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>
25 static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
26 #define BPF_PROG_TYPE(_id, _name, prog_ctx_type, kern_ctx_type) \
27 [_id] = & _name ## _verifier_ops,
28 #define BPF_MAP_TYPE(_id, _ops)
29 #include <linux/bpf_types.h>
34 /* bpf_check() is a static code analyzer that walks eBPF program
35 * instruction by instruction and updates register/stack state.
36 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
38 * The first pass is depth-first-search to check that the program is a DAG.
39 * It rejects the following programs:
40 * - larger than BPF_MAXINSNS insns
41 * - if loop is present (detected via back-edge)
42 * - unreachable insns exist (shouldn't be a forest. program = one function)
43 * - out of bounds or malformed jumps
44 * The second pass is all possible path descent from the 1st insn.
45 * Since it's analyzing all pathes through the program, the length of the
46 * analysis is limited to 64k insn, which may be hit even if total number of
47 * insn is less then 4K, but there are too many branches that change stack/regs.
48 * Number of 'branches to be analyzed' is limited to 1k
50 * On entry to each instruction, each register has a type, and the instruction
51 * changes the types of the registers depending on instruction semantics.
52 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
55 * All registers are 64-bit.
56 * R0 - return register
57 * R1-R5 argument passing registers
58 * R6-R9 callee saved registers
59 * R10 - frame pointer read-only
61 * At the start of BPF program the register R1 contains a pointer to bpf_context
62 * and has type PTR_TO_CTX.
64 * Verifier tracks arithmetic operations on pointers in case:
65 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
66 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
67 * 1st insn copies R10 (which has FRAME_PTR) type into R1
68 * and 2nd arithmetic instruction is pattern matched to recognize
69 * that it wants to construct a pointer to some element within stack.
70 * So after 2nd insn, the register R1 has type PTR_TO_STACK
71 * (and -20 constant is saved for further stack bounds checking).
72 * Meaning that this reg is a pointer to stack plus known immediate constant.
74 * Most of the time the registers have SCALAR_VALUE type, which
75 * means the register has some value, but it's not a valid pointer.
76 * (like pointer plus pointer becomes SCALAR_VALUE type)
78 * When verifier sees load or store instructions the type of base register
79 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK, PTR_TO_SOCKET. These are
80 * four pointer types recognized by check_mem_access() function.
82 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
83 * and the range of [ptr, ptr + map's value_size) is accessible.
85 * registers used to pass values to function calls are checked against
86 * function argument constraints.
88 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
89 * It means that the register type passed to this function must be
90 * PTR_TO_STACK and it will be used inside the function as
91 * 'pointer to map element key'
93 * For example the argument constraints for bpf_map_lookup_elem():
94 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
95 * .arg1_type = ARG_CONST_MAP_PTR,
96 * .arg2_type = ARG_PTR_TO_MAP_KEY,
98 * ret_type says that this function returns 'pointer to map elem value or null'
99 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
100 * 2nd argument should be a pointer to stack, which will be used inside
101 * the helper function as a pointer to map element key.
103 * On the kernel side the helper function looks like:
104 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
106 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
107 * void *key = (void *) (unsigned long) r2;
110 * here kernel can access 'key' and 'map' pointers safely, knowing that
111 * [key, key + map->key_size) bytes are valid and were initialized on
112 * the stack of eBPF program.
115 * Corresponding eBPF program may look like:
116 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
117 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
118 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
119 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
120 * here verifier looks at prototype of map_lookup_elem() and sees:
121 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
122 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
124 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
125 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
126 * and were initialized prior to this call.
127 * If it's ok, then verifier allows this BPF_CALL insn and looks at
128 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
129 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
130 * returns ether pointer to map value or NULL.
132 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
133 * insn, the register holding that pointer in the true branch changes state to
134 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
135 * branch. See check_cond_jmp_op().
137 * After the call R0 is set to return type of the function and registers R1-R5
138 * are set to NOT_INIT to indicate that they are no longer readable.
140 * The following reference types represent a potential reference to a kernel
141 * resource which, after first being allocated, must be checked and freed by
143 * - PTR_TO_SOCKET_OR_NULL, PTR_TO_SOCKET
145 * When the verifier sees a helper call return a reference type, it allocates a
146 * pointer id for the reference and stores it in the current function state.
147 * Similar to the way that PTR_TO_MAP_VALUE_OR_NULL is converted into
148 * PTR_TO_MAP_VALUE, PTR_TO_SOCKET_OR_NULL becomes PTR_TO_SOCKET when the type
149 * passes through a NULL-check conditional. For the branch wherein the state is
150 * changed to CONST_IMM, the verifier releases the reference.
152 * For each helper function that allocates a reference, such as
153 * bpf_sk_lookup_tcp(), there is a corresponding release function, such as
154 * bpf_sk_release(). When a reference type passes into the release function,
155 * the verifier also releases the reference. If any unchecked or unreleased
156 * reference remains at the end of the program, the verifier rejects it.
159 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
160 struct bpf_verifier_stack_elem {
161 /* verifer state is 'st'
162 * before processing instruction 'insn_idx'
163 * and after processing instruction 'prev_insn_idx'
165 struct bpf_verifier_state st;
168 struct bpf_verifier_stack_elem *next;
171 #define BPF_COMPLEXITY_LIMIT_JMP_SEQ 8192
172 #define BPF_COMPLEXITY_LIMIT_STATES 64
174 #define BPF_MAP_KEY_POISON (1ULL << 63)
175 #define BPF_MAP_KEY_SEEN (1ULL << 62)
177 #define BPF_MAP_PTR_UNPRIV 1UL
178 #define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \
179 POISON_POINTER_DELTA))
180 #define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
182 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
184 return BPF_MAP_PTR(aux->map_ptr_state) == BPF_MAP_PTR_POISON;
187 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
189 return aux->map_ptr_state & BPF_MAP_PTR_UNPRIV;
192 static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
193 const struct bpf_map *map, bool unpriv)
195 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON & BPF_MAP_PTR_UNPRIV);
196 unpriv |= bpf_map_ptr_unpriv(aux);
197 aux->map_ptr_state = (unsigned long)map |
198 (unpriv ? BPF_MAP_PTR_UNPRIV : 0UL);
201 static bool bpf_map_key_poisoned(const struct bpf_insn_aux_data *aux)
203 return aux->map_key_state & BPF_MAP_KEY_POISON;
206 static bool bpf_map_key_unseen(const struct bpf_insn_aux_data *aux)
208 return !(aux->map_key_state & BPF_MAP_KEY_SEEN);
211 static u64 bpf_map_key_immediate(const struct bpf_insn_aux_data *aux)
213 return aux->map_key_state & ~(BPF_MAP_KEY_SEEN | BPF_MAP_KEY_POISON);
216 static void bpf_map_key_store(struct bpf_insn_aux_data *aux, u64 state)
218 bool poisoned = bpf_map_key_poisoned(aux);
220 aux->map_key_state = state | BPF_MAP_KEY_SEEN |
221 (poisoned ? BPF_MAP_KEY_POISON : 0ULL);
224 struct bpf_call_arg_meta {
225 struct bpf_map *map_ptr;
230 s64 msize_smax_value;
231 u64 msize_umax_value;
237 struct btf *btf_vmlinux;
239 static DEFINE_MUTEX(bpf_verifier_lock);
241 static const struct bpf_line_info *
242 find_linfo(const struct bpf_verifier_env *env, u32 insn_off)
244 const struct bpf_line_info *linfo;
245 const struct bpf_prog *prog;
249 nr_linfo = prog->aux->nr_linfo;
251 if (!nr_linfo || insn_off >= prog->len)
254 linfo = prog->aux->linfo;
255 for (i = 1; i < nr_linfo; i++)
256 if (insn_off < linfo[i].insn_off)
259 return &linfo[i - 1];
262 void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
267 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
269 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
270 "verifier log line truncated - local buffer too short\n");
272 n = min(log->len_total - log->len_used - 1, n);
275 if (log->level == BPF_LOG_KERNEL) {
276 pr_err("BPF:%s\n", log->kbuf);
279 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
285 /* log_level controls verbosity level of eBPF verifier.
286 * bpf_verifier_log_write() is used to dump the verification trace to the log,
287 * so the user can figure out what's wrong with the program
289 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
290 const char *fmt, ...)
294 if (!bpf_verifier_log_needed(&env->log))
298 bpf_verifier_vlog(&env->log, fmt, args);
301 EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
303 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
305 struct bpf_verifier_env *env = private_data;
308 if (!bpf_verifier_log_needed(&env->log))
312 bpf_verifier_vlog(&env->log, fmt, args);
316 __printf(2, 3) void bpf_log(struct bpf_verifier_log *log,
317 const char *fmt, ...)
321 if (!bpf_verifier_log_needed(log))
325 bpf_verifier_vlog(log, fmt, args);
329 static const char *ltrim(const char *s)
337 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env,
339 const char *prefix_fmt, ...)
341 const struct bpf_line_info *linfo;
343 if (!bpf_verifier_log_needed(&env->log))
346 linfo = find_linfo(env, insn_off);
347 if (!linfo || linfo == env->prev_linfo)
353 va_start(args, prefix_fmt);
354 bpf_verifier_vlog(&env->log, prefix_fmt, args);
359 ltrim(btf_name_by_offset(env->prog->aux->btf,
362 env->prev_linfo = linfo;
365 static bool type_is_pkt_pointer(enum bpf_reg_type type)
367 return type == PTR_TO_PACKET ||
368 type == PTR_TO_PACKET_META;
371 static bool type_is_sk_pointer(enum bpf_reg_type type)
373 return type == PTR_TO_SOCKET ||
374 type == PTR_TO_SOCK_COMMON ||
375 type == PTR_TO_TCP_SOCK ||
376 type == PTR_TO_XDP_SOCK;
379 static bool reg_type_may_be_null(enum bpf_reg_type type)
381 return type == PTR_TO_MAP_VALUE_OR_NULL ||
382 type == PTR_TO_SOCKET_OR_NULL ||
383 type == PTR_TO_SOCK_COMMON_OR_NULL ||
384 type == PTR_TO_TCP_SOCK_OR_NULL;
387 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
389 return reg->type == PTR_TO_MAP_VALUE &&
390 map_value_has_spin_lock(reg->map_ptr);
393 static bool reg_type_may_be_refcounted_or_null(enum bpf_reg_type type)
395 return type == PTR_TO_SOCKET ||
396 type == PTR_TO_SOCKET_OR_NULL ||
397 type == PTR_TO_TCP_SOCK ||
398 type == PTR_TO_TCP_SOCK_OR_NULL;
401 static bool arg_type_may_be_refcounted(enum bpf_arg_type type)
403 return type == ARG_PTR_TO_SOCK_COMMON;
406 /* Determine whether the function releases some resources allocated by another
407 * function call. The first reference type argument will be assumed to be
408 * released by release_reference().
410 static bool is_release_function(enum bpf_func_id func_id)
412 return func_id == BPF_FUNC_sk_release;
415 static bool is_acquire_function(enum bpf_func_id func_id)
417 return func_id == BPF_FUNC_sk_lookup_tcp ||
418 func_id == BPF_FUNC_sk_lookup_udp ||
419 func_id == BPF_FUNC_skc_lookup_tcp;
422 static bool is_ptr_cast_function(enum bpf_func_id func_id)
424 return func_id == BPF_FUNC_tcp_sock ||
425 func_id == BPF_FUNC_sk_fullsock;
428 /* string representation of 'enum bpf_reg_type' */
429 static const char * const reg_type_str[] = {
431 [SCALAR_VALUE] = "inv",
432 [PTR_TO_CTX] = "ctx",
433 [CONST_PTR_TO_MAP] = "map_ptr",
434 [PTR_TO_MAP_VALUE] = "map_value",
435 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
436 [PTR_TO_STACK] = "fp",
437 [PTR_TO_PACKET] = "pkt",
438 [PTR_TO_PACKET_META] = "pkt_meta",
439 [PTR_TO_PACKET_END] = "pkt_end",
440 [PTR_TO_FLOW_KEYS] = "flow_keys",
441 [PTR_TO_SOCKET] = "sock",
442 [PTR_TO_SOCKET_OR_NULL] = "sock_or_null",
443 [PTR_TO_SOCK_COMMON] = "sock_common",
444 [PTR_TO_SOCK_COMMON_OR_NULL] = "sock_common_or_null",
445 [PTR_TO_TCP_SOCK] = "tcp_sock",
446 [PTR_TO_TCP_SOCK_OR_NULL] = "tcp_sock_or_null",
447 [PTR_TO_TP_BUFFER] = "tp_buffer",
448 [PTR_TO_XDP_SOCK] = "xdp_sock",
449 [PTR_TO_BTF_ID] = "ptr_",
452 static char slot_type_char[] = {
453 [STACK_INVALID] = '?',
459 static void print_liveness(struct bpf_verifier_env *env,
460 enum bpf_reg_liveness live)
462 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
464 if (live & REG_LIVE_READ)
466 if (live & REG_LIVE_WRITTEN)
468 if (live & REG_LIVE_DONE)
472 static struct bpf_func_state *func(struct bpf_verifier_env *env,
473 const struct bpf_reg_state *reg)
475 struct bpf_verifier_state *cur = env->cur_state;
477 return cur->frame[reg->frameno];
480 const char *kernel_type_name(u32 id)
482 return btf_name_by_offset(btf_vmlinux,
483 btf_type_by_id(btf_vmlinux, id)->name_off);
486 static void print_verifier_state(struct bpf_verifier_env *env,
487 const struct bpf_func_state *state)
489 const struct bpf_reg_state *reg;
494 verbose(env, " frame%d:", state->frameno);
495 for (i = 0; i < MAX_BPF_REG; i++) {
496 reg = &state->regs[i];
500 verbose(env, " R%d", i);
501 print_liveness(env, reg->live);
502 verbose(env, "=%s", reg_type_str[t]);
503 if (t == SCALAR_VALUE && reg->precise)
505 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
506 tnum_is_const(reg->var_off)) {
507 /* reg->off should be 0 for SCALAR_VALUE */
508 verbose(env, "%lld", reg->var_off.value + reg->off);
510 if (t == PTR_TO_BTF_ID)
511 verbose(env, "%s", kernel_type_name(reg->btf_id));
512 verbose(env, "(id=%d", reg->id);
513 if (reg_type_may_be_refcounted_or_null(t))
514 verbose(env, ",ref_obj_id=%d", reg->ref_obj_id);
515 if (t != SCALAR_VALUE)
516 verbose(env, ",off=%d", reg->off);
517 if (type_is_pkt_pointer(t))
518 verbose(env, ",r=%d", reg->range);
519 else if (t == CONST_PTR_TO_MAP ||
520 t == PTR_TO_MAP_VALUE ||
521 t == PTR_TO_MAP_VALUE_OR_NULL)
522 verbose(env, ",ks=%d,vs=%d",
523 reg->map_ptr->key_size,
524 reg->map_ptr->value_size);
525 if (tnum_is_const(reg->var_off)) {
526 /* Typically an immediate SCALAR_VALUE, but
527 * could be a pointer whose offset is too big
530 verbose(env, ",imm=%llx", reg->var_off.value);
532 if (reg->smin_value != reg->umin_value &&
533 reg->smin_value != S64_MIN)
534 verbose(env, ",smin_value=%lld",
535 (long long)reg->smin_value);
536 if (reg->smax_value != reg->umax_value &&
537 reg->smax_value != S64_MAX)
538 verbose(env, ",smax_value=%lld",
539 (long long)reg->smax_value);
540 if (reg->umin_value != 0)
541 verbose(env, ",umin_value=%llu",
542 (unsigned long long)reg->umin_value);
543 if (reg->umax_value != U64_MAX)
544 verbose(env, ",umax_value=%llu",
545 (unsigned long long)reg->umax_value);
546 if (!tnum_is_unknown(reg->var_off)) {
549 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
550 verbose(env, ",var_off=%s", tn_buf);
556 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
557 char types_buf[BPF_REG_SIZE + 1];
561 for (j = 0; j < BPF_REG_SIZE; j++) {
562 if (state->stack[i].slot_type[j] != STACK_INVALID)
564 types_buf[j] = slot_type_char[
565 state->stack[i].slot_type[j]];
567 types_buf[BPF_REG_SIZE] = 0;
570 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
571 print_liveness(env, state->stack[i].spilled_ptr.live);
572 if (state->stack[i].slot_type[0] == STACK_SPILL) {
573 reg = &state->stack[i].spilled_ptr;
575 verbose(env, "=%s", reg_type_str[t]);
576 if (t == SCALAR_VALUE && reg->precise)
578 if (t == SCALAR_VALUE && tnum_is_const(reg->var_off))
579 verbose(env, "%lld", reg->var_off.value + reg->off);
581 verbose(env, "=%s", types_buf);
584 if (state->acquired_refs && state->refs[0].id) {
585 verbose(env, " refs=%d", state->refs[0].id);
586 for (i = 1; i < state->acquired_refs; i++)
587 if (state->refs[i].id)
588 verbose(env, ",%d", state->refs[i].id);
593 #define COPY_STATE_FN(NAME, COUNT, FIELD, SIZE) \
594 static int copy_##NAME##_state(struct bpf_func_state *dst, \
595 const struct bpf_func_state *src) \
599 if (WARN_ON_ONCE(dst->COUNT < src->COUNT)) { \
600 /* internal bug, make state invalid to reject the program */ \
601 memset(dst, 0, sizeof(*dst)); \
604 memcpy(dst->FIELD, src->FIELD, \
605 sizeof(*src->FIELD) * (src->COUNT / SIZE)); \
608 /* copy_reference_state() */
609 COPY_STATE_FN(reference, acquired_refs, refs, 1)
610 /* copy_stack_state() */
611 COPY_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
614 #define REALLOC_STATE_FN(NAME, COUNT, FIELD, SIZE) \
615 static int realloc_##NAME##_state(struct bpf_func_state *state, int size, \
618 u32 old_size = state->COUNT; \
619 struct bpf_##NAME##_state *new_##FIELD; \
620 int slot = size / SIZE; \
622 if (size <= old_size || !size) { \
625 state->COUNT = slot * SIZE; \
626 if (!size && old_size) { \
627 kfree(state->FIELD); \
628 state->FIELD = NULL; \
632 new_##FIELD = kmalloc_array(slot, sizeof(struct bpf_##NAME##_state), \
638 memcpy(new_##FIELD, state->FIELD, \
639 sizeof(*new_##FIELD) * (old_size / SIZE)); \
640 memset(new_##FIELD + old_size / SIZE, 0, \
641 sizeof(*new_##FIELD) * (size - old_size) / SIZE); \
643 state->COUNT = slot * SIZE; \
644 kfree(state->FIELD); \
645 state->FIELD = new_##FIELD; \
648 /* realloc_reference_state() */
649 REALLOC_STATE_FN(reference, acquired_refs, refs, 1)
650 /* realloc_stack_state() */
651 REALLOC_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
652 #undef REALLOC_STATE_FN
654 /* do_check() starts with zero-sized stack in struct bpf_verifier_state to
655 * make it consume minimal amount of memory. check_stack_write() access from
656 * the program calls into realloc_func_state() to grow the stack size.
657 * Note there is a non-zero 'parent' pointer inside bpf_verifier_state
658 * which realloc_stack_state() copies over. It points to previous
659 * bpf_verifier_state which is never reallocated.
661 static int realloc_func_state(struct bpf_func_state *state, int stack_size,
662 int refs_size, bool copy_old)
664 int err = realloc_reference_state(state, refs_size, copy_old);
667 return realloc_stack_state(state, stack_size, copy_old);
670 /* Acquire a pointer id from the env and update the state->refs to include
671 * this new pointer reference.
672 * On success, returns a valid pointer id to associate with the register
673 * On failure, returns a negative errno.
675 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
677 struct bpf_func_state *state = cur_func(env);
678 int new_ofs = state->acquired_refs;
681 err = realloc_reference_state(state, state->acquired_refs + 1, true);
685 state->refs[new_ofs].id = id;
686 state->refs[new_ofs].insn_idx = insn_idx;
691 /* release function corresponding to acquire_reference_state(). Idempotent. */
692 static int release_reference_state(struct bpf_func_state *state, int ptr_id)
696 last_idx = state->acquired_refs - 1;
697 for (i = 0; i < state->acquired_refs; i++) {
698 if (state->refs[i].id == ptr_id) {
699 if (last_idx && i != last_idx)
700 memcpy(&state->refs[i], &state->refs[last_idx],
701 sizeof(*state->refs));
702 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
703 state->acquired_refs--;
710 static int transfer_reference_state(struct bpf_func_state *dst,
711 struct bpf_func_state *src)
713 int err = realloc_reference_state(dst, src->acquired_refs, false);
716 err = copy_reference_state(dst, src);
722 static void free_func_state(struct bpf_func_state *state)
731 static void clear_jmp_history(struct bpf_verifier_state *state)
733 kfree(state->jmp_history);
734 state->jmp_history = NULL;
735 state->jmp_history_cnt = 0;
738 static void free_verifier_state(struct bpf_verifier_state *state,
743 for (i = 0; i <= state->curframe; i++) {
744 free_func_state(state->frame[i]);
745 state->frame[i] = NULL;
747 clear_jmp_history(state);
752 /* copy verifier state from src to dst growing dst stack space
753 * when necessary to accommodate larger src stack
755 static int copy_func_state(struct bpf_func_state *dst,
756 const struct bpf_func_state *src)
760 err = realloc_func_state(dst, src->allocated_stack, src->acquired_refs,
764 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
765 err = copy_reference_state(dst, src);
768 return copy_stack_state(dst, src);
771 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
772 const struct bpf_verifier_state *src)
774 struct bpf_func_state *dst;
775 u32 jmp_sz = sizeof(struct bpf_idx_pair) * src->jmp_history_cnt;
778 if (dst_state->jmp_history_cnt < src->jmp_history_cnt) {
779 kfree(dst_state->jmp_history);
780 dst_state->jmp_history = kmalloc(jmp_sz, GFP_USER);
781 if (!dst_state->jmp_history)
784 memcpy(dst_state->jmp_history, src->jmp_history, jmp_sz);
785 dst_state->jmp_history_cnt = src->jmp_history_cnt;
787 /* if dst has more stack frames then src frame, free them */
788 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
789 free_func_state(dst_state->frame[i]);
790 dst_state->frame[i] = NULL;
792 dst_state->speculative = src->speculative;
793 dst_state->curframe = src->curframe;
794 dst_state->active_spin_lock = src->active_spin_lock;
795 dst_state->branches = src->branches;
796 dst_state->parent = src->parent;
797 dst_state->first_insn_idx = src->first_insn_idx;
798 dst_state->last_insn_idx = src->last_insn_idx;
799 for (i = 0; i <= src->curframe; i++) {
800 dst = dst_state->frame[i];
802 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
805 dst_state->frame[i] = dst;
807 err = copy_func_state(dst, src->frame[i]);
814 static void update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
817 u32 br = --st->branches;
819 /* WARN_ON(br > 1) technically makes sense here,
820 * but see comment in push_stack(), hence:
822 WARN_ONCE((int)br < 0,
823 "BUG update_branch_counts:branches_to_explore=%d\n",
831 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
834 struct bpf_verifier_state *cur = env->cur_state;
835 struct bpf_verifier_stack_elem *elem, *head = env->head;
838 if (env->head == NULL)
842 err = copy_verifier_state(cur, &head->st);
847 *insn_idx = head->insn_idx;
849 *prev_insn_idx = head->prev_insn_idx;
851 free_verifier_state(&head->st, false);
858 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
859 int insn_idx, int prev_insn_idx,
862 struct bpf_verifier_state *cur = env->cur_state;
863 struct bpf_verifier_stack_elem *elem;
866 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
870 elem->insn_idx = insn_idx;
871 elem->prev_insn_idx = prev_insn_idx;
872 elem->next = env->head;
875 err = copy_verifier_state(&elem->st, cur);
878 elem->st.speculative |= speculative;
879 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
880 verbose(env, "The sequence of %d jumps is too complex.\n",
884 if (elem->st.parent) {
885 ++elem->st.parent->branches;
886 /* WARN_ON(branches > 2) technically makes sense here,
888 * 1. speculative states will bump 'branches' for non-branch
890 * 2. is_state_visited() heuristics may decide not to create
891 * a new state for a sequence of branches and all such current
892 * and cloned states will be pointing to a single parent state
893 * which might have large 'branches' count.
898 free_verifier_state(env->cur_state, true);
899 env->cur_state = NULL;
900 /* pop all elements and return */
901 while (!pop_stack(env, NULL, NULL));
905 #define CALLER_SAVED_REGS 6
906 static const int caller_saved[CALLER_SAVED_REGS] = {
907 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
910 static void __mark_reg_not_init(struct bpf_reg_state *reg);
912 /* Mark the unknown part of a register (variable offset or scalar value) as
913 * known to have the value @imm.
915 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
917 /* Clear id, off, and union(map_ptr, range) */
918 memset(((u8 *)reg) + sizeof(reg->type), 0,
919 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
920 reg->var_off = tnum_const(imm);
921 reg->smin_value = (s64)imm;
922 reg->smax_value = (s64)imm;
923 reg->umin_value = imm;
924 reg->umax_value = imm;
927 /* Mark the 'variable offset' part of a register as zero. This should be
928 * used only on registers holding a pointer type.
930 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
932 __mark_reg_known(reg, 0);
935 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
937 __mark_reg_known(reg, 0);
938 reg->type = SCALAR_VALUE;
941 static void mark_reg_known_zero(struct bpf_verifier_env *env,
942 struct bpf_reg_state *regs, u32 regno)
944 if (WARN_ON(regno >= MAX_BPF_REG)) {
945 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
946 /* Something bad happened, let's kill all regs */
947 for (regno = 0; regno < MAX_BPF_REG; regno++)
948 __mark_reg_not_init(regs + regno);
951 __mark_reg_known_zero(regs + regno);
954 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
956 return type_is_pkt_pointer(reg->type);
959 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
961 return reg_is_pkt_pointer(reg) ||
962 reg->type == PTR_TO_PACKET_END;
965 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
966 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
967 enum bpf_reg_type which)
969 /* The register can already have a range from prior markings.
970 * This is fine as long as it hasn't been advanced from its
973 return reg->type == which &&
976 tnum_equals_const(reg->var_off, 0);
979 /* Attempts to improve min/max values based on var_off information */
980 static void __update_reg_bounds(struct bpf_reg_state *reg)
982 /* min signed is max(sign bit) | min(other bits) */
983 reg->smin_value = max_t(s64, reg->smin_value,
984 reg->var_off.value | (reg->var_off.mask & S64_MIN));
985 /* max signed is min(sign bit) | max(other bits) */
986 reg->smax_value = min_t(s64, reg->smax_value,
987 reg->var_off.value | (reg->var_off.mask & S64_MAX));
988 reg->umin_value = max(reg->umin_value, reg->var_off.value);
989 reg->umax_value = min(reg->umax_value,
990 reg->var_off.value | reg->var_off.mask);
993 /* Uses signed min/max values to inform unsigned, and vice-versa */
994 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
996 /* Learn sign from signed bounds.
997 * If we cannot cross the sign boundary, then signed and unsigned bounds
998 * are the same, so combine. This works even in the negative case, e.g.
999 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1001 if (reg->smin_value >= 0 || reg->smax_value < 0) {
1002 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1004 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1008 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1009 * boundary, so we must be careful.
1011 if ((s64)reg->umax_value >= 0) {
1012 /* Positive. We can't learn anything from the smin, but smax
1013 * is positive, hence safe.
1015 reg->smin_value = reg->umin_value;
1016 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1018 } else if ((s64)reg->umin_value < 0) {
1019 /* Negative. We can't learn anything from the smax, but smin
1020 * is negative, hence safe.
1022 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1024 reg->smax_value = reg->umax_value;
1028 /* Attempts to improve var_off based on unsigned min/max information */
1029 static void __reg_bound_offset(struct bpf_reg_state *reg)
1031 reg->var_off = tnum_intersect(reg->var_off,
1032 tnum_range(reg->umin_value,
1036 static void __reg_bound_offset32(struct bpf_reg_state *reg)
1038 u64 mask = 0xffffFFFF;
1039 struct tnum range = tnum_range(reg->umin_value & mask,
1040 reg->umax_value & mask);
1041 struct tnum lo32 = tnum_cast(reg->var_off, 4);
1042 struct tnum hi32 = tnum_lshift(tnum_rshift(reg->var_off, 32), 32);
1044 reg->var_off = tnum_or(hi32, tnum_intersect(lo32, range));
1047 /* Reset the min/max bounds of a register */
1048 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
1050 reg->smin_value = S64_MIN;
1051 reg->smax_value = S64_MAX;
1052 reg->umin_value = 0;
1053 reg->umax_value = U64_MAX;
1056 /* Mark a register as having a completely unknown (scalar) value. */
1057 static void __mark_reg_unknown(struct bpf_reg_state *reg)
1060 * Clear type, id, off, and union(map_ptr, range) and
1061 * padding between 'type' and union
1063 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
1064 reg->type = SCALAR_VALUE;
1065 reg->var_off = tnum_unknown;
1067 __mark_reg_unbounded(reg);
1070 static void mark_reg_unknown(struct bpf_verifier_env *env,
1071 struct bpf_reg_state *regs, u32 regno)
1073 if (WARN_ON(regno >= MAX_BPF_REG)) {
1074 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
1075 /* Something bad happened, let's kill all regs except FP */
1076 for (regno = 0; regno < BPF_REG_FP; regno++)
1077 __mark_reg_not_init(regs + regno);
1081 __mark_reg_unknown(regs);
1082 /* constant backtracking is enabled for root without bpf2bpf calls */
1083 regs->precise = env->subprog_cnt > 1 || !env->allow_ptr_leaks ?
1087 static void __mark_reg_not_init(struct bpf_reg_state *reg)
1089 __mark_reg_unknown(reg);
1090 reg->type = NOT_INIT;
1093 static void mark_reg_not_init(struct bpf_verifier_env *env,
1094 struct bpf_reg_state *regs, u32 regno)
1096 if (WARN_ON(regno >= MAX_BPF_REG)) {
1097 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
1098 /* Something bad happened, let's kill all regs except FP */
1099 for (regno = 0; regno < BPF_REG_FP; regno++)
1100 __mark_reg_not_init(regs + regno);
1103 __mark_reg_not_init(regs + regno);
1106 #define DEF_NOT_SUBREG (0)
1107 static void init_reg_state(struct bpf_verifier_env *env,
1108 struct bpf_func_state *state)
1110 struct bpf_reg_state *regs = state->regs;
1113 for (i = 0; i < MAX_BPF_REG; i++) {
1114 mark_reg_not_init(env, regs, i);
1115 regs[i].live = REG_LIVE_NONE;
1116 regs[i].parent = NULL;
1117 regs[i].subreg_def = DEF_NOT_SUBREG;
1121 regs[BPF_REG_FP].type = PTR_TO_STACK;
1122 mark_reg_known_zero(env, regs, BPF_REG_FP);
1123 regs[BPF_REG_FP].frameno = state->frameno;
1125 /* 1st arg to a function */
1126 regs[BPF_REG_1].type = PTR_TO_CTX;
1127 mark_reg_known_zero(env, regs, BPF_REG_1);
1130 #define BPF_MAIN_FUNC (-1)
1131 static void init_func_state(struct bpf_verifier_env *env,
1132 struct bpf_func_state *state,
1133 int callsite, int frameno, int subprogno)
1135 state->callsite = callsite;
1136 state->frameno = frameno;
1137 state->subprogno = subprogno;
1138 init_reg_state(env, state);
1142 SRC_OP, /* register is used as source operand */
1143 DST_OP, /* register is used as destination operand */
1144 DST_OP_NO_MARK /* same as above, check only, don't mark */
1147 static int cmp_subprogs(const void *a, const void *b)
1149 return ((struct bpf_subprog_info *)a)->start -
1150 ((struct bpf_subprog_info *)b)->start;
1153 static int find_subprog(struct bpf_verifier_env *env, int off)
1155 struct bpf_subprog_info *p;
1157 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
1158 sizeof(env->subprog_info[0]), cmp_subprogs);
1161 return p - env->subprog_info;
1165 static int add_subprog(struct bpf_verifier_env *env, int off)
1167 int insn_cnt = env->prog->len;
1170 if (off >= insn_cnt || off < 0) {
1171 verbose(env, "call to invalid destination\n");
1174 ret = find_subprog(env, off);
1177 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
1178 verbose(env, "too many subprograms\n");
1181 env->subprog_info[env->subprog_cnt++].start = off;
1182 sort(env->subprog_info, env->subprog_cnt,
1183 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
1187 static int check_subprogs(struct bpf_verifier_env *env)
1189 int i, ret, subprog_start, subprog_end, off, cur_subprog = 0;
1190 struct bpf_subprog_info *subprog = env->subprog_info;
1191 struct bpf_insn *insn = env->prog->insnsi;
1192 int insn_cnt = env->prog->len;
1194 /* Add entry function. */
1195 ret = add_subprog(env, 0);
1199 /* determine subprog starts. The end is one before the next starts */
1200 for (i = 0; i < insn_cnt; i++) {
1201 if (insn[i].code != (BPF_JMP | BPF_CALL))
1203 if (insn[i].src_reg != BPF_PSEUDO_CALL)
1205 if (!env->allow_ptr_leaks) {
1206 verbose(env, "function calls to other bpf functions are allowed for root only\n");
1209 ret = add_subprog(env, i + insn[i].imm + 1);
1214 /* Add a fake 'exit' subprog which could simplify subprog iteration
1215 * logic. 'subprog_cnt' should not be increased.
