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 pathes 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 ether 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 #define COPY_STATE_FN(NAME, COUNT, FIELD, SIZE) \
741 static int copy_##NAME##_state(struct bpf_func_state *dst, \
742 const struct bpf_func_state *src) \
746 if (WARN_ON_ONCE(dst->COUNT < src->COUNT)) { \
747 /* internal bug, make state invalid to reject the program */ \
748 memset(dst, 0, sizeof(*dst)); \
751 memcpy(dst->FIELD, src->FIELD, \
752 sizeof(*src->FIELD) * (src->COUNT / SIZE)); \
755 /* copy_reference_state() */
756 COPY_STATE_FN(reference, acquired_refs, refs, 1)
757 /* copy_stack_state() */
758 COPY_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
761 #define REALLOC_STATE_FN(NAME, COUNT, FIELD, SIZE) \
762 static int realloc_##NAME##_state(struct bpf_func_state *state, int size, \
765 u32 old_size = state->COUNT; \
766 struct bpf_##NAME##_state *new_##FIELD; \
767 int slot = size / SIZE; \
769 if (size <= old_size || !size) { \
772 state->COUNT = slot * SIZE; \
773 if (!size && old_size) { \
774 kfree(state->FIELD); \
775 state->FIELD = NULL; \
779 new_##FIELD = kmalloc_array(slot, sizeof(struct bpf_##NAME##_state), \
785 memcpy(new_##FIELD, state->FIELD, \
786 sizeof(*new_##FIELD) * (old_size / SIZE)); \
787 memset(new_##FIELD + old_size / SIZE, 0, \
788 sizeof(*new_##FIELD) * (size - old_size) / SIZE); \
790 state->COUNT = slot * SIZE; \
791 kfree(state->FIELD); \
792 state->FIELD = new_##FIELD; \
795 /* realloc_reference_state() */
796 REALLOC_STATE_FN(reference, acquired_refs, refs, 1)
797 /* realloc_stack_state() */
798 REALLOC_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
799 #undef REALLOC_STATE_FN
801 /* do_check() starts with zero-sized stack in struct bpf_verifier_state to
802 * make it consume minimal amount of memory. check_stack_write() access from
803 * the program calls into realloc_func_state() to grow the stack size.
804 * Note there is a non-zero 'parent' pointer inside bpf_verifier_state
805 * which realloc_stack_state() copies over. It points to previous
806 * bpf_verifier_state which is never reallocated.
808 static int realloc_func_state(struct bpf_func_state *state, int stack_size,
809 int refs_size, bool copy_old)
811 int err = realloc_reference_state(state, refs_size, copy_old);
814 return realloc_stack_state(state, stack_size, copy_old);
817 /* Acquire a pointer id from the env and update the state->refs to include
818 * this new pointer reference.
819 * On success, returns a valid pointer id to associate with the register
820 * On failure, returns a negative errno.
822 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
824 struct bpf_func_state *state = cur_func(env);
825 int new_ofs = state->acquired_refs;
828 err = realloc_reference_state(state, state->acquired_refs + 1, true);
832 state->refs[new_ofs].id = id;
833 state->refs[new_ofs].insn_idx = insn_idx;
838 /* release function corresponding to acquire_reference_state(). Idempotent. */
839 static int release_reference_state(struct bpf_func_state *state, int ptr_id)
843 last_idx = state->acquired_refs - 1;
844 for (i = 0; i < state->acquired_refs; i++) {
845 if (state->refs[i].id == ptr_id) {
846 if (last_idx && i != last_idx)
847 memcpy(&state->refs[i], &state->refs[last_idx],
848 sizeof(*state->refs));
849 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
850 state->acquired_refs--;
857 static int transfer_reference_state(struct bpf_func_state *dst,
858 struct bpf_func_state *src)
860 int err = realloc_reference_state(dst, src->acquired_refs, false);
863 err = copy_reference_state(dst, src);
869 static void free_func_state(struct bpf_func_state *state)
878 static void clear_jmp_history(struct bpf_verifier_state *state)
880 kfree(state->jmp_history);
881 state->jmp_history = NULL;
882 state->jmp_history_cnt = 0;
885 static void free_verifier_state(struct bpf_verifier_state *state,
890 for (i = 0; i <= state->curframe; i++) {
891 free_func_state(state->frame[i]);
892 state->frame[i] = NULL;
894 clear_jmp_history(state);
899 /* copy verifier state from src to dst growing dst stack space
900 * when necessary to accommodate larger src stack
902 static int copy_func_state(struct bpf_func_state *dst,
903 const struct bpf_func_state *src)
907 err = realloc_func_state(dst, src->allocated_stack, src->acquired_refs,
911 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
912 err = copy_reference_state(dst, src);
915 return copy_stack_state(dst, src);
918 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
919 const struct bpf_verifier_state *src)
921 struct bpf_func_state *dst;
922 u32 jmp_sz = sizeof(struct bpf_idx_pair) * src->jmp_history_cnt;
925 if (dst_state->jmp_history_cnt < src->jmp_history_cnt) {
926 kfree(dst_state->jmp_history);
927 dst_state->jmp_history = kmalloc(jmp_sz, GFP_USER);
928 if (!dst_state->jmp_history)
931 memcpy(dst_state->jmp_history, src->jmp_history, jmp_sz);
932 dst_state->jmp_history_cnt = src->jmp_history_cnt;
934 /* if dst has more stack frames then src frame, free them */
935 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
936 free_func_state(dst_state->frame[i]);
937 dst_state->frame[i] = NULL;
939 dst_state->speculative = src->speculative;
940 dst_state->curframe = src->curframe;
941 dst_state->active_spin_lock = src->active_spin_lock;
942 dst_state->branches = src->branches;
943 dst_state->parent = src->parent;
944 dst_state->first_insn_idx = src->first_insn_idx;
945 dst_state->last_insn_idx = src->last_insn_idx;
946 for (i = 0; i <= src->curframe; i++) {
947 dst = dst_state->frame[i];
949 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
952 dst_state->frame[i] = dst;
954 err = copy_func_state(dst, src->frame[i]);
961 static void update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
964 u32 br = --st->branches;
966 /* WARN_ON(br > 1) technically makes sense here,
967 * but see comment in push_stack(), hence:
969 WARN_ONCE((int)br < 0,
970 "BUG update_branch_counts:branches_to_explore=%d\n",
978 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
979 int *insn_idx, bool pop_log)
981 struct bpf_verifier_state *cur = env->cur_state;
982 struct bpf_verifier_stack_elem *elem, *head = env->head;
985 if (env->head == NULL)
989 err = copy_verifier_state(cur, &head->st);
994 bpf_vlog_reset(&env->log, head->log_pos);
996 *insn_idx = head->insn_idx;
998 *prev_insn_idx = head->prev_insn_idx;
1000 free_verifier_state(&head->st, false);
1007 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
1008 int insn_idx, int prev_insn_idx,
1011 struct bpf_verifier_state *cur = env->cur_state;
1012 struct bpf_verifier_stack_elem *elem;
1015 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
1019 elem->insn_idx = insn_idx;
1020 elem->prev_insn_idx = prev_insn_idx;
1021 elem->next = env->head;
1022 elem->log_pos = env->log.len_used;
1025 err = copy_verifier_state(&elem->st, cur);
1028 elem->st.speculative |= speculative;
1029 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
1030 verbose(env, "The sequence of %d jumps is too complex.\n",
1034 if (elem->st.parent) {
1035 ++elem->st.parent->branches;
1036 /* WARN_ON(branches > 2) technically makes sense here,
1038 * 1. speculative states will bump 'branches' for non-branch
1040 * 2. is_state_visited() heuristics may decide not to create
1041 * a new state for a sequence of branches and all such current
1042 * and cloned states will be pointing to a single parent state
1043 * which might have large 'branches' count.
1048 free_verifier_state(env->cur_state, true);
1049 env->cur_state = NULL;
1050 /* pop all elements and return */
1051 while (!pop_stack(env, NULL, NULL, false));
1055 #define CALLER_SAVED_REGS 6
1056 static const int caller_saved[CALLER_SAVED_REGS] = {
1057 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
1060 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1061 struct bpf_reg_state *reg);
1063 /* This helper doesn't clear reg->id */
1064 static void ___mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1066 reg->var_off = tnum_const(imm);
1067 reg->smin_value = (s64)imm;
1068 reg->smax_value = (s64)imm;
1069 reg->umin_value = imm;
1070 reg->umax_value = imm;
1072 reg->s32_min_value = (s32)imm;
1073 reg->s32_max_value = (s32)imm;
1074 reg->u32_min_value = (u32)imm;
1075 reg->u32_max_value = (u32)imm;
1078 /* Mark the unknown part of a register (variable offset or scalar value) as
1079 * known to have the value @imm.
1081 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1083 /* Clear id, off, and union(map_ptr, range) */
1084 memset(((u8 *)reg) + sizeof(reg->type), 0,
1085 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
1086 ___mark_reg_known(reg, imm);
1089 static void __mark_reg32_known(struct bpf_reg_state *reg, u64 imm)
1091 reg->var_off = tnum_const_subreg(reg->var_off, imm);
1092 reg->s32_min_value = (s32)imm;
1093 reg->s32_max_value = (s32)imm;
1094 reg->u32_min_value = (u32)imm;
1095 reg->u32_max_value = (u32)imm;
1098 /* Mark the 'variable offset' part of a register as zero. This should be
1099 * used only on registers holding a pointer type.
1101 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
1103 __mark_reg_known(reg, 0);
1106 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
1108 __mark_reg_known(reg, 0);
1109 reg->type = SCALAR_VALUE;
1112 static void mark_reg_known_zero(struct bpf_verifier_env *env,
1113 struct bpf_reg_state *regs, u32 regno)
1115 if (WARN_ON(regno >= MAX_BPF_REG)) {
1116 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
1117 /* Something bad happened, let's kill all regs */
1118 for (regno = 0; regno < MAX_BPF_REG; regno++)
1119 __mark_reg_not_init(env, regs + regno);
1122 __mark_reg_known_zero(regs + regno);
1125 static void mark_ptr_not_null_reg(struct bpf_reg_state *reg)
1127 switch (reg->type) {
1128 case PTR_TO_MAP_VALUE_OR_NULL: {
1129 const struct bpf_map *map = reg->map_ptr;
1131 if (map->inner_map_meta) {
1132 reg->type = CONST_PTR_TO_MAP;
1133 reg->map_ptr = map->inner_map_meta;
1134 } else if (map->map_type == BPF_MAP_TYPE_XSKMAP) {
1135 reg->type = PTR_TO_XDP_SOCK;
1136 } else if (map->map_type == BPF_MAP_TYPE_SOCKMAP ||
1137 map->map_type == BPF_MAP_TYPE_SOCKHASH) {
1138 reg->type = PTR_TO_SOCKET;
1140 reg->type = PTR_TO_MAP_VALUE;
1144 case PTR_TO_SOCKET_OR_NULL:
1145 reg->type = PTR_TO_SOCKET;
1147 case PTR_TO_SOCK_COMMON_OR_NULL:
1148 reg->type = PTR_TO_SOCK_COMMON;
1150 case PTR_TO_TCP_SOCK_OR_NULL:
1151 reg->type = PTR_TO_TCP_SOCK;
1153 case PTR_TO_BTF_ID_OR_NULL:
1154 reg->type = PTR_TO_BTF_ID;
1156 case PTR_TO_MEM_OR_NULL:
1157 reg->type = PTR_TO_MEM;
1159 case PTR_TO_RDONLY_BUF_OR_NULL:
1160 reg->type = PTR_TO_RDONLY_BUF;
1162 case PTR_TO_RDWR_BUF_OR_NULL:
1163 reg->type = PTR_TO_RDWR_BUF;
1166 WARN_ONCE(1, "unknown nullable register type");
1170 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
1172 return type_is_pkt_pointer(reg->type);
1175 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
1177 return reg_is_pkt_pointer(reg) ||
1178 reg->type == PTR_TO_PACKET_END;
1181 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
1182 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
1183 enum bpf_reg_type which)
1185 /* The register can already have a range from prior markings.
1186 * This is fine as long as it hasn't been advanced from its
1189 return reg->type == which &&
1192 tnum_equals_const(reg->var_off, 0);
1195 /* Reset the min/max bounds of a register */
1196 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
1198 reg->smin_value = S64_MIN;
1199 reg->smax_value = S64_MAX;
1200 reg->umin_value = 0;
1201 reg->umax_value = U64_MAX;
1203 reg->s32_min_value = S32_MIN;
1204 reg->s32_max_value = S32_MAX;
1205 reg->u32_min_value = 0;
1206 reg->u32_max_value = U32_MAX;
1209 static void __mark_reg64_unbounded(struct bpf_reg_state *reg)
1211 reg->smin_value = S64_MIN;
1212 reg->smax_value = S64_MAX;
1213 reg->umin_value = 0;
1214 reg->umax_value = U64_MAX;
1217 static void __mark_reg32_unbounded(struct bpf_reg_state *reg)
1219 reg->s32_min_value = S32_MIN;
1220 reg->s32_max_value = S32_MAX;
1221 reg->u32_min_value = 0;
1222 reg->u32_max_value = U32_MAX;
1225 static void __update_reg32_bounds(struct bpf_reg_state *reg)
1227 struct tnum var32_off = tnum_subreg(reg->var_off);
1229 /* min signed is max(sign bit) | min(other bits) */
1230 reg->s32_min_value = max_t(s32, reg->s32_min_value,
1231 var32_off.value | (var32_off.mask & S32_MIN));
1232 /* max signed is min(sign bit) | max(other bits) */
1233 reg->s32_max_value = min_t(s32, reg->s32_max_value,
1234 var32_off.value | (var32_off.mask & S32_MAX));
1235 reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)var32_off.value);
1236 reg->u32_max_value = min(reg->u32_max_value,
1237 (u32)(var32_off.value | var32_off.mask));
1240 static void __update_reg64_bounds(struct bpf_reg_state *reg)
1242 /* min signed is max(sign bit) | min(other bits) */
1243 reg->smin_value = max_t(s64, reg->smin_value,
1244 reg->var_off.value | (reg->var_off.mask & S64_MIN));
1245 /* max signed is min(sign bit) | max(other bits) */
1246 reg->smax_value = min_t(s64, reg->smax_value,
1247 reg->var_off.value | (reg->var_off.mask & S64_MAX));
1248 reg->umin_value = max(reg->umin_value, reg->var_off.value);
1249 reg->umax_value = min(reg->umax_value,
1250 reg->var_off.value | reg->var_off.mask);
1253 static void __update_reg_bounds(struct bpf_reg_state *reg)
1255 __update_reg32_bounds(reg);
1256 __update_reg64_bounds(reg);
1259 /* Uses signed min/max values to inform unsigned, and vice-versa */
1260 static void __reg32_deduce_bounds(struct bpf_reg_state *reg)
1262 /* Learn sign from signed bounds.
1263 * If we cannot cross the sign boundary, then signed and unsigned bounds
1264 * are the same, so combine. This works even in the negative case, e.g.
1265 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1267 if (reg->s32_min_value >= 0 || reg->s32_max_value < 0) {
1268 reg->s32_min_value = reg->u32_min_value =
1269 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1270 reg->s32_max_value = reg->u32_max_value =
1271 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1274 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1275 * boundary, so we must be careful.
1277 if ((s32)reg->u32_max_value >= 0) {
1278 /* Positive. We can't learn anything from the smin, but smax
1279 * is positive, hence safe.
1281 reg->s32_min_value = reg->u32_min_value;
1282 reg->s32_max_value = reg->u32_max_value =
1283 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1284 } else if ((s32)reg->u32_min_value < 0) {
1285 /* Negative. We can't learn anything from the smax, but smin
1286 * is negative, hence safe.
1288 reg->s32_min_value = reg->u32_min_value =
1289 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1290 reg->s32_max_value = reg->u32_max_value;
1294 static void __reg64_deduce_bounds(struct bpf_reg_state *reg)
1296 /* Learn sign from signed bounds.
1297 * If we cannot cross the sign boundary, then signed and unsigned bounds
1298 * are the same, so combine. This works even in the negative case, e.g.
1299 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1301 if (reg->smin_value >= 0 || reg->smax_value < 0) {
1302 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1304 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1308 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1309 * boundary, so we must be careful.
1311 if ((s64)reg->umax_value >= 0) {
1312 /* Positive. We can't learn anything from the smin, but smax
1313 * is positive, hence safe.
1315 reg->smin_value = reg->umin_value;
1316 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1318 } else if ((s64)reg->umin_value < 0) {
1319 /* Negative. We can't learn anything from the smax, but smin
1320 * is negative, hence safe.
1322 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1324 reg->smax_value = reg->umax_value;
1328 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
1330 __reg32_deduce_bounds(reg);
1331 __reg64_deduce_bounds(reg);
1334 /* Attempts to improve var_off based on unsigned min/max information */
1335 static void __reg_bound_offset(struct bpf_reg_state *reg)
1337 struct tnum var64_off = tnum_intersect(reg->var_off,
1338 tnum_range(reg->umin_value,
1340 struct tnum var32_off = tnum_intersect(tnum_subreg(reg->var_off),
1341 tnum_range(reg->u32_min_value,
1342 reg->u32_max_value));
1344 reg->var_off = tnum_or(tnum_clear_subreg(var64_off), var32_off);
1347 static void __reg_assign_32_into_64(struct bpf_reg_state *reg)
1349 reg->umin_value = reg->u32_min_value;
1350 reg->umax_value = reg->u32_max_value;
1351 /* Attempt to pull 32-bit signed bounds into 64-bit bounds
1352 * but must be positive otherwise set to worse case bounds
1353 * and refine later from tnum.
1355 if (reg->s32_min_value >= 0 && reg->s32_max_value >= 0)
1356 reg->smax_value = reg->s32_max_value;
1358 reg->smax_value = U32_MAX;
1359 if (reg->s32_min_value >= 0)
1360 reg->smin_value = reg->s32_min_value;
1362 reg->smin_value = 0;
1365 static void __reg_combine_32_into_64(struct bpf_reg_state *reg)
1367 /* special case when 64-bit register has upper 32-bit register
1368 * zeroed. Typically happens after zext or <<32, >>32 sequence
1369 * allowing us to use 32-bit bounds directly,
1371 if (tnum_equals_const(tnum_clear_subreg(reg->var_off), 0)) {
1372 __reg_assign_32_into_64(reg);
1374 /* Otherwise the best we can do is push lower 32bit known and
1375 * unknown bits into register (var_off set from jmp logic)
1376 * then learn as much as possible from the 64-bit tnum
1377 * known and unknown bits. The previous smin/smax bounds are
1378 * invalid here because of jmp32 compare so mark them unknown
1379 * so they do not impact tnum bounds calculation.
1381 __mark_reg64_unbounded(reg);
1382 __update_reg_bounds(reg);
1385 /* Intersecting with the old var_off might have improved our bounds
1386 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1387 * then new var_off is (0; 0x7f...fc) which improves our umax.
1389 __reg_deduce_bounds(reg);
1390 __reg_bound_offset(reg);
1391 __update_reg_bounds(reg);
1394 static bool __reg64_bound_s32(s64 a)
1396 return a > S32_MIN && a < S32_MAX;
1399 static bool __reg64_bound_u32(u64 a)
1401 return a > U32_MIN && a < U32_MAX;
1404 static void __reg_combine_64_into_32(struct bpf_reg_state *reg)
1406 __mark_reg32_unbounded(reg);
1408 if (__reg64_bound_s32(reg->smin_value) && __reg64_bound_s32(reg->smax_value)) {
1409 reg->s32_min_value = (s32)reg->smin_value;
1410 reg->s32_max_value = (s32)reg->smax_value;
1412 if (__reg64_bound_u32(reg->umin_value) && __reg64_bound_u32(reg->umax_value)) {
1413 reg->u32_min_value = (u32)reg->umin_value;
1414 reg->u32_max_value = (u32)reg->umax_value;
1417 /* Intersecting with the old var_off might have improved our bounds
1418 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1419 * then new var_off is (0; 0x7f...fc) which improves our umax.
1421 __reg_deduce_bounds(reg);
1422 __reg_bound_offset(reg);
1423 __update_reg_bounds(reg);
1426 /* Mark a register as having a completely unknown (scalar) value. */
1427 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
1428 struct bpf_reg_state *reg)
1431 * Clear type, id, off, and union(map_ptr, range) and
1432 * padding between 'type' and union
1434 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
1435 reg->type = SCALAR_VALUE;
1436 reg->var_off = tnum_unknown;
1438 reg->precise = env->subprog_cnt > 1 || !env->bpf_capable;
1439 __mark_reg_unbounded(reg);
1442 static void mark_reg_unknown(struct bpf_verifier_env *env,
1443 struct bpf_reg_state *regs, u32 regno)
1445 if (WARN_ON(regno >= MAX_BPF_REG)) {
1446 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
1447 /* Something bad happened, let's kill all regs except FP */
1448 for (regno = 0; regno < BPF_REG_FP; regno++)
1449 __mark_reg_not_init(env, regs + regno);
1452 __mark_reg_unknown(env, regs + regno);
1455 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1456 struct bpf_reg_state *reg)
1458 __mark_reg_unknown(env, reg);
1459 reg->type = NOT_INIT;
1462 static void mark_reg_not_init(struct bpf_verifier_env *env,
1463 struct bpf_reg_state *regs, u32 regno)
1465 if (WARN_ON(regno >= MAX_BPF_REG)) {
1466 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
1467 /* Something bad happened, let's kill all regs except FP */
1468 for (regno = 0; regno < BPF_REG_FP; regno++)
1469 __mark_reg_not_init(env, regs + regno);
1472 __mark_reg_not_init(env, regs + regno);
1475 static void mark_btf_ld_reg(struct bpf_verifier_env *env,
1476 struct bpf_reg_state *regs, u32 regno,
1477 enum bpf_reg_type reg_type,
1478 struct btf *btf, u32 btf_id)
1480 if (reg_type == SCALAR_VALUE) {
1481 mark_reg_unknown(env, regs, regno);
1484 mark_reg_known_zero(env, regs, regno);
1485 regs[regno].type = PTR_TO_BTF_ID;
1486 regs[regno].btf = btf;
1487 regs[regno].btf_id = btf_id;
1490 #define DEF_NOT_SUBREG (0)
1491 static void init_reg_state(struct bpf_verifier_env *env,
1492 struct bpf_func_state *state)
1494 struct bpf_reg_state *regs = state->regs;
1497 for (i = 0; i < MAX_BPF_REG; i++) {
1498 mark_reg_not_init(env, regs, i);
1499 regs[i].live = REG_LIVE_NONE;
1500 regs[i].parent = NULL;
1501 regs[i].subreg_def = DEF_NOT_SUBREG;
1505 regs[BPF_REG_FP].type = PTR_TO_STACK;
1506 mark_reg_known_zero(env, regs, BPF_REG_FP);
1507 regs[BPF_REG_FP].frameno = state->frameno;
1510 #define BPF_MAIN_FUNC (-1)
1511 static void init_func_state(struct bpf_verifier_env *env,
1512 struct bpf_func_state *state,
1513 int callsite, int frameno, int subprogno)
1515 state->callsite = callsite;
1516 state->frameno = frameno;
1517 state->subprogno = subprogno;
1518 init_reg_state(env, state);
1522 SRC_OP, /* register is used as source operand */
1523 DST_OP, /* register is used as destination operand */
1524 DST_OP_NO_MARK /* same as above, check only, don't mark */
1527 static int cmp_subprogs(const void *a, const void *b)
1529 return ((struct bpf_subprog_info *)a)->start -
1530 ((struct bpf_subprog_info *)b)->start;
1533 static int find_subprog(struct bpf_verifier_env *env, int off)
1535 struct bpf_subprog_info *p;
1537 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
1538 sizeof(env->subprog_info[0]), cmp_subprogs);
1541 return p - env->subprog_info;
1545 static int add_subprog(struct bpf_verifier_env *env, int off)
1547 int insn_cnt = env->prog->len;
1550 if (off >= insn_cnt || off < 0) {
1551 verbose(env, "call to invalid destination\n");
1554 ret = find_subprog(env, off);
1557 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
1558 verbose(env, "too many subprograms\n");
1561 /* determine subprog starts. The end is one before the next starts */
1562 env->subprog_info[env->subprog_cnt++].start = off;
1563 sort(env->subprog_info, env->subprog_cnt,
1564 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
1565 return env->subprog_cnt - 1;
1568 struct bpf_kfunc_desc {
1569 struct btf_func_model func_model;
1574 #define MAX_KFUNC_DESCS 256
1575 struct bpf_kfunc_desc_tab {
1576 struct bpf_kfunc_desc descs[MAX_KFUNC_DESCS];
1580 static int kfunc_desc_cmp_by_id(const void *a, const void *b)
1582 const struct bpf_kfunc_desc *d0 = a;
1583 const struct bpf_kfunc_desc *d1 = b;
1585 /* func_id is not greater than BTF_MAX_TYPE */
1586 return d0->func_id - d1->func_id;
1589 static const struct bpf_kfunc_desc *
1590 find_kfunc_desc(const struct bpf_prog *prog, u32 func_id)
1592 struct bpf_kfunc_desc desc = {
1595 struct bpf_kfunc_desc_tab *tab;
1597 tab = prog->aux->kfunc_tab;
1598 return bsearch(&desc, tab->descs, tab->nr_descs,
1599 sizeof(tab->descs[0]), kfunc_desc_cmp_by_id);
1602 static int add_kfunc_call(struct bpf_verifier_env *env, u32 func_id)
1604 const struct btf_type *func, *func_proto;
1605 struct bpf_kfunc_desc_tab *tab;
1606 struct bpf_prog_aux *prog_aux;
1607 struct bpf_kfunc_desc *desc;
1608 const char *func_name;
1612 prog_aux = env->prog->aux;
1613 tab = prog_aux->kfunc_tab;
1616 verbose(env, "calling kernel function is not supported without CONFIG_DEBUG_INFO_BTF\n");
1620 if (!env->prog->jit_requested) {
1621 verbose(env, "JIT is required for calling kernel function\n");
1625 if (!bpf_jit_supports_kfunc_call()) {
1626 verbose(env, "JIT does not support calling kernel function\n");
1630 if (!env->prog->gpl_compatible) {
1631 verbose(env, "cannot call kernel function from non-GPL compatible program\n");
1635 tab = kzalloc(sizeof(*tab), GFP_KERNEL);
1638 prog_aux->kfunc_tab = tab;
1641 if (find_kfunc_desc(env->prog, func_id))
1644 if (tab->nr_descs == MAX_KFUNC_DESCS) {
1645 verbose(env, "too many different kernel function calls\n");
1649 func = btf_type_by_id(btf_vmlinux, func_id);
1650 if (!func || !btf_type_is_func(func)) {
1651 verbose(env, "kernel btf_id %u is not a function\n",
1655 func_proto = btf_type_by_id(btf_vmlinux, func->type);
1656 if (!func_proto || !btf_type_is_func_proto(func_proto)) {
1657 verbose(env, "kernel function btf_id %u does not have a valid func_proto\n",
1662 func_name = btf_name_by_offset(btf_vmlinux, func->name_off);
1663 addr = kallsyms_lookup_name(func_name);
1665 verbose(env, "cannot find address for kernel function %s\n",
1670 desc = &tab->descs[tab->nr_descs++];
1671 desc->func_id = func_id;
1672 desc->imm = BPF_CAST_CALL(addr) - __bpf_call_base;
1673 err = btf_distill_func_proto(&env->log, btf_vmlinux,
1674 func_proto, func_name,
1677 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
1678 kfunc_desc_cmp_by_id, NULL);
1682 static int kfunc_desc_cmp_by_imm(const void *a, const void *b)
1684 const struct bpf_kfunc_desc *d0 = a;
1685 const struct bpf_kfunc_desc *d1 = b;
1687 if (d0->imm > d1->imm)
1689 else if (d0->imm < d1->imm)
1694 static void sort_kfunc_descs_by_imm(struct bpf_prog *prog)
1696 struct bpf_kfunc_desc_tab *tab;
1698 tab = prog->aux->kfunc_tab;
1702 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
1703 kfunc_desc_cmp_by_imm, NULL);
1706 bool bpf_prog_has_kfunc_call(const struct bpf_prog *prog)
1708 return !!prog->aux->kfunc_tab;
1711 const struct btf_func_model *
1712 bpf_jit_find_kfunc_model(const struct bpf_prog *prog,
1713 const struct bpf_insn *insn)
1715 const struct bpf_kfunc_desc desc = {
1718 const struct bpf_kfunc_desc *res;
1719 struct bpf_kfunc_desc_tab *tab;
1721 tab = prog->aux->kfunc_tab;
1722 res = bsearch(&desc, tab->descs, tab->nr_descs,
1723 sizeof(tab->descs[0]), kfunc_desc_cmp_by_imm);
1725 return res ? &res->func_model : NULL;
1728 static int add_subprog_and_kfunc(struct bpf_verifier_env *env)
1730 struct bpf_subprog_info *subprog = env->subprog_info;
1731 struct bpf_insn *insn = env->prog->insnsi;
1732 int i, ret, insn_cnt = env->prog->len;
1734 /* Add entry function. */
1735 ret = add_subprog(env, 0);
1739 for (i = 0; i < insn_cnt; i++, insn++) {
1740 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn) &&
1741 !bpf_pseudo_kfunc_call(insn))
1744 if (!env->bpf_capable) {
1745 verbose(env, "loading/calling other bpf or kernel functions are allowed for CAP_BPF and CAP_SYS_ADMIN\n");
1749 if (bpf_pseudo_func(insn)) {
1750 ret = add_subprog(env, i + insn->imm + 1);
1752 /* remember subprog */
1754 } else if (bpf_pseudo_call(insn)) {
1755 ret = add_subprog(env, i + insn->imm + 1);
1757 ret = add_kfunc_call(env, insn->imm);
1764 /* Add a fake 'exit' subprog which could simplify subprog iteration
1765 * logic. 'subprog_cnt' should not be increased.
1767 subprog[env->subprog_cnt].start = insn_cnt;
1769 if (env->log.level & BPF_LOG_LEVEL2)
1770 for (i = 0; i < env->subprog_cnt; i++)
1771 verbose(env, "func#%d @%d\n", i, subprog[i].start);
1776 static int check_subprogs(struct bpf_verifier_env *env)
1778 int i, subprog_start, subprog_end, off, cur_subprog = 0;
1779 struct bpf_subprog_info *subprog = env->subprog_info;
1780 struct bpf_insn *insn = env->prog->insnsi;
1781 int insn_cnt = env->prog->len;
1783 /* now check that all jumps are within the same subprog */
1784 subprog_start = subprog[cur_subprog].start;
1785 subprog_end = subprog[cur_subprog + 1].start;
1786 for (i = 0; i < insn_cnt; i++) {
1787 u8 code = insn[i].code;
1789 if (code == (BPF_JMP | BPF_CALL) &&
1790 insn[i].imm == BPF_FUNC_tail_call &&
1791 insn[i].src_reg != BPF_PSEUDO_CALL)
1792 subprog[cur_subprog].has_tail_call = true;
1793 if (BPF_CLASS(code) == BPF_LD &&
1794 (BPF_MODE(code) == BPF_ABS || BPF_MODE(code) == BPF_IND))
1795 subprog[cur_subprog].has_ld_abs = true;
1796 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
1798 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
1800 off = i + insn[i].off + 1;
1801 if (off < subprog_start || off >= subprog_end) {
1802 verbose(env, "jump out of range from insn %d to %d\n", i, off);
1806 if (i == subprog_end - 1) {
1807 /* to avoid fall-through from one subprog into another
1808 * the last insn of the subprog should be either exit
1809 * or unconditional jump back
1811 if (code != (BPF_JMP | BPF_EXIT) &&
1812 code != (BPF_JMP | BPF_JA)) {
1813 verbose(env, "last insn is not an exit or jmp\n");
1816 subprog_start = subprog_end;
1818 if (cur_subprog < env->subprog_cnt)
1819 subprog_end = subprog[cur_subprog + 1].start;
1825 /* Parentage chain of this register (or stack slot) should take care of all
1826 * issues like callee-saved registers, stack slot allocation time, etc.
1828 static int mark_reg_read(struct bpf_verifier_env *env,
1829 const struct bpf_reg_state *state,
1830 struct bpf_reg_state *parent, u8 flag)
1832 bool writes = parent == state->parent; /* Observe write marks */
1836 /* if read wasn't screened by an earlier write ... */
1837 if (writes && state->live & REG_LIVE_WRITTEN)
1839 if (parent->live & REG_LIVE_DONE) {
1840 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
1841 reg_type_str[parent->type],
1842 parent->var_off.value, parent->off);
1845 /* The first condition is more likely to be true than the
1846 * second, checked it first.
1848 if ((parent->live & REG_LIVE_READ) == flag ||
1849 parent->live & REG_LIVE_READ64)
1850 /* The parentage chain never changes and
1851 * this parent was already marked as LIVE_READ.
1852 * There is no need to keep walking the chain again and
1853 * keep re-marking all parents as LIVE_READ.
1854 * This case happens when the same register is read
1855 * multiple times without writes into it in-between.
1856 * Also, if parent has the stronger REG_LIVE_READ64 set,
1857 * then no need to set the weak REG_LIVE_READ32.
1860 /* ... then we depend on parent's value */
1861 parent->live |= flag;
1862 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
1863 if (flag == REG_LIVE_READ64)
1864 parent->live &= ~REG_LIVE_READ32;
1866 parent = state->parent;
1871 if (env->longest_mark_read_walk < cnt)
1872 env->longest_mark_read_walk = cnt;
1876 /* This function is supposed to be used by the following 32-bit optimization
1877 * code only. It returns TRUE if the source or destination register operates
1878 * on 64-bit, otherwise return FALSE.
1880 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
1881 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
1886 class = BPF_CLASS(code);
1888 if (class == BPF_JMP) {
1889 /* BPF_EXIT for "main" will reach here. Return TRUE
1894 if (op == BPF_CALL) {
1895 /* BPF to BPF call will reach here because of marking
1896 * caller saved clobber with DST_OP_NO_MARK for which we
1897 * don't care the register def because they are anyway
1898 * marked as NOT_INIT already.
1900 if (insn->src_reg == BPF_PSEUDO_CALL)
1902 /* Helper call will reach here because of arg type
1903 * check, conservatively return TRUE.
1912 if (class == BPF_ALU64 || class == BPF_JMP ||
1913 /* BPF_END always use BPF_ALU class. */
1914 (class == BPF_ALU && op == BPF_END && insn->imm == 64))
1917 if (class == BPF_ALU || class == BPF_JMP32)
1920 if (class == BPF_LDX) {
1922 return BPF_SIZE(code) == BPF_DW;
1923 /* LDX source must be ptr. */
1927 if (class == BPF_STX) {
1928 /* BPF_STX (including atomic variants) has multiple source
1929 * operands, one of which is a ptr. Check whether the caller is
1932 if (t == SRC_OP && reg->type != SCALAR_VALUE)
1934 return BPF_SIZE(code) == BPF_DW;
1937 if (class == BPF_LD) {
1938 u8 mode = BPF_MODE(code);
1941 if (mode == BPF_IMM)
1944 /* Both LD_IND and LD_ABS return 32-bit data. */
1948 /* Implicit ctx ptr. */
1949 if (regno == BPF_REG_6)
1952 /* Explicit source could be any width. */
1956 if (class == BPF_ST)
1957 /* The only source register for BPF_ST is a ptr. */
1960 /* Conservatively return true at default. */
1964 /* Return the regno defined by the insn, or -1. */
1965 static int insn_def_regno(const struct bpf_insn *insn)
1967 switch (BPF_CLASS(insn->code)) {
1973 if (BPF_MODE(insn->code) == BPF_ATOMIC &&
1974 (insn->imm & BPF_FETCH)) {
1975 if (insn->imm == BPF_CMPXCHG)
1978 return insn->src_reg;
1983 return insn->dst_reg;
1987 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
1988 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
1990 int dst_reg = insn_def_regno(insn);
1995 return !is_reg64(env, insn, dst_reg, NULL, DST_OP);
1998 static void mark_insn_zext(struct bpf_verifier_env *env,
1999 struct bpf_reg_state *reg)
2001 s32 def_idx = reg->subreg_def;
2003 if (def_idx == DEF_NOT_SUBREG)
2006 env->insn_aux_data[def_idx - 1].zext_dst = true;
2007 /* The dst will be zero extended, so won't be sub-register anymore. */
2008 reg->subreg_def = DEF_NOT_SUBREG;
2011 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
2012 enum reg_arg_type t)
2014 struct bpf_verifier_state *vstate = env->cur_state;
2015 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2016 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
2017 struct bpf_reg_state *reg, *regs = state->regs;
2020 if (regno >= MAX_BPF_REG) {
2021 verbose(env, "R%d is invalid\n", regno);
2026 rw64 = is_reg64(env, insn, regno, reg, t);
2028 /* check whether register used as source operand can be read */
2029 if (reg->type == NOT_INIT) {
2030 verbose(env, "R%d !read_ok\n", regno);
2033 /* We don't need to worry about FP liveness because it's read-only */
2034 if (regno == BPF_REG_FP)
2038 mark_insn_zext(env, reg);
2040 return mark_reg_read(env, reg, reg->parent,
2041 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
2043 /* check whether register used as dest operand can be written to */
2044 if (regno == BPF_REG_FP) {
2045 verbose(env, "frame pointer is read only\n");
2048 reg->live |= REG_LIVE_WRITTEN;
2049 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
2051 mark_reg_unknown(env, regs, regno);
2056 /* for any branch, call, exit record the history of jmps in the given state */
2057 static int push_jmp_history(struct bpf_verifier_env *env,
2058 struct bpf_verifier_state *cur)
2060 u32 cnt = cur->jmp_history_cnt;
2061 struct bpf_idx_pair *p;
2064 p = krealloc(cur->jmp_history, cnt * sizeof(*p), GFP_USER);
2067 p[cnt - 1].idx = env->insn_idx;
2068 p[cnt - 1].prev_idx = env->prev_insn_idx;
2069 cur->jmp_history = p;
2070 cur->jmp_history_cnt = cnt;
2074 /* Backtrack one insn at a time. If idx is not at the top of recorded
2075 * history then previous instruction came from straight line execution.
2077 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
2082 if (cnt && st->jmp_history[cnt - 1].idx == i) {
2083 i = st->jmp_history[cnt - 1].prev_idx;
2091 static const char *disasm_kfunc_name(void *data, const struct bpf_insn *insn)
2093 const struct btf_type *func;
2095 if (insn->src_reg != BPF_PSEUDO_KFUNC_CALL)
2098 func = btf_type_by_id(btf_vmlinux, insn->imm);
2099 return btf_name_by_offset(btf_vmlinux, func->name_off);
2102 /* For given verifier state backtrack_insn() is called from the last insn to
2103 * the first insn. Its purpose is to compute a bitmask of registers and
2104 * stack slots that needs precision in the parent verifier state.
2106 static int backtrack_insn(struct bpf_verifier_env *env, int idx,
2107 u32 *reg_mask, u64 *stack_mask)
2109 const struct bpf_insn_cbs cbs = {
2110 .cb_call = disasm_kfunc_name,
2111 .cb_print = verbose,
2112 .private_data = env,
2114 struct bpf_insn *insn = env->prog->insnsi + idx;
2115 u8 class = BPF_CLASS(insn->code);
2116 u8 opcode = BPF_OP(insn->code);
2117 u8 mode = BPF_MODE(insn->code);
2118 u32 dreg = 1u << insn->dst_reg;
2119 u32 sreg = 1u << insn->src_reg;
2122 if (insn->code == 0)
2124 if (env->log.level & BPF_LOG_LEVEL) {
2125 verbose(env, "regs=%x stack=%llx before ", *reg_mask, *stack_mask);
2126 verbose(env, "%d: ", idx);
2127 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
2130 if (class == BPF_ALU || class == BPF_ALU64) {
2131 if (!(*reg_mask & dreg))
2133 if (opcode == BPF_MOV) {
2134 if (BPF_SRC(insn->code) == BPF_X) {
2136 * dreg needs precision after this insn
2137 * sreg needs precision before this insn
2143 * dreg needs precision after this insn.
2144 * Corresponding register is already marked
2145 * as precise=true in this verifier state.
2146 * No further markings in parent are necessary
2151 if (BPF_SRC(insn->code) == BPF_X) {
2153 * both dreg and sreg need precision
2158 * dreg still needs precision before this insn
2161 } else if (class == BPF_LDX) {
2162 if (!(*reg_mask & dreg))
2166 /* scalars can only be spilled into stack w/o losing precision.
2167 * Load from any other memory can be zero extended.
2168 * The desire to keep that precision is already indicated
2169 * by 'precise' mark in corresponding register of this state.
2170 * No further tracking necessary.
2172 if (insn->src_reg != BPF_REG_FP)
2174 if (BPF_SIZE(insn->code) != BPF_DW)
2177 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
2178 * that [fp - off] slot contains scalar that needs to be
2179 * tracked with precision
2181 spi = (-insn->off - 1) / BPF_REG_SIZE;
2183 verbose(env, "BUG spi %d\n", spi);
2184 WARN_ONCE(1, "verifier backtracking bug");
2187 *stack_mask |= 1ull << spi;
2188 } else if (class == BPF_STX || class == BPF_ST) {
2189 if (*reg_mask & dreg)
2190 /* stx & st shouldn't be using _scalar_ dst_reg
2191 * to access memory. It means backtracking
2192 * encountered a case of pointer subtraction.
2195 /* scalars can only be spilled into stack */
2196 if (insn->dst_reg != BPF_REG_FP)
2198 if (BPF_SIZE(insn->code) != BPF_DW)
2200 spi = (-insn->off - 1) / BPF_REG_SIZE;
2202 verbose(env, "BUG spi %d\n", spi);
2203 WARN_ONCE(1, "verifier backtracking bug");
2206 if (!(*stack_mask & (1ull << spi)))
2208 *stack_mask &= ~(1ull << spi);
2209 if (class == BPF_STX)
2211 } else if (class == BPF_JMP || class == BPF_JMP32) {
2212 if (opcode == BPF_CALL) {
2213 if (insn->src_reg == BPF_PSEUDO_CALL)
2215 /* regular helper call sets R0 */
2217 if (*reg_mask & 0x3f) {
2218 /* if backtracing was looking for registers R1-R5
2219 * they should have been found already.
2221 verbose(env, "BUG regs %x\n", *reg_mask);
2222 WARN_ONCE(1, "verifier backtracking bug");
2225 } else if (opcode == BPF_EXIT) {
2228 } else if (class == BPF_LD) {
2229 if (!(*reg_mask & dreg))
2232 /* It's ld_imm64 or ld_abs or ld_ind.
