1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
3 * Copyright (c) 2016 Facebook
4 * Copyright (c) 2018 Covalent IO, Inc. http://covalent.io
6 #include <uapi/linux/btf.h>
7 #include <linux/kernel.h>
8 #include <linux/types.h>
9 #include <linux/slab.h>
10 #include <linux/bpf.h>
11 #include <linux/btf.h>
12 #include <linux/bpf_verifier.h>
13 #include <linux/filter.h>
14 #include <net/netlink.h>
15 #include <linux/file.h>
16 #include <linux/vmalloc.h>
17 #include <linux/stringify.h>
18 #include <linux/bsearch.h>
19 #include <linux/sort.h>
20 #include <linux/perf_event.h>
21 #include <linux/ctype.h>
22 #include <linux/error-injection.h>
23 #include <linux/bpf_lsm.h>
24 #include <linux/btf_ids.h>
28 static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
29 #define BPF_PROG_TYPE(_id, _name, prog_ctx_type, kern_ctx_type) \
30 [_id] = & _name ## _verifier_ops,
31 #define BPF_MAP_TYPE(_id, _ops)
32 #define BPF_LINK_TYPE(_id, _name)
33 #include <linux/bpf_types.h>
39 /* bpf_check() is a static code analyzer that walks eBPF program
40 * instruction by instruction and updates register/stack state.
41 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
43 * The first pass is depth-first-search to check that the program is a DAG.
44 * It rejects the following programs:
45 * - larger than BPF_MAXINSNS insns
46 * - if loop is present (detected via back-edge)
47 * - unreachable insns exist (shouldn't be a forest. program = one function)
48 * - out of bounds or malformed jumps
49 * The second pass is all possible path descent from the 1st insn.
50 * Since it's analyzing all paths through the program, the length of the
51 * analysis is limited to 64k insn, which may be hit even if total number of
52 * insn is less then 4K, but there are too many branches that change stack/regs.
53 * Number of 'branches to be analyzed' is limited to 1k
55 * On entry to each instruction, each register has a type, and the instruction
56 * changes the types of the registers depending on instruction semantics.
57 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
60 * All registers are 64-bit.
61 * R0 - return register
62 * R1-R5 argument passing registers
63 * R6-R9 callee saved registers
64 * R10 - frame pointer read-only
66 * At the start of BPF program the register R1 contains a pointer to bpf_context
67 * and has type PTR_TO_CTX.
69 * Verifier tracks arithmetic operations on pointers in case:
70 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
71 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
72 * 1st insn copies R10 (which has FRAME_PTR) type into R1
73 * and 2nd arithmetic instruction is pattern matched to recognize
74 * that it wants to construct a pointer to some element within stack.
75 * So after 2nd insn, the register R1 has type PTR_TO_STACK
76 * (and -20 constant is saved for further stack bounds checking).
77 * Meaning that this reg is a pointer to stack plus known immediate constant.
79 * Most of the time the registers have SCALAR_VALUE type, which
80 * means the register has some value, but it's not a valid pointer.
81 * (like pointer plus pointer becomes SCALAR_VALUE type)
83 * When verifier sees load or store instructions the type of base register
84 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK, PTR_TO_SOCKET. These are
85 * four pointer types recognized by check_mem_access() function.
87 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
88 * and the range of [ptr, ptr + map's value_size) is accessible.
90 * registers used to pass values to function calls are checked against
91 * function argument constraints.
93 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
94 * It means that the register type passed to this function must be
95 * PTR_TO_STACK and it will be used inside the function as
96 * 'pointer to map element key'
98 * For example the argument constraints for bpf_map_lookup_elem():
99 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
100 * .arg1_type = ARG_CONST_MAP_PTR,
101 * .arg2_type = ARG_PTR_TO_MAP_KEY,
103 * ret_type says that this function returns 'pointer to map elem value or null'
104 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
105 * 2nd argument should be a pointer to stack, which will be used inside
106 * the helper function as a pointer to map element key.
108 * On the kernel side the helper function looks like:
109 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
111 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
112 * void *key = (void *) (unsigned long) r2;
115 * here kernel can access 'key' and 'map' pointers safely, knowing that
116 * [key, key + map->key_size) bytes are valid and were initialized on
117 * the stack of eBPF program.
120 * Corresponding eBPF program may look like:
121 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
122 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
123 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
124 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
125 * here verifier looks at prototype of map_lookup_elem() and sees:
126 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
127 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
129 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
130 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
131 * and were initialized prior to this call.
132 * If it's ok, then verifier allows this BPF_CALL insn and looks at
133 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
134 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
135 * returns either pointer to map value or NULL.
137 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
138 * insn, the register holding that pointer in the true branch changes state to
139 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
140 * branch. See check_cond_jmp_op().
142 * After the call R0 is set to return type of the function and registers R1-R5
143 * are set to NOT_INIT to indicate that they are no longer readable.
145 * The following reference types represent a potential reference to a kernel
146 * resource which, after first being allocated, must be checked and freed by
148 * - PTR_TO_SOCKET_OR_NULL, PTR_TO_SOCKET
150 * When the verifier sees a helper call return a reference type, it allocates a
151 * pointer id for the reference and stores it in the current function state.
152 * Similar to the way that PTR_TO_MAP_VALUE_OR_NULL is converted into
153 * PTR_TO_MAP_VALUE, PTR_TO_SOCKET_OR_NULL becomes PTR_TO_SOCKET when the type
154 * passes through a NULL-check conditional. For the branch wherein the state is
155 * changed to CONST_IMM, the verifier releases the reference.
157 * For each helper function that allocates a reference, such as
158 * bpf_sk_lookup_tcp(), there is a corresponding release function, such as
159 * bpf_sk_release(). When a reference type passes into the release function,
160 * the verifier also releases the reference. If any unchecked or unreleased
161 * reference remains at the end of the program, the verifier rejects it.
164 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
165 struct bpf_verifier_stack_elem {
166 /* verifer state is 'st'
167 * before processing instruction 'insn_idx'
168 * and after processing instruction 'prev_insn_idx'
170 struct bpf_verifier_state st;
173 struct bpf_verifier_stack_elem *next;
174 /* length of verifier log at the time this state was pushed on stack */
178 #define BPF_COMPLEXITY_LIMIT_JMP_SEQ 8192
179 #define BPF_COMPLEXITY_LIMIT_STATES 64
181 #define BPF_MAP_KEY_POISON (1ULL << 63)
182 #define BPF_MAP_KEY_SEEN (1ULL << 62)
184 #define BPF_MAP_PTR_UNPRIV 1UL
185 #define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \
186 POISON_POINTER_DELTA))
187 #define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
189 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
191 return BPF_MAP_PTR(aux->map_ptr_state) == BPF_MAP_PTR_POISON;
194 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
196 return aux->map_ptr_state & BPF_MAP_PTR_UNPRIV;
199 static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
200 const struct bpf_map *map, bool unpriv)
202 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON & BPF_MAP_PTR_UNPRIV);
203 unpriv |= bpf_map_ptr_unpriv(aux);
204 aux->map_ptr_state = (unsigned long)map |
205 (unpriv ? BPF_MAP_PTR_UNPRIV : 0UL);
208 static bool bpf_map_key_poisoned(const struct bpf_insn_aux_data *aux)
210 return aux->map_key_state & BPF_MAP_KEY_POISON;
213 static bool bpf_map_key_unseen(const struct bpf_insn_aux_data *aux)
215 return !(aux->map_key_state & BPF_MAP_KEY_SEEN);
218 static u64 bpf_map_key_immediate(const struct bpf_insn_aux_data *aux)
220 return aux->map_key_state & ~(BPF_MAP_KEY_SEEN | BPF_MAP_KEY_POISON);
223 static void bpf_map_key_store(struct bpf_insn_aux_data *aux, u64 state)
225 bool poisoned = bpf_map_key_poisoned(aux);
227 aux->map_key_state = state | BPF_MAP_KEY_SEEN |
228 (poisoned ? BPF_MAP_KEY_POISON : 0ULL);
231 static bool bpf_pseudo_call(const struct bpf_insn *insn)
233 return insn->code == (BPF_JMP | BPF_CALL) &&
234 insn->src_reg == BPF_PSEUDO_CALL;
237 static bool bpf_pseudo_kfunc_call(const struct bpf_insn *insn)
239 return insn->code == (BPF_JMP | BPF_CALL) &&
240 insn->src_reg == BPF_PSEUDO_KFUNC_CALL;
243 static bool bpf_pseudo_func(const struct bpf_insn *insn)
245 return insn->code == (BPF_LD | BPF_IMM | BPF_DW) &&
246 insn->src_reg == BPF_PSEUDO_FUNC;
249 struct bpf_call_arg_meta {
250 struct bpf_map *map_ptr;
266 struct btf *btf_vmlinux;
268 static DEFINE_MUTEX(bpf_verifier_lock);
270 static const struct bpf_line_info *
271 find_linfo(const struct bpf_verifier_env *env, u32 insn_off)
273 const struct bpf_line_info *linfo;
274 const struct bpf_prog *prog;
278 nr_linfo = prog->aux->nr_linfo;
280 if (!nr_linfo || insn_off >= prog->len)
283 linfo = prog->aux->linfo;
284 for (i = 1; i < nr_linfo; i++)
285 if (insn_off < linfo[i].insn_off)
288 return &linfo[i - 1];
291 void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
296 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
298 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
299 "verifier log line truncated - local buffer too short\n");
301 n = min(log->len_total - log->len_used - 1, n);
304 if (log->level == BPF_LOG_KERNEL) {
305 pr_err("BPF:%s\n", log->kbuf);
308 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
314 static void bpf_vlog_reset(struct bpf_verifier_log *log, u32 new_pos)
318 if (!bpf_verifier_log_needed(log))
321 log->len_used = new_pos;
322 if (put_user(zero, log->ubuf + new_pos))
326 /* log_level controls verbosity level of eBPF verifier.
327 * bpf_verifier_log_write() is used to dump the verification trace to the log,
328 * so the user can figure out what's wrong with the program
330 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
331 const char *fmt, ...)
335 if (!bpf_verifier_log_needed(&env->log))
339 bpf_verifier_vlog(&env->log, fmt, args);
342 EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
344 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
346 struct bpf_verifier_env *env = private_data;
349 if (!bpf_verifier_log_needed(&env->log))
353 bpf_verifier_vlog(&env->log, fmt, args);
357 __printf(2, 3) void bpf_log(struct bpf_verifier_log *log,
358 const char *fmt, ...)
362 if (!bpf_verifier_log_needed(log))
366 bpf_verifier_vlog(log, fmt, args);
370 static const char *ltrim(const char *s)
378 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env,
380 const char *prefix_fmt, ...)
382 const struct bpf_line_info *linfo;
384 if (!bpf_verifier_log_needed(&env->log))
387 linfo = find_linfo(env, insn_off);
388 if (!linfo || linfo == env->prev_linfo)
394 va_start(args, prefix_fmt);
395 bpf_verifier_vlog(&env->log, prefix_fmt, args);
400 ltrim(btf_name_by_offset(env->prog->aux->btf,
403 env->prev_linfo = linfo;
406 static void verbose_invalid_scalar(struct bpf_verifier_env *env,
407 struct bpf_reg_state *reg,
408 struct tnum *range, const char *ctx,
409 const char *reg_name)
413 verbose(env, "At %s the register %s ", ctx, reg_name);
414 if (!tnum_is_unknown(reg->var_off)) {
415 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
416 verbose(env, "has value %s", tn_buf);
418 verbose(env, "has unknown scalar value");
420 tnum_strn(tn_buf, sizeof(tn_buf), *range);
421 verbose(env, " should have been in %s\n", tn_buf);
424 static bool type_is_pkt_pointer(enum bpf_reg_type type)
426 return type == PTR_TO_PACKET ||
427 type == PTR_TO_PACKET_META;
430 static bool type_is_sk_pointer(enum bpf_reg_type type)
432 return type == PTR_TO_SOCKET ||
433 type == PTR_TO_SOCK_COMMON ||
434 type == PTR_TO_TCP_SOCK ||
435 type == PTR_TO_XDP_SOCK;
438 static bool reg_type_not_null(enum bpf_reg_type type)
440 return type == PTR_TO_SOCKET ||
441 type == PTR_TO_TCP_SOCK ||
442 type == PTR_TO_MAP_VALUE ||
443 type == PTR_TO_MAP_KEY ||
444 type == PTR_TO_SOCK_COMMON;
447 static bool reg_type_may_be_null(enum bpf_reg_type type)
449 return type == PTR_TO_MAP_VALUE_OR_NULL ||
450 type == PTR_TO_SOCKET_OR_NULL ||
451 type == PTR_TO_SOCK_COMMON_OR_NULL ||
452 type == PTR_TO_TCP_SOCK_OR_NULL ||
453 type == PTR_TO_BTF_ID_OR_NULL ||
454 type == PTR_TO_MEM_OR_NULL ||
455 type == PTR_TO_RDONLY_BUF_OR_NULL ||
456 type == PTR_TO_RDWR_BUF_OR_NULL;
459 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
461 return reg->type == PTR_TO_MAP_VALUE &&
462 map_value_has_spin_lock(reg->map_ptr);
465 static bool reg_type_may_be_refcounted_or_null(enum bpf_reg_type type)
467 return type == PTR_TO_SOCKET ||
468 type == PTR_TO_SOCKET_OR_NULL ||
469 type == PTR_TO_TCP_SOCK ||
470 type == PTR_TO_TCP_SOCK_OR_NULL ||
471 type == PTR_TO_MEM ||
472 type == PTR_TO_MEM_OR_NULL;
475 static bool arg_type_may_be_refcounted(enum bpf_arg_type type)
477 return type == ARG_PTR_TO_SOCK_COMMON;
480 static bool arg_type_may_be_null(enum bpf_arg_type type)
482 return type == ARG_PTR_TO_MAP_VALUE_OR_NULL ||
483 type == ARG_PTR_TO_MEM_OR_NULL ||
484 type == ARG_PTR_TO_CTX_OR_NULL ||
485 type == ARG_PTR_TO_SOCKET_OR_NULL ||
486 type == ARG_PTR_TO_ALLOC_MEM_OR_NULL ||
487 type == ARG_PTR_TO_STACK_OR_NULL;
490 /* Determine whether the function releases some resources allocated by another
491 * function call. The first reference type argument will be assumed to be
492 * released by release_reference().
494 static bool is_release_function(enum bpf_func_id func_id)
496 return func_id == BPF_FUNC_sk_release ||
497 func_id == BPF_FUNC_ringbuf_submit ||
498 func_id == BPF_FUNC_ringbuf_discard;
501 static bool may_be_acquire_function(enum bpf_func_id func_id)
503 return func_id == BPF_FUNC_sk_lookup_tcp ||
504 func_id == BPF_FUNC_sk_lookup_udp ||
505 func_id == BPF_FUNC_skc_lookup_tcp ||
506 func_id == BPF_FUNC_map_lookup_elem ||
507 func_id == BPF_FUNC_ringbuf_reserve;
510 static bool is_acquire_function(enum bpf_func_id func_id,
511 const struct bpf_map *map)
513 enum bpf_map_type map_type = map ? map->map_type : BPF_MAP_TYPE_UNSPEC;
515 if (func_id == BPF_FUNC_sk_lookup_tcp ||
516 func_id == BPF_FUNC_sk_lookup_udp ||
517 func_id == BPF_FUNC_skc_lookup_tcp ||
518 func_id == BPF_FUNC_ringbuf_reserve)
521 if (func_id == BPF_FUNC_map_lookup_elem &&
522 (map_type == BPF_MAP_TYPE_SOCKMAP ||
523 map_type == BPF_MAP_TYPE_SOCKHASH))
529 static bool is_ptr_cast_function(enum bpf_func_id func_id)
531 return func_id == BPF_FUNC_tcp_sock ||
532 func_id == BPF_FUNC_sk_fullsock ||
533 func_id == BPF_FUNC_skc_to_tcp_sock ||
534 func_id == BPF_FUNC_skc_to_tcp6_sock ||
535 func_id == BPF_FUNC_skc_to_udp6_sock ||
536 func_id == BPF_FUNC_skc_to_tcp_timewait_sock ||
537 func_id == BPF_FUNC_skc_to_tcp_request_sock;
540 static bool is_cmpxchg_insn(const struct bpf_insn *insn)
542 return BPF_CLASS(insn->code) == BPF_STX &&
543 BPF_MODE(insn->code) == BPF_ATOMIC &&
544 insn->imm == BPF_CMPXCHG;
547 /* string representation of 'enum bpf_reg_type' */
548 static const char * const reg_type_str[] = {
550 [SCALAR_VALUE] = "inv",
551 [PTR_TO_CTX] = "ctx",
552 [CONST_PTR_TO_MAP] = "map_ptr",
553 [PTR_TO_MAP_VALUE] = "map_value",
554 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
555 [PTR_TO_STACK] = "fp",
556 [PTR_TO_PACKET] = "pkt",
557 [PTR_TO_PACKET_META] = "pkt_meta",
558 [PTR_TO_PACKET_END] = "pkt_end",
559 [PTR_TO_FLOW_KEYS] = "flow_keys",
560 [PTR_TO_SOCKET] = "sock",
561 [PTR_TO_SOCKET_OR_NULL] = "sock_or_null",
562 [PTR_TO_SOCK_COMMON] = "sock_common",
563 [PTR_TO_SOCK_COMMON_OR_NULL] = "sock_common_or_null",
564 [PTR_TO_TCP_SOCK] = "tcp_sock",
565 [PTR_TO_TCP_SOCK_OR_NULL] = "tcp_sock_or_null",
566 [PTR_TO_TP_BUFFER] = "tp_buffer",
567 [PTR_TO_XDP_SOCK] = "xdp_sock",
568 [PTR_TO_BTF_ID] = "ptr_",
569 [PTR_TO_BTF_ID_OR_NULL] = "ptr_or_null_",
570 [PTR_TO_PERCPU_BTF_ID] = "percpu_ptr_",
571 [PTR_TO_MEM] = "mem",
572 [PTR_TO_MEM_OR_NULL] = "mem_or_null",
573 [PTR_TO_RDONLY_BUF] = "rdonly_buf",
574 [PTR_TO_RDONLY_BUF_OR_NULL] = "rdonly_buf_or_null",
575 [PTR_TO_RDWR_BUF] = "rdwr_buf",
576 [PTR_TO_RDWR_BUF_OR_NULL] = "rdwr_buf_or_null",
577 [PTR_TO_FUNC] = "func",
578 [PTR_TO_MAP_KEY] = "map_key",
581 static char slot_type_char[] = {
582 [STACK_INVALID] = '?',
588 static void print_liveness(struct bpf_verifier_env *env,
589 enum bpf_reg_liveness live)
591 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
593 if (live & REG_LIVE_READ)
595 if (live & REG_LIVE_WRITTEN)
597 if (live & REG_LIVE_DONE)
601 static struct bpf_func_state *func(struct bpf_verifier_env *env,
602 const struct bpf_reg_state *reg)
604 struct bpf_verifier_state *cur = env->cur_state;
606 return cur->frame[reg->frameno];
609 static const char *kernel_type_name(const struct btf* btf, u32 id)
611 return btf_name_by_offset(btf, btf_type_by_id(btf, id)->name_off);
614 static void print_verifier_state(struct bpf_verifier_env *env,
615 const struct bpf_func_state *state)
617 const struct bpf_reg_state *reg;
622 verbose(env, " frame%d:", state->frameno);
623 for (i = 0; i < MAX_BPF_REG; i++) {
624 reg = &state->regs[i];
628 verbose(env, " R%d", i);
629 print_liveness(env, reg->live);
630 verbose(env, "=%s", reg_type_str[t]);
631 if (t == SCALAR_VALUE && reg->precise)
633 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
634 tnum_is_const(reg->var_off)) {
635 /* reg->off should be 0 for SCALAR_VALUE */
636 verbose(env, "%lld", reg->var_off.value + reg->off);
638 if (t == PTR_TO_BTF_ID ||
639 t == PTR_TO_BTF_ID_OR_NULL ||
640 t == PTR_TO_PERCPU_BTF_ID)
641 verbose(env, "%s", kernel_type_name(reg->btf, reg->btf_id));
642 verbose(env, "(id=%d", reg->id);
643 if (reg_type_may_be_refcounted_or_null(t))
644 verbose(env, ",ref_obj_id=%d", reg->ref_obj_id);
645 if (t != SCALAR_VALUE)
646 verbose(env, ",off=%d", reg->off);
647 if (type_is_pkt_pointer(t))
648 verbose(env, ",r=%d", reg->range);
649 else if (t == CONST_PTR_TO_MAP ||
650 t == PTR_TO_MAP_KEY ||
651 t == PTR_TO_MAP_VALUE ||
652 t == PTR_TO_MAP_VALUE_OR_NULL)
653 verbose(env, ",ks=%d,vs=%d",
654 reg->map_ptr->key_size,
655 reg->map_ptr->value_size);
656 if (tnum_is_const(reg->var_off)) {
657 /* Typically an immediate SCALAR_VALUE, but
658 * could be a pointer whose offset is too big
661 verbose(env, ",imm=%llx", reg->var_off.value);
663 if (reg->smin_value != reg->umin_value &&
664 reg->smin_value != S64_MIN)
665 verbose(env, ",smin_value=%lld",
666 (long long)reg->smin_value);
667 if (reg->smax_value != reg->umax_value &&
668 reg->smax_value != S64_MAX)
669 verbose(env, ",smax_value=%lld",
670 (long long)reg->smax_value);
671 if (reg->umin_value != 0)
672 verbose(env, ",umin_value=%llu",
673 (unsigned long long)reg->umin_value);
674 if (reg->umax_value != U64_MAX)
675 verbose(env, ",umax_value=%llu",
676 (unsigned long long)reg->umax_value);
677 if (!tnum_is_unknown(reg->var_off)) {
680 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
681 verbose(env, ",var_off=%s", tn_buf);
683 if (reg->s32_min_value != reg->smin_value &&
684 reg->s32_min_value != S32_MIN)
685 verbose(env, ",s32_min_value=%d",
686 (int)(reg->s32_min_value));
687 if (reg->s32_max_value != reg->smax_value &&
688 reg->s32_max_value != S32_MAX)
689 verbose(env, ",s32_max_value=%d",
690 (int)(reg->s32_max_value));
691 if (reg->u32_min_value != reg->umin_value &&
692 reg->u32_min_value != U32_MIN)
693 verbose(env, ",u32_min_value=%d",
694 (int)(reg->u32_min_value));
695 if (reg->u32_max_value != reg->umax_value &&
696 reg->u32_max_value != U32_MAX)
697 verbose(env, ",u32_max_value=%d",
698 (int)(reg->u32_max_value));
703 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
704 char types_buf[BPF_REG_SIZE + 1];
708 for (j = 0; j < BPF_REG_SIZE; j++) {
709 if (state->stack[i].slot_type[j] != STACK_INVALID)
711 types_buf[j] = slot_type_char[
712 state->stack[i].slot_type[j]];
714 types_buf[BPF_REG_SIZE] = 0;
717 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
718 print_liveness(env, state->stack[i].spilled_ptr.live);
719 if (state->stack[i].slot_type[0] == STACK_SPILL) {
720 reg = &state->stack[i].spilled_ptr;
722 verbose(env, "=%s", reg_type_str[t]);
723 if (t == SCALAR_VALUE && reg->precise)
725 if (t == SCALAR_VALUE && tnum_is_const(reg->var_off))
726 verbose(env, "%lld", reg->var_off.value + reg->off);
728 verbose(env, "=%s", types_buf);
731 if (state->acquired_refs && state->refs[0].id) {
732 verbose(env, " refs=%d", state->refs[0].id);
733 for (i = 1; i < state->acquired_refs; i++)
734 if (state->refs[i].id)
735 verbose(env, ",%d", state->refs[i].id);
740 /* copy array src of length n * size bytes to dst. dst is reallocated if it's too
741 * small to hold src. This is different from krealloc since we don't want to preserve
742 * the contents of dst.
744 * Leaves dst untouched if src is NULL or length is zero. Returns NULL if memory could
747 static void *copy_array(void *dst, const void *src, size_t n, size_t size, gfp_t flags)
751 if (ZERO_OR_NULL_PTR(src))
754 if (unlikely(check_mul_overflow(n, size, &bytes)))
757 if (ksize(dst) < bytes) {
759 dst = kmalloc_track_caller(bytes, flags);
764 memcpy(dst, src, bytes);
766 return dst ? dst : ZERO_SIZE_PTR;
769 /* resize an array from old_n items to new_n items. the array is reallocated if it's too
770 * small to hold new_n items. new items are zeroed out if the array grows.
772 * Contrary to krealloc_array, does not free arr if new_n is zero.
774 static void *realloc_array(void *arr, size_t old_n, size_t new_n, size_t size)
776 if (!new_n || old_n == new_n)
779 arr = krealloc_array(arr, new_n, size, GFP_KERNEL);
784 memset(arr + old_n * size, 0, (new_n - old_n) * size);
787 return arr ? arr : ZERO_SIZE_PTR;
790 static int copy_reference_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
792 dst->refs = copy_array(dst->refs, src->refs, src->acquired_refs,
793 sizeof(struct bpf_reference_state), GFP_KERNEL);
797 dst->acquired_refs = src->acquired_refs;
801 static int copy_stack_state(struct bpf_func_state *dst, const struct bpf_func_state *src)
803 size_t n = src->allocated_stack / BPF_REG_SIZE;
805 dst->stack = copy_array(dst->stack, src->stack, n, sizeof(struct bpf_stack_state),
810 dst->allocated_stack = src->allocated_stack;
814 static int resize_reference_state(struct bpf_func_state *state, size_t n)
816 state->refs = realloc_array(state->refs, state->acquired_refs, n,
817 sizeof(struct bpf_reference_state));
821 state->acquired_refs = n;
825 static int grow_stack_state(struct bpf_func_state *state, int size)
827 size_t old_n = state->allocated_stack / BPF_REG_SIZE, n = size / BPF_REG_SIZE;
832 state->stack = realloc_array(state->stack, old_n, n, sizeof(struct bpf_stack_state));
836 state->allocated_stack = size;
840 /* Acquire a pointer id from the env and update the state->refs to include
841 * this new pointer reference.
842 * On success, returns a valid pointer id to associate with the register
843 * On failure, returns a negative errno.
845 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
847 struct bpf_func_state *state = cur_func(env);
848 int new_ofs = state->acquired_refs;
851 err = resize_reference_state(state, state->acquired_refs + 1);
855 state->refs[new_ofs].id = id;
856 state->refs[new_ofs].insn_idx = insn_idx;
861 /* release function corresponding to acquire_reference_state(). Idempotent. */
862 static int release_reference_state(struct bpf_func_state *state, int ptr_id)
866 last_idx = state->acquired_refs - 1;
867 for (i = 0; i < state->acquired_refs; i++) {
868 if (state->refs[i].id == ptr_id) {
869 if (last_idx && i != last_idx)
870 memcpy(&state->refs[i], &state->refs[last_idx],
871 sizeof(*state->refs));
872 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
873 state->acquired_refs--;
880 static void free_func_state(struct bpf_func_state *state)
889 static void clear_jmp_history(struct bpf_verifier_state *state)
891 kfree(state->jmp_history);
892 state->jmp_history = NULL;
893 state->jmp_history_cnt = 0;
896 static void free_verifier_state(struct bpf_verifier_state *state,
901 for (i = 0; i <= state->curframe; i++) {
902 free_func_state(state->frame[i]);
903 state->frame[i] = NULL;
905 clear_jmp_history(state);
910 /* copy verifier state from src to dst growing dst stack space
911 * when necessary to accommodate larger src stack
913 static int copy_func_state(struct bpf_func_state *dst,
914 const struct bpf_func_state *src)
918 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
919 err = copy_reference_state(dst, src);
922 return copy_stack_state(dst, src);
925 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
926 const struct bpf_verifier_state *src)
928 struct bpf_func_state *dst;
931 dst_state->jmp_history = copy_array(dst_state->jmp_history, src->jmp_history,
932 src->jmp_history_cnt, sizeof(struct bpf_idx_pair),
934 if (!dst_state->jmp_history)
936 dst_state->jmp_history_cnt = src->jmp_history_cnt;
938 /* if dst has more stack frames then src frame, free them */
939 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
940 free_func_state(dst_state->frame[i]);
941 dst_state->frame[i] = NULL;
943 dst_state->speculative = src->speculative;
944 dst_state->curframe = src->curframe;
945 dst_state->active_spin_lock = src->active_spin_lock;
946 dst_state->branches = src->branches;
947 dst_state->parent = src->parent;
948 dst_state->first_insn_idx = src->first_insn_idx;
949 dst_state->last_insn_idx = src->last_insn_idx;
950 for (i = 0; i <= src->curframe; i++) {
951 dst = dst_state->frame[i];
953 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
956 dst_state->frame[i] = dst;
958 err = copy_func_state(dst, src->frame[i]);
965 static void update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
968 u32 br = --st->branches;
970 /* WARN_ON(br > 1) technically makes sense here,
971 * but see comment in push_stack(), hence:
973 WARN_ONCE((int)br < 0,
974 "BUG update_branch_counts:branches_to_explore=%d\n",
982 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
983 int *insn_idx, bool pop_log)
985 struct bpf_verifier_state *cur = env->cur_state;
986 struct bpf_verifier_stack_elem *elem, *head = env->head;
989 if (env->head == NULL)
993 err = copy_verifier_state(cur, &head->st);
998 bpf_vlog_reset(&env->log, head->log_pos);
1000 *insn_idx = head->insn_idx;
1002 *prev_insn_idx = head->prev_insn_idx;
1004 free_verifier_state(&head->st, false);
1011 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
1012 int insn_idx, int prev_insn_idx,
1015 struct bpf_verifier_state *cur = env->cur_state;
1016 struct bpf_verifier_stack_elem *elem;
1019 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
1023 elem->insn_idx = insn_idx;
1024 elem->prev_insn_idx = prev_insn_idx;
1025 elem->next = env->head;
1026 elem->log_pos = env->log.len_used;
1029 err = copy_verifier_state(&elem->st, cur);
1032 elem->st.speculative |= speculative;
1033 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
1034 verbose(env, "The sequence of %d jumps is too complex.\n",
1038 if (elem->st.parent) {
1039 ++elem->st.parent->branches;
1040 /* WARN_ON(branches > 2) technically makes sense here,
1042 * 1. speculative states will bump 'branches' for non-branch
1044 * 2. is_state_visited() heuristics may decide not to create
1045 * a new state for a sequence of branches and all such current
1046 * and cloned states will be pointing to a single parent state
1047 * which might have large 'branches' count.
1052 free_verifier_state(env->cur_state, true);
1053 env->cur_state = NULL;
1054 /* pop all elements and return */
1055 while (!pop_stack(env, NULL, NULL, false));
1059 #define CALLER_SAVED_REGS 6
1060 static const int caller_saved[CALLER_SAVED_REGS] = {
1061 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
1064 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1065 struct bpf_reg_state *reg);
1067 /* This helper doesn't clear reg->id */
1068 static void ___mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1070 reg->var_off = tnum_const(imm);
1071 reg->smin_value = (s64)imm;
1072 reg->smax_value = (s64)imm;
1073 reg->umin_value = imm;
1074 reg->umax_value = imm;
1076 reg->s32_min_value = (s32)imm;
1077 reg->s32_max_value = (s32)imm;
1078 reg->u32_min_value = (u32)imm;
1079 reg->u32_max_value = (u32)imm;
1082 /* Mark the unknown part of a register (variable offset or scalar value) as
1083 * known to have the value @imm.
1085 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
1087 /* Clear id, off, and union(map_ptr, range) */
1088 memset(((u8 *)reg) + sizeof(reg->type), 0,
1089 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
1090 ___mark_reg_known(reg, imm);
1093 static void __mark_reg32_known(struct bpf_reg_state *reg, u64 imm)
1095 reg->var_off = tnum_const_subreg(reg->var_off, imm);
1096 reg->s32_min_value = (s32)imm;
1097 reg->s32_max_value = (s32)imm;
1098 reg->u32_min_value = (u32)imm;
1099 reg->u32_max_value = (u32)imm;
1102 /* Mark the 'variable offset' part of a register as zero. This should be
1103 * used only on registers holding a pointer type.
1105 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
1107 __mark_reg_known(reg, 0);
1110 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
1112 __mark_reg_known(reg, 0);
1113 reg->type = SCALAR_VALUE;
1116 static void mark_reg_known_zero(struct bpf_verifier_env *env,
1117 struct bpf_reg_state *regs, u32 regno)
1119 if (WARN_ON(regno >= MAX_BPF_REG)) {
1120 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
1121 /* Something bad happened, let's kill all regs */
1122 for (regno = 0; regno < MAX_BPF_REG; regno++)
1123 __mark_reg_not_init(env, regs + regno);
1126 __mark_reg_known_zero(regs + regno);
1129 static void mark_ptr_not_null_reg(struct bpf_reg_state *reg)
1131 switch (reg->type) {
1132 case PTR_TO_MAP_VALUE_OR_NULL: {
1133 const struct bpf_map *map = reg->map_ptr;
1135 if (map->inner_map_meta) {
1136 reg->type = CONST_PTR_TO_MAP;
1137 reg->map_ptr = map->inner_map_meta;
1138 } else if (map->map_type == BPF_MAP_TYPE_XSKMAP) {
1139 reg->type = PTR_TO_XDP_SOCK;
1140 } else if (map->map_type == BPF_MAP_TYPE_SOCKMAP ||
1141 map->map_type == BPF_MAP_TYPE_SOCKHASH) {
1142 reg->type = PTR_TO_SOCKET;
1144 reg->type = PTR_TO_MAP_VALUE;
1148 case PTR_TO_SOCKET_OR_NULL:
1149 reg->type = PTR_TO_SOCKET;
1151 case PTR_TO_SOCK_COMMON_OR_NULL:
1152 reg->type = PTR_TO_SOCK_COMMON;
1154 case PTR_TO_TCP_SOCK_OR_NULL:
1155 reg->type = PTR_TO_TCP_SOCK;
1157 case PTR_TO_BTF_ID_OR_NULL:
1158 reg->type = PTR_TO_BTF_ID;
1160 case PTR_TO_MEM_OR_NULL:
1161 reg->type = PTR_TO_MEM;
1163 case PTR_TO_RDONLY_BUF_OR_NULL:
1164 reg->type = PTR_TO_RDONLY_BUF;
1166 case PTR_TO_RDWR_BUF_OR_NULL:
1167 reg->type = PTR_TO_RDWR_BUF;
1170 WARN_ONCE(1, "unknown nullable register type");
1174 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
1176 return type_is_pkt_pointer(reg->type);
1179 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
1181 return reg_is_pkt_pointer(reg) ||
1182 reg->type == PTR_TO_PACKET_END;
1185 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
1186 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
1187 enum bpf_reg_type which)
1189 /* The register can already have a range from prior markings.
1190 * This is fine as long as it hasn't been advanced from its
1193 return reg->type == which &&
1196 tnum_equals_const(reg->var_off, 0);
1199 /* Reset the min/max bounds of a register */
1200 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
1202 reg->smin_value = S64_MIN;
1203 reg->smax_value = S64_MAX;
1204 reg->umin_value = 0;
1205 reg->umax_value = U64_MAX;
1207 reg->s32_min_value = S32_MIN;
1208 reg->s32_max_value = S32_MAX;
1209 reg->u32_min_value = 0;
1210 reg->u32_max_value = U32_MAX;
1213 static void __mark_reg64_unbounded(struct bpf_reg_state *reg)
1215 reg->smin_value = S64_MIN;
1216 reg->smax_value = S64_MAX;
1217 reg->umin_value = 0;
1218 reg->umax_value = U64_MAX;
1221 static void __mark_reg32_unbounded(struct bpf_reg_state *reg)
1223 reg->s32_min_value = S32_MIN;
1224 reg->s32_max_value = S32_MAX;
1225 reg->u32_min_value = 0;
1226 reg->u32_max_value = U32_MAX;
1229 static void __update_reg32_bounds(struct bpf_reg_state *reg)
1231 struct tnum var32_off = tnum_subreg(reg->var_off);
1233 /* min signed is max(sign bit) | min(other bits) */
1234 reg->s32_min_value = max_t(s32, reg->s32_min_value,
1235 var32_off.value | (var32_off.mask & S32_MIN));
1236 /* max signed is min(sign bit) | max(other bits) */
1237 reg->s32_max_value = min_t(s32, reg->s32_max_value,
1238 var32_off.value | (var32_off.mask & S32_MAX));
1239 reg->u32_min_value = max_t(u32, reg->u32_min_value, (u32)var32_off.value);
1240 reg->u32_max_value = min(reg->u32_max_value,
1241 (u32)(var32_off.value | var32_off.mask));
1244 static void __update_reg64_bounds(struct bpf_reg_state *reg)
1246 /* min signed is max(sign bit) | min(other bits) */
1247 reg->smin_value = max_t(s64, reg->smin_value,
1248 reg->var_off.value | (reg->var_off.mask & S64_MIN));
1249 /* max signed is min(sign bit) | max(other bits) */
1250 reg->smax_value = min_t(s64, reg->smax_value,
1251 reg->var_off.value | (reg->var_off.mask & S64_MAX));
1252 reg->umin_value = max(reg->umin_value, reg->var_off.value);
1253 reg->umax_value = min(reg->umax_value,
1254 reg->var_off.value | reg->var_off.mask);
1257 static void __update_reg_bounds(struct bpf_reg_state *reg)
1259 __update_reg32_bounds(reg);
1260 __update_reg64_bounds(reg);
1263 /* Uses signed min/max values to inform unsigned, and vice-versa */
1264 static void __reg32_deduce_bounds(struct bpf_reg_state *reg)
1266 /* Learn sign from signed bounds.
1267 * If we cannot cross the sign boundary, then signed and unsigned bounds
1268 * are the same, so combine. This works even in the negative case, e.g.
1269 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1271 if (reg->s32_min_value >= 0 || reg->s32_max_value < 0) {
1272 reg->s32_min_value = reg->u32_min_value =
1273 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1274 reg->s32_max_value = reg->u32_max_value =
1275 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1278 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1279 * boundary, so we must be careful.
1281 if ((s32)reg->u32_max_value >= 0) {
1282 /* Positive. We can't learn anything from the smin, but smax
1283 * is positive, hence safe.
1285 reg->s32_min_value = reg->u32_min_value;
1286 reg->s32_max_value = reg->u32_max_value =
1287 min_t(u32, reg->s32_max_value, reg->u32_max_value);
1288 } else if ((s32)reg->u32_min_value < 0) {
1289 /* Negative. We can't learn anything from the smax, but smin
1290 * is negative, hence safe.
1292 reg->s32_min_value = reg->u32_min_value =
1293 max_t(u32, reg->s32_min_value, reg->u32_min_value);
1294 reg->s32_max_value = reg->u32_max_value;
1298 static void __reg64_deduce_bounds(struct bpf_reg_state *reg)
1300 /* Learn sign from signed bounds.
1301 * If we cannot cross the sign boundary, then signed and unsigned bounds
1302 * are the same, so combine. This works even in the negative case, e.g.
1303 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1305 if (reg->smin_value >= 0 || reg->smax_value < 0) {
1306 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1308 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1312 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1313 * boundary, so we must be careful.
1315 if ((s64)reg->umax_value >= 0) {
1316 /* Positive. We can't learn anything from the smin, but smax
1317 * is positive, hence safe.
1319 reg->smin_value = reg->umin_value;
1320 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1322 } else if ((s64)reg->umin_value < 0) {
1323 /* Negative. We can't learn anything from the smax, but smin
1324 * is negative, hence safe.
1326 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1328 reg->smax_value = reg->umax_value;
1332 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
1334 __reg32_deduce_bounds(reg);
1335 __reg64_deduce_bounds(reg);
1338 /* Attempts to improve var_off based on unsigned min/max information */
1339 static void __reg_bound_offset(struct bpf_reg_state *reg)
1341 struct tnum var64_off = tnum_intersect(reg->var_off,
1342 tnum_range(reg->umin_value,
1344 struct tnum var32_off = tnum_intersect(tnum_subreg(reg->var_off),
1345 tnum_range(reg->u32_min_value,
1346 reg->u32_max_value));
1348 reg->var_off = tnum_or(tnum_clear_subreg(var64_off), var32_off);
1351 static void __reg_assign_32_into_64(struct bpf_reg_state *reg)
1353 reg->umin_value = reg->u32_min_value;
1354 reg->umax_value = reg->u32_max_value;
1355 /* Attempt to pull 32-bit signed bounds into 64-bit bounds
1356 * but must be positive otherwise set to worse case bounds
1357 * and refine later from tnum.
1359 if (reg->s32_min_value >= 0 && reg->s32_max_value >= 0)
1360 reg->smax_value = reg->s32_max_value;
1362 reg->smax_value = U32_MAX;
1363 if (reg->s32_min_value >= 0)
1364 reg->smin_value = reg->s32_min_value;
1366 reg->smin_value = 0;
1369 static void __reg_combine_32_into_64(struct bpf_reg_state *reg)
1371 /* special case when 64-bit register has upper 32-bit register
1372 * zeroed. Typically happens after zext or <<32, >>32 sequence
1373 * allowing us to use 32-bit bounds directly,
1375 if (tnum_equals_const(tnum_clear_subreg(reg->var_off), 0)) {
1376 __reg_assign_32_into_64(reg);
1378 /* Otherwise the best we can do is push lower 32bit known and
1379 * unknown bits into register (var_off set from jmp logic)
1380 * then learn as much as possible from the 64-bit tnum
1381 * known and unknown bits. The previous smin/smax bounds are
1382 * invalid here because of jmp32 compare so mark them unknown
1383 * so they do not impact tnum bounds calculation.
1385 __mark_reg64_unbounded(reg);
1386 __update_reg_bounds(reg);
1389 /* Intersecting with the old var_off might have improved our bounds
1390 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1391 * then new var_off is (0; 0x7f...fc) which improves our umax.
1393 __reg_deduce_bounds(reg);
1394 __reg_bound_offset(reg);
1395 __update_reg_bounds(reg);
1398 static bool __reg64_bound_s32(s64 a)
1400 return a > S32_MIN && a < S32_MAX;
1403 static bool __reg64_bound_u32(u64 a)
1405 return a > U32_MIN && a < U32_MAX;
1408 static void __reg_combine_64_into_32(struct bpf_reg_state *reg)
1410 __mark_reg32_unbounded(reg);
1412 if (__reg64_bound_s32(reg->smin_value) && __reg64_bound_s32(reg->smax_value)) {
1413 reg->s32_min_value = (s32)reg->smin_value;
1414 reg->s32_max_value = (s32)reg->smax_value;
1416 if (__reg64_bound_u32(reg->umin_value) && __reg64_bound_u32(reg->umax_value)) {
1417 reg->u32_min_value = (u32)reg->umin_value;
1418 reg->u32_max_value = (u32)reg->umax_value;
1421 /* Intersecting with the old var_off might have improved our bounds
1422 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
1423 * then new var_off is (0; 0x7f...fc) which improves our umax.
1425 __reg_deduce_bounds(reg);
1426 __reg_bound_offset(reg);
1427 __update_reg_bounds(reg);
1430 /* Mark a register as having a completely unknown (scalar) value. */
1431 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
1432 struct bpf_reg_state *reg)
1435 * Clear type, id, off, and union(map_ptr, range) and
1436 * padding between 'type' and union
1438 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
1439 reg->type = SCALAR_VALUE;
1440 reg->var_off = tnum_unknown;
1442 reg->precise = env->subprog_cnt > 1 || !env->bpf_capable;
1443 __mark_reg_unbounded(reg);
1446 static void mark_reg_unknown(struct bpf_verifier_env *env,
1447 struct bpf_reg_state *regs, u32 regno)
1449 if (WARN_ON(regno >= MAX_BPF_REG)) {
1450 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
1451 /* Something bad happened, let's kill all regs except FP */
1452 for (regno = 0; regno < BPF_REG_FP; regno++)
1453 __mark_reg_not_init(env, regs + regno);
1456 __mark_reg_unknown(env, regs + regno);
1459 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1460 struct bpf_reg_state *reg)
1462 __mark_reg_unknown(env, reg);
1463 reg->type = NOT_INIT;
1466 static void mark_reg_not_init(struct bpf_verifier_env *env,
1467 struct bpf_reg_state *regs, u32 regno)
1469 if (WARN_ON(regno >= MAX_BPF_REG)) {
1470 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
1471 /* Something bad happened, let's kill all regs except FP */
1472 for (regno = 0; regno < BPF_REG_FP; regno++)
1473 __mark_reg_not_init(env, regs + regno);
1476 __mark_reg_not_init(env, regs + regno);
1479 static void mark_btf_ld_reg(struct bpf_verifier_env *env,
1480 struct bpf_reg_state *regs, u32 regno,
1481 enum bpf_reg_type reg_type,
1482 struct btf *btf, u32 btf_id)
1484 if (reg_type == SCALAR_VALUE) {
1485 mark_reg_unknown(env, regs, regno);
1488 mark_reg_known_zero(env, regs, regno);
1489 regs[regno].type = PTR_TO_BTF_ID;
1490 regs[regno].btf = btf;
1491 regs[regno].btf_id = btf_id;
1494 #define DEF_NOT_SUBREG (0)
1495 static void init_reg_state(struct bpf_verifier_env *env,
1496 struct bpf_func_state *state)
1498 struct bpf_reg_state *regs = state->regs;
1501 for (i = 0; i < MAX_BPF_REG; i++) {
1502 mark_reg_not_init(env, regs, i);
1503 regs[i].live = REG_LIVE_NONE;
1504 regs[i].parent = NULL;
1505 regs[i].subreg_def = DEF_NOT_SUBREG;
1509 regs[BPF_REG_FP].type = PTR_TO_STACK;
1510 mark_reg_known_zero(env, regs, BPF_REG_FP);
1511 regs[BPF_REG_FP].frameno = state->frameno;
1514 #define BPF_MAIN_FUNC (-1)
1515 static void init_func_state(struct bpf_verifier_env *env,
1516 struct bpf_func_state *state,
1517 int callsite, int frameno, int subprogno)
1519 state->callsite = callsite;
1520 state->frameno = frameno;
1521 state->subprogno = subprogno;
1522 init_reg_state(env, state);
1526 SRC_OP, /* register is used as source operand */
1527 DST_OP, /* register is used as destination operand */
1528 DST_OP_NO_MARK /* same as above, check only, don't mark */
1531 static int cmp_subprogs(const void *a, const void *b)
1533 return ((struct bpf_subprog_info *)a)->start -
1534 ((struct bpf_subprog_info *)b)->start;
1537 static int find_subprog(struct bpf_verifier_env *env, int off)
1539 struct bpf_subprog_info *p;
1541 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
1542 sizeof(env->subprog_info[0]), cmp_subprogs);
1545 return p - env->subprog_info;
1549 static int add_subprog(struct bpf_verifier_env *env, int off)
1551 int insn_cnt = env->prog->len;
1554 if (off >= insn_cnt || off < 0) {
1555 verbose(env, "call to invalid destination\n");
1558 ret = find_subprog(env, off);
1561 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
1562 verbose(env, "too many subprograms\n");
1565 /* determine subprog starts. The end is one before the next starts */
1566 env->subprog_info[env->subprog_cnt++].start = off;
1567 sort(env->subprog_info, env->subprog_cnt,
1568 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
1569 return env->subprog_cnt - 1;
1572 struct bpf_kfunc_desc {
1573 struct btf_func_model func_model;
1578 #define MAX_KFUNC_DESCS 256
1579 struct bpf_kfunc_desc_tab {
1580 struct bpf_kfunc_desc descs[MAX_KFUNC_DESCS];
1584 static int kfunc_desc_cmp_by_id(const void *a, const void *b)
1586 const struct bpf_kfunc_desc *d0 = a;
1587 const struct bpf_kfunc_desc *d1 = b;
1589 /* func_id is not greater than BTF_MAX_TYPE */
1590 return d0->func_id - d1->func_id;
1593 static const struct bpf_kfunc_desc *
1594 find_kfunc_desc(const struct bpf_prog *prog, u32 func_id)
1596 struct bpf_kfunc_desc desc = {
1599 struct bpf_kfunc_desc_tab *tab;
1601 tab = prog->aux->kfunc_tab;
1602 return bsearch(&desc, tab->descs, tab->nr_descs,
1603 sizeof(tab->descs[0]), kfunc_desc_cmp_by_id);
1606 static int add_kfunc_call(struct bpf_verifier_env *env, u32 func_id)
1608 const struct btf_type *func, *func_proto;
1609 struct bpf_kfunc_desc_tab *tab;
1610 struct bpf_prog_aux *prog_aux;
1611 struct bpf_kfunc_desc *desc;
1612 const char *func_name;
1616 prog_aux = env->prog->aux;
1617 tab = prog_aux->kfunc_tab;
1620 verbose(env, "calling kernel function is not supported without CONFIG_DEBUG_INFO_BTF\n");
1624 if (!env->prog->jit_requested) {
1625 verbose(env, "JIT is required for calling kernel function\n");
1629 if (!bpf_jit_supports_kfunc_call()) {
1630 verbose(env, "JIT does not support calling kernel function\n");
1634 if (!env->prog->gpl_compatible) {
1635 verbose(env, "cannot call kernel function from non-GPL compatible program\n");
1639 tab = kzalloc(sizeof(*tab), GFP_KERNEL);
1642 prog_aux->kfunc_tab = tab;
1645 if (find_kfunc_desc(env->prog, func_id))
1648 if (tab->nr_descs == MAX_KFUNC_DESCS) {
1649 verbose(env, "too many different kernel function calls\n");
1653 func = btf_type_by_id(btf_vmlinux, func_id);
1654 if (!func || !btf_type_is_func(func)) {
1655 verbose(env, "kernel btf_id %u is not a function\n",
1659 func_proto = btf_type_by_id(btf_vmlinux, func->type);
1660 if (!func_proto || !btf_type_is_func_proto(func_proto)) {
1661 verbose(env, "kernel function btf_id %u does not have a valid func_proto\n",
1666 func_name = btf_name_by_offset(btf_vmlinux, func->name_off);
1667 addr = kallsyms_lookup_name(func_name);
1669 verbose(env, "cannot find address for kernel function %s\n",
1674 desc = &tab->descs[tab->nr_descs++];
1675 desc->func_id = func_id;
1676 desc->imm = BPF_CAST_CALL(addr) - __bpf_call_base;
1677 err = btf_distill_func_proto(&env->log, btf_vmlinux,
1678 func_proto, func_name,
1681 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
1682 kfunc_desc_cmp_by_id, NULL);
1686 static int kfunc_desc_cmp_by_imm(const void *a, const void *b)
1688 const struct bpf_kfunc_desc *d0 = a;
1689 const struct bpf_kfunc_desc *d1 = b;
1691 if (d0->imm > d1->imm)
1693 else if (d0->imm < d1->imm)
1698 static void sort_kfunc_descs_by_imm(struct bpf_prog *prog)
1700 struct bpf_kfunc_desc_tab *tab;
1702 tab = prog->aux->kfunc_tab;
1706 sort(tab->descs, tab->nr_descs, sizeof(tab->descs[0]),
1707 kfunc_desc_cmp_by_imm, NULL);
1710 bool bpf_prog_has_kfunc_call(const struct bpf_prog *prog)
1712 return !!prog->aux->kfunc_tab;
1715 const struct btf_func_model *
1716 bpf_jit_find_kfunc_model(const struct bpf_prog *prog,
1717 const struct bpf_insn *insn)
1719 const struct bpf_kfunc_desc desc = {
1722 const struct bpf_kfunc_desc *res;
1723 struct bpf_kfunc_desc_tab *tab;
1725 tab = prog->aux->kfunc_tab;
1726 res = bsearch(&desc, tab->descs, tab->nr_descs,
1727 sizeof(tab->descs[0]), kfunc_desc_cmp_by_imm);
1729 return res ? &res->func_model : NULL;
1732 static int add_subprog_and_kfunc(struct bpf_verifier_env *env)
1734 struct bpf_subprog_info *subprog = env->subprog_info;
1735 struct bpf_insn *insn = env->prog->insnsi;
1736 int i, ret, insn_cnt = env->prog->len;
1738 /* Add entry function. */
1739 ret = add_subprog(env, 0);
1743 for (i = 0; i < insn_cnt; i++, insn++) {
1744 if (!bpf_pseudo_func(insn) && !bpf_pseudo_call(insn) &&
1745 !bpf_pseudo_kfunc_call(insn))
1748 if (!env->bpf_capable) {
1749 verbose(env, "loading/calling other bpf or kernel functions are allowed for CAP_BPF and CAP_SYS_ADMIN\n");
1753 if (bpf_pseudo_func(insn)) {
1754 ret = add_subprog(env, i + insn->imm + 1);
1756 /* remember subprog */
1758 } else if (bpf_pseudo_call(insn)) {
1759 ret = add_subprog(env, i + insn->imm + 1);
1761 ret = add_kfunc_call(env, insn->imm);
1768 /* Add a fake 'exit' subprog which could simplify subprog iteration
1769 * logic. 'subprog_cnt' should not be increased.
