1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
3 * Copyright (c) 2016 Facebook
4 * Copyright (c) 2018 Covalent IO, Inc. http://covalent.io
6 #include <uapi/linux/btf.h>
7 #include <linux/kernel.h>
8 #include <linux/types.h>
9 #include <linux/slab.h>
10 #include <linux/bpf.h>
11 #include <linux/btf.h>
12 #include <linux/bpf_verifier.h>
13 #include <linux/filter.h>
14 #include <net/netlink.h>
15 #include <linux/file.h>
16 #include <linux/vmalloc.h>
17 #include <linux/stringify.h>
18 #include <linux/bsearch.h>
19 #include <linux/sort.h>
20 #include <linux/perf_event.h>
21 #include <linux/ctype.h>
22 #include <linux/error-injection.h>
23 #include <linux/bpf_lsm.h>
27 static const struct bpf_verifier_ops * const bpf_verifier_ops[] = {
28 #define BPF_PROG_TYPE(_id, _name, prog_ctx_type, kern_ctx_type) \
29 [_id] = & _name ## _verifier_ops,
30 #define BPF_MAP_TYPE(_id, _ops)
31 #include <linux/bpf_types.h>
36 /* bpf_check() is a static code analyzer that walks eBPF program
37 * instruction by instruction and updates register/stack state.
38 * All paths of conditional branches are analyzed until 'bpf_exit' insn.
40 * The first pass is depth-first-search to check that the program is a DAG.
41 * It rejects the following programs:
42 * - larger than BPF_MAXINSNS insns
43 * - if loop is present (detected via back-edge)
44 * - unreachable insns exist (shouldn't be a forest. program = one function)
45 * - out of bounds or malformed jumps
46 * The second pass is all possible path descent from the 1st insn.
47 * Since it's analyzing all pathes through the program, the length of the
48 * analysis is limited to 64k insn, which may be hit even if total number of
49 * insn is less then 4K, but there are too many branches that change stack/regs.
50 * Number of 'branches to be analyzed' is limited to 1k
52 * On entry to each instruction, each register has a type, and the instruction
53 * changes the types of the registers depending on instruction semantics.
54 * If instruction is BPF_MOV64_REG(BPF_REG_1, BPF_REG_5), then type of R5 is
57 * All registers are 64-bit.
58 * R0 - return register
59 * R1-R5 argument passing registers
60 * R6-R9 callee saved registers
61 * R10 - frame pointer read-only
63 * At the start of BPF program the register R1 contains a pointer to bpf_context
64 * and has type PTR_TO_CTX.
66 * Verifier tracks arithmetic operations on pointers in case:
67 * BPF_MOV64_REG(BPF_REG_1, BPF_REG_10),
68 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, -20),
69 * 1st insn copies R10 (which has FRAME_PTR) type into R1
70 * and 2nd arithmetic instruction is pattern matched to recognize
71 * that it wants to construct a pointer to some element within stack.
72 * So after 2nd insn, the register R1 has type PTR_TO_STACK
73 * (and -20 constant is saved for further stack bounds checking).
74 * Meaning that this reg is a pointer to stack plus known immediate constant.
76 * Most of the time the registers have SCALAR_VALUE type, which
77 * means the register has some value, but it's not a valid pointer.
78 * (like pointer plus pointer becomes SCALAR_VALUE type)
80 * When verifier sees load or store instructions the type of base register
81 * can be: PTR_TO_MAP_VALUE, PTR_TO_CTX, PTR_TO_STACK, PTR_TO_SOCKET. These are
82 * four pointer types recognized by check_mem_access() function.
84 * PTR_TO_MAP_VALUE means that this register is pointing to 'map element value'
85 * and the range of [ptr, ptr + map's value_size) is accessible.
87 * registers used to pass values to function calls are checked against
88 * function argument constraints.
90 * ARG_PTR_TO_MAP_KEY is one of such argument constraints.
91 * It means that the register type passed to this function must be
92 * PTR_TO_STACK and it will be used inside the function as
93 * 'pointer to map element key'
95 * For example the argument constraints for bpf_map_lookup_elem():
96 * .ret_type = RET_PTR_TO_MAP_VALUE_OR_NULL,
97 * .arg1_type = ARG_CONST_MAP_PTR,
98 * .arg2_type = ARG_PTR_TO_MAP_KEY,
100 * ret_type says that this function returns 'pointer to map elem value or null'
101 * function expects 1st argument to be a const pointer to 'struct bpf_map' and
102 * 2nd argument should be a pointer to stack, which will be used inside
103 * the helper function as a pointer to map element key.
105 * On the kernel side the helper function looks like:
106 * u64 bpf_map_lookup_elem(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
108 * struct bpf_map *map = (struct bpf_map *) (unsigned long) r1;
109 * void *key = (void *) (unsigned long) r2;
112 * here kernel can access 'key' and 'map' pointers safely, knowing that
113 * [key, key + map->key_size) bytes are valid and were initialized on
114 * the stack of eBPF program.
117 * Corresponding eBPF program may look like:
118 * BPF_MOV64_REG(BPF_REG_2, BPF_REG_10), // after this insn R2 type is FRAME_PTR
119 * BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, -4), // after this insn R2 type is PTR_TO_STACK
120 * BPF_LD_MAP_FD(BPF_REG_1, map_fd), // after this insn R1 type is CONST_PTR_TO_MAP
121 * BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0, BPF_FUNC_map_lookup_elem),
122 * here verifier looks at prototype of map_lookup_elem() and sees:
123 * .arg1_type == ARG_CONST_MAP_PTR and R1->type == CONST_PTR_TO_MAP, which is ok,
124 * Now verifier knows that this map has key of R1->map_ptr->key_size bytes
126 * Then .arg2_type == ARG_PTR_TO_MAP_KEY and R2->type == PTR_TO_STACK, ok so far,
127 * Now verifier checks that [R2, R2 + map's key_size) are within stack limits
128 * and were initialized prior to this call.
129 * If it's ok, then verifier allows this BPF_CALL insn and looks at
130 * .ret_type which is RET_PTR_TO_MAP_VALUE_OR_NULL, so it sets
131 * R0->type = PTR_TO_MAP_VALUE_OR_NULL which means bpf_map_lookup_elem() function
132 * returns ether pointer to map value or NULL.
134 * When type PTR_TO_MAP_VALUE_OR_NULL passes through 'if (reg != 0) goto +off'
135 * insn, the register holding that pointer in the true branch changes state to
136 * PTR_TO_MAP_VALUE and the same register changes state to CONST_IMM in the false
137 * branch. See check_cond_jmp_op().
139 * After the call R0 is set to return type of the function and registers R1-R5
140 * are set to NOT_INIT to indicate that they are no longer readable.
142 * The following reference types represent a potential reference to a kernel
143 * resource which, after first being allocated, must be checked and freed by
145 * - PTR_TO_SOCKET_OR_NULL, PTR_TO_SOCKET
147 * When the verifier sees a helper call return a reference type, it allocates a
148 * pointer id for the reference and stores it in the current function state.
149 * Similar to the way that PTR_TO_MAP_VALUE_OR_NULL is converted into
150 * PTR_TO_MAP_VALUE, PTR_TO_SOCKET_OR_NULL becomes PTR_TO_SOCKET when the type
151 * passes through a NULL-check conditional. For the branch wherein the state is
152 * changed to CONST_IMM, the verifier releases the reference.
154 * For each helper function that allocates a reference, such as
155 * bpf_sk_lookup_tcp(), there is a corresponding release function, such as
156 * bpf_sk_release(). When a reference type passes into the release function,
157 * the verifier also releases the reference. If any unchecked or unreleased
158 * reference remains at the end of the program, the verifier rejects it.
161 /* verifier_state + insn_idx are pushed to stack when branch is encountered */
162 struct bpf_verifier_stack_elem {
163 /* verifer state is 'st'
164 * before processing instruction 'insn_idx'
165 * and after processing instruction 'prev_insn_idx'
167 struct bpf_verifier_state st;
170 struct bpf_verifier_stack_elem *next;
173 #define BPF_COMPLEXITY_LIMIT_JMP_SEQ 8192
174 #define BPF_COMPLEXITY_LIMIT_STATES 64
176 #define BPF_MAP_KEY_POISON (1ULL << 63)
177 #define BPF_MAP_KEY_SEEN (1ULL << 62)
179 #define BPF_MAP_PTR_UNPRIV 1UL
180 #define BPF_MAP_PTR_POISON ((void *)((0xeB9FUL << 1) + \
181 POISON_POINTER_DELTA))
182 #define BPF_MAP_PTR(X) ((struct bpf_map *)((X) & ~BPF_MAP_PTR_UNPRIV))
184 static bool bpf_map_ptr_poisoned(const struct bpf_insn_aux_data *aux)
186 return BPF_MAP_PTR(aux->map_ptr_state) == BPF_MAP_PTR_POISON;
189 static bool bpf_map_ptr_unpriv(const struct bpf_insn_aux_data *aux)
191 return aux->map_ptr_state & BPF_MAP_PTR_UNPRIV;
194 static void bpf_map_ptr_store(struct bpf_insn_aux_data *aux,
195 const struct bpf_map *map, bool unpriv)
197 BUILD_BUG_ON((unsigned long)BPF_MAP_PTR_POISON & BPF_MAP_PTR_UNPRIV);
198 unpriv |= bpf_map_ptr_unpriv(aux);
199 aux->map_ptr_state = (unsigned long)map |
200 (unpriv ? BPF_MAP_PTR_UNPRIV : 0UL);
203 static bool bpf_map_key_poisoned(const struct bpf_insn_aux_data *aux)
205 return aux->map_key_state & BPF_MAP_KEY_POISON;
208 static bool bpf_map_key_unseen(const struct bpf_insn_aux_data *aux)
210 return !(aux->map_key_state & BPF_MAP_KEY_SEEN);
213 static u64 bpf_map_key_immediate(const struct bpf_insn_aux_data *aux)
215 return aux->map_key_state & ~(BPF_MAP_KEY_SEEN | BPF_MAP_KEY_POISON);
218 static void bpf_map_key_store(struct bpf_insn_aux_data *aux, u64 state)
220 bool poisoned = bpf_map_key_poisoned(aux);
222 aux->map_key_state = state | BPF_MAP_KEY_SEEN |
223 (poisoned ? BPF_MAP_KEY_POISON : 0ULL);
226 struct bpf_call_arg_meta {
227 struct bpf_map *map_ptr;
238 struct btf *btf_vmlinux;
240 static DEFINE_MUTEX(bpf_verifier_lock);
242 static const struct bpf_line_info *
243 find_linfo(const struct bpf_verifier_env *env, u32 insn_off)
245 const struct bpf_line_info *linfo;
246 const struct bpf_prog *prog;
250 nr_linfo = prog->aux->nr_linfo;
252 if (!nr_linfo || insn_off >= prog->len)
255 linfo = prog->aux->linfo;
256 for (i = 1; i < nr_linfo; i++)
257 if (insn_off < linfo[i].insn_off)
260 return &linfo[i - 1];
263 void bpf_verifier_vlog(struct bpf_verifier_log *log, const char *fmt,
268 n = vscnprintf(log->kbuf, BPF_VERIFIER_TMP_LOG_SIZE, fmt, args);
270 WARN_ONCE(n >= BPF_VERIFIER_TMP_LOG_SIZE - 1,
271 "verifier log line truncated - local buffer too short\n");
273 n = min(log->len_total - log->len_used - 1, n);
276 if (log->level == BPF_LOG_KERNEL) {
277 pr_err("BPF:%s\n", log->kbuf);
280 if (!copy_to_user(log->ubuf + log->len_used, log->kbuf, n + 1))
286 /* log_level controls verbosity level of eBPF verifier.
287 * bpf_verifier_log_write() is used to dump the verification trace to the log,
288 * so the user can figure out what's wrong with the program
290 __printf(2, 3) void bpf_verifier_log_write(struct bpf_verifier_env *env,
291 const char *fmt, ...)
295 if (!bpf_verifier_log_needed(&env->log))
299 bpf_verifier_vlog(&env->log, fmt, args);
302 EXPORT_SYMBOL_GPL(bpf_verifier_log_write);
304 __printf(2, 3) static void verbose(void *private_data, const char *fmt, ...)
306 struct bpf_verifier_env *env = private_data;
309 if (!bpf_verifier_log_needed(&env->log))
313 bpf_verifier_vlog(&env->log, fmt, args);
317 __printf(2, 3) void bpf_log(struct bpf_verifier_log *log,
318 const char *fmt, ...)
322 if (!bpf_verifier_log_needed(log))
326 bpf_verifier_vlog(log, fmt, args);
330 static const char *ltrim(const char *s)
338 __printf(3, 4) static void verbose_linfo(struct bpf_verifier_env *env,
340 const char *prefix_fmt, ...)
342 const struct bpf_line_info *linfo;
344 if (!bpf_verifier_log_needed(&env->log))
347 linfo = find_linfo(env, insn_off);
348 if (!linfo || linfo == env->prev_linfo)
354 va_start(args, prefix_fmt);
355 bpf_verifier_vlog(&env->log, prefix_fmt, args);
360 ltrim(btf_name_by_offset(env->prog->aux->btf,
363 env->prev_linfo = linfo;
366 static bool type_is_pkt_pointer(enum bpf_reg_type type)
368 return type == PTR_TO_PACKET ||
369 type == PTR_TO_PACKET_META;
372 static bool type_is_sk_pointer(enum bpf_reg_type type)
374 return type == PTR_TO_SOCKET ||
375 type == PTR_TO_SOCK_COMMON ||
376 type == PTR_TO_TCP_SOCK ||
377 type == PTR_TO_XDP_SOCK;
380 static bool reg_type_may_be_null(enum bpf_reg_type type)
382 return type == PTR_TO_MAP_VALUE_OR_NULL ||
383 type == PTR_TO_SOCKET_OR_NULL ||
384 type == PTR_TO_SOCK_COMMON_OR_NULL ||
385 type == PTR_TO_TCP_SOCK_OR_NULL;
388 static bool reg_may_point_to_spin_lock(const struct bpf_reg_state *reg)
390 return reg->type == PTR_TO_MAP_VALUE &&
391 map_value_has_spin_lock(reg->map_ptr);
394 static bool reg_type_may_be_refcounted_or_null(enum bpf_reg_type type)
396 return type == PTR_TO_SOCKET ||
397 type == PTR_TO_SOCKET_OR_NULL ||
398 type == PTR_TO_TCP_SOCK ||
399 type == PTR_TO_TCP_SOCK_OR_NULL;
402 static bool arg_type_may_be_refcounted(enum bpf_arg_type type)
404 return type == ARG_PTR_TO_SOCK_COMMON;
407 /* Determine whether the function releases some resources allocated by another
408 * function call. The first reference type argument will be assumed to be
409 * released by release_reference().
411 static bool is_release_function(enum bpf_func_id func_id)
413 return func_id == BPF_FUNC_sk_release;
416 static bool is_acquire_function(enum bpf_func_id func_id)
418 return func_id == BPF_FUNC_sk_lookup_tcp ||
419 func_id == BPF_FUNC_sk_lookup_udp ||
420 func_id == BPF_FUNC_skc_lookup_tcp;
423 static bool is_ptr_cast_function(enum bpf_func_id func_id)
425 return func_id == BPF_FUNC_tcp_sock ||
426 func_id == BPF_FUNC_sk_fullsock;
429 /* string representation of 'enum bpf_reg_type' */
430 static const char * const reg_type_str[] = {
432 [SCALAR_VALUE] = "inv",
433 [PTR_TO_CTX] = "ctx",
434 [CONST_PTR_TO_MAP] = "map_ptr",
435 [PTR_TO_MAP_VALUE] = "map_value",
436 [PTR_TO_MAP_VALUE_OR_NULL] = "map_value_or_null",
437 [PTR_TO_STACK] = "fp",
438 [PTR_TO_PACKET] = "pkt",
439 [PTR_TO_PACKET_META] = "pkt_meta",
440 [PTR_TO_PACKET_END] = "pkt_end",
441 [PTR_TO_FLOW_KEYS] = "flow_keys",
442 [PTR_TO_SOCKET] = "sock",
443 [PTR_TO_SOCKET_OR_NULL] = "sock_or_null",
444 [PTR_TO_SOCK_COMMON] = "sock_common",
445 [PTR_TO_SOCK_COMMON_OR_NULL] = "sock_common_or_null",
446 [PTR_TO_TCP_SOCK] = "tcp_sock",
447 [PTR_TO_TCP_SOCK_OR_NULL] = "tcp_sock_or_null",
448 [PTR_TO_TP_BUFFER] = "tp_buffer",
449 [PTR_TO_XDP_SOCK] = "xdp_sock",
450 [PTR_TO_BTF_ID] = "ptr_",
453 static char slot_type_char[] = {
454 [STACK_INVALID] = '?',
460 static void print_liveness(struct bpf_verifier_env *env,
461 enum bpf_reg_liveness live)
463 if (live & (REG_LIVE_READ | REG_LIVE_WRITTEN | REG_LIVE_DONE))
465 if (live & REG_LIVE_READ)
467 if (live & REG_LIVE_WRITTEN)
469 if (live & REG_LIVE_DONE)
473 static struct bpf_func_state *func(struct bpf_verifier_env *env,
474 const struct bpf_reg_state *reg)
476 struct bpf_verifier_state *cur = env->cur_state;
478 return cur->frame[reg->frameno];
481 const char *kernel_type_name(u32 id)
483 return btf_name_by_offset(btf_vmlinux,
484 btf_type_by_id(btf_vmlinux, id)->name_off);
487 static void print_verifier_state(struct bpf_verifier_env *env,
488 const struct bpf_func_state *state)
490 const struct bpf_reg_state *reg;
495 verbose(env, " frame%d:", state->frameno);
496 for (i = 0; i < MAX_BPF_REG; i++) {
497 reg = &state->regs[i];
501 verbose(env, " R%d", i);
502 print_liveness(env, reg->live);
503 verbose(env, "=%s", reg_type_str[t]);
504 if (t == SCALAR_VALUE && reg->precise)
506 if ((t == SCALAR_VALUE || t == PTR_TO_STACK) &&
507 tnum_is_const(reg->var_off)) {
508 /* reg->off should be 0 for SCALAR_VALUE */
509 verbose(env, "%lld", reg->var_off.value + reg->off);
511 if (t == PTR_TO_BTF_ID)
512 verbose(env, "%s", kernel_type_name(reg->btf_id));
513 verbose(env, "(id=%d", reg->id);
514 if (reg_type_may_be_refcounted_or_null(t))
515 verbose(env, ",ref_obj_id=%d", reg->ref_obj_id);
516 if (t != SCALAR_VALUE)
517 verbose(env, ",off=%d", reg->off);
518 if (type_is_pkt_pointer(t))
519 verbose(env, ",r=%d", reg->range);
520 else if (t == CONST_PTR_TO_MAP ||
521 t == PTR_TO_MAP_VALUE ||
522 t == PTR_TO_MAP_VALUE_OR_NULL)
523 verbose(env, ",ks=%d,vs=%d",
524 reg->map_ptr->key_size,
525 reg->map_ptr->value_size);
526 if (tnum_is_const(reg->var_off)) {
527 /* Typically an immediate SCALAR_VALUE, but
528 * could be a pointer whose offset is too big
531 verbose(env, ",imm=%llx", reg->var_off.value);
533 if (reg->smin_value != reg->umin_value &&
534 reg->smin_value != S64_MIN)
535 verbose(env, ",smin_value=%lld",
536 (long long)reg->smin_value);
537 if (reg->smax_value != reg->umax_value &&
538 reg->smax_value != S64_MAX)
539 verbose(env, ",smax_value=%lld",
540 (long long)reg->smax_value);
541 if (reg->umin_value != 0)
542 verbose(env, ",umin_value=%llu",
543 (unsigned long long)reg->umin_value);
544 if (reg->umax_value != U64_MAX)
545 verbose(env, ",umax_value=%llu",
546 (unsigned long long)reg->umax_value);
547 if (!tnum_is_unknown(reg->var_off)) {
550 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
551 verbose(env, ",var_off=%s", tn_buf);
557 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
558 char types_buf[BPF_REG_SIZE + 1];
562 for (j = 0; j < BPF_REG_SIZE; j++) {
563 if (state->stack[i].slot_type[j] != STACK_INVALID)
565 types_buf[j] = slot_type_char[
566 state->stack[i].slot_type[j]];
568 types_buf[BPF_REG_SIZE] = 0;
571 verbose(env, " fp%d", (-i - 1) * BPF_REG_SIZE);
572 print_liveness(env, state->stack[i].spilled_ptr.live);
573 if (state->stack[i].slot_type[0] == STACK_SPILL) {
574 reg = &state->stack[i].spilled_ptr;
576 verbose(env, "=%s", reg_type_str[t]);
577 if (t == SCALAR_VALUE && reg->precise)
579 if (t == SCALAR_VALUE && tnum_is_const(reg->var_off))
580 verbose(env, "%lld", reg->var_off.value + reg->off);
582 verbose(env, "=%s", types_buf);
585 if (state->acquired_refs && state->refs[0].id) {
586 verbose(env, " refs=%d", state->refs[0].id);
587 for (i = 1; i < state->acquired_refs; i++)
588 if (state->refs[i].id)
589 verbose(env, ",%d", state->refs[i].id);
594 #define COPY_STATE_FN(NAME, COUNT, FIELD, SIZE) \
595 static int copy_##NAME##_state(struct bpf_func_state *dst, \
596 const struct bpf_func_state *src) \
600 if (WARN_ON_ONCE(dst->COUNT < src->COUNT)) { \
601 /* internal bug, make state invalid to reject the program */ \
602 memset(dst, 0, sizeof(*dst)); \
605 memcpy(dst->FIELD, src->FIELD, \
606 sizeof(*src->FIELD) * (src->COUNT / SIZE)); \
609 /* copy_reference_state() */
610 COPY_STATE_FN(reference, acquired_refs, refs, 1)
611 /* copy_stack_state() */
612 COPY_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
615 #define REALLOC_STATE_FN(NAME, COUNT, FIELD, SIZE) \
616 static int realloc_##NAME##_state(struct bpf_func_state *state, int size, \
619 u32 old_size = state->COUNT; \
620 struct bpf_##NAME##_state *new_##FIELD; \
621 int slot = size / SIZE; \
623 if (size <= old_size || !size) { \
626 state->COUNT = slot * SIZE; \
627 if (!size && old_size) { \
628 kfree(state->FIELD); \
629 state->FIELD = NULL; \
633 new_##FIELD = kmalloc_array(slot, sizeof(struct bpf_##NAME##_state), \
639 memcpy(new_##FIELD, state->FIELD, \
640 sizeof(*new_##FIELD) * (old_size / SIZE)); \
641 memset(new_##FIELD + old_size / SIZE, 0, \
642 sizeof(*new_##FIELD) * (size - old_size) / SIZE); \
644 state->COUNT = slot * SIZE; \
645 kfree(state->FIELD); \
646 state->FIELD = new_##FIELD; \
649 /* realloc_reference_state() */
650 REALLOC_STATE_FN(reference, acquired_refs, refs, 1)
651 /* realloc_stack_state() */
652 REALLOC_STATE_FN(stack, allocated_stack, stack, BPF_REG_SIZE)
653 #undef REALLOC_STATE_FN
655 /* do_check() starts with zero-sized stack in struct bpf_verifier_state to
656 * make it consume minimal amount of memory. check_stack_write() access from
657 * the program calls into realloc_func_state() to grow the stack size.
658 * Note there is a non-zero 'parent' pointer inside bpf_verifier_state
659 * which realloc_stack_state() copies over. It points to previous
660 * bpf_verifier_state which is never reallocated.
662 static int realloc_func_state(struct bpf_func_state *state, int stack_size,
663 int refs_size, bool copy_old)
665 int err = realloc_reference_state(state, refs_size, copy_old);
668 return realloc_stack_state(state, stack_size, copy_old);
671 /* Acquire a pointer id from the env and update the state->refs to include
672 * this new pointer reference.
673 * On success, returns a valid pointer id to associate with the register
674 * On failure, returns a negative errno.
676 static int acquire_reference_state(struct bpf_verifier_env *env, int insn_idx)
678 struct bpf_func_state *state = cur_func(env);
679 int new_ofs = state->acquired_refs;
682 err = realloc_reference_state(state, state->acquired_refs + 1, true);
686 state->refs[new_ofs].id = id;
687 state->refs[new_ofs].insn_idx = insn_idx;
692 /* release function corresponding to acquire_reference_state(). Idempotent. */
693 static int release_reference_state(struct bpf_func_state *state, int ptr_id)
697 last_idx = state->acquired_refs - 1;
698 for (i = 0; i < state->acquired_refs; i++) {
699 if (state->refs[i].id == ptr_id) {
700 if (last_idx && i != last_idx)
701 memcpy(&state->refs[i], &state->refs[last_idx],
702 sizeof(*state->refs));
703 memset(&state->refs[last_idx], 0, sizeof(*state->refs));
704 state->acquired_refs--;
711 static int transfer_reference_state(struct bpf_func_state *dst,
712 struct bpf_func_state *src)
714 int err = realloc_reference_state(dst, src->acquired_refs, false);
717 err = copy_reference_state(dst, src);
723 static void free_func_state(struct bpf_func_state *state)
732 static void clear_jmp_history(struct bpf_verifier_state *state)
734 kfree(state->jmp_history);
735 state->jmp_history = NULL;
736 state->jmp_history_cnt = 0;
739 static void free_verifier_state(struct bpf_verifier_state *state,
744 for (i = 0; i <= state->curframe; i++) {
745 free_func_state(state->frame[i]);
746 state->frame[i] = NULL;
748 clear_jmp_history(state);
753 /* copy verifier state from src to dst growing dst stack space
754 * when necessary to accommodate larger src stack
756 static int copy_func_state(struct bpf_func_state *dst,
757 const struct bpf_func_state *src)
761 err = realloc_func_state(dst, src->allocated_stack, src->acquired_refs,
765 memcpy(dst, src, offsetof(struct bpf_func_state, acquired_refs));
766 err = copy_reference_state(dst, src);
769 return copy_stack_state(dst, src);
772 static int copy_verifier_state(struct bpf_verifier_state *dst_state,
773 const struct bpf_verifier_state *src)
775 struct bpf_func_state *dst;
776 u32 jmp_sz = sizeof(struct bpf_idx_pair) * src->jmp_history_cnt;
779 if (dst_state->jmp_history_cnt < src->jmp_history_cnt) {
780 kfree(dst_state->jmp_history);
781 dst_state->jmp_history = kmalloc(jmp_sz, GFP_USER);
782 if (!dst_state->jmp_history)
785 memcpy(dst_state->jmp_history, src->jmp_history, jmp_sz);
786 dst_state->jmp_history_cnt = src->jmp_history_cnt;
788 /* if dst has more stack frames then src frame, free them */
789 for (i = src->curframe + 1; i <= dst_state->curframe; i++) {
790 free_func_state(dst_state->frame[i]);
791 dst_state->frame[i] = NULL;
793 dst_state->speculative = src->speculative;
794 dst_state->curframe = src->curframe;
795 dst_state->active_spin_lock = src->active_spin_lock;
796 dst_state->branches = src->branches;
797 dst_state->parent = src->parent;
798 dst_state->first_insn_idx = src->first_insn_idx;
799 dst_state->last_insn_idx = src->last_insn_idx;
800 for (i = 0; i <= src->curframe; i++) {
801 dst = dst_state->frame[i];
803 dst = kzalloc(sizeof(*dst), GFP_KERNEL);
806 dst_state->frame[i] = dst;
808 err = copy_func_state(dst, src->frame[i]);
815 static void update_branch_counts(struct bpf_verifier_env *env, struct bpf_verifier_state *st)
818 u32 br = --st->branches;
820 /* WARN_ON(br > 1) technically makes sense here,
821 * but see comment in push_stack(), hence:
823 WARN_ONCE((int)br < 0,
824 "BUG update_branch_counts:branches_to_explore=%d\n",
832 static int pop_stack(struct bpf_verifier_env *env, int *prev_insn_idx,
835 struct bpf_verifier_state *cur = env->cur_state;
836 struct bpf_verifier_stack_elem *elem, *head = env->head;
839 if (env->head == NULL)
843 err = copy_verifier_state(cur, &head->st);
848 *insn_idx = head->insn_idx;
850 *prev_insn_idx = head->prev_insn_idx;
852 free_verifier_state(&head->st, false);
859 static struct bpf_verifier_state *push_stack(struct bpf_verifier_env *env,
860 int insn_idx, int prev_insn_idx,
863 struct bpf_verifier_state *cur = env->cur_state;
864 struct bpf_verifier_stack_elem *elem;
867 elem = kzalloc(sizeof(struct bpf_verifier_stack_elem), GFP_KERNEL);
871 elem->insn_idx = insn_idx;
872 elem->prev_insn_idx = prev_insn_idx;
873 elem->next = env->head;
876 err = copy_verifier_state(&elem->st, cur);
879 elem->st.speculative |= speculative;
880 if (env->stack_size > BPF_COMPLEXITY_LIMIT_JMP_SEQ) {
881 verbose(env, "The sequence of %d jumps is too complex.\n",
885 if (elem->st.parent) {
886 ++elem->st.parent->branches;
887 /* WARN_ON(branches > 2) technically makes sense here,
889 * 1. speculative states will bump 'branches' for non-branch
891 * 2. is_state_visited() heuristics may decide not to create
892 * a new state for a sequence of branches and all such current
893 * and cloned states will be pointing to a single parent state
894 * which might have large 'branches' count.