2233 * For ld_imm64 no further tracking of precision
2234 * into parent is necessary
2236 if (mode == BPF_IND || mode == BPF_ABS)
2237 /* to be analyzed */
2243 /* the scalar precision tracking algorithm:
2244 * . at the start all registers have precise=false.
2245 * . scalar ranges are tracked as normal through alu and jmp insns.
2246 * . once precise value of the scalar register is used in:
2247 * . ptr + scalar alu
2248 * . if (scalar cond K|scalar)
2249 * . helper_call(.., scalar, ...) where ARG_CONST is expected
2250 * backtrack through the verifier states and mark all registers and
2251 * stack slots with spilled constants that these scalar regisers
2252 * should be precise.
2253 * . during state pruning two registers (or spilled stack slots)
2254 * are equivalent if both are not precise.
2256 * Note the verifier cannot simply walk register parentage chain,
2257 * since many different registers and stack slots could have been
2258 * used to compute single precise scalar.
2260 * The approach of starting with precise=true for all registers and then
2261 * backtrack to mark a register as not precise when the verifier detects
2262 * that program doesn't care about specific value (e.g., when helper
2263 * takes register as ARG_ANYTHING parameter) is not safe.
2265 * It's ok to walk single parentage chain of the verifier states.
2266 * It's possible that this backtracking will go all the way till 1st insn.
2267 * All other branches will be explored for needing precision later.
2269 * The backtracking needs to deal with cases like:
2270 * 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)
2273 * if r5 > 0x79f goto pc+7
2274 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
2277 * call bpf_perf_event_output#25
2278 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
2282 * call foo // uses callee's r6 inside to compute r0
2286 * to track above reg_mask/stack_mask needs to be independent for each frame.
2288 * Also if parent's curframe > frame where backtracking started,
2289 * the verifier need to mark registers in both frames, otherwise callees
2290 * may incorrectly prune callers. This is similar to
2291 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
2293 * For now backtracking falls back into conservative marking.
2295 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
2296 struct bpf_verifier_state *st)
2298 struct bpf_func_state *func;
2299 struct bpf_reg_state *reg;
2302 /* big hammer: mark all scalars precise in this path.
2303 * pop_stack may still get !precise scalars.
2305 for (; st; st = st->parent)
2306 for (i = 0; i <= st->curframe; i++) {
2307 func = st->frame[i];
2308 for (j = 0; j < BPF_REG_FP; j++) {
2309 reg = &func->regs[j];
2310 if (reg->type != SCALAR_VALUE)
2312 reg->precise = true;
2314 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
2315 if (func->stack[j].slot_type[0] != STACK_SPILL)
2317 reg = &func->stack[j].spilled_ptr;
2318 if (reg->type != SCALAR_VALUE)
2320 reg->precise = true;
2325 static int __mark_chain_precision(struct bpf_verifier_env *env, int regno,
2328 struct bpf_verifier_state *st = env->cur_state;
2329 int first_idx = st->first_insn_idx;
2330 int last_idx = env->insn_idx;
2331 struct bpf_func_state *func;
2332 struct bpf_reg_state *reg;
2333 u32 reg_mask = regno >= 0 ? 1u << regno : 0;
2334 u64 stack_mask = spi >= 0 ? 1ull << spi : 0;
2335 bool skip_first = true;
2336 bool new_marks = false;
2339 if (!env->bpf_capable)
2342 func = st->frame[st->curframe];
2344 reg = &func->regs[regno];
2345 if (reg->type != SCALAR_VALUE) {
2346 WARN_ONCE(1, "backtracing misuse");
2353 reg->precise = true;
2357 if (func->stack[spi].slot_type[0] != STACK_SPILL) {
2361 reg = &func->stack[spi].spilled_ptr;
2362 if (reg->type != SCALAR_VALUE) {
2370 reg->precise = true;
2376 if (!reg_mask && !stack_mask)
2379 DECLARE_BITMAP(mask, 64);
2380 u32 history = st->jmp_history_cnt;
2382 if (env->log.level & BPF_LOG_LEVEL)
2383 verbose(env, "last_idx %d first_idx %d\n", last_idx, first_idx);
2384 for (i = last_idx;;) {
2389 err = backtrack_insn(env, i, ®_mask, &stack_mask);
2391 if (err == -ENOTSUPP) {
2392 mark_all_scalars_precise(env, st);
2397 if (!reg_mask && !stack_mask)
2398 /* Found assignment(s) into tracked register in this state.
2399 * Since this state is already marked, just return.
2400 * Nothing to be tracked further in the parent state.
2405 i = get_prev_insn_idx(st, i, &history);
2406 if (i >= env->prog->len) {
2407 /* This can happen if backtracking reached insn 0
2408 * and there are still reg_mask or stack_mask
2410 * It means the backtracking missed the spot where
2411 * particular register was initialized with a constant.
2413 verbose(env, "BUG backtracking idx %d\n", i);
2414 WARN_ONCE(1, "verifier backtracking bug");
2423 func = st->frame[st->curframe];
2424 bitmap_from_u64(mask, reg_mask);
2425 for_each_set_bit(i, mask, 32) {
2426 reg = &func->regs[i];
2427 if (reg->type != SCALAR_VALUE) {
2428 reg_mask &= ~(1u << i);
2433 reg->precise = true;
2436 bitmap_from_u64(mask, stack_mask);
2437 for_each_set_bit(i, mask, 64) {
2438 if (i >= func->allocated_stack / BPF_REG_SIZE) {
2439 /* the sequence of instructions:
2441 * 3: (7b) *(u64 *)(r3 -8) = r0
2442 * 4: (79) r4 = *(u64 *)(r10 -8)
2443 * doesn't contain jmps. It's backtracked
2444 * as a single block.
2445 * During backtracking insn 3 is not recognized as
2446 * stack access, so at the end of backtracking
2447 * stack slot fp-8 is still marked in stack_mask.
2448 * However the parent state may not have accessed
2449 * fp-8 and it's "unallocated" stack space.
2450 * In such case fallback to conservative.
2452 mark_all_scalars_precise(env, st);
2456 if (func->stack[i].slot_type[0] != STACK_SPILL) {
2457 stack_mask &= ~(1ull << i);
2460 reg = &func->stack[i].spilled_ptr;
2461 if (reg->type != SCALAR_VALUE) {
2462 stack_mask &= ~(1ull << i);
2467 reg->precise = true;
2469 if (env->log.level & BPF_LOG_LEVEL) {
2470 print_verifier_state(env, func);
2471 verbose(env, "parent %s regs=%x stack=%llx marks\n",
2472 new_marks ? "didn't have" : "already had",
2473 reg_mask, stack_mask);
2476 if (!reg_mask && !stack_mask)
2481 last_idx = st->last_insn_idx;
2482 first_idx = st->first_insn_idx;
2487 static int mark_chain_precision(struct bpf_verifier_env *env, int regno)
2489 return __mark_chain_precision(env, regno, -1);
2492 static int mark_chain_precision_stack(struct bpf_verifier_env *env, int spi)
2494 return __mark_chain_precision(env, -1, spi);
2497 static bool is_spillable_regtype(enum bpf_reg_type type)
2500 case PTR_TO_MAP_VALUE:
2501 case PTR_TO_MAP_VALUE_OR_NULL:
2505 case PTR_TO_PACKET_META:
2506 case PTR_TO_PACKET_END:
2507 case PTR_TO_FLOW_KEYS:
2508 case CONST_PTR_TO_MAP:
2510 case PTR_TO_SOCKET_OR_NULL:
2511 case PTR_TO_SOCK_COMMON:
2512 case PTR_TO_SOCK_COMMON_OR_NULL:
2513 case PTR_TO_TCP_SOCK:
2514 case PTR_TO_TCP_SOCK_OR_NULL:
2515 case PTR_TO_XDP_SOCK:
2517 case PTR_TO_BTF_ID_OR_NULL:
2518 case PTR_TO_RDONLY_BUF:
2519 case PTR_TO_RDONLY_BUF_OR_NULL:
2520 case PTR_TO_RDWR_BUF:
2521 case PTR_TO_RDWR_BUF_OR_NULL:
2522 case PTR_TO_PERCPU_BTF_ID:
2524 case PTR_TO_MEM_OR_NULL:
2526 case PTR_TO_MAP_KEY:
2533 /* Does this register contain a constant zero? */
2534 static bool register_is_null(struct bpf_reg_state *reg)
2536 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
2539 static bool register_is_const(struct bpf_reg_state *reg)
2541 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
2544 static bool __is_scalar_unbounded(struct bpf_reg_state *reg)
2546 return tnum_is_unknown(reg->var_off) &&
2547 reg->smin_value == S64_MIN && reg->smax_value == S64_MAX &&
2548 reg->umin_value == 0 && reg->umax_value == U64_MAX &&
2549 reg->s32_min_value == S32_MIN && reg->s32_max_value == S32_MAX &&
2550 reg->u32_min_value == 0 && reg->u32_max_value == U32_MAX;
2553 static bool register_is_bounded(struct bpf_reg_state *reg)
2555 return reg->type == SCALAR_VALUE && !__is_scalar_unbounded(reg);
2558 static bool __is_pointer_value(bool allow_ptr_leaks,
2559 const struct bpf_reg_state *reg)
2561 if (allow_ptr_leaks)
2564 return reg->type != SCALAR_VALUE;
2567 static void save_register_state(struct bpf_func_state *state,
2568 int spi, struct bpf_reg_state *reg)
2572 state->stack[spi].spilled_ptr = *reg;
2573 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2575 for (i = 0; i < BPF_REG_SIZE; i++)
2576 state->stack[spi].slot_type[i] = STACK_SPILL;
2579 /* check_stack_{read,write}_fixed_off functions track spill/fill of registers,
2580 * stack boundary and alignment are checked in check_mem_access()
2582 static int check_stack_write_fixed_off(struct bpf_verifier_env *env,
2583 /* stack frame we're writing to */
2584 struct bpf_func_state *state,
2585 int off, int size, int value_regno,
2588 struct bpf_func_state *cur; /* state of the current function */
2589 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
2590 u32 dst_reg = env->prog->insnsi[insn_idx].dst_reg;
2591 struct bpf_reg_state *reg = NULL;
2593 err = realloc_func_state(state, round_up(slot + 1, BPF_REG_SIZE),
2594 state->acquired_refs, true);
2597 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
2598 * so it's aligned access and [off, off + size) are within stack limits
2600 if (!env->allow_ptr_leaks &&
2601 state->stack[spi].slot_type[0] == STACK_SPILL &&
2602 size != BPF_REG_SIZE) {
2603 verbose(env, "attempt to corrupt spilled pointer on stack\n");
2607 cur = env->cur_state->frame[env->cur_state->curframe];
2608 if (value_regno >= 0)
2609 reg = &cur->regs[value_regno];
2610 if (!env->bypass_spec_v4) {
2611 bool sanitize = reg && is_spillable_regtype(reg->type);
2613 for (i = 0; i < size; i++) {
2614 if (state->stack[spi].slot_type[i] == STACK_INVALID) {
2621 env->insn_aux_data[insn_idx].sanitize_stack_spill = true;
2624 if (reg && size == BPF_REG_SIZE && register_is_bounded(reg) &&
2625 !register_is_null(reg) && env->bpf_capable) {
2626 if (dst_reg != BPF_REG_FP) {
2627 /* The backtracking logic can only recognize explicit
2628 * stack slot address like [fp - 8]. Other spill of
2629 * scalar via different register has to be conervative.
2630 * Backtrack from here and mark all registers as precise
2631 * that contributed into 'reg' being a constant.
2633 err = mark_chain_precision(env, value_regno);
2637 save_register_state(state, spi, reg);
2638 } else if (reg && is_spillable_regtype(reg->type)) {
2639 /* register containing pointer is being spilled into stack */
2640 if (size != BPF_REG_SIZE) {
2641 verbose_linfo(env, insn_idx, "; ");
2642 verbose(env, "invalid size of register spill\n");
2645 if (state != cur && reg->type == PTR_TO_STACK) {
2646 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
2649 save_register_state(state, spi, reg);
2651 u8 type = STACK_MISC;
2653 /* regular write of data into stack destroys any spilled ptr */
2654 state->stack[spi].spilled_ptr.type = NOT_INIT;
2655 /* Mark slots as STACK_MISC if they belonged to spilled ptr. */
2656 if (state->stack[spi].slot_type[0] == STACK_SPILL)
2657 for (i = 0; i < BPF_REG_SIZE; i++)
2658 state->stack[spi].slot_type[i] = STACK_MISC;
2660 /* only mark the slot as written if all 8 bytes were written
2661 * otherwise read propagation may incorrectly stop too soon
2662 * when stack slots are partially written.
2663 * This heuristic means that read propagation will be
2664 * conservative, since it will add reg_live_read marks
2665 * to stack slots all the way to first state when programs
2666 * writes+reads less than 8 bytes
2668 if (size == BPF_REG_SIZE)
2669 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2671 /* when we zero initialize stack slots mark them as such */
2672 if (reg && register_is_null(reg)) {
2673 /* backtracking doesn't work for STACK_ZERO yet. */
2674 err = mark_chain_precision(env, value_regno);
2680 /* Mark slots affected by this stack write. */
2681 for (i = 0; i < size; i++)
2682 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
2688 /* Write the stack: 'stack[ptr_regno + off] = value_regno'. 'ptr_regno' is
2689 * known to contain a variable offset.
2690 * This function checks whether the write is permitted and conservatively
2691 * tracks the effects of the write, considering that each stack slot in the
2692 * dynamic range is potentially written to.
2694 * 'off' includes 'regno->off'.
2695 * 'value_regno' can be -1, meaning that an unknown value is being written to
2698 * Spilled pointers in range are not marked as written because we don't know
2699 * what's going to be actually written. This means that read propagation for
2700 * future reads cannot be terminated by this write.
2702 * For privileged programs, uninitialized stack slots are considered
2703 * initialized by this write (even though we don't know exactly what offsets
2704 * are going to be written to). The idea is that we don't want the verifier to
2705 * reject future reads that access slots written to through variable offsets.
2707 static int check_stack_write_var_off(struct bpf_verifier_env *env,
2708 /* func where register points to */
2709 struct bpf_func_state *state,
2710 int ptr_regno, int off, int size,
2711 int value_regno, int insn_idx)
2713 struct bpf_func_state *cur; /* state of the current function */
2714 int min_off, max_off;
2716 struct bpf_reg_state *ptr_reg = NULL, *value_reg = NULL;
2717 bool writing_zero = false;
2718 /* set if the fact that we're writing a zero is used to let any
2719 * stack slots remain STACK_ZERO
2721 bool zero_used = false;
2723 cur = env->cur_state->frame[env->cur_state->curframe];
2724 ptr_reg = &cur->regs[ptr_regno];
2725 min_off = ptr_reg->smin_value + off;
2726 max_off = ptr_reg->smax_value + off + size;
2727 if (value_regno >= 0)
2728 value_reg = &cur->regs[value_regno];
2729 if (value_reg && register_is_null(value_reg))
2730 writing_zero = true;
2732 err = realloc_func_state(state, round_up(-min_off, BPF_REG_SIZE),
2733 state->acquired_refs, true);
2738 /* Variable offset writes destroy any spilled pointers in range. */
2739 for (i = min_off; i < max_off; i++) {
2740 u8 new_type, *stype;
2744 spi = slot / BPF_REG_SIZE;
2745 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
2747 if (!env->allow_ptr_leaks
2748 && *stype != NOT_INIT
2749 && *stype != SCALAR_VALUE) {
2750 /* Reject the write if there's are spilled pointers in
2751 * range. If we didn't reject here, the ptr status
2752 * would be erased below (even though not all slots are
2753 * actually overwritten), possibly opening the door to
2756 verbose(env, "spilled ptr in range of var-offset stack write; insn %d, ptr off: %d",
2761 /* Erase all spilled pointers. */
2762 state->stack[spi].spilled_ptr.type = NOT_INIT;
2764 /* Update the slot type. */
2765 new_type = STACK_MISC;
2766 if (writing_zero && *stype == STACK_ZERO) {
2767 new_type = STACK_ZERO;
2770 /* If the slot is STACK_INVALID, we check whether it's OK to
2771 * pretend that it will be initialized by this write. The slot
2772 * might not actually be written to, and so if we mark it as
2773 * initialized future reads might leak uninitialized memory.
2774 * For privileged programs, we will accept such reads to slots
2775 * that may or may not be written because, if we're reject
2776 * them, the error would be too confusing.
2778 if (*stype == STACK_INVALID && !env->allow_uninit_stack) {
2779 verbose(env, "uninit stack in range of var-offset write prohibited for !root; insn %d, off: %d",
2786 /* backtracking doesn't work for STACK_ZERO yet. */
2787 err = mark_chain_precision(env, value_regno);
2794 /* When register 'dst_regno' is assigned some values from stack[min_off,
2795 * max_off), we set the register's type according to the types of the
2796 * respective stack slots. If all the stack values are known to be zeros, then
2797 * so is the destination reg. Otherwise, the register is considered to be
2798 * SCALAR. This function does not deal with register filling; the caller must
2799 * ensure that all spilled registers in the stack range have been marked as
2802 static void mark_reg_stack_read(struct bpf_verifier_env *env,
2803 /* func where src register points to */
2804 struct bpf_func_state *ptr_state,
2805 int min_off, int max_off, int dst_regno)
2807 struct bpf_verifier_state *vstate = env->cur_state;
2808 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2813 for (i = min_off; i < max_off; i++) {
2815 spi = slot / BPF_REG_SIZE;
2816 stype = ptr_state->stack[spi].slot_type;
2817 if (stype[slot % BPF_REG_SIZE] != STACK_ZERO)
2821 if (zeros == max_off - min_off) {
2822 /* any access_size read into register is zero extended,
2823 * so the whole register == const_zero
2825 __mark_reg_const_zero(&state->regs[dst_regno]);
2826 /* backtracking doesn't support STACK_ZERO yet,
2827 * so mark it precise here, so that later
2828 * backtracking can stop here.
2829 * Backtracking may not need this if this register
2830 * doesn't participate in pointer adjustment.
2831 * Forward propagation of precise flag is not
2832 * necessary either. This mark is only to stop
2833 * backtracking. Any register that contributed
2834 * to const 0 was marked precise before spill.
2836 state->regs[dst_regno].precise = true;
2838 /* have read misc data from the stack */
2839 mark_reg_unknown(env, state->regs, dst_regno);
2841 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
2844 /* Read the stack at 'off' and put the results into the register indicated by
2845 * 'dst_regno'. It handles reg filling if the addressed stack slot is a
2848 * 'dst_regno' can be -1, meaning that the read value is not going to a
2851 * The access is assumed to be within the current stack bounds.
2853 static int check_stack_read_fixed_off(struct bpf_verifier_env *env,
2854 /* func where src register points to */
2855 struct bpf_func_state *reg_state,
2856 int off, int size, int dst_regno)
2858 struct bpf_verifier_state *vstate = env->cur_state;
2859 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2860 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
2861 struct bpf_reg_state *reg;
2864 stype = reg_state->stack[spi].slot_type;
2865 reg = ®_state->stack[spi].spilled_ptr;
2867 if (stype[0] == STACK_SPILL) {
2868 if (size != BPF_REG_SIZE) {
2869 if (reg->type != SCALAR_VALUE) {
2870 verbose_linfo(env, env->insn_idx, "; ");
2871 verbose(env, "invalid size of register fill\n");
2874 if (dst_regno >= 0) {
2875 mark_reg_unknown(env, state->regs, dst_regno);
2876 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
2878 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2881 for (i = 1; i < BPF_REG_SIZE; i++) {
2882 if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) {
2883 verbose(env, "corrupted spill memory\n");
2888 if (dst_regno >= 0) {
2889 /* restore register state from stack */
2890 state->regs[dst_regno] = *reg;
2891 /* mark reg as written since spilled pointer state likely
2892 * has its liveness marks cleared by is_state_visited()
2893 * which resets stack/reg liveness for state transitions
2895 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
2896 } else if (__is_pointer_value(env->allow_ptr_leaks, reg)) {
2897 /* If dst_regno==-1, the caller is asking us whether
2898 * it is acceptable to use this value as a SCALAR_VALUE
2900 * We must not allow unprivileged callers to do that
2901 * with spilled pointers.
2903 verbose(env, "leaking pointer from stack off %d\n",
2907 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2911 for (i = 0; i < size; i++) {
2912 type = stype[(slot - i) % BPF_REG_SIZE];
2913 if (type == STACK_MISC)
2915 if (type == STACK_ZERO)
2917 verbose(env, "invalid read from stack off %d+%d size %d\n",
2921 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2923 mark_reg_stack_read(env, reg_state, off, off + size, dst_regno);
2928 enum stack_access_src {
2929 ACCESS_DIRECT = 1, /* the access is performed by an instruction */
2930 ACCESS_HELPER = 2, /* the access is performed by a helper */
2933 static int check_stack_range_initialized(struct bpf_verifier_env *env,
2934 int regno, int off, int access_size,
2935 bool zero_size_allowed,
2936 enum stack_access_src type,
2937 struct bpf_call_arg_meta *meta);
2939 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
2941 return cur_regs(env) + regno;
2944 /* Read the stack at 'ptr_regno + off' and put the result into the register
2946 * 'off' includes the pointer register's fixed offset(i.e. 'ptr_regno.off'),
2947 * but not its variable offset.
2948 * 'size' is assumed to be <= reg size and the access is assumed to be aligned.
2950 * As opposed to check_stack_read_fixed_off, this function doesn't deal with
2951 * filling registers (i.e. reads of spilled register cannot be detected when
2952 * the offset is not fixed). We conservatively mark 'dst_regno' as containing
2953 * SCALAR_VALUE. That's why we assert that the 'ptr_regno' has a variable
2954 * offset; for a fixed offset check_stack_read_fixed_off should be used
2957 static int check_stack_read_var_off(struct bpf_verifier_env *env,
2958 int ptr_regno, int off, int size, int dst_regno)
2960 /* The state of the source register. */
2961 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
2962 struct bpf_func_state *ptr_state = func(env, reg);
2964 int min_off, max_off;
2966 /* Note that we pass a NULL meta, so raw access will not be permitted.
2968 err = check_stack_range_initialized(env, ptr_regno, off, size,
2969 false, ACCESS_DIRECT, NULL);
2973 min_off = reg->smin_value + off;
2974 max_off = reg->smax_value + off;
2975 mark_reg_stack_read(env, ptr_state, min_off, max_off + size, dst_regno);
2979 /* check_stack_read dispatches to check_stack_read_fixed_off or
2980 * check_stack_read_var_off.
2982 * The caller must ensure that the offset falls within the allocated stack
2985 * 'dst_regno' is a register which will receive the value from the stack. It
2986 * can be -1, meaning that the read value is not going to a register.
2988 static int check_stack_read(struct bpf_verifier_env *env,
2989 int ptr_regno, int off, int size,
2992 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
2993 struct bpf_func_state *state = func(env, reg);
2995 /* Some accesses are only permitted with a static offset. */
2996 bool var_off = !tnum_is_const(reg->var_off);
2998 /* The offset is required to be static when reads don't go to a
2999 * register, in order to not leak pointers (see
3000 * check_stack_read_fixed_off).
3002 if (dst_regno < 0 && var_off) {
3005 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3006 verbose(env, "variable offset stack pointer cannot be passed into helper function; var_off=%s off=%d size=%d\n",
3010 /* Variable offset is prohibited for unprivileged mode for simplicity
3011 * since it requires corresponding support in Spectre masking for stack
3012 * ALU. See also retrieve_ptr_limit().
3014 if (!env->bypass_spec_v1 && var_off) {
3017 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3018 verbose(env, "R%d variable offset stack access prohibited for !root, var_off=%s\n",
3024 off += reg->var_off.value;
3025 err = check_stack_read_fixed_off(env, state, off, size,
3028 /* Variable offset stack reads need more conservative handling
3029 * than fixed offset ones. Note that dst_regno >= 0 on this
3032 err = check_stack_read_var_off(env, ptr_regno, off, size,
3039 /* check_stack_write dispatches to check_stack_write_fixed_off or
3040 * check_stack_write_var_off.
3042 * 'ptr_regno' is the register used as a pointer into the stack.
3043 * 'off' includes 'ptr_regno->off', but not its variable offset (if any).
3044 * 'value_regno' is the register whose value we're writing to the stack. It can
3045 * be -1, meaning that we're not writing from a register.
3047 * The caller must ensure that the offset falls within the maximum stack size.
3049 static int check_stack_write(struct bpf_verifier_env *env,
3050 int ptr_regno, int off, int size,
3051 int value_regno, int insn_idx)
3053 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
3054 struct bpf_func_state *state = func(env, reg);
3057 if (tnum_is_const(reg->var_off)) {
3058 off += reg->var_off.value;
3059 err = check_stack_write_fixed_off(env, state, off, size,
3060 value_regno, insn_idx);
3062 /* Variable offset stack reads need more conservative handling
3063 * than fixed offset ones.
3065 err = check_stack_write_var_off(env, state,
3066 ptr_regno, off, size,
3067 value_regno, insn_idx);
3072 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
3073 int off, int size, enum bpf_access_type type)
3075 struct bpf_reg_state *regs = cur_regs(env);
3076 struct bpf_map *map = regs[regno].map_ptr;
3077 u32 cap = bpf_map_flags_to_cap(map);
3079 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
3080 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
3081 map->value_size, off, size);
3085 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
3086 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
3087 map->value_size, off, size);
3094 /* check read/write into memory region (e.g., map value, ringbuf sample, etc) */
3095 static int __check_mem_access(struct bpf_verifier_env *env, int regno,
3096 int off, int size, u32 mem_size,
3097 bool zero_size_allowed)
3099 bool size_ok = size > 0 || (size == 0 && zero_size_allowed);
3100 struct bpf_reg_state *reg;
3102 if (off >= 0 && size_ok && (u64)off + size <= mem_size)
3105 reg = &cur_regs(env)[regno];
3106 switch (reg->type) {
3107 case PTR_TO_MAP_KEY:
3108 verbose(env, "invalid access to map key, key_size=%d off=%d size=%d\n",
3109 mem_size, off, size);
3111 case PTR_TO_MAP_VALUE:
3112 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
3113 mem_size, off, size);
3116 case PTR_TO_PACKET_META:
3117 case PTR_TO_PACKET_END:
3118 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
3119 off, size, regno, reg->id, off, mem_size);
3123 verbose(env, "invalid access to memory, mem_size=%u off=%d size=%d\n",
3124 mem_size, off, size);
3130 /* check read/write into a memory region with possible variable offset */
3131 static int check_mem_region_access(struct bpf_verifier_env *env, u32 regno,
3132 int off, int size, u32 mem_size,
3133 bool zero_size_allowed)
3135 struct bpf_verifier_state *vstate = env->cur_state;
3136 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3137 struct bpf_reg_state *reg = &state->regs[regno];
3140 /* We may have adjusted the register pointing to memory region, so we
3141 * need to try adding each of min_value and max_value to off
3142 * to make sure our theoretical access will be safe.
3144 if (env->log.level & BPF_LOG_LEVEL)
3145 print_verifier_state(env, state);
3147 /* The minimum value is only important with signed
3148 * comparisons where we can't assume the floor of a
3149 * value is 0. If we are using signed variables for our
3150 * index'es we need to make sure that whatever we use
3151 * will have a set floor within our range.
3153 if (reg->smin_value < 0 &&
3154 (reg->smin_value == S64_MIN ||
3155 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
3156 reg->smin_value + off < 0)) {
3157 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3161 err = __check_mem_access(env, regno, reg->smin_value + off, size,
3162 mem_size, zero_size_allowed);
3164 verbose(env, "R%d min value is outside of the allowed memory range\n",
3169 /* If we haven't set a max value then we need to bail since we can't be
3170 * sure we won't do bad things.
3171 * If reg->umax_value + off could overflow, treat that as unbounded too.
3173 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
3174 verbose(env, "R%d unbounded memory access, make sure to bounds check any such access\n",
3178 err = __check_mem_access(env, regno, reg->umax_value + off, size,
3179 mem_size, zero_size_allowed);
3181 verbose(env, "R%d max value is outside of the allowed memory range\n",
3189 /* check read/write into a map element with possible variable offset */
3190 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
3191 int off, int size, bool zero_size_allowed)
3193 struct bpf_verifier_state *vstate = env->cur_state;
3194 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3195 struct bpf_reg_state *reg = &state->regs[regno];
3196 struct bpf_map *map = reg->map_ptr;
3199 err = check_mem_region_access(env, regno, off, size, map->value_size,
3204 if (map_value_has_spin_lock(map)) {
3205 u32 lock = map->spin_lock_off;
3207 /* if any part of struct bpf_spin_lock can be touched by
3208 * load/store reject this program.
3209 * To check that [x1, x2) overlaps with [y1, y2)
3210 * it is sufficient to check x1 < y2 && y1 < x2.
3212 if (reg->smin_value + off < lock + sizeof(struct bpf_spin_lock) &&
3213 lock < reg->umax_value + off + size) {
3214 verbose(env, "bpf_spin_lock cannot be accessed directly by load/store\n");
3221 #define MAX_PACKET_OFF 0xffff
3223 static enum bpf_prog_type resolve_prog_type(struct bpf_prog *prog)
3225 return prog->aux->dst_prog ? prog->aux->dst_prog->type : prog->type;
3228 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
3229 const struct bpf_call_arg_meta *meta,
3230 enum bpf_access_type t)
3232 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
3234 switch (prog_type) {
3235 /* Program types only with direct read access go here! */
3236 case BPF_PROG_TYPE_LWT_IN:
3237 case BPF_PROG_TYPE_LWT_OUT:
3238 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
3239 case BPF_PROG_TYPE_SK_REUSEPORT:
3240 case BPF_PROG_TYPE_FLOW_DISSECTOR:
3241 case BPF_PROG_TYPE_CGROUP_SKB:
3246 /* Program types with direct read + write access go here! */
3247 case BPF_PROG_TYPE_SCHED_CLS:
3248 case BPF_PROG_TYPE_SCHED_ACT:
3249 case BPF_PROG_TYPE_XDP:
3250 case BPF_PROG_TYPE_LWT_XMIT:
3251 case BPF_PROG_TYPE_SK_SKB:
3252 case BPF_PROG_TYPE_SK_MSG:
3254 return meta->pkt_access;
3256 env->seen_direct_write = true;
3259 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
3261 env->seen_direct_write = true;
3270 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
3271 int size, bool zero_size_allowed)
3273 struct bpf_reg_state *regs = cur_regs(env);
3274 struct bpf_reg_state *reg = ®s[regno];
3277 /* We may have added a variable offset to the packet pointer; but any
3278 * reg->range we have comes after that. We are only checking the fixed
3282 /* We don't allow negative numbers, because we aren't tracking enough
3283 * detail to prove they're safe.
3285 if (reg->smin_value < 0) {
3286 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3291 err = reg->range < 0 ? -EINVAL :
3292 __check_mem_access(env, regno, off, size, reg->range,
3295 verbose(env, "R%d offset is outside of the packet\n", regno);
3299 /* __check_mem_access has made sure "off + size - 1" is within u16.
3300 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
3301 * otherwise find_good_pkt_pointers would have refused to set range info
3302 * that __check_mem_access would have rejected this pkt access.
3303 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
3305 env->prog->aux->max_pkt_offset =
3306 max_t(u32, env->prog->aux->max_pkt_offset,
3307 off + reg->umax_value + size - 1);
3312 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
3313 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
3314 enum bpf_access_type t, enum bpf_reg_type *reg_type,
3315 struct btf **btf, u32 *btf_id)
3317 struct bpf_insn_access_aux info = {
3318 .reg_type = *reg_type,
3322 if (env->ops->is_valid_access &&
3323 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
3324 /* A non zero info.ctx_field_size indicates that this field is a
3325 * candidate for later verifier transformation to load the whole
3326 * field and then apply a mask when accessed with a narrower
3327 * access than actual ctx access size. A zero info.ctx_field_size
3328 * will only allow for whole field access and rejects any other
3329 * type of narrower access.
3331 *reg_type = info.reg_type;
3333 if (*reg_type == PTR_TO_BTF_ID || *reg_type == PTR_TO_BTF_ID_OR_NULL) {
3335 *btf_id = info.btf_id;
3337 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
3339 /* remember the offset of last byte accessed in ctx */
3340 if (env->prog->aux->max_ctx_offset < off + size)
3341 env->prog->aux->max_ctx_offset = off + size;
3345 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
3349 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
3352 if (size < 0 || off < 0 ||
3353 (u64)off + size > sizeof(struct bpf_flow_keys)) {
3354 verbose(env, "invalid access to flow keys off=%d size=%d\n",
3361 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
3362 u32 regno, int off, int size,
3363 enum bpf_access_type t)
3365 struct bpf_reg_state *regs = cur_regs(env);
3366 struct bpf_reg_state *reg = ®s[regno];
3367 struct bpf_insn_access_aux info = {};
3370 if (reg->smin_value < 0) {
3371 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3376 switch (reg->type) {
3377 case PTR_TO_SOCK_COMMON:
3378 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
3381 valid = bpf_sock_is_valid_access(off, size, t, &info);
3383 case PTR_TO_TCP_SOCK:
3384 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
3386 case PTR_TO_XDP_SOCK:
3387 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
3395 env->insn_aux_data[insn_idx].ctx_field_size =
3396 info.ctx_field_size;
3400 verbose(env, "R%d invalid %s access off=%d size=%d\n",
3401 regno, reg_type_str[reg->type], off, size);
3406 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
3408 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
3411 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
3413 const struct bpf_reg_state *reg = reg_state(env, regno);
3415 return reg->type == PTR_TO_CTX;
3418 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
3420 const struct bpf_reg_state *reg = reg_state(env, regno);
3422 return type_is_sk_pointer(reg->type);
3425 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
3427 const struct bpf_reg_state *reg = reg_state(env, regno);
3429 return type_is_pkt_pointer(reg->type);
3432 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
3434 const struct bpf_reg_state *reg = reg_state(env, regno);
3436 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
3437 return reg->type == PTR_TO_FLOW_KEYS;
3440 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
3441 const struct bpf_reg_state *reg,
3442 int off, int size, bool strict)
3444 struct tnum reg_off;
3447 /* Byte size accesses are always allowed. */
3448 if (!strict || size == 1)
3451 /* For platforms that do not have a Kconfig enabling
3452 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
3453 * NET_IP_ALIGN is universally set to '2'. And on platforms
3454 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
3455 * to this code only in strict mode where we want to emulate
3456 * the NET_IP_ALIGN==2 checking. Therefore use an
3457 * unconditional IP align value of '2'.
3461 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
3462 if (!tnum_is_aligned(reg_off, size)) {
3465 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3467 "misaligned packet access off %d+%s+%d+%d size %d\n",
3468 ip_align, tn_buf, reg->off, off, size);
3475 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
3476 const struct bpf_reg_state *reg,
3477 const char *pointer_desc,
3478 int off, int size, bool strict)
3480 struct tnum reg_off;
3482 /* Byte size accesses are always allowed. */
3483 if (!strict || size == 1)
3486 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
3487 if (!tnum_is_aligned(reg_off, size)) {
3490 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3491 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
3492 pointer_desc, tn_buf, reg->off, off, size);
3499 static int check_ptr_alignment(struct bpf_verifier_env *env,
3500 const struct bpf_reg_state *reg, int off,
3501 int size, bool strict_alignment_once)
3503 bool strict = env->strict_alignment || strict_alignment_once;
3504 const char *pointer_desc = "";
3506 switch (reg->type) {
3508 case PTR_TO_PACKET_META:
3509 /* Special case, because of NET_IP_ALIGN. Given metadata sits
3510 * right in front, treat it the very same way.
3512 return check_pkt_ptr_alignment(env, reg, off, size, strict);
3513 case PTR_TO_FLOW_KEYS:
3514 pointer_desc = "flow keys ";
3516 case PTR_TO_MAP_KEY:
3517 pointer_desc = "key ";
3519 case PTR_TO_MAP_VALUE:
3520 pointer_desc = "value ";
3523 pointer_desc = "context ";
3526 pointer_desc = "stack ";
3527 /* The stack spill tracking logic in check_stack_write_fixed_off()
3528 * and check_stack_read_fixed_off() relies on stack accesses being
3534 pointer_desc = "sock ";
3536 case PTR_TO_SOCK_COMMON:
3537 pointer_desc = "sock_common ";
3539 case PTR_TO_TCP_SOCK:
3540 pointer_desc = "tcp_sock ";
3542 case PTR_TO_XDP_SOCK:
3543 pointer_desc = "xdp_sock ";
3548 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
3552 static int update_stack_depth(struct bpf_verifier_env *env,
3553 const struct bpf_func_state *func,
3556 u16 stack = env->subprog_info[func->subprogno].stack_depth;
3561 /* update known max for given subprogram */
3562 env->subprog_info[func->subprogno].stack_depth = -off;
3566 /* starting from main bpf function walk all instructions of the function
3567 * and recursively walk all callees that given function can call.
3568 * Ignore jump and exit insns.
3569 * Since recursion is prevented by check_cfg() this algorithm
3570 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
3572 static int check_max_stack_depth(struct bpf_verifier_env *env)
3574 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
3575 struct bpf_subprog_info *subprog = env->subprog_info;
3576 struct bpf_insn *insn = env->prog->insnsi;
3577 bool tail_call_reachable = false;
3578 int ret_insn[MAX_CALL_FRAMES];
3579 int ret_prog[MAX_CALL_FRAMES];
3583 /* protect against potential stack overflow that might happen when
3584 * bpf2bpf calls get combined with tailcalls. Limit the caller's stack
3585 * depth for such case down to 256 so that the worst case scenario
3586 * would result in 8k stack size (32 which is tailcall limit * 256 =
3589 * To get the idea what might happen, see an example:
3590 * func1 -> sub rsp, 128
3591 * subfunc1 -> sub rsp, 256
3592 * tailcall1 -> add rsp, 256
3593 * func2 -> sub rsp, 192 (total stack size = 128 + 192 = 320)
3594 * subfunc2 -> sub rsp, 64
3595 * subfunc22 -> sub rsp, 128
3596 * tailcall2 -> add rsp, 128
3597 * func3 -> sub rsp, 32 (total stack size 128 + 192 + 64 + 32 = 416)
3599 * tailcall will unwind the current stack frame but it will not get rid
3600 * of caller's stack as shown on the example above.
3602 if (idx && subprog[idx].has_tail_call && depth >= 256) {
3604 "tail_calls are not allowed when call stack of previous frames is %d bytes. Too large\n",
3608 /* round up to 32-bytes, since this is granularity
3609 * of interpreter stack size
3611 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
3612 if (depth > MAX_BPF_STACK) {
3613 verbose(env, "combined stack size of %d calls is %d. Too large\n",
3618 subprog_end = subprog[idx + 1].start;
3619 for (; i < subprog_end; i++) {
3620 if (!bpf_pseudo_call(insn + i) && !bpf_pseudo_func(insn + i))
3622 /* remember insn and function to return to */
3623 ret_insn[frame] = i + 1;
3624 ret_prog[frame] = idx;
3626 /* find the callee */
3627 i = i + insn[i].imm + 1;
3628 idx = find_subprog(env, i);
3630 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3635 if (subprog[idx].has_tail_call)
3636 tail_call_reachable = true;
3639 if (frame >= MAX_CALL_FRAMES) {
3640 verbose(env, "the call stack of %d frames is too deep !\n",
3646 /* if tail call got detected across bpf2bpf calls then mark each of the
3647 * currently present subprog frames as tail call reachable subprogs;
3648 * this info will be utilized by JIT so that we will be preserving the
3649 * tail call counter throughout bpf2bpf calls combined with tailcalls
3651 if (tail_call_reachable)
3652 for (j = 0; j < frame; j++)
3653 subprog[ret_prog[j]].tail_call_reachable = true;
3654 if (subprog[0].tail_call_reachable)
3655 env->prog->aux->tail_call_reachable = true;
3657 /* end of for() loop means the last insn of the 'subprog'
3658 * was reached. Doesn't matter whether it was JA or EXIT
3662 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
3664 i = ret_insn[frame];
3665 idx = ret_prog[frame];
3669 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
3670 static int get_callee_stack_depth(struct bpf_verifier_env *env,
3671 const struct bpf_insn *insn, int idx)
3673 int start = idx + insn->imm + 1, subprog;
3675 subprog = find_subprog(env, start);
3677 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3681 return env->subprog_info[subprog].stack_depth;
3685 int check_ctx_reg(struct bpf_verifier_env *env,
3686 const struct bpf_reg_state *reg, int regno)
3688 /* Access to ctx or passing it to a helper is only allowed in
3689 * its original, unmodified form.
3693 verbose(env, "dereference of modified ctx ptr R%d off=%d disallowed\n",
3698 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3701 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3702 verbose(env, "variable ctx access var_off=%s disallowed\n", tn_buf);
3709 static int __check_buffer_access(struct bpf_verifier_env *env,
3710 const char *buf_info,
3711 const struct bpf_reg_state *reg,
3712 int regno, int off, int size)
3716 "R%d invalid %s buffer access: off=%d, size=%d\n",
3717 regno, buf_info, off, size);
3720 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3723 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3725 "R%d invalid variable buffer offset: off=%d, var_off=%s\n",
3726 regno, off, tn_buf);
3733 static int check_tp_buffer_access(struct bpf_verifier_env *env,
3734 const struct bpf_reg_state *reg,
3735 int regno, int off, int size)
3739 err = __check_buffer_access(env, "tracepoint", reg, regno, off, size);
3743 if (off + size > env->prog->aux->max_tp_access)
3744 env->prog->aux->max_tp_access = off + size;
3749 static int check_buffer_access(struct bpf_verifier_env *env,
3750 const struct bpf_reg_state *reg,
3751 int regno, int off, int size,
3752 bool zero_size_allowed,
3753 const char *buf_info,
3758 err = __check_buffer_access(env, buf_info, reg, regno, off, size);
3762 if (off + size > *max_access)
3763 *max_access = off + size;
3768 /* BPF architecture zero extends alu32 ops into 64-bit registesr */
3769 static void zext_32_to_64(struct bpf_reg_state *reg)
3771 reg->var_off = tnum_subreg(reg->var_off);
3772 __reg_assign_32_into_64(reg);
3775 /* truncate register to smaller size (in bytes)
3776 * must be called with size < BPF_REG_SIZE
3778 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
3782 /* clear high bits in bit representation */
3783 reg->var_off = tnum_cast(reg->var_off, size);
3785 /* fix arithmetic bounds */
3786 mask = ((u64)1 << (size * 8)) - 1;
3787 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
3788 reg->umin_value &= mask;
3789 reg->umax_value &= mask;
3791 reg->umin_value = 0;
3792 reg->umax_value = mask;
3794 reg->smin_value = reg->umin_value;
3795 reg->smax_value = reg->umax_value;
3797 /* If size is smaller than 32bit register the 32bit register
3798 * values are also truncated so we push 64-bit bounds into
3799 * 32-bit bounds. Above were truncated < 32-bits already.
