_const(reg->var_off))
8107 return !tnum_equals_const(reg->var_off, val);
8110 if ((~reg->var_off.mask & reg->var_off.value) & val)
8112 if (!((reg->var_off.mask | reg->var_off.value) & val))
8116 if (reg->umin_value > val)
8118 else if (reg->umax_value <= val)
8122 if (reg->smin_value > sval)
8124 else if (reg->smax_value <= sval)
8128 if (reg->umax_value < val)
8130 else if (reg->umin_value >= val)
8134 if (reg->smax_value < sval)
8136 else if (reg->smin_value >= sval)
8140 if (reg->umin_value >= val)
8142 else if (reg->umax_value < val)
8146 if (reg->smin_value >= sval)
8148 else if (reg->smax_value < sval)
8152 if (reg->umax_value <= val)
8154 else if (reg->umin_value > val)
8158 if (reg->smax_value <= sval)
8160 else if (reg->smin_value > sval)
8168 /* compute branch direction of the expression "if (reg opcode val) goto target;"
8170 * 1 - branch will be taken and "goto target" will be executed
8171 * 0 - branch will not be taken and fall-through to next insn
8172 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value
8175 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
8178 if (__is_pointer_value(false, reg)) {
8179 if (!reg_type_not_null(reg->type))
8182 /* If pointer is valid tests against zero will fail so we can
8183 * use this to direct branch taken.
8199 return is_branch32_taken(reg, val, opcode);
8200 return is_branch64_taken(reg, val, opcode);
8203 static int flip_opcode(u32 opcode)
8205 /* How can we transform "a <op> b" into "b <op> a"? */
8206 static const u8 opcode_flip[16] = {
8207 /* these stay the same */
8208 [BPF_JEQ >> 4] = BPF_JEQ,
8209 [BPF_JNE >> 4] = BPF_JNE,
8210 [BPF_JSET >> 4] = BPF_JSET,
8211 /* these swap "lesser" and "greater" (L and G in the opcodes) */
8212 [BPF_JGE >> 4] = BPF_JLE,
8213 [BPF_JGT >> 4] = BPF_JLT,
8214 [BPF_JLE >> 4] = BPF_JGE,
8215 [BPF_JLT >> 4] = BPF_JGT,
8216 [BPF_JSGE >> 4] = BPF_JSLE,
8217 [BPF_JSGT >> 4] = BPF_JSLT,
8218 [BPF_JSLE >> 4] = BPF_JSGE,
8219 [BPF_JSLT >> 4] = BPF_JSGT
8221 return opcode_flip[opcode >> 4];
8224 static int is_pkt_ptr_branch_taken(struct bpf_reg_state *dst_reg,
8225 struct bpf_reg_state *src_reg,
8228 struct bpf_reg_state *pkt;
8230 if (src_reg->type == PTR_TO_PACKET_END) {
8232 } else if (dst_reg->type == PTR_TO_PACKET_END) {
8234 opcode = flip_opcode(opcode);
8239 if (pkt->range >= 0)
8244 /* pkt <= pkt_end */
8248 if (pkt->range == BEYOND_PKT_END)
8249 /* pkt has at last one extra byte beyond pkt_end */
8250 return opcode == BPF_JGT;
8256 /* pkt >= pkt_end */
8257 if (pkt->range == BEYOND_PKT_END || pkt->range == AT_PKT_END)
8258 return opcode == BPF_JGE;
8264 /* Adjusts the register min/max values in the case that the dst_reg is the
8265 * variable register that we are working on, and src_reg is a constant or we're
8266 * simply doing a BPF_K check.
8267 * In JEQ/JNE cases we also adjust the var_off values.
8269 static void reg_set_min_max(struct bpf_reg_state *true_reg,
8270 struct bpf_reg_state *false_reg,
8272 u8 opcode, bool is_jmp32)
8274 struct tnum false_32off = tnum_subreg(false_reg->var_off);
8275 struct tnum false_64off = false_reg->var_off;
8276 struct tnum true_32off = tnum_subreg(true_reg->var_off);
8277 struct tnum true_64off = true_reg->var_off;
8278 s64 sval = (s64)val;
8279 s32 sval32 = (s32)val32;
8281 /* If the dst_reg is a pointer, we can't learn anything about its
8282 * variable offset from the compare (unless src_reg were a pointer into
8283 * the same object, but we don't bother with that.
8284 * Since false_reg and true_reg have the same type by construction, we
8285 * only need to check one of them for pointerness.
8287 if (__is_pointer_value(false, false_reg))
8294 struct bpf_reg_state *reg =
8295 opcode == BPF_JEQ ? true_reg : false_reg;
8297 /* JEQ/JNE comparison doesn't change the register equivalence.
8299 * if (r1 == 42) goto label;
8301 * label: // here both r1 and r2 are known to be 42.
8303 * Hence when marking register as known preserve it's ID.
8306 __mark_reg32_known(reg, val32);
8308 ___mark_reg_known(reg, val);
8313 false_32off = tnum_and(false_32off, tnum_const(~val32));
8314 if (is_power_of_2(val32))
8315 true_32off = tnum_or(true_32off,
8318 false_64off = tnum_and(false_64off, tnum_const(~val));
8319 if (is_power_of_2(val))
8320 true_64off = tnum_or(true_64off,
8328 u32 false_umax = opcode == BPF_JGT ? val32 : val32 - 1;
8329 u32 true_umin = opcode == BPF_JGT ? val32 + 1 : val32;
8331 false_reg->u32_max_value = min(false_reg->u32_max_value,
8333 true_reg->u32_min_value = max(true_reg->u32_min_value,
8336 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
8337 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
8339 false_reg->umax_value = min(false_reg->umax_value, false_umax);
8340 true_reg->umin_value = max(true_reg->umin_value, true_umin);
8348 s32 false_smax = opcode == BPF_JSGT ? sval32 : sval32 - 1;
8349 s32 true_smin = opcode == BPF_JSGT ? sval32 + 1 : sval32;
8351 false_reg->s32_max_value = min(false_reg->s32_max_value, false_smax);
8352 true_reg->s32_min_value = max(true_reg->s32_min_value, true_smin);
8354 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
8355 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
8357 false_reg->smax_value = min(false_reg->smax_value, false_smax);
8358 true_reg->smin_value = max(true_reg->smin_value, true_smin);
8366 u32 false_umin = opcode == BPF_JLT ? val32 : val32 + 1;
8367 u32 true_umax = opcode == BPF_JLT ? val32 - 1 : val32;
8369 false_reg->u32_min_value = max(false_reg->u32_min_value,
8371 true_reg->u32_max_value = min(true_reg->u32_max_value,
8374 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
8375 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
8377 false_reg->umin_value = max(false_reg->umin_value, false_umin);
8378 true_reg->umax_value = min(true_reg->umax_value, true_umax);
8386 s32 false_smin = opcode == BPF_JSLT ? sval32 : sval32 + 1;
8387 s32 true_smax = opcode == BPF_JSLT ? sval32 - 1 : sval32;
8389 false_reg->s32_min_value = max(false_reg->s32_min_value, false_smin);
8390 true_reg->s32_max_value = min(true_reg->s32_max_value, true_smax);
8392 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
8393 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
8395 false_reg->smin_value = max(false_reg->smin_value, false_smin);
8396 true_reg->smax_value = min(true_reg->smax_value, true_smax);
8405 false_reg->var_off = tnum_or(tnum_clear_subreg(false_64off),
8406 tnum_subreg(false_32off));
8407 true_reg->var_off = tnum_or(tnum_clear_subreg(true_64off),
8408 tnum_subreg(true_32off));
8409 __reg_combine_32_into_64(false_reg);
8410 __reg_combine_32_into_64(true_reg);
8412 false_reg->var_off = false_64off;
8413 true_reg->var_off = true_64off;
8414 __reg_combine_64_into_32(false_reg);
8415 __reg_combine_64_into_32(true_reg);
8419 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
8422 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
8423 struct bpf_reg_state *false_reg,
8425 u8 opcode, bool is_jmp32)
8427 opcode = flip_opcode(opcode);
8428 /* This uses zero as "not present in table"; luckily the zero opcode,
8429 * BPF_JA, can't get here.
8432 reg_set_min_max(true_reg, false_reg, val, val32, opcode, is_jmp32);
8435 /* Regs are known to be equal, so intersect their min/max/var_off */
8436 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
8437 struct bpf_reg_state *dst_reg)
8439 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
8440 dst_reg->umin_value);
8441 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
8442 dst_reg->umax_value);
8443 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
8444 dst_reg->smin_value);
8445 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
8446 dst_reg->smax_value);
8447 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
8449 /* We might have learned new bounds from the var_off. */
8450 __update_reg_bounds(src_reg);
8451 __update_reg_bounds(dst_reg);
8452 /* We might have learned something about the sign bit. */
8453 __reg_deduce_bounds(src_reg);
8454 __reg_deduce_bounds(dst_reg);
8455 /* We might have learned some bits from the bounds. */
8456 __reg_bound_offset(src_reg);
8457 __reg_bound_offset(dst_reg);
8458 /* Intersecting with the old var_off might have improved our bounds
8459 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
8460 * then new var_off is (0; 0x7f...fc) which improves our umax.
8462 __update_reg_bounds(src_reg);
8463 __update_reg_bounds(dst_reg);
8466 static void reg_combine_min_max(struct bpf_reg_state *true_src,
8467 struct bpf_reg_state *true_dst,
8468 struct bpf_reg_state *false_src,
8469 struct bpf_reg_state *false_dst,
8474 __reg_combine_min_max(true_src, true_dst);
8477 __reg_combine_min_max(false_src, false_dst);
8482 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
8483 struct bpf_reg_state *reg, u32 id,
8486 if (reg_type_may_be_null(reg->type) && reg->id == id &&
8487 !WARN_ON_ONCE(!reg->id)) {
8488 /* Old offset (both fixed and variable parts) should
8489 * have been known-zero, because we don't allow pointer
8490 * arithmetic on pointers that might be NULL.
8492 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
8493 !tnum_equals_const(reg->var_off, 0) ||
8495 __mark_reg_known_zero(reg);
8499 reg->type = SCALAR_VALUE;
8500 /* We don't need id and ref_obj_id from this point
8501 * onwards anymore, thus we should better reset it,
8502 * so that state pruning has chances to take effect.
8505 reg->ref_obj_id = 0;
8510 mark_ptr_not_null_reg(reg);
8512 if (!reg_may_point_to_spin_lock(reg)) {
8513 /* For not-NULL ptr, reg->ref_obj_id will be reset
8514 * in release_reg_references().
8516 * reg->id is still used by spin_lock ptr. Other
8517 * than spin_lock ptr type, reg->id can be reset.
8524 static void __mark_ptr_or_null_regs(struct bpf_func_state *state, u32 id,
8527 struct bpf_reg_state *reg;
8530 for (i = 0; i < MAX_BPF_REG; i++)
8531 mark_ptr_or_null_reg(state, &state->regs[i], id, is_null);
8533 bpf_for_each_spilled_reg(i, state, reg) {
8536 mark_ptr_or_null_reg(state, reg, id, is_null);
8540 /* The logic is similar to find_good_pkt_pointers(), both could eventually
8541 * be folded together at some point.
8543 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
8546 struct bpf_func_state *state = vstate->frame[vstate->curframe];
8547 struct bpf_reg_state *regs = state->regs;
8548 u32 ref_obj_id = regs[regno].ref_obj_id;
8549 u32 id = regs[regno].id;
8552 if (ref_obj_id && ref_obj_id == id && is_null)
8553 /* regs[regno] is in the " == NULL" branch.
8554 * No one could have freed the reference state before
8555 * doing the NULL check.
8557 WARN_ON_ONCE(release_reference_state(state, id));
8559 for (i = 0; i <= vstate->curframe; i++)
8560 __mark_ptr_or_null_regs(vstate->frame[i], id, is_null);
8563 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
8564 struct bpf_reg_state *dst_reg,
8565 struct bpf_reg_state *src_reg,
8566 struct bpf_verifier_state *this_branch,
8567 struct bpf_verifier_state *other_branch)
8569 if (BPF_SRC(insn->code) != BPF_X)
8572 /* Pointers are always 64-bit. */
8573 if (BPF_CLASS(insn->code) == BPF_JMP32)
8576 switch (BPF_OP(insn->code)) {
8578 if ((dst_reg->type == PTR_TO_PACKET &&
8579 src_reg->type == PTR_TO_PACKET_END) ||
8580 (dst_reg->type == PTR_TO_PACKET_META &&
8581 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
8582 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
8583 find_good_pkt_pointers(this_branch, dst_reg,
8584 dst_reg->type, false);
8585 mark_pkt_end(other_branch, insn->dst_reg, true);
8586 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
8587 src_reg->type == PTR_TO_PACKET) ||
8588 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
8589 src_reg->type == PTR_TO_PACKET_META)) {
8590 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
8591 find_good_pkt_pointers(other_branch, src_reg,
8592 src_reg->type, true);
8593 mark_pkt_end(this_branch, insn->src_reg, false);
8599 if ((dst_reg->type == PTR_TO_PACKET &&
8600 src_reg->type == PTR_TO_PACKET_END) ||
8601 (dst_reg->type == PTR_TO_PACKET_META &&
8602 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
8603 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
8604 find_good_pkt_pointers(other_branch, dst_reg,
8605 dst_reg->type, true);
8606 mark_pkt_end(this_branch, insn->dst_reg, false);
8607 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
8608 src_reg->type == PTR_TO_PACKET) ||
8609 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
8610 src_reg->type == PTR_TO_PACKET_META)) {
8611 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
8612 find_good_pkt_pointers(this_branch, src_reg,
8613 src_reg->type, false);
8614 mark_pkt_end(other_branch, insn->src_reg, true);
8620 if ((dst_reg->type == PTR_TO_PACKET &&
8621 src_reg->type == PTR_TO_PACKET_END) ||
8622 (dst_reg->type == PTR_TO_PACKET_META &&
8623 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
8624 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
8625 find_good_pkt_pointers(this_branch, dst_reg,
8626 dst_reg->type, true);
8627 mark_pkt_end(other_branch, insn->dst_reg, false);
8628 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
8629 src_reg->type == PTR_TO_PACKET) ||
8630 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
8631 src_reg->type == PTR_TO_PACKET_META)) {
8632 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
8633 find_good_pkt_pointers(other_branch, src_reg,
8634 src_reg->type, false);
8635 mark_pkt_end(this_branch, insn->src_reg, true);
8641 if ((dst_reg->type == PTR_TO_PACKET &&
8642 src_reg->type == PTR_TO_PACKET_END) ||
8643 (dst_reg->type == PTR_TO_PACKET_META &&
8644 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
8645 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
8646 find_good_pkt_pointers(other_branch, dst_reg,
8647 dst_reg->type, false);
8648 mark_pkt_end(this_branch, insn->dst_reg, true);
8649 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
8650 src_reg->type == PTR_TO_PACKET) ||
8651 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
8652 src_reg->type == PTR_TO_PACKET_META)) {
8653 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
8654 find_good_pkt_pointers(this_branch, src_reg,
8655 src_reg->type, true);
8656 mark_pkt_end(other_branch, insn->src_reg, false);
8668 static void find_equal_scalars(struct bpf_verifier_state *vstate,
8669 struct bpf_reg_state *known_reg)
8671 struct bpf_func_state *state;
8672 struct bpf_reg_state *reg;
8675 for (i = 0; i <= vstate->curframe; i++) {
8676 state = vstate->frame[i];
8677 for (j = 0; j < MAX_BPF_REG; j++) {
8678 reg = &state->regs[j];
8679 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
8683 bpf_for_each_spilled_reg(j, state, reg) {
8686 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
8692 static int check_cond_jmp_op(struct bpf_verifier_env *env,
8693 struct bpf_insn *insn, int *insn_idx)
8695 struct bpf_verifier_state *this_branch = env->cur_state;
8696 struct bpf_verifier_state *other_branch;
8697 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
8698 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
8699 u8 opcode = BPF_OP(insn->code);
8704 /* Only conditional jumps are expected to reach here. */
8705 if (opcode == BPF_JA || opcode > BPF_JSLE) {
8706 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
8710 if (BPF_SRC(insn->code) == BPF_X) {
8711 if (insn->imm != 0) {
8712 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
8716 /* check src1 operand */
8717 err = check_reg_arg(env, insn->src_reg, SRC_OP);
8721 if (is_pointer_value(env, insn->src_reg)) {
8722 verbose(env, "R%d pointer comparison prohibited\n",
8726 src_reg = ®s[insn->src_reg];
8728 if (insn->src_reg != BPF_REG_0) {
8729 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
8734 /* check src2 operand */
8735 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
8739 dst_reg = ®s[insn->dst_reg];
8740 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
8742 if (BPF_SRC(insn->code) == BPF_K) {
8743 pred = is_branch_taken(dst_reg, insn->imm, opcode, is_jmp32);
8744 } else if (src_reg->type == SCALAR_VALUE &&
8745 is_jmp32 && tnum_is_const(tnum_subreg(src_reg->var_off))) {
8746 pred = is_branch_taken(dst_reg,
8747 tnum_subreg(src_reg->var_off).value,
8750 } else if (src_reg->type == SCALAR_VALUE &&
8751 !is_jmp32 && tnum_is_const(src_reg->var_off)) {
8752 pred = is_branch_taken(dst_reg,
8753 src_reg->var_off.value,
8756 } else if (reg_is_pkt_pointer_any(dst_reg) &&
8757 reg_is_pkt_pointer_any(src_reg) &&
8759 pred = is_pkt_ptr_branch_taken(dst_reg, src_reg, opcode);
8763 /* If we get here with a dst_reg pointer type it is because
8764 * above is_branch_taken() special cased the 0 comparison.
8766 if (!__is_pointer_value(false, dst_reg))
8767 err = mark_chain_precision(env, insn->dst_reg);
8768 if (BPF_SRC(insn->code) == BPF_X && !err &&
8769 !__is_pointer_value(false, src_reg))
8770 err = mark_chain_precision(env, insn->src_reg);
8776 /* Only follow the goto, ignore fall-through. If needed, push
8777 * the fall-through branch for simulation under speculative
8780 if (!env->bypass_spec_v1 &&
8781 !sanitize_speculative_path(env, insn, *insn_idx + 1,
8784 *insn_idx += insn->off;
8786 } else if (pred == 0) {
8787 /* Only follow the fall-through branch, since that's where the
8788 * program will go. If needed, push the goto branch for
8789 * simulation under speculative execution.