899 free_verifier_state(env->cur_state, true);
900 env->cur_state = NULL;
901 /* pop all elements and return */
902 while (!pop_stack(env, NULL, NULL));
906 #define CALLER_SAVED_REGS 6
907 static const int caller_saved[CALLER_SAVED_REGS] = {
908 BPF_REG_0, BPF_REG_1, BPF_REG_2, BPF_REG_3, BPF_REG_4, BPF_REG_5
911 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
912 struct bpf_reg_state *reg);
914 /* Mark the unknown part of a register (variable offset or scalar value) as
915 * known to have the value @imm.
917 static void __mark_reg_known(struct bpf_reg_state *reg, u64 imm)
919 /* Clear id, off, and union(map_ptr, range) */
920 memset(((u8 *)reg) + sizeof(reg->type), 0,
921 offsetof(struct bpf_reg_state, var_off) - sizeof(reg->type));
922 reg->var_off = tnum_const(imm);
923 reg->smin_value = (s64)imm;
924 reg->smax_value = (s64)imm;
925 reg->umin_value = imm;
926 reg->umax_value = imm;
929 /* Mark the 'variable offset' part of a register as zero. This should be
930 * used only on registers holding a pointer type.
932 static void __mark_reg_known_zero(struct bpf_reg_state *reg)
934 __mark_reg_known(reg, 0);
937 static void __mark_reg_const_zero(struct bpf_reg_state *reg)
939 __mark_reg_known(reg, 0);
940 reg->type = SCALAR_VALUE;
943 static void mark_reg_known_zero(struct bpf_verifier_env *env,
944 struct bpf_reg_state *regs, u32 regno)
946 if (WARN_ON(regno >= MAX_BPF_REG)) {
947 verbose(env, "mark_reg_known_zero(regs, %u)\n", regno);
948 /* Something bad happened, let's kill all regs */
949 for (regno = 0; regno < MAX_BPF_REG; regno++)
950 __mark_reg_not_init(env, regs + regno);
953 __mark_reg_known_zero(regs + regno);
956 static bool reg_is_pkt_pointer(const struct bpf_reg_state *reg)
958 return type_is_pkt_pointer(reg->type);
961 static bool reg_is_pkt_pointer_any(const struct bpf_reg_state *reg)
963 return reg_is_pkt_pointer(reg) ||
964 reg->type == PTR_TO_PACKET_END;
967 /* Unmodified PTR_TO_PACKET[_META,_END] register from ctx access. */
968 static bool reg_is_init_pkt_pointer(const struct bpf_reg_state *reg,
969 enum bpf_reg_type which)
971 /* The register can already have a range from prior markings.
972 * This is fine as long as it hasn't been advanced from its
975 return reg->type == which &&
978 tnum_equals_const(reg->var_off, 0);
981 /* Attempts to improve min/max values based on var_off information */
982 static void __update_reg_bounds(struct bpf_reg_state *reg)
984 /* min signed is max(sign bit) | min(other bits) */
985 reg->smin_value = max_t(s64, reg->smin_value,
986 reg->var_off.value | (reg->var_off.mask & S64_MIN));
987 /* max signed is min(sign bit) | max(other bits) */
988 reg->smax_value = min_t(s64, reg->smax_value,
989 reg->var_off.value | (reg->var_off.mask & S64_MAX));
990 reg->umin_value = max(reg->umin_value, reg->var_off.value);
991 reg->umax_value = min(reg->umax_value,
992 reg->var_off.value | reg->var_off.mask);
995 /* Uses signed min/max values to inform unsigned, and vice-versa */
996 static void __reg_deduce_bounds(struct bpf_reg_state *reg)
998 /* Learn sign from signed bounds.
999 * If we cannot cross the sign boundary, then signed and unsigned bounds
1000 * are the same, so combine. This works even in the negative case, e.g.
1001 * -3 s<= x s<= -1 implies 0xf...fd u<= x u<= 0xf...ff.
1003 if (reg->smin_value >= 0 || reg->smax_value < 0) {
1004 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1006 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1010 /* Learn sign from unsigned bounds. Signed bounds cross the sign
1011 * boundary, so we must be careful.
1013 if ((s64)reg->umax_value >= 0) {
1014 /* Positive. We can't learn anything from the smin, but smax
1015 * is positive, hence safe.
1017 reg->smin_value = reg->umin_value;
1018 reg->smax_value = reg->umax_value = min_t(u64, reg->smax_value,
1020 } else if ((s64)reg->umin_value < 0) {
1021 /* Negative. We can't learn anything from the smax, but smin
1022 * is negative, hence safe.
1024 reg->smin_value = reg->umin_value = max_t(u64, reg->smin_value,
1026 reg->smax_value = reg->umax_value;
1030 /* Attempts to improve var_off based on unsigned min/max information */
1031 static void __reg_bound_offset(struct bpf_reg_state *reg)
1033 reg->var_off = tnum_intersect(reg->var_off,
1034 tnum_range(reg->umin_value,
1038 /* Reset the min/max bounds of a register */
1039 static void __mark_reg_unbounded(struct bpf_reg_state *reg)
1041 reg->smin_value = S64_MIN;
1042 reg->smax_value = S64_MAX;
1043 reg->umin_value = 0;
1044 reg->umax_value = U64_MAX;
1047 /* Mark a register as having a completely unknown (scalar) value. */
1048 static void __mark_reg_unknown(const struct bpf_verifier_env *env,
1049 struct bpf_reg_state *reg)
1052 * Clear type, id, off, and union(map_ptr, range) and
1053 * padding between 'type' and union
1055 memset(reg, 0, offsetof(struct bpf_reg_state, var_off));
1056 reg->type = SCALAR_VALUE;
1057 reg->var_off = tnum_unknown;
1059 reg->precise = env->subprog_cnt > 1 || !env->allow_ptr_leaks ?
1061 __mark_reg_unbounded(reg);
1064 static void mark_reg_unknown(struct bpf_verifier_env *env,
1065 struct bpf_reg_state *regs, u32 regno)
1067 if (WARN_ON(regno >= MAX_BPF_REG)) {
1068 verbose(env, "mark_reg_unknown(regs, %u)\n", regno);
1069 /* Something bad happened, let's kill all regs except FP */
1070 for (regno = 0; regno < BPF_REG_FP; regno++)
1071 __mark_reg_not_init(env, regs + regno);
1074 __mark_reg_unknown(env, regs + regno);
1077 static void __mark_reg_not_init(const struct bpf_verifier_env *env,
1078 struct bpf_reg_state *reg)
1080 __mark_reg_unknown(env, reg);
1081 reg->type = NOT_INIT;
1084 static void mark_reg_not_init(struct bpf_verifier_env *env,
1085 struct bpf_reg_state *regs, u32 regno)
1087 if (WARN_ON(regno >= MAX_BPF_REG)) {
1088 verbose(env, "mark_reg_not_init(regs, %u)\n", regno);
1089 /* Something bad happened, let's kill all regs except FP */
1090 for (regno = 0; regno < BPF_REG_FP; regno++)
1091 __mark_reg_not_init(env, regs + regno);
1094 __mark_reg_not_init(env, regs + regno);
1097 #define DEF_NOT_SUBREG (0)
1098 static void init_reg_state(struct bpf_verifier_env *env,
1099 struct bpf_func_state *state)
1101 struct bpf_reg_state *regs = state->regs;
1104 for (i = 0; i < MAX_BPF_REG; i++) {
1105 mark_reg_not_init(env, regs, i);
1106 regs[i].live = REG_LIVE_NONE;
1107 regs[i].parent = NULL;
1108 regs[i].subreg_def = DEF_NOT_SUBREG;
1112 regs[BPF_REG_FP].type = PTR_TO_STACK;
1113 mark_reg_known_zero(env, regs, BPF_REG_FP);
1114 regs[BPF_REG_FP].frameno = state->frameno;
1117 #define BPF_MAIN_FUNC (-1)
1118 static void init_func_state(struct bpf_verifier_env *env,
1119 struct bpf_func_state *state,
1120 int callsite, int frameno, int subprogno)
1122 state->callsite = callsite;
1123 state->frameno = frameno;
1124 state->subprogno = subprogno;
1125 init_reg_state(env, state);
1129 SRC_OP, /* register is used as source operand */
1130 DST_OP, /* register is used as destination operand */
1131 DST_OP_NO_MARK /* same as above, check only, don't mark */
1134 static int cmp_subprogs(const void *a, const void *b)
1136 return ((struct bpf_subprog_info *)a)->start -
1137 ((struct bpf_subprog_info *)b)->start;
1140 static int find_subprog(struct bpf_verifier_env *env, int off)
1142 struct bpf_subprog_info *p;
1144 p = bsearch(&off, env->subprog_info, env->subprog_cnt,
1145 sizeof(env->subprog_info[0]), cmp_subprogs);
1148 return p - env->subprog_info;
1152 static int add_subprog(struct bpf_verifier_env *env, int off)
1154 int insn_cnt = env->prog->len;
1157 if (off >= insn_cnt || off < 0) {
1158 verbose(env, "call to invalid destination\n");
1161 ret = find_subprog(env, off);
1164 if (env->subprog_cnt >= BPF_MAX_SUBPROGS) {
1165 verbose(env, "too many subprograms\n");
1168 env->subprog_info[env->subprog_cnt++].start = off;
1169 sort(env->subprog_info, env->subprog_cnt,
1170 sizeof(env->subprog_info[0]), cmp_subprogs, NULL);
1174 static int check_subprogs(struct bpf_verifier_env *env)
1176 int i, ret, subprog_start, subprog_end, off, cur_subprog = 0;
1177 struct bpf_subprog_info *subprog = env->subprog_info;
1178 struct bpf_insn *insn = env->prog->insnsi;
1179 int insn_cnt = env->prog->len;
1181 /* Add entry function. */
1182 ret = add_subprog(env, 0);
1186 /* determine subprog starts. The end is one before the next starts */
1187 for (i = 0; i < insn_cnt; i++) {
1188 if (insn[i].code != (BPF_JMP | BPF_CALL))
1190 if (insn[i].src_reg != BPF_PSEUDO_CALL)
1192 if (!env->allow_ptr_leaks) {
1193 verbose(env, "function calls to other bpf functions are allowed for root only\n");
1196 ret = add_subprog(env, i + insn[i].imm + 1);
1201 /* Add a fake 'exit' subprog which could simplify subprog iteration
1202 * logic. 'subprog_cnt' should not be increased.
1204 subprog[env->subprog_cnt].start = insn_cnt;
1206 if (env->log.level & BPF_LOG_LEVEL2)
1207 for (i = 0; i < env->subprog_cnt; i++)
1208 verbose(env, "func#%d @%d\n", i, subprog[i].start);
1210 /* now check that all jumps are within the same subprog */
1211 subprog_start = subprog[cur_subprog].start;
1212 subprog_end = subprog[cur_subprog + 1].start;
1213 for (i = 0; i < insn_cnt; i++) {
1214 u8 code = insn[i].code;
1216 if (BPF_CLASS(code) != BPF_JMP && BPF_CLASS(code) != BPF_JMP32)
1218 if (BPF_OP(code) == BPF_EXIT || BPF_OP(code) == BPF_CALL)
1220 off = i + insn[i].off + 1;
1221 if (off < subprog_start || off >= subprog_end) {
1222 verbose(env, "jump out of range from insn %d to %d\n", i, off);
1226 if (i == subprog_end - 1) {
1227 /* to avoid fall-through from one subprog into another
1228 * the last insn of the subprog should be either exit
1229 * or unconditional jump back
1231 if (code != (BPF_JMP | BPF_EXIT) &&
1232 code != (BPF_JMP | BPF_JA)) {
1233 verbose(env, "last insn is not an exit or jmp\n");
1236 subprog_start = subprog_end;
1238 if (cur_subprog < env->subprog_cnt)
1239 subprog_end = subprog[cur_subprog + 1].start;
1245 /* Parentage chain of this register (or stack slot) should take care of all
1246 * issues like callee-saved registers, stack slot allocation time, etc.
1248 static int mark_reg_read(struct bpf_verifier_env *env,
1249 const struct bpf_reg_state *state,
1250 struct bpf_reg_state *parent, u8 flag)
1252 bool writes = parent == state->parent; /* Observe write marks */
1256 /* if read wasn't screened by an earlier write ... */
1257 if (writes && state->live & REG_LIVE_WRITTEN)
1259 if (parent->live & REG_LIVE_DONE) {
1260 verbose(env, "verifier BUG type %s var_off %lld off %d\n",
1261 reg_type_str[parent->type],
1262 parent->var_off.value, parent->off);
1265 /* The first condition is more likely to be true than the
1266 * second, checked it first.
1268 if ((parent->live & REG_LIVE_READ) == flag ||
1269 parent->live & REG_LIVE_READ64)
1270 /* The parentage chain never changes and
1271 * this parent was already marked as LIVE_READ.
1272 * There is no need to keep walking the chain again and
1273 * keep re-marking all parents as LIVE_READ.
1274 * This case happens when the same register is read
1275 * multiple times without writes into it in-between.
1276 * Also, if parent has the stronger REG_LIVE_READ64 set,
1277 * then no need to set the weak REG_LIVE_READ32.
1280 /* ... then we depend on parent's value */
1281 parent->live |= flag;
1282 /* REG_LIVE_READ64 overrides REG_LIVE_READ32. */
1283 if (flag == REG_LIVE_READ64)
1284 parent->live &= ~REG_LIVE_READ32;
1286 parent = state->parent;
1291 if (env->longest_mark_read_walk < cnt)
1292 env->longest_mark_read_walk = cnt;
1296 /* This function is supposed to be used by the following 32-bit optimization
1297 * code only. It returns TRUE if the source or destination register operates
1298 * on 64-bit, otherwise return FALSE.
1300 static bool is_reg64(struct bpf_verifier_env *env, struct bpf_insn *insn,
1301 u32 regno, struct bpf_reg_state *reg, enum reg_arg_type t)
1306 class = BPF_CLASS(code);
1308 if (class == BPF_JMP) {
1309 /* BPF_EXIT for "main" will reach here. Return TRUE
1314 if (op == BPF_CALL) {
1315 /* BPF to BPF call will reach here because of marking
1316 * caller saved clobber with DST_OP_NO_MARK for which we
1317 * don't care the register def because they are anyway
1318 * marked as NOT_INIT already.
1320 if (insn->src_reg == BPF_PSEUDO_CALL)
1322 /* Helper call will reach here because of arg type
1323 * check, conservatively return TRUE.
1332 if (class == BPF_ALU64 || class == BPF_JMP ||
1333 /* BPF_END always use BPF_ALU class. */
1334 (class == BPF_ALU && op == BPF_END && insn->imm == 64))
1337 if (class == BPF_ALU || class == BPF_JMP32)
1340 if (class == BPF_LDX) {
1342 return BPF_SIZE(code) == BPF_DW;
1343 /* LDX source must be ptr. */
1347 if (class == BPF_STX) {
1348 if (reg->type != SCALAR_VALUE)
1350 return BPF_SIZE(code) == BPF_DW;
1353 if (class == BPF_LD) {
1354 u8 mode = BPF_MODE(code);
1357 if (mode == BPF_IMM)
1360 /* Both LD_IND and LD_ABS return 32-bit data. */
1364 /* Implicit ctx ptr. */
1365 if (regno == BPF_REG_6)
1368 /* Explicit source could be any width. */
1372 if (class == BPF_ST)
1373 /* The only source register for BPF_ST is a ptr. */
1376 /* Conservatively return true at default. */
1380 /* Return TRUE if INSN doesn't have explicit value define. */
1381 static bool insn_no_def(struct bpf_insn *insn)
1383 u8 class = BPF_CLASS(insn->code);
1385 return (class == BPF_JMP || class == BPF_JMP32 ||
1386 class == BPF_STX || class == BPF_ST);
1389 /* Return TRUE if INSN has defined any 32-bit value explicitly. */
1390 static bool insn_has_def32(struct bpf_verifier_env *env, struct bpf_insn *insn)
1392 if (insn_no_def(insn))
1395 return !is_reg64(env, insn, insn->dst_reg, NULL, DST_OP);
1398 static void mark_insn_zext(struct bpf_verifier_env *env,
1399 struct bpf_reg_state *reg)
1401 s32 def_idx = reg->subreg_def;
1403 if (def_idx == DEF_NOT_SUBREG)
1406 env->insn_aux_data[def_idx - 1].zext_dst = true;
1407 /* The dst will be zero extended, so won't be sub-register anymore. */
1408 reg->subreg_def = DEF_NOT_SUBREG;
1411 static int check_reg_arg(struct bpf_verifier_env *env, u32 regno,
1412 enum reg_arg_type t)
1414 struct bpf_verifier_state *vstate = env->cur_state;
1415 struct bpf_func_state *state = vstate->frame[vstate->curframe];
1416 struct bpf_insn *insn = env->prog->insnsi + env->insn_idx;
1417 struct bpf_reg_state *reg, *regs = state->regs;
1420 if (regno >= MAX_BPF_REG) {
1421 verbose(env, "R%d is invalid\n", regno);
1426 rw64 = is_reg64(env, insn, regno, reg, t);
1428 /* check whether register used as source operand can be read */
1429 if (reg->type == NOT_INIT) {
1430 verbose(env, "R%d !read_ok\n", regno);
1433 /* We don't need to worry about FP liveness because it's read-only */
1434 if (regno == BPF_REG_FP)
1438 mark_insn_zext(env, reg);
1440 return mark_reg_read(env, reg, reg->parent,
1441 rw64 ? REG_LIVE_READ64 : REG_LIVE_READ32);
1443 /* check whether register used as dest operand can be written to */
1444 if (regno == BPF_REG_FP) {
1445 verbose(env, "frame pointer is read only\n");
1448 reg->live |= REG_LIVE_WRITTEN;
1449 reg->subreg_def = rw64 ? DEF_NOT_SUBREG : env->insn_idx + 1;
1451 mark_reg_unknown(env, regs, regno);
1456 /* for any branch, call, exit record the history of jmps in the given state */
1457 static int push_jmp_history(struct bpf_verifier_env *env,
1458 struct bpf_verifier_state *cur)
1460 u32 cnt = cur->jmp_history_cnt;
1461 struct bpf_idx_pair *p;
1464 p = krealloc(cur->jmp_history, cnt * sizeof(*p), GFP_USER);
1467 p[cnt - 1].idx = env->insn_idx;
1468 p[cnt - 1].prev_idx = env->prev_insn_idx;
1469 cur->jmp_history = p;
1470 cur->jmp_history_cnt = cnt;
1474 /* Backtrack one insn at a time. If idx is not at the top of recorded
1475 * history then previous instruction came from straight line execution.
1477 static int get_prev_insn_idx(struct bpf_verifier_state *st, int i,
1482 if (cnt && st->jmp_history[cnt - 1].idx == i) {
1483 i = st->jmp_history[cnt - 1].prev_idx;
1491 /* For given verifier state backtrack_insn() is called from the last insn to
1492 * the first insn. Its purpose is to compute a bitmask of registers and
1493 * stack slots that needs precision in the parent verifier state.
1495 static int backtrack_insn(struct bpf_verifier_env *env, int idx,
1496 u32 *reg_mask, u64 *stack_mask)
1498 const struct bpf_insn_cbs cbs = {
1499 .cb_print = verbose,
1500 .private_data = env,
1502 struct bpf_insn *insn = env->prog->insnsi + idx;
1503 u8 class = BPF_CLASS(insn->code);
1504 u8 opcode = BPF_OP(insn->code);
1505 u8 mode = BPF_MODE(insn->code);
1506 u32 dreg = 1u << insn->dst_reg;
1507 u32 sreg = 1u << insn->src_reg;
1510 if (insn->code == 0)
1512 if (env->log.level & BPF_LOG_LEVEL) {
1513 verbose(env, "regs=%x stack=%llx before ", *reg_mask, *stack_mask);
1514 verbose(env, "%d: ", idx);
1515 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
1518 if (class == BPF_ALU || class == BPF_ALU64) {
1519 if (!(*reg_mask & dreg))
1521 if (opcode == BPF_MOV) {
1522 if (BPF_SRC(insn->code) == BPF_X) {
1524 * dreg needs precision after this insn
1525 * sreg needs precision before this insn
1531 * dreg needs precision after this insn.
1532 * Corresponding register is already marked
1533 * as precise=true in this verifier state.
1534 * No further markings in parent are necessary
1539 if (BPF_SRC(insn->code) == BPF_X) {
1541 * both dreg and sreg need precision
1546 * dreg still needs precision before this insn
1549 } else if (class == BPF_LDX) {
1550 if (!(*reg_mask & dreg))
1554 /* scalars can only be spilled into stack w/o losing precision.
1555 * Load from any other memory can be zero extended.
1556 * The desire to keep that precision is already indicated
1557 * by 'precise' mark in corresponding register of this state.
1558 * No further tracking necessary.
1560 if (insn->src_reg != BPF_REG_FP)
1562 if (BPF_SIZE(insn->code) != BPF_DW)
1565 /* dreg = *(u64 *)[fp - off] was a fill from the stack.
1566 * that [fp - off] slot contains scalar that needs to be
1567 * tracked with precision
1569 spi = (-insn->off - 1) / BPF_REG_SIZE;
1571 verbose(env, "BUG spi %d\n", spi);
1572 WARN_ONCE(1, "verifier backtracking bug");
1575 *stack_mask |= 1ull << spi;
1576 } else if (class == BPF_STX || class == BPF_ST) {
1577 if (*reg_mask & dreg)
1578 /* stx & st shouldn't be using _scalar_ dst_reg
1579 * to access memory. It means backtracking
1580 * encountered a case of pointer subtraction.
1583 /* scalars can only be spilled into stack */
1584 if (insn->dst_reg != BPF_REG_FP)
1586 if (BPF_SIZE(insn->code) != BPF_DW)
1588 spi = (-insn->off - 1) / BPF_REG_SIZE;
1590 verbose(env, "BUG spi %d\n", spi);
1591 WARN_ONCE(1, "verifier backtracking bug");
1594 if (!(*stack_mask & (1ull << spi)))
1596 *stack_mask &= ~(1ull << spi);
1597 if (class == BPF_STX)
1599 } else if (class == BPF_JMP || class == BPF_JMP32) {
1600 if (opcode == BPF_CALL) {
1601 if (insn->src_reg == BPF_PSEUDO_CALL)
1603 /* regular helper call sets R0 */
1605 if (*reg_mask & 0x3f) {
1606 /* if backtracing was looking for registers R1-R5
1607 * they should have been found already.
1609 verbose(env, "BUG regs %x\n", *reg_mask);
1610 WARN_ONCE(1, "verifier backtracking bug");
1613 } else if (opcode == BPF_EXIT) {
1616 } else if (class == BPF_LD) {
1617 if (!(*reg_mask & dreg))
1620 /* It's ld_imm64 or ld_abs or ld_ind.
1621 * For ld_imm64 no further tracking of precision
1622 * into parent is necessary
1624 if (mode == BPF_IND || mode == BPF_ABS)
1625 /* to be analyzed */
1631 /* the scalar precision tracking algorithm:
1632 * . at the start all registers have precise=false.
1633 * . scalar ranges are tracked as normal through alu and jmp insns.
1634 * . once precise value of the scalar register is used in:
1635 * . ptr + scalar alu
1636 * . if (scalar cond K|scalar)
1637 * . helper_call(.., scalar, ...) where ARG_CONST is expected
1638 * backtrack through the verifier states and mark all registers and
1639 * stack slots with spilled constants that these scalar regisers
1640 * should be precise.
1641 * . during state pruning two registers (or spilled stack slots)
1642 * are equivalent if both are not precise.
1644 * Note the verifier cannot simply walk register parentage chain,
1645 * since many different registers and stack slots could have been
1646 * used to compute single precise scalar.
1648 * The approach of starting with precise=true for all registers and then
1649 * backtrack to mark a register as not precise when the verifier detects
1650 * that program doesn't care about specific value (e.g., when helper
1651 * takes register as ARG_ANYTHING parameter) is not safe.
1653 * It's ok to walk single parentage chain of the verifier states.
1654 * It's possible that this backtracking will go all the way till 1st insn.
1655 * All other branches will be explored for needing precision later.
1657 * The backtracking needs to deal with cases like:
1658 * 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)
1661 * if r5 > 0x79f goto pc+7
1662 * R5_w=inv(id=0,umax_value=1951,var_off=(0x0; 0x7ff))
1665 * call bpf_perf_event_output#25
1666 * where .arg5_type = ARG_CONST_SIZE_OR_ZERO
1670 * call foo // uses callee's r6 inside to compute r0
1674 * to track above reg_mask/stack_mask needs to be independent for each frame.
1676 * Also if parent's curframe > frame where backtracking started,
1677 * the verifier need to mark registers in both frames, otherwise callees
1678 * may incorrectly prune callers. This is similar to
1679 * commit 7640ead93924 ("bpf: verifier: make sure callees don't prune with caller differences")
1681 * For now backtracking falls back into conservative marking.
1683 static void mark_all_scalars_precise(struct bpf_verifier_env *env,
1684 struct bpf_verifier_state *st)
1686 struct bpf_func_state *func;
1687 struct bpf_reg_state *reg;
1690 /* big hammer: mark all scalars precise in this path.
1691 * pop_stack may still get !precise scalars.
1693 for (; st; st = st->parent)
1694 for (i = 0; i <= st->curframe; i++) {
1695 func = st->frame[i];
1696 for (j = 0; j < BPF_REG_FP; j++) {
1697 reg = &func->regs[j];
1698 if (reg->type != SCALAR_VALUE)
1700 reg->precise = true;
1702 for (j = 0; j < func->allocated_stack / BPF_REG_SIZE; j++) {
1703 if (func->stack[j].slot_type[0] != STACK_SPILL)
1705 reg = &func->stack[j].spilled_ptr;
1706 if (reg->type != SCALAR_VALUE)
1708 reg->precise = true;
1713 static int __mark_chain_precision(struct bpf_verifier_env *env, int regno,
1716 struct bpf_verifier_state *st = env->cur_state;
1717 int first_idx = st->first_insn_idx;
1718 int last_idx = env->insn_idx;
1719 struct bpf_func_state *func;
1720 struct bpf_reg_state *reg;
1721 u32 reg_mask = regno >= 0 ? 1u << regno : 0;
1722 u64 stack_mask = spi >= 0 ? 1ull << spi : 0;
1723 bool skip_first = true;
1724 bool new_marks = false;
1727 if (!env->allow_ptr_leaks)
1728 /* backtracking is root only for now */
1731 func = st->frame[st->curframe];
1733 reg = &func->regs[regno];
1734 if (reg->type != SCALAR_VALUE) {
1735 WARN_ONCE(1, "backtracing misuse");
1742 reg->precise = true;
1746 if (func->stack[spi].slot_type[0] != STACK_SPILL) {
1750 reg = &func->stack[spi].spilled_ptr;
1751 if (reg->type != SCALAR_VALUE) {
1759 reg->precise = true;
1765 if (!reg_mask && !stack_mask)
1768 DECLARE_BITMAP(mask, 64);
1769 u32 history = st->jmp_history_cnt;
1771 if (env->log.level & BPF_LOG_LEVEL)
1772 verbose(env, "last_idx %d first_idx %d\n", last_idx, first_idx);
1773 for (i = last_idx;;) {
1778 err = backtrack_insn(env, i, ®_mask, &stack_mask);
1780 if (err == -ENOTSUPP) {
1781 mark_all_scalars_precise(env, st);
1786 if (!reg_mask && !stack_mask)
1787 /* Found assignment(s) into tracked register in this state.
1788 * Since this state is already marked, just return.
1789 * Nothing to be tracked further in the parent state.
1794 i = get_prev_insn_idx(st, i, &history);
1795 if (i >= env->prog->len) {
1796 /* This can happen if backtracking reached insn 0
1797 * and there are still reg_mask or stack_mask
1799 * It means the backtracking missed the spot where
1800 * particular register was initialized with a constant.
1802 verbose(env, "BUG backtracking idx %d\n", i);
1803 WARN_ONCE(1, "verifier backtracking bug");
1812 func = st->frame[st->curframe];
1813 bitmap_from_u64(mask, reg_mask);
1814 for_each_set_bit(i, mask, 32) {
1815 reg = &func->regs[i];
1816 if (reg->type != SCALAR_VALUE) {
1817 reg_mask &= ~(1u << i);
1822 reg->precise = true;
1825 bitmap_from_u64(mask, stack_mask);
1826 for_each_set_bit(i, mask, 64) {
1827 if (i >= func->allocated_stack / BPF_REG_SIZE) {
1828 /* the sequence of instructions:
1830 * 3: (7b) *(u64 *)(r3 -8) = r0
1831 * 4: (79) r4 = *(u64 *)(r10 -8)
1832 * doesn't contain jmps. It's backtracked
1833 * as a single block.
1834 * During backtracking insn 3 is not recognized as
1835 * stack access, so at the end of backtracking
1836 * stack slot fp-8 is still marked in stack_mask.
1837 * However the parent state may not have accessed
1838 * fp-8 and it's "unallocated" stack space.
1839 * In such case fallback to conservative.