1771 subprog[env->subprog_cnt].start = insn_cnt;
1773 if (env->log.level & BPF_LOG_LEVEL2)
1774 for (i = 0; i < env->subprog_cnt; i++)
1775 verbose(env, "func#%d @%d\n", i, subprog[i].start);
1780 static int check_subprogs(struct bpf_verifier_env *env)
1782 int i, subprog_start, subprog_end, off, cur_subprog = 0;
1783 struct bpf_subprog_info *subprog = env->subprog_info;
1784 struct bpf_insn *insn = env->prog->insnsi;
1785 int insn_cnt = env->prog->len;
1787 /* now check that all jumps are within the same subprog */
1788 subprog_start = subprog[cur_subprog].start;
1789 subprog_end = subprog[cur_subprog + 1].start;
1790 for (i = 0; i < insn_cnt; i++) {
1791 u8 code = insn[i].code;
1793 if (code == (BPF_JMP | BPF_CALL) &&
1794 insn[i].imm == BPF_FUNC_tail_call &&
1795 insn[i].src_reg != BPF_PSEUDO_CALL)
1796 subprog[cur_subprog].has_tail_call = true;
1797 if (BPF_CLASS(code) == BPF_LD &&
1798 (BPF_MODE(code) == BPF_ABS || BPF_MODE(code) == BPF_IND))
1799 subprog[cur_subprog].has_ld_abs = true;
1800 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
1802 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
1804 off = i + insn[i].off + 1;
1805 if (off < subprog_start || off >= subprog_end) {
1806 verbose(env, "jump out of range from insn %d to %d\n", i, off);
1810 if (i == subprog_end - 1) {
1811 /* to avoid fall-through from one subprog into another
1812 * the last insn of the subprog should be either exit
1813 * or unconditional jump back
1815 if (code != (BPF_JMP | BPF_EXIT) &&
1816 code != (BPF_JMP | BPF_JA)) {
1817 verbose(env, "last insn is not an exit or jmp\n");
1820 subprog_start = subprog_end;
1822 if (cur_subprog < env->subprog_cnt)
1823 subprog_end = subprog[cur_subprog + 1].start;
1829 /* Parentage chain of this register (or stack slot) should take care of all
1830 * issues like callee-saved registers, stack slot allocation time, etc.
1832 static int mark_reg_read(struct bpf_verifier_env *env,
1833 const struct bpf_reg_state *state,
1834 struct bpf_reg_state *parent, u8 flag)
1836 bool writes = parent == state->parent; /* Observe write marks */
1840 /* if read wasn't screened by an earlier write ... */
1841 if (writes && state->live & REG_LIVE_WRITTEN)
1843 if (parent->live & REG_LIVE_DONE) {
1844 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
1845 reg_type_str[parent->type],
1846 parent->var_off.value, parent->off);
1849 /* The first condition is more likely to be true than the
1850 * second, checked it first.
1852 if ((parent->live & REG_LIVE_READ) == flag ||
1853 parent->live & REG_LIVE_READ64)
1854 /* The parentage chain never changes and
1855 * this parent was already marked as LIVE_READ.
1856 * There is no need to keep walking the chain again and
1857 * keep re-marking all parents as LIVE_READ.
1858 * This case happens when the same register is read
1859 * multiple times without writes into it in-between.
1860 * Also, if parent has the stronger REG_LIVE_READ64 set,
1861 * then no need to set the weak REG_LIVE_READ32.
1864 /* ... then we depend on parent's value */
1865 parent->live |= flag;
1866 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
1867 if (flag == REG_LIVE_READ64)
1868 parent->live &= ~REG_LIVE_READ32;
1870 parent = state->parent;
1875 if (env->longest_mark_read_walk < cnt)
1876 env->longest_mark_read_walk = cnt;
1880 /* This function is supposed to be used by the following 32-bit optimization
1881 * code only. It returns TRUE if the source or destination register operates
1882 * on 64-bit, otherwise return FALSE.
1884 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
1885 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
1890 class = BPF_CLASS(code);
1892 if (class == BPF_JMP) {
1893 /* BPF_EXIT for "main" will reach here. Return TRUE
1898 if (op == BPF_CALL) {
1899 /* BPF to BPF call will reach here because of marking
1900 * caller saved clobber with DST_OP_NO_MARK for which we
1901 * don't care the register def because they are anyway
1902 * marked as NOT_INIT already.
1904 if (insn->src_reg == BPF_PSEUDO_CALL)
1906 /* Helper call will reach here because of arg type
1907 * check, conservatively return TRUE.
1916 if (class == BPF_ALU64 || class == BPF_JMP ||
1917 /* BPF_END always use BPF_ALU class. */
1918 (class == BPF_ALU && op == BPF_END && insn->imm == 64))
1921 if (class == BPF_ALU || class == BPF_JMP32)
1924 if (class == BPF_LDX) {
1926 return BPF_SIZE(code) == BPF_DW;
1927 /* LDX source must be ptr. */
1931 if (class == BPF_STX) {
1932 /* BPF_STX (including atomic variants) has multiple source
1933 * operands, one of which is a ptr. Check whether the caller is
1936 if (t == SRC_OP && reg->type != SCALAR_VALUE)
1938 return BPF_SIZE(code) == BPF_DW;
1941 if (class == BPF_LD) {
1942 u8 mode = BPF_MODE(code);
1945 if (mode == BPF_IMM)
1948 /* Both LD_IND and LD_ABS return 32-bit data. */
1952 /* Implicit ctx ptr. */
1953 if (regno == BPF_REG_6)
1956 /* Explicit source could be any width. */
1960 if (class == BPF_ST)
1961 /* The only source register for BPF_ST is a ptr. */
1964 /* Conservatively return true at default. */
1968 /* Return the regno defined by the insn, or -1. */
1969 static int insn_def_regno(const struct bpf_insn *insn)
1971 switch (BPF_CLASS(insn->code)) {
1977 if (BPF_MODE(insn->code) == BPF_ATOMIC &&
1978 (insn->imm & BPF_FETCH)) {
1979 if (insn->imm == BPF_CMPXCHG)
1982 return insn->src_reg;
1987 return insn->dst_reg;
1991 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
1992 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
1994 int dst_reg = insn_def_regno(insn);
1999 return !is_reg64(env, insn, dst_reg, NULL, DST_OP);
2002 static void mark_insn_zext(struct bpf_verifier_env *env,
2003 struct bpf_reg_state *reg)
2005 s32 def_idx = reg->subreg_def;
2007 if (def_idx == DEF_NOT_SUBREG)
2010 env->insn_aux_data[def_idx - 1].zext_dst = true;
2011 /* The dst will be zero extended, so won't be sub-register anymore. */
2012 reg->subreg_def = DEF_NOT_SUBREG;
2015 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
2016 enum reg_arg_type t)
2018 struct bpf_verifier_state *vstate = env->cur_state;
2019 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2020 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
2021 struct bpf_reg_state *reg, *regs = state->regs;
2024 if (regno >= MAX_BPF_REG) {
2025 verbose(env, "R%d is invalid\n", regno);
2030 rw64 = is_reg64(env, insn, regno, reg, t);
2032 /* check whether register used as source operand can be read */
2033 if (reg->type == NOT_INIT) {
2034 verbose(env, "R%d !read_ok\n", regno);
2037 /* We don't need to worry about FP liveness because it's read-only */
2038 if (regno == BPF_REG_FP)
2042 mark_insn_zext(env, reg);
2044 return mark_reg_read(env, reg, reg->parent,
2045 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
2047 /* check whether register used as dest operand can be written to */
2048 if (regno == BPF_REG_FP) {
2049 verbose(env, "frame pointer is read only\n");
2052 reg->live |= REG_LIVE_WRITTEN;
2053 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
2055 mark_reg_unknown(env, regs, regno);
2060 /* for any branch, call, exit record the history of jmps in the given state */
2061 static int push_jmp_history(struct bpf_verifier_env *env,
2062 struct bpf_verifier_state *cur)
2064 u32 cnt = cur->jmp_history_cnt;
2065 struct bpf_idx_pair *p;
2068 p = krealloc(cur->jmp_history, cnt * sizeof(*p), GFP_USER);
2071 p[cnt - 1].idx = env->insn_idx;
2072 p[cnt - 1].prev_idx = env->prev_insn_idx;
2073 cur->jmp_history = p;
2074 cur->jmp_history_cnt = cnt;
2078 /* Backtrack one insn at a time. If idx is not at the top of recorded
2079 * history then previous instruction came from straight line execution.
2081 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
2086 if (cnt && st->jmp_history[cnt - 1].idx == i) {
2087 i = st->jmp_history[cnt - 1].prev_idx;
2095 static const char *disasm_kfunc_name(void *data, const struct bpf_insn *insn)
2097 const struct btf_type *func;
2099 if (insn->src_reg != BPF_PSEUDO_KFUNC_CALL)
2102 func = btf_type_by_id(btf_vmlinux, insn->imm);
2103 return btf_name_by_offset(btf_vmlinux, func->name_off);
2106 /* For given verifier state backtrack_insn() is called from the last insn to
2107 * the first insn. Its purpose is to compute a bitmask of registers and
2108 * stack slots that needs precision in the parent verifier state.
2110 static int backtrack_insn(struct bpf_verifier_env *env, int idx,
2111 u32 *reg_mask, u64 *stack_mask)
2113 const struct bpf_insn_cbs cbs = {
2114 .cb_call = disasm_kfunc_name,
2115 .cb_print = verbose,
2116 .private_data = env,
2118 struct bpf_insn *insn = env->prog->insnsi + idx;
2119 u8 class = BPF_CLASS(insn->code);
2120 u8 opcode = BPF_OP(insn->code);
2121 u8 mode = BPF_MODE(insn->code);
2122 u32 dreg = 1u << insn->dst_reg;
2123 u32 sreg = 1u << insn->src_reg;
2126 if (insn->code == 0)
2128 if (env->log.level & BPF_LOG_LEVEL) {
2129 verbose(env, "regs=%x stack=%llx before ", *reg_mask, *stack_mask);
2130 verbose(env, "%d: ", idx);
2131 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
2134 if (class == BPF_ALU || class == BPF_ALU64) {
2135 if (!(*reg_mask & dreg))
2137 if (opcode == BPF_MOV) {
2138 if (BPF_SRC(insn->code) == BPF_X) {
2140 * dreg needs precision after this insn
2141 * sreg needs precision before this insn
2147 * dreg needs precision after this insn.
2148 * Corresponding register is already marked
2149 * as precise=true in this verifier state.
2150 * No further markings in parent are necessary
2155 if (BPF_SRC(insn->code) == BPF_X) {
2157 * both dreg and sreg need precision
2162 * dreg still needs precision before this insn
2165 } else if (class == BPF_LDX) {
2166 if (!(*reg_mask & dreg))
2170 /* scalars can only be spilled into stack w/o losing precision.
2171 * Load from any other memory can be zero extended.
2172 * The desire to keep that precision is already indicated
2173 * by 'precise' mark in corresponding register of this state.
2174 * No further tracking necessary.
2176 if (insn->src_reg != BPF_REG_FP)
2178 if (BPF_SIZE(insn->code) != BPF_DW)
2181 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
2182 * that [fp - off] slot contains scalar that needs to be
2183 * tracked with precision
2185 spi = (-insn->off - 1) / BPF_REG_SIZE;
2187 verbose(env, "BUG spi %d\n", spi);
2188 WARN_ONCE(1, "verifier backtracking bug");
2191 *stack_mask |= 1ull << spi;
2192 } else if (class == BPF_STX || class == BPF_ST) {
2193 if (*reg_mask & dreg)
2194 /* stx & st shouldn't be using _scalar_ dst_reg
2195 * to access memory. It means backtracking
2196 * encountered a case of pointer subtraction.
2199 /* scalars can only be spilled into stack */
2200 if (insn->dst_reg != BPF_REG_FP)
2202 if (BPF_SIZE(insn->code) != BPF_DW)
2204 spi = (-insn->off - 1) / BPF_REG_SIZE;
2206 verbose(env, "BUG spi %d\n", spi);
2207 WARN_ONCE(1, "verifier backtracking bug");
2210 if (!(*stack_mask & (1ull << spi)))
2212 *stack_mask &= ~(1ull << spi);
2213 if (class == BPF_STX)
2215 } else if (class == BPF_JMP || class == BPF_JMP32) {
2216 if (opcode == BPF_CALL) {
2217 if (insn->src_reg == BPF_PSEUDO_CALL)
2219 /* regular helper call sets R0 */
2221 if (*reg_mask & 0x3f) {
2222 /* if backtracing was looking for registers R1-R5
2223 * they should have been found already.
2225 verbose(env, "BUG regs %x\n", *reg_mask);
2226 WARN_ONCE(1, "verifier backtracking bug");
2229 } else if (opcode == BPF_EXIT) {
2232 } else if (class == BPF_LD) {
2233 if (!(*reg_mask & dreg))
2236 /* It's ld_imm64 or ld_abs or ld_ind.
2237 * For ld_imm64 no further tracking of precision
2238 * into parent is necessary
2240 if (mode == BPF_IND || mode == BPF_ABS)
2241 /* to be analyzed */
2247 /* the scalar precision tracking algorithm:
2248 * . at the start all registers have precise=false.
2249 * . scalar ranges are tracked as normal through alu and jmp insns.
2250 * . once precise value of the scalar register is used in:
2251 * . ptr + scalar alu
2252 * . if (scalar cond K|scalar)
2253 * . helper_call(.., scalar, ...) where ARG_CONST is expected
2254 * backtrack through the verifier states and mark all registers and
2255 * stack slots with spilled constants that these scalar regisers
2256 * should be precise.
2257 * . during state pruning two registers (or spilled stack slots)
2258 * are equivalent if both are not precise.
2260 * Note the verifier cannot simply walk register parentage chain,
2261 * since many different registers and stack slots could have been
2262 * used to compute single precise scalar.
2264 * The approach of starting with precise=true for all registers and then
2265 * backtrack to mark a register as not precise when the verifier detects
2266 * that program doesn't care about specific value (e.g., when helper
2267 * takes register as ARG_ANYTHING parameter) is not safe.
2269 * It's ok to walk single parentage chain of the verifier states.
2270 * It's possible that this backtracking will go all the way till 1st insn.
2271 * All other branches will be explored for needing precision later.
2273 * The backtracking needs to deal with cases like:
2274 * R8=map_value(id=0,off=0,ks=4,vs=1952,imm=0) R9_w=map_value(id=0,off=40,ks=4,vs=1952,imm=0)
2277 * if r5 > 0x79f goto pc+7
2278 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
2281 * call bpf_perf_event_output#25
2282 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
2286 * call foo // uses callee's r6 inside to compute r0
2290 * to track above reg_mask/stack_mask needs to be independent for each frame.
2292 * Also if parent's curframe > frame where backtracking started,
2293 * the verifier need to mark registers in both frames, otherwise callees
2294 * may incorrectly prune callers. This is similar to
2295 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
2297 * For now backtracking falls back into conservative marking.
2299 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
2300 struct bpf_verifier_state *st)
2302 struct bpf_func_state *func;
2303 struct bpf_reg_state *reg;
2306 /* big hammer: mark all scalars precise in this path.
2307 * pop_stack may still get !precise scalars.
2309 for (; st; st = st->parent)
2310 for (i = 0; i <= st->curframe; i++) {
2311 func = st->frame[i];
2312 for (j = 0; j < BPF_REG_FP; j++) {
2313 reg = &func->regs[j];
2314 if (reg->type != SCALAR_VALUE)
2316 reg->precise = true;
2318 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
2319 if (func->stack[j].slot_type[0] != STACK_SPILL)
2321 reg = &func->stack[j].spilled_ptr;
2322 if (reg->type != SCALAR_VALUE)
2324 reg->precise = true;
2329 static int __mark_chain_precision(struct bpf_verifier_env *env, int regno,
2332 struct bpf_verifier_state *st = env->cur_state;
2333 int first_idx = st->first_insn_idx;
2334 int last_idx = env->insn_idx;
2335 struct bpf_func_state *func;
2336 struct bpf_reg_state *reg;
2337 u32 reg_mask = regno >= 0 ? 1u << regno : 0;
2338 u64 stack_mask = spi >= 0 ? 1ull << spi : 0;
2339 bool skip_first = true;
2340 bool new_marks = false;
2343 if (!env->bpf_capable)
2346 func = st->frame[st->curframe];
2348 reg = &func->regs[regno];
2349 if (reg->type != SCALAR_VALUE) {
2350 WARN_ONCE(1, "backtracing misuse");
2357 reg->precise = true;
2361 if (func->stack[spi].slot_type[0] != STACK_SPILL) {
2365 reg = &func->stack[spi].spilled_ptr;
2366 if (reg->type != SCALAR_VALUE) {
2374 reg->precise = true;
2380 if (!reg_mask && !stack_mask)
2383 DECLARE_BITMAP(mask, 64);
2384 u32 history = st->jmp_history_cnt;
2386 if (env->log.level & BPF_LOG_LEVEL)
2387 verbose(env, "last_idx %d first_idx %d\n", last_idx, first_idx);
2388 for (i = last_idx;;) {
2393 err = backtrack_insn(env, i, ®_mask, &stack_mask);
2395 if (err == -ENOTSUPP) {
2396 mark_all_scalars_precise(env, st);
2401 if (!reg_mask && !stack_mask)
2402 /* Found assignment(s) into tracked register in this state.
2403 * Since this state is already marked, just return.
2404 * Nothing to be tracked further in the parent state.
2409 i = get_prev_insn_idx(st, i, &history);
2410 if (i >= env->prog->len) {
2411 /* This can happen if backtracking reached insn 0
2412 * and there are still reg_mask or stack_mask
2414 * It means the backtracking missed the spot where
2415 * particular register was initialized with a constant.
2417 verbose(env, "BUG backtracking idx %d\n", i);
2418 WARN_ONCE(1, "verifier backtracking bug");
2427 func = st->frame[st->curframe];
2428 bitmap_from_u64(mask, reg_mask);
2429 for_each_set_bit(i, mask, 32) {
2430 reg = &func->regs[i];
2431 if (reg->type != SCALAR_VALUE) {
2432 reg_mask &= ~(1u << i);
2437 reg->precise = true;
2440 bitmap_from_u64(mask, stack_mask);
2441 for_each_set_bit(i, mask, 64) {
2442 if (i >= func->allocated_stack / BPF_REG_SIZE) {
2443 /* the sequence of instructions:
2445 * 3: (7b) *(u64 *)(r3 -8) = r0
2446 * 4: (79) r4 = *(u64 *)(r10 -8)
2447 * doesn't contain jmps. It's backtracked
2448 * as a single block.
2449 * During backtracking insn 3 is not recognized as
2450 * stack access, so at the end of backtracking
2451 * stack slot fp-8 is still marked in stack_mask.
2452 * However the parent state may not have accessed
2453 * fp-8 and it's "unallocated" stack space.
2454 * In such case fallback to conservative.
2456 mark_all_scalars_precise(env, st);
2460 if (func->stack[i].slot_type[0] != STACK_SPILL) {
2461 stack_mask &= ~(1ull << i);
2464 reg = &func->stack[i].spilled_ptr;
2465 if (reg->type != SCALAR_VALUE) {
2466 stack_mask &= ~(1ull << i);
2471 reg->precise = true;
2473 if (env->log.level & BPF_LOG_LEVEL) {
2474 print_verifier_state(env, func);
2475 verbose(env, "parent %s regs=%x stack=%llx marks\n",
2476 new_marks ? "didn't have" : "already had",
2477 reg_mask, stack_mask);
2480 if (!reg_mask && !stack_mask)
2485 last_idx = st->last_insn_idx;
2486 first_idx = st->first_insn_idx;
2491 static int mark_chain_precision(struct bpf_verifier_env *env, int regno)
2493 return __mark_chain_precision(env, regno, -1);
2496 static int mark_chain_precision_stack(struct bpf_verifier_env *env, int spi)
2498 return __mark_chain_precision(env, -1, spi);
2501 static bool is_spillable_regtype(enum bpf_reg_type type)
2504 case PTR_TO_MAP_VALUE:
2505 case PTR_TO_MAP_VALUE_OR_NULL:
2509 case PTR_TO_PACKET_META:
2510 case PTR_TO_PACKET_END:
2511 case PTR_TO_FLOW_KEYS:
2512 case CONST_PTR_TO_MAP:
2514 case PTR_TO_SOCKET_OR_NULL:
2515 case PTR_TO_SOCK_COMMON:
2516 case PTR_TO_SOCK_COMMON_OR_NULL:
2517 case PTR_TO_TCP_SOCK:
2518 case PTR_TO_TCP_SOCK_OR_NULL:
2519 case PTR_TO_XDP_SOCK:
2521 case PTR_TO_BTF_ID_OR_NULL:
2522 case PTR_TO_RDONLY_BUF:
2523 case PTR_TO_RDONLY_BUF_OR_NULL:
2524 case PTR_TO_RDWR_BUF:
2525 case PTR_TO_RDWR_BUF_OR_NULL:
2526 case PTR_TO_PERCPU_BTF_ID:
2528 case PTR_TO_MEM_OR_NULL:
2530 case PTR_TO_MAP_KEY:
2537 /* Does this register contain a constant zero? */
2538 static bool register_is_null(struct bpf_reg_state *reg)
2540 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
2543 static bool register_is_const(struct bpf_reg_state *reg)
2545 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
2548 static bool __is_scalar_unbounded(struct bpf_reg_state *reg)
2550 return tnum_is_unknown(reg->var_off) &&
2551 reg->smin_value == S64_MIN && reg->smax_value == S64_MAX &&
2552 reg->umin_value == 0 && reg->umax_value == U64_MAX &&
2553 reg->s32_min_value == S32_MIN && reg->s32_max_value == S32_MAX &&
2554 reg->u32_min_value == 0 && reg->u32_max_value == U32_MAX;
2557 static bool register_is_bounded(struct bpf_reg_state *reg)
2559 return reg->type == SCALAR_VALUE && !__is_scalar_unbounded(reg);
2562 static bool __is_pointer_value(bool allow_ptr_leaks,
2563 const struct bpf_reg_state *reg)
2565 if (allow_ptr_leaks)
2568 return reg->type != SCALAR_VALUE;
2571 static void save_register_state(struct bpf_func_state *state,
2572 int spi, struct bpf_reg_state *reg)
2576 state->stack[spi].spilled_ptr = *reg;
2577 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2579 for (i = 0; i < BPF_REG_SIZE; i++)
2580 state->stack[spi].slot_type[i] = STACK_SPILL;
2583 /* check_stack_{read,write}_fixed_off functions track spill/fill of registers,
2584 * stack boundary and alignment are checked in check_mem_access()
2586 static int check_stack_write_fixed_off(struct bpf_verifier_env *env,
2587 /* stack frame we're writing to */
2588 struct bpf_func_state *state,
2589 int off, int size, int value_regno,
2592 struct bpf_func_state *cur; /* state of the current function */
2593 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
2594 u32 dst_reg = env->prog->insnsi[insn_idx].dst_reg;
2595 struct bpf_reg_state *reg = NULL;
2597 err = grow_stack_state(state, round_up(slot + 1, BPF_REG_SIZE));
2600 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
2601 * so it's aligned access and [off, off + size) are within stack limits
2603 if (!env->allow_ptr_leaks &&
2604 state->stack[spi].slot_type[0] == STACK_SPILL &&
2605 size != BPF_REG_SIZE) {
2606 verbose(env, "attempt to corrupt spilled pointer on stack\n");
2610 cur = env->cur_state->frame[env->cur_state->curframe];
2611 if (value_regno >= 0)
2612 reg = &cur->regs[value_regno];
2614 if (reg && size == BPF_REG_SIZE && register_is_bounded(reg) &&
2615 !register_is_null(reg) && env->bpf_capable) {
2616 if (dst_reg != BPF_REG_FP) {
2617 /* The backtracking logic can only recognize explicit
2618 * stack slot address like [fp - 8]. Other spill of
2619 * scalar via different register has to be conservative.
2620 * Backtrack from here and mark all registers as precise
2621 * that contributed into 'reg' being a constant.
2623 err = mark_chain_precision(env, value_regno);
2627 save_register_state(state, spi, reg);
2628 } else if (reg && is_spillable_regtype(reg->type)) {
2629 /* register containing pointer is being spilled into stack */
2630 if (size != BPF_REG_SIZE) {
2631 verbose_linfo(env, insn_idx, "; ");
2632 verbose(env, "invalid size of register spill\n");
2636 if (state != cur && reg->type == PTR_TO_STACK) {
2637 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
2641 if (!env->bypass_spec_v4) {
2642 bool sanitize = false;
2644 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
2645 register_is_const(&state->stack[spi].spilled_ptr))
2647 for (i = 0; i < BPF_REG_SIZE; i++)
2648 if (state->stack[spi].slot_type[i] == STACK_MISC) {
2653 int *poff = &env->insn_aux_data[insn_idx].sanitize_stack_off;
2654 int soff = (-spi - 1) * BPF_REG_SIZE;
2656 /* detected reuse of integer stack slot with a pointer
2657 * which means either llvm is reusing stack slot or
2658 * an attacker is trying to exploit CVE-2018-3639
2659 * (speculative store bypass)
2660 * Have to sanitize that slot with preemptive
2663 if (*poff && *poff != soff) {
2664 /* disallow programs where single insn stores
2665 * into two different stack slots, since verifier
2666 * cannot sanitize them
2669 "insn %d cannot access two stack slots fp%d and fp%d",
2670 insn_idx, *poff, soff);
2676 save_register_state(state, spi, reg);
2678 u8 type = STACK_MISC;
2680 /* regular write of data into stack destroys any spilled ptr */
2681 state->stack[spi].spilled_ptr.type = NOT_INIT;
2682 /* Mark slots as STACK_MISC if they belonged to spilled ptr. */
2683 if (state->stack[spi].slot_type[0] == STACK_SPILL)
2684 for (i = 0; i < BPF_REG_SIZE; i++)
2685 state->stack[spi].slot_type[i] = STACK_MISC;
2687 /* only mark the slot as written if all 8 bytes were written
2688 * otherwise read propagation may incorrectly stop too soon
2689 * when stack slots are partially written.
2690 * This heuristic means that read propagation will be
2691 * conservative, since it will add reg_live_read marks
2692 * to stack slots all the way to first state when programs
2693 * writes+reads less than 8 bytes
2695 if (size == BPF_REG_SIZE)
2696 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2698 /* when we zero initialize stack slots mark them as such */
2699 if (reg && register_is_null(reg)) {
2700 /* backtracking doesn't work for STACK_ZERO yet. */
2701 err = mark_chain_precision(env, value_regno);
2707 /* Mark slots affected by this stack write. */
2708 for (i = 0; i < size; i++)
2709 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
2715 /* Write the stack: 'stack[ptr_regno + off] = value_regno'. 'ptr_regno' is
2716 * known to contain a variable offset.
2717 * This function checks whether the write is permitted and conservatively
2718 * tracks the effects of the write, considering that each stack slot in the
2719 * dynamic range is potentially written to.
2721 * 'off' includes 'regno->off'.
2722 * 'value_regno' can be -1, meaning that an unknown value is being written to
2725 * Spilled pointers in range are not marked as written because we don't know
2726 * what's going to be actually written. This means that read propagation for
2727 * future reads cannot be terminated by this write.
2729 * For privileged programs, uninitialized stack slots are considered
2730 * initialized by this write (even though we don't know exactly what offsets
2731 * are going to be written to). The idea is that we don't want the verifier to
2732 * reject future reads that access slots written to through variable offsets.
2734 static int check_stack_write_var_off(struct bpf_verifier_env *env,
2735 /* func where register points to */
2736 struct bpf_func_state *state,
2737 int ptr_regno, int off, int size,
2738 int value_regno, int insn_idx)
2740 struct bpf_func_state *cur; /* state of the current function */
2741 int min_off, max_off;
2743 struct bpf_reg_state *ptr_reg = NULL, *value_reg = NULL;
2744 bool writing_zero = false;
2745 /* set if the fact that we're writing a zero is used to let any
2746 * stack slots remain STACK_ZERO
2748 bool zero_used = false;
2750 cur = env->cur_state->frame[env->cur_state->curframe];
2751 ptr_reg = &cur->regs[ptr_regno];
2752 min_off = ptr_reg->smin_value + off;
2753 max_off = ptr_reg->smax_value + off + size;
2754 if (value_regno >= 0)
2755 value_reg = &cur->regs[value_regno];
2756 if (value_reg && register_is_null(value_reg))
2757 writing_zero = true;
2759 err = grow_stack_state(state, round_up(-min_off, BPF_REG_SIZE));
2764 /* Variable offset writes destroy any spilled pointers in range. */
2765 for (i = min_off; i < max_off; i++) {
2766 u8 new_type, *stype;
2770 spi = slot / BPF_REG_SIZE;
2771 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
2773 if (!env->allow_ptr_leaks
2774 && *stype != NOT_INIT
2775 && *stype != SCALAR_VALUE) {
2776 /* Reject the write if there's are spilled pointers in
2777 * range. If we didn't reject here, the ptr status
2778 * would be erased below (even though not all slots are
2779 * actually overwritten), possibly opening the door to
2782 verbose(env, "spilled ptr in range of var-offset stack write; insn %d, ptr off: %d",
2787 /* Erase all spilled pointers. */
2788 state->stack[spi].spilled_ptr.type = NOT_INIT;
2790 /* Update the slot type. */
2791 new_type = STACK_MISC;
2792 if (writing_zero && *stype == STACK_ZERO) {
2793 new_type = STACK_ZERO;
2796 /* If the slot is STACK_INVALID, we check whether it's OK to
2797 * pretend that it will be initialized by this write. The slot
2798 * might not actually be written to, and so if we mark it as
2799 * initialized future reads might leak uninitialized memory.
2800 * For privileged programs, we will accept such reads to slots
2801 * that may or may not be written because, if we're reject
2802 * them, the error would be too confusing.
2804 if (*stype == STACK_INVALID && !env->allow_uninit_stack) {
2805 verbose(env, "uninit stack in range of var-offset write prohibited for !root; insn %d, off: %d",
2812 /* backtracking doesn't work for STACK_ZERO yet. */
2813 err = mark_chain_precision(env, value_regno);
2820 /* When register 'dst_regno' is assigned some values from stack[min_off,
2821 * max_off), we set the register's type according to the types of the
2822 * respective stack slots. If all the stack values are known to be zeros, then
2823 * so is the destination reg. Otherwise, the register is considered to be
2824 * SCALAR. This function does not deal with register filling; the caller must
2825 * ensure that all spilled registers in the stack range have been marked as
2828 static void mark_reg_stack_read(struct bpf_verifier_env *env,
2829 /* func where src register points to */
2830 struct bpf_func_state *ptr_state,
2831 int min_off, int max_off, int dst_regno)
2833 struct bpf_verifier_state *vstate = env->cur_state;
2834 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2839 for (i = min_off; i < max_off; i++) {
2841 spi = slot / BPF_REG_SIZE;
2842 stype = ptr_state->stack[spi].slot_type;
2843 if (stype[slot % BPF_REG_SIZE] != STACK_ZERO)
2847 if (zeros == max_off - min_off) {
2848 /* any access_size read into register is zero extended,
2849 * so the whole register == const_zero
2851 __mark_reg_const_zero(&state->regs[dst_regno]);
2852 /* backtracking doesn't support STACK_ZERO yet,
2853 * so mark it precise here, so that later
2854 * backtracking can stop here.
2855 * Backtracking may not need this if this register
2856 * doesn't participate in pointer adjustment.
2857 * Forward propagation of precise flag is not
2858 * necessary either. This mark is only to stop
2859 * backtracking. Any register that contributed
2860 * to const 0 was marked precise before spill.
2862 state->regs[dst_regno].precise = true;
2864 /* have read misc data from the stack */
2865 mark_reg_unknown(env, state->regs, dst_regno);
2867 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
2870 /* Read the stack at 'off' and put the results into the register indicated by
2871 * 'dst_regno'. It handles reg filling if the addressed stack slot is a
2874 * 'dst_regno' can be -1, meaning that the read value is not going to a
2877 * The access is assumed to be within the current stack bounds.
2879 static int check_stack_read_fixed_off(struct bpf_verifier_env *env,
2880 /* func where src register points to */
2881 struct bpf_func_state *reg_state,
2882 int off, int size, int dst_regno)
2884 struct bpf_verifier_state *vstate = env->cur_state;
2885 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2886 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
2887 struct bpf_reg_state *reg;
2890 stype = reg_state->stack[spi].slot_type;
2891 reg = ®_state->stack[spi].spilled_ptr;
2893 if (stype[0] == STACK_SPILL) {
2894 if (size != BPF_REG_SIZE) {
2895 if (reg->type != SCALAR_VALUE) {
2896 verbose_linfo(env, env->insn_idx, "; ");
2897 verbose(env, "invalid size of register fill\n");
2900 if (dst_regno >= 0) {
2901 mark_reg_unknown(env, state->regs, dst_regno);
2902 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
2904 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2907 for (i = 1; i < BPF_REG_SIZE; i++) {
2908 if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) {
2909 verbose(env, "corrupted spill memory\n");
2914 if (dst_regno >= 0) {
2915 /* restore register state from stack */
2916 state->regs[dst_regno] = *reg;
2917 /* mark reg as written since spilled pointer state likely
2918 * has its liveness marks cleared by is_state_visited()
2919 * which resets stack/reg liveness for state transitions
2921 state->regs[dst_regno].live |= REG_LIVE_WRITTEN;
2922 } else if (__is_pointer_value(env->allow_ptr_leaks, reg)) {
2923 /* If dst_regno==-1, the caller is asking us whether
2924 * it is acceptable to use this value as a SCALAR_VALUE
2926 * We must not allow unprivileged callers to do that
2927 * with spilled pointers.
2929 verbose(env, "leaking pointer from stack off %d\n",
2933 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2937 for (i = 0; i < size; i++) {
2938 type = stype[(slot - i) % BPF_REG_SIZE];
2939 if (type == STACK_MISC)
2941 if (type == STACK_ZERO)
2943 verbose(env, "invalid read from stack off %d+%d size %d\n",
2947 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2949 mark_reg_stack_read(env, reg_state, off, off + size, dst_regno);
2954 enum stack_access_src {
2955 ACCESS_DIRECT = 1, /* the access is performed by an instruction */
2956 ACCESS_HELPER = 2, /* the access is performed by a helper */
2959 static int check_stack_range_initialized(struct bpf_verifier_env *env,
2960 int regno, int off, int access_size,
2961 bool zero_size_allowed,
2962 enum stack_access_src type,
2963 struct bpf_call_arg_meta *meta);
2965 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
2967 return cur_regs(env) + regno;
2970 /* Read the stack at 'ptr_regno + off' and put the result into the register
2972 * 'off' includes the pointer register's fixed offset(i.e. 'ptr_regno.off'),
2973 * but not its variable offset.
2974 * 'size' is assumed to be <= reg size and the access is assumed to be aligned.
2976 * As opposed to check_stack_read_fixed_off, this function doesn't deal with
2977 * filling registers (i.e. reads of spilled register cannot be detected when
2978 * the offset is not fixed). We conservatively mark 'dst_regno' as containing
2979 * SCALAR_VALUE. That's why we assert that the 'ptr_regno' has a variable
2980 * offset; for a fixed offset check_stack_read_fixed_off should be used
2983 static int check_stack_read_var_off(struct bpf_verifier_env *env,
2984 int ptr_regno, int off, int size, int dst_regno)
2986 /* The state of the source register. */
2987 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
2988 struct bpf_func_state *ptr_state = func(env, reg);
2990 int min_off, max_off;
2992 /* Note that we pass a NULL meta, so raw access will not be permitted.
2994 err = check_stack_range_initialized(env, ptr_regno, off, size,
2995 false, ACCESS_DIRECT, NULL);
2999 min_off = reg->smin_value + off;
3000 max_off = reg->smax_value + off;
3001 mark_reg_stack_read(env, ptr_state, min_off, max_off + size, dst_regno);
3005 /* check_stack_read dispatches to check_stack_read_fixed_off or
3006 * check_stack_read_var_off.
3008 * The caller must ensure that the offset falls within the allocated stack
3011 * 'dst_regno' is a register which will receive the value from the stack. It
3012 * can be -1, meaning that the read value is not going to a register.
3014 static int check_stack_read(struct bpf_verifier_env *env,
3015 int ptr_regno, int off, int size,
3018 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
3019 struct bpf_func_state *state = func(env, reg);
3021 /* Some accesses are only permitted with a static offset. */
3022 bool var_off = !tnum_is_const(reg->var_off);
3024 /* The offset is required to be static when reads don't go to a
3025 * register, in order to not leak pointers (see
3026 * check_stack_read_fixed_off).
3028 if (dst_regno < 0 && var_off) {
3031 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3032 verbose(env, "variable offset stack pointer cannot be passed into helper function; var_off=%s off=%d size=%d\n",
3036 /* Variable offset is prohibited for unprivileged mode for simplicity
3037 * since it requires corresponding support in Spectre masking for stack
3038 * ALU. See also retrieve_ptr_limit().
3040 if (!env->bypass_spec_v1 && var_off) {
3043 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3044 verbose(env, "R%d variable offset stack access prohibited for !root, var_off=%s\n",
3050 off += reg->var_off.value;
3051 err = check_stack_read_fixed_off(env, state, off, size,
3054 /* Variable offset stack reads need more conservative handling
3055 * than fixed offset ones. Note that dst_regno >= 0 on this
3058 err = check_stack_read_var_off(env, ptr_regno, off, size,
3065 /* check_stack_write dispatches to check_stack_write_fixed_off or
3066 * check_stack_write_var_off.
3068 * 'ptr_regno' is the register used as a pointer into the stack.
3069 * 'off' includes 'ptr_regno->off', but not its variable offset (if any).
3070 * 'value_regno' is the register whose value we're writing to the stack. It can
3071 * be -1, meaning that we're not writing from a register.
3073 * The caller must ensure that the offset falls within the maximum stack size.
3075 static int check_stack_write(struct bpf_verifier_env *env,
3076 int ptr_regno, int off, int size,
3077 int value_regno, int insn_idx)
3079 struct bpf_reg_state *reg = reg_state(env, ptr_regno);
3080 struct bpf_func_state *state = func(env, reg);
3083 if (tnum_is_const(reg->var_off)) {
3084 off += reg->var_off.value;
3085 err = check_stack_write_fixed_off(env, state, off, size,
3086 value_regno, insn_idx);
3088 /* Variable offset stack reads need more conservative handling
3089 * than fixed offset ones.
3091 err = check_stack_write_var_off(env, state,
3092 ptr_regno, off, size,
3093 value_regno, insn_idx);
3098 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
3099 int off, int size, enum bpf_access_type type)
3101 struct bpf_reg_state *regs = cur_regs(env);
3102 struct bpf_map *map = regs[regno].map_ptr;
3103 u32 cap = bpf_map_flags_to_cap(map);
3105 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
3106 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
3107 map->value_size, off, size);
3111 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
3112 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
3113 map->value_size, off, size);
3120 /* check read/write into memory region (e.g., map value, ringbuf sample, etc) */
3121 static int __check_mem_access(struct bpf_verifier_env *env, int regno,
3122 int off, int size, u32 mem_size,
3123 bool zero_size_allowed)
3125 bool size_ok = size > 0 || (size == 0 && zero_size_allowed);
3126 struct bpf_reg_state *reg;
3128 if (off >= 0 && size_ok && (u64)off + size <= mem_size)
3131 reg = &cur_regs(env)[regno];
3132 switch (reg->type) {
3133 case PTR_TO_MAP_KEY:
3134 verbose(env, "invalid access to map key, key_size=%d off=%d size=%d\n",
3135 mem_size, off, size);
3137 case PTR_TO_MAP_VALUE:
3138 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
3139 mem_size, off, size);
3142 case PTR_TO_PACKET_META:
3143 case PTR_TO_PACKET_END:
3144 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
3145 off, size, regno, reg->id, off, mem_size);
3149 verbose(env, "invalid access to memory, mem_size=%u off=%d size=%d\n",
3150 mem_size, off, size);
3156 /* check read/write into a memory region with possible variable offset */
3157 static int check_mem_region_access(struct bpf_verifier_env *env, u32 regno,
3158 int off, int size, u32 mem_size,
3159 bool zero_size_allowed)
3161 struct bpf_verifier_state *vstate = env->cur_state;
3162 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3163 struct bpf_reg_state *reg = &state->regs[regno];
3166 /* We may have adjusted the register pointing to memory region, so we
3167 * need to try adding each of min_value and max_value to off
3168 * to make sure our theoretical access will be safe.
3170 if (env->log.level & BPF_LOG_LEVEL)
3171 print_verifier_state(env, state);
3173 /* The minimum value is only important with signed
3174 * comparisons where we can't assume the floor of a
3175 * value is 0. If we are using signed variables for our
3176 * index'es we need to make sure that whatever we use
3177 * will have a set floor within our range.
3179 if (reg->smin_value < 0 &&
3180 (reg->smin_value == S64_MIN ||
3181 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
3182 reg->smin_value + off < 0)) {
3183 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3187 err = __check_mem_access(env, regno, reg->smin_value + off, size,
3188 mem_size, zero_size_allowed);
3190 verbose(env, "R%d min value is outside of the allowed memory range\n",
3195 /* If we haven't set a max value then we need to bail since we can't be
3196 * sure we won't do bad things.
3197 * If reg->umax_value + off could overflow, treat that as unbounded too.
3199 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
3200 verbose(env, "R%d unbounded memory access, make sure to bounds check any such access\n",
3204 err = __check_mem_access(env, regno, reg->umax_value + off, size,
3205 mem_size, zero_size_allowed);
3207 verbose(env, "R%d max value is outside of the allowed memory range\n",
3215 /* check read/write into a map element with possible variable offset */
3216 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
3217 int off, int size, bool zero_size_allowed)
3219 struct bpf_verifier_state *vstate = env->cur_state;
3220 struct bpf_func_state *state = vstate->frame[vstate->curframe];
3221 struct bpf_reg_state *reg = &state->regs[regno];
3222 struct bpf_map *map = reg->map_ptr;
3225 err = check_mem_region_access(env, regno, off, size, map->value_size,
3230 if (map_value_has_spin_lock(map)) {
3231 u32 lock = map->spin_lock_off;
3233 /* if any part of struct bpf_spin_lock can be touched by
3234 * load/store reject this program.
3235 * To check that [x1, x2) overlaps with [y1, y2)
3236 * it is sufficient to check x1 < y2 && y1 < x2.