1217 subprog[env->subprog_cnt].start = insn_cnt;
1219 if (env->log.level & BPF_LOG_LEVEL2)
1220 for (i = 0; i < env->subprog_cnt; i++)
1221 verbose(env, "func#%d @%d\n", i, subprog[i].start);
1223 /* now check that all jumps are within the same subprog */
1224 subprog_start = subprog[cur_subprog].start;
1225 subprog_end = subprog[cur_subprog + 1].start;
1226 for (i = 0; i < insn_cnt; i++) {
1227 u8 code = insn[i].code;
1229 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
1231 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
1233 off = i + insn[i].off + 1;
1234 if (off < subprog_start || off >= subprog_end) {
1235 verbose(env, "jump out of range from insn %d to %d\n", i, off);
1239 if (i == subprog_end - 1) {
1240 /* to avoid fall-through from one subprog into another
1241 * the last insn of the subprog should be either exit
1242 * or unconditional jump back
1244 if (code != (BPF_JMP | BPF_EXIT) &&
1245 code != (BPF_JMP | BPF_JA)) {
1246 verbose(env, "last insn is not an exit or jmp\n");
1249 subprog_start = subprog_end;
1251 if (cur_subprog < env->subprog_cnt)
1252 subprog_end = subprog[cur_subprog + 1].start;
1258 /* Parentage chain of this register (or stack slot) should take care of all
1259 * issues like callee-saved registers, stack slot allocation time, etc.
1261 static int mark_reg_read(struct bpf_verifier_env *env,
1262 const struct bpf_reg_state *state,
1263 struct bpf_reg_state *parent, u8 flag)
1265 bool writes = parent == state->parent; /* Observe write marks */
1269 /* if read wasn't screened by an earlier write ... */
1270 if (writes && state->live & REG_LIVE_WRITTEN)
1272 if (parent->live & REG_LIVE_DONE) {
1273 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
1274 reg_type_str[parent->type],
1275 parent->var_off.value, parent->off);
1278 /* The first condition is more likely to be true than the
1279 * second, checked it first.
1281 if ((parent->live & REG_LIVE_READ) == flag ||
1282 parent->live & REG_LIVE_READ64)
1283 /* The parentage chain never changes and
1284 * this parent was already marked as LIVE_READ.
1285 * There is no need to keep walking the chain again and
1286 * keep re-marking all parents as LIVE_READ.
1287 * This case happens when the same register is read
1288 * multiple times without writes into it in-between.
1289 * Also, if parent has the stronger REG_LIVE_READ64 set,
1290 * then no need to set the weak REG_LIVE_READ32.
1293 /* ... then we depend on parent's value */
1294 parent->live |= flag;
1295 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
1296 if (flag == REG_LIVE_READ64)
1297 parent->live &= ~REG_LIVE_READ32;
1299 parent = state->parent;
1304 if (env->longest_mark_read_walk < cnt)
1305 env->longest_mark_read_walk = cnt;
1309 /* This function is supposed to be used by the following 32-bit optimization
1310 * code only. It returns TRUE if the source or destination register operates
1311 * on 64-bit, otherwise return FALSE.
1313 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
1314 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
1319 class = BPF_CLASS(code);
1321 if (class == BPF_JMP) {
1322 /* BPF_EXIT for "main" will reach here. Return TRUE
1327 if (op == BPF_CALL) {
1328 /* BPF to BPF call will reach here because of marking
1329 * caller saved clobber with DST_OP_NO_MARK for which we
1330 * don't care the register def because they are anyway
1331 * marked as NOT_INIT already.
1333 if (insn->src_reg == BPF_PSEUDO_CALL)
1335 /* Helper call will reach here because of arg type
1336 * check, conservatively return TRUE.
1345 if (class == BPF_ALU64 || class == BPF_JMP ||
1346 /* BPF_END always use BPF_ALU class. */
1347 (class == BPF_ALU && op == BPF_END && insn->imm == 64))
1350 if (class == BPF_ALU || class == BPF_JMP32)
1353 if (class == BPF_LDX) {
1355 return BPF_SIZE(code) == BPF_DW;
1356 /* LDX source must be ptr. */
1360 if (class == BPF_STX) {
1361 if (reg->type != SCALAR_VALUE)
1363 return BPF_SIZE(code) == BPF_DW;
1366 if (class == BPF_LD) {
1367 u8 mode = BPF_MODE(code);
1370 if (mode == BPF_IMM)
1373 /* Both LD_IND and LD_ABS return 32-bit data. */
1377 /* Implicit ctx ptr. */
1378 if (regno == BPF_REG_6)
1381 /* Explicit source could be any width. */
1385 if (class == BPF_ST)
1386 /* The only source register for BPF_ST is a ptr. */
1389 /* Conservatively return true at default. */
1393 /* Return TRUE if INSN doesn't have explicit value define. */
1394 static bool insn_no_def(struct bpf_insn *insn)
1396 u8 class = BPF_CLASS(insn->code);
1398 return (class == BPF_JMP || class == BPF_JMP32 ||
1399 class == BPF_STX || class == BPF_ST);
1402 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
1403 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
1405 if (insn_no_def(insn))
1408 return !is_reg64(env, insn, insn->dst_reg, NULL, DST_OP);
1411 static void mark_insn_zext(struct bpf_verifier_env *env,
1412 struct bpf_reg_state *reg)
1414 s32 def_idx = reg->subreg_def;
1416 if (def_idx == DEF_NOT_SUBREG)
1419 env->insn_aux_data[def_idx - 1].zext_dst = true;
1420 /* The dst will be zero extended, so won't be sub-register anymore. */
1421 reg->subreg_def = DEF_NOT_SUBREG;
1424 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
1425 enum reg_arg_type t)
1427 struct bpf_verifier_state *vstate = env->cur_state;
1428 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1429 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
1430 struct bpf_reg_state *reg, *regs = state->regs;
1433 if (regno >= MAX_BPF_REG) {
1434 verbose(env, "R%d is invalid\n", regno);
1439 rw64 = is_reg64(env, insn, regno, reg, t);
1441 /* check whether register used as source operand can be read */
1442 if (reg->type == NOT_INIT) {
1443 verbose(env, "R%d !read_ok\n", regno);
1446 /* We don't need to worry about FP liveness because it's read-only */
1447 if (regno == BPF_REG_FP)
1451 mark_insn_zext(env, reg);
1453 return mark_reg_read(env, reg, reg->parent,
1454 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
1456 /* check whether register used as dest operand can be written to */
1457 if (regno == BPF_REG_FP) {
1458 verbose(env, "frame pointer is read only\n");
1461 reg->live |= REG_LIVE_WRITTEN;
1462 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
1464 mark_reg_unknown(env, regs, regno);
1469 /* for any branch, call, exit record the history of jmps in the given state */
1470 static int push_jmp_history(struct bpf_verifier_env *env,
1471 struct bpf_verifier_state *cur)
1473 u32 cnt = cur->jmp_history_cnt;
1474 struct bpf_idx_pair *p;
1477 p = krealloc(cur->jmp_history, cnt * sizeof(*p), GFP_USER);
1480 p[cnt - 1].idx = env->insn_idx;
1481 p[cnt - 1].prev_idx = env->prev_insn_idx;
1482 cur->jmp_history = p;
1483 cur->jmp_history_cnt = cnt;
1487 /* Backtrack one insn at a time. If idx is not at the top of recorded
1488 * history then previous instruction came from straight line execution.
1490 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
1495 if (cnt && st->jmp_history[cnt - 1].idx == i) {
1496 i = st->jmp_history[cnt - 1].prev_idx;
1504 /* For given verifier state backtrack_insn() is called from the last insn to
1505 * the first insn. Its purpose is to compute a bitmask of registers and
1506 * stack slots that needs precision in the parent verifier state.
1508 static int backtrack_insn(struct bpf_verifier_env *env, int idx,
1509 u32 *reg_mask, u64 *stack_mask)
1511 const struct bpf_insn_cbs cbs = {
1512 .cb_print = verbose,
1513 .private_data = env,
1515 struct bpf_insn *insn = env->prog->insnsi + idx;
1516 u8 class = BPF_CLASS(insn->code);
1517 u8 opcode = BPF_OP(insn->code);
1518 u8 mode = BPF_MODE(insn->code);
1519 u32 dreg = 1u << insn->dst_reg;
1520 u32 sreg = 1u << insn->src_reg;
1523 if (insn->code == 0)
1525 if (env->log.level & BPF_LOG_LEVEL) {
1526 verbose(env, "regs=%x stack=%llx before ", *reg_mask, *stack_mask);
1527 verbose(env, "%d: ", idx);
1528 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
1531 if (class == BPF_ALU || class == BPF_ALU64) {
1532 if (!(*reg_mask & dreg))
1534 if (opcode == BPF_MOV) {
1535 if (BPF_SRC(insn->code) == BPF_X) {
1537 * dreg needs precision after this insn
1538 * sreg needs precision before this insn
1544 * dreg needs precision after this insn.
1545 * Corresponding register is already marked
1546 * as precise=true in this verifier state.
1547 * No further markings in parent are necessary
1552 if (BPF_SRC(insn->code) == BPF_X) {
1554 * both dreg and sreg need precision
1559 * dreg still needs precision before this insn
1562 } else if (class == BPF_LDX) {
1563 if (!(*reg_mask & dreg))
1567 /* scalars can only be spilled into stack w/o losing precision.
1568 * Load from any other memory can be zero extended.
1569 * The desire to keep that precision is already indicated
1570 * by 'precise' mark in corresponding register of this state.
1571 * No further tracking necessary.
1573 if (insn->src_reg != BPF_REG_FP)
1575 if (BPF_SIZE(insn->code) != BPF_DW)
1578 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
1579 * that [fp - off] slot contains scalar that needs to be
1580 * tracked with precision
1582 spi = (-insn->off - 1) / BPF_REG_SIZE;
1584 verbose(env, "BUG spi %d\n", spi);
1585 WARN_ONCE(1, "verifier backtracking bug");
1588 *stack_mask |= 1ull << spi;
1589 } else if (class == BPF_STX || class == BPF_ST) {
1590 if (*reg_mask & dreg)
1591 /* stx & st shouldn't be using _scalar_ dst_reg
1592 * to access memory. It means backtracking
1593 * encountered a case of pointer subtraction.
1596 /* scalars can only be spilled into stack */
1597 if (insn->dst_reg != BPF_REG_FP)
1599 if (BPF_SIZE(insn->code) != BPF_DW)
1601 spi = (-insn->off - 1) / BPF_REG_SIZE;
1603 verbose(env, "BUG spi %d\n", spi);
1604 WARN_ONCE(1, "verifier backtracking bug");
1607 if (!(*stack_mask & (1ull << spi)))
1609 *stack_mask &= ~(1ull << spi);
1610 if (class == BPF_STX)
1612 } else if (class == BPF_JMP || class == BPF_JMP32) {
1613 if (opcode == BPF_CALL) {
1614 if (insn->src_reg == BPF_PSEUDO_CALL)
1616 /* regular helper call sets R0 */
1618 if (*reg_mask & 0x3f) {
1619 /* if backtracing was looking for registers R1-R5
1620 * they should have been found already.
1622 verbose(env, "BUG regs %x\n", *reg_mask);
1623 WARN_ONCE(1, "verifier backtracking bug");
1626 } else if (opcode == BPF_EXIT) {
1629 } else if (class == BPF_LD) {
1630 if (!(*reg_mask & dreg))
1633 /* It's ld_imm64 or ld_abs or ld_ind.
1634 * For ld_imm64 no further tracking of precision
1635 * into parent is necessary
1637 if (mode == BPF_IND || mode == BPF_ABS)
1638 /* to be analyzed */
1644 /* the scalar precision tracking algorithm:
1645 * . at the start all registers have precise=false.
1646 * . scalar ranges are tracked as normal through alu and jmp insns.
1647 * . once precise value of the scalar register is used in:
1648 * . ptr + scalar alu
1649 * . if (scalar cond K|scalar)
1650 * . helper_call(.., scalar, ...) where ARG_CONST is expected
1651 * backtrack through the verifier states and mark all registers and
1652 * stack slots with spilled constants that these scalar regisers
1653 * should be precise.
1654 * . during state pruning two registers (or spilled stack slots)
1655 * are equivalent if both are not precise.
1657 * Note the verifier cannot simply walk register parentage chain,
1658 * since many different registers and stack slots could have been
1659 * used to compute single precise scalar.
1661 * The approach of starting with precise=true for all registers and then
1662 * backtrack to mark a register as not precise when the verifier detects
1663 * that program doesn't care about specific value (e.g., when helper
1664 * takes register as ARG_ANYTHING parameter) is not safe.
1666 * It's ok to walk single parentage chain of the verifier states.
1667 * It's possible that this backtracking will go all the way till 1st insn.
1668 * All other branches will be explored for needing precision later.
1670 * The backtracking needs to deal with cases like:
1671 * 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)
1674 * if r5 > 0x79f goto pc+7
1675 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
1678 * call bpf_perf_event_output#25
1679 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
1683 * call foo // uses callee's r6 inside to compute r0
1687 * to track above reg_mask/stack_mask needs to be independent for each frame.
1689 * Also if parent's curframe > frame where backtracking started,
1690 * the verifier need to mark registers in both frames, otherwise callees
1691 * may incorrectly prune callers. This is similar to
1692 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
1694 * For now backtracking falls back into conservative marking.
1696 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
1697 struct bpf_verifier_state *st)
1699 struct bpf_func_state *func;
1700 struct bpf_reg_state *reg;
1703 /* big hammer: mark all scalars precise in this path.
1704 * pop_stack may still get !precise scalars.
1706 for (; st; st = st->parent)
1707 for (i = 0; i <= st->curframe; i++) {
1708 func = st->frame[i];
1709 for (j = 0; j < BPF_REG_FP; j++) {
1710 reg = &func->regs[j];
1711 if (reg->type != SCALAR_VALUE)
1713 reg->precise = true;
1715 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
1716 if (func->stack[j].slot_type[0] != STACK_SPILL)
1718 reg = &func->stack[j].spilled_ptr;
1719 if (reg->type != SCALAR_VALUE)
1721 reg->precise = true;
1726 static int __mark_chain_precision(struct bpf_verifier_env *env, int regno,
1729 struct bpf_verifier_state *st = env->cur_state;
1730 int first_idx = st->first_insn_idx;
1731 int last_idx = env->insn_idx;
1732 struct bpf_func_state *func;
1733 struct bpf_reg_state *reg;
1734 u32 reg_mask = regno >= 0 ? 1u << regno : 0;
1735 u64 stack_mask = spi >= 0 ? 1ull << spi : 0;
1736 bool skip_first = true;
1737 bool new_marks = false;
1740 if (!env->allow_ptr_leaks)
1741 /* backtracking is root only for now */
1744 func = st->frame[st->curframe];
1746 reg = &func->regs[regno];
1747 if (reg->type != SCALAR_VALUE) {
1748 WARN_ONCE(1, "backtracing misuse");
1755 reg->precise = true;
1759 if (func->stack[spi].slot_type[0] != STACK_SPILL) {
1763 reg = &func->stack[spi].spilled_ptr;
1764 if (reg->type != SCALAR_VALUE) {
1772 reg->precise = true;
1778 if (!reg_mask && !stack_mask)
1781 DECLARE_BITMAP(mask, 64);
1782 u32 history = st->jmp_history_cnt;
1784 if (env->log.level & BPF_LOG_LEVEL)
1785 verbose(env, "last_idx %d first_idx %d\n", last_idx, first_idx);
1786 for (i = last_idx;;) {
1791 err = backtrack_insn(env, i, ®_mask, &stack_mask);
1793 if (err == -ENOTSUPP) {
1794 mark_all_scalars_precise(env, st);
1799 if (!reg_mask && !stack_mask)
1800 /* Found assignment(s) into tracked register in this state.
1801 * Since this state is already marked, just return.
1802 * Nothing to be tracked further in the parent state.
1807 i = get_prev_insn_idx(st, i, &history);
1808 if (i >= env->prog->len) {
1809 /* This can happen if backtracking reached insn 0
1810 * and there are still reg_mask or stack_mask
1812 * It means the backtracking missed the spot where
1813 * particular register was initialized with a constant.
1815 verbose(env, "BUG backtracking idx %d\n", i);
1816 WARN_ONCE(1, "verifier backtracking bug");
1825 func = st->frame[st->curframe];
1826 bitmap_from_u64(mask, reg_mask);
1827 for_each_set_bit(i, mask, 32) {
1828 reg = &func->regs[i];
1829 if (reg->type != SCALAR_VALUE) {
1830 reg_mask &= ~(1u << i);
1835 reg->precise = true;
1838 bitmap_from_u64(mask, stack_mask);
1839 for_each_set_bit(i, mask, 64) {
1840 if (i >= func->allocated_stack / BPF_REG_SIZE) {
1841 /* the sequence of instructions:
1843 * 3: (7b) *(u64 *)(r3 -8) = r0
1844 * 4: (79) r4 = *(u64 *)(r10 -8)
1845 * doesn't contain jmps. It's backtracked
1846 * as a single block.
1847 * During backtracking insn 3 is not recognized as
1848 * stack access, so at the end of backtracking
1849 * stack slot fp-8 is still marked in stack_mask.
1850 * However the parent state may not have accessed
1851 * fp-8 and it's "unallocated" stack space.
1852 * In such case fallback to conservative.
1854 mark_all_scalars_precise(env, st);
1858 if (func->stack[i].slot_type[0] != STACK_SPILL) {
1859 stack_mask &= ~(1ull << i);
1862 reg = &func->stack[i].spilled_ptr;
1863 if (reg->type != SCALAR_VALUE) {
1864 stack_mask &= ~(1ull << i);
1869 reg->precise = true;
1871 if (env->log.level & BPF_LOG_LEVEL) {
1872 print_verifier_state(env, func);
1873 verbose(env, "parent %s regs=%x stack=%llx marks\n",
1874 new_marks ? "didn't have" : "already had",
1875 reg_mask, stack_mask);
1878 if (!reg_mask && !stack_mask)
1883 last_idx = st->last_insn_idx;
1884 first_idx = st->first_insn_idx;
1889 static int mark_chain_precision(struct bpf_verifier_env *env, int regno)
1891 return __mark_chain_precision(env, regno, -1);
1894 static int mark_chain_precision_stack(struct bpf_verifier_env *env, int spi)
1896 return __mark_chain_precision(env, -1, spi);
1899 static bool is_spillable_regtype(enum bpf_reg_type type)
1902 case PTR_TO_MAP_VALUE:
1903 case PTR_TO_MAP_VALUE_OR_NULL:
1907 case PTR_TO_PACKET_META:
1908 case PTR_TO_PACKET_END:
1909 case PTR_TO_FLOW_KEYS:
1910 case CONST_PTR_TO_MAP:
1912 case PTR_TO_SOCKET_OR_NULL:
1913 case PTR_TO_SOCK_COMMON:
1914 case PTR_TO_SOCK_COMMON_OR_NULL:
1915 case PTR_TO_TCP_SOCK:
1916 case PTR_TO_TCP_SOCK_OR_NULL:
1917 case PTR_TO_XDP_SOCK:
1924 /* Does this register contain a constant zero? */
1925 static bool register_is_null(struct bpf_reg_state *reg)
1927 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
1930 static bool register_is_const(struct bpf_reg_state *reg)
1932 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
1935 static void save_register_state(struct bpf_func_state *state,
1936 int spi, struct bpf_reg_state *reg)
1940 state->stack[spi].spilled_ptr = *reg;
1941 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1943 for (i = 0; i < BPF_REG_SIZE; i++)
1944 state->stack[spi].slot_type[i] = STACK_SPILL;
1947 /* check_stack_read/write functions track spill/fill of registers,
1948 * stack boundary and alignment are checked in check_mem_access()
1950 static int check_stack_write(struct bpf_verifier_env *env,
1951 struct bpf_func_state *state, /* func where register points to */
1952 int off, int size, int value_regno, int insn_idx)
1954 struct bpf_func_state *cur; /* state of the current function */
1955 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
1956 u32 dst_reg = env->prog->insnsi[insn_idx].dst_reg;
1957 struct bpf_reg_state *reg = NULL;
1959 err = realloc_func_state(state, round_up(slot + 1, BPF_REG_SIZE),
1960 state->acquired_refs, true);
1963 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
1964 * so it's aligned access and [off, off + size) are within stack limits
1966 if (!env->allow_ptr_leaks &&
1967 state->stack[spi].slot_type[0] == STACK_SPILL &&
1968 size != BPF_REG_SIZE) {
1969 verbose(env, "attempt to corrupt spilled pointer on stack\n");
1973 cur = env->cur_state->frame[env->cur_state->curframe];
1974 if (value_regno >= 0)
1975 reg = &cur->regs[value_regno];
1977 if (reg && size == BPF_REG_SIZE && register_is_const(reg) &&
1978 !register_is_null(reg) && env->allow_ptr_leaks) {
1979 if (dst_reg != BPF_REG_FP) {
1980 /* The backtracking logic can only recognize explicit
1981 * stack slot address like [fp - 8]. Other spill of
1982 * scalar via different register has to be conervative.
1983 * Backtrack from here and mark all registers as precise
1984 * that contributed into 'reg' being a constant.
1986 err = mark_chain_precision(env, value_regno);
1990 save_register_state(state, spi, reg);
1991 } else if (reg && is_spillable_regtype(reg->type)) {
1992 /* register containing pointer is being spilled into stack */
1993 if (size != BPF_REG_SIZE) {
1994 verbose_linfo(env, insn_idx, "; ");
1995 verbose(env, "invalid size of register spill\n");
1999 if (state != cur && reg->type == PTR_TO_STACK) {
2000 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
2004 if (!env->allow_ptr_leaks) {
2005 bool sanitize = false;
2007 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
2008 register_is_const(&state->stack[spi].spilled_ptr))
2010 for (i = 0; i < BPF_REG_SIZE; i++)
2011 if (state->stack[spi].slot_type[i] == STACK_MISC) {
2016 int *poff = &env->insn_aux_data[insn_idx].sanitize_stack_off;
2017 int soff = (-spi - 1) * BPF_REG_SIZE;
2019 /* detected reuse of integer stack slot with a pointer
2020 * which means either llvm is reusing stack slot or
2021 * an attacker is trying to exploit CVE-2018-3639
2022 * (speculative store bypass)
2023 * Have to sanitize that slot with preemptive
2026 if (*poff && *poff != soff) {
2027 /* disallow programs where single insn stores
2028 * into two different stack slots, since verifier
2029 * cannot sanitize them
2032 "insn %d cannot access two stack slots fp%d and fp%d",
2033 insn_idx, *poff, soff);
2039 save_register_state(state, spi, reg);
2041 u8 type = STACK_MISC;
2043 /* regular write of data into stack destroys any spilled ptr */
2044 state->stack[spi].spilled_ptr.type = NOT_INIT;
2045 /* Mark slots as STACK_MISC if they belonged to spilled ptr. */
2046 if (state->stack[spi].slot_type[0] == STACK_SPILL)
2047 for (i = 0; i < BPF_REG_SIZE; i++)
2048 state->stack[spi].slot_type[i] = STACK_MISC;
2050 /* only mark the slot as written if all 8 bytes were written
2051 * otherwise read propagation may incorrectly stop too soon
2052 * when stack slots are partially written.
2053 * This heuristic means that read propagation will be
2054 * conservative, since it will add reg_live_read marks
2055 * to stack slots all the way to first state when programs
2056 * writes+reads less than 8 bytes
2058 if (size == BPF_REG_SIZE)
2059 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2061 /* when we zero initialize stack slots mark them as such */
2062 if (reg && register_is_null(reg)) {
2063 /* backtracking doesn't work for STACK_ZERO yet. */
2064 err = mark_chain_precision(env, value_regno);
2070 /* Mark slots affected by this stack write. */
2071 for (i = 0; i < size; i++)
2072 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
2078 static int check_stack_read(struct bpf_verifier_env *env,
2079 struct bpf_func_state *reg_state /* func where register points to */,
2080 int off, int size, int value_regno)
2082 struct bpf_verifier_state *vstate = env->cur_state;
2083 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2084 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
2085 struct bpf_reg_state *reg;
2088 if (reg_state->allocated_stack <= slot) {
2089 verbose(env, "invalid read from stack off %d+0 size %d\n",
2093 stype = reg_state->stack[spi].slot_type;
2094 reg = ®_state->stack[spi].spilled_ptr;
2096 if (stype[0] == STACK_SPILL) {
2097 if (size != BPF_REG_SIZE) {
2098 if (reg->type != SCALAR_VALUE) {
2099 verbose_linfo(env, env->insn_idx, "; ");
2100 verbose(env, "invalid size of register fill\n");
2103 if (value_regno >= 0) {
2104 mark_reg_unknown(env, state->regs, value_regno);
2105 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
2107 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2110 for (i = 1; i < BPF_REG_SIZE; i++) {
2111 if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) {
2112 verbose(env, "corrupted spill memory\n");
2117 if (value_regno >= 0) {
2118 /* restore register state from stack */
2119 state->regs[value_regno] = *reg;
2120 /* mark reg as written since spilled pointer state likely
2121 * has its liveness marks cleared by is_state_visited()
2122 * which resets stack/reg liveness for state transitions
2124 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
2126 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2130 for (i = 0; i < size; i++) {
2131 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_MISC)
2133 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_ZERO) {
2137 verbose(env, "invalid read from stack off %d+%d size %d\n",
2141 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2142 if (value_regno >= 0) {
2143 if (zeros == size) {
2144 /* any size read into register is zero extended,
2145 * so the whole register == const_zero
2147 __mark_reg_const_zero(&state->regs[value_regno]);
2148 /* backtracking doesn't support STACK_ZERO yet,
2149 * so mark it precise here, so that later
2150 * backtracking can stop here.
2151 * Backtracking may not need this if this register
2152 * doesn't participate in pointer adjustment.
2153 * Forward propagation of precise flag is not
2154 * necessary either. This mark is only to stop
2155 * backtracking. Any register that contributed
2156 * to const 0 was marked precise before spill.
2158 state->regs[value_regno].precise = true;
2160 /* have read misc data from the stack */
2161 mark_reg_unknown(env, state->regs, value_regno);
2163 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
2169 static int check_stack_access(struct bpf_verifier_env *env,
2170 const struct bpf_reg_state *reg,
2173 /* Stack accesses must be at a fixed offset, so that we
2174 * can determine what type of data were returned. See
2175 * check_stack_read().
2177 if (!tnum_is_const(reg->var_off)) {
2180 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2181 verbose(env, "variable stack access var_off=%s off=%d size=%d\n",
2186 if (off >= 0 || off < -MAX_BPF_STACK) {
2187 verbose(env, "invalid stack off=%d size=%d\n", off, size);
2194 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
2195 int off, int size, enum bpf_access_type type)
2197 struct bpf_reg_state *regs = cur_regs(env);
2198 struct bpf_map *map = regs[regno].map_ptr;
2199 u32 cap = bpf_map_flags_to_cap(map);
2201 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
2202 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
2203 map->value_size, off, size);
2207 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
2208 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
2209 map->value_size, off, size);
2216 /* check read/write into map element returned by bpf_map_lookup_elem() */
2217 static int __check_map_access(struct bpf_verifier_env *env, u32 regno, int off,
2218 int size, bool zero_size_allowed)
2220 struct bpf_reg_state *regs = cur_regs(env);
2221 struct bpf_map *map = regs[regno].map_ptr;
2223 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
2224 off + size > map->value_size) {
2225 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
2226 map->value_size, off, size);
2232 /* check read/write into a map element with possible variable offset */
2233 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
2234 int off, int size, bool zero_size_allowed)
2236 struct bpf_verifier_state *vstate = env->cur_state;
2237 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2238 struct bpf_reg_state *reg = &state->regs[regno];
2241 /* We may have adjusted the register to this map value, so we
2242 * need to try adding each of min_value and max_value to off
2243 * to make sure our theoretical access will be safe.
2245 if (env->log.level & BPF_LOG_LEVEL)
2246 print_verifier_state(env, state);
2248 /* The minimum value is only important with signed
2249 * comparisons where we can't assume the floor of a
2250 * value is 0. If we are using signed variables for our
2251 * index'es we need to make sure that whatever we use
2252 * will have a set floor within our range.
2254 if (reg->smin_value < 0 &&
2255 (reg->smin_value == S64_MIN ||
2256 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
2257 reg->smin_value + off < 0)) {
2258 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2262 err = __check_map_access(env, regno, reg->smin_value + off, size,
2265 verbose(env, "R%d min value is outside of the array range\n",
2270 /* If we haven't set a max value then we need to bail since we can't be
2271 * sure we won't do bad things.
2272 * If reg->umax_value + off could overflow, treat that as unbounded too.
2274 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
2275 verbose(env, "R%d unbounded memory access, make sure to bounds check any array access into a map\n",
2279 err = __check_map_access(env, regno, reg->umax_value + off, size,
2282 verbose(env, "R%d max value is outside of the array range\n",
2285 if (map_value_has_spin_lock(reg->map_ptr)) {
2286 u32 lock = reg->map_ptr->spin_lock_off;
2288 /* if any part of struct bpf_spin_lock can be touched by
2289 * load/store reject this program.
2290 * To check that [x1, x2) overlaps with [y1, y2)
2291 * it is sufficient to check x1 < y2 && y1 < x2.
2293 if (reg->smin_value + off < lock + sizeof(struct bpf_spin_lock) &&
2294 lock < reg->umax_value + off + size) {
2295 verbose(env, "bpf_spin_lock cannot be accessed directly by load/store\n");
2302 #define MAX_PACKET_OFF 0xffff
2304 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
2305 const struct bpf_call_arg_meta *meta,
2306 enum bpf_access_type t)
2308 switch (env->prog->type) {
2309 /* Program types only with direct read access go here! */
2310 case BPF_PROG_TYPE_LWT_IN:
2311 case BPF_PROG_TYPE_LWT_OUT:
2312 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
2313 case BPF_PROG_TYPE_SK_REUSEPORT:
2314 case BPF_PROG_TYPE_FLOW_DISSECTOR:
2315 case BPF_PROG_TYPE_CGROUP_SKB:
2320 /* Program types with direct read + write access go here! */
2321 case BPF_PROG_TYPE_SCHED_CLS:
2322 case BPF_PROG_TYPE_SCHED_ACT:
2323 case BPF_PROG_TYPE_XDP:
2324 case BPF_PROG_TYPE_LWT_XMIT:
2325 case BPF_PROG_TYPE_SK_SKB:
2326 case BPF_PROG_TYPE_SK_MSG:
2328 return meta->pkt_access;
2330 env->seen_direct_write = true;
2333 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
2335 env->seen_direct_write = true;
2344 static int __check_packet_access(struct bpf_verifier_env *env, u32 regno,
2345 int off, int size, bool zero_size_allowed)
2347 struct bpf_reg_state *regs = cur_regs(env);
2348 struct bpf_reg_state *reg = ®s[regno];
2350 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
2351 (u64)off + size > reg->range) {
2352 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
2353 off, size, regno, reg->id, reg->off, reg->range);
2359 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
2360 int size, bool zero_size_allowed)
2362 struct bpf_reg_state *regs = cur_regs(env);
2363 struct bpf_reg_state *reg = ®s[regno];
2366 /* We may have added a variable offset to the packet pointer; but any
2367 * reg->range we have comes after that. We are only checking the fixed
2371 /* We don't allow negative numbers, because we aren't tracking enough
2372 * detail to prove they're safe.
2374 if (reg->smin_value < 0) {
2375 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2379 err = __check_packet_access(env, regno, off, size, zero_size_allowed);
2381 verbose(env, "R%d offset is outside of the packet\n", regno);
2385 /* __check_packet_access has made sure "off + size - 1" is within u16.