1841 mark_all_scalars_precise(env, st);
1845 if (func->stack[i].slot_type[0] != STACK_SPILL) {
1846 stack_mask &= ~(1ull << i);
1849 reg = &func->stack[i].spilled_ptr;
1850 if (reg->type != SCALAR_VALUE) {
1851 stack_mask &= ~(1ull << i);
1856 reg->precise = true;
1858 if (env->log.level & BPF_LOG_LEVEL) {
1859 print_verifier_state(env, func);
1860 verbose(env, "parent %s regs=%x stack=%llx marks\n",
1861 new_marks ? "didn't have" : "already had",
1862 reg_mask, stack_mask);
1865 if (!reg_mask && !stack_mask)
1870 last_idx = st->last_insn_idx;
1871 first_idx = st->first_insn_idx;
1876 static int mark_chain_precision(struct bpf_verifier_env *env, int regno)
1878 return __mark_chain_precision(env, regno, -1);
1881 static int mark_chain_precision_stack(struct bpf_verifier_env *env, int spi)
1883 return __mark_chain_precision(env, -1, spi);
1886 static bool is_spillable_regtype(enum bpf_reg_type type)
1889 case PTR_TO_MAP_VALUE:
1890 case PTR_TO_MAP_VALUE_OR_NULL:
1894 case PTR_TO_PACKET_META:
1895 case PTR_TO_PACKET_END:
1896 case PTR_TO_FLOW_KEYS:
1897 case CONST_PTR_TO_MAP:
1899 case PTR_TO_SOCKET_OR_NULL:
1900 case PTR_TO_SOCK_COMMON:
1901 case PTR_TO_SOCK_COMMON_OR_NULL:
1902 case PTR_TO_TCP_SOCK:
1903 case PTR_TO_TCP_SOCK_OR_NULL:
1904 case PTR_TO_XDP_SOCK:
1912 /* Does this register contain a constant zero? */
1913 static bool register_is_null(struct bpf_reg_state *reg)
1915 return reg->type == SCALAR_VALUE && tnum_equals_const(reg->var_off, 0);
1918 static bool register_is_const(struct bpf_reg_state *reg)
1920 return reg->type == SCALAR_VALUE && tnum_is_const(reg->var_off);
1923 static void save_register_state(struct bpf_func_state *state,
1924 int spi, struct bpf_reg_state *reg)
1928 state->stack[spi].spilled_ptr = *reg;
1929 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
1931 for (i = 0; i < BPF_REG_SIZE; i++)
1932 state->stack[spi].slot_type[i] = STACK_SPILL;
1935 /* check_stack_read/write functions track spill/fill of registers,
1936 * stack boundary and alignment are checked in check_mem_access()
1938 static int check_stack_write(struct bpf_verifier_env *env,
1939 struct bpf_func_state *state, /* func where register points to */
1940 int off, int size, int value_regno, int insn_idx)
1942 struct bpf_func_state *cur; /* state of the current function */
1943 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE, err;
1944 u32 dst_reg = env->prog->insnsi[insn_idx].dst_reg;
1945 struct bpf_reg_state *reg = NULL;
1947 err = realloc_func_state(state, round_up(slot + 1, BPF_REG_SIZE),
1948 state->acquired_refs, true);
1951 /* caller checked that off % size == 0 and -MAX_BPF_STACK <= off < 0,
1952 * so it's aligned access and [off, off + size) are within stack limits
1954 if (!env->allow_ptr_leaks &&
1955 state->stack[spi].slot_type[0] == STACK_SPILL &&
1956 size != BPF_REG_SIZE) {
1957 verbose(env, "attempt to corrupt spilled pointer on stack\n");
1961 cur = env->cur_state->frame[env->cur_state->curframe];
1962 if (value_regno >= 0)
1963 reg = &cur->regs[value_regno];
1965 if (reg && size == BPF_REG_SIZE && register_is_const(reg) &&
1966 !register_is_null(reg) && env->allow_ptr_leaks) {
1967 if (dst_reg != BPF_REG_FP) {
1968 /* The backtracking logic can only recognize explicit
1969 * stack slot address like [fp - 8]. Other spill of
1970 * scalar via different register has to be conervative.
1971 * Backtrack from here and mark all registers as precise
1972 * that contributed into 'reg' being a constant.
1974 err = mark_chain_precision(env, value_regno);
1978 save_register_state(state, spi, reg);
1979 } else if (reg && is_spillable_regtype(reg->type)) {
1980 /* register containing pointer is being spilled into stack */
1981 if (size != BPF_REG_SIZE) {
1982 verbose_linfo(env, insn_idx, "; ");
1983 verbose(env, "invalid size of register spill\n");
1987 if (state != cur && reg->type == PTR_TO_STACK) {
1988 verbose(env, "cannot spill pointers to stack into stack frame of the caller\n");
1992 if (!env->allow_ptr_leaks) {
1993 bool sanitize = false;
1995 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
1996 register_is_const(&state->stack[spi].spilled_ptr))
1998 for (i = 0; i < BPF_REG_SIZE; i++)
1999 if (state->stack[spi].slot_type[i] == STACK_MISC) {
2004 int *poff = &env->insn_aux_data[insn_idx].sanitize_stack_off;
2005 int soff = (-spi - 1) * BPF_REG_SIZE;
2007 /* detected reuse of integer stack slot with a pointer
2008 * which means either llvm is reusing stack slot or
2009 * an attacker is trying to exploit CVE-2018-3639
2010 * (speculative store bypass)
2011 * Have to sanitize that slot with preemptive
2014 if (*poff && *poff != soff) {
2015 /* disallow programs where single insn stores
2016 * into two different stack slots, since verifier
2017 * cannot sanitize them
2020 "insn %d cannot access two stack slots fp%d and fp%d",
2021 insn_idx, *poff, soff);
2027 save_register_state(state, spi, reg);
2029 u8 type = STACK_MISC;
2031 /* regular write of data into stack destroys any spilled ptr */
2032 state->stack[spi].spilled_ptr.type = NOT_INIT;
2033 /* Mark slots as STACK_MISC if they belonged to spilled ptr. */
2034 if (state->stack[spi].slot_type[0] == STACK_SPILL)
2035 for (i = 0; i < BPF_REG_SIZE; i++)
2036 state->stack[spi].slot_type[i] = STACK_MISC;
2038 /* only mark the slot as written if all 8 bytes were written
2039 * otherwise read propagation may incorrectly stop too soon
2040 * when stack slots are partially written.
2041 * This heuristic means that read propagation will be
2042 * conservative, since it will add reg_live_read marks
2043 * to stack slots all the way to first state when programs
2044 * writes+reads less than 8 bytes
2046 if (size == BPF_REG_SIZE)
2047 state->stack[spi].spilled_ptr.live |= REG_LIVE_WRITTEN;
2049 /* when we zero initialize stack slots mark them as such */
2050 if (reg && register_is_null(reg)) {
2051 /* backtracking doesn't work for STACK_ZERO yet. */
2052 err = mark_chain_precision(env, value_regno);
2058 /* Mark slots affected by this stack write. */
2059 for (i = 0; i < size; i++)
2060 state->stack[spi].slot_type[(slot - i) % BPF_REG_SIZE] =
2066 static int check_stack_read(struct bpf_verifier_env *env,
2067 struct bpf_func_state *reg_state /* func where register points to */,
2068 int off, int size, int value_regno)
2070 struct bpf_verifier_state *vstate = env->cur_state;
2071 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2072 int i, slot = -off - 1, spi = slot / BPF_REG_SIZE;
2073 struct bpf_reg_state *reg;
2076 if (reg_state->allocated_stack <= slot) {
2077 verbose(env, "invalid read from stack off %d+0 size %d\n",
2081 stype = reg_state->stack[spi].slot_type;
2082 reg = ®_state->stack[spi].spilled_ptr;
2084 if (stype[0] == STACK_SPILL) {
2085 if (size != BPF_REG_SIZE) {
2086 if (reg->type != SCALAR_VALUE) {
2087 verbose_linfo(env, env->insn_idx, "; ");
2088 verbose(env, "invalid size of register fill\n");
2091 if (value_regno >= 0) {
2092 mark_reg_unknown(env, state->regs, value_regno);
2093 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
2095 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2098 for (i = 1; i < BPF_REG_SIZE; i++) {
2099 if (stype[(slot - i) % BPF_REG_SIZE] != STACK_SPILL) {
2100 verbose(env, "corrupted spill memory\n");
2105 if (value_regno >= 0) {
2106 /* restore register state from stack */
2107 state->regs[value_regno] = *reg;
2108 /* mark reg as written since spilled pointer state likely
2109 * has its liveness marks cleared by is_state_visited()
2110 * which resets stack/reg liveness for state transitions
2112 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
2114 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2118 for (i = 0; i < size; i++) {
2119 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_MISC)
2121 if (stype[(slot - i) % BPF_REG_SIZE] == STACK_ZERO) {
2125 verbose(env, "invalid read from stack off %d+%d size %d\n",
2129 mark_reg_read(env, reg, reg->parent, REG_LIVE_READ64);
2130 if (value_regno >= 0) {
2131 if (zeros == size) {
2132 /* any size read into register is zero extended,
2133 * so the whole register == const_zero
2135 __mark_reg_const_zero(&state->regs[value_regno]);
2136 /* backtracking doesn't support STACK_ZERO yet,
2137 * so mark it precise here, so that later
2138 * backtracking can stop here.
2139 * Backtracking may not need this if this register
2140 * doesn't participate in pointer adjustment.
2141 * Forward propagation of precise flag is not
2142 * necessary either. This mark is only to stop
2143 * backtracking. Any register that contributed
2144 * to const 0 was marked precise before spill.
2146 state->regs[value_regno].precise = true;
2148 /* have read misc data from the stack */
2149 mark_reg_unknown(env, state->regs, value_regno);
2151 state->regs[value_regno].live |= REG_LIVE_WRITTEN;
2157 static int check_stack_access(struct bpf_verifier_env *env,
2158 const struct bpf_reg_state *reg,
2161 /* Stack accesses must be at a fixed offset, so that we
2162 * can determine what type of data were returned. See
2163 * check_stack_read().
2165 if (!tnum_is_const(reg->var_off)) {
2168 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2169 verbose(env, "variable stack access var_off=%s off=%d size=%d\n",
2174 if (off >= 0 || off < -MAX_BPF_STACK) {
2175 verbose(env, "invalid stack off=%d size=%d\n", off, size);
2182 static int check_map_access_type(struct bpf_verifier_env *env, u32 regno,
2183 int off, int size, enum bpf_access_type type)
2185 struct bpf_reg_state *regs = cur_regs(env);
2186 struct bpf_map *map = regs[regno].map_ptr;
2187 u32 cap = bpf_map_flags_to_cap(map);
2189 if (type == BPF_WRITE && !(cap & BPF_MAP_CAN_WRITE)) {
2190 verbose(env, "write into map forbidden, value_size=%d off=%d size=%d\n",
2191 map->value_size, off, size);
2195 if (type == BPF_READ && !(cap & BPF_MAP_CAN_READ)) {
2196 verbose(env, "read from map forbidden, value_size=%d off=%d size=%d\n",
2197 map->value_size, off, size);
2204 /* check read/write into map element returned by bpf_map_lookup_elem() */
2205 static int __check_map_access(struct bpf_verifier_env *env, u32 regno, int off,
2206 int size, bool zero_size_allowed)
2208 struct bpf_reg_state *regs = cur_regs(env);
2209 struct bpf_map *map = regs[regno].map_ptr;
2211 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
2212 off + size > map->value_size) {
2213 verbose(env, "invalid access to map value, value_size=%d off=%d size=%d\n",
2214 map->value_size, off, size);
2220 /* check read/write into a map element with possible variable offset */
2221 static int check_map_access(struct bpf_verifier_env *env, u32 regno,
2222 int off, int size, bool zero_size_allowed)
2224 struct bpf_verifier_state *vstate = env->cur_state;
2225 struct bpf_func_state *state = vstate->frame[vstate->curframe];
2226 struct bpf_reg_state *reg = &state->regs[regno];
2229 /* We may have adjusted the register to this map value, so we
2230 * need to try adding each of min_value and max_value to off
2231 * to make sure our theoretical access will be safe.
2233 if (env->log.level & BPF_LOG_LEVEL)
2234 print_verifier_state(env, state);
2236 /* The minimum value is only important with signed
2237 * comparisons where we can't assume the floor of a
2238 * value is 0. If we are using signed variables for our
2239 * index'es we need to make sure that whatever we use
2240 * will have a set floor within our range.
2242 if (reg->smin_value < 0 &&
2243 (reg->smin_value == S64_MIN ||
2244 (off + reg->smin_value != (s64)(s32)(off + reg->smin_value)) ||
2245 reg->smin_value + off < 0)) {
2246 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2250 err = __check_map_access(env, regno, reg->smin_value + off, size,
2253 verbose(env, "R%d min value is outside of the array range\n",
2258 /* If we haven't set a max value then we need to bail since we can't be
2259 * sure we won't do bad things.
2260 * If reg->umax_value + off could overflow, treat that as unbounded too.
2262 if (reg->umax_value >= BPF_MAX_VAR_OFF) {
2263 verbose(env, "R%d unbounded memory access, make sure to bounds check any array access into a map\n",
2267 err = __check_map_access(env, regno, reg->umax_value + off, size,
2270 verbose(env, "R%d max value is outside of the array range\n",
2273 if (map_value_has_spin_lock(reg->map_ptr)) {
2274 u32 lock = reg->map_ptr->spin_lock_off;
2276 /* if any part of struct bpf_spin_lock can be touched by
2277 * load/store reject this program.
2278 * To check that [x1, x2) overlaps with [y1, y2)
2279 * it is sufficient to check x1 < y2 && y1 < x2.
2281 if (reg->smin_value + off < lock + sizeof(struct bpf_spin_lock) &&
2282 lock < reg->umax_value + off + size) {
2283 verbose(env, "bpf_spin_lock cannot be accessed directly by load/store\n");
2290 #define MAX_PACKET_OFF 0xffff
2292 static bool may_access_direct_pkt_data(struct bpf_verifier_env *env,
2293 const struct bpf_call_arg_meta *meta,
2294 enum bpf_access_type t)
2296 switch (env->prog->type) {
2297 /* Program types only with direct read access go here! */
2298 case BPF_PROG_TYPE_LWT_IN:
2299 case BPF_PROG_TYPE_LWT_OUT:
2300 case BPF_PROG_TYPE_LWT_SEG6LOCAL:
2301 case BPF_PROG_TYPE_SK_REUSEPORT:
2302 case BPF_PROG_TYPE_FLOW_DISSECTOR:
2303 case BPF_PROG_TYPE_CGROUP_SKB:
2308 /* Program types with direct read + write access go here! */
2309 case BPF_PROG_TYPE_SCHED_CLS:
2310 case BPF_PROG_TYPE_SCHED_ACT:
2311 case BPF_PROG_TYPE_XDP:
2312 case BPF_PROG_TYPE_LWT_XMIT:
2313 case BPF_PROG_TYPE_SK_SKB:
2314 case BPF_PROG_TYPE_SK_MSG:
2316 return meta->pkt_access;
2318 env->seen_direct_write = true;
2321 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
2323 env->seen_direct_write = true;
2332 static int __check_packet_access(struct bpf_verifier_env *env, u32 regno,
2333 int off, int size, bool zero_size_allowed)
2335 struct bpf_reg_state *regs = cur_regs(env);
2336 struct bpf_reg_state *reg = ®s[regno];
2338 if (off < 0 || size < 0 || (size == 0 && !zero_size_allowed) ||
2339 (u64)off + size > reg->range) {
2340 verbose(env, "invalid access to packet, off=%d size=%d, R%d(id=%d,off=%d,r=%d)\n",
2341 off, size, regno, reg->id, reg->off, reg->range);
2347 static int check_packet_access(struct bpf_verifier_env *env, u32 regno, int off,
2348 int size, bool zero_size_allowed)
2350 struct bpf_reg_state *regs = cur_regs(env);
2351 struct bpf_reg_state *reg = ®s[regno];
2354 /* We may have added a variable offset to the packet pointer; but any
2355 * reg->range we have comes after that. We are only checking the fixed
2359 /* We don't allow negative numbers, because we aren't tracking enough
2360 * detail to prove they're safe.
2362 if (reg->smin_value < 0) {
2363 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2367 err = __check_packet_access(env, regno, off, size, zero_size_allowed);
2369 verbose(env, "R%d offset is outside of the packet\n", regno);
2373 /* __check_packet_access has made sure "off + size - 1" is within u16.
2374 * reg->umax_value can't be bigger than MAX_PACKET_OFF which is 0xffff,
2375 * otherwise find_good_pkt_pointers would have refused to set range info
2376 * that __check_packet_access would have rejected this pkt access.
2377 * Therefore, "off + reg->umax_value + size - 1" won't overflow u32.
2379 env->prog->aux->max_pkt_offset =
2380 max_t(u32, env->prog->aux->max_pkt_offset,
2381 off + reg->umax_value + size - 1);
2386 /* check access to 'struct bpf_context' fields. Supports fixed offsets only */
2387 static int check_ctx_access(struct bpf_verifier_env *env, int insn_idx, int off, int size,
2388 enum bpf_access_type t, enum bpf_reg_type *reg_type,
2391 struct bpf_insn_access_aux info = {
2392 .reg_type = *reg_type,
2396 if (env->ops->is_valid_access &&
2397 env->ops->is_valid_access(off, size, t, env->prog, &info)) {
2398 /* A non zero info.ctx_field_size indicates that this field is a
2399 * candidate for later verifier transformation to load the whole
2400 * field and then apply a mask when accessed with a narrower
2401 * access than actual ctx access size. A zero info.ctx_field_size
2402 * will only allow for whole field access and rejects any other
2403 * type of narrower access.
2405 *reg_type = info.reg_type;
2407 if (*reg_type == PTR_TO_BTF_ID)
2408 *btf_id = info.btf_id;
2410 env->insn_aux_data[insn_idx].ctx_field_size = info.ctx_field_size;
2411 /* remember the offset of last byte accessed in ctx */
2412 if (env->prog->aux->max_ctx_offset < off + size)
2413 env->prog->aux->max_ctx_offset = off + size;
2417 verbose(env, "invalid bpf_context access off=%d size=%d\n", off, size);
2421 static int check_flow_keys_access(struct bpf_verifier_env *env, int off,
2424 if (size < 0 || off < 0 ||
2425 (u64)off + size > sizeof(struct bpf_flow_keys)) {
2426 verbose(env, "invalid access to flow keys off=%d size=%d\n",
2433 static int check_sock_access(struct bpf_verifier_env *env, int insn_idx,
2434 u32 regno, int off, int size,
2435 enum bpf_access_type t)
2437 struct bpf_reg_state *regs = cur_regs(env);
2438 struct bpf_reg_state *reg = ®s[regno];
2439 struct bpf_insn_access_aux info = {};
2442 if (reg->smin_value < 0) {
2443 verbose(env, "R%d min value is negative, either use unsigned index or do a if (index >=0) check.\n",
2448 switch (reg->type) {
2449 case PTR_TO_SOCK_COMMON:
2450 valid = bpf_sock_common_is_valid_access(off, size, t, &info);
2453 valid = bpf_sock_is_valid_access(off, size, t, &info);
2455 case PTR_TO_TCP_SOCK:
2456 valid = bpf_tcp_sock_is_valid_access(off, size, t, &info);
2458 case PTR_TO_XDP_SOCK:
2459 valid = bpf_xdp_sock_is_valid_access(off, size, t, &info);
2467 env->insn_aux_data[insn_idx].ctx_field_size =
2468 info.ctx_field_size;
2472 verbose(env, "R%d invalid %s access off=%d size=%d\n",
2473 regno, reg_type_str[reg->type], off, size);
2478 static bool __is_pointer_value(bool allow_ptr_leaks,
2479 const struct bpf_reg_state *reg)
2481 if (allow_ptr_leaks)
2484 return reg->type != SCALAR_VALUE;
2487 static struct bpf_reg_state *reg_state(struct bpf_verifier_env *env, int regno)
2489 return cur_regs(env) + regno;
2492 static bool is_pointer_value(struct bpf_verifier_env *env, int regno)
2494 return __is_pointer_value(env->allow_ptr_leaks, reg_state(env, regno));
2497 static bool is_ctx_reg(struct bpf_verifier_env *env, int regno)
2499 const struct bpf_reg_state *reg = reg_state(env, regno);
2501 return reg->type == PTR_TO_CTX;
2504 static bool is_sk_reg(struct bpf_verifier_env *env, int regno)
2506 const struct bpf_reg_state *reg = reg_state(env, regno);
2508 return type_is_sk_pointer(reg->type);
2511 static bool is_pkt_reg(struct bpf_verifier_env *env, int regno)
2513 const struct bpf_reg_state *reg = reg_state(env, regno);
2515 return type_is_pkt_pointer(reg->type);
2518 static bool is_flow_key_reg(struct bpf_verifier_env *env, int regno)
2520 const struct bpf_reg_state *reg = reg_state(env, regno);
2522 /* Separate to is_ctx_reg() since we still want to allow BPF_ST here. */
2523 return reg->type == PTR_TO_FLOW_KEYS;
2526 static int check_pkt_ptr_alignment(struct bpf_verifier_env *env,
2527 const struct bpf_reg_state *reg,
2528 int off, int size, bool strict)
2530 struct tnum reg_off;
2533 /* Byte size accesses are always allowed. */
2534 if (!strict || size == 1)
2537 /* For platforms that do not have a Kconfig enabling
2538 * CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS the value of
2539 * NET_IP_ALIGN is universally set to '2'. And on platforms
2540 * that do set CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS, we get
2541 * to this code only in strict mode where we want to emulate
2542 * the NET_IP_ALIGN==2 checking. Therefore use an
2543 * unconditional IP align value of '2'.
2547 reg_off = tnum_add(reg->var_off, tnum_const(ip_align + reg->off + off));
2548 if (!tnum_is_aligned(reg_off, size)) {
2551 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2553 "misaligned packet access off %d+%s+%d+%d size %d\n",
2554 ip_align, tn_buf, reg->off, off, size);
2561 static int check_generic_ptr_alignment(struct bpf_verifier_env *env,
2562 const struct bpf_reg_state *reg,
2563 const char *pointer_desc,
2564 int off, int size, bool strict)
2566 struct tnum reg_off;
2568 /* Byte size accesses are always allowed. */
2569 if (!strict || size == 1)
2572 reg_off = tnum_add(reg->var_off, tnum_const(reg->off + off));
2573 if (!tnum_is_aligned(reg_off, size)) {
2576 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2577 verbose(env, "misaligned %saccess off %s+%d+%d size %d\n",
2578 pointer_desc, tn_buf, reg->off, off, size);
2585 static int check_ptr_alignment(struct bpf_verifier_env *env,
2586 const struct bpf_reg_state *reg, int off,
2587 int size, bool strict_alignment_once)
2589 bool strict = env->strict_alignment || strict_alignment_once;
2590 const char *pointer_desc = "";
2592 switch (reg->type) {
2594 case PTR_TO_PACKET_META:
2595 /* Special case, because of NET_IP_ALIGN. Given metadata sits
2596 * right in front, treat it the very same way.
2598 return check_pkt_ptr_alignment(env, reg, off, size, strict);
2599 case PTR_TO_FLOW_KEYS:
2600 pointer_desc = "flow keys ";
2602 case PTR_TO_MAP_VALUE:
2603 pointer_desc = "value ";
2606 pointer_desc = "context ";
2609 pointer_desc = "stack ";
2610 /* The stack spill tracking logic in check_stack_write()
2611 * and check_stack_read() relies on stack accesses being
2617 pointer_desc = "sock ";
2619 case PTR_TO_SOCK_COMMON:
2620 pointer_desc = "sock_common ";
2622 case PTR_TO_TCP_SOCK:
2623 pointer_desc = "tcp_sock ";
2625 case PTR_TO_XDP_SOCK:
2626 pointer_desc = "xdp_sock ";
2631 return check_generic_ptr_alignment(env, reg, pointer_desc, off, size,
2635 static int update_stack_depth(struct bpf_verifier_env *env,
2636 const struct bpf_func_state *func,
2639 u16 stack = env->subprog_info[func->subprogno].stack_depth;
2644 /* update known max for given subprogram */
2645 env->subprog_info[func->subprogno].stack_depth = -off;
2649 /* starting from main bpf function walk all instructions of the function
2650 * and recursively walk all callees that given function can call.
2651 * Ignore jump and exit insns.
2652 * Since recursion is prevented by check_cfg() this algorithm
2653 * only needs a local stack of MAX_CALL_FRAMES to remember callsites
2655 static int check_max_stack_depth(struct bpf_verifier_env *env)
2657 int depth = 0, frame = 0, idx = 0, i = 0, subprog_end;
2658 struct bpf_subprog_info *subprog = env->subprog_info;
2659 struct bpf_insn *insn = env->prog->insnsi;
2660 int ret_insn[MAX_CALL_FRAMES];
2661 int ret_prog[MAX_CALL_FRAMES];
2664 /* round up to 32-bytes, since this is granularity
2665 * of interpreter stack size
2667 depth += round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
2668 if (depth > MAX_BPF_STACK) {
2669 verbose(env, "combined stack size of %d calls is %d. Too large\n",
2674 subprog_end = subprog[idx + 1].start;
2675 for (; i < subprog_end; i++) {
2676 if (insn[i].code != (BPF_JMP | BPF_CALL))
2678 if (insn[i].src_reg != BPF_PSEUDO_CALL)
2680 /* remember insn and function to return to */
2681 ret_insn[frame] = i + 1;
2682 ret_prog[frame] = idx;
2684 /* find the callee */
2685 i = i + insn[i].imm + 1;
2686 idx = find_subprog(env, i);
2688 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
2693 if (frame >= MAX_CALL_FRAMES) {
2694 verbose(env, "the call stack of %d frames is too deep !\n",
2700 /* end of for() loop means the last insn of the 'subprog'
2701 * was reached. Doesn't matter whether it was JA or EXIT
2705 depth -= round_up(max_t(u32, subprog[idx].stack_depth, 1), 32);
2707 i = ret_insn[frame];
2708 idx = ret_prog[frame];
2712 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
2713 static int get_callee_stack_depth(struct bpf_verifier_env *env,
2714 const struct bpf_insn *insn, int idx)
2716 int start = idx + insn->imm + 1, subprog;
2718 subprog = find_subprog(env, start);
2720 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
2724 return env->subprog_info[subprog].stack_depth;
2728 int check_ctx_reg(struct bpf_verifier_env *env,
2729 const struct bpf_reg_state *reg, int regno)
2731 /* Access to ctx or passing it to a helper is only allowed in
2732 * its original, unmodified form.
2736 verbose(env, "dereference of modified ctx ptr R%d off=%d disallowed\n",
2741 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
2744 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2745 verbose(env, "variable ctx access var_off=%s disallowed\n", tn_buf);
2752 static int check_tp_buffer_access(struct bpf_verifier_env *env,
2753 const struct bpf_reg_state *reg,
2754 int regno, int off, int size)
2758 "R%d invalid tracepoint buffer access: off=%d, size=%d",
2762 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
2765 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2767 "R%d invalid variable buffer offset: off=%d, var_off=%s",
2768 regno, off, tn_buf);
2771 if (off + size > env->prog->aux->max_tp_access)
2772 env->prog->aux->max_tp_access = off + size;
2778 /* truncate register to smaller size (in bytes)
2779 * must be called with size < BPF_REG_SIZE
2781 static void coerce_reg_to_size(struct bpf_reg_state *reg, int size)
2785 /* clear high bits in bit representation */
2786 reg->var_off = tnum_cast(reg->var_off, size);
2788 /* fix arithmetic bounds */
2789 mask = ((u64)1 << (size * 8)) - 1;
2790 if ((reg->umin_value & ~mask) == (reg->umax_value & ~mask)) {
2791 reg->umin_value &= mask;
2792 reg->umax_value &= mask;
2794 reg->umin_value = 0;
2795 reg->umax_value = mask;
2797 reg->smin_value = reg->umin_value;
2798 reg->smax_value = reg->umax_value;
2801 static bool bpf_map_is_rdonly(const struct bpf_map *map)
2803 return (map->map_flags & BPF_F_RDONLY_PROG) && map->frozen;
2806 static int bpf_map_direct_read(struct bpf_map *map, int off, int size, u64 *val)
2812 err = map->ops->map_direct_value_addr(map, &addr, off);
2815 ptr = (void *)(long)addr + off;
2819 *val = (u64)*(u8 *)ptr;
2822 *val = (u64)*(u16 *)ptr;
2825 *val = (u64)*(u32 *)ptr;
2836 static int check_ptr_to_btf_access(struct bpf_verifier_env *env,
2837 struct bpf_reg_state *regs,
2838 int regno, int off, int size,
2839 enum bpf_access_type atype,
2842 struct bpf_reg_state *reg = regs + regno;
2843 const struct btf_type *t = btf_type_by_id(btf_vmlinux, reg->btf_id);
2844 const char *tname = btf_name_by_offset(btf_vmlinux, t->name_off);
2850 "R%d is ptr_%s invalid negative access: off=%d\n",
2854 if (!tnum_is_const(reg->var_off) || reg->var_off.value) {
2857 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
2859 "R%d is ptr_%s invalid variable offset: off=%d, var_off=%s\n",
2860 regno, tname, off, tn_buf);
2864 if (env->ops->btf_struct_access) {
2865 ret = env->ops->btf_struct_access(&env->log, t, off, size,
2868 if (atype != BPF_READ) {
2869 verbose(env, "only read is supported\n");
2873 ret = btf_struct_access(&env->log, t, off, size, atype,
2880 if (atype == BPF_READ) {
2881 if (ret == SCALAR_VALUE) {
2882 mark_reg_unknown(env, regs, value_regno);
2885 mark_reg_known_zero(env, regs, value_regno);
2886 regs[value_regno].type = PTR_TO_BTF_ID;
2887 regs[value_regno].btf_id = btf_id;
2893 /* check whether memory at (regno + off) is accessible for t = (read | write)
2894 * if t==write, value_regno is a register which value is stored into memory
2895 * if t==read, value_regno is a register which will receive the value from memory
2896 * if t==write && value_regno==-1, some unknown value is stored into memory
2897 * if t==read && value_regno==-1, don't care what we read from memory
2899 static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regno,
2900 int off, int bpf_size, enum bpf_access_type t,
2901 int value_regno, bool strict_alignment_once)
2903 struct bpf_reg_state *regs = cur_regs(env);
2904 struct bpf_reg_state *reg = regs + regno;
2905 struct bpf_func_state *state;
2908 size = bpf_size_to_bytes(bpf_size);
2912 /* alignment checks will add in reg->off themselves */
2913 err = check_ptr_alignment(env, reg, off, size, strict_alignment_once);
2917 /* for access checks, reg->off is just part of off */
2920 if (reg->type == PTR_TO_MAP_VALUE) {
2921 if (t == BPF_WRITE && value_regno >= 0 &&
2922 is_pointer_value(env, value_regno)) {
2923 verbose(env, "R%d leaks addr into map\n", value_regno);
2926 err = check_map_access_type(env, regno, off, size, t);
2929 err = check_map_access(env, regno, off, size, false);
2930 if (!err && t == BPF_READ && value_regno >= 0) {
2931 struct bpf_map *map = reg->map_ptr;
2933 /* if map is read-only, track its contents as scalars */
2934 if (tnum_is_const(reg->var_off) &&
2935 bpf_map_is_rdonly(map) &&
2936 map->ops->map_direct_value_addr) {
2937 int map_off = off + reg->var_off.value;
2940 err = bpf_map_direct_read(map, map_off, size,
2945 regs[value_regno].type = SCALAR_VALUE;
2946 __mark_reg_known(®s[value_regno], val);
2948 mark_reg_unknown(env, regs, value_regno);
2951 } else if (reg->type == PTR_TO_CTX) {
2952 enum bpf_reg_type reg_type = SCALAR_VALUE;
2955 if (t == BPF_WRITE && value_regno >= 0 &&
2956 is_pointer_value(env, value_regno)) {
2957 verbose(env, "R%d leaks addr into ctx\n", value_regno);
2961 err = check_ctx_reg(env, reg, regno);
2965 err = check_ctx_access(env, insn_idx, off, size, t, ®_type, &btf_id);
2967 verbose_linfo(env, insn_idx, "; ");
2968 if (!err && t == BPF_READ && value_regno >= 0) {
2969 /* ctx access returns either a scalar, or a
2970 * PTR_TO_PACKET[_META,_END]. In the latter
2971 * case, we know the offset is zero.