3803 __reg_combine_64_into_32(reg);
3806 static bool bpf_map_is_rdonly(const struct bpf_map *map)
3808 return (map->map_flags & BPF_F_RDONLY_PROG) && map->frozen;
3811 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val)
3817 err = map->ops->map_direct_value_addr(map, &addr, off);
3820 ptr = (void *)(long)addr + off;
3824 *val = (u64)*(u8 *)ptr;
3827 *val = (u64)*(u16 *)ptr;
3830 *val = (u64)*(u32 *)ptr;
3841 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
3842 struct bpf_reg_state *regs,
3843 int regno, int off, int size,
3844 enum bpf_access_type atype,
3847 struct bpf_reg_state *reg = regs + regno;
3848 const struct btf_type *t = btf_type_by_id(reg->btf, reg->btf_id);
3849 const char *tname = btf_name_by_offset(reg->btf, t->name_off);
3855 "R%d is ptr_%s invalid negative access: off=%d\n",
3859 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3862 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3864 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
3865 regno, tname, off, tn_buf);
3869 if (env->ops->btf_struct_access) {
3870 ret = env->ops->btf_struct_access(&env->log, reg->btf, t,
3871 off, size, atype, &btf_id);
3873 if (atype != BPF_READ) {
3874 verbose(env, "only read is supported\n");
3878 ret = btf_struct_access(&env->log, reg->btf, t, off, size,
3885 if (atype == BPF_READ && value_regno >= 0)
3886 mark_btf_ld_reg(env, regs, value_regno, ret, reg->btf, btf_id);
3891 static int check_ptr_to_map_access(struct bpf_verifier_env *env,
3892 struct bpf_reg_state *regs,
3893 int regno, int off, int size,
3894 enum bpf_access_type atype,
3897 struct bpf_reg_state *reg = regs + regno;
3898 struct bpf_map *map = reg->map_ptr;
3899 const struct btf_type *t;
3905 verbose(env, "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n");
3909 if (!map->ops->map_btf_id || !*map->ops->map_btf_id) {
3910 verbose(env, "map_ptr access not supported for map type %d\n",
3915 t = btf_type_by_id(btf_vmlinux, *map->ops->map_btf_id);
3916 tname = btf_name_by_offset(btf_vmlinux, t->name_off);
3918 if (!env->allow_ptr_to_map_access) {
3920 "%s access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
3926 verbose(env, "R%d is %s invalid negative access: off=%d\n",
3931 if (atype != BPF_READ) {
3932 verbose(env, "only read from %s is supported\n", tname);
3936 ret = btf_struct_access(&env->log, btf_vmlinux, t, off, size, atype, &btf_id);
3940 if (value_regno >= 0)
3941 mark_btf_ld_reg(env, regs, value_regno, ret, btf_vmlinux, btf_id);
3946 /* Check that the stack access at the given offset is within bounds. The
3947 * maximum valid offset is -1.
3949 * The minimum valid offset is -MAX_BPF_STACK for writes, and
3950 * -state->allocated_stack for reads.
3952 static int check_stack_slot_within_bounds(int off,
3953 struct bpf_func_state *state,
3954 enum bpf_access_type t)
3959 min_valid_off = -MAX_BPF_STACK;
3961 min_valid_off = -state->allocated_stack;
3963 if (off < min_valid_off || off > -1)
3968 /* Check that the stack access at 'regno + off' falls within the maximum stack
3971 * 'off' includes `regno->offset`, but not its dynamic part (if any).
3973 static int check_stack_access_within_bounds(
3974 struct bpf_verifier_env *env,
3975 int regno, int off, int access_size,
3976 enum stack_access_src src, enum bpf_access_type type)
3978 struct bpf_reg_state *regs = cur_regs(env);
3979 struct bpf_reg_state *reg = regs + regno;
3980 struct bpf_func_state *state = func(env, reg);
3981 int min_off, max_off;
3985 if (src == ACCESS_HELPER)
3986 /* We don't know if helpers are reading or writing (or both). */
3987 err_extra = " indirect access to";
3988 else if (type == BPF_READ)
3989 err_extra = " read from";
3991 err_extra = " write to";
3993 if (tnum_is_const(reg->var_off)) {
3994 min_off = reg->var_off.value + off;
3995 if (access_size > 0)
3996 max_off = min_off + access_size - 1;
4000 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
4001 reg->smin_value <= -BPF_MAX_VAR_OFF) {
4002 verbose(env, "invalid unbounded variable-offset%s stack R%d\n",
4006 min_off = reg->smin_value + off;
4007 if (access_size > 0)
4008 max_off = reg->smax_value + off + access_size - 1;
4013 err = check_stack_slot_within_bounds(min_off, state, type);
4015 err = check_stack_slot_within_bounds(max_off, state, type);
4018 if (tnum_is_const(reg->var_off)) {
4019 verbose(env, "invalid%s stack R%d off=%d size=%d\n",
4020 err_extra, regno, off, access_size);
4024 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4025 verbose(env, "invalid variable-offset%s stack R%d var_off=%s size=%d\n",
4026 err_extra, regno, tn_buf, access_size);
4032 /* check whether memory at (regno + off) is accessible for t = (read | write)
4033 * if t==write, value_regno is a register which value is stored into memory
4034 * if t==read, value_regno is a register which will receive the value from memory
4035 * if t==write && value_regno==-1, some unknown value is stored into memory
4036 * if t==read && value_regno==-1, don't care what we read from memory
4038 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
4039 int off, int bpf_size, enum bpf_access_type t,
4040 int value_regno, bool strict_alignment_once)
4042 struct bpf_reg_state *regs = cur_regs(env);
4043 struct bpf_reg_state *reg = regs + regno;
4044 struct bpf_func_state *state;
4047 size = bpf_size_to_bytes(bpf_size);
4051 /* alignment checks will add in reg->off themselves */
4052 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
4056 /* for access checks, reg->off is just part of off */
4059 if (reg->type == PTR_TO_MAP_KEY) {
4060 if (t == BPF_WRITE) {
4061 verbose(env, "write to change key R%d not allowed\n", regno);
4065 err = check_mem_region_access(env, regno, off, size,
4066 reg->map_ptr->key_size, false);
4069 if (value_regno >= 0)
4070 mark_reg_unknown(env, regs, value_regno);
4071 } else if (reg->type == PTR_TO_MAP_VALUE) {
4072 if (t == BPF_WRITE && value_regno >= 0 &&
4073 is_pointer_value(env, value_regno)) {
4074 verbose(env, "R%d leaks addr into map\n", value_regno);
4077 err = check_map_access_type(env, regno, off, size, t);
4080 err = check_map_access(env, regno, off, size, false);
4081 if (!err && t == BPF_READ && value_regno >= 0) {
4082 struct bpf_map *map = reg->map_ptr;
4084 /* if map is read-only, track its contents as scalars */
4085 if (tnum_is_const(reg->var_off) &&
4086 bpf_map_is_rdonly(map) &&
4087 map->ops->map_direct_value_addr) {
4088 int map_off = off + reg->var_off.value;
4091 err = bpf_map_direct_read(map, map_off, size,
4096 regs[value_regno].type = SCALAR_VALUE;
4097 __mark_reg_known(®s[value_regno], val);
4099 mark_reg_unknown(env, regs, value_regno);
4102 } else if (reg->type == PTR_TO_MEM) {
4103 if (t == BPF_WRITE && value_regno >= 0 &&
4104 is_pointer_value(env, value_regno)) {
4105 verbose(env, "R%d leaks addr into mem\n", value_regno);
4108 err = check_mem_region_access(env, regno, off, size,
4109 reg->mem_size, false);
4110 if (!err && t == BPF_READ && value_regno >= 0)
4111 mark_reg_unknown(env, regs, value_regno);
4112 } else if (reg->type == PTR_TO_CTX) {
4113 enum bpf_reg_type reg_type = SCALAR_VALUE;
4114 struct btf *btf = NULL;
4117 if (t == BPF_WRITE && value_regno >= 0 &&
4118 is_pointer_value(env, value_regno)) {
4119 verbose(env, "R%d leaks addr into ctx\n", value_regno);
4123 err = check_ctx_reg(env, reg, regno);
4127 err = check_ctx_access(env, insn_idx, off, size, t, ®_type, &btf, &btf_id);
4129 verbose_linfo(env, insn_idx, "; ");
4130 if (!err && t == BPF_READ && value_regno >= 0) {
4131 /* ctx access returns either a scalar, or a
4132 * PTR_TO_PACKET[_META,_END]. In the latter
4133 * case, we know the offset is zero.
4135 if (reg_type == SCALAR_VALUE) {
4136 mark_reg_unknown(env, regs, value_regno);
4138 mark_reg_known_zero(env, regs,
4140 if (reg_type_may_be_null(reg_type))
4141 regs[value_regno].id = ++env->id_gen;
4142 /* A load of ctx field could have different
4143 * actual load size with the one encoded in the
4144 * insn. When the dst is PTR, it is for sure not
4147 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
4148 if (reg_type == PTR_TO_BTF_ID ||
4149 reg_type == PTR_TO_BTF_ID_OR_NULL) {
4150 regs[value_regno].btf = btf;
4151 regs[value_regno].btf_id = btf_id;
4154 regs[value_regno].type = reg_type;
4157 } else if (reg->type == PTR_TO_STACK) {
4158 /* Basic bounds checks. */
4159 err = check_stack_access_within_bounds(env, regno, off, size, ACCESS_DIRECT, t);
4163 state = func(env, reg);
4164 err = update_stack_depth(env, state, off);
4169 err = check_stack_read(env, regno, off, size,
4172 err = check_stack_write(env, regno, off, size,
4173 value_regno, insn_idx);
4174 } else if (reg_is_pkt_pointer(reg)) {
4175 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
4176 verbose(env, "cannot write into packet\n");
4179 if (t == BPF_WRITE && value_regno >= 0 &&
4180 is_pointer_value(env, value_regno)) {
4181 verbose(env, "R%d leaks addr into packet\n",
4185 err = check_packet_access(env, regno, off, size, false);
4186 if (!err && t == BPF_READ && value_regno >= 0)
4187 mark_reg_unknown(env, regs, value_regno);
4188 } else if (reg->type == PTR_TO_FLOW_KEYS) {
4189 if (t == BPF_WRITE && value_regno >= 0 &&
4190 is_pointer_value(env, value_regno)) {
4191 verbose(env, "R%d leaks addr into flow keys\n",
4196 err = check_flow_keys_access(env, off, size);
4197 if (!err && t == BPF_READ && value_regno >= 0)
4198 mark_reg_unknown(env, regs, value_regno);
4199 } else if (type_is_sk_pointer(reg->type)) {
4200 if (t == BPF_WRITE) {
4201 verbose(env, "R%d cannot write into %s\n",
4202 regno, reg_type_str[reg->type]);
4205 err = check_sock_access(env, insn_idx, regno, off, size, t);
4206 if (!err && value_regno >= 0)
4207 mark_reg_unknown(env, regs, value_regno);
4208 } else if (reg->type == PTR_TO_TP_BUFFER) {
4209 err = check_tp_buffer_access(env, reg, regno, off, size);
4210 if (!err && t == BPF_READ && value_regno >= 0)
4211 mark_reg_unknown(env, regs, value_regno);
4212 } else if (reg->type == PTR_TO_BTF_ID) {
4213 err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
4215 } else if (reg->type == CONST_PTR_TO_MAP) {
4216 err = check_ptr_to_map_access(env, regs, regno, off, size, t,
4218 } else if (reg->type == PTR_TO_RDONLY_BUF) {
4219 if (t == BPF_WRITE) {
4220 verbose(env, "R%d cannot write into %s\n",
4221 regno, reg_type_str[reg->type]);
4224 err = check_buffer_access(env, reg, regno, off, size, false,
4226 &env->prog->aux->max_rdonly_access);
4227 if (!err && value_regno >= 0)
4228 mark_reg_unknown(env, regs, value_regno);
4229 } else if (reg->type == PTR_TO_RDWR_BUF) {
4230 err = check_buffer_access(env, reg, regno, off, size, false,
4232 &env->prog->aux->max_rdwr_access);
4233 if (!err && t == BPF_READ && value_regno >= 0)
4234 mark_reg_unknown(env, regs, value_regno);
4236 verbose(env, "R%d invalid mem access '%s'\n", regno,
4237 reg_type_str[reg->type]);
4241 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
4242 regs[value_regno].type == SCALAR_VALUE) {
4243 /* b/h/w load zero-extends, mark upper bits as known 0 */
4244 coerce_reg_to_size(®s[value_regno], size);
4249 static int check_atomic(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
4254 switch (insn->imm) {
4256 case BPF_ADD | BPF_FETCH:
4258 case BPF_AND | BPF_FETCH:
4260 case BPF_OR | BPF_FETCH:
4262 case BPF_XOR | BPF_FETCH:
4267 verbose(env, "BPF_ATOMIC uses invalid atomic opcode %02x\n", insn->imm);
4271 if (BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) {
4272 verbose(env, "invalid atomic operand size\n");
4276 /* check src1 operand */
4277 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4281 /* check src2 operand */
4282 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4286 if (insn->imm == BPF_CMPXCHG) {
4287 /* Check comparison of R0 with memory location */
4288 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
4293 if (is_pointer_value(env, insn->src_reg)) {
4294 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
4298 if (is_ctx_reg(env, insn->dst_reg) ||
4299 is_pkt_reg(env, insn->dst_reg) ||
4300 is_flow_key_reg(env, insn->dst_reg) ||
4301 is_sk_reg(env, insn->dst_reg)) {
4302 verbose(env, "BPF_ATOMIC stores into R%d %s is not allowed\n",
4304 reg_type_str[reg_state(env, insn->dst_reg)->type]);
4308 if (insn->imm & BPF_FETCH) {
4309 if (insn->imm == BPF_CMPXCHG)
4310 load_reg = BPF_REG_0;
4312 load_reg = insn->src_reg;
4314 /* check and record load of old value */
4315 err = check_reg_arg(env, load_reg, DST_OP);
4319 /* This instruction accesses a memory location but doesn't
4320 * actually load it into a register.
4325 /* check whether we can read the memory */
4326 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4327 BPF_SIZE(insn->code), BPF_READ, load_reg, true);
4331 /* check whether we can write into the same memory */
4332 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4333 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
4340 /* When register 'regno' is used to read the stack (either directly or through
4341 * a helper function) make sure that it's within stack boundary and, depending
4342 * on the access type, that all elements of the stack are initialized.
4344 * 'off' includes 'regno->off', but not its dynamic part (if any).
4346 * All registers that have been spilled on the stack in the slots within the
4347 * read offsets are marked as read.
4349 static int check_stack_range_initialized(
4350 struct bpf_verifier_env *env, int regno, int off,
4351 int access_size, bool zero_size_allowed,
4352 enum stack_access_src type, struct bpf_call_arg_meta *meta)
4354 struct bpf_reg_state *reg = reg_state(env, regno);
4355 struct bpf_func_state *state = func(env, reg);
4356 int err, min_off, max_off, i, j, slot, spi;
4357 char *err_extra = type == ACCESS_HELPER ? " indirect" : "";
4358 enum bpf_access_type bounds_check_type;
4359 /* Some accesses can write anything into the stack, others are
4362 bool clobber = false;
4364 if (access_size == 0 && !zero_size_allowed) {
4365 verbose(env, "invalid zero-sized read\n");
4369 if (type == ACCESS_HELPER) {
4370 /* The bounds checks for writes are more permissive than for
4371 * reads. However, if raw_mode is not set, we'll do extra
4374 bounds_check_type = BPF_WRITE;
4377 bounds_check_type = BPF_READ;
4379 err = check_stack_access_within_bounds(env, regno, off, access_size,
4380 type, bounds_check_type);
4385 if (tnum_is_const(reg->var_off)) {
4386 min_off = max_off = reg->var_off.value + off;
4388 /* Variable offset is prohibited for unprivileged mode for
4389 * simplicity since it requires corresponding support in
4390 * Spectre masking for stack ALU.
4391 * See also retrieve_ptr_limit().
4393 if (!env->bypass_spec_v1) {
4396 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4397 verbose(env, "R%d%s variable offset stack access prohibited for !root, var_off=%s\n",
4398 regno, err_extra, tn_buf);
4401 /* Only initialized buffer on stack is allowed to be accessed
4402 * with variable offset. With uninitialized buffer it's hard to
4403 * guarantee that whole memory is marked as initialized on
4404 * helper return since specific bounds are unknown what may
4405 * cause uninitialized stack leaking.
4407 if (meta && meta->raw_mode)
4410 min_off = reg->smin_value + off;
4411 max_off = reg->smax_value + off;
4414 if (meta && meta->raw_mode) {
4415 meta->access_size = access_size;
4416 meta->regno = regno;
4420 for (i = min_off; i < max_off + access_size; i++) {
4424 spi = slot / BPF_REG_SIZE;
4425 if (state->allocated_stack <= slot)
4427 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
4428 if (*stype == STACK_MISC)
4430 if (*stype == STACK_ZERO) {
4432 /* helper can write anything into the stack */
4433 *stype = STACK_MISC;
4438 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
4439 state->stack[spi].spilled_ptr.type == PTR_TO_BTF_ID)
4442 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
4443 (state->stack[spi].spilled_ptr.type == SCALAR_VALUE ||
4444 env->allow_ptr_leaks)) {
4446 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
4447 for (j = 0; j < BPF_REG_SIZE; j++)
4448 state->stack[spi].slot_type[j] = STACK_MISC;
4454 if (tnum_is_const(reg->var_off)) {
4455 verbose(env, "invalid%s read from stack R%d off %d+%d size %d\n",
4456 err_extra, regno, min_off, i - min_off, access_size);
4460 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4461 verbose(env, "invalid%s read from stack R%d var_off %s+%d size %d\n",
4462 err_extra, regno, tn_buf, i - min_off, access_size);
4466 /* reading any byte out of 8-byte 'spill_slot' will cause
4467 * the whole slot to be marked as 'read'
4469 mark_reg_read(env, &state->stack[spi].spilled_ptr,
4470 state->stack[spi].spilled_ptr.parent,
4473 return update_stack_depth(env, state, min_off);
4476 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
4477 int access_size, bool zero_size_allowed,
4478 struct bpf_call_arg_meta *meta)
4480 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4482 switch (reg->type) {
4484 case PTR_TO_PACKET_META:
4485 return check_packet_access(env, regno, reg->off, access_size,
4487 case PTR_TO_MAP_KEY:
4488 return check_mem_region_access(env, regno, reg->off, access_size,
4489 reg->map_ptr->key_size, false);
4490 case PTR_TO_MAP_VALUE:
4491 if (check_map_access_type(env, regno, reg->off, access_size,
4492 meta && meta->raw_mode ? BPF_WRITE :
4495 return check_map_access(env, regno, reg->off, access_size,
4498 return check_mem_region_access(env, regno, reg->off,
4499 access_size, reg->mem_size,
4501 case PTR_TO_RDONLY_BUF:
4502 if (meta && meta->raw_mode)
4504 return check_buffer_access(env, reg, regno, reg->off,
4505 access_size, zero_size_allowed,
4507 &env->prog->aux->max_rdonly_access);
4508 case PTR_TO_RDWR_BUF:
4509 return check_buffer_access(env, reg, regno, reg->off,
4510 access_size, zero_size_allowed,
4512 &env->prog->aux->max_rdwr_access);
4514 return check_stack_range_initialized(
4516 regno, reg->off, access_size,
4517 zero_size_allowed, ACCESS_HELPER, meta);
4518 default: /* scalar_value or invalid ptr */
4519 /* Allow zero-byte read from NULL, regardless of pointer type */
4520 if (zero_size_allowed && access_size == 0 &&
4521 register_is_null(reg))
4524 verbose(env, "R%d type=%s expected=%s\n", regno,
4525 reg_type_str[reg->type],
4526 reg_type_str[PTR_TO_STACK]);
4531 int check_mem_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
4532 u32 regno, u32 mem_size)
4534 if (register_is_null(reg))
4537 if (reg_type_may_be_null(reg->type)) {
4538 /* Assuming that the register contains a value check if the memory
4539 * access is safe. Temporarily save and restore the register's state as
4540 * the conversion shouldn't be visible to a caller.
4542 const struct bpf_reg_state saved_reg = *reg;
4545 mark_ptr_not_null_reg(reg);
4546 rv = check_helper_mem_access(env, regno, mem_size, true, NULL);
4551 return check_helper_mem_access(env, regno, mem_size, true, NULL);
4554 /* Implementation details:
4555 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL
4556 * Two bpf_map_lookups (even with the same key) will have different reg->id.
4557 * For traditional PTR_TO_MAP_VALUE the verifier clears reg->id after
4558 * value_or_null->value transition, since the verifier only cares about
4559 * the range of access to valid map value pointer and doesn't care about actual
4560 * address of the map element.
4561 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
4562 * reg->id > 0 after value_or_null->value transition. By doing so
4563 * two bpf_map_lookups will be considered two different pointers that
4564 * point to different bpf_spin_locks.
4565 * The verifier allows taking only one bpf_spin_lock at a time to avoid
4567 * Since only one bpf_spin_lock is allowed the checks are simpler than
4568 * reg_is_refcounted() logic. The verifier needs to remember only
4569 * one spin_lock instead of array of acquired_refs.
4570 * cur_state->active_spin_lock remembers which map value element got locked
4571 * and clears it after bpf_spin_unlock.
4573 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
4576 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4577 struct bpf_verifier_state *cur = env->cur_state;
4578 bool is_const = tnum_is_const(reg->var_off);
4579 struct bpf_map *map = reg->map_ptr;
4580 u64 val = reg->var_off.value;
4584 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
4590 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
4594 if (!map_value_has_spin_lock(map)) {
4595 if (map->spin_lock_off == -E2BIG)
4597 "map '%s' has more than one 'struct bpf_spin_lock'\n",
4599 else if (map->spin_lock_off == -ENOENT)
4601 "map '%s' doesn't have 'struct bpf_spin_lock'\n",
4605 "map '%s' is not a struct type or bpf_spin_lock is mangled\n",
4609 if (map->spin_lock_off != val + reg->off) {
4610 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock'\n",
4615 if (cur->active_spin_lock) {
4617 "Locking two bpf_spin_locks are not allowed\n");
4620 cur->active_spin_lock = reg->id;
4622 if (!cur->active_spin_lock) {
4623 verbose(env, "bpf_spin_unlock without taking a lock\n");
4626 if (cur->active_spin_lock != reg->id) {
4627 verbose(env, "bpf_spin_unlock of different lock\n");
4630 cur->active_spin_lock = 0;
4635 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
4637 return type == ARG_PTR_TO_MEM ||
4638 type == ARG_PTR_TO_MEM_OR_NULL ||
4639 type == ARG_PTR_TO_UNINIT_MEM;
4642 static bool arg_type_is_mem_size(enum bpf_arg_type type)
4644 return type == ARG_CONST_SIZE ||
4645 type == ARG_CONST_SIZE_OR_ZERO;
4648 static bool arg_type_is_alloc_size(enum bpf_arg_type type)
4650 return type == ARG_CONST_ALLOC_SIZE_OR_ZERO;
4653 static bool arg_type_is_int_ptr(enum bpf_arg_type type)
4655 return type == ARG_PTR_TO_INT ||
4656 type == ARG_PTR_TO_LONG;
4659 static int int_ptr_type_to_size(enum bpf_arg_type type)
4661 if (type == ARG_PTR_TO_INT)
4663 else if (type == ARG_PTR_TO_LONG)
4669 static int resolve_map_arg_type(struct bpf_verifier_env *env,
4670 const struct bpf_call_arg_meta *meta,
4671 enum bpf_arg_type *arg_type)
4673 if (!meta->map_ptr) {
4674 /* kernel subsystem misconfigured verifier */
4675 verbose(env, "invalid map_ptr to access map->type\n");
4679 switch (meta->map_ptr->map_type) {
4680 case BPF_MAP_TYPE_SOCKMAP:
4681 case BPF_MAP_TYPE_SOCKHASH:
4682 if (*arg_type == ARG_PTR_TO_MAP_VALUE) {
4683 *arg_type = ARG_PTR_TO_BTF_ID_SOCK_COMMON;
4685 verbose(env, "invalid arg_type for sockmap/sockhash\n");
4696 struct bpf_reg_types {
4697 const enum bpf_reg_type types[10];
4701 static const struct bpf_reg_types map_key_value_types = {
4711 static const struct bpf_reg_types sock_types = {
4721 static const struct bpf_reg_types btf_id_sock_common_types = {
4729 .btf_id = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
4733 static const struct bpf_reg_types mem_types = {
4746 static const struct bpf_reg_types int_ptr_types = {
4756 static const struct bpf_reg_types fullsock_types = { .types = { PTR_TO_SOCKET } };
4757 static const struct bpf_reg_types scalar_types = { .types = { SCALAR_VALUE } };
4758 static const struct bpf_reg_types context_types = { .types = { PTR_TO_CTX } };
4759 static const struct bpf_reg_types alloc_mem_types = { .types = { PTR_TO_MEM } };
4760 static const struct bpf_reg_types const_map_ptr_types = { .types = { CONST_PTR_TO_MAP } };
4761 static const struct bpf_reg_types btf_ptr_types = { .types = { PTR_TO_BTF_ID } };
4762 static const struct bpf_reg_types spin_lock_types = { .types = { PTR_TO_MAP_VALUE } };
4763 static const struct bpf_reg_types percpu_btf_ptr_types = { .types = { PTR_TO_PERCPU_BTF_ID } };
4764 static const struct bpf_reg_types func_ptr_types = { .types = { PTR_TO_FUNC } };
4765 static const struct bpf_reg_types stack_ptr_types = { .types = { PTR_TO_STACK } };
4766 static const struct bpf_reg_types const_str_ptr_types = { .types = { PTR_TO_MAP_VALUE } };
4768 static const struct bpf_reg_types *compatible_reg_types[__BPF_ARG_TYPE_MAX] = {
4769 [ARG_PTR_TO_MAP_KEY] = &map_key_value_types,
4770 [ARG_PTR_TO_MAP_VALUE] = &map_key_value_types,
4771 [ARG_PTR_TO_UNINIT_MAP_VALUE] = &map_key_value_types,
4772 [ARG_PTR_TO_MAP_VALUE_OR_NULL] = &map_key_value_types,
4773 [ARG_CONST_SIZE] = &scalar_types,
4774 [ARG_CONST_SIZE_OR_ZERO] = &scalar_types,
4775 [ARG_CONST_ALLOC_SIZE_OR_ZERO] = &scalar_types,
4776 [ARG_CONST_MAP_PTR] = &const_map_ptr_types,
4777 [ARG_PTR_TO_CTX] = &context_types,
4778 [ARG_PTR_TO_CTX_OR_NULL] = &context_types,
4779 [ARG_PTR_TO_SOCK_COMMON] = &sock_types,
4781 [ARG_PTR_TO_BTF_ID_SOCK_COMMON] = &btf_id_sock_common_types,
4783 [ARG_PTR_TO_SOCKET] = &fullsock_types,
4784 [ARG_PTR_TO_SOCKET_OR_NULL] = &fullsock_types,
4785 [ARG_PTR_TO_BTF_ID] = &btf_ptr_types,
4786 [ARG_PTR_TO_SPIN_LOCK] = &spin_lock_types,
4787 [ARG_PTR_TO_MEM] = &mem_types,
4788 [ARG_PTR_TO_MEM_OR_NULL] = &mem_types,
4789 [ARG_PTR_TO_UNINIT_MEM] = &mem_types,
4790 [ARG_PTR_TO_ALLOC_MEM] = &alloc_mem_types,
4791 [ARG_PTR_TO_ALLOC_MEM_OR_NULL] = &alloc_mem_types,
4792 [ARG_PTR_TO_INT] = &int_ptr_types,
4793 [ARG_PTR_TO_LONG] = &int_ptr_types,
4794 [ARG_PTR_TO_PERCPU_BTF_ID] = &percpu_btf_ptr_types,
4795 [ARG_PTR_TO_FUNC] = &func_ptr_types,
4796 [ARG_PTR_TO_STACK_OR_NULL] = &stack_ptr_types,
4797 [ARG_PTR_TO_CONST_STR] = &const_str_ptr_types,
4800 static int check_reg_type(struct bpf_verifier_env *env, u32 regno,
4801 enum bpf_arg_type arg_type,
4802 const u32 *arg_btf_id)
4804 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4805 enum bpf_reg_type expected, type = reg->type;
4806 const struct bpf_reg_types *compatible;
4809 compatible = compatible_reg_types[arg_type];
4811 verbose(env, "verifier internal error: unsupported arg type %d\n", arg_type);
4815 for (i = 0; i < ARRAY_SIZE(compatible->types); i++) {
4816 expected = compatible->types[i];
4817 if (expected == NOT_INIT)
4820 if (type == expected)
4824 verbose(env, "R%d type=%s expected=", regno, reg_type_str[type]);
4825 for (j = 0; j + 1 < i; j++)
4826 verbose(env, "%s, ", reg_type_str[compatible->types[j]]);
4827 verbose(env, "%s\n", reg_type_str[compatible->types[j]]);
4831 if (type == PTR_TO_BTF_ID) {
4833 if (!compatible->btf_id) {
4834 verbose(env, "verifier internal error: missing arg compatible BTF ID\n");
4837 arg_btf_id = compatible->btf_id;
4840 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
4841 btf_vmlinux, *arg_btf_id)) {
4842 verbose(env, "R%d is of type %s but %s is expected\n",
4843 regno, kernel_type_name(reg->btf, reg->btf_id),
4844 kernel_type_name(btf_vmlinux, *arg_btf_id));
4848 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
4849 verbose(env, "R%d is a pointer to in-kernel struct with non-zero offset\n",
4858 static int check_func_arg(struct bpf_verifier_env *env, u32 arg,
4859 struct bpf_call_arg_meta *meta,
4860 const struct bpf_func_proto *fn)
4862 u32 regno = BPF_REG_1 + arg;
4863 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4864 enum bpf_arg_type arg_type = fn->arg_type[arg];
4865 enum bpf_reg_type type = reg->type;
4868 if (arg_type == ARG_DONTCARE)
4871 err = check_reg_arg(env, regno, SRC_OP);
4875 if (arg_type == ARG_ANYTHING) {
4876 if (is_pointer_value(env, regno)) {
4877 verbose(env, "R%d leaks addr into helper function\n",
4884 if (type_is_pkt_pointer(type) &&
4885 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
4886 verbose(env, "helper access to the packet is not allowed\n");
4890 if (arg_type == ARG_PTR_TO_MAP_VALUE ||
4891 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE ||
4892 arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL) {
4893 err = resolve_map_arg_type(env, meta, &arg_type);
4898 if (register_is_null(reg) && arg_type_may_be_null(arg_type))
4899 /* A NULL register has a SCALAR_VALUE type, so skip
4902 goto skip_type_check;
4904 err = check_reg_type(env, regno, arg_type, fn->arg_btf_id[arg]);
4908 if (type == PTR_TO_CTX) {
4909 err = check_ctx_reg(env, reg, regno);
4915 if (reg->ref_obj_id) {
4916 if (meta->ref_obj_id) {
4917 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
4918 regno, reg->ref_obj_id,
4922 meta->ref_obj_id = reg->ref_obj_id;
4925 if (arg_type == ARG_CONST_MAP_PTR) {
4926 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
4927 meta->map_ptr = reg->map_ptr;
4928 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
4929 /* bpf_map_xxx(..., map_ptr, ..., key) call:
4930 * check that [key, key + map->key_size) are within
4931 * stack limits and initialized
4933 if (!meta->map_ptr) {
4934 /* in function declaration map_ptr must come before
4935 * map_key, so that it's verified and known before
4936 * we have to check map_key here. Otherwise it means
4937 * that kernel subsystem misconfigured verifier
4939 verbose(env, "invalid map_ptr to access map->key\n");
4942 err = check_helper_mem_access(env, regno,
4943 meta->map_ptr->key_size, false,
4945 } else if (arg_type == ARG_PTR_TO_MAP_VALUE ||
4946 (arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL &&
4947 !register_is_null(reg)) ||
4948 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
4949 /* bpf_map_xxx(..., map_ptr, ..., value) call:
4950 * check [value, value + map->value_size) validity
4952 if (!meta->map_ptr) {
4953 /* kernel subsystem misconfigured verifier */
4954 verbose(env, "invalid map_ptr to access map->value\n");
4957 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE);
4958 err = check_helper_mem_access(env, regno,
4959 meta->map_ptr->value_size, false,
4961 } else if (arg_type == ARG_PTR_TO_PERCPU_BTF_ID) {
4963 verbose(env, "Helper has invalid btf_id in R%d\n", regno);
4966 meta->ret_btf = reg->btf;
4967 meta->ret_btf_id = reg->btf_id;
4968 } else if (arg_type == ARG_PTR_TO_SPIN_LOCK) {
4969 if (meta->func_id == BPF_FUNC_spin_lock) {
4970 if (process_spin_lock(env, regno, true))
4972 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
4973 if (process_spin_lock(env, regno, false))
4976 verbose(env, "verifier internal error\n");
4979 } else if (arg_type == ARG_PTR_TO_FUNC) {
4980 meta->subprogno = reg->subprogno;
4981 } else if (arg_type_is_mem_ptr(arg_type)) {
4982 /* The access to this pointer is only checked when we hit the
4983 * next is_mem_size argument below.
4985 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MEM);
4986 } else if (arg_type_is_mem_size(arg_type)) {
4987 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
4989 /* This is used to refine r0 return value bounds for helpers
4990 * that enforce this value as an upper bound on return values.
4991 * See do_refine_retval_range() for helpers that can refine
4992 * the return value. C type of helper is u32 so we pull register
4993 * bound from umax_value however, if negative verifier errors
4994 * out. Only upper bounds can be learned because retval is an
4995 * int type and negative retvals are allowed.
4997 meta->msize_max_value = reg->umax_value;
4999 /* The register is SCALAR_VALUE; the access check
5000 * happens using its boundaries.
5002 if (!tnum_is_const(reg->var_off))
5003 /* For unprivileged variable accesses, disable raw
5004 * mode so that the program is required to
5005 * initialize all the memory that the helper could
5006 * just partially fill up.
5010 if (reg->smin_value < 0) {
5011 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
5016 if (reg->umin_value == 0) {
5017 err = check_helper_mem_access(env, regno - 1, 0,
5024 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
5025 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
5029 err = check_helper_mem_access(env, regno - 1,
5031 zero_size_allowed, meta);
5033 err = mark_chain_precision(env, regno);
5034 } else if (arg_type_is_alloc_size(arg_type)) {
5035 if (!tnum_is_const(reg->var_off)) {
5036 verbose(env, "R%d is not a known constant'\n",
5040 meta->mem_size = reg->var_off.value;
5041 } else if (arg_type_is_int_ptr(arg_type)) {
5042 int size = int_ptr_type_to_size(arg_type);
5044 err = check_helper_mem_access(env, regno, size, false, meta);
5047 err = check_ptr_alignment(env, reg, 0, size, true);
5048 } else if (arg_type == ARG_PTR_TO_CONST_STR) {
5049 struct bpf_map *map = reg->map_ptr;
5054 if (!bpf_map_is_rdonly(map)) {
5055 verbose(env, "R%d does not point to a readonly map'\n", regno);
5059 if (!tnum_is_const(reg->var_off)) {
5060 verbose(env, "R%d is not a constant address'\n", regno);
5064 if (!map->ops->map_direct_value_addr) {
5065 verbose(env, "no direct value access support for this map type\n");
5069 err = check_map_access(env, regno, reg->off,
5070 map->value_size - reg->off, false);
5074 map_off = reg->off + reg->var_off.value;
5075 err = map->ops->map_direct_value_addr(map, &map_addr, map_off);
5077 verbose(env, "direct value access on string failed\n");
5081 str_ptr = (char *)(long)(map_addr);
5082 if (!strnchr(str_ptr + map_off, map->value_size - map_off, 0)) {
5083 verbose(env, "string is not zero-terminated\n");
5091 static bool may_update_sockmap(struct bpf_verifier_env *env, int func_id)
5093 enum bpf_attach_type eatype = env->prog->expected_attach_type;
5094 enum bpf_prog_type type = resolve_prog_type(env->prog);
5096 if (func_id != BPF_FUNC_map_update_elem)
5099 /* It's not possible to get access to a locked struct sock in these
5100 * contexts, so updating is safe.
5103 case BPF_PROG_TYPE_TRACING:
5104 if (eatype == BPF_TRACE_ITER)
5107 case BPF_PROG_TYPE_SOCKET_FILTER:
5108 case BPF_PROG_TYPE_SCHED_CLS:
5109 case BPF_PROG_TYPE_SCHED_ACT:
5110 case BPF_PROG_TYPE_XDP:
5111 case BPF_PROG_TYPE_SK_REUSEPORT:
5112 case BPF_PROG_TYPE_FLOW_DISSECTOR:
5113 case BPF_PROG_TYPE_SK_LOOKUP:
5119 verbose(env, "cannot update sockmap in this context\n");
5123 static bool allow_tail_call_in_subprogs(struct bpf_verifier_env *env)
5125 return env->prog->jit_requested && IS_ENABLED(CONFIG_X86_64);
5128 static int check_map_func_compatibility(struct bpf_verifier_env *env,
5129 struct bpf_map *map, int func_id)
5134 /* We need a two way check, first is from map perspective ... */
5135 switch (map->map_type) {
5136 case BPF_MAP_TYPE_PROG_ARRAY:
5137 if (func_id != BPF_FUNC_tail_call)
5140 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
5141 if (func_id != BPF_FUNC_perf_event_read &&
5142 func_id != BPF_FUNC_perf_event_output &&
5143 func_id != BPF_FUNC_skb_output &&
5144 func_id != BPF_FUNC_perf_event_read_value &&
5145 func_id != BPF_FUNC_xdp_output)
5148 case BPF_MAP_TYPE_RINGBUF:
5149 if (func_id != BPF_FUNC_ringbuf_output &&
5150 func_id != BPF_FUNC_ringbuf_reserve &&
5151 func_id != BPF_FUNC_ringbuf_query)
5154 case BPF_MAP_TYPE_STACK_TRACE:
5155 if (func_id != BPF_FUNC_get_stackid)
5158 case BPF_MAP_TYPE_CGROUP_ARRAY:
5159 if (func_id != BPF_FUNC_skb_under_cgroup &&
5160 func_id != BPF_FUNC_current_task_under_cgroup)
5163 case BPF_MAP_TYPE_CGROUP_STORAGE:
5164 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
5165 if (func_id != BPF_FUNC_get_local_storage)
5168 case BPF_MAP_TYPE_DEVMAP:
5169 case BPF_MAP_TYPE_DEVMAP_HASH:
5170 if (func_id != BPF_FUNC_redirect_map &&
5171 func_id != BPF_FUNC_map_lookup_elem)
5174 /* Restrict bpf side of cpumap and xskmap, open when use-cases
5177 case BPF_MAP_TYPE_CPUMAP:
5178 if (func_id != BPF_FUNC_redirect_map)
5181 case BPF_MAP_TYPE_XSKMAP:
5182 if (func_id != BPF_FUNC_redirect_map &&
5183 func_id != BPF_FUNC_map_lookup_elem)
5186 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
5187 case BPF_MAP_TYPE_HASH_OF_MAPS:
5188 if (func_id != BPF_FUNC_map_lookup_elem)
5191 case BPF_MAP_TYPE_SOCKMAP:
5192 if (func_id != BPF_FUNC_sk_redirect_map &&
5193 func_id != BPF_FUNC_sock_map_update &&
5194 func_id != BPF_FUNC_map_delete_elem &&
5195 func_id != BPF_FUNC_msg_redirect_map &&
5196 func_id != BPF_FUNC_sk_select_reuseport &&
5197 func_id != BPF_FUNC_map_lookup_elem &&
5198 !may_update_sockmap(env, func_id))
5201 case BPF_MAP_TYPE_SOCKHASH:
5202 if (func_id != BPF_FUNC_sk_redirect_hash &&
5203 func_id != BPF_FUNC_sock_hash_update &&
5204 func_id != BPF_FUNC_map_delete_elem &&
5205 func_id != BPF_FUNC_msg_redirect_hash &&
5206 func_id != BPF_FUNC_sk_select_reuseport &&
5207 func_id != BPF_FUNC_map_lookup_elem &&
5208 !may_update_sockmap(env, func_id))
5211 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
5212 if (func_id != BPF_FUNC_sk_select_reuseport)
5215 case BPF_MAP_TYPE_QUEUE:
5216 case BPF_MAP_TYPE_STACK:
5217 if (func_id != BPF_FUNC_map_peek_elem &&
5218 func_id != BPF_FUNC_map_pop_elem &&
5219 func_id != BPF_FUNC_map_push_elem)
5222 case BPF_MAP_TYPE_SK_STORAGE:
5223 if (func_id != BPF_FUNC_sk_storage_get &&
5224 func_id != BPF_FUNC_sk_storage_delete)
5227 case BPF_MAP_TYPE_INODE_STORAGE:
5228 if (func_id != BPF_FUNC_inode_storage_get &&
5229 func_id != BPF_FUNC_inode_storage_delete)
5232 case BPF_MAP_TYPE_TASK_STORAGE:
5233 if (func_id != BPF_FUNC_task_storage_get &&
5234 func_id != BPF_FUNC_task_storage_delete)
5241 /* ... and second from the function itself. */
5243 case BPF_FUNC_tail_call:
5244 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
5246 if (env->subprog_cnt > 1 && !allow_tail_call_in_subprogs(env)) {
5247 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
5251 case BPF_FUNC_perf_event_read:
5252 case BPF_FUNC_perf_event_output:
5253 case BPF_FUNC_perf_event_read_value:
5254 case BPF_FUNC_skb_output:
5255 case BPF_FUNC_xdp_output:
5256 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
5259 case BPF_FUNC_ringbuf_output:
5260 case BPF_FUNC_ringbuf_reserve:
5261 case BPF_FUNC_ringbuf_query:
5262 if (map->map_type != BPF_MAP_TYPE_RINGBUF)
5265 case BPF_FUNC_get_stackid:
5266 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
5269 case BPF_FUNC_current_task_under_cgroup:
5270 case BPF_FUNC_skb_under_cgroup:
5271 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
5274 case BPF_FUNC_redirect_map:
5275 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
5276 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
5277 map->map_type != BPF_MAP_TYPE_CPUMAP &&
5278 map->map_type != BPF_MAP_TYPE_XSKMAP)
5281 case BPF_FUNC_sk_redirect_map:
5282 case BPF_FUNC_msg_redirect_map:
5283 case BPF_FUNC_sock_map_update:
5284 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
5287 case BPF_FUNC_sk_redirect_hash:
5288 case BPF_FUNC_msg_redirect_hash:
5289 case BPF_FUNC_sock_hash_update:
5290 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
5293 case BPF_FUNC_get_local_storage:
5294 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
5295 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
5298 case BPF_FUNC_sk_select_reuseport:
5299 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
5300 map->map_type != BPF_MAP_TYPE_SOCKMAP &&
5301 map->map_type != BPF_MAP_TYPE_SOCKHASH)
5304 case BPF_FUNC_map_peek_elem:
5305 case BPF_FUNC_map_pop_elem:
5306 case BPF_FUNC_map_push_elem:
5307 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
5308 map->map_type != BPF_MAP_TYPE_STACK)
5311 case BPF_FUNC_sk_storage_get:
5312 case BPF_FUNC_sk_storage_delete:
5313 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
5316 case BPF_FUNC_inode_storage_get:
5317 case BPF_FUNC_inode_storage_delete:
5318 if (map->map_type != BPF_MAP_TYPE_INODE_STORAGE)
5321 case BPF_FUNC_task_storage_get:
5322 case BPF_FUNC_task_storage_delete:
5323 if (map->map_type != BPF_MAP_TYPE_TASK_STORAGE)
5332 verbose(env, "cannot pass map_type %d into func %s#%d\n",
5333 map->map_type, func_id_name(func_id), func_id);
5337 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
5341 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
5343 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
5345 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
5347 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
5349 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
5352 /* We only support one arg being in raw mode at the moment,
5353 * which is sufficient for the helper functions we have
5359 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
5360 enum bpf_arg_type arg_next)
5362 return (arg_type_is_mem_ptr(arg_curr) &&
5363 !arg_type_is_mem_size(arg_next)) ||
5364 (!arg_type_is_mem_ptr(arg_curr) &&
5365 arg_type_is_mem_size(arg_next));
5368 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
5370 /* bpf_xxx(..., buf, len) call will access 'len'
5371 * bytes from memory 'buf'. Both arg types need
5372 * to be paired, so make sure there's no buggy
5373 * helper function specification.