2386 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
2387 * otherwise find_good_pkt_pointers would have refused to set range info
2388 * that __check_packet_access would have rejected this pkt access.
2389 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
2391 env->prog->aux->max_pkt_offset =
2392 max_t(u32, env->prog->aux->max_pkt_offset,
2393 off + reg->umax_value + size - 1);
2398 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
2399 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
2400 enum bpf_access_type t, enum bpf_reg_type *reg_type,
2403 struct bpf_insn_access_aux info = {
2404 .reg_type = *reg_type,
2408 if (env->ops->is_valid_access &&
2409 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
2410 /* A non zero info.ctx_field_size indicates that this field is a
2411 * candidate for later verifier transformation to load the whole
2412 * field and then apply a mask when accessed with a narrower
2413 * access than actual ctx access size. A zero info.ctx_field_size
2414 * will only allow for whole field access and rejects any other
2415 * type of narrower access.
2417 *reg_type = info.reg_type;
2419 if (*reg_type == PTR_TO_BTF_ID)
2420 *btf_id = info.btf_id;
2422 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
2423 /* remember the offset of last byte accessed in ctx */
2424 if (env->prog->aux->max_ctx_offset < off + size)
2425 env->prog->aux->max_ctx_offset = off + size;
2429 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
2433 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
2436 if (size < 0 || off < 0 ||
2437 (u64)off + size > sizeof(struct bpf_flow_keys)) {
2438 verbose(env, "invalid access to flow keys off=%d size=%d\n",
2445 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
2446 u32 regno, int off, int size,
2447 enum bpf_access_type t)
2449 struct bpf_reg_state *regs = cur_regs(env);
2450 struct bpf_reg_state *reg = ®s[regno];
2451 struct bpf_insn_access_aux info = {};
2454 if (reg->smin_value < 0) {
2455 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2460 switch (reg->type) {
2461 case PTR_TO_SOCK_COMMON:
2462 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
2465 valid = bpf_sock_is_valid_access(off, size, t, &info);
2467 case PTR_TO_TCP_SOCK:
2468 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
2470 case PTR_TO_XDP_SOCK:
2471 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
2479 env->insn_aux_data[insn_idx].ctx_field_size =
2480 info.ctx_field_size;
2484 verbose(env, "R%d invalid %s access off=%d size=%d\n",
2485 regno, reg_type_str[reg->type], off, size);
2490 static bool __is_pointer_value(bool allow_ptr_leaks,
2491 const struct bpf_reg_state *reg)
2493 if (allow_ptr_leaks)
2496 return reg->type != SCALAR_VALUE;
2499 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
2501 return cur_regs(env) + regno;
2504 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
2506 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
2509 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
2511 const struct bpf_reg_state *reg = reg_state(env, regno);
2513 return reg->type == PTR_TO_CTX;
2516 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
2518 const struct bpf_reg_state *reg = reg_state(env, regno);
2520 return type_is_sk_pointer(reg->type);
2523 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
2525 const struct bpf_reg_state *reg = reg_state(env, regno);
2527 return type_is_pkt_pointer(reg->type);
2530 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
2532 const struct bpf_reg_state *reg = reg_state(env, regno);
2534 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
2535 return reg->type == PTR_TO_FLOW_KEYS;
2538 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
2539 const struct bpf_reg_state *reg,
2540 int off, int size, bool strict)
2542 struct tnum reg_off;
2545 /* Byte size accesses are always allowed. */
2546 if (!strict || size == 1)
2549 /* For platforms that do not have a Kconfig enabling
2550 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
2551 * NET_IP_ALIGN is universally set to '2'. And on platforms
2552 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
2553 * to this code only in strict mode where we want to emulate
2554 * the NET_IP_ALIGN==2 checking. Therefore use an
2555 * unconditional IP align value of '2'.
2559 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
2560 if (!tnum_is_aligned(reg_off, size)) {
2563 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2565 "misaligned packet access off %d+%s+%d+%d size %d\n",
2566 ip_align, tn_buf, reg->off, off, size);
2573 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
2574 const struct bpf_reg_state *reg,
2575 const char *pointer_desc,
2576 int off, int size, bool strict)
2578 struct tnum reg_off;
2580 /* Byte size accesses are always allowed. */
2581 if (!strict || size == 1)
2584 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
2585 if (!tnum_is_aligned(reg_off, size)) {
2588 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2589 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
2590 pointer_desc, tn_buf, reg->off, off, size);
2597 static int check_ptr_alignment(struct bpf_verifier_env *env,
2598 const struct bpf_reg_state *reg, int off,
2599 int size, bool strict_alignment_once)
2601 bool strict = env->strict_alignment || strict_alignment_once;
2602 const char *pointer_desc = "";
2604 switch (reg->type) {
2606 case PTR_TO_PACKET_META:
2607 /* Special case, because of NET_IP_ALIGN. Given metadata sits
2608 * right in front, treat it the very same way.
2610 return check_pkt_ptr_alignment(env, reg, off, size, strict);
2611 case PTR_TO_FLOW_KEYS:
2612 pointer_desc = "flow keys ";
2614 case PTR_TO_MAP_VALUE:
2615 pointer_desc = "value ";
2618 pointer_desc = "context ";
2621 pointer_desc = "stack ";
2622 /* The stack spill tracking logic in check_stack_write()
2623 * and check_stack_read() relies on stack accesses being
2629 pointer_desc = "sock ";
2631 case PTR_TO_SOCK_COMMON:
2632 pointer_desc = "sock_common ";
2634 case PTR_TO_TCP_SOCK:
2635 pointer_desc = "tcp_sock ";
2637 case PTR_TO_XDP_SOCK:
2638 pointer_desc = "xdp_sock ";
2643 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
2647 static int update_stack_depth(struct bpf_verifier_env *env,
2648 const struct bpf_func_state *func,
2651 u16 stack = env->subprog_info[func->subprogno].stack_depth;
2656 /* update known max for given subprogram */
2657 env->subprog_info[func->subprogno].stack_depth = -off;
2661 /* starting from main bpf function walk all instructions of the function
2662 * and recursively walk all callees that given function can call.
2663 * Ignore jump and exit insns.
2664 * Since recursion is prevented by check_cfg() this algorithm
2665 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
2667 static int check_max_stack_depth(struct bpf_verifier_env *env)
2669 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
2670 struct bpf_subprog_info *subprog = env->subprog_info;
2671 struct bpf_insn *insn = env->prog->insnsi;
2672 int ret_insn[MAX_CALL_FRAMES];
2673 int ret_prog[MAX_CALL_FRAMES];
2676 /* round up to 32-bytes, since this is granularity
2677 * of interpreter stack size
2679 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
2680 if (depth > MAX_BPF_STACK) {
2681 verbose(env, "combined stack size of %d calls is %d. Too large\n",
2686 subprog_end = subprog[idx + 1].start;
2687 for (; i < subprog_end; i++) {
2688 if (insn[i].code != (BPF_JMP | BPF_CALL))
2690 if (insn[i].src_reg != BPF_PSEUDO_CALL)
2692 /* remember insn and function to return to */
2693 ret_insn[frame] = i + 1;
2694 ret_prog[frame] = idx;
2696 /* find the callee */
2697 i = i + insn[i].imm + 1;
2698 idx = find_subprog(env, i);
2700 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
2705 if (frame >= MAX_CALL_FRAMES) {
2706 verbose(env, "the call stack of %d frames is too deep !\n",
2712 /* end of for() loop means the last insn of the 'subprog'
2713 * was reached. Doesn't matter whether it was JA or EXIT
2717 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
2719 i = ret_insn[frame];
2720 idx = ret_prog[frame];
2724 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
2725 static int get_callee_stack_depth(struct bpf_verifier_env *env,
2726 const struct bpf_insn *insn, int idx)
2728 int start = idx + insn->imm + 1, subprog;
2730 subprog = find_subprog(env, start);
2732 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
2736 return env->subprog_info[subprog].stack_depth;
2740 static int check_ctx_reg(struct bpf_verifier_env *env,
2741 const struct bpf_reg_state *reg, int regno)
2743 /* Access to ctx or passing it to a helper is only allowed in
2744 * its original, unmodified form.
2748 verbose(env, "dereference of modified ctx ptr R%d off=%d disallowed\n",
2753 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
2756 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2757 verbose(env, "variable ctx access var_off=%s disallowed\n", tn_buf);
2764 static int check_tp_buffer_access(struct bpf_verifier_env *env,
2765 const struct bpf_reg_state *reg,
2766 int regno, int off, int size)
2770 "R%d invalid tracepoint buffer access: off=%d, size=%d",
2774 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
2777 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2779 "R%d invalid variable buffer offset: off=%d, var_off=%s",
2780 regno, off, tn_buf);
2783 if (off + size > env->prog->aux->max_tp_access)
2784 env->prog->aux->max_tp_access = off + size;
2790 /* truncate register to smaller size (in bytes)
2791 * must be called with size < BPF_REG_SIZE
2793 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
2797 /* clear high bits in bit representation */
2798 reg->var_off = tnum_cast(reg->var_off, size);
2800 /* fix arithmetic bounds */
2801 mask = ((u64)1 << (size * 8)) - 1;
2802 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
2803 reg->umin_value &= mask;
2804 reg->umax_value &= mask;
2806 reg->umin_value = 0;
2807 reg->umax_value = mask;
2809 reg->smin_value = reg->umin_value;
2810 reg->smax_value = reg->umax_value;
2813 static bool bpf_map_is_rdonly(const struct bpf_map *map)
2815 return (map->map_flags & BPF_F_RDONLY_PROG) && map->frozen;
2818 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val)
2824 err = map->ops->map_direct_value_addr(map, &addr, off);
2827 ptr = (void *)(long)addr + off;
2831 *val = (u64)*(u8 *)ptr;
2834 *val = (u64)*(u16 *)ptr;
2837 *val = (u64)*(u32 *)ptr;
2848 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
2849 struct bpf_reg_state *regs,
2850 int regno, int off, int size,
2851 enum bpf_access_type atype,
2854 struct bpf_reg_state *reg = regs + regno;
2855 const struct btf_type *t = btf_type_by_id(btf_vmlinux, reg->btf_id);
2856 const char *tname = btf_name_by_offset(btf_vmlinux, t->name_off);
2860 if (atype != BPF_READ) {
2861 verbose(env, "only read is supported\n");
2867 "R%d is ptr_%s invalid negative access: off=%d\n",
2871 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
2874 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2876 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
2877 regno, tname, off, tn_buf);
2881 ret = btf_struct_access(&env->log, t, off, size, atype, &btf_id);
2885 if (ret == SCALAR_VALUE) {
2886 mark_reg_unknown(env, regs, value_regno);
2889 mark_reg_known_zero(env, regs, value_regno);
2890 regs[value_regno].type = PTR_TO_BTF_ID;
2891 regs[value_regno].btf_id = btf_id;
2895 /* check whether memory at (regno + off) is accessible for t = (read | write)
2896 * if t==write, value_regno is a register which value is stored into memory
2897 * if t==read, value_regno is a register which will receive the value from memory
2898 * if t==write && value_regno==-1, some unknown value is stored into memory
2899 * if t==read && value_regno==-1, don't care what we read from memory
2901 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
2902 int off, int bpf_size, enum bpf_access_type t,
2903 int value_regno, bool strict_alignment_once)
2905 struct bpf_reg_state *regs = cur_regs(env);
2906 struct bpf_reg_state *reg = regs + regno;
2907 struct bpf_func_state *state;
2910 size = bpf_size_to_bytes(bpf_size);
2914 /* alignment checks will add in reg->off themselves */
2915 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
2919 /* for access checks, reg->off is just part of off */
2922 if (reg->type == PTR_TO_MAP_VALUE) {
2923 if (t == BPF_WRITE && value_regno >= 0 &&
2924 is_pointer_value(env, value_regno)) {
2925 verbose(env, "R%d leaks addr into map\n", value_regno);
2928 err = check_map_access_type(env, regno, off, size, t);
2931 err = check_map_access(env, regno, off, size, false);
2932 if (!err && t == BPF_READ && value_regno >= 0) {
2933 struct bpf_map *map = reg->map_ptr;
2935 /* if map is read-only, track its contents as scalars */
2936 if (tnum_is_const(reg->var_off) &&
2937 bpf_map_is_rdonly(map) &&
2938 map->ops->map_direct_value_addr) {
2939 int map_off = off + reg->var_off.value;
2942 err = bpf_map_direct_read(map, map_off, size,
2947 regs[value_regno].type = SCALAR_VALUE;
2948 __mark_reg_known(®s[value_regno], val);
2950 mark_reg_unknown(env, regs, value_regno);
2953 } else if (reg->type == PTR_TO_CTX) {
2954 enum bpf_reg_type reg_type = SCALAR_VALUE;
2957 if (t == BPF_WRITE && value_regno >= 0 &&
2958 is_pointer_value(env, value_regno)) {
2959 verbose(env, "R%d leaks addr into ctx\n", value_regno);
2963 err = check_ctx_reg(env, reg, regno);
2967 err = check_ctx_access(env, insn_idx, off, size, t, ®_type, &btf_id);
2969 verbose_linfo(env, insn_idx, "; ");
2970 if (!err && t == BPF_READ && value_regno >= 0) {
2971 /* ctx access returns either a scalar, or a
2972 * PTR_TO_PACKET[_META,_END]. In the latter
2973 * case, we know the offset is zero.
2975 if (reg_type == SCALAR_VALUE) {
2976 mark_reg_unknown(env, regs, value_regno);
2978 mark_reg_known_zero(env, regs,
2980 if (reg_type_may_be_null(reg_type))
2981 regs[value_regno].id = ++env->id_gen;
2982 /* A load of ctx field could have different
2983 * actual load size with the one encoded in the
2984 * insn. When the dst is PTR, it is for sure not
2987 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
2988 if (reg_type == PTR_TO_BTF_ID)
2989 regs[value_regno].btf_id = btf_id;
2991 regs[value_regno].type = reg_type;
2994 } else if (reg->type == PTR_TO_STACK) {
2995 off += reg->var_off.value;
2996 err = check_stack_access(env, reg, off, size);
3000 state = func(env, reg);
3001 err = update_stack_depth(env, state, off);
3006 err = check_stack_write(env, state, off, size,
3007 value_regno, insn_idx);
3009 err = check_stack_read(env, state, off, size,
3011 } else if (reg_is_pkt_pointer(reg)) {
3012 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
3013 verbose(env, "cannot write into packet\n");
3016 if (t == BPF_WRITE && value_regno >= 0 &&
3017 is_pointer_value(env, value_regno)) {
3018 verbose(env, "R%d leaks addr into packet\n",
3022 err = check_packet_access(env, regno, off, size, false);
3023 if (!err && t == BPF_READ && value_regno >= 0)
3024 mark_reg_unknown(env, regs, value_regno);
3025 } else if (reg->type == PTR_TO_FLOW_KEYS) {
3026 if (t == BPF_WRITE && value_regno >= 0 &&
3027 is_pointer_value(env, value_regno)) {
3028 verbose(env, "R%d leaks addr into flow keys\n",
3033 err = check_flow_keys_access(env, off, size);
3034 if (!err && t == BPF_READ && value_regno >= 0)
3035 mark_reg_unknown(env, regs, value_regno);
3036 } else if (type_is_sk_pointer(reg->type)) {
3037 if (t == BPF_WRITE) {
3038 verbose(env, "R%d cannot write into %s\n",
3039 regno, reg_type_str[reg->type]);
3042 err = check_sock_access(env, insn_idx, regno, off, size, t);
3043 if (!err && value_regno >= 0)
3044 mark_reg_unknown(env, regs, value_regno);
3045 } else if (reg->type == PTR_TO_TP_BUFFER) {
3046 err = check_tp_buffer_access(env, reg, regno, off, size);
3047 if (!err && t == BPF_READ && value_regno >= 0)
3048 mark_reg_unknown(env, regs, value_regno);
3049 } else if (reg->type == PTR_TO_BTF_ID) {
3050 err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
3053 verbose(env, "R%d invalid mem access '%s'\n", regno,
3054 reg_type_str[reg->type]);
3058 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
3059 regs[value_regno].type == SCALAR_VALUE) {
3060 /* b/h/w load zero-extends, mark upper bits as known 0 */
3061 coerce_reg_to_size(®s[value_regno], size);
3066 static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
3070 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
3072 verbose(env, "BPF_XADD uses reserved fields\n");
3076 /* check src1 operand */
3077 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3081 /* check src2 operand */
3082 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3086 if (is_pointer_value(env, insn->src_reg)) {
3087 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
3091 if (is_ctx_reg(env, insn->dst_reg) ||
3092 is_pkt_reg(env, insn->dst_reg) ||
3093 is_flow_key_reg(env, insn->dst_reg) ||
3094 is_sk_reg(env, insn->dst_reg)) {
3095 verbose(env, "BPF_XADD stores into R%d %s is not allowed\n",
3097 reg_type_str[reg_state(env, insn->dst_reg)->type]);
3101 /* check whether atomic_add can read the memory */
3102 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3103 BPF_SIZE(insn->code), BPF_READ, -1, true);
3107 /* check whether atomic_add can write into the same memory */
3108 return check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3109 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
3112 static int __check_stack_boundary(struct bpf_verifier_env *env, u32 regno,
3113 int off, int access_size,
3114 bool zero_size_allowed)
3116 struct bpf_reg_state *reg = reg_state(env, regno);
3118 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
3119 access_size < 0 || (access_size == 0 && !zero_size_allowed)) {
3120 if (tnum_is_const(reg->var_off)) {
3121 verbose(env, "invalid stack type R%d off=%d access_size=%d\n",
3122 regno, off, access_size);
3126 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3127 verbose(env, "invalid stack type R%d var_off=%s access_size=%d\n",
3128 regno, tn_buf, access_size);
3135 /* when register 'regno' is passed into function that will read 'access_size'
3136 * bytes from that pointer, make sure that it's within stack boundary
3137 * and all elements of stack are initialized.
3138 * Unlike most pointer bounds-checking functions, this one doesn't take an
3139 * 'off' argument, so it has to add in reg->off itself.
3141 static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
3142 int access_size, bool zero_size_allowed,
3143 struct bpf_call_arg_meta *meta)
3145 struct bpf_reg_state *reg = reg_state(env, regno);
3146 struct bpf_func_state *state = func(env, reg);
3147 int err, min_off, max_off, i, j, slot, spi;
3149 if (reg->type != PTR_TO_STACK) {
3150 /* Allow zero-byte read from NULL, regardless of pointer type */
3151 if (zero_size_allowed && access_size == 0 &&
3152 register_is_null(reg))
3155 verbose(env, "R%d type=%s expected=%s\n", regno,
3156 reg_type_str[reg->type],
3157 reg_type_str[PTR_TO_STACK]);
3161 if (tnum_is_const(reg->var_off)) {
3162 min_off = max_off = reg->var_off.value + reg->off;
3163 err = __check_stack_boundary(env, regno, min_off, access_size,
3168 /* Variable offset is prohibited for unprivileged mode for
3169 * simplicity since it requires corresponding support in
3170 * Spectre masking for stack ALU.
3171 * See also retrieve_ptr_limit().
3173 if (!env->allow_ptr_leaks) {
3176 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3177 verbose(env, "R%d indirect variable offset stack access prohibited for !root, var_off=%s\n",
3181 /* Only initialized buffer on stack is allowed to be accessed
3182 * with variable offset. With uninitialized buffer it's hard to
3183 * guarantee that whole memory is marked as initialized on
3184 * helper return since specific bounds are unknown what may
3185 * cause uninitialized stack leaking.
3187 if (meta && meta->raw_mode)
3190 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
3191 reg->smax_value <= -BPF_MAX_VAR_OFF) {
3192 verbose(env, "R%d unbounded indirect variable offset stack access\n",
3196 min_off = reg->smin_value + reg->off;
3197 max_off = reg->smax_value + reg->off;
3198 err = __check_stack_boundary(env, regno, min_off, access_size,
3201 verbose(env, "R%d min value is outside of stack bound\n",
3205 err = __check_stack_boundary(env, regno, max_off, access_size,
3208 verbose(env, "R%d max value is outside of stack bound\n",
3214 if (meta && meta->raw_mode) {
3215 meta->access_size = access_size;
3216 meta->regno = regno;
3220 for (i = min_off; i < max_off + access_size; i++) {
3224 spi = slot / BPF_REG_SIZE;
3225 if (state->allocated_stack <= slot)
3227 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
3228 if (*stype == STACK_MISC)
3230 if (*stype == STACK_ZERO) {
3231 /* helper can write anything into the stack */
3232 *stype = STACK_MISC;
3235 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
3236 state->stack[spi].spilled_ptr.type == SCALAR_VALUE) {
3237 __mark_reg_unknown(&state->stack[spi].spilled_ptr);
3238 for (j = 0; j < BPF_REG_SIZE; j++)
3239 state->stack[spi].slot_type[j] = STACK_MISC;
3244 if (tnum_is_const(reg->var_off)) {
3245 verbose(env, "invalid indirect read from stack off %d+%d size %d\n",
3246 min_off, i - min_off, access_size);
3250 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3251 verbose(env, "invalid indirect read from stack var_off %s+%d size %d\n",
3252 tn_buf, i - min_off, access_size);
3256 /* reading any byte out of 8-byte 'spill_slot' will cause
3257 * the whole slot to be marked as 'read'
3259 mark_reg_read(env, &state->stack[spi].spilled_ptr,
3260 state->stack[spi].spilled_ptr.parent,
3263 return update_stack_depth(env, state, min_off);
3266 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
3267 int access_size, bool zero_size_allowed,
3268 struct bpf_call_arg_meta *meta)
3270 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
3272 switch (reg->type) {
3274 case PTR_TO_PACKET_META:
3275 return check_packet_access(env, regno, reg->off, access_size,
3277 case PTR_TO_MAP_VALUE:
3278 if (check_map_access_type(env, regno, reg->off, access_size,
3279 meta && meta->raw_mode ? BPF_WRITE :
3282 return check_map_access(env, regno, reg->off, access_size,
3284 default: /* scalar_value|ptr_to_stack or invalid ptr */
3285 return check_stack_boundary(env, regno, access_size,
3286 zero_size_allowed, meta);
3290 /* Implementation details:
3291 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL
3292 * Two bpf_map_lookups (even with the same key) will have different reg->id.
3293 * For traditional PTR_TO_MAP_VALUE the verifier clears reg->id after
3294 * value_or_null->value transition, since the verifier only cares about
3295 * the range of access to valid map value pointer and doesn't care about actual
3296 * address of the map element.
3297 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
3298 * reg->id > 0 after value_or_null->value transition. By doing so
3299 * two bpf_map_lookups will be considered two different pointers that
3300 * point to different bpf_spin_locks.
3301 * The verifier allows taking only one bpf_spin_lock at a time to avoid
3303 * Since only one bpf_spin_lock is allowed the checks are simpler than
3304 * reg_is_refcounted() logic. The verifier needs to remember only
3305 * one spin_lock instead of array of acquired_refs.
3306 * cur_state->active_spin_lock remembers which map value element got locked
3307 * and clears it after bpf_spin_unlock.
3309 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
3312 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
3313 struct bpf_verifier_state *cur = env->cur_state;
3314 bool is_const = tnum_is_const(reg->var_off);
3315 struct bpf_map *map = reg->map_ptr;
3316 u64 val = reg->var_off.value;
3318 if (reg->type != PTR_TO_MAP_VALUE) {
3319 verbose(env, "R%d is not a pointer to map_value\n", regno);
3324 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
3330 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
3334 if (!map_value_has_spin_lock(map)) {
3335 if (map->spin_lock_off == -E2BIG)
3337 "map '%s' has more than one 'struct bpf_spin_lock'\n",
3339 else if (map->spin_lock_off == -ENOENT)
3341 "map '%s' doesn't have 'struct bpf_spin_lock'\n",
3345 "map '%s' is not a struct type or bpf_spin_lock is mangled\n",
3349 if (map->spin_lock_off != val + reg->off) {
3350 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock'\n",
3355 if (cur->active_spin_lock) {
3357 "Locking two bpf_spin_locks are not allowed\n");
3360 cur->active_spin_lock = reg->id;
3362 if (!cur->active_spin_lock) {
3363 verbose(env, "bpf_spin_unlock without taking a lock\n");
3366 if (cur->active_spin_lock != reg->id) {
3367 verbose(env, "bpf_spin_unlock of different lock\n");
3370 cur->active_spin_lock = 0;
3375 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
3377 return type == ARG_PTR_TO_MEM ||
3378 type == ARG_PTR_TO_MEM_OR_NULL ||
3379 type == ARG_PTR_TO_UNINIT_MEM;
3382 static bool arg_type_is_mem_size(enum bpf_arg_type type)
3384 return type == ARG_CONST_SIZE ||
3385 type == ARG_CONST_SIZE_OR_ZERO;
3388 static bool arg_type_is_int_ptr(enum bpf_arg_type type)
3390 return type == ARG_PTR_TO_INT ||
3391 type == ARG_PTR_TO_LONG;
3394 static int int_ptr_type_to_size(enum bpf_arg_type type)
3396 if (type == ARG_PTR_TO_INT)
3398 else if (type == ARG_PTR_TO_LONG)
3404 static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
3405 enum bpf_arg_type arg_type,
3406 struct bpf_call_arg_meta *meta)
3408 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
3409 enum bpf_reg_type expected_type, type = reg->type;
3412 if (arg_type == ARG_DONTCARE)
3415 err = check_reg_arg(env, regno, SRC_OP);
3419 if (arg_type == ARG_ANYTHING) {
3420 if (is_pointer_value(env, regno)) {
3421 verbose(env, "R%d leaks addr into helper function\n",
3428 if (type_is_pkt_pointer(type) &&
3429 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
3430 verbose(env, "helper access to the packet is not allowed\n");
3434 if (arg_type == ARG_PTR_TO_MAP_KEY ||
3435 arg_type == ARG_PTR_TO_MAP_VALUE ||
3436 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE ||
3437 arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL) {
3438 expected_type = PTR_TO_STACK;
3439 if (register_is_null(reg) &&
3440 arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL)
3441 /* final test in check_stack_boundary() */;
3442 else if (!type_is_pkt_pointer(type) &&
3443 type != PTR_TO_MAP_VALUE &&
3444 type != expected_type)
3446 } else if (arg_type == ARG_CONST_SIZE ||
3447 arg_type == ARG_CONST_SIZE_OR_ZERO) {
3448 expected_type = SCALAR_VALUE;
3449 if (type != expected_type)
3451 } else if (arg_type == ARG_CONST_MAP_PTR) {
3452 expected_type = CONST_PTR_TO_MAP;
3453 if (type != expected_type)
3455 } else if (arg_type == ARG_PTR_TO_CTX) {
3456 expected_type = PTR_TO_CTX;
3457 if (type != expected_type)
3459 err = check_ctx_reg(env, reg, regno);
3462 } else if (arg_type == ARG_PTR_TO_SOCK_COMMON) {
3463 expected_type = PTR_TO_SOCK_COMMON;
3464 /* Any sk pointer can be ARG_PTR_TO_SOCK_COMMON */
3465 if (!type_is_sk_pointer(type))
3467 if (reg->ref_obj_id) {
3468 if (meta->ref_obj_id) {
3469 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
3470 regno, reg->ref_obj_id,
3474 meta->ref_obj_id = reg->ref_obj_id;
3476 } else if (arg_type == ARG_PTR_TO_SOCKET) {
3477 expected_type = PTR_TO_SOCKET;
3478 if (type != expected_type)
3480 } else if (arg_type == ARG_PTR_TO_BTF_ID) {
3481 expected_type = PTR_TO_BTF_ID;
3482 if (type != expected_type)
3484 if (reg->btf_id != meta->btf_id) {
3485 verbose(env, "Helper has type %s got %s in R%d\n",
3486 kernel_type_name(meta->btf_id),
3487 kernel_type_name(reg->btf_id), regno);
3491 if (!tnum_is_const(reg->var_off) || reg->var_off.value || reg->off) {
3492 verbose(env, "R%d is a pointer to in-kernel struct with non-zero offset\n",
3496 } else if (arg_type == ARG_PTR_TO_SPIN_LOCK) {
3497 if (meta->func_id == BPF_FUNC_spin_lock) {
3498 if (process_spin_lock(env, regno, true))
3500 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
3501 if (process_spin_lock(env, regno, false))
3504 verbose(env, "verifier internal error\n");
3507 } else if (arg_type_is_mem_ptr(arg_type)) {
3508 expected_type = PTR_TO_STACK;
3509 /* One exception here. In case function allows for NULL to be
3510 * passed in as argument, it's a SCALAR_VALUE type. Final test
3511 * happens during stack boundary checking.
3513 if (register_is_null(reg) &&
3514 arg_type == ARG_PTR_TO_MEM_OR_NULL)
3515 /* final test in check_stack_boundary() */;
3516 else if (!type_is_pkt_pointer(type) &&
3517 type != PTR_TO_MAP_VALUE &&
3518 type != expected_type)
3520 meta->raw_mode = arg_type == ARG_PTR_TO_UNINIT_MEM;
3521 } else if (arg_type_is_int_ptr(arg_type)) {
3522 expected_type = PTR_TO_STACK;
3523 if (!type_is_pkt_pointer(type) &&
3524 type != PTR_TO_MAP_VALUE &&
3525 type != expected_type)
3528 verbose(env, "unsupported arg_type %d\n", arg_type);
3532 if (arg_type == ARG_CONST_MAP_PTR) {
3533 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
3534 meta->map_ptr = reg->map_ptr;
3535 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
3536 /* bpf_map_xxx(..., map_ptr, ..., key) call:
3537 * check that [key, key + map->key_size) are within
3538 * stack limits and initialized
3540 if (!meta->map_ptr) {
3541 /* in function declaration map_ptr must come before
3542 * map_key, so that it's verified and known before
3543 * we have to check map_key here. Otherwise it means
3544 * that kernel subsystem misconfigured verifier
3546 verbose(env, "invalid map_ptr to access map->key\n");
3549 err = check_helper_mem_access(env, regno,
3550 meta->map_ptr->key_size, false,
3552 } else if (arg_type == ARG_PTR_TO_MAP_VALUE ||
3553 (arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL &&
3554 !register_is_null(reg)) ||
3555 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
3556 /* bpf_map_xxx(..., map_ptr, ..., value) call:
3557 * check [value, value + map->value_size) validity
3559 if (!meta->map_ptr) {
3560 /* kernel subsystem misconfigured verifier */
3561 verbose(env, "invalid map_ptr to access map->value\n");
3564 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE);
3565 err = check_helper_mem_access(env, regno,
3566 meta->map_ptr->value_size, false,
3568 } else if (arg_type_is_mem_size(arg_type)) {
3569 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
3571 /* remember the mem_size which may be used later
3572 * to refine return values.
3574 meta->msize_smax_value = reg->smax_value;
3575 meta->msize_umax_value = reg->umax_value;
3577 /* The register is SCALAR_VALUE; the access check
3578 * happens using its boundaries.
3580 if (!tnum_is_const(reg->var_off))
3581 /* For unprivileged variable accesses, disable raw
3582 * mode so that the program is required to
3583 * initialize all the memory that the helper could
3584 * just partially fill up.