2973 if (reg_type == SCALAR_VALUE) {
2974 mark_reg_unknown(env, regs, value_regno);
2976 mark_reg_known_zero(env, regs,
2978 if (reg_type_may_be_null(reg_type))
2979 regs[value_regno].id = ++env->id_gen;
2980 /* A load of ctx field could have different
2981 * actual load size with the one encoded in the
2982 * insn. When the dst is PTR, it is for sure not
2985 regs[value_regno].subreg_def = DEF_NOT_SUBREG;
2986 if (reg_type == PTR_TO_BTF_ID)
2987 regs[value_regno].btf_id = btf_id;
2989 regs[value_regno].type = reg_type;
2992 } else if (reg->type == PTR_TO_STACK) {
2993 off += reg->var_off.value;
2994 err = check_stack_access(env, reg, off, size);
2998 state = func(env, reg);
2999 err = update_stack_depth(env, state, off);
3004 err = check_stack_write(env, state, off, size,
3005 value_regno, insn_idx);
3007 err = check_stack_read(env, state, off, size,
3009 } else if (reg_is_pkt_pointer(reg)) {
3010 if (t == BPF_WRITE && !may_access_direct_pkt_data(env, NULL, t)) {
3011 verbose(env, "cannot write into packet\n");
3014 if (t == BPF_WRITE && value_regno >= 0 &&
3015 is_pointer_value(env, value_regno)) {
3016 verbose(env, "R%d leaks addr into packet\n",
3020 err = check_packet_access(env, regno, off, size, false);
3021 if (!err && t == BPF_READ && value_regno >= 0)
3022 mark_reg_unknown(env, regs, value_regno);
3023 } else if (reg->type == PTR_TO_FLOW_KEYS) {
3024 if (t == BPF_WRITE && value_regno >= 0 &&
3025 is_pointer_value(env, value_regno)) {
3026 verbose(env, "R%d leaks addr into flow keys\n",
3031 err = check_flow_keys_access(env, off, size);
3032 if (!err && t == BPF_READ && value_regno >= 0)
3033 mark_reg_unknown(env, regs, value_regno);
3034 } else if (type_is_sk_pointer(reg->type)) {
3035 if (t == BPF_WRITE) {
3036 verbose(env, "R%d cannot write into %s\n",
3037 regno, reg_type_str[reg->type]);
3040 err = check_sock_access(env, insn_idx, regno, off, size, t);
3041 if (!err && value_regno >= 0)
3042 mark_reg_unknown(env, regs, value_regno);
3043 } else if (reg->type == PTR_TO_TP_BUFFER) {
3044 err = check_tp_buffer_access(env, reg, regno, off, size);
3045 if (!err && t == BPF_READ && value_regno >= 0)
3046 mark_reg_unknown(env, regs, value_regno);
3047 } else if (reg->type == PTR_TO_BTF_ID) {
3048 err = check_ptr_to_btf_access(env, regs, regno, off, size, t,
3051 verbose(env, "R%d invalid mem access '%s'\n", regno,
3052 reg_type_str[reg->type]);
3056 if (!err && size < BPF_REG_SIZE && value_regno >= 0 && t == BPF_READ &&
3057 regs[value_regno].type == SCALAR_VALUE) {
3058 /* b/h/w load zero-extends, mark upper bits as known 0 */
3059 coerce_reg_to_size(®s[value_regno], size);
3064 static int check_xadd(struct bpf_verifier_env *env, int insn_idx, struct bpf_insn *insn)
3068 if ((BPF_SIZE(insn->code) != BPF_W && BPF_SIZE(insn->code) != BPF_DW) ||
3070 verbose(env, "BPF_XADD uses reserved fields\n");
3074 /* check src1 operand */
3075 err = check_reg_arg(env, insn->src_reg, SRC_OP);
3079 /* check src2 operand */
3080 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
3084 if (is_pointer_value(env, insn->src_reg)) {
3085 verbose(env, "R%d leaks addr into mem\n", insn->src_reg);
3089 if (is_ctx_reg(env, insn->dst_reg) ||
3090 is_pkt_reg(env, insn->dst_reg) ||
3091 is_flow_key_reg(env, insn->dst_reg) ||
3092 is_sk_reg(env, insn->dst_reg)) {
3093 verbose(env, "BPF_XADD stores into R%d %s is not allowed\n",
3095 reg_type_str[reg_state(env, insn->dst_reg)->type]);
3099 /* check whether atomic_add can read the memory */
3100 err = check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3101 BPF_SIZE(insn->code), BPF_READ, -1, true);
3105 /* check whether atomic_add can write into the same memory */
3106 return check_mem_access(env, insn_idx, insn->dst_reg, insn->off,
3107 BPF_SIZE(insn->code), BPF_WRITE, -1, true);
3110 static int __check_stack_boundary(struct bpf_verifier_env *env, u32 regno,
3111 int off, int access_size,
3112 bool zero_size_allowed)
3114 struct bpf_reg_state *reg = reg_state(env, regno);
3116 if (off >= 0 || off < -MAX_BPF_STACK || off + access_size > 0 ||
3117 access_size < 0 || (access_size == 0 && !zero_size_allowed)) {
3118 if (tnum_is_const(reg->var_off)) {
3119 verbose(env, "invalid stack type R%d off=%d access_size=%d\n",
3120 regno, off, access_size);
3124 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3125 verbose(env, "invalid stack type R%d var_off=%s access_size=%d\n",
3126 regno, tn_buf, access_size);
3133 /* when register 'regno' is passed into function that will read 'access_size'
3134 * bytes from that pointer, make sure that it's within stack boundary
3135 * and all elements of stack are initialized.
3136 * Unlike most pointer bounds-checking functions, this one doesn't take an
3137 * 'off' argument, so it has to add in reg->off itself.
3139 static int check_stack_boundary(struct bpf_verifier_env *env, int regno,
3140 int access_size, bool zero_size_allowed,
3141 struct bpf_call_arg_meta *meta)
3143 struct bpf_reg_state *reg = reg_state(env, regno);
3144 struct bpf_func_state *state = func(env, reg);
3145 int err, min_off, max_off, i, j, slot, spi;
3147 if (reg->type != PTR_TO_STACK) {
3148 /* Allow zero-byte read from NULL, regardless of pointer type */
3149 if (zero_size_allowed && access_size == 0 &&
3150 register_is_null(reg))
3153 verbose(env, "R%d type=%s expected=%s\n", regno,
3154 reg_type_str[reg->type],
3155 reg_type_str[PTR_TO_STACK]);
3159 if (tnum_is_const(reg->var_off)) {
3160 min_off = max_off = reg->var_off.value + reg->off;
3161 err = __check_stack_boundary(env, regno, min_off, access_size,
3166 /* Variable offset is prohibited for unprivileged mode for
3167 * simplicity since it requires corresponding support in
3168 * Spectre masking for stack ALU.
3169 * See also retrieve_ptr_limit().
3171 if (!env->allow_ptr_leaks) {
3174 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3175 verbose(env, "R%d indirect variable offset stack access prohibited for !root, var_off=%s\n",
3179 /* Only initialized buffer on stack is allowed to be accessed
3180 * with variable offset. With uninitialized buffer it's hard to
3181 * guarantee that whole memory is marked as initialized on
3182 * helper return since specific bounds are unknown what may
3183 * cause uninitialized stack leaking.
3185 if (meta && meta->raw_mode)
3188 if (reg->smax_value >= BPF_MAX_VAR_OFF ||
3189 reg->smax_value <= -BPF_MAX_VAR_OFF) {
3190 verbose(env, "R%d unbounded indirect variable offset stack access\n",
3194 min_off = reg->smin_value + reg->off;
3195 max_off = reg->smax_value + reg->off;
3196 err = __check_stack_boundary(env, regno, min_off, access_size,
3199 verbose(env, "R%d min value is outside of stack bound\n",
3203 err = __check_stack_boundary(env, regno, max_off, access_size,
3206 verbose(env, "R%d max value is outside of stack bound\n",
3212 if (meta && meta->raw_mode) {
3213 meta->access_size = access_size;
3214 meta->regno = regno;
3218 for (i = min_off; i < max_off + access_size; i++) {
3222 spi = slot / BPF_REG_SIZE;
3223 if (state->allocated_stack <= slot)
3225 stype = &state->stack[spi].slot_type[slot % BPF_REG_SIZE];
3226 if (*stype == STACK_MISC)
3228 if (*stype == STACK_ZERO) {
3229 /* helper can write anything into the stack */
3230 *stype = STACK_MISC;
3233 if (state->stack[spi].slot_type[0] == STACK_SPILL &&
3234 state->stack[spi].spilled_ptr.type == SCALAR_VALUE) {
3235 __mark_reg_unknown(env, &state->stack[spi].spilled_ptr);
3236 for (j = 0; j < BPF_REG_SIZE; j++)
3237 state->stack[spi].slot_type[j] = STACK_MISC;
3242 if (tnum_is_const(reg->var_off)) {
3243 verbose(env, "invalid indirect read from stack off %d+%d size %d\n",
3244 min_off, i - min_off, access_size);
3248 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
3249 verbose(env, "invalid indirect read from stack var_off %s+%d size %d\n",
3250 tn_buf, i - min_off, access_size);
3254 /* reading any byte out of 8-byte 'spill_slot' will cause
3255 * the whole slot to be marked as 'read'
3257 mark_reg_read(env, &state->stack[spi].spilled_ptr,
3258 state->stack[spi].spilled_ptr.parent,
3261 return update_stack_depth(env, state, min_off);
3264 static int check_helper_mem_access(struct bpf_verifier_env *env, int regno,
3265 int access_size, bool zero_size_allowed,
3266 struct bpf_call_arg_meta *meta)
3268 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
3270 switch (reg->type) {
3272 case PTR_TO_PACKET_META:
3273 return check_packet_access(env, regno, reg->off, access_size,
3275 case PTR_TO_MAP_VALUE:
3276 if (check_map_access_type(env, regno, reg->off, access_size,
3277 meta && meta->raw_mode ? BPF_WRITE :
3280 return check_map_access(env, regno, reg->off, access_size,
3282 default: /* scalar_value|ptr_to_stack or invalid ptr */
3283 return check_stack_boundary(env, regno, access_size,
3284 zero_size_allowed, meta);
3288 /* Implementation details:
3289 * bpf_map_lookup returns PTR_TO_MAP_VALUE_OR_NULL
3290 * Two bpf_map_lookups (even with the same key) will have different reg->id.
3291 * For traditional PTR_TO_MAP_VALUE the verifier clears reg->id after
3292 * value_or_null->value transition, since the verifier only cares about
3293 * the range of access to valid map value pointer and doesn't care about actual
3294 * address of the map element.
3295 * For maps with 'struct bpf_spin_lock' inside map value the verifier keeps
3296 * reg->id > 0 after value_or_null->value transition. By doing so
3297 * two bpf_map_lookups will be considered two different pointers that
3298 * point to different bpf_spin_locks.
3299 * The verifier allows taking only one bpf_spin_lock at a time to avoid
3301 * Since only one bpf_spin_lock is allowed the checks are simpler than
3302 * reg_is_refcounted() logic. The verifier needs to remember only
3303 * one spin_lock instead of array of acquired_refs.
3304 * cur_state->active_spin_lock remembers which map value element got locked
3305 * and clears it after bpf_spin_unlock.
3307 static int process_spin_lock(struct bpf_verifier_env *env, int regno,
3310 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
3311 struct bpf_verifier_state *cur = env->cur_state;
3312 bool is_const = tnum_is_const(reg->var_off);
3313 struct bpf_map *map = reg->map_ptr;
3314 u64 val = reg->var_off.value;
3316 if (reg->type != PTR_TO_MAP_VALUE) {
3317 verbose(env, "R%d is not a pointer to map_value\n", regno);
3322 "R%d doesn't have constant offset. bpf_spin_lock has to be at the constant offset\n",
3328 "map '%s' has to have BTF in order to use bpf_spin_lock\n",
3332 if (!map_value_has_spin_lock(map)) {
3333 if (map->spin_lock_off == -E2BIG)
3335 "map '%s' has more than one 'struct bpf_spin_lock'\n",
3337 else if (map->spin_lock_off == -ENOENT)
3339 "map '%s' doesn't have 'struct bpf_spin_lock'\n",
3343 "map '%s' is not a struct type or bpf_spin_lock is mangled\n",
3347 if (map->spin_lock_off != val + reg->off) {
3348 verbose(env, "off %lld doesn't point to 'struct bpf_spin_lock'\n",
3353 if (cur->active_spin_lock) {
3355 "Locking two bpf_spin_locks are not allowed\n");
3358 cur->active_spin_lock = reg->id;
3360 if (!cur->active_spin_lock) {
3361 verbose(env, "bpf_spin_unlock without taking a lock\n");
3364 if (cur->active_spin_lock != reg->id) {
3365 verbose(env, "bpf_spin_unlock of different lock\n");
3368 cur->active_spin_lock = 0;
3373 static bool arg_type_is_mem_ptr(enum bpf_arg_type type)
3375 return type == ARG_PTR_TO_MEM ||
3376 type == ARG_PTR_TO_MEM_OR_NULL ||
3377 type == ARG_PTR_TO_UNINIT_MEM;
3380 static bool arg_type_is_mem_size(enum bpf_arg_type type)
3382 return type == ARG_CONST_SIZE ||
3383 type == ARG_CONST_SIZE_OR_ZERO;
3386 static bool arg_type_is_int_ptr(enum bpf_arg_type type)
3388 return type == ARG_PTR_TO_INT ||
3389 type == ARG_PTR_TO_LONG;
3392 static int int_ptr_type_to_size(enum bpf_arg_type type)
3394 if (type == ARG_PTR_TO_INT)
3396 else if (type == ARG_PTR_TO_LONG)
3402 static int check_func_arg(struct bpf_verifier_env *env, u32 regno,
3403 enum bpf_arg_type arg_type,
3404 struct bpf_call_arg_meta *meta)
3406 struct bpf_reg_state *regs = cur_regs(env), *reg = ®s[regno];
3407 enum bpf_reg_type expected_type, type = reg->type;
3410 if (arg_type == ARG_DONTCARE)
3413 err = check_reg_arg(env, regno, SRC_OP);
3417 if (arg_type == ARG_ANYTHING) {
3418 if (is_pointer_value(env, regno)) {
3419 verbose(env, "R%d leaks addr into helper function\n",
3426 if (type_is_pkt_pointer(type) &&
3427 !may_access_direct_pkt_data(env, meta, BPF_READ)) {
3428 verbose(env, "helper access to the packet is not allowed\n");
3432 if (arg_type == ARG_PTR_TO_MAP_KEY ||
3433 arg_type == ARG_PTR_TO_MAP_VALUE ||
3434 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE ||
3435 arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL) {
3436 expected_type = PTR_TO_STACK;
3437 if (register_is_null(reg) &&
3438 arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL)
3439 /* final test in check_stack_boundary() */;
3440 else if (!type_is_pkt_pointer(type) &&
3441 type != PTR_TO_MAP_VALUE &&
3442 type != expected_type)
3444 } else if (arg_type == ARG_CONST_SIZE ||
3445 arg_type == ARG_CONST_SIZE_OR_ZERO) {
3446 expected_type = SCALAR_VALUE;
3447 if (type != expected_type)
3449 } else if (arg_type == ARG_CONST_MAP_PTR) {
3450 expected_type = CONST_PTR_TO_MAP;
3451 if (type != expected_type)
3453 } else if (arg_type == ARG_PTR_TO_CTX ||
3454 arg_type == ARG_PTR_TO_CTX_OR_NULL) {
3455 expected_type = PTR_TO_CTX;
3456 if (!(register_is_null(reg) &&
3457 arg_type == ARG_PTR_TO_CTX_OR_NULL)) {
3458 if (type != expected_type)
3460 err = check_ctx_reg(env, reg, regno);
3464 } else if (arg_type == ARG_PTR_TO_SOCK_COMMON) {
3465 expected_type = PTR_TO_SOCK_COMMON;
3466 /* Any sk pointer can be ARG_PTR_TO_SOCK_COMMON */
3467 if (!type_is_sk_pointer(type))
3469 if (reg->ref_obj_id) {
3470 if (meta->ref_obj_id) {
3471 verbose(env, "verifier internal error: more than one arg with ref_obj_id R%d %u %u\n",
3472 regno, reg->ref_obj_id,
3476 meta->ref_obj_id = reg->ref_obj_id;
3478 } else if (arg_type == ARG_PTR_TO_SOCKET) {
3479 expected_type = PTR_TO_SOCKET;
3480 if (type != expected_type)
3482 } else if (arg_type == ARG_PTR_TO_BTF_ID) {
3483 expected_type = PTR_TO_BTF_ID;
3484 if (type != expected_type)
3486 if (reg->btf_id != meta->btf_id) {
3487 verbose(env, "Helper has type %s got %s in R%d\n",
3488 kernel_type_name(meta->btf_id),
3489 kernel_type_name(reg->btf_id), regno);
3493 if (!tnum_is_const(reg->var_off) || reg->var_off.value || reg->off) {
3494 verbose(env, "R%d is a pointer to in-kernel struct with non-zero offset\n",
3498 } else if (arg_type == ARG_PTR_TO_SPIN_LOCK) {
3499 if (meta->func_id == BPF_FUNC_spin_lock) {
3500 if (process_spin_lock(env, regno, true))
3502 } else if (meta->func_id == BPF_FUNC_spin_unlock) {
3503 if (process_spin_lock(env, regno, false))
3506 verbose(env, "verifier internal error\n");
3509 } else if (arg_type_is_mem_ptr(arg_type)) {
3510 expected_type = PTR_TO_STACK;
3511 /* One exception here. In case function allows for NULL to be
3512 * passed in as argument, it's a SCALAR_VALUE type. Final test
3513 * happens during stack boundary checking.
3515 if (register_is_null(reg) &&
3516 arg_type == ARG_PTR_TO_MEM_OR_NULL)
3517 /* final test in check_stack_boundary() */;
3518 else if (!type_is_pkt_pointer(type) &&
3519 type != PTR_TO_MAP_VALUE &&
3520 type != expected_type)
3522 meta->raw_mode = arg_type == ARG_PTR_TO_UNINIT_MEM;
3523 } else if (arg_type_is_int_ptr(arg_type)) {
3524 expected_type = PTR_TO_STACK;
3525 if (!type_is_pkt_pointer(type) &&
3526 type != PTR_TO_MAP_VALUE &&
3527 type != expected_type)
3530 verbose(env, "unsupported arg_type %d\n", arg_type);
3534 if (arg_type == ARG_CONST_MAP_PTR) {
3535 /* bpf_map_xxx(map_ptr) call: remember that map_ptr */
3536 meta->map_ptr = reg->map_ptr;
3537 } else if (arg_type == ARG_PTR_TO_MAP_KEY) {
3538 /* bpf_map_xxx(..., map_ptr, ..., key) call:
3539 * check that [key, key + map->key_size) are within
3540 * stack limits and initialized
3542 if (!meta->map_ptr) {
3543 /* in function declaration map_ptr must come before
3544 * map_key, so that it's verified and known before
3545 * we have to check map_key here. Otherwise it means
3546 * that kernel subsystem misconfigured verifier
3548 verbose(env, "invalid map_ptr to access map->key\n");
3551 err = check_helper_mem_access(env, regno,
3552 meta->map_ptr->key_size, false,
3554 } else if (arg_type == ARG_PTR_TO_MAP_VALUE ||
3555 (arg_type == ARG_PTR_TO_MAP_VALUE_OR_NULL &&
3556 !register_is_null(reg)) ||
3557 arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE) {
3558 /* bpf_map_xxx(..., map_ptr, ..., value) call:
3559 * check [value, value + map->value_size) validity
3561 if (!meta->map_ptr) {
3562 /* kernel subsystem misconfigured verifier */
3563 verbose(env, "invalid map_ptr to access map->value\n");
3566 meta->raw_mode = (arg_type == ARG_PTR_TO_UNINIT_MAP_VALUE);
3567 err = check_helper_mem_access(env, regno,
3568 meta->map_ptr->value_size, false,
3570 } else if (arg_type_is_mem_size(arg_type)) {
3571 bool zero_size_allowed = (arg_type == ARG_CONST_SIZE_OR_ZERO);
3573 /* This is used to refine r0 return value bounds for helpers
3574 * that enforce this value as an upper bound on return values.
3575 * See do_refine_retval_range() for helpers that can refine
3576 * the return value. C type of helper is u32 so we pull register
3577 * bound from umax_value however, if negative verifier errors
3578 * out. Only upper bounds can be learned because retval is an
3579 * int type and negative retvals are allowed.
3581 meta->msize_max_value = reg->umax_value;
3583 /* The register is SCALAR_VALUE; the access check
3584 * happens using its boundaries.
3586 if (!tnum_is_const(reg->var_off))
3587 /* For unprivileged variable accesses, disable raw
3588 * mode so that the program is required to
3589 * initialize all the memory that the helper could
3590 * just partially fill up.
3594 if (reg->smin_value < 0) {
3595 verbose(env, "R%d min value is negative, either use unsigned or 'var &= const'\n",
3600 if (reg->umin_value == 0) {
3601 err = check_helper_mem_access(env, regno - 1, 0,
3608 if (reg->umax_value >= BPF_MAX_VAR_SIZ) {
3609 verbose(env, "R%d unbounded memory access, use 'var &= const' or 'if (var < const)'\n",
3613 err = check_helper_mem_access(env, regno - 1,
3615 zero_size_allowed, meta);
3617 err = mark_chain_precision(env, regno);
3618 } else if (arg_type_is_int_ptr(arg_type)) {
3619 int size = int_ptr_type_to_size(arg_type);
3621 err = check_helper_mem_access(env, regno, size, false, meta);
3624 err = check_ptr_alignment(env, reg, 0, size, true);
3629 verbose(env, "R%d type=%s expected=%s\n", regno,
3630 reg_type_str[type], reg_type_str[expected_type]);
3634 static int check_map_func_compatibility(struct bpf_verifier_env *env,
3635 struct bpf_map *map, int func_id)
3640 /* We need a two way check, first is from map perspective ... */
3641 switch (map->map_type) {
3642 case BPF_MAP_TYPE_PROG_ARRAY:
3643 if (func_id != BPF_FUNC_tail_call)
3646 case BPF_MAP_TYPE_PERF_EVENT_ARRAY:
3647 if (func_id != BPF_FUNC_perf_event_read &&
3648 func_id != BPF_FUNC_perf_event_output &&
3649 func_id != BPF_FUNC_skb_output &&
3650 func_id != BPF_FUNC_perf_event_read_value &&
3651 func_id != BPF_FUNC_xdp_output)
3654 case BPF_MAP_TYPE_STACK_TRACE:
3655 if (func_id != BPF_FUNC_get_stackid)
3658 case BPF_MAP_TYPE_CGROUP_ARRAY:
3659 if (func_id != BPF_FUNC_skb_under_cgroup &&
3660 func_id != BPF_FUNC_current_task_under_cgroup)
3663 case BPF_MAP_TYPE_CGROUP_STORAGE:
3664 case BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE:
3665 if (func_id != BPF_FUNC_get_local_storage)
3668 case BPF_MAP_TYPE_DEVMAP:
3669 case BPF_MAP_TYPE_DEVMAP_HASH:
3670 if (func_id != BPF_FUNC_redirect_map &&
3671 func_id != BPF_FUNC_map_lookup_elem)
3674 /* Restrict bpf side of cpumap and xskmap, open when use-cases
3677 case BPF_MAP_TYPE_CPUMAP:
3678 if (func_id != BPF_FUNC_redirect_map)
3681 case BPF_MAP_TYPE_XSKMAP:
3682 if (func_id != BPF_FUNC_redirect_map &&
3683 func_id != BPF_FUNC_map_lookup_elem)
3686 case BPF_MAP_TYPE_ARRAY_OF_MAPS:
3687 case BPF_MAP_TYPE_HASH_OF_MAPS:
3688 if (func_id != BPF_FUNC_map_lookup_elem)
3691 case BPF_MAP_TYPE_SOCKMAP:
3692 if (func_id != BPF_FUNC_sk_redirect_map &&
3693 func_id != BPF_FUNC_sock_map_update &&
3694 func_id != BPF_FUNC_map_delete_elem &&
3695 func_id != BPF_FUNC_msg_redirect_map &&
3696 func_id != BPF_FUNC_sk_select_reuseport)
3699 case BPF_MAP_TYPE_SOCKHASH:
3700 if (func_id != BPF_FUNC_sk_redirect_hash &&
3701 func_id != BPF_FUNC_sock_hash_update &&
3702 func_id != BPF_FUNC_map_delete_elem &&
3703 func_id != BPF_FUNC_msg_redirect_hash &&
3704 func_id != BPF_FUNC_sk_select_reuseport)
3707 case BPF_MAP_TYPE_REUSEPORT_SOCKARRAY:
3708 if (func_id != BPF_FUNC_sk_select_reuseport)
3711 case BPF_MAP_TYPE_QUEUE:
3712 case BPF_MAP_TYPE_STACK:
3713 if (func_id != BPF_FUNC_map_peek_elem &&
3714 func_id != BPF_FUNC_map_pop_elem &&
3715 func_id != BPF_FUNC_map_push_elem)
3718 case BPF_MAP_TYPE_SK_STORAGE:
3719 if (func_id != BPF_FUNC_sk_storage_get &&
3720 func_id != BPF_FUNC_sk_storage_delete)
3727 /* ... and second from the function itself. */
3729 case BPF_FUNC_tail_call:
3730 if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
3732 if (env->subprog_cnt > 1) {
3733 verbose(env, "tail_calls are not allowed in programs with bpf-to-bpf calls\n");
3737 case BPF_FUNC_perf_event_read:
3738 case BPF_FUNC_perf_event_output:
3739 case BPF_FUNC_perf_event_read_value:
3740 case BPF_FUNC_skb_output:
3741 case BPF_FUNC_xdp_output:
3742 if (map->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY)
3745 case BPF_FUNC_get_stackid:
3746 if (map->map_type != BPF_MAP_TYPE_STACK_TRACE)
3749 case BPF_FUNC_current_task_under_cgroup:
3750 case BPF_FUNC_skb_under_cgroup:
3751 if (map->map_type != BPF_MAP_TYPE_CGROUP_ARRAY)
3754 case BPF_FUNC_redirect_map:
3755 if (map->map_type != BPF_MAP_TYPE_DEVMAP &&
3756 map->map_type != BPF_MAP_TYPE_DEVMAP_HASH &&
3757 map->map_type != BPF_MAP_TYPE_CPUMAP &&
3758 map->map_type != BPF_MAP_TYPE_XSKMAP)
3761 case BPF_FUNC_sk_redirect_map:
3762 case BPF_FUNC_msg_redirect_map:
3763 case BPF_FUNC_sock_map_update:
3764 if (map->map_type != BPF_MAP_TYPE_SOCKMAP)
3767 case BPF_FUNC_sk_redirect_hash:
3768 case BPF_FUNC_msg_redirect_hash:
3769 case BPF_FUNC_sock_hash_update:
3770 if (map->map_type != BPF_MAP_TYPE_SOCKHASH)
3773 case BPF_FUNC_get_local_storage:
3774 if (map->map_type != BPF_MAP_TYPE_CGROUP_STORAGE &&
3775 map->map_type != BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
3778 case BPF_FUNC_sk_select_reuseport:
3779 if (map->map_type != BPF_MAP_TYPE_REUSEPORT_SOCKARRAY &&
3780 map->map_type != BPF_MAP_TYPE_SOCKMAP &&
3781 map->map_type != BPF_MAP_TYPE_SOCKHASH)
3784 case BPF_FUNC_map_peek_elem:
3785 case BPF_FUNC_map_pop_elem:
3786 case BPF_FUNC_map_push_elem:
3787 if (map->map_type != BPF_MAP_TYPE_QUEUE &&
3788 map->map_type != BPF_MAP_TYPE_STACK)
3791 case BPF_FUNC_sk_storage_get:
3792 case BPF_FUNC_sk_storage_delete:
3793 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE)
3802 verbose(env, "cannot pass map_type %d into func %s#%d\n",
3803 map->map_type, func_id_name(func_id), func_id);
3807 static bool check_raw_mode_ok(const struct bpf_func_proto *fn)
3811 if (fn->arg1_type == ARG_PTR_TO_UNINIT_MEM)
3813 if (fn->arg2_type == ARG_PTR_TO_UNINIT_MEM)
3815 if (fn->arg3_type == ARG_PTR_TO_UNINIT_MEM)
3817 if (fn->arg4_type == ARG_PTR_TO_UNINIT_MEM)
3819 if (fn->arg5_type == ARG_PTR_TO_UNINIT_MEM)
3822 /* We only support one arg being in raw mode at the moment,
3823 * which is sufficient for the helper functions we have
3829 static bool check_args_pair_invalid(enum bpf_arg_type arg_curr,
3830 enum bpf_arg_type arg_next)
3832 return (arg_type_is_mem_ptr(arg_curr) &&
3833 !arg_type_is_mem_size(arg_next)) ||
3834 (!arg_type_is_mem_ptr(arg_curr) &&
3835 arg_type_is_mem_size(arg_next));
3838 static bool check_arg_pair_ok(const struct bpf_func_proto *fn)
3840 /* bpf_xxx(..., buf, len) call will access 'len'
3841 * bytes from memory 'buf'. Both arg types need
3842 * to be paired, so make sure there's no buggy
3843 * helper function specification.