8791 if (!env->bypass_spec_v1 &&
8792 !sanitize_speculative_path(env, insn,
8793 *insn_idx + insn->off + 1,
8799 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
8803 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
8805 /* detect if we are comparing against a constant value so we can adjust
8806 * our min/max values for our dst register.
8807 * this is only legit if both are scalars (or pointers to the same
8808 * object, I suppose, but we don't support that right now), because
8809 * otherwise the different base pointers mean the offsets aren't
8812 if (BPF_SRC(insn->code) == BPF_X) {
8813 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
8815 if (dst_reg->type == SCALAR_VALUE &&
8816 src_reg->type == SCALAR_VALUE) {
8817 if (tnum_is_const(src_reg->var_off) ||
8819 tnum_is_const(tnum_subreg(src_reg->var_off))))
8820 reg_set_min_max(&other_branch_regs[insn->dst_reg],
8822 src_reg->var_off.value,
8823 tnum_subreg(src_reg->var_off).value,
8825 else if (tnum_is_const(dst_reg->var_off) ||
8827 tnum_is_const(tnum_subreg(dst_reg->var_off))))
8828 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
8830 dst_reg->var_off.value,
8831 tnum_subreg(dst_reg->var_off).value,
8833 else if (!is_jmp32 &&
8834 (opcode == BPF_JEQ || opcode == BPF_JNE))
8835 /* Comparing for equality, we can combine knowledge */
8836 reg_combine_min_max(&other_branch_regs[insn->src_reg],
8837 &other_branch_regs[insn->dst_reg],
8838 src_reg, dst_reg, opcode);
8840 !WARN_ON_ONCE(src_reg->id != other_branch_regs[insn->src_reg].id)) {
8841 find_equal_scalars(this_branch, src_reg);
8842 find_equal_scalars(other_branch, &other_branch_regs[insn->src_reg]);
8846 } else if (dst_reg->type == SCALAR_VALUE) {
8847 reg_set_min_max(&other_branch_regs[insn->dst_reg],
8848 dst_reg, insn->imm, (u32)insn->imm,
8852 if (dst_reg->type == SCALAR_VALUE && dst_reg->id &&
8853 !WARN_ON_ONCE(dst_reg->id != other_branch_regs[insn->dst_reg].id)) {
8854 find_equal_scalars(this_branch, dst_reg);
8855 find_equal_scalars(other_branch, &other_branch_regs[insn->dst_reg]);
8858 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
8859 * NOTE: these optimizations below are related with pointer comparison
8860 * which will never be JMP32.
8862 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
8863 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
8864 reg_type_may_be_null(dst_reg->type)) {
8865 /* Mark all identical registers in each branch as either
8866 * safe or unknown depending R == 0 or R != 0 conditional.
8868 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
8870 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
8872 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
8873 this_branch, other_branch) &&
8874 is_pointer_value(env, insn->dst_reg)) {
8875 verbose(env, "R%d pointer comparison prohibited\n",
8879 if (env->log.level & BPF_LOG_LEVEL)
8880 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
8884 /* verify BPF_LD_IMM64 instruction */
8885 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
8887 struct bpf_insn_aux_data *aux = cur_aux(env);
8888 struct bpf_reg_state *regs = cur_regs(env);
8889 struct bpf_reg_state *dst_reg;
8890 struct bpf_map *map;
8893 if (BPF_SIZE(insn->code) != BPF_DW) {
8894 verbose(env, "invalid BPF_LD_IMM insn\n");
8897 if (insn->off != 0) {
8898 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
8902 err = check_reg_arg(env, insn->dst_reg, DST_OP);
8906 dst_reg = ®s[insn->dst_reg];
8907 if (insn->src_reg == 0) {
8908 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
8910 dst_reg->type = SCALAR_VALUE;
8911 __mark_reg_known(®s[insn->dst_reg], imm);
8915 if (insn->src_reg == BPF_PSEUDO_BTF_ID) {
8916 mark_reg_known_zero(env, regs, insn->dst_reg);
8918 dst_reg->type = aux->btf_var.reg_type;
8919 switch (dst_reg->type) {
8921 dst_reg->mem_size = aux->btf_var.mem_size;
8924 case PTR_TO_PERCPU_BTF_ID:
8925 dst_reg->btf = aux->btf_var.btf;
8926 dst_reg->btf_id = aux->btf_var.btf_id;
8929 verbose(env, "bpf verifier is misconfigured\n");
8935 if (insn->src_reg == BPF_PSEUDO_FUNC) {
8936 struct bpf_prog_aux *aux = env->prog->aux;
8937 u32 subprogno = insn[1].imm;
8939 if (!aux->func_info) {
8940 verbose(env, "missing btf func_info\n");
8943 if (aux->func_info_aux[subprogno].linkage != BTF_FUNC_STATIC) {
8944 verbose(env, "callback function not static\n");
8948 dst_reg->type = PTR_TO_FUNC;
8949 dst_reg->subprogno = subprogno;
8953 map = env->used_maps[aux->map_index];
8954 mark_reg_known_zero(env, regs, insn->dst_reg);
8955 dst_reg->map_ptr = map;
8957 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE ||
8958 insn->src_reg == BPF_PSEUDO_MAP_IDX_VALUE) {
8959 dst_reg->type = PTR_TO_MAP_VALUE;
8960 dst_reg->off = aux->map_off;
8961 if (map_value_has_spin_lock(map))
8962 dst_reg->id = ++env->id_gen;
8963 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD ||
8964 insn->src_reg == BPF_PSEUDO_MAP_IDX) {
8965 dst_reg->type = CONST_PTR_TO_MAP;
8967 verbose(env, "bpf verifier is misconfigured\n");
8974 static bool may_access_skb(enum bpf_prog_type type)
8977 case BPF_PROG_TYPE_SOCKET_FILTER:
8978 case BPF_PROG_TYPE_SCHED_CLS:
8979 case BPF_PROG_TYPE_SCHED_ACT:
8986 /* verify safety of LD_ABS|LD_IND instructions:
8987 * - they can only appear in the programs where ctx == skb
8988 * - since they are wrappers of function calls, they scratch R1-R5 registers,
8989 * preserve R6-R9, and store return value into R0
8992 * ctx == skb == R6 == CTX
8995 * SRC == any register
8996 * IMM == 32-bit immediate
8999 * R0 - 8/16/32-bit skb data converted to cpu endianness
9001 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
9003 struct bpf_reg_state *regs = cur_regs(env);
9004 static const int ctx_reg = BPF_REG_6;
9005 u8 mode = BPF_MODE(insn->code);
9008 if (!may_access_skb(resolve_prog_type(env->prog))) {
9009 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
9013 if (!env->ops->gen_ld_abs) {
9014 verbose(env, "bpf verifier is misconfigured\n");
9018 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
9019 BPF_SIZE(insn->code) == BPF_DW ||
9020 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
9021 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
9025 /* check whether implicit source operand (register R6) is readable */
9026 err = check_reg_arg(env, ctx_reg, SRC_OP);
9030 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
9031 * gen_ld_abs() may terminate the program at runtime, leading to
9034 err = check_reference_leak(env);
9036 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
9040 if (env->cur_state->active_spin_lock) {
9041 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
9045 if (regs[ctx_reg].type != PTR_TO_CTX) {
9047 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
9051 if (mode == BPF_IND) {
9052 /* check explicit source operand */
9053 err = check_reg_arg(env, insn->src_reg, SRC_OP);
9058 err = check_ctx_reg(env, ®s[ctx_reg], ctx_reg);
9062 /* reset caller saved regs to unreadable */
9063 for (i = 0; i < CALLER_SAVED_REGS; i++) {
9064 mark_reg_not_init(env, regs, caller_saved[i]);
9065 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
9068 /* mark destination R0 register as readable, since it contains
9069 * the value fetched from the packet.
9070 * Already marked as written above.
9072 mark_reg_unknown(env, regs, BPF_REG_0);
9073 /* ld_abs load up to 32-bit skb data. */
9074 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
9078 static int check_return_code(struct bpf_verifier_env *env)
9080 struct tnum enforce_attach_type_range = tnum_unknown;
9081 const struct bpf_prog *prog = env->prog;
9082 struct bpf_reg_state *reg;
9083 struct tnum range = tnum_range(0, 1);
9084 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
9086 const bool is_subprog = env->cur_state->frame[0]->subprogno;
9088 /* LSM and struct_ops func-ptr's return type could be "void" */
9090 (prog_type == BPF_PROG_TYPE_STRUCT_OPS ||
9091 prog_type == BPF_PROG_TYPE_LSM) &&
9092 !prog->aux->attach_func_proto->type)
9095 /* eBPF calling convention is such that R0 is used
9096 * to return the value from eBPF program.
9097 * Make sure that it's readable at this time
9098 * of bpf_exit, which means that program wrote
9099 * something into it earlier
9101 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
9105 if (is_pointer_value(env, BPF_REG_0)) {
9106 verbose(env, "R0 leaks addr as return value\n");
9110 reg = cur_regs(env) + BPF_REG_0;
9112 if (reg->type != SCALAR_VALUE) {
9113 verbose(env, "At subprogram exit the register R0 is not a scalar value (%s)\n",
9114 reg_type_str[reg->type]);
9120 switch (prog_type) {
9121 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
9122 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
9123 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG ||
9124 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETPEERNAME ||
9125 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETPEERNAME ||
9126 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETSOCKNAME ||
9127 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETSOCKNAME)
9128 range = tnum_range(1, 1);
9129 if (env->prog->expected_attach_type == BPF_CGROUP_INET4_BIND ||
9130 env->prog->expected_attach_type == BPF_CGROUP_INET6_BIND)
9131 range = tnum_range(0, 3);
9133 case BPF_PROG_TYPE_CGROUP_SKB:
9134 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
9135 range = tnum_range(0, 3);
9136 enforce_attach_type_range = tnum_range(2, 3);
9139 case BPF_PROG_TYPE_CGROUP_SOCK:
9140 case BPF_PROG_TYPE_SOCK_OPS:
9141 case BPF_PROG_TYPE_CGROUP_DEVICE:
9142 case BPF_PROG_TYPE_CGROUP_SYSCTL:
9143 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
9145 case BPF_PROG_TYPE_RAW_TRACEPOINT:
9146 if (!env->prog->aux->attach_btf_id)
9148 range = tnum_const(0);
9150 case BPF_PROG_TYPE_TRACING:
9151 switch (env->prog->expected_attach_type) {
9152 case BPF_TRACE_FENTRY:
9153 case BPF_TRACE_FEXIT:
9154 range = tnum_const(0);
9156 case BPF_TRACE_RAW_TP:
9157 case BPF_MODIFY_RETURN:
9159 case BPF_TRACE_ITER:
9165 case BPF_PROG_TYPE_SK_LOOKUP:
9166 range = tnum_range(SK_DROP, SK_PASS);
9168 case BPF_PROG_TYPE_EXT:
9169 /* freplace program can return anything as its return value
9170 * depends on the to-be-replaced kernel func or bpf program.
9176 if (reg->type != SCALAR_VALUE) {
9177 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
9178 reg_type_str[reg->type]);
9182 if (!tnum_in(range, reg->var_off)) {
9183 verbose_invalid_scalar(env, reg, &range, "program exit", "R0");
9187 if (!tnum_is_unknown(enforce_attach_type_range) &&
9188 tnum_in(enforce_attach_type_range, reg->var_off))
9189 env->prog->enforce_expected_attach_type = 1;
9193 /* non-recursive DFS pseudo code
9194 * 1 procedure DFS-iterative(G,v):
9195 * 2 label v as discovered
9196 * 3 let S be a stack
9198 * 5 while S is not empty
9200 * 7 if t is what we're looking for:
9202 * 9 for all edges e in G.adjacentEdges(t) do
9203 * 10 if edge e is already labelled
9204 * 11 continue with the next edge
9205 * 12 w <- G.adjacentVertex(t,e)
9206 * 13 if vertex w is not discovered and not explored
9207 * 14 label e as tree-edge
9208 * 15 label w as discovered
9211 * 18 else if vertex w is discovered
9212 * 19 label e as back-edge
9214 * 21 // vertex w is explored
9215 * 22 label e as forward- or cross-edge
9216 * 23 label t as explored
9221 * 0x11 - discovered and fall-through edge labelled
9222 * 0x12 - discovered and fall-through and branch edges labelled
9233 static u32 state_htab_size(struct bpf_verifier_env *env)
9235 return env->prog->len;
9238 static struct bpf_verifier_state_list **explored_state(
9239 struct bpf_verifier_env *env,
9242 struct bpf_verifier_state *cur = env->cur_state;
9243 struct bpf_func_state *state = cur->frame[cur->curframe];
9245 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
9248 static void init_explored_state(struct bpf_verifier_env *env, int idx)
9250 env->insn_aux_data[idx].prune_point = true;
9258 /* t, w, e - match pseudo-code above:
9259 * t - index of current instruction
9260 * w - next instruction
9263 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env,
9266 int *insn_stack = env->cfg.insn_stack;
9267 int *insn_state = env->cfg.insn_state;
9269 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
9270 return DONE_EXPLORING;
9272 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
9273 return DONE_EXPLORING;
9275 if (w < 0 || w >= env->prog->len) {
9276 verbose_linfo(env, t, "%d: ", t);
9277 verbose(env, "jump out of range from insn %d to %d\n", t, w);
9282 /* mark branch target for state pruning */
9283 init_explored_state(env, w);
9285 if (insn_state[w] == 0) {
9287 insn_state[t] = DISCOVERED | e;
9288 insn_state[w] = DISCOVERED;
9289 if (env->cfg.cur_stack >= env->prog->len)
9291 insn_stack[env->cfg.cur_stack++] = w;
9292 return KEEP_EXPLORING;
9293 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
9294 if (loop_ok && env->bpf_capable)
9295 return DONE_EXPLORING;
9296 verbose_linfo(env, t, "%d: ", t);
9297 verbose_linfo(env, w, "%d: ", w);
9298 verbose(env, "back-edge from insn %d to %d\n", t, w);
9300 } else if (insn_state[w] == EXPLORED) {
9301 /* forward- or cross-edge */
9302 insn_state[t] = DISCOVERED | e;
9304 verbose(env, "insn state internal bug\n");
9307 return DONE_EXPLORING;
9310 static int visit_func_call_insn(int t, int insn_cnt,
9311 struct bpf_insn *insns,
9312 struct bpf_verifier_env *env,
9317 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
9321 if (t + 1 < insn_cnt)
9322 init_explored_state(env, t + 1);
9324 init_explored_state(env, t);
9325 ret = push_insn(t, t + insns[t].imm + 1, BRANCH,
9331 /* Visits the instruction at index t and returns one of the following:
9332 * < 0 - an error occurred
9333 * DONE_EXPLORING - the instruction was fully explored
9334 * KEEP_EXPLORING - there is still work to be done before it is fully explored
9336 static int visit_insn(int t, int insn_cnt, struct bpf_verifier_env *env)
9338 struct bpf_insn *insns = env->prog->insnsi;
9341 if (bpf_pseudo_func(insns + t))
9342 return visit_func_call_insn(t, insn_cnt, insns, env, true);
9344 /* All non-branch instructions have a single fall-through edge. */
9345 if (BPF_CLASS(insns[t].code) != BPF_JMP &&
9346 BPF_CLASS(insns[t].code) != BPF_JMP32)
9347 return push_insn(t, t + 1, FALLTHROUGH, env, false);
9349 switch (BPF_OP(insns[t].code)) {
9351 return DONE_EXPLORING;
9354 return visit_func_call_insn(t, insn_cnt, insns, env,
9355 insns[t].src_reg == BPF_PSEUDO_CALL);
9358 if (BPF_SRC(insns[t].code) != BPF_K)
9361 /* unconditional jump with single edge */
9362 ret = push_insn(t, t + insns[t].off + 1, FALLTHROUGH, env,
9367 /* unconditional jmp is not a good pruning point,
9368 * but it's marked, since backtracking needs
9369 * to record jmp history in is_state_visited().
9371 init_explored_state(env, t + insns[t].off + 1);
9372 /* tell verifier to check for equivalent states
9373 * after every call and jump
9375 if (t + 1 < insn_cnt)
9376 init_explored_state(env, t + 1);
9381 /* conditional jump with two edges */
9382 init_explored_state(env, t);
9383 ret = push_insn(t, t + 1, FALLTHROUGH, env, true);
9387 return push_insn(t, t + insns[t].off + 1, BRANCH, env, true);
9391 /* non-recursive depth-first-search to detect loops in BPF program
9392 * loop == back-edge in directed graph
9394 static int check_cfg(struct bpf_verifier_env *env)
9396 int insn_cnt = env->prog->len;
9397 int *insn_stack, *insn_state;
9401 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
9405 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
9411 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
9412 insn_stack[0] = 0; /* 0 is the first instruction */
9413 env->cfg.cur_stack = 1;
9415 while (env->cfg.cur_stack > 0) {
9416 int t = insn_stack[env->cfg.cur_stack - 1];
9418 ret = visit_insn(t, insn_cnt, env);
9420 case DONE_EXPLORING:
9421 insn_state[t] = EXPLORED;
9422 env->cfg.cur_stack--;
9424 case KEEP_EXPLORING:
9428 verbose(env, "visit_insn internal bug\n");
9435 if (env->cfg.cur_stack < 0) {
9436 verbose(env, "pop stack internal bug\n");
9441 for (i = 0; i < insn_cnt; i++) {
9442 if (insn_state[i] != EXPLORED) {
9443 verbose(env, "unreachable insn %d\n", i);
9448 ret = 0; /* cfg looks good */
9453 env->cfg.insn_state = env->cfg.insn_stack = NULL;
9457 static int check_abnormal_return(struct bpf_verifier_env *env)
9461 for (i = 1; i < env->subprog_cnt; i++) {
9462 if (env->subprog_info[i].has_ld_abs) {
9463 verbose(env, "LD_ABS is not allowed in subprogs without BTF\n");
9466 if (env->subprog_info[i].has_tail_call) {
9467 verbose(env, "tail_call is not allowed in subprogs without BTF\n");
9474 /* The minimum supported BTF func info size */
9475 #define MIN_BPF_FUNCINFO_SIZE 8
9476 #define MAX_FUNCINFO_REC_SIZE 252
9478 static int check_btf_func(struct bpf_verifier_env *env,
9479 const union bpf_attr *attr,
9482 const struct btf_type *type, *func_proto, *ret_type;
9483 u32 i, nfuncs, urec_size, min_size;
9484 u32 krec_size = sizeof(struct bpf_func_info);
9485 struct bpf_func_info *krecord;
9486 struct bpf_func_info_aux *info_aux = NULL;
9487 struct bpf_prog *prog;
9488 const struct btf *btf;
9490 u32 prev_offset = 0;
9494 nfuncs = attr->func_info_cnt;
9496 if (check_abnormal_return(env))
9501 if (nfuncs != env->subprog_cnt) {
9502 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
9506 urec_size = attr->func_info_rec_size;
9507 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
9508 urec_size > MAX_FUNCINFO_REC_SIZE ||
9509 urec_size % sizeof(u32)) {
9510 verbose(env, "invalid func info rec size %u\n", urec_size);
9515 btf = prog->aux->btf;
9517 urecord = make_bpfptr(attr->func_info, uattr.is_kernel);
9518 min_size = min_t(u32, krec_size, urec_size);
9520 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
9523 info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN);
9527 for (i = 0; i < nfuncs; i++) {
9528 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
9530 if (ret == -E2BIG) {
9531 verbose(env, "nonzero tailing record in func info");
9532 /* set the size kernel expects so loader can zero
9533 * out the rest of the record.