5375 if (arg_type_is_mem_size(fn->arg1_type) ||
5376 arg_type_is_mem_ptr(fn->arg5_type) ||
5377 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
5378 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
5379 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
5380 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
5386 static bool check_refcount_ok(const struct bpf_func_proto *fn, int func_id)
5390 if (arg_type_may_be_refcounted(fn->arg1_type))
5392 if (arg_type_may_be_refcounted(fn->arg2_type))
5394 if (arg_type_may_be_refcounted(fn->arg3_type))
5396 if (arg_type_may_be_refcounted(fn->arg4_type))
5398 if (arg_type_may_be_refcounted(fn->arg5_type))
5401 /* A reference acquiring function cannot acquire
5402 * another refcounted ptr.
5404 if (may_be_acquire_function(func_id) && count)
5407 /* We only support one arg being unreferenced at the moment,
5408 * which is sufficient for the helper functions we have right now.
5413 static bool check_btf_id_ok(const struct bpf_func_proto *fn)
5417 for (i = 0; i < ARRAY_SIZE(fn->arg_type); i++) {
5418 if (fn->arg_type[i] == ARG_PTR_TO_BTF_ID && !fn->arg_btf_id[i])
5421 if (fn->arg_type[i] != ARG_PTR_TO_BTF_ID && fn->arg_btf_id[i])
5428 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
5430 return check_raw_mode_ok(fn) &&
5431 check_arg_pair_ok(fn) &&
5432 check_btf_id_ok(fn) &&
5433 check_refcount_ok(fn, func_id) ? 0 : -EINVAL;
5436 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
5437 * are now invalid, so turn them into unknown SCALAR_VALUE.
5439 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
5440 struct bpf_func_state *state)
5442 struct bpf_reg_state *regs = state->regs, *reg;
5445 for (i = 0; i < MAX_BPF_REG; i++)
5446 if (reg_is_pkt_pointer_any(®s[i]))
5447 mark_reg_unknown(env, regs, i);
5449 bpf_for_each_spilled_reg(i, state, reg) {
5452 if (reg_is_pkt_pointer_any(reg))
5453 __mark_reg_unknown(env, reg);
5457 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
5459 struct bpf_verifier_state *vstate = env->cur_state;
5462 for (i = 0; i <= vstate->curframe; i++)
5463 __clear_all_pkt_pointers(env, vstate->frame[i]);
5468 BEYOND_PKT_END = -2,
5471 static void mark_pkt_end(struct bpf_verifier_state *vstate, int regn, bool range_open)
5473 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5474 struct bpf_reg_state *reg = &state->regs[regn];
5476 if (reg->type != PTR_TO_PACKET)
5477 /* PTR_TO_PACKET_META is not supported yet */
5480 /* The 'reg' is pkt > pkt_end or pkt >= pkt_end.
5481 * How far beyond pkt_end it goes is unknown.
5482 * if (!range_open) it's the case of pkt >= pkt_end
5483 * if (range_open) it's the case of pkt > pkt_end
5484 * hence this pointer is at least 1 byte bigger than pkt_end
5487 reg->range = BEYOND_PKT_END;
5489 reg->range = AT_PKT_END;
5492 static void release_reg_references(struct bpf_verifier_env *env,
5493 struct bpf_func_state *state,
5496 struct bpf_reg_state *regs = state->regs, *reg;
5499 for (i = 0; i < MAX_BPF_REG; i++)
5500 if (regs[i].ref_obj_id == ref_obj_id)
5501 mark_reg_unknown(env, regs, i);
5503 bpf_for_each_spilled_reg(i, state, reg) {
5506 if (reg->ref_obj_id == ref_obj_id)
5507 __mark_reg_unknown(env, reg);
5511 /* The pointer with the specified id has released its reference to kernel
5512 * resources. Identify all copies of the same pointer and clear the reference.
5514 static int release_reference(struct bpf_verifier_env *env,
5517 struct bpf_verifier_state *vstate = env->cur_state;
5521 err = release_reference_state(cur_func(env), ref_obj_id);
5525 for (i = 0; i <= vstate->curframe; i++)
5526 release_reg_references(env, vstate->frame[i], ref_obj_id);
5531 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
5532 struct bpf_reg_state *regs)
5536 /* after the call registers r0 - r5 were scratched */
5537 for (i = 0; i < CALLER_SAVED_REGS; i++) {
5538 mark_reg_not_init(env, regs, caller_saved[i]);
5539 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
5543 typedef int (*set_callee_state_fn)(struct bpf_verifier_env *env,
5544 struct bpf_func_state *caller,
5545 struct bpf_func_state *callee,
5548 static int __check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
5549 int *insn_idx, int subprog,
5550 set_callee_state_fn set_callee_state_cb)
5552 struct bpf_verifier_state *state = env->cur_state;
5553 struct bpf_func_info_aux *func_info_aux;
5554 struct bpf_func_state *caller, *callee;
5556 bool is_global = false;
5558 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
5559 verbose(env, "the call stack of %d frames is too deep\n",
5560 state->curframe + 2);
5564 caller = state->frame[state->curframe];
5565 if (state->frame[state->curframe + 1]) {
5566 verbose(env, "verifier bug. Frame %d already allocated\n",
5567 state->curframe + 1);
5571 func_info_aux = env->prog->aux->func_info_aux;
5573 is_global = func_info_aux[subprog].linkage == BTF_FUNC_GLOBAL;
5574 err = btf_check_subprog_arg_match(env, subprog, caller->regs);
5579 verbose(env, "Caller passes invalid args into func#%d\n",
5583 if (env->log.level & BPF_LOG_LEVEL)
5585 "Func#%d is global and valid. Skipping.\n",
5587 clear_caller_saved_regs(env, caller->regs);
5589 /* All global functions return a 64-bit SCALAR_VALUE */
5590 mark_reg_unknown(env, caller->regs, BPF_REG_0);
5591 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
5593 /* continue with next insn after call */
5598 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
5601 state->frame[state->curframe + 1] = callee;
5603 /* callee cannot access r0, r6 - r9 for reading and has to write
5604 * into its own stack before reading from it.
5605 * callee can read/write into caller's stack
5607 init_func_state(env, callee,
5608 /* remember the callsite, it will be used by bpf_exit */
5609 *insn_idx /* callsite */,
5610 state->curframe + 1 /* frameno within this callchain */,
5611 subprog /* subprog number within this prog */);
5613 /* Transfer references to the callee */
5614 err = transfer_reference_state(callee, caller);
5618 err = set_callee_state_cb(env, caller, callee, *insn_idx);
5622 clear_caller_saved_regs(env, caller->regs);
5624 /* only increment it after check_reg_arg() finished */
5627 /* and go analyze first insn of the callee */
5628 *insn_idx = env->subprog_info[subprog].start - 1;
5630 if (env->log.level & BPF_LOG_LEVEL) {
5631 verbose(env, "caller:\n");
5632 print_verifier_state(env, caller);
5633 verbose(env, "callee:\n");
5634 print_verifier_state(env, callee);
5639 int map_set_for_each_callback_args(struct bpf_verifier_env *env,
5640 struct bpf_func_state *caller,
5641 struct bpf_func_state *callee)
5643 /* bpf_for_each_map_elem(struct bpf_map *map, void *callback_fn,
5644 * void *callback_ctx, u64 flags);
5645 * callback_fn(struct bpf_map *map, void *key, void *value,
5646 * void *callback_ctx);
5648 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
5650 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
5651 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
5652 callee->regs[BPF_REG_2].map_ptr = caller->regs[BPF_REG_1].map_ptr;
5654 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
5655 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
5656 callee->regs[BPF_REG_3].map_ptr = caller->regs[BPF_REG_1].map_ptr;
5658 /* pointer to stack or null */
5659 callee->regs[BPF_REG_4] = caller->regs[BPF_REG_3];
5662 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
5666 static int set_callee_state(struct bpf_verifier_env *env,
5667 struct bpf_func_state *caller,
5668 struct bpf_func_state *callee, int insn_idx)
5672 /* copy r1 - r5 args that callee can access. The copy includes parent
5673 * pointers, which connects us up to the liveness chain
5675 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
5676 callee->regs[i] = caller->regs[i];
5680 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
5683 int subprog, target_insn;
5685 target_insn = *insn_idx + insn->imm + 1;
5686 subprog = find_subprog(env, target_insn);
5688 verbose(env, "verifier bug. No program starts at insn %d\n",
5693 return __check_func_call(env, insn, insn_idx, subprog, set_callee_state);
5696 static int set_map_elem_callback_state(struct bpf_verifier_env *env,
5697 struct bpf_func_state *caller,
5698 struct bpf_func_state *callee,
5701 struct bpf_insn_aux_data *insn_aux = &env->insn_aux_data[insn_idx];
5702 struct bpf_map *map;
5705 if (bpf_map_ptr_poisoned(insn_aux)) {
5706 verbose(env, "tail_call abusing map_ptr\n");
5710 map = BPF_MAP_PTR(insn_aux->map_ptr_state);
5711 if (!map->ops->map_set_for_each_callback_args ||
5712 !map->ops->map_for_each_callback) {
5713 verbose(env, "callback function not allowed for map\n");
5717 err = map->ops->map_set_for_each_callback_args(env, caller, callee);
5721 callee->in_callback_fn = true;
5725 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
5727 struct bpf_verifier_state *state = env->cur_state;
5728 struct bpf_func_state *caller, *callee;
5729 struct bpf_reg_state *r0;
5732 callee = state->frame[state->curframe];
5733 r0 = &callee->regs[BPF_REG_0];
5734 if (r0->type == PTR_TO_STACK) {
5735 /* technically it's ok to return caller's stack pointer
5736 * (or caller's caller's pointer) back to the caller,
5737 * since these pointers are valid. Only current stack
5738 * pointer will be invalid as soon as function exits,
5739 * but let's be conservative
5741 verbose(env, "cannot return stack pointer to the caller\n");
5746 caller = state->frame[state->curframe];
5747 if (callee->in_callback_fn) {
5748 /* enforce R0 return value range [0, 1]. */
5749 struct tnum range = tnum_range(0, 1);
5751 if (r0->type != SCALAR_VALUE) {
5752 verbose(env, "R0 not a scalar value\n");
5755 if (!tnum_in(range, r0->var_off)) {
5756 verbose_invalid_scalar(env, r0, &range, "callback return", "R0");
5760 /* return to the caller whatever r0 had in the callee */
5761 caller->regs[BPF_REG_0] = *r0;
5764 /* Transfer references to the caller */
5765 err = transfer_reference_state(caller, callee);
5769 *insn_idx = callee->callsite + 1;
5770 if (env->log.level & BPF_LOG_LEVEL) {
5771 verbose(env, "returning from callee:\n");
5772 print_verifier_state(env, callee);
5773 verbose(env, "to caller at %d:\n", *insn_idx);
5774 print_verifier_state(env, caller);
5776 /* clear everything in the callee */
5777 free_func_state(callee);
5778 state->frame[state->curframe + 1] = NULL;
5782 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
5784 struct bpf_call_arg_meta *meta)
5786 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
5788 if (ret_type != RET_INTEGER ||
5789 (func_id != BPF_FUNC_get_stack &&
5790 func_id != BPF_FUNC_get_task_stack &&
5791 func_id != BPF_FUNC_probe_read_str &&
5792 func_id != BPF_FUNC_probe_read_kernel_str &&
5793 func_id != BPF_FUNC_probe_read_user_str))
5796 ret_reg->smax_value = meta->msize_max_value;
5797 ret_reg->s32_max_value = meta->msize_max_value;
5798 ret_reg->smin_value = -MAX_ERRNO;
5799 ret_reg->s32_min_value = -MAX_ERRNO;
5800 __reg_deduce_bounds(ret_reg);
5801 __reg_bound_offset(ret_reg);
5802 __update_reg_bounds(ret_reg);
5806 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
5807 int func_id, int insn_idx)
5809 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
5810 struct bpf_map *map = meta->map_ptr;
5812 if (func_id != BPF_FUNC_tail_call &&
5813 func_id != BPF_FUNC_map_lookup_elem &&
5814 func_id != BPF_FUNC_map_update_elem &&
5815 func_id != BPF_FUNC_map_delete_elem &&
5816 func_id != BPF_FUNC_map_push_elem &&
5817 func_id != BPF_FUNC_map_pop_elem &&
5818 func_id != BPF_FUNC_map_peek_elem &&
5819 func_id != BPF_FUNC_for_each_map_elem &&
5820 func_id != BPF_FUNC_redirect_map)
5824 verbose(env, "kernel subsystem misconfigured verifier\n");
5828 /* In case of read-only, some additional restrictions
5829 * need to be applied in order to prevent altering the
5830 * state of the map from program side.
5832 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
5833 (func_id == BPF_FUNC_map_delete_elem ||
5834 func_id == BPF_FUNC_map_update_elem ||
5835 func_id == BPF_FUNC_map_push_elem ||
5836 func_id == BPF_FUNC_map_pop_elem)) {
5837 verbose(env, "write into map forbidden\n");
5841 if (!BPF_MAP_PTR(aux->map_ptr_state))
5842 bpf_map_ptr_store(aux, meta->map_ptr,
5843 !meta->map_ptr->bypass_spec_v1);
5844 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
5845 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
5846 !meta->map_ptr->bypass_spec_v1);
5851 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
5852 int func_id, int insn_idx)
5854 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
5855 struct bpf_reg_state *regs = cur_regs(env), *reg;
5856 struct bpf_map *map = meta->map_ptr;
5861 if (func_id != BPF_FUNC_tail_call)
5863 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
5864 verbose(env, "kernel subsystem misconfigured verifier\n");
5868 range = tnum_range(0, map->max_entries - 1);
5869 reg = ®s[BPF_REG_3];
5871 if (!register_is_const(reg) || !tnum_in(range, reg->var_off)) {
5872 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
5876 err = mark_chain_precision(env, BPF_REG_3);
5880 val = reg->var_off.value;
5881 if (bpf_map_key_unseen(aux))
5882 bpf_map_key_store(aux, val);
5883 else if (!bpf_map_key_poisoned(aux) &&
5884 bpf_map_key_immediate(aux) != val)
5885 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
5889 static int check_reference_leak(struct bpf_verifier_env *env)
5891 struct bpf_func_state *state = cur_func(env);
5894 for (i = 0; i < state->acquired_refs; i++) {
5895 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
5896 state->refs[i].id, state->refs[i].insn_idx);
5898 return state->acquired_refs ? -EINVAL : 0;
5901 static int check_bpf_snprintf_call(struct bpf_verifier_env *env,
5902 struct bpf_reg_state *regs)
5904 struct bpf_reg_state *fmt_reg = ®s[BPF_REG_3];
5905 struct bpf_reg_state *data_len_reg = ®s[BPF_REG_5];
5906 struct bpf_map *fmt_map = fmt_reg->map_ptr;
5907 int err, fmt_map_off, num_args;
5911 /* data must be an array of u64 */
5912 if (data_len_reg->var_off.value % 8)
5914 num_args = data_len_reg->var_off.value / 8;
5916 /* fmt being ARG_PTR_TO_CONST_STR guarantees that var_off is const
5917 * and map_direct_value_addr is set.
5919 fmt_map_off = fmt_reg->off + fmt_reg->var_off.value;
5920 err = fmt_map->ops->map_direct_value_addr(fmt_map, &fmt_addr,
5923 verbose(env, "verifier bug\n");
5926 fmt = (char *)(long)fmt_addr + fmt_map_off;
5928 /* We are also guaranteed that fmt+fmt_map_off is NULL terminated, we
5929 * can focus on validating the format specifiers.
5931 err = bpf_bprintf_prepare(fmt, UINT_MAX, NULL, NULL, num_args);
5933 verbose(env, "Invalid format string\n");
5938 static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
5941 const struct bpf_func_proto *fn = NULL;
5942 struct bpf_reg_state *regs;
5943 struct bpf_call_arg_meta meta;
5944 int insn_idx = *insn_idx_p;
5946 int i, err, func_id;
5948 /* find function prototype */
5949 func_id = insn->imm;
5950 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
5951 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
5956 if (env->ops->get_func_proto)
5957 fn = env->ops->get_func_proto(func_id, env->prog);
5959 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
5964 /* eBPF programs must be GPL compatible to use GPL-ed functions */
5965 if (!env->prog->gpl_compatible && fn->gpl_only) {
5966 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
5970 if (fn->allowed && !fn->allowed(env->prog)) {
5971 verbose(env, "helper call is not allowed in probe\n");
5975 /* With LD_ABS/IND some JITs save/restore skb from r1. */
5976 changes_data = bpf_helper_changes_pkt_data(fn->func);
5977 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
5978 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
5979 func_id_name(func_id), func_id);
5983 memset(&meta, 0, sizeof(meta));
5984 meta.pkt_access = fn->pkt_access;
5986 err = check_func_proto(fn, func_id);
5988 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
5989 func_id_name(func_id), func_id);
5993 meta.func_id = func_id;
5995 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++) {
5996 err = check_func_arg(env, i, &meta, fn);
6001 err = record_func_map(env, &meta, func_id, insn_idx);
6005 err = record_func_key(env, &meta, func_id, insn_idx);
6009 /* Mark slots with STACK_MISC in case of raw mode, stack offset
6010 * is inferred from register state.
6012 for (i = 0; i < meta.access_size; i++) {
6013 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
6014 BPF_WRITE, -1, false);
6019 if (func_id == BPF_FUNC_tail_call) {
6020 err = check_reference_leak(env);
6022 verbose(env, "tail_call would lead to reference leak\n");
6025 } else if (is_release_function(func_id)) {
6026 err = release_reference(env, meta.ref_obj_id);
6028 verbose(env, "func %s#%d reference has not been acquired before\n",
6029 func_id_name(func_id), func_id);
6034 regs = cur_regs(env);
6036 /* check that flags argument in get_local_storage(map, flags) is 0,
6037 * this is required because get_local_storage() can't return an error.
6039 if (func_id == BPF_FUNC_get_local_storage &&
6040 !register_is_null(®s[BPF_REG_2])) {
6041 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
6045 if (func_id == BPF_FUNC_for_each_map_elem) {
6046 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
6047 set_map_elem_callback_state);
6052 if (func_id == BPF_FUNC_snprintf) {
6053 err = check_bpf_snprintf_call(env, regs);
6058 /* reset caller saved regs */
6059 for (i = 0; i < CALLER_SAVED_REGS; i++) {
6060 mark_reg_not_init(env, regs, caller_saved[i]);
6061 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
6064 /* helper call returns 64-bit value. */
6065 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
6067 /* update return register (already marked as written above) */
6068 if (fn->ret_type == RET_INTEGER) {
6069 /* sets type to SCALAR_VALUE */
6070 mark_reg_unknown(env, regs, BPF_REG_0);
6071 } else if (fn->ret_type == RET_VOID) {
6072 regs[BPF_REG_0].type = NOT_INIT;
6073 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL ||
6074 fn->ret_type == RET_PTR_TO_MAP_VALUE) {
6075 /* There is no offset yet applied, variable or fixed */
6076 mark_reg_known_zero(env, regs, BPF_REG_0);
6077 /* remember map_ptr, so that check_map_access()
6078 * can check 'value_size' boundary of memory access
6079 * to map element returned from bpf_map_lookup_elem()
6081 if (meta.map_ptr == NULL) {
6083 "kernel subsystem misconfigured verifier\n");
6086 regs[BPF_REG_0].map_ptr = meta.map_ptr;
6087 if (fn->ret_type == RET_PTR_TO_MAP_VALUE) {
6088 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE;
6089 if (map_value_has_spin_lock(meta.map_ptr))
6090 regs[BPF_REG_0].id = ++env->id_gen;
6092 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
6094 } else if (fn->ret_type == RET_PTR_TO_SOCKET_OR_NULL) {
6095 mark_reg_known_zero(env, regs, BPF_REG_0);
6096 regs[BPF_REG_0].type = PTR_TO_SOCKET_OR_NULL;
6097 } else if (fn->ret_type == RET_PTR_TO_SOCK_COMMON_OR_NULL) {
6098 mark_reg_known_zero(env, regs, BPF_REG_0);
6099 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON_OR_NULL;
6100 } else if (fn->ret_type == RET_PTR_TO_TCP_SOCK_OR_NULL) {
6101 mark_reg_known_zero(env, regs, BPF_REG_0);
6102 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK_OR_NULL;
6103 } else if (fn->ret_type == RET_PTR_TO_ALLOC_MEM_OR_NULL) {
6104 mark_reg_known_zero(env, regs, BPF_REG_0);
6105 regs[BPF_REG_0].type = PTR_TO_MEM_OR_NULL;
6106 regs[BPF_REG_0].mem_size = meta.mem_size;
6107 } else if (fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID_OR_NULL ||
6108 fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID) {
6109 const struct btf_type *t;
6111 mark_reg_known_zero(env, regs, BPF_REG_0);
6112 t = btf_type_skip_modifiers(meta.ret_btf, meta.ret_btf_id, NULL);
6113 if (!btf_type_is_struct(t)) {
6115 const struct btf_type *ret;
6118 /* resolve the type size of ksym. */
6119 ret = btf_resolve_size(meta.ret_btf, t, &tsize);
6121 tname = btf_name_by_offset(meta.ret_btf, t->name_off);
6122 verbose(env, "unable to resolve the size of type '%s': %ld\n",
6123 tname, PTR_ERR(ret));
6126 regs[BPF_REG_0].type =
6127 fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID ?
6128 PTR_TO_MEM : PTR_TO_MEM_OR_NULL;
6129 regs[BPF_REG_0].mem_size = tsize;
6131 regs[BPF_REG_0].type =
6132 fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID ?
6133 PTR_TO_BTF_ID : PTR_TO_BTF_ID_OR_NULL;
6134 regs[BPF_REG_0].btf = meta.ret_btf;
6135 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
6137 } else if (fn->ret_type == RET_PTR_TO_BTF_ID_OR_NULL ||
6138 fn->ret_type == RET_PTR_TO_BTF_ID) {
6141 mark_reg_known_zero(env, regs, BPF_REG_0);
6142 regs[BPF_REG_0].type = fn->ret_type == RET_PTR_TO_BTF_ID ?
6144 PTR_TO_BTF_ID_OR_NULL;
6145 ret_btf_id = *fn->ret_btf_id;
6146 if (ret_btf_id == 0) {
6147 verbose(env, "invalid return type %d of func %s#%d\n",
6148 fn->ret_type, func_id_name(func_id), func_id);
6151 /* current BPF helper definitions are only coming from
6152 * built-in code with type IDs from vmlinux BTF
6154 regs[BPF_REG_0].btf = btf_vmlinux;
6155 regs[BPF_REG_0].btf_id = ret_btf_id;
6157 verbose(env, "unknown return type %d of func %s#%d\n",
6158 fn->ret_type, func_id_name(func_id), func_id);
6162 if (reg_type_may_be_null(regs[BPF_REG_0].type))
6163 regs[BPF_REG_0].id = ++env->id_gen;
6165 if (is_ptr_cast_function(func_id)) {
6166 /* For release_reference() */
6167 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
6168 } else if (is_acquire_function(func_id, meta.map_ptr)) {
6169 int id = acquire_reference_state(env, insn_idx);
6173 /* For mark_ptr_or_null_reg() */
6174 regs[BPF_REG_0].id = id;
6175 /* For release_reference() */
6176 regs[BPF_REG_0].ref_obj_id = id;
6179 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
6181 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
6185 if ((func_id == BPF_FUNC_get_stack ||
6186 func_id == BPF_FUNC_get_task_stack) &&
6187 !env->prog->has_callchain_buf) {
6188 const char *err_str;
6190 #ifdef CONFIG_PERF_EVENTS
6191 err = get_callchain_buffers(sysctl_perf_event_max_stack);
6192 err_str = "cannot get callchain buffer for func %s#%d\n";
6195 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
6198 verbose(env, err_str, func_id_name(func_id), func_id);
6202 env->prog->has_callchain_buf = true;
6205 if (func_id == BPF_FUNC_get_stackid || func_id == BPF_FUNC_get_stack)
6206 env->prog->call_get_stack = true;
6209 clear_all_pkt_pointers(env);
6213 /* mark_btf_func_reg_size() is used when the reg size is determined by
6214 * the BTF func_proto's return value size and argument.
6216 static void mark_btf_func_reg_size(struct bpf_verifier_env *env, u32 regno,
6219 struct bpf_reg_state *reg = &cur_regs(env)[regno];
6221 if (regno == BPF_REG_0) {
6222 /* Function return value */
6223 reg->live |= REG_LIVE_WRITTEN;
6224 reg->subreg_def = reg_size == sizeof(u64) ?
6225 DEF_NOT_SUBREG : env->insn_idx + 1;
6227 /* Function argument */
6228 if (reg_size == sizeof(u64)) {
6229 mark_insn_zext(env, reg);
6230 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
6232 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ32);
6237 static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn)
6239 const struct btf_type *t, *func, *func_proto, *ptr_type;
6240 struct bpf_reg_state *regs = cur_regs(env);
6241 const char *func_name, *ptr_type_name;
6242 u32 i, nargs, func_id, ptr_type_id;
6243 const struct btf_param *args;
6246 func_id = insn->imm;
6247 func = btf_type_by_id(btf_vmlinux, func_id);
6248 func_name = btf_name_by_offset(btf_vmlinux, func->name_off);
6249 func_proto = btf_type_by_id(btf_vmlinux, func->type);
6251 if (!env->ops->check_kfunc_call ||
6252 !env->ops->check_kfunc_call(func_id)) {
6253 verbose(env, "calling kernel function %s is not allowed\n",
6258 /* Check the arguments */
6259 err = btf_check_kfunc_arg_match(env, btf_vmlinux, func_id, regs);
6263 for (i = 0; i < CALLER_SAVED_REGS; i++)
6264 mark_reg_not_init(env, regs, caller_saved[i]);
6266 /* Check return type */
6267 t = btf_type_skip_modifiers(btf_vmlinux, func_proto->type, NULL);
6268 if (btf_type_is_scalar(t)) {
6269 mark_reg_unknown(env, regs, BPF_REG_0);
6270 mark_btf_func_reg_size(env, BPF_REG_0, t->size);
6271 } else if (btf_type_is_ptr(t)) {
6272 ptr_type = btf_type_skip_modifiers(btf_vmlinux, t->type,
6274 if (!btf_type_is_struct(ptr_type)) {
6275 ptr_type_name = btf_name_by_offset(btf_vmlinux,
6276 ptr_type->name_off);
6277 verbose(env, "kernel function %s returns pointer type %s %s is not supported\n",
6278 func_name, btf_type_str(ptr_type),
6282 mark_reg_known_zero(env, regs, BPF_REG_0);
6283 regs[BPF_REG_0].btf = btf_vmlinux;
6284 regs[BPF_REG_0].type = PTR_TO_BTF_ID;
6285 regs[BPF_REG_0].btf_id = ptr_type_id;
6286 mark_btf_func_reg_size(env, BPF_REG_0, sizeof(void *));
6287 } /* else { add_kfunc_call() ensures it is btf_type_is_void(t) } */
6289 nargs = btf_type_vlen(func_proto);
6290 args = (const struct btf_param *)(func_proto + 1);
6291 for (i = 0; i < nargs; i++) {
6294 t = btf_type_skip_modifiers(btf_vmlinux, args[i].type, NULL);
6295 if (btf_type_is_ptr(t))
6296 mark_btf_func_reg_size(env, regno, sizeof(void *));
6298 /* scalar. ensured by btf_check_kfunc_arg_match() */
6299 mark_btf_func_reg_size(env, regno, t->size);
6305 static bool signed_add_overflows(s64 a, s64 b)
6307 /* Do the add in u64, where overflow is well-defined */
6308 s64 res = (s64)((u64)a + (u64)b);
6315 static bool signed_add32_overflows(s32 a, s32 b)
6317 /* Do the add in u32, where overflow is well-defined */
6318 s32 res = (s32)((u32)a + (u32)b);
6325 static bool signed_sub_overflows(s64 a, s64 b)
6327 /* Do the sub in u64, where overflow is well-defined */
6328 s64 res = (s64)((u64)a - (u64)b);
6335 static bool signed_sub32_overflows(s32 a, s32 b)
6337 /* Do the sub in u32, where overflow is well-defined */
6338 s32 res = (s32)((u32)a - (u32)b);
6345 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
6346 const struct bpf_reg_state *reg,
6347 enum bpf_reg_type type)
6349 bool known = tnum_is_const(reg->var_off);
6350 s64 val = reg->var_off.value;
6351 s64 smin = reg->smin_value;
6353 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
6354 verbose(env, "math between %s pointer and %lld is not allowed\n",
6355 reg_type_str[type], val);
6359 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
6360 verbose(env, "%s pointer offset %d is not allowed\n",
6361 reg_type_str[type], reg->off);
6365 if (smin == S64_MIN) {
6366 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
6367 reg_type_str[type]);
6371 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
6372 verbose(env, "value %lld makes %s pointer be out of bounds\n",
6373 smin, reg_type_str[type]);
6380 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
6382 return &env->insn_aux_data[env->insn_idx];
6393 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
6394 u32 *alu_limit, bool mask_to_left)
6396 u32 max = 0, ptr_limit = 0;
6398 switch (ptr_reg->type) {
6400 /* Offset 0 is out-of-bounds, but acceptable start for the
6401 * left direction, see BPF_REG_FP. Also, unknown scalar
6402 * offset where we would need to deal with min/max bounds is
6403 * currently prohibited for unprivileged.
6405 max = MAX_BPF_STACK + mask_to_left;
6406 ptr_limit = -(ptr_reg->var_off.value + ptr_reg->off);
6408 case PTR_TO_MAP_VALUE:
6409 max = ptr_reg->map_ptr->value_size;
6410 ptr_limit = (mask_to_left ?
6411 ptr_reg->smin_value :
6412 ptr_reg->umax_value) + ptr_reg->off;
6418 if (ptr_limit >= max)
6419 return REASON_LIMIT;
6420 *alu_limit = ptr_limit;
6424 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
6425 const struct bpf_insn *insn)
6427 return env->bypass_spec_v1 || BPF_SRC(insn->code) == BPF_K;
6430 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
6431 u32 alu_state, u32 alu_limit)
6433 /* If we arrived here from different branches with different
6434 * state or limits to sanitize, then this won't work.
6436 if (aux->alu_state &&
6437 (aux->alu_state != alu_state ||
6438 aux->alu_limit != alu_limit))
6439 return REASON_PATHS;
6441 /* Corresponding fixup done in do_misc_fixups(). */
6442 aux->alu_state = alu_state;
6443 aux->alu_limit = alu_limit;
6447 static int sanitize_val_alu(struct bpf_verifier_env *env,
6448 struct bpf_insn *insn)
6450 struct bpf_insn_aux_data *aux = cur_aux(env);
6452 if (can_skip_alu_sanitation(env, insn))
6455 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
6458 static bool sanitize_needed(u8 opcode)
6460 return opcode == BPF_ADD || opcode == BPF_SUB;
6463 struct bpf_sanitize_info {
6464 struct bpf_insn_aux_data aux;
6468 static struct bpf_verifier_state *
6469 sanitize_speculative_path(struct bpf_verifier_env *env,
6470 const struct bpf_insn *insn,
6471 u32 next_idx, u32 curr_idx)
6473 struct bpf_verifier_state *branch;
6474 struct bpf_reg_state *regs;
6476 branch = push_stack(env, next_idx, curr_idx, true);
6477 if (branch && insn) {
6478 regs = branch->frame[branch->curframe]->regs;
6479 if (BPF_SRC(insn->code) == BPF_K) {
6480 mark_reg_unknown(env, regs, insn->dst_reg);
6481 } else if (BPF_SRC(insn->code) == BPF_X) {
6482 mark_reg_unknown(env, regs, insn->dst_reg);
6483 mark_reg_unknown(env, regs, insn->src_reg);
6489 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
6490 struct bpf_insn *insn,
6491 const struct bpf_reg_state *ptr_reg,
6492 const struct bpf_reg_state *off_reg,
6493 struct bpf_reg_state *dst_reg,
6494 struct bpf_sanitize_info *info,
6495 const bool commit_window)
6497 struct bpf_insn_aux_data *aux = commit_window ? cur_aux(env) : &info->aux;
6498 struct bpf_verifier_state *vstate = env->cur_state;
6499 bool off_is_imm = tnum_is_const(off_reg->var_off);
6500 bool off_is_neg = off_reg->smin_value < 0;
6501 bool ptr_is_dst_reg = ptr_reg == dst_reg;
6502 u8 opcode = BPF_OP(insn->code);
6503 u32 alu_state, alu_limit;
6504 struct bpf_reg_state tmp;
6508 if (can_skip_alu_sanitation(env, insn))
6511 /* We already marked aux for masking from non-speculative
6512 * paths, thus we got here in the first place. We only care
6513 * to explore bad access from here.
6515 if (vstate->speculative)
6518 if (!commit_window) {
6519 if (!tnum_is_const(off_reg->var_off) &&
6520 (off_reg->smin_value < 0) != (off_reg->smax_value < 0))
6521 return REASON_BOUNDS;
6523 info->mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
6524 (opcode == BPF_SUB && !off_is_neg);
6527 err = retrieve_ptr_limit(ptr_reg, &alu_limit, info->mask_to_left);
6531 if (commit_window) {
6532 /* In commit phase we narrow the masking window based on
6533 * the observed pointer move after the simulated operation.
6535 alu_state = info->aux.alu_state;
6536 alu_limit = abs(info->aux.alu_limit - alu_limit);
6538 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
6539 alu_state |= off_is_imm ? BPF_ALU_IMMEDIATE : 0;
6540 alu_state |= ptr_is_dst_reg ?
6541 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
6543 /* Limit pruning on unknown scalars to enable deep search for
6544 * potential masking differences from other program paths.
6547 env->explore_alu_limits = true;
6550 err = update_alu_sanitation_state(aux, alu_state, alu_limit);
6554 /* If we're in commit phase, we're done here given we already
6555 * pushed the truncated dst_reg into the speculative verification
6558 * Also, when register is a known constant, we rewrite register-based
6559 * operation to immediate-based, and thus do not need masking (and as
6560 * a consequence, do not need to simulate the zero-truncation either).
6562 if (commit_window || off_is_imm)
6565 /* Simulate and find potential out-of-bounds access under
6566 * speculative execution from truncation as a result of
6567 * masking when off was not within expected range. If off
6568 * sits in dst, then we temporarily need to move ptr there
6569 * to simulate dst (== 0) +/-= ptr. Needed, for example,
6570 * for cases where we use K-based arithmetic in one direction
6571 * and truncated reg-based in the other in order to explore
6574 if (!ptr_is_dst_reg) {
6576 *dst_reg = *ptr_reg;
6578 ret = sanitize_speculative_path(env, NULL, env->insn_idx + 1,
6580 if (!ptr_is_dst_reg && ret)
6582 return !ret ? REASON_STACK : 0;
6585 static void sanitize_mark_insn_seen(struct bpf_verifier_env *env)
6587 struct bpf_verifier_state *vstate = env->cur_state;
6589 /* If we simulate paths under speculation, we don't update the
6590 * insn as 'seen' such that when we verify unreachable paths in
6591 * the non-speculative domain, sanitize_dead_code() can still
6592 * rewrite/sanitize them.
6594 if (!vstate->speculative)
6595 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
6598 static int sanitize_err(struct bpf_verifier_env *env,
6599 const struct bpf_insn *insn, int reason,
6600 const struct bpf_reg_state *off_reg,
6601 const struct bpf_reg_state *dst_reg)
6603 static const char *err = "pointer arithmetic with it prohibited for !root";
6604 const char *op = BPF_OP(insn->code) == BPF_ADD ? "add" : "sub";
6605 u32 dst = insn->dst_reg, src = insn->src_reg;
6609 verbose(env, "R%d has unknown scalar with mixed signed bounds, %s\n",
6610 off_reg == dst_reg ? dst : src, err);
6613 verbose(env, "R%d has pointer with unsupported alu operation, %s\n",
6614 off_reg == dst_reg ? src : dst, err);
6617 verbose(env, "R%d tried to %s from different maps, paths or scalars, %s\n",
6621 verbose(env, "R%d tried to %s beyond pointer bounds, %s\n",
6625 verbose(env, "R%d could not be pushed for speculative verification, %s\n",
6629 verbose(env, "verifier internal error: unknown reason (%d)\n",
6637 /* check that stack access falls within stack limits and that 'reg' doesn't
6638 * have a variable offset.
6640 * Variable offset is prohibited for unprivileged mode for simplicity since it
6641 * requires corresponding support in Spectre masking for stack ALU. See also
6642 * retrieve_ptr_limit().
6645 * 'off' includes 'reg->off'.
6647 static int check_stack_access_for_ptr_arithmetic(
6648 struct bpf_verifier_env *env,
6650 const struct bpf_reg_state *reg,
6653 if (!tnum_is_const(reg->var_off)) {
6656 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6657 verbose(env, "R%d variable stack access prohibited for !root, var_off=%s off=%d\n",
6658 regno, tn_buf, off);
6662 if (off >= 0 || off < -MAX_BPF_STACK) {
6663 verbose(env, "R%d stack pointer arithmetic goes out of range, "
6664 "prohibited for !root; off=%d\n", regno, off);
6671 static int sanitize_check_bounds(struct bpf_verifier_env *env,
6672 const struct bpf_insn *insn,
6673 const struct bpf_reg_state *dst_reg)
6675 u32 dst = insn->dst_reg;
6677 /* For unprivileged we require that resulting offset must be in bounds
6678 * in order to be able to sanitize access later on.
6680 if (env->bypass_spec_v1)
6683 switch (dst_reg->type) {
6685 if (check_stack_access_for_ptr_arithmetic(env, dst, dst_reg,
6686 dst_reg->off + dst_reg->var_off.value))
6689 case PTR_TO_MAP_VALUE:
6690 if (check_map_access(env, dst, dst_reg->off, 1, false)) {
6691 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
6692 "prohibited for !root\n", dst);
6703 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
6704 * Caller should also handle BPF_MOV case separately.
6705 * If we return -EACCES, caller may want to try again treating pointer as a
6706 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
6708 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
6709 struct bpf_insn *insn,
6710 const struct bpf_reg_state *ptr_reg,
6711 const struct bpf_reg_state *off_reg)
6713 struct bpf_verifier_state *vstate = env->cur_state;
6714 struct bpf_func_state *state = vstate->frame[vstate->curframe];
6715 struct bpf_reg_state *regs = state->regs, *dst_reg;
6716 bool known = tnum_is_const(off_reg->var_off);
6717 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
6718 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
6719 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
6720 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
6721 struct bpf_sanitize_info info = {};
6722 u8 opcode = BPF_OP(insn->code);
6723 u32 dst = insn->dst_reg;
6726 dst_reg = ®s[dst];
6728 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
6729 smin_val > smax_val || umin_val > umax_val) {
6730 /* Taint dst register if offset had invalid bounds derived from
6731 * e.g. dead branches.
6733 __mark_reg_unknown(env, dst_reg);
6737 if (BPF_CLASS(insn->code) != BPF_ALU64) {
6738 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
6739 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
6740 __mark_reg_unknown(env, dst_reg);
6745 "R%d 32-bit pointer arithmetic prohibited\n",
6750 switch (ptr_reg->type) {
6751 case PTR_TO_MAP_VALUE_OR_NULL:
6752 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
6753 dst, reg_type_str[ptr_reg->type]);
6755 case CONST_PTR_TO_MAP:
6756 /* smin_val represents the known value */
6757 if (known && smin_val == 0 && opcode == BPF_ADD)
6760 case PTR_TO_PACKET_END:
6762 case PTR_TO_SOCKET_OR_NULL:
6763 case PTR_TO_SOCK_COMMON:
6764 case PTR_TO_SOCK_COMMON_OR_NULL:
6765 case PTR_TO_TCP_SOCK:
6766 case PTR_TO_TCP_SOCK_OR_NULL:
6767 case PTR_TO_XDP_SOCK:
6768 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
6769 dst, reg_type_str[ptr_reg->type]);
6775 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
6776 * The id may be overwritten later if we create a new variable offset.