3845 if (arg_type_is_mem_size(fn->arg1_type) ||
3846 arg_type_is_mem_ptr(fn->arg5_type) ||
3847 check_args_pair_invalid(fn->arg1_type, fn->arg2_type) ||
3848 check_args_pair_invalid(fn->arg2_type, fn->arg3_type) ||
3849 check_args_pair_invalid(fn->arg3_type, fn->arg4_type) ||
3850 check_args_pair_invalid(fn->arg4_type, fn->arg5_type))
3856 static bool check_refcount_ok(const struct bpf_func_proto *fn, int func_id)
3860 if (arg_type_may_be_refcounted(fn->arg1_type))
3862 if (arg_type_may_be_refcounted(fn->arg2_type))
3864 if (arg_type_may_be_refcounted(fn->arg3_type))
3866 if (arg_type_may_be_refcounted(fn->arg4_type))
3868 if (arg_type_may_be_refcounted(fn->arg5_type))
3871 /* A reference acquiring function cannot acquire
3872 * another refcounted ptr.
3874 if (is_acquire_function(func_id) && count)
3877 /* We only support one arg being unreferenced at the moment,
3878 * which is sufficient for the helper functions we have right now.
3883 static int check_func_proto(const struct bpf_func_proto *fn, int func_id)
3885 return check_raw_mode_ok(fn) &&
3886 check_arg_pair_ok(fn) &&
3887 check_refcount_ok(fn, func_id) ? 0 : -EINVAL;
3890 /* Packet data might have moved, any old PTR_TO_PACKET[_META,_END]
3891 * are now invalid, so turn them into unknown SCALAR_VALUE.
3893 static void __clear_all_pkt_pointers(struct bpf_verifier_env *env,
3894 struct bpf_func_state *state)
3896 struct bpf_reg_state *regs = state->regs, *reg;
3899 for (i = 0; i < MAX_BPF_REG; i++)
3900 if (reg_is_pkt_pointer_any(®s[i]))
3901 mark_reg_unknown(env, regs, i);
3903 bpf_for_each_spilled_reg(i, state, reg) {
3906 if (reg_is_pkt_pointer_any(reg))
3907 __mark_reg_unknown(env, reg);
3911 static void clear_all_pkt_pointers(struct bpf_verifier_env *env)
3913 struct bpf_verifier_state *vstate = env->cur_state;
3916 for (i = 0; i <= vstate->curframe; i++)
3917 __clear_all_pkt_pointers(env, vstate->frame[i]);
3920 static void release_reg_references(struct bpf_verifier_env *env,
3921 struct bpf_func_state *state,
3924 struct bpf_reg_state *regs = state->regs, *reg;
3927 for (i = 0; i < MAX_BPF_REG; i++)
3928 if (regs[i].ref_obj_id == ref_obj_id)
3929 mark_reg_unknown(env, regs, i);
3931 bpf_for_each_spilled_reg(i, state, reg) {
3934 if (reg->ref_obj_id == ref_obj_id)
3935 __mark_reg_unknown(env, reg);
3939 /* The pointer with the specified id has released its reference to kernel
3940 * resources. Identify all copies of the same pointer and clear the reference.
3942 static int release_reference(struct bpf_verifier_env *env,
3945 struct bpf_verifier_state *vstate = env->cur_state;
3949 err = release_reference_state(cur_func(env), ref_obj_id);
3953 for (i = 0; i <= vstate->curframe; i++)
3954 release_reg_references(env, vstate->frame[i], ref_obj_id);
3959 static void clear_caller_saved_regs(struct bpf_verifier_env *env,
3960 struct bpf_reg_state *regs)
3964 /* after the call registers r0 - r5 were scratched */
3965 for (i = 0; i < CALLER_SAVED_REGS; i++) {
3966 mark_reg_not_init(env, regs, caller_saved[i]);
3967 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
3971 static int check_func_call(struct bpf_verifier_env *env, struct bpf_insn *insn,
3974 struct bpf_verifier_state *state = env->cur_state;
3975 struct bpf_func_info_aux *func_info_aux;
3976 struct bpf_func_state *caller, *callee;
3977 int i, err, subprog, target_insn;
3978 bool is_global = false;
3980 if (state->curframe + 1 >= MAX_CALL_FRAMES) {
3981 verbose(env, "the call stack of %d frames is too deep\n",
3982 state->curframe + 2);
3986 target_insn = *insn_idx + insn->imm;
3987 subprog = find_subprog(env, target_insn + 1);
3989 verbose(env, "verifier bug. No program starts at insn %d\n",
3994 caller = state->frame[state->curframe];
3995 if (state->frame[state->curframe + 1]) {
3996 verbose(env, "verifier bug. Frame %d already allocated\n",
3997 state->curframe + 1);
4001 func_info_aux = env->prog->aux->func_info_aux;
4003 is_global = func_info_aux[subprog].linkage == BTF_FUNC_GLOBAL;
4004 err = btf_check_func_arg_match(env, subprog, caller->regs);
4009 verbose(env, "Caller passes invalid args into func#%d\n",
4013 if (env->log.level & BPF_LOG_LEVEL)
4015 "Func#%d is global and valid. Skipping.\n",
4017 clear_caller_saved_regs(env, caller->regs);
4019 /* All global functions return SCALAR_VALUE */
4020 mark_reg_unknown(env, caller->regs, BPF_REG_0);
4022 /* continue with next insn after call */
4027 callee = kzalloc(sizeof(*callee), GFP_KERNEL);
4030 state->frame[state->curframe + 1] = callee;
4032 /* callee cannot access r0, r6 - r9 for reading and has to write
4033 * into its own stack before reading from it.
4034 * callee can read/write into caller's stack
4036 init_func_state(env, callee,
4037 /* remember the callsite, it will be used by bpf_exit */
4038 *insn_idx /* callsite */,
4039 state->curframe + 1 /* frameno within this callchain */,
4040 subprog /* subprog number within this prog */);
4042 /* Transfer references to the callee */
4043 err = transfer_reference_state(callee, caller);
4047 /* copy r1 - r5 args that callee can access. The copy includes parent
4048 * pointers, which connects us up to the liveness chain
4050 for (i = BPF_REG_1; i <= BPF_REG_5; i++)
4051 callee->regs[i] = caller->regs[i];
4053 clear_caller_saved_regs(env, caller->regs);
4055 /* only increment it after check_reg_arg() finished */
4058 /* and go analyze first insn of the callee */
4059 *insn_idx = target_insn;
4061 if (env->log.level & BPF_LOG_LEVEL) {
4062 verbose(env, "caller:\n");
4063 print_verifier_state(env, caller);
4064 verbose(env, "callee:\n");
4065 print_verifier_state(env, callee);
4070 static int prepare_func_exit(struct bpf_verifier_env *env, int *insn_idx)
4072 struct bpf_verifier_state *state = env->cur_state;
4073 struct bpf_func_state *caller, *callee;
4074 struct bpf_reg_state *r0;
4077 callee = state->frame[state->curframe];
4078 r0 = &callee->regs[BPF_REG_0];
4079 if (r0->type == PTR_TO_STACK) {
4080 /* technically it's ok to return caller's stack pointer
4081 * (or caller's caller's pointer) back to the caller,
4082 * since these pointers are valid. Only current stack
4083 * pointer will be invalid as soon as function exits,
4084 * but let's be conservative
4086 verbose(env, "cannot return stack pointer to the caller\n");
4091 caller = state->frame[state->curframe];
4092 /* return to the caller whatever r0 had in the callee */
4093 caller->regs[BPF_REG_0] = *r0;
4095 /* Transfer references to the caller */
4096 err = transfer_reference_state(caller, callee);
4100 *insn_idx = callee->callsite + 1;
4101 if (env->log.level & BPF_LOG_LEVEL) {
4102 verbose(env, "returning from callee:\n");
4103 print_verifier_state(env, callee);
4104 verbose(env, "to caller at %d:\n", *insn_idx);
4105 print_verifier_state(env, caller);
4107 /* clear everything in the callee */
4108 free_func_state(callee);
4109 state->frame[state->curframe + 1] = NULL;
4113 static void do_refine_retval_range(struct bpf_reg_state *regs, int ret_type,
4115 struct bpf_call_arg_meta *meta)
4117 struct bpf_reg_state *ret_reg = ®s[BPF_REG_0];
4119 if (ret_type != RET_INTEGER ||
4120 (func_id != BPF_FUNC_get_stack &&
4121 func_id != BPF_FUNC_probe_read_str))
4124 ret_reg->smax_value = meta->msize_max_value;
4125 __reg_deduce_bounds(ret_reg);
4126 __reg_bound_offset(ret_reg);
4127 __update_reg_bounds(ret_reg);
4131 record_func_map(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
4132 int func_id, int insn_idx)
4134 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
4135 struct bpf_map *map = meta->map_ptr;
4137 if (func_id != BPF_FUNC_tail_call &&
4138 func_id != BPF_FUNC_map_lookup_elem &&
4139 func_id != BPF_FUNC_map_update_elem &&
4140 func_id != BPF_FUNC_map_delete_elem &&
4141 func_id != BPF_FUNC_map_push_elem &&
4142 func_id != BPF_FUNC_map_pop_elem &&
4143 func_id != BPF_FUNC_map_peek_elem)
4147 verbose(env, "kernel subsystem misconfigured verifier\n");
4151 /* In case of read-only, some additional restrictions
4152 * need to be applied in order to prevent altering the
4153 * state of the map from program side.
4155 if ((map->map_flags & BPF_F_RDONLY_PROG) &&
4156 (func_id == BPF_FUNC_map_delete_elem ||
4157 func_id == BPF_FUNC_map_update_elem ||
4158 func_id == BPF_FUNC_map_push_elem ||
4159 func_id == BPF_FUNC_map_pop_elem)) {
4160 verbose(env, "write into map forbidden\n");
4164 if (!BPF_MAP_PTR(aux->map_ptr_state))
4165 bpf_map_ptr_store(aux, meta->map_ptr,
4166 meta->map_ptr->unpriv_array);
4167 else if (BPF_MAP_PTR(aux->map_ptr_state) != meta->map_ptr)
4168 bpf_map_ptr_store(aux, BPF_MAP_PTR_POISON,
4169 meta->map_ptr->unpriv_array);
4174 record_func_key(struct bpf_verifier_env *env, struct bpf_call_arg_meta *meta,
4175 int func_id, int insn_idx)
4177 struct bpf_insn_aux_data *aux = &env->insn_aux_data[insn_idx];
4178 struct bpf_reg_state *regs = cur_regs(env), *reg;
4179 struct bpf_map *map = meta->map_ptr;
4184 if (func_id != BPF_FUNC_tail_call)
4186 if (!map || map->map_type != BPF_MAP_TYPE_PROG_ARRAY) {
4187 verbose(env, "kernel subsystem misconfigured verifier\n");
4191 range = tnum_range(0, map->max_entries - 1);
4192 reg = ®s[BPF_REG_3];
4194 if (!register_is_const(reg) || !tnum_in(range, reg->var_off)) {
4195 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
4199 err = mark_chain_precision(env, BPF_REG_3);
4203 val = reg->var_off.value;
4204 if (bpf_map_key_unseen(aux))
4205 bpf_map_key_store(aux, val);
4206 else if (!bpf_map_key_poisoned(aux) &&
4207 bpf_map_key_immediate(aux) != val)
4208 bpf_map_key_store(aux, BPF_MAP_KEY_POISON);
4212 static int check_reference_leak(struct bpf_verifier_env *env)
4214 struct bpf_func_state *state = cur_func(env);
4217 for (i = 0; i < state->acquired_refs; i++) {
4218 verbose(env, "Unreleased reference id=%d alloc_insn=%d\n",
4219 state->refs[i].id, state->refs[i].insn_idx);
4221 return state->acquired_refs ? -EINVAL : 0;
4224 static int check_helper_call(struct bpf_verifier_env *env, int func_id, int insn_idx)
4226 const struct bpf_func_proto *fn = NULL;
4227 struct bpf_reg_state *regs;
4228 struct bpf_call_arg_meta meta;
4232 /* find function prototype */
4233 if (func_id < 0 || func_id >= __BPF_FUNC_MAX_ID) {
4234 verbose(env, "invalid func %s#%d\n", func_id_name(func_id),
4239 if (env->ops->get_func_proto)
4240 fn = env->ops->get_func_proto(func_id, env->prog);
4242 verbose(env, "unknown func %s#%d\n", func_id_name(func_id),
4247 /* eBPF programs must be GPL compatible to use GPL-ed functions */
4248 if (!env->prog->gpl_compatible && fn->gpl_only) {
4249 verbose(env, "cannot call GPL-restricted function from non-GPL compatible program\n");
4253 /* With LD_ABS/IND some JITs save/restore skb from r1. */
4254 changes_data = bpf_helper_changes_pkt_data(fn->func);
4255 if (changes_data && fn->arg1_type != ARG_PTR_TO_CTX) {
4256 verbose(env, "kernel subsystem misconfigured func %s#%d: r1 != ctx\n",
4257 func_id_name(func_id), func_id);
4261 memset(&meta, 0, sizeof(meta));
4262 meta.pkt_access = fn->pkt_access;
4264 err = check_func_proto(fn, func_id);
4266 verbose(env, "kernel subsystem misconfigured func %s#%d\n",
4267 func_id_name(func_id), func_id);
4271 meta.func_id = func_id;
4273 for (i = 0; i < 5; i++) {
4274 err = btf_resolve_helper_id(&env->log, fn, i);
4277 err = check_func_arg(env, BPF_REG_1 + i, fn->arg_type[i], &meta);
4282 err = record_func_map(env, &meta, func_id, insn_idx);
4286 err = record_func_key(env, &meta, func_id, insn_idx);
4290 /* Mark slots with STACK_MISC in case of raw mode, stack offset
4291 * is inferred from register state.
4293 for (i = 0; i < meta.access_size; i++) {
4294 err = check_mem_access(env, insn_idx, meta.regno, i, BPF_B,
4295 BPF_WRITE, -1, false);
4300 if (func_id == BPF_FUNC_tail_call) {
4301 err = check_reference_leak(env);
4303 verbose(env, "tail_call would lead to reference leak\n");
4306 } else if (is_release_function(func_id)) {
4307 err = release_reference(env, meta.ref_obj_id);
4309 verbose(env, "func %s#%d reference has not been acquired before\n",
4310 func_id_name(func_id), func_id);
4315 regs = cur_regs(env);
4317 /* check that flags argument in get_local_storage(map, flags) is 0,
4318 * this is required because get_local_storage() can't return an error.
4320 if (func_id == BPF_FUNC_get_local_storage &&
4321 !register_is_null(®s[BPF_REG_2])) {
4322 verbose(env, "get_local_storage() doesn't support non-zero flags\n");
4326 /* reset caller saved regs */
4327 for (i = 0; i < CALLER_SAVED_REGS; i++) {
4328 mark_reg_not_init(env, regs, caller_saved[i]);
4329 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
4332 /* helper call returns 64-bit value. */
4333 regs[BPF_REG_0].subreg_def = DEF_NOT_SUBREG;
4335 /* update return register (already marked as written above) */
4336 if (fn->ret_type == RET_INTEGER) {
4337 /* sets type to SCALAR_VALUE */
4338 mark_reg_unknown(env, regs, BPF_REG_0);
4339 } else if (fn->ret_type == RET_VOID) {
4340 regs[BPF_REG_0].type = NOT_INIT;
4341 } else if (fn->ret_type == RET_PTR_TO_MAP_VALUE_OR_NULL ||
4342 fn->ret_type == RET_PTR_TO_MAP_VALUE) {
4343 /* There is no offset yet applied, variable or fixed */
4344 mark_reg_known_zero(env, regs, BPF_REG_0);
4345 /* remember map_ptr, so that check_map_access()
4346 * can check 'value_size' boundary of memory access
4347 * to map element returned from bpf_map_lookup_elem()
4349 if (meta.map_ptr == NULL) {
4351 "kernel subsystem misconfigured verifier\n");
4354 regs[BPF_REG_0].map_ptr = meta.map_ptr;
4355 if (fn->ret_type == RET_PTR_TO_MAP_VALUE) {
4356 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE;
4357 if (map_value_has_spin_lock(meta.map_ptr))
4358 regs[BPF_REG_0].id = ++env->id_gen;
4360 regs[BPF_REG_0].type = PTR_TO_MAP_VALUE_OR_NULL;
4361 regs[BPF_REG_0].id = ++env->id_gen;
4363 } else if (fn->ret_type == RET_PTR_TO_SOCKET_OR_NULL) {
4364 mark_reg_known_zero(env, regs, BPF_REG_0);
4365 regs[BPF_REG_0].type = PTR_TO_SOCKET_OR_NULL;
4366 regs[BPF_REG_0].id = ++env->id_gen;
4367 } else if (fn->ret_type == RET_PTR_TO_SOCK_COMMON_OR_NULL) {
4368 mark_reg_known_zero(env, regs, BPF_REG_0);
4369 regs[BPF_REG_0].type = PTR_TO_SOCK_COMMON_OR_NULL;
4370 regs[BPF_REG_0].id = ++env->id_gen;
4371 } else if (fn->ret_type == RET_PTR_TO_TCP_SOCK_OR_NULL) {
4372 mark_reg_known_zero(env, regs, BPF_REG_0);
4373 regs[BPF_REG_0].type = PTR_TO_TCP_SOCK_OR_NULL;
4374 regs[BPF_REG_0].id = ++env->id_gen;
4376 verbose(env, "unknown return type %d of func %s#%d\n",
4377 fn->ret_type, func_id_name(func_id), func_id);
4381 if (is_ptr_cast_function(func_id)) {
4382 /* For release_reference() */
4383 regs[BPF_REG_0].ref_obj_id = meta.ref_obj_id;
4384 } else if (is_acquire_function(func_id)) {
4385 int id = acquire_reference_state(env, insn_idx);
4389 /* For mark_ptr_or_null_reg() */
4390 regs[BPF_REG_0].id = id;
4391 /* For release_reference() */
4392 regs[BPF_REG_0].ref_obj_id = id;
4395 do_refine_retval_range(regs, fn->ret_type, func_id, &meta);
4397 err = check_map_func_compatibility(env, meta.map_ptr, func_id);
4401 if (func_id == BPF_FUNC_get_stack && !env->prog->has_callchain_buf) {
4402 const char *err_str;
4404 #ifdef CONFIG_PERF_EVENTS
4405 err = get_callchain_buffers(sysctl_perf_event_max_stack);
4406 err_str = "cannot get callchain buffer for func %s#%d\n";
4409 err_str = "func %s#%d not supported without CONFIG_PERF_EVENTS\n";
4412 verbose(env, err_str, func_id_name(func_id), func_id);
4416 env->prog->has_callchain_buf = true;
4420 clear_all_pkt_pointers(env);
4424 static bool signed_add_overflows(s64 a, s64 b)
4426 /* Do the add in u64, where overflow is well-defined */
4427 s64 res = (s64)((u64)a + (u64)b);
4434 static bool signed_sub_overflows(s64 a, s64 b)
4436 /* Do the sub in u64, where overflow is well-defined */
4437 s64 res = (s64)((u64)a - (u64)b);
4444 static bool check_reg_sane_offset(struct bpf_verifier_env *env,
4445 const struct bpf_reg_state *reg,
4446 enum bpf_reg_type type)
4448 bool known = tnum_is_const(reg->var_off);
4449 s64 val = reg->var_off.value;
4450 s64 smin = reg->smin_value;
4452 if (known && (val >= BPF_MAX_VAR_OFF || val <= -BPF_MAX_VAR_OFF)) {
4453 verbose(env, "math between %s pointer and %lld is not allowed\n",
4454 reg_type_str[type], val);
4458 if (reg->off >= BPF_MAX_VAR_OFF || reg->off <= -BPF_MAX_VAR_OFF) {
4459 verbose(env, "%s pointer offset %d is not allowed\n",
4460 reg_type_str[type], reg->off);
4464 if (smin == S64_MIN) {
4465 verbose(env, "math between %s pointer and register with unbounded min value is not allowed\n",
4466 reg_type_str[type]);
4470 if (smin >= BPF_MAX_VAR_OFF || smin <= -BPF_MAX_VAR_OFF) {
4471 verbose(env, "value %lld makes %s pointer be out of bounds\n",
4472 smin, reg_type_str[type]);
4479 static struct bpf_insn_aux_data *cur_aux(struct bpf_verifier_env *env)
4481 return &env->insn_aux_data[env->insn_idx];
4484 static int retrieve_ptr_limit(const struct bpf_reg_state *ptr_reg,
4485 u32 *ptr_limit, u8 opcode, bool off_is_neg)
4487 bool mask_to_left = (opcode == BPF_ADD && off_is_neg) ||
4488 (opcode == BPF_SUB && !off_is_neg);
4491 switch (ptr_reg->type) {
4493 /* Indirect variable offset stack access is prohibited in
4494 * unprivileged mode so it's not handled here.
4496 off = ptr_reg->off + ptr_reg->var_off.value;
4498 *ptr_limit = MAX_BPF_STACK + off;
4502 case PTR_TO_MAP_VALUE:
4504 *ptr_limit = ptr_reg->umax_value + ptr_reg->off;
4506 off = ptr_reg->smin_value + ptr_reg->off;
4507 *ptr_limit = ptr_reg->map_ptr->value_size - off;
4515 static bool can_skip_alu_sanitation(const struct bpf_verifier_env *env,
4516 const struct bpf_insn *insn)
4518 return env->allow_ptr_leaks || BPF_SRC(insn->code) == BPF_K;
4521 static int update_alu_sanitation_state(struct bpf_insn_aux_data *aux,
4522 u32 alu_state, u32 alu_limit)
4524 /* If we arrived here from different branches with different
4525 * state or limits to sanitize, then this won't work.
4527 if (aux->alu_state &&
4528 (aux->alu_state != alu_state ||
4529 aux->alu_limit != alu_limit))
4532 /* Corresponding fixup done in fixup_bpf_calls(). */
4533 aux->alu_state = alu_state;
4534 aux->alu_limit = alu_limit;
4538 static int sanitize_val_alu(struct bpf_verifier_env *env,
4539 struct bpf_insn *insn)
4541 struct bpf_insn_aux_data *aux = cur_aux(env);
4543 if (can_skip_alu_sanitation(env, insn))
4546 return update_alu_sanitation_state(aux, BPF_ALU_NON_POINTER, 0);
4549 static int sanitize_ptr_alu(struct bpf_verifier_env *env,
4550 struct bpf_insn *insn,
4551 const struct bpf_reg_state *ptr_reg,
4552 struct bpf_reg_state *dst_reg,
4555 struct bpf_verifier_state *vstate = env->cur_state;
4556 struct bpf_insn_aux_data *aux = cur_aux(env);
4557 bool ptr_is_dst_reg = ptr_reg == dst_reg;
4558 u8 opcode = BPF_OP(insn->code);
4559 u32 alu_state, alu_limit;
4560 struct bpf_reg_state tmp;
4563 if (can_skip_alu_sanitation(env, insn))
4566 /* We already marked aux for masking from non-speculative
4567 * paths, thus we got here in the first place. We only care
4568 * to explore bad access from here.
4570 if (vstate->speculative)
4573 alu_state = off_is_neg ? BPF_ALU_NEG_VALUE : 0;
4574 alu_state |= ptr_is_dst_reg ?
4575 BPF_ALU_SANITIZE_SRC : BPF_ALU_SANITIZE_DST;
4577 if (retrieve_ptr_limit(ptr_reg, &alu_limit, opcode, off_is_neg))
4579 if (update_alu_sanitation_state(aux, alu_state, alu_limit))
4582 /* Simulate and find potential out-of-bounds access under
4583 * speculative execution from truncation as a result of
4584 * masking when off was not within expected range. If off
4585 * sits in dst, then we temporarily need to move ptr there
4586 * to simulate dst (== 0) +/-= ptr. Needed, for example,
4587 * for cases where we use K-based arithmetic in one direction
4588 * and truncated reg-based in the other in order to explore
4591 if (!ptr_is_dst_reg) {
4593 *dst_reg = *ptr_reg;
4595 ret = push_stack(env, env->insn_idx + 1, env->insn_idx, true);
4596 if (!ptr_is_dst_reg && ret)
4598 return !ret ? -EFAULT : 0;
4601 /* Handles arithmetic on a pointer and a scalar: computes new min/max and var_off.
4602 * Caller should also handle BPF_MOV case separately.
4603 * If we return -EACCES, caller may want to try again treating pointer as a
4604 * scalar. So we only emit a diagnostic if !env->allow_ptr_leaks.
4606 static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
4607 struct bpf_insn *insn,
4608 const struct bpf_reg_state *ptr_reg,
4609 const struct bpf_reg_state *off_reg)
4611 struct bpf_verifier_state *vstate = env->cur_state;
4612 struct bpf_func_state *state = vstate->frame[vstate->curframe];
4613 struct bpf_reg_state *regs = state->regs, *dst_reg;
4614 bool known = tnum_is_const(off_reg->var_off);
4615 s64 smin_val = off_reg->smin_value, smax_val = off_reg->smax_value,
4616 smin_ptr = ptr_reg->smin_value, smax_ptr = ptr_reg->smax_value;
4617 u64 umin_val = off_reg->umin_value, umax_val = off_reg->umax_value,
4618 umin_ptr = ptr_reg->umin_value, umax_ptr = ptr_reg->umax_value;
4619 u32 dst = insn->dst_reg, src = insn->src_reg;
4620 u8 opcode = BPF_OP(insn->code);
4623 dst_reg = ®s[dst];
4625 if ((known && (smin_val != smax_val || umin_val != umax_val)) ||
4626 smin_val > smax_val || umin_val > umax_val) {
4627 /* Taint dst register if offset had invalid bounds derived from
4628 * e.g. dead branches.
4630 __mark_reg_unknown(env, dst_reg);
4634 if (BPF_CLASS(insn->code) != BPF_ALU64) {
4635 /* 32-bit ALU ops on pointers produce (meaningless) scalars */
4637 "R%d 32-bit pointer arithmetic prohibited\n",
4642 switch (ptr_reg->type) {
4643 case PTR_TO_MAP_VALUE_OR_NULL:
4644 verbose(env, "R%d pointer arithmetic on %s prohibited, null-check it first\n",
4645 dst, reg_type_str[ptr_reg->type]);
4647 case CONST_PTR_TO_MAP:
4648 case PTR_TO_PACKET_END:
4650 case PTR_TO_SOCKET_OR_NULL:
4651 case PTR_TO_SOCK_COMMON:
4652 case PTR_TO_SOCK_COMMON_OR_NULL:
4653 case PTR_TO_TCP_SOCK:
4654 case PTR_TO_TCP_SOCK_OR_NULL:
4655 case PTR_TO_XDP_SOCK:
4656 verbose(env, "R%d pointer arithmetic on %s prohibited\n",
4657 dst, reg_type_str[ptr_reg->type]);
4659 case PTR_TO_MAP_VALUE:
4660 if (!env->allow_ptr_leaks && !known && (smin_val < 0) != (smax_val < 0)) {
4661 verbose(env, "R%d has unknown scalar with mixed signed bounds, pointer arithmetic with it prohibited for !root\n",
4662 off_reg == dst_reg ? dst : src);
4670 /* In case of 'scalar += pointer', dst_reg inherits pointer type and id.