3238 if (reg->smin_value + off < lock + sizeof(struct bpf_spin_lock) &&
3239 lock < reg->umax_value + off + size) {
3240 verbose(env, "bpf_spin_lock cannot be accessed directly by load/store\n");
3247 #define MAX_PACKET_OFF 0xffff
3249 static enum bpf_prog_type resolve_prog_type(struct bpf_prog *prog)
3251 return prog->aux->dst_prog ? prog->aux->dst_prog->type : prog->type;
3254 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
3255 const struct bpf_call_arg_meta *meta,
3256 enum bpf_access_type t)
3258 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
3260 switch (prog_type) {
3261 /* Program types only with direct read access go here! */
3262 case BPF_PROG_TYPE_LWT_IN:
3263 case BPF_PROG_TYPE_LWT_OUT:
3264 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
3265 case BPF_PROG_TYPE_SK_REUSEPORT:
3266 case BPF_PROG_TYPE_FLOW_DISSECTOR:
3267 case BPF_PROG_TYPE_CGROUP_SKB:
3272 /* Program types with direct read + write access go here! */
3273 case BPF_PROG_TYPE_SCHED_CLS:
3274 case BPF_PROG_TYPE_SCHED_ACT:
3275 case BPF_PROG_TYPE_XDP:
3276 case BPF_PROG_TYPE_LWT_XMIT:
3277 case BPF_PROG_TYPE_SK_SKB:
3278 case BPF_PROG_TYPE_SK_MSG:
3280 return meta->pkt_access;
3282 env->seen_direct_write = true;
3285 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
3287 env->seen_direct_write = true;
3296 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
3297 int size, bool zero_size_allowed)
3299 struct bpf_reg_state *regs = cur_regs(env);
3300 struct bpf_reg_state *reg = ®s[regno];
3303 /* We may have added a variable offset to the packet pointer; but any
3304 * reg->range we have comes after that. We are only checking the fixed
3308 /* We don't allow negative numbers, because we aren't tracking enough
3309 * detail to prove they're safe.
3311 if (reg->smin_value < 0) {
3312 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3317 err = reg->range < 0 ? -EINVAL :
3318 __check_mem_access(env, regno, off, size, reg->range,
3321 verbose(env, "R%d offset is outside of the packet\n", regno);
3325 /* __check_mem_access has made sure "off + size - 1" is within u16.
3326 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
3327 * otherwise find_good_pkt_pointers would have refused to set range info
3328 * that __check_mem_access would have rejected this pkt access.
3329 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
3331 env->prog->aux->max_pkt_offset =
3332 max_t(u32, env->prog->aux->max_pkt_offset,
3333 off + reg->umax_value + size - 1);
3338 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
3339 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
3340 enum bpf_access_type t, enum bpf_reg_type *reg_type,
3341 struct btf **btf, u32 *btf_id)
3343 struct bpf_insn_access_aux info = {
3344 .reg_type = *reg_type,
3348 if (env->ops->is_valid_access &&
3349 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
3350 /* A non zero info.ctx_field_size indicates that this field is a
3351 * candidate for later verifier transformation to load the whole
3352 * field and then apply a mask when accessed with a narrower
3353 * access than actual ctx access size. A zero info.ctx_field_size
3354 * will only allow for whole field access and rejects any other
3355 * type of narrower access.
3357 *reg_type = info.reg_type;
3359 if (*reg_type == PTR_TO_BTF_ID || *reg_type == PTR_TO_BTF_ID_OR_NULL) {
3361 *btf_id = info.btf_id;
3363 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
3365 /* remember the offset of last byte accessed in ctx */
3366 if (env->prog->aux->max_ctx_offset < off + size)
3367 env->prog->aux->max_ctx_offset = off + size;
3371 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
3375 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
3378 if (size < 0 || off < 0 ||
3379 (u64)off + size > sizeof(struct bpf_flow_keys)) {
3380 verbose(env, "invalid access to flow keys off=%d size=%d\n",
3387 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
3388 u32 regno, int off, int size,
3389 enum bpf_access_type t)
3391 struct bpf_reg_state *regs = cur_regs(env);
3392 struct bpf_reg_state *reg = ®s[regno];
3393 struct bpf_insn_access_aux info = {};
3396 if (reg->smin_value < 0) {
3397 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
3402 switch (reg->type) {
3403 case PTR_TO_SOCK_COMMON:
3404 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
3407 valid = bpf_sock_is_valid_access(off, size, t, &info);
3409 case PTR_TO_TCP_SOCK:
3410 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
3412 case PTR_TO_XDP_SOCK:
3413 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
3421 env->insn_aux_data[insn_idx].ctx_field_size =
3422 info.ctx_field_size;
3426 verbose(env, "R%d invalid %s access off=%d size=%d\n",
3427 regno, reg_type_str[reg->type], off, size);
3432 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
3434 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
3437 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
3439 const struct bpf_reg_state *reg = reg_state(env, regno);
3441 return reg->type == PTR_TO_CTX;
3444 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
3446 const struct bpf_reg_state *reg = reg_state(env, regno);
3448 return type_is_sk_pointer(reg->type);
3451 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
3453 const struct bpf_reg_state *reg = reg_state(env, regno);
3455 return type_is_pkt_pointer(reg->type);
3458 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
3460 const struct bpf_reg_state *reg = reg_state(env, regno);
3462 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
3463 return reg->type == PTR_TO_FLOW_KEYS;
3466 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
3467 const struct bpf_reg_state *reg,
3468 int off, int size, bool strict)
3470 struct tnum reg_off;
3473 /* Byte size accesses are always allowed. */
3474 if (!strict || size == 1)
3477 /* For platforms that do not have a Kconfig enabling
3478 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
3479 * NET_IP_ALIGN is universally set to '2'. And on platforms
3480 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
3481 * to this code only in strict mode where we want to emulate
3482 * the NET_IP_ALIGN==2 checking. Therefore use an
3483 * unconditional IP align value of '2'.
3487 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
3488 if (!tnum_is_aligned(reg_off, size)) {
3491 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3493 "misaligned packet access off %d+%s+%d+%d size %d\n",
3494 ip_align, tn_buf, reg->off, off, size);
3501 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
3502 const struct bpf_reg_state *reg,
3503 const char *pointer_desc,
3504 int off, int size, bool strict)
3506 struct tnum reg_off;
3508 /* Byte size accesses are always allowed. */
3509 if (!strict || size == 1)
3512 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
3513 if (!tnum_is_aligned(reg_off, size)) {
3516 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3517 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
3518 pointer_desc, tn_buf, reg->off, off, size);
3525 static int check_ptr_alignment(struct bpf_verifier_env *env,
3526 const struct bpf_reg_state *reg, int off,
3527 int size, bool strict_alignment_once)
3529 bool strict = env->strict_alignment || strict_alignment_once;
3530 const char *pointer_desc = "";
3532 switch (reg->type) {
3534 case PTR_TO_PACKET_META:
3535 /* Special case, because of NET_IP_ALIGN. Given metadata sits
3536 * right in front, treat it the very same way.
3538 return check_pkt_ptr_alignment(env, reg, off, size, strict);
3539 case PTR_TO_FLOW_KEYS:
3540 pointer_desc = "flow keys ";
3542 case PTR_TO_MAP_KEY:
3543 pointer_desc = "key ";
3545 case PTR_TO_MAP_VALUE:
3546 pointer_desc = "value ";
3549 pointer_desc = "context ";
3552 pointer_desc = "stack ";
3553 /* The stack spill tracking logic in check_stack_write_fixed_off()
3554 * and check_stack_read_fixed_off() relies on stack accesses being
3560 pointer_desc = "sock ";
3562 case PTR_TO_SOCK_COMMON:
3563 pointer_desc = "sock_common ";
3565 case PTR_TO_TCP_SOCK:
3566 pointer_desc = "tcp_sock ";
3568 case PTR_TO_XDP_SOCK:
3569 pointer_desc = "xdp_sock ";
3574 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
3578 static int update_stack_depth(struct bpf_verifier_env *env,
3579 const struct bpf_func_state *func,
3582 u16 stack = env->subprog_info[func->subprogno].stack_depth;
3587 /* update known max for given subprogram */
3588 env->subprog_info[func->subprogno].stack_depth = -off;
3592 /* starting from main bpf function walk all instructions of the function
3593 * and recursively walk all callees that given function can call.
3594 * Ignore jump and exit insns.
3595 * Since recursion is prevented by check_cfg() this algorithm
3596 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
3598 static int check_max_stack_depth(struct bpf_verifier_env *env)
3600 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
3601 struct bpf_subprog_info *subprog = env->subprog_info;
3602 struct bpf_insn *insn = env->prog->insnsi;
3603 bool tail_call_reachable = false;
3604 int ret_insn[MAX_CALL_FRAMES];
3605 int ret_prog[MAX_CALL_FRAMES];
3609 /* protect against potential stack overflow that might happen when
3610 * bpf2bpf calls get combined with tailcalls. Limit the caller's stack
3611 * depth for such case down to 256 so that the worst case scenario
3612 * would result in 8k stack size (32 which is tailcall limit * 256 =
3615 * To get the idea what might happen, see an example:
3616 * func1 -> sub rsp, 128
3617 * subfunc1 -> sub rsp, 256
3618 * tailcall1 -> add rsp, 256
3619 * func2 -> sub rsp, 192 (total stack size = 128 + 192 = 320)
3620 * subfunc2 -> sub rsp, 64
3621 * subfunc22 -> sub rsp, 128
3622 * tailcall2 -> add rsp, 128
3623 * func3 -> sub rsp, 32 (total stack size 128 + 192 + 64 + 32 = 416)
3625 * tailcall will unwind the current stack frame but it will not get rid
3626 * of caller's stack as shown on the example above.
3628 if (idx && subprog[idx].has_tail_call && depth >= 256) {
3630 "tail_calls are not allowed when call stack of previous frames is %d bytes. Too large\n",
3634 /* round up to 32-bytes, since this is granularity
3635 * of interpreter stack size
3637 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
3638 if (depth > MAX_BPF_STACK) {
3639 verbose(env, "combined stack size of %d calls is %d. Too large\n",
3644 subprog_end = subprog[idx + 1].start;
3645 for (; i < subprog_end; i++) {
3646 if (!bpf_pseudo_call(insn + i) && !bpf_pseudo_func(insn + i))
3648 /* remember insn and function to return to */
3649 ret_insn[frame] = i + 1;
3650 ret_prog[frame] = idx;
3652 /* find the callee */
3653 i = i + insn[i].imm + 1;
3654 idx = find_subprog(env, i);
3656 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3661 if (subprog[idx].has_tail_call)
3662 tail_call_reachable = true;
3665 if (frame >= MAX_CALL_FRAMES) {
3666 verbose(env, "the call stack of %d frames is too deep !\n",
3672 /* if tail call got detected across bpf2bpf calls then mark each of the
3673 * currently present subprog frames as tail call reachable subprogs;
3674 * this info will be utilized by JIT so that we will be preserving the
3675 * tail call counter throughout bpf2bpf calls combined with tailcalls
3677 if (tail_call_reachable)
3678 for (j = 0; j < frame; j++)
3679 subprog[ret_prog[j]].tail_call_reachable = true;
3681 /* end of for() loop means the last insn of the 'subprog'
3682 * was reached. Doesn't matter whether it was JA or EXIT
3686 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
3688 i = ret_insn[frame];
3689 idx = ret_prog[frame];
3693 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
3694 static int get_callee_stack_depth(struct bpf_verifier_env *env,
3695 const struct bpf_insn *insn, int idx)
3697 int start = idx + insn->imm + 1, subprog;
3699 subprog = find_subprog(env, start);
3701 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
3705 return env->subprog_info[subprog].stack_depth;
3709 int check_ctx_reg(struct bpf_verifier_env *env,
3710 const struct bpf_reg_state *reg, int regno)
3712 /* Access to ctx or passing it to a helper is only allowed in
3713 * its original, unmodified form.
3717 verbose(env, "dereference of modified ctx ptr R%d off=%d disallowed\n",
3722 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3725 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3726 verbose(env, "variable ctx access var_off=%s disallowed\n", tn_buf);
3733 static int __check_buffer_access(struct bpf_verifier_env *env,
3734 const char *buf_info,
3735 const struct bpf_reg_state *reg,
3736 int regno, int off, int size)
3740 "R%d invalid %s buffer access: off=%d, size=%d\n",
3741 regno, buf_info, off, size);
3744 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3747 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3749 "R%d invalid variable buffer offset: off=%d, var_off=%s\n",
3750 regno, off, tn_buf);
3757 static int check_tp_buffer_access(struct bpf_verifier_env *env,
3758 const struct bpf_reg_state *reg,
3759 int regno, int off, int size)
3763 err = __check_buffer_access(env, "tracepoint", reg, regno, off, size);
3767 if (off + size > env->prog->aux->max_tp_access)
3768 env->prog->aux->max_tp_access = off + size;
3773 static int check_buffer_access(struct bpf_verifier_env *env,
3774 const struct bpf_reg_state *reg,
3775 int regno, int off, int size,
3776 bool zero_size_allowed,
3777 const char *buf_info,
3782 err = __check_buffer_access(env, buf_info, reg, regno, off, size);
3786 if (off + size > *max_access)
3787 *max_access = off + size;
3792 /* BPF architecture zero extends alu32 ops into 64-bit registesr */
3793 static void zext_32_to_64(struct bpf_reg_state *reg)
3795 reg->var_off = tnum_subreg(reg->var_off);
3796 __reg_assign_32_into_64(reg);
3799 /* truncate register to smaller size (in bytes)
3800 * must be called with size < BPF_REG_SIZE
3802 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
3806 /* clear high bits in bit representation */
3807 reg->var_off = tnum_cast(reg->var_off, size);
3809 /* fix arithmetic bounds */
3810 mask = ((u64)1 << (size * 8)) - 1;
3811 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
3812 reg->umin_value &= mask;
3813 reg->umax_value &= mask;
3815 reg->umin_value = 0;
3816 reg->umax_value = mask;
3818 reg->smin_value = reg->umin_value;
3819 reg->smax_value = reg->umax_value;
3821 /* If size is smaller than 32bit register the 32bit register
3822 * values are also truncated so we push 64-bit bounds into
3823 * 32-bit bounds. Above were truncated < 32-bits already.
3827 __reg_combine_64_into_32(reg);
3830 static bool bpf_map_is_rdonly(const struct bpf_map *map)
3832 return (map->map_flags & BPF_F_RDONLY_PROG) && map->frozen;
3835 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val)
3841 err = map->ops->map_direct_value_addr(map, &addr, off);
3844 ptr = (void *)(long)addr + off;
3848 *val = (u64)*(u8 *)ptr;
3851 *val = (u64)*(u16 *)ptr;
3854 *val = (u64)*(u32 *)ptr;
3865 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
3866 struct bpf_reg_state *regs,
3867 int regno, int off, int size,
3868 enum bpf_access_type atype,
3871 struct bpf_reg_state *reg = regs + regno;
3872 const struct btf_type *t = btf_type_by_id(reg->btf, reg->btf_id);
3873 const char *tname = btf_name_by_offset(reg->btf, t->name_off);
3879 "R%d is ptr_%s invalid negative access: off=%d\n",
3883 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
3886 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3888 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
3889 regno, tname, off, tn_buf);
3893 if (env->ops->btf_struct_access) {
3894 ret = env->ops->btf_struct_access(&env->log, reg->btf, t,
3895 off, size, atype, &btf_id);
3897 if (atype != BPF_READ) {
3898 verbose(env, "only read is supported\n");
3902 ret = btf_struct_access(&env->log, reg->btf, t, off, size,
3909 if (atype == BPF_READ && value_regno >= 0)
3910 mark_btf_ld_reg(env, regs, value_regno, ret, reg->btf, btf_id);
3915 static int check_ptr_to_map_access(struct bpf_verifier_env *env,
3916 struct bpf_reg_state *regs,
3917 int regno, int off, int size,
3918 enum bpf_access_type atype,
3921 struct bpf_reg_state *reg = regs + regno;
3922 struct bpf_map *map = reg->map_ptr;
3923 const struct btf_type *t;
3929 verbose(env, "map_ptr access not supported without CONFIG_DEBUG_INFO_BTF\n");
3933 if (!map->ops->map_btf_id || !*map->ops->map_btf_id) {
3934 verbose(env, "map_ptr access not supported for map type %d\n",
3939 t = btf_type_by_id(btf_vmlinux, *map->ops->map_btf_id);
3940 tname = btf_name_by_offset(btf_vmlinux, t->name_off);
3942 if (!env->allow_ptr_to_map_access) {
3944 "%s access is allowed only to CAP_PERFMON and CAP_SYS_ADMIN\n",
3950 verbose(env, "R%d is %s invalid negative access: off=%d\n",
3955 if (atype != BPF_READ) {
3956 verbose(env, "only read from %s is supported\n", tname);
3960 ret = btf_struct_access(&env->log, btf_vmlinux, t, off, size, atype, &btf_id);
3964 if (value_regno >= 0)
3965 mark_btf_ld_reg(env, regs, value_regno, ret, btf_vmlinux, btf_id);
3970 /* Check that the stack access at the given offset is within bounds. The
3971 * maximum valid offset is -1.
3973 * The minimum valid offset is -MAX_BPF_STACK for writes, and
3974 * -state->allocated_stack for reads.
3976 static int check_stack_slot_within_bounds(int off,
3977 struct bpf_func_state *state,
3978 enum bpf_access_type t)
3983 min_valid_off = -MAX_BPF_STACK;
3985 min_valid_off = -state->allocated_stack;
3987 if (off < min_valid_off || off > -1)
3992 /* Check that the stack access at 'regno + off' falls within the maximum stack
3995 * 'off' includes `regno->offset`, but not its dynamic part (if any).
3997 static int check_stack_access_within_bounds(
3998 struct bpf_verifier_env *env,
3999 int regno, int off, int access_size,
4000 enum stack_access_src src, enum bpf_access_type type)
4002 struct bpf_reg_state *regs = cur_regs(env);
4003 struct bpf_reg_state *reg = regs + regno;
4004 struct bpf_func_state *state = func(env, reg);
4005 int min_off, max_off;
4009 if (src == ACCESS_HELPER)
4010 /* We don't know if helpers are reading or writing (or both). */
4011 err_extra = " indirect access to";
4012 else if (type == BPF_READ)
4013 err_extra = " read from";
4015 err_extra = " write to";
4017 if (tnum_is_const(reg->var_off)) {
4018 min_off = reg->var_off.value + off;
4019 if (access_size > 0)
4020 max_off = min_off + access_size - 1;
4024 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
4025 reg->smin_value <= -BPF_MAX_VAR_OFF) {
4026 verbose(env, "invalid unbounded variable-offset%s stack R%d\n",
4030 min_off = reg->smin_value + off;
4031 if (access_size > 0)
4032 max_off = reg->smax_value + off + access_size - 1;
4037 err = check_stack_slot_within_bounds(min_off, state, type);
4039 err = check_stack_slot_within_bounds(max_off, state, type);
4042 if (tnum_is_const(reg->var_off)) {
4043 verbose(env, "invalid%s stack R%d off=%d size=%d\n",
4044 err_extra, regno, off, access_size);
4048 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4049 verbose(env, "invalid variable-offset%s stack R%d var_off=%s size=%d\n",
4050 err_extra, regno, tn_buf, access_size);
4056 /* check whether memory at (regno + off) is accessible for t = (read | write)
4057 * if t==write, value_regno is a register which value is stored into memory
4058 * if t==read, value_regno is a register which will receive the value from memory
4059 * if t==write && value_regno==-1, some unknown value is stored into memory
4060 * if t==read && value_regno==-1, don't care what we read from memory
4062 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
4063 int off, int bpf_size, enum bpf_access_type t,
4064 int value_regno, bool strict_alignment_once)
4066 struct bpf_reg_state *regs = cur_regs(env);
4067 struct bpf_reg_state *reg = regs + regno;
4068 struct bpf_func_state *state;
4071 size = bpf_size_to_bytes(bpf_size);
4075 /* alignment checks will add in reg->off themselves */
4076 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
4080 /* for access checks, reg->off is just part of off */
4083 if (reg->type == PTR_TO_MAP_KEY) {
4084 if (t == BPF_WRITE) {
4085 verbose(env, "write to change key R%d not allowed\n", regno);
4089 err = check_mem_region_access(env, regno, off, size,
4090 reg->map_ptr->key_size, false);
4093 if (value_regno >= 0)
4094 mark_reg_unknown(env, regs, value_regno);
4095 } else if (reg->type == PTR_TO_MAP_VALUE) {
4096 if (t == BPF_WRITE && value_regno >= 0 &&
4097 is_pointer_value(env, value_regno)) {
4098 verbose(env, "R%d leaks addr into map\n", value_regno);
4101 err = check_map_access_type(env, regno, off, size, t);
4104 err = check_map_access(env, regno, off, size, false);
4105 if (!err && t == BPF_READ && value_regno >= 0) {
4106 struct bpf_map *map = reg->map_ptr;
4108 /* if map is read-only, track its contents as scalars */
4109 if (tnum_is_const(reg->var_off) &&
4110 bpf_map_is_rdonly(map) &&
4111 map->ops->map_direct_value_addr) {
4112 int map_off = off + reg->var_off.value;
4115 err = bpf_map_direct_read(map, map_off, size,
4120 regs[value_regno].type = SCALAR_VALUE;
4121 __mark_reg_known(®s[value_regno], val);
4123 mark_reg_unknown(env, regs, value_regno);
4126 } else if (reg->type == PTR_TO_MEM) {
4127 if (t == BPF_WRITE && value_regno >= 0 &&
4128 is_pointer_value(env, value_regno)) {
4129 verbose(env, "R%d leaks addr into mem\n", value_regno);
4132 err = check_mem_region_access(env, regno, off, size,
4133 reg->mem_size, false);
4134 if (!err && t == BPF_READ && value_regno >= 0)
4135 mark_reg_unknown(env, regs, value_regno);
4136 } else if (reg->type == PTR_TO_CTX) {
4137 enum bpf_reg_type reg_type = SCALAR_VALUE;
4138 struct btf *btf = NULL;
4141 if (t == BPF_WRITE && value_regno >= 0 &&
4142 is_pointer_value(env, value_regno)) {
4143 verbose(env, "R%d leaks addr into ctx\n", value_regno);
4147 err = check_ctx_reg(env, reg, regno);
4151 err = check_ctx_access(env, insn_idx, off, size, t, ®_type, &btf, &btf_id);
4153 verbose_linfo(env, insn_idx, "; ");
4154 if (!err && t == BPF_READ && value_regno >= 0) {
4155 /* ctx access returns either a scalar, or a
4156 * PTR_TO_PACKET[_META,_END]. In the latter
4157 * case, we know the offset is zero.
4159 if (reg_type == SCALAR_VALUE) {
4160 mark_reg_unknown(env, regs, value_regno);
4162 mark_reg_known_zero(env, regs,
4164 if (reg_type_may_be_null(reg_type))
4165 regs[value_regno].id = ++env->id_gen;
4166 /* A load of ctx field could have different
4167 * actual load size with the one encoded in the
4168 * insn. When the dst is PTR, it is for sure not
4171 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
4172 if (reg_type == PTR_TO_BTF_ID ||
4173 reg_type == PTR_TO_BTF_ID_OR_NULL) {
4174 regs[value_regno].btf = btf;
4175 regs[value_regno].btf_id = btf_id;
4178 regs[value_regno].type = reg_type;
4181 } else if (reg->type == PTR_TO_STACK) {
4182 /* Basic bounds checks. */
4183 err = check_stack_access_within_bounds(env, regno, off, size, ACCESS_DIRECT, t);
4187 state = func(env, reg);
4188 err = update_stack_depth(env, state, off);
4193 err = check_stack_read(env, regno, off, size,
4196 err = check_stack_write(env, regno, off, size,
4197 value_regno, insn_idx);
4198 } else if (reg_is_pkt_pointer(reg)) {
4199 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
4200 verbose(env, "cannot write into packet\n");
4203 if (t == BPF_WRITE && value_regno >= 0 &&
4204 is_pointer_value(env, value_regno)) {
4205 verbose(env, "R%d leaks addr into packet\n",
4209 err = check_packet_access(env, regno, off, size, false);
4210 if (!err && t == BPF_READ && value_regno >= 0)
4211 mark_reg_unknown(env, regs, value_regno);
4212 } else if (reg->type == PTR_TO_FLOW_KEYS) {
4213 if (t == BPF_WRITE && value_regno >= 0 &&
4214 is_pointer_value(env, value_regno)) {
4215 verbose(env, "R%d leaks addr into flow keys\n",
4220 err = check_flow_keys_access(env, off, size);
4221 if (!err && t == BPF_READ && value_regno >= 0)
4222 mark_reg_unknown(env, regs, value_regno);
4223 } else if (type_is_sk_pointer(reg->type)) {
4224 if (t == BPF_WRITE) {
4225 verbose(env, "R%d cannot write into %s\n",
4226 regno, reg_type_str[reg->type]);
4229 err = check_sock_access(env, insn_idx, regno, off, size, t);
4230 if (!err && value_regno >= 0)
4231 mark_reg_unknown(env, regs, value_regno);
4232 } else if (reg->type == PTR_TO_TP_BUFFER) {
4233 err = check_tp_buffer_access(env, reg, regno, off, size);
4234 if (!err && t == BPF_READ && value_regno >= 0)
4235 mark_reg_unknown(env, regs, value_regno);
4236 } else if (reg->type == PTR_TO_BTF_ID) {
4237 err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
4239 } else if (reg->type == CONST_PTR_TO_MAP) {
4240 err = check_ptr_to_map_access(env, regs, regno, off, size, t,
4242 } else if (reg->type == PTR_TO_RDONLY_BUF) {
4243 if (t == BPF_WRITE) {
4244 verbose(env, "R%d cannot write into %s\n",
4245 regno, reg_type_str[reg->type]);
4248 err = check_buffer_access(env, reg, regno, off, size, false,
4250 &env->prog->aux->max_rdonly_access);
4251 if (!err && value_regno >= 0)
4252 mark_reg_unknown(env, regs, value_regno);
4253 } else if (reg->type == PTR_TO_RDWR_BUF) {
4254 err = check_buffer_access(env, reg, regno, off, size, false,
4256 &env->prog->aux->max_rdwr_access);
4257 if (!err && t == BPF_READ && value_regno >= 0)
4258 mark_reg_unknown(env, regs, value_regno);
4260 verbose(env, "R%d invalid mem access '%s'\n", regno,
4261 reg_type_str[reg->type]);
4265 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
4266 regs[value_regno].type == SCALAR_VALUE) {
4267 /* b/h/w load zero-extends, mark upper bits as known 0 */
4268 coerce_reg_to_size(®s[value_regno], size);
4273 static int check_atomic(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
4278 switch (insn->imm) {
4280 case BPF_ADD | BPF_FETCH:
4282 case BPF_AND | BPF_FETCH:
4284 case BPF_OR | BPF_FETCH:
4286 case BPF_XOR | BPF_FETCH:
4291 verbose(env, "BPF_ATOMIC uses invalid atomic opcode %02x\n", insn->imm);
4295 if (BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) {
4296 verbose(env, "invalid atomic operand size\n");
4300 /* check src1 operand */
4301 err = check_reg_arg(env, insn->src_reg, SRC_OP);
4305 /* check src2 operand */
4306 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
4310 if (insn->imm == BPF_CMPXCHG) {
4311 /* Check comparison of R0 with memory location */
4312 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
4317 if (is_pointer_value(env, insn->src_reg)) {
4318 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
4322 if (is_ctx_reg(env, insn->dst_reg) ||
4323 is_pkt_reg(env, insn->dst_reg) ||
4324 is_flow_key_reg(env, insn->dst_reg) ||
4325 is_sk_reg(env, insn->dst_reg)) {
4326 verbose(env, "BPF_ATOMIC stores into R%d %s is not allowed\n",
4328 reg_type_str[reg_state(env, insn->dst_reg)->type]);
4332 if (insn->imm & BPF_FETCH) {
4333 if (insn->imm == BPF_CMPXCHG)
4334 load_reg = BPF_REG_0;
4336 load_reg = insn->src_reg;
4338 /* check and record load of old value */
4339 err = check_reg_arg(env, load_reg, DST_OP);
4343 /* This instruction accesses a memory location but doesn't
4344 * actually load it into a register.
4349 /* check whether we can read the memory */
4350 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4351 BPF_SIZE(insn->code), BPF_READ, load_reg, true);
4355 /* check whether we can write into the same memory */
4356 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
4357 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
4364 /* When register 'regno' is used to read the stack (either directly or through
4365 * a helper function) make sure that it's within stack boundary and, depending
4366 * on the access type, that all elements of the stack are initialized.
4368 * 'off' includes 'regno->off', but not its dynamic part (if any).
4370 * All registers that have been spilled on the stack in the slots within the
4371 * read offsets are marked as read.
4373 static int check_stack_range_initialized(
4374 struct bpf_verifier_env *env, int regno, int off,
4375 int access_size, bool zero_size_allowed,
4376 enum stack_access_src type, struct bpf_call_arg_meta *meta)
4378 struct bpf_reg_state *reg = reg_state(env, regno);
4379 struct bpf_func_state *state = func(env, reg);
4380 int err, min_off, max_off, i, j, slot, spi;
4381 char *err_extra = type == ACCESS_HELPER ? " indirect" : "";
4382 enum bpf_access_type bounds_check_type;
4383 /* Some accesses can write anything into the stack, others are
4386 bool clobber = false;
4388 if (access_size == 0 && !zero_size_allowed) {
4389 verbose(env, "invalid zero-sized read\n");
4393 if (type == ACCESS_HELPER) {
4394 /* The bounds checks for writes are more permissive than for
4395 * reads. However, if raw_mode is not set, we'll do extra
4398 bounds_check_type = BPF_WRITE;
4401 bounds_check_type = BPF_READ;
4403 err = check_stack_access_within_bounds(env, regno, off, access_size,
4404 type, bounds_check_type);
4409 if (tnum_is_const(reg->var_off)) {
4410 min_off = max_off = reg->var_off.value + off;
4412 /* Variable offset is prohibited for unprivileged mode for
4413 * simplicity since it requires corresponding support in
4414 * Spectre masking for stack ALU.
4415 * See also retrieve_ptr_limit().
4417 if (!env->bypass_spec_v1) {
4420 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4421 verbose(env, "R%d%s variable offset stack access prohibited for !root, var_off=%s\n",
4422 regno, err_extra, tn_buf);
4425 /* Only initialized buffer on stack is allowed to be accessed
4426 * with variable offset. With uninitialized buffer it's hard to
4427 * guarantee that whole memory is marked as initialized on
4428 * helper return since specific bounds are unknown what may
4429 * cause uninitialized stack leaking.
4431 if (meta && meta->raw_mode)
4434 min_off = reg->smin_value + off;
4435 max_off = reg->smax_value + off;
4438 if (meta && meta->raw_mode) {
4439 meta->access_size = access_size;
4440 meta->regno = regno;
4444 for (i = min_off; i < max_off + access_size; i++) {
4448 spi = slot / BPF_REG_SIZE;
4449 if (state->allocated_stack <= slot)
4451 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
4452 if (*stype == STACK_MISC)
4454 if (*stype == STACK_ZERO) {
4456 /* helper can write anything into the stack */
4457 *stype = STACK_MISC;
4462 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
4463 state->stack[spi].spilled_ptr.type == PTR_TO_BTF_ID)
4466 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
4467 (state->stack[spi].spilled_ptr.type == SCALAR_VALUE ||
4468 env->allow_ptr_leaks)) {
4470 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
4471 for (j = 0; j < BPF_REG_SIZE; j++)
4472 state->stack[spi].slot_type[j] = STACK_MISC;
4478 if (tnum_is_const(reg->var_off)) {
4479 verbose(env, "invalid%s read from stack R%d off %d+%d size %d\n",
4480 err_extra, regno, min_off, i - min_off, access_size);
4484 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
4485 verbose(env, "invalid%s read from stack R%d var_off %s+%d size %d\n",
4486 err_extra, regno, tn_buf, i - min_off, access_size);
4490 /* reading any byte out of 8-byte 'spill_slot' will cause
4491 * the whole slot to be marked as 'read'
4493 mark_reg_read(env, &state->stack[spi].spilled_ptr,
4494 state->stack[spi].spilled_ptr.parent,
4497 return update_stack_depth(env, state, min_off);
4500 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
4501 int access_size, bool zero_size_allowed,
4502 struct bpf_call_arg_meta *meta)
4504 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4506 switch (reg->type) {
4508 case PTR_TO_PACKET_META:
4509 return check_packet_access(env, regno, reg->off, access_size,
4511 case PTR_TO_MAP_KEY:
4512 return check_mem_region_access(env, regno, reg->off, access_size,
4513 reg->map_ptr->key_size, false);
4514 case PTR_TO_MAP_VALUE:
4515 if (check_map_access_type(env, regno, reg->off, access_size,
4516 meta && meta->raw_mode ? BPF_WRITE :
4519 return check_map_access(env, regno, reg->off, access_size,
4522 return check_mem_region_access(env, regno, reg->off,
4523 access_size, reg->mem_size,
4525 case PTR_TO_RDONLY_BUF:
4526 if (meta && meta->raw_mode)
4528 return check_buffer_access(env, reg, regno, reg->off,
4529 access_size, zero_size_allowed,
4531 &env->prog->aux->max_rdonly_access);
4532 case PTR_TO_RDWR_BUF:
4533 return check_buffer_access(env, reg, regno, reg->off,
4534 access_size, zero_size_allowed,
4536 &env->prog->aux->max_rdwr_access);
4538 return check_stack_range_initialized(
4540 regno, reg->off, access_size,
4541 zero_size_allowed, ACCESS_HELPER, meta);
4542 default: /* scalar_value or invalid ptr */
4543 /* Allow zero-byte read from NULL, regardless of pointer type */
4544 if (zero_size_allowed && access_size == 0 &&
4545 register_is_null(reg))
4548 verbose(env, "R%d type=%s expected=%s\n", regno,
4549 reg_type_str[reg->type],
4550 reg_type_str[PTR_TO_STACK]);
4555 int check_mem_reg(struct bpf_verifier_env *env, struct bpf_reg_state *reg,
4556 u32 regno, u32 mem_size)
4558 if (register_is_null(reg))
4561 if (reg_type_may_be_null(reg->type)) {
4562 /* Assuming that the register contains a value check if the memory
4563 * access is safe. Temporarily save and restore the register's state as
4564 * the conversion shouldn't be visible to a caller.
4566 const struct bpf_reg_state saved_reg = *reg;
4569 mark_ptr_not_null_reg(reg);
4570 rv = check_helper_mem_access(env, regno, mem_size, true, NULL);
4575 return check_helper_mem_access(env, regno, mem_size, true, NULL);
4578 /* Implementation details:
4579 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL
4580 * Two bpf_map_lookups (even with the same key) will have different reg->id.
4581 * For traditional PTR_TO_MAP_VALUE the verifier clears reg->id after
4582 * value_or_null->value transition, since the verifier only cares about
4583 * the range of access to valid map value pointer and doesn't care about actual
4584 * address of the map element.
4585 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
4586 * reg->id > 0 after value_or_null->value transition. By doing so
4587 * two bpf_map_lookups will be considered two different pointers that
4588 * point to different bpf_spin_locks.
4589 * The verifier allows taking only one bpf_spin_lock at a time to avoid
4591 * Since only one bpf_spin_lock is allowed the checks are simpler than
4592 * reg_is_refcounted() logic. The verifier needs to remember only
4593 * one spin_lock instead of array of acquired_refs.
4594 * cur_state->active_spin_lock remembers which map value element got locked
4595 * and clears it after bpf_spin_unlock.
4597 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
4600 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4601 struct bpf_verifier_state *cur = env->cur_state;
4602 bool is_const = tnum_is_const(reg->var_off);
4603 struct bpf_map *map = reg->map_ptr;
4604 u64 val = reg->var_off.value;
4608 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
4614 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
4618 if (!map_value_has_spin_lock(map)) {
4619 if (map->spin_lock_off == -E2BIG)
4621 "map '%s' has more than one 'struct bpf_spin_lock'\n",
4623 else if (map->spin_lock_off == -ENOENT)
4625 "map '%s' doesn't have 'struct bpf_spin_lock'\n",
4629 "map '%s' is not a struct type or bpf_spin_lock is mangled\n",
4633 if (map->spin_lock_off != val + reg->off) {
4634 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock'\n",
4639 if (cur->active_spin_lock) {
4641 "Locking two bpf_spin_locks are not allowed\n");
4644 cur->active_spin_lock = reg->id;
4646 if (!cur->active_spin_lock) {
4647 verbose(env, "bpf_spin_unlock without taking a lock\n");
4650 if (cur->active_spin_lock != reg->id) {
4651 verbose(env, "bpf_spin_unlock of different lock\n");
4654 cur->active_spin_lock = 0;
4659 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
4661 return type == ARG_PTR_TO_MEM ||
4662 type == ARG_PTR_TO_MEM_OR_NULL ||
4663 type == ARG_PTR_TO_UNINIT_MEM;
4666 static bool arg_type_is_mem_size(enum bpf_arg_type type)
4668 return type == ARG_CONST_SIZE ||
4669 type == ARG_CONST_SIZE_OR_ZERO;
4672 static bool arg_type_is_alloc_size(enum bpf_arg_type type)
4674 return type == ARG_CONST_ALLOC_SIZE_OR_ZERO;
4677 static bool arg_type_is_int_ptr(enum bpf_arg_type type)
4679 return type == ARG_PTR_TO_INT ||
4680 type == ARG_PTR_TO_LONG;
4683 static int int_ptr_type_to_size(enum bpf_arg_type type)
4685 if (type == ARG_PTR_TO_INT)
4687 else if (type == ARG_PTR_TO_LONG)
4693 static int resolve_map_arg_type(struct bpf_verifier_env *env,
4694 const struct bpf_call_arg_meta *meta,
4695 enum bpf_arg_type *arg_type)
4697 if (!meta->map_ptr) {
4698 /* kernel subsystem misconfigured verifier */
4699 verbose(env, "invalid map_ptr to access map->type\n");
4703 switch (meta->map_ptr->map_type) {
4704 case BPF_MAP_TYPE_SOCKMAP:
4705 case BPF_MAP_TYPE_SOCKHASH:
4706 if (*arg_type == ARG_PTR_TO_MAP_VALUE) {
4707 *arg_type = ARG_PTR_TO_BTF_ID_SOCK_COMMON;
4709 verbose(env, "invalid arg_type for sockmap/sockhash\n");
4720 struct bpf_reg_types {
4721 const enum bpf_reg_type types[10];
4725 static const struct bpf_reg_types map_key_value_types = {
4735 static const struct bpf_reg_types sock_types = {
4745 static const struct bpf_reg_types btf_id_sock_common_types = {
4753 .btf_id = &btf_sock_ids[BTF_SOCK_TYPE_SOCK_COMMON],
4757 static const struct bpf_reg_types mem_types = {
4770 static const struct bpf_reg_types int_ptr_types = {
4780 static const struct bpf_reg_types fullsock_types = { .types = { PTR_TO_SOCKET } };
4781 static const struct bpf_reg_types scalar_types = { .types = { SCALAR_VALUE } };
4782 static const struct bpf_reg_types context_types = { .types = { PTR_TO_CTX } };
4783 static const struct bpf_reg_types alloc_mem_types = { .types = { PTR_TO_MEM } };
4784 static const struct bpf_reg_types const_map_ptr_types = { .types = { CONST_PTR_TO_MAP } };
4785 static const struct bpf_reg_types btf_ptr_types = { .types = { PTR_TO_BTF_ID } };
4786 static const struct bpf_reg_types spin_lock_types = { .types = { PTR_TO_MAP_VALUE } };
4787 static const struct bpf_reg_types percpu_btf_ptr_types = { .types = { PTR_TO_PERCPU_BTF_ID } };
4788 static const struct bpf_reg_types func_ptr_types = { .types = { PTR_TO_FUNC } };
4789 static const struct bpf_reg_types stack_ptr_types = { .types = { PTR_TO_STACK } };
4790 static const struct bpf_reg_types const_str_ptr_types = { .types = { PTR_TO_MAP_VALUE } };
4792 static const struct bpf_reg_types *compatible_reg_types[__BPF_ARG_TYPE_MAX] = {
4793 [ARG_PTR_TO_MAP_KEY] = &map_key_value_types,
4794 [ARG_PTR_TO_MAP_VALUE] = &map_key_value_types,
4795 [ARG_PTR_TO_UNINIT_MAP_VALUE] = &map_key_value_types,
4796 [ARG_PTR_TO_MAP_VALUE_OR_NULL] = &map_key_value_types,
4797 [ARG_CONST_SIZE] = &scalar_types,
4798 [ARG_CONST_SIZE_OR_ZERO] = &scalar_types,
4799 [ARG_CONST_ALLOC_SIZE_OR_ZERO] = &scalar_types,
4800 [ARG_CONST_MAP_PTR] = &const_map_ptr_types,
4801 [ARG_PTR_TO_CTX] = &context_types,
4802 [ARG_PTR_TO_CTX_OR_NULL] = &context_types,
4803 [ARG_PTR_TO_SOCK_COMMON] = &sock_types,
4805 [ARG_PTR_TO_BTF_ID_SOCK_COMMON] = &btf_id_sock_common_types,
4807 [ARG_PTR_TO_SOCKET] = &fullsock_types,
4808 [ARG_PTR_TO_SOCKET_OR_NULL] = &fullsock_types,
4809 [ARG_PTR_TO_BTF_ID] = &btf_ptr_types,
4810 [ARG_PTR_TO_SPIN_LOCK] = &spin_lock_types,
4811 [ARG_PTR_TO_MEM] = &mem_types,
4812 [ARG_PTR_TO_MEM_OR_NULL] = &mem_types,
4813 [ARG_PTR_TO_UNINIT_MEM] = &mem_types,
4814 [ARG_PTR_TO_ALLOC_MEM] = &alloc_mem_types,
4815 [ARG_PTR_TO_ALLOC_MEM_OR_NULL] = &alloc_mem_types,
4816 [ARG_PTR_TO_INT] = &int_ptr_types,
4817 [ARG_PTR_TO_LONG] = &int_ptr_types,
4818 [ARG_PTR_TO_PERCPU_BTF_ID] = &percpu_btf_ptr_types,
4819 [ARG_PTR_TO_FUNC] = &func_ptr_types,
4820 [ARG_PTR_TO_STACK_OR_NULL] = &stack_ptr_types,
4821 [ARG_PTR_TO_CONST_STR] = &const_str_ptr_types,
4824 static int check_reg_type(struct bpf_verifier_env *env, u32 regno,
4825 enum bpf_arg_type arg_type,
4826 const u32 *arg_btf_id)
4828 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4829 enum bpf_reg_type expected, type = reg->type;
4830 const struct bpf_reg_types *compatible;
4833 compatible = compatible_reg_types[arg_type];
4835 verbose(env, "verifier internal error: unsupported arg type %d\n", arg_type);
4839 for (i = 0; i < ARRAY_SIZE(compatible->types); i++) {
4840 expected = compatible->types[i];
4841 if (expected == NOT_INIT)
4844 if (type == expected)
4848 verbose(env, "R%d type=%s expected=", regno, reg_type_str[type]);
4849 for (j = 0; j + 1 < i; j++)
4850 verbose(env, "%s, ", reg_type_str[compatible->types[j]]);
4851 verbose(env, "%s\n", reg_type_str[compatible->types[j]]);
4855 if (type == PTR_TO_BTF_ID) {
4857 if (!compatible->btf_id) {
4858 verbose(env, "verifier internal error: missing arg compatible BTF ID\n");
4861 arg_btf_id = compatible->btf_id;
4864 if (!btf_struct_ids_match(&env->log, reg->btf, reg->btf_id, reg->off,
4865 btf_vmlinux, *arg_btf_id)) {
4866 verbose(env, "R%d is of type %s but %s is expected\n",
4867 regno, kernel_type_name(reg->btf, reg->btf_id),
4868 kernel_type_name(btf_vmlinux, *arg_btf_id));
4872 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
4873 verbose(env, "R%d is a pointer to in-kernel struct with non-zero offset\n",
4882 static int check_func_arg(struct bpf_verifier_env *env, u32 arg,
4883 struct bpf_call_arg_meta *meta,
4884 const struct bpf_func_proto *fn)
4886 u32 regno = BPF_REG_1 + arg;
4887 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
4888 enum bpf_arg_type arg_type = fn->arg_type[arg];
4889 enum bpf_reg_type type = reg->type;
4892 if (arg_type == ARG_DONTCARE)
4895 err = check_reg_arg(env, regno, SRC_OP);
4899 if (arg_type == ARG_ANYTHING) {
4900 if (is_pointer_value(env, regno)) {
4901 verbose(env, "R%d leaks addr into helper function\n",
4908 if (type_is_pkt_pointer(type) &&
4909 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
4910 verbose(env, "helper access to the packet is not allowed\n");
4914 if (arg_type == ARG_PTR_TO_MAP_VALUE ||
4915 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE ||
4916 arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL) {
4917 err = resolve_map_arg_type(env, meta, &arg_type);
4922 if (register_is_null(reg) && arg_type_may_be_null(arg_type))
4923 /* A NULL register has a SCALAR_VALUE type, so skip
4926 goto skip_type_check;
4928 err = check_reg_type(env, regno, arg_type, fn->arg_btf_id[arg]);
4932 if (type == PTR_TO_CTX) {
4933 err = check_ctx_reg(env, reg, regno);
4939 if (reg->ref_obj_id) {
4940 if (meta->ref_obj_id) {
4941 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
4942 regno, reg->ref_obj_id,
4946 meta->ref_obj_id = reg->ref_obj_id;
4949 if (arg_type == ARG_CONST_MAP_PTR) {
4950 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
4951 meta->map_ptr = reg->map_ptr;
4952 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
4953 /* bpf_map_xxx(..., map_ptr, ..., key) call:
4954 * check that [key, key + map->key_size) are within
4955 * stack limits and initialized
4957 if (!meta->map_ptr) {
4958 /* in function declaration map_ptr must come before
4959 * map_key, so that it's verified and known before
4960 * we have to check map_key here. Otherwise it means
4961 * that kernel subsystem misconfigured verifier
4963 verbose(env, "invalid map_ptr to access map->key\n");
4966 err = check_helper_mem_access(env, regno,
4967 meta->map_ptr->key_size, false,
4969 } else if (arg_type == ARG_PTR_TO_MAP_VALUE ||
4970 (arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL &&
4971 !register_is_null(reg)) ||
4972 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
4973 /* bpf_map_xxx(..., map_ptr, ..., value) call:
4974 * check [value, value + map->value_size) validity
4976 if (!meta->map_ptr) {
4977 /* kernel subsystem misconfigured verifier */
4978 verbose(env, "invalid map_ptr to access map->value\n");
4981 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE);
4982 err = check_helper_mem_access(env, regno,
4983 meta->map_ptr->value_size, false,
4985 } else if (arg_type == ARG_PTR_TO_PERCPU_BTF_ID) {
4987 verbose(env, "Helper has invalid btf_id in R%d\n", regno);
4990 meta->ret_btf = reg->btf;
4991 meta->ret_btf_id = reg->btf_id;
4992 } else if (arg_type == ARG_PTR_TO_SPIN_LOCK) {
4993 if (meta->func_id == BPF_FUNC_spin_lock) {
4994 if (process_spin_lock(env, regno, true))
4996 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
4997 if (process_spin_lock(env, regno, false))
5000 verbose(env, "verifier internal error\n");
5003 } else if (arg_type == ARG_PTR_TO_FUNC) {
5004 meta->subprogno = reg->subprogno;
5005 } else if (arg_type_is_mem_ptr(arg_type)) {
5006 /* The access to this pointer is only checked when we hit the
5007 * next is_mem_size argument below.
5009 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MEM);
5010 } else if (arg_type_is_mem_size(arg_type)) {
5011 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
5013 /* This is used to refine r0 return value bounds for helpers
5014 * that enforce this value as an upper bound on return values.
5015 * See do_refine_retval_range() for helpers that can refine
5016 * the return value. C type of helper is u32 so we pull register
5017 * bound from umax_value however, if negative verifier errors
5018 * out. Only upper bounds can be learned because retval is an
5019 * int type and negative retvals are allowed.
5021 meta->msize_max_value = reg->umax_value;
5023 /* The register is SCALAR_VALUE; the access check
5024 * happens using its boundaries.
5026 if (!tnum_is_const(reg->var_off))
5027 /* For unprivileged variable accesses, disable raw
5028 * mode so that the program is required to
5029 * initialize all the memory that the helper could
5030 * just partially fill up.