9535 if (copy_to_bpfptr_offset(uattr,
9536 offsetof(union bpf_attr, func_info_rec_size),
9537 &min_size, sizeof(min_size)))
9543 if (copy_from_bpfptr(&krecord[i], urecord, min_size)) {
9548 /* check insn_off */
9551 if (krecord[i].insn_off) {
9553 "nonzero insn_off %u for the first func info record",
9554 krecord[i].insn_off);
9557 } else if (krecord[i].insn_off <= prev_offset) {
9559 "same or smaller insn offset (%u) than previous func info record (%u)",
9560 krecord[i].insn_off, prev_offset);
9564 if (env->subprog_info[i].start != krecord[i].insn_off) {
9565 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
9570 type = btf_type_by_id(btf, krecord[i].type_id);
9571 if (!type || !btf_type_is_func(type)) {
9572 verbose(env, "invalid type id %d in func info",
9573 krecord[i].type_id);
9576 info_aux[i].linkage = BTF_INFO_VLEN(type->info);
9578 func_proto = btf_type_by_id(btf, type->type);
9579 if (unlikely(!func_proto || !btf_type_is_func_proto(func_proto)))
9580 /* btf_func_check() already verified it during BTF load */
9582 ret_type = btf_type_skip_modifiers(btf, func_proto->type, NULL);
9584 btf_type_is_small_int(ret_type) || btf_type_is_enum(ret_type);
9585 if (i && !scalar_return && env->subprog_info[i].has_ld_abs) {
9586 verbose(env, "LD_ABS is only allowed in functions that return 'int'.\n");
9589 if (i && !scalar_return && env->subprog_info[i].has_tail_call) {
9590 verbose(env, "tail_call is only allowed in functions that return 'int'.\n");
9594 prev_offset = krecord[i].insn_off;
9595 bpfptr_add(&urecord, urec_size);
9598 prog->aux->func_info = krecord;
9599 prog->aux->func_info_cnt = nfuncs;
9600 prog->aux->func_info_aux = info_aux;
9609 static void adjust_btf_func(struct bpf_verifier_env *env)
9611 struct bpf_prog_aux *aux = env->prog->aux;
9614 if (!aux->func_info)
9617 for (i = 0; i < env->subprog_cnt; i++)
9618 aux->func_info[i].insn_off = env->subprog_info[i].start;
9621 #define MIN_BPF_LINEINFO_SIZE (offsetof(struct bpf_line_info, line_col) + \
9622 sizeof(((struct bpf_line_info *)(0))->line_col))
9623 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
9625 static int check_btf_line(struct bpf_verifier_env *env,
9626 const union bpf_attr *attr,
9629 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
9630 struct bpf_subprog_info *sub;
9631 struct bpf_line_info *linfo;
9632 struct bpf_prog *prog;
9633 const struct btf *btf;
9637 nr_linfo = attr->line_info_cnt;
9641 rec_size = attr->line_info_rec_size;
9642 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
9643 rec_size > MAX_LINEINFO_REC_SIZE ||
9644 rec_size & (sizeof(u32) - 1))
9647 /* Need to zero it in case the userspace may
9648 * pass in a smaller bpf_line_info object.
9650 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
9651 GFP_KERNEL | __GFP_NOWARN);
9656 btf = prog->aux->btf;
9659 sub = env->subprog_info;
9660 ulinfo = make_bpfptr(attr->line_info, uattr.is_kernel);
9661 expected_size = sizeof(struct bpf_line_info);
9662 ncopy = min_t(u32, expected_size, rec_size);
9663 for (i = 0; i < nr_linfo; i++) {
9664 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
9666 if (err == -E2BIG) {
9667 verbose(env, "nonzero tailing record in line_info");
9668 if (copy_to_bpfptr_offset(uattr,
9669 offsetof(union bpf_attr, line_info_rec_size),
9670 &expected_size, sizeof(expected_size)))
9676 if (copy_from_bpfptr(&linfo[i], ulinfo, ncopy)) {
9682 * Check insn_off to ensure
9683 * 1) strictly increasing AND
9684 * 2) bounded by prog->len
9686 * The linfo[0].insn_off == 0 check logically falls into
9687 * the later "missing bpf_line_info for func..." case
9688 * because the first linfo[0].insn_off must be the
9689 * first sub also and the first sub must have
9690 * subprog_info[0].start == 0.
9692 if ((i && linfo[i].insn_off <= prev_offset) ||
9693 linfo[i].insn_off >= prog->len) {
9694 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
9695 i, linfo[i].insn_off, prev_offset,
9701 if (!prog->insnsi[linfo[i].insn_off].code) {
9703 "Invalid insn code at line_info[%u].insn_off\n",
9709 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
9710 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
9711 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
9716 if (s != env->subprog_cnt) {
9717 if (linfo[i].insn_off == sub[s].start) {
9718 sub[s].linfo_idx = i;
9720 } else if (sub[s].start < linfo[i].insn_off) {
9721 verbose(env, "missing bpf_line_info for func#%u\n", s);
9727 prev_offset = linfo[i].insn_off;
9728 bpfptr_add(&ulinfo, rec_size);
9731 if (s != env->subprog_cnt) {
9732 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
9733 env->subprog_cnt - s, s);
9738 prog->aux->linfo = linfo;
9739 prog->aux->nr_linfo = nr_linfo;
9748 static int check_btf_info(struct bpf_verifier_env *env,
9749 const union bpf_attr *attr,
9755 if (!attr->func_info_cnt && !attr->line_info_cnt) {
9756 if (check_abnormal_return(env))
9761 btf = btf_get_by_fd(attr->prog_btf_fd);
9763 return PTR_ERR(btf);
9764 if (btf_is_kernel(btf)) {
9768 env->prog->aux->btf = btf;
9770 err = check_btf_func(env, attr, uattr);
9774 err = check_btf_line(env, attr, uattr);
9781 /* check %cur's range satisfies %old's */
9782 static bool range_within(struct bpf_reg_state *old,
9783 struct bpf_reg_state *cur)
9785 return old->umin_value <= cur->umin_value &&
9786 old->umax_value >= cur->umax_value &&
9787 old->smin_value <= cur->smin_value &&
9788 old->smax_value >= cur->smax_value &&
9789 old->u32_min_value <= cur->u32_min_value &&
9790 old->u32_max_value >= cur->u32_max_value &&
9791 old->s32_min_value <= cur->s32_min_value &&
9792 old->s32_max_value >= cur->s32_max_value;
9795 /* If in the old state two registers had the same id, then they need to have
9796 * the same id in the new state as well. But that id could be different from
9797 * the old state, so we need to track the mapping from old to new ids.
9798 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
9799 * regs with old id 5 must also have new id 9 for the new state to be safe. But
9800 * regs with a different old id could still have new id 9, we don't care about
9802 * So we look through our idmap to see if this old id has been seen before. If
9803 * so, we require the new id to match; otherwise, we add the id pair to the map.
9805 static bool check_ids(u32 old_id, u32 cur_id, struct bpf_id_pair *idmap)
9809 for (i = 0; i < BPF_ID_MAP_SIZE; i++) {
9810 if (!idmap[i].old) {
9811 /* Reached an empty slot; haven't seen this id before */
9812 idmap[i].old = old_id;
9813 idmap[i].cur = cur_id;
9816 if (idmap[i].old == old_id)
9817 return idmap[i].cur == cur_id;
9819 /* We ran out of idmap slots, which should be impossible */
9824 static void clean_func_state(struct bpf_verifier_env *env,
9825 struct bpf_func_state *st)
9827 enum bpf_reg_liveness live;
9830 for (i = 0; i < BPF_REG_FP; i++) {
9831 live = st->regs[i].live;
9832 /* liveness must not touch this register anymore */
9833 st->regs[i].live |= REG_LIVE_DONE;
9834 if (!(live & REG_LIVE_READ))
9835 /* since the register is unused, clear its state
9836 * to make further comparison simpler
9838 __mark_reg_not_init(env, &st->regs[i]);
9841 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
9842 live = st->stack[i].spilled_ptr.live;
9843 /* liveness must not touch this stack slot anymore */
9844 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
9845 if (!(live & REG_LIVE_READ)) {
9846 __mark_reg_not_init(env, &st->stack[i].spilled_ptr);
9847 for (j = 0; j < BPF_REG_SIZE; j++)
9848 st->stack[i].slot_type[j] = STACK_INVALID;
9853 static void clean_verifier_state(struct bpf_verifier_env *env,
9854 struct bpf_verifier_state *st)
9858 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
9859 /* all regs in this state in all frames were already marked */
9862 for (i = 0; i <= st->curframe; i++)
9863 clean_func_state(env, st->frame[i]);
9866 /* the parentage chains form a tree.
9867 * the verifier states are added to state lists at given insn and
9868 * pushed into state stack for future exploration.
9869 * when the verifier reaches bpf_exit insn some of the verifer states
9870 * stored in the state lists have their final liveness state already,
9871 * but a lot of states will get revised from liveness point of view when
9872 * the verifier explores other branches.
9875 * 2: if r1 == 100 goto pc+1
9878 * when the verifier reaches exit insn the register r0 in the state list of
9879 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
9880 * of insn 2 and goes exploring further. At the insn 4 it will walk the
9881 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
9883 * Since the verifier pushes the branch states as it sees them while exploring
9884 * the program the condition of walking the branch instruction for the second
9885 * time means that all states below this branch were already explored and
9886 * their final liveness marks are already propagated.
9887 * Hence when the verifier completes the search of state list in is_state_visited()
9888 * we can call this clean_live_states() function to mark all liveness states
9889 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
9891 * This function also clears the registers and stack for states that !READ
9892 * to simplify state merging.
9894 * Important note here that walking the same branch instruction in the callee
9895 * doesn't meant that the states are DONE. The verifier has to compare
9898 static void clean_live_states(struct bpf_verifier_env *env, int insn,
9899 struct bpf_verifier_state *cur)
9901 struct bpf_verifier_state_list *sl;
9904 sl = *explored_state(env, insn);
9906 if (sl->state.branches)
9908 if (sl->state.insn_idx != insn ||
9909 sl->state.curframe != cur->curframe)
9911 for (i = 0; i <= cur->curframe; i++)
9912 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
9914 clean_verifier_state(env, &sl->state);
9920 /* Returns true if (rold safe implies rcur safe) */
9921 static bool regsafe(struct bpf_verifier_env *env, struct bpf_reg_state *rold,
9922 struct bpf_reg_state *rcur, struct bpf_id_pair *idmap)
9926 if (!(rold->live & REG_LIVE_READ))
9927 /* explored state didn't use this */
9930 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
9932 if (rold->type == PTR_TO_STACK)
9933 /* two stack pointers are equal only if they're pointing to
9934 * the same stack frame, since fp-8 in foo != fp-8 in bar
9936 return equal && rold->frameno == rcur->frameno;
9941 if (rold->type == NOT_INIT)
9942 /* explored state can't have used this */
9944 if (rcur->type == NOT_INIT)
9946 switch (rold->type) {
9948 if (env->explore_alu_limits)
9950 if (rcur->type == SCALAR_VALUE) {
9951 if (!rold->precise && !rcur->precise)
9953 /* new val must satisfy old val knowledge */
9954 return range_within(rold, rcur) &&
9955 tnum_in(rold->var_off, rcur->var_off);
9957 /* We're trying to use a pointer in place of a scalar.
9958 * Even if the scalar was unbounded, this could lead to
9959 * pointer leaks because scalars are allowed to leak
9960 * while pointers are not. We could make this safe in
9961 * special cases if root is calling us, but it's
9962 * probably not worth the hassle.
9966 case PTR_TO_MAP_KEY:
9967 case PTR_TO_MAP_VALUE:
9968 /* If the new min/max/var_off satisfy the old ones and
9969 * everything else matches, we are OK.
9970 * 'id' is not compared, since it's only used for maps with
9971 * bpf_spin_lock inside map element and in such cases if
9972 * the rest of the prog is valid for one map element then
9973 * it's valid for all map elements regardless of the key
9974 * used in bpf_map_lookup()
9976 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
9977 range_within(rold, rcur) &&
9978 tnum_in(rold->var_off, rcur->var_off);
9979 case PTR_TO_MAP_VALUE_OR_NULL:
9980 /* a PTR_TO_MAP_VALUE could be safe to use as a
9981 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
9982 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
9983 * checked, doing so could have affected others with the same
9984 * id, and we can't check for that because we lost the id when
9985 * we converted to a PTR_TO_MAP_VALUE.
9987 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
9989 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
9991 /* Check our ids match any regs they're supposed to */
9992 return check_ids(rold->id, rcur->id, idmap);
9993 case PTR_TO_PACKET_META:
9995 if (rcur->type != rold->type)
9997 /* We must have at least as much range as the old ptr
9998 * did, so that any accesses which were safe before are
9999 * still safe. This is true even if old range < old off,
10000 * since someone could have accessed through (ptr - k), or
10001 * even done ptr -= k in a register, to get a safe access.
10003 if (rold->range > rcur->range)
10005 /* If the offsets don't match, we can't trust our alignment;
10006 * nor can we be sure that we won't fall out of range.
10008 if (rold->off != rcur->off)
10010 /* id relations must be preserved */
10011 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
10013 /* new val must satisfy old val knowledge */
10014 return range_within(rold, rcur) &&
10015 tnum_in(rold->var_off, rcur->var_off);
10017 case CONST_PTR_TO_MAP:
10018 case PTR_TO_PACKET_END:
10019 case PTR_TO_FLOW_KEYS:
10020 case PTR_TO_SOCKET:
10021 case PTR_TO_SOCKET_OR_NULL:
10022 case PTR_TO_SOCK_COMMON:
10023 case PTR_TO_SOCK_COMMON_OR_NULL:
10024 case PTR_TO_TCP_SOCK:
10025 case PTR_TO_TCP_SOCK_OR_NULL:
10026 case PTR_TO_XDP_SOCK:
10027 /* Only valid matches are exact, which memcmp() above
10028 * would have accepted
10031 /* Don't know what's going on, just say it's not safe */
10035 /* Shouldn't get here; if we do, say it's not safe */
10040 static bool stacksafe(struct bpf_verifier_env *env, struct bpf_func_state *old,
10041 struct bpf_func_state *cur, struct bpf_id_pair *idmap)
10045 /* walk slots of the explored stack and ignore any additional
10046 * slots in the current stack, since explored(safe) state
10049 for (i = 0; i < old->allocated_stack; i++) {
10050 spi = i / BPF_REG_SIZE;
10052 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
10053 i += BPF_REG_SIZE - 1;
10054 /* explored state didn't use this */
10058 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
10061 /* explored stack has more populated slots than current stack
10062 * and these slots were used
10064 if (i >= cur->allocated_stack)
10067 /* if old state was safe with misc data in the stack
10068 * it will be safe with zero-initialized stack.
10069 * The opposite is not true
10071 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
10072 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
10074 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
10075 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
10076 /* Ex: old explored (safe) state has STACK_SPILL in
10077 * this stack slot, but current has STACK_MISC ->
10078 * this verifier states are not equivalent,
10079 * return false to continue verification of this path
10082 if (i % BPF_REG_SIZE)
10084 if (old->stack[spi].slot_type[0] != STACK_SPILL)
10086 if (!regsafe(env, &old->stack[spi].spilled_ptr,
10087 &cur->stack[spi].spilled_ptr, idmap))
10088 /* when explored and current stack slot are both storing
10089 * spilled registers, check that stored pointers types
10090 * are the same as well.
10091 * Ex: explored safe path could have stored
10092 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
10093 * but current path has stored:
10094 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
10095 * such verifier states are not equivalent.
10096 * return false to continue verification of this path
10103 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
10105 if (old->acquired_refs != cur->acquired_refs)
10107 return !memcmp(old->refs, cur->refs,
10108 sizeof(*old->refs) * old->acquired_refs);
10111 /* compare two verifier states
10113 * all states stored in state_list are known to be valid, since
10114 * verifier reached 'bpf_exit' instruction through them
10116 * this function is called when verifier exploring different branches of
10117 * execution popped from the state stack. If it sees an old state that has
10118 * more strict register state and more strict stack state then this execution
10119 * branch doesn't need to be explored further, since verifier already
10120 * concluded that more strict state leads to valid finish.
10122 * Therefore two states are equivalent if register state is more conservative
10123 * and explored stack state is more conservative than the current one.
10126 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
10127 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
10129 * In other words if current stack state (one being explored) has more
10130 * valid slots than old one that already passed validation, it means
10131 * the verifier can stop exploring and conclude that current state is valid too
10133 * Similarly with registers. If explored state has register type as invalid
10134 * whereas register type in current state is meaningful, it means that
10135 * the current state will reach 'bpf_exit' instruction safely
10137 static bool func_states_equal(struct bpf_verifier_env *env, struct bpf_func_state *old,
10138 struct bpf_func_state *cur)
10142 memset(env->idmap_scratch, 0, sizeof(env->idmap_scratch));
10143 for (i = 0; i < MAX_BPF_REG; i++)
10144 if (!regsafe(env, &old->regs[i], &cur->regs[i],
10145 env->idmap_scratch))
10148 if (!stacksafe(env, old, cur, env->idmap_scratch))
10151 if (!refsafe(old, cur))
10157 static bool states_equal(struct bpf_verifier_env *env,
10158 struct bpf_verifier_state *old,
10159 struct bpf_verifier_state *cur)
10163 if (old->curframe != cur->curframe)
10166 /* Verification state from speculative execution simulation
10167 * must never prune a non-speculative execution one.