4671 * The id may be overwritten later if we create a new variable offset.
4673 dst_reg->type = ptr_reg->type;
4674 dst_reg->id = ptr_reg->id;
4676 if (!check_reg_sane_offset(env, off_reg, ptr_reg->type) ||
4677 !check_reg_sane_offset(env, ptr_reg, ptr_reg->type))
4682 ret = sanitize_ptr_alu(env, insn, ptr_reg, dst_reg, smin_val < 0);
4684 verbose(env, "R%d tried to add from different maps or paths\n", dst);
4687 /* We can take a fixed offset as long as it doesn't overflow
4688 * the s32 'off' field
4690 if (known && (ptr_reg->off + smin_val ==
4691 (s64)(s32)(ptr_reg->off + smin_val))) {
4692 /* pointer += K. Accumulate it into fixed offset */
4693 dst_reg->smin_value = smin_ptr;
4694 dst_reg->smax_value = smax_ptr;
4695 dst_reg->umin_value = umin_ptr;
4696 dst_reg->umax_value = umax_ptr;
4697 dst_reg->var_off = ptr_reg->var_off;
4698 dst_reg->off = ptr_reg->off + smin_val;
4699 dst_reg->raw = ptr_reg->raw;
4702 /* A new variable offset is created. Note that off_reg->off
4703 * == 0, since it's a scalar.
4704 * dst_reg gets the pointer type and since some positive
4705 * integer value was added to the pointer, give it a new 'id'
4706 * if it's a PTR_TO_PACKET.
4707 * this creates a new 'base' pointer, off_reg (variable) gets
4708 * added into the variable offset, and we copy the fixed offset
4711 if (signed_add_overflows(smin_ptr, smin_val) ||
4712 signed_add_overflows(smax_ptr, smax_val)) {
4713 dst_reg->smin_value = S64_MIN;
4714 dst_reg->smax_value = S64_MAX;
4716 dst_reg->smin_value = smin_ptr + smin_val;
4717 dst_reg->smax_value = smax_ptr + smax_val;
4719 if (umin_ptr + umin_val < umin_ptr ||
4720 umax_ptr + umax_val < umax_ptr) {
4721 dst_reg->umin_value = 0;
4722 dst_reg->umax_value = U64_MAX;
4724 dst_reg->umin_value = umin_ptr + umin_val;
4725 dst_reg->umax_value = umax_ptr + umax_val;
4727 dst_reg->var_off = tnum_add(ptr_reg->var_off, off_reg->var_off);
4728 dst_reg->off = ptr_reg->off;
4729 dst_reg->raw = ptr_reg->raw;
4730 if (reg_is_pkt_pointer(ptr_reg)) {
4731 dst_reg->id = ++env->id_gen;
4732 /* something was added to pkt_ptr, set range to zero */
4737 ret = sanitize_ptr_alu(env, insn, ptr_reg, dst_reg, smin_val < 0);
4739 verbose(env, "R%d tried to sub from different maps or paths\n", dst);
4742 if (dst_reg == off_reg) {
4743 /* scalar -= pointer. Creates an unknown scalar */
4744 verbose(env, "R%d tried to subtract pointer from scalar\n",
4748 /* We don't allow subtraction from FP, because (according to
4749 * test_verifier.c test "invalid fp arithmetic", JITs might not
4750 * be able to deal with it.
4752 if (ptr_reg->type == PTR_TO_STACK) {
4753 verbose(env, "R%d subtraction from stack pointer prohibited\n",
4757 if (known && (ptr_reg->off - smin_val ==
4758 (s64)(s32)(ptr_reg->off - smin_val))) {
4759 /* pointer -= K. Subtract it from fixed offset */
4760 dst_reg->smin_value = smin_ptr;
4761 dst_reg->smax_value = smax_ptr;
4762 dst_reg->umin_value = umin_ptr;
4763 dst_reg->umax_value = umax_ptr;
4764 dst_reg->var_off = ptr_reg->var_off;
4765 dst_reg->id = ptr_reg->id;
4766 dst_reg->off = ptr_reg->off - smin_val;
4767 dst_reg->raw = ptr_reg->raw;
4770 /* A new variable offset is created. If the subtrahend is known
4771 * nonnegative, then any reg->range we had before is still good.
4773 if (signed_sub_overflows(smin_ptr, smax_val) ||
4774 signed_sub_overflows(smax_ptr, smin_val)) {
4775 /* Overflow possible, we know nothing */
4776 dst_reg->smin_value = S64_MIN;
4777 dst_reg->smax_value = S64_MAX;
4779 dst_reg->smin_value = smin_ptr - smax_val;
4780 dst_reg->smax_value = smax_ptr - smin_val;
4782 if (umin_ptr < umax_val) {
4783 /* Overflow possible, we know nothing */
4784 dst_reg->umin_value = 0;
4785 dst_reg->umax_value = U64_MAX;
4787 /* Cannot overflow (as long as bounds are consistent) */
4788 dst_reg->umin_value = umin_ptr - umax_val;
4789 dst_reg->umax_value = umax_ptr - umin_val;
4791 dst_reg->var_off = tnum_sub(ptr_reg->var_off, off_reg->var_off);
4792 dst_reg->off = ptr_reg->off;
4793 dst_reg->raw = ptr_reg->raw;
4794 if (reg_is_pkt_pointer(ptr_reg)) {
4795 dst_reg->id = ++env->id_gen;
4796 /* something was added to pkt_ptr, set range to zero */
4804 /* bitwise ops on pointers are troublesome, prohibit. */
4805 verbose(env, "R%d bitwise operator %s on pointer prohibited\n",
4806 dst, bpf_alu_string[opcode >> 4]);
4809 /* other operators (e.g. MUL,LSH) produce non-pointer results */
4810 verbose(env, "R%d pointer arithmetic with %s operator prohibited\n",
4811 dst, bpf_alu_string[opcode >> 4]);
4815 if (!check_reg_sane_offset(env, dst_reg, ptr_reg->type))
4818 __update_reg_bounds(dst_reg);
4819 __reg_deduce_bounds(dst_reg);
4820 __reg_bound_offset(dst_reg);
4822 /* For unprivileged we require that resulting offset must be in bounds
4823 * in order to be able to sanitize access later on.
4825 if (!env->allow_ptr_leaks) {
4826 if (dst_reg->type == PTR_TO_MAP_VALUE &&
4827 check_map_access(env, dst, dst_reg->off, 1, false)) {
4828 verbose(env, "R%d pointer arithmetic of map value goes out of range, "
4829 "prohibited for !root\n", dst);
4831 } else if (dst_reg->type == PTR_TO_STACK &&
4832 check_stack_access(env, dst_reg, dst_reg->off +
4833 dst_reg->var_off.value, 1)) {
4834 verbose(env, "R%d stack pointer arithmetic goes out of range, "
4835 "prohibited for !root\n", dst);
4843 static void scalar_min_max_add(struct bpf_reg_state *dst_reg,
4844 struct bpf_reg_state *src_reg)
4846 s64 smin_val = src_reg->smin_value;
4847 s64 smax_val = src_reg->smax_value;
4848 u64 umin_val = src_reg->umin_value;
4849 u64 umax_val = src_reg->umax_value;
4851 if (signed_add_overflows(dst_reg->smin_value, smin_val) ||
4852 signed_add_overflows(dst_reg->smax_value, smax_val)) {
4853 dst_reg->smin_value = S64_MIN;
4854 dst_reg->smax_value = S64_MAX;
4856 dst_reg->smin_value += smin_val;
4857 dst_reg->smax_value += smax_val;
4859 if (dst_reg->umin_value + umin_val < umin_val ||
4860 dst_reg->umax_value + umax_val < umax_val) {
4861 dst_reg->umin_value = 0;
4862 dst_reg->umax_value = U64_MAX;
4864 dst_reg->umin_value += umin_val;
4865 dst_reg->umax_value += umax_val;
4867 dst_reg->var_off = tnum_add(dst_reg->var_off, src_reg->var_off);
4870 static void scalar_min_max_sub(struct bpf_reg_state *dst_reg,
4871 struct bpf_reg_state *src_reg)
4873 s64 smin_val = src_reg->smin_value;
4874 s64 smax_val = src_reg->smax_value;
4875 u64 umin_val = src_reg->umin_value;
4876 u64 umax_val = src_reg->umax_value;
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 static void scalar_min_max_mul(struct bpf_reg_state *dst_reg,
4900 struct bpf_reg_state *src_reg)
4902 s64 smin_val = src_reg->smin_value;
4903 u64 umin_val = src_reg->umin_value;
4904 u64 umax_val = src_reg->umax_value;
4906 dst_reg->var_off = tnum_mul(dst_reg->var_off, src_reg->var_off);
4907 if (smin_val < 0 || dst_reg->smin_value < 0) {
4908 /* Ain't nobody got time to multiply that sign */
4909 __mark_reg_unbounded(dst_reg);
4910 __update_reg_bounds(dst_reg);
4913 /* Both values are positive, so we can work with unsigned and
4914 * copy the result to signed (unless it exceeds S64_MAX).
4916 if (umax_val > U32_MAX || dst_reg->umax_value > U32_MAX) {
4917 /* Potential overflow, we know nothing */
4918 __mark_reg_unbounded(dst_reg);
4919 /* (except what we can learn from the var_off) */
4920 __update_reg_bounds(dst_reg);
4923 dst_reg->umin_value *= umin_val;
4924 dst_reg->umax_value *= umax_val;
4925 if (dst_reg->umax_value > S64_MAX) {
4926 /* Overflow possible, we know nothing */
4927 dst_reg->smin_value = S64_MIN;
4928 dst_reg->smax_value = S64_MAX;
4930 dst_reg->smin_value = dst_reg->umin_value;
4931 dst_reg->smax_value = dst_reg->umax_value;
4935 static void scalar_min_max_and(struct bpf_reg_state *dst_reg,
4936 struct bpf_reg_state *src_reg)
4938 s64 smin_val = src_reg->smin_value;
4939 u64 umax_val = src_reg->umax_value;
4941 /* We get our minimum from the var_off, since that's inherently
4942 * bitwise. Our maximum is the minimum of the operands' maxima.
4944 dst_reg->var_off = tnum_and(dst_reg->var_off, src_reg->var_off);
4945 dst_reg->umin_value = dst_reg->var_off.value;
4946 dst_reg->umax_value = min(dst_reg->umax_value, umax_val);
4947 if (dst_reg->smin_value < 0 || smin_val < 0) {
4948 /* Lose signed bounds when ANDing negative numbers,
4949 * ain't nobody got time for that.
4951 dst_reg->smin_value = S64_MIN;
4952 dst_reg->smax_value = S64_MAX;
4954 /* ANDing two positives gives a positive, so safe to
4955 * cast result into s64.
4957 dst_reg->smin_value = dst_reg->umin_value;
4958 dst_reg->smax_value = dst_reg->umax_value;
4960 /* We may learn something more from the var_off */
4961 __update_reg_bounds(dst_reg);
4964 static void scalar_min_max_or(struct bpf_reg_state *dst_reg,
4965 struct bpf_reg_state *src_reg)
4967 s64 smin_val = src_reg->smin_value;
4968 u64 umin_val = src_reg->umin_value;
4970 /* We get our maximum from the var_off, and our minimum is the
4971 * maximum of the operands' minima
4973 dst_reg->var_off = tnum_or(dst_reg->var_off, src_reg->var_off);
4974 dst_reg->umin_value = max(dst_reg->umin_value, umin_val);
4975 dst_reg->umax_value = dst_reg->var_off.value | dst_reg->var_off.mask;
4976 if (dst_reg->smin_value < 0 || smin_val < 0) {
4977 /* Lose signed bounds when ORing negative numbers,
4978 * ain't nobody got time for that.
4980 dst_reg->smin_value = S64_MIN;
4981 dst_reg->smax_value = S64_MAX;
4983 /* ORing two positives gives a positive, so safe to
4984 * cast result into s64.
4986 dst_reg->smin_value = dst_reg->umin_value;
4987 dst_reg->smax_value = dst_reg->umax_value;
4989 /* We may learn something more from the var_off */
4990 __update_reg_bounds(dst_reg);
4993 static void scalar_min_max_lsh(struct bpf_reg_state *dst_reg,
4994 struct bpf_reg_state *src_reg)
4996 u64 umax_val = src_reg->umax_value;
4997 u64 umin_val = src_reg->umin_value;
4999 /* We lose all sign bit information (except what we can pick
5002 dst_reg->smin_value = S64_MIN;
5003 dst_reg->smax_value = S64_MAX;
5004 /* If we might shift our top bit out, then we know nothing */
5005 if (dst_reg->umax_value > 1ULL << (63 - umax_val)) {
5006 dst_reg->umin_value = 0;
5007 dst_reg->umax_value = U64_MAX;
5009 dst_reg->umin_value <<= umin_val;
5010 dst_reg->umax_value <<= umax_val;
5012 dst_reg->var_off = tnum_lshift(dst_reg->var_off, umin_val);
5013 /* We may learn something more from the var_off */
5014 __update_reg_bounds(dst_reg);
5017 static void scalar_min_max_rsh(struct bpf_reg_state *dst_reg,
5018 struct bpf_reg_state *src_reg)
5020 u64 umax_val = src_reg->umax_value;
5021 u64 umin_val = src_reg->umin_value;
5023 /* BPF_RSH is an unsigned shift. If the value in dst_reg might
5024 * be negative, then either:
5025 * 1) src_reg might be zero, so the sign bit of the result is
5026 * unknown, so we lose our signed bounds
5027 * 2) it's known negative, thus the unsigned bounds capture the
5029 * 3) the signed bounds cross zero, so they tell us nothing
5031 * If the value in dst_reg is known nonnegative, then again the
5032 * unsigned bounts capture the signed bounds.
5033 * Thus, in all cases it suffices to blow away our signed bounds
5034 * and rely on inferring new ones from the unsigned bounds and
5035 * var_off of the result.
5037 dst_reg->smin_value = S64_MIN;
5038 dst_reg->smax_value = S64_MAX;
5039 dst_reg->var_off = tnum_rshift(dst_reg->var_off, umin_val);
5040 dst_reg->umin_value >>= umax_val;
5041 dst_reg->umax_value >>= umin_val;
5042 /* We may learn something more from the var_off */
5043 __update_reg_bounds(dst_reg);
5046 static void scalar_min_max_arsh(struct bpf_reg_state *dst_reg,
5047 struct bpf_reg_state *src_reg,
5050 u64 umin_val = src_reg->umin_value;
5052 /* Upon reaching here, src_known is true and
5053 * umax_val is equal to umin_val.
5055 if (insn_bitness == 32) {
5056 dst_reg->smin_value = (u32)(((s32)dst_reg->smin_value) >> umin_val);
5057 dst_reg->smax_value = (u32)(((s32)dst_reg->smax_value) >> umin_val);
5059 dst_reg->smin_value >>= umin_val;
5060 dst_reg->smax_value >>= umin_val;
5063 dst_reg->var_off = tnum_arshift(dst_reg->var_off, umin_val,
5066 /* blow away the dst_reg umin_value/umax_value and rely on
5067 * dst_reg var_off to refine the result.
5069 dst_reg->umin_value = 0;
5070 dst_reg->umax_value = U64_MAX;
5071 __update_reg_bounds(dst_reg);
5074 /* WARNING: This function does calculations on 64-bit values, but the actual
5075 * execution may occur on 32-bit values. Therefore, things like bitshifts
5076 * need extra checks in the 32-bit case.
5078 static int adjust_scalar_min_max_vals(struct bpf_verifier_env *env,
5079 struct bpf_insn *insn,
5080 struct bpf_reg_state *dst_reg,
5081 struct bpf_reg_state src_reg)
5083 struct bpf_reg_state *regs = cur_regs(env);
5084 u8 opcode = BPF_OP(insn->code);
5085 bool src_known, dst_known;
5086 s64 smin_val, smax_val;
5087 u64 umin_val, umax_val;
5088 u64 insn_bitness = (BPF_CLASS(insn->code) == BPF_ALU64) ? 64 : 32;
5089 u32 dst = insn->dst_reg;
5092 if (insn_bitness == 32) {
5093 /* Relevant for 32-bit RSH: Information can propagate towards
5094 * LSB, so it isn't sufficient to only truncate the output to
5097 coerce_reg_to_size(dst_reg, 4);
5098 coerce_reg_to_size(&src_reg, 4);
5101 smin_val = src_reg.smin_value;
5102 smax_val = src_reg.smax_value;
5103 umin_val = src_reg.umin_value;
5104 umax_val = src_reg.umax_value;
5105 src_known = tnum_is_const(src_reg.var_off);
5106 dst_known = tnum_is_const(dst_reg->var_off);
5108 if ((src_known && (smin_val != smax_val || umin_val != umax_val)) ||
5109 smin_val > smax_val || umin_val > umax_val) {
5110 /* Taint dst register if offset had invalid bounds derived from
5111 * e.g. dead branches.
5113 __mark_reg_unknown(env, dst_reg);
5118 opcode != BPF_ADD && opcode != BPF_SUB && opcode != BPF_AND) {
5119 __mark_reg_unknown(env, dst_reg);
5125 ret = sanitize_val_alu(env, insn);
5127 verbose(env, "R%d tried to add from different pointers or scalars\n", dst);
5130 scalar_min_max_add(dst_reg, &src_reg);
5133 ret = sanitize_val_alu(env, insn);
5135 verbose(env, "R%d tried to sub from different pointers or scalars\n", dst);
5138 scalar_min_max_sub(dst_reg, &src_reg);
5141 scalar_min_max_mul(dst_reg, &src_reg);
5144 if (src_known && dst_known) {
5145 __mark_reg_known(dst_reg, dst_reg->var_off.value &
5146 src_reg.var_off.value);
5149 scalar_min_max_and(dst_reg, &src_reg);
5152 if (src_known && dst_known) {
5153 __mark_reg_known(dst_reg, dst_reg->var_off.value |
5154 src_reg.var_off.value);
5157 scalar_min_max_or(dst_reg, &src_reg);
5160 if (umax_val >= insn_bitness) {
5161 /* Shifts greater than 31 or 63 are undefined.
5162 * This includes shifts by a negative number.
5164 mark_reg_unknown(env, regs, insn->dst_reg);
5167 scalar_min_max_lsh(dst_reg, &src_reg);
5170 if (umax_val >= insn_bitness) {
5171 /* Shifts greater than 31 or 63 are undefined.
5172 * This includes shifts by a negative number.
5174 mark_reg_unknown(env, regs, insn->dst_reg);
5177 scalar_min_max_rsh(dst_reg, &src_reg);
5180 if (umax_val >= insn_bitness) {
5181 /* Shifts greater than 31 or 63 are undefined.
5182 * This includes shifts by a negative number.
5184 mark_reg_unknown(env, regs, insn->dst_reg);
5187 scalar_min_max_arsh(dst_reg, &src_reg, insn_bitness);
5190 mark_reg_unknown(env, regs, insn->dst_reg);
5194 if (BPF_CLASS(insn->code) != BPF_ALU64) {
5195 /* 32-bit ALU ops are (32,32)->32 */
5196 coerce_reg_to_size(dst_reg, 4);
5199 __update_reg_bounds(dst_reg);
5200 __reg_deduce_bounds(dst_reg);
5201 __reg_bound_offset(dst_reg);
5205 /* Handles ALU ops other than BPF_END, BPF_NEG and BPF_MOV: computes new min/max
5208 static int adjust_reg_min_max_vals(struct bpf_verifier_env *env,
5209 struct bpf_insn *insn)
5211 struct bpf_verifier_state *vstate = env->cur_state;
5212 struct bpf_func_state *state = vstate->frame[vstate->curframe];
5213 struct bpf_reg_state *regs = state->regs, *dst_reg, *src_reg;
5214 struct bpf_reg_state *ptr_reg = NULL, off_reg = {0};
5215 u8 opcode = BPF_OP(insn->code);
5218 dst_reg = ®s[insn->dst_reg];
5220 if (dst_reg->type != SCALAR_VALUE)
5222 if (BPF_SRC(insn->code) == BPF_X) {
5223 src_reg = ®s[insn->src_reg];
5224 if (src_reg->type != SCALAR_VALUE) {
5225 if (dst_reg->type != SCALAR_VALUE) {
5226 /* Combining two pointers by any ALU op yields
5227 * an arbitrary scalar. Disallow all math except
5228 * pointer subtraction
5230 if (opcode == BPF_SUB && env->allow_ptr_leaks) {
5231 mark_reg_unknown(env, regs, insn->dst_reg);
5234 verbose(env, "R%d pointer %s pointer prohibited\n",
5236 bpf_alu_string[opcode >> 4]);
5239 /* scalar += pointer
5240 * This is legal, but we have to reverse our
5241 * src/dest handling in computing the range
5243 err = mark_chain_precision(env, insn->dst_reg);
5246 return adjust_ptr_min_max_vals(env, insn,
5249 } else if (ptr_reg) {
5250 /* pointer += scalar */
5251 err = mark_chain_precision(env, insn->src_reg);
5254 return adjust_ptr_min_max_vals(env, insn,
5258 /* Pretend the src is a reg with a known value, since we only
5259 * need to be able to read from this state.
5261 off_reg.type = SCALAR_VALUE;
5262 __mark_reg_known(&off_reg, insn->imm);
5264 if (ptr_reg) /* pointer += K */
5265 return adjust_ptr_min_max_vals(env, insn,
5269 /* Got here implies adding two SCALAR_VALUEs */
5270 if (WARN_ON_ONCE(ptr_reg)) {
5271 print_verifier_state(env, state);
5272 verbose(env, "verifier internal error: unexpected ptr_reg\n");
5275 if (WARN_ON(!src_reg)) {
5276 print_verifier_state(env, state);
5277 verbose(env, "verifier internal error: no src_reg\n");
5280 return adjust_scalar_min_max_vals(env, insn, dst_reg, *src_reg);
5283 /* check validity of 32-bit and 64-bit arithmetic operations */
5284 static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
5286 struct bpf_reg_state *regs = cur_regs(env);
5287 u8 opcode = BPF_OP(insn->code);
5290 if (opcode == BPF_END || opcode == BPF_NEG) {
5291 if (opcode == BPF_NEG) {
5292 if (BPF_SRC(insn->code) != 0 ||
5293 insn->src_reg != BPF_REG_0 ||
5294 insn->off != 0 || insn->imm != 0) {
5295 verbose(env, "BPF_NEG uses reserved fields\n");
5299 if (insn->src_reg != BPF_REG_0 || insn->off != 0 ||
5300 (insn->imm != 16 && insn->imm != 32 && insn->imm != 64) ||
5301 BPF_CLASS(insn->code) == BPF_ALU64) {
5302 verbose(env, "BPF_END uses reserved fields\n");
5307 /* check src operand */
5308 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
5312 if (is_pointer_value(env, insn->dst_reg)) {
5313 verbose(env, "R%d pointer arithmetic prohibited\n",
5318 /* check dest operand */
5319 err = check_reg_arg(env, insn->dst_reg, DST_OP);
5323 } else if (opcode == BPF_MOV) {
5325 if (BPF_SRC(insn->code) == BPF_X) {
5326 if (insn->imm != 0 || insn->off != 0) {
5327 verbose(env, "BPF_MOV uses reserved fields\n");
5331 /* check src operand */
5332 err = check_reg_arg(env, insn->src_reg, SRC_OP);
5336 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
5337 verbose(env, "BPF_MOV uses reserved fields\n");
5342 /* check dest operand, mark as required later */
5343 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
5347 if (BPF_SRC(insn->code) == BPF_X) {
5348 struct bpf_reg_state *src_reg = regs + insn->src_reg;
5349 struct bpf_reg_state *dst_reg = regs + insn->dst_reg;
5351 if (BPF_CLASS(insn->code) == BPF_ALU64) {
5353 * copy register state to dest reg
5355 *dst_reg = *src_reg;
5356 dst_reg->live |= REG_LIVE_WRITTEN;
5357 dst_reg->subreg_def = DEF_NOT_SUBREG;
5360 if (is_pointer_value(env, insn->src_reg)) {
5362 "R%d partial copy of pointer\n",
5365 } else if (src_reg->type == SCALAR_VALUE) {
5366 *dst_reg = *src_reg;
5367 dst_reg->live |= REG_LIVE_WRITTEN;
5368 dst_reg->subreg_def = env->insn_idx + 1;
5370 mark_reg_unknown(env, regs,
5373 coerce_reg_to_size(dst_reg, 4);
5377 * remember the value we stored into this reg
5379 /* clear any state __mark_reg_known doesn't set */
5380 mark_reg_unknown(env, regs, insn->dst_reg);
5381 regs[insn->dst_reg].type = SCALAR_VALUE;
5382 if (BPF_CLASS(insn->code) == BPF_ALU64) {
5383 __mark_reg_known(regs + insn->dst_reg,
5386 __mark_reg_known(regs + insn->dst_reg,
5391 } else if (opcode > BPF_END) {
5392 verbose(env, "invalid BPF_ALU opcode %x\n", opcode);
5395 } else { /* all other ALU ops: and, sub, xor, add, ... */
5397 if (BPF_SRC(insn->code) == BPF_X) {
5398 if (insn->imm != 0 || insn->off != 0) {
5399 verbose(env, "BPF_ALU uses reserved fields\n");
5402 /* check src1 operand */
5403 err = check_reg_arg(env, insn->src_reg, SRC_OP);
5407 if (insn->src_reg != BPF_REG_0 || insn->off != 0) {
5408 verbose(env, "BPF_ALU uses reserved fields\n");
5413 /* check src2 operand */
5414 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
5418 if ((opcode == BPF_MOD || opcode == BPF_DIV) &&
5419 BPF_SRC(insn->code) == BPF_K && insn->imm == 0) {
5420 verbose(env, "div by zero\n");
5424 if ((opcode == BPF_LSH || opcode == BPF_RSH ||
5425 opcode == BPF_ARSH) && BPF_SRC(insn->code) == BPF_K) {
5426 int size = BPF_CLASS(insn->code) == BPF_ALU64 ? 64 : 32;
5428 if (insn->imm < 0 || insn->imm >= size) {
5429 verbose(env, "invalid shift %d\n", insn->imm);
5434 /* check dest operand */
5435 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
5439 return adjust_reg_min_max_vals(env, insn);
5445 static void __find_good_pkt_pointers(struct bpf_func_state *state,
5446 struct bpf_reg_state *dst_reg,
5447 enum bpf_reg_type type, u16 new_range)
5449 struct bpf_reg_state *reg;
5452 for (i = 0; i < MAX_BPF_REG; i++) {
5453 reg = &state->regs[i];
5454 if (reg->type == type && reg->id == dst_reg->id)
5455 /* keep the maximum range already checked */
5456 reg->range = max(reg->range, new_range);
5459 bpf_for_each_spilled_reg(i, state, reg) {
5462 if (reg->type == type && reg->id == dst_reg->id)
5463 reg->range = max(reg->range, new_range);
5467 static void find_good_pkt_pointers(struct bpf_verifier_state *vstate,
5468 struct bpf_reg_state *dst_reg,
5469 enum bpf_reg_type type,
5470 bool range_right_open)
5475 if (dst_reg->off < 0 ||
5476 (dst_reg->off == 0 && range_right_open))
5477 /* This doesn't give us any range */
5480 if (dst_reg->umax_value > MAX_PACKET_OFF ||
5481 dst_reg->umax_value + dst_reg->off > MAX_PACKET_OFF)
5482 /* Risk of overflow. For instance, ptr + (1<<63) may be less
5483 * than pkt_end, but that's because it's also less than pkt.
5487 new_range = dst_reg->off;
5488 if (range_right_open)
5491 /* Examples for register markings:
5493 * pkt_data in dst register:
5497 * if (r2 > pkt_end) goto <handle exception>
5502 * if (r2 < pkt_end) goto <access okay>
5503 * <handle exception>
5506 * r2 == dst_reg, pkt_end == src_reg
5507 * r2=pkt(id=n,off=8,r=0)
5508 * r3=pkt(id=n,off=0,r=0)
5510 * pkt_data in src register:
5514 * if (pkt_end >= r2) goto <access okay>
5515 * <handle exception>
5519 * if (pkt_end <= r2) goto <handle exception>
5523 * pkt_end == dst_reg, r2 == src_reg
5524 * r2=pkt(id=n,off=8,r=0)
5525 * r3=pkt(id=n,off=0,r=0)
5527 * Find register r3 and mark its range as r3=pkt(id=n,off=0,r=8)
5528 * or r3=pkt(id=n,off=0,r=8-1), so that range of bytes [r3, r3 + 8)
5529 * and [r3, r3 + 8-1) respectively is safe to access depending on
5533 /* If our ids match, then we must have the same max_value. And we
5534 * don't care about the other reg's fixed offset, since if it's too big
5535 * the range won't allow anything.
5536 * dst_reg->off is known < MAX_PACKET_OFF, therefore it fits in a u16.