6778 dst_reg->type = ptr_reg->type;
6779 dst_reg->id = ptr_reg->id;
6781 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
6782 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
6785 /* pointer types do not carry 32-bit bounds at the moment. */
6786 __mark_reg32_unbounded(dst_reg);
6788 if (sanitize_needed(opcode)) {
6789 ret = sanitize_ptr_alu(env, insn, ptr_reg, off_reg, dst_reg,
6792 return sanitize_err(env, insn, ret, off_reg, dst_reg);
6797 /* We can take a fixed offset as long as it doesn't overflow
6798 * the s32 'off' field
6800 if (known && (ptr_reg->off + smin_val ==
6801 (s64)(s32)(ptr_reg->off + smin_val))) {
6802 /* pointer += K. Accumulate it into fixed offset */
6803 dst_reg->smin_value = smin_ptr;
6804 dst_reg->smax_value = smax_ptr;
6805 dst_reg->umin_value = umin_ptr;
6806 dst_reg->umax_value = umax_ptr;
6807 dst_reg->var_off = ptr_reg->var_off;
6808 dst_reg->off = ptr_reg->off + smin_val;
6809 dst_reg->raw = ptr_reg->raw;
6812 /* A new variable offset is created. Note that off_reg->off
6813 * == 0, since it's a scalar.
6814 * dst_reg gets the pointer type and since some positive
6815 * integer value was added to the pointer, give it a new 'id'
6816 * if it's a PTR_TO_PACKET.
6817 * this creates a new 'base' pointer, off_reg (variable) gets
6818 * added into the variable offset, and we copy the fixed offset
6821 if (signed_add_overflows(smin_ptr, smin_val) ||
6822 signed_add_overflows(smax_ptr, smax_val)) {
6823 dst_reg->smin_value = S64_MIN;
6824 dst_reg->smax_value = S64_MAX;
6826 dst_reg->smin_value = smin_ptr + smin_val;
6827 dst_reg->smax_value = smax_ptr + smax_val;
6829 if (umin_ptr + umin_val < umin_ptr ||
6830 umax_ptr + umax_val < umax_ptr) {
6831 dst_reg->umin_value = 0;
6832 dst_reg->umax_value = U64_MAX;
6834 dst_reg->umin_value = umin_ptr + umin_val;
6835 dst_reg->umax_value = umax_ptr + umax_val;
6837 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
6838 dst_reg->off = ptr_reg->off;
6839 dst_reg->raw = ptr_reg->raw;
6840 if (reg_is_pkt_pointer(ptr_reg)) {
6841 dst_reg->id = ++env->id_gen;
6842 /* something was added to pkt_ptr, set range to zero */
6843 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
6847 if (dst_reg == off_reg) {
6848 /* scalar -= pointer. Creates an unknown scalar */
6849 verbose(env, "R%d tried to subtract pointer from scalar\n",
6853 /* We don't allow subtraction from FP, because (according to
6854 * test_verifier.c test "invalid fp arithmetic", JITs might not
6855 * be able to deal with it.
6857 if (ptr_reg->type == PTR_TO_STACK) {
6858 verbose(env, "R%d subtraction from stack pointer prohibited\n",
6862 if (known && (ptr_reg->off - smin_val ==
6863 (s64)(s32)(ptr_reg->off - smin_val))) {
6864 /* pointer -= K. Subtract it from fixed offset */
6865 dst_reg->smin_value = smin_ptr;
6866 dst_reg->smax_value = smax_ptr;
6867 dst_reg->umin_value = umin_ptr;
6868 dst_reg->umax_value = umax_ptr;
6869 dst_reg->var_off = ptr_reg->var_off;
6870 dst_reg->id = ptr_reg->id;
6871 dst_reg->off = ptr_reg->off - smin_val;
6872 dst_reg->raw = ptr_reg->raw;
6875 /* A new variable offset is created. If the subtrahend is known
6876 * nonnegative, then any reg->range we had before is still good.
6878 if (signed_sub_overflows(smin_ptr, smax_val) ||
6879 signed_sub_overflows(smax_ptr, smin_val)) {
6880 /* Overflow possible, we know nothing */
6881 dst_reg->smin_value = S64_MIN;
6882 dst_reg->smax_value = S64_MAX;
6884 dst_reg->smin_value = smin_ptr - smax_val;
6885 dst_reg->smax_value = smax_ptr - smin_val;
6887 if (umin_ptr < umax_val) {
6888 /* Overflow possible, we know nothing */
6889 dst_reg->umin_value = 0;
6890 dst_reg->umax_value = U64_MAX;
6892 /* Cannot overflow (as long as bounds are consistent) */
6893 dst_reg->umin_value = umin_ptr - umax_val;
6894 dst_reg->umax_value = umax_ptr - umin_val;
6896 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
6897 dst_reg->off = ptr_reg->off;
6898 dst_reg->raw = ptr_reg->raw;
6899 if (reg_is_pkt_pointer(ptr_reg)) {
6900 dst_reg->id = ++env->id_gen;
6901 /* something was added to pkt_ptr, set range to zero */
6903 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
6909 /* bitwise ops on pointers are troublesome, prohibit. */
6910 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
6911 dst, bpf_alu_string[opcode >> 4]);
6914 /* other operators (e.g. MUL,LSH) produce non-pointer results */
6915 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
6916 dst, bpf_alu_string[opcode >> 4]);
6920 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
6923 __update_reg_bounds(dst_reg);
6924 __reg_deduce_bounds(dst_reg);
6925 __reg_bound_offset(dst_reg);
6927 if (sanitize_check_bounds(env, insn, dst_reg) < 0)
6929 if (sanitize_needed(opcode)) {
6930 ret = sanitize_ptr_alu(env, insn, dst_reg, off_reg, dst_reg,
6933 return sanitize_err(env, insn, ret, off_reg, dst_reg);
6939 static void scalar32_min_max_add(struct bpf_reg_state *dst_reg,
6940 struct bpf_reg_state *src_reg)
6942 s32 smin_val = src_reg->s32_min_value;
6943 s32 smax_val = src_reg->s32_max_value;
6944 u32 umin_val = src_reg->u32_min_value;
6945 u32 umax_val = src_reg->u32_max_value;
6947 if (signed_add32_overflows(dst_reg->s32_min_value, smin_val) ||
6948 signed_add32_overflows(dst_reg->s32_max_value, smax_val)) {
6949 dst_reg->s32_min_value = S32_MIN;
6950 dst_reg->s32_max_value = S32_MAX;
6952 dst_reg->s32_min_value += smin_val;
6953 dst_reg->s32_max_value += smax_val;
6955 if (dst_reg->u32_min_value + umin_val < umin_val ||
6956 dst_reg->u32_max_value + umax_val < umax_val) {
6957 dst_reg->u32_min_value = 0;
6958 dst_reg->u32_max_value = U32_MAX;
6960 dst_reg->u32_min_value += umin_val;
6961 dst_reg->u32_max_value += umax_val;
6965 static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
6966 struct bpf_reg_state *src_reg)
6968 s64 smin_val = src_reg->smin_value;
6969 s64 smax_val = src_reg->smax_value;
6970 u64 umin_val = src_reg->umin_value;
6971 u64 umax_val = src_reg->umax_value;
6973 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
6974 signed_add_overflows(dst_reg->smax_value, smax_val)) {
6975 dst_reg->smin_value = S64_MIN;
6976 dst_reg->smax_value = S64_MAX;
6978 dst_reg->smin_value += smin_val;
6979 dst_reg->smax_value += smax_val;
6981 if (dst_reg->umin_value + umin_val < umin_val ||
6982 dst_reg->umax_value + umax_val < umax_val) {
6983 dst_reg->umin_value = 0;
6984 dst_reg->umax_value = U64_MAX;
6986 dst_reg->umin_value += umin_val;
6987 dst_reg->umax_value += umax_val;
6991 static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg,
6992 struct bpf_reg_state *src_reg)
6994 s32 smin_val = src_reg->s32_min_value;
6995 s32 smax_val = src_reg->s32_max_value;
6996 u32 umin_val = src_reg->u32_min_value;
6997 u32 umax_val = src_reg->u32_max_value;
6999 if (signed_sub32_overflows(dst_reg->s32_min_value, smax_val) ||
7000 signed_sub32_overflows(dst_reg->s32_max_value, smin_val)) {
7001 /* Overflow possible, we know nothing */
7002 dst_reg->s32_min_value = S32_MIN;
7003 dst_reg->s32_max_value = S32_MAX;
7005 dst_reg->s32_min_value -= smax_val;
7006 dst_reg->s32_max_value -= smin_val;
7008 if (dst_reg->u32_min_value < umax_val) {
7009 /* Overflow possible, we know nothing */
7010 dst_reg->u32_min_value = 0;
7011 dst_reg->u32_max_value = U32_MAX;
7013 /* Cannot overflow (as long as bounds are consistent) */
7014 dst_reg->u32_min_value -= umax_val;
7015 dst_reg->u32_max_value -= umin_val;
7019 static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
7020 struct bpf_reg_state *src_reg)
7022 s64 smin_val = src_reg->smin_value;
7023 s64 smax_val = src_reg->smax_value;
7024 u64 umin_val = src_reg->umin_value;
7025 u64 umax_val = src_reg->umax_value;
7027 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
7028 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
7029 /* Overflow possible, we know nothing */
7030 dst_reg->smin_value = S64_MIN;
7031 dst_reg->smax_value = S64_MAX;
7033 dst_reg->smin_value -= smax_val;
7034 dst_reg->smax_value -= smin_val;
7036 if (dst_reg->umin_value < umax_val) {
7037 /* Overflow possible, we know nothing */
7038 dst_reg->umin_value = 0;
7039 dst_reg->umax_value = U64_MAX;
7041 /* Cannot overflow (as long as bounds are consistent) */
7042 dst_reg->umin_value -= umax_val;
7043 dst_reg->umax_value -= umin_val;
7047 static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg,
7048 struct bpf_reg_state *src_reg)
7050 s32 smin_val = src_reg->s32_min_value;
7051 u32 umin_val = src_reg->u32_min_value;
7052 u32 umax_val = src_reg->u32_max_value;
7054 if (smin_val < 0 || dst_reg->s32_min_value < 0) {
7055 /* Ain't nobody got time to multiply that sign */
7056 __mark_reg32_unbounded(dst_reg);
7059 /* Both values are positive, so we can work with unsigned and
7060 * copy the result to signed (unless it exceeds S32_MAX).
7062 if (umax_val > U16_MAX || dst_reg->u32_max_value > U16_MAX) {
7063 /* Potential overflow, we know nothing */
7064 __mark_reg32_unbounded(dst_reg);
7067 dst_reg->u32_min_value *= umin_val;
7068 dst_reg->u32_max_value *= umax_val;
7069 if (dst_reg->u32_max_value > S32_MAX) {
7070 /* Overflow possible, we know nothing */
7071 dst_reg->s32_min_value = S32_MIN;
7072 dst_reg->s32_max_value = S32_MAX;
7074 dst_reg->s32_min_value = dst_reg->u32_min_value;
7075 dst_reg->s32_max_value = dst_reg->u32_max_value;
7079 static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
7080 struct bpf_reg_state *src_reg)
7082 s64 smin_val = src_reg->smin_value;
7083 u64 umin_val = src_reg->umin_value;
7084 u64 umax_val = src_reg->umax_value;
7086 if (smin_val < 0 || dst_reg->smin_value < 0) {
7087 /* Ain't nobody got time to multiply that sign */
7088 __mark_reg64_unbounded(dst_reg);
7091 /* Both values are positive, so we can work with unsigned and
7092 * copy the result to signed (unless it exceeds S64_MAX).
7094 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
7095 /* Potential overflow, we know nothing */
7096 __mark_reg64_unbounded(dst_reg);
7099 dst_reg->umin_value *= umin_val;
7100 dst_reg->umax_value *= umax_val;
7101 if (dst_reg->umax_value > S64_MAX) {
7102 /* Overflow possible, we know nothing */
7103 dst_reg->smin_value = S64_MIN;
7104 dst_reg->smax_value = S64_MAX;
7106 dst_reg->smin_value = dst_reg->umin_value;
7107 dst_reg->smax_value = dst_reg->umax_value;
7111 static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
7112 struct bpf_reg_state *src_reg)
7114 bool src_known = tnum_subreg_is_const(src_reg->var_off);
7115 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
7116 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
7117 s32 smin_val = src_reg->s32_min_value;
7118 u32 umax_val = src_reg->u32_max_value;
7120 if (src_known && dst_known) {
7121 __mark_reg32_known(dst_reg, var32_off.value);
7125 /* We get our minimum from the var_off, since that's inherently
7126 * bitwise. Our maximum is the minimum of the operands' maxima.
7128 dst_reg->u32_min_value = var32_off.value;
7129 dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val);
7130 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
7131 /* Lose signed bounds when ANDing negative numbers,
7132 * ain't nobody got time for that.
7134 dst_reg->s32_min_value = S32_MIN;
7135 dst_reg->s32_max_value = S32_MAX;
7137 /* ANDing two positives gives a positive, so safe to
7138 * cast result into s64.
7140 dst_reg->s32_min_value = dst_reg->u32_min_value;
7141 dst_reg->s32_max_value = dst_reg->u32_max_value;
7145 static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
7146 struct bpf_reg_state *src_reg)
7148 bool src_known = tnum_is_const(src_reg->var_off);
7149 bool dst_known = tnum_is_const(dst_reg->var_off);
7150 s64 smin_val = src_reg->smin_value;
7151 u64 umax_val = src_reg->umax_value;
7153 if (src_known && dst_known) {
7154 __mark_reg_known(dst_reg, dst_reg->var_off.value);
7158 /* We get our minimum from the var_off, since that's inherently
7159 * bitwise. Our maximum is the minimum of the operands' maxima.
7161 dst_reg->umin_value = dst_reg->var_off.value;
7162 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
7163 if (dst_reg->smin_value < 0 || smin_val < 0) {
7164 /* Lose signed bounds when ANDing negative numbers,
7165 * ain't nobody got time for that.
7167 dst_reg->smin_value = S64_MIN;
7168 dst_reg->smax_value = S64_MAX;
7170 /* ANDing two positives gives a positive, so safe to
7171 * cast result into s64.
7173 dst_reg->smin_value = dst_reg->umin_value;
7174 dst_reg->smax_value = dst_reg->umax_value;
7176 /* We may learn something more from the var_off */
7177 __update_reg_bounds(dst_reg);
7180 static void scalar32_min_max_or(struct bpf_reg_state *dst_reg,
7181 struct bpf_reg_state *src_reg)
7183 bool src_known = tnum_subreg_is_const(src_reg->var_off);
7184 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
7185 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
7186 s32 smin_val = src_reg->s32_min_value;
7187 u32 umin_val = src_reg->u32_min_value;
7189 if (src_known && dst_known) {
7190 __mark_reg32_known(dst_reg, var32_off.value);
7194 /* We get our maximum from the var_off, and our minimum is the
7195 * maximum of the operands' minima
7197 dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val);
7198 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
7199 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
7200 /* Lose signed bounds when ORing negative numbers,
7201 * ain't nobody got time for that.
7203 dst_reg->s32_min_value = S32_MIN;
7204 dst_reg->s32_max_value = S32_MAX;
7206 /* ORing two positives gives a positive, so safe to
7207 * cast result into s64.
7209 dst_reg->s32_min_value = dst_reg->u32_min_value;
7210 dst_reg->s32_max_value = dst_reg->u32_max_value;
7214 static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
7215 struct bpf_reg_state *src_reg)
7217 bool src_known = tnum_is_const(src_reg->var_off);
7218 bool dst_known = tnum_is_const(dst_reg->var_off);
7219 s64 smin_val = src_reg->smin_value;
7220 u64 umin_val = src_reg->umin_value;
7222 if (src_known && dst_known) {
7223 __mark_reg_known(dst_reg, dst_reg->var_off.value);
7227 /* We get our maximum from the var_off, and our minimum is the
7228 * maximum of the operands' minima
7230 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
7231 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
7232 if (dst_reg->smin_value < 0 || smin_val < 0) {
7233 /* Lose signed bounds when ORing negative numbers,
7234 * ain't nobody got time for that.
7236 dst_reg->smin_value = S64_MIN;
7237 dst_reg->smax_value = S64_MAX;
7239 /* ORing two positives gives a positive, so safe to
7240 * cast result into s64.
7242 dst_reg->smin_value = dst_reg->umin_value;
7243 dst_reg->smax_value = dst_reg->umax_value;
7245 /* We may learn something more from the var_off */
7246 __update_reg_bounds(dst_reg);
7249 static void scalar32_min_max_xor(struct bpf_reg_state *dst_reg,
7250 struct bpf_reg_state *src_reg)
7252 bool src_known = tnum_subreg_is_const(src_reg->var_off);
7253 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
7254 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
7255 s32 smin_val = src_reg->s32_min_value;
7257 if (src_known && dst_known) {
7258 __mark_reg32_known(dst_reg, var32_off.value);
7262 /* We get both minimum and maximum from the var32_off. */
7263 dst_reg->u32_min_value = var32_off.value;
7264 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
7266 if (dst_reg->s32_min_value >= 0 && smin_val >= 0) {
7267 /* XORing two positive sign numbers gives a positive,
7268 * so safe to cast u32 result into s32.
7270 dst_reg->s32_min_value = dst_reg->u32_min_value;
7271 dst_reg->s32_max_value = dst_reg->u32_max_value;
7273 dst_reg->s32_min_value = S32_MIN;
7274 dst_reg->s32_max_value = S32_MAX;
7278 static void scalar_min_max_xor(struct bpf_reg_state *dst_reg,
7279 struct bpf_reg_state *src_reg)
7281 bool src_known = tnum_is_const(src_reg->var_off);
7282 bool dst_known = tnum_is_const(dst_reg->var_off);
7283 s64 smin_val = src_reg->smin_value;
7285 if (src_known && dst_known) {
7286 /* dst_reg->var_off.value has been updated earlier */
7287 __mark_reg_known(dst_reg, dst_reg->var_off.value);
7291 /* We get both minimum and maximum from the var_off. */
7292 dst_reg->umin_value = dst_reg->var_off.value;
7293 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
7295 if (dst_reg->smin_value >= 0 && smin_val >= 0) {
7296 /* XORing two positive sign numbers gives a positive,
7297 * so safe to cast u64 result into s64.
7299 dst_reg->smin_value = dst_reg->umin_value;
7300 dst_reg->smax_value = dst_reg->umax_value;
7302 dst_reg->smin_value = S64_MIN;
7303 dst_reg->smax_value = S64_MAX;
7306 __update_reg_bounds(dst_reg);
7309 static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
7310 u64 umin_val, u64 umax_val)
7312 /* We lose all sign bit information (except what we can pick
7315 dst_reg->s32_min_value = S32_MIN;
7316 dst_reg->s32_max_value = S32_MAX;
7317 /* If we might shift our top bit out, then we know nothing */
7318 if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) {
7319 dst_reg->u32_min_value = 0;
7320 dst_reg->u32_max_value = U32_MAX;
7322 dst_reg->u32_min_value <<= umin_val;
7323 dst_reg->u32_max_value <<= umax_val;
7327 static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
7328 struct bpf_reg_state *src_reg)
7330 u32 umax_val = src_reg->u32_max_value;
7331 u32 umin_val = src_reg->u32_min_value;
7332 /* u32 alu operation will zext upper bits */
7333 struct tnum subreg = tnum_subreg(dst_reg->var_off);
7335 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
7336 dst_reg->var_off = tnum_subreg(tnum_lshift(subreg, umin_val));
7337 /* Not required but being careful mark reg64 bounds as unknown so
7338 * that we are forced to pick them up from tnum and zext later and
7339 * if some path skips this step we are still safe.
7341 __mark_reg64_unbounded(dst_reg);
7342 __update_reg32_bounds(dst_reg);
7345 static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg,
7346 u64 umin_val, u64 umax_val)
7348 /* Special case <<32 because it is a common compiler pattern to sign
7349 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
7350 * positive we know this shift will also be positive so we can track
7351 * bounds correctly. Otherwise we lose all sign bit information except
7352 * what we can pick up from var_off. Perhaps we can generalize this
7353 * later to shifts of any length.
7355 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0)
7356 dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32;
7358 dst_reg->smax_value = S64_MAX;
7360 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0)
7361 dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32;
7363 dst_reg->smin_value = S64_MIN;
7365 /* If we might shift our top bit out, then we know nothing */
7366 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
7367 dst_reg->umin_value = 0;
7368 dst_reg->umax_value = U64_MAX;
7370 dst_reg->umin_value <<= umin_val;
7371 dst_reg->umax_value <<= umax_val;
7375 static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
7376 struct bpf_reg_state *src_reg)
7378 u64 umax_val = src_reg->umax_value;
7379 u64 umin_val = src_reg->umin_value;
7381 /* scalar64 calc uses 32bit unshifted bounds so must be called first */
7382 __scalar64_min_max_lsh(dst_reg, umin_val, umax_val);
7383 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
7385 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
7386 /* We may learn something more from the var_off */
7387 __update_reg_bounds(dst_reg);
7390 static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg,
7391 struct bpf_reg_state *src_reg)
7393 struct tnum subreg = tnum_subreg(dst_reg->var_off);
7394 u32 umax_val = src_reg->u32_max_value;
7395 u32 umin_val = src_reg->u32_min_value;
7397 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
7398 * be negative, then either:
7399 * 1) src_reg might be zero, so the sign bit of the result is
7400 * unknown, so we lose our signed bounds
7401 * 2) it's known negative, thus the unsigned bounds capture the
7403 * 3) the signed bounds cross zero, so they tell us nothing
7405 * If the value in dst_reg is known nonnegative, then again the
7406 * unsigned bounds capture the signed bounds.
7407 * Thus, in all cases it suffices to blow away our signed bounds
7408 * and rely on inferring new ones from the unsigned bounds and
7409 * var_off of the result.
7411 dst_reg->s32_min_value = S32_MIN;
7412 dst_reg->s32_max_value = S32_MAX;
7414 dst_reg->var_off = tnum_rshift(subreg, umin_val);
7415 dst_reg->u32_min_value >>= umax_val;
7416 dst_reg->u32_max_value >>= umin_val;
7418 __mark_reg64_unbounded(dst_reg);
7419 __update_reg32_bounds(dst_reg);
7422 static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
7423 struct bpf_reg_state *src_reg)
7425 u64 umax_val = src_reg->umax_value;
7426 u64 umin_val = src_reg->umin_value;
7428 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
7429 * be negative, then either:
7430 * 1) src_reg might be zero, so the sign bit of the result is
7431 * unknown, so we lose our signed bounds
7432 * 2) it's known negative, thus the unsigned bounds capture the
7434 * 3) the signed bounds cross zero, so they tell us nothing
7436 * If the value in dst_reg is known nonnegative, then again the
7437 * unsigned bounds capture the signed bounds.
7438 * Thus, in all cases it suffices to blow away our signed bounds
7439 * and rely on inferring new ones from the unsigned bounds and
7440 * var_off of the result.
7442 dst_reg->smin_value = S64_MIN;
7443 dst_reg->smax_value = S64_MAX;
7444 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
7445 dst_reg->umin_value >>= umax_val;
7446 dst_reg->umax_value >>= umin_val;
7448 /* Its not easy to operate on alu32 bounds here because it depends
7449 * on bits being shifted in. Take easy way out and mark unbounded
7450 * so we can recalculate later from tnum.
7452 __mark_reg32_unbounded(dst_reg);
7453 __update_reg_bounds(dst_reg);
7456 static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg,
7457 struct bpf_reg_state *src_reg)
7459 u64 umin_val = src_reg->u32_min_value;
7461 /* Upon reaching here, src_known is true and
7462 * umax_val is equal to umin_val.
7464 dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val);
7465 dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val);
7467 dst_reg->var_off = tnum_arshift(tnum_subreg(dst_reg->var_off), umin_val, 32);
7469 /* blow away the dst_reg umin_value/umax_value and rely on
7470 * dst_reg var_off to refine the result.
7472 dst_reg->u32_min_value = 0;
7473 dst_reg->u32_max_value = U32_MAX;
7475 __mark_reg64_unbounded(dst_reg);
7476 __update_reg32_bounds(dst_reg);
7479 static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
7480 struct bpf_reg_state *src_reg)
7482 u64 umin_val = src_reg->umin_value;
7484 /* Upon reaching here, src_known is true and umax_val is equal
7487 dst_reg->smin_value >>= umin_val;
7488 dst_reg->smax_value >>= umin_val;
7490 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val, 64);
7492 /* blow away the dst_reg umin_value/umax_value and rely on
7493 * dst_reg var_off to refine the result.
7495 dst_reg->umin_value = 0;
7496 dst_reg->umax_value = U64_MAX;
7498 /* Its not easy to operate on alu32 bounds here because it depends
7499 * on bits being shifted in from upper 32-bits. Take easy way out
7500 * and mark unbounded so we can recalculate later from tnum.
7502 __mark_reg32_unbounded(dst_reg);
7503 __update_reg_bounds(dst_reg);
7506 /* WARNING: This function does calculations on 64-bit values, but the actual
7507 * execution may occur on 32-bit values. Therefore, things like bitshifts
7508 * need extra checks in the 32-bit case.
7510 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
7511 struct bpf_insn *insn,
7512 struct bpf_reg_state *dst_reg,
7513 struct bpf_reg_state src_reg)
7515 struct bpf_reg_state *regs = cur_regs(env);
7516 u8 opcode = BPF_OP(insn->code);
7518 s64 smin_val, smax_val;
7519 u64 umin_val, umax_val;
7520 s32 s32_min_val, s32_max_val;
7521 u32 u32_min_val, u32_max_val;
7522 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
7523 bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
7526 smin_val = src_reg.smin_value;
7527 smax_val = src_reg.smax_value;
7528 umin_val = src_reg.umin_value;
7529 umax_val = src_reg.umax_value;
7531 s32_min_val = src_reg.s32_min_value;
7532 s32_max_val = src_reg.s32_max_value;
7533 u32_min_val = src_reg.u32_min_value;
7534 u32_max_val = src_reg.u32_max_value;
7537 src_known = tnum_subreg_is_const(src_reg.var_off);
7539 (s32_min_val != s32_max_val || u32_min_val != u32_max_val)) ||
7540 s32_min_val > s32_max_val || u32_min_val > u32_max_val) {
7541 /* Taint dst register if offset had invalid bounds
7542 * derived from e.g. dead branches.
7544 __mark_reg_unknown(env, dst_reg);
7548 src_known = tnum_is_const(src_reg.var_off);
7550 (smin_val != smax_val || umin_val != umax_val)) ||
7551 smin_val > smax_val || umin_val > umax_val) {
7552 /* Taint dst register if offset had invalid bounds
7553 * derived from e.g. dead branches.
7555 __mark_reg_unknown(env, dst_reg);
7561 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
7562 __mark_reg_unknown(env, dst_reg);
7566 if (sanitize_needed(opcode)) {
7567 ret = sanitize_val_alu(env, insn);
7569 return sanitize_err(env, insn, ret, NULL, NULL);
7572 /* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
7573 * There are two classes of instructions: The first class we track both
7574 * alu32 and alu64 sign/unsigned bounds independently this provides the
7575 * greatest amount of precision when alu operations are mixed with jmp32
7576 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
7577 * and BPF_OR. This is possible because these ops have fairly easy to
7578 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
7579 * See alu32 verifier tests for examples. The second class of
7580 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
7581 * with regards to tracking sign/unsigned bounds because the bits may
7582 * cross subreg boundaries in the alu64 case. When this happens we mark
7583 * the reg unbounded in the subreg bound space and use the resulting
7584 * tnum to calculate an approximation of the sign/unsigned bounds.
7588 scalar32_min_max_add(dst_reg, &src_reg);
7589 scalar_min_max_add(dst_reg, &src_reg);
7590 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
7593 scalar32_min_max_sub(dst_reg, &src_reg);
7594 scalar_min_max_sub(dst_reg, &src_reg);
7595 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
7598 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
7599 scalar32_min_max_mul(dst_reg, &src_reg);
7600 scalar_min_max_mul(dst_reg, &src_reg);
7603 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
7604 scalar32_min_max_and(dst_reg, &src_reg);
7605 scalar_min_max_and(dst_reg, &src_reg);
7608 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
7609 scalar32_min_max_or(dst_reg, &src_reg);
7610 scalar_min_max_or(dst_reg, &src_reg);
7613 dst_reg->var_off = tnum_xor(dst_reg->var_off, src_reg.var_off);
7614 scalar32_min_max_xor(dst_reg, &src_reg);
7615 scalar_min_max_xor(dst_reg, &src_reg);
7618 if (umax_val >= insn_bitness) {
7619 /* Shifts greater than 31 or 63 are undefined.
7620 * This includes shifts by a negative number.
7622 mark_reg_unknown(env, regs, insn->dst_reg);
7626 scalar32_min_max_lsh(dst_reg, &src_reg);
7628 scalar_min_max_lsh(dst_reg, &src_reg);
7631 if (umax_val >= insn_bitness) {
7632 /* Shifts greater than 31 or 63 are undefined.
7633 * This includes shifts by a negative number.
7635 mark_reg_unknown(env, regs, insn->dst_reg);
7639 scalar32_min_max_rsh(dst_reg, &src_reg);
7641 scalar_min_max_rsh(dst_reg, &src_reg);
7644 if (umax_val >= insn_bitness) {
7645 /* Shifts greater than 31 or 63 are undefined.
7646 * This includes shifts by a negative number.
7648 mark_reg_unknown(env, regs, insn->dst_reg);
7652 scalar32_min_max_arsh(dst_reg, &src_reg);
7654 scalar_min_max_arsh(dst_reg, &src_reg);
7657 mark_reg_unknown(env, regs, insn->dst_reg);
7661 /* ALU32 ops are zero extended into 64bit register */
7663 zext_32_to_64(dst_reg);
7665 __update_reg_bounds(dst_reg);
7666 __reg_deduce_bounds(dst_reg);
7667 __reg_bound_offset(dst_reg);
7671 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
7674 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
7675 struct bpf_insn *insn)
7677 struct bpf_verifier_state *vstate = env->cur_state;
7678 struct bpf_func_state *state = vstate->frame[vstate->curframe];
7679 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
7680 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
7681 u8 opcode = BPF_OP(insn->code);
7684 dst_reg = ®s[insn->dst_reg];
7686 if (dst_reg->type != SCALAR_VALUE)
7689 /* Make sure ID is cleared otherwise dst_reg min/max could be
7690 * incorrectly propagated into other registers by find_equal_scalars()
7693 if (BPF_SRC(insn->code) == BPF_X) {
7694 src_reg = ®s[insn->src_reg];
7695 if (src_reg->type != SCALAR_VALUE) {
7696 if (dst_reg->type != SCALAR_VALUE) {
7697 /* Combining two pointers by any ALU op yields
7698 * an arbitrary scalar. Disallow all math except
7699 * pointer subtraction
7701 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
7702 mark_reg_unknown(env, regs, insn->dst_reg);
7705 verbose(env, "R%d pointer %s pointer prohibited\n",
7707 bpf_alu_string[opcode >> 4]);
7710 /* scalar += pointer
7711 * This is legal, but we have to reverse our
7712 * src/dest handling in computing the range
7714 err = mark_chain_precision(env, insn->dst_reg);
7717 return adjust_ptr_min_max_vals(env, insn,
7720 } else if (ptr_reg) {
7721 /* pointer += scalar */
7722 err = mark_chain_precision(env, insn->src_reg);
7725 return adjust_ptr_min_max_vals(env, insn,
7729 /* Pretend the src is a reg with a known value, since we only
7730 * need to be able to read from this state.
7732 off_reg.type = SCALAR_VALUE;
7733 __mark_reg_known(&off_reg, insn->imm);
7735 if (ptr_reg) /* pointer += K */
7736 return adjust_ptr_min_max_vals(env, insn,
7740 /* Got here implies adding two SCALAR_VALUEs */
7741 if (WARN_ON_ONCE(ptr_reg)) {
7742 print_verifier_state(env, state);
7743 verbose(env, "verifier internal error: unexpected ptr_reg\n");
7746 if (WARN_ON(!src_reg)) {
7747 print_verifier_state(env, state);
7748 verbose(env, "verifier internal error: no src_reg\n");
7751 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
7754 /* check validity of 32-bit and 64-bit arithmetic operations */
7755 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
7757 struct bpf_reg_state *regs = cur_regs(env);
7758 u8 opcode = BPF_OP(insn->code);
7761 if (opcode == BPF_END || opcode == BPF_NEG) {
7762 if (opcode == BPF_NEG) {
7763 if (BPF_SRC(insn->code) != 0 ||
7764 insn->src_reg != BPF_REG_0 ||
7765 insn->off != 0 || insn->imm != 0) {
7766 verbose(env, "BPF_NEG uses reserved fields\n");
7770 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
7771 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
7772 BPF_CLASS(insn->code) == BPF_ALU64) {
7773 verbose(env, "BPF_END uses reserved fields\n");
7778 /* check src operand */
7779 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
7783 if (is_pointer_value(env, insn->dst_reg)) {
7784 verbose(env, "R%d pointer arithmetic prohibited\n",
7789 /* check dest operand */
7790 err = check_reg_arg(env, insn->dst_reg, DST_OP);
7794 } else if (opcode == BPF_MOV) {
7796 if (BPF_SRC(insn->code) == BPF_X) {
7797 if (insn->imm != 0 || insn->off != 0) {
7798 verbose(env, "BPF_MOV uses reserved fields\n");
7802 /* check src operand */
7803 err = check_reg_arg(env, insn->src_reg, SRC_OP);
7807 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
7808 verbose(env, "BPF_MOV uses reserved fields\n");
7813 /* check dest operand, mark as required later */
7814 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
7818 if (BPF_SRC(insn->code) == BPF_X) {
7819 struct bpf_reg_state *src_reg = regs + insn->src_reg;
7820 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
7822 if (BPF_CLASS(insn->code) == BPF_ALU64) {
7824 * copy register state to dest reg
7826 if (src_reg->type == SCALAR_VALUE && !src_reg->id)
7827 /* Assign src and dst registers the same ID
7828 * that will be used by find_equal_scalars()
7829 * to propagate min/max range.
7831 src_reg->id = ++env->id_gen;
7832 *dst_reg = *src_reg;
7833 dst_reg->live |= REG_LIVE_WRITTEN;
7834 dst_reg->subreg_def = DEF_NOT_SUBREG;
7837 if (is_pointer_value(env, insn->src_reg)) {
7839 "R%d partial copy of pointer\n",
7842 } else if (src_reg->type == SCALAR_VALUE) {
7843 *dst_reg = *src_reg;
7844 /* Make sure ID is cleared otherwise
7845 * dst_reg min/max could be incorrectly
7846 * propagated into src_reg by find_equal_scalars()
7849 dst_reg->live |= REG_LIVE_WRITTEN;
7850 dst_reg->subreg_def = env->insn_idx + 1;
7852 mark_reg_unknown(env, regs,
7855 zext_32_to_64(dst_reg);
7859 * remember the value we stored into this reg
7861 /* clear any state __mark_reg_known doesn't set */
7862 mark_reg_unknown(env, regs, insn->dst_reg);
7863 regs[insn->dst_reg].type = SCALAR_VALUE;
7864 if (BPF_CLASS(insn->code) == BPF_ALU64) {
7865 __mark_reg_known(regs + insn->dst_reg,
7868 __mark_reg_known(regs + insn->dst_reg,
7873 } else if (opcode > BPF_END) {
7874 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
7877 } else { /* all other ALU ops: and, sub, xor, add, ... */
7879 if (BPF_SRC(insn->code) == BPF_X) {
7880 if (insn->imm != 0 || insn->off != 0) {
7881 verbose(env, "BPF_ALU uses reserved fields\n");
7884 /* check src1 operand */
7885 err = check_reg_arg(env, insn->src_reg, SRC_OP);
7889 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
7890 verbose(env, "BPF_ALU uses reserved fields\n");
7895 /* check src2 operand */
7896 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
7900 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
7901 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
7902 verbose(env, "div by zero\n");
7906 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
7907 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
7908 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
7910 if (insn->imm < 0 || insn->imm >= size) {
7911 verbose(env, "invalid shift %d\n", insn->imm);
7916 /* check dest operand */
7917 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
7921 return adjust_reg_min_max_vals(env, insn);
7927 static void __find_good_pkt_pointers(struct bpf_func_state *state,
7928 struct bpf_reg_state *dst_reg,
7929 enum bpf_reg_type type, int new_range)
7931 struct bpf_reg_state *reg;
7934 for (i = 0; i < MAX_BPF_REG; i++) {
7935 reg = &state->regs[i];
7936 if (reg->type == type && reg->id == dst_reg->id)
7937 /* keep the maximum range already checked */
7938 reg->range = max(reg->range, new_range);
7941 bpf_for_each_spilled_reg(i, state, reg) {
7944 if (reg->type == type && reg->id == dst_reg->id)
7945 reg->range = max(reg->range, new_range);
7949 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
7950 struct bpf_reg_state *dst_reg,
7951 enum bpf_reg_type type,
7952 bool range_right_open)
7956 if (dst_reg->off < 0 ||
7957 (dst_reg->off == 0 && range_right_open))
7958 /* This doesn't give us any range */
7961 if (dst_reg->umax_value > MAX_PACKET_OFF ||
7962 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
7963 /* Risk of overflow. For instance, ptr + (1<<63) may be less
7964 * than pkt_end, but that's because it's also less than pkt.
7968 new_range = dst_reg->off;
7969 if (range_right_open)
7972 /* Examples for register markings:
7974 * pkt_data in dst register:
7978 * if (r2 > pkt_end) goto <handle exception>
7983 * if (r2 < pkt_end) goto <access okay>
7984 * <handle exception>
7987 * r2 == dst_reg, pkt_end == src_reg
7988 * r2=pkt(id=n,off=8,r=0)
7989 * r3=pkt(id=n,off=0,r=0)
7991 * pkt_data in src register:
7995 * if (pkt_end >= r2) goto <access okay>
7996 * <handle exception>
8000 * if (pkt_end <= r2) goto <handle exception>
8004 * pkt_end == dst_reg, r2 == src_reg
8005 * r2=pkt(id=n,off=8,r=0)
8006 * r3=pkt(id=n,off=0,r=0)
8008 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
8009 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
8010 * and [r3, r3 + 8-1) respectively is safe to access depending on
8014 /* If our ids match, then we must have the same max_value. And we
8015 * don't care about the other reg's fixed offset, since if it's too big
8016 * the range won't allow anything.
8017 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
8019 for (i = 0; i <= vstate->curframe; i++)
8020 __find_good_pkt_pointers(vstate->frame[i], dst_reg, type,
8024 static int is_branch32_taken(struct bpf_reg_state *reg, u32 val, u8 opcode)
8026 struct tnum subreg = tnum_subreg(reg->var_off);
8027 s32 sval = (s32)val;
8031 if (tnum_is_const(subreg))
8032 return !!tnum_equals_const(subreg, val);
8035 if (tnum_is_const(subreg))
8036 return !tnum_equals_const(subreg, val);
8039 if ((~subreg.mask & subreg.value) & val)
8041 if (!((subreg.mask | subreg.value) & val))
8045 if (reg->u32_min_value > val)
8047 else if (reg->u32_max_value <= val)
8051 if (reg->s32_min_value > sval)
8053 else if (reg->s32_max_value <= sval)
8057 if (reg->u32_max_value < val)
8059 else if (reg->u32_min_value >= val)
8063 if (reg->s32_max_value < sval)
8065 else if (reg->s32_min_value >= sval)
8069 if (reg->u32_min_value >= val)
8071 else if (reg->u32_max_value < val)
8075 if (reg->s32_min_value >= sval)
8077 else if (reg->s32_max_value < sval)
8081 if (reg->u32_max_value <= val)
8083 else if (reg->u32_min_value > val)
8087 if (reg->s32_max_value <= sval)
8089 else if (reg->s32_min_value > sval)
8098 static int is_branch64_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
8100 s64 sval = (s64)val;
8104 if (tnum_is_const(reg->var_off))
8105 return !!tnum_equals_const(reg->var_off, val);
8108 if (tnum_is_const(reg->var_off))
8109 return !tnum_equals_const(reg->var_off, val);
8112 if ((~reg->var_off.mask & reg->var_off.value) & val)
8114 if (!((reg->var_off.mask | reg->var_off.value) & val))
8118 if (reg->umin_value > val)
8120 else if (reg->umax_value <= val)
8124 if (reg->smin_value > sval)
8126 else if (reg->smax_value <= sval)
8130 if (reg->umax_value < val)
8132 else if (reg->umin_value >= val)
8136 if (reg->smax_value < sval)
8138 else if (reg->smin_value >= sval)
8142 if (reg->umin_value >= val)
8144 else if (reg->umax_value < val)
8148 if (reg->smin_value >= sval)
8150 else if (reg->smax_value < sval)
8154 if (reg->umax_value <= val)
8156 else if (reg->umin_value > val)
8160 if (reg->smax_value <= sval)
8162 else if (reg->smin_value > sval)
8170 /* compute branch direction of the expression "if (reg opcode val) goto target;"
8172 * 1 - branch will be taken and "goto target" will be executed
8173 * 0 - branch will not be taken and fall-through to next insn
8174 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value
8177 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
8180 if (__is_pointer_value(false, reg)) {
8181 if (!reg_type_not_null(reg->type))
8184 /* If pointer is valid tests against zero will fail so we can
8185 * use this to direct branch taken.
8201 return is_branch32_taken(reg, val, opcode);
8202 return is_branch64_taken(reg, val, opcode);
8205 static int flip_opcode(u32 opcode)
8207 /* How can we transform "a <op> b" into "b <op> a"? */
8208 static const u8 opcode_flip[16] = {
8209 /* these stay the same */
8210 [BPF_JEQ >> 4] = BPF_JEQ,
8211 [BPF_JNE >> 4] = BPF_JNE,
8212 [BPF_JSET >> 4] = BPF_JSET,
8213 /* these swap "lesser" and "greater" (L and G in the opcodes) */
8214 [BPF_JGE >> 4] = BPF_JLE,
8215 [BPF_JGT >> 4] = BPF_JLT,
8216 [BPF_JLE >> 4] = BPF_JGE,
8217 [BPF_JLT >> 4] = BPF_JGT,
8218 [BPF_JSGE >> 4] = BPF_JSLE,
8219 [BPF_JSGT >> 4] = BPF_JSLT,
8220 [BPF_JSLE >> 4] = BPF_JSGE,
8221 [BPF_JSLT >> 4] = BPF_JSGT
8223 return opcode_flip[opcode >> 4];
8226 static int is_pkt_ptr_branch_taken(struct bpf_reg_state *dst_reg,
8227 struct bpf_reg_state *src_reg,
8230 struct bpf_reg_state *pkt;
8232 if (src_reg->type == PTR_TO_PACKET_END) {
8234 } else if (dst_reg->type == PTR_TO_PACKET_END) {
8236 opcode = flip_opcode(opcode);
8241 if (pkt->range >= 0)
8246 /* pkt <= pkt_end */
8250 if (pkt->range == BEYOND_PKT_END)
8251 /* pkt has at last one extra byte beyond pkt_end */
8252 return opcode == BPF_JGT;
8258 /* pkt >= pkt_end */
8259 if (pkt->range == BEYOND_PKT_END || pkt->range == AT_PKT_END)
8260 return opcode == BPF_JGE;
8266 /* Adjusts the register min/max values in the case that the dst_reg is the
8267 * variable register that we are working on, and src_reg is a constant or we're
8268 * simply doing a BPF_K check.