5034 if (reg->smin_value < 0) {
5035 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
5040 if (reg->umin_value == 0) {
5041 err = check_helper_mem_access(env, regno - 1, 0,
5048 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
5049 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
5053 err = check_helper_mem_access(env, regno - 1,
5055 zero_size_allowed, meta);
5057 err = mark_chain_precision(env, regno);
5058 } else if (arg_type_is_alloc_size(arg_type)) {
5059 if (!tnum_is_const(reg->var_off)) {
5060 verbose(env, "R%d is not a known constant'\n",
5064 meta->mem_size = reg->var_off.value;
5065 } else if (arg_type_is_int_ptr(arg_type)) {
5066 int size = int_ptr_type_to_size(arg_type);
5068 err = check_helper_mem_access(env, regno, size, false, meta);
5071 err = check_ptr_alignment(env, reg, 0, size, true);
5072 } else if (arg_type == ARG_PTR_TO_CONST_STR) {
5073 struct bpf_map *map = reg->map_ptr;
5078 if (!bpf_map_is_rdonly(map)) {
5079 verbose(env, "R%d does not point to a readonly map'\n", regno);
5083 if (!tnum_is_const(reg->var_off)) {
5084 verbose(env, "R%d is not a constant address'\n", regno);
5088 if (!map->ops->map_direct_value_addr) {
5089 verbose(env, "no direct value access support for this map type\n");
5093 err = check_map_access(env, regno, reg->off,
5094 map->value_size - reg->off, false);
5098 map_off = reg->off + reg->var_off.value;
5099 err = map->ops->map_direct_value_addr(map, &map_addr, map_off);
5101 verbose(env, "direct value access on string failed\n");
5105 str_ptr = (char *)(long)(map_addr);
5106 if (!strnchr(str_ptr + map_off, map->value_size - map_off, 0)) {
5107 verbose(env, "string is not zero-terminated\n");
5115 static bool may_update_sockmap(struct bpf_verifier_env *env, int func_id)
5117 enum bpf_attach_type eatype = env->prog->expected_attach_type;
5118 enum bpf_prog_type type = resolve_prog_type(env->prog);
5120 if (func_id != BPF_FUNC_map_update_elem)
5123 /* It's not possible to get access to a locked struct sock in these
5124 * contexts, so updating is safe.
5127 case BPF_PROG_TYPE_TRACING:
5128 if (eatype == BPF_TRACE_ITER)
5131 case BPF_PROG_TYPE_SOCKET_FILTER:
5132 case BPF_PROG_TYPE_SCHED_CLS:
5133 case BPF_PROG_TYPE_SCHED_ACT:
5134 case BPF_PROG_TYPE_XDP:
5135 case BPF_PROG_TYPE_SK_REUSEPORT:
5136 case BPF_PROG_TYPE_FLOW_DISSECTOR:
5137 case BPF_PROG_TYPE_SK_LOOKUP:
5143 verbose(env, "cannot update sockmap in this context\n");
5147 static bool allow_tail_call_in_subprogs(struct bpf_verifier_env *env)
5149 return env->prog->jit_requested && IS_ENABLED(CONFIG_X86_64);
5152 static int check_map_func_compatibility(struct bpf_verifier_env *env,
5153 struct bpf_map *map, int func_id)
5158 /* We need a two way check, first is from map perspective ... */
5159 switch (map->map_type) {
5160 case BPF_MAP_TYPE_PROG_ARRAY:
5161 if (func_id != BPF_FUNC_tail_call)
5164 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
5165 if (func_id != BPF_FUNC_perf_event_read &&
5166 func_id != BPF_FUNC_perf_event_output &&
5167 func_id != BPF_FUNC_skb_output &&
5168 func_id != BPF_FUNC_perf_event_read_value &&
5169 func_id != BPF_FUNC_xdp_output)
5172 case BPF_MAP_TYPE_RINGBUF:
5173 if (func_id != BPF_FUNC_ringbuf_output &&
5174 func_id != BPF_FUNC_ringbuf_reserve &&
5175 func_id != BPF_FUNC_ringbuf_submit &&
5176 func_id != BPF_FUNC_ringbuf_discard &&
5177 func_id != BPF_FUNC_ringbuf_query)
5180 case BPF_MAP_TYPE_STACK_TRACE:
5181 if (func_id != BPF_FUNC_get_stackid)
5184 case BPF_MAP_TYPE_CGROUP_ARRAY:
5185 if (func_id != BPF_FUNC_skb_under_cgroup &&
5186 func_id != BPF_FUNC_current_task_under_cgroup)
5189 case BPF_MAP_TYPE_CGROUP_STORAGE:
5190 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
5191 if (func_id != BPF_FUNC_get_local_storage)
5194 case BPF_MAP_TYPE_DEVMAP:
5195 case BPF_MAP_TYPE_DEVMAP_HASH:
5196 if (func_id != BPF_FUNC_redirect_map &&
5197 func_id != BPF_FUNC_map_lookup_elem)
5200 /* Restrict bpf side of cpumap and xskmap, open when use-cases
5203 case BPF_MAP_TYPE_CPUMAP:
5204 if (func_id != BPF_FUNC_redirect_map)
5207 case BPF_MAP_TYPE_XSKMAP:
5208 if (func_id != BPF_FUNC_redirect_map &&
5209 func_id != BPF_FUNC_map_lookup_elem)
5212 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
5213 case BPF_MAP_TYPE_HASH_OF_MAPS:
5214 if (func_id != BPF_FUNC_map_lookup_elem)
5217 case BPF_MAP_TYPE_SOCKMAP:
5218 if (func_id != BPF_FUNC_sk_redirect_map &&
5219 func_id != BPF_FUNC_sock_map_update &&
5220 func_id != BPF_FUNC_map_delete_elem &&
5221 func_id != BPF_FUNC_msg_redirect_map &&
5222 func_id != BPF_FUNC_sk_select_reuseport &&
5223 func_id != BPF_FUNC_map_lookup_elem &&
5224 !may_update_sockmap(env, func_id))
5227 case BPF_MAP_TYPE_SOCKHASH:
5228 if (func_id != BPF_FUNC_sk_redirect_hash &&
5229 func_id != BPF_FUNC_sock_hash_update &&
5230 func_id != BPF_FUNC_map_delete_elem &&
5231 func_id != BPF_FUNC_msg_redirect_hash &&
5232 func_id != BPF_FUNC_sk_select_reuseport &&
5233 func_id != BPF_FUNC_map_lookup_elem &&
5234 !may_update_sockmap(env, func_id))
5237 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
5238 if (func_id != BPF_FUNC_sk_select_reuseport)
5241 case BPF_MAP_TYPE_QUEUE:
5242 case BPF_MAP_TYPE_STACK:
5243 if (func_id != BPF_FUNC_map_peek_elem &&
5244 func_id != BPF_FUNC_map_pop_elem &&
5245 func_id != BPF_FUNC_map_push_elem)
5248 case BPF_MAP_TYPE_SK_STORAGE:
5249 if (func_id != BPF_FUNC_sk_storage_get &&
5250 func_id != BPF_FUNC_sk_storage_delete)
5253 case BPF_MAP_TYPE_INODE_STORAGE:
5254 if (func_id != BPF_FUNC_inode_storage_get &&
5255 func_id != BPF_FUNC_inode_storage_delete)
5258 case BPF_MAP_TYPE_TASK_STORAGE:
5259 if (func_id != BPF_FUNC_task_storage_get &&
5260 func_id != BPF_FUNC_task_storage_delete)
5267 /* ... and second from the function itself. */
5269 case BPF_FUNC_tail_call:
5270 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
5272 if (env->subprog_cnt > 1 && !allow_tail_call_in_subprogs(env)) {
5273 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
5277 case BPF_FUNC_perf_event_read:
5278 case BPF_FUNC_perf_event_output:
5279 case BPF_FUNC_perf_event_read_value:
5280 case BPF_FUNC_skb_output:
5281 case BPF_FUNC_xdp_output:
5282 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
5285 case BPF_FUNC_get_stackid:
5286 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
5289 case BPF_FUNC_current_task_under_cgroup:
5290 case BPF_FUNC_skb_under_cgroup:
5291 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
5294 case BPF_FUNC_redirect_map:
5295 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
5296 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
5297 map->map_type != BPF_MAP_TYPE_CPUMAP &&
5298 map->map_type != BPF_MAP_TYPE_XSKMAP)
5301 case BPF_FUNC_sk_redirect_map:
5302 case BPF_FUNC_msg_redirect_map:
5303 case BPF_FUNC_sock_map_update:
5304 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
5307 case BPF_FUNC_sk_redirect_hash:
5308 case BPF_FUNC_msg_redirect_hash:
5309 case BPF_FUNC_sock_hash_update:
5310 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
5313 case BPF_FUNC_get_local_storage:
5314 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
5315 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
5318 case BPF_FUNC_sk_select_reuseport:
5319 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
5320 map->map_type != BPF_MAP_TYPE_SOCKMAP &&
5321 map->map_type != BPF_MAP_TYPE_SOCKHASH)
5324 case BPF_FUNC_map_peek_elem:
5325 case BPF_FUNC_map_pop_elem:
5326 case BPF_FUNC_map_push_elem:
5327 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
5328 map->map_type != BPF_MAP_TYPE_STACK)
5331 case BPF_FUNC_sk_storage_get:
5332 case BPF_FUNC_sk_storage_delete:
5333 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
5336 case BPF_FUNC_inode_storage_get:
5337 case BPF_FUNC_inode_storage_delete:
5338 if (map->map_type != BPF_MAP_TYPE_INODE_STORAGE)
5341 case BPF_FUNC_task_storage_get:
5342 case BPF_FUNC_task_storage_delete:
5343 if (map->map_type != BPF_MAP_TYPE_TASK_STORAGE)
5352 verbose(env, "cannot pass map_type %d into func %s#%d\n",
5353 map->map_type, func_id_name(func_id), func_id);
5357 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
5361 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
5363 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
5365 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
5367 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
5369 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
5372 /* We only support one arg being in raw mode at the moment,
5373 * which is sufficient for the helper functions we have
5379 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
5380 enum bpf_arg_type arg_next)
5382 return (arg_type_is_mem_ptr(arg_curr) &&
5383 !arg_type_is_mem_size(arg_next)) ||
5384 (!arg_type_is_mem_ptr(arg_curr) &&
5385 arg_type_is_mem_size(arg_next));
5388 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
5390 /* bpf_xxx(..., buf, len) call will access 'len'
5391 * bytes from memory 'buf'. Both arg types need
5392 * to be paired, so make sure there's no buggy
5393 * helper function specification.
5395 if (arg_type_is_mem_size(fn->arg1_type) ||
5396 arg_type_is_mem_ptr(fn->arg5_type) ||
5397 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
5398 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
5399 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
5400 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
5406 static bool check_refcount_ok(const struct bpf_func_proto *fn, int func_id)
5410 if (arg_type_may_be_refcounted(fn->arg1_type))
5412 if (arg_type_may_be_refcounted(fn->arg2_type))
5414 if (arg_type_may_be_refcounted(fn->arg3_type))
5416 if (arg_type_may_be_refcounted(fn->arg4_type))
5418 if (arg_type_may_be_refcounted(fn->arg5_type))
5421 /* A reference acquiring function cannot acquire
5422 * another refcounted ptr.
5424 if (may_be_acquire_function(func_id) && count)
5427 /* We only support one arg being unreferenced at the moment,
5428 * which is sufficient for the helper functions we have right now.
5433 static bool check_btf_id_ok(const struct bpf_func_proto *fn)
5437 for (i = 0; i < ARRAY_SIZE(fn->arg_type); i++) {
5438 if (fn->arg_type[i] == ARG_PTR_TO_BTF_ID && !fn->arg_btf_id[i])
5441 if (fn->arg_type[i] != ARG_PTR_TO_BTF_ID && fn->arg_btf_id[i])
5448 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
5450 return check_raw_mode_ok(fn) &&
5451 check_arg_pair_ok(fn) &&
5452 check_btf_id_ok(fn) &&
5453 check_refcount_ok(fn, func_id) ? 0 : -EINVAL;
5456 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
5457 * are now invalid, so turn them into unknown SCALAR_VALUE.
5459 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
5460 struct bpf_func_state *state)
5462 struct bpf_reg_state *regs = state->regs, *reg;
5465 for (i = 0; i < MAX_BPF_REG; i++)
5466 if (reg_is_pkt_pointer_any(®s[i]))
5467 mark_reg_unknown(env, regs, i);
5469 bpf_for_each_spilled_reg(i, state, reg) {
5472 if (reg_is_pkt_pointer_any(reg))
5473 __mark_reg_unknown(env, reg);
5477 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
5479 struct bpf_verifier_state *vstate = env->cur_state;
5482 for (i = 0; i <= vstate->curframe; i++)
5483 __clear_all_pkt_pointers(env, vstate->frame[i]);
5488 BEYOND_PKT_END = -2,
5491 static void mark_pkt_end(struct bpf_verifier_state *vstate, int regn, bool range_open)
5493 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5494 struct bpf_reg_state *reg = &state->regs[regn];
5496 if (reg->type != PTR_TO_PACKET)
5497 /* PTR_TO_PACKET_META is not supported yet */
5500 /* The 'reg' is pkt > pkt_end or pkt >= pkt_end.
5501 * How far beyond pkt_end it goes is unknown.
5502 * if (!range_open) it's the case of pkt >= pkt_end
5503 * if (range_open) it's the case of pkt > pkt_end
5504 * hence this pointer is at least 1 byte bigger than pkt_end
5507 reg->range = BEYOND_PKT_END;
5509 reg->range = AT_PKT_END;
5512 static void release_reg_references(struct bpf_verifier_env *env,
5513 struct bpf_func_state *state,
5516 struct bpf_reg_state *regs = state->regs, *reg;
5519 for (i = 0; i < MAX_BPF_REG; i++)
5520 if (regs[i].ref_obj_id == ref_obj_id)
5521 mark_reg_unknown(env, regs, i);
5523 bpf_for_each_spilled_reg(i, state, reg) {
5526 if (reg->ref_obj_id == ref_obj_id)
5527 __mark_reg_unknown(env, reg);
5531 /* The pointer with the specified id has released its reference to kernel
5532 * resources. Identify all copies of the same pointer and clear the reference.
5534 static int release_reference(struct bpf_verifier_env *env,
5537 struct bpf_verifier_state *vstate = env->cur_state;
5541 err = release_reference_state(cur_func(env), ref_obj_id);
5545 for (i = 0; i <= vstate->curframe; i++)
5546 release_reg_references(env, vstate->frame[i], ref_obj_id);
5551 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
5552 struct bpf_reg_state *regs)
5556 /* after the call registers r0 - r5 were scratched */
5557 for (i = 0; i < CALLER_SAVED_REGS; i++) {
5558 mark_reg_not_init(env, regs, caller_saved[i]);
5559 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
5563 typedef int (*set_callee_state_fn)(struct bpf_verifier_env *env,
5564 struct bpf_func_state *caller,
5565 struct bpf_func_state *callee,
5568 static int __check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
5569 int *insn_idx, int subprog,
5570 set_callee_state_fn set_callee_state_cb)
5572 struct bpf_verifier_state *state = env->cur_state;
5573 struct bpf_func_info_aux *func_info_aux;
5574 struct bpf_func_state *caller, *callee;
5576 bool is_global = false;
5578 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
5579 verbose(env, "the call stack of %d frames is too deep\n",
5580 state->curframe + 2);
5584 caller = state->frame[state->curframe];
5585 if (state->frame[state->curframe + 1]) {
5586 verbose(env, "verifier bug. Frame %d already allocated\n",
5587 state->curframe + 1);
5591 func_info_aux = env->prog->aux->func_info_aux;
5593 is_global = func_info_aux[subprog].linkage == BTF_FUNC_GLOBAL;
5594 err = btf_check_subprog_arg_match(env, subprog, caller->regs);
5599 verbose(env, "Caller passes invalid args into func#%d\n",
5603 if (env->log.level & BPF_LOG_LEVEL)
5605 "Func#%d is global and valid. Skipping.\n",
5607 clear_caller_saved_regs(env, caller->regs);
5609 /* All global functions return a 64-bit SCALAR_VALUE */
5610 mark_reg_unknown(env, caller->regs, BPF_REG_0);
5611 caller->regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
5613 /* continue with next insn after call */
5618 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
5621 state->frame[state->curframe + 1] = callee;
5623 /* callee cannot access r0, r6 - r9 for reading and has to write
5624 * into its own stack before reading from it.
5625 * callee can read/write into caller's stack
5627 init_func_state(env, callee,
5628 /* remember the callsite, it will be used by bpf_exit */
5629 *insn_idx /* callsite */,
5630 state->curframe + 1 /* frameno within this callchain */,
5631 subprog /* subprog number within this prog */);
5633 /* Transfer references to the callee */
5634 err = copy_reference_state(callee, caller);
5638 err = set_callee_state_cb(env, caller, callee, *insn_idx);
5642 clear_caller_saved_regs(env, caller->regs);
5644 /* only increment it after check_reg_arg() finished */
5647 /* and go analyze first insn of the callee */
5648 *insn_idx = env->subprog_info[subprog].start - 1;
5650 if (env->log.level & BPF_LOG_LEVEL) {
5651 verbose(env, "caller:\n");
5652 print_verifier_state(env, caller);
5653 verbose(env, "callee:\n");
5654 print_verifier_state(env, callee);
5659 int map_set_for_each_callback_args(struct bpf_verifier_env *env,
5660 struct bpf_func_state *caller,
5661 struct bpf_func_state *callee)
5663 /* bpf_for_each_map_elem(struct bpf_map *map, void *callback_fn,
5664 * void *callback_ctx, u64 flags);
5665 * callback_fn(struct bpf_map *map, void *key, void *value,
5666 * void *callback_ctx);
5668 callee->regs[BPF_REG_1] = caller->regs[BPF_REG_1];
5670 callee->regs[BPF_REG_2].type = PTR_TO_MAP_KEY;
5671 __mark_reg_known_zero(&callee->regs[BPF_REG_2]);
5672 callee->regs[BPF_REG_2].map_ptr = caller->regs[BPF_REG_1].map_ptr;
5674 callee->regs[BPF_REG_3].type = PTR_TO_MAP_VALUE;
5675 __mark_reg_known_zero(&callee->regs[BPF_REG_3]);
5676 callee->regs[BPF_REG_3].map_ptr = caller->regs[BPF_REG_1].map_ptr;
5678 /* pointer to stack or null */
5679 callee->regs[BPF_REG_4] = caller->regs[BPF_REG_3];
5682 __mark_reg_not_init(env, &callee->regs[BPF_REG_5]);
5686 static int set_callee_state(struct bpf_verifier_env *env,
5687 struct bpf_func_state *caller,
5688 struct bpf_func_state *callee, int insn_idx)
5692 /* copy r1 - r5 args that callee can access. The copy includes parent
5693 * pointers, which connects us up to the liveness chain
5695 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
5696 callee->regs[i] = caller->regs[i];
5700 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
5703 int subprog, target_insn;
5705 target_insn = *insn_idx + insn->imm + 1;
5706 subprog = find_subprog(env, target_insn);
5708 verbose(env, "verifier bug. No program starts at insn %d\n",
5713 return __check_func_call(env, insn, insn_idx, subprog, set_callee_state);
5716 static int set_map_elem_callback_state(struct bpf_verifier_env *env,
5717 struct bpf_func_state *caller,
5718 struct bpf_func_state *callee,
5721 struct bpf_insn_aux_data *insn_aux = &env->insn_aux_data[insn_idx];
5722 struct bpf_map *map;
5725 if (bpf_map_ptr_poisoned(insn_aux)) {
5726 verbose(env, "tail_call abusing map_ptr\n");
5730 map = BPF_MAP_PTR(insn_aux->map_ptr_state);
5731 if (!map->ops->map_set_for_each_callback_args ||
5732 !map->ops->map_for_each_callback) {
5733 verbose(env, "callback function not allowed for map\n");
5737 err = map->ops->map_set_for_each_callback_args(env, caller, callee);
5741 callee->in_callback_fn = true;
5745 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
5747 struct bpf_verifier_state *state = env->cur_state;
5748 struct bpf_func_state *caller, *callee;
5749 struct bpf_reg_state *r0;
5752 callee = state->frame[state->curframe];
5753 r0 = &callee->regs[BPF_REG_0];
5754 if (r0->type == PTR_TO_STACK) {
5755 /* technically it's ok to return caller's stack pointer
5756 * (or caller's caller's pointer) back to the caller,
5757 * since these pointers are valid. Only current stack
5758 * pointer will be invalid as soon as function exits,
5759 * but let's be conservative
5761 verbose(env, "cannot return stack pointer to the caller\n");
5766 caller = state->frame[state->curframe];
5767 if (callee->in_callback_fn) {
5768 /* enforce R0 return value range [0, 1]. */
5769 struct tnum range = tnum_range(0, 1);
5771 if (r0->type != SCALAR_VALUE) {
5772 verbose(env, "R0 not a scalar value\n");
5775 if (!tnum_in(range, r0->var_off)) {
5776 verbose_invalid_scalar(env, r0, &range, "callback return", "R0");
5780 /* return to the caller whatever r0 had in the callee */
5781 caller->regs[BPF_REG_0] = *r0;
5784 /* Transfer references to the caller */
5785 err = copy_reference_state(caller, callee);
5789 *insn_idx = callee->callsite + 1;
5790 if (env->log.level & BPF_LOG_LEVEL) {
5791 verbose(env, "returning from callee:\n");
5792 print_verifier_state(env, callee);
5793 verbose(env, "to caller at %d:\n", *insn_idx);
5794 print_verifier_state(env, caller);
5796 /* clear everything in the callee */
5797 free_func_state(callee);
5798 state->frame[state->curframe + 1] = NULL;
5802 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
5804 struct bpf_call_arg_meta *meta)
5806 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
5808 if (ret_type != RET_INTEGER ||
5809 (func_id != BPF_FUNC_get_stack &&
5810 func_id != BPF_FUNC_get_task_stack &&
5811 func_id != BPF_FUNC_probe_read_str &&
5812 func_id != BPF_FUNC_probe_read_kernel_str &&
5813 func_id != BPF_FUNC_probe_read_user_str))
5816 ret_reg->smax_value = meta->msize_max_value;
5817 ret_reg->s32_max_value = meta->msize_max_value;
5818 ret_reg->smin_value = -MAX_ERRNO;
5819 ret_reg->s32_min_value = -MAX_ERRNO;
5820 __reg_deduce_bounds(ret_reg);
5821 __reg_bound_offset(ret_reg);
5822 __update_reg_bounds(ret_reg);
5826 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
5827 int func_id, int insn_idx)
5829 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
5830 struct bpf_map *map = meta->map_ptr;
5832 if (func_id != BPF_FUNC_tail_call &&
5833 func_id != BPF_FUNC_map_lookup_elem &&
5834 func_id != BPF_FUNC_map_update_elem &&
5835 func_id != BPF_FUNC_map_delete_elem &&
5836 func_id != BPF_FUNC_map_push_elem &&
5837 func_id != BPF_FUNC_map_pop_elem &&
5838 func_id != BPF_FUNC_map_peek_elem &&
5839 func_id != BPF_FUNC_for_each_map_elem &&
5840 func_id != BPF_FUNC_redirect_map)
5844 verbose(env, "kernel subsystem misconfigured verifier\n");
5848 /* In case of read-only, some additional restrictions
5849 * need to be applied in order to prevent altering the
5850 * state of the map from program side.
5852 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
5853 (func_id == BPF_FUNC_map_delete_elem ||
5854 func_id == BPF_FUNC_map_update_elem ||
5855 func_id == BPF_FUNC_map_push_elem ||
5856 func_id == BPF_FUNC_map_pop_elem)) {
5857 verbose(env, "write into map forbidden\n");
5861 if (!BPF_MAP_PTR(aux->map_ptr_state))
5862 bpf_map_ptr_store(aux, meta->map_ptr,
5863 !meta->map_ptr->bypass_spec_v1);
5864 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
5865 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
5866 !meta->map_ptr->bypass_spec_v1);
5871 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
5872 int func_id, int insn_idx)
5874 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
5875 struct bpf_reg_state *regs = cur_regs(env), *reg;
5876 struct bpf_map *map = meta->map_ptr;
5881 if (func_id != BPF_FUNC_tail_call)
5883 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
5884 verbose(env, "kernel subsystem misconfigured verifier\n");
5888 range = tnum_range(0, map->max_entries - 1);
5889 reg = ®s[BPF_REG_3];
5891 if (!register_is_const(reg) || !tnum_in(range, reg->var_off)) {
5892 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
5896 err = mark_chain_precision(env, BPF_REG_3);
5900 val = reg->var_off.value;
5901 if (bpf_map_key_unseen(aux))
5902 bpf_map_key_store(aux, val);
5903 else if (!bpf_map_key_poisoned(aux) &&
5904 bpf_map_key_immediate(aux) != val)
5905 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
5909 static int check_reference_leak(struct bpf_verifier_env *env)
5911 struct bpf_func_state *state = cur_func(env);
5914 for (i = 0; i < state->acquired_refs; i++) {
5915 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
5916 state->refs[i].id, state->refs[i].insn_idx);
5918 return state->acquired_refs ? -EINVAL : 0;
5921 static int check_bpf_snprintf_call(struct bpf_verifier_env *env,
5922 struct bpf_reg_state *regs)
5924 struct bpf_reg_state *fmt_reg = ®s[BPF_REG_3];
5925 struct bpf_reg_state *data_len_reg = ®s[BPF_REG_5];
5926 struct bpf_map *fmt_map = fmt_reg->map_ptr;
5927 int err, fmt_map_off, num_args;
5931 /* data must be an array of u64 */
5932 if (data_len_reg->var_off.value % 8)
5934 num_args = data_len_reg->var_off.value / 8;
5936 /* fmt being ARG_PTR_TO_CONST_STR guarantees that var_off is const
5937 * and map_direct_value_addr is set.
5939 fmt_map_off = fmt_reg->off + fmt_reg->var_off.value;
5940 err = fmt_map->ops->map_direct_value_addr(fmt_map, &fmt_addr,
5943 verbose(env, "verifier bug\n");
5946 fmt = (char *)(long)fmt_addr + fmt_map_off;
5948 /* We are also guaranteed that fmt+fmt_map_off is NULL terminated, we
5949 * can focus on validating the format specifiers.
5951 err = bpf_bprintf_prepare(fmt, UINT_MAX, NULL, NULL, num_args);
5953 verbose(env, "Invalid format string\n");
5958 static int check_helper_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
5961 const struct bpf_func_proto *fn = NULL;
5962 struct bpf_reg_state *regs;
5963 struct bpf_call_arg_meta meta;
5964 int insn_idx = *insn_idx_p;
5966 int i, err, func_id;
5968 /* find function prototype */
5969 func_id = insn->imm;
5970 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
5971 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
5976 if (env->ops->get_func_proto)
5977 fn = env->ops->get_func_proto(func_id, env->prog);
5979 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
5984 /* eBPF programs must be GPL compatible to use GPL-ed functions */
5985 if (!env->prog->gpl_compatible && fn->gpl_only) {
5986 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
5990 if (fn->allowed && !fn->allowed(env->prog)) {
5991 verbose(env, "helper call is not allowed in probe\n");
5995 /* With LD_ABS/IND some JITs save/restore skb from r1. */
5996 changes_data = bpf_helper_changes_pkt_data(fn->func);
5997 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
5998 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
5999 func_id_name(func_id), func_id);
6003 memset(&meta, 0, sizeof(meta));
6004 meta.pkt_access = fn->pkt_access;
6006 err = check_func_proto(fn, func_id);
6008 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
6009 func_id_name(func_id), func_id);
6013 meta.func_id = func_id;
6015 for (i = 0; i < MAX_BPF_FUNC_REG_ARGS; i++) {
6016 err = check_func_arg(env, i, &meta, fn);
6021 err = record_func_map(env, &meta, func_id, insn_idx);
6025 err = record_func_key(env, &meta, func_id, insn_idx);
6029 /* Mark slots with STACK_MISC in case of raw mode, stack offset
6030 * is inferred from register state.
6032 for (i = 0; i < meta.access_size; i++) {
6033 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
6034 BPF_WRITE, -1, false);
6039 if (func_id == BPF_FUNC_tail_call) {
6040 err = check_reference_leak(env);
6042 verbose(env, "tail_call would lead to reference leak\n");
6045 } else if (is_release_function(func_id)) {
6046 err = release_reference(env, meta.ref_obj_id);
6048 verbose(env, "func %s#%d reference has not been acquired before\n",
6049 func_id_name(func_id), func_id);
6054 regs = cur_regs(env);
6056 /* check that flags argument in get_local_storage(map, flags) is 0,
6057 * this is required because get_local_storage() can't return an error.
6059 if (func_id == BPF_FUNC_get_local_storage &&
6060 !register_is_null(®s[BPF_REG_2])) {
6061 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
6065 if (func_id == BPF_FUNC_for_each_map_elem) {
6066 err = __check_func_call(env, insn, insn_idx_p, meta.subprogno,
6067 set_map_elem_callback_state);
6072 if (func_id == BPF_FUNC_snprintf) {
6073 err = check_bpf_snprintf_call(env, regs);
6078 /* reset caller saved regs */
6079 for (i = 0; i < CALLER_SAVED_REGS; i++) {
6080 mark_reg_not_init(env, regs, caller_saved[i]);
6081 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
6084 /* helper call returns 64-bit value. */
6085 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
6087 /* update return register (already marked as written above) */
6088 if (fn->ret_type == RET_INTEGER) {
6089 /* sets type to SCALAR_VALUE */
6090 mark_reg_unknown(env, regs, BPF_REG_0);
6091 } else if (fn->ret_type == RET_VOID) {
6092 regs[BPF_REG_0].type = NOT_INIT;
6093 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL ||
6094 fn->ret_type == RET_PTR_TO_MAP_VALUE) {
6095 /* There is no offset yet applied, variable or fixed */
6096 mark_reg_known_zero(env, regs, BPF_REG_0);
6097 /* remember map_ptr, so that check_map_access()
6098 * can check 'value_size' boundary of memory access
6099 * to map element returned from bpf_map_lookup_elem()
6101 if (meta.map_ptr == NULL) {
6103 "kernel subsystem misconfigured verifier\n");
6106 regs[BPF_REG_0].map_ptr = meta.map_ptr;
6107 if (fn->ret_type == RET_PTR_TO_MAP_VALUE) {
6108 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE;
6109 if (map_value_has_spin_lock(meta.map_ptr))
6110 regs[BPF_REG_0].id = ++env->id_gen;
6112 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
6114 } else if (fn->ret_type == RET_PTR_TO_SOCKET_OR_NULL) {
6115 mark_reg_known_zero(env, regs, BPF_REG_0);
6116 regs[BPF_REG_0].type = PTR_TO_SOCKET_OR_NULL;
6117 } else if (fn->ret_type == RET_PTR_TO_SOCK_COMMON_OR_NULL) {
6118 mark_reg_known_zero(env, regs, BPF_REG_0);
6119 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON_OR_NULL;
6120 } else if (fn->ret_type == RET_PTR_TO_TCP_SOCK_OR_NULL) {
6121 mark_reg_known_zero(env, regs, BPF_REG_0);
6122 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK_OR_NULL;
6123 } else if (fn->ret_type == RET_PTR_TO_ALLOC_MEM_OR_NULL) {
6124 mark_reg_known_zero(env, regs, BPF_REG_0);
6125 regs[BPF_REG_0].type = PTR_TO_MEM_OR_NULL;
6126 regs[BPF_REG_0].mem_size = meta.mem_size;
6127 } else if (fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID_OR_NULL ||
6128 fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID) {
6129 const struct btf_type *t;
6131 mark_reg_known_zero(env, regs, BPF_REG_0);
6132 t = btf_type_skip_modifiers(meta.ret_btf, meta.ret_btf_id, NULL);
6133 if (!btf_type_is_struct(t)) {
6135 const struct btf_type *ret;
6138 /* resolve the type size of ksym. */
6139 ret = btf_resolve_size(meta.ret_btf, t, &tsize);
6141 tname = btf_name_by_offset(meta.ret_btf, t->name_off);
6142 verbose(env, "unable to resolve the size of type '%s': %ld\n",
6143 tname, PTR_ERR(ret));
6146 regs[BPF_REG_0].type =
6147 fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID ?
6148 PTR_TO_MEM : PTR_TO_MEM_OR_NULL;
6149 regs[BPF_REG_0].mem_size = tsize;
6151 regs[BPF_REG_0].type =
6152 fn->ret_type == RET_PTR_TO_MEM_OR_BTF_ID ?
6153 PTR_TO_BTF_ID : PTR_TO_BTF_ID_OR_NULL;
6154 regs[BPF_REG_0].btf = meta.ret_btf;
6155 regs[BPF_REG_0].btf_id = meta.ret_btf_id;
6157 } else if (fn->ret_type == RET_PTR_TO_BTF_ID_OR_NULL ||
6158 fn->ret_type == RET_PTR_TO_BTF_ID) {
6161 mark_reg_known_zero(env, regs, BPF_REG_0);
6162 regs[BPF_REG_0].type = fn->ret_type == RET_PTR_TO_BTF_ID ?
6164 PTR_TO_BTF_ID_OR_NULL;
6165 ret_btf_id = *fn->ret_btf_id;
6166 if (ret_btf_id == 0) {
6167 verbose(env, "invalid return type %d of func %s#%d\n",
6168 fn->ret_type, func_id_name(func_id), func_id);
6171 /* current BPF helper definitions are only coming from
6172 * built-in code with type IDs from vmlinux BTF
6174 regs[BPF_REG_0].btf = btf_vmlinux;
6175 regs[BPF_REG_0].btf_id = ret_btf_id;
6177 verbose(env, "unknown return type %d of func %s#%d\n",
6178 fn->ret_type, func_id_name(func_id), func_id);
6182 if (reg_type_may_be_null(regs[BPF_REG_0].type))
6183 regs[BPF_REG_0].id = ++env->id_gen;
6185 if (is_ptr_cast_function(func_id)) {
6186 /* For release_reference() */
6187 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
6188 } else if (is_acquire_function(func_id, meta.map_ptr)) {
6189 int id = acquire_reference_state(env, insn_idx);
6193 /* For mark_ptr_or_null_reg() */
6194 regs[BPF_REG_0].id = id;
6195 /* For release_reference() */
6196 regs[BPF_REG_0].ref_obj_id = id;
6199 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
6201 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
6205 if ((func_id == BPF_FUNC_get_stack ||
6206 func_id == BPF_FUNC_get_task_stack) &&
6207 !env->prog->has_callchain_buf) {
6208 const char *err_str;
6210 #ifdef CONFIG_PERF_EVENTS
6211 err = get_callchain_buffers(sysctl_perf_event_max_stack);
6212 err_str = "cannot get callchain buffer for func %s#%d\n";
6215 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
6218 verbose(env, err_str, func_id_name(func_id), func_id);
6222 env->prog->has_callchain_buf = true;
6225 if (func_id == BPF_FUNC_get_stackid || func_id == BPF_FUNC_get_stack)
6226 env->prog->call_get_stack = true;
6229 clear_all_pkt_pointers(env);
6233 /* mark_btf_func_reg_size() is used when the reg size is determined by
6234 * the BTF func_proto's return value size and argument.
6236 static void mark_btf_func_reg_size(struct bpf_verifier_env *env, u32 regno,
6239 struct bpf_reg_state *reg = &cur_regs(env)[regno];
6241 if (regno == BPF_REG_0) {
6242 /* Function return value */
6243 reg->live |= REG_LIVE_WRITTEN;
6244 reg->subreg_def = reg_size == sizeof(u64) ?
6245 DEF_NOT_SUBREG : env->insn_idx + 1;
6247 /* Function argument */
6248 if (reg_size == sizeof(u64)) {
6249 mark_insn_zext(env, reg);
6250 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
6252 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ32);
6257 static int check_kfunc_call(struct bpf_verifier_env *env, struct bpf_insn *insn)
6259 const struct btf_type *t, *func, *func_proto, *ptr_type;
6260 struct bpf_reg_state *regs = cur_regs(env);
6261 const char *func_name, *ptr_type_name;
6262 u32 i, nargs, func_id, ptr_type_id;
6263 const struct btf_param *args;
6266 func_id = insn->imm;
6267 func = btf_type_by_id(btf_vmlinux, func_id);
6268 func_name = btf_name_by_offset(btf_vmlinux, func->name_off);
6269 func_proto = btf_type_by_id(btf_vmlinux, func->type);
6271 if (!env->ops->check_kfunc_call ||
6272 !env->ops->check_kfunc_call(func_id)) {
6273 verbose(env, "calling kernel function %s is not allowed\n",
6278 /* Check the arguments */
6279 err = btf_check_kfunc_arg_match(env, btf_vmlinux, func_id, regs);
6283 for (i = 0; i < CALLER_SAVED_REGS; i++)
6284 mark_reg_not_init(env, regs, caller_saved[i]);
6286 /* Check return type */
6287 t = btf_type_skip_modifiers(btf_vmlinux, func_proto->type, NULL);
6288 if (btf_type_is_scalar(t)) {
6289 mark_reg_unknown(env, regs, BPF_REG_0);
6290 mark_btf_func_reg_size(env, BPF_REG_0, t->size);
6291 } else if (btf_type_is_ptr(t)) {
6292 ptr_type = btf_type_skip_modifiers(btf_vmlinux, t->type,
6294 if (!btf_type_is_struct(ptr_type)) {
6295 ptr_type_name = btf_name_by_offset(btf_vmlinux,
6296 ptr_type->name_off);
6297 verbose(env, "kernel function %s returns pointer type %s %s is not supported\n",
6298 func_name, btf_type_str(ptr_type),
6302 mark_reg_known_zero(env, regs, BPF_REG_0);
6303 regs[BPF_REG_0].btf = btf_vmlinux;
6304 regs[BPF_REG_0].type = PTR_TO_BTF_ID;
6305 regs[BPF_REG_0].btf_id = ptr_type_id;
6306 mark_btf_func_reg_size(env, BPF_REG_0, sizeof(void *));
6307 } /* else { add_kfunc_call() ensures it is btf_type_is_void(t) } */
6309 nargs = btf_type_vlen(func_proto);
6310 args = (const struct btf_param *)(func_proto + 1);
6311 for (i = 0; i < nargs; i++) {
6314 t = btf_type_skip_modifiers(btf_vmlinux, args[i].type, NULL);
6315 if (btf_type_is_ptr(t))
6316 mark_btf_func_reg_size(env, regno, sizeof(void *));
6318 /* scalar. ensured by btf_check_kfunc_arg_match() */
6319 mark_btf_func_reg_size(env, regno, t->size);
6325 static bool signed_add_overflows(s64 a, s64 b)
6327 /* Do the add in u64, where overflow is well-defined */
6328 s64 res = (s64)((u64)a + (u64)b);
6335 static bool signed_add32_overflows(s32 a, s32 b)
6337 /* Do the add in u32, where overflow is well-defined */
6338 s32 res = (s32)((u32)a + (u32)b);
6345 static bool signed_sub_overflows(s64 a, s64 b)
6347 /* Do the sub in u64, where overflow is well-defined */
6348 s64 res = (s64)((u64)a - (u64)b);
6355 static bool signed_sub32_overflows(s32 a, s32 b)
6357 /* Do the sub in u32, where overflow is well-defined */
6358 s32 res = (s32)((u32)a - (u32)b);
6365 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
6366 const struct bpf_reg_state *reg,
6367 enum bpf_reg_type type)
6369 bool known = tnum_is_const(reg->var_off);
6370 s64 val = reg->var_off.value;
6371 s64 smin = reg->smin_value;
6373 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
6374 verbose(env, "math between %s pointer and %lld is not allowed\n",
6375 reg_type_str[type], val);
6379 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
6380 verbose(env, "%s pointer offset %d is not allowed\n",
6381 reg_type_str[type], reg->off);
6385 if (smin == S64_MIN) {
6386 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
6387 reg_type_str[type]);
6391 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
6392 verbose(env, "value %lld makes %s pointer be out of bounds\n",
6393 smin, reg_type_str[type]);
6400 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
6402 return &env->insn_aux_data[env->insn_idx];
6413 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
6414 u32 *alu_limit, bool mask_to_left)
6416 u32 max = 0, ptr_limit = 0;
6418 switch (ptr_reg->type) {
6420 /* Offset 0 is out-of-bounds, but acceptable start for the
6421 * left direction, see BPF_REG_FP. Also, unknown scalar
6422 * offset where we would need to deal with min/max bounds is
6423 * currently prohibited for unprivileged.
6425 max = MAX_BPF_STACK + mask_to_left;
6426 ptr_limit = -(ptr_reg->var_off.value + ptr_reg->off);
6428 case PTR_TO_MAP_VALUE:
6429 max = ptr_reg->map_ptr->value_size;
6430 ptr_limit = (mask_to_left ?
6431 ptr_reg->smin_value :
6432 ptr_reg->umax_value) + ptr_reg->off;
6438 if (ptr_limit >= max)
6439 return REASON_LIMIT;
6440 *alu_limit = ptr_limit;
6444 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
6445 const struct bpf_insn *insn)
6447 return env->bypass_spec_v1 || BPF_SRC(insn->code) == BPF_K;
6450 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
6451 u32 alu_state, u32 alu_limit)
6453 /* If we arrived here from different branches with different
6454 * state or limits to sanitize, then this won't work.
6456 if (aux->alu_state &&
6457 (aux->alu_state != alu_state ||
6458 aux->alu_limit != alu_limit))
6459 return REASON_PATHS;
6461 /* Corresponding fixup done in do_misc_fixups(). */
6462 aux->alu_state = alu_state;
6463 aux->alu_limit = alu_limit;
6467 static int sanitize_val_alu(struct bpf_verifier_env *env,
6468 struct bpf_insn *insn)
6470 struct bpf_insn_aux_data *aux = cur_aux(env);
6472 if (can_skip_alu_sanitation(env, insn))
6475 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
6478 static bool sanitize_needed(u8 opcode)
6480 return opcode == BPF_ADD || opcode == BPF_SUB;
6483 struct bpf_sanitize_info {
6484 struct bpf_insn_aux_data aux;
6488 static struct bpf_verifier_state *
6489 sanitize_speculative_path(struct bpf_verifier_env *env,
6490 const struct bpf_insn *insn,
6491 u32 next_idx, u32 curr_idx)
6493 struct bpf_verifier_state *branch;
6494 struct bpf_reg_state *regs;
6496 branch = push_stack(env, next_idx, curr_idx, true);
6497 if (branch && insn) {
6498 regs = branch->frame[branch->curframe]->regs;
6499 if (BPF_SRC(insn->code) == BPF_K) {
6500 mark_reg_unknown(env, regs, insn->dst_reg);
6501 } else if (BPF_SRC(insn->code) == BPF_X) {
6502 mark_reg_unknown(env, regs, insn->dst_reg);
6503 mark_reg_unknown(env, regs, insn->src_reg);
6509 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
6510 struct bpf_insn *insn,
6511 const struct bpf_reg_state *ptr_reg,
6512 const struct bpf_reg_state *off_reg,
6513 struct bpf_reg_state *dst_reg,
6514 struct bpf_sanitize_info *info,
6515 const bool commit_window)
6517 struct bpf_insn_aux_data *aux = commit_window ? cur_aux(env) : &info->aux;
6518 struct bpf_verifier_state *vstate = env->cur_state;
6519 bool off_is_imm = tnum_is_const(off_reg->var_off);
6520 bool off_is_neg = off_reg->smin_value < 0;
6521 bool ptr_is_dst_reg = ptr_reg == dst_reg;
6522 u8 opcode = BPF_OP(insn->code);
6523 u32 alu_state, alu_limit;
6524 struct bpf_reg_state tmp;
6528 if (can_skip_alu_sanitation(env, insn))
6531 /* We already marked aux for masking from non-speculative
6532 * paths, thus we got here in the first place. We only care
6533 * to explore bad access from here.
6535 if (vstate->speculative)
6538 if (!commit_window) {
6539 if (!tnum_is_const(off_reg->var_off) &&
6540 (off_reg->smin_value < 0) != (off_reg->smax_value < 0))
6541 return REASON_BOUNDS;
6543 info->mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
6544 (opcode == BPF_SUB && !off_is_neg);
6547 err = retrieve_ptr_limit(ptr_reg, &alu_limit, info->mask_to_left);
6551 if (commit_window) {
6552 /* In commit phase we narrow the masking window based on
6553 * the observed pointer move after the simulated operation.
6555 alu_state = info->aux.alu_state;
6556 alu_limit = abs(info->aux.alu_limit - alu_limit);
6558 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
6559 alu_state |= off_is_imm ? BPF_ALU_IMMEDIATE : 0;
6560 alu_state |= ptr_is_dst_reg ?
6561 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
6564 err = update_alu_sanitation_state(aux, alu_state, alu_limit);
6568 /* If we're in commit phase, we're done here given we already
6569 * pushed the truncated dst_reg into the speculative verification
6572 * Also, when register is a known constant, we rewrite register-based
6573 * operation to immediate-based, and thus do not need masking (and as
6574 * a consequence, do not need to simulate the zero-truncation either).
6576 if (commit_window || off_is_imm)
6579 /* Simulate and find potential out-of-bounds access under
6580 * speculative execution from truncation as a result of
6581 * masking when off was not within expected range. If off
6582 * sits in dst, then we temporarily need to move ptr there
6583 * to simulate dst (== 0) +/-= ptr. Needed, for example,
6584 * for cases where we use K-based arithmetic in one direction
6585 * and truncated reg-based in the other in order to explore
6588 if (!ptr_is_dst_reg) {
6590 *dst_reg = *ptr_reg;
6592 ret = sanitize_speculative_path(env, NULL, env->insn_idx + 1,
6594 if (!ptr_is_dst_reg && ret)
6596 return !ret ? REASON_STACK : 0;
6599 static void sanitize_mark_insn_seen(struct bpf_verifier_env *env)
6601 struct bpf_verifier_state *vstate = env->cur_state;
6603 /* If we simulate paths under speculation, we don't update the
6604 * insn as 'seen' such that when we verify unreachable paths in
6605 * the non-speculative domain, sanitize_dead_code() can still
6606 * rewrite/sanitize them.
6608 if (!vstate->speculative)
6609 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
6612 static int sanitize_err(struct bpf_verifier_env *env,
6613 const struct bpf_insn *insn, int reason,
6614 const struct bpf_reg_state *off_reg,
6615 const struct bpf_reg_state *dst_reg)
6617 static const char *err = "pointer arithmetic with it prohibited for !root";
6618 const char *op = BPF_OP(insn->code) == BPF_ADD ? "add" : "sub";
6619 u32 dst = insn->dst_reg, src = insn->src_reg;
6623 verbose(env, "R%d has unknown scalar with mixed signed bounds, %s\n",
6624 off_reg == dst_reg ? dst : src, err);
6627 verbose(env, "R%d has pointer with unsupported alu operation, %s\n",
6628 off_reg == dst_reg ? src : dst, err);
6631 verbose(env, "R%d tried to %s from different maps, paths or scalars, %s\n",
6635 verbose(env, "R%d tried to %s beyond pointer bounds, %s\n",
6639 verbose(env, "R%d could not be pushed for speculative verification, %s\n",
6643 verbose(env, "verifier internal error: unknown reason (%d)\n",
6651 /* check that stack access falls within stack limits and that 'reg' doesn't
6652 * have a variable offset.
6654 * Variable offset is prohibited for unprivileged mode for simplicity since it
6655 * requires corresponding support in Spectre masking for stack ALU. See also
6656 * retrieve_ptr_limit().
6659 * 'off' includes 'reg->off'.
6661 static int check_stack_access_for_ptr_arithmetic(
6662 struct bpf_verifier_env *env,
6664 const struct bpf_reg_state *reg,
6667 if (!tnum_is_const(reg->var_off)) {
6670 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6671 verbose(env, "R%d variable stack access prohibited for !root, var_off=%s off=%d\n",
6672 regno, tn_buf, off);
6676 if (off >= 0 || off < -MAX_BPF_STACK) {
6677 verbose(env, "R%d stack pointer arithmetic goes out of range, "
6678 "prohibited for !root; off=%d\n", regno, off);
6685 static int sanitize_check_bounds(struct bpf_verifier_env *env,
6686 const struct bpf_insn *insn,
6687 const struct bpf_reg_state *dst_reg)
6689 u32 dst = insn->dst_reg;
6691 /* For unprivileged we require that resulting offset must be in bounds
6692 * in order to be able to sanitize access later on.