5538 for (i = 0; i <= vstate->curframe; i++)
5539 __find_good_pkt_pointers(vstate->frame[i], dst_reg, type,
5543 /* compute branch direction of the expression "if (reg opcode val) goto target;"
5545 * 1 - branch will be taken and "goto target" will be executed
5546 * 0 - branch will not be taken and fall-through to next insn
5547 * -1 - unknown. Example: "if (reg < 5)" is unknown when register value range [0,10]
5549 static int is_branch_taken(struct bpf_reg_state *reg, u64 val, u8 opcode,
5552 struct bpf_reg_state reg_lo;
5555 if (__is_pointer_value(false, reg))
5561 /* For JMP32, only low 32 bits are compared, coerce_reg_to_size
5562 * could truncate high bits and update umin/umax according to
5563 * information of low bits.
5565 coerce_reg_to_size(reg, 4);
5566 /* smin/smax need special handling. For example, after coerce,
5567 * if smin_value is 0x00000000ffffffffLL, the value is -1 when
5568 * used as operand to JMP32. It is a negative number from s32's
5569 * point of view, while it is a positive number when seen as
5570 * s64. The smin/smax are kept as s64, therefore, when used with
5571 * JMP32, they need to be transformed into s32, then sign
5572 * extended back to s64.
5574 * Also, smin/smax were copied from umin/umax. If umin/umax has
5575 * different sign bit, then min/max relationship doesn't
5576 * maintain after casting into s32, for this case, set smin/smax
5579 if ((reg->umax_value ^ reg->umin_value) &
5581 reg->smin_value = S32_MIN;
5582 reg->smax_value = S32_MAX;
5584 reg->smin_value = (s64)(s32)reg->smin_value;
5585 reg->smax_value = (s64)(s32)reg->smax_value;
5588 sval = (s64)(s32)val;
5595 if (tnum_is_const(reg->var_off))
5596 return !!tnum_equals_const(reg->var_off, val);
5599 if (tnum_is_const(reg->var_off))
5600 return !tnum_equals_const(reg->var_off, val);
5603 if ((~reg->var_off.mask & reg->var_off.value) & val)
5605 if (!((reg->var_off.mask | reg->var_off.value) & val))
5609 if (reg->umin_value > val)
5611 else if (reg->umax_value <= val)
5615 if (reg->smin_value > sval)
5617 else if (reg->smax_value < sval)
5621 if (reg->umax_value < val)
5623 else if (reg->umin_value >= val)
5627 if (reg->smax_value < sval)
5629 else if (reg->smin_value >= sval)
5633 if (reg->umin_value >= val)
5635 else if (reg->umax_value < val)
5639 if (reg->smin_value >= sval)
5641 else if (reg->smax_value < sval)
5645 if (reg->umax_value <= val)
5647 else if (reg->umin_value > val)
5651 if (reg->smax_value <= sval)
5653 else if (reg->smin_value > sval)
5661 /* Generate min value of the high 32-bit from TNUM info. */
5662 static u64 gen_hi_min(struct tnum var)
5664 return var.value & ~0xffffffffULL;
5667 /* Generate max value of the high 32-bit from TNUM info. */
5668 static u64 gen_hi_max(struct tnum var)
5670 return (var.value | var.mask) & ~0xffffffffULL;
5673 /* Return true if VAL is compared with a s64 sign extended from s32, and they
5674 * are with the same signedness.
5676 static bool cmp_val_with_extended_s64(s64 sval, struct bpf_reg_state *reg)
5678 return ((s32)sval >= 0 &&
5679 reg->smin_value >= 0 && reg->smax_value <= S32_MAX) ||
5681 reg->smax_value <= 0 && reg->smin_value >= S32_MIN);
5684 /* Constrain the possible values of @reg with unsigned upper bound @bound.
5685 * If @is_exclusive, @bound is an exclusive limit, otherwise it is inclusive.
5686 * If @is_jmp32, @bound is a 32-bit value that only constrains the low 32 bits
5689 static void set_upper_bound(struct bpf_reg_state *reg, u64 bound, bool is_jmp32,
5693 /* There are no values for `reg` that make `reg<0` true. */
5699 /* Constrain the register's value in the tnum representation.
5700 * For 64-bit comparisons this happens later in
5701 * __reg_bound_offset(), but for 32-bit comparisons, we can be
5702 * more precise than what can be derived from the updated
5705 struct tnum t = tnum_range(0, bound);
5707 t.mask |= ~0xffffffffULL; /* upper half is unknown */
5708 reg->var_off = tnum_intersect(reg->var_off, t);
5710 /* Compute the 64-bit bound from the 32-bit bound. */
5711 bound += gen_hi_max(reg->var_off);
5713 reg->umax_value = min(reg->umax_value, bound);
5716 /* Constrain the possible values of @reg with unsigned lower bound @bound.
5717 * If @is_exclusive, @bound is an exclusive limit, otherwise it is inclusive.
5718 * If @is_jmp32, @bound is a 32-bit value that only constrains the low 32 bits
5721 static void set_lower_bound(struct bpf_reg_state *reg, u64 bound, bool is_jmp32,
5725 /* There are no values for `reg` that make `reg>MAX` true. */
5726 if (bound == (is_jmp32 ? U32_MAX : U64_MAX))
5731 /* Constrain the register's value in the tnum representation.
5732 * For 64-bit comparisons this happens later in
5733 * __reg_bound_offset(), but for 32-bit comparisons, we can be
5734 * more precise than what can be derived from the updated
5737 struct tnum t = tnum_range(bound, U32_MAX);
5739 t.mask |= ~0xffffffffULL; /* upper half is unknown */
5740 reg->var_off = tnum_intersect(reg->var_off, t);
5742 /* Compute the 64-bit bound from the 32-bit bound. */
5743 bound += gen_hi_min(reg->var_off);
5745 reg->umin_value = max(reg->umin_value, bound);
5748 /* Adjusts the register min/max values in the case that the dst_reg is the
5749 * variable register that we are working on, and src_reg is a constant or we're
5750 * simply doing a BPF_K check.
5751 * In JEQ/JNE cases we also adjust the var_off values.
5753 static void reg_set_min_max(struct bpf_reg_state *true_reg,
5754 struct bpf_reg_state *false_reg, u64 val,
5755 u8 opcode, bool is_jmp32)
5759 /* If the dst_reg is a pointer, we can't learn anything about its
5760 * variable offset from the compare (unless src_reg were a pointer into
5761 * the same object, but we don't bother with that.
5762 * Since false_reg and true_reg have the same type by construction, we
5763 * only need to check one of them for pointerness.
5765 if (__is_pointer_value(false, false_reg))
5768 val = is_jmp32 ? (u32)val : val;
5769 sval = is_jmp32 ? (s64)(s32)val : (s64)val;
5775 struct bpf_reg_state *reg =
5776 opcode == BPF_JEQ ? true_reg : false_reg;
5778 /* For BPF_JEQ, if this is false we know nothing Jon Snow, but
5779 * if it is true we know the value for sure. Likewise for
5783 u64 old_v = reg->var_off.value;
5784 u64 hi_mask = ~0xffffffffULL;
5786 reg->var_off.value = (old_v & hi_mask) | val;
5787 reg->var_off.mask &= hi_mask;
5789 __mark_reg_known(reg, val);
5794 false_reg->var_off = tnum_and(false_reg->var_off,
5796 if (is_power_of_2(val))
5797 true_reg->var_off = tnum_or(true_reg->var_off,
5803 set_upper_bound(false_reg, val, is_jmp32, opcode == BPF_JGE);
5804 set_lower_bound(true_reg, val, is_jmp32, opcode == BPF_JGT);
5810 s64 false_smax = opcode == BPF_JSGT ? sval : sval - 1;
5811 s64 true_smin = opcode == BPF_JSGT ? sval + 1 : sval;
5813 /* If the full s64 was not sign-extended from s32 then don't
5814 * deduct further info.
5816 if (is_jmp32 && !cmp_val_with_extended_s64(sval, false_reg))
5818 false_reg->smax_value = min(false_reg->smax_value, false_smax);
5819 true_reg->smin_value = max(true_reg->smin_value, true_smin);
5825 set_lower_bound(false_reg, val, is_jmp32, opcode == BPF_JLE);
5826 set_upper_bound(true_reg, val, is_jmp32, opcode == BPF_JLT);
5832 s64 false_smin = opcode == BPF_JSLT ? sval : sval + 1;
5833 s64 true_smax = opcode == BPF_JSLT ? sval - 1 : sval;
5835 if (is_jmp32 && !cmp_val_with_extended_s64(sval, false_reg))
5837 false_reg->smin_value = max(false_reg->smin_value, false_smin);
5838 true_reg->smax_value = min(true_reg->smax_value, true_smax);
5845 __reg_deduce_bounds(false_reg);
5846 __reg_deduce_bounds(true_reg);
5847 /* We might have learned some bits from the bounds. */
5848 __reg_bound_offset(false_reg);
5849 __reg_bound_offset(true_reg);
5850 /* Intersecting with the old var_off might have improved our bounds
5851 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
5852 * then new var_off is (0; 0x7f...fc) which improves our umax.
5854 __update_reg_bounds(false_reg);
5855 __update_reg_bounds(true_reg);
5858 /* Same as above, but for the case that dst_reg holds a constant and src_reg is
5861 static void reg_set_min_max_inv(struct bpf_reg_state *true_reg,
5862 struct bpf_reg_state *false_reg, u64 val,
5863 u8 opcode, bool is_jmp32)
5865 /* How can we transform "a <op> b" into "b <op> a"? */
5866 static const u8 opcode_flip[16] = {
5867 /* these stay the same */
5868 [BPF_JEQ >> 4] = BPF_JEQ,
5869 [BPF_JNE >> 4] = BPF_JNE,
5870 [BPF_JSET >> 4] = BPF_JSET,
5871 /* these swap "lesser" and "greater" (L and G in the opcodes) */
5872 [BPF_JGE >> 4] = BPF_JLE,
5873 [BPF_JGT >> 4] = BPF_JLT,
5874 [BPF_JLE >> 4] = BPF_JGE,
5875 [BPF_JLT >> 4] = BPF_JGT,
5876 [BPF_JSGE >> 4] = BPF_JSLE,
5877 [BPF_JSGT >> 4] = BPF_JSLT,
5878 [BPF_JSLE >> 4] = BPF_JSGE,
5879 [BPF_JSLT >> 4] = BPF_JSGT
5881 opcode = opcode_flip[opcode >> 4];
5882 /* This uses zero as "not present in table"; luckily the zero opcode,
5883 * BPF_JA, can't get here.
5886 reg_set_min_max(true_reg, false_reg, val, opcode, is_jmp32);
5889 /* Regs are known to be equal, so intersect their min/max/var_off */
5890 static void __reg_combine_min_max(struct bpf_reg_state *src_reg,
5891 struct bpf_reg_state *dst_reg)
5893 src_reg->umin_value = dst_reg->umin_value = max(src_reg->umin_value,
5894 dst_reg->umin_value);
5895 src_reg->umax_value = dst_reg->umax_value = min(src_reg->umax_value,
5896 dst_reg->umax_value);
5897 src_reg->smin_value = dst_reg->smin_value = max(src_reg->smin_value,
5898 dst_reg->smin_value);
5899 src_reg->smax_value = dst_reg->smax_value = min(src_reg->smax_value,
5900 dst_reg->smax_value);
5901 src_reg->var_off = dst_reg->var_off = tnum_intersect(src_reg->var_off,
5903 /* We might have learned new bounds from the var_off. */
5904 __update_reg_bounds(src_reg);
5905 __update_reg_bounds(dst_reg);
5906 /* We might have learned something about the sign bit. */
5907 __reg_deduce_bounds(src_reg);
5908 __reg_deduce_bounds(dst_reg);
5909 /* We might have learned some bits from the bounds. */
5910 __reg_bound_offset(src_reg);
5911 __reg_bound_offset(dst_reg);
5912 /* Intersecting with the old var_off might have improved our bounds
5913 * slightly. e.g. if umax was 0x7f...f and var_off was (0; 0xf...fc),
5914 * then new var_off is (0; 0x7f...fc) which improves our umax.
5916 __update_reg_bounds(src_reg);
5917 __update_reg_bounds(dst_reg);
5920 static void reg_combine_min_max(struct bpf_reg_state *true_src,
5921 struct bpf_reg_state *true_dst,
5922 struct bpf_reg_state *false_src,
5923 struct bpf_reg_state *false_dst,
5928 __reg_combine_min_max(true_src, true_dst);
5931 __reg_combine_min_max(false_src, false_dst);
5936 static void mark_ptr_or_null_reg(struct bpf_func_state *state,
5937 struct bpf_reg_state *reg, u32 id,
5940 if (reg_type_may_be_null(reg->type) && reg->id == id) {
5941 /* Old offset (both fixed and variable parts) should
5942 * have been known-zero, because we don't allow pointer
5943 * arithmetic on pointers that might be NULL.
5945 if (WARN_ON_ONCE(reg->smin_value || reg->smax_value ||
5946 !tnum_equals_const(reg->var_off, 0) ||
5948 __mark_reg_known_zero(reg);
5952 reg->type = SCALAR_VALUE;
5953 } else if (reg->type == PTR_TO_MAP_VALUE_OR_NULL) {
5954 if (reg->map_ptr->inner_map_meta) {
5955 reg->type = CONST_PTR_TO_MAP;
5956 reg->map_ptr = reg->map_ptr->inner_map_meta;
5957 } else if (reg->map_ptr->map_type ==
5958 BPF_MAP_TYPE_XSKMAP) {
5959 reg->type = PTR_TO_XDP_SOCK;
5961 reg->type = PTR_TO_MAP_VALUE;
5963 } else if (reg->type == PTR_TO_SOCKET_OR_NULL) {
5964 reg->type = PTR_TO_SOCKET;
5965 } else if (reg->type == PTR_TO_SOCK_COMMON_OR_NULL) {
5966 reg->type = PTR_TO_SOCK_COMMON;
5967 } else if (reg->type == PTR_TO_TCP_SOCK_OR_NULL) {
5968 reg->type = PTR_TO_TCP_SOCK;
5971 /* We don't need id and ref_obj_id from this point
5972 * onwards anymore, thus we should better reset it,
5973 * so that state pruning has chances to take effect.
5976 reg->ref_obj_id = 0;
5977 } else if (!reg_may_point_to_spin_lock(reg)) {
5978 /* For not-NULL ptr, reg->ref_obj_id will be reset
5979 * in release_reg_references().
5981 * reg->id is still used by spin_lock ptr. Other
5982 * than spin_lock ptr type, reg->id can be reset.
5989 static void __mark_ptr_or_null_regs(struct bpf_func_state *state, u32 id,
5992 struct bpf_reg_state *reg;
5995 for (i = 0; i < MAX_BPF_REG; i++)
5996 mark_ptr_or_null_reg(state, &state->regs[i], id, is_null);
5998 bpf_for_each_spilled_reg(i, state, reg) {
6001 mark_ptr_or_null_reg(state, reg, id, is_null);
6005 /* The logic is similar to find_good_pkt_pointers(), both could eventually
6006 * be folded together at some point.
6008 static void mark_ptr_or_null_regs(struct bpf_verifier_state *vstate, u32 regno,
6011 struct bpf_func_state *state = vstate->frame[vstate->curframe];
6012 struct bpf_reg_state *regs = state->regs;
6013 u32 ref_obj_id = regs[regno].ref_obj_id;
6014 u32 id = regs[regno].id;
6017 if (ref_obj_id && ref_obj_id == id && is_null)
6018 /* regs[regno] is in the " == NULL" branch.
6019 * No one could have freed the reference state before
6020 * doing the NULL check.
6022 WARN_ON_ONCE(release_reference_state(state, id));
6024 for (i = 0; i <= vstate->curframe; i++)
6025 __mark_ptr_or_null_regs(vstate->frame[i], id, is_null);
6028 static bool try_match_pkt_pointers(const struct bpf_insn *insn,
6029 struct bpf_reg_state *dst_reg,
6030 struct bpf_reg_state *src_reg,
6031 struct bpf_verifier_state *this_branch,
6032 struct bpf_verifier_state *other_branch)
6034 if (BPF_SRC(insn->code) != BPF_X)
6037 /* Pointers are always 64-bit. */
6038 if (BPF_CLASS(insn->code) == BPF_JMP32)
6041 switch (BPF_OP(insn->code)) {
6043 if ((dst_reg->type == PTR_TO_PACKET &&
6044 src_reg->type == PTR_TO_PACKET_END) ||
6045 (dst_reg->type == PTR_TO_PACKET_META &&
6046 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
6047 /* pkt_data' > pkt_end, pkt_meta' > pkt_data */
6048 find_good_pkt_pointers(this_branch, dst_reg,
6049 dst_reg->type, false);
6050 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
6051 src_reg->type == PTR_TO_PACKET) ||
6052 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
6053 src_reg->type == PTR_TO_PACKET_META)) {
6054 /* pkt_end > pkt_data', pkt_data > pkt_meta' */
6055 find_good_pkt_pointers(other_branch, src_reg,
6056 src_reg->type, true);
6062 if ((dst_reg->type == PTR_TO_PACKET &&
6063 src_reg->type == PTR_TO_PACKET_END) ||
6064 (dst_reg->type == PTR_TO_PACKET_META &&
6065 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
6066 /* pkt_data' < pkt_end, pkt_meta' < pkt_data */
6067 find_good_pkt_pointers(other_branch, dst_reg,
6068 dst_reg->type, true);
6069 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
6070 src_reg->type == PTR_TO_PACKET) ||
6071 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
6072 src_reg->type == PTR_TO_PACKET_META)) {
6073 /* pkt_end < pkt_data', pkt_data > pkt_meta' */
6074 find_good_pkt_pointers(this_branch, src_reg,
6075 src_reg->type, false);
6081 if ((dst_reg->type == PTR_TO_PACKET &&
6082 src_reg->type == PTR_TO_PACKET_END) ||
6083 (dst_reg->type == PTR_TO_PACKET_META &&
6084 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
6085 /* pkt_data' >= pkt_end, pkt_meta' >= pkt_data */
6086 find_good_pkt_pointers(this_branch, dst_reg,
6087 dst_reg->type, true);
6088 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
6089 src_reg->type == PTR_TO_PACKET) ||
6090 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
6091 src_reg->type == PTR_TO_PACKET_META)) {
6092 /* pkt_end >= pkt_data', pkt_data >= pkt_meta' */
6093 find_good_pkt_pointers(other_branch, src_reg,
6094 src_reg->type, false);
6100 if ((dst_reg->type == PTR_TO_PACKET &&
6101 src_reg->type == PTR_TO_PACKET_END) ||
6102 (dst_reg->type == PTR_TO_PACKET_META &&
6103 reg_is_init_pkt_pointer(src_reg, PTR_TO_PACKET))) {
6104 /* pkt_data' <= pkt_end, pkt_meta' <= pkt_data */
6105 find_good_pkt_pointers(other_branch, dst_reg,
6106 dst_reg->type, false);
6107 } else if ((dst_reg->type == PTR_TO_PACKET_END &&
6108 src_reg->type == PTR_TO_PACKET) ||
6109 (reg_is_init_pkt_pointer(dst_reg, PTR_TO_PACKET) &&
6110 src_reg->type == PTR_TO_PACKET_META)) {
6111 /* pkt_end <= pkt_data', pkt_data <= pkt_meta' */
6112 find_good_pkt_pointers(this_branch, src_reg,
6113 src_reg->type, true);
6125 static int check_cond_jmp_op(struct bpf_verifier_env *env,
6126 struct bpf_insn *insn, int *insn_idx)
6128 struct bpf_verifier_state *this_branch = env->cur_state;
6129 struct bpf_verifier_state *other_branch;
6130 struct bpf_reg_state *regs = this_branch->frame[this_branch->curframe]->regs;
6131 struct bpf_reg_state *dst_reg, *other_branch_regs, *src_reg = NULL;
6132 u8 opcode = BPF_OP(insn->code);
6137 /* Only conditional jumps are expected to reach here. */
6138 if (opcode == BPF_JA || opcode > BPF_JSLE) {
6139 verbose(env, "invalid BPF_JMP/JMP32 opcode %x\n", opcode);
6143 if (BPF_SRC(insn->code) == BPF_X) {
6144 if (insn->imm != 0) {
6145 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
6149 /* check src1 operand */
6150 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6154 if (is_pointer_value(env, insn->src_reg)) {
6155 verbose(env, "R%d pointer comparison prohibited\n",
6159 src_reg = ®s[insn->src_reg];
6161 if (insn->src_reg != BPF_REG_0) {
6162 verbose(env, "BPF_JMP/JMP32 uses reserved fields\n");
6167 /* check src2 operand */
6168 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
6172 dst_reg = ®s[insn->dst_reg];
6173 is_jmp32 = BPF_CLASS(insn->code) == BPF_JMP32;
6175 if (BPF_SRC(insn->code) == BPF_K)
6176 pred = is_branch_taken(dst_reg, insn->imm,
6178 else if (src_reg->type == SCALAR_VALUE &&
6179 tnum_is_const(src_reg->var_off))
6180 pred = is_branch_taken(dst_reg, src_reg->var_off.value,
6183 err = mark_chain_precision(env, insn->dst_reg);
6184 if (BPF_SRC(insn->code) == BPF_X && !err)
6185 err = mark_chain_precision(env, insn->src_reg);
6190 /* only follow the goto, ignore fall-through */
6191 *insn_idx += insn->off;
6193 } else if (pred == 0) {
6194 /* only follow fall-through branch, since
6195 * that's where the program will go
6200 other_branch = push_stack(env, *insn_idx + insn->off + 1, *insn_idx,
6204 other_branch_regs = other_branch->frame[other_branch->curframe]->regs;
6206 /* detect if we are comparing against a constant value so we can adjust
6207 * our min/max values for our dst register.
6208 * this is only legit if both are scalars (or pointers to the same
6209 * object, I suppose, but we don't support that right now), because
6210 * otherwise the different base pointers mean the offsets aren't
6213 if (BPF_SRC(insn->code) == BPF_X) {
6214 struct bpf_reg_state *src_reg = ®s[insn->src_reg];
6215 struct bpf_reg_state lo_reg0 = *dst_reg;
6216 struct bpf_reg_state lo_reg1 = *src_reg;
6217 struct bpf_reg_state *src_lo, *dst_lo;
6221 coerce_reg_to_size(dst_lo, 4);
6222 coerce_reg_to_size(src_lo, 4);
6224 if (dst_reg->type == SCALAR_VALUE &&
6225 src_reg->type == SCALAR_VALUE) {
6226 if (tnum_is_const(src_reg->var_off) ||
6227 (is_jmp32 && tnum_is_const(src_lo->var_off)))
6228 reg_set_min_max(&other_branch_regs[insn->dst_reg],
6231 ? src_lo->var_off.value
6232 : src_reg->var_off.value,
6234 else if (tnum_is_const(dst_reg->var_off) ||
6235 (is_jmp32 && tnum_is_const(dst_lo->var_off)))
6236 reg_set_min_max_inv(&other_branch_regs[insn->src_reg],
6239 ? dst_lo->var_off.value
6240 : dst_reg->var_off.value,
6242 else if (!is_jmp32 &&
6243 (opcode == BPF_JEQ || opcode == BPF_JNE))
6244 /* Comparing for equality, we can combine knowledge */
6245 reg_combine_min_max(&other_branch_regs[insn->src_reg],
6246 &other_branch_regs[insn->dst_reg],
6247 src_reg, dst_reg, opcode);
6249 } else if (dst_reg->type == SCALAR_VALUE) {
6250 reg_set_min_max(&other_branch_regs[insn->dst_reg],
6251 dst_reg, insn->imm, opcode, is_jmp32);
6254 /* detect if R == 0 where R is returned from bpf_map_lookup_elem().
6255 * NOTE: these optimizations below are related with pointer comparison
6256 * which will never be JMP32.
6258 if (!is_jmp32 && BPF_SRC(insn->code) == BPF_K &&
6259 insn->imm == 0 && (opcode == BPF_JEQ || opcode == BPF_JNE) &&
6260 reg_type_may_be_null(dst_reg->type)) {
6261 /* Mark all identical registers in each branch as either
6262 * safe or unknown depending R == 0 or R != 0 conditional.
6264 mark_ptr_or_null_regs(this_branch, insn->dst_reg,
6266 mark_ptr_or_null_regs(other_branch, insn->dst_reg,
6268 } else if (!try_match_pkt_pointers(insn, dst_reg, ®s[insn->src_reg],
6269 this_branch, other_branch) &&
6270 is_pointer_value(env, insn->dst_reg)) {
6271 verbose(env, "R%d pointer comparison prohibited\n",
6275 if (env->log.level & BPF_LOG_LEVEL)
6276 print_verifier_state(env, this_branch->frame[this_branch->curframe]);
6280 /* verify BPF_LD_IMM64 instruction */
6281 static int check_ld_imm(struct bpf_verifier_env *env, struct bpf_insn *insn)
6283 struct bpf_insn_aux_data *aux = cur_aux(env);
6284 struct bpf_reg_state *regs = cur_regs(env);
6285 struct bpf_map *map;
6288 if (BPF_SIZE(insn->code) != BPF_DW) {
6289 verbose(env, "invalid BPF_LD_IMM insn\n");
6292 if (insn->off != 0) {
6293 verbose(env, "BPF_LD_IMM64 uses reserved fields\n");
6297 err = check_reg_arg(env, insn->dst_reg, DST_OP);
6301 if (insn->src_reg == 0) {
6302 u64 imm = ((u64)(insn + 1)->imm << 32) | (u32)insn->imm;
6304 regs[insn->dst_reg].type = SCALAR_VALUE;
6305 __mark_reg_known(®s[insn->dst_reg], imm);
6309 map = env->used_maps[aux->map_index];
6310 mark_reg_known_zero(env, regs, insn->dst_reg);
6311 regs[insn->dst_reg].map_ptr = map;
6313 if (insn->src_reg == BPF_PSEUDO_MAP_VALUE) {
6314 regs[insn->dst_reg].type = PTR_TO_MAP_VALUE;
6315 regs[insn->dst_reg].off = aux->map_off;
6316 if (map_value_has_spin_lock(map))
6317 regs[insn->dst_reg].id = ++env->id_gen;
6318 } else if (insn->src_reg == BPF_PSEUDO_MAP_FD) {
6319 regs[insn->dst_reg].type = CONST_PTR_TO_MAP;
6321 verbose(env, "bpf verifier is misconfigured\n");
6328 static bool may_access_skb(enum bpf_prog_type type)
6331 case BPF_PROG_TYPE_SOCKET_FILTER:
6332 case BPF_PROG_TYPE_SCHED_CLS:
6333 case BPF_PROG_TYPE_SCHED_ACT:
6340 /* verify safety of LD_ABS|LD_IND instructions:
6341 * - they can only appear in the programs where ctx == skb
6342 * - since they are wrappers of function calls, they scratch R1-R5 registers,
6343 * preserve R6-R9, and store return value into R0
6346 * ctx == skb == R6 == CTX
6349 * SRC == any register
6350 * IMM == 32-bit immediate
6353 * R0 - 8/16/32-bit skb data converted to cpu endianness
6355 static int check_ld_abs(struct bpf_verifier_env *env, struct bpf_insn *insn)
6357 struct bpf_reg_state *regs = cur_regs(env);
6358 static const int ctx_reg = BPF_REG_6;
6359 u8 mode = BPF_MODE(insn->code);
6362 if (!may_access_skb(env->prog->type)) {
6363 verbose(env, "BPF_LD_[ABS|IND] instructions not allowed for this program type\n");
6367 if (!env->ops->gen_ld_abs) {
6368 verbose(env, "bpf verifier is misconfigured\n");
6372 if (env->subprog_cnt > 1) {
6373 /* when program has LD_ABS insn JITs and interpreter assume
6374 * that r1 == ctx == skb which is not the case for callees
6375 * that can have arbitrary arguments. It's problematic
6376 * for main prog as well since JITs would need to analyze
6377 * all functions in order to make proper register save/restore
6378 * decisions in the main prog. Hence disallow LD_ABS with calls
6380 verbose(env, "BPF_LD_[ABS|IND] instructions cannot be mixed with bpf-to-bpf calls\n");
6384 if (insn->dst_reg != BPF_REG_0 || insn->off != 0 ||
6385 BPF_SIZE(insn->code) == BPF_DW ||
6386 (mode == BPF_ABS && insn->src_reg != BPF_REG_0)) {
6387 verbose(env, "BPF_LD_[ABS|IND] uses reserved fields\n");
6391 /* check whether implicit source operand (register R6) is readable */
6392 err = check_reg_arg(env, ctx_reg, SRC_OP);
6396 /* Disallow usage of BPF_LD_[ABS|IND] with reference tracking, as
6397 * gen_ld_abs() may terminate the program at runtime, leading to
6400 err = check_reference_leak(env);
6402 verbose(env, "BPF_LD_[ABS|IND] cannot be mixed with socket references\n");
6406 if (env->cur_state->active_spin_lock) {
6407 verbose(env, "BPF_LD_[ABS|IND] cannot be used inside bpf_spin_lock-ed region\n");
6411 if (regs[ctx_reg].type != PTR_TO_CTX) {
6413 "at the time of BPF_LD_ABS|IND R6 != pointer to skb\n");
6417 if (mode == BPF_IND) {
6418 /* check explicit source operand */
6419 err = check_reg_arg(env, insn->src_reg, SRC_OP);
6424 err = check_ctx_reg(env, ®s[ctx_reg], ctx_reg);
6428 /* reset caller saved regs to unreadable */
6429 for (i = 0; i < CALLER_SAVED_REGS; i++) {
6430 mark_reg_not_init(env, regs, caller_saved[i]);
6431 check_reg_arg(env, caller_saved[i], DST_OP_NO_MARK);
6434 /* mark destination R0 register as readable, since it contains
6435 * the value fetched from the packet.