8269 * In JEQ/JNE cases we also adjust the var_off values.
8271 static void reg_set_min_max(struct bpf_reg_state *true_reg,
8272 struct bpf_reg_state *false_reg,
8274 u8 opcode, bool is_jmp32)
8276 struct tnum false_32off = tnum_subreg(false_reg->var_off);
8277 struct tnum false_64off = false_reg->var_off;
8278 struct tnum true_32off = tnum_subreg(true_reg->var_off);
8279 struct tnum true_64off = true_reg->var_off;
8280 s64 sval = (s64)val;
8281 s32 sval32 = (s32)val32;
8283 /* If the dst_reg is a pointer, we can't learn anything about its
8284 * variable offset from the compare (unless src_reg were a pointer into
8285 * the same object, but we don't bother with that.
8286 * Since false_reg and true_reg have the same type by construction, we
8287 * only need to check one of them for pointerness.
8289 if (__is_pointer_value(false, false_reg))
8296 struct bpf_reg_state *reg =
8297 opcode == BPF_JEQ ? true_reg : false_reg;
8299 /* JEQ/JNE comparison doesn't change the register equivalence.
8301 * if (r1 == 42) goto label;
8303 * label: // here both r1 and r2 are known to be 42.
8305 * Hence when marking register as known preserve it's ID.
8308 __mark_reg32_known(reg, val32);
8310 ___mark_reg_known(reg, val);
8315 false_32off = tnum_and(false_32off, tnum_const(~val32));
8316 if (is_power_of_2(val32))
8317 true_32off = tnum_or(true_32off,
8320 false_64off = tnum_and(false_64off, tnum_const(~val));
8321 if (is_power_of_2(val))
8322 true_64off = tnum_or(true_64off,
8330 u32 false_umax = opcode == BPF_JGT ? val32 : val32 - 1;
8331 u32 true_umin = opcode == BPF_JGT ? val32 + 1 : val32;
8333 false_reg->u32_max_value = min(false_reg->u32_max_value,
8335 true_reg->u32_min_value = max(true_reg->u32_min_value,
8338 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
8339 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
8341 false_reg->umax_value = min(false_reg->umax_value, false_umax);
8342 true_reg->umin_value = max(true_reg->umin_value, true_umin);
8350 s32 false_smax = opcode == BPF_JSGT ? sval32 : sval32 - 1;
8351 s32 true_smin = opcode == BPF_JSGT ? sval32 + 1 : sval32;
8353 false_reg->s32_max_value = min(false_reg->s32_max_value, false_smax);
8354 true_reg->s32_min_value = max(true_reg->s32_min_value, true_smin);
8356 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
8357 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
8359 false_reg->smax_value = min(false_reg->smax_value, false_smax);
8360 true_reg->smin_value = max(true_reg->smin_value, true_smin);
8368 u32 false_umin = opcode == BPF_JLT ? val32 : val32 + 1;
8369 u32 true_umax = opcode == BPF_JLT ? val32 - 1 : val32;
8371 false_reg->u32_min_value = max(false_reg->u32_min_value,
8373 true_reg->u32_max_value = min(true_reg->u32_max_value,
8376 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
8377 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
8379 false_reg->umin_value = max(false_reg->umin_value, false_umin);
8380 true_reg->umax_value = min(true_reg->umax_value, true_umax);
8388 s32 false_smin = opcode == BPF_JSLT ? sval32 : sval32 + 1;
8389 s32 true_smax = opcode == BPF_JSLT ? sval32 - 1 : sval32;
8391 false_reg->s32_min_value = max(false_reg->s32_min_value, false_smin);
8392 true_reg->s32_max_value = min(true_reg->s32_max_value, true_smax);
8394 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
8395 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
8397 false_reg->smin_value = max(false_reg->smin_value, false_smin);
8398 true_reg->smax_value = min(true_reg->smax_value, true_smax);
8407 false_reg->var_off = tnum_or(tnum_clear_subreg(false_64off),
8408 tnum_subreg(false_32off));
8409 true_reg->var_off = tnum_or(tnum_clear_subreg(true_64off),
8410 tnum_subreg(true_32off));
8411 __reg_combine_32_into_64(false_reg);
8412 __reg_combine_32_into_64(true_reg);
8414 false_reg->var_off = false_64off;
8415 true_reg->var_off = true_64off;
8416 __reg_combine_64_into_32(false_reg);
8417 __reg_combine_64_into_32(true_reg);
8421 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
8424 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
8425 struct bpf_reg_state *false_reg,
8427 u8 opcode, bool is_jmp32)
8429 opcode = flip_opcode(opcode);
8430 /* This uses zero as "not present in table"; luckily the zero opcode,
8431 * BPF_JA, can't get here.
8434 reg_set_min_max(true_reg, false_reg, val, val32, opcode, is_jmp32);
8437 /* Regs are known to be equal, so intersect their min/max/var_off */
8438 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
8439 struct bpf_reg_state *dst_reg)
8441 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
8442 dst_reg->umin_value);
8443 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
8444 dst_reg->umax_value);
8445 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
8446 dst_reg->smin_value);
8447 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
8448 dst_reg->smax_value);
8449 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
8451 /* We might have learned new bounds from the var_off. */
8452 __update_reg_bounds(src_reg);
8453 __update_reg_bounds(dst_reg);
8454 /* We might have learned something about the sign bit. */
8455 __reg_deduce_bounds(src_reg);
8456 __reg_deduce_bounds(dst_reg);
8457 /* We might have learned some bits from the bounds. */
8458 __reg_bound_offset(src_reg);
8459 __reg_bound_offset(dst_reg);
8460 /* Intersecting with the old var_off might have improved our bounds
8461 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
8462 * then new var_off is (0; 0x7f...fc) which improves our umax.
8464 __update_reg_bounds(src_reg);
8465 __update_reg_bounds(dst_reg);
8468 static void reg_combine_min_max(struct bpf_reg_state *true_src,
8469 struct bpf_reg_state *true_dst,
8470 struct bpf_reg_state *false_src,
8471 struct bpf_reg_state *false_dst,
8476 __reg_combine_min_max(true_src, true_dst);
8479 __reg_combine_min_max(false_src, false_dst);
8484 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
8485 struct bpf_reg_state *reg, u32 id,
8488 if (reg_type_may_be_null(reg->type) && reg->id == id &&
8489 !WARN_ON_ONCE(!reg->id)) {
8490 /* Old offset (both fixed and variable parts) should
8491 * have been known-zero, because we don't allow pointer
8492 * arithmetic on pointers that might be NULL.
8494 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
8495 !tnum_equals_const(reg->var_off, 0) ||
8497 __mark_reg_known_zero(reg);
8501 reg->type = SCALAR_VALUE;
8502 /* We don't need id and ref_obj_id from this point
8503 * onwards anymore, thus we should better reset it,
8504 * so that state pruning has chances to take effect.
8507 reg->ref_obj_id = 0;
8512 mark_ptr_not_null_reg(reg);
8514 if (!reg_may_point_to_spin_lock(reg)) {
8515 /* For not-NULL ptr, reg->ref_obj_id will be reset
8516 * in release_reg_references().
8518 * reg->id is still used by spin_lock ptr. Other
8519 * than spin_lock ptr type, reg->id can be reset.
8526 static void __mark_ptr_or_null_regs(struct bpf_func_state *state, u32 id,
8529 struct bpf_reg_state *reg;
8532 for (i = 0; i < MAX_BPF_REG; i++)
8533 mark_ptr_or_null_reg(state, &state->regs[i], id, is_null);
8535 bpf_for_each_spilled_reg(i, state, reg) {
8538 mark_ptr_or_null_reg(state, reg, id, is_null);
8542 /* The logic is similar to find_good_pkt_pointers(), both could eventually
8543 * be folded together at some point.
8545 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
8548 struct bpf_func_state *state = vstate->frame[vstate->curframe];
8549 struct bpf_reg_state *regs = state->regs;
8550 u32 ref_obj_id = regs[regno].ref_obj_id;
8551 u32 id = regs[regno].id;
8554 if (ref_obj_id && ref_obj_id == id && is_null)
8555 /* regs[regno] is in the " == NULL" branch.
8556 * No one could have freed the reference state before
8557 * doing the NULL check.
8559 WARN_ON_ONCE(release_reference_state(state, id));
8561 for (i = 0; i <= vstate->curframe; i++)
8562 __mark_ptr_or_null_regs(vstate->frame[i], id, is_null);
8565 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
8566 struct bpf_reg_state *dst_reg,
8567 struct bpf_reg_state *src_reg,
8568 struct bpf_verifier_state *this_branch,
8569 struct bpf_verifier_state *other_branch)
8571 if (BPF_SRC(insn->code) != BPF_X)
8574 /* Pointers are always 64-bit. */
8575 if (BPF_CLASS(insn->code) == BPF_JMP32)
8578 switch (BPF_OP(insn->code)) {
8580 if ((dst_reg->type == PTR_TO_PACKET &&
8581 src_reg->type == PTR_TO_PACKET_END) ||
8582 (dst_reg->type == PTR_TO_PACKET_META &&
8583 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
8584 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
8585 find_good_pkt_pointers(this_branch, dst_reg,
8586 dst_reg->type, false);
8587 mark_pkt_end(other_branch, insn->dst_reg, true);
8588 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
8589 src_reg->type == PTR_TO_PACKET) ||
8590 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
8591 src_reg->type == PTR_TO_PACKET_META)) {
8592 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
8593 find_good_pkt_pointers(other_branch, src_reg,
8594 src_reg->type, true);
8595 mark_pkt_end(this_branch, insn->src_reg, false);
8601 if ((dst_reg->type == PTR_TO_PACKET &&
8602 src_reg->type == PTR_TO_PACKET_END) ||
8603 (dst_reg->type == PTR_TO_PACKET_META &&
8604 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
8605 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
8606 find_good_pkt_pointers(other_branch, dst_reg,
8607 dst_reg->type, true);
8608 mark_pkt_end(this_branch, insn->dst_reg, false);
8609 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
8610 src_reg->type == PTR_TO_PACKET) ||
8611 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
8612 src_reg->type == PTR_TO_PACKET_META)) {
8613 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
8614 find_good_pkt_pointers(this_branch, src_reg,
8615 src_reg->type, false);
8616 mark_pkt_end(other_branch, insn->src_reg, true);
8622 if ((dst_reg->type == PTR_TO_PACKET &&
8623 src_reg->type == PTR_TO_PACKET_END) ||
8624 (dst_reg->type == PTR_TO_PACKET_META &&
8625 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
8626 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
8627 find_good_pkt_pointers(this_branch, dst_reg,
8628 dst_reg->type, true);
8629 mark_pkt_end(other_branch, insn->dst_reg, false);
8630 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
8631 src_reg->type == PTR_TO_PACKET) ||
8632 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
8633 src_reg->type == PTR_TO_PACKET_META)) {
8634 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
8635 find_good_pkt_pointers(other_branch, src_reg,
8636 src_reg->type, false);
8637 mark_pkt_end(this_branch, insn->src_reg, true);
8643 if ((dst_reg->type == PTR_TO_PACKET &&
8644 src_reg->type == PTR_TO_PACKET_END) ||
8645 (dst_reg->type == PTR_TO_PACKET_META &&
8646 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
8647 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
8648 find_good_pkt_pointers(other_branch, dst_reg,
8649 dst_reg->type, false);
8650 mark_pkt_end(this_branch, insn->dst_reg, true);
8651 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
8652 src_reg->type == PTR_TO_PACKET) ||
8653 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
8654 src_reg->type == PTR_TO_PACKET_META)) {
8655 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
8656 find_good_pkt_pointers(this_branch, src_reg,
8657 src_reg->type, true);
8658 mark_pkt_end(other_branch, insn->src_reg, false);
8670 static void find_equal_scalars(struct bpf_verifier_state *vstate,
8671 struct bpf_reg_state *known_reg)
8673 struct bpf_func_state *state;
8674 struct bpf_reg_state *reg;
8677 for (i = 0; i <= vstate->curframe; i++) {
8678 state = vstate->frame[i];
8679 for (j = 0; j < MAX_BPF_REG; j++) {
8680 reg = &state->regs[j];
8681 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
8685 bpf_for_each_spilled_reg(j, state, reg) {
8688 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
8694 static int check_cond_jmp_op(struct bpf_verifier_env *env,
8695 struct bpf_insn *insn, int *insn_idx)
8697 struct bpf_verifier_state *this_branch = env->cur_state;
8698 struct bpf_verifier_state *other_branch;
8699 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
8700 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
8701 u8 opcode = BPF_OP(insn->code);
8706 /* Only conditional jumps are expected to reach here. */
8707 if (opcode == BPF_JA || opcode > BPF_JSLE) {
8708 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
8712 if (BPF_SRC(insn->code) == BPF_X) {
8713 if (insn->imm != 0) {
8714 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
8718 /* check src1 operand */
8719 err = check_reg_arg(env, insn->src_reg, SRC_OP);
8723 if (is_pointer_value(env, insn->src_reg)) {
8724 verbose(env, "R%d pointer comparison prohibited\n",
8728 src_reg = ®s[insn->src_reg];
8730 if (insn->src_reg != BPF_REG_0) {
8731 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
8736 /* check src2 operand */
8737 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
8741 dst_reg = ®s[insn->dst_reg];
8742 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
8744 if (BPF_SRC(insn->code) == BPF_K) {
8745 pred = is_branch_taken(dst_reg, insn->imm, opcode, is_jmp32);
8746 } else if (src_reg->type == SCALAR_VALUE &&
8747 is_jmp32 && tnum_is_const(tnum_subreg(src_reg->var_off))) {
8748 pred = is_branch_taken(dst_reg,
8749 tnum_subreg(src_reg->var_off).value,
8752 } else if (src_reg->type == SCALAR_VALUE &&
8753 !is_jmp32 && tnum_is_const(src_reg->var_off)) {
8754 pred = is_branch_taken(dst_reg,
8755 src_reg->var_off.value,
8758 } else if (reg_is_pkt_pointer_any(dst_reg) &&
8759 reg_is_pkt_pointer_any(src_reg) &&
8761 pred = is_pkt_ptr_branch_taken(dst_reg, src_reg, opcode);
8765 /* If we get here with a dst_reg pointer type it is because
8766 * above is_branch_taken() special cased the 0 comparison.
8768 if (!__is_pointer_value(false, dst_reg))
8769 err = mark_chain_precision(env, insn->dst_reg);
8770 if (BPF_SRC(insn->code) == BPF_X && !err &&
8771 !__is_pointer_value(false, src_reg))
8772 err = mark_chain_precision(env, insn->src_reg);
8778 /* Only follow the goto, ignore fall-through. If needed, push
8779 * the fall-through branch for simulation under speculative
8782 if (!env->bypass_spec_v1 &&
8783 !sanitize_speculative_path(env, insn, *insn_idx + 1,
8786 *insn_idx += insn->off;
8788 } else if (pred == 0) {
8789 /* Only follow the fall-through branch, since that's where the
8790 * program will go. If needed, push the goto branch for
8791 * simulation under speculative execution.
8793 if (!env->bypass_spec_v1 &&
8794 !sanitize_speculative_path(env, insn,
8795 *insn_idx + insn->off + 1,
8801 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
8805 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
8807 /* detect if we are comparing against a constant value so we can adjust
8808 * our min/max values for our dst register.
8809 * this is only legit if both are scalars (or pointers to the same
8810 * object, I suppose, but we don't support that right now), because
8811 * otherwise the different base pointers mean the offsets aren't
8814 if (BPF_SRC(insn->code) == BPF_X) {
8815 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
8817 if (dst_reg->type == SCALAR_VALUE &&
8818 src_reg->type == SCALAR_VALUE) {
8819 if (tnum_is_const(src_reg->var_off) ||
8821 tnum_is_const(tnum_subreg(src_reg->var_off))))
8822 reg_set_min_max(&other_branch_regs[insn->dst_reg],
8824 src_reg->var_off.value,
8825 tnum_subreg(src_reg->var_off).value,
8827 else if (tnum_is_const(dst_reg->var_off) ||
8829 tnum_is_const(tnum_subreg(dst_reg->var_off))))
8830 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
8832 dst_reg->var_off.value,
8833 tnum_subreg(dst_reg->var_off).value,
8835 else if (!is_jmp32 &&
8836 (opcode == BPF_JEQ || opcode == BPF_JNE))
8837 /* Comparing for equality, we can combine knowledge */
8838 reg_combine_min_max(&other_branch_regs[insn->src_reg],
8839 &other_branch_regs[insn->dst_reg],
8840 src_reg, dst_reg, opcode);
8842 !WARN_ON_ONCE(src_reg->id != other_branch_regs[insn->src_reg].id)) {
8843 find_equal_scalars(this_branch, src_reg);
8844 find_equal_scalars(other_branch, &other_branch_regs[insn->src_reg]);
8848 } else if (dst_reg->type == SCALAR_VALUE) {
8849 reg_set_min_max(&other_branch_regs[insn->dst_reg],
8850 dst_reg, insn->imm, (u32)insn->imm,
8854 if (dst_reg->type == SCALAR_VALUE && dst_reg->id &&
8855 !WARN_ON_ONCE(dst_reg->id != other_branch_regs[insn->dst_reg].id)) {
8856 find_equal_scalars(this_branch, dst_reg);
8857 find_equal_scalars(other_branch, &other_branch_regs[insn->dst_reg]);
8860 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
8861 * NOTE: these optimizations below are related with pointer comparison
8862 * which will never be JMP32.
8864 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
8865 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
8866 reg_type_may_be_null(dst_reg->type)) {
8867 /* Mark all identical registers in each branch as either
8868 * safe or unknown depending R == 0 or R != 0 conditional.
8870 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
8872 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
8874 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
8875 this_branch, other_branch) &&
8876 is_pointer_value(env, insn->dst_reg)) {
8877 verbose(env, "R%d pointer comparison prohibited\n",
8881 if (env->log.level & BPF_LOG_LEVEL)
8882 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
8886 /* verify BPF_LD_IMM64 instruction */
8887 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
8889 struct bpf_insn_aux_data *aux = cur_aux(env);
8890 struct bpf_reg_state *regs = cur_regs(env);
8891 struct bpf_reg_state *dst_reg;
8892 struct bpf_map *map;
8895 if (BPF_SIZE(insn->code) != BPF_DW) {
8896 verbose(env, "invalid BPF_LD_IMM insn\n");
8899 if (insn->off != 0) {
8900 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
8904 err = check_reg_arg(env, insn->dst_reg, DST_OP);
8908 dst_reg = ®s[insn->dst_reg];
8909 if (insn->src_reg == 0) {
8910 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
8912 dst_reg->type = SCALAR_VALUE;
8913 __mark_reg_known(®s[insn->dst_reg], imm);
8917 if (insn->src_reg == BPF_PSEUDO_BTF_ID) {
8918 mark_reg_known_zero(env, regs, insn->dst_reg);
8920 dst_reg->type = aux->btf_var.reg_type;
8921 switch (dst_reg->type) {
8923 dst_reg->mem_size = aux->btf_var.mem_size;
8926 case PTR_TO_PERCPU_BTF_ID:
8927 dst_reg->btf = aux->btf_var.btf;
8928 dst_reg->btf_id = aux->btf_var.btf_id;
8931 verbose(env, "bpf verifier is misconfigured\n");
8937 if (insn->src_reg == BPF_PSEUDO_FUNC) {
8938 struct bpf_prog_aux *aux = env->prog->aux;
8939 u32 subprogno = insn[1].imm;
8941 if (!aux->func_info) {
8942 verbose(env, "missing btf func_info\n");
8945 if (aux->func_info_aux[subprogno].linkage != BTF_FUNC_STATIC) {
8946 verbose(env, "callback function not static\n");
8950 dst_reg->type = PTR_TO_FUNC;
8951 dst_reg->subprogno = subprogno;
8955 map = env->used_maps[aux->map_index];
8956 mark_reg_known_zero(env, regs, insn->dst_reg);
8957 dst_reg->map_ptr = map;
8959 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE) {
8960 dst_reg->type = PTR_TO_MAP_VALUE;
8961 dst_reg->off = aux->map_off;
8962 if (map_value_has_spin_lock(map))
8963 dst_reg->id = ++env->id_gen;
8964 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD) {
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 convetion 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,
9480 union bpf_attr __user *uattr)
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;
9489 void __user *urecord;
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 = u64_to_user_ptr(attr->func_info);
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 (put_user(min_size, &uattr->func_info_rec_size))
9541 if (copy_from_user(&krecord[i], urecord, min_size)) {
9546 /* check insn_off */
9549 if (krecord[i].insn_off) {
9551 "nonzero insn_off %u for the first func info record",
9552 krecord[i].insn_off);
9555 } else if (krecord[i].insn_off <= prev_offset) {
9557 "same or smaller insn offset (%u) than previous func info record (%u)",
9558 krecord[i].insn_off, prev_offset);
9562 if (env->subprog_info[i].start != krecord[i].insn_off) {
9563 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
9568 type = btf_type_by_id(btf, krecord[i].type_id);
9569 if (!type || !btf_type_is_func(type)) {
9570 verbose(env, "invalid type id %d in func info",
9571 krecord[i].type_id);
9574 info_aux[i].linkage = BTF_INFO_VLEN(type->info);
9576 func_proto = btf_type_by_id(btf, type->type);
9577 if (unlikely(!func_proto || !btf_type_is_func_proto(func_proto)))
9578 /* btf_func_check() already verified it during BTF load */
9580 ret_type = btf_type_skip_modifiers(btf, func_proto->type, NULL);
9582 btf_type_is_small_int(ret_type) || btf_type_is_enum(ret_type);
9583 if (i && !scalar_return && env->subprog_info[i].has_ld_abs) {
9584 verbose(env, "LD_ABS is only allowed in functions that return 'int'.\n");
9587 if (i && !scalar_return && env->subprog_info[i].has_tail_call) {
9588 verbose(env, "tail_call is only allowed in functions that return 'int'.\n");
9592 prev_offset = krecord[i].insn_off;
9593 urecord += urec_size;
9596 prog->aux->func_info = krecord;
9597 prog->aux->func_info_cnt = nfuncs;
9598 prog->aux->func_info_aux = info_aux;
9607 static void adjust_btf_func(struct bpf_verifier_env *env)
9609 struct bpf_prog_aux *aux = env->prog->aux;
9612 if (!aux->func_info)
9615 for (i = 0; i < env->subprog_cnt; i++)
9616 aux->func_info[i].insn_off = env->subprog_info[i].start;
9619 #define MIN_BPF_LINEINFO_SIZE (offsetof(struct bpf_line_info, line_col) + \
9620 sizeof(((struct bpf_line_info *)(0))->line_col))
9621 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
9623 static int check_btf_line(struct bpf_verifier_env *env,
9624 const union bpf_attr *attr,
9625 union bpf_attr __user *uattr)
9627 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
9628 struct bpf_subprog_info *sub;
9629 struct bpf_line_info *linfo;
9630 struct bpf_prog *prog;
9631 const struct btf *btf;
9632 void __user *ulinfo;
9635 nr_linfo = attr->line_info_cnt;
9639 rec_size = attr->line_info_rec_size;
9640 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
9641 rec_size > MAX_LINEINFO_REC_SIZE ||
9642 rec_size & (sizeof(u32) - 1))
9645 /* Need to zero it in case the userspace may
9646 * pass in a smaller bpf_line_info object.
9648 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
9649 GFP_KERNEL | __GFP_NOWARN);
9654 btf = prog->aux->btf;
9657 sub = env->subprog_info;
9658 ulinfo = u64_to_user_ptr(attr->line_info);
9659 expected_size = sizeof(struct bpf_line_info);
9660 ncopy = min_t(u32, expected_size, rec_size);
9661 for (i = 0; i < nr_linfo; i++) {
9662 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
9664 if (err == -E2BIG) {
9665 verbose(env, "nonzero tailing record in line_info");
9666 if (put_user(expected_size,
9667 &uattr->line_info_rec_size))
9673 if (copy_from_user(&linfo[i], ulinfo, ncopy)) {
9679 * Check insn_off to ensure
9680 * 1) strictly increasing AND
9681 * 2) bounded by prog->len
9683 * The linfo[0].insn_off == 0 check logically falls into
9684 * the later "missing bpf_line_info for func..." case
9685 * because the first linfo[0].insn_off must be the
9686 * first sub also and the first sub must have
9687 * subprog_info[0].start == 0.
9689 if ((i && linfo[i].insn_off <= prev_offset) ||
9690 linfo[i].insn_off >= prog->len) {
9691 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
9692 i, linfo[i].insn_off, prev_offset,
9698 if (!prog->insnsi[linfo[i].insn_off].code) {
9700 "Invalid insn code at line_info[%u].insn_off\n",
9706 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
9707 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
9708 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
9713 if (s != env->subprog_cnt) {
9714 if (linfo[i].insn_off == sub[s].start) {
9715 sub[s].linfo_idx = i;
9717 } else if (sub[s].start < linfo[i].insn_off) {
9718 verbose(env, "missing bpf_line_info for func#%u\n", s);
9724 prev_offset = linfo[i].insn_off;
9728 if (s != env->subprog_cnt) {
9729 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
9730 env->subprog_cnt - s, s);
9735 prog->aux->linfo = linfo;
9736 prog->aux->nr_linfo = nr_linfo;
9745 static int check_btf_info(struct bpf_verifier_env *env,
9746 const union bpf_attr *attr,
9747 union bpf_attr __user *uattr)
9752 if (!attr->func_info_cnt && !attr->line_info_cnt) {
9753 if (check_abnormal_return(env))
9758 btf = btf_get_by_fd(attr->prog_btf_fd);
9760 return PTR_ERR(btf);
9761 if (btf_is_kernel(btf)) {
9765 env->prog->aux->btf = btf;
9767 err = check_btf_func(env, attr, uattr);
9771 err = check_btf_line(env, attr, uattr);
9778 /* check %cur's range satisfies %old's */
9779 static bool range_within(struct bpf_reg_state *old,
9780 struct bpf_reg_state *cur)
9782 return old->umin_value <= cur->umin_value &&
9783 old->umax_value >= cur->umax_value &&
9784 old->smin_value <= cur->smin_value &&
9785 old->smax_value >= cur->smax_value &&
9786 old->u32_min_value <= cur->u32_min_value &&
9787 old->u32_max_value >= cur->u32_max_value &&
9788 old->s32_min_value <= cur->s32_min_value &&
9789 old->s32_max_value >= cur->s32_max_value;
9792 /* If in the old state two registers had the same id, then they need to have
9793 * the same id in the new state as well. But that id could be different from
9794 * the old state, so we need to track the mapping from old to new ids.
9795 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
9796 * regs with old id 5 must also have new id 9 for the new state to be safe. But
9797 * regs with a different old id could still have new id 9, we don't care about
9799 * So we look through our idmap to see if this old id has been seen before. If
9800 * so, we require the new id to match; otherwise, we add the id pair to the map.
9802 static bool check_ids(u32 old_id, u32 cur_id, struct bpf_id_pair *idmap)
9806 for (i = 0; i < BPF_ID_MAP_SIZE; i++) {
9807 if (!idmap[i].old) {
9808 /* Reached an empty slot; haven't seen this id before */
9809 idmap[i].old = old_id;
9810 idmap[i].cur = cur_id;
9813 if (idmap[i].old == old_id)
9814 return idmap[i].cur == cur_id;
9816 /* We ran out of idmap slots, which should be impossible */
9821 static void clean_func_state(struct bpf_verifier_env *env,
9822 struct bpf_func_state *st)
9824 enum bpf_reg_liveness live;
9827 for (i = 0; i < BPF_REG_FP; i++) {
9828 live = st->regs[i].live;
9829 /* liveness must not touch this register anymore */
9830 st->regs[i].live |= REG_LIVE_DONE;
9831 if (!(live & REG_LIVE_READ))
9832 /* since the register is unused, clear its state
9833 * to make further comparison simpler
9835 __mark_reg_not_init(env, &st->regs[i]);
9838 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
9839 live = st->stack[i].spilled_ptr.live;
9840 /* liveness must not touch this stack slot anymore */
9841 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
9842 if (!(live & REG_LIVE_READ)) {
9843 __mark_reg_not_init(env, &st->stack[i].spilled_ptr);
9844 for (j = 0; j < BPF_REG_SIZE; j++)
9845 st->stack[i].slot_type[j] = STACK_INVALID;
9850 static void clean_verifier_state(struct bpf_verifier_env *env,
9851 struct bpf_verifier_state *st)
9855 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
9856 /* all regs in this state in all frames were already marked */
9859 for (i = 0; i <= st->curframe; i++)
9860 clean_func_state(env, st->frame[i]);
9863 /* the parentage chains form a tree.
9864 * the verifier states are added to state lists at given insn and
9865 * pushed into state stack for future exploration.
9866 * when the verifier reaches bpf_exit insn some of the verifer states
9867 * stored in the state lists have their final liveness state already,
9868 * but a lot of states will get revised from liveness point of view when
9869 * the verifier explores other branches.
9872 * 2: if r1 == 100 goto pc+1
9875 * when the verifier reaches exit insn the register r0 in the state list of
9876 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
9877 * of insn 2 and goes exploring further. At the insn 4 it will walk the
9878 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
9880 * Since the verifier pushes the branch states as it sees them while exploring
9881 * the program the condition of walking the branch instruction for the second
9882 * time means that all states below this branch were already explored and
9883 * their final liveness markes are already propagated.
9884 * Hence when the verifier completes the search of state list in is_state_visited()
9885 * we can call this clean_live_states() function to mark all liveness states
9886 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
9888 * This function also clears the registers and stack for states that !READ
9889 * to simplify state merging.
9891 * Important note here that walking the same branch instruction in the callee
9892 * doesn't meant that the states are DONE. The verifier has to compare
9895 static void clean_live_states(struct bpf_verifier_env *env, int insn,
9896 struct bpf_verifier_state *cur)
9898 struct bpf_verifier_state_list *sl;
9901 sl = *explored_state(env, insn);
9903 if (sl->state.branches)
9905 if (sl->state.insn_idx != insn ||
9906 sl->state.curframe != cur->curframe)
9908 for (i = 0; i <= cur->curframe; i++)
9909 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
9911 clean_verifier_state(env, &sl->state);
9917 /* Returns true if (rold safe implies rcur safe) */
9918 static bool regsafe(struct bpf_verifier_env *env, struct bpf_reg_state *rold,
9919 struct bpf_reg_state *rcur, struct bpf_id_pair *idmap)
9923 if (!(rold->live & REG_LIVE_READ))
9924 /* explored state didn't use this */
9927 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
9929 if (rold->type == PTR_TO_STACK)
9930 /* two stack pointers are equal only if they're pointing to
9931 * the same stack frame, since fp-8 in foo != fp-8 in bar
9933 return equal && rold->frameno == rcur->frameno;
9938 if (rold->type == NOT_INIT)
9939 /* explored state can't have used this */
9941 if (rcur->type == NOT_INIT)
9943 switch (rold->type) {
9945 if (env->explore_alu_limits)
9947 if (rcur->type == SCALAR_VALUE) {
9948 if (!rold->precise && !rcur->precise)
9950 /* new val must satisfy old val knowledge */
9951 return range_within(rold, rcur) &&
9952 tnum_in(rold->var_off, rcur->var_off);
9954 /* We're trying to use a pointer in place of a scalar.
9955 * Even if the scalar was unbounded, this could lead to
9956 * pointer leaks because scalars are allowed to leak
9957 * while pointers are not. We could make this safe in
9958 * special cases if root is calling us, but it's
9959 * probably not worth the hassle.
9963 case PTR_TO_MAP_KEY:
9964 case PTR_TO_MAP_VALUE:
9965 /* If the new min/max/var_off satisfy the old ones and
9966 * everything else matches, we are OK.
9967 * 'id' is not compared, since it's only used for maps with
9968 * bpf_spin_lock inside map element and in such cases if
9969 * the rest of the prog is valid for one map element then
9970 * it's valid for all map elements regardless of the key
9971 * used in bpf_map_lookup()
9973 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
9974 range_within(rold, rcur) &&
9975 tnum_in(rold->var_off, rcur->var_off);
9976 case PTR_TO_MAP_VALUE_OR_NULL:
9977 /* a PTR_TO_MAP_VALUE could be safe to use as a
9978 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
9979 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
9980 * checked, doing so could have affected others with the same
9981 * id, and we can't check for that because we lost the id when
9982 * we converted to a PTR_TO_MAP_VALUE.
9984 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
9986 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
9988 /* Check our ids match any regs they're supposed to */
9989 return check_ids(rold->id, rcur->id, idmap);
9990 case PTR_TO_PACKET_META:
9992 if (rcur->type != rold->type)
9994 /* We must have at least as much range as the old ptr
9995 * did, so that any accesses which were safe before are
9996 * still safe. This is true even if old range < old off,
9997 * since someone could have accessed through (ptr - k), or
9998 * even done ptr -= k in a register, to get a safe access.
10000 if (rold->range > rcur->range)
10002 /* If the offsets don't match, we can't trust our alignment;
10003 * nor can we be sure that we won't fall out of range.
10005 if (rold->off != rcur->off)
10007 /* id relations must be preserved */
10008 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
10010 /* new val must satisfy old val knowledge */
10011 return range_within(rold, rcur) &&
10012 tnum_in(rold->var_off, rcur->var_off);
10014 case CONST_PTR_TO_MAP:
10015 case PTR_TO_PACKET_END:
10016 case PTR_TO_FLOW_KEYS:
10017 case PTR_TO_SOCKET:
10018 case PTR_TO_SOCKET_OR_NULL:
10019 case PTR_TO_SOCK_COMMON:
10020 case PTR_TO_SOCK_COMMON_OR_NULL:
10021 case PTR_TO_TCP_SOCK:
10022 case PTR_TO_TCP_SOCK_OR_NULL:
10023 case PTR_TO_XDP_SOCK:
10024 /* Only valid matches are exact, which memcmp() above
10025 * would have accepted
10028 /* Don't know what's going on, just say it's not safe */
10032 /* Shouldn't get here; if we do, say it's not safe */
10037 static bool stacksafe(struct bpf_verifier_env *env, struct bpf_func_state *old,
10038 struct bpf_func_state *cur, struct bpf_id_pair *idmap)
10042 /* walk slots of the explored stack and ignore any additional
10043 * slots in the current stack, since explored(safe) state
10046 for (i = 0; i < old->allocated_stack; i++) {
10047 spi = i / BPF_REG_SIZE;
10049 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
10050 i += BPF_REG_SIZE - 1;
10051 /* explored state didn't use this */
10055 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
10058 /* explored stack has more populated slots than current stack
10059 * and these slots were used
10061 if (i >= cur->allocated_stack)
10064 /* if old state was safe with misc data in the stack
10065 * it will be safe with zero-initialized stack.
10066 * The opposite is not true
10068 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
10069 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
10071 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
10072 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
10073 /* Ex: old explored (safe) state has STACK_SPILL in
10074 * this stack slot, but current has STACK_MISC ->
10075 * this verifier states are not equivalent,
10076 * return false to continue verification of this path
10079 if (i % BPF_REG_SIZE)
10081 if (old->stack[spi].slot_type[0] != STACK_SPILL)
10083 if (!regsafe(env, &old->stack[spi].spilled_ptr,
10084 &cur->stack[spi].spilled_ptr, idmap))
10085 /* when explored and current stack slot are both storing
10086 * spilled registers, check that stored pointers types
10087 * are the same as well.
10088 * Ex: explored safe path could have stored
10089 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
10090 * but current path has stored:
10091 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
10092 * such verifier states are not equivalent.
10093 * return false to continue verification of this path
10100 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
10102 if (old->acquired_refs != cur->acquired_refs)
10104 return !memcmp(old->refs, cur->refs,
10105 sizeof(*old->refs) * old->acquired_refs);
10108 /* compare two verifier states
10110 * all states stored in state_list are known to be valid, since
10111 * verifier reached 'bpf_exit' instruction through them
10113 * this function is called when verifier exploring different branches of
10114 * execution popped from the state stack. If it sees an old state that has
10115 * more strict register state and more strict stack state then this execution
10116 * branch doesn't need to be explored further, since verifier already
10117 * concluded that more strict state leads to valid finish.
10119 * Therefore two states are equivalent if register state is more conservative
10120 * and explored stack state is more conservative than the current one.
10123 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
10124 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
10126 * In other words if current stack state (one being explored) has more
10127 * valid slots than old one that already passed validation, it means
10128 * the verifier can stop exploring and conclude that current state is valid too
10130 * Similarly with registers. If explored state has register type as invalid
10131 * whereas register type in current state is meaningful, it means that
10132 * the current state will reach 'bpf_exit' instruction safely
10134 static bool func_states_equal(struct bpf_verifier_env *env, struct bpf_func_state *old,
10135 struct bpf_func_state *cur)
10139 memset(env->idmap_scratch, 0, sizeof(env->idmap_scratch));
10140 for (i = 0; i < MAX_BPF_REG; i++)
10141 if (!regsafe(env, &old->regs[i], &cur->regs[i],
10142 env->idmap_scratch))
10145 if (!stacksafe(env, old, cur, env->idmap_scratch))
10148 if (!refsafe(old, cur))
10154 static bool states_equal(struct bpf_verifier_env *env,
10155 struct bpf_verifier_state *old,
10156 struct bpf_verifier_state *cur)
10160 if (old->curframe != cur->curframe)
10163 /* Verification state from speculative execution simulation
10164 * must never prune a non-speculative execution one.
10166 if (old->speculative && !cur->speculative)
10169 if (old->active_spin_lock != cur->active_spin_lock)
10172 /* for states to be equal callsites have to be the same
10173 * and all frame states need to be equivalent
10175 for (i = 0; i <= old->curframe; i++) {
10176 if (old->frame[i]->callsite != cur->frame[i]->callsite)
10178 if (!func_states_equal(env, old->frame[i], cur->frame[i]))
10184 /* Return 0 if no propagation happened. Return negative error code if error
10185 * happened. Otherwise, return the propagated bit.
10187 static int propagate_liveness_reg(struct bpf_verifier_env *env,
10188 struct bpf_reg_state *reg,
10189 struct bpf_reg_state *parent_reg)
10191 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
10192 u8 flag = reg->live & REG_LIVE_READ;
10195 /* When comes here, read flags of PARENT_REG or REG could be any of
10196 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
10197 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
10199 if (parent_flag == REG_LIVE_READ64 ||
10200 /* Or if there is no read flag from REG. */
10202 /* Or if the read flag from REG is the same as PARENT_REG. */
10203 parent_flag == flag)
10206 err = mark_reg_read(env, reg, parent_reg, flag);
10213 /* A write screens off any subsequent reads; but write marks come from the
10214 * straight-line code between a state and its parent. When we arrive at an
10215 * equivalent state (jump target or such) we didn't arrive by the straight-line
10216 * code, so read marks in the state must propagate to the parent regardless
10217 * of the state's write marks. That's what 'parent == state->parent' comparison
10218 * in mark_reg_read() is for.
10220 static int propagate_liveness(struct bpf_verifier_env *env,
10221 const struct bpf_verifier_state *vstate,
10222 struct bpf_verifier_state *vparent)
10224 struct bpf_reg_state *state_reg, *parent_reg;
10225 struct bpf_func_state *state, *parent;
10226 int i, frame, err = 0;
10228 if (vparent->curframe != vstate->curframe) {
10229 WARN(1, "propagate_live: parent frame %d current frame %d\n",
10230 vparent->curframe, vstate->curframe);
10233 /* Propagate read liveness of registers... */
10234 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
10235 for (frame = 0; frame <= vstate->curframe; frame++) {
10236 parent = vparent->frame[frame];
10237 state = vstate->frame[frame];
10238 parent_reg = parent->regs;
10239 state_reg = state->regs;
10240 /* We don't need to worry about FP liveness, it's read-only */
10241 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
10242 err = propagate_liveness_reg(env, &state_reg[i],
10246 if (err == REG_LIVE_READ64)
10247 mark_insn_zext(env, &parent_reg[i]);
10250 /* Propagate stack slots. */
10251 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
10252 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
10253 parent_reg = &parent->stack[i].spilled_ptr;
10254 state_reg = &state->stack[i].spilled_ptr;
10255 err = propagate_liveness_reg(env, state_reg,
10264 /* find precise scalars in the previous equivalent state and
10265 * propagate them into the current state
10267 static int propagate_precision(struct bpf_verifier_env *env,
10268 const struct bpf_verifier_state *old)
10270 struct bpf_reg_state *state_reg;
10271 struct bpf_func_state *state;
10274 state = old->frame[old->curframe];
10275 state_reg = state->regs;
10276 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
10277 if (state_reg->type != SCALAR_VALUE ||
10278 !state_reg->precise)
10280 if (env->log.level & BPF_LOG_LEVEL2)
10281 verbose(env, "propagating r%d\n", i);
10282 err = mark_chain_precision(env, i);
10287 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
10288 if (state->stack[i].slot_type[0] != STACK_SPILL)
10290 state_reg = &state->stack[i].spilled_ptr;
10291 if (state_reg->type != SCALAR_VALUE ||
10292 !state_reg->precise)
10294 if (env->log.level & BPF_LOG_LEVEL2)
10295 verbose(env, "propagating fp%d\n",
10296 (-i - 1) * BPF_REG_SIZE);
10297 err = mark_chain_precision_stack(env, i);
10304 static bool states_maybe_looping(struct bpf_verifier_state *old,
10305 struct bpf_verifier_state *cur)
10307 struct bpf_func_state *fold, *fcur;
10308 int i, fr = cur->curframe;
10310 if (old->curframe != fr)
10313 fold = old->frame[fr];
10314 fcur = cur->frame[fr];
10315 for (i = 0; i < MAX_BPF_REG; i++)
10316 if (memcmp(&fold->regs[i], &fcur->regs[i],
10317 offsetof(struct bpf_reg_state, parent)))
10323 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
10325 struct bpf_verifier_state_list *new_sl;
10326 struct bpf_verifier_state_list *sl, **pprev;
10327 struct bpf_verifier_state *cur = env->cur_state, *new;
10328 int i, j, err, states_cnt = 0;
10329 bool add_new_state = env->test_state_freq ? true : false;
10331 cur->last_insn_idx = env->prev_insn_idx;
10332 if (!env->insn_aux_data[insn_idx].prune_point)
10333 /* this 'insn_idx' instruction wasn't marked, so we will not
10334 * be doing state search here
10338 /* bpf progs typically have pruning point every 4 instructions
10339 * http://vger.kernel.org/bpfconf2019.html#session-1
10340 * Do not add new state for future pruning if the verifier hasn't seen
10341 * at least 2 jumps and at least 8 instructions.