6694 if (env->bypass_spec_v1)
6697 switch (dst_reg->type) {
6699 if (check_stack_access_for_ptr_arithmetic(env, dst, dst_reg,
6700 dst_reg->off + dst_reg->var_off.value))
6703 case PTR_TO_MAP_VALUE:
6704 if (check_map_access(env, dst, dst_reg->off, 1, false)) {
6705 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
6706 "prohibited for !root\n", dst);
6717 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
6718 * Caller should also handle BPF_MOV case separately.
6719 * If we return -EACCES, caller may want to try again treating pointer as a
6720 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
6722 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
6723 struct bpf_insn *insn,
6724 const struct bpf_reg_state *ptr_reg,
6725 const struct bpf_reg_state *off_reg)
6727 struct bpf_verifier_state *vstate = env->cur_state;
6728 struct bpf_func_state *state = vstate->frame[vstate->curframe];
6729 struct bpf_reg_state *regs = state->regs, *dst_reg;
6730 bool known = tnum_is_const(off_reg->var_off);
6731 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
6732 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
6733 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
6734 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
6735 struct bpf_sanitize_info info = {};
6736 u8 opcode = BPF_OP(insn->code);
6737 u32 dst = insn->dst_reg;
6740 dst_reg = ®s[dst];
6742 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
6743 smin_val > smax_val || umin_val > umax_val) {
6744 /* Taint dst register if offset had invalid bounds derived from
6745 * e.g. dead branches.
6747 __mark_reg_unknown(env, dst_reg);
6751 if (BPF_CLASS(insn->code) != BPF_ALU64) {
6752 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
6753 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
6754 __mark_reg_unknown(env, dst_reg);
6759 "R%d 32-bit pointer arithmetic prohibited\n",
6764 switch (ptr_reg->type) {
6765 case PTR_TO_MAP_VALUE_OR_NULL:
6766 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
6767 dst, reg_type_str[ptr_reg->type]);
6769 case CONST_PTR_TO_MAP:
6770 /* smin_val represents the known value */
6771 if (known && smin_val == 0 && opcode == BPF_ADD)
6774 case PTR_TO_PACKET_END:
6776 case PTR_TO_SOCKET_OR_NULL:
6777 case PTR_TO_SOCK_COMMON:
6778 case PTR_TO_SOCK_COMMON_OR_NULL:
6779 case PTR_TO_TCP_SOCK:
6780 case PTR_TO_TCP_SOCK_OR_NULL:
6781 case PTR_TO_XDP_SOCK:
6782 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
6783 dst, reg_type_str[ptr_reg->type]);
6789 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
6790 * The id may be overwritten later if we create a new variable offset.
6792 dst_reg->type = ptr_reg->type;
6793 dst_reg->id = ptr_reg->id;
6795 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
6796 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
6799 /* pointer types do not carry 32-bit bounds at the moment. */
6800 __mark_reg32_unbounded(dst_reg);
6802 if (sanitize_needed(opcode)) {
6803 ret = sanitize_ptr_alu(env, insn, ptr_reg, off_reg, dst_reg,
6806 return sanitize_err(env, insn, ret, off_reg, dst_reg);
6811 /* We can take a fixed offset as long as it doesn't overflow
6812 * the s32 'off' field
6814 if (known && (ptr_reg->off + smin_val ==
6815 (s64)(s32)(ptr_reg->off + smin_val))) {
6816 /* pointer += K. Accumulate it into fixed offset */
6817 dst_reg->smin_value = smin_ptr;
6818 dst_reg->smax_value = smax_ptr;
6819 dst_reg->umin_value = umin_ptr;
6820 dst_reg->umax_value = umax_ptr;
6821 dst_reg->var_off = ptr_reg->var_off;
6822 dst_reg->off = ptr_reg->off + smin_val;
6823 dst_reg->raw = ptr_reg->raw;
6826 /* A new variable offset is created. Note that off_reg->off
6827 * == 0, since it's a scalar.
6828 * dst_reg gets the pointer type and since some positive
6829 * integer value was added to the pointer, give it a new 'id'
6830 * if it's a PTR_TO_PACKET.
6831 * this creates a new 'base' pointer, off_reg (variable) gets
6832 * added into the variable offset, and we copy the fixed offset
6835 if (signed_add_overflows(smin_ptr, smin_val) ||
6836 signed_add_overflows(smax_ptr, smax_val)) {
6837 dst_reg->smin_value = S64_MIN;
6838 dst_reg->smax_value = S64_MAX;
6840 dst_reg->smin_value = smin_ptr + smin_val;
6841 dst_reg->smax_value = smax_ptr + smax_val;
6843 if (umin_ptr + umin_val < umin_ptr ||
6844 umax_ptr + umax_val < umax_ptr) {
6845 dst_reg->umin_value = 0;
6846 dst_reg->umax_value = U64_MAX;
6848 dst_reg->umin_value = umin_ptr + umin_val;
6849 dst_reg->umax_value = umax_ptr + umax_val;
6851 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
6852 dst_reg->off = ptr_reg->off;
6853 dst_reg->raw = ptr_reg->raw;
6854 if (reg_is_pkt_pointer(ptr_reg)) {
6855 dst_reg->id = ++env->id_gen;
6856 /* something was added to pkt_ptr, set range to zero */
6857 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
6861 if (dst_reg == off_reg) {
6862 /* scalar -= pointer. Creates an unknown scalar */
6863 verbose(env, "R%d tried to subtract pointer from scalar\n",
6867 /* We don't allow subtraction from FP, because (according to
6868 * test_verifier.c test "invalid fp arithmetic", JITs might not
6869 * be able to deal with it.
6871 if (ptr_reg->type == PTR_TO_STACK) {
6872 verbose(env, "R%d subtraction from stack pointer prohibited\n",
6876 if (known && (ptr_reg->off - smin_val ==
6877 (s64)(s32)(ptr_reg->off - smin_val))) {
6878 /* pointer -= K. Subtract it from fixed offset */
6879 dst_reg->smin_value = smin_ptr;
6880 dst_reg->smax_value = smax_ptr;
6881 dst_reg->umin_value = umin_ptr;
6882 dst_reg->umax_value = umax_ptr;
6883 dst_reg->var_off = ptr_reg->var_off;
6884 dst_reg->id = ptr_reg->id;
6885 dst_reg->off = ptr_reg->off - smin_val;
6886 dst_reg->raw = ptr_reg->raw;
6889 /* A new variable offset is created. If the subtrahend is known
6890 * nonnegative, then any reg->range we had before is still good.
6892 if (signed_sub_overflows(smin_ptr, smax_val) ||
6893 signed_sub_overflows(smax_ptr, smin_val)) {
6894 /* Overflow possible, we know nothing */
6895 dst_reg->smin_value = S64_MIN;
6896 dst_reg->smax_value = S64_MAX;
6898 dst_reg->smin_value = smin_ptr - smax_val;
6899 dst_reg->smax_value = smax_ptr - smin_val;
6901 if (umin_ptr < umax_val) {
6902 /* Overflow possible, we know nothing */
6903 dst_reg->umin_value = 0;
6904 dst_reg->umax_value = U64_MAX;
6906 /* Cannot overflow (as long as bounds are consistent) */
6907 dst_reg->umin_value = umin_ptr - umax_val;
6908 dst_reg->umax_value = umax_ptr - umin_val;
6910 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
6911 dst_reg->off = ptr_reg->off;
6912 dst_reg->raw = ptr_reg->raw;
6913 if (reg_is_pkt_pointer(ptr_reg)) {
6914 dst_reg->id = ++env->id_gen;
6915 /* something was added to pkt_ptr, set range to zero */
6917 memset(&dst_reg->raw, 0, sizeof(dst_reg->raw));
6923 /* bitwise ops on pointers are troublesome, prohibit. */
6924 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
6925 dst, bpf_alu_string[opcode >> 4]);
6928 /* other operators (e.g. MUL,LSH) produce non-pointer results */
6929 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
6930 dst, bpf_alu_string[opcode >> 4]);
6934 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
6937 __update_reg_bounds(dst_reg);
6938 __reg_deduce_bounds(dst_reg);
6939 __reg_bound_offset(dst_reg);
6941 if (sanitize_check_bounds(env, insn, dst_reg) < 0)
6943 if (sanitize_needed(opcode)) {
6944 ret = sanitize_ptr_alu(env, insn, dst_reg, off_reg, dst_reg,
6947 return sanitize_err(env, insn, ret, off_reg, dst_reg);
6953 static void scalar32_min_max_add(struct bpf_reg_state *dst_reg,
6954 struct bpf_reg_state *src_reg)
6956 s32 smin_val = src_reg->s32_min_value;
6957 s32 smax_val = src_reg->s32_max_value;
6958 u32 umin_val = src_reg->u32_min_value;
6959 u32 umax_val = src_reg->u32_max_value;
6961 if (signed_add32_overflows(dst_reg->s32_min_value, smin_val) ||
6962 signed_add32_overflows(dst_reg->s32_max_value, smax_val)) {
6963 dst_reg->s32_min_value = S32_MIN;
6964 dst_reg->s32_max_value = S32_MAX;
6966 dst_reg->s32_min_value += smin_val;
6967 dst_reg->s32_max_value += smax_val;
6969 if (dst_reg->u32_min_value + umin_val < umin_val ||
6970 dst_reg->u32_max_value + umax_val < umax_val) {
6971 dst_reg->u32_min_value = 0;
6972 dst_reg->u32_max_value = U32_MAX;
6974 dst_reg->u32_min_value += umin_val;
6975 dst_reg->u32_max_value += umax_val;
6979 static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
6980 struct bpf_reg_state *src_reg)
6982 s64 smin_val = src_reg->smin_value;
6983 s64 smax_val = src_reg->smax_value;
6984 u64 umin_val = src_reg->umin_value;
6985 u64 umax_val = src_reg->umax_value;
6987 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
6988 signed_add_overflows(dst_reg->smax_value, smax_val)) {
6989 dst_reg->smin_value = S64_MIN;
6990 dst_reg->smax_value = S64_MAX;
6992 dst_reg->smin_value += smin_val;
6993 dst_reg->smax_value += smax_val;
6995 if (dst_reg->umin_value + umin_val < umin_val ||
6996 dst_reg->umax_value + umax_val < umax_val) {
6997 dst_reg->umin_value = 0;
6998 dst_reg->umax_value = U64_MAX;
7000 dst_reg->umin_value += umin_val;
7001 dst_reg->umax_value += umax_val;
7005 static void scalar32_min_max_sub(struct bpf_reg_state *dst_reg,
7006 struct bpf_reg_state *src_reg)
7008 s32 smin_val = src_reg->s32_min_value;
7009 s32 smax_val = src_reg->s32_max_value;
7010 u32 umin_val = src_reg->u32_min_value;
7011 u32 umax_val = src_reg->u32_max_value;
7013 if (signed_sub32_overflows(dst_reg->s32_min_value, smax_val) ||
7014 signed_sub32_overflows(dst_reg->s32_max_value, smin_val)) {
7015 /* Overflow possible, we know nothing */
7016 dst_reg->s32_min_value = S32_MIN;
7017 dst_reg->s32_max_value = S32_MAX;
7019 dst_reg->s32_min_value -= smax_val;
7020 dst_reg->s32_max_value -= smin_val;
7022 if (dst_reg->u32_min_value < umax_val) {
7023 /* Overflow possible, we know nothing */
7024 dst_reg->u32_min_value = 0;
7025 dst_reg->u32_max_value = U32_MAX;
7027 /* Cannot overflow (as long as bounds are consistent) */
7028 dst_reg->u32_min_value -= umax_val;
7029 dst_reg->u32_max_value -= umin_val;
7033 static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
7034 struct bpf_reg_state *src_reg)
7036 s64 smin_val = src_reg->smin_value;
7037 s64 smax_val = src_reg->smax_value;
7038 u64 umin_val = src_reg->umin_value;
7039 u64 umax_val = src_reg->umax_value;
7041 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
7042 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
7043 /* Overflow possible, we know nothing */
7044 dst_reg->smin_value = S64_MIN;
7045 dst_reg->smax_value = S64_MAX;
7047 dst_reg->smin_value -= smax_val;
7048 dst_reg->smax_value -= smin_val;
7050 if (dst_reg->umin_value < umax_val) {
7051 /* Overflow possible, we know nothing */
7052 dst_reg->umin_value = 0;
7053 dst_reg->umax_value = U64_MAX;
7055 /* Cannot overflow (as long as bounds are consistent) */
7056 dst_reg->umin_value -= umax_val;
7057 dst_reg->umax_value -= umin_val;
7061 static void scalar32_min_max_mul(struct bpf_reg_state *dst_reg,
7062 struct bpf_reg_state *src_reg)
7064 s32 smin_val = src_reg->s32_min_value;
7065 u32 umin_val = src_reg->u32_min_value;
7066 u32 umax_val = src_reg->u32_max_value;
7068 if (smin_val < 0 || dst_reg->s32_min_value < 0) {
7069 /* Ain't nobody got time to multiply that sign */
7070 __mark_reg32_unbounded(dst_reg);
7073 /* Both values are positive, so we can work with unsigned and
7074 * copy the result to signed (unless it exceeds S32_MAX).
7076 if (umax_val > U16_MAX || dst_reg->u32_max_value > U16_MAX) {
7077 /* Potential overflow, we know nothing */
7078 __mark_reg32_unbounded(dst_reg);
7081 dst_reg->u32_min_value *= umin_val;
7082 dst_reg->u32_max_value *= umax_val;
7083 if (dst_reg->u32_max_value > S32_MAX) {
7084 /* Overflow possible, we know nothing */
7085 dst_reg->s32_min_value = S32_MIN;
7086 dst_reg->s32_max_value = S32_MAX;
7088 dst_reg->s32_min_value = dst_reg->u32_min_value;
7089 dst_reg->s32_max_value = dst_reg->u32_max_value;
7093 static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
7094 struct bpf_reg_state *src_reg)
7096 s64 smin_val = src_reg->smin_value;
7097 u64 umin_val = src_reg->umin_value;
7098 u64 umax_val = src_reg->umax_value;
7100 if (smin_val < 0 || dst_reg->smin_value < 0) {
7101 /* Ain't nobody got time to multiply that sign */
7102 __mark_reg64_unbounded(dst_reg);
7105 /* Both values are positive, so we can work with unsigned and
7106 * copy the result to signed (unless it exceeds S64_MAX).
7108 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
7109 /* Potential overflow, we know nothing */
7110 __mark_reg64_unbounded(dst_reg);
7113 dst_reg->umin_value *= umin_val;
7114 dst_reg->umax_value *= umax_val;
7115 if (dst_reg->umax_value > S64_MAX) {
7116 /* Overflow possible, we know nothing */
7117 dst_reg->smin_value = S64_MIN;
7118 dst_reg->smax_value = S64_MAX;
7120 dst_reg->smin_value = dst_reg->umin_value;
7121 dst_reg->smax_value = dst_reg->umax_value;
7125 static void scalar32_min_max_and(struct bpf_reg_state *dst_reg,
7126 struct bpf_reg_state *src_reg)
7128 bool src_known = tnum_subreg_is_const(src_reg->var_off);
7129 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
7130 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
7131 s32 smin_val = src_reg->s32_min_value;
7132 u32 umax_val = src_reg->u32_max_value;
7134 if (src_known && dst_known) {
7135 __mark_reg32_known(dst_reg, var32_off.value);
7139 /* We get our minimum from the var_off, since that's inherently
7140 * bitwise. Our maximum is the minimum of the operands' maxima.
7142 dst_reg->u32_min_value = var32_off.value;
7143 dst_reg->u32_max_value = min(dst_reg->u32_max_value, umax_val);
7144 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
7145 /* Lose signed bounds when ANDing negative numbers,
7146 * ain't nobody got time for that.
7148 dst_reg->s32_min_value = S32_MIN;
7149 dst_reg->s32_max_value = S32_MAX;
7151 /* ANDing two positives gives a positive, so safe to
7152 * cast result into s64.
7154 dst_reg->s32_min_value = dst_reg->u32_min_value;
7155 dst_reg->s32_max_value = dst_reg->u32_max_value;
7159 static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
7160 struct bpf_reg_state *src_reg)
7162 bool src_known = tnum_is_const(src_reg->var_off);
7163 bool dst_known = tnum_is_const(dst_reg->var_off);
7164 s64 smin_val = src_reg->smin_value;
7165 u64 umax_val = src_reg->umax_value;
7167 if (src_known && dst_known) {
7168 __mark_reg_known(dst_reg, dst_reg->var_off.value);
7172 /* We get our minimum from the var_off, since that's inherently
7173 * bitwise. Our maximum is the minimum of the operands' maxima.
7175 dst_reg->umin_value = dst_reg->var_off.value;
7176 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
7177 if (dst_reg->smin_value < 0 || smin_val < 0) {
7178 /* Lose signed bounds when ANDing negative numbers,
7179 * ain't nobody got time for that.
7181 dst_reg->smin_value = S64_MIN;
7182 dst_reg->smax_value = S64_MAX;
7184 /* ANDing two positives gives a positive, so safe to
7185 * cast result into s64.
7187 dst_reg->smin_value = dst_reg->umin_value;
7188 dst_reg->smax_value = dst_reg->umax_value;
7190 /* We may learn something more from the var_off */
7191 __update_reg_bounds(dst_reg);
7194 static void scalar32_min_max_or(struct bpf_reg_state *dst_reg,
7195 struct bpf_reg_state *src_reg)
7197 bool src_known = tnum_subreg_is_const(src_reg->var_off);
7198 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
7199 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
7200 s32 smin_val = src_reg->s32_min_value;
7201 u32 umin_val = src_reg->u32_min_value;
7203 if (src_known && dst_known) {
7204 __mark_reg32_known(dst_reg, var32_off.value);
7208 /* We get our maximum from the var_off, and our minimum is the
7209 * maximum of the operands' minima
7211 dst_reg->u32_min_value = max(dst_reg->u32_min_value, umin_val);
7212 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
7213 if (dst_reg->s32_min_value < 0 || smin_val < 0) {
7214 /* Lose signed bounds when ORing negative numbers,
7215 * ain't nobody got time for that.
7217 dst_reg->s32_min_value = S32_MIN;
7218 dst_reg->s32_max_value = S32_MAX;
7220 /* ORing two positives gives a positive, so safe to
7221 * cast result into s64.
7223 dst_reg->s32_min_value = dst_reg->u32_min_value;
7224 dst_reg->s32_max_value = dst_reg->u32_max_value;
7228 static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
7229 struct bpf_reg_state *src_reg)
7231 bool src_known = tnum_is_const(src_reg->var_off);
7232 bool dst_known = tnum_is_const(dst_reg->var_off);
7233 s64 smin_val = src_reg->smin_value;
7234 u64 umin_val = src_reg->umin_value;
7236 if (src_known && dst_known) {
7237 __mark_reg_known(dst_reg, dst_reg->var_off.value);
7241 /* We get our maximum from the var_off, and our minimum is the
7242 * maximum of the operands' minima
7244 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
7245 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
7246 if (dst_reg->smin_value < 0 || smin_val < 0) {
7247 /* Lose signed bounds when ORing negative numbers,
7248 * ain't nobody got time for that.
7250 dst_reg->smin_value = S64_MIN;
7251 dst_reg->smax_value = S64_MAX;
7253 /* ORing two positives gives a positive, so safe to
7254 * cast result into s64.
7256 dst_reg->smin_value = dst_reg->umin_value;
7257 dst_reg->smax_value = dst_reg->umax_value;
7259 /* We may learn something more from the var_off */
7260 __update_reg_bounds(dst_reg);
7263 static void scalar32_min_max_xor(struct bpf_reg_state *dst_reg,
7264 struct bpf_reg_state *src_reg)
7266 bool src_known = tnum_subreg_is_const(src_reg->var_off);
7267 bool dst_known = tnum_subreg_is_const(dst_reg->var_off);
7268 struct tnum var32_off = tnum_subreg(dst_reg->var_off);
7269 s32 smin_val = src_reg->s32_min_value;
7271 if (src_known && dst_known) {
7272 __mark_reg32_known(dst_reg, var32_off.value);
7276 /* We get both minimum and maximum from the var32_off. */
7277 dst_reg->u32_min_value = var32_off.value;
7278 dst_reg->u32_max_value = var32_off.value | var32_off.mask;
7280 if (dst_reg->s32_min_value >= 0 && smin_val >= 0) {
7281 /* XORing two positive sign numbers gives a positive,
7282 * so safe to cast u32 result into s32.
7284 dst_reg->s32_min_value = dst_reg->u32_min_value;
7285 dst_reg->s32_max_value = dst_reg->u32_max_value;
7287 dst_reg->s32_min_value = S32_MIN;
7288 dst_reg->s32_max_value = S32_MAX;
7292 static void scalar_min_max_xor(struct bpf_reg_state *dst_reg,
7293 struct bpf_reg_state *src_reg)
7295 bool src_known = tnum_is_const(src_reg->var_off);
7296 bool dst_known = tnum_is_const(dst_reg->var_off);
7297 s64 smin_val = src_reg->smin_value;
7299 if (src_known && dst_known) {
7300 /* dst_reg->var_off.value has been updated earlier */
7301 __mark_reg_known(dst_reg, dst_reg->var_off.value);
7305 /* We get both minimum and maximum from the var_off. */
7306 dst_reg->umin_value = dst_reg->var_off.value;
7307 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
7309 if (dst_reg->smin_value >= 0 && smin_val >= 0) {
7310 /* XORing two positive sign numbers gives a positive,
7311 * so safe to cast u64 result into s64.
7313 dst_reg->smin_value = dst_reg->umin_value;
7314 dst_reg->smax_value = dst_reg->umax_value;
7316 dst_reg->smin_value = S64_MIN;
7317 dst_reg->smax_value = S64_MAX;
7320 __update_reg_bounds(dst_reg);
7323 static void __scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
7324 u64 umin_val, u64 umax_val)
7326 /* We lose all sign bit information (except what we can pick
7329 dst_reg->s32_min_value = S32_MIN;
7330 dst_reg->s32_max_value = S32_MAX;
7331 /* If we might shift our top bit out, then we know nothing */
7332 if (umax_val > 31 || dst_reg->u32_max_value > 1ULL << (31 - umax_val)) {
7333 dst_reg->u32_min_value = 0;
7334 dst_reg->u32_max_value = U32_MAX;
7336 dst_reg->u32_min_value <<= umin_val;
7337 dst_reg->u32_max_value <<= umax_val;
7341 static void scalar32_min_max_lsh(struct bpf_reg_state *dst_reg,
7342 struct bpf_reg_state *src_reg)
7344 u32 umax_val = src_reg->u32_max_value;
7345 u32 umin_val = src_reg->u32_min_value;
7346 /* u32 alu operation will zext upper bits */
7347 struct tnum subreg = tnum_subreg(dst_reg->var_off);
7349 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
7350 dst_reg->var_off = tnum_subreg(tnum_lshift(subreg, umin_val));
7351 /* Not required but being careful mark reg64 bounds as unknown so
7352 * that we are forced to pick them up from tnum and zext later and
7353 * if some path skips this step we are still safe.
7355 __mark_reg64_unbounded(dst_reg);
7356 __update_reg32_bounds(dst_reg);
7359 static void __scalar64_min_max_lsh(struct bpf_reg_state *dst_reg,
7360 u64 umin_val, u64 umax_val)
7362 /* Special case <<32 because it is a common compiler pattern to sign
7363 * extend subreg by doing <<32 s>>32. In this case if 32bit bounds are
7364 * positive we know this shift will also be positive so we can track
7365 * bounds correctly. Otherwise we lose all sign bit information except
7366 * what we can pick up from var_off. Perhaps we can generalize this
7367 * later to shifts of any length.
7369 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_max_value >= 0)
7370 dst_reg->smax_value = (s64)dst_reg->s32_max_value << 32;
7372 dst_reg->smax_value = S64_MAX;
7374 if (umin_val == 32 && umax_val == 32 && dst_reg->s32_min_value >= 0)
7375 dst_reg->smin_value = (s64)dst_reg->s32_min_value << 32;
7377 dst_reg->smin_value = S64_MIN;
7379 /* If we might shift our top bit out, then we know nothing */
7380 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
7381 dst_reg->umin_value = 0;
7382 dst_reg->umax_value = U64_MAX;
7384 dst_reg->umin_value <<= umin_val;
7385 dst_reg->umax_value <<= umax_val;
7389 static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
7390 struct bpf_reg_state *src_reg)
7392 u64 umax_val = src_reg->umax_value;
7393 u64 umin_val = src_reg->umin_value;
7395 /* scalar64 calc uses 32bit unshifted bounds so must be called first */
7396 __scalar64_min_max_lsh(dst_reg, umin_val, umax_val);
7397 __scalar32_min_max_lsh(dst_reg, umin_val, umax_val);
7399 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
7400 /* We may learn something more from the var_off */
7401 __update_reg_bounds(dst_reg);
7404 static void scalar32_min_max_rsh(struct bpf_reg_state *dst_reg,
7405 struct bpf_reg_state *src_reg)
7407 struct tnum subreg = tnum_subreg(dst_reg->var_off);
7408 u32 umax_val = src_reg->u32_max_value;
7409 u32 umin_val = src_reg->u32_min_value;
7411 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
7412 * be negative, then either:
7413 * 1) src_reg might be zero, so the sign bit of the result is
7414 * unknown, so we lose our signed bounds
7415 * 2) it's known negative, thus the unsigned bounds capture the
7417 * 3) the signed bounds cross zero, so they tell us nothing
7419 * If the value in dst_reg is known nonnegative, then again the
7420 * unsigned bounds capture the signed bounds.
7421 * Thus, in all cases it suffices to blow away our signed bounds
7422 * and rely on inferring new ones from the unsigned bounds and
7423 * var_off of the result.
7425 dst_reg->s32_min_value = S32_MIN;
7426 dst_reg->s32_max_value = S32_MAX;
7428 dst_reg->var_off = tnum_rshift(subreg, umin_val);
7429 dst_reg->u32_min_value >>= umax_val;
7430 dst_reg->u32_max_value >>= umin_val;
7432 __mark_reg64_unbounded(dst_reg);
7433 __update_reg32_bounds(dst_reg);
7436 static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
7437 struct bpf_reg_state *src_reg)
7439 u64 umax_val = src_reg->umax_value;
7440 u64 umin_val = src_reg->umin_value;
7442 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
7443 * be negative, then either:
7444 * 1) src_reg might be zero, so the sign bit of the result is
7445 * unknown, so we lose our signed bounds
7446 * 2) it's known negative, thus the unsigned bounds capture the
7448 * 3) the signed bounds cross zero, so they tell us nothing
7450 * If the value in dst_reg is known nonnegative, then again the
7451 * unsigned bounds capture the signed bounds.
7452 * Thus, in all cases it suffices to blow away our signed bounds
7453 * and rely on inferring new ones from the unsigned bounds and
7454 * var_off of the result.
7456 dst_reg->smin_value = S64_MIN;
7457 dst_reg->smax_value = S64_MAX;
7458 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
7459 dst_reg->umin_value >>= umax_val;
7460 dst_reg->umax_value >>= umin_val;
7462 /* Its not easy to operate on alu32 bounds here because it depends
7463 * on bits being shifted in. Take easy way out and mark unbounded
7464 * so we can recalculate later from tnum.
7466 __mark_reg32_unbounded(dst_reg);
7467 __update_reg_bounds(dst_reg);
7470 static void scalar32_min_max_arsh(struct bpf_reg_state *dst_reg,
7471 struct bpf_reg_state *src_reg)
7473 u64 umin_val = src_reg->u32_min_value;
7475 /* Upon reaching here, src_known is true and
7476 * umax_val is equal to umin_val.
7478 dst_reg->s32_min_value = (u32)(((s32)dst_reg->s32_min_value) >> umin_val);
7479 dst_reg->s32_max_value = (u32)(((s32)dst_reg->s32_max_value) >> umin_val);
7481 dst_reg->var_off = tnum_arshift(tnum_subreg(dst_reg->var_off), umin_val, 32);
7483 /* blow away the dst_reg umin_value/umax_value and rely on
7484 * dst_reg var_off to refine the result.
7486 dst_reg->u32_min_value = 0;
7487 dst_reg->u32_max_value = U32_MAX;
7489 __mark_reg64_unbounded(dst_reg);
7490 __update_reg32_bounds(dst_reg);
7493 static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
7494 struct bpf_reg_state *src_reg)
7496 u64 umin_val = src_reg->umin_value;
7498 /* Upon reaching here, src_known is true and umax_val is equal
7501 dst_reg->smin_value >>= umin_val;
7502 dst_reg->smax_value >>= umin_val;
7504 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val, 64);
7506 /* blow away the dst_reg umin_value/umax_value and rely on
7507 * dst_reg var_off to refine the result.
7509 dst_reg->umin_value = 0;
7510 dst_reg->umax_value = U64_MAX;
7512 /* Its not easy to operate on alu32 bounds here because it depends
7513 * on bits being shifted in from upper 32-bits. Take easy way out
7514 * and mark unbounded so we can recalculate later from tnum.
7516 __mark_reg32_unbounded(dst_reg);
7517 __update_reg_bounds(dst_reg);
7520 /* WARNING: This function does calculations on 64-bit values, but the actual
7521 * execution may occur on 32-bit values. Therefore, things like bitshifts
7522 * need extra checks in the 32-bit case.
7524 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
7525 struct bpf_insn *insn,
7526 struct bpf_reg_state *dst_reg,
7527 struct bpf_reg_state src_reg)
7529 struct bpf_reg_state *regs = cur_regs(env);
7530 u8 opcode = BPF_OP(insn->code);
7532 s64 smin_val, smax_val;
7533 u64 umin_val, umax_val;
7534 s32 s32_min_val, s32_max_val;
7535 u32 u32_min_val, u32_max_val;
7536 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
7537 bool alu32 = (BPF_CLASS(insn->code) != BPF_ALU64);
7540 smin_val = src_reg.smin_value;
7541 smax_val = src_reg.smax_value;
7542 umin_val = src_reg.umin_value;
7543 umax_val = src_reg.umax_value;
7545 s32_min_val = src_reg.s32_min_value;
7546 s32_max_val = src_reg.s32_max_value;
7547 u32_min_val = src_reg.u32_min_value;
7548 u32_max_val = src_reg.u32_max_value;
7551 src_known = tnum_subreg_is_const(src_reg.var_off);
7553 (s32_min_val != s32_max_val || u32_min_val != u32_max_val)) ||
7554 s32_min_val > s32_max_val || u32_min_val > u32_max_val) {
7555 /* Taint dst register if offset had invalid bounds
7556 * derived from e.g. dead branches.
7558 __mark_reg_unknown(env, dst_reg);
7562 src_known = tnum_is_const(src_reg.var_off);
7564 (smin_val != smax_val || umin_val != umax_val)) ||
7565 smin_val > smax_val || umin_val > umax_val) {
7566 /* Taint dst register if offset had invalid bounds
7567 * derived from e.g. dead branches.
7569 __mark_reg_unknown(env, dst_reg);
7575 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
7576 __mark_reg_unknown(env, dst_reg);
7580 if (sanitize_needed(opcode)) {
7581 ret = sanitize_val_alu(env, insn);
7583 return sanitize_err(env, insn, ret, NULL, NULL);
7586 /* Calculate sign/unsigned bounds and tnum for alu32 and alu64 bit ops.
7587 * There are two classes of instructions: The first class we track both
7588 * alu32 and alu64 sign/unsigned bounds independently this provides the
7589 * greatest amount of precision when alu operations are mixed with jmp32
7590 * operations. These operations are BPF_ADD, BPF_SUB, BPF_MUL, BPF_ADD,
7591 * and BPF_OR. This is possible because these ops have fairly easy to
7592 * understand and calculate behavior in both 32-bit and 64-bit alu ops.
7593 * See alu32 verifier tests for examples. The second class of
7594 * operations, BPF_LSH, BPF_RSH, and BPF_ARSH, however are not so easy
7595 * with regards to tracking sign/unsigned bounds because the bits may
7596 * cross subreg boundaries in the alu64 case. When this happens we mark
7597 * the reg unbounded in the subreg bound space and use the resulting
7598 * tnum to calculate an approximation of the sign/unsigned bounds.
7602 scalar32_min_max_add(dst_reg, &src_reg);
7603 scalar_min_max_add(dst_reg, &src_reg);
7604 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
7607 scalar32_min_max_sub(dst_reg, &src_reg);
7608 scalar_min_max_sub(dst_reg, &src_reg);
7609 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
7612 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
7613 scalar32_min_max_mul(dst_reg, &src_reg);
7614 scalar_min_max_mul(dst_reg, &src_reg);
7617 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
7618 scalar32_min_max_and(dst_reg, &src_reg);
7619 scalar_min_max_and(dst_reg, &src_reg);
7622 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
7623 scalar32_min_max_or(dst_reg, &src_reg);
7624 scalar_min_max_or(dst_reg, &src_reg);
7627 dst_reg->var_off = tnum_xor(dst_reg->var_off, src_reg.var_off);
7628 scalar32_min_max_xor(dst_reg, &src_reg);
7629 scalar_min_max_xor(dst_reg, &src_reg);
7632 if (umax_val >= insn_bitness) {
7633 /* Shifts greater than 31 or 63 are undefined.
7634 * This includes shifts by a negative number.
7636 mark_reg_unknown(env, regs, insn->dst_reg);
7640 scalar32_min_max_lsh(dst_reg, &src_reg);
7642 scalar_min_max_lsh(dst_reg, &src_reg);
7645 if (umax_val >= insn_bitness) {
7646 /* Shifts greater than 31 or 63 are undefined.
7647 * This includes shifts by a negative number.
7649 mark_reg_unknown(env, regs, insn->dst_reg);
7653 scalar32_min_max_rsh(dst_reg, &src_reg);
7655 scalar_min_max_rsh(dst_reg, &src_reg);
7658 if (umax_val >= insn_bitness) {
7659 /* Shifts greater than 31 or 63 are undefined.
7660 * This includes shifts by a negative number.
7662 mark_reg_unknown(env, regs, insn->dst_reg);
7666 scalar32_min_max_arsh(dst_reg, &src_reg);
7668 scalar_min_max_arsh(dst_reg, &src_reg);
7671 mark_reg_unknown(env, regs, insn->dst_reg);
7675 /* ALU32 ops are zero extended into 64bit register */
7677 zext_32_to_64(dst_reg);
7679 __update_reg_bounds(dst_reg);
7680 __reg_deduce_bounds(dst_reg);
7681 __reg_bound_offset(dst_reg);
7685 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
7688 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
7689 struct bpf_insn *insn)
7691 struct bpf_verifier_state *vstate = env->cur_state;
7692 struct bpf_func_state *state = vstate->frame[vstate->curframe];
7693 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
7694 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
7695 u8 opcode = BPF_OP(insn->code);
7698 dst_reg = ®s[insn->dst_reg];
7700 if (dst_reg->type != SCALAR_VALUE)
7703 /* Make sure ID is cleared otherwise dst_reg min/max could be
7704 * incorrectly propagated into other registers by find_equal_scalars()
7707 if (BPF_SRC(insn->code) == BPF_X) {
7708 src_reg = ®s[insn->src_reg];
7709 if (src_reg->type != SCALAR_VALUE) {
7710 if (dst_reg->type != SCALAR_VALUE) {
7711 /* Combining two pointers by any ALU op yields
7712 * an arbitrary scalar. Disallow all math except
7713 * pointer subtraction
7715 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
7716 mark_reg_unknown(env, regs, insn->dst_reg);
7719 verbose(env, "R%d pointer %s pointer prohibited\n",
7721 bpf_alu_string[opcode >> 4]);
7724 /* scalar += pointer
7725 * This is legal, but we have to reverse our
7726 * src/dest handling in computing the range
7728 err = mark_chain_precision(env, insn->dst_reg);
7731 return adjust_ptr_min_max_vals(env, insn,
7734 } else if (ptr_reg) {
7735 /* pointer += scalar */
7736 err = mark_chain_precision(env, insn->src_reg);
7739 return adjust_ptr_min_max_vals(env, insn,
7743 /* Pretend the src is a reg with a known value, since we only
7744 * need to be able to read from this state.
7746 off_reg.type = SCALAR_VALUE;
7747 __mark_reg_known(&off_reg, insn->imm);
7749 if (ptr_reg) /* pointer += K */
7750 return adjust_ptr_min_max_vals(env, insn,
7754 /* Got here implies adding two SCALAR_VALUEs */
7755 if (WARN_ON_ONCE(ptr_reg)) {
7756 print_verifier_state(env, state);
7757 verbose(env, "verifier internal error: unexpected ptr_reg\n");
7760 if (WARN_ON(!src_reg)) {
7761 print_verifier_state(env, state);
7762 verbose(env, "verifier internal error: no src_reg\n");
7765 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
7768 /* check validity of 32-bit and 64-bit arithmetic operations */
7769 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
7771 struct bpf_reg_state *regs = cur_regs(env);
7772 u8 opcode = BPF_OP(insn->code);
7775 if (opcode == BPF_END || opcode == BPF_NEG) {
7776 if (opcode == BPF_NEG) {
7777 if (BPF_SRC(insn->code) != 0 ||
7778 insn->src_reg != BPF_REG_0 ||
7779 insn->off != 0 || insn->imm != 0) {
7780 verbose(env, "BPF_NEG uses reserved fields\n");
7784 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
7785 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
7786 BPF_CLASS(insn->code) == BPF_ALU64) {
7787 verbose(env, "BPF_END uses reserved fields\n");
7792 /* check src operand */
7793 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
7797 if (is_pointer_value(env, insn->dst_reg)) {
7798 verbose(env, "R%d pointer arithmetic prohibited\n",
7803 /* check dest operand */
7804 err = check_reg_arg(env, insn->dst_reg, DST_OP);
7808 } else if (opcode == BPF_MOV) {
7810 if (BPF_SRC(insn->code) == BPF_X) {
7811 if (insn->imm != 0 || insn->off != 0) {
7812 verbose(env, "BPF_MOV uses reserved fields\n");
7816 /* check src operand */
7817 err = check_reg_arg(env, insn->src_reg, SRC_OP);
7821 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
7822 verbose(env, "BPF_MOV uses reserved fields\n");
7827 /* check dest operand, mark as required later */
7828 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
7832 if (BPF_SRC(insn->code) == BPF_X) {
7833 struct bpf_reg_state *src_reg = regs + insn->src_reg;
7834 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
7836 if (BPF_CLASS(insn->code) == BPF_ALU64) {
7838 * copy register state to dest reg
7840 if (src_reg->type == SCALAR_VALUE && !src_reg->id)
7841 /* Assign src and dst registers the same ID
7842 * that will be used by find_equal_scalars()
7843 * to propagate min/max range.
7845 src_reg->id = ++env->id_gen;
7846 *dst_reg = *src_reg;
7847 dst_reg->live |= REG_LIVE_WRITTEN;
7848 dst_reg->subreg_def = DEF_NOT_SUBREG;
7851 if (is_pointer_value(env, insn->src_reg)) {
7853 "R%d partial copy of pointer\n",
7856 } else if (src_reg->type == SCALAR_VALUE) {
7857 *dst_reg = *src_reg;
7858 /* Make sure ID is cleared otherwise
7859 * dst_reg min/max could be incorrectly
7860 * propagated into src_reg by find_equal_scalars()
7863 dst_reg->live |= REG_LIVE_WRITTEN;
7864 dst_reg->subreg_def = env->insn_idx + 1;
7866 mark_reg_unknown(env, regs,
7869 zext_32_to_64(dst_reg);
7873 * remember the value we stored into this reg
7875 /* clear any state __mark_reg_known doesn't set */
7876 mark_reg_unknown(env, regs, insn->dst_reg);
7877 regs[insn->dst_reg].type = SCALAR_VALUE;
7878 if (BPF_CLASS(insn->code) == BPF_ALU64) {
7879 __mark_reg_known(regs + insn->dst_reg,
7882 __mark_reg_known(regs + insn->dst_reg,
7887 } else if (opcode > BPF_END) {
7888 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
7891 } else { /* all other ALU ops: and, sub, xor, add, ... */
7893 if (BPF_SRC(insn->code) == BPF_X) {
7894 if (insn->imm != 0 || insn->off != 0) {
7895 verbose(env, "BPF_ALU uses reserved fields\n");
7898 /* check src1 operand */
7899 err = check_reg_arg(env, insn->src_reg, SRC_OP);
7903 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
7904 verbose(env, "BPF_ALU uses reserved fields\n");
7909 /* check src2 operand */
7910 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
7914 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
7915 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
7916 verbose(env, "div by zero\n");
7920 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
7921 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
7922 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
7924 if (insn->imm < 0 || insn->imm >= size) {
7925 verbose(env, "invalid shift %d\n", insn->imm);
7930 /* check dest operand */
7931 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
7935 return adjust_reg_min_max_vals(env, insn);
7941 static void __find_good_pkt_pointers(struct bpf_func_state *state,
7942 struct bpf_reg_state *dst_reg,
7943 enum bpf_reg_type type, int new_range)
7945 struct bpf_reg_state *reg;
7948 for (i = 0; i < MAX_BPF_REG; i++) {
7949 reg = &state->regs[i];
7950 if (reg->type == type && reg->id == dst_reg->id)
7951 /* keep the maximum range already checked */
7952 reg->range = max(reg->range, new_range);
7955 bpf_for_each_spilled_reg(i, state, reg) {
7958 if (reg->type == type && reg->id == dst_reg->id)
7959 reg->range = max(reg->range, new_range);
7963 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
7964 struct bpf_reg_state *dst_reg,
7965 enum bpf_reg_type type,
7966 bool range_right_open)
7970 if (dst_reg->off < 0 ||
7971 (dst_reg->off == 0 && range_right_open))
7972 /* This doesn't give us any range */
7975 if (dst_reg->umax_value > MAX_PACKET_OFF ||
7976 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
7977 /* Risk of overflow. For instance, ptr + (1<<63) may be less
7978 * than pkt_end, but that's because it's also less than pkt.
7982 new_range = dst_reg->off;
7983 if (range_right_open)
7986 /* Examples for register markings:
7988 * pkt_data in dst register:
7992 * if (r2 > pkt_end) goto <handle exception>
7997 * if (r2 < pkt_end) goto <access okay>
7998 * <handle exception>
8001 * r2 == dst_reg, pkt_end == src_reg
8002 * r2=pkt(id=n,off=8,r=0)
8003 * r3=pkt(id=n,off=0,r=0)
8005 * pkt_data in src register:
8009 * if (pkt_end >= r2) goto <access okay>
8010 * <handle exception>
8014 * if (pkt_end <= r2) goto <handle exception>
8018 * pkt_end == dst_reg, r2 == src_reg
8019 * r2=pkt(id=n,off=8,r=0)
8020 * r3=pkt(id=n,off=0,r=0)
8022 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
8023 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
8024 * and [r3, r3 + 8-1) respectively is safe to access depending on
8028 /* If our ids match, then we must have the same max_value. And we
8029 * don't care about the other reg's fixed offset, since if it's too big
8030 * the range won't allow anything.
8031 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
8033 for (i = 0; i <= vstate->curframe; i++)
8034 __find_good_pkt_pointers(vstate->frame[i], dst_reg, type,
8038 static int is_branch32_taken(struct bpf_reg_state *reg, u32 val, u8 opcode)
8040 struct tnum subreg = tnum_subreg(reg->var_off);
8041 s32 sval = (s32)val;
8045 if (tnum_is_const(subreg))
8046 return !!tnum_equals_const(subreg, val);
8049 if (tnum_is_const(subreg))
8050 return !tnum_equals_const(subreg, val);
8053 if ((~subreg.mask & subreg.value) & val)
8055 if (!((subreg.mask | subreg.value) & val))
8059 if (reg->u32_min_value > val)
8061 else if (reg->u32_max_value <= val)
8065 if (reg->s32_min_value > sval)
8067 else if (reg->s32_max_value <= sval)
8071 if (reg->u32_max_value < val)
8073 else if (reg->u32_min_value >= val)
8077 if (reg->s32_max_value < sval)
8079 else if (reg->s32_min_value >= sval)
8083 if (reg->u32_min_value >= val)
8085 else if (reg->u32_max_value < val)
8089 if (reg->s32_min_value >= sval)
8091 else if (reg->s32_max_value < sval)
8095 if (reg->u32_max_value <= val)
8097 else if (reg->u32_min_value > val)
8101 if (reg->s32_max_value <= sval)
8103 else if (reg->s32_min_value > sval)
8112 static int is_branch64_taken(struct bpf_reg_state *reg, u64 val, u8 opcode)
8114 s64 sval = (s64)val;
8118 if (tnum_is_const(reg->var_off))
8119 return !!tnum_equals_const(reg->var_off, val);
8122 if (tnum_is_const(reg->var_off))
8123 return !tnum_equals_const(reg->var_off, val);
8126 if ((~reg->var_off.mask & reg->var_off.value) & val)
8128 if (!((reg->var_off.mask | reg->var_off.value) & val))
8132 if (reg->umin_value > val)
8134 else if (reg->umax_value <= val)
8138 if (reg->smin_value > sval)
8140 else if (reg->smax_value <= sval)
8144 if (reg->umax_value < val)
8146 else if (reg->umin_value >= val)
8150 if (reg->smax_value < sval)
8152 else if (reg->smin_value >= sval)
8156 if (reg->umin_value >= val)
8158 else if (reg->umax_value < val)
8162 if (reg->smin_value >= sval)
8164 else if (reg->smax_value < sval)
8168 if (reg->umax_value <= val)
8170 else if (reg->umin_value > val)
8174 if (reg->smax_value <= sval)
8176 else if (reg->smin_value > sval)
8184 /* compute branch direction of the expression "if (reg opcode val) goto target;"
8186 * 1 - branch will be taken and "goto target" will be executed
8187 * 0 - branch will not be taken and fall-through to next insn
8188 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value
8191 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
8194 if (__is_pointer_value(false, reg)) {
8195 if (!reg_type_not_null(reg->type))
8198 /* If pointer is valid tests against zero will fail so we can
8199 * use this to direct branch taken.
8215 return is_branch32_taken(reg, val, opcode);
8216 return is_branch64_taken(reg, val, opcode);
8219 static int flip_opcode(u32 opcode)
8221 /* How can we transform "a <op> b" into "b <op> a"? */
8222 static const u8 opcode_flip[16] = {
8223 /* these stay the same */
8224 [BPF_JEQ >> 4] = BPF_JEQ,
8225 [BPF_JNE >> 4] = BPF_JNE,
8226 [BPF_JSET >> 4] = BPF_JSET,
8227 /* these swap "lesser" and "greater" (L and G in the opcodes) */
8228 [BPF_JGE >> 4] = BPF_JLE,
8229 [BPF_JGT >> 4] = BPF_JLT,
8230 [BPF_JLE >> 4] = BPF_JGE,
8231 [BPF_JLT >> 4] = BPF_JGT,
8232 [BPF_JSGE >> 4] = BPF_JSLE,
8233 [BPF_JSGT >> 4] = BPF_JSLT,
8234 [BPF_JSLE >> 4] = BPF_JSGE,
8235 [BPF_JSLT >> 4] = BPF_JSGT
8237 return opcode_flip[opcode >> 4];
8240 static int is_pkt_ptr_branch_taken(struct bpf_reg_state *dst_reg,
8241 struct bpf_reg_state *src_reg,
8244 struct bpf_reg_state *pkt;
8246 if (src_reg->type == PTR_TO_PACKET_END) {
8248 } else if (dst_reg->type == PTR_TO_PACKET_END) {
8250 opcode = flip_opcode(opcode);
8255 if (pkt->range >= 0)
8260 /* pkt <= pkt_end */
8264 if (pkt->range == BEYOND_PKT_END)
8265 /* pkt has at last one extra byte beyond pkt_end */
8266 return opcode == BPF_JGT;
8272 /* pkt >= pkt_end */
8273 if (pkt->range == BEYOND_PKT_END || pkt->range == AT_PKT_END)
8274 return opcode == BPF_JGE;
8280 /* Adjusts the register min/max values in the case that the dst_reg is the
8281 * variable register that we are working on, and src_reg is a constant or we're
8282 * simply doing a BPF_K check.