10169 if (old->speculative && !cur->speculative)
10172 if (old->active_spin_lock != cur->active_spin_lock)
10175 /* for states to be equal callsites have to be the same
10176 * and all frame states need to be equivalent
10178 for (i = 0; i <= old->curframe; i++) {
10179 if (old->frame[i]->callsite != cur->frame[i]->callsite)
10181 if (!func_states_equal(env, old->frame[i], cur->frame[i]))
10187 /* Return 0 if no propagation happened. Return negative error code if error
10188 * happened. Otherwise, return the propagated bit.
10190 static int propagate_liveness_reg(struct bpf_verifier_env *env,
10191 struct bpf_reg_state *reg,
10192 struct bpf_reg_state *parent_reg)
10194 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
10195 u8 flag = reg->live & REG_LIVE_READ;
10198 /* When comes here, read flags of PARENT_REG or REG could be any of
10199 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
10200 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
10202 if (parent_flag == REG_LIVE_READ64 ||
10203 /* Or if there is no read flag from REG. */
10205 /* Or if the read flag from REG is the same as PARENT_REG. */
10206 parent_flag == flag)
10209 err = mark_reg_read(env, reg, parent_reg, flag);
10216 /* A write screens off any subsequent reads; but write marks come from the
10217 * straight-line code between a state and its parent. When we arrive at an
10218 * equivalent state (jump target or such) we didn't arrive by the straight-line
10219 * code, so read marks in the state must propagate to the parent regardless
10220 * of the state's write marks. That's what 'parent == state->parent' comparison
10221 * in mark_reg_read() is for.
10223 static int propagate_liveness(struct bpf_verifier_env *env,
10224 const struct bpf_verifier_state *vstate,
10225 struct bpf_verifier_state *vparent)
10227 struct bpf_reg_state *state_reg, *parent_reg;
10228 struct bpf_func_state *state, *parent;
10229 int i, frame, err = 0;
10231 if (vparent->curframe != vstate->curframe) {
10232 WARN(1, "propagate_live: parent frame %d current frame %d\n",
10233 vparent->curframe, vstate->curframe);
10236 /* Propagate read liveness of registers... */
10237 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
10238 for (frame = 0; frame <= vstate->curframe; frame++) {
10239 parent = vparent->frame[frame];
10240 state = vstate->frame[frame];
10241 parent_reg = parent->regs;
10242 state_reg = state->regs;
10243 /* We don't need to worry about FP liveness, it's read-only */
10244 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
10245 err = propagate_liveness_reg(env, &state_reg[i],
10249 if (err == REG_LIVE_READ64)
10250 mark_insn_zext(env, &parent_reg[i]);
10253 /* Propagate stack slots. */
10254 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
10255 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
10256 parent_reg = &parent->stack[i].spilled_ptr;
10257 state_reg = &state->stack[i].spilled_ptr;
10258 err = propagate_liveness_reg(env, state_reg,
10267 /* find precise scalars in the previous equivalent state and
10268 * propagate them into the current state
10270 static int propagate_precision(struct bpf_verifier_env *env,
10271 const struct bpf_verifier_state *old)
10273 struct bpf_reg_state *state_reg;
10274 struct bpf_func_state *state;
10277 state = old->frame[old->curframe];
10278 state_reg = state->regs;
10279 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
10280 if (state_reg->type != SCALAR_VALUE ||
10281 !state_reg->precise)
10283 if (env->log.level & BPF_LOG_LEVEL2)
10284 verbose(env, "propagating r%d\n", i);
10285 err = mark_chain_precision(env, i);
10290 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
10291 if (state->stack[i].slot_type[0] != STACK_SPILL)
10293 state_reg = &state->stack[i].spilled_ptr;
10294 if (state_reg->type != SCALAR_VALUE ||
10295 !state_reg->precise)
10297 if (env->log.level & BPF_LOG_LEVEL2)
10298 verbose(env, "propagating fp%d\n",
10299 (-i - 1) * BPF_REG_SIZE);
10300 err = mark_chain_precision_stack(env, i);
10307 static bool states_maybe_looping(struct bpf_verifier_state *old,
10308 struct bpf_verifier_state *cur)
10310 struct bpf_func_state *fold, *fcur;
10311 int i, fr = cur->curframe;
10313 if (old->curframe != fr)
10316 fold = old->frame[fr];
10317 fcur = cur->frame[fr];
10318 for (i = 0; i < MAX_BPF_REG; i++)
10319 if (memcmp(&fold->regs[i], &fcur->regs[i],
10320 offsetof(struct bpf_reg_state, parent)))
10326 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
10328 struct bpf_verifier_state_list *new_sl;
10329 struct bpf_verifier_state_list *sl, **pprev;
10330 struct bpf_verifier_state *cur = env->cur_state, *new;
10331 int i, j, err, states_cnt = 0;
10332 bool add_new_state = env->test_state_freq ? true : false;
10334 cur->last_insn_idx = env->prev_insn_idx;
10335 if (!env->insn_aux_data[insn_idx].prune_point)
10336 /* this 'insn_idx' instruction wasn't marked, so we will not
10337 * be doing state search here
10341 /* bpf progs typically have pruning point every 4 instructions
10342 * http://vger.kernel.org/bpfconf2019.html#session-1
10343 * Do not add new state for future pruning if the verifier hasn't seen
10344 * at least 2 jumps and at least 8 instructions.
10345 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
10346 * In tests that amounts to up to 50% reduction into total verifier
10347 * memory consumption and 20% verifier time speedup.
10349 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
10350 env->insn_processed - env->prev_insn_processed >= 8)
10351 add_new_state = true;
10353 pprev = explored_state(env, insn_idx);
10356 clean_live_states(env, insn_idx, cur);
10360 if (sl->state.insn_idx != insn_idx)
10362 if (sl->state.branches) {
10363 if (states_maybe_looping(&sl->state, cur) &&
10364 states_equal(env, &sl->state, cur)) {
10365 verbose_linfo(env, insn_idx, "; ");
10366 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
10369 /* if the verifier is processing a loop, avoid adding new state
10370 * too often, since different loop iterations have distinct
10371 * states and may not help future pruning.
10372 * This threshold shouldn't be too low to make sure that
10373 * a loop with large bound will be rejected quickly.
10374 * The most abusive loop will be:
10376 * if r1 < 1000000 goto pc-2
10377 * 1M insn_procssed limit / 100 == 10k peak states.
10378 * This threshold shouldn't be too high either, since states
10379 * at the end of the loop are likely to be useful in pruning.
10381 if (env->jmps_processed - env->prev_jmps_processed < 20 &&
10382 env->insn_processed - env->prev_insn_processed < 100)
10383 add_new_state = false;
10386 if (states_equal(env, &sl->state, cur)) {
10388 /* reached equivalent register/stack state,
10389 * prune the search.
10390 * Registers read by the continuation are read by us.
10391 * If we have any write marks in env->cur_state, they
10392 * will prevent corresponding reads in the continuation
10393 * from reaching our parent (an explored_state). Our
10394 * own state will get the read marks recorded, but
10395 * they'll be immediately forgotten as we're pruning
10396 * this state and will pop a new one.
10398 err = propagate_liveness(env, &sl->state, cur);
10400 /* if previous state reached the exit with precision and
10401 * current state is equivalent to it (except precsion marks)
10402 * the precision needs to be propagated back in
10403 * the current state.
10405 err = err ? : push_jmp_history(env, cur);
10406 err = err ? : propagate_precision(env, &sl->state);
10412 /* when new state is not going to be added do not increase miss count.
10413 * Otherwise several loop iterations will remove the state
10414 * recorded earlier. The goal of these heuristics is to have
10415 * states from some iterations of the loop (some in the beginning
10416 * and some at the end) to help pruning.
10420 /* heuristic to determine whether this state is beneficial
10421 * to keep checking from state equivalence point of view.
10422 * Higher numbers increase max_states_per_insn and verification time,
10423 * but do not meaningfully decrease insn_processed.
10425 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
10426 /* the state is unlikely to be useful. Remove it to
10427 * speed up verification
10430 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
10431 u32 br = sl->state.branches;
10434 "BUG live_done but branches_to_explore %d\n",
10436 free_verifier_state(&sl->state, false);
10438 env->peak_states--;
10440 /* cannot free this state, since parentage chain may
10441 * walk it later. Add it for free_list instead to
10442 * be freed at the end of verification
10444 sl->next = env->free_list;
10445 env->free_list = sl;
10455 if (env->max_states_per_insn < states_cnt)
10456 env->max_states_per_insn = states_cnt;
10458 if (!env->bpf_capable && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
10459 return push_jmp_history(env, cur);
10461 if (!add_new_state)
10462 return push_jmp_history(env, cur);
10464 /* There were no equivalent states, remember the current one.
10465 * Technically the current state is not proven to be safe yet,
10466 * but it will either reach outer most bpf_exit (which means it's safe)
10467 * or it will be rejected. When there are no loops the verifier won't be
10468 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
10469 * again on the way to bpf_exit.
10470 * When looping the sl->state.branches will be > 0 and this state
10471 * will not be considered for equivalence until branches == 0.
10473 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
10476 env->total_states++;
10477 env->peak_states++;
10478 env->prev_jmps_processed = env->jmps_processed;
10479 env->prev_insn_processed = env->insn_processed;
10481 /* add new state to the head of linked list */
10482 new = &new_sl->state;
10483 err = copy_verifier_state(new, cur);
10485 free_verifier_state(new, false);
10489 new->insn_idx = insn_idx;
10490 WARN_ONCE(new->branches != 1,
10491 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
10494 cur->first_insn_idx = insn_idx;
10495 clear_jmp_history(cur);
10496 new_sl->next = *explored_state(env, insn_idx);
10497 *explored_state(env, insn_idx) = new_sl;
10498 /* connect new state to parentage chain. Current frame needs all
10499 * registers connected. Only r6 - r9 of the callers are alive (pushed
10500 * to the stack implicitly by JITs) so in callers' frames connect just
10501 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
10502 * the state of the call instruction (with WRITTEN set), and r0 comes
10503 * from callee with its full parentage chain, anyway.
10505 /* clear write marks in current state: the writes we did are not writes
10506 * our child did, so they don't screen off its reads from us.
10507 * (There are no read marks in current state, because reads always mark
10508 * their parent and current state never has children yet. Only
10509 * explored_states can get read marks.)
10511 for (j = 0; j <= cur->curframe; j++) {
10512 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
10513 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
10514 for (i = 0; i < BPF_REG_FP; i++)
10515 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
10518 /* all stack frames are accessible from callee, clear them all */
10519 for (j = 0; j <= cur->curframe; j++) {
10520 struct bpf_func_state *frame = cur->frame[j];
10521 struct bpf_func_state *newframe = new->frame[j];
10523 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
10524 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
10525 frame->stack[i].spilled_ptr.parent =
10526 &newframe->stack[i].spilled_ptr;
10532 /* Return true if it's OK to have the same insn return a different type. */
10533 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
10537 case PTR_TO_SOCKET:
10538 case PTR_TO_SOCKET_OR_NULL:
10539 case PTR_TO_SOCK_COMMON:
10540 case PTR_TO_SOCK_COMMON_OR_NULL:
10541 case PTR_TO_TCP_SOCK:
10542 case PTR_TO_TCP_SOCK_OR_NULL:
10543 case PTR_TO_XDP_SOCK:
10544 case PTR_TO_BTF_ID:
10545 case PTR_TO_BTF_ID_OR_NULL:
10552 /* If an instruction was previously used with particular pointer types, then we
10553 * need to be careful to avoid cases such as the below, where it may be ok
10554 * for one branch accessing the pointer, but not ok for the other branch:
10559 * R1 = some_other_valid_ptr;
10562 * R2 = *(u32 *)(R1 + 0);
10564 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
10566 return src != prev && (!reg_type_mismatch_ok(src) ||
10567 !reg_type_mismatch_ok(prev));
10570 static int do_check(struct bpf_verifier_env *env)
10572 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
10573 struct bpf_verifier_state *state = env->cur_state;
10574 struct bpf_insn *insns = env->prog->insnsi;
10575 struct bpf_reg_state *regs;
10576 int insn_cnt = env->prog->len;
10577 bool do_print_state = false;
10578 int prev_insn_idx = -1;
10581 struct bpf_insn *insn;
10585 env->prev_insn_idx = prev_insn_idx;
10586 if (env->insn_idx >= insn_cnt) {
10587 verbose(env, "invalid insn idx %d insn_cnt %d\n",
10588 env->insn_idx, insn_cnt);
10592 insn = &insns[env->insn_idx];
10593 class = BPF_CLASS(insn->code);
10595 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
10597 "BPF program is too large. Processed %d insn\n",
10598 env->insn_processed);
10602 err = is_state_visited(env, env->insn_idx);
10606 /* found equivalent state, can prune the search */
10607 if (env->log.level & BPF_LOG_LEVEL) {
10608 if (do_print_state)
10609 verbose(env, "\nfrom %d to %d%s: safe\n",
10610 env->prev_insn_idx, env->insn_idx,
10611 env->cur_state->speculative ?
10612 " (speculative execution)" : "");
10614 verbose(env, "%d: safe\n", env->insn_idx);
10616 goto process_bpf_exit;
10619 if (signal_pending(current))
10622 if (need_resched())
10625 if (env->log.level & BPF_LOG_LEVEL2 ||
10626 (env->log.level & BPF_LOG_LEVEL && do_print_state)) {
10627 if (env->log.level & BPF_LOG_LEVEL2)
10628 verbose(env, "%d:", env->insn_idx);
10630 verbose(env, "\nfrom %d to %d%s:",
10631 env->prev_insn_idx, env->insn_idx,
10632 env->cur_state->speculative ?
10633 " (speculative execution)" : "");
10634 print_verifier_state(env, state->frame[state->curframe]);
10635 do_print_state = false;
10638 if (env->log.level & BPF_LOG_LEVEL) {
10639 const struct bpf_insn_cbs cbs = {
10640 .cb_call = disasm_kfunc_name,
10641 .cb_print = verbose,
10642 .private_data = env,
10645 verbose_linfo(env, env->insn_idx, "; ");
10646 verbose(env, "%d: ", env->insn_idx);
10647 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
10650 if (bpf_prog_is_dev_bound(env->prog->aux)) {
10651 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
10652 env->prev_insn_idx);
10657 regs = cur_regs(env);
10658 sanitize_mark_insn_seen(env);
10659 prev_insn_idx = env->insn_idx;
10661 if (class == BPF_ALU || class == BPF_ALU64) {
10662 err = check_alu_op(env, insn);
10666 } else if (class == BPF_LDX) {
10667 enum bpf_reg_type *prev_src_type, src_reg_type;
10669 /* check for reserved fields is already done */
10671 /* check src operand */
10672 err = check_reg_arg(env, insn->src_reg, SRC_OP);
10676 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
10680 src_reg_type = regs[insn->src_reg].type;
10682 /* check that memory (src_reg + off) is readable,
10683 * the state of dst_reg will be updated by this func
10685 err = check_mem_access(env, env->insn_idx, insn->src_reg,
10686 insn->off, BPF_SIZE(insn->code),
10687 BPF_READ, insn->dst_reg, false);
10691 prev_src_type = &env->insn_aux_data[env->insn_idx].ptr_type;
10693 if (*prev_src_type == NOT_INIT) {
10694 /* saw a valid insn
10695 * dst_reg = *(u32 *)(src_reg + off)
10696 * save type to validate intersecting paths
10698 *prev_src_type = src_reg_type;
10700 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
10701 /* ABuser program is trying to use the same insn
10702 * dst_reg = *(u32*) (src_reg + off)
10703 * with different pointer types:
10704 * src_reg == ctx in one branch and
10705 * src_reg == stack|map in some other branch.