3588 if (reg->smin_value < 0) {
3589 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
3594 if (reg->umin_value == 0) {
3595 err = check_helper_mem_access(env, regno - 1, 0,
3602 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
3603 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
3607 err = check_helper_mem_access(env, regno - 1,
3609 zero_size_allowed, meta);
3611 err = mark_chain_precision(env, regno);
3612 } else if (arg_type_is_int_ptr(arg_type)) {
3613 int size = int_ptr_type_to_size(arg_type);
3615 err = check_helper_mem_access(env, regno, size, false, meta);
3618 err = check_ptr_alignment(env, reg, 0, size, true);
3623 verbose(env, "R%d type=%s expected=%s\n", regno,
3624 reg_type_str[type], reg_type_str[expected_type]);
3628 static int check_map_func_compatibility(struct bpf_verifier_env *env,
3629 struct bpf_map *map, int func_id)
3634 /* We need a two way check, first is from map perspective ... */
3635 switch (map->map_type) {
3636 case BPF_MAP_TYPE_PROG_ARRAY:
3637 if (func_id != BPF_FUNC_tail_call)
3640 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
3641 if (func_id != BPF_FUNC_perf_event_read &&
3642 func_id != BPF_FUNC_perf_event_output &&
3643 func_id != BPF_FUNC_skb_output &&
3644 func_id != BPF_FUNC_perf_event_read_value)
3647 case BPF_MAP_TYPE_STACK_TRACE:
3648 if (func_id != BPF_FUNC_get_stackid)
3651 case BPF_MAP_TYPE_CGROUP_ARRAY:
3652 if (func_id != BPF_FUNC_skb_under_cgroup &&
3653 func_id != BPF_FUNC_current_task_under_cgroup)
3656 case BPF_MAP_TYPE_CGROUP_STORAGE:
3657 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
3658 if (func_id != BPF_FUNC_get_local_storage)
3661 case BPF_MAP_TYPE_DEVMAP:
3662 case BPF_MAP_TYPE_DEVMAP_HASH:
3663 if (func_id != BPF_FUNC_redirect_map &&
3664 func_id != BPF_FUNC_map_lookup_elem)
3667 /* Restrict bpf side of cpumap and xskmap, open when use-cases
3670 case BPF_MAP_TYPE_CPUMAP:
3671 if (func_id != BPF_FUNC_redirect_map)
3674 case BPF_MAP_TYPE_XSKMAP:
3675 if (func_id != BPF_FUNC_redirect_map &&
3676 func_id != BPF_FUNC_map_lookup_elem)
3679 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
3680 case BPF_MAP_TYPE_HASH_OF_MAPS:
3681 if (func_id != BPF_FUNC_map_lookup_elem)
3684 case BPF_MAP_TYPE_SOCKMAP:
3685 if (func_id != BPF_FUNC_sk_redirect_map &&
3686 func_id != BPF_FUNC_sock_map_update &&
3687 func_id != BPF_FUNC_map_delete_elem &&
3688 func_id != BPF_FUNC_msg_redirect_map)
3691 case BPF_MAP_TYPE_SOCKHASH:
3692 if (func_id != BPF_FUNC_sk_redirect_hash &&
3693 func_id != BPF_FUNC_sock_hash_update &&
3694 func_id != BPF_FUNC_map_delete_elem &&
3695 func_id != BPF_FUNC_msg_redirect_hash)
3698 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
3699 if (func_id != BPF_FUNC_sk_select_reuseport)
3702 case BPF_MAP_TYPE_QUEUE:
3703 case BPF_MAP_TYPE_STACK:
3704 if (func_id != BPF_FUNC_map_peek_elem &&
3705 func_id != BPF_FUNC_map_pop_elem &&
3706 func_id != BPF_FUNC_map_push_elem)
3709 case BPF_MAP_TYPE_SK_STORAGE:
3710 if (func_id != BPF_FUNC_sk_storage_get &&
3711 func_id != BPF_FUNC_sk_storage_delete)
3718 /* ... and second from the function itself. */
3720 case BPF_FUNC_tail_call:
3721 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
3723 if (env->subprog_cnt > 1) {
3724 verbose(env, "tail_calls are not allowed in programs with bpf-to-bpf calls\n");
3728 case BPF_FUNC_perf_event_read:
3729 case BPF_FUNC_perf_event_output:
3730 case BPF_FUNC_perf_event_read_value:
3731 case BPF_FUNC_skb_output:
3732 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
3735 case BPF_FUNC_get_stackid:
3736 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
3739 case BPF_FUNC_current_task_under_cgroup:
3740 case BPF_FUNC_skb_under_cgroup:
3741 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
3744 case BPF_FUNC_redirect_map:
3745 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
3746 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
3747 map->map_type != BPF_MAP_TYPE_CPUMAP &&
3748 map->map_type != BPF_MAP_TYPE_XSKMAP)
3751 case BPF_FUNC_sk_redirect_map:
3752 case BPF_FUNC_msg_redirect_map:
3753 case BPF_FUNC_sock_map_update:
3754 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
3757 case BPF_FUNC_sk_redirect_hash:
3758 case BPF_FUNC_msg_redirect_hash:
3759 case BPF_FUNC_sock_hash_update:
3760 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
3763 case BPF_FUNC_get_local_storage:
3764 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
3765 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
3768 case BPF_FUNC_sk_select_reuseport:
3769 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY)
3772 case BPF_FUNC_map_peek_elem:
3773 case BPF_FUNC_map_pop_elem:
3774 case BPF_FUNC_map_push_elem:
3775 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
3776 map->map_type != BPF_MAP_TYPE_STACK)
3779 case BPF_FUNC_sk_storage_get:
3780 case BPF_FUNC_sk_storage_delete:
3781 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
3790 verbose(env, "cannot pass map_type %d into func %s#%d\n",
3791 map->map_type, func_id_name(func_id), func_id);
3795 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
3799 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
3801 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
3803 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
3805 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
3807 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
3810 /* We only support one arg being in raw mode at the moment,
3811 * which is sufficient for the helper functions we have
3817 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
3818 enum bpf_arg_type arg_next)
3820 return (arg_type_is_mem_ptr(arg_curr) &&
3821 !arg_type_is_mem_size(arg_next)) ||
3822 (!arg_type_is_mem_ptr(arg_curr) &&
3823 arg_type_is_mem_size(arg_next));
3826 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
3828 /* bpf_xxx(..., buf, len) call will access 'len'
3829 * bytes from memory 'buf'. Both arg types need
3830 * to be paired, so make sure there's no buggy
3831 * helper function specification.
3833 if (arg_type_is_mem_size(fn->arg1_type) ||
3834 arg_type_is_mem_ptr(fn->arg5_type) ||
3835 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
3836 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
3837 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
3838 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
3844 static bool check_refcount_ok(const struct bpf_func_proto *fn, int func_id)
3848 if (arg_type_may_be_refcounted(fn->arg1_type))
3850 if (arg_type_may_be_refcounted(fn->arg2_type))
3852 if (arg_type_may_be_refcounted(fn->arg3_type))
3854 if (arg_type_may_be_refcounted(fn->arg4_type))
3856 if (arg_type_may_be_refcounted(fn->arg5_type))
3859 /* A reference acquiring function cannot acquire
3860 * another refcounted ptr.
3862 if (is_acquire_function(func_id) && count)
3865 /* We only support one arg being unreferenced at the moment,
3866 * which is sufficient for the helper functions we have right now.
3871 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
3873 return check_raw_mode_ok(fn) &&
3874 check_arg_pair_ok(fn) &&
3875 check_refcount_ok(fn, func_id) ? 0 : -EINVAL;
3878 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
3879 * are now invalid, so turn them into unknown SCALAR_VALUE.
3881 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
3882 struct bpf_func_state *state)
3884 struct bpf_reg_state *regs = state->regs, *reg;
3887 for (i = 0; i < MAX_BPF_REG; i++)
3888 if (reg_is_pkt_pointer_any(®s[i]))
3889 mark_reg_unknown(env, regs, i);
3891 bpf_for_each_spilled_reg(i, state, reg) {
3894 if (reg_is_pkt_pointer_any(reg))
3895 __mark_reg_unknown(reg);
3899 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
3901 struct bpf_verifier_state *vstate = env->cur_state;
3904 for (i = 0; i <= vstate->curframe; i++)
3905 __clear_all_pkt_pointers(env, vstate->frame[i]);
3908 static void release_reg_references(struct bpf_verifier_env *env,
3909 struct bpf_func_state *state,
3912 struct bpf_reg_state *regs = state->regs, *reg;
3915 for (i = 0; i < MAX_BPF_REG; i++)
3916 if (regs[i].ref_obj_id == ref_obj_id)
3917 mark_reg_unknown(env, regs, i);
3919 bpf_for_each_spilled_reg(i, state, reg) {
3922 if (reg->ref_obj_id == ref_obj_id)
3923 __mark_reg_unknown(reg);
3927 /* The pointer with the specified id has released its reference to kernel
3928 * resources. Identify all copies of the same pointer and clear the reference.
3930 static int release_reference(struct bpf_verifier_env *env,
3933 struct bpf_verifier_state *vstate = env->cur_state;
3937 err = release_reference_state(cur_func(env), ref_obj_id);
3941 for (i = 0; i <= vstate->curframe; i++)
3942 release_reg_references(env, vstate->frame[i], ref_obj_id);
3947 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
3950 struct bpf_verifier_state *state = env->cur_state;
3951 struct bpf_func_state *caller, *callee;
3952 int i, err, subprog, target_insn;
3954 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
3955 verbose(env, "the call stack of %d frames is too deep\n",
3956 state->curframe + 2);
3960 target_insn = *insn_idx + insn->imm;
3961 subprog = find_subprog(env, target_insn + 1);
3963 verbose(env, "verifier bug. No program starts at insn %d\n",
3968 caller = state->frame[state->curframe];
3969 if (state->frame[state->curframe + 1]) {
3970 verbose(env, "verifier bug. Frame %d already allocated\n",
3971 state->curframe + 1);
3975 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
3978 state->frame[state->curframe + 1] = callee;
3980 /* callee cannot access r0, r6 - r9 for reading and has to write
3981 * into its own stack before reading from it.
3982 * callee can read/write into caller's stack
3984 init_func_state(env, callee,
3985 /* remember the callsite, it will be used by bpf_exit */
3986 *insn_idx /* callsite */,
3987 state->curframe + 1 /* frameno within this callchain */,
3988 subprog /* subprog number within this prog */);
3990 /* Transfer references to the callee */
3991 err = transfer_reference_state(callee, caller);
3995 /* copy r1 - r5 args that callee can access. The copy includes parent
3996 * pointers, which connects us up to the liveness chain
3998 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
3999 callee->regs[i] = caller->regs[i];
4001 /* after the call registers r0 - r5 were scratched */
4002 for (i = 0; i < CALLER_SAVED_REGS; i++) {
4003 mark_reg_not_init(env, caller->regs, caller_saved[i]);
4004 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
4007 /* only increment it after check_reg_arg() finished */
4010 if (btf_check_func_arg_match(env, subprog))
4013 /* and go analyze first insn of the callee */
4014 *insn_idx = target_insn;
4016 if (env->log.level & BPF_LOG_LEVEL) {
4017 verbose(env, "caller:\n");
4018 print_verifier_state(env, caller);
4019 verbose(env, "callee:\n");
4020 print_verifier_state(env, callee);
4025 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
4027 struct bpf_verifier_state *state = env->cur_state;
4028 struct bpf_func_state *caller, *callee;
4029 struct bpf_reg_state *r0;
4032 callee = state->frame[state->curframe];
4033 r0 = &callee->regs[BPF_REG_0];
4034 if (r0->type == PTR_TO_STACK) {
4035 /* technically it's ok to return caller's stack pointer
4036 * (or caller's caller's pointer) back to the caller,
4037 * since these pointers are valid. Only current stack
4038 * pointer will be invalid as soon as function exits,
4039 * but let's be conservative
4041 verbose(env, "cannot return stack pointer to the caller\n");
4046 caller = state->frame[state->curframe];
4047 /* return to the caller whatever r0 had in the callee */
4048 caller->regs[BPF_REG_0] = *r0;
4050 /* Transfer references to the caller */
4051 err = transfer_reference_state(caller, callee);
4055 *insn_idx = callee->callsite + 1;
4056 if (env->log.level & BPF_LOG_LEVEL) {
4057 verbose(env, "returning from callee:\n");
4058 print_verifier_state(env, callee);
4059 verbose(env, "to caller at %d:\n", *insn_idx);
4060 print_verifier_state(env, caller);
4062 /* clear everything in the callee */
4063 free_func_state(callee);
4064 state->frame[state->curframe + 1] = NULL;
4068 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
4070 struct bpf_call_arg_meta *meta)
4072 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
4074 if (ret_type != RET_INTEGER ||
4075 (func_id != BPF_FUNC_get_stack &&
4076 func_id != BPF_FUNC_probe_read_str))
4079 ret_reg->smax_value = meta->msize_smax_value;
4080 ret_reg->umax_value = meta->msize_umax_value;
4081 __reg_deduce_bounds(ret_reg);
4082 __reg_bound_offset(ret_reg);
4086 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
4087 int func_id, int insn_idx)
4089 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
4090 struct bpf_map *map = meta->map_ptr;
4092 if (func_id != BPF_FUNC_tail_call &&
4093 func_id != BPF_FUNC_map_lookup_elem &&
4094 func_id != BPF_FUNC_map_update_elem &&
4095 func_id != BPF_FUNC_map_delete_elem &&
4096 func_id != BPF_FUNC_map_push_elem &&
4097 func_id != BPF_FUNC_map_pop_elem &&
4098 func_id != BPF_FUNC_map_peek_elem)
4102 verbose(env, "kernel subsystem misconfigured verifier\n");
4106 /* In case of read-only, some additional restrictions
4107 * need to be applied in order to prevent altering the
4108 * state of the map from program side.
4110 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
4111 (func_id == BPF_FUNC_map_delete_elem ||
4112 func_id == BPF_FUNC_map_update_elem ||
4113 func_id == BPF_FUNC_map_push_elem ||
4114 func_id == BPF_FUNC_map_pop_elem)) {
4115 verbose(env, "write into map forbidden\n");
4119 if (!BPF_MAP_PTR(aux->map_ptr_state))
4120 bpf_map_ptr_store(aux, meta->map_ptr,
4121 meta->map_ptr->unpriv_array);
4122 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
4123 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
4124 meta->map_ptr->unpriv_array);
4129 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
4130 int func_id, int insn_idx)
4132 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
4133 struct bpf_reg_state *regs = cur_regs(env), *reg;
4134 struct bpf_map *map = meta->map_ptr;
4139 if (func_id != BPF_FUNC_tail_call)
4141 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
4142 verbose(env, "kernel subsystem misconfigured verifier\n");
4146 range = tnum_range(0, map->max_entries - 1);
4147 reg = ®s[BPF_REG_3];
4149 if (!register_is_const(reg) || !tnum_in(range, reg->var_off)) {
4150 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
4154 err = mark_chain_precision(env, BPF_REG_3);
4158 val = reg->var_off.value;
4159 if (bpf_map_key_unseen(aux))
4160 bpf_map_key_store(aux, val);
4161 else if (!bpf_map_key_poisoned(aux) &&
4162 bpf_map_key_immediate(aux) != val)
4163 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
4167 static int check_reference_leak(struct bpf_verifier_env *env)
4169 struct bpf_func_state *state = cur_func(env);
4172 for (i = 0; i < state->acquired_refs; i++) {
4173 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
4174 state->refs[i].id, state->refs[i].insn_idx);
4176 return state->acquired_refs ? -EINVAL : 0;
4179 static int check_helper_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
4181 const struct bpf_func_proto *fn = NULL;
4182 struct bpf_reg_state *regs;
4183 struct bpf_call_arg_meta meta;
4187 /* find function prototype */
4188 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
4189 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
4194 if (env->ops->get_func_proto)
4195 fn = env->ops->get_func_proto(func_id, env->prog);
4197 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
4202 /* eBPF programs must be GPL compatible to use GPL-ed functions */
4203 if (!env->prog->gpl_compatible && fn->gpl_only) {
4204 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
4208 /* With LD_ABS/IND some JITs save/restore skb from r1. */
4209 changes_data = bpf_helper_changes_pkt_data(fn->func);
4210 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
4211 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
4212 func_id_name(func_id), func_id);
4216 memset(&meta, 0, sizeof(meta));
4217 meta.pkt_access = fn->pkt_access;
4219 err = check_func_proto(fn, func_id);
4221 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
4222 func_id_name(func_id), func_id);
4226 meta.func_id = func_id;
4228 for (i = 0; i < 5; i++) {
4229 err = btf_resolve_helper_id(&env->log, fn, i);
4232 err = check_func_arg(env, BPF_REG_1 + i, fn->arg_type[i], &meta);
4237 err = record_func_map(env, &meta, func_id, insn_idx);
4241 err = record_func_key(env, &meta, func_id, insn_idx);
4245 /* Mark slots with STACK_MISC in case of raw mode, stack offset
4246 * is inferred from register state.
4248 for (i = 0; i < meta.access_size; i++) {
4249 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
4250 BPF_WRITE, -1, false);
4255 if (func_id == BPF_FUNC_tail_call) {
4256 err = check_reference_leak(env);
4258 verbose(env, "tail_call would lead to reference leak\n");
4261 } else if (is_release_function(func_id)) {
4262 err = release_reference(env, meta.ref_obj_id);
4264 verbose(env, "func %s#%d reference has not been acquired before\n",
4265 func_id_name(func_id), func_id);
4270 regs = cur_regs(env);
4272 /* check that flags argument in get_local_storage(map, flags) is 0,
4273 * this is required because get_local_storage() can't return an error.
4275 if (func_id == BPF_FUNC_get_local_storage &&
4276 !register_is_null(®s[BPF_REG_2])) {
4277 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
4281 /* reset caller saved regs */
4282 for (i = 0; i < CALLER_SAVED_REGS; i++) {
4283 mark_reg_not_init(env, regs, caller_saved[i]);
4284 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
4287 /* helper call returns 64-bit value. */
4288 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
4290 /* update return register (already marked as written above) */
4291 if (fn->ret_type == RET_INTEGER) {
4292 /* sets type to SCALAR_VALUE */
4293 mark_reg_unknown(env, regs, BPF_REG_0);
4294 } else if (fn->ret_type == RET_VOID) {
4295 regs[BPF_REG_0].type = NOT_INIT;
4296 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL ||
4297 fn->ret_type == RET_PTR_TO_MAP_VALUE) {
4298 /* There is no offset yet applied, variable or fixed */
4299 mark_reg_known_zero(env, regs, BPF_REG_0);
4300 /* remember map_ptr, so that check_map_access()
4301 * can check 'value_size' boundary of memory access
4302 * to map element returned from bpf_map_lookup_elem()
4304 if (meta.map_ptr == NULL) {
4306 "kernel subsystem misconfigured verifier\n");
4309 regs[BPF_REG_0].map_ptr = meta.map_ptr;
4310 if (fn->ret_type == RET_PTR_TO_MAP_VALUE) {
4311 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE;
4312 if (map_value_has_spin_lock(meta.map_ptr))
4313 regs[BPF_REG_0].id = ++env->id_gen;
4315 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
4316 regs[BPF_REG_0].id = ++env->id_gen;
4318 } else if (fn->ret_type == RET_PTR_TO_SOCKET_OR_NULL) {
4319 mark_reg_known_zero(env, regs, BPF_REG_0);
4320 regs[BPF_REG_0].type = PTR_TO_SOCKET_OR_NULL;
4321 regs[BPF_REG_0].id = ++env->id_gen;
4322 } else if (fn->ret_type == RET_PTR_TO_SOCK_COMMON_OR_NULL) {
4323 mark_reg_known_zero(env, regs, BPF_REG_0);
4324 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON_OR_NULL;
4325 regs[BPF_REG_0].id = ++env->id_gen;
4326 } else if (fn->ret_type == RET_PTR_TO_TCP_SOCK_OR_NULL) {
4327 mark_reg_known_zero(env, regs, BPF_REG_0);
4328 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK_OR_NULL;
4329 regs[BPF_REG_0].id = ++env->id_gen;
4331 verbose(env, "unknown return type %d of func %s#%d\n",
4332 fn->ret_type, func_id_name(func_id), func_id);
4336 if (is_ptr_cast_function(func_id)) {
4337 /* For release_reference() */
4338 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
4339 } else if (is_acquire_function(func_id)) {
4340 int id = acquire_reference_state(env, insn_idx);
4344 /* For mark_ptr_or_null_reg() */
4345 regs[BPF_REG_0].id = id;
4346 /* For release_reference() */
4347 regs[BPF_REG_0].ref_obj_id = id;
4350 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
4352 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
4356 if (func_id == BPF_FUNC_get_stack && !env->prog->has_callchain_buf) {
4357 const char *err_str;
4359 #ifdef CONFIG_PERF_EVENTS
4360 err = get_callchain_buffers(sysctl_perf_event_max_stack);
4361 err_str = "cannot get callchain buffer for func %s#%d\n";
4364 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
4367 verbose(env, err_str, func_id_name(func_id), func_id);
4371 env->prog->has_callchain_buf = true;
4375 clear_all_pkt_pointers(env);
4379 static bool signed_add_overflows(s64 a, s64 b)
4381 /* Do the add in u64, where overflow is well-defined */
4382 s64 res = (s64)((u64)a + (u64)b);
4389 static bool signed_sub_overflows(s64 a, s64 b)
4391 /* Do the sub in u64, where overflow is well-defined */
4392 s64 res = (s64)((u64)a - (u64)b);
4399 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
4400 const struct bpf_reg_state *reg,
4401 enum bpf_reg_type type)
4403 bool known = tnum_is_const(reg->var_off);
4404 s64 val = reg->var_off.value;
4405 s64 smin = reg->smin_value;
4407 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
4408 verbose(env, "math between %s pointer and %lld is not allowed\n",
4409 reg_type_str[type], val);
4413 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
4414 verbose(env, "%s pointer offset %d is not allowed\n",
4415 reg_type_str[type], reg->off);
4419 if (smin == S64_MIN) {
4420 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
4421 reg_type_str[type]);
4425 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
4426 verbose(env, "value %lld makes %s pointer be out of bounds\n",
4427 smin, reg_type_str[type]);
4434 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
4436 return &env->insn_aux_data[env->insn_idx];
4439 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
4440 u32 *ptr_limit, u8 opcode, bool off_is_neg)
4442 bool mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
4443 (opcode == BPF_SUB && !off_is_neg);
4446 switch (ptr_reg->type) {
4448 /* Indirect variable offset stack access is prohibited in
4449 * unprivileged mode so it's not handled here.
4451 off = ptr_reg->off + ptr_reg->var_off.value;
4453 *ptr_limit = MAX_BPF_STACK + off;
4457 case PTR_TO_MAP_VALUE:
4459 *ptr_limit = ptr_reg->umax_value + ptr_reg->off;
4461 off = ptr_reg->smin_value + ptr_reg->off;
4462 *ptr_limit = ptr_reg->map_ptr->value_size - off;
4470 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
4471 const struct bpf_insn *insn)
4473 return env->allow_ptr_leaks || BPF_SRC(insn->code) == BPF_K;
4476 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
4477 u32 alu_state, u32 alu_limit)
4479 /* If we arrived here from different branches with different
4480 * state or limits to sanitize, then this won't work.
4482 if (aux->alu_state &&
4483 (aux->alu_state != alu_state ||
4484 aux->alu_limit != alu_limit))
4487 /* Corresponding fixup done in fixup_bpf_calls(). */
4488 aux->alu_state = alu_state;
4489 aux->alu_limit = alu_limit;
4493 static int sanitize_val_alu(struct bpf_verifier_env *env,
4494 struct bpf_insn *insn)
4496 struct bpf_insn_aux_data *aux = cur_aux(env);
4498 if (can_skip_alu_sanitation(env, insn))
4501 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
4504 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
4505 struct bpf_insn *insn,
4506 const struct bpf_reg_state *ptr_reg,
4507 struct bpf_reg_state *dst_reg,
4510 struct bpf_verifier_state *vstate = env->cur_state;
4511 struct bpf_insn_aux_data *aux = cur_aux(env);
4512 bool ptr_is_dst_reg = ptr_reg == dst_reg;
4513 u8 opcode = BPF_OP(insn->code);
4514 u32 alu_state, alu_limit;
4515 struct bpf_reg_state tmp;
4518 if (can_skip_alu_sanitation(env, insn))
4521 /* We already marked aux for masking from non-speculative
4522 * paths, thus we got here in the first place. We only care
4523 * to explore bad access from here.
4525 if (vstate->speculative)
4528 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
4529 alu_state |= ptr_is_dst_reg ?
4530 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
4532 if (retrieve_ptr_limit(ptr_reg, &alu_limit, opcode, off_is_neg))
4534 if (update_alu_sanitation_state(aux, alu_state, alu_limit))
4537 /* Simulate and find potential out-of-bounds access under
4538 * speculative execution from truncation as a result of
4539 * masking when off was not within expected range. If off
4540 * sits in dst, then we temporarily need to move ptr there
4541 * to simulate dst (== 0) +/-= ptr. Needed, for example,
4542 * for cases where we use K-based arithmetic in one direction
4543 * and truncated reg-based in the other in order to explore
4546 if (!ptr_is_dst_reg) {
4548 *dst_reg = *ptr_reg;
4550 ret = push_stack(env, env->insn_idx + 1, env->insn_idx, true);
4551 if (!ptr_is_dst_reg && ret)
4553 return !ret ? -EFAULT : 0;
4556 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
4557 * Caller should also handle BPF_MOV case separately.
4558 * If we return -EACCES, caller may want to try again treating pointer as a
4559 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
4561 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
4562 struct bpf_insn *insn,
4563 const struct bpf_reg_state *ptr_reg,
4564 const struct bpf_reg_state *off_reg)
4566 struct bpf_verifier_state *vstate = env->cur_state;
4567 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4568 struct bpf_reg_state *regs = state->regs, *dst_reg;
4569 bool known = tnum_is_const(off_reg->var_off);
4570 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
4571 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
4572 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
4573 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
4574 u32 dst = insn->dst_reg, src = insn->src_reg;
4575 u8 opcode = BPF_OP(insn->code);
4578 dst_reg = ®s[dst];
4580 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
4581 smin_val > smax_val || umin_val > umax_val) {
4582 /* Taint dst register if offset had invalid bounds derived from
4583 * e.g. dead branches.
4585 __mark_reg_unknown(dst_reg);
4589 if (BPF_CLASS(insn->code) != BPF_ALU64) {
4590 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
4592 "R%d 32-bit pointer arithmetic prohibited\n",
4597 switch (ptr_reg->type) {
4598 case PTR_TO_MAP_VALUE_OR_NULL:
4599 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
4600 dst, reg_type_str[ptr_reg->type]);
4602 case CONST_PTR_TO_MAP:
4603 case PTR_TO_PACKET_END:
4605 case PTR_TO_SOCKET_OR_NULL:
4606 case PTR_TO_SOCK_COMMON:
4607 case PTR_TO_SOCK_COMMON_OR_NULL:
4608 case PTR_TO_TCP_SOCK:
4609 case PTR_TO_TCP_SOCK_OR_NULL:
4610 case PTR_TO_XDP_SOCK:
4611 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
4612 dst, reg_type_str[ptr_reg->type]);
4614 case PTR_TO_MAP_VALUE:
4615 if (!env->allow_ptr_leaks && !known && (smin_val < 0) != (smax_val < 0)) {
4616 verbose(env, "R%d has unknown scalar with mixed signed bounds, pointer arithmetic with it prohibited for !root\n",
4617 off_reg == dst_reg ? dst : src);
4625 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
4626 * The id may be overwritten later if we create a new variable offset.
4628 dst_reg->type = ptr_reg->type;
4629 dst_reg->id = ptr_reg->id;
4631 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
4632 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
4637 ret = sanitize_ptr_alu(env, insn, ptr_reg, dst_reg, smin_val < 0);
4639 verbose(env, "R%d tried to add from different maps or paths\n", dst);
4642 /* We can take a fixed offset as long as it doesn't overflow
4643 * the s32 'off' field
4645 if (known && (ptr_reg->off + smin_val ==
4646 (s64)(s32)(ptr_reg->off + smin_val))) {
4647 /* pointer += K. Accumulate it into fixed offset */
4648 dst_reg->smin_value = smin_ptr;
4649 dst_reg->smax_value = smax_ptr;
4650 dst_reg->umin_value = umin_ptr;
4651 dst_reg->umax_value = umax_ptr;
4652 dst_reg->var_off = ptr_reg->var_off;
4653 dst_reg->off = ptr_reg->off + smin_val;
4654 dst_reg->raw = ptr_reg->raw;
4657 /* A new variable offset is created. Note that off_reg->off
4658 * == 0, since it's a scalar.
4659 * dst_reg gets the pointer type and since some positive
4660 * integer value was added to the pointer, give it a new 'id'
4661 * if it's a PTR_TO_PACKET.
4662 * this creates a new 'base' pointer, off_reg (variable) gets
4663 * added into the variable offset, and we copy the fixed offset
4666 if (signed_add_overflows(smin_ptr, smin_val) ||
4667 signed_add_overflows(smax_ptr, smax_val)) {
4668 dst_reg->smin_value = S64_MIN;
4669 dst_reg->smax_value = S64_MAX;
4671 dst_reg->smin_value = smin_ptr + smin_val;
4672 dst_reg->smax_value = smax_ptr + smax_val;
4674 if (umin_ptr + umin_val < umin_ptr ||
4675 umax_ptr + umax_val < umax_ptr) {
4676 dst_reg->umin_value = 0;
4677 dst_reg->umax_value = U64_MAX;
4679 dst_reg->umin_value = umin_ptr + umin_val;
4680 dst_reg->umax_value = umax_ptr + umax_val;
4682 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
4683 dst_reg->off = ptr_reg->off;
4684 dst_reg->raw = ptr_reg->raw;
4685 if (reg_is_pkt_pointer(ptr_reg)) {
4686 dst_reg->id = ++env->id_gen;
4687 /* something was added to pkt_ptr, set range to zero */
4692 ret = sanitize_ptr_alu(env, insn, ptr_reg, dst_reg, smin_val < 0);
4694 verbose(env, "R%d tried to sub from different maps or paths\n", dst);
4697 if (dst_reg == off_reg) {
4698 /* scalar -= pointer. Creates an unknown scalar */
4699 verbose(env, "R%d tried to subtract pointer from scalar\n",
4703 /* We don't allow subtraction from FP, because (according to
4704 * test_verifier.c test "invalid fp arithmetic", JITs might not
4705 * be able to deal with it.
4707 if (ptr_reg->type == PTR_TO_STACK) {
4708 verbose(env, "R%d subtraction from stack pointer prohibited\n",
4712 if (known && (ptr_reg->off - smin_val ==
4713 (s64)(s32)(ptr_reg->off - smin_val))) {
4714 /* pointer -= K. Subtract it from fixed offset */
4715 dst_reg->smin_value = smin_ptr;
4716 dst_reg->smax_value = smax_ptr;
4717 dst_reg->umin_value = umin_ptr;
4718 dst_reg->umax_value = umax_ptr;
4719 dst_reg->var_off = ptr_reg->var_off;
4720 dst_reg->id = ptr_reg->id;
4721 dst_reg->off = ptr_reg->off - smin_val;
4722 dst_reg->raw = ptr_reg->raw;
4725 /* A new variable offset is created. If the subtrahend is known
4726 * nonnegative, then any reg->range we had before is still good.
4728 if (signed_sub_overflows(smin_ptr, smax_val) ||
4729 signed_sub_overflows(smax_ptr, smin_val)) {
4730 /* Overflow possible, we know nothing */
4731 dst_reg->smin_value = S64_MIN;
4732 dst_reg->smax_value = S64_MAX;
4734 dst_reg->smin_value = smin_ptr - smax_val;
4735 dst_reg->smax_value = smax_ptr - smin_val;
4737 if (umin_ptr < umax_val) {
4738 /* Overflow possible, we know nothing */
4739 dst_reg->umin_value = 0;
4740 dst_reg->umax_value = U64_MAX;
4742 /* Cannot overflow (as long as bounds are consistent) */
4743 dst_reg->umin_value = umin_ptr - umax_val;
4744 dst_reg->umax_value = umax_ptr - umin_val;
4746 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
4747 dst_reg->off = ptr_reg->off;
4748 dst_reg->raw = ptr_reg->raw;
4749 if (reg_is_pkt_pointer(ptr_reg)) {
4750 dst_reg->id = ++env->id_gen;
4751 /* something was added to pkt_ptr, set range to zero */
4759 /* bitwise ops on pointers are troublesome, prohibit. */
4760 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
4761 dst, bpf_alu_string[opcode >> 4]);
4764 /* other operators (e.g. MUL,LSH) produce non-pointer results */
4765 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
4766 dst, bpf_alu_string[opcode >> 4]);
4770 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
4773 __update_reg_bounds(dst_reg);
4774 __reg_deduce_bounds(dst_reg);
4775 __reg_bound_offset(dst_reg);
4777 /* For unprivileged we require that resulting offset must be in bounds
4778 * in order to be able to sanitize access later on.
4780 if (!env->allow_ptr_leaks) {
4781 if (dst_reg->type == PTR_TO_MAP_VALUE &&
4782 check_map_access(env, dst, dst_reg->off, 1, false)) {
4783 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
4784 "prohibited for !root\n", dst);
4786 } else if (dst_reg->type == PTR_TO_STACK &&
4787 check_stack_access(env, dst_reg, dst_reg->off +
4788 dst_reg->var_off.value, 1)) {
4789 verbose(env, "R%d stack pointer arithmetic goes out of range, "
4790 "prohibited for !root\n", dst);
4798 /* WARNING: This function does calculations on 64-bit values, but the actual
4799 * execution may occur on 32-bit values. Therefore, things like bitshifts
4800 * need extra checks in the 32-bit case.