8283 * In JEQ/JNE cases we also adjust the var_off values.
8285 static void reg_set_min_max(struct bpf_reg_state *true_reg,
8286 struct bpf_reg_state *false_reg,
8288 u8 opcode, bool is_jmp32)
8290 struct tnum false_32off = tnum_subreg(false_reg->var_off);
8291 struct tnum false_64off = false_reg->var_off;
8292 struct tnum true_32off = tnum_subreg(true_reg->var_off);
8293 struct tnum true_64off = true_reg->var_off;
8294 s64 sval = (s64)val;
8295 s32 sval32 = (s32)val32;
8297 /* If the dst_reg is a pointer, we can't learn anything about its
8298 * variable offset from the compare (unless src_reg were a pointer into
8299 * the same object, but we don't bother with that.
8300 * Since false_reg and true_reg have the same type by construction, we
8301 * only need to check one of them for pointerness.
8303 if (__is_pointer_value(false, false_reg))
8310 struct bpf_reg_state *reg =
8311 opcode == BPF_JEQ ? true_reg : false_reg;
8313 /* JEQ/JNE comparison doesn't change the register equivalence.
8315 * if (r1 == 42) goto label;
8317 * label: // here both r1 and r2 are known to be 42.
8319 * Hence when marking register as known preserve it's ID.
8322 __mark_reg32_known(reg, val32);
8324 ___mark_reg_known(reg, val);
8329 false_32off = tnum_and(false_32off, tnum_const(~val32));
8330 if (is_power_of_2(val32))
8331 true_32off = tnum_or(true_32off,
8334 false_64off = tnum_and(false_64off, tnum_const(~val));
8335 if (is_power_of_2(val))
8336 true_64off = tnum_or(true_64off,
8344 u32 false_umax = opcode == BPF_JGT ? val32 : val32 - 1;
8345 u32 true_umin = opcode == BPF_JGT ? val32 + 1 : val32;
8347 false_reg->u32_max_value = min(false_reg->u32_max_value,
8349 true_reg->u32_min_value = max(true_reg->u32_min_value,
8352 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
8353 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
8355 false_reg->umax_value = min(false_reg->umax_value, false_umax);
8356 true_reg->umin_value = max(true_reg->umin_value, true_umin);
8364 s32 false_smax = opcode == BPF_JSGT ? sval32 : sval32 - 1;
8365 s32 true_smin = opcode == BPF_JSGT ? sval32 + 1 : sval32;
8367 false_reg->s32_max_value = min(false_reg->s32_max_value, false_smax);
8368 true_reg->s32_min_value = max(true_reg->s32_min_value, true_smin);
8370 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
8371 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
8373 false_reg->smax_value = min(false_reg->smax_value, false_smax);
8374 true_reg->smin_value = max(true_reg->smin_value, true_smin);
8382 u32 false_umin = opcode == BPF_JLT ? val32 : val32 + 1;
8383 u32 true_umax = opcode == BPF_JLT ? val32 - 1 : val32;
8385 false_reg->u32_min_value = max(false_reg->u32_min_value,
8387 true_reg->u32_max_value = min(true_reg->u32_max_value,
8390 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
8391 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
8393 false_reg->umin_value = max(false_reg->umin_value, false_umin);
8394 true_reg->umax_value = min(true_reg->umax_value, true_umax);
8402 s32 false_smin = opcode == BPF_JSLT ? sval32 : sval32 + 1;
8403 s32 true_smax = opcode == BPF_JSLT ? sval32 - 1 : sval32;
8405 false_reg->s32_min_value = max(false_reg->s32_min_value, false_smin);
8406 true_reg->s32_max_value = min(true_reg->s32_max_value, true_smax);
8408 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
8409 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
8411 false_reg->smin_value = max(false_reg->smin_value, false_smin);
8412 true_reg->smax_value = min(true_reg->smax_value, true_smax);
8421 false_reg->var_off = tnum_or(tnum_clear_subreg(false_64off),
8422 tnum_subreg(false_32off));
8423 true_reg->var_off = tnum_or(tnum_clear_subreg(true_64off),
8424 tnum_subreg(true_32off));
8425 __reg_combine_32_into_64(false_reg);
8426 __reg_combine_32_into_64(true_reg);
8428 false_reg->var_off = false_64off;
8429 true_reg->var_off = true_64off;
8430 __reg_combine_64_into_32(false_reg);
8431 __reg_combine_64_into_32(true_reg);
8435 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
8438 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
8439 struct bpf_reg_state *false_reg,
8441 u8 opcode, bool is_jmp32)
8443 opcode = flip_opcode(opcode);
8444 /* This uses zero as "not present in table"; luckily the zero opcode,
8445 * BPF_JA, can't get here.
8448 reg_set_min_max(true_reg, false_reg, val, val32, opcode, is_jmp32);
8451 /* Regs are known to be equal, so intersect their min/max/var_off */
8452 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
8453 struct bpf_reg_state *dst_reg)
8455 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
8456 dst_reg->umin_value);
8457 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
8458 dst_reg->umax_value);
8459 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
8460 dst_reg->smin_value);
8461 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
8462 dst_reg->smax_value);
8463 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
8465 /* We might have learned new bounds from the var_off. */
8466 __update_reg_bounds(src_reg);
8467 __update_reg_bounds(dst_reg);
8468 /* We might have learned something about the sign bit. */
8469 __reg_deduce_bounds(src_reg);
8470 __reg_deduce_bounds(dst_reg);
8471 /* We might have learned some bits from the bounds. */
8472 __reg_bound_offset(src_reg);
8473 __reg_bound_offset(dst_reg);
8474 /* Intersecting with the old var_off might have improved our bounds
8475 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
8476 * then new var_off is (0; 0x7f...fc) which improves our umax.
8478 __update_reg_bounds(src_reg);
8479 __update_reg_bounds(dst_reg);
8482 static void reg_combine_min_max(struct bpf_reg_state *true_src,
8483 struct bpf_reg_state *true_dst,
8484 struct bpf_reg_state *false_src,
8485 struct bpf_reg_state *false_dst,
8490 __reg_combine_min_max(true_src, true_dst);
8493 __reg_combine_min_max(false_src, false_dst);
8498 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
8499 struct bpf_reg_state *reg, u32 id,
8502 if (reg_type_may_be_null(reg->type) && reg->id == id &&
8503 !WARN_ON_ONCE(!reg->id)) {
8504 /* Old offset (both fixed and variable parts) should
8505 * have been known-zero, because we don't allow pointer
8506 * arithmetic on pointers that might be NULL.
8508 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
8509 !tnum_equals_const(reg->var_off, 0) ||
8511 __mark_reg_known_zero(reg);
8515 reg->type = SCALAR_VALUE;
8516 /* We don't need id and ref_obj_id from this point
8517 * onwards anymore, thus we should better reset it,
8518 * so that state pruning has chances to take effect.
8521 reg->ref_obj_id = 0;
8526 mark_ptr_not_null_reg(reg);
8528 if (!reg_may_point_to_spin_lock(reg)) {
8529 /* For not-NULL ptr, reg->ref_obj_id will be reset
8530 * in release_reg_references().
8532 * reg->id is still used by spin_lock ptr. Other
8533 * than spin_lock ptr type, reg->id can be reset.
8540 static void __mark_ptr_or_null_regs(struct bpf_func_state *state, u32 id,
8543 struct bpf_reg_state *reg;
8546 for (i = 0; i < MAX_BPF_REG; i++)
8547 mark_ptr_or_null_reg(state, &state->regs[i], id, is_null);
8549 bpf_for_each_spilled_reg(i, state, reg) {
8552 mark_ptr_or_null_reg(state, reg, id, is_null);
8556 /* The logic is similar to find_good_pkt_pointers(), both could eventually
8557 * be folded together at some point.
8559 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
8562 struct bpf_func_state *state = vstate->frame[vstate->curframe];
8563 struct bpf_reg_state *regs = state->regs;
8564 u32 ref_obj_id = regs[regno].ref_obj_id;
8565 u32 id = regs[regno].id;
8568 if (ref_obj_id && ref_obj_id == id && is_null)
8569 /* regs[regno] is in the " == NULL" branch.
8570 * No one could have freed the reference state before
8571 * doing the NULL check.
8573 WARN_ON_ONCE(release_reference_state(state, id));
8575 for (i = 0; i <= vstate->curframe; i++)
8576 __mark_ptr_or_null_regs(vstate->frame[i], id, is_null);
8579 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
8580 struct bpf_reg_state *dst_reg,
8581 struct bpf_reg_state *src_reg,
8582 struct bpf_verifier_state *this_branch,
8583 struct bpf_verifier_state *other_branch)
8585 if (BPF_SRC(insn->code) != BPF_X)
8588 /* Pointers are always 64-bit. */
8589 if (BPF_CLASS(insn->code) == BPF_JMP32)
8592 switch (BPF_OP(insn->code)) {
8594 if ((dst_reg->type == PTR_TO_PACKET &&
8595 src_reg->type == PTR_TO_PACKET_END) ||
8596 (dst_reg->type == PTR_TO_PACKET_META &&
8597 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
8598 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
8599 find_good_pkt_pointers(this_branch, dst_reg,
8600 dst_reg->type, false);
8601 mark_pkt_end(other_branch, insn->dst_reg, true);
8602 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
8603 src_reg->type == PTR_TO_PACKET) ||
8604 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
8605 src_reg->type == PTR_TO_PACKET_META)) {
8606 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
8607 find_good_pkt_pointers(other_branch, src_reg,
8608 src_reg->type, true);
8609 mark_pkt_end(this_branch, insn->src_reg, false);
8615 if ((dst_reg->type == PTR_TO_PACKET &&
8616 src_reg->type == PTR_TO_PACKET_END) ||
8617 (dst_reg->type == PTR_TO_PACKET_META &&
8618 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
8619 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
8620 find_good_pkt_pointers(other_branch, dst_reg,
8621 dst_reg->type, true);
8622 mark_pkt_end(this_branch, insn->dst_reg, false);
8623 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
8624 src_reg->type == PTR_TO_PACKET) ||
8625 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
8626 src_reg->type == PTR_TO_PACKET_META)) {
8627 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
8628 find_good_pkt_pointers(this_branch, src_reg,
8629 src_reg->type, false);
8630 mark_pkt_end(other_branch, insn->src_reg, true);
8636 if ((dst_reg->type == PTR_TO_PACKET &&
8637 src_reg->type == PTR_TO_PACKET_END) ||
8638 (dst_reg->type == PTR_TO_PACKET_META &&
8639 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
8640 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
8641 find_good_pkt_pointers(this_branch, dst_reg,
8642 dst_reg->type, true);
8643 mark_pkt_end(other_branch, insn->dst_reg, false);
8644 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
8645 src_reg->type == PTR_TO_PACKET) ||
8646 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
8647 src_reg->type == PTR_TO_PACKET_META)) {
8648 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
8649 find_good_pkt_pointers(other_branch, src_reg,
8650 src_reg->type, false);
8651 mark_pkt_end(this_branch, insn->src_reg, true);
8657 if ((dst_reg->type == PTR_TO_PACKET &&
8658 src_reg->type == PTR_TO_PACKET_END) ||
8659 (dst_reg->type == PTR_TO_PACKET_META &&
8660 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
8661 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
8662 find_good_pkt_pointers(other_branch, dst_reg,
8663 dst_reg->type, false);
8664 mark_pkt_end(this_branch, insn->dst_reg, true);
8665 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
8666 src_reg->type == PTR_TO_PACKET) ||
8667 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
8668 src_reg->type == PTR_TO_PACKET_META)) {
8669 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
8670 find_good_pkt_pointers(this_branch, src_reg,
8671 src_reg->type, true);
8672 mark_pkt_end(other_branch, insn->src_reg, false);
8684 static void find_equal_scalars(struct bpf_verifier_state *vstate,
8685 struct bpf_reg_state *known_reg)
8687 struct bpf_func_state *state;
8688 struct bpf_reg_state *reg;
8691 for (i = 0; i <= vstate->curframe; i++) {
8692 state = vstate->frame[i];
8693 for (j = 0; j < MAX_BPF_REG; j++) {
8694 reg = &state->regs[j];
8695 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
8699 bpf_for_each_spilled_reg(j, state, reg) {
8702 if (reg->type == SCALAR_VALUE && reg->id == known_reg->id)
8708 static int check_cond_jmp_op(struct bpf_verifier_env *env,
8709 struct bpf_insn *insn, int *insn_idx)
8711 struct bpf_verifier_state *this_branch = env->cur_state;
8712 struct bpf_verifier_state *other_branch;
8713 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
8714 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
8715 u8 opcode = BPF_OP(insn->code);
8720 /* Only conditional jumps are expected to reach here. */
8721 if (opcode == BPF_JA || opcode > BPF_JSLE) {
8722 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
8726 if (BPF_SRC(insn->code) == BPF_X) {
8727 if (insn->imm != 0) {
8728 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
8732 /* check src1 operand */
8733 err = check_reg_arg(env, insn->src_reg, SRC_OP);
8737 if (is_pointer_value(env, insn->src_reg)) {
8738 verbose(env, "R%d pointer comparison prohibited\n",
8742 src_reg = ®s[insn->src_reg];
8744 if (insn->src_reg != BPF_REG_0) {
8745 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
8750 /* check src2 operand */
8751 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
8755 dst_reg = ®s[insn->dst_reg];
8756 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
8758 if (BPF_SRC(insn->code) == BPF_K) {
8759 pred = is_branch_taken(dst_reg, insn->imm, opcode, is_jmp32);
8760 } else if (src_reg->type == SCALAR_VALUE &&
8761 is_jmp32 && tnum_is_const(tnum_subreg(src_reg->var_off))) {
8762 pred = is_branch_taken(dst_reg,
8763 tnum_subreg(src_reg->var_off).value,
8766 } else if (src_reg->type == SCALAR_VALUE &&
8767 !is_jmp32 && tnum_is_const(src_reg->var_off)) {
8768 pred = is_branch_taken(dst_reg,
8769 src_reg->var_off.value,
8772 } else if (reg_is_pkt_pointer_any(dst_reg) &&
8773 reg_is_pkt_pointer_any(src_reg) &&
8775 pred = is_pkt_ptr_branch_taken(dst_reg, src_reg, opcode);
8779 /* If we get here with a dst_reg pointer type it is because
8780 * above is_branch_taken() special cased the 0 comparison.
8782 if (!__is_pointer_value(false, dst_reg))
8783 err = mark_chain_precision(env, insn->dst_reg);
8784 if (BPF_SRC(insn->code) == BPF_X && !err &&
8785 !__is_pointer_value(false, src_reg))
8786 err = mark_chain_precision(env, insn->src_reg);
8792 /* Only follow the goto, ignore fall-through. If needed, push
8793 * the fall-through branch for simulation under speculative
8796 if (!env->bypass_spec_v1 &&
8797 !sanitize_speculative_path(env, insn, *insn_idx + 1,
8800 *insn_idx += insn->off;
8802 } else if (pred == 0) {
8803 /* Only follow the fall-through branch, since that's where the
8804 * program will go. If needed, push the goto branch for
8805 * simulation under speculative execution.
8807 if (!env->bypass_spec_v1 &&
8808 !sanitize_speculative_path(env, insn,
8809 *insn_idx + insn->off + 1,
8815 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
8819 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
8821 /* detect if we are comparing against a constant value so we can adjust
8822 * our min/max values for our dst register.
8823 * this is only legit if both are scalars (or pointers to the same
8824 * object, I suppose, but we don't support that right now), because
8825 * otherwise the different base pointers mean the offsets aren't
8828 if (BPF_SRC(insn->code) == BPF_X) {
8829 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
8831 if (dst_reg->type == SCALAR_VALUE &&
8832 src_reg->type == SCALAR_VALUE) {
8833 if (tnum_is_const(src_reg->var_off) ||
8835 tnum_is_const(tnum_subreg(src_reg->var_off))))
8836 reg_set_min_max(&other_branch_regs[insn->dst_reg],
8838 src_reg->var_off.value,
8839 tnum_subreg(src_reg->var_off).value,
8841 else if (tnum_is_const(dst_reg->var_off) ||
8843 tnum_is_const(tnum_subreg(dst_reg->var_off))))
8844 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
8846 dst_reg->var_off.value,
8847 tnum_subreg(dst_reg->var_off).value,
8849 else if (!is_jmp32 &&
8850 (opcode == BPF_JEQ || opcode == BPF_JNE))
8851 /* Comparing for equality, we can combine knowledge */
8852 reg_combine_min_max(&other_branch_regs[insn->src_reg],
8853 &other_branch_regs[insn->dst_reg],
8854 src_reg, dst_reg, opcode);
8856 !WARN_ON_ONCE(src_reg->id != other_branch_regs[insn->src_reg].id)) {
8857 find_equal_scalars(this_branch, src_reg);
8858 find_equal_scalars(other_branch, &other_branch_regs[insn->src_reg]);
8862 } else if (dst_reg->type == SCALAR_VALUE) {
8863 reg_set_min_max(&other_branch_regs[insn->dst_reg],
8864 dst_reg, insn->imm, (u32)insn->imm,
8868 if (dst_reg->type == SCALAR_VALUE && dst_reg->id &&
8869 !WARN_ON_ONCE(dst_reg->id != other_branch_regs[insn->dst_reg].id)) {
8870 find_equal_scalars(this_branch, dst_reg);
8871 find_equal_scalars(other_branch, &other_branch_regs[insn->dst_reg]);
8874 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
8875 * NOTE: these optimizations below are related with pointer comparison
8876 * which will never be JMP32.
8878 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
8879 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
8880 reg_type_may_be_null(dst_reg->type)) {
8881 /* Mark all identical registers in each branch as either
8882 * safe or unknown depending R == 0 or R != 0 conditional.
8884 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
8886 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
8888 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
8889 this_branch, other_branch) &&
8890 is_pointer_value(env, insn->dst_reg)) {
8891 verbose(env, "R%d pointer comparison prohibited\n",
8895 if (env->log.level & BPF_LOG_LEVEL)
8896 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
8900 /* verify BPF_LD_IMM64 instruction */
8901 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
8903 struct bpf_insn_aux_data *aux = cur_aux(env);
8904 struct bpf_reg_state *regs = cur_regs(env);
8905 struct bpf_reg_state *dst_reg;
8906 struct bpf_map *map;
8909 if (BPF_SIZE(insn->code) != BPF_DW) {
8910 verbose(env, "invalid BPF_LD_IMM insn\n");
8913 if (insn->off != 0) {
8914 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
8918 err = check_reg_arg(env, insn->dst_reg, DST_OP);
8922 dst_reg = ®s[insn->dst_reg];
8923 if (insn->src_reg == 0) {
8924 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
8926 dst_reg->type = SCALAR_VALUE;
8927 __mark_reg_known(®s[insn->dst_reg], imm);
8931 if (insn->src_reg == BPF_PSEUDO_BTF_ID) {
8932 mark_reg_known_zero(env, regs, insn->dst_reg);
8934 dst_reg->type = aux->btf_var.reg_type;
8935 switch (dst_reg->type) {
8937 dst_reg->mem_size = aux->btf_var.mem_size;
8940 case PTR_TO_PERCPU_BTF_ID:
8941 dst_reg->btf = aux->btf_var.btf;
8942 dst_reg->btf_id = aux->btf_var.btf_id;
8945 verbose(env, "bpf verifier is misconfigured\n");
8951 if (insn->src_reg == BPF_PSEUDO_FUNC) {
8952 struct bpf_prog_aux *aux = env->prog->aux;
8953 u32 subprogno = insn[1].imm;
8955 if (!aux->func_info) {
8956 verbose(env, "missing btf func_info\n");
8959 if (aux->func_info_aux[subprogno].linkage != BTF_FUNC_STATIC) {
8960 verbose(env, "callback function not static\n");
8964 dst_reg->type = PTR_TO_FUNC;
8965 dst_reg->subprogno = subprogno;
8969 map = env->used_maps[aux->map_index];
8970 mark_reg_known_zero(env, regs, insn->dst_reg);
8971 dst_reg->map_ptr = map;
8973 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE ||
8974 insn->src_reg == BPF_PSEUDO_MAP_IDX_VALUE) {
8975 dst_reg->type = PTR_TO_MAP_VALUE;
8976 dst_reg->off = aux->map_off;
8977 if (map_value_has_spin_lock(map))
8978 dst_reg->id = ++env->id_gen;
8979 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD ||
8980 insn->src_reg == BPF_PSEUDO_MAP_IDX) {
8981 dst_reg->type = CONST_PTR_TO_MAP;
8983 verbose(env, "bpf verifier is misconfigured\n");
8990 static bool may_access_skb(enum bpf_prog_type type)
8993 case BPF_PROG_TYPE_SOCKET_FILTER:
8994 case BPF_PROG_TYPE_SCHED_CLS:
8995 case BPF_PROG_TYPE_SCHED_ACT:
9002 /* verify safety of LD_ABS|LD_IND instructions:
9003 * - they can only appear in the programs where ctx == skb
9004 * - since they are wrappers of function calls, they scratch R1-R5 registers,
9005 * preserve R6-R9, and store return value into R0
9008 * ctx == skb == R6 == CTX
9011 * SRC == any register
9012 * IMM == 32-bit immediate
9015 * R0 - 8/16/32-bit skb data converted to cpu endianness
9017 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
9019 struct bpf_reg_state *regs = cur_regs(env);
9020 static const int ctx_reg = BPF_REG_6;
9021 u8 mode = BPF_MODE(insn->code);
9024 if (!may_access_skb(resolve_prog_type(env->prog))) {
9025 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
9029 if (!env->ops->gen_ld_abs) {
9030 verbose(env, "bpf verifier is misconfigured\n");
9034 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
9035 BPF_SIZE(insn->code) == BPF_DW ||
9036 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
9037 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
9041 /* check whether implicit source operand (register R6) is readable */
9042 err = check_reg_arg(env, ctx_reg, SRC_OP);
9046 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
9047 * gen_ld_abs() may terminate the program at runtime, leading to
9050 err = check_reference_leak(env);
9052 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
9056 if (env->cur_state->active_spin_lock) {
9057 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
9061 if (regs[ctx_reg].type != PTR_TO_CTX) {
9063 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
9067 if (mode == BPF_IND) {
9068 /* check explicit source operand */
9069 err = check_reg_arg(env, insn->src_reg, SRC_OP);
9074 err = check_ctx_reg(env, ®s[ctx_reg], ctx_reg);
9078 /* reset caller saved regs to unreadable */
9079 for (i = 0; i < CALLER_SAVED_REGS; i++) {
9080 mark_reg_not_init(env, regs, caller_saved[i]);
9081 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
9084 /* mark destination R0 register as readable, since it contains
9085 * the value fetched from the packet.
9086 * Already marked as written above.
9088 mark_reg_unknown(env, regs, BPF_REG_0);
9089 /* ld_abs load up to 32-bit skb data. */
9090 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
9094 static int check_return_code(struct bpf_verifier_env *env)
9096 struct tnum enforce_attach_type_range = tnum_unknown;
9097 const struct bpf_prog *prog = env->prog;
9098 struct bpf_reg_state *reg;
9099 struct tnum range = tnum_range(0, 1);
9100 enum bpf_prog_type prog_type = resolve_prog_type(env->prog);
9102 const bool is_subprog = env->cur_state->frame[0]->subprogno;
9104 /* LSM and struct_ops func-ptr's return type could be "void" */
9106 (prog_type == BPF_PROG_TYPE_STRUCT_OPS ||
9107 prog_type == BPF_PROG_TYPE_LSM) &&
9108 !prog->aux->attach_func_proto->type)
9111 /* eBPF calling convention is such that R0 is used
9112 * to return the value from eBPF program.
9113 * Make sure that it's readable at this time
9114 * of bpf_exit, which means that program wrote
9115 * something into it earlier
9117 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
9121 if (is_pointer_value(env, BPF_REG_0)) {
9122 verbose(env, "R0 leaks addr as return value\n");
9126 reg = cur_regs(env) + BPF_REG_0;
9128 if (reg->type != SCALAR_VALUE) {
9129 verbose(env, "At subprogram exit the register R0 is not a scalar value (%s)\n",
9130 reg_type_str[reg->type]);
9136 switch (prog_type) {
9137 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
9138 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
9139 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG ||
9140 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETPEERNAME ||
9141 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETPEERNAME ||
9142 env->prog->expected_attach_type == BPF_CGROUP_INET4_GETSOCKNAME ||
9143 env->prog->expected_attach_type == BPF_CGROUP_INET6_GETSOCKNAME)
9144 range = tnum_range(1, 1);
9145 if (env->prog->expected_attach_type == BPF_CGROUP_INET4_BIND ||
9146 env->prog->expected_attach_type == BPF_CGROUP_INET6_BIND)
9147 range = tnum_range(0, 3);
9149 case BPF_PROG_TYPE_CGROUP_SKB:
9150 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
9151 range = tnum_range(0, 3);
9152 enforce_attach_type_range = tnum_range(2, 3);
9155 case BPF_PROG_TYPE_CGROUP_SOCK:
9156 case BPF_PROG_TYPE_SOCK_OPS:
9157 case BPF_PROG_TYPE_CGROUP_DEVICE:
9158 case BPF_PROG_TYPE_CGROUP_SYSCTL:
9159 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
9161 case BPF_PROG_TYPE_RAW_TRACEPOINT:
9162 if (!env->prog->aux->attach_btf_id)
9164 range = tnum_const(0);
9166 case BPF_PROG_TYPE_TRACING:
9167 switch (env->prog->expected_attach_type) {
9168 case BPF_TRACE_FENTRY:
9169 case BPF_TRACE_FEXIT:
9170 range = tnum_const(0);
9172 case BPF_TRACE_RAW_TP:
9173 case BPF_MODIFY_RETURN:
9175 case BPF_TRACE_ITER:
9181 case BPF_PROG_TYPE_SK_LOOKUP:
9182 range = tnum_range(SK_DROP, SK_PASS);
9184 case BPF_PROG_TYPE_EXT:
9185 /* freplace program can return anything as its return value
9186 * depends on the to-be-replaced kernel func or bpf program.
9192 if (reg->type != SCALAR_VALUE) {
9193 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
9194 reg_type_str[reg->type]);
9198 if (!tnum_in(range, reg->var_off)) {
9199 verbose_invalid_scalar(env, reg, &range, "program exit", "R0");
9203 if (!tnum_is_unknown(enforce_attach_type_range) &&
9204 tnum_in(enforce_attach_type_range, reg->var_off))
9205 env->prog->enforce_expected_attach_type = 1;
9209 /* non-recursive DFS pseudo code
9210 * 1 procedure DFS-iterative(G,v):
9211 * 2 label v as discovered
9212 * 3 let S be a stack
9214 * 5 while S is not empty
9216 * 7 if t is what we're looking for:
9218 * 9 for all edges e in G.adjacentEdges(t) do
9219 * 10 if edge e is already labelled
9220 * 11 continue with the next edge
9221 * 12 w <- G.adjacentVertex(t,e)
9222 * 13 if vertex w is not discovered and not explored
9223 * 14 label e as tree-edge
9224 * 15 label w as discovered
9227 * 18 else if vertex w is discovered
9228 * 19 label e as back-edge
9230 * 21 // vertex w is explored
9231 * 22 label e as forward- or cross-edge
9232 * 23 label t as explored
9237 * 0x11 - discovered and fall-through edge labelled
9238 * 0x12 - discovered and fall-through and branch edges labelled
9249 static u32 state_htab_size(struct bpf_verifier_env *env)
9251 return env->prog->len;
9254 static struct bpf_verifier_state_list **explored_state(
9255 struct bpf_verifier_env *env,
9258 struct bpf_verifier_state *cur = env->cur_state;
9259 struct bpf_func_state *state = cur->frame[cur->curframe];
9261 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
9264 static void init_explored_state(struct bpf_verifier_env *env, int idx)
9266 env->insn_aux_data[idx].prune_point = true;
9274 /* t, w, e - match pseudo-code above:
9275 * t - index of current instruction
9276 * w - next instruction
9279 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env,
9282 int *insn_stack = env->cfg.insn_stack;
9283 int *insn_state = env->cfg.insn_state;
9285 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
9286 return DONE_EXPLORING;
9288 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
9289 return DONE_EXPLORING;
9291 if (w < 0 || w >= env->prog->len) {
9292 verbose_linfo(env, t, "%d: ", t);
9293 verbose(env, "jump out of range from insn %d to %d\n", t, w);
9298 /* mark branch target for state pruning */
9299 init_explored_state(env, w);
9301 if (insn_state[w] == 0) {
9303 insn_state[t] = DISCOVERED | e;
9304 insn_state[w] = DISCOVERED;
9305 if (env->cfg.cur_stack >= env->prog->len)
9307 insn_stack[env->cfg.cur_stack++] = w;
9308 return KEEP_EXPLORING;
9309 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
9310 if (loop_ok && env->bpf_capable)
9311 return DONE_EXPLORING;
9312 verbose_linfo(env, t, "%d: ", t);
9313 verbose_linfo(env, w, "%d: ", w);
9314 verbose(env, "back-edge from insn %d to %d\n", t, w);
9316 } else if (insn_state[w] == EXPLORED) {
9317 /* forward- or cross-edge */
9318 insn_state[t] = DISCOVERED | e;
9320 verbose(env, "insn state internal bug\n");
9323 return DONE_EXPLORING;
9326 static int visit_func_call_insn(int t, int insn_cnt,
9327 struct bpf_insn *insns,
9328 struct bpf_verifier_env *env,
9333 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
9337 if (t + 1 < insn_cnt)
9338 init_explored_state(env, t + 1);
9340 init_explored_state(env, t);
9341 ret = push_insn(t, t + insns[t].imm + 1, BRANCH,
9347 /* Visits the instruction at index t and returns one of the following:
9348 * < 0 - an error occurred
9349 * DONE_EXPLORING - the instruction was fully explored
9350 * KEEP_EXPLORING - there is still work to be done before it is fully explored
9352 static int visit_insn(int t, int insn_cnt, struct bpf_verifier_env *env)
9354 struct bpf_insn *insns = env->prog->insnsi;
9357 if (bpf_pseudo_func(insns + t))
9358 return visit_func_call_insn(t, insn_cnt, insns, env, true);
9360 /* All non-branch instructions have a single fall-through edge. */
9361 if (BPF_CLASS(insns[t].code) != BPF_JMP &&
9362 BPF_CLASS(insns[t].code) != BPF_JMP32)
9363 return push_insn(t, t + 1, FALLTHROUGH, env, false);
9365 switch (BPF_OP(insns[t].code)) {
9367 return DONE_EXPLORING;
9370 return visit_func_call_insn(t, insn_cnt, insns, env,
9371 insns[t].src_reg == BPF_PSEUDO_CALL);
9374 if (BPF_SRC(insns[t].code) != BPF_K)
9377 /* unconditional jump with single edge */
9378 ret = push_insn(t, t + insns[t].off + 1, FALLTHROUGH, env,
9383 /* unconditional jmp is not a good pruning point,
9384 * but it's marked, since backtracking needs
9385 * to record jmp history in is_state_visited().
9387 init_explored_state(env, t + insns[t].off + 1);
9388 /* tell verifier to check for equivalent states
9389 * after every call and jump
9391 if (t + 1 < insn_cnt)
9392 init_explored_state(env, t + 1);
9397 /* conditional jump with two edges */
9398 init_explored_state(env, t);
9399 ret = push_insn(t, t + 1, FALLTHROUGH, env, true);
9403 return push_insn(t, t + insns[t].off + 1, BRANCH, env, true);
9407 /* non-recursive depth-first-search to detect loops in BPF program
9408 * loop == back-edge in directed graph
9410 static int check_cfg(struct bpf_verifier_env *env)
9412 int insn_cnt = env->prog->len;
9413 int *insn_stack, *insn_state;
9417 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
9421 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
9427 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
9428 insn_stack[0] = 0; /* 0 is the first instruction */
9429 env->cfg.cur_stack = 1;
9431 while (env->cfg.cur_stack > 0) {
9432 int t = insn_stack[env->cfg.cur_stack - 1];
9434 ret = visit_insn(t, insn_cnt, env);
9436 case DONE_EXPLORING:
9437 insn_state[t] = EXPLORED;
9438 env->cfg.cur_stack--;
9440 case KEEP_EXPLORING:
9444 verbose(env, "visit_insn internal bug\n");
9451 if (env->cfg.cur_stack < 0) {
9452 verbose(env, "pop stack internal bug\n");
9457 for (i = 0; i < insn_cnt; i++) {
9458 if (insn_state[i] != EXPLORED) {
9459 verbose(env, "unreachable insn %d\n", i);
9464 ret = 0; /* cfg looks good */
9469 env->cfg.insn_state = env->cfg.insn_stack = NULL;
9473 static int check_abnormal_return(struct bpf_verifier_env *env)
9477 for (i = 1; i < env->subprog_cnt; i++) {
9478 if (env->subprog_info[i].has_ld_abs) {
9479 verbose(env, "LD_ABS is not allowed in subprogs without BTF\n");
9482 if (env->subprog_info[i].has_tail_call) {
9483 verbose(env, "tail_call is not allowed in subprogs without BTF\n");
9490 /* The minimum supported BTF func info size */
9491 #define MIN_BPF_FUNCINFO_SIZE 8
9492 #define MAX_FUNCINFO_REC_SIZE 252
9494 static int check_btf_func(struct bpf_verifier_env *env,
9495 const union bpf_attr *attr,
9498 const struct btf_type *type, *func_proto, *ret_type;
9499 u32 i, nfuncs, urec_size, min_size;
9500 u32 krec_size = sizeof(struct bpf_func_info);
9501 struct bpf_func_info *krecord;
9502 struct bpf_func_info_aux *info_aux = NULL;
9503 struct bpf_prog *prog;
9504 const struct btf *btf;
9506 u32 prev_offset = 0;
9510 nfuncs = attr->func_info_cnt;
9512 if (check_abnormal_return(env))
9517 if (nfuncs != env->subprog_cnt) {
9518 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
9522 urec_size = attr->func_info_rec_size;
9523 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
9524 urec_size > MAX_FUNCINFO_REC_SIZE ||
9525 urec_size % sizeof(u32)) {
9526 verbose(env, "invalid func info rec size %u\n", urec_size);
9531 btf = prog->aux->btf;
9533 urecord = make_bpfptr(attr->func_info, uattr.is_kernel);
9534 min_size = min_t(u32, krec_size, urec_size);
9536 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
9539 info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN);
9543 for (i = 0; i < nfuncs; i++) {
9544 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
9546 if (ret == -E2BIG) {
9547 verbose(env, "nonzero tailing record in func info");
9548 /* set the size kernel expects so loader can zero
9549 * out the rest of the record.
9551 if (copy_to_bpfptr_offset(uattr,
9552 offsetof(union bpf_attr, func_info_rec_size),
9553 &min_size, sizeof(min_size)))
9559 if (copy_from_bpfptr(&krecord[i], urecord, min_size)) {
9564 /* check insn_off */
9567 if (krecord[i].insn_off) {
9569 "nonzero insn_off %u for the first func info record",
9570 krecord[i].insn_off);
9573 } else if (krecord[i].insn_off <= prev_offset) {
9575 "same or smaller insn offset (%u) than previous func info record (%u)",
9576 krecord[i].insn_off, prev_offset);
9580 if (env->subprog_info[i].start != krecord[i].insn_off) {
9581 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
9586 type = btf_type_by_id(btf, krecord[i].type_id);
9587 if (!type || !btf_type_is_func(type)) {
9588 verbose(env, "invalid type id %d in func info",
9589 krecord[i].type_id);
9592 info_aux[i].linkage = BTF_INFO_VLEN(type->info);
9594 func_proto = btf_type_by_id(btf, type->type);
9595 if (unlikely(!func_proto || !btf_type_is_func_proto(func_proto)))
9596 /* btf_func_check() already verified it during BTF load */
9598 ret_type = btf_type_skip_modifiers(btf, func_proto->type, NULL);
9600 btf_type_is_small_int(ret_type) || btf_type_is_enum(ret_type);
9601 if (i && !scalar_return && env->subprog_info[i].has_ld_abs) {
9602 verbose(env, "LD_ABS is only allowed in functions that return 'int'.\n");
9605 if (i && !scalar_return && env->subprog_info[i].has_tail_call) {
9606 verbose(env, "tail_call is only allowed in functions that return 'int'.\n");
9610 prev_offset = krecord[i].insn_off;
9611 bpfptr_add(&urecord, urec_size);
9614 prog->aux->func_info = krecord;
9615 prog->aux->func_info_cnt = nfuncs;
9616 prog->aux->func_info_aux = info_aux;
9625 static void adjust_btf_func(struct bpf_verifier_env *env)
9627 struct bpf_prog_aux *aux = env->prog->aux;
9630 if (!aux->func_info)
9633 for (i = 0; i < env->subprog_cnt; i++)
9634 aux->func_info[i].insn_off = env->subprog_info[i].start;
9637 #define MIN_BPF_LINEINFO_SIZE (offsetof(struct bpf_line_info, line_col) + \
9638 sizeof(((struct bpf_line_info *)(0))->line_col))
9639 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
9641 static int check_btf_line(struct bpf_verifier_env *env,
9642 const union bpf_attr *attr,
9645 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
9646 struct bpf_subprog_info *sub;
9647 struct bpf_line_info *linfo;
9648 struct bpf_prog *prog;
9649 const struct btf *btf;
9653 nr_linfo = attr->line_info_cnt;
9657 rec_size = attr->line_info_rec_size;
9658 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
9659 rec_size > MAX_LINEINFO_REC_SIZE ||
9660 rec_size & (sizeof(u32) - 1))
9663 /* Need to zero it in case the userspace may
9664 * pass in a smaller bpf_line_info object.
9666 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
9667 GFP_KERNEL | __GFP_NOWARN);
9672 btf = prog->aux->btf;
9675 sub = env->subprog_info;
9676 ulinfo = make_bpfptr(attr->line_info, uattr.is_kernel);
9677 expected_size = sizeof(struct bpf_line_info);
9678 ncopy = min_t(u32, expected_size, rec_size);
9679 for (i = 0; i < nr_linfo; i++) {
9680 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
9682 if (err == -E2BIG) {
9683 verbose(env, "nonzero tailing record in line_info");
9684 if (copy_to_bpfptr_offset(uattr,
9685 offsetof(union bpf_attr, line_info_rec_size),
9686 &expected_size, sizeof(expected_size)))
9692 if (copy_from_bpfptr(&linfo[i], ulinfo, ncopy)) {
9698 * Check insn_off to ensure
9699 * 1) strictly increasing AND
9700 * 2) bounded by prog->len
9702 * The linfo[0].insn_off == 0 check logically falls into
9703 * the later "missing bpf_line_info for func..." case
9704 * because the first linfo[0].insn_off must be the
9705 * first sub also and the first sub must have
9706 * subprog_info[0].start == 0.
9708 if ((i && linfo[i].insn_off <= prev_offset) ||
9709 linfo[i].insn_off >= prog->len) {
9710 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
9711 i, linfo[i].insn_off, prev_offset,
9717 if (!prog->insnsi[linfo[i].insn_off].code) {
9719 "Invalid insn code at line_info[%u].insn_off\n",
9725 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
9726 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
9727 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
9732 if (s != env->subprog_cnt) {
9733 if (linfo[i].insn_off == sub[s].start) {
9734 sub[s].linfo_idx = i;
9736 } else if (sub[s].start < linfo[i].insn_off) {
9737 verbose(env, "missing bpf_line_info for func#%u\n", s);
9743 prev_offset = linfo[i].insn_off;
9744 bpfptr_add(&ulinfo, rec_size);
9747 if (s != env->subprog_cnt) {
9748 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
9749 env->subprog_cnt - s, s);
9754 prog->aux->linfo = linfo;
9755 prog->aux->nr_linfo = nr_linfo;
9764 static int check_btf_info(struct bpf_verifier_env *env,
9765 const union bpf_attr *attr,
9771 if (!attr->func_info_cnt && !attr->line_info_cnt) {
9772 if (check_abnormal_return(env))
9777 btf = btf_get_by_fd(attr->prog_btf_fd);
9779 return PTR_ERR(btf);
9780 if (btf_is_kernel(btf)) {
9784 env->prog->aux->btf = btf;
9786 err = check_btf_func(env, attr, uattr);
9790 err = check_btf_line(env, attr, uattr);
9797 /* check %cur's range satisfies %old's */
9798 static bool range_within(struct bpf_reg_state *old,
9799 struct bpf_reg_state *cur)
9801 return old->umin_value <= cur->umin_value &&
9802 old->umax_value >= cur->umax_value &&
9803 old->smin_value <= cur->smin_value &&
9804 old->smax_value >= cur->smax_value &&
9805 old->u32_min_value <= cur->u32_min_value &&
9806 old->u32_max_value >= cur->u32_max_value &&
9807 old->s32_min_value <= cur->s32_min_value &&
9808 old->s32_max_value >= cur->s32_max_value;
9811 /* If in the old state two registers had the same id, then they need to have
9812 * the same id in the new state as well. But that id could be different from
9813 * the old state, so we need to track the mapping from old to new ids.
9814 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
9815 * regs with old id 5 must also have new id 9 for the new state to be safe. But
9816 * regs with a different old id could still have new id 9, we don't care about
9818 * So we look through our idmap to see if this old id has been seen before. If
9819 * so, we require the new id to match; otherwise, we add the id pair to the map.
9821 static bool check_ids(u32 old_id, u32 cur_id, struct bpf_id_pair *idmap)
9825 for (i = 0; i < BPF_ID_MAP_SIZE; i++) {
9826 if (!idmap[i].old) {
9827 /* Reached an empty slot; haven't seen this id before */
9828 idmap[i].old = old_id;
9829 idmap[i].cur = cur_id;
9832 if (idmap[i].old == old_id)
9833 return idmap[i].cur == cur_id;
9835 /* We ran out of idmap slots, which should be impossible */
9840 static void clean_func_state(struct bpf_verifier_env *env,
9841 struct bpf_func_state *st)
9843 enum bpf_reg_liveness live;
9846 for (i = 0; i < BPF_REG_FP; i++) {
9847 live = st->regs[i].live;
9848 /* liveness must not touch this register anymore */
9849 st->regs[i].live |= REG_LIVE_DONE;
9850 if (!(live & REG_LIVE_READ))
9851 /* since the register is unused, clear its state
9852 * to make further comparison simpler
9854 __mark_reg_not_init(env, &st->regs[i]);
9857 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
9858 live = st->stack[i].spilled_ptr.live;
9859 /* liveness must not touch this stack slot anymore */
9860 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
9861 if (!(live & REG_LIVE_READ)) {
9862 __mark_reg_not_init(env, &st->stack[i].spilled_ptr);
9863 for (j = 0; j < BPF_REG_SIZE; j++)
9864 st->stack[i].slot_type[j] = STACK_INVALID;
9869 static void clean_verifier_state(struct bpf_verifier_env *env,
9870 struct bpf_verifier_state *st)
9874 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
9875 /* all regs in this state in all frames were already marked */
9878 for (i = 0; i <= st->curframe; i++)
9879 clean_func_state(env, st->frame[i]);
9882 /* the parentage chains form a tree.
9883 * the verifier states are added to state lists at given insn and
9884 * pushed into state stack for future exploration.
9885 * when the verifier reaches bpf_exit insn some of the verifer states
9886 * stored in the state lists have their final liveness state already,
9887 * but a lot of states will get revised from liveness point of view when
9888 * the verifier explores other branches.
9891 * 2: if r1 == 100 goto pc+1
9894 * when the verifier reaches exit insn the register r0 in the state list of
9895 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
9896 * of insn 2 and goes exploring further. At the insn 4 it will walk the
9897 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
9899 * Since the verifier pushes the branch states as it sees them while exploring
9900 * the program the condition of walking the branch instruction for the second
9901 * time means that all states below this branch were already explored and
9902 * their final liveness marks are already propagated.
9903 * Hence when the verifier completes the search of state list in is_state_visited()
9904 * we can call this clean_live_states() function to mark all liveness states
9905 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
9907 * This function also clears the registers and stack for states that !READ
9908 * to simplify state merging.
9910 * Important note here that walking the same branch instruction in the callee
9911 * doesn't meant that the states are DONE. The verifier has to compare
9914 static void clean_live_states(struct bpf_verifier_env *env, int insn,
9915 struct bpf_verifier_state *cur)
9917 struct bpf_verifier_state_list *sl;
9920 sl = *explored_state(env, insn);
9922 if (sl->state.branches)
9924 if (sl->state.insn_idx != insn ||
9925 sl->state.curframe != cur->curframe)
9927 for (i = 0; i <= cur->curframe; i++)
9928 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
9930 clean_verifier_state(env, &sl->state);
9936 /* Returns true if (rold safe implies rcur safe) */
9937 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
9938 struct bpf_id_pair *idmap)
9942 if (!(rold->live & REG_LIVE_READ))
9943 /* explored state didn't use this */
9946 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
9948 if (rold->type == PTR_TO_STACK)
9949 /* two stack pointers are equal only if they're pointing to
9950 * the same stack frame, since fp-8 in foo != fp-8 in bar
9952 return equal && rold->frameno == rcur->frameno;
9957 if (rold->type == NOT_INIT)
9958 /* explored state can't have used this */
9960 if (rcur->type == NOT_INIT)
9962 switch (rold->type) {
9964 if (rcur->type == SCALAR_VALUE) {
9965 if (!rold->precise && !rcur->precise)
9967 /* new val must satisfy old val knowledge */
9968 return range_within(rold, rcur) &&
9969 tnum_in(rold->var_off, rcur->var_off);
9971 /* We're trying to use a pointer in place of a scalar.
9972 * Even if the scalar was unbounded, this could lead to
9973 * pointer leaks because scalars are allowed to leak
9974 * while pointers are not. We could make this safe in
9975 * special cases if root is calling us, but it's
9976 * probably not worth the hassle.
9980 case PTR_TO_MAP_KEY:
9981 case PTR_TO_MAP_VALUE:
9982 /* If the new min/max/var_off satisfy the old ones and
9983 * everything else matches, we are OK.
9984 * 'id' is not compared, since it's only used for maps with
9985 * bpf_spin_lock inside map element and in such cases if
9986 * the rest of the prog is valid for one map element then
9987 * it's valid for all map elements regardless of the key
9988 * used in bpf_map_lookup()
9990 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
9991 range_within(rold, rcur) &&
9992 tnum_in(rold->var_off, rcur->var_off);
9993 case PTR_TO_MAP_VALUE_OR_NULL:
9994 /* a PTR_TO_MAP_VALUE could be safe to use as a
9995 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
9996 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
9997 * checked, doing so could have affected others with the same
9998 * id, and we can't check for that because we lost the id when
9999 * we converted to a PTR_TO_MAP_VALUE.
10001 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
10003 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
10005 /* Check our ids match any regs they're supposed to */
10006 return check_ids(rold->id, rcur->id, idmap);
10007 case PTR_TO_PACKET_META:
10008 case PTR_TO_PACKET:
10009 if (rcur->type != rold->type)
10011 /* We must have at least as much range as the old ptr
10012 * did, so that any accesses which were safe before are
10013 * still safe. This is true even if old range < old off,
10014 * since someone could have accessed through (ptr - k), or
10015 * even done ptr -= k in a register, to get a safe access.
10017 if (rold->range > rcur->range)
10019 /* If the offsets don't match, we can't trust our alignment;
10020 * nor can we be sure that we won't fall out of range.