10708 verbose(env, "same insn cannot be used with different pointers\n");
10712 } else if (class == BPF_STX) {
10713 enum bpf_reg_type *prev_dst_type, dst_reg_type;
10715 if (BPF_MODE(insn->code) == BPF_ATOMIC) {
10716 err = check_atomic(env, env->insn_idx, insn);
10723 if (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0) {
10724 verbose(env, "BPF_STX uses reserved fields\n");
10728 /* check src1 operand */
10729 err = check_reg_arg(env, insn->src_reg, SRC_OP);
10732 /* check src2 operand */
10733 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
10737 dst_reg_type = regs[insn->dst_reg].type;
10739 /* check that memory (dst_reg + off) is writeable */
10740 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
10741 insn->off, BPF_SIZE(insn->code),
10742 BPF_WRITE, insn->src_reg, false);
10746 prev_dst_type = &env->insn_aux_data[env->insn_idx].ptr_type;
10748 if (*prev_dst_type == NOT_INIT) {
10749 *prev_dst_type = dst_reg_type;
10750 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
10751 verbose(env, "same insn cannot be used with different pointers\n");
10755 } else if (class == BPF_ST) {
10756 if (BPF_MODE(insn->code) != BPF_MEM ||
10757 insn->src_reg != BPF_REG_0) {
10758 verbose(env, "BPF_ST uses reserved fields\n");
10761 /* check src operand */
10762 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
10766 if (is_ctx_reg(env, insn->dst_reg)) {
10767 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
10769 reg_type_str[reg_state(env, insn->dst_reg)->type]);
10773 /* check that memory (dst_reg + off) is writeable */
10774 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
10775 insn->off, BPF_SIZE(insn->code),
10776 BPF_WRITE, -1, false);
10780 } else if (class == BPF_JMP || class == BPF_JMP32) {
10781 u8 opcode = BPF_OP(insn->code);
10783 env->jmps_processed++;
10784 if (opcode == BPF_CALL) {
10785 if (BPF_SRC(insn->code) != BPF_K ||
10787 (insn->src_reg != BPF_REG_0 &&
10788 insn->src_reg != BPF_PSEUDO_CALL &&
10789 insn->src_reg != BPF_PSEUDO_KFUNC_CALL) ||
10790 insn->dst_reg != BPF_REG_0 ||
10791 class == BPF_JMP32) {
10792 verbose(env, "BPF_CALL uses reserved fields\n");
10796 if (env->cur_state->active_spin_lock &&
10797 (insn->src_reg == BPF_PSEUDO_CALL ||
10798 insn->imm != BPF_FUNC_spin_unlock)) {
10799 verbose(env, "function calls are not allowed while holding a lock\n");
10802 if (insn->src_reg == BPF_PSEUDO_CALL)
10803 err = check_func_call(env, insn, &env->insn_idx);
10804 else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL)
10805 err = check_kfunc_call(env, insn);
10807 err = check_helper_call(env, insn, &env->insn_idx);
10810 } else if (opcode == BPF_JA) {
10811 if (BPF_SRC(insn->code) != BPF_K ||
10813 insn->src_reg != BPF_REG_0 ||
10814 insn->dst_reg != BPF_REG_0 ||
10815 class == BPF_JMP32) {
10816 verbose(env, "BPF_JA uses reserved fields\n");
10820 env->insn_idx += insn->off + 1;
10823 } else if (opcode == BPF_EXIT) {
10824 if (BPF_SRC(insn->code) != BPF_K ||
10826 insn->src_reg != BPF_REG_0 ||
10827 insn->dst_reg != BPF_REG_0 ||
10828 class == BPF_JMP32) {
10829 verbose(env, "BPF_EXIT uses reserved fields\n");
10833 if (env->cur_state->active_spin_lock) {
10834 verbose(env, "bpf_spin_unlock is missing\n");
10838 if (state->curframe) {
10839 /* exit from nested function */
10840 err = prepare_func_exit(env, &env->insn_idx);
10843 do_print_state = true;
10847 err = check_reference_leak(env);
10851 err = check_return_code(env);
10855 update_branch_counts(env, env->cur_state);
10856 err = pop_stack(env, &prev_insn_idx,
10857 &env->insn_idx, pop_log);
10859 if (err != -ENOENT)
10863 do_print_state = true;
10867 err = check_cond_jmp_op(env, insn, &env->insn_idx);
10871 } else if (class == BPF_LD) {
10872 u8 mode = BPF_MODE(insn->code);
10874 if (mode == BPF_ABS || mode == BPF_IND) {
10875 err = check_ld_abs(env, insn);
10879 } else if (mode == BPF_IMM) {
10880 err = check_ld_imm(env, insn);
10885 sanitize_mark_insn_seen(env);
10887 verbose(env, "invalid BPF_LD mode\n");
10891 verbose(env, "unknown insn class %d\n", class);
10901 static int find_btf_percpu_datasec(struct btf *btf)
10903 const struct btf_type *t;
10908 * Both vmlinux and module each have their own ".data..percpu"
10909 * DATASECs in BTF. So for module's case, we need to skip vmlinux BTF
10910 * types to look at only module's own BTF types.
10912 n = btf_nr_types(btf);
10913 if (btf_is_module(btf))
10914 i = btf_nr_types(btf_vmlinux);
10918 for(; i < n; i++) {
10919 t = btf_type_by_id(btf, i);
10920 if (BTF_INFO_KIND(t->info) != BTF_KIND_DATASEC)
10923 tname = btf_name_by_offset(btf, t->name_off);
10924 if (!strcmp(tname, ".data..percpu"))
10931 /* replace pseudo btf_id with kernel symbol address */
10932 static int check_pseudo_btf_id(struct bpf_verifier_env *env,
10933 struct bpf_insn *insn,
10934 struct bpf_insn_aux_data *aux)
10936 const struct btf_var_secinfo *vsi;
10937 const struct btf_type *datasec;
10938 struct btf_mod_pair *btf_mod;
10939 const struct btf_type *t;
10940 const char *sym_name;
10941 bool percpu = false;
10942 u32 type, id = insn->imm;
10946 int i, btf_fd, err;
10948 btf_fd = insn[1].imm;
10950 btf = btf_get_by_fd(btf_fd);
10952 verbose(env, "invalid module BTF object FD specified.\n");
10956 if (!btf_vmlinux) {
10957 verbose(env, "kernel is missing BTF, make sure CONFIG_DEBUG_INFO_BTF=y is specified in Kconfig.\n");
10964 t = btf_type_by_id(btf, id);
10966 verbose(env, "ldimm64 insn specifies invalid btf_id %d.\n", id);
10971 if (!btf_type_is_var(t)) {
10972 verbose(env, "pseudo btf_id %d in ldimm64 isn't KIND_VAR.\n", id);
10977 sym_name = btf_name_by_offset(btf, t->name_off);
10978 addr = kallsyms_lookup_name(sym_name);
10980 verbose(env, "ldimm64 failed to find the address for kernel symbol '%s'.\n",
10986 datasec_id = find_btf_percpu_datasec(btf);
10987 if (datasec_id > 0) {
10988 datasec = btf_type_by_id(btf, datasec_id);
10989 for_each_vsi(i, datasec, vsi) {
10990 if (vsi->type == id) {
10997 insn[0].imm = (u32)addr;
10998 insn[1].imm = addr >> 32;
11001 t = btf_type_skip_modifiers(btf, type, NULL);
11003 aux->btf_var.reg_type = PTR_TO_PERCPU_BTF_ID;
11004 aux->btf_var.btf = btf;
11005 aux->btf_var.btf_id = type;
11006 } else if (!btf_type_is_struct(t)) {
11007 const struct btf_type *ret;
11011 /* resolve the type size of ksym. */
11012 ret = btf_resolve_size(btf, t, &tsize);
11014 tname = btf_name_by_offset(btf, t->name_off);
11015 verbose(env, "ldimm64 unable to resolve the size of type '%s': %ld\n",
11016 tname, PTR_ERR(ret));
11020 aux->btf_var.reg_type = PTR_TO_MEM;
11021 aux->btf_var.mem_size = tsize;
11023 aux->btf_var.reg_type = PTR_TO_BTF_ID;
11024 aux->btf_var.btf = btf;
11025 aux->btf_var.btf_id = type;
11028 /* check whether we recorded this BTF (and maybe module) already */
11029 for (i = 0; i < env->used_btf_cnt; i++) {
11030 if (env->used_btfs[i].btf == btf) {
11036 if (env->used_btf_cnt >= MAX_USED_BTFS) {
11041 btf_mod = &env->used_btfs[env->used_btf_cnt];
11042 btf_mod->btf = btf;
11043 btf_mod->module = NULL;
11045 /* if we reference variables from kernel module, bump its refcount */
11046 if (btf_is_module(btf)) {
11047 btf_mod->module = btf_try_get_module(btf);
11048 if (!btf_mod->module) {
11054 env->used_btf_cnt++;
11062 static int check_map_prealloc(struct bpf_map *map)
11064 return (map->map_type != BPF_MAP_TYPE_HASH &&
11065 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
11066 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
11067 !(map->map_flags & BPF_F_NO_PREALLOC);
11070 static bool is_tracing_prog_type(enum bpf_prog_type type)
11073 case BPF_PROG_TYPE_KPROBE:
11074 case BPF_PROG_TYPE_TRACEPOINT:
11075 case BPF_PROG_TYPE_PERF_EVENT:
11076 case BPF_PROG_TYPE_RAW_TRACEPOINT:
11083 static bool is_preallocated_map(struct bpf_map *map)
11085 if (!check_map_prealloc(map))
11087 if (map->inner_map_meta && !check_map_prealloc(map->inner_map_meta))
11092 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
11093 struct bpf_map *map,
11094 struct bpf_prog *prog)
11097 enum bpf_prog_type prog_type = resolve_prog_type(prog);
11099 * Validate that trace type programs use preallocated hash maps.
11101 * For programs attached to PERF events this is mandatory as the
11102 * perf NMI can hit any arbitrary code sequence.
11104 * All other trace types using preallocated hash maps are unsafe as
11105 * well because tracepoint or kprobes can be inside locked regions
11106 * of the memory allocator or at a place where a recursion into the
11107 * memory allocator would see inconsistent state.
11109 * On RT enabled kernels run-time allocation of all trace type
11110 * programs is strictly prohibited due to lock type constraints. On
11111 * !RT kernels it is allowed for backwards compatibility reasons for
11112 * now, but warnings are emitted so developers are made aware of
11113 * the unsafety and can fix their programs before this is enforced.
11115 if (is_tracing_prog_type(prog_type) && !is_preallocated_map(map)) {
11116 if (prog_type == BPF_PROG_TYPE_PERF_EVENT) {
11117 verbose(env, "perf_event programs can only use preallocated hash map\n");
11120 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
11121 verbose(env, "trace type programs can only use preallocated hash map\n");
11124 WARN_ONCE(1, "trace type BPF program uses run-time allocation\n");
11125 verbose(env, "trace type programs with run-time allocated hash maps are unsafe. Switch to preallocated hash maps.\n");
11128 if (map_value_has_spin_lock(map)) {
11129 if (prog_type == BPF_PROG_TYPE_SOCKET_FILTER) {
11130 verbose(env, "socket filter progs cannot use bpf_spin_lock yet\n");
11134 if (is_tracing_prog_type(prog_type)) {
11135 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
11139 if (prog->aux->sleepable) {
11140 verbose(env, "sleepable progs cannot use bpf_spin_lock yet\n");
11145 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
11146 !bpf_offload_prog_map_match(prog, map)) {
11147 verbose(env, "offload device mismatch between prog and map\n");
11151 if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
11152 verbose(env, "bpf_struct_ops map cannot be used in prog\n");
11156 if (prog->aux->sleepable)
11157 switch (map->map_type) {
11158 case BPF_MAP_TYPE_HASH:
11159 case BPF_MAP_TYPE_LRU_HASH:
11160 case BPF_MAP_TYPE_ARRAY:
11161 case BPF_MAP_TYPE_PERCPU_HASH:
11162 case BPF_MAP_TYPE_PERCPU_ARRAY:
11163 case BPF_MAP_TYPE_LRU_PERCPU_HASH:
11164 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
11165 case BPF_MAP_TYPE_HASH_OF_MAPS:
11166 if (!is_preallocated_map(map)) {
11168 "Sleepable programs can only use preallocated maps\n");
11172 case BPF_MAP_TYPE_RINGBUF:
11176 "Sleepable programs can only use array, hash, and ringbuf maps\n");
11183 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
11185 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
11186 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
11189 /* find and rewrite pseudo imm in ld_imm64 instructions:
11191 * 1. if it accesses map FD, replace it with actual map pointer.
11192 * 2. if it accesses btf_id of a VAR, replace it with pointer to the var.
11194 * NOTE: btf_vmlinux is required for converting pseudo btf_id.
11196 static int resolve_pseudo_ldimm64(struct bpf_verifier_env *env)
11198 struct bpf_insn *insn = env->prog->insnsi;
11199 int insn_cnt = env->prog->len;
11202 err = bpf_prog_calc_tag(env->prog);
11206 for (i = 0; i < insn_cnt; i++, insn++) {
11207 if (BPF_CLASS(insn->code) == BPF_LDX &&
11208 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
11209 verbose(env, "BPF_LDX uses reserved fields\n");
11213 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
11214 struct bpf_insn_aux_data *aux;
11215 struct bpf_map *map;
11220 if (i == insn_cnt - 1 || insn[1].code != 0 ||
11221 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
11222 insn[1].off != 0) {
11223 verbose(env, "invalid bpf_ld_imm64 insn\n");
11227 if (insn[0].src_reg == 0)
11228 /* valid generic load 64-bit imm */
11231 if (insn[0].src_reg == BPF_PSEUDO_BTF_ID) {
11232 aux = &env->insn_aux_data[i];
11233 err = check_pseudo_btf_id(env, insn, aux);
11239 if (insn[0].src_reg == BPF_PSEUDO_FUNC) {
11240 aux = &env->insn_aux_data[i];
11241 aux->ptr_type = PTR_TO_FUNC;
11245 /* In final convert_pseudo_ld_imm64() step, this is
11246 * converted into regular 64-bit imm load insn.
11248 switch (insn[0].src_reg) {
11249 case BPF_PSEUDO_MAP_VALUE:
11250 case BPF_PSEUDO_MAP_IDX_VALUE:
11252 case BPF_PSEUDO_MAP_FD:
11253 case BPF_PSEUDO_MAP_IDX:
11254 if (insn[1].imm == 0)
11258 verbose(env, "unrecognized bpf_ld_imm64 insn\n");
11262 switch (insn[0].src_reg) {
11263 case BPF_PSEUDO_MAP_IDX_VALUE:
11264 case BPF_PSEUDO_MAP_IDX:
11265 if (bpfptr_is_null(env->fd_array)) {
11266 verbose(env, "fd_idx without fd_array is invalid\n");
11269 if (copy_from_bpfptr_offset(&fd, env->fd_array,
11270 insn[0].imm * sizeof(fd),
11280 map = __bpf_map_get(f);
11282 verbose(env, "fd %d is not pointing to valid bpf_map\n",
11284 return PTR_ERR(map);
11287 err = check_map_prog_compatibility(env, map, env->prog);
11293 aux = &env->insn_aux_data[i];
11294 if (insn[0].src_reg == BPF_PSEUDO_MAP_FD ||
11295 insn[0].src_reg == BPF_PSEUDO_MAP_IDX) {
11296 addr = (unsigned long)map;
11298 u32 off = insn[1].imm;
11300 if (off >= BPF_MAX_VAR_OFF) {
11301 verbose(env, "direct value offset of %u is not allowed\n", off);
11306 if (!map->ops->map_direct_value_addr) {
11307 verbose(env, "no direct value access support for this map type\n");
11312 err = map->ops->map_direct_value_addr(map, &addr, off);
11314 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
11315 map->value_size, off);
11320 aux->map_off = off;
11324 insn[0].imm = (u32)addr;
11325 insn[1].imm = addr >> 32;
11327 /* check whether we recorded this map already */
11328 for (j = 0; j < env->used_map_cnt; j++) {
11329 if (env->used_maps[j] == map) {
11330 aux->map_index = j;
11336 if (env->used_map_cnt >= MAX_USED_MAPS) {
11341 /* hold the map. If the program is rejected by verifier,
11342 * the map will be released by release_maps() or it
11343 * will be used by the valid program until it's unloaded
11344 * and all maps are released in free_used_maps()
11348 aux->map_index = env->used_map_cnt;
11349 env->used_maps[env->used_map_cnt++] = map;
11351 if (bpf_map_is_cgroup_storage(map) &&
11352 bpf_cgroup_storage_assign(env->prog->aux, map)) {
11353 verbose(env, "only one cgroup storage of each type is allowed\n");
11365 /* Basic sanity check before we invest more work here. */
11366 if (!bpf_opcode_in_insntable(insn->code)) {
11367 verbose(env, "unknown opcode %02x\n", insn->code);
11372 /* now all pseudo BPF_LD_IMM64 instructions load valid
11373 * 'struct bpf_map *' into a register instead of user map_fd.
11374 * These pointers will be used later by verifier to validate map access.
11379 /* drop refcnt of maps used by the rejected program */
11380 static void release_maps(struct bpf_verifier_env *env)
11382 __bpf_free_used_maps(env->prog->aux, env->used_maps,
11383 env->used_map_cnt);
11386 /* drop refcnt of maps used by the rejected program */
11387 static void release_btfs(struct bpf_verifier_env *env)
11389 __bpf_free_used_btfs(env->prog->aux, env->used_btfs,
11390 env->used_btf_cnt);
11393 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
11394 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
11396 struct bpf_insn *insn = env->prog->insnsi;
11397 int insn_cnt = env->prog->len;
11400 for (i = 0; i < insn_cnt; i++, insn++) {
11401 if (insn->code != (BPF_LD | BPF_IMM | BPF_DW))
11403 if (insn->src_reg == BPF_PSEUDO_FUNC)
11409 /* single env->prog->insni[off] instruction was replaced with the range
11410 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
11411 * [0, off) and [off, end) to new locations, so the patched range stays zero
11413 static int adjust_insn_aux_data(struct bpf_verifier_env *env,
11414 struct bpf_prog *new_prog, u32 off, u32 cnt)
11416 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
11417 struct bpf_insn *insn = new_prog->insnsi;
11418 u32 old_seen = old_data[off].seen;
11422 /* aux info at OFF always needs adjustment, no matter fast path
11423 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
11424 * original insn at old prog.
11426 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
11430 prog_len = new_prog->len;
11431 new_data = vzalloc(array_size(prog_len,
11432 sizeof(struct bpf_insn_aux_data)));
11435 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
11436 memcpy(new_data + off + cnt - 1, old_data + off,
11437 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
11438 for (i = off; i < off + cnt - 1; i++) {
11439 /* Expand insni[off]'s seen count to the patched range. */
11440 new_data[i].seen = old_seen;
11441 new_data[i].zext_dst = insn_has_def32(env, insn + i);
11443 env->insn_aux_data = new_data;
11448 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
11454 /* NOTE: fake 'exit' subprog should be updated as well. */
11455 for (i = 0; i <= env->subprog_cnt; i++) {
11456 if (env->subprog_info[i].start <= off)
11458 env->subprog_info[i].start += len - 1;
11462 static void adjust_poke_descs(struct bpf_prog *prog, u32 off, u32 len)
11464 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
11465 int i, sz = prog->aux->size_poke_tab;
11466 struct bpf_jit_poke_descriptor *desc;
11468 for (i = 0; i < sz; i++) {
11470 if (desc->insn_idx <= off)
11472 desc->insn_idx += len - 1;
11476 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
11477 const struct bpf_insn *patch, u32 len)
11479 struct bpf_prog *new_prog;
11481 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
11482 if (IS_ERR(new_prog)) {
11483 if (PTR_ERR(new_prog) == -ERANGE)
11485 "insn %d cannot be patched due to 16-bit range\n",
11486 env->insn_aux_data[off].orig_idx);
11489 if (adjust_insn_aux_data(env, new_prog, off, len))
11491 adjust_subprog_starts(env, off, len);
11492 adjust_poke_descs(new_prog, off, len);
11496 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
11501 /* find first prog starting at or after off (first to remove) */
11502 for (i = 0; i < env->subprog_cnt; i++)
11503 if (env->subprog_info[i].start >= off)
11505 /* find first prog starting at or after off + cnt (first to stay) */
11506 for (j = i; j < env->subprog_cnt; j++)
11507 if (env->subprog_info[j].start >= off + cnt)
11509 /* if j doesn't start exactly at off + cnt, we are just removing
11510 * the front of previous prog
11512 if (env->subprog_info[j].start != off + cnt)
11516 struct bpf_prog_aux *aux = env->prog->aux;
11519 /* move fake 'exit' subprog as well */
11520 move = env->subprog_cnt + 1 - j;
11522 memmove(env->subprog_info + i,
11523 env->subprog_info + j,
11524 sizeof(*env->subprog_info) * move);
11525 env->subprog_cnt -= j - i;
11527 /* remove func_info */
11528 if (aux->func_info) {
11529 move = aux->func_info_cnt - j;
11531 memmove(aux->func_info + i,
11532 aux->func_info + j,
11533 sizeof(*aux->func_info) * move);
11534 aux->func_info_cnt -= j - i;
11535 /* func_info->insn_off is set after all code rewrites,
11536 * in adjust_btf_func() - no need to adjust
11540 /* convert i from "first prog to remove" to "first to adjust" */
11541 if (env->subprog_info[i].start == off)
11545 /* update fake 'exit' subprog as well */
11546 for (; i <= env->subprog_cnt; i++)
11547 env->subprog_info[i].start -= cnt;
11552 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
11555 struct bpf_prog *prog = env->prog;
11556 u32 i, l_off, l_cnt, nr_linfo;
11557 struct bpf_line_info *linfo;
11559 nr_linfo = prog->aux->nr_linfo;
11563 linfo = prog->aux->linfo;
11565 /* find first line info to remove, count lines to be removed */
11566 for (i = 0; i < nr_linfo; i++)
11567 if (linfo[i].insn_off >= off)
11572 for (; i < nr_linfo; i++)
11573 if (linfo[i].insn_off < off + cnt)
11578 /* First live insn doesn't match first live linfo, it needs to "inherit"
11579 * last removed linfo. prog is already modified, so prog->len == off
11580 * means no live instructions after (tail of the program was removed).