10342 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
10343 * In tests that amounts to up to 50% reduction into total verifier
10344 * memory consumption and 20% verifier time speedup.
10346 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
10347 env->insn_processed - env->prev_insn_processed >= 8)
10348 add_new_state = true;
10350 pprev = explored_state(env, insn_idx);
10353 clean_live_states(env, insn_idx, cur);
10357 if (sl->state.insn_idx != insn_idx)
10359 if (sl->state.branches) {
10360 if (states_maybe_looping(&sl->state, cur) &&
10361 states_equal(env, &sl->state, cur)) {
10362 verbose_linfo(env, insn_idx, "; ");
10363 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
10366 /* if the verifier is processing a loop, avoid adding new state
10367 * too often, since different loop iterations have distinct
10368 * states and may not help future pruning.
10369 * This threshold shouldn't be too low to make sure that
10370 * a loop with large bound will be rejected quickly.
10371 * The most abusive loop will be:
10373 * if r1 < 1000000 goto pc-2
10374 * 1M insn_procssed limit / 100 == 10k peak states.
10375 * This threshold shouldn't be too high either, since states
10376 * at the end of the loop are likely to be useful in pruning.
10378 if (env->jmps_processed - env->prev_jmps_processed < 20 &&
10379 env->insn_processed - env->prev_insn_processed < 100)
10380 add_new_state = false;
10383 if (states_equal(env, &sl->state, cur)) {
10385 /* reached equivalent register/stack state,
10386 * prune the search.
10387 * Registers read by the continuation are read by us.
10388 * If we have any write marks in env->cur_state, they
10389 * will prevent corresponding reads in the continuation
10390 * from reaching our parent (an explored_state). Our
10391 * own state will get the read marks recorded, but
10392 * they'll be immediately forgotten as we're pruning
10393 * this state and will pop a new one.
10395 err = propagate_liveness(env, &sl->state, cur);
10397 /* if previous state reached the exit with precision and
10398 * current state is equivalent to it (except precsion marks)
10399 * the precision needs to be propagated back in
10400 * the current state.
10402 err = err ? : push_jmp_history(env, cur);
10403 err = err ? : propagate_precision(env, &sl->state);
10409 /* when new state is not going to be added do not increase miss count.
10410 * Otherwise several loop iterations will remove the state
10411 * recorded earlier. The goal of these heuristics is to have
10412 * states from some iterations of the loop (some in the beginning
10413 * and some at the end) to help pruning.
10417 /* heuristic to determine whether this state is beneficial
10418 * to keep checking from state equivalence point of view.
10419 * Higher numbers increase max_states_per_insn and verification time,
10420 * but do not meaningfully decrease insn_processed.
10422 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
10423 /* the state is unlikely to be useful. Remove it to
10424 * speed up verification
10427 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
10428 u32 br = sl->state.branches;
10431 "BUG live_done but branches_to_explore %d\n",
10433 free_verifier_state(&sl->state, false);
10435 env->peak_states--;
10437 /* cannot free this state, since parentage chain may
10438 * walk it later. Add it for free_list instead to
10439 * be freed at the end of verification
10441 sl->next = env->free_list;
10442 env->free_list = sl;
10452 if (env->max_states_per_insn < states_cnt)
10453 env->max_states_per_insn = states_cnt;
10455 if (!env->bpf_capable && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
10456 return push_jmp_history(env, cur);
10458 if (!add_new_state)
10459 return push_jmp_history(env, cur);
10461 /* There were no equivalent states, remember the current one.
10462 * Technically the current state is not proven to be safe yet,
10463 * but it will either reach outer most bpf_exit (which means it's safe)
10464 * or it will be rejected. When there are no loops the verifier won't be
10465 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
10466 * again on the way to bpf_exit.
10467 * When looping the sl->state.branches will be > 0 and this state
10468 * will not be considered for equivalence until branches == 0.
10470 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
10473 env->total_states++;
10474 env->peak_states++;
10475 env->prev_jmps_processed = env->jmps_processed;
10476 env->prev_insn_processed = env->insn_processed;
10478 /* add new state to the head of linked list */
10479 new = &new_sl->state;
10480 err = copy_verifier_state(new, cur);
10482 free_verifier_state(new, false);
10486 new->insn_idx = insn_idx;
10487 WARN_ONCE(new->branches != 1,
10488 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
10491 cur->first_insn_idx = insn_idx;
10492 clear_jmp_history(cur);
10493 new_sl->next = *explored_state(env, insn_idx);
10494 *explored_state(env, insn_idx) = new_sl;
10495 /* connect new state to parentage chain. Current frame needs all
10496 * registers connected. Only r6 - r9 of the callers are alive (pushed
10497 * to the stack implicitly by JITs) so in callers' frames connect just
10498 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
10499 * the state of the call instruction (with WRITTEN set), and r0 comes
10500 * from callee with its full parentage chain, anyway.
10502 /* clear write marks in current state: the writes we did are not writes
10503 * our child did, so they don't screen off its reads from us.
10504 * (There are no read marks in current state, because reads always mark
10505 * their parent and current state never has children yet. Only
10506 * explored_states can get read marks.)
10508 for (j = 0; j <= cur->curframe; j++) {
10509 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
10510 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
10511 for (i = 0; i < BPF_REG_FP; i++)
10512 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
10515 /* all stack frames are accessible from callee, clear them all */
10516 for (j = 0; j <= cur->curframe; j++) {
10517 struct bpf_func_state *frame = cur->frame[j];
10518 struct bpf_func_state *newframe = new->frame[j];
10520 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
10521 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
10522 frame->stack[i].spilled_ptr.parent =
10523 &newframe->stack[i].spilled_ptr;
10529 /* Return true if it's OK to have the same insn return a different type. */
10530 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
10534 case PTR_TO_SOCKET:
10535 case PTR_TO_SOCKET_OR_NULL:
10536 case PTR_TO_SOCK_COMMON:
10537 case PTR_TO_SOCK_COMMON_OR_NULL:
10538 case PTR_TO_TCP_SOCK:
10539 case PTR_TO_TCP_SOCK_OR_NULL:
10540 case PTR_TO_XDP_SOCK:
10541 case PTR_TO_BTF_ID:
10542 case PTR_TO_BTF_ID_OR_NULL:
10549 /* If an instruction was previously used with particular pointer types, then we
10550 * need to be careful to avoid cases such as the below, where it may be ok
10551 * for one branch accessing the pointer, but not ok for the other branch:
10556 * R1 = some_other_valid_ptr;
10559 * R2 = *(u32 *)(R1 + 0);
10561 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
10563 return src != prev && (!reg_type_mismatch_ok(src) ||
10564 !reg_type_mismatch_ok(prev));
10567 static int do_check(struct bpf_verifier_env *env)
10569 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
10570 struct bpf_verifier_state *state = env->cur_state;
10571 struct bpf_insn *insns = env->prog->insnsi;
10572 struct bpf_reg_state *regs;
10573 int insn_cnt = env->prog->len;
10574 bool do_print_state = false;
10575 int prev_insn_idx = -1;
10578 struct bpf_insn *insn;
10582 env->prev_insn_idx = prev_insn_idx;
10583 if (env->insn_idx >= insn_cnt) {
10584 verbose(env, "invalid insn idx %d insn_cnt %d\n",
10585 env->insn_idx, insn_cnt);
10589 insn = &insns[env->insn_idx];
10590 class = BPF_CLASS(insn->code);
10592 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
10594 "BPF program is too large. Processed %d insn\n",
10595 env->insn_processed);
10599 err = is_state_visited(env, env->insn_idx);
10603 /* found equivalent state, can prune the search */
10604 if (env->log.level & BPF_LOG_LEVEL) {
10605 if (do_print_state)
10606 verbose(env, "\nfrom %d to %d%s: safe\n",
10607 env->prev_insn_idx, env->insn_idx,
10608 env->cur_state->speculative ?
10609 " (speculative execution)" : "");
10611 verbose(env, "%d: safe\n", env->insn_idx);
10613 goto process_bpf_exit;
10616 if (signal_pending(current))
10619 if (need_resched())
10622 if (env->log.level & BPF_LOG_LEVEL2 ||
10623 (env->log.level & BPF_LOG_LEVEL && do_print_state)) {
10624 if (env->log.level & BPF_LOG_LEVEL2)
10625 verbose(env, "%d:", env->insn_idx);
10627 verbose(env, "\nfrom %d to %d%s:",
10628 env->prev_insn_idx, env->insn_idx,
10629 env->cur_state->speculative ?
10630 " (speculative execution)" : "");
10631 print_verifier_state(env, state->frame[state->curframe]);
10632 do_print_state = false;
10635 if (env->log.level & BPF_LOG_LEVEL) {
10636 const struct bpf_insn_cbs cbs = {
10637 .cb_call = disasm_kfunc_name,
10638 .cb_print = verbose,
10639 .private_data = env,
10642 verbose_linfo(env, env->insn_idx, "; ");
10643 verbose(env, "%d: ", env->insn_idx);
10644 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
10647 if (bpf_prog_is_dev_bound(env->prog->aux)) {
10648 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
10649 env->prev_insn_idx);
10654 regs = cur_regs(env);
10655 sanitize_mark_insn_seen(env);
10656 prev_insn_idx = env->insn_idx;
10658 if (class == BPF_ALU || class == BPF_ALU64) {
10659 err = check_alu_op(env, insn);
10663 } else if (class == BPF_LDX) {
10664 enum bpf_reg_type *prev_src_type, src_reg_type;
10666 /* check for reserved fields is already done */
10668 /* check src operand */
10669 err = check_reg_arg(env, insn->src_reg, SRC_OP);
10673 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
10677 src_reg_type = regs[insn->src_reg].type;
10679 /* check that memory (src_reg + off) is readable,
10680 * the state of dst_reg will be updated by this func
10682 err = check_mem_access(env, env->insn_idx, insn->src_reg,
10683 insn->off, BPF_SIZE(insn->code),
10684 BPF_READ, insn->dst_reg, false);
10688 prev_src_type = &env->insn_aux_data[env->insn_idx].ptr_type;
10690 if (*prev_src_type == NOT_INIT) {
10691 /* saw a valid insn
10692 * dst_reg = *(u32 *)(src_reg + off)
10693 * save type to validate intersecting paths
10695 *prev_src_type = src_reg_type;
10697 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
10698 /* ABuser program is trying to use the same insn
10699 * dst_reg = *(u32*) (src_reg + off)
10700 * with different pointer types:
10701 * src_reg == ctx in one branch and
10702 * src_reg == stack|map in some other branch.
10705 verbose(env, "same insn cannot be used with different pointers\n");
10709 } else if (class == BPF_STX) {
10710 enum bpf_reg_type *prev_dst_type, dst_reg_type;
10712 if (BPF_MODE(insn->code) == BPF_ATOMIC) {
10713 err = check_atomic(env, env->insn_idx, insn);
10720 if (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0) {
10721 verbose(env, "BPF_STX uses reserved fields\n");
10725 /* check src1 operand */
10726 err = check_reg_arg(env, insn->src_reg, SRC_OP);
10729 /* check src2 operand */
10730 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
10734 dst_reg_type = regs[insn->dst_reg].type;
10736 /* check that memory (dst_reg + off) is writeable */
10737 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
10738 insn->off, BPF_SIZE(insn->code),
10739 BPF_WRITE, insn->src_reg, false);
10743 prev_dst_type = &env->insn_aux_data[env->insn_idx].ptr_type;
10745 if (*prev_dst_type == NOT_INIT) {
10746 *prev_dst_type = dst_reg_type;
10747 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
10748 verbose(env, "same insn cannot be used with different pointers\n");
10752 } else if (class == BPF_ST) {
10753 if (BPF_MODE(insn->code) != BPF_MEM ||
10754 insn->src_reg != BPF_REG_0) {
10755 verbose(env, "BPF_ST uses reserved fields\n");
10758 /* check src operand */
10759 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
10763 if (is_ctx_reg(env, insn->dst_reg)) {
10764 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
10766 reg_type_str[reg_state(env, insn->dst_reg)->type]);
10770 /* check that memory (dst_reg + off) is writeable */
10771 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
10772 insn->off, BPF_SIZE(insn->code),
10773 BPF_WRITE, -1, false);
10777 } else if (class == BPF_JMP || class == BPF_JMP32) {
10778 u8 opcode = BPF_OP(insn->code);
10780 env->jmps_processed++;
10781 if (opcode == BPF_CALL) {
10782 if (BPF_SRC(insn->code) != BPF_K ||
10784 (insn->src_reg != BPF_REG_0 &&
10785 insn->src_reg != BPF_PSEUDO_CALL &&
10786 insn->src_reg != BPF_PSEUDO_KFUNC_CALL) ||
10787 insn->dst_reg != BPF_REG_0 ||
10788 class == BPF_JMP32) {
10789 verbose(env, "BPF_CALL uses reserved fields\n");
10793 if (env->cur_state->active_spin_lock &&
10794 (insn->src_reg == BPF_PSEUDO_CALL ||
10795 insn->imm != BPF_FUNC_spin_unlock)) {
10796 verbose(env, "function calls are not allowed while holding a lock\n");
10799 if (insn->src_reg == BPF_PSEUDO_CALL)
10800 err = check_func_call(env, insn, &env->insn_idx);
10801 else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL)
10802 err = check_kfunc_call(env, insn);
10804 err = check_helper_call(env, insn, &env->insn_idx);
10807 } else if (opcode == BPF_JA) {
10808 if (BPF_SRC(insn->code) != BPF_K ||
10810 insn->src_reg != BPF_REG_0 ||
10811 insn->dst_reg != BPF_REG_0 ||
10812 class == BPF_JMP32) {
10813 verbose(env, "BPF_JA uses reserved fields\n");
10817 env->insn_idx += insn->off + 1;
10820 } else if (opcode == BPF_EXIT) {
10821 if (BPF_SRC(insn->code) != BPF_K ||
10823 insn->src_reg != BPF_REG_0 ||
10824 insn->dst_reg != BPF_REG_0 ||
10825 class == BPF_JMP32) {
10826 verbose(env, "BPF_EXIT uses reserved fields\n");
10830 if (env->cur_state->active_spin_lock) {
10831 verbose(env, "bpf_spin_unlock is missing\n");
10835 if (state->curframe) {
10836 /* exit from nested function */
10837 err = prepare_func_exit(env, &env->insn_idx);
10840 do_print_state = true;
10844 err = check_reference_leak(env);
10848 err = check_return_code(env);
10852 update_branch_counts(env, env->cur_state);
10853 err = pop_stack(env, &prev_insn_idx,
10854 &env->insn_idx, pop_log);
10856 if (err != -ENOENT)
10860 do_print_state = true;
10864 err = check_cond_jmp_op(env, insn, &env->insn_idx);
10868 } else if (class == BPF_LD) {
10869 u8 mode = BPF_MODE(insn->code);
10871 if (mode == BPF_ABS || mode == BPF_IND) {
10872 err = check_ld_abs(env, insn);
10876 } else if (mode == BPF_IMM) {
10877 err = check_ld_imm(env, insn);
10882 sanitize_mark_insn_seen(env);
10884 verbose(env, "invalid BPF_LD mode\n");
10888 verbose(env, "unknown insn class %d\n", class);
10898 static int find_btf_percpu_datasec(struct btf *btf)
10900 const struct btf_type *t;
10905 * Both vmlinux and module each have their own ".data..percpu"
10906 * DATASECs in BTF. So for module's case, we need to skip vmlinux BTF
10907 * types to look at only module's own BTF types.
10909 n = btf_nr_types(btf);
10910 if (btf_is_module(btf))
10911 i = btf_nr_types(btf_vmlinux);
10915 for(; i < n; i++) {
10916 t = btf_type_by_id(btf, i);
10917 if (BTF_INFO_KIND(t->info) != BTF_KIND_DATASEC)
10920 tname = btf_name_by_offset(btf, t->name_off);
10921 if (!strcmp(tname, ".data..percpu"))
10928 /* replace pseudo btf_id with kernel symbol address */
10929 static int check_pseudo_btf_id(struct bpf_verifier_env *env,
10930 struct bpf_insn *insn,
10931 struct bpf_insn_aux_data *aux)
10933 const struct btf_var_secinfo *vsi;
10934 const struct btf_type *datasec;
10935 struct btf_mod_pair *btf_mod;
10936 const struct btf_type *t;
10937 const char *sym_name;
10938 bool percpu = false;
10939 u32 type, id = insn->imm;
10943 int i, btf_fd, err;
10945 btf_fd = insn[1].imm;
10947 btf = btf_get_by_fd(btf_fd);
10949 verbose(env, "invalid module BTF object FD specified.\n");
10953 if (!btf_vmlinux) {
10954 verbose(env, "kernel is missing BTF, make sure CONFIG_DEBUG_INFO_BTF=y is specified in Kconfig.\n");
10961 t = btf_type_by_id(btf, id);
10963 verbose(env, "ldimm64 insn specifies invalid btf_id %d.\n", id);
10968 if (!btf_type_is_var(t)) {
10969 verbose(env, "pseudo btf_id %d in ldimm64 isn't KIND_VAR.\n", id);
10974 sym_name = btf_name_by_offset(btf, t->name_off);
10975 addr = kallsyms_lookup_name(sym_name);
10977 verbose(env, "ldimm64 failed to find the address for kernel symbol '%s'.\n",
10983 datasec_id = find_btf_percpu_datasec(btf);
10984 if (datasec_id > 0) {
10985 datasec = btf_type_by_id(btf, datasec_id);
10986 for_each_vsi(i, datasec, vsi) {
10987 if (vsi->type == id) {
10994 insn[0].imm = (u32)addr;
10995 insn[1].imm = addr >> 32;
10998 t = btf_type_skip_modifiers(btf, type, NULL);
11000 aux->btf_var.reg_type = PTR_TO_PERCPU_BTF_ID;
11001 aux->btf_var.btf = btf;
11002 aux->btf_var.btf_id = type;
11003 } else if (!btf_type_is_struct(t)) {
11004 const struct btf_type *ret;
11008 /* resolve the type size of ksym. */
11009 ret = btf_resolve_size(btf, t, &tsize);
11011 tname = btf_name_by_offset(btf, t->name_off);
11012 verbose(env, "ldimm64 unable to resolve the size of type '%s': %ld\n",
11013 tname, PTR_ERR(ret));
11017 aux->btf_var.reg_type = PTR_TO_MEM;
11018 aux->btf_var.mem_size = tsize;
11020 aux->btf_var.reg_type = PTR_TO_BTF_ID;
11021 aux->btf_var.btf = btf;
11022 aux->btf_var.btf_id = type;
11025 /* check whether we recorded this BTF (and maybe module) already */
11026 for (i = 0; i < env->used_btf_cnt; i++) {
11027 if (env->used_btfs[i].btf == btf) {
11033 if (env->used_btf_cnt >= MAX_USED_BTFS) {
11038 btf_mod = &env->used_btfs[env->used_btf_cnt];
11039 btf_mod->btf = btf;
11040 btf_mod->module = NULL;
11042 /* if we reference variables from kernel module, bump its refcount */
11043 if (btf_is_module(btf)) {
11044 btf_mod->module = btf_try_get_module(btf);
11045 if (!btf_mod->module) {
11051 env->used_btf_cnt++;
11059 static int check_map_prealloc(struct bpf_map *map)
11061 return (map->map_type != BPF_MAP_TYPE_HASH &&
11062 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
11063 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
11064 !(map->map_flags & BPF_F_NO_PREALLOC);
11067 static bool is_tracing_prog_type(enum bpf_prog_type type)
11070 case BPF_PROG_TYPE_KPROBE:
11071 case BPF_PROG_TYPE_TRACEPOINT:
11072 case BPF_PROG_TYPE_PERF_EVENT:
11073 case BPF_PROG_TYPE_RAW_TRACEPOINT:
11080 static bool is_preallocated_map(struct bpf_map *map)
11082 if (!check_map_prealloc(map))
11084 if (map->inner_map_meta && !check_map_prealloc(map->inner_map_meta))
11089 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
11090 struct bpf_map *map,
11091 struct bpf_prog *prog)
11094 enum bpf_prog_type prog_type = resolve_prog_type(prog);
11096 * Validate that trace type programs use preallocated hash maps.
11098 * For programs attached to PERF events this is mandatory as the
11099 * perf NMI can hit any arbitrary code sequence.
11101 * All other trace types using preallocated hash maps are unsafe as
11102 * well because tracepoint or kprobes can be inside locked regions
11103 * of the memory allocator or at a place where a recursion into the
11104 * memory allocator would see inconsistent state.
11106 * On RT enabled kernels run-time allocation of all trace type
11107 * programs is strictly prohibited due to lock type constraints. On
11108 * !RT kernels it is allowed for backwards compatibility reasons for
11109 * now, but warnings are emitted so developers are made aware of
11110 * the unsafety and can fix their programs before this is enforced.
11112 if (is_tracing_prog_type(prog_type) && !is_preallocated_map(map)) {
11113 if (prog_type == BPF_PROG_TYPE_PERF_EVENT) {
11114 verbose(env, "perf_event programs can only use preallocated hash map\n");
11117 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
11118 verbose(env, "trace type programs can only use preallocated hash map\n");
11121 WARN_ONCE(1, "trace type BPF program uses run-time allocation\n");
11122 verbose(env, "trace type programs with run-time allocated hash maps are unsafe. Switch to preallocated hash maps.\n");
11125 if (map_value_has_spin_lock(map)) {
11126 if (prog_type == BPF_PROG_TYPE_SOCKET_FILTER) {
11127 verbose(env, "socket filter progs cannot use bpf_spin_lock yet\n");
11131 if (is_tracing_prog_type(prog_type)) {
11132 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
11136 if (prog->aux->sleepable) {
11137 verbose(env, "sleepable progs cannot use bpf_spin_lock yet\n");
11142 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
11143 !bpf_offload_prog_map_match(prog, map)) {
11144 verbose(env, "offload device mismatch between prog and map\n");
11148 if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
11149 verbose(env, "bpf_struct_ops map cannot be used in prog\n");
11153 if (prog->aux->sleepable)
11154 switch (map->map_type) {
11155 case BPF_MAP_TYPE_HASH:
11156 case BPF_MAP_TYPE_LRU_HASH:
11157 case BPF_MAP_TYPE_ARRAY:
11158 case BPF_MAP_TYPE_PERCPU_HASH:
11159 case BPF_MAP_TYPE_PERCPU_ARRAY:
11160 case BPF_MAP_TYPE_LRU_PERCPU_HASH:
11161 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
11162 case BPF_MAP_TYPE_HASH_OF_MAPS:
11163 if (!is_preallocated_map(map)) {
11165 "Sleepable programs can only use preallocated maps\n");
11169 case BPF_MAP_TYPE_RINGBUF:
11173 "Sleepable programs can only use array, hash, and ringbuf maps\n");
11180 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
11182 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
11183 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
11186 /* find and rewrite pseudo imm in ld_imm64 instructions:
11188 * 1. if it accesses map FD, replace it with actual map pointer.
11189 * 2. if it accesses btf_id of a VAR, replace it with pointer to the var.
11191 * NOTE: btf_vmlinux is required for converting pseudo btf_id.
11193 static int resolve_pseudo_ldimm64(struct bpf_verifier_env *env)
11195 struct bpf_insn *insn = env->prog->insnsi;
11196 int insn_cnt = env->prog->len;
11199 err = bpf_prog_calc_tag(env->prog);
11203 for (i = 0; i < insn_cnt; i++, insn++) {
11204 if (BPF_CLASS(insn->code) == BPF_LDX &&
11205 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
11206 verbose(env, "BPF_LDX uses reserved fields\n");
11210 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
11211 struct bpf_insn_aux_data *aux;
11212 struct bpf_map *map;
11216 if (i == insn_cnt - 1 || insn[1].code != 0 ||
11217 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
11218 insn[1].off != 0) {
11219 verbose(env, "invalid bpf_ld_imm64 insn\n");
11223 if (insn[0].src_reg == 0)
11224 /* valid generic load 64-bit imm */
11227 if (insn[0].src_reg == BPF_PSEUDO_BTF_ID) {
11228 aux = &env->insn_aux_data[i];
11229 err = check_pseudo_btf_id(env, insn, aux);
11235 if (insn[0].src_reg == BPF_PSEUDO_FUNC) {
11236 aux = &env->insn_aux_data[i];
11237 aux->ptr_type = PTR_TO_FUNC;
11241 /* In final convert_pseudo_ld_imm64() step, this is
11242 * converted into regular 64-bit imm load insn.
11244 if ((insn[0].src_reg != BPF_PSEUDO_MAP_FD &&
11245 insn[0].src_reg != BPF_PSEUDO_MAP_VALUE) ||
11246 (insn[0].src_reg == BPF_PSEUDO_MAP_FD &&
11247 insn[1].imm != 0)) {
11249 "unrecognized bpf_ld_imm64 insn\n");
11253 f = fdget(insn[0].imm);
11254 map = __bpf_map_get(f);
11256 verbose(env, "fd %d is not pointing to valid bpf_map\n",
11258 return PTR_ERR(map);
11261 err = check_map_prog_compatibility(env, map, env->prog);
11267 aux = &env->insn_aux_data[i];
11268 if (insn->src_reg == BPF_PSEUDO_MAP_FD) {
11269 addr = (unsigned long)map;
11271 u32 off = insn[1].imm;
11273 if (off >= BPF_MAX_VAR_OFF) {
11274 verbose(env, "direct value offset of %u is not allowed\n", off);
11279 if (!map->ops->map_direct_value_addr) {
11280 verbose(env, "no direct value access support for this map type\n");
11285 err = map->ops->map_direct_value_addr(map, &addr, off);
11287 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
11288 map->value_size, off);
11293 aux->map_off = off;
11297 insn[0].imm = (u32)addr;
11298 insn[1].imm = addr >> 32;
11300 /* check whether we recorded this map already */
11301 for (j = 0; j < env->used_map_cnt; j++) {
11302 if (env->used_maps[j] == map) {
11303 aux->map_index = j;
11309 if (env->used_map_cnt >= MAX_USED_MAPS) {
11314 /* hold the map. If the program is rejected by verifier,
11315 * the map will be released by release_maps() or it
11316 * will be used by the valid program until it's unloaded
11317 * and all maps are released in free_used_maps()
11321 aux->map_index = env->used_map_cnt;
11322 env->used_maps[env->used_map_cnt++] = map;
11324 if (bpf_map_is_cgroup_storage(map) &&
11325 bpf_cgroup_storage_assign(env->prog->aux, map)) {
11326 verbose(env, "only one cgroup storage of each type is allowed\n");
11338 /* Basic sanity check before we invest more work here. */
11339 if (!bpf_opcode_in_insntable(insn->code)) {
11340 verbose(env, "unknown opcode %02x\n", insn->code);
11345 /* now all pseudo BPF_LD_IMM64 instructions load valid
11346 * 'struct bpf_map *' into a register instead of user map_fd.
11347 * These pointers will be used later by verifier to validate map access.
11352 /* drop refcnt of maps used by the rejected program */
11353 static void release_maps(struct bpf_verifier_env *env)
11355 __bpf_free_used_maps(env->prog->aux, env->used_maps,
11356 env->used_map_cnt);
11359 /* drop refcnt of maps used by the rejected program */
11360 static void release_btfs(struct bpf_verifier_env *env)
11362 __bpf_free_used_btfs(env->prog->aux, env->used_btfs,
11363 env->used_btf_cnt);
11366 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
11367 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
11369 struct bpf_insn *insn = env->prog->insnsi;
11370 int insn_cnt = env->prog->len;
11373 for (i = 0; i < insn_cnt; i++, insn++) {
11374 if (insn->code != (BPF_LD | BPF_IMM | BPF_DW))
11376 if (insn->src_reg == BPF_PSEUDO_FUNC)
11382 /* single env->prog->insni[off] instruction was replaced with the range
11383 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
11384 * [0, off) and [off, end) to new locations, so the patched range stays zero
11386 static int adjust_insn_aux_data(struct bpf_verifier_env *env,
11387 struct bpf_prog *new_prog, u32 off, u32 cnt)
11389 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
11390 struct bpf_insn *insn = new_prog->insnsi;
11391 u32 old_seen = old_data[off].seen;
11395 /* aux info at OFF always needs adjustment, no matter fast path
11396 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
11397 * original insn at old prog.
11399 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
11403 prog_len = new_prog->len;
11404 new_data = vzalloc(array_size(prog_len,
11405 sizeof(struct bpf_insn_aux_data)));
11408 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
11409 memcpy(new_data + off + cnt - 1, old_data + off,
11410 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
11411 for (i = off; i < off + cnt - 1; i++) {
11412 /* Expand insni[off]'s seen count to the patched range. */
11413 new_data[i].seen = old_seen;
11414 new_data[i].zext_dst = insn_has_def32(env, insn + i);
11416 env->insn_aux_data = new_data;
11421 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
11427 /* NOTE: fake 'exit' subprog should be updated as well. */
11428 for (i = 0; i <= env->subprog_cnt; i++) {
11429 if (env->subprog_info[i].start <= off)
11431 env->subprog_info[i].start += len - 1;
11435 static void adjust_poke_descs(struct bpf_prog *prog, u32 off, u32 len)
11437 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
11438 int i, sz = prog->aux->size_poke_tab;
11439 struct bpf_jit_poke_descriptor *desc;
11441 for (i = 0; i < sz; i++) {
11443 if (desc->insn_idx <= off)
11445 desc->insn_idx += len - 1;
11449 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
11450 const struct bpf_insn *patch, u32 len)
11452 struct bpf_prog *new_prog;
11454 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
11455 if (IS_ERR(new_prog)) {
11456 if (PTR_ERR(new_prog) == -ERANGE)
11458 "insn %d cannot be patched due to 16-bit range\n",
11459 env->insn_aux_data[off].orig_idx);
11462 if (adjust_insn_aux_data(env, new_prog, off, len))
11464 adjust_subprog_starts(env, off, len);
11465 adjust_poke_descs(new_prog, off, len);
11469 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
11474 /* find first prog starting at or after off (first to remove) */
11475 for (i = 0; i < env->subprog_cnt; i++)
11476 if (env->subprog_info[i].start >= off)
11478 /* find first prog starting at or after off + cnt (first to stay) */
11479 for (j = i; j < env->subprog_cnt; j++)
11480 if (env->subprog_info[j].start >= off + cnt)
11482 /* if j doesn't start exactly at off + cnt, we are just removing
11483 * the front of previous prog
11485 if (env->subprog_info[j].start != off + cnt)
11489 struct bpf_prog_aux *aux = env->prog->aux;
11492 /* move fake 'exit' subprog as well */
11493 move = env->subprog_cnt + 1 - j;
11495 memmove(env->subprog_info + i,
11496 env->subprog_info + j,
11497 sizeof(*env->subprog_info) * move);
11498 env->subprog_cnt -= j - i;
11500 /* remove func_info */
11501 if (aux->func_info) {
11502 move = aux->func_info_cnt - j;
11504 memmove(aux->func_info + i,
11505 aux->func_info + j,
11506 sizeof(*aux->func_info) * move);
11507 aux->func_info_cnt -= j - i;
11508 /* func_info->insn_off is set after all code rewrites,
11509 * in adjust_btf_func() - no need to adjust
11513 /* convert i from "first prog to remove" to "first to adjust" */
11514 if (env->subprog_info[i].start == off)
11518 /* update fake 'exit' subprog as well */
11519 for (; i <= env->subprog_cnt; i++)
11520 env->subprog_info[i].start -= cnt;
11525 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
11528 struct bpf_prog *prog = env->prog;
11529 u32 i, l_off, l_cnt, nr_linfo;
11530 struct bpf_line_info *linfo;
11532 nr_linfo = prog->aux->nr_linfo;
11536 linfo = prog->aux->linfo;
11538 /* find first line info to remove, count lines to be removed */
11539 for (i = 0; i < nr_linfo; i++)
11540 if (linfo[i].insn_off >= off)
11545 for (; i < nr_linfo; i++)
11546 if (linfo[i].insn_off < off + cnt)
11551 /* First live insn doesn't match first live linfo, it needs to "inherit"
11552 * last removed linfo. prog is already modified, so prog->len == off
11553 * means no live instructions after (tail of the program was removed).
11555 if (prog->len != off && l_cnt &&
11556 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
11558 linfo[--i].insn_off = off + cnt;
11561 /* remove the line info which refer to the removed instructions */
11563 memmove(linfo + l_off, linfo + i,
11564 sizeof(*linfo) * (nr_linfo - i));
11566 prog->aux->nr_linfo -= l_cnt;
11567 nr_linfo = prog->aux->nr_linfo;
11570 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
11571 for (i = l_off; i < nr_linfo; i++)
11572 linfo[i].insn_off -= cnt;
11574 /* fix up all subprogs (incl. 'exit') which start >= off */
11575 for (i = 0; i <= env->subprog_cnt; i++)
11576 if (env->subprog_info[i].linfo_idx > l_off) {
11577 /* program may have started in the removed region but
11578 * may not be fully removed
11580 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
11581 env->subprog_info[i].linfo_idx -= l_cnt;
11583 env->subprog_info[i].linfo_idx = l_off;
11589 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
11591 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
11592 unsigned int orig_prog_len = env->prog->len;
11595 if (bpf_prog_is_dev_bound(env->prog->aux))
11596 bpf_prog_offload_remove_insns(env, off, cnt);
11598 err = bpf_remove_insns(env->prog, off, cnt);
11602 err = adjust_subprog_starts_after_remove(env, off, cnt);
11606 err = bpf_adj_linfo_after_remove(env, off, cnt);
11610 memmove(aux_data + off, aux_data + off + cnt,
11611 sizeof(*aux_data) * (orig_prog_len - off - cnt));
11616 /* The verifier does more data flow analysis than llvm and will not
11617 * explore branches that are dead at run time. Malicious programs can
11618 * have dead code too. Therefore replace all dead at-run-time code
11621 * Just nops are not optimal, e.g. if they would sit at the end of the
11622 * program and through another bug we would manage to jump there, then
11623 * we'd execute beyond program memory otherwise. Returning exception
11624 * code also wouldn't work since we can have subprogs where the dead
11625 * code could be located.
11627 static void sanitize_dead_code(struct bpf_verifier_env *env)
11629 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
11630 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
11631 struct bpf_insn *insn = env->prog->insnsi;
11632 const int insn_cnt = env->prog->len;
11635 for (i = 0; i < insn_cnt; i++) {
11636 if (aux_data[i].seen)
11638 memcpy(insn + i, &trap, sizeof(trap));
11639 aux_data[i].zext_dst = false;
11643 static bool insn_is_cond_jump(u8 code)
11647 if (BPF_CLASS(code) == BPF_JMP32)
11650 if (BPF_CLASS(code) != BPF_JMP)
11654 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
11657 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
11659 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
11660 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
11661 struct bpf_insn *insn = env->prog->insnsi;
11662 const int insn_cnt = env->prog->len;
11665 for (i = 0; i < insn_cnt; i++, insn++) {
11666 if (!insn_is_cond_jump(insn->code))
11669 if (!aux_data[i + 1].seen)
11670 ja.off = insn->off;
11671 else if (!aux_data[i + 1 + insn->off].seen)
11676 if (bpf_prog_is_dev_bound(env->prog->aux))
11677 bpf_prog_offload_replace_insn(env, i, &ja);
11679 memcpy(insn, &ja, sizeof(ja));
11683 static int opt_remove_dead_code(struct bpf_verifier_env *env)
11685 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
11686 int insn_cnt = env->prog->len;
11689 for (i = 0; i < insn_cnt; i++) {
11693 while (i + j < insn_cnt && !aux_data[i + j].seen)
11698 err = verifier_remove_insns(env, i, j);
11701 insn_cnt = env->prog->len;
11707 static int opt_remove_nops(struct bpf_verifier_env *env)
11709 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
11710 struct bpf_insn *insn = env->prog->insnsi;
11711 int insn_cnt = env->prog->len;
11714 for (i = 0; i < insn_cnt; i++) {
11715 if (memcmp(&insn[i], &ja, sizeof(ja)))
11718 err = verifier_remove_insns(env, i, 1);
11728 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
11729 const union bpf_attr *attr)
11731 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
11732 struct bpf_insn_aux_data *aux = env->insn_aux_data;
11733 int i, patch_len, delta = 0, len = env->prog->len;
11734 struct bpf_insn *insns = env->prog->insnsi;
11735 struct bpf_prog *new_prog;
11738 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
11739 zext_patch[1] = BPF_ZEXT_REG(0);
11740 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
11741 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
11742 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
11743 for (i = 0; i < len; i++) {
11744 int adj_idx = i + delta;
11745 struct bpf_insn insn;
11748 insn = insns[adj_idx];
11749 load_reg = insn_def_regno(&insn);
11750 if (!aux[adj_idx].zext_dst) {
11758 class = BPF_CLASS(code);
11759 if (load_reg == -1)
11762 /* NOTE: arg "reg" (the fourth one) is only used for
11763 * BPF_STX + SRC_OP, so it is safe to pass NULL
11766 if (is_reg64(env, &insn, load_reg, NULL, DST_OP)) {
11767 if (class == BPF_LD &&
11768 BPF_MODE(code) == BPF_IMM)
11773 /* ctx load could be transformed into wider load. */
11774 if (class == BPF_LDX &&
11775 aux[adj_idx].ptr_type == PTR_TO_CTX)
11778 imm_rnd = get_random_int();
11779 rnd_hi32_patch[0] = insn;
11780 rnd_hi32_patch[1].imm = imm_rnd;
11781 rnd_hi32_patch[3].dst_reg = load_reg;
11782 patch = rnd_hi32_patch;
11784 goto apply_patch_buffer;
11787 /* Add in an zero-extend instruction if a) the JIT has requested
11788 * it or b) it's a CMPXCHG.
11790 * The latter is because: BPF_CMPXCHG always loads a value into
11791 * R0, therefore always zero-extends. However some archs'
11792 * equivalent instruction only does this load when the
11793 * comparison is successful. This detail of CMPXCHG is
11794 * orthogonal to the general zero-extension behaviour of the
11795 * CPU, so it's treated independently of bpf_jit_needs_zext.
11797 if (!bpf_jit_needs_zext() && !is_cmpxchg_insn(&insn))
11800 if (WARN_ON(load_reg == -1)) {
11801 verbose(env, "verifier bug. zext_dst is set, but no reg is defined\n");
11805 zext_patch[0] = insn;
11806 zext_patch[1].dst_reg = load_reg;
11807 zext_patch[1].src_reg = load_reg;
11808 patch = zext_patch;
11810 apply_patch_buffer:
11811 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
11814 env->prog = new_prog;
11815 insns = new_prog->insnsi;
11816 aux = env->insn_aux_data;
11817 delta += patch_len - 1;
11823 /* convert load instructions that access fields of a context type into a
11824 * sequence of instructions that access fields of the underlying structure:
11825 * struct __sk_buff -> struct sk_buff
11826 * struct bpf_sock_ops -> struct sock
11828 static int convert_ctx_accesses(struct bpf_verifier_env *env)
11830 const struct bpf_verifier_ops *ops = env->ops;
11831 int i, cnt, size, ctx_field_size, delta = 0;
11832 const int insn_cnt = env->prog->len;
11833 struct bpf_insn insn_buf[16], *insn;
11834 u32 target_size, size_default, off;
11835 struct bpf_prog *new_prog;
11836 enum bpf_access_type type;
11837 bool is_narrower_load;
11839 if (ops->gen_prologue || env->seen_direct_write) {
11840 if (!ops->gen_prologue) {
11841 verbose(env, "bpf verifier is misconfigured\n");
11844 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
11846 if (cnt >= ARRAY_SIZE(insn_buf)) {
11847 verbose(env, "bpf verifier is misconfigured\n");
11850 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
11854 env->prog = new_prog;
11859 if (bpf_prog_is_dev_bound(env->prog->aux))
11862 insn = env->prog->insnsi + delta;
11864 for (i = 0; i < insn_cnt; i++, insn++) {
11865 bpf_convert_ctx_access_t convert_ctx_access;
11868 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
11869 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
11870 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
11871 insn->code == (BPF_LDX | BPF_MEM | BPF_DW)) {
11874 } else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
11875 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
11876 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
11877 insn->code == (BPF_STX | BPF_MEM | BPF_DW) ||
11878 insn->code == (BPF_ST | BPF_MEM | BPF_B) ||
11879 insn->code == (BPF_ST | BPF_MEM | BPF_H) ||
11880 insn->code == (BPF_ST | BPF_MEM | BPF_W) ||
11881 insn->code == (BPF_ST | BPF_MEM | BPF_DW)) {
11883 ctx_access = BPF_CLASS(insn->code) == BPF_STX;
11888 if (type == BPF_WRITE &&
11889 env->insn_aux_data[i + delta].sanitize_stack_spill) {
11890 struct bpf_insn patch[] = {
11895 cnt = ARRAY_SIZE(patch);
11896 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
11901 env->prog = new_prog;
11902 insn = new_prog->insnsi + i + delta;
11909 switch (env->insn_aux_data[i + delta].ptr_type) {
11911 if (!ops->convert_ctx_access)
11913 convert_ctx_access = ops->convert_ctx_access;
11915 case PTR_TO_SOCKET:
11916 case PTR_TO_SOCK_COMMON:
11917 convert_ctx_access = bpf_sock_convert_ctx_access;
11919 case PTR_TO_TCP_SOCK:
11920 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
11922 case PTR_TO_XDP_SOCK:
11923 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
11925 case PTR_TO_BTF_ID:
11926 if (type == BPF_READ) {
11927 insn->code = BPF_LDX | BPF_PROBE_MEM |
11928 BPF_SIZE((insn)->code);
11929 env->prog->aux->num_exentries++;
11930 } else if (resolve_prog_type(env->prog) != BPF_PROG_TYPE_STRUCT_OPS) {
11931 verbose(env, "Writes through BTF pointers are not allowed\n");
11939 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
11940 size = BPF_LDST_BYTES(insn);
11942 /* If the read access is a narrower load of the field,
11943 * convert to a 4/8-byte load, to minimum program type specific
11944 * convert_ctx_access changes. If conversion is successful,
11945 * we will apply proper mask to the result.