4802 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
4803 struct bpf_insn *insn,
4804 struct bpf_reg_state *dst_reg,
4805 struct bpf_reg_state src_reg)
4807 struct bpf_reg_state *regs = cur_regs(env);
4808 u8 opcode = BPF_OP(insn->code);
4809 bool src_known, dst_known;
4810 s64 smin_val, smax_val;
4811 u64 umin_val, umax_val;
4812 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
4813 u32 dst = insn->dst_reg;
4816 if (insn_bitness == 32) {
4817 /* Relevant for 32-bit RSH: Information can propagate towards
4818 * LSB, so it isn't sufficient to only truncate the output to
4821 coerce_reg_to_size(dst_reg, 4);
4822 coerce_reg_to_size(&src_reg, 4);
4825 smin_val = src_reg.smin_value;
4826 smax_val = src_reg.smax_value;
4827 umin_val = src_reg.umin_value;
4828 umax_val = src_reg.umax_value;
4829 src_known = tnum_is_const(src_reg.var_off);
4830 dst_known = tnum_is_const(dst_reg->var_off);
4832 if ((src_known && (smin_val != smax_val || umin_val != umax_val)) ||
4833 smin_val > smax_val || umin_val > umax_val) {
4834 /* Taint dst register if offset had invalid bounds derived from
4835 * e.g. dead branches.
4837 __mark_reg_unknown(dst_reg);
4842 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
4843 __mark_reg_unknown(dst_reg);
4849 ret = sanitize_val_alu(env, insn);
4851 verbose(env, "R%d tried to add from different pointers or scalars\n", dst);
4854 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
4855 signed_add_overflows(dst_reg->smax_value, smax_val)) {
4856 dst_reg->smin_value = S64_MIN;
4857 dst_reg->smax_value = S64_MAX;
4859 dst_reg->smin_value += smin_val;
4860 dst_reg->smax_value += smax_val;
4862 if (dst_reg->umin_value + umin_val < umin_val ||
4863 dst_reg->umax_value + umax_val < umax_val) {
4864 dst_reg->umin_value = 0;
4865 dst_reg->umax_value = U64_MAX;
4867 dst_reg->umin_value += umin_val;
4868 dst_reg->umax_value += umax_val;
4870 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg.var_off);
4873 ret = sanitize_val_alu(env, insn);
4875 verbose(env, "R%d tried to sub from different pointers or scalars\n", dst);
4878 if (signed_sub_overflows(dst_reg->smin_value, smax_val) ||
4879 signed_sub_overflows(dst_reg->smax_value, smin_val)) {
4880 /* Overflow possible, we know nothing */
4881 dst_reg->smin_value = S64_MIN;
4882 dst_reg->smax_value = S64_MAX;
4884 dst_reg->smin_value -= smax_val;
4885 dst_reg->smax_value -= smin_val;
4887 if (dst_reg->umin_value < umax_val) {
4888 /* Overflow possible, we know nothing */
4889 dst_reg->umin_value = 0;
4890 dst_reg->umax_value = U64_MAX;
4892 /* Cannot overflow (as long as bounds are consistent) */
4893 dst_reg->umin_value -= umax_val;
4894 dst_reg->umax_value -= umin_val;
4896 dst_reg->var_off = tnum_sub(dst_reg->var_off, src_reg.var_off);
4899 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg.var_off);
4900 if (smin_val < 0 || dst_reg->smin_value < 0) {
4901 /* Ain't nobody got time to multiply that sign */
4902 __mark_reg_unbounded(dst_reg);
4903 __update_reg_bounds(dst_reg);
4906 /* Both values are positive, so we can work with unsigned and
4907 * copy the result to signed (unless it exceeds S64_MAX).
4909 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
4910 /* Potential overflow, we know nothing */
4911 __mark_reg_unbounded(dst_reg);
4912 /* (except what we can learn from the var_off) */
4913 __update_reg_bounds(dst_reg);
4916 dst_reg->umin_value *= umin_val;
4917 dst_reg->umax_value *= umax_val;
4918 if (dst_reg->umax_value > S64_MAX) {
4919 /* Overflow possible, we know nothing */
4920 dst_reg->smin_value = S64_MIN;
4921 dst_reg->smax_value = S64_MAX;
4923 dst_reg->smin_value = dst_reg->umin_value;
4924 dst_reg->smax_value = dst_reg->umax_value;
4928 if (src_known && dst_known) {
4929 __mark_reg_known(dst_reg, dst_reg->var_off.value &
4930 src_reg.var_off.value);
4933 /* We get our minimum from the var_off, since that's inherently
4934 * bitwise. Our maximum is the minimum of the operands' maxima.
4936 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg.var_off);
4937 dst_reg->umin_value = dst_reg->var_off.value;
4938 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
4939 if (dst_reg->smin_value < 0 || smin_val < 0) {
4940 /* Lose signed bounds when ANDing negative numbers,
4941 * ain't nobody got time for that.
4943 dst_reg->smin_value = S64_MIN;
4944 dst_reg->smax_value = S64_MAX;
4946 /* ANDing two positives gives a positive, so safe to
4947 * cast result into s64.
4949 dst_reg->smin_value = dst_reg->umin_value;
4950 dst_reg->smax_value = dst_reg->umax_value;
4952 /* We may learn something more from the var_off */
4953 __update_reg_bounds(dst_reg);
4956 if (src_known && dst_known) {
4957 __mark_reg_known(dst_reg, dst_reg->var_off.value |
4958 src_reg.var_off.value);
4961 /* We get our maximum from the var_off, and our minimum is the
4962 * maximum of the operands' minima
4964 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg.var_off);
4965 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
4966 dst_reg->umax_value = dst_reg->var_off.value |
4967 dst_reg->var_off.mask;
4968 if (dst_reg->smin_value < 0 || smin_val < 0) {
4969 /* Lose signed bounds when ORing negative numbers,
4970 * ain't nobody got time for that.
4972 dst_reg->smin_value = S64_MIN;
4973 dst_reg->smax_value = S64_MAX;
4975 /* ORing two positives gives a positive, so safe to
4976 * cast result into s64.
4978 dst_reg->smin_value = dst_reg->umin_value;
4979 dst_reg->smax_value = dst_reg->umax_value;
4981 /* We may learn something more from the var_off */
4982 __update_reg_bounds(dst_reg);
4985 if (umax_val >= insn_bitness) {
4986 /* Shifts greater than 31 or 63 are undefined.
4987 * This includes shifts by a negative number.
4989 mark_reg_unknown(env, regs, insn->dst_reg);
4992 /* We lose all sign bit information (except what we can pick
4995 dst_reg->smin_value = S64_MIN;
4996 dst_reg->smax_value = S64_MAX;
4997 /* If we might shift our top bit out, then we know nothing */
4998 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
4999 dst_reg->umin_value = 0;
5000 dst_reg->umax_value = U64_MAX;
5002 dst_reg->umin_value <<= umin_val;
5003 dst_reg->umax_value <<= umax_val;
5005 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
5006 /* We may learn something more from the var_off */
5007 __update_reg_bounds(dst_reg);
5010 if (umax_val >= insn_bitness) {
5011 /* Shifts greater than 31 or 63 are undefined.
5012 * This includes shifts by a negative number.
5014 mark_reg_unknown(env, regs, insn->dst_reg);
5017 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
5018 * be negative, then either:
5019 * 1) src_reg might be zero, so the sign bit of the result is
5020 * unknown, so we lose our signed bounds
5021 * 2) it's known negative, thus the unsigned bounds capture the
5023 * 3) the signed bounds cross zero, so they tell us nothing
5025 * If the value in dst_reg is known nonnegative, then again the
5026 * unsigned bounts capture the signed bounds.
5027 * Thus, in all cases it suffices to blow away our signed bounds
5028 * and rely on inferring new ones from the unsigned bounds and
5029 * var_off of the result.
5031 dst_reg->smin_value = S64_MIN;
5032 dst_reg->smax_value = S64_MAX;
5033 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
5034 dst_reg->umin_value >>= umax_val;
5035 dst_reg->umax_value >>= umin_val;
5036 /* We may learn something more from the var_off */
5037 __update_reg_bounds(dst_reg);
5040 if (umax_val >= insn_bitness) {
5041 /* Shifts greater than 31 or 63 are undefined.
5042 * This includes shifts by a negative number.
5044 mark_reg_unknown(env, regs, insn->dst_reg);
5048 /* Upon reaching here, src_known is true and
5049 * umax_val is equal to umin_val.
5051 dst_reg->smin_value >>= umin_val;
5052 dst_reg->smax_value >>= umin_val;
5053 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val);
5055 /* blow away the dst_reg umin_value/umax_value and rely on
5056 * dst_reg var_off to refine the result.
5058 dst_reg->umin_value = 0;
5059 dst_reg->umax_value = U64_MAX;
5060 __update_reg_bounds(dst_reg);
5063 mark_reg_unknown(env, regs, insn->dst_reg);
5067 if (BPF_CLASS(insn->code) != BPF_ALU64) {
5068 /* 32-bit ALU ops are (32,32)->32 */
5069 coerce_reg_to_size(dst_reg, 4);
5072 __reg_deduce_bounds(dst_reg);
5073 __reg_bound_offset(dst_reg);
5077 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
5080 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
5081 struct bpf_insn *insn)
5083 struct bpf_verifier_state *vstate = env->cur_state;
5084 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5085 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
5086 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
5087 u8 opcode = BPF_OP(insn->code);
5090 dst_reg = ®s[insn->dst_reg];
5092 if (dst_reg->type != SCALAR_VALUE)
5094 if (BPF_SRC(insn->code) == BPF_X) {
5095 src_reg = ®s[insn->src_reg];
5096 if (src_reg->type != SCALAR_VALUE) {
5097 if (dst_reg->type != SCALAR_VALUE) {
5098 /* Combining two pointers by any ALU op yields
5099 * an arbitrary scalar. Disallow all math except
5100 * pointer subtraction
5102 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
5103 mark_reg_unknown(env, regs, insn->dst_reg);
5106 verbose(env, "R%d pointer %s pointer prohibited\n",
5108 bpf_alu_string[opcode >> 4]);
5111 /* scalar += pointer
5112 * This is legal, but we have to reverse our
5113 * src/dest handling in computing the range
5115 err = mark_chain_precision(env, insn->dst_reg);
5118 return adjust_ptr_min_max_vals(env, insn,
5121 } else if (ptr_reg) {
5122 /* pointer += scalar */
5123 err = mark_chain_precision(env, insn->src_reg);
5126 return adjust_ptr_min_max_vals(env, insn,
5130 /* Pretend the src is a reg with a known value, since we only
5131 * need to be able to read from this state.
5133 off_reg.type = SCALAR_VALUE;
5134 __mark_reg_known(&off_reg, insn->imm);
5136 if (ptr_reg) /* pointer += K */
5137 return adjust_ptr_min_max_vals(env, insn,
5141 /* Got here implies adding two SCALAR_VALUEs */
5142 if (WARN_ON_ONCE(ptr_reg)) {
5143 print_verifier_state(env, state);
5144 verbose(env, "verifier internal error: unexpected ptr_reg\n");
5147 if (WARN_ON(!src_reg)) {
5148 print_verifier_state(env, state);
5149 verbose(env, "verifier internal error: no src_reg\n");
5152 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
5155 /* check validity of 32-bit and 64-bit arithmetic operations */
5156 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
5158 struct bpf_reg_state *regs = cur_regs(env);
5159 u8 opcode = BPF_OP(insn->code);
5162 if (opcode == BPF_END || opcode == BPF_NEG) {
5163 if (opcode == BPF_NEG) {
5164 if (BPF_SRC(insn->code) != 0 ||
5165 insn->src_reg != BPF_REG_0 ||
5166 insn->off != 0 || insn->imm != 0) {
5167 verbose(env, "BPF_NEG uses reserved fields\n");
5171 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
5172 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
5173 BPF_CLASS(insn->code) == BPF_ALU64) {
5174 verbose(env, "BPF_END uses reserved fields\n");
5179 /* check src operand */
5180 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
5184 if (is_pointer_value(env, insn->dst_reg)) {
5185 verbose(env, "R%d pointer arithmetic prohibited\n",
5190 /* check dest operand */
5191 err = check_reg_arg(env, insn->dst_reg, DST_OP);
5195 } else if (opcode == BPF_MOV) {
5197 if (BPF_SRC(insn->code) == BPF_X) {
5198 if (insn->imm != 0 || insn->off != 0) {
5199 verbose(env, "BPF_MOV uses reserved fields\n");
5203 /* check src operand */
5204 err = check_reg_arg(env, insn->src_reg, SRC_OP);
5208 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
5209 verbose(env, "BPF_MOV uses reserved fields\n");
5214 /* check dest operand, mark as required later */
5215 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
5219 if (BPF_SRC(insn->code) == BPF_X) {
5220 struct bpf_reg_state *src_reg = regs + insn->src_reg;
5221 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
5223 if (BPF_CLASS(insn->code) == BPF_ALU64) {
5225 * copy register state to dest reg
5227 *dst_reg = *src_reg;
5228 dst_reg->live |= REG_LIVE_WRITTEN;
5229 dst_reg->subreg_def = DEF_NOT_SUBREG;
5232 if (is_pointer_value(env, insn->src_reg)) {
5234 "R%d partial copy of pointer\n",
5237 } else if (src_reg->type == SCALAR_VALUE) {
5238 *dst_reg = *src_reg;
5239 dst_reg->live |= REG_LIVE_WRITTEN;
5240 dst_reg->subreg_def = env->insn_idx + 1;
5242 mark_reg_unknown(env, regs,
5245 coerce_reg_to_size(dst_reg, 4);
5249 * remember the value we stored into this reg
5251 /* clear any state __mark_reg_known doesn't set */
5252 mark_reg_unknown(env, regs, insn->dst_reg);
5253 regs[insn->dst_reg].type = SCALAR_VALUE;
5254 if (BPF_CLASS(insn->code) == BPF_ALU64) {
5255 __mark_reg_known(regs + insn->dst_reg,
5258 __mark_reg_known(regs + insn->dst_reg,
5263 } else if (opcode > BPF_END) {
5264 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
5267 } else { /* all other ALU ops: and, sub, xor, add, ... */
5269 if (BPF_SRC(insn->code) == BPF_X) {
5270 if (insn->imm != 0 || insn->off != 0) {
5271 verbose(env, "BPF_ALU uses reserved fields\n");
5274 /* check src1 operand */
5275 err = check_reg_arg(env, insn->src_reg, SRC_OP);
5279 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
5280 verbose(env, "BPF_ALU uses reserved fields\n");
5285 /* check src2 operand */
5286 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
5290 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
5291 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
5292 verbose(env, "div by zero\n");
5296 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
5297 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
5298 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
5300 if (insn->imm < 0 || insn->imm >= size) {
5301 verbose(env, "invalid shift %d\n", insn->imm);
5306 /* check dest operand */
5307 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
5311 return adjust_reg_min_max_vals(env, insn);
5317 static void __find_good_pkt_pointers(struct bpf_func_state *state,
5318 struct bpf_reg_state *dst_reg,
5319 enum bpf_reg_type type, u16 new_range)
5321 struct bpf_reg_state *reg;
5324 for (i = 0; i < MAX_BPF_REG; i++) {
5325 reg = &state->regs[i];
5326 if (reg->type == type && reg->id == dst_reg->id)
5327 /* keep the maximum range already checked */
5328 reg->range = max(reg->range, new_range);
5331 bpf_for_each_spilled_reg(i, state, reg) {
5334 if (reg->type == type && reg->id == dst_reg->id)
5335 reg->range = max(reg->range, new_range);
5339 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
5340 struct bpf_reg_state *dst_reg,
5341 enum bpf_reg_type type,
5342 bool range_right_open)
5347 if (dst_reg->off < 0 ||
5348 (dst_reg->off == 0 && range_right_open))
5349 /* This doesn't give us any range */
5352 if (dst_reg->umax_value > MAX_PACKET_OFF ||
5353 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
5354 /* Risk of overflow. For instance, ptr + (1<<63) may be less
5355 * than pkt_end, but that's because it's also less than pkt.
5359 new_range = dst_reg->off;
5360 if (range_right_open)
5363 /* Examples for register markings:
5365 * pkt_data in dst register:
5369 * if (r2 > pkt_end) goto <handle exception>
5374 * if (r2 < pkt_end) goto <access okay>
5375 * <handle exception>
5378 * r2 == dst_reg, pkt_end == src_reg
5379 * r2=pkt(id=n,off=8,r=0)
5380 * r3=pkt(id=n,off=0,r=0)
5382 * pkt_data in src register:
5386 * if (pkt_end >= r2) goto <access okay>
5387 * <handle exception>
5391 * if (pkt_end <= r2) goto <handle exception>
5395 * pkt_end == dst_reg, r2 == src_reg
5396 * r2=pkt(id=n,off=8,r=0)
5397 * r3=pkt(id=n,off=0,r=0)
5399 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
5400 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
5401 * and [r3, r3 + 8-1) respectively is safe to access depending on
5405 /* If our ids match, then we must have the same max_value. And we
5406 * don't care about the other reg's fixed offset, since if it's too big
5407 * the range won't allow anything.
5408 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
5410 for (i = 0; i <= vstate->curframe; i++)
5411 __find_good_pkt_pointers(vstate->frame[i], dst_reg, type,
5415 /* compute branch direction of the expression "if (reg opcode val) goto target;"
5417 * 1 - branch will be taken and "goto target" will be executed
5418 * 0 - branch will not be taken and fall-through to next insn
5419 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value range [0,10]
5421 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
5424 struct bpf_reg_state reg_lo;
5427 if (__is_pointer_value(false, reg))
5433 /* For JMP32, only low 32 bits are compared, coerce_reg_to_size
5434 * could truncate high bits and update umin/umax according to
5435 * information of low bits.
5437 coerce_reg_to_size(reg, 4);
5438 /* smin/smax need special handling. For example, after coerce,
5439 * if smin_value is 0x00000000ffffffffLL, the value is -1 when
5440 * used as operand to JMP32. It is a negative number from s32's
5441 * point of view, while it is a positive number when seen as
5442 * s64. The smin/smax are kept as s64, therefore, when used with
5443 * JMP32, they need to be transformed into s32, then sign
5444 * extended back to s64.
5446 * Also, smin/smax were copied from umin/umax. If umin/umax has
5447 * different sign bit, then min/max relationship doesn't
5448 * maintain after casting into s32, for this case, set smin/smax
5451 if ((reg->umax_value ^ reg->umin_value) &
5453 reg->smin_value = S32_MIN;
5454 reg->smax_value = S32_MAX;
5456 reg->smin_value = (s64)(s32)reg->smin_value;
5457 reg->smax_value = (s64)(s32)reg->smax_value;
5460 sval = (s64)(s32)val;
5467 if (tnum_is_const(reg->var_off))
5468 return !!tnum_equals_const(reg->var_off, val);
5471 if (tnum_is_const(reg->var_off))
5472 return !tnum_equals_const(reg->var_off, val);
5475 if ((~reg->var_off.mask & reg->var_off.value) & val)
5477 if (!((reg->var_off.mask | reg->var_off.value) & val))
5481 if (reg->umin_value > val)
5483 else if (reg->umax_value <= val)
5487 if (reg->smin_value > sval)
5489 else if (reg->smax_value < sval)
5493 if (reg->umax_value < val)
5495 else if (reg->umin_value >= val)
5499 if (reg->smax_value < sval)
5501 else if (reg->smin_value >= sval)
5505 if (reg->umin_value >= val)
5507 else if (reg->umax_value < val)
5511 if (reg->smin_value >= sval)
5513 else if (reg->smax_value < sval)
5517 if (reg->umax_value <= val)
5519 else if (reg->umin_value > val)
5523 if (reg->smax_value <= sval)
5525 else if (reg->smin_value > sval)
5533 /* Generate min value of the high 32-bit from TNUM info. */
5534 static u64 gen_hi_min(struct tnum var)
5536 return var.value & ~0xffffffffULL;
5539 /* Generate max value of the high 32-bit from TNUM info. */
5540 static u64 gen_hi_max(struct tnum var)
5542 return (var.value | var.mask) & ~0xffffffffULL;
5545 /* Return true if VAL is compared with a s64 sign extended from s32, and they
5546 * are with the same signedness.
5548 static bool cmp_val_with_extended_s64(s64 sval, struct bpf_reg_state *reg)
5550 return ((s32)sval >= 0 &&
5551 reg->smin_value >= 0 && reg->smax_value <= S32_MAX) ||
5553 reg->smax_value <= 0 && reg->smin_value >= S32_MIN);
5556 /* Adjusts the register min/max values in the case that the dst_reg is the
5557 * variable register that we are working on, and src_reg is a constant or we're
5558 * simply doing a BPF_K check.
5559 * In JEQ/JNE cases we also adjust the var_off values.
5561 static void reg_set_min_max(struct bpf_reg_state *true_reg,
5562 struct bpf_reg_state *false_reg, u64 val,
5563 u8 opcode, bool is_jmp32)
5567 /* If the dst_reg is a pointer, we can't learn anything about its
5568 * variable offset from the compare (unless src_reg were a pointer into
5569 * the same object, but we don't bother with that.
5570 * Since false_reg and true_reg have the same type by construction, we
5571 * only need to check one of them for pointerness.
5573 if (__is_pointer_value(false, false_reg))
5576 val = is_jmp32 ? (u32)val : val;
5577 sval = is_jmp32 ? (s64)(s32)val : (s64)val;
5583 struct bpf_reg_state *reg =
5584 opcode == BPF_JEQ ? true_reg : false_reg;
5586 /* For BPF_JEQ, if this is false we know nothing Jon Snow, but
5587 * if it is true we know the value for sure. Likewise for
5591 u64 old_v = reg->var_off.value;
5592 u64 hi_mask = ~0xffffffffULL;
5594 reg->var_off.value = (old_v & hi_mask) | val;
5595 reg->var_off.mask &= hi_mask;
5597 __mark_reg_known(reg, val);
5602 false_reg->var_off = tnum_and(false_reg->var_off,
5604 if (is_power_of_2(val))
5605 true_reg->var_off = tnum_or(true_reg->var_off,
5611 u64 false_umax = opcode == BPF_JGT ? val : val - 1;
5612 u64 true_umin = opcode == BPF_JGT ? val + 1 : val;
5615 false_umax += gen_hi_max(false_reg->var_off);
5616 true_umin += gen_hi_min(true_reg->var_off);
5618 false_reg->umax_value = min(false_reg->umax_value, false_umax);
5619 true_reg->umin_value = max(true_reg->umin_value, true_umin);
5625 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
5626 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
5628 /* If the full s64 was not sign-extended from s32 then don't
5629 * deduct further info.
5631 if (is_jmp32 && !cmp_val_with_extended_s64(sval, false_reg))
5633 false_reg->smax_value = min(false_reg->smax_value, false_smax);
5634 true_reg->smin_value = max(true_reg->smin_value, true_smin);
5640 u64 false_umin = opcode == BPF_JLT ? val : val + 1;
5641 u64 true_umax = opcode == BPF_JLT ? val - 1 : val;
5644 false_umin += gen_hi_min(false_reg->var_off);
5645 true_umax += gen_hi_max(true_reg->var_off);
5647 false_reg->umin_value = max(false_reg->umin_value, false_umin);
5648 true_reg->umax_value = min(true_reg->umax_value, true_umax);
5654 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
5655 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
5657 if (is_jmp32 && !cmp_val_with_extended_s64(sval, false_reg))
5659 false_reg->smin_value = max(false_reg->smin_value, false_smin);
5660 true_reg->smax_value = min(true_reg->smax_value, true_smax);
5667 __reg_deduce_bounds(false_reg);
5668 __reg_deduce_bounds(true_reg);
5669 /* We might have learned some bits from the bounds. */
5670 __reg_bound_offset(false_reg);
5671 __reg_bound_offset(true_reg);
5673 __reg_bound_offset32(false_reg);
5674 __reg_bound_offset32(true_reg);
5676 /* Intersecting with the old var_off might have improved our bounds
5677 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
5678 * then new var_off is (0; 0x7f...fc) which improves our umax.
5680 __update_reg_bounds(false_reg);
5681 __update_reg_bounds(true_reg);
5684 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
5687 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
5688 struct bpf_reg_state *false_reg, u64 val,
5689 u8 opcode, bool is_jmp32)
5693 if (__is_pointer_value(false, false_reg))
5696 val = is_jmp32 ? (u32)val : val;
5697 sval = is_jmp32 ? (s64)(s32)val : (s64)val;
5703 struct bpf_reg_state *reg =
5704 opcode == BPF_JEQ ? true_reg : false_reg;
5707 u64 old_v = reg->var_off.value;
5708 u64 hi_mask = ~0xffffffffULL;
5710 reg->var_off.value = (old_v & hi_mask) | val;
5711 reg->var_off.mask &= hi_mask;
5713 __mark_reg_known(reg, val);
5718 false_reg->var_off = tnum_and(false_reg->var_off,
5720 if (is_power_of_2(val))
5721 true_reg->var_off = tnum_or(true_reg->var_off,
5727 u64 false_umin = opcode == BPF_JGT ? val : val + 1;
5728 u64 true_umax = opcode == BPF_JGT ? val - 1 : val;
5731 false_umin += gen_hi_min(false_reg->var_off);
5732 true_umax += gen_hi_max(true_reg->var_off);
5734 false_reg->umin_value = max(false_reg->umin_value, false_umin);
5735 true_reg->umax_value = min(true_reg->umax_value, true_umax);
5741 s64 false_smin = opcode == BPF_JSGT ? sval : sval + 1;
5742 s64 true_smax = opcode == BPF_JSGT ? sval - 1 : sval;
5744 if (is_jmp32 && !cmp_val_with_extended_s64(sval, false_reg))
5746 false_reg->smin_value = max(false_reg->smin_value, false_smin);
5747 true_reg->smax_value = min(true_reg->smax_value, true_smax);
5753 u64 false_umax = opcode == BPF_JLT ? val : val - 1;
5754 u64 true_umin = opcode == BPF_JLT ? val + 1 : val;
5757 false_umax += gen_hi_max(false_reg->var_off);
5758 true_umin += gen_hi_min(true_reg->var_off);
5760 false_reg->umax_value = min(false_reg->umax_value, false_umax);
5761 true_reg->umin_value = max(true_reg->umin_value, true_umin);
5767 s64 false_smax = opcode == BPF_JSLT ? sval : sval - 1;
5768 s64 true_smin = opcode == BPF_JSLT ? sval + 1 : sval;
5770 if (is_jmp32 && !cmp_val_with_extended_s64(sval, false_reg))
5772 false_reg->smax_value = min(false_reg->smax_value, false_smax);
5773 true_reg->smin_value = max(true_reg->smin_value, true_smin);
5780 __reg_deduce_bounds(false_reg);
5781 __reg_deduce_bounds(true_reg);
5782 /* We might have learned some bits from the bounds. */
5783 __reg_bound_offset(false_reg);
5784 __reg_bound_offset(true_reg);
5786 __reg_bound_offset32(false_reg);
5787 __reg_bound_offset32(true_reg);
5789 /* Intersecting with the old var_off might have improved our bounds
5790 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
5791 * then new var_off is (0; 0x7f...fc) which improves our umax.
5793 __update_reg_bounds(false_reg);
5794 __update_reg_bounds(true_reg);
5797 /* Regs are known to be equal, so intersect their min/max/var_off */
5798 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
5799 struct bpf_reg_state *dst_reg)
5801 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
5802 dst_reg->umin_value);
5803 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
5804 dst_reg->umax_value);
5805 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
5806 dst_reg->smin_value);
5807 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
5808 dst_reg->smax_value);
5809 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
5811 /* We might have learned new bounds from the var_off. */
5812 __update_reg_bounds(src_reg);
5813 __update_reg_bounds(dst_reg);
5814 /* We might have learned something about the sign bit. */
5815 __reg_deduce_bounds(src_reg);
5816 __reg_deduce_bounds(dst_reg);
5817 /* We might have learned some bits from the bounds. */
5818 __reg_bound_offset(src_reg);
5819 __reg_bound_offset(dst_reg);
5820 /* Intersecting with the old var_off might have improved our bounds
5821 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
5822 * then new var_off is (0; 0x7f...fc) which improves our umax.
5824 __update_reg_bounds(src_reg);
5825 __update_reg_bounds(dst_reg);
5828 static void reg_combine_min_max(struct bpf_reg_state *true_src,
5829 struct bpf_reg_state *true_dst,
5830 struct bpf_reg_state *false_src,
5831 struct bpf_reg_state *false_dst,
5836 __reg_combine_min_max(true_src, true_dst);
5839 __reg_combine_min_max(false_src, false_dst);
5844 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
5845 struct bpf_reg_state *reg, u32 id,
5848 if (reg_type_may_be_null(reg->type) && reg->id == id) {
5849 /* Old offset (both fixed and variable parts) should
5850 * have been known-zero, because we don't allow pointer
5851 * arithmetic on pointers that might be NULL.
5853 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
5854 !tnum_equals_const(reg->var_off, 0) ||
5856 __mark_reg_known_zero(reg);
5860 reg->type = SCALAR_VALUE;
5861 } else if (reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
5862 if (reg->map_ptr->inner_map_meta) {
5863 reg->type = CONST_PTR_TO_MAP;
5864 reg->map_ptr = reg->map_ptr->inner_map_meta;
5865 } else if (reg->map_ptr->map_type ==
5866 BPF_MAP_TYPE_XSKMAP) {
5867 reg->type = PTR_TO_XDP_SOCK;
5869 reg->type = PTR_TO_MAP_VALUE;
5871 } else if (reg->type == PTR_TO_SOCKET_OR_NULL) {
5872 reg->type = PTR_TO_SOCKET;
5873 } else if (reg->type == PTR_TO_SOCK_COMMON_OR_NULL) {
5874 reg->type = PTR_TO_SOCK_COMMON;
5875 } else if (reg->type == PTR_TO_TCP_SOCK_OR_NULL) {
5876 reg->type = PTR_TO_TCP_SOCK;
5879 /* We don't need id and ref_obj_id from this point
5880 * onwards anymore, thus we should better reset it,
5881 * so that state pruning has chances to take effect.
5884 reg->ref_obj_id = 0;
5885 } else if (!reg_may_point_to_spin_lock(reg)) {
5886 /* For not-NULL ptr, reg->ref_obj_id will be reset
5887 * in release_reg_references().
5889 * reg->id is still used by spin_lock ptr. Other
5890 * than spin_lock ptr type, reg->id can be reset.
5897 static void __mark_ptr_or_null_regs(struct bpf_func_state *state, u32 id,
5900 struct bpf_reg_state *reg;
5903 for (i = 0; i < MAX_BPF_REG; i++)
5904 mark_ptr_or_null_reg(state, &state->regs[i], id, is_null);
5906 bpf_for_each_spilled_reg(i, state, reg) {
5909 mark_ptr_or_null_reg(state, reg, id, is_null);
5913 /* The logic is similar to find_good_pkt_pointers(), both could eventually
5914 * be folded together at some point.
5916 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
5919 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5920 struct bpf_reg_state *regs = state->regs;
5921 u32 ref_obj_id = regs[regno].ref_obj_id;
5922 u32 id = regs[regno].id;
5925 if (ref_obj_id && ref_obj_id == id && is_null)
5926 /* regs[regno] is in the " == NULL" branch.
5927 * No one could have freed the reference state before
5928 * doing the NULL check.