10022 if (rold->off != rcur->off)
10024 /* id relations must be preserved */
10025 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
10027 /* new val must satisfy old val knowledge */
10028 return range_within(rold, rcur) &&
10029 tnum_in(rold->var_off, rcur->var_off);
10031 case CONST_PTR_TO_MAP:
10032 case PTR_TO_PACKET_END:
10033 case PTR_TO_FLOW_KEYS:
10034 case PTR_TO_SOCKET:
10035 case PTR_TO_SOCKET_OR_NULL:
10036 case PTR_TO_SOCK_COMMON:
10037 case PTR_TO_SOCK_COMMON_OR_NULL:
10038 case PTR_TO_TCP_SOCK:
10039 case PTR_TO_TCP_SOCK_OR_NULL:
10040 case PTR_TO_XDP_SOCK:
10041 /* Only valid matches are exact, which memcmp() above
10042 * would have accepted
10045 /* Don't know what's going on, just say it's not safe */
10049 /* Shouldn't get here; if we do, say it's not safe */
10054 static bool stacksafe(struct bpf_func_state *old,
10055 struct bpf_func_state *cur,
10056 struct bpf_id_pair *idmap)
10060 /* walk slots of the explored stack and ignore any additional
10061 * slots in the current stack, since explored(safe) state
10064 for (i = 0; i < old->allocated_stack; i++) {
10065 spi = i / BPF_REG_SIZE;
10067 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
10068 i += BPF_REG_SIZE - 1;
10069 /* explored state didn't use this */
10073 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
10076 /* explored stack has more populated slots than current stack
10077 * and these slots were used
10079 if (i >= cur->allocated_stack)
10082 /* if old state was safe with misc data in the stack
10083 * it will be safe with zero-initialized stack.
10084 * The opposite is not true
10086 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
10087 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
10089 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
10090 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
10091 /* Ex: old explored (safe) state has STACK_SPILL in
10092 * this stack slot, but current has STACK_MISC ->
10093 * this verifier states are not equivalent,
10094 * return false to continue verification of this path
10097 if (i % BPF_REG_SIZE)
10099 if (old->stack[spi].slot_type[0] != STACK_SPILL)
10101 if (!regsafe(&old->stack[spi].spilled_ptr,
10102 &cur->stack[spi].spilled_ptr,
10104 /* when explored and current stack slot are both storing
10105 * spilled registers, check that stored pointers types
10106 * are the same as well.
10107 * Ex: explored safe path could have stored
10108 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
10109 * but current path has stored:
10110 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
10111 * such verifier states are not equivalent.
10112 * return false to continue verification of this path
10119 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
10121 if (old->acquired_refs != cur->acquired_refs)
10123 return !memcmp(old->refs, cur->refs,
10124 sizeof(*old->refs) * old->acquired_refs);
10127 /* compare two verifier states
10129 * all states stored in state_list are known to be valid, since
10130 * verifier reached 'bpf_exit' instruction through them
10132 * this function is called when verifier exploring different branches of
10133 * execution popped from the state stack. If it sees an old state that has
10134 * more strict register state and more strict stack state then this execution
10135 * branch doesn't need to be explored further, since verifier already
10136 * concluded that more strict state leads to valid finish.
10138 * Therefore two states are equivalent if register state is more conservative
10139 * and explored stack state is more conservative than the current one.
10142 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
10143 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
10145 * In other words if current stack state (one being explored) has more
10146 * valid slots than old one that already passed validation, it means
10147 * the verifier can stop exploring and conclude that current state is valid too
10149 * Similarly with registers. If explored state has register type as invalid
10150 * whereas register type in current state is meaningful, it means that
10151 * the current state will reach 'bpf_exit' instruction safely
10153 static bool func_states_equal(struct bpf_verifier_env *env, struct bpf_func_state *old,
10154 struct bpf_func_state *cur)
10158 memset(env->idmap_scratch, 0, sizeof(env->idmap_scratch));
10159 for (i = 0; i < MAX_BPF_REG; i++)
10160 if (!regsafe(&old->regs[i], &cur->regs[i], env->idmap_scratch))
10163 if (!stacksafe(old, cur, env->idmap_scratch))
10166 if (!refsafe(old, cur))
10172 static bool states_equal(struct bpf_verifier_env *env,
10173 struct bpf_verifier_state *old,
10174 struct bpf_verifier_state *cur)
10178 if (old->curframe != cur->curframe)
10181 /* Verification state from speculative execution simulation
10182 * must never prune a non-speculative execution one.
10184 if (old->speculative && !cur->speculative)
10187 if (old->active_spin_lock != cur->active_spin_lock)
10190 /* for states to be equal callsites have to be the same
10191 * and all frame states need to be equivalent
10193 for (i = 0; i <= old->curframe; i++) {
10194 if (old->frame[i]->callsite != cur->frame[i]->callsite)
10196 if (!func_states_equal(env, old->frame[i], cur->frame[i]))
10202 /* Return 0 if no propagation happened. Return negative error code if error
10203 * happened. Otherwise, return the propagated bit.
10205 static int propagate_liveness_reg(struct bpf_verifier_env *env,
10206 struct bpf_reg_state *reg,
10207 struct bpf_reg_state *parent_reg)
10209 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
10210 u8 flag = reg->live & REG_LIVE_READ;
10213 /* When comes here, read flags of PARENT_REG or REG could be any of
10214 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
10215 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
10217 if (parent_flag == REG_LIVE_READ64 ||
10218 /* Or if there is no read flag from REG. */
10220 /* Or if the read flag from REG is the same as PARENT_REG. */
10221 parent_flag == flag)
10224 err = mark_reg_read(env, reg, parent_reg, flag);
10231 /* A write screens off any subsequent reads; but write marks come from the
10232 * straight-line code between a state and its parent. When we arrive at an
10233 * equivalent state (jump target or such) we didn't arrive by the straight-line
10234 * code, so read marks in the state must propagate to the parent regardless
10235 * of the state's write marks. That's what 'parent == state->parent' comparison
10236 * in mark_reg_read() is for.
10238 static int propagate_liveness(struct bpf_verifier_env *env,
10239 const struct bpf_verifier_state *vstate,
10240 struct bpf_verifier_state *vparent)
10242 struct bpf_reg_state *state_reg, *parent_reg;
10243 struct bpf_func_state *state, *parent;
10244 int i, frame, err = 0;
10246 if (vparent->curframe != vstate->curframe) {
10247 WARN(1, "propagate_live: parent frame %d current frame %d\n",
10248 vparent->curframe, vstate->curframe);
10251 /* Propagate read liveness of registers... */
10252 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
10253 for (frame = 0; frame <= vstate->curframe; frame++) {
10254 parent = vparent->frame[frame];
10255 state = vstate->frame[frame];
10256 parent_reg = parent->regs;
10257 state_reg = state->regs;
10258 /* We don't need to worry about FP liveness, it's read-only */
10259 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
10260 err = propagate_liveness_reg(env, &state_reg[i],
10264 if (err == REG_LIVE_READ64)
10265 mark_insn_zext(env, &parent_reg[i]);
10268 /* Propagate stack slots. */
10269 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
10270 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
10271 parent_reg = &parent->stack[i].spilled_ptr;
10272 state_reg = &state->stack[i].spilled_ptr;
10273 err = propagate_liveness_reg(env, state_reg,
10282 /* find precise scalars in the previous equivalent state and
10283 * propagate them into the current state
10285 static int propagate_precision(struct bpf_verifier_env *env,
10286 const struct bpf_verifier_state *old)
10288 struct bpf_reg_state *state_reg;
10289 struct bpf_func_state *state;
10292 state = old->frame[old->curframe];
10293 state_reg = state->regs;
10294 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
10295 if (state_reg->type != SCALAR_VALUE ||
10296 !state_reg->precise)
10298 if (env->log.level & BPF_LOG_LEVEL2)
10299 verbose(env, "propagating r%d\n", i);
10300 err = mark_chain_precision(env, i);
10305 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
10306 if (state->stack[i].slot_type[0] != STACK_SPILL)
10308 state_reg = &state->stack[i].spilled_ptr;
10309 if (state_reg->type != SCALAR_VALUE ||
10310 !state_reg->precise)
10312 if (env->log.level & BPF_LOG_LEVEL2)
10313 verbose(env, "propagating fp%d\n",
10314 (-i - 1) * BPF_REG_SIZE);
10315 err = mark_chain_precision_stack(env, i);
10322 static bool states_maybe_looping(struct bpf_verifier_state *old,
10323 struct bpf_verifier_state *cur)
10325 struct bpf_func_state *fold, *fcur;
10326 int i, fr = cur->curframe;
10328 if (old->curframe != fr)
10331 fold = old->frame[fr];
10332 fcur = cur->frame[fr];
10333 for (i = 0; i < MAX_BPF_REG; i++)
10334 if (memcmp(&fold->regs[i], &fcur->regs[i],
10335 offsetof(struct bpf_reg_state, parent)))
10341 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
10343 struct bpf_verifier_state_list *new_sl;
10344 struct bpf_verifier_state_list *sl, **pprev;
10345 struct bpf_verifier_state *cur = env->cur_state, *new;
10346 int i, j, err, states_cnt = 0;
10347 bool add_new_state = env->test_state_freq ? true : false;
10349 cur->last_insn_idx = env->prev_insn_idx;
10350 if (!env->insn_aux_data[insn_idx].prune_point)
10351 /* this 'insn_idx' instruction wasn't marked, so we will not
10352 * be doing state search here
10356 /* bpf progs typically have pruning point every 4 instructions
10357 * http://vger.kernel.org/bpfconf2019.html#session-1
10358 * Do not add new state for future pruning if the verifier hasn't seen
10359 * at least 2 jumps and at least 8 instructions.
10360 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
10361 * In tests that amounts to up to 50% reduction into total verifier
10362 * memory consumption and 20% verifier time speedup.
10364 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
10365 env->insn_processed - env->prev_insn_processed >= 8)
10366 add_new_state = true;
10368 pprev = explored_state(env, insn_idx);
10371 clean_live_states(env, insn_idx, cur);
10375 if (sl->state.insn_idx != insn_idx)
10377 if (sl->state.branches) {
10378 if (states_maybe_looping(&sl->state, cur) &&
10379 states_equal(env, &sl->state, cur)) {
10380 verbose_linfo(env, insn_idx, "; ");
10381 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
10384 /* if the verifier is processing a loop, avoid adding new state
10385 * too often, since different loop iterations have distinct
10386 * states and may not help future pruning.
10387 * This threshold shouldn't be too low to make sure that
10388 * a loop with large bound will be rejected quickly.
10389 * The most abusive loop will be:
10391 * if r1 < 1000000 goto pc-2
10392 * 1M insn_procssed limit / 100 == 10k peak states.
10393 * This threshold shouldn't be too high either, since states
10394 * at the end of the loop are likely to be useful in pruning.
10396 if (env->jmps_processed - env->prev_jmps_processed < 20 &&
10397 env->insn_processed - env->prev_insn_processed < 100)
10398 add_new_state = false;
10401 if (states_equal(env, &sl->state, cur)) {
10403 /* reached equivalent register/stack state,
10404 * prune the search.
10405 * Registers read by the continuation are read by us.
10406 * If we have any write marks in env->cur_state, they
10407 * will prevent corresponding reads in the continuation
10408 * from reaching our parent (an explored_state). Our
10409 * own state will get the read marks recorded, but
10410 * they'll be immediately forgotten as we're pruning
10411 * this state and will pop a new one.
10413 err = propagate_liveness(env, &sl->state, cur);
10415 /* if previous state reached the exit with precision and
10416 * current state is equivalent to it (except precsion marks)
10417 * the precision needs to be propagated back in
10418 * the current state.
10420 err = err ? : push_jmp_history(env, cur);
10421 err = err ? : propagate_precision(env, &sl->state);
10427 /* when new state is not going to be added do not increase miss count.
10428 * Otherwise several loop iterations will remove the state
10429 * recorded earlier. The goal of these heuristics is to have
10430 * states from some iterations of the loop (some in the beginning
10431 * and some at the end) to help pruning.
10435 /* heuristic to determine whether this state is beneficial
10436 * to keep checking from state equivalence point of view.
10437 * Higher numbers increase max_states_per_insn and verification time,
10438 * but do not meaningfully decrease insn_processed.
10440 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
10441 /* the state is unlikely to be useful. Remove it to
10442 * speed up verification
10445 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
10446 u32 br = sl->state.branches;
10449 "BUG live_done but branches_to_explore %d\n",
10451 free_verifier_state(&sl->state, false);
10453 env->peak_states--;
10455 /* cannot free this state, since parentage chain may
10456 * walk it later. Add it for free_list instead to
10457 * be freed at the end of verification
10459 sl->next = env->free_list;
10460 env->free_list = sl;
10470 if (env->max_states_per_insn < states_cnt)
10471 env->max_states_per_insn = states_cnt;
10473 if (!env->bpf_capable && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
10474 return push_jmp_history(env, cur);
10476 if (!add_new_state)
10477 return push_jmp_history(env, cur);
10479 /* There were no equivalent states, remember the current one.
10480 * Technically the current state is not proven to be safe yet,
10481 * but it will either reach outer most bpf_exit (which means it's safe)
10482 * or it will be rejected. When there are no loops the verifier won't be
10483 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
10484 * again on the way to bpf_exit.
10485 * When looping the sl->state.branches will be > 0 and this state
10486 * will not be considered for equivalence until branches == 0.
10488 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
10491 env->total_states++;
10492 env->peak_states++;
10493 env->prev_jmps_processed = env->jmps_processed;
10494 env->prev_insn_processed = env->insn_processed;
10496 /* add new state to the head of linked list */
10497 new = &new_sl->state;
10498 err = copy_verifier_state(new, cur);
10500 free_verifier_state(new, false);
10504 new->insn_idx = insn_idx;
10505 WARN_ONCE(new->branches != 1,
10506 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
10509 cur->first_insn_idx = insn_idx;
10510 clear_jmp_history(cur);
10511 new_sl->next = *explored_state(env, insn_idx);
10512 *explored_state(env, insn_idx) = new_sl;
10513 /* connect new state to parentage chain. Current frame needs all
10514 * registers connected. Only r6 - r9 of the callers are alive (pushed
10515 * to the stack implicitly by JITs) so in callers' frames connect just
10516 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
10517 * the state of the call instruction (with WRITTEN set), and r0 comes
10518 * from callee with its full parentage chain, anyway.
10520 /* clear write marks in current state: the writes we did are not writes
10521 * our child did, so they don't screen off its reads from us.
10522 * (There are no read marks in current state, because reads always mark
10523 * their parent and current state never has children yet. Only
10524 * explored_states can get read marks.)
10526 for (j = 0; j <= cur->curframe; j++) {
10527 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
10528 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
10529 for (i = 0; i < BPF_REG_FP; i++)
10530 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
10533 /* all stack frames are accessible from callee, clear them all */
10534 for (j = 0; j <= cur->curframe; j++) {
10535 struct bpf_func_state *frame = cur->frame[j];
10536 struct bpf_func_state *newframe = new->frame[j];
10538 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
10539 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
10540 frame->stack[i].spilled_ptr.parent =
10541 &newframe->stack[i].spilled_ptr;
10547 /* Return true if it's OK to have the same insn return a different type. */
10548 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
10552 case PTR_TO_SOCKET:
10553 case PTR_TO_SOCKET_OR_NULL:
10554 case PTR_TO_SOCK_COMMON:
10555 case PTR_TO_SOCK_COMMON_OR_NULL:
10556 case PTR_TO_TCP_SOCK:
10557 case PTR_TO_TCP_SOCK_OR_NULL:
10558 case PTR_TO_XDP_SOCK:
10559 case PTR_TO_BTF_ID:
10560 case PTR_TO_BTF_ID_OR_NULL:
10567 /* If an instruction was previously used with particular pointer types, then we
10568 * need to be careful to avoid cases such as the below, where it may be ok
10569 * for one branch accessing the pointer, but not ok for the other branch:
10574 * R1 = some_other_valid_ptr;
10577 * R2 = *(u32 *)(R1 + 0);
10579 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
10581 return src != prev && (!reg_type_mismatch_ok(src) ||
10582 !reg_type_mismatch_ok(prev));
10585 static int do_check(struct bpf_verifier_env *env)
10587 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
10588 struct bpf_verifier_state *state = env->cur_state;
10589 struct bpf_insn *insns = env->prog->insnsi;
10590 struct bpf_reg_state *regs;
10591 int insn_cnt = env->prog->len;
10592 bool do_print_state = false;
10593 int prev_insn_idx = -1;
10596 struct bpf_insn *insn;
10600 env->prev_insn_idx = prev_insn_idx;
10601 if (env->insn_idx >= insn_cnt) {
10602 verbose(env, "invalid insn idx %d insn_cnt %d\n",
10603 env->insn_idx, insn_cnt);
10607 insn = &insns[env->insn_idx];
10608 class = BPF_CLASS(insn->code);
10610 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
10612 "BPF program is too large. Processed %d insn\n",
10613 env->insn_processed);
10617 err = is_state_visited(env, env->insn_idx);
10621 /* found equivalent state, can prune the search */
10622 if (env->log.level & BPF_LOG_LEVEL) {
10623 if (do_print_state)
10624 verbose(env, "\nfrom %d to %d%s: safe\n",
10625 env->prev_insn_idx, env->insn_idx,
10626 env->cur_state->speculative ?
10627 " (speculative execution)" : "");
10629 verbose(env, "%d: safe\n", env->insn_idx);
10631 goto process_bpf_exit;
10634 if (signal_pending(current))
10637 if (need_resched())
10640 if (env->log.level & BPF_LOG_LEVEL2 ||
10641 (env->log.level & BPF_LOG_LEVEL && do_print_state)) {
10642 if (env->log.level & BPF_LOG_LEVEL2)
10643 verbose(env, "%d:", env->insn_idx);
10645 verbose(env, "\nfrom %d to %d%s:",
10646 env->prev_insn_idx, env->insn_idx,
10647 env->cur_state->speculative ?
10648 " (speculative execution)" : "");
10649 print_verifier_state(env, state->frame[state->curframe]);
10650 do_print_state = false;
10653 if (env->log.level & BPF_LOG_LEVEL) {
10654 const struct bpf_insn_cbs cbs = {
10655 .cb_call = disasm_kfunc_name,
10656 .cb_print = verbose,
10657 .private_data = env,
10660 verbose_linfo(env, env->insn_idx, "; ");
10661 verbose(env, "%d: ", env->insn_idx);
10662 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
10665 if (bpf_prog_is_dev_bound(env->prog->aux)) {
10666 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
10667 env->prev_insn_idx);
10672 regs = cur_regs(env);
10673 sanitize_mark_insn_seen(env);
10674 prev_insn_idx = env->insn_idx;
10676 if (class == BPF_ALU || class == BPF_ALU64) {
10677 err = check_alu_op(env, insn);
10681 } else if (class == BPF_LDX) {
10682 enum bpf_reg_type *prev_src_type, src_reg_type;
10684 /* check for reserved fields is already done */
10686 /* check src operand */
10687 err = check_reg_arg(env, insn->src_reg, SRC_OP);
10691 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
10695 src_reg_type = regs[insn->src_reg].type;
10697 /* check that memory (src_reg + off) is readable,
10698 * the state of dst_reg will be updated by this func
10700 err = check_mem_access(env, env->insn_idx, insn->src_reg,
10701 insn->off, BPF_SIZE(insn->code),
10702 BPF_READ, insn->dst_reg, false);
10706 prev_src_type = &env->insn_aux_data[env->insn_idx].ptr_type;
10708 if (*prev_src_type == NOT_INIT) {
10709 /* saw a valid insn
10710 * dst_reg = *(u32 *)(src_reg + off)
10711 * save type to validate intersecting paths
10713 *prev_src_type = src_reg_type;
10715 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
10716 /* ABuser program is trying to use the same insn
10717 * dst_reg = *(u32*) (src_reg + off)
10718 * with different pointer types:
10719 * src_reg == ctx in one branch and
10720 * src_reg == stack|map in some other branch.
10723 verbose(env, "same insn cannot be used with different pointers\n");
10727 } else if (class == BPF_STX) {
10728 enum bpf_reg_type *prev_dst_type, dst_reg_type;
10730 if (BPF_MODE(insn->code) == BPF_ATOMIC) {
10731 err = check_atomic(env, env->insn_idx, insn);
10738 if (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0) {
10739 verbose(env, "BPF_STX uses reserved fields\n");
10743 /* check src1 operand */
10744 err = check_reg_arg(env, insn->src_reg, SRC_OP);
10747 /* check src2 operand */
10748 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
10752 dst_reg_type = regs[insn->dst_reg].type;
10754 /* check that memory (dst_reg + off) is writeable */
10755 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
10756 insn->off, BPF_SIZE(insn->code),
10757 BPF_WRITE, insn->src_reg, false);
10761 prev_dst_type = &env->insn_aux_data[env->insn_idx].ptr_type;
10763 if (*prev_dst_type == NOT_INIT) {
10764 *prev_dst_type = dst_reg_type;
10765 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
10766 verbose(env, "same insn cannot be used with different pointers\n");
10770 } else if (class == BPF_ST) {
10771 if (BPF_MODE(insn->code) != BPF_MEM ||
10772 insn->src_reg != BPF_REG_0) {
10773 verbose(env, "BPF_ST uses reserved fields\n");
10776 /* check src operand */
10777 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
10781 if (is_ctx_reg(env, insn->dst_reg)) {
10782 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
10784 reg_type_str[reg_state(env, insn->dst_reg)->type]);
10788 /* check that memory (dst_reg + off) is writeable */
10789 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
10790 insn->off, BPF_SIZE(insn->code),
10791 BPF_WRITE, -1, false);
10795 } else if (class == BPF_JMP || class == BPF_JMP32) {
10796 u8 opcode = BPF_OP(insn->code);
10798 env->jmps_processed++;
10799 if (opcode == BPF_CALL) {
10800 if (BPF_SRC(insn->code) != BPF_K ||
10802 (insn->src_reg != BPF_REG_0 &&
10803 insn->src_reg != BPF_PSEUDO_CALL &&
10804 insn->src_reg != BPF_PSEUDO_KFUNC_CALL) ||
10805 insn->dst_reg != BPF_REG_0 ||
10806 class == BPF_JMP32) {
10807 verbose(env, "BPF_CALL uses reserved fields\n");
10811 if (env->cur_state->active_spin_lock &&
10812 (insn->src_reg == BPF_PSEUDO_CALL ||
10813 insn->imm != BPF_FUNC_spin_unlock)) {
10814 verbose(env, "function calls are not allowed while holding a lock\n");
10817 if (insn->src_reg == BPF_PSEUDO_CALL)
10818 err = check_func_call(env, insn, &env->insn_idx);
10819 else if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL)
10820 err = check_kfunc_call(env, insn);
10822 err = check_helper_call(env, insn, &env->insn_idx);
10825 } else if (opcode == BPF_JA) {
10826 if (BPF_SRC(insn->code) != BPF_K ||
10828 insn->src_reg != BPF_REG_0 ||
10829 insn->dst_reg != BPF_REG_0 ||
10830 class == BPF_JMP32) {
10831 verbose(env, "BPF_JA uses reserved fields\n");
10835 env->insn_idx += insn->off + 1;
10838 } else if (opcode == BPF_EXIT) {
10839 if (BPF_SRC(insn->code) != BPF_K ||
10841 insn->src_reg != BPF_REG_0 ||
10842 insn->dst_reg != BPF_REG_0 ||
10843 class == BPF_JMP32) {
10844 verbose(env, "BPF_EXIT uses reserved fields\n");
10848 if (env->cur_state->active_spin_lock) {
10849 verbose(env, "bpf_spin_unlock is missing\n");
10853 if (state->curframe) {
10854 /* exit from nested function */
10855 err = prepare_func_exit(env, &env->insn_idx);
10858 do_print_state = true;
10862 err = check_reference_leak(env);
10866 err = check_return_code(env);
10870 update_branch_counts(env, env->cur_state);
10871 err = pop_stack(env, &prev_insn_idx,
10872 &env->insn_idx, pop_log);
10874 if (err != -ENOENT)
10878 do_print_state = true;
10882 err = check_cond_jmp_op(env, insn, &env->insn_idx);
10886 } else if (class == BPF_LD) {
10887 u8 mode = BPF_MODE(insn->code);
10889 if (mode == BPF_ABS || mode == BPF_IND) {
10890 err = check_ld_abs(env, insn);
10894 } else if (mode == BPF_IMM) {
10895 err = check_ld_imm(env, insn);
10900 sanitize_mark_insn_seen(env);
10902 verbose(env, "invalid BPF_LD mode\n");
10906 verbose(env, "unknown insn class %d\n", class);
10916 static int find_btf_percpu_datasec(struct btf *btf)
10918 const struct btf_type *t;
10923 * Both vmlinux and module each have their own ".data..percpu"
10924 * DATASECs in BTF. So for module's case, we need to skip vmlinux BTF
10925 * types to look at only module's own BTF types.
10927 n = btf_nr_types(btf);
10928 if (btf_is_module(btf))
10929 i = btf_nr_types(btf_vmlinux);
10933 for(; i < n; i++) {
10934 t = btf_type_by_id(btf, i);
10935 if (BTF_INFO_KIND(t->info) != BTF_KIND_DATASEC)
10938 tname = btf_name_by_offset(btf, t->name_off);
10939 if (!strcmp(tname, ".data..percpu"))
10946 /* replace pseudo btf_id with kernel symbol address */
10947 static int check_pseudo_btf_id(struct bpf_verifier_env *env,
10948 struct bpf_insn *insn,
10949 struct bpf_insn_aux_data *aux)
10951 const struct btf_var_secinfo *vsi;
10952 const struct btf_type *datasec;
10953 struct btf_mod_pair *btf_mod;
10954 const struct btf_type *t;
10955 const char *sym_name;
10956 bool percpu = false;
10957 u32 type, id = insn->imm;
10961 int i, btf_fd, err;
10963 btf_fd = insn[1].imm;
10965 btf = btf_get_by_fd(btf_fd);
10967 verbose(env, "invalid module BTF object FD specified.\n");
10971 if (!btf_vmlinux) {
10972 verbose(env, "kernel is missing BTF, make sure CONFIG_DEBUG_INFO_BTF=y is specified in Kconfig.\n");
10979 t = btf_type_by_id(btf, id);
10981 verbose(env, "ldimm64 insn specifies invalid btf_id %d.\n", id);
10986 if (!btf_type_is_var(t)) {
10987 verbose(env, "pseudo btf_id %d in ldimm64 isn't KIND_VAR.\n", id);
10992 sym_name = btf_name_by_offset(btf, t->name_off);
10993 addr = kallsyms_lookup_name(sym_name);
10995 verbose(env, "ldimm64 failed to find the address for kernel symbol '%s'.\n",
11001 datasec_id = find_btf_percpu_datasec(btf);
11002 if (datasec_id > 0) {
11003 datasec = btf_type_by_id(btf, datasec_id);
11004 for_each_vsi(i, datasec, vsi) {
11005 if (vsi->type == id) {
11012 insn[0].imm = (u32)addr;
11013 insn[1].imm = addr >> 32;
11016 t = btf_type_skip_modifiers(btf, type, NULL);
11018 aux->btf_var.reg_type = PTR_TO_PERCPU_BTF_ID;
11019 aux->btf_var.btf = btf;
11020 aux->btf_var.btf_id = type;
11021 } else if (!btf_type_is_struct(t)) {
11022 const struct btf_type *ret;
11026 /* resolve the type size of ksym. */
11027 ret = btf_resolve_size(btf, t, &tsize);
11029 tname = btf_name_by_offset(btf, t->name_off);
11030 verbose(env, "ldimm64 unable to resolve the size of type '%s': %ld\n",
11031 tname, PTR_ERR(ret));
11035 aux->btf_var.reg_type = PTR_TO_MEM;
11036 aux->btf_var.mem_size = tsize;
11038 aux->btf_var.reg_type = PTR_TO_BTF_ID;
11039 aux->btf_var.btf = btf;
11040 aux->btf_var.btf_id = type;
11043 /* check whether we recorded this BTF (and maybe module) already */
11044 for (i = 0; i < env->used_btf_cnt; i++) {
11045 if (env->used_btfs[i].btf == btf) {
11051 if (env->used_btf_cnt >= MAX_USED_BTFS) {
11056 btf_mod = &env->used_btfs[env->used_btf_cnt];
11057 btf_mod->btf = btf;
11058 btf_mod->module = NULL;
11060 /* if we reference variables from kernel module, bump its refcount */
11061 if (btf_is_module(btf)) {
11062 btf_mod->module = btf_try_get_module(btf);
11063 if (!btf_mod->module) {
11069 env->used_btf_cnt++;
11077 static int check_map_prealloc(struct bpf_map *map)
11079 return (map->map_type != BPF_MAP_TYPE_HASH &&
11080 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
11081 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
11082 !(map->map_flags & BPF_F_NO_PREALLOC);
11085 static bool is_tracing_prog_type(enum bpf_prog_type type)
11088 case BPF_PROG_TYPE_KPROBE:
11089 case BPF_PROG_TYPE_TRACEPOINT:
11090 case BPF_PROG_TYPE_PERF_EVENT:
11091 case BPF_PROG_TYPE_RAW_TRACEPOINT:
11098 static bool is_preallocated_map(struct bpf_map *map)
11100 if (!check_map_prealloc(map))
11102 if (map->inner_map_meta && !check_map_prealloc(map->inner_map_meta))
11107 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
11108 struct bpf_map *map,
11109 struct bpf_prog *prog)
11112 enum bpf_prog_type prog_type = resolve_prog_type(prog);
11114 * Validate that trace type programs use preallocated hash maps.
11116 * For programs attached to PERF events this is mandatory as the
11117 * perf NMI can hit any arbitrary code sequence.
11119 * All other trace types using preallocated hash maps are unsafe as
11120 * well because tracepoint or kprobes can be inside locked regions
11121 * of the memory allocator or at a place where a recursion into the
11122 * memory allocator would see inconsistent state.
11124 * On RT enabled kernels run-time allocation of all trace type
11125 * programs is strictly prohibited due to lock type constraints. On
11126 * !RT kernels it is allowed for backwards compatibility reasons for
11127 * now, but warnings are emitted so developers are made aware of
11128 * the unsafety and can fix their programs before this is enforced.
11130 if (is_tracing_prog_type(prog_type) && !is_preallocated_map(map)) {
11131 if (prog_type == BPF_PROG_TYPE_PERF_EVENT) {
11132 verbose(env, "perf_event programs can only use preallocated hash map\n");
11135 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
11136 verbose(env, "trace type programs can only use preallocated hash map\n");
11139 WARN_ONCE(1, "trace type BPF program uses run-time allocation\n");
11140 verbose(env, "trace type programs with run-time allocated hash maps are unsafe. Switch to preallocated hash maps.\n");
11143 if (map_value_has_spin_lock(map)) {
11144 if (prog_type == BPF_PROG_TYPE_SOCKET_FILTER) {
11145 verbose(env, "socket filter progs cannot use bpf_spin_lock yet\n");
11149 if (is_tracing_prog_type(prog_type)) {
11150 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
11154 if (prog->aux->sleepable) {
11155 verbose(env, "sleepable progs cannot use bpf_spin_lock yet\n");
11160 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
11161 !bpf_offload_prog_map_match(prog, map)) {
11162 verbose(env, "offload device mismatch between prog and map\n");
11166 if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
11167 verbose(env, "bpf_struct_ops map cannot be used in prog\n");
11171 if (prog->aux->sleepable)
11172 switch (map->map_type) {
11173 case BPF_MAP_TYPE_HASH:
11174 case BPF_MAP_TYPE_LRU_HASH:
11175 case BPF_MAP_TYPE_ARRAY:
11176 case BPF_MAP_TYPE_PERCPU_HASH:
11177 case BPF_MAP_TYPE_PERCPU_ARRAY:
11178 case BPF_MAP_TYPE_LRU_PERCPU_HASH:
11179 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
11180 case BPF_MAP_TYPE_HASH_OF_MAPS:
11181 if (!is_preallocated_map(map)) {
11183 "Sleepable programs can only use preallocated maps\n");
11187 case BPF_MAP_TYPE_RINGBUF:
11191 "Sleepable programs can only use array, hash, and ringbuf maps\n");
11198 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
11200 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
11201 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
11204 /* find and rewrite pseudo imm in ld_imm64 instructions:
11206 * 1. if it accesses map FD, replace it with actual map pointer.
11207 * 2. if it accesses btf_id of a VAR, replace it with pointer to the var.
11209 * NOTE: btf_vmlinux is required for converting pseudo btf_id.
11211 static int resolve_pseudo_ldimm64(struct bpf_verifier_env *env)
11213 struct bpf_insn *insn = env->prog->insnsi;
11214 int insn_cnt = env->prog->len;
11217 err = bpf_prog_calc_tag(env->prog);
11221 for (i = 0; i < insn_cnt; i++, insn++) {
11222 if (BPF_CLASS(insn->code) == BPF_LDX &&
11223 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
11224 verbose(env, "BPF_LDX uses reserved fields\n");
11228 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
11229 struct bpf_insn_aux_data *aux;
11230 struct bpf_map *map;
11235 if (i == insn_cnt - 1 || insn[1].code != 0 ||
11236 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
11237 insn[1].off != 0) {
11238 verbose(env, "invalid bpf_ld_imm64 insn\n");
11242 if (insn[0].src_reg == 0)
11243 /* valid generic load 64-bit imm */
11246 if (insn[0].src_reg == BPF_PSEUDO_BTF_ID) {
11247 aux = &env->insn_aux_data[i];
11248 err = check_pseudo_btf_id(env, insn, aux);
11254 if (insn[0].src_reg == BPF_PSEUDO_FUNC) {
11255 aux = &env->insn_aux_data[i];
11256 aux->ptr_type = PTR_TO_FUNC;
11260 /* In final convert_pseudo_ld_imm64() step, this is
11261 * converted into regular 64-bit imm load insn.
11263 switch (insn[0].src_reg) {
11264 case BPF_PSEUDO_MAP_VALUE:
11265 case BPF_PSEUDO_MAP_IDX_VALUE:
11267 case BPF_PSEUDO_MAP_FD:
11268 case BPF_PSEUDO_MAP_IDX:
11269 if (insn[1].imm == 0)
11273 verbose(env, "unrecognized bpf_ld_imm64 insn\n");
11277 switch (insn[0].src_reg) {
11278 case BPF_PSEUDO_MAP_IDX_VALUE:
11279 case BPF_PSEUDO_MAP_IDX:
11280 if (bpfptr_is_null(env->fd_array)) {
11281 verbose(env, "fd_idx without fd_array is invalid\n");
11284 if (copy_from_bpfptr_offset(&fd, env->fd_array,
11285 insn[0].imm * sizeof(fd),
11295 map = __bpf_map_get(f);
11297 verbose(env, "fd %d is not pointing to valid bpf_map\n",
11299 return PTR_ERR(map);
11302 err = check_map_prog_compatibility(env, map, env->prog);
11308 aux = &env->insn_aux_data[i];
11309 if (insn[0].src_reg == BPF_PSEUDO_MAP_FD ||
11310 insn[0].src_reg == BPF_PSEUDO_MAP_IDX) {
11311 addr = (unsigned long)map;
11313 u32 off = insn[1].imm;
11315 if (off >= BPF_MAX_VAR_OFF) {
11316 verbose(env, "direct value offset of %u is not allowed\n", off);
11321 if (!map->ops->map_direct_value_addr) {
11322 verbose(env, "no direct value access support for this map type\n");
11327 err = map->ops->map_direct_value_addr(map, &addr, off);
11329 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
11330 map->value_size, off);
11335 aux->map_off = off;
11339 insn[0].imm = (u32)addr;
11340 insn[1].imm = addr >> 32;
11342 /* check whether we recorded this map already */
11343 for (j = 0; j < env->used_map_cnt; j++) {
11344 if (env->used_maps[j] == map) {
11345 aux->map_index = j;
11351 if (env->used_map_cnt >= MAX_USED_MAPS) {
11356 /* hold the map. If the program is rejected by verifier,
11357 * the map will be released by release_maps() or it
11358 * will be used by the valid program until it's unloaded
11359 * and all maps are released in free_used_maps()
11363 aux->map_index = env->used_map_cnt;
11364 env->used_maps[env->used_map_cnt++] = map;
11366 if (bpf_map_is_cgroup_storage(map) &&
11367 bpf_cgroup_storage_assign(env->prog->aux, map)) {
11368 verbose(env, "only one cgroup storage of each type is allowed\n");
11380 /* Basic sanity check before we invest more work here. */
11381 if (!bpf_opcode_in_insntable(insn->code)) {
11382 verbose(env, "unknown opcode %02x\n", insn->code);
11387 /* now all pseudo BPF_LD_IMM64 instructions load valid
11388 * 'struct bpf_map *' into a register instead of user map_fd.
11389 * These pointers will be used later by verifier to validate map access.
11394 /* drop refcnt of maps used by the rejected program */
11395 static void release_maps(struct bpf_verifier_env *env)
11397 __bpf_free_used_maps(env->prog->aux, env->used_maps,
11398 env->used_map_cnt);
11401 /* drop refcnt of maps used by the rejected program */
11402 static void release_btfs(struct bpf_verifier_env *env)
11404 __bpf_free_used_btfs(env->prog->aux, env->used_btfs,
11405 env->used_btf_cnt);
11408 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
11409 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
11411 struct bpf_insn *insn = env->prog->insnsi;
11412 int insn_cnt = env->prog->len;
11415 for (i = 0; i < insn_cnt; i++, insn++) {
11416 if (insn->code != (BPF_LD | BPF_IMM | BPF_DW))
11418 if (insn->src_reg == BPF_PSEUDO_FUNC)
11424 /* single env->prog->insni[off] instruction was replaced with the range
11425 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
11426 * [0, off) and [off, end) to new locations, so the patched range stays zero
11428 static int adjust_insn_aux_data(struct bpf_verifier_env *env,
11429 struct bpf_prog *new_prog, u32 off, u32 cnt)
11431 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
11432 struct bpf_insn *insn = new_prog->insnsi;
11433 u32 old_seen = old_data[off].seen;
11437 /* aux info at OFF always needs adjustment, no matter fast path
11438 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
11439 * original insn at old prog.
11441 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
11445 prog_len = new_prog->len;
11446 new_data = vzalloc(array_size(prog_len,
11447 sizeof(struct bpf_insn_aux_data)));
11450 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
11451 memcpy(new_data + off + cnt - 1, old_data + off,
11452 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
11453 for (i = off; i < off + cnt - 1; i++) {
11454 /* Expand insni[off]'s seen count to the patched range. */
11455 new_data[i].seen = old_seen;
11456 new_data[i].zext_dst = insn_has_def32(env, insn + i);
11458 env->insn_aux_data = new_data;
11463 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
11469 /* NOTE: fake 'exit' subprog should be updated as well. */
11470 for (i = 0; i <= env->subprog_cnt; i++) {
11471 if (env->subprog_info[i].start <= off)
11473 env->subprog_info[i].start += len - 1;
11477 static void adjust_poke_descs(struct bpf_prog *prog, u32 off, u32 len)
11479 struct bpf_jit_poke_descriptor *tab = prog->aux->poke_tab;
11480 int i, sz = prog->aux->size_poke_tab;
11481 struct bpf_jit_poke_descriptor *desc;
11483 for (i = 0; i < sz; i++) {
11485 if (desc->insn_idx <= off)
11487 desc->insn_idx += len - 1;
11491 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
11492 const struct bpf_insn *patch, u32 len)
11494 struct bpf_prog *new_prog;
11496 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
11497 if (IS_ERR(new_prog)) {
11498 if (PTR_ERR(new_prog) == -ERANGE)
11500 "insn %d cannot be patched due to 16-bit range\n",
11501 env->insn_aux_data[off].orig_idx);
11504 if (adjust_insn_aux_data(env, new_prog, off, len))
11506 adjust_subprog_starts(env, off, len);
11507 adjust_poke_descs(new_prog, off, len);
11511 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
11516 /* find first prog starting at or after off (first to remove) */
11517 for (i = 0; i < env->subprog_cnt; i++)
11518 if (env->subprog_info[i].start >= off)
11520 /* find first prog starting at or after off + cnt (first to stay) */
11521 for (j = i; j < env->subprog_cnt; j++)
11522 if (env->subprog_info[j].start >= off + cnt)
11524 /* if j doesn't start exactly at off + cnt, we are just removing
11525 * the front of previous prog
11527 if (env->subprog_info[j].start != off + cnt)
11531 struct bpf_prog_aux *aux = env->prog->aux;
11534 /* move fake 'exit' subprog as well */
11535 move = env->subprog_cnt + 1 - j;
11537 memmove(env->subprog_info + i,
11538 env->subprog_info + j,
11539 sizeof(*env->subprog_info) * move);
11540 env->subprog_cnt -= j - i;
11542 /* remove func_info */
11543 if (aux->func_info) {
11544 move = aux->func_info_cnt - j;
11546 memmove(aux->func_info + i,
11547 aux->func_info + j,
11548 sizeof(*aux->func_info) * move);
11549 aux->func_info_cnt -= j - i;
11550 /* func_info->insn_off is set after all code rewrites,
11551 * in adjust_btf_func() - no need to adjust
11555 /* convert i from "first prog to remove" to "first to adjust" */
11556 if (env->subprog_info[i].start == off)
11560 /* update fake 'exit' subprog as well */
11561 for (; i <= env->subprog_cnt; i++)
11562 env->subprog_info[i].start -= cnt;
11567 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
11570 struct bpf_prog *prog = env->prog;
11571 u32 i, l_off, l_cnt, nr_linfo;
11572 struct bpf_line_info *linfo;
11574 nr_linfo = prog->aux->nr_linfo;
11578 linfo = prog->aux->linfo;
11580 /* find first line info to remove, count lines to be removed */
11581 for (i = 0; i < nr_linfo; i++)
11582 if (linfo[i].insn_off >= off)
11587 for (; i < nr_linfo; i++)
11588 if (linfo[i].insn_off < off + cnt)
11593 /* First live insn doesn't match first live linfo, it needs to "inherit"
11594 * last removed linfo. prog is already modified, so prog->len == off
11595 * means no live instructions after (tail of the program was removed).
11597 if (prog->len != off && l_cnt &&
11598 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
11600 linfo[--i].insn_off = off + cnt;
11603 /* remove the line info which refer to the removed instructions */
11605 memmove(linfo + l_off, linfo + i,
11606 sizeof(*linfo) * (nr_linfo - i));
11608 prog->aux->nr_linfo -= l_cnt;
11609 nr_linfo = prog->aux->nr_linfo;
11612 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
11613 for (i = l_off; i < nr_linfo; i++)
11614 linfo[i].insn_off -= cnt;
11616 /* fix up all subprogs (incl. 'exit') which start >= off */
11617 for (i = 0; i <= env->subprog_cnt; i++)
11618 if (env->subprog_info[i].linfo_idx > l_off) {
11619 /* program may have started in the removed region but
11620 * may not be fully removed
11622 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
11623 env->subprog_info[i].linfo_idx -= l_cnt;
11625 env->subprog_info[i].linfo_idx = l_off;
11631 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
11633 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
11634 unsigned int orig_prog_len = env->prog->len;
11637 if (bpf_prog_is_dev_bound(env->prog->aux))
11638 bpf_prog_offload_remove_insns(env, off, cnt);
11640 err = bpf_remove_insns(env->prog, off, cnt);
11644 err = adjust_subprog_starts_after_remove(env, off, cnt);
11648 err = bpf_adj_linfo_after_remove(env, off, cnt);
11652 memmove(aux_data + off, aux_data + off + cnt,
11653 sizeof(*aux_data) * (orig_prog_len - off - cnt));
11658 /* The verifier does more data flow analysis than llvm and will not
11659 * explore branches that are dead at run time. Malicious programs can
11660 * have dead code too. Therefore replace all dead at-run-time code
11663 * Just nops are not optimal, e.g. if they would sit at the end of the
11664 * program and through another bug we would manage to jump there, then
11665 * we'd execute beyond program memory otherwise. Returning exception
11666 * code also wouldn't work since we can have subprogs where the dead
11667 * code could be located.
11669 static void sanitize_dead_code(struct bpf_verifier_env *env)
11671 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
11672 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
11673 struct bpf_insn *insn = env->prog->insnsi;
11674 const int insn_cnt = env->prog->len;
11677 for (i = 0; i < insn_cnt; i++) {
11678 if (aux_data[i].seen)
11680 memcpy(insn + i, &trap, sizeof(trap));
11684 static bool insn_is_cond_jump(u8 code)
11688 if (BPF_CLASS(code) == BPF_JMP32)
11691 if (BPF_CLASS(code) != BPF_JMP)
11695 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
11698 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
11700 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
11701 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
11702 struct bpf_insn *insn = env->prog->insnsi;
11703 const int insn_cnt = env->prog->len;
11706 for (i = 0; i < insn_cnt; i++, insn++) {
11707 if (!insn_is_cond_jump(insn->code))
11710 if (!aux_data[i + 1].seen)
11711 ja.off = insn->off;
11712 else if (!aux_data[i + 1 + insn->off].seen)
11717 if (bpf_prog_is_dev_bound(env->prog->aux))
11718 bpf_prog_offload_replace_insn(env, i, &ja);
11720 memcpy(insn, &ja, sizeof(ja));
11724 static int opt_remove_dead_code(struct bpf_verifier_env *env)
11726 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
11727 int insn_cnt = env->prog->len;
11730 for (i = 0; i < insn_cnt; i++) {
11734 while (i + j < insn_cnt && !aux_data[i + j].seen)
11739 err = verifier_remove_insns(env, i, j);
11742 insn_cnt = env->prog->len;
11748 static int opt_remove_nops(struct bpf_verifier_env *env)
11750 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
11751 struct bpf_insn *insn = env->prog->insnsi;
11752 int insn_cnt = env->prog->len;
11755 for (i = 0; i < insn_cnt; i++) {
11756 if (memcmp(&insn[i], &ja, sizeof(ja)))
11759 err = verifier_remove_insns(env, i, 1);
11769 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
11770 const union bpf_attr *attr)
11772 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
11773 struct bpf_insn_aux_data *aux = env->insn_aux_data;
11774 int i, patch_len, delta = 0, len = env->prog->len;
11775 struct bpf_insn *insns = env->prog->insnsi;
11776 struct bpf_prog *new_prog;
11779 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
11780 zext_patch[1] = BPF_ZEXT_REG(0);
11781 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
11782 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
11783 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
11784 for (i = 0; i < len; i++) {
11785 int adj_idx = i + delta;
11786 struct bpf_insn insn;
11789 insn = insns[adj_idx];
11790 load_reg = insn_def_regno(&insn);
11791 if (!aux[adj_idx].zext_dst) {
11799 class = BPF_CLASS(code);
11800 if (load_reg == -1)
11803 /* NOTE: arg "reg" (the fourth one) is only used for
11804 * BPF_STX + SRC_OP, so it is safe to pass NULL
11807 if (is_reg64(env, &insn, load_reg, NULL, DST_OP)) {
11808 if (class == BPF_LD &&
11809 BPF_MODE(code) == BPF_IMM)
11814 /* ctx load could be transformed into wider load. */
11815 if (class == BPF_LDX &&
11816 aux[adj_idx].ptr_type == PTR_TO_CTX)
11819 imm_rnd = get_random_int();
11820 rnd_hi32_patch[0] = insn;
11821 rnd_hi32_patch[1].imm = imm_rnd;
11822 rnd_hi32_patch[3].dst_reg = load_reg;
11823 patch = rnd_hi32_patch;
11825 goto apply_patch_buffer;
11828 /* Add in an zero-extend instruction if a) the JIT has requested
11829 * it or b) it's a CMPXCHG.
11831 * The latter is because: BPF_CMPXCHG always loads a value into
11832 * R0, therefore always zero-extends. However some archs'
11833 * equivalent instruction only does this load when the
11834 * comparison is successful. This detail of CMPXCHG is
11835 * orthogonal to the general zero-extension behaviour of the
11836 * CPU, so it's treated independently of bpf_jit_needs_zext.