11947 is_narrower_load = size < ctx_field_size;
11948 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
11950 if (is_narrower_load) {
11953 if (type == BPF_WRITE) {
11954 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
11959 if (ctx_field_size == 4)
11961 else if (ctx_field_size == 8)
11962 size_code = BPF_DW;
11964 insn->off = off & ~(size_default - 1);
11965 insn->code = BPF_LDX | BPF_MEM | size_code;
11969 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
11971 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
11972 (ctx_field_size && !target_size)) {
11973 verbose(env, "bpf verifier is misconfigured\n");
11977 if (is_narrower_load && size < target_size) {
11978 u8 shift = bpf_ctx_narrow_access_offset(
11979 off, size, size_default) * 8;
11980 if (ctx_field_size <= 4) {
11982 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
11985 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
11986 (1 << size * 8) - 1);
11989 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
11992 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
11993 (1ULL << size * 8) - 1);
11997 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12003 /* keep walking new program and skip insns we just inserted */
12004 env->prog = new_prog;
12005 insn = new_prog->insnsi + i + delta;
12011 static int jit_subprogs(struct bpf_verifier_env *env)
12013 struct bpf_prog *prog = env->prog, **func, *tmp;
12014 int i, j, subprog_start, subprog_end = 0, len, subprog;
12015 struct bpf_map *map_ptr;
12016 struct bpf_insn *insn;
12017 void *old_bpf_func;
12018 int err, num_exentries;
12020 if (env->subprog_cnt <= 1)
12023 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
12024 if (bpf_pseudo_func(insn)) {
12025 env->insn_aux_data[i].call_imm = insn->imm;
12026 /* subprog is encoded in insn[1].imm */
12030 if (!bpf_pseudo_call(insn))
12032 /* Upon error here we cannot fall back to interpreter but
12033 * need a hard reject of the program. Thus -EFAULT is
12034 * propagated in any case.
12036 subprog = find_subprog(env, i + insn->imm + 1);
12038 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
12039 i + insn->imm + 1);
12042 /* temporarily remember subprog id inside insn instead of
12043 * aux_data, since next loop will split up all insns into funcs
12045 insn->off = subprog;
12046 /* remember original imm in case JIT fails and fallback
12047 * to interpreter will be needed
12049 env->insn_aux_data[i].call_imm = insn->imm;
12050 /* point imm to __bpf_call_base+1 from JITs point of view */
12054 err = bpf_prog_alloc_jited_linfo(prog);
12056 goto out_undo_insn;
12059 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
12061 goto out_undo_insn;
12063 for (i = 0; i < env->subprog_cnt; i++) {
12064 subprog_start = subprog_end;
12065 subprog_end = env->subprog_info[i + 1].start;
12067 len = subprog_end - subprog_start;
12068 /* BPF_PROG_RUN doesn't call subprogs directly,
12069 * hence main prog stats include the runtime of subprogs.
12070 * subprogs don't have IDs and not reachable via prog_get_next_id
12071 * func[i]->stats will never be accessed and stays NULL
12073 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
12076 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
12077 len * sizeof(struct bpf_insn));
12078 func[i]->type = prog->type;
12079 func[i]->len = len;
12080 if (bpf_prog_calc_tag(func[i]))
12082 func[i]->is_func = 1;
12083 func[i]->aux->func_idx = i;
12084 /* Below members will be freed only at prog->aux */
12085 func[i]->aux->btf = prog->aux->btf;
12086 func[i]->aux->func_info = prog->aux->func_info;
12087 func[i]->aux->poke_tab = prog->aux->poke_tab;
12088 func[i]->aux->size_poke_tab = prog->aux->size_poke_tab;
12090 for (j = 0; j < prog->aux->size_poke_tab; j++) {
12091 struct bpf_jit_poke_descriptor *poke;
12093 poke = &prog->aux->poke_tab[j];
12094 if (poke->insn_idx < subprog_end &&
12095 poke->insn_idx >= subprog_start)
12096 poke->aux = func[i]->aux;
12099 /* Use bpf_prog_F_tag to indicate functions in stack traces.
12100 * Long term would need debug info to populate names
12102 func[i]->aux->name[0] = 'F';
12103 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
12104 func[i]->jit_requested = 1;
12105 func[i]->aux->kfunc_tab = prog->aux->kfunc_tab;
12106 func[i]->aux->linfo = prog->aux->linfo;
12107 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
12108 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
12109 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
12111 insn = func[i]->insnsi;
12112 for (j = 0; j < func[i]->len; j++, insn++) {
12113 if (BPF_CLASS(insn->code) == BPF_LDX &&
12114 BPF_MODE(insn->code) == BPF_PROBE_MEM)
12117 func[i]->aux->num_exentries = num_exentries;
12118 func[i]->aux->tail_call_reachable = env->subprog_info[i].tail_call_reachable;
12119 func[i] = bpf_int_jit_compile(func[i]);
12120 if (!func[i]->jited) {
12127 /* at this point all bpf functions were successfully JITed
12128 * now populate all bpf_calls with correct addresses and
12129 * run last pass of JIT
12131 for (i = 0; i < env->subprog_cnt; i++) {
12132 insn = func[i]->insnsi;
12133 for (j = 0; j < func[i]->len; j++, insn++) {
12134 if (bpf_pseudo_func(insn)) {
12135 subprog = insn[1].imm;
12136 insn[0].imm = (u32)(long)func[subprog]->bpf_func;
12137 insn[1].imm = ((u64)(long)func[subprog]->bpf_func) >> 32;
12140 if (!bpf_pseudo_call(insn))
12142 subprog = insn->off;
12143 insn->imm = BPF_CAST_CALL(func[subprog]->bpf_func) -
12147 /* we use the aux data to keep a list of the start addresses
12148 * of the JITed images for each function in the program
12150 * for some architectures, such as powerpc64, the imm field
12151 * might not be large enough to hold the offset of the start
12152 * address of the callee's JITed image from __bpf_call_base
12154 * in such cases, we can lookup the start address of a callee
12155 * by using its subprog id, available from the off field of
12156 * the call instruction, as an index for this list
12158 func[i]->aux->func = func;
12159 func[i]->aux->func_cnt = env->subprog_cnt;
12161 for (i = 0; i < env->subprog_cnt; i++) {
12162 old_bpf_func = func[i]->bpf_func;
12163 tmp = bpf_int_jit_compile(func[i]);
12164 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
12165 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
12172 /* finally lock prog and jit images for all functions and
12173 * populate kallsysm
12175 for (i = 0; i < env->subprog_cnt; i++) {
12176 bpf_prog_lock_ro(func[i]);
12177 bpf_prog_kallsyms_add(func[i]);
12180 /* Last step: make now unused interpreter insns from main
12181 * prog consistent for later dump requests, so they can
12182 * later look the same as if they were interpreted only.
12184 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
12185 if (bpf_pseudo_func(insn)) {
12186 insn[0].imm = env->insn_aux_data[i].call_imm;
12187 insn[1].imm = find_subprog(env, i + insn[0].imm + 1);
12190 if (!bpf_pseudo_call(insn))
12192 insn->off = env->insn_aux_data[i].call_imm;
12193 subprog = find_subprog(env, i + insn->off + 1);
12194 insn->imm = subprog;
12198 prog->bpf_func = func[0]->bpf_func;
12199 prog->aux->func = func;
12200 prog->aux->func_cnt = env->subprog_cnt;
12201 bpf_prog_jit_attempt_done(prog);
12204 /* We failed JIT'ing, so at this point we need to unregister poke
12205 * descriptors from subprogs, so that kernel is not attempting to
12206 * patch it anymore as we're freeing the subprog JIT memory.
12208 for (i = 0; i < prog->aux->size_poke_tab; i++) {
12209 map_ptr = prog->aux->poke_tab[i].tail_call.map;
12210 map_ptr->ops->map_poke_untrack(map_ptr, prog->aux);
12212 /* At this point we're guaranteed that poke descriptors are not
12213 * live anymore. We can just unlink its descriptor table as it's
12214 * released with the main prog.
12216 for (i = 0; i < env->subprog_cnt; i++) {
12219 func[i]->aux->poke_tab = NULL;
12220 bpf_jit_free(func[i]);
12224 /* cleanup main prog to be interpreted */
12225 prog->jit_requested = 0;
12226 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
12227 if (!bpf_pseudo_call(insn))
12230 insn->imm = env->insn_aux_data[i].call_imm;
12232 bpf_prog_jit_attempt_done(prog);
12236 static int fixup_call_args(struct bpf_verifier_env *env)
12238 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
12239 struct bpf_prog *prog = env->prog;
12240 struct bpf_insn *insn = prog->insnsi;
12241 bool has_kfunc_call = bpf_prog_has_kfunc_call(prog);
12246 if (env->prog->jit_requested &&
12247 !bpf_prog_is_dev_bound(env->prog->aux)) {
12248 err = jit_subprogs(env);
12251 if (err == -EFAULT)
12254 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
12255 if (has_kfunc_call) {
12256 verbose(env, "calling kernel functions are not allowed in non-JITed programs\n");
12259 if (env->subprog_cnt > 1 && env->prog->aux->tail_call_reachable) {
12260 /* When JIT fails the progs with bpf2bpf calls and tail_calls
12261 * have to be rejected, since interpreter doesn't support them yet.
12263 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
12266 for (i = 0; i < prog->len; i++, insn++) {
12267 if (bpf_pseudo_func(insn)) {
12268 /* When JIT fails the progs with callback calls
12269 * have to be rejected, since interpreter doesn't support them yet.
12271 verbose(env, "callbacks are not allowed in non-JITed programs\n");
12275 if (!bpf_pseudo_call(insn))
12277 depth = get_callee_stack_depth(env, insn, i);
12280 bpf_patch_call_args(insn, depth);
12287 static int fixup_kfunc_call(struct bpf_verifier_env *env,
12288 struct bpf_insn *insn)
12290 const struct bpf_kfunc_desc *desc;
12292 /* insn->imm has the btf func_id. Replace it with
12293 * an address (relative to __bpf_base_call).
12295 desc = find_kfunc_desc(env->prog, insn->imm);
12297 verbose(env, "verifier internal error: kernel function descriptor not found for func_id %u\n",
12302 insn->imm = desc->imm;
12307 /* Do various post-verification rewrites in a single program pass.
12308 * These rewrites simplify JIT and interpreter implementations.
12310 static int do_misc_fixups(struct bpf_verifier_env *env)
12312 struct bpf_prog *prog = env->prog;
12313 bool expect_blinding = bpf_jit_blinding_enabled(prog);
12314 struct bpf_insn *insn = prog->insnsi;
12315 const struct bpf_func_proto *fn;
12316 const int insn_cnt = prog->len;
12317 const struct bpf_map_ops *ops;
12318 struct bpf_insn_aux_data *aux;
12319 struct bpf_insn insn_buf[16];
12320 struct bpf_prog *new_prog;
12321 struct bpf_map *map_ptr;
12322 int i, ret, cnt, delta = 0;
12324 for (i = 0; i < insn_cnt; i++, insn++) {
12325 /* Make divide-by-zero exceptions impossible. */
12326 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
12327 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
12328 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
12329 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
12330 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
12331 bool isdiv = BPF_OP(insn->code) == BPF_DIV;
12332 struct bpf_insn *patchlet;
12333 struct bpf_insn chk_and_div[] = {
12334 /* [R,W]x div 0 -> 0 */
12335 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
12336 BPF_JNE | BPF_K, insn->src_reg,
12338 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
12339 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
12342 struct bpf_insn chk_and_mod[] = {
12343 /* [R,W]x mod 0 -> [R,W]x */
12344 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
12345 BPF_JEQ | BPF_K, insn->src_reg,
12346 0, 1 + (is64 ? 0 : 1), 0),
12348 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
12349 BPF_MOV32_REG(insn->dst_reg, insn->dst_reg),
12352 patchlet = isdiv ? chk_and_div : chk_and_mod;
12353 cnt = isdiv ? ARRAY_SIZE(chk_and_div) :
12354 ARRAY_SIZE(chk_and_mod) - (is64 ? 2 : 0);
12356 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
12361 env->prog = prog = new_prog;
12362 insn = new_prog->insnsi + i + delta;
12366 /* Implement LD_ABS and LD_IND with a rewrite, if supported by the program type. */
12367 if (BPF_CLASS(insn->code) == BPF_LD &&
12368 (BPF_MODE(insn->code) == BPF_ABS ||
12369 BPF_MODE(insn->code) == BPF_IND)) {
12370 cnt = env->ops->gen_ld_abs(insn, insn_buf);
12371 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
12372 verbose(env, "bpf verifier is misconfigured\n");
12376 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12381 env->prog = prog = new_prog;
12382 insn = new_prog->insnsi + i + delta;
12386 /* Rewrite pointer arithmetic to mitigate speculation attacks. */
12387 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
12388 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
12389 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
12390 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
12391 struct bpf_insn *patch = &insn_buf[0];
12392 bool issrc, isneg, isimm;
12395 aux = &env->insn_aux_data[i + delta];
12396 if (!aux->alu_state ||
12397 aux->alu_state == BPF_ALU_NON_POINTER)
12400 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
12401 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
12402 BPF_ALU_SANITIZE_SRC;
12403 isimm = aux->alu_state & BPF_ALU_IMMEDIATE;
12405 off_reg = issrc ? insn->src_reg : insn->dst_reg;
12407 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
12410 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
12411 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
12412 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
12413 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
12414 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
12415 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
12416 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX, off_reg);
12419 *patch++ = BPF_MOV64_REG(insn->dst_reg, insn->src_reg);
12420 insn->src_reg = BPF_REG_AX;
12422 insn->code = insn->code == code_add ?
12423 code_sub : code_add;
12425 if (issrc && isneg && !isimm)
12426 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
12427 cnt = patch - insn_buf;
12429 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12434 env->prog = prog = new_prog;
12435 insn = new_prog->insnsi + i + delta;
12439 if (insn->code != (BPF_JMP | BPF_CALL))
12441 if (insn->src_reg == BPF_PSEUDO_CALL)
12443 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
12444 ret = fixup_kfunc_call(env, insn);
12450 if (insn->imm == BPF_FUNC_get_route_realm)
12451 prog->dst_needed = 1;
12452 if (insn->imm == BPF_FUNC_get_prandom_u32)
12453 bpf_user_rnd_init_once();
12454 if (insn->imm == BPF_FUNC_override_return)
12455 prog->kprobe_override = 1;
12456 if (insn->imm == BPF_FUNC_tail_call) {
12457 /* If we tail call into other programs, we
12458 * cannot make any assumptions since they can
12459 * be replaced dynamically during runtime in
12460 * the program array.
12462 prog->cb_access = 1;
12463 if (!allow_tail_call_in_subprogs(env))
12464 prog->aux->stack_depth = MAX_BPF_STACK;
12465 prog->aux->max_pkt_offset = MAX_PACKET_OFF;
12467 /* mark bpf_tail_call as different opcode to avoid
12468 * conditional branch in the interpeter for every normal
12469 * call and to prevent accidental JITing by JIT compiler
12470 * that doesn't support bpf_tail_call yet
12473 insn->code = BPF_JMP | BPF_TAIL_CALL;
12475 aux = &env->insn_aux_data[i + delta];
12476 if (env->bpf_capable && !expect_blinding &&
12477 prog->jit_requested &&
12478 !bpf_map_key_poisoned(aux) &&
12479 !bpf_map_ptr_poisoned(aux) &&
12480 !bpf_map_ptr_unpriv(aux)) {
12481 struct bpf_jit_poke_descriptor desc = {
12482 .reason = BPF_POKE_REASON_TAIL_CALL,
12483 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
12484 .tail_call.key = bpf_map_key_immediate(aux),
12485 .insn_idx = i + delta,
12488 ret = bpf_jit_add_poke_descriptor(prog, &desc);
12490 verbose(env, "adding tail call poke descriptor failed\n");
12494 insn->imm = ret + 1;
12498 if (!bpf_map_ptr_unpriv(aux))
12501 /* instead of changing every JIT dealing with tail_call
12502 * emit two extra insns:
12503 * if (index >= max_entries) goto out;
12504 * index &= array->index_mask;
12505 * to avoid out-of-bounds cpu speculation
12507 if (bpf_map_ptr_poisoned(aux)) {
12508 verbose(env, "tail_call abusing map_ptr\n");
12512 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
12513 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
12514 map_ptr->max_entries, 2);
12515 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
12516 container_of(map_ptr,
12519 insn_buf[2] = *insn;
12521 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12526 env->prog = prog = new_prog;
12527 insn = new_prog->insnsi + i + delta;
12531 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
12532 * and other inlining handlers are currently limited to 64 bit
12535 if (prog->jit_requested && BITS_PER_LONG == 64 &&
12536 (insn->imm == BPF_FUNC_map_lookup_elem ||
12537 insn->imm == BPF_FUNC_map_update_elem ||
12538 insn->imm == BPF_FUNC_map_delete_elem ||
12539 insn->imm == BPF_FUNC_map_push_elem ||
12540 insn->imm == BPF_FUNC_map_pop_elem ||
12541 insn->imm == BPF_FUNC_map_peek_elem ||
12542 insn->imm == BPF_FUNC_redirect_map)) {
12543 aux = &env->insn_aux_data[i + delta];
12544 if (bpf_map_ptr_poisoned(aux))
12545 goto patch_call_imm;
12547 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
12548 ops = map_ptr->ops;
12549 if (insn->imm == BPF_FUNC_map_lookup_elem &&
12550 ops->map_gen_lookup) {
12551 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
12552 if (cnt == -EOPNOTSUPP)
12553 goto patch_map_ops_generic;
12554 if (cnt <= 0 || cnt >= ARRAY_SIZE(insn_buf)) {
12555 verbose(env, "bpf verifier is misconfigured\n");
12559 new_prog = bpf_patch_insn_data(env, i + delta,
12565 env->prog = prog = new_prog;
12566 insn = new_prog->insnsi + i + delta;
12570 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
12571 (void *(*)(struct bpf_map *map, void *key))NULL));
12572 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
12573 (int (*)(struct bpf_map *map, void *key))NULL));
12574 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
12575 (int (*)(struct bpf_map *map, void *key, void *value,
12577 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
12578 (int (*)(struct bpf_map *map, void *value,
12580 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
12581 (int (*)(struct bpf_map *map, void *value))NULL));
12582 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
12583 (int (*)(struct bpf_map *map, void *value))NULL));
12584 BUILD_BUG_ON(!__same_type(ops->map_redirect,
12585 (int (*)(struct bpf_map *map, u32 ifindex, u64 flags))NULL));
12587 patch_map_ops_generic:
12588 switch (insn->imm) {
12589 case BPF_FUNC_map_lookup_elem:
12590 insn->imm = BPF_CAST_CALL(ops->map_lookup_elem) -
12593 case BPF_FUNC_map_update_elem:
12594 insn->imm = BPF_CAST_CALL(ops->map_update_elem) -
12597 case BPF_FUNC_map_delete_elem:
12598 insn->imm = BPF_CAST_CALL(ops->map_delete_elem) -
12601 case BPF_FUNC_map_push_elem:
12602 insn->imm = BPF_CAST_CALL(ops->map_push_elem) -
12605 case BPF_FUNC_map_pop_elem:
12606 insn->imm = BPF_CAST_CALL(ops->map_pop_elem) -
12609 case BPF_FUNC_map_peek_elem:
12610 insn->imm = BPF_CAST_CALL(ops->map_peek_elem) -
12613 case BPF_FUNC_redirect_map:
12614 insn->imm = BPF_CAST_CALL(ops->map_redirect) -
12619 goto patch_call_imm;
12622 /* Implement bpf_jiffies64 inline. */
12623 if (prog->jit_requested && BITS_PER_LONG == 64 &&
12624 insn->imm == BPF_FUNC_jiffies64) {
12625 struct bpf_insn ld_jiffies_addr[2] = {
12626 BPF_LD_IMM64(BPF_REG_0,
12627 (unsigned long)&jiffies),
12630 insn_buf[0] = ld_jiffies_addr[0];
12631 insn_buf[1] = ld_jiffies_addr[1];
12632 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
12636 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
12642 env->prog = prog = new_prog;
12643 insn = new_prog->insnsi + i + delta;
12648 fn = env->ops->get_func_proto(insn->imm, env->prog);
12649 /* all functions that have prototype and verifier allowed
12650 * programs to call them, must be real in-kernel functions
12654 "kernel subsystem misconfigured func %s#%d\n",
12655 func_id_name(insn->imm), insn->imm);
12658 insn->imm = fn->func - __bpf_call_base;
12661 /* Since poke tab is now finalized, publish aux to tracker. */
12662 for (i = 0; i < prog->aux->size_poke_tab; i++) {
12663 map_ptr = prog->aux->poke_tab[i].tail_call.map;
12664 if (!map_ptr->ops->map_poke_track ||
12665 !map_ptr->ops->map_poke_untrack ||
12666 !map_ptr->ops->map_poke_run) {
12667 verbose(env, "bpf verifier is misconfigured\n");
12671 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
12673 verbose(env, "tracking tail call prog failed\n");
12678 sort_kfunc_descs_by_imm(env->prog);
12683 static void free_states(struct bpf_verifier_env *env)
12685 struct bpf_verifier_state_list *sl, *sln;
12688 sl = env->free_list;
12691 free_verifier_state(&sl->state, false);
12695 env->free_list = NULL;
12697 if (!env->explored_states)
12700 for (i = 0; i < state_htab_size(env); i++) {
12701 sl = env->explored_states[i];
12705 free_verifier_state(&sl->state, false);
12709 env->explored_states[i] = NULL;
12713 static int do_check_common(struct bpf_verifier_env *env, int subprog)
12715 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
12716 struct bpf_verifier_state *state;
12717 struct bpf_reg_state *regs;
12720 env->prev_linfo = NULL;
12723 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
12726 state->curframe = 0;
12727 state->speculative = false;
12728 state->branches = 1;
12729 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
12730 if (!state->frame[0]) {
12734 env->cur_state = state;
12735 init_func_state(env, state->frame[0],
12736 BPF_MAIN_FUNC /* callsite */,
12740 regs = state->frame[state->curframe]->regs;
12741 if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
12742 ret = btf_prepare_func_args(env, subprog, regs);
12745 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
12746 if (regs[i].type == PTR_TO_CTX)
12747 mark_reg_known_zero(env, regs, i);
12748 else if (regs[i].type == SCALAR_VALUE)
12749 mark_reg_unknown(env, regs, i);
12750 else if (regs[i].type == PTR_TO_MEM_OR_NULL) {
12751 const u32 mem_size = regs[i].mem_size;
12753 mark_reg_known_zero(env, regs, i);
12754 regs[i].mem_size = mem_size;
12755 regs[i].id = ++env->id_gen;
12759 /* 1st arg to a function */
12760 regs[BPF_REG_1].type = PTR_TO_CTX;
12761 mark_reg_known_zero(env, regs, BPF_REG_1);
12762 ret = btf_check_subprog_arg_match(env, subprog, regs);
12763 if (ret == -EFAULT)
12764 /* unlikely verifier bug. abort.
12765 * ret == 0 and ret < 0 are sadly acceptable for
12766 * main() function due to backward compatibility.
12767 * Like socket filter program may be written as:
12768 * int bpf_prog(struct pt_regs *ctx)
12769 * and never dereference that ctx in the program.
12770 * 'struct pt_regs' is a type mismatch for socket
12771 * filter that should be using 'struct __sk_buff'.
12776 ret = do_check(env);
12778 /* check for NULL is necessary, since cur_state can be freed inside
12779 * do_check() under memory pressure.
12781 if (env->cur_state) {
12782 free_verifier_state(env->cur_state, true);
12783 env->cur_state = NULL;
12785 while (!pop_stack(env, NULL, NULL, false));
12786 if (!ret && pop_log)
12787 bpf_vlog_reset(&env->log, 0);
12792 /* Verify all global functions in a BPF program one by one based on their BTF.
12793 * All global functions must pass verification. Otherwise the whole program is rejected.
12804 * foo() will be verified first for R1=any_scalar_value. During verification it
12805 * will be assumed that bar() already verified successfully and call to bar()
12806 * from foo() will be checked for type match only. Later bar() will be verified
12807 * independently to check that it's safe for R1=any_scalar_value.
12809 static int do_check_subprogs(struct bpf_verifier_env *env)
12811 struct bpf_prog_aux *aux = env->prog->aux;
12814 if (!aux->func_info)
12817 for (i = 1; i < env->subprog_cnt; i++) {
12818 if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL)
12820 env->insn_idx = env->subprog_info[i].start;
12821 WARN_ON_ONCE(env->insn_idx == 0);
12822 ret = do_check_common(env, i);
12825 } else if (env->log.level & BPF_LOG_LEVEL) {
12827 "Func#%d is safe for any args that match its prototype\n",
12834 static int do_check_main(struct bpf_verifier_env *env)
12839 ret = do_check_common(env, 0);
12841 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
12846 static void print_verification_stats(struct bpf_verifier_env *env)
12850 if (env->log.level & BPF_LOG_STATS) {
12851 verbose(env, "verification time %lld usec\n",
12852 div_u64(env->verification_time, 1000));
12853 verbose(env, "stack depth ");
12854 for (i = 0; i < env->subprog_cnt; i++) {
12855 u32 depth = env->subprog_info[i].stack_depth;
12857 verbose(env, "%d", depth);
12858 if (i + 1 < env->subprog_cnt)
12861 verbose(env, "\n");
12863 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
12864 "total_states %d peak_states %d mark_read %d\n",
12865 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
12866 env->max_states_per_insn, env->total_states,
12867 env->peak_states, env->longest_mark_read_walk);
12870 static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
12872 const struct btf_type *t, *func_proto;
12873 const struct bpf_struct_ops *st_ops;
12874 const struct btf_member *member;
12875 struct bpf_prog *prog = env->prog;
12876 u32 btf_id, member_idx;
12879 if (!prog->gpl_compatible) {
12880 verbose(env, "struct ops programs must have a GPL compatible license\n");
12884 btf_id = prog->aux->attach_btf_id;
12885 st_ops = bpf_struct_ops_find(btf_id);
12887 verbose(env, "attach_btf_id %u is not a supported struct\n",
12893 member_idx = prog->expected_attach_type;
12894 if (member_idx >= btf_type_vlen(t)) {
12895 verbose(env, "attach to invalid member idx %u of struct %s\n",
12896 member_idx, st_ops->name);
12900 member = &btf_type_member(t)[member_idx];
12901 mname = btf_name_by_offset(btf_vmlinux, member->name_off);
12902 func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type,
12905 verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
12906 mname, member_idx, st_ops->name);
12910 if (st_ops->check_member) {
12911 int err = st_ops->check_member(t, member);
12914 verbose(env, "attach to unsupported member %s of struct %s\n",
12915 mname, st_ops->name);
12920 prog->aux->attach_func_proto = func_proto;
12921 prog->aux->attach_func_name = mname;
12922 env->ops = st_ops->verifier_ops;
12926 #define SECURITY_PREFIX "security_"
12928 static int check_attach_modify_return(unsigned long addr, const char *func_name)
12930 if (within_error_injection_list(addr) ||
12931 !strncmp(SECURITY_PREFIX, func_name, sizeof(SECURITY_PREFIX) - 1))
12937 /* list of non-sleepable functions that are otherwise on
12938 * ALLOW_ERROR_INJECTION list
12940 BTF_SET_START(btf_non_sleepable_error_inject)
12941 /* Three functions below can be called from sleepable and non-sleepable context.
12942 * Assume non-sleepable from bpf safety point of view.
12944 BTF_ID(func, __add_to_page_cache_locked)
12945 BTF_ID(func, should_fail_alloc_page)
12946 BTF_ID(func, should_failslab)
12947 BTF_SET_END(btf_non_sleepable_error_inject)
12949 static int check_non_sleepable_error_inject(u32 btf_id)
12951 return btf_id_set_contains(&btf_non_sleepable_error_inject, btf_id);
12954 int bpf_check_attach_target(struct bpf_verifier_log *log,
12955 const struct bpf_prog *prog,
12956 const struct bpf_prog *tgt_prog,
12958 struct bpf_attach_target_info *tgt_info)
12960 bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
12961 const char prefix[] = "btf_trace_";
12962 int ret = 0, subprog = -1, i;
12963 const struct btf_type *t;
12964 bool conservative = true;
12970 bpf_log(log, "Tracing programs must provide btf_id\n");
12973 btf = tgt_prog ? tgt_prog->aux->btf : prog->aux->attach_btf;
12976 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
12979 t = btf_type_by_id(btf, btf_id);
12981 bpf_log(log, "attach_btf_id %u is invalid\n", btf_id);
12984 tname = btf_name_by_offset(btf, t->name_off);
12986 bpf_log(log, "attach_btf_id %u doesn't have a name\n", btf_id);
12990 struct bpf_prog_aux *aux = tgt_prog->aux;
12992 for (i = 0; i < aux->func_info_cnt; i++)
12993 if (aux->func_info[i].type_id == btf_id) {
12997 if (subprog == -1) {
12998 bpf_log(log, "Subprog %s doesn't exist\n", tname);
13001 conservative = aux->func_info_aux[subprog].unreliable;
13002 if (prog_extension) {
13003 if (conservative) {
13005 "Cannot replace static functions\n");
13008 if (!prog->jit_requested) {
13010 "Extension programs should be JITed\n");
13014 if (!tgt_prog->jited) {
13015 bpf_log(log, "Can attach to only JITed progs\n");
13018 if (tgt_prog->type == prog->type) {
13019 /* Cannot fentry/fexit another fentry/fexit program.
13020 * Cannot attach program extension to another extension.
13021 * It's ok to attach fentry/fexit to extension program.
13023 bpf_log(log, "Cannot recursively attach\n");
13026 if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
13028 (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
13029 tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
13030 /* Program extensions can extend all program types
13031 * except fentry/fexit. The reason is the following.
13032 * The fentry/fexit programs are used for performance
13033 * analysis, stats and can be attached to any program
13034 * type except themselves. When extension program is
13035 * replacing XDP function it is necessary to allow
13036 * performance analysis of all functions. Both original
13037 * XDP program and its program extension. Hence
13038 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
13039 * allowed. If extending of fentry/fexit was allowed it
13040 * would be possible to create long call chain
13041 * fentry->extension->fentry->extension beyond
13042 * reasonable stack size. Hence extending fentry is not
13045 bpf_log(log, "Cannot extend fentry/fexit\n");
13049 if (prog_extension) {
13050 bpf_log(log, "Cannot replace kernel functions\n");
13055 switch (prog->expected_attach_type) {
13056 case BPF_TRACE_RAW_TP:
13059 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
13062 if (!btf_type_is_typedef(t)) {
13063 bpf_log(log, "attach_btf_id %u is not a typedef\n",
13067 if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
13068 bpf_log(log, "attach_btf_id %u points to wrong type name %s\n",
13072 tname += sizeof(prefix) - 1;
13073 t = btf_type_by_id(btf, t->type);
13074 if (!btf_type_is_ptr(t))
13075 /* should never happen in valid vmlinux build */
13077 t = btf_type_by_id(btf, t->type);
13078 if (!btf_type_is_func_proto(t))
13079 /* should never happen in valid vmlinux build */
13083 case BPF_TRACE_ITER:
13084 if (!btf_type_is_func(t)) {
13085 bpf_log(log, "attach_btf_id %u is not a function\n",
13089 t = btf_type_by_id(btf, t->type);
13090 if (!btf_type_is_func_proto(t))
13092 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
13097 if (!prog_extension)
13100 case BPF_MODIFY_RETURN:
13102 case BPF_TRACE_FENTRY:
13103 case BPF_TRACE_FEXIT:
13104 if (!btf_type_is_func(t)) {
13105 bpf_log(log, "attach_btf_id %u is not a function\n",
13109 if (prog_extension &&
13110 btf_check_type_match(log, prog, btf, t))
13112 t = btf_type_by_id(btf, t->type);
13113 if (!btf_type_is_func_proto(t))
13116 if ((prog->aux->saved_dst_prog_type || prog->aux->saved_dst_attach_type) &&
13117 (!tgt_prog || prog->aux->saved_dst_prog_type != tgt_prog->type ||
13118 prog->aux->saved_dst_attach_type != tgt_prog->expected_attach_type))
13121 if (tgt_prog && conservative)
13124 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
13130 addr = (long) tgt_prog->bpf_func;
13132 addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
13134 addr = kallsyms_lookup_name(tname);
13137 "The address of function %s cannot be found\n",
13143 if (prog->aux->sleepable) {
13145 switch (prog->type) {
13146 case BPF_PROG_TYPE_TRACING:
13147 /* fentry/fexit/fmod_ret progs can be sleepable only if they are
13148 * attached to ALLOW_ERROR_INJECTION and are not in denylist.
13150 if (!check_non_sleepable_error_inject(btf_id) &&
13151 within_error_injection_list(addr))
13154 case BPF_PROG_TYPE_LSM:
13155 /* LSM progs check that they are attached to bpf_lsm_*() funcs.
13156 * Only some of them are sleepable.
13158 if (bpf_lsm_is_sleepable_hook(btf_id))
13165 bpf_log(log, "%s is not sleepable\n", tname);
13168 } else if (prog->expected_attach_type == BPF_MODIFY_RETURN) {
13170 bpf_log(log, "can't modify return codes of BPF programs\n");
13173 ret = check_attach_modify_return(addr, tname);
13175 bpf_log(log, "%s() is not modifiable\n", tname);
13182 tgt_info->tgt_addr = addr;
13183 tgt_info->tgt_name = tname;
13184 tgt_info->tgt_type = t;
13188 BTF_SET_START(btf_id_deny)
13191 BTF_ID(func, migrate_disable)
13192 BTF_ID(func, migrate_enable)
13194 #if !defined CONFIG_PREEMPT_RCU && !defined CONFIG_TINY_RCU
13195 BTF_ID(func, rcu_read_unlock_strict)
13197 BTF_SET_END(btf_id_deny)
13199 static int check_attach_btf_id(struct bpf_verifier_env *env)
13201 struct bpf_prog *prog = env->prog;
13202 struct bpf_prog *tgt_prog = prog->aux->dst_prog;
13203 struct bpf_attach_target_info tgt_info = {};
13204 u32 btf_id = prog->aux->attach_btf_id;
13205 struct bpf_trampoline *tr;
13209 if (prog->aux->sleepable && prog->type != BPF_PROG_TYPE_TRACING &&
13210 prog->type != BPF_PROG_TYPE_LSM) {
13211 verbose(env, "Only fentry/fexit/fmod_ret and lsm programs can be sleepable\n");
13215 if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
13216 return check_struct_ops_btf_id(env);
13218 if (prog->type != BPF_PROG_TYPE_TRACING &&
13219 prog->type != BPF_PROG_TYPE_LSM &&
13220 prog->type != BPF_PROG_TYPE_EXT)
13223 ret = bpf_check_attach_target(&env->log, prog, tgt_prog, btf_id, &tgt_info);
13227 if (tgt_prog && prog->type == BPF_PROG_TYPE_EXT) {
13228 /* to make freplace equivalent to their targets, they need to
13229 * inherit env->ops and expected_attach_type for the rest of the
13232 env->ops = bpf_verifier_ops[tgt_prog->type];
13233 prog->expected_attach_type = tgt_prog->expected_attach_type;
13236 /* store info about the attachment target that will be used later */
13237 prog->aux->attach_func_proto = tgt_info.tgt_type;
13238 prog->aux->attach_func_name = tgt_info.tgt_name;
13241 prog->aux->saved_dst_prog_type = tgt_prog->type;
13242 prog->aux->saved_dst_attach_type = tgt_prog->expected_attach_type;
13245 if (prog->expected_attach_type == BPF_TRACE_RAW_TP) {
13246 prog->aux->attach_btf_trace = true;
13248 } else if (prog->expected_attach_type == BPF_TRACE_ITER) {
13249 if (!bpf_iter_prog_supported(prog))
13254 if (prog->type == BPF_PROG_TYPE_LSM) {
13255 ret = bpf_lsm_verify_prog(&env->log, prog);
13258 } else if (prog->type == BPF_PROG_TYPE_TRACING &&
13259 btf_id_set_contains(&btf_id_deny, btf_id)) {
13263 key = bpf_trampoline_compute_key(tgt_prog, prog->aux->attach_btf, btf_id);
13264 tr = bpf_trampoline_get(key, &tgt_info);
13268 prog->aux->dst_trampoline = tr;
13272 struct btf *bpf_get_btf_vmlinux(void)
13274 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
13275 mutex_lock(&bpf_verifier_lock);
13277 btf_vmlinux = btf_parse_vmlinux();
13278 mutex_unlock(&bpf_verifier_lock);
13280 return btf_vmlinux;
13283 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr,
13284 union bpf_attr __user *uattr)
13286 u64 start_time = ktime_get_ns();
13287 struct bpf_verifier_env *env;
13288 struct bpf_verifier_log *log;
13289 int i, len, ret = -EINVAL;
13292 /* no program is valid */
13293 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
13296 /* 'struct bpf_verifier_env' can be global, but since it's not small,
13297 * allocate/free it every time bpf_check() is called
13299 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
13304 len = (*prog)->len;
13305 env->insn_aux_data =
13306 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
13308 if (!env->insn_aux_data)
13310 for (i = 0; i < len; i++)
13311 env->insn_aux_data[i].orig_idx = i;
13313 env->ops = bpf_verifier_ops[env->prog->type];
13314 is_priv = bpf_capable();
13316 bpf_get_btf_vmlinux();
13318 /* grab the mutex to protect few globals used by verifier */
13320 mutex_lock(&bpf_verifier_lock);
13322 if (attr->log_level || attr->log_buf || attr->log_size) {
13323 /* user requested verbose verifier output
13324 * and supplied buffer to store the verification trace
13326 log->level = attr->log_level;
13327 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
13328 log->len_total = attr->log_size;
13331 /* log attributes have to be sane */
13332 if (log->len_total < 128 || log->len_total > UINT_MAX >> 2 ||
13333 !log->level || !log->ubuf || log->level & ~BPF_LOG_MASK)
13337 if (IS_ERR(btf_vmlinux)) {
13338 /* Either gcc or pahole or kernel are broken. */
13339 verbose(env, "in-kernel BTF is malformed\n");
13340 ret = PTR_ERR(btf_vmlinux);
13341 goto skip_full_check;
13344 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
13345 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
13346 env->strict_alignment = true;
13347 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
13348 env->strict_alignment = false;
13350 env->allow_ptr_leaks = bpf_allow_ptr_leaks();
13351 env->allow_uninit_stack = bpf_allow_uninit_stack();
13352 env->allow_ptr_to_map_access = bpf_allow_ptr_to_map_access();
13353 env->bypass_spec_v1 = bpf_bypass_spec_v1();
13354 env->bypass_spec_v4 = bpf_bypass_spec_v4();
13355 env->bpf_capable = bpf_capable();
13358 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
13360 env->explored_states = kvcalloc(state_htab_size(env),
13361 sizeof(struct bpf_verifier_state_list *),
13364 if (!env->explored_states)
13365 goto skip_full_check;
13367 ret = add_subprog_and_kfunc(env);
13369 goto skip_full_check;
13371 ret = check_subprogs(env);
13373 goto skip_full_check;
13375 ret = check_btf_info(env, attr, uattr);
13377 goto skip_full_check;
13379 ret = check_attach_btf_id(env);
13381 goto skip_full_check;
13383 ret = resolve_pseudo_ldimm64(env);
13385 goto skip_full_check;
13387 if (bpf_prog_is_dev_bound(env->prog->aux)) {
13388 ret = bpf_prog_offload_verifier_prep(env->prog);
13390 goto skip_full_check;
13393 ret = check_cfg(env);
13395 goto skip_full_check;
13397 ret = do_check_subprogs(env);
13398 ret = ret ?: do_check_main(env);
13400 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
13401 ret = bpf_prog_offload_finalize(env);
13404 kvfree(env->explored_states);
13407 ret = check_max_stack_depth(env);
13409 /* instruction rewrites happen after this point */
13412 opt_hard_wire_dead_code_branches(env);
13414 ret = opt_remove_dead_code(env);
13416 ret = opt_remove_nops(env);
13419 sanitize_dead_code(env);
13423 /* program is valid, convert *(u32*)(ctx + off) accesses */
13424 ret = convert_ctx_accesses(env);
13427 ret = do_misc_fixups(env);
13429 /* do 32-bit optimization after insn patching has done so those patched
13430 * insns could be handled correctly.
13432 if (ret == 0 && !bpf_prog_is_dev_bound(env->prog->aux)) {
13433 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
13434 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
13439 ret = fixup_call_args(env);
13441 env->verification_time = ktime_get_ns() - start_time;
13442 print_verification_stats(env);
13444 if (log->level && bpf_verifier_log_full(log))
13446 if (log->level && !log->ubuf) {
13448 goto err_release_maps;
13452 goto err_release_maps;
13454 if (env->used_map_cnt) {
13455 /* if program passed verifier, update used_maps in bpf_prog_info */
13456 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
13457 sizeof(env->used_maps[0]),
13460 if (!env->prog->aux->used_maps) {
13462 goto err_release_maps;
13465 memcpy(env->prog->aux->used_maps, env->used_maps,
13466 sizeof(env->used_maps[0]) * env->used_map_cnt);
13467 env->prog->aux->used_map_cnt = env->used_map_cnt;
13469 if (env->used_btf_cnt) {
13470 /* if program passed verifier, update used_btfs in bpf_prog_aux */
13471 env->prog->aux->used_btfs = kmalloc_array(env->used_btf_cnt,
13472 sizeof(env->used_btfs[0]),
13474 if (!env->prog->aux->used_btfs) {
13476 goto err_release_maps;
13479 memcpy(env->prog->aux->used_btfs, env->used_btfs,
13480 sizeof(env->used_btfs[0]) * env->used_btf_cnt);
13481 env->prog->aux->used_btf_cnt = env->used_btf_cnt;
13483 if (env->used_map_cnt || env->used_btf_cnt) {
13484 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
13485 * bpf_ld_imm64 instructions
13487 convert_pseudo_ld_imm64(env);
13490 adjust_btf_func(env);
13493 if (!env->prog->aux->used_maps)
13494 /* if we didn't copy map pointers into bpf_prog_info, release
13495 * them now. Otherwise free_used_maps() will release them.
13498 if (!env->prog->aux->used_btfs)
13501 /* extension progs temporarily inherit the attach_type of their targets
13502 for verification purposes, so set it back to zero before returning
13504 if (env->prog->type == BPF_PROG_TYPE_EXT)
13505 env->prog->expected_attach_type = 0;
13510 mutex_unlock(&bpf_verifier_lock);
13511 vfree(env->insn_aux_data);