5930 WARN_ON_ONCE(release_reference_state(state, id));
5932 for (i = 0; i <= vstate->curframe; i++)
5933 __mark_ptr_or_null_regs(vstate->frame[i], id, is_null);
5936 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
5937 struct bpf_reg_state *dst_reg,
5938 struct bpf_reg_state *src_reg,
5939 struct bpf_verifier_state *this_branch,
5940 struct bpf_verifier_state *other_branch)
5942 if (BPF_SRC(insn->code) != BPF_X)
5945 /* Pointers are always 64-bit. */
5946 if (BPF_CLASS(insn->code) == BPF_JMP32)
5949 switch (BPF_OP(insn->code)) {
5951 if ((dst_reg->type == PTR_TO_PACKET &&
5952 src_reg->type == PTR_TO_PACKET_END) ||
5953 (dst_reg->type == PTR_TO_PACKET_META &&
5954 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
5955 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
5956 find_good_pkt_pointers(this_branch, dst_reg,
5957 dst_reg->type, false);
5958 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
5959 src_reg->type == PTR_TO_PACKET) ||
5960 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
5961 src_reg->type == PTR_TO_PACKET_META)) {
5962 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
5963 find_good_pkt_pointers(other_branch, src_reg,
5964 src_reg->type, true);
5970 if ((dst_reg->type == PTR_TO_PACKET &&
5971 src_reg->type == PTR_TO_PACKET_END) ||
5972 (dst_reg->type == PTR_TO_PACKET_META &&
5973 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
5974 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
5975 find_good_pkt_pointers(other_branch, dst_reg,
5976 dst_reg->type, true);
5977 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
5978 src_reg->type == PTR_TO_PACKET) ||
5979 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
5980 src_reg->type == PTR_TO_PACKET_META)) {
5981 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
5982 find_good_pkt_pointers(this_branch, src_reg,
5983 src_reg->type, false);
5989 if ((dst_reg->type == PTR_TO_PACKET &&
5990 src_reg->type == PTR_TO_PACKET_END) ||
5991 (dst_reg->type == PTR_TO_PACKET_META &&
5992 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
5993 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
5994 find_good_pkt_pointers(this_branch, dst_reg,
5995 dst_reg->type, true);
5996 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
5997 src_reg->type == PTR_TO_PACKET) ||
5998 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
5999 src_reg->type == PTR_TO_PACKET_META)) {
6000 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
6001 find_good_pkt_pointers(other_branch, src_reg,
6002 src_reg->type, false);
6008 if ((dst_reg->type == PTR_TO_PACKET &&
6009 src_reg->type == PTR_TO_PACKET_END) ||
6010 (dst_reg->type == PTR_TO_PACKET_META &&
6011 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
6012 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
6013 find_good_pkt_pointers(other_branch, dst_reg,
6014 dst_reg->type, false);
6015 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
6016 src_reg->type == PTR_TO_PACKET) ||
6017 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
6018 src_reg->type == PTR_TO_PACKET_META)) {
6019 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
6020 find_good_pkt_pointers(this_branch, src_reg,
6021 src_reg->type, true);
6033 static int check_cond_jmp_op(struct bpf_verifier_env *env,
6034 struct bpf_insn *insn, int *insn_idx)
6036 struct bpf_verifier_state *this_branch = env->cur_state;
6037 struct bpf_verifier_state *other_branch;
6038 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
6039 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
6040 u8 opcode = BPF_OP(insn->code);
6045 /* Only conditional jumps are expected to reach here. */
6046 if (opcode == BPF_JA || opcode > BPF_JSLE) {
6047 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
6051 if (BPF_SRC(insn->code) == BPF_X) {
6052 if (insn->imm != 0) {
6053 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
6057 /* check src1 operand */
6058 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6062 if (is_pointer_value(env, insn->src_reg)) {
6063 verbose(env, "R%d pointer comparison prohibited\n",
6067 src_reg = ®s[insn->src_reg];
6069 if (insn->src_reg != BPF_REG_0) {
6070 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
6075 /* check src2 operand */
6076 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
6080 dst_reg = ®s[insn->dst_reg];
6081 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
6083 if (BPF_SRC(insn->code) == BPF_K)
6084 pred = is_branch_taken(dst_reg, insn->imm,
6086 else if (src_reg->type == SCALAR_VALUE &&
6087 tnum_is_const(src_reg->var_off))
6088 pred = is_branch_taken(dst_reg, src_reg->var_off.value,
6091 err = mark_chain_precision(env, insn->dst_reg);
6092 if (BPF_SRC(insn->code) == BPF_X && !err)
6093 err = mark_chain_precision(env, insn->src_reg);
6098 /* only follow the goto, ignore fall-through */
6099 *insn_idx += insn->off;
6101 } else if (pred == 0) {
6102 /* only follow fall-through branch, since
6103 * that's where the program will go
6108 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
6112 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
6114 /* detect if we are comparing against a constant value so we can adjust
6115 * our min/max values for our dst register.
6116 * this is only legit if both are scalars (or pointers to the same
6117 * object, I suppose, but we don't support that right now), because
6118 * otherwise the different base pointers mean the offsets aren't
6121 if (BPF_SRC(insn->code) == BPF_X) {
6122 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
6123 struct bpf_reg_state lo_reg0 = *dst_reg;
6124 struct bpf_reg_state lo_reg1 = *src_reg;
6125 struct bpf_reg_state *src_lo, *dst_lo;
6129 coerce_reg_to_size(dst_lo, 4);
6130 coerce_reg_to_size(src_lo, 4);
6132 if (dst_reg->type == SCALAR_VALUE &&
6133 src_reg->type == SCALAR_VALUE) {
6134 if (tnum_is_const(src_reg->var_off) ||
6135 (is_jmp32 && tnum_is_const(src_lo->var_off)))
6136 reg_set_min_max(&other_branch_regs[insn->dst_reg],
6139 ? src_lo->var_off.value
6140 : src_reg->var_off.value,
6142 else if (tnum_is_const(dst_reg->var_off) ||
6143 (is_jmp32 && tnum_is_const(dst_lo->var_off)))
6144 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
6147 ? dst_lo->var_off.value
6148 : dst_reg->var_off.value,
6150 else if (!is_jmp32 &&
6151 (opcode == BPF_JEQ || opcode == BPF_JNE))
6152 /* Comparing for equality, we can combine knowledge */
6153 reg_combine_min_max(&other_branch_regs[insn->src_reg],
6154 &other_branch_regs[insn->dst_reg],
6155 src_reg, dst_reg, opcode);
6157 } else if (dst_reg->type == SCALAR_VALUE) {
6158 reg_set_min_max(&other_branch_regs[insn->dst_reg],
6159 dst_reg, insn->imm, opcode, is_jmp32);
6162 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
6163 * NOTE: these optimizations below are related with pointer comparison
6164 * which will never be JMP32.
6166 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
6167 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
6168 reg_type_may_be_null(dst_reg->type)) {
6169 /* Mark all identical registers in each branch as either
6170 * safe or unknown depending R == 0 or R != 0 conditional.
6172 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
6174 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
6176 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
6177 this_branch, other_branch) &&
6178 is_pointer_value(env, insn->dst_reg)) {
6179 verbose(env, "R%d pointer comparison prohibited\n",
6183 if (env->log.level & BPF_LOG_LEVEL)
6184 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
6188 /* verify BPF_LD_IMM64 instruction */
6189 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
6191 struct bpf_insn_aux_data *aux = cur_aux(env);
6192 struct bpf_reg_state *regs = cur_regs(env);
6193 struct bpf_map *map;
6196 if (BPF_SIZE(insn->code) != BPF_DW) {
6197 verbose(env, "invalid BPF_LD_IMM insn\n");
6200 if (insn->off != 0) {
6201 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
6205 err = check_reg_arg(env, insn->dst_reg, DST_OP);
6209 if (insn->src_reg == 0) {
6210 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
6212 regs[insn->dst_reg].type = SCALAR_VALUE;
6213 __mark_reg_known(®s[insn->dst_reg], imm);
6217 map = env->used_maps[aux->map_index];
6218 mark_reg_known_zero(env, regs, insn->dst_reg);
6219 regs[insn->dst_reg].map_ptr = map;
6221 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE) {
6222 regs[insn->dst_reg].type = PTR_TO_MAP_VALUE;
6223 regs[insn->dst_reg].off = aux->map_off;
6224 if (map_value_has_spin_lock(map))
6225 regs[insn->dst_reg].id = ++env->id_gen;
6226 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD) {
6227 regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
6229 verbose(env, "bpf verifier is misconfigured\n");
6236 static bool may_access_skb(enum bpf_prog_type type)
6239 case BPF_PROG_TYPE_SOCKET_FILTER:
6240 case BPF_PROG_TYPE_SCHED_CLS:
6241 case BPF_PROG_TYPE_SCHED_ACT:
6248 /* verify safety of LD_ABS|LD_IND instructions:
6249 * - they can only appear in the programs where ctx == skb
6250 * - since they are wrappers of function calls, they scratch R1-R5 registers,
6251 * preserve R6-R9, and store return value into R0
6254 * ctx == skb == R6 == CTX
6257 * SRC == any register
6258 * IMM == 32-bit immediate
6261 * R0 - 8/16/32-bit skb data converted to cpu endianness
6263 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
6265 struct bpf_reg_state *regs = cur_regs(env);
6266 u8 mode = BPF_MODE(insn->code);
6269 if (!may_access_skb(env->prog->type)) {
6270 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
6274 if (!env->ops->gen_ld_abs) {
6275 verbose(env, "bpf verifier is misconfigured\n");
6279 if (env->subprog_cnt > 1) {
6280 /* when program has LD_ABS insn JITs and interpreter assume
6281 * that r1 == ctx == skb which is not the case for callees
6282 * that can have arbitrary arguments. It's problematic
6283 * for main prog as well since JITs would need to analyze
6284 * all functions in order to make proper register save/restore
6285 * decisions in the main prog. Hence disallow LD_ABS with calls
6287 verbose(env, "BPF_LD_[ABS|IND] instructions cannot be mixed with bpf-to-bpf calls\n");
6291 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
6292 BPF_SIZE(insn->code) == BPF_DW ||
6293 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
6294 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
6298 /* check whether implicit source operand (register R6) is readable */
6299 err = check_reg_arg(env, BPF_REG_6, SRC_OP);
6303 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
6304 * gen_ld_abs() may terminate the program at runtime, leading to
6307 err = check_reference_leak(env);
6309 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
6313 if (env->cur_state->active_spin_lock) {
6314 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
6318 if (regs[BPF_REG_6].type != PTR_TO_CTX) {
6320 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
6324 if (mode == BPF_IND) {
6325 /* check explicit source operand */
6326 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6331 /* reset caller saved regs to unreadable */
6332 for (i = 0; i < CALLER_SAVED_REGS; i++) {
6333 mark_reg_not_init(env, regs, caller_saved[i]);
6334 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
6337 /* mark destination R0 register as readable, since it contains
6338 * the value fetched from the packet.
6339 * Already marked as written above.
6341 mark_reg_unknown(env, regs, BPF_REG_0);
6342 /* ld_abs load up to 32-bit skb data. */
6343 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
6347 static int check_return_code(struct bpf_verifier_env *env)
6349 struct tnum enforce_attach_type_range = tnum_unknown;
6350 struct bpf_reg_state *reg;
6351 struct tnum range = tnum_range(0, 1);
6353 switch (env->prog->type) {
6354 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
6355 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
6356 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG)
6357 range = tnum_range(1, 1);
6359 case BPF_PROG_TYPE_CGROUP_SKB:
6360 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
6361 range = tnum_range(0, 3);
6362 enforce_attach_type_range = tnum_range(2, 3);
6365 case BPF_PROG_TYPE_CGROUP_SOCK:
6366 case BPF_PROG_TYPE_SOCK_OPS:
6367 case BPF_PROG_TYPE_CGROUP_DEVICE:
6368 case BPF_PROG_TYPE_CGROUP_SYSCTL:
6369 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
6371 case BPF_PROG_TYPE_RAW_TRACEPOINT:
6372 if (!env->prog->aux->attach_btf_id)
6374 range = tnum_const(0);
6380 reg = cur_regs(env) + BPF_REG_0;
6381 if (reg->type != SCALAR_VALUE) {
6382 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
6383 reg_type_str[reg->type]);
6387 if (!tnum_in(range, reg->var_off)) {
6390 verbose(env, "At program exit the register R0 ");
6391 if (!tnum_is_unknown(reg->var_off)) {
6392 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6393 verbose(env, "has value %s", tn_buf);
6395 verbose(env, "has unknown scalar value");
6397 tnum_strn(tn_buf, sizeof(tn_buf), range);
6398 verbose(env, " should have been in %s\n", tn_buf);
6402 if (!tnum_is_unknown(enforce_attach_type_range) &&
6403 tnum_in(enforce_attach_type_range, reg->var_off))
6404 env->prog->enforce_expected_attach_type = 1;
6408 /* non-recursive DFS pseudo code
6409 * 1 procedure DFS-iterative(G,v):
6410 * 2 label v as discovered
6411 * 3 let S be a stack
6413 * 5 while S is not empty
6415 * 7 if t is what we're looking for:
6417 * 9 for all edges e in G.adjacentEdges(t) do
6418 * 10 if edge e is already labelled
6419 * 11 continue with the next edge
6420 * 12 w <- G.adjacentVertex(t,e)
6421 * 13 if vertex w is not discovered and not explored
6422 * 14 label e as tree-edge
6423 * 15 label w as discovered
6426 * 18 else if vertex w is discovered
6427 * 19 label e as back-edge
6429 * 21 // vertex w is explored
6430 * 22 label e as forward- or cross-edge
6431 * 23 label t as explored
6436 * 0x11 - discovered and fall-through edge labelled
6437 * 0x12 - discovered and fall-through and branch edges labelled
6448 static u32 state_htab_size(struct bpf_verifier_env *env)
6450 return env->prog->len;
6453 static struct bpf_verifier_state_list **explored_state(
6454 struct bpf_verifier_env *env,
6457 struct bpf_verifier_state *cur = env->cur_state;
6458 struct bpf_func_state *state = cur->frame[cur->curframe];
6460 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
6463 static void init_explored_state(struct bpf_verifier_env *env, int idx)
6465 env->insn_aux_data[idx].prune_point = true;
6468 /* t, w, e - match pseudo-code above:
6469 * t - index of current instruction
6470 * w - next instruction
6473 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env,
6476 int *insn_stack = env->cfg.insn_stack;
6477 int *insn_state = env->cfg.insn_state;
6479 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
6482 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
6485 if (w < 0 || w >= env->prog->len) {
6486 verbose_linfo(env, t, "%d: ", t);
6487 verbose(env, "jump out of range from insn %d to %d\n", t, w);
6492 /* mark branch target for state pruning */
6493 init_explored_state(env, w);
6495 if (insn_state[w] == 0) {
6497 insn_state[t] = DISCOVERED | e;
6498 insn_state[w] = DISCOVERED;
6499 if (env->cfg.cur_stack >= env->prog->len)
6501 insn_stack[env->cfg.cur_stack++] = w;
6503 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
6504 if (loop_ok && env->allow_ptr_leaks)
6506 verbose_linfo(env, t, "%d: ", t);
6507 verbose_linfo(env, w, "%d: ", w);
6508 verbose(env, "back-edge from insn %d to %d\n", t, w);
6510 } else if (insn_state[w] == EXPLORED) {
6511 /* forward- or cross-edge */
6512 insn_state[t] = DISCOVERED | e;
6514 verbose(env, "insn state internal bug\n");
6520 /* non-recursive depth-first-search to detect loops in BPF program
6521 * loop == back-edge in directed graph
6523 static int check_cfg(struct bpf_verifier_env *env)
6525 struct bpf_insn *insns = env->prog->insnsi;
6526 int insn_cnt = env->prog->len;
6527 int *insn_stack, *insn_state;
6531 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
6535 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
6541 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
6542 insn_stack[0] = 0; /* 0 is the first instruction */
6543 env->cfg.cur_stack = 1;
6546 if (env->cfg.cur_stack == 0)
6548 t = insn_stack[env->cfg.cur_stack - 1];
6550 if (BPF_CLASS(insns[t].code) == BPF_JMP ||
6551 BPF_CLASS(insns[t].code) == BPF_JMP32) {
6552 u8 opcode = BPF_OP(insns[t].code);
6554 if (opcode == BPF_EXIT) {
6556 } else if (opcode == BPF_CALL) {
6557 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
6562 if (t + 1 < insn_cnt)
6563 init_explored_state(env, t + 1);
6564 if (insns[t].src_reg == BPF_PSEUDO_CALL) {
6565 init_explored_state(env, t);
6566 ret = push_insn(t, t + insns[t].imm + 1, BRANCH,
6573 } else if (opcode == BPF_JA) {
6574 if (BPF_SRC(insns[t].code) != BPF_K) {
6578 /* unconditional jump with single edge */
6579 ret = push_insn(t, t + insns[t].off + 1,
6580 FALLTHROUGH, env, true);
6585 /* unconditional jmp is not a good pruning point,
6586 * but it's marked, since backtracking needs
6587 * to record jmp history in is_state_visited().
6589 init_explored_state(env, t + insns[t].off + 1);
6590 /* tell verifier to check for equivalent states
6591 * after every call and jump
6593 if (t + 1 < insn_cnt)
6594 init_explored_state(env, t + 1);
6596 /* conditional jump with two edges */
6597 init_explored_state(env, t);
6598 ret = push_insn(t, t + 1, FALLTHROUGH, env, true);
6604 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env, true);
6611 /* all other non-branch instructions with single
6614 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
6622 insn_state[t] = EXPLORED;
6623 if (env->cfg.cur_stack-- <= 0) {
6624 verbose(env, "pop stack internal bug\n");
6631 for (i = 0; i < insn_cnt; i++) {
6632 if (insn_state[i] != EXPLORED) {
6633 verbose(env, "unreachable insn %d\n", i);
6638 ret = 0; /* cfg looks good */
6643 env->cfg.insn_state = env->cfg.insn_stack = NULL;
6647 /* The minimum supported BTF func info size */
6648 #define MIN_BPF_FUNCINFO_SIZE 8
6649 #define MAX_FUNCINFO_REC_SIZE 252
6651 static int check_btf_func(struct bpf_verifier_env *env,
6652 const union bpf_attr *attr,
6653 union bpf_attr __user *uattr)
6655 u32 i, nfuncs, urec_size, min_size;
6656 u32 krec_size = sizeof(struct bpf_func_info);
6657 struct bpf_func_info *krecord;
6658 struct bpf_func_info_aux *info_aux = NULL;
6659 const struct btf_type *type;
6660 struct bpf_prog *prog;
6661 const struct btf *btf;
6662 void __user *urecord;
6663 u32 prev_offset = 0;
6666 nfuncs = attr->func_info_cnt;
6670 if (nfuncs != env->subprog_cnt) {
6671 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
6675 urec_size = attr->func_info_rec_size;
6676 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
6677 urec_size > MAX_FUNCINFO_REC_SIZE ||
6678 urec_size % sizeof(u32)) {
6679 verbose(env, "invalid func info rec size %u\n", urec_size);
6684 btf = prog->aux->btf;
6686 urecord = u64_to_user_ptr(attr->func_info);
6687 min_size = min_t(u32, krec_size, urec_size);
6689 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
6692 info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN);
6696 for (i = 0; i < nfuncs; i++) {
6697 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
6699 if (ret == -E2BIG) {
6700 verbose(env, "nonzero tailing record in func info");
6701 /* set the size kernel expects so loader can zero
6702 * out the rest of the record.
6704 if (put_user(min_size, &uattr->func_info_rec_size))
6710 if (copy_from_user(&krecord[i], urecord, min_size)) {
6715 /* check insn_off */
6717 if (krecord[i].insn_off) {
6719 "nonzero insn_off %u for the first func info record",
6720 krecord[i].insn_off);
6724 } else if (krecord[i].insn_off <= prev_offset) {
6726 "same or smaller insn offset (%u) than previous func info record (%u)",
6727 krecord[i].insn_off, prev_offset);
6732 if (env->subprog_info[i].start != krecord[i].insn_off) {
6733 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
6739 type = btf_type_by_id(btf, krecord[i].type_id);
6740 if (!type || BTF_INFO_KIND(type->info) != BTF_KIND_FUNC) {
6741 verbose(env, "invalid type id %d in func info",
6742 krecord[i].type_id);
6746 prev_offset = krecord[i].insn_off;
6747 urecord += urec_size;
6750 prog->aux->func_info = krecord;
6751 prog->aux->func_info_cnt = nfuncs;
6752 prog->aux->func_info_aux = info_aux;
6761 static void adjust_btf_func(struct bpf_verifier_env *env)
6763 struct bpf_prog_aux *aux = env->prog->aux;
6766 if (!aux->func_info)
6769 for (i = 0; i < env->subprog_cnt; i++)
6770 aux->func_info[i].insn_off = env->subprog_info[i].start;
6773 #define MIN_BPF_LINEINFO_SIZE (offsetof(struct bpf_line_info, line_col) + \
6774 sizeof(((struct bpf_line_info *)(0))->line_col))
6775 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
6777 static int check_btf_line(struct bpf_verifier_env *env,
6778 const union bpf_attr *attr,
6779 union bpf_attr __user *uattr)
6781 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
6782 struct bpf_subprog_info *sub;
6783 struct bpf_line_info *linfo;
6784 struct bpf_prog *prog;
6785 const struct btf *btf;
6786 void __user *ulinfo;
6789 nr_linfo = attr->line_info_cnt;
6793 rec_size = attr->line_info_rec_size;
6794 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
6795 rec_size > MAX_LINEINFO_REC_SIZE ||
6796 rec_size & (sizeof(u32) - 1))
6799 /* Need to zero it in case the userspace may
6800 * pass in a smaller bpf_line_info object.
6802 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
6803 GFP_KERNEL | __GFP_NOWARN);
6808 btf = prog->aux->btf;
6811 sub = env->subprog_info;
6812 ulinfo = u64_to_user_ptr(attr->line_info);
6813 expected_size = sizeof(struct bpf_line_info);
6814 ncopy = min_t(u32, expected_size, rec_size);
6815 for (i = 0; i < nr_linfo; i++) {
6816 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
6818 if (err == -E2BIG) {
6819 verbose(env, "nonzero tailing record in line_info");
6820 if (put_user(expected_size,
6821 &uattr->line_info_rec_size))
6827 if (copy_from_user(&linfo[i], ulinfo, ncopy)) {
6833 * Check insn_off to ensure
6834 * 1) strictly increasing AND
6835 * 2) bounded by prog->len
6837 * The linfo[0].insn_off == 0 check logically falls into
6838 * the later "missing bpf_line_info for func..." case
6839 * because the first linfo[0].insn_off must be the
6840 * first sub also and the first sub must have
6841 * subprog_info[0].start == 0.
6843 if ((i && linfo[i].insn_off <= prev_offset) ||
6844 linfo[i].insn_off >= prog->len) {
6845 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
6846 i, linfo[i].insn_off, prev_offset,
6852 if (!prog->insnsi[linfo[i].insn_off].code) {
6854 "Invalid insn code at line_info[%u].insn_off\n",
6860 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
6861 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
6862 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
6867 if (s != env->subprog_cnt) {
6868 if (linfo[i].insn_off == sub[s].start) {
6869 sub[s].linfo_idx = i;
6871 } else if (sub[s].start < linfo[i].insn_off) {
6872 verbose(env, "missing bpf_line_info for func#%u\n", s);
6878 prev_offset = linfo[i].insn_off;
6882 if (s != env->subprog_cnt) {
6883 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
6884 env->subprog_cnt - s, s);
6889 prog->aux->linfo = linfo;
6890 prog->aux->nr_linfo = nr_linfo;
6899 static int check_btf_info(struct bpf_verifier_env *env,
6900 const union bpf_attr *attr,
6901 union bpf_attr __user *uattr)
6906 if (!attr->func_info_cnt && !attr->line_info_cnt)
6909 btf = btf_get_by_fd(attr->prog_btf_fd);
6911 return PTR_ERR(btf);
6912 env->prog->aux->btf = btf;
6914 err = check_btf_func(env, attr, uattr);
6918 err = check_btf_line(env, attr, uattr);
6925 /* check %cur's range satisfies %old's */
6926 static bool range_within(struct bpf_reg_state *old,
6927 struct bpf_reg_state *cur)
6929 return old->umin_value <= cur->umin_value &&
6930 old->umax_value >= cur->umax_value &&
6931 old->smin_value <= cur->smin_value &&
6932 old->smax_value >= cur->smax_value;
6935 /* Maximum number of register states that can exist at once */
6936 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
6942 /* If in the old state two registers had the same id, then they need to have
6943 * the same id in the new state as well. But that id could be different from
6944 * the old state, so we need to track the mapping from old to new ids.
6945 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
6946 * regs with old id 5 must also have new id 9 for the new state to be safe. But
6947 * regs with a different old id could still have new id 9, we don't care about
6949 * So we look through our idmap to see if this old id has been seen before. If
6950 * so, we require the new id to match; otherwise, we add the id pair to the map.
6952 static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
6956 for (i = 0; i < ID_MAP_SIZE; i++) {
6957 if (!idmap[i].old) {
6958 /* Reached an empty slot; haven't seen this id before */
6959 idmap[i].old = old_id;
6960 idmap[i].cur = cur_id;
6963 if (idmap[i].old == old_id)
6964 return idmap[i].cur == cur_id;
6966 /* We ran out of idmap slots, which should be impossible */
6971 static void clean_func_state(struct bpf_verifier_env *env,
6972 struct bpf_func_state *st)
6974 enum bpf_reg_liveness live;
6977 for (i = 0; i < BPF_REG_FP; i++) {
6978 live = st->regs[i].live;
6979 /* liveness must not touch this register anymore */
6980 st->regs[i].live |= REG_LIVE_DONE;
6981 if (!(live & REG_LIVE_READ))
6982 /* since the register is unused, clear its state
6983 * to make further comparison simpler
6985 __mark_reg_not_init(&st->regs[i]);
6988 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
6989 live = st->stack[i].spilled_ptr.live;
6990 /* liveness must not touch this stack slot anymore */
6991 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
6992 if (!(live & REG_LIVE_READ)) {
6993 __mark_reg_not_init(&st->stack[i].spilled_ptr);
6994 for (j = 0; j < BPF_REG_SIZE; j++)
6995 st->stack[i].slot_type[j] = STACK_INVALID;
7000 static void clean_verifier_state(struct bpf_verifier_env *env,
7001 struct bpf_verifier_state *st)
7005 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
7006 /* all regs in this state in all frames were already marked */
7009 for (i = 0; i <= st->curframe; i++)
7010 clean_func_state(env, st->frame[i]);
7013 /* the parentage chains form a tree.
7014 * the verifier states are added to state lists at given insn and
7015 * pushed into state stack for future exploration.
7016 * when the verifier reaches bpf_exit insn some of the verifer states
7017 * stored in the state lists have their final liveness state already,
7018 * but a lot of states will get revised from liveness point of view when
7019 * the verifier explores other branches.
7022 * 2: if r1 == 100 goto pc+1
7025 * when the verifier reaches exit insn the register r0 in the state list of
7026 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
7027 * of insn 2 and goes exploring further. At the insn 4 it will walk the
7028 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
7030 * Since the verifier pushes the branch states as it sees them while exploring
7031 * the program the condition of walking the branch instruction for the second
7032 * time means that all states below this branch were already explored and
7033 * their final liveness markes are already propagated.
7034 * Hence when the verifier completes the search of state list in is_state_visited()
7035 * we can call this clean_live_states() function to mark all liveness states
7036 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
7038 * This function also clears the registers and stack for states that !READ
7039 * to simplify state merging.
7041 * Important note here that walking the same branch instruction in the callee
7042 * doesn't meant that the states are DONE. The verifier has to compare
7045 static void clean_live_states(struct bpf_verifier_env *env, int insn,
7046 struct bpf_verifier_state *cur)
7048 struct bpf_verifier_state_list *sl;
7051 sl = *explored_state(env, insn);
7053 if (sl->state.branches)
7055 if (sl->state.insn_idx != insn ||
7056 sl->state.curframe != cur->curframe)
7058 for (i = 0; i <= cur->curframe; i++)
7059 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
7061 clean_verifier_state(env, &sl->state);
7067 /* Returns true if (rold safe implies rcur safe) */
7068 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
7069 struct idpair *idmap)
7073 if (!(rold->live & REG_LIVE_READ))
7074 /* explored state didn't use this */
7077 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
7079 if (rold->type == PTR_TO_STACK)
7080 /* two stack pointers are equal only if they're pointing to
7081 * the same stack frame, since fp-8 in foo != fp-8 in bar
7083 return equal && rold->frameno == rcur->frameno;
7088 if (rold->type == NOT_INIT)
7089 /* explored state can't have used this */
7091 if (rcur->type == NOT_INIT)
7093 switch (rold->type) {
7095 if (rcur->type == SCALAR_VALUE) {
7096 if (!rold->precise && !rcur->precise)
7098 /* new val must satisfy old val knowledge */
7099 return range_within(rold, rcur) &&
7100 tnum_in(rold->var_off, rcur->var_off);
7102 /* We're trying to use a pointer in place of a scalar.
7103 * Even if the scalar was unbounded, this could lead to
7104 * pointer leaks because scalars are allowed to leak
7105 * while pointers are not. We could make this safe in
7106 * special cases if root is calling us, but it's
7107 * probably not worth the hassle.
7111 case PTR_TO_MAP_VALUE:
7112 /* If the new min/max/var_off satisfy the old ones and
7113 * everything else matches, we are OK.
7114 * 'id' is not compared, since it's only used for maps with
7115 * bpf_spin_lock inside map element and in such cases if
7116 * the rest of the prog is valid for one map element then
7117 * it's valid for all map elements regardless of the key
7118 * used in bpf_map_lookup()
7120 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
7121 range_within(rold, rcur) &&
7122 tnum_in(rold->var_off, rcur->var_off);
7123 case PTR_TO_MAP_VALUE_OR_NULL:
7124 /* a PTR_TO_MAP_VALUE could be safe to use as a
7125 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
7126 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
7127 * checked, doing so could have affected others with the same
7128 * id, and we can't check for that because we lost the id when
7129 * we converted to a PTR_TO_MAP_VALUE.
7131 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
7133 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
7135 /* Check our ids match any regs they're supposed to */
7136 return check_ids(rold->id, rcur->id, idmap);
7137 case PTR_TO_PACKET_META:
7139 if (rcur->type != rold->type)
7141 /* We must have at least as much range as the old ptr
7142 * did, so that any accesses which were safe before are
7143 * still safe. This is true even if old range < old off,
7144 * since someone could have accessed through (ptr - k), or
7145 * even done ptr -= k in a register, to get a safe access.
7147 if (rold->range > rcur->range)
7149 /* If the offsets don't match, we can't trust our alignment;
7150 * nor can we be sure that we won't fall out of range.
7152 if (rold->off != rcur->off)
7154 /* id relations must be preserved */
7155 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
7157 /* new val must satisfy old val knowledge */
7158 return range_within(rold, rcur) &&
7159 tnum_in(rold->var_off, rcur->var_off);
7161 case CONST_PTR_TO_MAP:
7162 case PTR_TO_PACKET_END:
7163 case PTR_TO_FLOW_KEYS:
7165 case PTR_TO_SOCKET_OR_NULL:
7166 case PTR_TO_SOCK_COMMON:
7167 case PTR_TO_SOCK_COMMON_OR_NULL:
7168 case PTR_TO_TCP_SOCK:
7169 case PTR_TO_TCP_SOCK_OR_NULL:
7170 case PTR_TO_XDP_SOCK:
7171 /* Only valid matches are exact, which memcmp() above
7172 * would have accepted
7175 /* Don't know what's going on, just say it's not safe */
7179 /* Shouldn't get here; if we do, say it's not safe */
7184 static bool stacksafe(struct bpf_func_state *old,
7185 struct bpf_func_state *cur,
7186 struct idpair *idmap)
7190 /* walk slots of the explored stack and ignore any additional
7191 * slots in the current stack, since explored(safe) state
7194 for (i = 0; i < old->allocated_stack; i++) {
7195 spi = i / BPF_REG_SIZE;
7197 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
7198 i += BPF_REG_SIZE - 1;
7199 /* explored state didn't use this */
7203 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
7206 /* explored stack has more populated slots than current stack
7207 * and these slots were used
7209 if (i >= cur->allocated_stack)
7212 /* if old state was safe with misc data in the stack
7213 * it will be safe with zero-initialized stack.
7214 * The opposite is not true
7216 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
7217 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
7219 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
7220 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
7221 /* Ex: old explored (safe) state has STACK_SPILL in
7222 * this stack slot, but current has has STACK_MISC ->
7223 * this verifier states are not equivalent,
7224 * return false to continue verification of this path
7227 if (i % BPF_REG_SIZE)
7229 if (old->stack[spi].slot_type[0] != STACK_SPILL)
7231 if (!regsafe(&old->stack[spi].spilled_ptr,
7232 &cur->stack[spi].spilled_ptr,
7234 /* when explored and current stack slot are both storing
7235 * spilled registers, check that stored pointers types
7236 * are the same as well.
7237 * Ex: explored safe path could have stored
7238 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
7239 * but current path has stored:
7240 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
7241 * such verifier states are not equivalent.
7242 * return false to continue verification of this path
7249 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
7251 if (old->acquired_refs != cur->acquired_refs)
7253 return !memcmp(old->refs, cur->refs,
7254 sizeof(*old->refs) * old->acquired_refs);
7257 /* compare two verifier states
7259 * all states stored in state_list are known to be valid, since
7260 * verifier reached 'bpf_exit' instruction through them
7262 * this function is called when verifier exploring different branches of
7263 * execution popped from the state stack. If it sees an old state that has
7264 * more strict register state and more strict stack state then this execution
7265 * branch doesn't need to be explored further, since verifier already
7266 * concluded that more strict state leads to valid finish.
7268 * Therefore two states are equivalent if register state is more conservative
7269 * and explored stack state is more conservative than the current one.