11582 if (prog->len != off && l_cnt &&
11583 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
11585 linfo[--i].insn_off = off + cnt;
11588 /* remove the line info which refer to the removed instructions */
11590 memmove(linfo + l_off, linfo + i,
11591 sizeof(*linfo) * (nr_linfo - i));
11593 prog->aux->nr_linfo -= l_cnt;
11594 nr_linfo = prog->aux->nr_linfo;
11597 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
11598 for (i = l_off; i < nr_linfo; i++)
11599 linfo[i].insn_off -= cnt;
11601 /* fix up all subprogs (incl. 'exit') which start >= off */
11602 for (i = 0; i <= env->subprog_cnt; i++)
11603 if (env->subprog_info[i].linfo_idx > l_off) {
11604 /* program may have started in the removed region but
11605 * may not be fully removed
11607 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
11608 env->subprog_info[i].linfo_idx -= l_cnt;
11610 env->subprog_info[i].linfo_idx = l_off;
11616 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
11618 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
11619 unsigned int orig_prog_len = env->prog->len;
11622 if (bpf_prog_is_dev_bound(env->prog->aux))
11623 bpf_prog_offload_remove_insns(env, off, cnt);
11625 err = bpf_remove_insns(env->prog, off, cnt);
11629 err = adjust_subprog_starts_after_remove(env, off, cnt);
11633 err = bpf_adj_linfo_after_remove(env, off, cnt);
11637 memmove(aux_data + off, aux_data + off + cnt,
11638 sizeof(*aux_data) * (orig_prog_len - off - cnt));
11643 /* The verifier does more data flow analysis than llvm and will not
11644 * explore branches that are dead at run time. Malicious programs can
11645 * have dead code too. Therefore replace all dead at-run-time code
11648 * Just nops are not optimal, e.g. if they would sit at the end of the
11649 * program and through another bug we would manage to jump there, then
11650 * we'd execute beyond program memory otherwise. Returning exception
11651 * code also wouldn't work since we can have subprogs where the dead
11652 * code could be located.
11654 static void sanitize_dead_code(struct bpf_verifier_env *env)
11656 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
11657 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
11658 struct bpf_insn *insn = env->prog->insnsi;
11659 const int insn_cnt = env->prog->len;
11662 for (i = 0; i < insn_cnt; i++) {
11663 if (aux_data[i].seen)
11665 memcpy(insn + i, &trap, sizeof(trap));
11666 aux_data[i].zext_dst = false;
11670 static bool insn_is_cond_jump(u8 code)
11674 if (BPF_CLASS(code) == BPF_JMP32)
11677 if (BPF_CLASS(code) != BPF_JMP)
11681 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
11684 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
11686 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
11687 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
11688 struct bpf_insn *insn = env->prog->insnsi;
11689 const int insn_cnt = env->prog->len;
11692 for (i = 0; i < insn_cnt; i++, insn++) {
11693 if (!insn_is_cond_jump(insn->code))
11696 if (!aux_data[i + 1].seen)
11697 ja.off = insn->off;
11698 else if (!aux_data[i + 1 + insn->off].seen)
11703 if (bpf_prog_is_dev_bound(env->prog->aux))
11704 bpf_prog_offload_replace_insn(env, i, &ja);
11706 memcpy(insn, &ja, sizeof(ja));
11710 static int opt_remove_dead_code(struct bpf_verifier_env *env)
11712 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
11713 int insn_cnt = env->prog->len;
11716 for (i = 0; i < insn_cnt; i++) {
11720 while (i + j < insn_cnt && !aux_data[i + j].seen)
11725 err = verifier_remove_insns(env, i, j);
11728 insn_cnt = env->prog->len;
11734 static int opt_remove_nops(struct bpf_verifier_env *env)
11736 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
11737 struct bpf_insn *insn = env->prog->insnsi;
11738 int insn_cnt = env->prog->len;
11741 for (i = 0; i < insn_cnt; i++) {
11742 if (memcmp(&insn[i], &ja, sizeof(ja)))
11745 err = verifier_remove_insns(env, i, 1);
11755 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
11756 const union bpf_attr *attr)
11758 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
11759 struct bpf_insn_aux_data *aux = env->insn_aux_data;
11760 int i, patch_len, delta = 0, len = env->prog->len;
11761 struct bpf_insn *insns = env->prog->insnsi;
11762 struct bpf_prog *new_prog;
11765 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
11766 zext_patch[1] = BPF_ZEXT_REG(0);
11767 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
11768 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
11769 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
11770 for (i = 0; i < len; i++) {
11771 int adj_idx = i + delta;
11772 struct bpf_insn insn;
11775 insn = insns[adj_idx];
11776 load_reg = insn_def_regno(&insn);
11777 if (!aux[adj_idx].zext_dst) {
11785 class = BPF_CLASS(code);
11786 if (load_reg == -1)
11789 /* NOTE: arg "reg" (the fourth one) is only used for
11790 * BPF_STX + SRC_OP, so it is safe to pass NULL
11793 if (is_reg64(env, &insn, load_reg, NULL, DST_OP)) {
11794 if (class == BPF_LD &&
11795 BPF_MODE(code) == BPF_IMM)
11800 /* ctx load could be transformed into wider load. */
11801 if (class == BPF_LDX &&
11802 aux[adj_idx].ptr_type == PTR_TO_CTX)
11805 imm_rnd = get_random_int();
11806 rnd_hi32_patch[0] = insn;
11807 rnd_hi32_patch[1].imm = imm_rnd;
11808 rnd_hi32_patch[3].dst_reg = load_reg;
11809 patch = rnd_hi32_patch;
11811 goto apply_patch_buffer;
11814 /* Add in an zero-extend instruction if a) the JIT has requested
11815 * it or b) it's a CMPXCHG.
11817 * The latter is because: BPF_CMPXCHG always loads a value into
11818 * R0, therefore always zero-extends. However some archs'
11819 * equivalent instruction only does this load when the
11820 * comparison is successful. This detail of CMPXCHG is
11821 * orthogonal to the general zero-extension behaviour of the
11822 * CPU, so it's treated independently of bpf_jit_needs_zext.
11824 if (!bpf_jit_needs_zext() && !is_cmpxchg_insn(&insn))
11827 if (WARN_ON(load_reg == -1)) {
11828 verbose(env, "verifier bug. zext_dst is set, but no reg is defined\n");
11832 zext_patch[0] = insn;
11833 zext_patch[1].dst_reg = load_reg;
11834 zext_patch[1].src_reg = load_reg;
11835 patch = zext_patch;
11837 apply_patch_buffer:
11838 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
11841 env->prog = new_prog;
11842 insns = new_prog->insnsi;
11843 aux = env->insn_aux_data;
11844 delta += patch_len - 1;
11850 /* convert load instructions that access fields of a context type into a
11851 * sequence of instructions that access fields of the underlying structure:
11852 * struct __sk_buff -> struct sk_buff
11853 * struct bpf_sock_ops -> struct sock
11855 static int convert_ctx_accesses(struct bpf_verifier_env *env)
11857 const struct bpf_verifier_ops *ops = env->ops;
11858 int i, cnt, size, ctx_field_size, delta = 0;
11859 const int insn_cnt = env->prog->len;
11860 struct bpf_insn insn_buf[16], *insn;
11861 u32 target_size, size_default, off;
11862 struct bpf_prog *new_prog;
11863 enum bpf_access_type type;
11864 bool is_narrower_load;
11866 if (ops->gen_prologue || env->seen_direct_write) {
11867 if (!ops->gen_prologue) {
11868 verbose(env, "bpf verifier is misconfigured\n");
11871 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
11873 if (cnt >= ARRAY_SIZE(insn_buf)) {
11874 verbose(env, "bpf verifier is misconfigured\n");
11877 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
11881 env->prog = new_prog;
11886 if (bpf_prog_is_dev_bound(env->prog->aux))
11889 insn = env->prog->insnsi + delta;
11891 for (i = 0; i < insn_cnt; i++, insn++) {
11892 bpf_convert_ctx_access_t convert_ctx_access;
11895 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
11896 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
11897 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
11898 insn->code == (BPF_LDX | BPF_MEM | BPF_DW)) {
11901 } else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
11902 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
11903 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
11904 insn->code == (BPF_STX | BPF_MEM | BPF_DW) ||
11905 insn->code == (BPF_ST | BPF_MEM | BPF_B) ||
11906 insn->code == (BPF_ST | BPF_MEM | BPF_H) ||
11907 insn->code == (BPF_ST | BPF_MEM | BPF_W) ||
11908 insn->code == (BPF_ST | BPF_MEM | BPF_DW)) {
11910 ctx_access = BPF_CLASS(insn->code) == BPF_STX;
11915 if (type == BPF_WRITE &&
11916 env->insn_aux_data[i + delta].sanitize_stack_spill) {
11917 struct bpf_insn patch[] = {
11922 cnt = ARRAY_SIZE(patch);
11923 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
11928 env->prog = new_prog;
11929 insn = new_prog->insnsi + i + delta;
11936 switch (env->insn_aux_data[i + delta].ptr_type) {
11938 if (!ops->convert_ctx_access)
11940 convert_ctx_access = ops->convert_ctx_access;
11942 case PTR_TO_SOCKET:
11943 case PTR_TO_SOCK_COMMON:
11944 convert_ctx_access = bpf_sock_convert_ctx_access;
11946 case PTR_TO_TCP_SOCK:
11947 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
11949 case PTR_TO_XDP_SOCK:
11950 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
11952 case PTR_TO_BTF_ID:
11953 if (type == BPF_READ) {
11954 insn->code = BPF_LDX | BPF_PROBE_MEM |
11955 BPF_SIZE((insn)->code);
11956 env->prog->aux->num_exentries++;
11957 } else if (resolve_prog_type(env->prog) != BPF_PROG_TYPE_STRUCT_OPS) {
11958 verbose(env, "Writes through BTF pointers are not allowed\n");
11966 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
11967 size = BPF_LDST_BYTES(insn);
11969 /* If the read access is a narrower load of the field,
11970 * convert to a 4/8-byte load, to minimum program type specific
11971 * convert_ctx_access changes. If conversion is successful,
11972 * we will apply proper mask to the result.
11974 is_narrower_load = size < ctx_field_size;
11975 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
11977 if (is_narrower_load) {
11980 if (type == BPF_WRITE) {
11981 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
11986 if (ctx_field_size == 4)
11988 else if (ctx_field_size == 8)
11989 size_code = BPF_DW;
11991 insn->off = off & ~(size_default - 1);
11992 insn->code = BPF_LDX | BPF_MEM | size_code;
11996 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
11998 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
11999 (ctx_field_size && !target_size)) {
12000 verbose(env, "bpf verifier is misconfigured\n");
12004 if (is_narrower_load && size < target_size) {
12005 u8 shift = bpf_ctx_narrow_access_offset(
12006 off, size, size_default) * 8;
12007 if (ctx_field_size <= 4) {
12009 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
12012 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
12013 (1 << size * 8) - 1);
12016 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
12019 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
12020 (1ULL << size * 8) - 1);
12024 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12030 /* keep walking new program and skip insns we just inserted */
12031 env->prog = new_prog;
12032 insn = new_prog->insnsi + i + delta;
12038 static int jit_subprogs(struct bpf_verifier_env *env)
12040 struct bpf_prog *prog = env->prog, **func, *tmp;
12041 int i, j, subprog_start, subprog_end = 0, len, subprog;
12042 struct bpf_map *map_ptr;
12043 struct bpf_insn *insn;
12044 void *old_bpf_func;
12045 int err, num_exentries;
12047 if (env->subprog_cnt <= 1)
12050 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
12051 if (bpf_pseudo_func(insn)) {
12052 env->insn_aux_data[i].call_imm = insn->imm;
12053 /* subprog is encoded in insn[1].imm */
12057 if (!bpf_pseudo_call(insn))
12059 /* Upon error here we cannot fall back to interpreter but
12060 * need a hard reject of the program. Thus -EFAULT is
12061 * propagated in any case.
12063 subprog = find_subprog(env, i + insn->imm + 1);
12065 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
12066 i + insn->imm + 1);
12069 /* temporarily remember subprog id inside insn instead of
12070 * aux_data, since next loop will split up all insns into funcs
12072 insn->off = subprog;
12073 /* remember original imm in case JIT fails and fallback
12074 * to interpreter will be needed
12076 env->insn_aux_data[i].call_imm = insn->imm;
12077 /* point imm to __bpf_call_base+1 from JITs point of view */
12081 err = bpf_prog_alloc_jited_linfo(prog);
12083 goto out_undo_insn;
12086 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
12088 goto out_undo_insn;
12090 for (i = 0; i < env->subprog_cnt; i++) {
12091 subprog_start = subprog_end;
12092 subprog_end = env->subprog_info[i + 1].start;
12094 len = subprog_end - subprog_start;
12095 /* BPF_PROG_RUN doesn't call subprogs directly,
12096 * hence main prog stats include the runtime of subprogs.
12097 * subprogs don't have IDs and not reachable via prog_get_next_id
12098 * func[i]->stats will never be accessed and stays NULL
12100 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
12103 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
12104 len * sizeof(struct bpf_insn));
12105 func[i]->type = prog->type;
12106 func[i]->len = len;
12107 if (bpf_prog_calc_tag(func[i]))
12109 func[i]->is_func = 1;
12110 func[i]->aux->func_idx = i;
12111 /* Below members will be freed only at prog->aux */
12112 func[i]->aux->btf = prog->aux->btf;
12113 func[i]->aux->func_info = prog->aux->func_info;
12114 func[i]->aux->poke_tab = prog->aux->poke_tab;
12115 func[i]->aux->size_poke_tab = prog->aux->size_poke_tab;
12117 for (j = 0; j < prog->aux->size_poke_tab; j++) {
12118 struct bpf_jit_poke_descriptor *poke;
12120 poke = &prog->aux->poke_tab[j];
12121 if (poke->insn_idx < subprog_end &&
12122 poke->insn_idx >= subprog_start)
12123 poke->aux = func[i]->aux;
12126 /* Use bpf_prog_F_tag to indicate functions in stack traces.
12127 * Long term would need debug info to populate names
12129 func[i]->aux->name[0] = 'F';
12130 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
12131 func[i]->jit_requested = 1;
12132 func[i]->aux->kfunc_tab = prog->aux->kfunc_tab;
12133 func[i]->aux->linfo = prog->aux->linfo;
12134 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
12135 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
12136 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
12138 insn = func[i]->insnsi;
12139 for (j = 0; j < func[i]->len; j++, insn++) {
12140 if (BPF_CLASS(insn->code) == BPF_LDX &&
12141 BPF_MODE(insn->code) == BPF_PROBE_MEM)
12144 func[i]->aux->num_exentries = num_exentries;
12145 func[i]->aux->tail_call_reachable = env->subprog_info[i].tail_call_reachable;
12146 func[i] = bpf_int_jit_compile(func[i]);
12147 if (!func[i]->jited) {
12154 /* at this point all bpf functions were successfully JITed
12155 * now populate all bpf_calls with correct addresses and
12156 * run last pass of JIT
12158 for (i = 0; i < env->subprog_cnt; i++) {
12159 insn = func[i]->insnsi;
12160 for (j = 0; j < func[i]->len; j++, insn++) {
12161 if (bpf_pseudo_func(insn)) {
12162 subprog = insn[1].imm;
12163 insn[0].imm = (u32)(long)func[subprog]->bpf_func;
12164 insn[1].imm = ((u64)(long)func[subprog]->bpf_func) >> 32;
12167 if (!bpf_pseudo_call(insn))
12169 subprog = insn->off;
12170 insn->imm = BPF_CAST_CALL(func[subprog]->bpf_func) -
12174 /* we use the aux data to keep a list of the start addresses
12175 * of the JITed images for each function in the program
12177 * for some architectures, such as powerpc64, the imm field
12178 * might not be large enough to hold the offset of the start
12179 * address of the callee's JITed image from __bpf_call_base
12181 * in such cases, we can lookup the start address of a callee
12182 * by using its subprog id, available from the off field of
12183 * the call instruction, as an index for this list
12185 func[i]->aux->func = func;
12186 func[i]->aux->func_cnt = env->subprog_cnt;
12188 for (i = 0; i < env->subprog_cnt; i++) {
12189 old_bpf_func = func[i]->bpf_func;
12190 tmp = bpf_int_jit_compile(func[i]);
12191 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
12192 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
12199 /* finally lock prog and jit images for all functions and
12200 * populate kallsysm
12202 for (i = 0; i < env->subprog_cnt; i++) {
12203 bpf_prog_lock_ro(func[i]);
12204 bpf_prog_kallsyms_add(func[i]);
12207 /* Last step: make now unused interpreter insns from main
12208 * prog consistent for later dump requests, so they can
12209 * later look the same as if they were interpreted only.