11838 if (!bpf_jit_needs_zext() && !is_cmpxchg_insn(&insn))
11841 if (WARN_ON(load_reg == -1)) {
11842 verbose(env, "verifier bug. zext_dst is set, but no reg is defined\n");
11846 zext_patch[0] = insn;
11847 zext_patch[1].dst_reg = load_reg;
11848 zext_patch[1].src_reg = load_reg;
11849 patch = zext_patch;
11851 apply_patch_buffer:
11852 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
11855 env->prog = new_prog;
11856 insns = new_prog->insnsi;
11857 aux = env->insn_aux_data;
11858 delta += patch_len - 1;
11864 /* convert load instructions that access fields of a context type into a
11865 * sequence of instructions that access fields of the underlying structure:
11866 * struct __sk_buff -> struct sk_buff
11867 * struct bpf_sock_ops -> struct sock
11869 static int convert_ctx_accesses(struct bpf_verifier_env *env)
11871 const struct bpf_verifier_ops *ops = env->ops;
11872 int i, cnt, size, ctx_field_size, delta = 0;
11873 const int insn_cnt = env->prog->len;
11874 struct bpf_insn insn_buf[16], *insn;
11875 u32 target_size, size_default, off;
11876 struct bpf_prog *new_prog;
11877 enum bpf_access_type type;
11878 bool is_narrower_load;
11880 if (ops->gen_prologue || env->seen_direct_write) {
11881 if (!ops->gen_prologue) {
11882 verbose(env, "bpf verifier is misconfigured\n");
11885 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
11887 if (cnt >= ARRAY_SIZE(insn_buf)) {
11888 verbose(env, "bpf verifier is misconfigured\n");
11891 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
11895 env->prog = new_prog;
11900 if (bpf_prog_is_dev_bound(env->prog->aux))
11903 insn = env->prog->insnsi + delta;
11905 for (i = 0; i < insn_cnt; i++, insn++) {
11906 bpf_convert_ctx_access_t convert_ctx_access;
11908 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
11909 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
11910 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
11911 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
11913 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
11914 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
11915 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
11916 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
11921 if (type == BPF_WRITE &&
11922 env->insn_aux_data[i + delta].sanitize_stack_off) {
11923 struct bpf_insn patch[] = {
11924 /* Sanitize suspicious stack slot with zero.
11925 * There are no memory dependencies for this store,
11926 * since it's only using frame pointer and immediate
11929 BPF_ST_MEM(BPF_DW, BPF_REG_FP,
11930 env->insn_aux_data[i + delta].sanitize_stack_off,
11932 /* the original STX instruction will immediately
11933 * overwrite the same stack slot with appropriate value
11938 cnt = ARRAY_SIZE(patch);
11939 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
11944 env->prog = new_prog;
11945 insn = new_prog->insnsi + i + delta;
11949 switch (env->insn_aux_data[i + delta].ptr_type) {
11951 if (!ops->convert_ctx_access)
11953 convert_ctx_access = ops->convert_ctx_access;
11955 case PTR_TO_SOCKET:
11956 case PTR_TO_SOCK_COMMON:
11957 convert_ctx_access = bpf_sock_convert_ctx_access;
11959 case PTR_TO_TCP_SOCK:
11960 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
11962 case PTR_TO_XDP_SOCK:
11963 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
11965 case PTR_TO_BTF_ID:
11966 if (type == BPF_READ) {
11967 insn->code = BPF_LDX | BPF_PROBE_MEM |
11968 BPF_SIZE((insn)->code);
11969 env->prog->aux->num_exentries++;
11970 } else if (resolve_prog_type(env->prog) != BPF_PROG_TYPE_STRUCT_OPS) {
11971 verbose(env, "Writes through BTF pointers are not allowed\n");
11979 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
11980 size = BPF_LDST_BYTES(insn);
11982 /* If the read access is a narrower load of the field,
11983 * convert to a 4/8-byte load, to minimum program type specific
11984 * convert_ctx_access changes. If conversion is successful,
11985 * we will apply proper mask to the result.
11987 is_narrower_load = size < ctx_field_size;
11988 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
11990 if (is_narrower_load) {
11993 if (type == BPF_WRITE) {
11994 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
11999 if (ctx_field_size == 4)
12001 else if (ctx_field_size == 8)
12002 size_code = BPF_DW;
12004 insn->off = off & ~(size_default - 1);
12005 insn->code = BPF_LDX | BPF_MEM | size_code;
12009 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
12011 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
12012 (ctx_field_size && !target_size)) {
12013 verbose(env, "bpf verifier is misconfigured\n");
12017 if (is_narrower_load && size < target_size) {
12018 u8 shift = bpf_ctx_narrow_access_offset(
12019 off, size, size_default) * 8;
12020 if (ctx_field_size <= 4) {
12022 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
12025 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
12026 (1 << size * 8) - 1);
12029 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
12032 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
12033 (1ULL << size * 8) - 1);
12037 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12043 /* keep walking new program and skip insns we just inserted */
12044 env->prog = new_prog;
12045 insn = new_prog->insnsi + i + delta;
12051 static int jit_subprogs(struct bpf_verifier_env *env)
12053 struct bpf_prog *prog = env->prog, **func, *tmp;
12054 int i, j, subprog_start, subprog_end = 0, len, subprog;
12055 struct bpf_map *map_ptr;
12056 struct bpf_insn *insn;
12057 void *old_bpf_func;
12058 int err, num_exentries;
12060 if (env->subprog_cnt <= 1)
12063 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
12064 if (bpf_pseudo_func(insn)) {
12065 env->insn_aux_data[i].call_imm = insn->imm;
12066 /* subprog is encoded in insn[1].imm */
12070 if (!bpf_pseudo_call(insn))
12072 /* Upon error here we cannot fall back to interpreter but
12073 * need a hard reject of the program. Thus -EFAULT is
12074 * propagated in any case.
12076 subprog = find_subprog(env, i + insn->imm + 1);
12078 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
12079 i + insn->imm + 1);
12082 /* temporarily remember subprog id inside insn instead of
12083 * aux_data, since next loop will split up all insns into funcs
12085 insn->off = subprog;
12086 /* remember original imm in case JIT fails and fallback
12087 * to interpreter will be needed
12089 env->insn_aux_data[i].call_imm = insn->imm;
12090 /* point imm to __bpf_call_base+1 from JITs point of view */
12094 err = bpf_prog_alloc_jited_linfo(prog);
12096 goto out_undo_insn;
12099 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
12101 goto out_undo_insn;
12103 for (i = 0; i < env->subprog_cnt; i++) {
12104 subprog_start = subprog_end;
12105 subprog_end = env->subprog_info[i + 1].start;
12107 len = subprog_end - subprog_start;
12108 /* BPF_PROG_RUN doesn't call subprogs directly,
12109 * hence main prog stats include the runtime of subprogs.
12110 * subprogs don't have IDs and not reachable via prog_get_next_id
12111 * func[i]->stats will never be accessed and stays NULL
12113 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
12116 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
12117 len * sizeof(struct bpf_insn));
12118 func[i]->type = prog->type;
12119 func[i]->len = len;
12120 if (bpf_prog_calc_tag(func[i]))
12122 func[i]->is_func = 1;
12123 func[i]->aux->func_idx = i;
12124 /* the btf and func_info will be freed only at prog->aux */
12125 func[i]->aux->btf = prog->aux->btf;
12126 func[i]->aux->func_info = prog->aux->func_info;
12128 for (j = 0; j < prog->aux->size_poke_tab; j++) {
12129 u32 insn_idx = prog->aux->poke_tab[j].insn_idx;
12132 if (!(insn_idx >= subprog_start &&
12133 insn_idx <= subprog_end))
12136 ret = bpf_jit_add_poke_descriptor(func[i],
12137 &prog->aux->poke_tab[j]);
12139 verbose(env, "adding tail call poke descriptor failed\n");
12143 func[i]->insnsi[insn_idx - subprog_start].imm = ret + 1;
12145 map_ptr = func[i]->aux->poke_tab[ret].tail_call.map;
12146 ret = map_ptr->ops->map_poke_track(map_ptr, func[i]->aux);
12148 verbose(env, "tracking tail call prog failed\n");
12153 /* Use bpf_prog_F_tag to indicate functions in stack traces.
12154 * Long term would need debug info to populate names
12156 func[i]->aux->name[0] = 'F';
12157 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
12158 func[i]->jit_requested = 1;
12159 func[i]->aux->kfunc_tab = prog->aux->kfunc_tab;
12160 func[i]->aux->linfo = prog->aux->linfo;
12161 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
12162 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
12163 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
12165 insn = func[i]->insnsi;
12166 for (j = 0; j < func[i]->len; j++, insn++) {
12167 if (BPF_CLASS(insn->code) == BPF_LDX &&
12168 BPF_MODE(insn->code) == BPF_PROBE_MEM)
12171 func[i]->aux->num_exentries = num_exentries;
12172 func[i]->aux->tail_call_reachable = env->subprog_info[i].tail_call_reachable;
12173 func[i] = bpf_int_jit_compile(func[i]);
12174 if (!func[i]->jited) {
12181 /* Untrack main program's aux structs so that during map_poke_run()
12182 * we will not stumble upon the unfilled poke descriptors; each
12183 * of the main program's poke descs got distributed across subprogs
12184 * and got tracked onto map, so we are sure that none of them will
12185 * be missed after the operation below
12187 for (i = 0; i < prog->aux->size_poke_tab; i++) {
12188 map_ptr = prog->aux->poke_tab[i].tail_call.map;
12190 map_ptr->ops->map_poke_untrack(map_ptr, prog->aux);
12193 /* at this point all bpf functions were successfully JITed
12194 * now populate all bpf_calls with correct addresses and
12195 * run last pass of JIT
12197 for (i = 0; i < env->subprog_cnt; i++) {
12198 insn = func[i]->insnsi;
12199 for (j = 0; j < func[i]->len; j++, insn++) {
12200 if (bpf_pseudo_func(insn)) {
12201 subprog = insn[1].imm;
12202 insn[0].imm = (u32)(long)func[subprog]->bpf_func;
12203 insn[1].imm = ((u64)(long)func[subprog]->bpf_func) >> 32;
12206 if (!bpf_pseudo_call(insn))
12208 subprog = insn->off;
12209 insn->imm = BPF_CAST_CALL(func[subprog]->bpf_func) -
12213 /* we use the aux data to keep a list of the start addresses
12214 * of the JITed images for each function in the program
12216 * for some architectures, such as powerpc64, the imm field
12217 * might not be large enough to hold the offset of the start
12218 * address of the callee's JITed image from __bpf_call_base
12220 * in such cases, we can lookup the start address of a callee
12221 * by using its subprog id, available from the off field of
12222 * the call instruction, as an index for this list
12224 func[i]->aux->func = func;
12225 func[i]->aux->func_cnt = env->subprog_cnt;
12227 for (i = 0; i < env->subprog_cnt; i++) {
12228 old_bpf_func = func[i]->bpf_func;
12229 tmp = bpf_int_jit_compile(func[i]);
12230 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
12231 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
12238 /* finally lock prog and jit images for all functions and
12239 * populate kallsysm
12241 for (i = 0; i < env->subprog_cnt; i++) {
12242 bpf_prog_lock_ro(func[i]);
12243 bpf_prog_kallsyms_add(func[i]);
12246 /* Last step: make now unused interpreter insns from main
12247 * prog consistent for later dump requests, so they can
12248 * later look the same as if they were interpreted only.
12250 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
12251 if (bpf_pseudo_func(insn)) {
12252 insn[0].imm = env->insn_aux_data[i].call_imm;
12253 insn[1].imm = find_subprog(env, i + insn[0].imm + 1);
12256 if (!bpf_pseudo_call(insn))
12258 insn->off = env->insn_aux_data[i].call_imm;
12259 subprog = find_subprog(env, i + insn->off + 1);
12260 insn->imm = subprog;
12264 prog->bpf_func = func[0]->bpf_func;
12265 prog->aux->func = func;
12266 prog->aux->func_cnt = env->subprog_cnt;
12267 bpf_prog_jit_attempt_done(prog);
12270 for (i = 0; i < env->subprog_cnt; i++) {
12274 for (j = 0; j < func[i]->aux->size_poke_tab; j++) {
12275 map_ptr = func[i]->aux->poke_tab[j].tail_call.map;
12276 map_ptr->ops->map_poke_untrack(map_ptr, func[i]->aux);
12278 bpf_jit_free(func[i]);
12282 /* cleanup main prog to be interpreted */
12283 prog->jit_requested = 0;
12284 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
12285 if (!bpf_pseudo_call(insn))
12288 insn->imm = env->insn_aux_data[i].call_imm;
12290 bpf_prog_jit_attempt_done(prog);
12294 static int fixup_call_args(struct bpf_verifier_env *env)
12296 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
12297 struct bpf_prog *prog = env->prog;
12298 struct bpf_insn *insn = prog->insnsi;
12299 bool has_kfunc_call = bpf_prog_has_kfunc_call(prog);
12304 if (env->prog->jit_requested &&
12305 !bpf_prog_is_dev_bound(env->prog->aux)) {
12306 err = jit_subprogs(env);
12309 if (err == -EFAULT)
12312 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
12313 if (has_kfunc_call) {
12314 verbose(env, "calling kernel functions are not allowed in non-JITed programs\n");
12317 if (env->subprog_cnt > 1 && env->prog->aux->tail_call_reachable) {
12318 /* When JIT fails the progs with bpf2bpf calls and tail_calls
12319 * have to be rejected, since interpreter doesn't support them yet.
12321 verbose(env, "tail_calls are not allowed in non-JITed programs with bpf-to-bpf calls\n");
12324 for (i = 0; i < prog->len; i++, insn++) {
12325 if (bpf_pseudo_func(insn)) {
12326 /* When JIT fails the progs with callback calls
12327 * have to be rejected, since interpreter doesn't support them yet.
12329 verbose(env, "callbacks are not allowed in non-JITed programs\n");
12333 if (!bpf_pseudo_call(insn))
12335 depth = get_callee_stack_depth(env, insn, i);
12338 bpf_patch_call_args(insn, depth);
12345 static int fixup_kfunc_call(struct bpf_verifier_env *env,
12346 struct bpf_insn *insn)
12348 const struct bpf_kfunc_desc *desc;
12350 /* insn->imm has the btf func_id. Replace it with
12351 * an address (relative to __bpf_base_call).
12353 desc = find_kfunc_desc(env->prog, insn->imm);
12355 verbose(env, "verifier internal error: kernel function descriptor not found for func_id %u\n",
12360 insn->imm = desc->imm;
12365 /* Do various post-verification rewrites in a single program pass.
12366 * These rewrites simplify JIT and interpreter implementations.
12368 static int do_misc_fixups(struct bpf_verifier_env *env)
12370 struct bpf_prog *prog = env->prog;
12371 bool expect_blinding = bpf_jit_blinding_enabled(prog);
12372 struct bpf_insn *insn = prog->insnsi;
12373 const struct bpf_func_proto *fn;
12374 const int insn_cnt = prog->len;
12375 const struct bpf_map_ops *ops;
12376 struct bpf_insn_aux_data *aux;
12377 struct bpf_insn insn_buf[16];
12378 struct bpf_prog *new_prog;
12379 struct bpf_map *map_ptr;
12380 int i, ret, cnt, delta = 0;
12382 for (i = 0; i < insn_cnt; i++, insn++) {
12383 /* Make divide-by-zero exceptions impossible. */
12384 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
12385 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
12386 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
12387 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
12388 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
12389 bool isdiv = BPF_OP(insn->code) == BPF_DIV;
12390 struct bpf_insn *patchlet;
12391 struct bpf_insn chk_and_div[] = {
12392 /* [R,W]x div 0 -> 0 */
12393 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
12394 BPF_JNE | BPF_K, insn->src_reg,
12396 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
12397 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
12400 struct bpf_insn chk_and_mod[] = {
12401 /* [R,W]x mod 0 -> [R,W]x */
12402 BPF_RAW_INSN((is64 ? BPF_JMP : BPF_JMP32) |
12403 BPF_JEQ | BPF_K, insn->src_reg,
12404 0, 1 + (is64 ? 0 : 1), 0),
12406 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
12407 BPF_MOV32_REG(insn->dst_reg, insn->dst_reg),
12410 patchlet = isdiv ? chk_and_div : chk_and_mod;
12411 cnt = isdiv ? ARRAY_SIZE(chk_and_div) :
12412 ARRAY_SIZE(chk_and_mod) - (is64 ? 2 : 0);
12414 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
12419 env->prog = prog = new_prog;
12420 insn = new_prog->insnsi + i + delta;
12424 /* Implement LD_ABS and LD_IND with a rewrite, if supported by the program type. */
12425 if (BPF_CLASS(insn->code) == BPF_LD &&
12426 (BPF_MODE(insn->code) == BPF_ABS ||
12427 BPF_MODE(insn->code) == BPF_IND)) {
12428 cnt = env->ops->gen_ld_abs(insn, insn_buf);
12429 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
12430 verbose(env, "bpf verifier is misconfigured\n");
12434 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12439 env->prog = prog = new_prog;
12440 insn = new_prog->insnsi + i + delta;
12444 /* Rewrite pointer arithmetic to mitigate speculation attacks. */
12445 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
12446 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
12447 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
12448 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
12449 struct bpf_insn *patch = &insn_buf[0];
12450 bool issrc, isneg, isimm;
12453 aux = &env->insn_aux_data[i + delta];
12454 if (!aux->alu_state ||
12455 aux->alu_state == BPF_ALU_NON_POINTER)
12458 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
12459 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
12460 BPF_ALU_SANITIZE_SRC;
12461 isimm = aux->alu_state & BPF_ALU_IMMEDIATE;
12463 off_reg = issrc ? insn->src_reg : insn->dst_reg;
12465 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
12468 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
12469 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit);
12470 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
12471 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
12472 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
12473 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
12474 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX, off_reg);
12477 *patch++ = BPF_MOV64_REG(insn->dst_reg, insn->src_reg);
12478 insn->src_reg = BPF_REG_AX;
12480 insn->code = insn->code == code_add ?
12481 code_sub : code_add;
12483 if (issrc && isneg && !isimm)
12484 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
12485 cnt = patch - insn_buf;
12487 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12492 env->prog = prog = new_prog;
12493 insn = new_prog->insnsi + i + delta;
12497 if (insn->code != (BPF_JMP | BPF_CALL))
12499 if (insn->src_reg == BPF_PSEUDO_CALL)
12501 if (insn->src_reg == BPF_PSEUDO_KFUNC_CALL) {
12502 ret = fixup_kfunc_call(env, insn);
12508 if (insn->imm == BPF_FUNC_get_route_realm)
12509 prog->dst_needed = 1;
12510 if (insn->imm == BPF_FUNC_get_prandom_u32)
12511 bpf_user_rnd_init_once();
12512 if (insn->imm == BPF_FUNC_override_return)
12513 prog->kprobe_override = 1;
12514 if (insn->imm == BPF_FUNC_tail_call) {
12515 /* If we tail call into other programs, we
12516 * cannot make any assumptions since they can
12517 * be replaced dynamically during runtime in
12518 * the program array.
12520 prog->cb_access = 1;
12521 if (!allow_tail_call_in_subprogs(env))
12522 prog->aux->stack_depth = MAX_BPF_STACK;
12523 prog->aux->max_pkt_offset = MAX_PACKET_OFF;
12525 /* mark bpf_tail_call as different opcode to avoid
12526 * conditional branch in the interpreter for every normal
12527 * call and to prevent accidental JITing by JIT compiler
12528 * that doesn't support bpf_tail_call yet
12531 insn->code = BPF_JMP | BPF_TAIL_CALL;
12533 aux = &env->insn_aux_data[i + delta];
12534 if (env->bpf_capable && !expect_blinding &&
12535 prog->jit_requested &&
12536 !bpf_map_key_poisoned(aux) &&
12537 !bpf_map_ptr_poisoned(aux) &&
12538 !bpf_map_ptr_unpriv(aux)) {
12539 struct bpf_jit_poke_descriptor desc = {
12540 .reason = BPF_POKE_REASON_TAIL_CALL,
12541 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
12542 .tail_call.key = bpf_map_key_immediate(aux),
12543 .insn_idx = i + delta,
12546 ret = bpf_jit_add_poke_descriptor(prog, &desc);
12548 verbose(env, "adding tail call poke descriptor failed\n");
12552 insn->imm = ret + 1;
12556 if (!bpf_map_ptr_unpriv(aux))
12559 /* instead of changing every JIT dealing with tail_call
12560 * emit two extra insns:
12561 * if (index >= max_entries) goto out;
12562 * index &= array->index_mask;
12563 * to avoid out-of-bounds cpu speculation
12565 if (bpf_map_ptr_poisoned(aux)) {
12566 verbose(env, "tail_call abusing map_ptr\n");
12570 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
12571 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
12572 map_ptr->max_entries, 2);
12573 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
12574 container_of(map_ptr,
12577 insn_buf[2] = *insn;
12579 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
12584 env->prog = prog = new_prog;
12585 insn = new_prog->insnsi + i + delta;
12589 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
12590 * and other inlining handlers are currently limited to 64 bit
12593 if (prog->jit_requested && BITS_PER_LONG == 64 &&
12594 (insn->imm == BPF_FUNC_map_lookup_elem ||
12595 insn->imm == BPF_FUNC_map_update_elem ||
12596 insn->imm == BPF_FUNC_map_delete_elem ||
12597 insn->imm == BPF_FUNC_map_push_elem ||
12598 insn->imm == BPF_FUNC_map_pop_elem ||
12599 insn->imm == BPF_FUNC_map_peek_elem ||
12600 insn->imm == BPF_FUNC_redirect_map)) {
12601 aux = &env->insn_aux_data[i + delta];
12602 if (bpf_map_ptr_poisoned(aux))
12603 goto patch_call_imm;
12605 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
12606 ops = map_ptr->ops;
12607 if (insn->imm == BPF_FUNC_map_lookup_elem &&
12608 ops->map_gen_lookup) {
12609 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
12610 if (cnt == -EOPNOTSUPP)
12611 goto patch_map_ops_generic;
12612 if (cnt <= 0 || cnt >= ARRAY_SIZE(insn_buf)) {
12613 verbose(env, "bpf verifier is misconfigured\n");
12617 new_prog = bpf_patch_insn_data(env, i + delta,
12623 env->prog = prog = new_prog;
12624 insn = new_prog->insnsi + i + delta;
12628 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
12629 (void *(*)(struct bpf_map *map, void *key))NULL));
12630 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
12631 (int (*)(struct bpf_map *map, void *key))NULL));
12632 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
12633 (int (*)(struct bpf_map *map, void *key, void *value,
12635 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
12636 (int (*)(struct bpf_map *map, void *value,
12638 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
12639 (int (*)(struct bpf_map *map, void *value))NULL));
12640 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
12641 (int (*)(struct bpf_map *map, void *value))NULL));
12642 BUILD_BUG_ON(!__same_type(ops->map_redirect,
12643 (int (*)(struct bpf_map *map, u32 ifindex, u64 flags))NULL));
12645 patch_map_ops_generic:
12646 switch (insn->imm) {
12647 case BPF_FUNC_map_lookup_elem:
12648 insn->imm = BPF_CAST_CALL(ops->map_lookup_elem) -
12651 case BPF_FUNC_map_update_elem:
12652 insn->imm = BPF_CAST_CALL(ops->map_update_elem) -
12655 case BPF_FUNC_map_delete_elem:
12656 insn->imm = BPF_CAST_CALL(ops->map_delete_elem) -
12659 case BPF_FUNC_map_push_elem:
12660 insn->imm = BPF_CAST_CALL(ops->map_push_elem) -
12663 case BPF_FUNC_map_pop_elem:
12664 insn->imm = BPF_CAST_CALL(ops->map_pop_elem) -
12667 case BPF_FUNC_map_peek_elem:
12668 insn->imm = BPF_CAST_CALL(ops->map_peek_elem) -
12671 case BPF_FUNC_redirect_map:
12672 insn->imm = BPF_CAST_CALL(ops->map_redirect) -
12677 goto patch_call_imm;
12680 /* Implement bpf_jiffies64 inline. */
12681 if (prog->jit_requested && BITS_PER_LONG == 64 &&
12682 insn->imm == BPF_FUNC_jiffies64) {
12683 struct bpf_insn ld_jiffies_addr[2] = {
12684 BPF_LD_IMM64(BPF_REG_0,
12685 (unsigned long)&jiffies),
12688 insn_buf[0] = ld_jiffies_addr[0];
12689 insn_buf[1] = ld_jiffies_addr[1];
12690 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
12694 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
12700 env->prog = prog = new_prog;
12701 insn = new_prog->insnsi + i + delta;
12706 fn = env->ops->get_func_proto(insn->imm, env->prog);
12707 /* all functions that have prototype and verifier allowed
12708 * programs to call them, must be real in-kernel functions
12712 "kernel subsystem misconfigured func %s#%d\n",
12713 func_id_name(insn->imm), insn->imm);
12716 insn->imm = fn->func - __bpf_call_base;
12719 /* Since poke tab is now finalized, publish aux to tracker. */
12720 for (i = 0; i < prog->aux->size_poke_tab; i++) {
12721 map_ptr = prog->aux->poke_tab[i].tail_call.map;
12722 if (!map_ptr->ops->map_poke_track ||
12723 !map_ptr->ops->map_poke_untrack ||
12724 !map_ptr->ops->map_poke_run) {
12725 verbose(env, "bpf verifier is misconfigured\n");
12729 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
12731 verbose(env, "tracking tail call prog failed\n");
12736 sort_kfunc_descs_by_imm(env->prog);
12741 static void free_states(struct bpf_verifier_env *env)
12743 struct bpf_verifier_state_list *sl, *sln;
12746 sl = env->free_list;
12749 free_verifier_state(&sl->state, false);
12753 env->free_list = NULL;
12755 if (!env->explored_states)
12758 for (i = 0; i < state_htab_size(env); i++) {
12759 sl = env->explored_states[i];
12763 free_verifier_state(&sl->state, false);
12767 env->explored_states[i] = NULL;
12771 /* The verifier is using insn_aux_data[] to store temporary data during
12772 * verification and to store information for passes that run after the
12773 * verification like dead code sanitization. do_check_common() for subprogram N
12774 * may analyze many other subprograms. sanitize_insn_aux_data() clears all
12775 * temporary data after do_check_common() finds that subprogram N cannot be
12776 * verified independently. pass_cnt counts the number of times
12777 * do_check_common() was run and insn->aux->seen tells the pass number
12778 * insn_aux_data was touched. These variables are compared to clear temporary
12779 * data from failed pass. For testing and experiments do_check_common() can be
12780 * run multiple times even when prior attempt to verify is unsuccessful.
12782 * Note that special handling is needed on !env->bypass_spec_v1 if this is
12783 * ever called outside of error path with subsequent program rejection.
12785 static void sanitize_insn_aux_data(struct bpf_verifier_env *env)
12787 struct bpf_insn *insn = env->prog->insnsi;
12788 struct bpf_insn_aux_data *aux;
12791 for (i = 0; i < env->prog->len; i++) {
12792 class = BPF_CLASS(insn[i].code);
12793 if (class != BPF_LDX && class != BPF_STX)
12795 aux = &env->insn_aux_data[i];
12796 if (aux->seen != env->pass_cnt)
12798 memset(aux, 0, offsetof(typeof(*aux), orig_idx));
12802 static int do_check_common(struct bpf_verifier_env *env, int subprog)
12804 bool pop_log = !(env->log.level & BPF_LOG_LEVEL2);
12805 struct bpf_verifier_state *state;
12806 struct bpf_reg_state *regs;
12809 env->prev_linfo = NULL;
12812 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
12815 state->curframe = 0;
12816 state->speculative = false;
12817 state->branches = 1;
12818 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
12819 if (!state->frame[0]) {
12823 env->cur_state = state;
12824 init_func_state(env, state->frame[0],
12825 BPF_MAIN_FUNC /* callsite */,
12829 regs = state->frame[state->curframe]->regs;
12830 if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
12831 ret = btf_prepare_func_args(env, subprog, regs);
12834 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
12835 if (regs[i].type == PTR_TO_CTX)
12836 mark_reg_known_zero(env, regs, i);
12837 else if (regs[i].type == SCALAR_VALUE)
12838 mark_reg_unknown(env, regs, i);
12839 else if (regs[i].type == PTR_TO_MEM_OR_NULL) {
12840 const u32 mem_size = regs[i].mem_size;
12842 mark_reg_known_zero(env, regs, i);
12843 regs[i].mem_size = mem_size;
12844 regs[i].id = ++env->id_gen;
12848 /* 1st arg to a function */
12849 regs[BPF_REG_1].type = PTR_TO_CTX;
12850 mark_reg_known_zero(env, regs, BPF_REG_1);
12851 ret = btf_check_subprog_arg_match(env, subprog, regs);
12852 if (ret == -EFAULT)
12853 /* unlikely verifier bug. abort.
12854 * ret == 0 and ret < 0 are sadly acceptable for
12855 * main() function due to backward compatibility.
12856 * Like socket filter program may be written as:
12857 * int bpf_prog(struct pt_regs *ctx)
12858 * and never dereference that ctx in the program.
12859 * 'struct pt_regs' is a type mismatch for socket
12860 * filter that should be using 'struct __sk_buff'.
12865 ret = do_check(env);
12867 /* check for NULL is necessary, since cur_state can be freed inside
12868 * do_check() under memory pressure.
12870 if (env->cur_state) {
12871 free_verifier_state(env->cur_state, true);
12872 env->cur_state = NULL;
12874 while (!pop_stack(env, NULL, NULL, false));
12875 if (!ret && pop_log)
12876 bpf_vlog_reset(&env->log, 0);
12879 /* clean aux data in case subprog was rejected */
12880 sanitize_insn_aux_data(env);
12884 /* Verify all global functions in a BPF program one by one based on their BTF.
12885 * All global functions must pass verification. Otherwise the whole program is rejected.
12896 * foo() will be verified first for R1=any_scalar_value. During verification it
12897 * will be assumed that bar() already verified successfully and call to bar()
12898 * from foo() will be checked for type match only. Later bar() will be verified
12899 * independently to check that it's safe for R1=any_scalar_value.
12901 static int do_check_subprogs(struct bpf_verifier_env *env)
12903 struct bpf_prog_aux *aux = env->prog->aux;
12906 if (!aux->func_info)
12909 for (i = 1; i < env->subprog_cnt; i++) {
12910 if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL)
12912 env->insn_idx = env->subprog_info[i].start;
12913 WARN_ON_ONCE(env->insn_idx == 0);
12914 ret = do_check_common(env, i);
12917 } else if (env->log.level & BPF_LOG_LEVEL) {
12919 "Func#%d is safe for any args that match its prototype\n",
12926 static int do_check_main(struct bpf_verifier_env *env)
12931 ret = do_check_common(env, 0);
12933 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
12938 static void print_verification_stats(struct bpf_verifier_env *env)
12942 if (env->log.level & BPF_LOG_STATS) {
12943 verbose(env, "verification time %lld usec\n",
12944 div_u64(env->verification_time, 1000));
12945 verbose(env, "stack depth ");
12946 for (i = 0; i < env->subprog_cnt; i++) {
12947 u32 depth = env->subprog_info[i].stack_depth;
12949 verbose(env, "%d", depth);
12950 if (i + 1 < env->subprog_cnt)
12953 verbose(env, "\n");
12955 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
12956 "total_states %d peak_states %d mark_read %d\n",
12957 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
12958 env->max_states_per_insn, env->total_states,
12959 env->peak_states, env->longest_mark_read_walk);
12962 static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
12964 const struct btf_type *t, *func_proto;
12965 const struct bpf_struct_ops *st_ops;
12966 const struct btf_member *member;
12967 struct bpf_prog *prog = env->prog;
12968 u32 btf_id, member_idx;
12971 if (!prog->gpl_compatible) {
12972 verbose(env, "struct ops programs must have a GPL compatible license\n");
12976 btf_id = prog->aux->attach_btf_id;
12977 st_ops = bpf_struct_ops_find(btf_id);
12979 verbose(env, "attach_btf_id %u is not a supported struct\n",
12985 member_idx = prog->expected_attach_type;
12986 if (member_idx >= btf_type_vlen(t)) {
12987 verbose(env, "attach to invalid member idx %u of struct %s\n",
12988 member_idx, st_ops->name);
12992 member = &btf_type_member(t)[member_idx];
12993 mname = btf_name_by_offset(btf_vmlinux, member->name_off);
12994 func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type,
12997 verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
12998 mname, member_idx, st_ops->name);
13002 if (st_ops->check_member) {
13003 int err = st_ops->check_member(t, member);
13006 verbose(env, "attach to unsupported member %s of struct %s\n",
13007 mname, st_ops->name);
13012 prog->aux->attach_func_proto = func_proto;
13013 prog->aux->attach_func_name = mname;
13014 env->ops = st_ops->verifier_ops;
13018 #define SECURITY_PREFIX "security_"
13020 static int check_attach_modify_return(unsigned long addr, const char *func_name)
13022 if (within_error_injection_list(addr) ||
13023 !strncmp(SECURITY_PREFIX, func_name, sizeof(SECURITY_PREFIX) - 1))
13029 /* list of non-sleepable functions that are otherwise on
13030 * ALLOW_ERROR_INJECTION list
13032 BTF_SET_START(btf_non_sleepable_error_inject)
13033 /* Three functions below can be called from sleepable and non-sleepable context.
13034 * Assume non-sleepable from bpf safety point of view.
13036 BTF_ID(func, __add_to_page_cache_locked)
13037 BTF_ID(func, should_fail_alloc_page)
13038 BTF_ID(func, should_failslab)
13039 BTF_SET_END(btf_non_sleepable_error_inject)
13041 static int check_non_sleepable_error_inject(u32 btf_id)
13043 return btf_id_set_contains(&btf_non_sleepable_error_inject, btf_id);
13046 int bpf_check_attach_target(struct bpf_verifier_log *log,
13047 const struct bpf_prog *prog,
13048 const struct bpf_prog *tgt_prog,
13050 struct bpf_attach_target_info *tgt_info)
13052 bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
13053 const char prefix[] = "btf_trace_";
13054 int ret = 0, subprog = -1, i;
13055 const struct btf_type *t;
13056 bool conservative = true;
13062 bpf_log(log, "Tracing programs must provide btf_id\n");
13065 btf = tgt_prog ? tgt_prog->aux->btf : prog->aux->attach_btf;
13068 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
13071 t = btf_type_by_id(btf, btf_id);
13073 bpf_log(log, "attach_btf_id %u is invalid\n", btf_id);
13076 tname = btf_name_by_offset(btf, t->name_off);
13078 bpf_log(log, "attach_btf_id %u doesn't have a name\n", btf_id);
13082 struct bpf_prog_aux *aux = tgt_prog->aux;
13084 for (i = 0; i < aux->func_info_cnt; i++)
13085 if (aux->func_info[i].type_id == btf_id) {
13089 if (subprog == -1) {
13090 bpf_log(log, "Subprog %s doesn't exist\n", tname);
13093 conservative = aux->func_info_aux[subprog].unreliable;
13094 if (prog_extension) {
13095 if (conservative) {
13097 "Cannot replace static functions\n");
13100 if (!prog->jit_requested) {
13102 "Extension programs should be JITed\n");
13106 if (!tgt_prog->jited) {
13107 bpf_log(log, "Can attach to only JITed progs\n");
13110 if (tgt_prog->type == prog->type) {
13111 /* Cannot fentry/fexit another fentry/fexit program.
13112 * Cannot attach program extension to another extension.
13113 * It's ok to attach fentry/fexit to extension program.
13115 bpf_log(log, "Cannot recursively attach\n");
13118 if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
13120 (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
13121 tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
13122 /* Program extensions can extend all program types
13123 * except fentry/fexit. The reason is the following.
13124 * The fentry/fexit programs are used for performance
13125 * analysis, stats and can be attached to any program
13126 * type except themselves. When extension program is
13127 * replacing XDP function it is necessary to allow
13128 * performance analysis of all functions. Both original
13129 * XDP program and its program extension. Hence
13130 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
13131 * allowed. If extending of fentry/fexit was allowed it
13132 * would be possible to create long call chain
13133 * fentry->extension->fentry->extension beyond
13134 * reasonable stack size. Hence extending fentry is not
13137 bpf_log(log, "Cannot extend fentry/fexit\n");
13141 if (prog_extension) {
13142 bpf_log(log, "Cannot replace kernel functions\n");
13147 switch (prog->expected_attach_type) {
13148 case BPF_TRACE_RAW_TP:
13151 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
13154 if (!btf_type_is_typedef(t)) {
13155 bpf_log(log, "attach_btf_id %u is not a typedef\n",
13159 if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
13160 bpf_log(log, "attach_btf_id %u points to wrong type name %s\n",
13164 tname += sizeof(prefix) - 1;
13165 t = btf_type_by_id(btf, t->type);
13166 if (!btf_type_is_ptr(t))
13167 /* should never happen in valid vmlinux build */
13169 t = btf_type_by_id(btf, t->type);
13170 if (!btf_type_is_func_proto(t))
13171 /* should never happen in valid vmlinux build */
13175 case BPF_TRACE_ITER:
13176 if (!btf_type_is_func(t)) {
13177 bpf_log(log, "attach_btf_id %u is not a function\n",
13181 t = btf_type_by_id(btf, t->type);
13182 if (!btf_type_is_func_proto(t))
13184 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
13189 if (!prog_extension)
13192 case BPF_MODIFY_RETURN:
13194 case BPF_TRACE_FENTRY:
13195 case BPF_TRACE_FEXIT:
13196 if (!btf_type_is_func(t)) {
13197 bpf_log(log, "attach_btf_id %u is not a function\n",
13201 if (prog_extension &&
13202 btf_check_type_match(log, prog, btf, t))
13204 t = btf_type_by_id(btf, t->type);
13205 if (!btf_type_is_func_proto(t))
13208 if ((prog->aux->saved_dst_prog_type || prog->aux->saved_dst_attach_type) &&
13209 (!tgt_prog || prog->aux->saved_dst_prog_type != tgt_prog->type ||
13210 prog->aux->saved_dst_attach_type != tgt_prog->expected_attach_type))
13213 if (tgt_prog && conservative)
13216 ret = btf_distill_func_proto(log, btf, t, tname, &tgt_info->fmodel);
13222 addr = (long) tgt_prog->bpf_func;
13224 addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
13226 addr = kallsyms_lookup_name(tname);
13229 "The address of function %s cannot be found\n",
13235 if (prog->aux->sleepable) {
13237 switch (prog->type) {
13238 case BPF_PROG_TYPE_TRACING:
13239 /* fentry/fexit/fmod_ret progs can be sleepable only if they are
13240 * attached to ALLOW_ERROR_INJECTION and are not in denylist.
13242 if (!check_non_sleepable_error_inject(btf_id) &&
13243 within_error_injection_list(addr))
13246 case BPF_PROG_TYPE_LSM:
13247 /* LSM progs check that they are attached to bpf_lsm_*() funcs.
13248 * Only some of them are sleepable.
13250 if (bpf_lsm_is_sleepable_hook(btf_id))
13257 bpf_log(log, "%s is not sleepable\n", tname);
13260 } else if (prog->expected_attach_type == BPF_MODIFY_RETURN) {
13262 bpf_log(log, "can't modify return codes of BPF programs\n");
13265 ret = check_attach_modify_return(addr, tname);
13267 bpf_log(log, "%s() is not modifiable\n", tname);
13274 tgt_info->tgt_addr = addr;
13275 tgt_info->tgt_name = tname;
13276 tgt_info->tgt_type = t;
13280 BTF_SET_START(btf_id_deny)
13283 BTF_ID(func, migrate_disable)
13284 BTF_ID(func, migrate_enable)
13286 #if !defined CONFIG_PREEMPT_RCU && !defined CONFIG_TINY_RCU
13287 BTF_ID(func, rcu_read_unlock_strict)
13289 BTF_SET_END(btf_id_deny)
13291 static int check_attach_btf_id(struct bpf_verifier_env *env)
13293 struct bpf_prog *prog = env->prog;
13294 struct bpf_prog *tgt_prog = prog->aux->dst_prog;
13295 struct bpf_attach_target_info tgt_info = {};
13296 u32 btf_id = prog->aux->attach_btf_id;
13297 struct bpf_trampoline *tr;
13301 if (prog->type == BPF_PROG_TYPE_SYSCALL) {
13302 if (prog->aux->sleepable)
13303 /* attach_btf_id checked to be zero already */
13305 verbose(env, "Syscall programs can only be sleepable\n");
13309 if (prog->aux->sleepable && prog->type != BPF_PROG_TYPE_TRACING &&
13310 prog->type != BPF_PROG_TYPE_LSM) {
13311 verbose(env, "Only fentry/fexit/fmod_ret and lsm programs can be sleepable\n");
13315 if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
13316 return check_struct_ops_btf_id(env);
13318 if (prog->type != BPF_PROG_TYPE_TRACING &&
13319 prog->type != BPF_PROG_TYPE_LSM &&
13320 prog->type != BPF_PROG_TYPE_EXT)
13323 ret = bpf_check_attach_target(&env->log, prog, tgt_prog, btf_id, &tgt_info);
13327 if (tgt_prog && prog->type == BPF_PROG_TYPE_EXT) {
13328 /* to make freplace equivalent to their targets, they need to
13329 * inherit env->ops and expected_attach_type for the rest of the
13332 env->ops = bpf_verifier_ops[tgt_prog->type];
13333 prog->expected_attach_type = tgt_prog->expected_attach_type;
13336 /* store info about the attachment target that will be used later */
13337 prog->aux->attach_func_proto = tgt_info.tgt_type;
13338 prog->aux->attach_func_name = tgt_info.tgt_name;
13341 prog->aux->saved_dst_prog_type = tgt_prog->type;
13342 prog->aux->saved_dst_attach_type = tgt_prog->expected_attach_type;
13345 if (prog->expected_attach_type == BPF_TRACE_RAW_TP) {
13346 prog->aux->attach_btf_trace = true;
13348 } else if (prog->expected_attach_type == BPF_TRACE_ITER) {
13349 if (!bpf_iter_prog_supported(prog))
13354 if (prog->type == BPF_PROG_TYPE_LSM) {
13355 ret = bpf_lsm_verify_prog(&env->log, prog);
13358 } else if (prog->type == BPF_PROG_TYPE_TRACING &&
13359 btf_id_set_contains(&btf_id_deny, btf_id)) {
13363 key = bpf_trampoline_compute_key(tgt_prog, prog->aux->attach_btf, btf_id);
13364 tr = bpf_trampoline_get(key, &tgt_info);
13368 prog->aux->dst_trampoline = tr;
13372 struct btf *bpf_get_btf_vmlinux(void)
13374 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
13375 mutex_lock(&bpf_verifier_lock);
13377 btf_vmlinux = btf_parse_vmlinux();
13378 mutex_unlock(&bpf_verifier_lock);
13380 return btf_vmlinux;
13383 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr, bpfptr_t uattr)
13385 u64 start_time = ktime_get_ns();
13386 struct bpf_verifier_env *env;
13387 struct bpf_verifier_log *log;
13388 int i, len, ret = -EINVAL;
13391 /* no program is valid */
13392 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
13395 /* 'struct bpf_verifier_env' can be global, but since it's not small,
13396 * allocate/free it every time bpf_check() is called
13398 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
13403 len = (*prog)->len;
13404 env->insn_aux_data =
13405 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
13407 if (!env->insn_aux_data)
13409 for (i = 0; i < len; i++)
13410 env->insn_aux_data[i].orig_idx = i;
13412 env->ops = bpf_verifier_ops[env->prog->type];
13413 env->fd_array = make_bpfptr(attr->fd_array, uattr.is_kernel);
13414 is_priv = bpf_capable();
13416 bpf_get_btf_vmlinux();
13418 /* grab the mutex to protect few globals used by verifier */
13420 mutex_lock(&bpf_verifier_lock);
13422 if (attr->log_level || attr->log_buf || attr->log_size) {
13423 /* user requested verbose verifier output
13424 * and supplied buffer to store the verification trace
13426 log->level = attr->log_level;
13427 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
13428 log->len_total = attr->log_size;
13431 /* log attributes have to be sane */
13432 if (log->len_total < 128 || log->len_total > UINT_MAX >> 2 ||
13433 !log->level || !log->ubuf || log->level & ~BPF_LOG_MASK)
13437 if (IS_ERR(btf_vmlinux)) {
13438 /* Either gcc or pahole or kernel are broken. */
13439 verbose(env, "in-kernel BTF is malformed\n");
13440 ret = PTR_ERR(btf_vmlinux);
13441 goto skip_full_check;
13444 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
13445 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
13446 env->strict_alignment = true;
13447 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
13448 env->strict_alignment = false;
13450 env->allow_ptr_leaks = bpf_allow_ptr_leaks();
13451 env->allow_uninit_stack = bpf_allow_uninit_stack();
13452 env->allow_ptr_to_map_access = bpf_allow_ptr_to_map_access();
13453 env->bypass_spec_v1 = bpf_bypass_spec_v1();
13454 env->bypass_spec_v4 = bpf_bypass_spec_v4();
13455 env->bpf_capable = bpf_capable();
13458 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
13460 env->explored_states = kvcalloc(state_htab_size(env),
13461 sizeof(struct bpf_verifier_state_list *),
13464 if (!env->explored_states)
13465 goto skip_full_check;
13467 ret = add_subprog_and_kfunc(env);
13469 goto skip_full_check;
13471 ret = check_subprogs(env);
13473 goto skip_full_check;
13475 ret = check_btf_info(env, attr, uattr);
13477 goto skip_full_check;
13479 ret = check_attach_btf_id(env);
13481 goto skip_full_check;
13483 ret = resolve_pseudo_ldimm64(env);
13485 goto skip_full_check;
13487 if (bpf_prog_is_dev_bound(env->prog->aux)) {
13488 ret = bpf_prog_offload_verifier_prep(env->prog);
13490 goto skip_full_check;
13493 ret = check_cfg(env);
13495 goto skip_full_check;
13497 ret = do_check_subprogs(env);
13498 ret = ret ?: do_check_main(env);
13500 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
13501 ret = bpf_prog_offload_finalize(env);
13504 kvfree(env->explored_states);
13507 ret = check_max_stack_depth(env);
13509 /* instruction rewrites happen after this point */
13512 opt_hard_wire_dead_code_branches(env);
13514 ret = opt_remove_dead_code(env);
13516 ret = opt_remove_nops(env);
13519 sanitize_dead_code(env);
13523 /* program is valid, convert *(u32*)(ctx + off) accesses */
13524 ret = convert_ctx_accesses(env);
13527 ret = do_misc_fixups(env);
13529 /* do 32-bit optimization after insn patching has done so those patched
13530 * insns could be handled correctly.
13532 if (ret == 0 && !bpf_prog_is_dev_bound(env->prog->aux)) {
13533 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
13534 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
13539 ret = fixup_call_args(env);
13541 env->verification_time = ktime_get_ns() - start_time;
13542 print_verification_stats(env);
13544 if (log->level && bpf_verifier_log_full(log))
13546 if (log->level && !log->ubuf) {
13548 goto err_release_maps;
13552 goto err_release_maps;
13554 if (env->used_map_cnt) {
13555 /* if program passed verifier, update used_maps in bpf_prog_info */
13556 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
13557 sizeof(env->used_maps[0]),
13560 if (!env->prog->aux->used_maps) {
13562 goto err_release_maps;
13565 memcpy(env->prog->aux->used_maps, env->used_maps,
13566 sizeof(env->used_maps[0]) * env->used_map_cnt);
13567 env->prog->aux->used_map_cnt = env->used_map_cnt;
13569 if (env->used_btf_cnt) {
13570 /* if program passed verifier, update used_btfs in bpf_prog_aux */
13571 env->prog->aux->used_btfs = kmalloc_array(env->used_btf_cnt,
13572 sizeof(env->used_btfs[0]),
13574 if (!env->prog->aux->used_btfs) {
13576 goto err_release_maps;
13579 memcpy(env->prog->aux->used_btfs, env->used_btfs,
13580 sizeof(env->used_btfs[0]) * env->used_btf_cnt);
13581 env->prog->aux->used_btf_cnt = env->used_btf_cnt;
13583 if (env->used_map_cnt || env->used_btf_cnt) {
13584 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
13585 * bpf_ld_imm64 instructions
13587 convert_pseudo_ld_imm64(env);
13590 adjust_btf_func(env);
13593 if (!env->prog->aux->used_maps)
13594 /* if we didn't copy map pointers into bpf_prog_info, release
13595 * them now. Otherwise free_used_maps() will release them.
13598 if (!env->prog->aux->used_btfs)
13601 /* extension progs temporarily inherit the attach_type of their targets
13602 for verification purposes, so set it back to zero before returning
13604 if (env->prog->type == BPF_PROG_TYPE_EXT)
13605 env->prog->expected_attach_type = 0;
13610 mutex_unlock(&bpf_verifier_lock);
13611 vfree(env->insn_aux_data);