7272 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
7273 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
7275 * In other words if current stack state (one being explored) has more
7276 * valid slots than old one that already passed validation, it means
7277 * the verifier can stop exploring and conclude that current state is valid too
7279 * Similarly with registers. If explored state has register type as invalid
7280 * whereas register type in current state is meaningful, it means that
7281 * the current state will reach 'bpf_exit' instruction safely
7283 static bool func_states_equal(struct bpf_func_state *old,
7284 struct bpf_func_state *cur)
7286 struct idpair *idmap;
7290 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
7291 /* If we failed to allocate the idmap, just say it's not safe */
7295 for (i = 0; i < MAX_BPF_REG; i++) {
7296 if (!regsafe(&old->regs[i], &cur->regs[i], idmap))
7300 if (!stacksafe(old, cur, idmap))
7303 if (!refsafe(old, cur))
7311 static bool states_equal(struct bpf_verifier_env *env,
7312 struct bpf_verifier_state *old,
7313 struct bpf_verifier_state *cur)
7317 if (old->curframe != cur->curframe)
7320 /* Verification state from speculative execution simulation
7321 * must never prune a non-speculative execution one.
7323 if (old->speculative && !cur->speculative)
7326 if (old->active_spin_lock != cur->active_spin_lock)
7329 /* for states to be equal callsites have to be the same
7330 * and all frame states need to be equivalent
7332 for (i = 0; i <= old->curframe; i++) {
7333 if (old->frame[i]->callsite != cur->frame[i]->callsite)
7335 if (!func_states_equal(old->frame[i], cur->frame[i]))
7341 /* Return 0 if no propagation happened. Return negative error code if error
7342 * happened. Otherwise, return the propagated bit.
7344 static int propagate_liveness_reg(struct bpf_verifier_env *env,
7345 struct bpf_reg_state *reg,
7346 struct bpf_reg_state *parent_reg)
7348 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
7349 u8 flag = reg->live & REG_LIVE_READ;
7352 /* When comes here, read flags of PARENT_REG or REG could be any of
7353 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
7354 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
7356 if (parent_flag == REG_LIVE_READ64 ||
7357 /* Or if there is no read flag from REG. */
7359 /* Or if the read flag from REG is the same as PARENT_REG. */
7360 parent_flag == flag)
7363 err = mark_reg_read(env, reg, parent_reg, flag);
7370 /* A write screens off any subsequent reads; but write marks come from the
7371 * straight-line code between a state and its parent. When we arrive at an
7372 * equivalent state (jump target or such) we didn't arrive by the straight-line
7373 * code, so read marks in the state must propagate to the parent regardless
7374 * of the state's write marks. That's what 'parent == state->parent' comparison
7375 * in mark_reg_read() is for.
7377 static int propagate_liveness(struct bpf_verifier_env *env,
7378 const struct bpf_verifier_state *vstate,
7379 struct bpf_verifier_state *vparent)
7381 struct bpf_reg_state *state_reg, *parent_reg;
7382 struct bpf_func_state *state, *parent;
7383 int i, frame, err = 0;
7385 if (vparent->curframe != vstate->curframe) {
7386 WARN(1, "propagate_live: parent frame %d current frame %d\n",
7387 vparent->curframe, vstate->curframe);
7390 /* Propagate read liveness of registers... */
7391 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
7392 for (frame = 0; frame <= vstate->curframe; frame++) {
7393 parent = vparent->frame[frame];
7394 state = vstate->frame[frame];
7395 parent_reg = parent->regs;
7396 state_reg = state->regs;
7397 /* We don't need to worry about FP liveness, it's read-only */
7398 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
7399 err = propagate_liveness_reg(env, &state_reg[i],
7403 if (err == REG_LIVE_READ64)
7404 mark_insn_zext(env, &parent_reg[i]);
7407 /* Propagate stack slots. */
7408 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
7409 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
7410 parent_reg = &parent->stack[i].spilled_ptr;
7411 state_reg = &state->stack[i].spilled_ptr;
7412 err = propagate_liveness_reg(env, state_reg,
7421 /* find precise scalars in the previous equivalent state and
7422 * propagate them into the current state
7424 static int propagate_precision(struct bpf_verifier_env *env,
7425 const struct bpf_verifier_state *old)
7427 struct bpf_reg_state *state_reg;
7428 struct bpf_func_state *state;
7431 state = old->frame[old->curframe];
7432 state_reg = state->regs;
7433 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
7434 if (state_reg->type != SCALAR_VALUE ||
7435 !state_reg->precise)
7437 if (env->log.level & BPF_LOG_LEVEL2)
7438 verbose(env, "propagating r%d\n", i);
7439 err = mark_chain_precision(env, i);
7444 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
7445 if (state->stack[i].slot_type[0] != STACK_SPILL)
7447 state_reg = &state->stack[i].spilled_ptr;
7448 if (state_reg->type != SCALAR_VALUE ||
7449 !state_reg->precise)
7451 if (env->log.level & BPF_LOG_LEVEL2)
7452 verbose(env, "propagating fp%d\n",
7453 (-i - 1) * BPF_REG_SIZE);
7454 err = mark_chain_precision_stack(env, i);
7461 static bool states_maybe_looping(struct bpf_verifier_state *old,
7462 struct bpf_verifier_state *cur)
7464 struct bpf_func_state *fold, *fcur;
7465 int i, fr = cur->curframe;
7467 if (old->curframe != fr)
7470 fold = old->frame[fr];
7471 fcur = cur->frame[fr];
7472 for (i = 0; i < MAX_BPF_REG; i++)
7473 if (memcmp(&fold->regs[i], &fcur->regs[i],
7474 offsetof(struct bpf_reg_state, parent)))
7480 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
7482 struct bpf_verifier_state_list *new_sl;
7483 struct bpf_verifier_state_list *sl, **pprev;
7484 struct bpf_verifier_state *cur = env->cur_state, *new;
7485 int i, j, err, states_cnt = 0;
7486 bool add_new_state = env->test_state_freq ? true : false;
7488 cur->last_insn_idx = env->prev_insn_idx;
7489 if (!env->insn_aux_data[insn_idx].prune_point)
7490 /* this 'insn_idx' instruction wasn't marked, so we will not
7491 * be doing state search here
7495 /* bpf progs typically have pruning point every 4 instructions
7496 * http://vger.kernel.org/bpfconf2019.html#session-1
7497 * Do not add new state for future pruning if the verifier hasn't seen
7498 * at least 2 jumps and at least 8 instructions.
7499 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
7500 * In tests that amounts to up to 50% reduction into total verifier
7501 * memory consumption and 20% verifier time speedup.
7503 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
7504 env->insn_processed - env->prev_insn_processed >= 8)
7505 add_new_state = true;
7507 pprev = explored_state(env, insn_idx);
7510 clean_live_states(env, insn_idx, cur);
7514 if (sl->state.insn_idx != insn_idx)
7516 if (sl->state.branches) {
7517 if (states_maybe_looping(&sl->state, cur) &&
7518 states_equal(env, &sl->state, cur)) {
7519 verbose_linfo(env, insn_idx, "; ");
7520 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
7523 /* if the verifier is processing a loop, avoid adding new state
7524 * too often, since different loop iterations have distinct
7525 * states and may not help future pruning.
7526 * This threshold shouldn't be too low to make sure that
7527 * a loop with large bound will be rejected quickly.
7528 * The most abusive loop will be:
7530 * if r1 < 1000000 goto pc-2
7531 * 1M insn_procssed limit / 100 == 10k peak states.
7532 * This threshold shouldn't be too high either, since states
7533 * at the end of the loop are likely to be useful in pruning.
7535 if (env->jmps_processed - env->prev_jmps_processed < 20 &&
7536 env->insn_processed - env->prev_insn_processed < 100)
7537 add_new_state = false;
7540 if (states_equal(env, &sl->state, cur)) {
7542 /* reached equivalent register/stack state,
7544 * Registers read by the continuation are read by us.
7545 * If we have any write marks in env->cur_state, they
7546 * will prevent corresponding reads in the continuation
7547 * from reaching our parent (an explored_state). Our
7548 * own state will get the read marks recorded, but
7549 * they'll be immediately forgotten as we're pruning
7550 * this state and will pop a new one.
7552 err = propagate_liveness(env, &sl->state, cur);
7554 /* if previous state reached the exit with precision and
7555 * current state is equivalent to it (except precsion marks)
7556 * the precision needs to be propagated back in
7557 * the current state.
7559 err = err ? : push_jmp_history(env, cur);
7560 err = err ? : propagate_precision(env, &sl->state);
7566 /* when new state is not going to be added do not increase miss count.
7567 * Otherwise several loop iterations will remove the state
7568 * recorded earlier. The goal of these heuristics is to have
7569 * states from some iterations of the loop (some in the beginning
7570 * and some at the end) to help pruning.
7574 /* heuristic to determine whether this state is beneficial
7575 * to keep checking from state equivalence point of view.
7576 * Higher numbers increase max_states_per_insn and verification time,
7577 * but do not meaningfully decrease insn_processed.
7579 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
7580 /* the state is unlikely to be useful. Remove it to
7581 * speed up verification
7584 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
7585 u32 br = sl->state.branches;
7588 "BUG live_done but branches_to_explore %d\n",
7590 free_verifier_state(&sl->state, false);
7594 /* cannot free this state, since parentage chain may
7595 * walk it later. Add it for free_list instead to
7596 * be freed at the end of verification
7598 sl->next = env->free_list;
7599 env->free_list = sl;
7609 if (env->max_states_per_insn < states_cnt)
7610 env->max_states_per_insn = states_cnt;
7612 if (!env->allow_ptr_leaks && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
7613 return push_jmp_history(env, cur);
7616 return push_jmp_history(env, cur);
7618 /* There were no equivalent states, remember the current one.
7619 * Technically the current state is not proven to be safe yet,
7620 * but it will either reach outer most bpf_exit (which means it's safe)
7621 * or it will be rejected. When there are no loops the verifier won't be
7622 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
7623 * again on the way to bpf_exit.
7624 * When looping the sl->state.branches will be > 0 and this state
7625 * will not be considered for equivalence until branches == 0.
7627 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
7630 env->total_states++;
7632 env->prev_jmps_processed = env->jmps_processed;
7633 env->prev_insn_processed = env->insn_processed;
7635 /* add new state to the head of linked list */
7636 new = &new_sl->state;
7637 err = copy_verifier_state(new, cur);
7639 free_verifier_state(new, false);
7643 new->insn_idx = insn_idx;
7644 WARN_ONCE(new->branches != 1,
7645 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
7648 cur->first_insn_idx = insn_idx;
7649 clear_jmp_history(cur);
7650 new_sl->next = *explored_state(env, insn_idx);
7651 *explored_state(env, insn_idx) = new_sl;
7652 /* connect new state to parentage chain. Current frame needs all
7653 * registers connected. Only r6 - r9 of the callers are alive (pushed
7654 * to the stack implicitly by JITs) so in callers' frames connect just
7655 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
7656 * the state of the call instruction (with WRITTEN set), and r0 comes
7657 * from callee with its full parentage chain, anyway.
7659 /* clear write marks in current state: the writes we did are not writes
7660 * our child did, so they don't screen off its reads from us.
7661 * (There are no read marks in current state, because reads always mark
7662 * their parent and current state never has children yet. Only
7663 * explored_states can get read marks.)
7665 for (j = 0; j <= cur->curframe; j++) {
7666 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
7667 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
7668 for (i = 0; i < BPF_REG_FP; i++)
7669 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
7672 /* all stack frames are accessible from callee, clear them all */
7673 for (j = 0; j <= cur->curframe; j++) {
7674 struct bpf_func_state *frame = cur->frame[j];
7675 struct bpf_func_state *newframe = new->frame[j];
7677 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
7678 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
7679 frame->stack[i].spilled_ptr.parent =
7680 &newframe->stack[i].spilled_ptr;
7686 /* Return true if it's OK to have the same insn return a different type. */
7687 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
7692 case PTR_TO_SOCKET_OR_NULL:
7693 case PTR_TO_SOCK_COMMON:
7694 case PTR_TO_SOCK_COMMON_OR_NULL:
7695 case PTR_TO_TCP_SOCK:
7696 case PTR_TO_TCP_SOCK_OR_NULL:
7697 case PTR_TO_XDP_SOCK:
7705 /* If an instruction was previously used with particular pointer types, then we
7706 * need to be careful to avoid cases such as the below, where it may be ok
7707 * for one branch accessing the pointer, but not ok for the other branch:
7712 * R1 = some_other_valid_ptr;
7715 * R2 = *(u32 *)(R1 + 0);
7717 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
7719 return src != prev && (!reg_type_mismatch_ok(src) ||
7720 !reg_type_mismatch_ok(prev));
7723 static int do_check(struct bpf_verifier_env *env)
7725 struct bpf_verifier_state *state;
7726 struct bpf_insn *insns = env->prog->insnsi;
7727 struct bpf_reg_state *regs;
7728 int insn_cnt = env->prog->len;
7729 bool do_print_state = false;
7730 int prev_insn_idx = -1;
7732 env->prev_linfo = NULL;
7734 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
7737 state->curframe = 0;
7738 state->speculative = false;
7739 state->branches = 1;
7740 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
7741 if (!state->frame[0]) {
7745 env->cur_state = state;
7746 init_func_state(env, state->frame[0],
7747 BPF_MAIN_FUNC /* callsite */,
7749 0 /* subprogno, zero == main subprog */);
7751 if (btf_check_func_arg_match(env, 0))
7755 struct bpf_insn *insn;
7759 env->prev_insn_idx = prev_insn_idx;
7760 if (env->insn_idx >= insn_cnt) {
7761 verbose(env, "invalid insn idx %d insn_cnt %d\n",
7762 env->insn_idx, insn_cnt);
7766 insn = &insns[env->insn_idx];
7767 class = BPF_CLASS(insn->code);
7769 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
7771 "BPF program is too large. Processed %d insn\n",
7772 env->insn_processed);
7776 err = is_state_visited(env, env->insn_idx);
7780 /* found equivalent state, can prune the search */
7781 if (env->log.level & BPF_LOG_LEVEL) {
7783 verbose(env, "\nfrom %d to %d%s: safe\n",
7784 env->prev_insn_idx, env->insn_idx,
7785 env->cur_state->speculative ?
7786 " (speculative execution)" : "");
7788 verbose(env, "%d: safe\n", env->insn_idx);
7790 goto process_bpf_exit;
7793 if (signal_pending(current))
7799 if (env->log.level & BPF_LOG_LEVEL2 ||
7800 (env->log.level & BPF_LOG_LEVEL && do_print_state)) {
7801 if (env->log.level & BPF_LOG_LEVEL2)
7802 verbose(env, "%d:", env->insn_idx);
7804 verbose(env, "\nfrom %d to %d%s:",
7805 env->prev_insn_idx, env->insn_idx,
7806 env->cur_state->speculative ?
7807 " (speculative execution)" : "");
7808 print_verifier_state(env, state->frame[state->curframe]);
7809 do_print_state = false;
7812 if (env->log.level & BPF_LOG_LEVEL) {
7813 const struct bpf_insn_cbs cbs = {
7814 .cb_print = verbose,
7815 .private_data = env,
7818 verbose_linfo(env, env->insn_idx, "; ");
7819 verbose(env, "%d: ", env->insn_idx);
7820 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
7823 if (bpf_prog_is_dev_bound(env->prog->aux)) {
7824 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
7825 env->prev_insn_idx);
7830 regs = cur_regs(env);
7831 env->insn_aux_data[env->insn_idx].seen = true;
7832 prev_insn_idx = env->insn_idx;
7834 if (class == BPF_ALU || class == BPF_ALU64) {
7835 err = check_alu_op(env, insn);
7839 } else if (class == BPF_LDX) {
7840 enum bpf_reg_type *prev_src_type, src_reg_type;
7842 /* check for reserved fields is already done */
7844 /* check src operand */
7845 err = check_reg_arg(env, insn->src_reg, SRC_OP);
7849 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
7853 src_reg_type = regs[insn->src_reg].type;
7855 /* check that memory (src_reg + off) is readable,
7856 * the state of dst_reg will be updated by this func
7858 err = check_mem_access(env, env->insn_idx, insn->src_reg,
7859 insn->off, BPF_SIZE(insn->code),
7860 BPF_READ, insn->dst_reg, false);
7864 prev_src_type = &env->insn_aux_data[env->insn_idx].ptr_type;
7866 if (*prev_src_type == NOT_INIT) {
7868 * dst_reg = *(u32 *)(src_reg + off)
7869 * save type to validate intersecting paths
7871 *prev_src_type = src_reg_type;
7873 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
7874 /* ABuser program is trying to use the same insn
7875 * dst_reg = *(u32*) (src_reg + off)
7876 * with different pointer types:
7877 * src_reg == ctx in one branch and
7878 * src_reg == stack|map in some other branch.
7881 verbose(env, "same insn cannot be used with different pointers\n");
7885 } else if (class == BPF_STX) {
7886 enum bpf_reg_type *prev_dst_type, dst_reg_type;
7888 if (BPF_MODE(insn->code) == BPF_XADD) {
7889 err = check_xadd(env, env->insn_idx, insn);
7896 /* check src1 operand */
7897 err = check_reg_arg(env, insn->src_reg, SRC_OP);
7900 /* check src2 operand */
7901 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
7905 dst_reg_type = regs[insn->dst_reg].type;
7907 /* check that memory (dst_reg + off) is writeable */
7908 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
7909 insn->off, BPF_SIZE(insn->code),
7910 BPF_WRITE, insn->src_reg, false);
7914 prev_dst_type = &env->insn_aux_data[env->insn_idx].ptr_type;
7916 if (*prev_dst_type == NOT_INIT) {
7917 *prev_dst_type = dst_reg_type;
7918 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
7919 verbose(env, "same insn cannot be used with different pointers\n");
7923 } else if (class == BPF_ST) {
7924 if (BPF_MODE(insn->code) != BPF_MEM ||
7925 insn->src_reg != BPF_REG_0) {
7926 verbose(env, "BPF_ST uses reserved fields\n");
7929 /* check src operand */
7930 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
7934 if (is_ctx_reg(env, insn->dst_reg)) {
7935 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
7937 reg_type_str[reg_state(env, insn->dst_reg)->type]);
7941 /* check that memory (dst_reg + off) is writeable */
7942 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
7943 insn->off, BPF_SIZE(insn->code),
7944 BPF_WRITE, -1, false);
7948 } else if (class == BPF_JMP || class == BPF_JMP32) {
7949 u8 opcode = BPF_OP(insn->code);
7951 env->jmps_processed++;
7952 if (opcode == BPF_CALL) {
7953 if (BPF_SRC(insn->code) != BPF_K ||
7955 (insn->src_reg != BPF_REG_0 &&
7956 insn->src_reg != BPF_PSEUDO_CALL) ||
7957 insn->dst_reg != BPF_REG_0 ||
7958 class == BPF_JMP32) {
7959 verbose(env, "BPF_CALL uses reserved fields\n");
7963 if (env->cur_state->active_spin_lock &&
7964 (insn->src_reg == BPF_PSEUDO_CALL ||
7965 insn->imm != BPF_FUNC_spin_unlock)) {
7966 verbose(env, "function calls are not allowed while holding a lock\n");
7969 if (insn->src_reg == BPF_PSEUDO_CALL)
7970 err = check_func_call(env, insn, &env->insn_idx);
7972 err = check_helper_call(env, insn->imm, env->insn_idx);
7976 } else if (opcode == BPF_JA) {
7977 if (BPF_SRC(insn->code) != BPF_K ||
7979 insn->src_reg != BPF_REG_0 ||
7980 insn->dst_reg != BPF_REG_0 ||
7981 class == BPF_JMP32) {
7982 verbose(env, "BPF_JA uses reserved fields\n");
7986 env->insn_idx += insn->off + 1;
7989 } else if (opcode == BPF_EXIT) {
7990 if (BPF_SRC(insn->code) != BPF_K ||
7992 insn->src_reg != BPF_REG_0 ||
7993 insn->dst_reg != BPF_REG_0 ||
7994 class == BPF_JMP32) {
7995 verbose(env, "BPF_EXIT uses reserved fields\n");
7999 if (env->cur_state->active_spin_lock) {
8000 verbose(env, "bpf_spin_unlock is missing\n");
8004 if (state->curframe) {
8005 /* exit from nested function */
8006 err = prepare_func_exit(env, &env->insn_idx);
8009 do_print_state = true;
8013 err = check_reference_leak(env);
8017 /* eBPF calling convetion is such that R0 is used
8018 * to return the value from eBPF program.
8019 * Make sure that it's readable at this time
8020 * of bpf_exit, which means that program wrote
8021 * something into it earlier
8023 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
8027 if (is_pointer_value(env, BPF_REG_0)) {
8028 verbose(env, "R0 leaks addr as return value\n");
8032 err = check_return_code(env);
8036 update_branch_counts(env, env->cur_state);
8037 err = pop_stack(env, &prev_insn_idx,
8044 do_print_state = true;
8048 err = check_cond_jmp_op(env, insn, &env->insn_idx);
8052 } else if (class == BPF_LD) {
8053 u8 mode = BPF_MODE(insn->code);
8055 if (mode == BPF_ABS || mode == BPF_IND) {
8056 err = check_ld_abs(env, insn);
8060 } else if (mode == BPF_IMM) {
8061 err = check_ld_imm(env, insn);
8066 env->insn_aux_data[env->insn_idx].seen = true;
8068 verbose(env, "invalid BPF_LD mode\n");
8072 verbose(env, "unknown insn class %d\n", class);
8079 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
8083 static int check_map_prealloc(struct bpf_map *map)
8085 return (map->map_type != BPF_MAP_TYPE_HASH &&
8086 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
8087 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
8088 !(map->map_flags & BPF_F_NO_PREALLOC);
8091 static bool is_tracing_prog_type(enum bpf_prog_type type)
8094 case BPF_PROG_TYPE_KPROBE:
8095 case BPF_PROG_TYPE_TRACEPOINT:
8096 case BPF_PROG_TYPE_PERF_EVENT:
8097 case BPF_PROG_TYPE_RAW_TRACEPOINT:
8104 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
8105 struct bpf_map *map,
8106 struct bpf_prog *prog)
8109 /* Make sure that BPF_PROG_TYPE_PERF_EVENT programs only use
8110 * preallocated hash maps, since doing memory allocation
8111 * in overflow_handler can crash depending on where nmi got
8114 if (prog->type == BPF_PROG_TYPE_PERF_EVENT) {
8115 if (!check_map_prealloc(map)) {
8116 verbose(env, "perf_event programs can only use preallocated hash map\n");
8119 if (map->inner_map_meta &&
8120 !check_map_prealloc(map->inner_map_meta)) {
8121 verbose(env, "perf_event programs can only use preallocated inner hash map\n");
8126 if ((is_tracing_prog_type(prog->type) ||
8127 prog->type == BPF_PROG_TYPE_SOCKET_FILTER) &&
8128 map_value_has_spin_lock(map)) {
8129 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
8133 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
8134 !bpf_offload_prog_map_match(prog, map)) {
8135 verbose(env, "offload device mismatch between prog and map\n");
8142 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
8144 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
8145 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
8148 /* look for pseudo eBPF instructions that access map FDs and
8149 * replace them with actual map pointers
8151 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env *env)
8153 struct bpf_insn *insn = env->prog->insnsi;
8154 int insn_cnt = env->prog->len;
8157 err = bpf_prog_calc_tag(env->prog);
8161 for (i = 0; i < insn_cnt; i++, insn++) {
8162 if (BPF_CLASS(insn->code) == BPF_LDX &&
8163 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
8164 verbose(env, "BPF_LDX uses reserved fields\n");
8168 if (BPF_CLASS(insn->code) == BPF_STX &&
8169 ((BPF_MODE(insn->code) != BPF_MEM &&
8170 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
8171 verbose(env, "BPF_STX uses reserved fields\n");
8175 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
8176 struct bpf_insn_aux_data *aux;
8177 struct bpf_map *map;
8181 if (i == insn_cnt - 1 || insn[1].code != 0 ||
8182 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
8184 verbose(env, "invalid bpf_ld_imm64 insn\n");
8188 if (insn[0].src_reg == 0)
8189 /* valid generic load 64-bit imm */
8192 /* In final convert_pseudo_ld_imm64() step, this is
8193 * converted into regular 64-bit imm load insn.
8195 if ((insn[0].src_reg != BPF_PSEUDO_MAP_FD &&
8196 insn[0].src_reg != BPF_PSEUDO_MAP_VALUE) ||
8197 (insn[0].src_reg == BPF_PSEUDO_MAP_FD &&
8198 insn[1].imm != 0)) {
8200 "unrecognized bpf_ld_imm64 insn\n");
8204 f = fdget(insn[0].imm);
8205 map = __bpf_map_get(f);
8207 verbose(env, "fd %d is not pointing to valid bpf_map\n",
8209 return PTR_ERR(map);
8212 err = check_map_prog_compatibility(env, map, env->prog);
8218 aux = &env->insn_aux_data[i];
8219 if (insn->src_reg == BPF_PSEUDO_MAP_FD) {
8220 addr = (unsigned long)map;
8222 u32 off = insn[1].imm;
8224 if (off >= BPF_MAX_VAR_OFF) {
8225 verbose(env, "direct value offset of %u is not allowed\n", off);
8230 if (!map->ops->map_direct_value_addr) {
8231 verbose(env, "no direct value access support for this map type\n");
8236 err = map->ops->map_direct_value_addr(map, &addr, off);
8238 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
8239 map->value_size, off);
8248 insn[0].imm = (u32)addr;
8249 insn[1].imm = addr >> 32;
8251 /* check whether we recorded this map already */
8252 for (j = 0; j < env->used_map_cnt; j++) {
8253 if (env->used_maps[j] == map) {
8260 if (env->used_map_cnt >= MAX_USED_MAPS) {
8265 /* hold the map. If the program is rejected by verifier,
8266 * the map will be released by release_maps() or it
8267 * will be used by the valid program until it's unloaded
8268 * and all maps are released in free_used_maps()
8272 aux->map_index = env->used_map_cnt;
8273 env->used_maps[env->used_map_cnt++] = map;
8275 if (bpf_map_is_cgroup_storage(map) &&
8276 bpf_cgroup_storage_assign(env->prog->aux, map)) {
8277 verbose(env, "only one cgroup storage of each type is allowed\n");
8289 /* Basic sanity check before we invest more work here. */
8290 if (!bpf_opcode_in_insntable(insn->code)) {
8291 verbose(env, "unknown opcode %02x\n", insn->code);
8296 /* now all pseudo BPF_LD_IMM64 instructions load valid
8297 * 'struct bpf_map *' into a register instead of user map_fd.
8298 * These pointers will be used later by verifier to validate map access.
8303 /* drop refcnt of maps used by the rejected program */
8304 static void release_maps(struct bpf_verifier_env *env)
8306 __bpf_free_used_maps(env->prog->aux, env->used_maps,
8310 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
8311 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
8313 struct bpf_insn *insn = env->prog->insnsi;
8314 int insn_cnt = env->prog->len;
8317 for (i = 0; i < insn_cnt; i++, insn++)
8318 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
8322 /* single env->prog->insni[off] instruction was replaced with the range
8323 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
8324 * [0, off) and [off, end) to new locations, so the patched range stays zero
8326 static int adjust_insn_aux_data(struct bpf_verifier_env *env,
8327 struct bpf_prog *new_prog, u32 off, u32 cnt)
8329 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
8330 struct bpf_insn *insn = new_prog->insnsi;
8334 /* aux info at OFF always needs adjustment, no matter fast path
8335 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
8336 * original insn at old prog.
8338 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
8342 prog_len = new_prog->len;
8343 new_data = vzalloc(array_size(prog_len,
8344 sizeof(struct bpf_insn_aux_data)));
8347 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
8348 memcpy(new_data + off + cnt - 1, old_data + off,
8349 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
8350 for (i = off; i < off + cnt - 1; i++) {
8351 new_data[i].seen = true;
8352 new_data[i].zext_dst = insn_has_def32(env, insn + i);
8354 env->insn_aux_data = new_data;
8359 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
8365 /* NOTE: fake 'exit' subprog should be updated as well. */
8366 for (i = 0; i <= env->subprog_cnt; i++) {
8367 if (env->subprog_info[i].start <= off)
8369 env->subprog_info[i].start += len - 1;
8373 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
8374 const struct bpf_insn *patch, u32 len)
8376 struct bpf_prog *new_prog;
8378 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
8379 if (IS_ERR(new_prog)) {
8380 if (PTR_ERR(new_prog) == -ERANGE)
8382 "insn %d cannot be patched due to 16-bit range\n",
8383 env->insn_aux_data[off].orig_idx);
8386 if (adjust_insn_aux_data(env, new_prog, off, len))
8388 adjust_subprog_starts(env, off, len);
8392 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
8397 /* find first prog starting at or after off (first to remove) */
8398 for (i = 0; i < env->subprog_cnt; i++)
8399 if (env->subprog_info[i].start >= off)
8401 /* find first prog starting at or after off + cnt (first to stay) */
8402 for (j = i; j < env->subprog_cnt; j++)
8403 if (env->subprog_info[j].start >= off + cnt)
8405 /* if j doesn't start exactly at off + cnt, we are just removing
8406 * the front of previous prog
8408 if (env->subprog_info[j].start != off + cnt)
8412 struct bpf_prog_aux *aux = env->prog->aux;
8415 /* move fake 'exit' subprog as well */
8416 move = env->subprog_cnt + 1 - j;
8418 memmove(env->subprog_info + i,
8419 env->subprog_info + j,
8420 sizeof(*env->subprog_info) * move);
8421 env->subprog_cnt -= j - i;
8423 /* remove func_info */
8424 if (aux->func_info) {
8425 move = aux->func_info_cnt - j;
8427 memmove(aux->func_info + i,
8429 sizeof(*aux->func_info) * move);
8430 aux->func_info_cnt -= j - i;
8431 /* func_info->insn_off is set after all code rewrites,
8432 * in adjust_btf_func() - no need to adjust
8436 /* convert i from "first prog to remove" to "first to adjust" */
8437 if (env->subprog_info[i].start == off)
8441 /* update fake 'exit' subprog as well */
8442 for (; i <= env->subprog_cnt; i++)
8443 env->subprog_info[i].start -= cnt;
8448 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
8451 struct bpf_prog *prog = env->prog;
8452 u32 i, l_off, l_cnt, nr_linfo;
8453 struct bpf_line_info *linfo;
8455 nr_linfo = prog->aux->nr_linfo;
8459 linfo = prog->aux->linfo;
8461 /* find first line info to remove, count lines to be removed */
8462 for (i = 0; i < nr_linfo; i++)
8463 if (linfo[i].insn_off >= off)
8468 for (; i < nr_linfo; i++)
8469 if (linfo[i].insn_off < off + cnt)
8474 /* First live insn doesn't match first live linfo, it needs to "inherit"
8475 * last removed linfo. prog is already modified, so prog->len == off
8476 * means no live instructions after (tail of the program was removed).
8478 if (prog->len != off && l_cnt &&
8479 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
8481 linfo[--i].insn_off = off + cnt;
8484 /* remove the line info which refer to the removed instructions */
8486 memmove(linfo + l_off, linfo + i,
8487 sizeof(*linfo) * (nr_linfo - i));
8489 prog->aux->nr_linfo -= l_cnt;
8490 nr_linfo = prog->aux->nr_linfo;
8493 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
8494 for (i = l_off; i < nr_linfo; i++)
8495 linfo[i].insn_off -= cnt;
8497 /* fix up all subprogs (incl. 'exit') which start >= off */
8498 for (i = 0; i <= env->subprog_cnt; i++)
8499 if (env->subprog_info[i].linfo_idx > l_off) {
8500 /* program may have started in the removed region but
8501 * may not be fully removed
8503 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
8504 env->subprog_info[i].linfo_idx -= l_cnt;
8506 env->subprog_info[i].linfo_idx = l_off;
8512 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
8514 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
8515 unsigned int orig_prog_len = env->prog->len;
8518 if (bpf_prog_is_dev_bound(env->prog->aux))
8519 bpf_prog_offload_remove_insns(env, off, cnt);
8521 err = bpf_remove_insns(env->prog, off, cnt);
8525 err = adjust_subprog_starts_after_remove(env, off, cnt);
8529 err = bpf_adj_linfo_after_remove(env, off, cnt);
8533 memmove(aux_data + off, aux_data + off + cnt,
8534 sizeof(*aux_data) * (orig_prog_len - off - cnt));
8539 /* The verifier does more data flow analysis than llvm and will not
8540 * explore branches that are dead at run time. Malicious programs can
8541 * have dead code too. Therefore replace all dead at-run-time code
8544 * Just nops are not optimal, e.g. if they would sit at the end of the
8545 * program and through another bug we would manage to jump there, then
8546 * we'd execute beyond program memory otherwise. Returning exception
8547 * code also wouldn't work since we can have subprogs where the dead
8548 * code could be located.