12211 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
12212 if (bpf_pseudo_func(insn)) {
12213 insn[0].imm = env->insn_aux_data[i].call_imm;
12214 insn[1].imm = find_subprog(env, i + insn[0].imm + 1);
12217 if (!bpf_pseudo_call(insn))
12219 insn->off = env->insn_aux_data[i].call_imm;
12220 subprog = find_subprog(env, i + insn->off + 1);
12221 insn->imm = subprog;
12225 prog->bpf_func = func[0]->bpf_func;
12226 prog->aux->func = func;
12227 prog->aux->func_cnt = env->subprog_cnt;
12228 bpf_prog_jit_attempt_done(prog);
12231 /* We failed JIT'ing, so at this point we need to unregister poke
12232 * descriptors from subprogs, so that kernel is not attempting to
12233 * patch it anymore as we're freeing the subprog JIT memory.
12235 for (i = 0; i < prog->aux->size_poke_tab; i++) {
12236 map_ptr = prog->aux->poke_tab[i].tail_call.map;
12237 map_ptr->ops->map_poke_untrack(map_ptr, prog->aux);
12239 /* At this point we're guaranteed that poke descriptors are not
12240 * live anymore. We can just unlink its descriptor table as it's
12241 * released with the main prog.
12243 for (i = 0; i < env->subprog_cnt; i++) {
12246 func[i]->aux->poke_tab = NULL;
12247 bpf_jit_free(func[i]);
12251 /* cleanup main prog to be interpreted */
12252 prog->jit_requested = 0;
12253 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
12254 if (!bpf_pseudo_call(insn))
12257 insn->imm = env->insn_aux_data[i].call_imm;
12259 bpf_prog_jit_attempt_done(prog);
12263 static int fixup_call_args(struct bpf_verifier_env *env)
12265 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
12266 struct bpf_prog *prog = env->prog;
12267 struct bpf_insn *insn = prog->insnsi;
12268 bool has_kfunc_call = bpf_prog_has_kfunc_call(prog);
12273 if (env->prog->jit_requested &&
12274 !bpf_prog_is_dev_bound(env->prog->aux)) {
12275 err = jit_subprogs(env);
12278 if (err == -EFAULT)
12281 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
12282 if (has_kfunc_call) {
12283 verbose(env, "calling kernel functions are not allowed in non-JITed programs\n");
12286 if (env->subprog_cnt > 1 && env->prog->aux->tail_call_reachable) {
12287 /* When JIT fails the progs with bpf2bpf calls and tail_calls
12288 * have to be rejected, since interpreter doesn't support them yet.
12290 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
12293 for (i = 0; i < prog->len; i++, insn++) {
12294 if (bpf_pseudo_func(insn)) {
12295 /* When JIT fails the progs with callback calls
12296 * have to be rejected, since interpreter doesn't support them yet.
12298 verbose(env, "callbacks are not allowed in non-JITed programs\n");
12302 if (!bpf_pseudo_call(insn))
12304 depth = get_callee_stack_depth(env, insn, i);
12307 bpf_patch_call_args(insn, depth);
12314 static int fixup_kfunc_call(struct bpf_verifier_env *env,
12315 struct bpf_insn *insn)
12317 const struct bpf_kfunc_desc *desc;
12319 /* insn->imm has the btf func_id. Replace it with
12320 * an address (relative to __bpf_base_call).
12322 desc = find_kfunc_desc(env->prog, insn->imm);
12324 verbose(env, "verifier internal error: kernel function descriptor not found for func_id %u\n",
12329 insn->imm = desc->imm;
12334 /* Do various post-verification rewrites in a single program pass.
12335 * These rewrites simplify JIT and interpreter implementations.
12337 static int do_misc_fixups(struct bpf_verifier_env *env)
12339 struct bpf_prog *prog = env->prog;
12340 bool expect_blinding = bpf_jit_blinding_enabled(prog);
12341 struct bpf_insn *insn = prog->insnsi;
12342 const struct bpf_func_proto *fn;
12343 const int insn_cnt = prog->len;
12344 const struct bpf_map_ops *ops;
12345 struct bpf_insn_aux_data *aux;
12346 struct bpf_insn insn_buf[16];
12347 struct bpf_prog *new_prog;
12348 struct bpf_map *map_ptr;
12349 int i, ret, cnt, delta = 0;
12351 for (i = 0; i < insn_cnt; i++, insn++) {
12352 /* Make divide-by-zero exceptions impossible. */
12353 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
12354 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
12355 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
12356 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
12357 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
12358 bool isdiv = BPF_OP(insn->code) == BPF_DIV;
12359 struct bpf_insn *patchlet;
12360 struct bpf_insn chk_and_div[] = {
12361 /* [R,W]x div 0 -> 0 */
12362 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
12363 BPF_JNE | BPF_K, insn->src_reg,
12365 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
12366 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
12369 struct bpf_insn chk_and_mod[] = {
12370 /* [R,W]x mod 0 -> [R,W]x */
12371 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
12372 BPF_JEQ | BPF_K, insn->src_reg,
12373 0, 1 + (is64 ? 0 : 1), 0),
12375 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
12376 BPF_MOV32_REG(insn->dst_reg, insn->dst_reg),
12379 patchlet = isdiv ? chk_and_div : chk_and_mod;
12380 cnt = isdiv ? ARRAY_SIZE(chk_and_div) :
12381 ARRAY_SIZE(chk_and_mod) - (is64 ? 2 : 0);
12383 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
12388 env->prog = prog = new_prog;
12389 insn = new_prog->insnsi + i + delta;
12393 /* Implement LD_ABS and LD_IND with a rewrite, if supported by the program type. */
12394 if (BPF_CLASS(insn->code) == BPF_LD &&
12395 (BPF_MODE(insn->code) == BPF_ABS ||
12396 BPF_MODE(insn->code) == BPF_IND)) {
12397 cnt = env->ops->gen_ld_abs(insn, insn_buf);
12398 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
12399 verbose(env, "bpf verifier is misconfigured\n");
12403 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12408 env->prog = prog = new_prog;
12409 insn = new_prog->insnsi + i + delta;
12413 /* Rewrite pointer arithmetic to mitigate speculation attacks. */
12414 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
12415 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
12416 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
12417 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
12418 struct bpf_insn *patch = &insn_buf[0];
12419 bool issrc, isneg, isimm;
12422 aux = &env->insn_aux_data[i + delta];
12423 if (!aux->alu_state ||
12424 aux->alu_state == BPF_ALU_NON_POINTER)
12427 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
12428 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
12429 BPF_ALU_SANITIZE_SRC;
12430 isimm = aux->alu_state & BPF_ALU_IMMEDIATE;
12432 off_reg = issrc ? insn->src_reg : insn->dst_reg;
12434 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
12437 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
12438 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
12439 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
12440 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
12441 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
12442 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
12443 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX, off_reg);
12446 *patch++ = BPF_MOV64_REG(insn->dst_reg, insn->src_reg);
12447 insn->src_reg = BPF_REG_AX;
12449 insn->code = insn->code == code_add ?
12450 code_sub : code_add;
12452 if (issrc && isneg && !isimm)
12453 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
12454 cnt = patch - insn_buf;
12456 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12461 env->prog = prog = new_prog;
12462 insn = new_prog->insnsi + i + delta;
12466 if (insn->code != (BPF_JMP | BPF_CALL))
12468 if (insn->src_reg == BPF_PSEUDO_CALL)
12470 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
12471 ret = fixup_kfunc_call(env, insn);
12477 if (insn->imm == BPF_FUNC_get_route_realm)
12478 prog->dst_needed = 1;
12479 if (insn->imm == BPF_FUNC_get_prandom_u32)
12480 bpf_user_rnd_init_once();
12481 if (insn->imm == BPF_FUNC_override_return)
12482 prog->kprobe_override = 1;
12483 if (insn->imm == BPF_FUNC_tail_call) {
12484 /* If we tail call into other programs, we
12485 * cannot make any assumptions since they can
12486 * be replaced dynamically during runtime in
12487 * the program array.
12489 prog->cb_access = 1;
12490 if (!allow_tail_call_in_subprogs(env))
12491 prog->aux->stack_depth = MAX_BPF_STACK;
12492 prog->aux->max_pkt_offset = MAX_PACKET_OFF;
12494 /* mark bpf_tail_call as different opcode to avoid
12495 * conditional branch in the interpreter for every normal
12496 * call and to prevent accidental JITing by JIT compiler
12497 * that doesn't support bpf_tail_call yet
12500 insn->code = BPF_JMP | BPF_TAIL_CALL;
12502 aux = &env->insn_aux_data[i + delta];
12503 if (env->bpf_capable && !expect_blinding &&
12504 prog->jit_requested &&
12505 !bpf_map_key_poisoned(aux) &&
12506 !bpf_map_ptr_poisoned(aux) &&
12507 !bpf_map_ptr_unpriv(aux)) {
12508 struct bpf_jit_poke_descriptor desc = {
12509 .reason = BPF_POKE_REASON_TAIL_CALL,
12510 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
12511 .tail_call.key = bpf_map_key_immediate(aux),
12512 .insn_idx = i + delta,
12515 ret = bpf_jit_add_poke_descriptor(prog, &desc);
12517 verbose(env, "adding tail call poke descriptor failed\n");
12521 insn->imm = ret + 1;
12525 if (!bpf_map_ptr_unpriv(aux))
12528 /* instead of changing every JIT dealing with tail_call
12529 * emit two extra insns:
12530 * if (index >= max_entries) goto out;
12531 * index &= array->index_mask;
12532 * to avoid out-of-bounds cpu speculation
12534 if (bpf_map_ptr_poisoned(aux)) {
12535 verbose(env, "tail_call abusing map_ptr\n");
12539 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
12540 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
12541 map_ptr->max_entries, 2);
12542 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
12543 container_of(map_ptr,
12546 insn_buf[2] = *insn;
12548 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12553 env->prog = prog = new_prog;
12554 insn = new_prog->insnsi + i + delta;
12558 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
12559 * and other inlining handlers are currently limited to 64 bit
12562 if (prog->jit_requested && BITS_PER_LONG == 64 &&
12563 (insn->imm == BPF_FUNC_map_lookup_elem ||
12564 insn->imm == BPF_FUNC_map_update_elem ||
12565 insn->imm == BPF_FUNC_map_delete_elem ||
12566 insn->imm == BPF_FUNC_map_push_elem ||
12567 insn->imm == BPF_FUNC_map_pop_elem ||
12568 insn->imm == BPF_FUNC_map_peek_elem ||
12569 insn->imm == BPF_FUNC_redirect_map)) {
12570 aux = &env->insn_aux_data[i + delta];
12571 if (bpf_map_ptr_poisoned(aux))
12572 goto patch_call_imm;
12574 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
12575 ops = map_ptr->ops;
12576 if (insn->imm == BPF_FUNC_map_lookup_elem &&
12577 ops->map_gen_lookup) {
12578 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
12579 if (cnt == -EOPNOTSUPP)
12580 goto patch_map_ops_generic;
12581 if (cnt <= 0 || cnt >= ARRAY_SIZE(insn_buf)) {
12582 verbose(env, "bpf verifier is misconfigured\n");
12586 new_prog = bpf_patch_insn_data(env, i + delta,
12592 env->prog = prog = new_prog;
12593 insn = new_prog->insnsi + i + delta;
12597 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
12598 (void *(*)(struct bpf_map *map, void *key))NULL));
12599 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
12600 (int (*)(struct bpf_map *map, void *key))NULL));
12601 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
12602 (int (*)(struct bpf_map *map, void *key, void *value,
12604 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
12605 (int (*)(struct bpf_map *map, void *value,
12607 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
12608 (int (*)(struct bpf_map *map, void *value))NULL));
12609 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
12610 (int (*)(struct bpf_map *map, void *value))NULL));
12611 BUILD_BUG_ON(!__same_type(ops->map_redirect,
12612 (int (*)(struct bpf_map *map, u32 ifindex, u64 flags))NULL));
12614 patch_map_ops_generic:
12615 switch (insn->imm) {
12616 case BPF_FUNC_map_lookup_elem:
12617 insn->imm = BPF_CAST_CALL(ops->map_lookup_elem) -
12620 case BPF_FUNC_map_update_elem:
12621 insn->imm = BPF_CAST_CALL(ops->map_update_elem) -
12624 case BPF_FUNC_map_delete_elem:
12625 insn->imm = BPF_CAST_CALL(ops->map_delete_elem) -
12628 case BPF_FUNC_map_push_elem:
12629 insn->imm = BPF_CAST_CALL(ops->map_push_elem) -
12632 case BPF_FUNC_map_pop_elem:
12633 insn->imm = BPF_CAST_CALL(ops->map_pop_elem) -
12636 case BPF_FUNC_map_peek_elem:
12637 insn->imm = BPF_CAST_CALL(ops->map_peek_elem) -
12640 case BPF_FUNC_redirect_map:
12641 insn->imm = BPF_CAST_CALL(ops->map_redirect) -
12646 goto patch_call_imm;
12649 /* Implement bpf_jiffies64 inline. */
12650 if (prog->jit_requested && BITS_PER_LONG == 64 &&
12651 insn->imm == BPF_FUNC_jiffies64) {
12652 struct bpf_insn ld_jiffies_addr[2] = {
12653 BPF_LD_IMM64(BPF_REG_0,
12654 (unsigned long)&jiffies),
12657 insn_buf[0] = ld_jiffies_addr[0];
12658 insn_buf[1] = ld_jiffies_addr[1];
12659 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
12663 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
12669 env->prog = prog = new_prog;
12670 insn = new_prog->insnsi + i + delta;
12675 fn = env->ops->get_func_proto(insn->imm, env->prog);
12676 /* all functions that have prototype and verifier allowed
12677 * programs to call them, must be real in-kernel functions
12681 "kernel subsystem misconfigured func %s#%d\n",
12682 func_id_name(insn->imm), insn->imm);
12685 insn->imm = fn->func - __bpf_call_base;
12688 /* Since poke tab is now finalized, publish aux to tracker. */
12689 for (i = 0; i < prog->aux->size_poke_tab; i++) {
12690 map_ptr = prog->aux->poke_tab[i].tail_call.map;
12691 if (!map_ptr->ops->map_poke_track ||
12692 !map_ptr->ops->map_poke_untrack ||
12693 !map_ptr->ops->map_poke_run) {
12694 verbose(env, "bpf verifier is misconfigured\n");
12698 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
12700 verbose(env, "tracking tail call prog failed\n");
12705 sort_kfunc_descs_by_imm(env->prog);
12710 static void free_states(struct bpf_verifier_env *env)
12712 struct bpf_verifier_state_list *sl, *sln;
12715 sl = env->free_list;
12718 free_verifier_state(&sl->state, false);
12722 env->free_list = NULL;
12724 if (!env->explored_states)
12727 for (i = 0; i < state_htab_size(env); i++) {
12728 sl = env->explored_states[i];
12732 free_verifier_state(&sl->state, false);
12736 env->explored_states[i] = NULL;
12740 static int do_check_common(struct bpf_verifier_env *env, int subprog)
12742 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
12743 struct bpf_verifier_state *state;
12744 struct bpf_reg_state *regs;
12747 env->prev_linfo = NULL;
12750 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
12753 state->curframe = 0;
12754 state->speculative = false;
12755 state->branches = 1;
12756 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
12757 if (!state->frame[0]) {
12761 env->cur_state = state;
12762 init_func_state(env, state->frame[0],
12763 BPF_MAIN_FUNC /* callsite */,
12767 regs = state->frame[state->curframe]->regs;
12768 if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
12769 ret = btf_prepare_func_args(env, subprog, regs);
12772 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
12773 if (regs[i].type == PTR_TO_CTX)
12774 mark_reg_known_zero(env, regs, i);
12775 else if (regs[i].type == SCALAR_VALUE)
12776 mark_reg_unknown(env, regs, i);
12777 else if (regs[i].type == PTR_TO_MEM_OR_NULL) {
12778 const u32 mem_size = regs[i].mem_size;
12780 mark_reg_known_zero(env, regs, i);
12781 regs[i].mem_size = mem_size;
12782 regs[i].id = ++env->id_gen;
12786 /* 1st arg to a function */
12787 regs[BPF_REG_1].type = PTR_TO_CTX;
12788 mark_reg_known_zero(env, regs, BPF_REG_1);
12789 ret = btf_check_subprog_arg_match(env, subprog, regs);
12790 if (ret == -EFAULT)
12791 /* unlikely verifier bug. abort.
12792 * ret == 0 and ret < 0 are sadly acceptable for
12793 * main() function due to backward compatibility.
12794 * Like socket filter program may be written as:
12795 * int bpf_prog(struct pt_regs *ctx)
12796 * and never dereference that ctx in the program.
12797 * 'struct pt_regs' is a type mismatch for socket
12798 * filter that should be using 'struct __sk_buff'.
12803 ret = do_check(env);
12805 /* check for NULL is necessary, since cur_state can be freed inside
12806 * do_check() under memory pressure.
12808 if (env->cur_state) {
12809 free_verifier_state(env->cur_state, true);
12810 env->cur_state = NULL;
12812 while (!pop_stack(env, NULL, NULL, false));
12813 if (!ret && pop_log)
12814 bpf_vlog_reset(&env->log, 0);
12819 /* Verify all global functions in a BPF program one by one based on their BTF.
12820 * All global functions must pass verification. Otherwise the whole program is rejected.
12831 * foo() will be verified first for R1=any_scalar_value. During verification it
12832 * will be assumed that bar() already verified successfully and call to bar()
12833 * from foo() will be checked for type match only. Later bar() will be verified
12834 * independently to check that it's safe for R1=any_scalar_value.