8550 static void sanitize_dead_code(struct bpf_verifier_env *env)
8552 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
8553 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
8554 struct bpf_insn *insn = env->prog->insnsi;
8555 const int insn_cnt = env->prog->len;
8558 for (i = 0; i < insn_cnt; i++) {
8559 if (aux_data[i].seen)
8561 memcpy(insn + i, &trap, sizeof(trap));
8565 static bool insn_is_cond_jump(u8 code)
8569 if (BPF_CLASS(code) == BPF_JMP32)
8572 if (BPF_CLASS(code) != BPF_JMP)
8576 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
8579 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
8581 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
8582 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
8583 struct bpf_insn *insn = env->prog->insnsi;
8584 const int insn_cnt = env->prog->len;
8587 for (i = 0; i < insn_cnt; i++, insn++) {
8588 if (!insn_is_cond_jump(insn->code))
8591 if (!aux_data[i + 1].seen)
8593 else if (!aux_data[i + 1 + insn->off].seen)
8598 if (bpf_prog_is_dev_bound(env->prog->aux))
8599 bpf_prog_offload_replace_insn(env, i, &ja);
8601 memcpy(insn, &ja, sizeof(ja));
8605 static int opt_remove_dead_code(struct bpf_verifier_env *env)
8607 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
8608 int insn_cnt = env->prog->len;
8611 for (i = 0; i < insn_cnt; i++) {
8615 while (i + j < insn_cnt && !aux_data[i + j].seen)
8620 err = verifier_remove_insns(env, i, j);
8623 insn_cnt = env->prog->len;
8629 static int opt_remove_nops(struct bpf_verifier_env *env)
8631 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
8632 struct bpf_insn *insn = env->prog->insnsi;
8633 int insn_cnt = env->prog->len;
8636 for (i = 0; i < insn_cnt; i++) {
8637 if (memcmp(&insn[i], &ja, sizeof(ja)))
8640 err = verifier_remove_insns(env, i, 1);
8650 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
8651 const union bpf_attr *attr)
8653 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
8654 struct bpf_insn_aux_data *aux = env->insn_aux_data;
8655 int i, patch_len, delta = 0, len = env->prog->len;
8656 struct bpf_insn *insns = env->prog->insnsi;
8657 struct bpf_prog *new_prog;
8660 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
8661 zext_patch[1] = BPF_ZEXT_REG(0);
8662 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
8663 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
8664 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
8665 for (i = 0; i < len; i++) {
8666 int adj_idx = i + delta;
8667 struct bpf_insn insn;
8669 insn = insns[adj_idx];
8670 if (!aux[adj_idx].zext_dst) {
8678 class = BPF_CLASS(code);
8679 if (insn_no_def(&insn))
8682 /* NOTE: arg "reg" (the fourth one) is only used for
8683 * BPF_STX which has been ruled out in above
8684 * check, it is safe to pass NULL here.
8686 if (is_reg64(env, &insn, insn.dst_reg, NULL, DST_OP)) {
8687 if (class == BPF_LD &&
8688 BPF_MODE(code) == BPF_IMM)
8693 /* ctx load could be transformed into wider load. */
8694 if (class == BPF_LDX &&
8695 aux[adj_idx].ptr_type == PTR_TO_CTX)
8698 imm_rnd = get_random_int();
8699 rnd_hi32_patch[0] = insn;
8700 rnd_hi32_patch[1].imm = imm_rnd;
8701 rnd_hi32_patch[3].dst_reg = insn.dst_reg;
8702 patch = rnd_hi32_patch;
8704 goto apply_patch_buffer;
8707 if (!bpf_jit_needs_zext())
8710 zext_patch[0] = insn;
8711 zext_patch[1].dst_reg = insn.dst_reg;
8712 zext_patch[1].src_reg = insn.dst_reg;
8716 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
8719 env->prog = new_prog;
8720 insns = new_prog->insnsi;
8721 aux = env->insn_aux_data;
8722 delta += patch_len - 1;
8728 /* convert load instructions that access fields of a context type into a
8729 * sequence of instructions that access fields of the underlying structure:
8730 * struct __sk_buff -> struct sk_buff
8731 * struct bpf_sock_ops -> struct sock
8733 static int convert_ctx_accesses(struct bpf_verifier_env *env)
8735 const struct bpf_verifier_ops *ops = env->ops;
8736 int i, cnt, size, ctx_field_size, delta = 0;
8737 const int insn_cnt = env->prog->len;
8738 struct bpf_insn insn_buf[16], *insn;
8739 u32 target_size, size_default, off;
8740 struct bpf_prog *new_prog;
8741 enum bpf_access_type type;
8742 bool is_narrower_load;
8744 if (ops->gen_prologue || env->seen_direct_write) {
8745 if (!ops->gen_prologue) {
8746 verbose(env, "bpf verifier is misconfigured\n");
8749 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
8751 if (cnt >= ARRAY_SIZE(insn_buf)) {
8752 verbose(env, "bpf verifier is misconfigured\n");
8755 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
8759 env->prog = new_prog;
8764 if (bpf_prog_is_dev_bound(env->prog->aux))
8767 insn = env->prog->insnsi + delta;
8769 for (i = 0; i < insn_cnt; i++, insn++) {
8770 bpf_convert_ctx_access_t convert_ctx_access;
8772 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
8773 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
8774 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
8775 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
8777 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
8778 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
8779 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
8780 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
8785 if (type == BPF_WRITE &&
8786 env->insn_aux_data[i + delta].sanitize_stack_off) {
8787 struct bpf_insn patch[] = {
8788 /* Sanitize suspicious stack slot with zero.
8789 * There are no memory dependencies for this store,
8790 * since it's only using frame pointer and immediate
8793 BPF_ST_MEM(BPF_DW, BPF_REG_FP,
8794 env->insn_aux_data[i + delta].sanitize_stack_off,
8796 /* the original STX instruction will immediately
8797 * overwrite the same stack slot with appropriate value
8802 cnt = ARRAY_SIZE(patch);
8803 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
8808 env->prog = new_prog;
8809 insn = new_prog->insnsi + i + delta;
8813 switch (env->insn_aux_data[i + delta].ptr_type) {
8815 if (!ops->convert_ctx_access)
8817 convert_ctx_access = ops->convert_ctx_access;
8820 case PTR_TO_SOCK_COMMON:
8821 convert_ctx_access = bpf_sock_convert_ctx_access;
8823 case PTR_TO_TCP_SOCK:
8824 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
8826 case PTR_TO_XDP_SOCK:
8827 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
8830 if (type == BPF_WRITE) {
8831 verbose(env, "Writes through BTF pointers are not allowed\n");
8834 insn->code = BPF_LDX | BPF_PROBE_MEM | BPF_SIZE((insn)->code);
8835 env->prog->aux->num_exentries++;
8841 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
8842 size = BPF_LDST_BYTES(insn);
8844 /* If the read access is a narrower load of the field,
8845 * convert to a 4/8-byte load, to minimum program type specific
8846 * convert_ctx_access changes. If conversion is successful,
8847 * we will apply proper mask to the result.
8849 is_narrower_load = size < ctx_field_size;
8850 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
8852 if (is_narrower_load) {
8855 if (type == BPF_WRITE) {
8856 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
8861 if (ctx_field_size == 4)
8863 else if (ctx_field_size == 8)
8866 insn->off = off & ~(size_default - 1);
8867 insn->code = BPF_LDX | BPF_MEM | size_code;
8871 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
8873 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
8874 (ctx_field_size && !target_size)) {
8875 verbose(env, "bpf verifier is misconfigured\n");
8879 if (is_narrower_load && size < target_size) {
8880 u8 shift = bpf_ctx_narrow_access_offset(
8881 off, size, size_default) * 8;
8882 if (ctx_field_size <= 4) {
8884 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
8887 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
8888 (1 << size * 8) - 1);
8891 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
8894 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
8895 (1ULL << size * 8) - 1);
8899 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
8905 /* keep walking new program and skip insns we just inserted */
8906 env->prog = new_prog;
8907 insn = new_prog->insnsi + i + delta;
8913 static int jit_subprogs(struct bpf_verifier_env *env)
8915 struct bpf_prog *prog = env->prog, **func, *tmp;
8916 int i, j, subprog_start, subprog_end = 0, len, subprog;
8917 struct bpf_insn *insn;
8921 if (env->subprog_cnt <= 1)
8924 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
8925 if (insn->code != (BPF_JMP | BPF_CALL) ||
8926 insn->src_reg != BPF_PSEUDO_CALL)
8928 /* Upon error here we cannot fall back to interpreter but
8929 * need a hard reject of the program. Thus -EFAULT is
8930 * propagated in any case.
8932 subprog = find_subprog(env, i + insn->imm + 1);
8934 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
8938 /* temporarily remember subprog id inside insn instead of
8939 * aux_data, since next loop will split up all insns into funcs
8941 insn->off = subprog;
8942 /* remember original imm in case JIT fails and fallback
8943 * to interpreter will be needed
8945 env->insn_aux_data[i].call_imm = insn->imm;
8946 /* point imm to __bpf_call_base+1 from JITs point of view */
8950 err = bpf_prog_alloc_jited_linfo(prog);
8955 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
8959 for (i = 0; i < env->subprog_cnt; i++) {
8960 subprog_start = subprog_end;
8961 subprog_end = env->subprog_info[i + 1].start;
8963 len = subprog_end - subprog_start;
8964 /* BPF_PROG_RUN doesn't call subprogs directly,
8965 * hence main prog stats include the runtime of subprogs.
8966 * subprogs don't have IDs and not reachable via prog_get_next_id
8967 * func[i]->aux->stats will never be accessed and stays NULL
8969 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
8972 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
8973 len * sizeof(struct bpf_insn));
8974 func[i]->type = prog->type;
8976 if (bpf_prog_calc_tag(func[i]))
8978 func[i]->is_func = 1;
8979 func[i]->aux->func_idx = i;
8980 /* the btf and func_info will be freed only at prog->aux */
8981 func[i]->aux->btf = prog->aux->btf;
8982 func[i]->aux->func_info = prog->aux->func_info;
8984 /* Use bpf_prog_F_tag to indicate functions in stack traces.
8985 * Long term would need debug info to populate names
8987 func[i]->aux->name[0] = 'F';
8988 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
8989 func[i]->jit_requested = 1;
8990 func[i]->aux->linfo = prog->aux->linfo;
8991 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
8992 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
8993 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
8994 func[i] = bpf_int_jit_compile(func[i]);
8995 if (!func[i]->jited) {
9001 /* at this point all bpf functions were successfully JITed
9002 * now populate all bpf_calls with correct addresses and
9003 * run last pass of JIT
9005 for (i = 0; i < env->subprog_cnt; i++) {
9006 insn = func[i]->insnsi;
9007 for (j = 0; j < func[i]->len; j++, insn++) {
9008 if (insn->code != (BPF_JMP | BPF_CALL) ||
9009 insn->src_reg != BPF_PSEUDO_CALL)
9011 subprog = insn->off;
9012 insn->imm = BPF_CAST_CALL(func[subprog]->bpf_func) -
9016 /* we use the aux data to keep a list of the start addresses
9017 * of the JITed images for each function in the program
9019 * for some architectures, such as powerpc64, the imm field
9020 * might not be large enough to hold the offset of the start
9021 * address of the callee's JITed image from __bpf_call_base
9023 * in such cases, we can lookup the start address of a callee
9024 * by using its subprog id, available from the off field of
9025 * the call instruction, as an index for this list
9027 func[i]->aux->func = func;
9028 func[i]->aux->func_cnt = env->subprog_cnt;
9030 for (i = 0; i < env->subprog_cnt; i++) {
9031 old_bpf_func = func[i]->bpf_func;
9032 tmp = bpf_int_jit_compile(func[i]);
9033 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
9034 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
9041 /* finally lock prog and jit images for all functions and
9044 for (i = 0; i < env->subprog_cnt; i++) {
9045 bpf_prog_lock_ro(func[i]);
9046 bpf_prog_kallsyms_add(func[i]);
9049 /* Last step: make now unused interpreter insns from main
9050 * prog consistent for later dump requests, so they can
9051 * later look the same as if they were interpreted only.
9053 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
9054 if (insn->code != (BPF_JMP | BPF_CALL) ||
9055 insn->src_reg != BPF_PSEUDO_CALL)
9057 insn->off = env->insn_aux_data[i].call_imm;
9058 subprog = find_subprog(env, i + insn->off + 1);
9059 insn->imm = subprog;
9063 prog->bpf_func = func[0]->bpf_func;
9064 prog->aux->func = func;
9065 prog->aux->func_cnt = env->subprog_cnt;
9066 bpf_prog_free_unused_jited_linfo(prog);
9069 for (i = 0; i < env->subprog_cnt; i++)
9071 bpf_jit_free(func[i]);
9074 /* cleanup main prog to be interpreted */
9075 prog->jit_requested = 0;
9076 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
9077 if (insn->code != (BPF_JMP | BPF_CALL) ||
9078 insn->src_reg != BPF_PSEUDO_CALL)
9081 insn->imm = env->insn_aux_data[i].call_imm;
9083 bpf_prog_free_jited_linfo(prog);
9087 static int fixup_call_args(struct bpf_verifier_env *env)
9089 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
9090 struct bpf_prog *prog = env->prog;
9091 struct bpf_insn *insn = prog->insnsi;
9096 if (env->prog->jit_requested &&
9097 !bpf_prog_is_dev_bound(env->prog->aux)) {
9098 err = jit_subprogs(env);
9104 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
9105 for (i = 0; i < prog->len; i++, insn++) {
9106 if (insn->code != (BPF_JMP | BPF_CALL) ||
9107 insn->src_reg != BPF_PSEUDO_CALL)
9109 depth = get_callee_stack_depth(env, insn, i);
9112 bpf_patch_call_args(insn, depth);
9119 /* fixup insn->imm field of bpf_call instructions
9120 * and inline eligible helpers as explicit sequence of BPF instructions
9122 * this function is called after eBPF program passed verification
9124 static int fixup_bpf_calls(struct bpf_verifier_env *env)
9126 struct bpf_prog *prog = env->prog;
9127 bool expect_blinding = bpf_jit_blinding_enabled(prog);
9128 struct bpf_insn *insn = prog->insnsi;
9129 const struct bpf_func_proto *fn;
9130 const int insn_cnt = prog->len;
9131 const struct bpf_map_ops *ops;
9132 struct bpf_insn_aux_data *aux;
9133 struct bpf_insn insn_buf[16];
9134 struct bpf_prog *new_prog;
9135 struct bpf_map *map_ptr;
9136 int i, ret, cnt, delta = 0;
9138 for (i = 0; i < insn_cnt; i++, insn++) {
9139 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
9140 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
9141 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
9142 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
9143 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
9144 struct bpf_insn mask_and_div[] = {
9145 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
9147 BPF_JMP_IMM(BPF_JNE, insn->src_reg, 0, 2),
9148 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
9149 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
9152 struct bpf_insn mask_and_mod[] = {
9153 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
9154 /* Rx mod 0 -> Rx */
9155 BPF_JMP_IMM(BPF_JEQ, insn->src_reg, 0, 1),
9158 struct bpf_insn *patchlet;
9160 if (insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
9161 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
9162 patchlet = mask_and_div + (is64 ? 1 : 0);
9163 cnt = ARRAY_SIZE(mask_and_div) - (is64 ? 1 : 0);
9165 patchlet = mask_and_mod + (is64 ? 1 : 0);
9166 cnt = ARRAY_SIZE(mask_and_mod) - (is64 ? 1 : 0);
9169 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
9174 env->prog = prog = new_prog;
9175 insn = new_prog->insnsi + i + delta;
9179 if (BPF_CLASS(insn->code) == BPF_LD &&
9180 (BPF_MODE(insn->code) == BPF_ABS ||
9181 BPF_MODE(insn->code) == BPF_IND)) {
9182 cnt = env->ops->gen_ld_abs(insn, insn_buf);
9183 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
9184 verbose(env, "bpf verifier is misconfigured\n");
9188 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
9193 env->prog = prog = new_prog;
9194 insn = new_prog->insnsi + i + delta;
9198 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
9199 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
9200 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
9201 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
9202 struct bpf_insn insn_buf[16];
9203 struct bpf_insn *patch = &insn_buf[0];
9207 aux = &env->insn_aux_data[i + delta];
9208 if (!aux->alu_state ||
9209 aux->alu_state == BPF_ALU_NON_POINTER)
9212 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
9213 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
9214 BPF_ALU_SANITIZE_SRC;
9216 off_reg = issrc ? insn->src_reg : insn->dst_reg;
9218 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
9219 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit - 1);
9220 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
9221 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
9222 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
9223 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
9225 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX,
9227 insn->src_reg = BPF_REG_AX;
9229 *patch++ = BPF_ALU64_REG(BPF_AND, off_reg,
9233 insn->code = insn->code == code_add ?
9234 code_sub : code_add;
9237 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
9238 cnt = patch - insn_buf;
9240 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
9245 env->prog = prog = new_prog;
9246 insn = new_prog->insnsi + i + delta;
9250 if (insn->code != (BPF_JMP | BPF_CALL))
9252 if (insn->src_reg == BPF_PSEUDO_CALL)
9255 if (insn->imm == BPF_FUNC_get_route_realm)
9256 prog->dst_needed = 1;
9257 if (insn->imm == BPF_FUNC_get_prandom_u32)
9258 bpf_user_rnd_init_once();
9259 if (insn->imm == BPF_FUNC_override_return)
9260 prog->kprobe_override = 1;
9261 if (insn->imm == BPF_FUNC_tail_call) {
9262 /* If we tail call into other programs, we
9263 * cannot make any assumptions since they can
9264 * be replaced dynamically during runtime in
9265 * the program array.
9267 prog->cb_access = 1;
9268 env->prog->aux->stack_depth = MAX_BPF_STACK;
9269 env->prog->aux->max_pkt_offset = MAX_PACKET_OFF;
9271 /* mark bpf_tail_call as different opcode to avoid
9272 * conditional branch in the interpeter for every normal
9273 * call and to prevent accidental JITing by JIT compiler
9274 * that doesn't support bpf_tail_call yet
9277 insn->code = BPF_JMP | BPF_TAIL_CALL;
9279 aux = &env->insn_aux_data[i + delta];
9280 if (env->allow_ptr_leaks && !expect_blinding &&
9281 prog->jit_requested &&
9282 !bpf_map_key_poisoned(aux) &&
9283 !bpf_map_ptr_poisoned(aux) &&
9284 !bpf_map_ptr_unpriv(aux)) {
9285 struct bpf_jit_poke_descriptor desc = {
9286 .reason = BPF_POKE_REASON_TAIL_CALL,
9287 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
9288 .tail_call.key = bpf_map_key_immediate(aux),
9291 ret = bpf_jit_add_poke_descriptor(prog, &desc);
9293 verbose(env, "adding tail call poke descriptor failed\n");
9297 insn->imm = ret + 1;
9301 if (!bpf_map_ptr_unpriv(aux))
9304 /* instead of changing every JIT dealing with tail_call
9305 * emit two extra insns:
9306 * if (index >= max_entries) goto out;
9307 * index &= array->index_mask;
9308 * to avoid out-of-bounds cpu speculation
9310 if (bpf_map_ptr_poisoned(aux)) {
9311 verbose(env, "tail_call abusing map_ptr\n");
9315 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
9316 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
9317 map_ptr->max_entries, 2);
9318 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
9319 container_of(map_ptr,
9322 insn_buf[2] = *insn;
9324 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
9329 env->prog = prog = new_prog;
9330 insn = new_prog->insnsi + i + delta;
9334 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
9335 * and other inlining handlers are currently limited to 64 bit
9338 if (prog->jit_requested && BITS_PER_LONG == 64 &&
9339 (insn->imm == BPF_FUNC_map_lookup_elem ||
9340 insn->imm == BPF_FUNC_map_update_elem ||
9341 insn->imm == BPF_FUNC_map_delete_elem ||
9342 insn->imm == BPF_FUNC_map_push_elem ||
9343 insn->imm == BPF_FUNC_map_pop_elem ||
9344 insn->imm == BPF_FUNC_map_peek_elem)) {
9345 aux = &env->insn_aux_data[i + delta];
9346 if (bpf_map_ptr_poisoned(aux))
9347 goto patch_call_imm;
9349 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
9351 if (insn->imm == BPF_FUNC_map_lookup_elem &&
9352 ops->map_gen_lookup) {
9353 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
9354 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
9355 verbose(env, "bpf verifier is misconfigured\n");
9359 new_prog = bpf_patch_insn_data(env, i + delta,
9365 env->prog = prog = new_prog;
9366 insn = new_prog->insnsi + i + delta;
9370 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
9371 (void *(*)(struct bpf_map *map, void *key))NULL));
9372 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
9373 (int (*)(struct bpf_map *map, void *key))NULL));
9374 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
9375 (int (*)(struct bpf_map *map, void *key, void *value,
9377 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
9378 (int (*)(struct bpf_map *map, void *value,
9380 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
9381 (int (*)(struct bpf_map *map, void *value))NULL));
9382 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
9383 (int (*)(struct bpf_map *map, void *value))NULL));
9385 switch (insn->imm) {
9386 case BPF_FUNC_map_lookup_elem:
9387 insn->imm = BPF_CAST_CALL(ops->map_lookup_elem) -
9390 case BPF_FUNC_map_update_elem:
9391 insn->imm = BPF_CAST_CALL(ops->map_update_elem) -
9394 case BPF_FUNC_map_delete_elem:
9395 insn->imm = BPF_CAST_CALL(ops->map_delete_elem) -
9398 case BPF_FUNC_map_push_elem:
9399 insn->imm = BPF_CAST_CALL(ops->map_push_elem) -
9402 case BPF_FUNC_map_pop_elem:
9403 insn->imm = BPF_CAST_CALL(ops->map_pop_elem) -
9406 case BPF_FUNC_map_peek_elem:
9407 insn->imm = BPF_CAST_CALL(ops->map_peek_elem) -
9412 goto patch_call_imm;
9416 fn = env->ops->get_func_proto(insn->imm, env->prog);
9417 /* all functions that have prototype and verifier allowed
9418 * programs to call them, must be real in-kernel functions
9422 "kernel subsystem misconfigured func %s#%d\n",
9423 func_id_name(insn->imm), insn->imm);
9426 insn->imm = fn->func - __bpf_call_base;
9429 /* Since poke tab is now finalized, publish aux to tracker. */
9430 for (i = 0; i < prog->aux->size_poke_tab; i++) {
9431 map_ptr = prog->aux->poke_tab[i].tail_call.map;
9432 if (!map_ptr->ops->map_poke_track ||
9433 !map_ptr->ops->map_poke_untrack ||
9434 !map_ptr->ops->map_poke_run) {
9435 verbose(env, "bpf verifier is misconfigured\n");
9439 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
9441 verbose(env, "tracking tail call prog failed\n");
9449 static void free_states(struct bpf_verifier_env *env)
9451 struct bpf_verifier_state_list *sl, *sln;
9454 sl = env->free_list;
9457 free_verifier_state(&sl->state, false);
9462 if (!env->explored_states)
9465 for (i = 0; i < state_htab_size(env); i++) {
9466 sl = env->explored_states[i];
9470 free_verifier_state(&sl->state, false);
9476 kvfree(env->explored_states);
9479 static void print_verification_stats(struct bpf_verifier_env *env)
9483 if (env->log.level & BPF_LOG_STATS) {
9484 verbose(env, "verification time %lld usec\n",
9485 div_u64(env->verification_time, 1000));
9486 verbose(env, "stack depth ");
9487 for (i = 0; i < env->subprog_cnt; i++) {
9488 u32 depth = env->subprog_info[i].stack_depth;
9490 verbose(env, "%d", depth);
9491 if (i + 1 < env->subprog_cnt)
9496 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
9497 "total_states %d peak_states %d mark_read %d\n",
9498 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
9499 env->max_states_per_insn, env->total_states,
9500 env->peak_states, env->longest_mark_read_walk);
9503 static int check_attach_btf_id(struct bpf_verifier_env *env)
9505 struct bpf_prog *prog = env->prog;
9506 struct bpf_prog *tgt_prog = prog->aux->linked_prog;
9507 u32 btf_id = prog->aux->attach_btf_id;
9508 const char prefix[] = "btf_trace_";
9509 int ret = 0, subprog = -1, i;
9510 struct bpf_trampoline *tr;
9511 const struct btf_type *t;
9512 bool conservative = true;
9518 if (prog->type != BPF_PROG_TYPE_TRACING)
9522 verbose(env, "Tracing programs must provide btf_id\n");
9525 btf = bpf_prog_get_target_btf(prog);
9528 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
9531 t = btf_type_by_id(btf, btf_id);
9533 verbose(env, "attach_btf_id %u is invalid\n", btf_id);
9536 tname = btf_name_by_offset(btf, t->name_off);
9538 verbose(env, "attach_btf_id %u doesn't have a name\n", btf_id);
9542 struct bpf_prog_aux *aux = tgt_prog->aux;
9544 for (i = 0; i < aux->func_info_cnt; i++)
9545 if (aux->func_info[i].type_id == btf_id) {
9549 if (subprog == -1) {
9550 verbose(env, "Subprog %s doesn't exist\n", tname);
9553 conservative = aux->func_info_aux[subprog].unreliable;
9554 key = ((u64)aux->id) << 32 | btf_id;
9559 switch (prog->expected_attach_type) {
9560 case BPF_TRACE_RAW_TP:
9563 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
9566 if (!btf_type_is_typedef(t)) {
9567 verbose(env, "attach_btf_id %u is not a typedef\n",
9571 if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
9572 verbose(env, "attach_btf_id %u points to wrong type name %s\n",
9576 tname += sizeof(prefix) - 1;
9577 t = btf_type_by_id(btf, t->type);
9578 if (!btf_type_is_ptr(t))
9579 /* should never happen in valid vmlinux build */
9581 t = btf_type_by_id(btf, t->type);
9582 if (!btf_type_is_func_proto(t))
9583 /* should never happen in valid vmlinux build */
9586 /* remember two read only pointers that are valid for
9587 * the life time of the kernel
9589 prog->aux->attach_func_name = tname;
9590 prog->aux->attach_func_proto = t;
9591 prog->aux->attach_btf_trace = true;
9593 case BPF_TRACE_FENTRY:
9594 case BPF_TRACE_FEXIT:
9595 if (!btf_type_is_func(t)) {
9596 verbose(env, "attach_btf_id %u is not a function\n",
9600 t = btf_type_by_id(btf, t->type);
9601 if (!btf_type_is_func_proto(t))
9603 tr = bpf_trampoline_lookup(key);
9606 prog->aux->attach_func_name = tname;
9607 /* t is either vmlinux type or another program's type */
9608 prog->aux->attach_func_proto = t;
9609 mutex_lock(&tr->mutex);
9610 if (tr->func.addr) {
9611 prog->aux->trampoline = tr;
9614 if (tgt_prog && conservative) {
9615 prog->aux->attach_func_proto = NULL;
9618 ret = btf_distill_func_proto(&env->log, btf, t,
9619 tname, &tr->func.model);
9623 if (!tgt_prog->jited) {
9625 verbose(env, "Can trace only JITed BPF progs\n");
9629 if (tgt_prog->type == BPF_PROG_TYPE_TRACING) {
9630 /* prevent cycles */
9631 verbose(env, "Cannot recursively attach\n");
9636 addr = (long) tgt_prog->bpf_func;
9638 addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
9640 addr = kallsyms_lookup_name(tname);
9643 "The address of function %s cannot be found\n",
9649 tr->func.addr = (void *)addr;
9650 prog->aux->trampoline = tr;
9652 mutex_unlock(&tr->mutex);
9654 bpf_trampoline_put(tr);
9661 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr,
9662 union bpf_attr __user *uattr)
9664 u64 start_time = ktime_get_ns();
9665 struct bpf_verifier_env *env;
9666 struct bpf_verifier_log *log;
9667 int i, len, ret = -EINVAL;
9670 /* no program is valid */
9671 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
9674 /* 'struct bpf_verifier_env' can be global, but since it's not small,
9675 * allocate/free it every time bpf_check() is called
9677 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
9683 env->insn_aux_data =
9684 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
9686 if (!env->insn_aux_data)
9688 for (i = 0; i < len; i++)
9689 env->insn_aux_data[i].orig_idx = i;
9691 env->ops = bpf_verifier_ops[env->prog->type];
9692 is_priv = capable(CAP_SYS_ADMIN);
9694 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
9695 mutex_lock(&bpf_verifier_lock);
9697 btf_vmlinux = btf_parse_vmlinux();
9698 mutex_unlock(&bpf_verifier_lock);
9701 /* grab the mutex to protect few globals used by verifier */
9703 mutex_lock(&bpf_verifier_lock);
9705 if (attr->log_level || attr->log_buf || attr->log_size) {
9706 /* user requested verbose verifier output
9707 * and supplied buffer to store the verification trace
9709 log->level = attr->log_level;
9710 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
9711 log->len_total = attr->log_size;
9714 /* log attributes have to be sane */
9715 if (log->len_total < 128 || log->len_total > UINT_MAX >> 2 ||
9716 !log->level || !log->ubuf || log->level & ~BPF_LOG_MASK)
9720 if (IS_ERR(btf_vmlinux)) {
9721 /* Either gcc or pahole or kernel are broken. */
9722 verbose(env, "in-kernel BTF is malformed\n");
9723 ret = PTR_ERR(btf_vmlinux);
9724 goto skip_full_check;
9727 ret = check_attach_btf_id(env);
9729 goto skip_full_check;
9731 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
9732 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
9733 env->strict_alignment = true;
9734 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
9735 env->strict_alignment = false;
9737 env->allow_ptr_leaks = is_priv;
9740 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
9742 ret = replace_map_fd_with_map_ptr(env);
9744 goto skip_full_check;
9746 if (bpf_prog_is_dev_bound(env->prog->aux)) {
9747 ret = bpf_prog_offload_verifier_prep(env->prog);
9749 goto skip_full_check;
9752 env->explored_states = kvcalloc(state_htab_size(env),
9753 sizeof(struct bpf_verifier_state_list *),
9756 if (!env->explored_states)
9757 goto skip_full_check;
9759 ret = check_subprogs(env);
9761 goto skip_full_check;
9763 ret = check_btf_info(env, attr, uattr);
9765 goto skip_full_check;
9767 ret = check_cfg(env);
9769 goto skip_full_check;
9771 ret = do_check(env);
9772 if (env->cur_state) {
9773 free_verifier_state(env->cur_state, true);
9774 env->cur_state = NULL;
9777 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
9778 ret = bpf_prog_offload_finalize(env);
9781 while (!pop_stack(env, NULL, NULL));
9785 ret = check_max_stack_depth(env);
9787 /* instruction rewrites happen after this point */
9790 opt_hard_wire_dead_code_branches(env);
9792 ret = opt_remove_dead_code(env);
9794 ret = opt_remove_nops(env);
9797 sanitize_dead_code(env);
9801 /* program is valid, convert *(u32*)(ctx + off) accesses */
9802 ret = convert_ctx_accesses(env);
9805 ret = fixup_bpf_calls(env);
9807 /* do 32-bit optimization after insn patching has done so those patched
9808 * insns could be handled correctly.
9810 if (ret == 0 && !bpf_prog_is_dev_bound(env->prog->aux)) {
9811 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
9812 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
9817 ret = fixup_call_args(env);
9819 env->verification_time = ktime_get_ns() - start_time;
9820 print_verification_stats(env);
9822 if (log->level && bpf_verifier_log_full(log))
9824 if (log->level && !log->ubuf) {
9826 goto err_release_maps;
9829 if (ret == 0 && env->used_map_cnt) {
9830 /* if program passed verifier, update used_maps in bpf_prog_info */
9831 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
9832 sizeof(env->used_maps[0]),
9835 if (!env->prog->aux->used_maps) {
9837 goto err_release_maps;
9840 memcpy(env->prog->aux->used_maps, env->used_maps,
9841 sizeof(env->used_maps[0]) * env->used_map_cnt);
9842 env->prog->aux->used_map_cnt = env->used_map_cnt;
9844 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
9845 * bpf_ld_imm64 instructions
9847 convert_pseudo_ld_imm64(env);
9851 adjust_btf_func(env);
9854 if (!env->prog->aux->used_maps)
9855 /* if we didn't copy map pointers into bpf_prog_info, release
9856 * them now. Otherwise free_used_maps() will release them.
9862 mutex_unlock(&bpf_verifier_lock);
9863 vfree(env->insn_aux_data);