12836 static int do_check_subprogs(struct bpf_verifier_env *env)
12838 struct bpf_prog_aux *aux = env->prog->aux;
12841 if (!aux->func_info)
12844 for (i = 1; i < env->subprog_cnt; i++) {
12845 if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL)
12847 env->insn_idx = env->subprog_info[i].start;
12848 WARN_ON_ONCE(env->insn_idx == 0);
12849 ret = do_check_common(env, i);
12852 } else if (env->log.level & BPF_LOG_LEVEL) {
12854 "Func#%d is safe for any args that match its prototype\n",
12861 static int do_check_main(struct bpf_verifier_env *env)
12866 ret = do_check_common(env, 0);
12868 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
12873 static void print_verification_stats(struct bpf_verifier_env *env)
12877 if (env->log.level & BPF_LOG_STATS) {
12878 verbose(env, "verification time %lld usec\n",
12879 div_u64(env->verification_time, 1000));
12880 verbose(env, "stack depth ");
12881 for (i = 0; i < env->subprog_cnt; i++) {
12882 u32 depth = env->subprog_info[i].stack_depth;
12884 verbose(env, "%d", depth);
12885 if (i + 1 < env->subprog_cnt)
12888 verbose(env, "\n");
12890 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
12891 "total_states %d peak_states %d mark_read %d\n",
12892 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
12893 env->max_states_per_insn, env->total_states,
12894 env->peak_states, env->longest_mark_read_walk);
12897 static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
12899 const struct btf_type *t, *func_proto;
12900 const struct bpf_struct_ops *st_ops;
12901 const struct btf_member *member;
12902 struct bpf_prog *prog = env->prog;
12903 u32 btf_id, member_idx;
12906 if (!prog->gpl_compatible) {
12907 verbose(env, "struct ops programs must have a GPL compatible license\n");
12911 btf_id = prog->aux->attach_btf_id;
12912 st_ops = bpf_struct_ops_find(btf_id);
12914 verbose(env, "attach_btf_id %u is not a supported struct\n",
12920 member_idx = prog->expected_attach_type;
12921 if (member_idx >= btf_type_vlen(t)) {
12922 verbose(env, "attach to invalid member idx %u of struct %s\n",
12923 member_idx, st_ops->name);
12927 member = &btf_type_member(t)[member_idx];
12928 mname = btf_name_by_offset(btf_vmlinux, member->name_off);
12929 func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type,
12932 verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
12933 mname, member_idx, st_ops->name);
12937 if (st_ops->check_member) {
12938 int err = st_ops->check_member(t, member);
12941 verbose(env, "attach to unsupported member %s of struct %s\n",
12942 mname, st_ops->name);
12947 prog->aux->attach_func_proto = func_proto;
12948 prog->aux->attach_func_name = mname;
12949 env->ops = st_ops->verifier_ops;
12953 #define SECURITY_PREFIX "security_"
12955 static int check_attach_modify_return(unsigned long addr, const char *func_name)
12957 if (within_error_injection_list(addr) ||
12958 !strncmp(SECURITY_PREFIX, func_name, sizeof(SECURITY_PREFIX) - 1))
12964 /* list of non-sleepable functions that are otherwise on
12965 * ALLOW_ERROR_INJECTION list
12967 BTF_SET_START(btf_non_sleepable_error_inject)
12968 /* Three functions below can be called from sleepable and non-sleepable context.
12969 * Assume non-sleepable from bpf safety point of view.
12971 BTF_ID(func, __add_to_page_cache_locked)
12972 BTF_ID(func, should_fail_alloc_page)
12973 BTF_ID(func, should_failslab)
12974 BTF_SET_END(btf_non_sleepable_error_inject)
12976 static int check_non_sleepable_error_inject(u32 btf_id)
12978 return btf_id_set_contains(&btf_non_sleepable_error_inject, btf_id);
12981 int bpf_check_attach_target(struct bpf_verifier_log *log,
12982 const struct bpf_prog *prog,
12983 const struct bpf_prog *tgt_prog,
12985 struct bpf_attach_target_info *tgt_info)
12987 bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
12988 const char prefix[] = "btf_trace_";
12989 int ret = 0, subprog = -1, i;
12990 const struct btf_type *t;
12991 bool conservative = true;
12997 bpf_log(log, "Tracing programs must provide btf_id\n");
13000 btf = tgt_prog ? tgt_prog->aux->btf : prog->aux->attach_btf;
13003 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
13006 t = btf_type_by_id(btf, btf_id);
13008 bpf_log(log, "attach_btf_id %u is invalid\n", btf_id);
13011 tname = btf_name_by_offset(btf, t->name_off);
13013 bpf_log(log, "attach_btf_id %u doesn't have a name\n", btf_id);
13017 struct bpf_prog_aux *aux = tgt_prog->aux;
13019 for (i = 0; i < aux->func_info_cnt; i++)
13020 if (aux->func_info[i].type_id == btf_id) {
13024 if (subprog == -1) {
13025 bpf_log(log, "Subprog %s doesn't exist\n", tname);
13028 conservative = aux->func_info_aux[subprog].unreliable;
13029 if (prog_extension) {
13030 if (conservative) {
13032 "Cannot replace static functions\n");
13035 if (!prog->jit_requested) {
13037 "Extension programs should be JITed\n");
13041 if (!tgt_prog->jited) {
13042 bpf_log(log, "Can attach to only JITed progs\n");
13045 if (tgt_prog->type == prog->type) {
13046 /* Cannot fentry/fexit another fentry/fexit program.
13047 * Cannot attach program extension to another extension.
13048 * It's ok to attach fentry/fexit to extension program.
13050 bpf_log(log, "Cannot recursively attach\n");
13053 if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
13055 (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
13056 tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
13057 /* Program extensions can extend all program types
13058 * except fentry/fexit. The reason is the following.
13059 * The fentry/fexit programs are used for performance
13060 * analysis, stats and can be attached to any program
13061 * type except themselves. When extension program is
13062 * replacing XDP function it is necessary to allow
13063 * performance analysis of all functions. Both original
13064 * XDP program and its program extension. Hence
13065 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
13066 * allowed. If extending of fentry/fexit was allowed it
13067 * would be possible to create long call chain
13068 * fentry->extension->fentry->extension beyond
13069 * reasonable stack size. Hence extending fentry is not
13072 bpf_log(log, "Cannot extend fentry/fexit\n");
13076 if (prog_extension) {
13077 bpf_log(log, "Cannot replace kernel functions\n");
13082 switch (prog->expected_attach_type) {
13083 case BPF_TRACE_RAW_TP:
13086 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
13089 if (!btf_type_is_typedef(t)) {
13090 bpf_log(log, "attach_btf_id %u is not a typedef\n",
13094 if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
13095 bpf_log(log, "attach_btf_id %u points to wrong type name %s\n",
13099 tname += sizeof(prefix) - 1;
13100 t = btf_type_by_id(btf, t->type);
13101 if (!btf_type_is_ptr(t))
13102 /* should never happen in valid vmlinux build */
13104 t = btf_type_by_id(btf, t->type);
13105 if (!btf_type_is_func_proto(t))
13106 /* should never happen in valid vmlinux build */
13110 case BPF_TRACE_ITER:
13111 if (!btf_type_is_func(t)) {
13112 bpf_log(log, "attach_btf_id %u is not a function\n",
13116 t = btf_type_by_id(btf, t->type);
13117 if (!btf_type_is_func_proto(t))
13119 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
13124 if (!prog_extension)
13127 case BPF_MODIFY_RETURN:
13129 case BPF_TRACE_FENTRY:
13130 case BPF_TRACE_FEXIT:
13131 if (!btf_type_is_func(t)) {
13132 bpf_log(log, "attach_btf_id %u is not a function\n",
13136 if (prog_extension &&
13137 btf_check_type_match(log, prog, btf, t))
13139 t = btf_type_by_id(btf, t->type);
13140 if (!btf_type_is_func_proto(t))
13143 if ((prog->aux->saved_dst_prog_type || prog->aux->saved_dst_attach_type) &&
13144 (!tgt_prog || prog->aux->saved_dst_prog_type != tgt_prog->type ||
13145 prog->aux->saved_dst_attach_type != tgt_prog->expected_attach_type))
13148 if (tgt_prog && conservative)
13151 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
13157 addr = (long) tgt_prog->bpf_func;
13159 addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
13161 addr = kallsyms_lookup_name(tname);
13164 "The address of function %s cannot be found\n",
13170 if (prog->aux->sleepable) {
13172 switch (prog->type) {
13173 case BPF_PROG_TYPE_TRACING:
13174 /* fentry/fexit/fmod_ret progs can be sleepable only if they are
13175 * attached to ALLOW_ERROR_INJECTION and are not in denylist.
13177 if (!check_non_sleepable_error_inject(btf_id) &&
13178 within_error_injection_list(addr))
13181 case BPF_PROG_TYPE_LSM:
13182 /* LSM progs check that they are attached to bpf_lsm_*() funcs.
13183 * Only some of them are sleepable.
13185 if (bpf_lsm_is_sleepable_hook(btf_id))
13192 bpf_log(log, "%s is not sleepable\n", tname);
13195 } else if (prog->expected_attach_type == BPF_MODIFY_RETURN) {
13197 bpf_log(log, "can't modify return codes of BPF programs\n");
13200 ret = check_attach_modify_return(addr, tname);
13202 bpf_log(log, "%s() is not modifiable\n", tname);
13209 tgt_info->tgt_addr = addr;
13210 tgt_info->tgt_name = tname;
13211 tgt_info->tgt_type = t;
13215 BTF_SET_START(btf_id_deny)
13218 BTF_ID(func, migrate_disable)
13219 BTF_ID(func, migrate_enable)
13221 #if !defined CONFIG_PREEMPT_RCU && !defined CONFIG_TINY_RCU
13222 BTF_ID(func, rcu_read_unlock_strict)
13224 BTF_SET_END(btf_id_deny)
13226 static int check_attach_btf_id(struct bpf_verifier_env *env)
13228 struct bpf_prog *prog = env->prog;
13229 struct bpf_prog *tgt_prog = prog->aux->dst_prog;
13230 struct bpf_attach_target_info tgt_info = {};
13231 u32 btf_id = prog->aux->attach_btf_id;
13232 struct bpf_trampoline *tr;
13236 if (prog->type == BPF_PROG_TYPE_SYSCALL) {
13237 if (prog->aux->sleepable)
13238 /* attach_btf_id checked to be zero already */
13240 verbose(env, "Syscall programs can only be sleepable\n");
13244 if (prog->aux->sleepable && prog->type != BPF_PROG_TYPE_TRACING &&
13245 prog->type != BPF_PROG_TYPE_LSM) {
13246 verbose(env, "Only fentry/fexit/fmod_ret and lsm programs can be sleepable\n");
13250 if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
13251 return check_struct_ops_btf_id(env);
13253 if (prog->type != BPF_PROG_TYPE_TRACING &&
13254 prog->type != BPF_PROG_TYPE_LSM &&
13255 prog->type != BPF_PROG_TYPE_EXT)
13258 ret = bpf_check_attach_target(&env->log, prog, tgt_prog, btf_id, &tgt_info);
13262 if (tgt_prog && prog->type == BPF_PROG_TYPE_EXT) {
13263 /* to make freplace equivalent to their targets, they need to
13264 * inherit env->ops and expected_attach_type for the rest of the
13267 env->ops = bpf_verifier_ops[tgt_prog->type];
13268 prog->expected_attach_type = tgt_prog->expected_attach_type;
13271 /* store info about the attachment target that will be used later */
13272 prog->aux->attach_func_proto = tgt_info.tgt_type;
13273 prog->aux->attach_func_name = tgt_info.tgt_name;
13276 prog->aux->saved_dst_prog_type = tgt_prog->type;
13277 prog->aux->saved_dst_attach_type = tgt_prog->expected_attach_type;
13280 if (prog->expected_attach_type == BPF_TRACE_RAW_TP) {
13281 prog->aux->attach_btf_trace = true;
13283 } else if (prog->expected_attach_type == BPF_TRACE_ITER) {
13284 if (!bpf_iter_prog_supported(prog))
13289 if (prog->type == BPF_PROG_TYPE_LSM) {
13290 ret = bpf_lsm_verify_prog(&env->log, prog);
13293 } else if (prog->type == BPF_PROG_TYPE_TRACING &&
13294 btf_id_set_contains(&btf_id_deny, btf_id)) {
13298 key = bpf_trampoline_compute_key(tgt_prog, prog->aux->attach_btf, btf_id);
13299 tr = bpf_trampoline_get(key, &tgt_info);
13303 prog->aux->dst_trampoline = tr;
13307 struct btf *bpf_get_btf_vmlinux(void)
13309 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
13310 mutex_lock(&bpf_verifier_lock);
13312 btf_vmlinux = btf_parse_vmlinux();
13313 mutex_unlock(&bpf_verifier_lock);
13315 return btf_vmlinux;
13318 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr, bpfptr_t uattr)
13320 u64 start_time = ktime_get_ns();
13321 struct bpf_verifier_env *env;
13322 struct bpf_verifier_log *log;
13323 int i, len, ret = -EINVAL;
13326 /* no program is valid */
13327 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
13330 /* 'struct bpf_verifier_env' can be global, but since it's not small,
13331 * allocate/free it every time bpf_check() is called
13333 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
13338 len = (*prog)->len;
13339 env->insn_aux_data =
13340 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
13342 if (!env->insn_aux_data)
13344 for (i = 0; i < len; i++)
13345 env->insn_aux_data[i].orig_idx = i;
13347 env->ops = bpf_verifier_ops[env->prog->type];
13348 env->fd_array = make_bpfptr(attr->fd_array, uattr.is_kernel);
13349 is_priv = bpf_capable();
13351 bpf_get_btf_vmlinux();
13353 /* grab the mutex to protect few globals used by verifier */
13355 mutex_lock(&bpf_verifier_lock);
13357 if (attr->log_level || attr->log_buf || attr->log_size) {
13358 /* user requested verbose verifier output
13359 * and supplied buffer to store the verification trace
13361 log->level = attr->log_level;
13362 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
13363 log->len_total = attr->log_size;
13366 /* log attributes have to be sane */
13367 if (log->len_total < 128 || log->len_total > UINT_MAX >> 2 ||
13368 !log->level || !log->ubuf || log->level & ~BPF_LOG_MASK)
13372 if (IS_ERR(btf_vmlinux)) {
13373 /* Either gcc or pahole or kernel are broken. */
13374 verbose(env, "in-kernel BTF is malformed\n");
13375 ret = PTR_ERR(btf_vmlinux);
13376 goto skip_full_check;
13379 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
13380 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
13381 env->strict_alignment = true;
13382 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
13383 env->strict_alignment = false;
13385 env->allow_ptr_leaks = bpf_allow_ptr_leaks();
13386 env->allow_uninit_stack = bpf_allow_uninit_stack();
13387 env->allow_ptr_to_map_access = bpf_allow_ptr_to_map_access();
13388 env->bypass_spec_v1 = bpf_bypass_spec_v1();
13389 env->bypass_spec_v4 = bpf_bypass_spec_v4();
13390 env->bpf_capable = bpf_capable();
13393 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
13395 env->explored_states = kvcalloc(state_htab_size(env),
13396 sizeof(struct bpf_verifier_state_list *),
13399 if (!env->explored_states)
13400 goto skip_full_check;
13402 ret = add_subprog_and_kfunc(env);
13404 goto skip_full_check;
13406 ret = check_subprogs(env);
13408 goto skip_full_check;
13410 ret = check_btf_info(env, attr, uattr);
13412 goto skip_full_check;
13414 ret = check_attach_btf_id(env);
13416 goto skip_full_check;
13418 ret = resolve_pseudo_ldimm64(env);
13420 goto skip_full_check;
13422 if (bpf_prog_is_dev_bound(env->prog->aux)) {
13423 ret = bpf_prog_offload_verifier_prep(env->prog);
13425 goto skip_full_check;
13428 ret = check_cfg(env);
13430 goto skip_full_check;
13432 ret = do_check_subprogs(env);
13433 ret = ret ?: do_check_main(env);
13435 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
13436 ret = bpf_prog_offload_finalize(env);
13439 kvfree(env->explored_states);
13442 ret = check_max_stack_depth(env);
13444 /* instruction rewrites happen after this point */
13447 opt_hard_wire_dead_code_branches(env);
13449 ret = opt_remove_dead_code(env);
13451 ret = opt_remove_nops(env);
13454 sanitize_dead_code(env);
13458 /* program is valid, convert *(u32*)(ctx + off) accesses */
13459 ret = convert_ctx_accesses(env);
13462 ret = do_misc_fixups(env);
13464 /* do 32-bit optimization after insn patching has done so those patched
13465 * insns could be handled correctly.
13467 if (ret == 0 && !bpf_prog_is_dev_bound(env->prog->aux)) {
13468 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
13469 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
13474 ret = fixup_call_args(env);
13476 env->verification_time = ktime_get_ns() - start_time;
13477 print_verification_stats(env);
13479 if (log->level && bpf_verifier_log_full(log))
13481 if (log->level && !log->ubuf) {
13483 goto err_release_maps;
13487 goto err_release_maps;
13489 if (env->used_map_cnt) {
13490 /* if program passed verifier, update used_maps in bpf_prog_info */
13491 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
13492 sizeof(env->used_maps[0]),
13495 if (!env->prog->aux->used_maps) {
13497 goto err_release_maps;
13500 memcpy(env->prog->aux->used_maps, env->used_maps,
13501 sizeof(env->used_maps[0]) * env->used_map_cnt);
13502 env->prog->aux->used_map_cnt = env->used_map_cnt;
13504 if (env->used_btf_cnt) {
13505 /* if program passed verifier, update used_btfs in bpf_prog_aux */
13506 env->prog->aux->used_btfs = kmalloc_array(env->used_btf_cnt,
13507 sizeof(env->used_btfs[0]),
13509 if (!env->prog->aux->used_btfs) {
13511 goto err_release_maps;
13514 memcpy(env->prog->aux->used_btfs, env->used_btfs,
13515 sizeof(env->used_btfs[0]) * env->used_btf_cnt);
13516 env->prog->aux->used_btf_cnt = env->used_btf_cnt;
13518 if (env->used_map_cnt || env->used_btf_cnt) {
13519 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
13520 * bpf_ld_imm64 instructions
13522 convert_pseudo_ld_imm64(env);
13525 adjust_btf_func(env);
13528 if (!env->prog->aux->used_maps)
13529 /* if we didn't copy map pointers into bpf_prog_info, release
13530 * them now. Otherwise free_used_maps() will release them.
13533 if (!env->prog->aux->used_btfs)
13536 /* extension progs temporarily inherit the attach_type of their targets
13537 for verification purposes, so set it back to zero before returning
13539 if (env->prog->type == BPF_PROG_TYPE_EXT)
13540 env->prog->expected_attach_type = 0;
13545 mutex_unlock(&bpf_verifier_lock);
13546 vfree(env->insn_aux_data);