6436 * Already marked as written above.
6438 mark_reg_unknown(env, regs, BPF_REG_0);
6439 /* ld_abs load up to 32-bit skb data. */
6440 regs[BPF_REG_0].subreg_def = env->insn_idx + 1;
6444 static int check_return_code(struct bpf_verifier_env *env)
6446 struct tnum enforce_attach_type_range = tnum_unknown;
6447 const struct bpf_prog *prog = env->prog;
6448 struct bpf_reg_state *reg;
6449 struct tnum range = tnum_range(0, 1);
6452 /* LSM and struct_ops func-ptr's return type could be "void" */
6453 if ((env->prog->type == BPF_PROG_TYPE_STRUCT_OPS ||
6454 env->prog->type == BPF_PROG_TYPE_LSM) &&
6455 !prog->aux->attach_func_proto->type)
6458 /* eBPF calling convetion is such that R0 is used
6459 * to return the value from eBPF program.
6460 * Make sure that it's readable at this time
6461 * of bpf_exit, which means that program wrote
6462 * something into it earlier
6464 err = check_reg_arg(env, BPF_REG_0, SRC_OP);
6468 if (is_pointer_value(env, BPF_REG_0)) {
6469 verbose(env, "R0 leaks addr as return value\n");
6473 switch (env->prog->type) {
6474 case BPF_PROG_TYPE_CGROUP_SOCK_ADDR:
6475 if (env->prog->expected_attach_type == BPF_CGROUP_UDP4_RECVMSG ||
6476 env->prog->expected_attach_type == BPF_CGROUP_UDP6_RECVMSG)
6477 range = tnum_range(1, 1);
6479 case BPF_PROG_TYPE_CGROUP_SKB:
6480 if (env->prog->expected_attach_type == BPF_CGROUP_INET_EGRESS) {
6481 range = tnum_range(0, 3);
6482 enforce_attach_type_range = tnum_range(2, 3);
6485 case BPF_PROG_TYPE_CGROUP_SOCK:
6486 case BPF_PROG_TYPE_SOCK_OPS:
6487 case BPF_PROG_TYPE_CGROUP_DEVICE:
6488 case BPF_PROG_TYPE_CGROUP_SYSCTL:
6489 case BPF_PROG_TYPE_CGROUP_SOCKOPT:
6491 case BPF_PROG_TYPE_RAW_TRACEPOINT:
6492 if (!env->prog->aux->attach_btf_id)
6494 range = tnum_const(0);
6500 reg = cur_regs(env) + BPF_REG_0;
6501 if (reg->type != SCALAR_VALUE) {
6502 verbose(env, "At program exit the register R0 is not a known value (%s)\n",
6503 reg_type_str[reg->type]);
6507 if (!tnum_in(range, reg->var_off)) {
6510 verbose(env, "At program exit the register R0 ");
6511 if (!tnum_is_unknown(reg->var_off)) {
6512 tnum_strn(tn_buf, sizeof(tn_buf), reg->var_off);
6513 verbose(env, "has value %s", tn_buf);
6515 verbose(env, "has unknown scalar value");
6517 tnum_strn(tn_buf, sizeof(tn_buf), range);
6518 verbose(env, " should have been in %s\n", tn_buf);
6522 if (!tnum_is_unknown(enforce_attach_type_range) &&
6523 tnum_in(enforce_attach_type_range, reg->var_off))
6524 env->prog->enforce_expected_attach_type = 1;
6528 /* non-recursive DFS pseudo code
6529 * 1 procedure DFS-iterative(G,v):
6530 * 2 label v as discovered
6531 * 3 let S be a stack
6533 * 5 while S is not empty
6535 * 7 if t is what we're looking for:
6537 * 9 for all edges e in G.adjacentEdges(t) do
6538 * 10 if edge e is already labelled
6539 * 11 continue with the next edge
6540 * 12 w <- G.adjacentVertex(t,e)
6541 * 13 if vertex w is not discovered and not explored
6542 * 14 label e as tree-edge
6543 * 15 label w as discovered
6546 * 18 else if vertex w is discovered
6547 * 19 label e as back-edge
6549 * 21 // vertex w is explored
6550 * 22 label e as forward- or cross-edge
6551 * 23 label t as explored
6556 * 0x11 - discovered and fall-through edge labelled
6557 * 0x12 - discovered and fall-through and branch edges labelled
6568 static u32 state_htab_size(struct bpf_verifier_env *env)
6570 return env->prog->len;
6573 static struct bpf_verifier_state_list **explored_state(
6574 struct bpf_verifier_env *env,
6577 struct bpf_verifier_state *cur = env->cur_state;
6578 struct bpf_func_state *state = cur->frame[cur->curframe];
6580 return &env->explored_states[(idx ^ state->callsite) % state_htab_size(env)];
6583 static void init_explored_state(struct bpf_verifier_env *env, int idx)
6585 env->insn_aux_data[idx].prune_point = true;
6588 /* t, w, e - match pseudo-code above:
6589 * t - index of current instruction
6590 * w - next instruction
6593 static int push_insn(int t, int w, int e, struct bpf_verifier_env *env,
6596 int *insn_stack = env->cfg.insn_stack;
6597 int *insn_state = env->cfg.insn_state;
6599 if (e == FALLTHROUGH && insn_state[t] >= (DISCOVERED | FALLTHROUGH))
6602 if (e == BRANCH && insn_state[t] >= (DISCOVERED | BRANCH))
6605 if (w < 0 || w >= env->prog->len) {
6606 verbose_linfo(env, t, "%d: ", t);
6607 verbose(env, "jump out of range from insn %d to %d\n", t, w);
6612 /* mark branch target for state pruning */
6613 init_explored_state(env, w);
6615 if (insn_state[w] == 0) {
6617 insn_state[t] = DISCOVERED | e;
6618 insn_state[w] = DISCOVERED;
6619 if (env->cfg.cur_stack >= env->prog->len)
6621 insn_stack[env->cfg.cur_stack++] = w;
6623 } else if ((insn_state[w] & 0xF0) == DISCOVERED) {
6624 if (loop_ok && env->allow_ptr_leaks)
6626 verbose_linfo(env, t, "%d: ", t);
6627 verbose_linfo(env, w, "%d: ", w);
6628 verbose(env, "back-edge from insn %d to %d\n", t, w);
6630 } else if (insn_state[w] == EXPLORED) {
6631 /* forward- or cross-edge */
6632 insn_state[t] = DISCOVERED | e;
6634 verbose(env, "insn state internal bug\n");
6640 /* non-recursive depth-first-search to detect loops in BPF program
6641 * loop == back-edge in directed graph
6643 static int check_cfg(struct bpf_verifier_env *env)
6645 struct bpf_insn *insns = env->prog->insnsi;
6646 int insn_cnt = env->prog->len;
6647 int *insn_stack, *insn_state;
6651 insn_state = env->cfg.insn_state = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
6655 insn_stack = env->cfg.insn_stack = kvcalloc(insn_cnt, sizeof(int), GFP_KERNEL);
6661 insn_state[0] = DISCOVERED; /* mark 1st insn as discovered */
6662 insn_stack[0] = 0; /* 0 is the first instruction */
6663 env->cfg.cur_stack = 1;
6666 if (env->cfg.cur_stack == 0)
6668 t = insn_stack[env->cfg.cur_stack - 1];
6670 if (BPF_CLASS(insns[t].code) == BPF_JMP ||
6671 BPF_CLASS(insns[t].code) == BPF_JMP32) {
6672 u8 opcode = BPF_OP(insns[t].code);
6674 if (opcode == BPF_EXIT) {
6676 } else if (opcode == BPF_CALL) {
6677 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
6682 if (t + 1 < insn_cnt)
6683 init_explored_state(env, t + 1);
6684 if (insns[t].src_reg == BPF_PSEUDO_CALL) {
6685 init_explored_state(env, t);
6686 ret = push_insn(t, t + insns[t].imm + 1, BRANCH,
6693 } else if (opcode == BPF_JA) {
6694 if (BPF_SRC(insns[t].code) != BPF_K) {
6698 /* unconditional jump with single edge */
6699 ret = push_insn(t, t + insns[t].off + 1,
6700 FALLTHROUGH, env, true);
6705 /* unconditional jmp is not a good pruning point,
6706 * but it's marked, since backtracking needs
6707 * to record jmp history in is_state_visited().
6709 init_explored_state(env, t + insns[t].off + 1);
6710 /* tell verifier to check for equivalent states
6711 * after every call and jump
6713 if (t + 1 < insn_cnt)
6714 init_explored_state(env, t + 1);
6716 /* conditional jump with two edges */
6717 init_explored_state(env, t);
6718 ret = push_insn(t, t + 1, FALLTHROUGH, env, true);
6724 ret = push_insn(t, t + insns[t].off + 1, BRANCH, env, true);
6731 /* all other non-branch instructions with single
6734 ret = push_insn(t, t + 1, FALLTHROUGH, env, false);
6742 insn_state[t] = EXPLORED;
6743 if (env->cfg.cur_stack-- <= 0) {
6744 verbose(env, "pop stack internal bug\n");
6751 for (i = 0; i < insn_cnt; i++) {
6752 if (insn_state[i] != EXPLORED) {
6753 verbose(env, "unreachable insn %d\n", i);
6758 ret = 0; /* cfg looks good */
6763 env->cfg.insn_state = env->cfg.insn_stack = NULL;
6767 /* The minimum supported BTF func info size */
6768 #define MIN_BPF_FUNCINFO_SIZE 8
6769 #define MAX_FUNCINFO_REC_SIZE 252
6771 static int check_btf_func(struct bpf_verifier_env *env,
6772 const union bpf_attr *attr,
6773 union bpf_attr __user *uattr)
6775 u32 i, nfuncs, urec_size, min_size;
6776 u32 krec_size = sizeof(struct bpf_func_info);
6777 struct bpf_func_info *krecord;
6778 struct bpf_func_info_aux *info_aux = NULL;
6779 const struct btf_type *type;
6780 struct bpf_prog *prog;
6781 const struct btf *btf;
6782 void __user *urecord;
6783 u32 prev_offset = 0;
6786 nfuncs = attr->func_info_cnt;
6790 if (nfuncs != env->subprog_cnt) {
6791 verbose(env, "number of funcs in func_info doesn't match number of subprogs\n");
6795 urec_size = attr->func_info_rec_size;
6796 if (urec_size < MIN_BPF_FUNCINFO_SIZE ||
6797 urec_size > MAX_FUNCINFO_REC_SIZE ||
6798 urec_size % sizeof(u32)) {
6799 verbose(env, "invalid func info rec size %u\n", urec_size);
6804 btf = prog->aux->btf;
6806 urecord = u64_to_user_ptr(attr->func_info);
6807 min_size = min_t(u32, krec_size, urec_size);
6809 krecord = kvcalloc(nfuncs, krec_size, GFP_KERNEL | __GFP_NOWARN);
6812 info_aux = kcalloc(nfuncs, sizeof(*info_aux), GFP_KERNEL | __GFP_NOWARN);
6816 for (i = 0; i < nfuncs; i++) {
6817 ret = bpf_check_uarg_tail_zero(urecord, krec_size, urec_size);
6819 if (ret == -E2BIG) {
6820 verbose(env, "nonzero tailing record in func info");
6821 /* set the size kernel expects so loader can zero
6822 * out the rest of the record.
6824 if (put_user(min_size, &uattr->func_info_rec_size))
6830 if (copy_from_user(&krecord[i], urecord, min_size)) {
6835 /* check insn_off */
6837 if (krecord[i].insn_off) {
6839 "nonzero insn_off %u for the first func info record",
6840 krecord[i].insn_off);
6844 } else if (krecord[i].insn_off <= prev_offset) {
6846 "same or smaller insn offset (%u) than previous func info record (%u)",
6847 krecord[i].insn_off, prev_offset);
6852 if (env->subprog_info[i].start != krecord[i].insn_off) {
6853 verbose(env, "func_info BTF section doesn't match subprog layout in BPF program\n");
6859 type = btf_type_by_id(btf, krecord[i].type_id);
6860 if (!type || !btf_type_is_func(type)) {
6861 verbose(env, "invalid type id %d in func info",
6862 krecord[i].type_id);
6866 info_aux[i].linkage = BTF_INFO_VLEN(type->info);
6867 prev_offset = krecord[i].insn_off;
6868 urecord += urec_size;
6871 prog->aux->func_info = krecord;
6872 prog->aux->func_info_cnt = nfuncs;
6873 prog->aux->func_info_aux = info_aux;
6882 static void adjust_btf_func(struct bpf_verifier_env *env)
6884 struct bpf_prog_aux *aux = env->prog->aux;
6887 if (!aux->func_info)
6890 for (i = 0; i < env->subprog_cnt; i++)
6891 aux->func_info[i].insn_off = env->subprog_info[i].start;
6894 #define MIN_BPF_LINEINFO_SIZE (offsetof(struct bpf_line_info, line_col) + \
6895 sizeof(((struct bpf_line_info *)(0))->line_col))
6896 #define MAX_LINEINFO_REC_SIZE MAX_FUNCINFO_REC_SIZE
6898 static int check_btf_line(struct bpf_verifier_env *env,
6899 const union bpf_attr *attr,
6900 union bpf_attr __user *uattr)
6902 u32 i, s, nr_linfo, ncopy, expected_size, rec_size, prev_offset = 0;
6903 struct bpf_subprog_info *sub;
6904 struct bpf_line_info *linfo;
6905 struct bpf_prog *prog;
6906 const struct btf *btf;
6907 void __user *ulinfo;
6910 nr_linfo = attr->line_info_cnt;
6914 rec_size = attr->line_info_rec_size;
6915 if (rec_size < MIN_BPF_LINEINFO_SIZE ||
6916 rec_size > MAX_LINEINFO_REC_SIZE ||
6917 rec_size & (sizeof(u32) - 1))
6920 /* Need to zero it in case the userspace may
6921 * pass in a smaller bpf_line_info object.
6923 linfo = kvcalloc(nr_linfo, sizeof(struct bpf_line_info),
6924 GFP_KERNEL | __GFP_NOWARN);
6929 btf = prog->aux->btf;
6932 sub = env->subprog_info;
6933 ulinfo = u64_to_user_ptr(attr->line_info);
6934 expected_size = sizeof(struct bpf_line_info);
6935 ncopy = min_t(u32, expected_size, rec_size);
6936 for (i = 0; i < nr_linfo; i++) {
6937 err = bpf_check_uarg_tail_zero(ulinfo, expected_size, rec_size);
6939 if (err == -E2BIG) {
6940 verbose(env, "nonzero tailing record in line_info");
6941 if (put_user(expected_size,
6942 &uattr->line_info_rec_size))
6948 if (copy_from_user(&linfo[i], ulinfo, ncopy)) {
6954 * Check insn_off to ensure
6955 * 1) strictly increasing AND
6956 * 2) bounded by prog->len
6958 * The linfo[0].insn_off == 0 check logically falls into
6959 * the later "missing bpf_line_info for func..." case
6960 * because the first linfo[0].insn_off must be the
6961 * first sub also and the first sub must have
6962 * subprog_info[0].start == 0.
6964 if ((i && linfo[i].insn_off <= prev_offset) ||
6965 linfo[i].insn_off >= prog->len) {
6966 verbose(env, "Invalid line_info[%u].insn_off:%u (prev_offset:%u prog->len:%u)\n",
6967 i, linfo[i].insn_off, prev_offset,
6973 if (!prog->insnsi[linfo[i].insn_off].code) {
6975 "Invalid insn code at line_info[%u].insn_off\n",
6981 if (!btf_name_by_offset(btf, linfo[i].line_off) ||
6982 !btf_name_by_offset(btf, linfo[i].file_name_off)) {
6983 verbose(env, "Invalid line_info[%u].line_off or .file_name_off\n", i);
6988 if (s != env->subprog_cnt) {
6989 if (linfo[i].insn_off == sub[s].start) {
6990 sub[s].linfo_idx = i;
6992 } else if (sub[s].start < linfo[i].insn_off) {
6993 verbose(env, "missing bpf_line_info for func#%u\n", s);
6999 prev_offset = linfo[i].insn_off;
7003 if (s != env->subprog_cnt) {
7004 verbose(env, "missing bpf_line_info for %u funcs starting from func#%u\n",
7005 env->subprog_cnt - s, s);
7010 prog->aux->linfo = linfo;
7011 prog->aux->nr_linfo = nr_linfo;
7020 static int check_btf_info(struct bpf_verifier_env *env,
7021 const union bpf_attr *attr,
7022 union bpf_attr __user *uattr)
7027 if (!attr->func_info_cnt && !attr->line_info_cnt)
7030 btf = btf_get_by_fd(attr->prog_btf_fd);
7032 return PTR_ERR(btf);
7033 env->prog->aux->btf = btf;
7035 err = check_btf_func(env, attr, uattr);
7039 err = check_btf_line(env, attr, uattr);
7046 /* check %cur's range satisfies %old's */
7047 static bool range_within(struct bpf_reg_state *old,
7048 struct bpf_reg_state *cur)
7050 return old->umin_value <= cur->umin_value &&
7051 old->umax_value >= cur->umax_value &&
7052 old->smin_value <= cur->smin_value &&
7053 old->smax_value >= cur->smax_value;
7056 /* Maximum number of register states that can exist at once */
7057 #define ID_MAP_SIZE (MAX_BPF_REG + MAX_BPF_STACK / BPF_REG_SIZE)
7063 /* If in the old state two registers had the same id, then they need to have
7064 * the same id in the new state as well. But that id could be different from
7065 * the old state, so we need to track the mapping from old to new ids.
7066 * Once we have seen that, say, a reg with old id 5 had new id 9, any subsequent
7067 * regs with old id 5 must also have new id 9 for the new state to be safe. But
7068 * regs with a different old id could still have new id 9, we don't care about
7070 * So we look through our idmap to see if this old id has been seen before. If
7071 * so, we require the new id to match; otherwise, we add the id pair to the map.
7073 static bool check_ids(u32 old_id, u32 cur_id, struct idpair *idmap)
7077 for (i = 0; i < ID_MAP_SIZE; i++) {
7078 if (!idmap[i].old) {
7079 /* Reached an empty slot; haven't seen this id before */
7080 idmap[i].old = old_id;
7081 idmap[i].cur = cur_id;
7084 if (idmap[i].old == old_id)
7085 return idmap[i].cur == cur_id;
7087 /* We ran out of idmap slots, which should be impossible */
7092 static void clean_func_state(struct bpf_verifier_env *env,
7093 struct bpf_func_state *st)
7095 enum bpf_reg_liveness live;
7098 for (i = 0; i < BPF_REG_FP; i++) {
7099 live = st->regs[i].live;
7100 /* liveness must not touch this register anymore */
7101 st->regs[i].live |= REG_LIVE_DONE;
7102 if (!(live & REG_LIVE_READ))
7103 /* since the register is unused, clear its state
7104 * to make further comparison simpler
7106 __mark_reg_not_init(env, &st->regs[i]);
7109 for (i = 0; i < st->allocated_stack / BPF_REG_SIZE; i++) {
7110 live = st->stack[i].spilled_ptr.live;
7111 /* liveness must not touch this stack slot anymore */
7112 st->stack[i].spilled_ptr.live |= REG_LIVE_DONE;
7113 if (!(live & REG_LIVE_READ)) {
7114 __mark_reg_not_init(env, &st->stack[i].spilled_ptr);
7115 for (j = 0; j < BPF_REG_SIZE; j++)
7116 st->stack[i].slot_type[j] = STACK_INVALID;
7121 static void clean_verifier_state(struct bpf_verifier_env *env,
7122 struct bpf_verifier_state *st)
7126 if (st->frame[0]->regs[0].live & REG_LIVE_DONE)
7127 /* all regs in this state in all frames were already marked */
7130 for (i = 0; i <= st->curframe; i++)
7131 clean_func_state(env, st->frame[i]);
7134 /* the parentage chains form a tree.
7135 * the verifier states are added to state lists at given insn and
7136 * pushed into state stack for future exploration.
7137 * when the verifier reaches bpf_exit insn some of the verifer states
7138 * stored in the state lists have their final liveness state already,
7139 * but a lot of states will get revised from liveness point of view when
7140 * the verifier explores other branches.
7143 * 2: if r1 == 100 goto pc+1
7146 * when the verifier reaches exit insn the register r0 in the state list of
7147 * insn 2 will be seen as !REG_LIVE_READ. Then the verifier pops the other_branch
7148 * of insn 2 and goes exploring further. At the insn 4 it will walk the
7149 * parentage chain from insn 4 into insn 2 and will mark r0 as REG_LIVE_READ.
7151 * Since the verifier pushes the branch states as it sees them while exploring
7152 * the program the condition of walking the branch instruction for the second
7153 * time means that all states below this branch were already explored and
7154 * their final liveness markes are already propagated.
7155 * Hence when the verifier completes the search of state list in is_state_visited()
7156 * we can call this clean_live_states() function to mark all liveness states
7157 * as REG_LIVE_DONE to indicate that 'parent' pointers of 'struct bpf_reg_state'
7159 * This function also clears the registers and stack for states that !READ
7160 * to simplify state merging.
7162 * Important note here that walking the same branch instruction in the callee
7163 * doesn't meant that the states are DONE. The verifier has to compare
7166 static void clean_live_states(struct bpf_verifier_env *env, int insn,
7167 struct bpf_verifier_state *cur)
7169 struct bpf_verifier_state_list *sl;
7172 sl = *explored_state(env, insn);
7174 if (sl->state.branches)
7176 if (sl->state.insn_idx != insn ||
7177 sl->state.curframe != cur->curframe)
7179 for (i = 0; i <= cur->curframe; i++)
7180 if (sl->state.frame[i]->callsite != cur->frame[i]->callsite)
7182 clean_verifier_state(env, &sl->state);
7188 /* Returns true if (rold safe implies rcur safe) */
7189 static bool regsafe(struct bpf_reg_state *rold, struct bpf_reg_state *rcur,
7190 struct idpair *idmap)
7194 if (!(rold->live & REG_LIVE_READ))
7195 /* explored state didn't use this */
7198 equal = memcmp(rold, rcur, offsetof(struct bpf_reg_state, parent)) == 0;
7200 if (rold->type == PTR_TO_STACK)
7201 /* two stack pointers are equal only if they're pointing to
7202 * the same stack frame, since fp-8 in foo != fp-8 in bar
7204 return equal && rold->frameno == rcur->frameno;
7209 if (rold->type == NOT_INIT)
7210 /* explored state can't have used this */
7212 if (rcur->type == NOT_INIT)
7214 switch (rold->type) {
7216 if (rcur->type == SCALAR_VALUE) {
7217 if (!rold->precise && !rcur->precise)
7219 /* new val must satisfy old val knowledge */
7220 return range_within(rold, rcur) &&
7221 tnum_in(rold->var_off, rcur->var_off);
7223 /* We're trying to use a pointer in place of a scalar.
7224 * Even if the scalar was unbounded, this could lead to
7225 * pointer leaks because scalars are allowed to leak
7226 * while pointers are not. We could make this safe in
7227 * special cases if root is calling us, but it's
7228 * probably not worth the hassle.
7232 case PTR_TO_MAP_VALUE:
7233 /* If the new min/max/var_off satisfy the old ones and
7234 * everything else matches, we are OK.
7235 * 'id' is not compared, since it's only used for maps with
7236 * bpf_spin_lock inside map element and in such cases if
7237 * the rest of the prog is valid for one map element then
7238 * it's valid for all map elements regardless of the key
7239 * used in bpf_map_lookup()
7241 return memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)) == 0 &&
7242 range_within(rold, rcur) &&
7243 tnum_in(rold->var_off, rcur->var_off);
7244 case PTR_TO_MAP_VALUE_OR_NULL:
7245 /* a PTR_TO_MAP_VALUE could be safe to use as a
7246 * PTR_TO_MAP_VALUE_OR_NULL into the same map.
7247 * However, if the old PTR_TO_MAP_VALUE_OR_NULL then got NULL-
7248 * checked, doing so could have affected others with the same
7249 * id, and we can't check for that because we lost the id when
7250 * we converted to a PTR_TO_MAP_VALUE.
7252 if (rcur->type != PTR_TO_MAP_VALUE_OR_NULL)
7254 if (memcmp(rold, rcur, offsetof(struct bpf_reg_state, id)))
7256 /* Check our ids match any regs they're supposed to */
7257 return check_ids(rold->id, rcur->id, idmap);
7258 case PTR_TO_PACKET_META:
7260 if (rcur->type != rold->type)
7262 /* We must have at least as much range as the old ptr
7263 * did, so that any accesses which were safe before are
7264 * still safe. This is true even if old range < old off,
7265 * since someone could have accessed through (ptr - k), or
7266 * even done ptr -= k in a register, to get a safe access.
7268 if (rold->range > rcur->range)
7270 /* If the offsets don't match, we can't trust our alignment;
7271 * nor can we be sure that we won't fall out of range.
7273 if (rold->off != rcur->off)
7275 /* id relations must be preserved */
7276 if (rold->id && !check_ids(rold->id, rcur->id, idmap))
7278 /* new val must satisfy old val knowledge */
7279 return range_within(rold, rcur) &&
7280 tnum_in(rold->var_off, rcur->var_off);
7282 case CONST_PTR_TO_MAP:
7283 case PTR_TO_PACKET_END:
7284 case PTR_TO_FLOW_KEYS:
7286 case PTR_TO_SOCKET_OR_NULL:
7287 case PTR_TO_SOCK_COMMON:
7288 case PTR_TO_SOCK_COMMON_OR_NULL:
7289 case PTR_TO_TCP_SOCK:
7290 case PTR_TO_TCP_SOCK_OR_NULL:
7291 case PTR_TO_XDP_SOCK:
7292 /* Only valid matches are exact, which memcmp() above
7293 * would have accepted
7296 /* Don't know what's going on, just say it's not safe */
7300 /* Shouldn't get here; if we do, say it's not safe */
7305 static bool stacksafe(struct bpf_func_state *old,
7306 struct bpf_func_state *cur,
7307 struct idpair *idmap)
7311 /* walk slots of the explored stack and ignore any additional
7312 * slots in the current stack, since explored(safe) state
7315 for (i = 0; i < old->allocated_stack; i++) {
7316 spi = i / BPF_REG_SIZE;
7318 if (!(old->stack[spi].spilled_ptr.live & REG_LIVE_READ)) {
7319 i += BPF_REG_SIZE - 1;
7320 /* explored state didn't use this */
7324 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_INVALID)
7327 /* explored stack has more populated slots than current stack
7328 * and these slots were used
7330 if (i >= cur->allocated_stack)
7333 /* if old state was safe with misc data in the stack
7334 * it will be safe with zero-initialized stack.
7335 * The opposite is not true
7337 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_MISC &&
7338 cur->stack[spi].slot_type[i % BPF_REG_SIZE] == STACK_ZERO)
7340 if (old->stack[spi].slot_type[i % BPF_REG_SIZE] !=
7341 cur->stack[spi].slot_type[i % BPF_REG_SIZE])
7342 /* Ex: old explored (safe) state has STACK_SPILL in
7343 * this stack slot, but current has has STACK_MISC ->
7344 * this verifier states are not equivalent,
7345 * return false to continue verification of this path
7348 if (i % BPF_REG_SIZE)
7350 if (old->stack[spi].slot_type[0] != STACK_SPILL)
7352 if (!regsafe(&old->stack[spi].spilled_ptr,
7353 &cur->stack[spi].spilled_ptr,
7355 /* when explored and current stack slot are both storing
7356 * spilled registers, check that stored pointers types
7357 * are the same as well.
7358 * Ex: explored safe path could have stored
7359 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -8}
7360 * but current path has stored:
7361 * (bpf_reg_state) {.type = PTR_TO_STACK, .off = -16}
7362 * such verifier states are not equivalent.
7363 * return false to continue verification of this path
7370 static bool refsafe(struct bpf_func_state *old, struct bpf_func_state *cur)
7372 if (old->acquired_refs != cur->acquired_refs)
7374 return !memcmp(old->refs, cur->refs,
7375 sizeof(*old->refs) * old->acquired_refs);
7378 /* compare two verifier states
7380 * all states stored in state_list are known to be valid, since
7381 * verifier reached 'bpf_exit' instruction through them
7383 * this function is called when verifier exploring different branches of
7384 * execution popped from the state stack. If it sees an old state that has
7385 * more strict register state and more strict stack state then this execution
7386 * branch doesn't need to be explored further, since verifier already
7387 * concluded that more strict state leads to valid finish.
7389 * Therefore two states are equivalent if register state is more conservative
7390 * and explored stack state is more conservative than the current one.
7393 * (slot1=INV slot2=MISC) == (slot1=MISC slot2=MISC)
7394 * (slot1=MISC slot2=MISC) != (slot1=INV slot2=MISC)
7396 * In other words if current stack state (one being explored) has more
7397 * valid slots than old one that already passed validation, it means
7398 * the verifier can stop exploring and conclude that current state is valid too
7400 * Similarly with registers. If explored state has register type as invalid
7401 * whereas register type in current state is meaningful, it means that
7402 * the current state will reach 'bpf_exit' instruction safely
7404 static bool func_states_equal(struct bpf_func_state *old,
7405 struct bpf_func_state *cur)
7407 struct idpair *idmap;
7411 idmap = kcalloc(ID_MAP_SIZE, sizeof(struct idpair), GFP_KERNEL);
7412 /* If we failed to allocate the idmap, just say it's not safe */
7416 for (i = 0; i < MAX_BPF_REG; i++) {
7417 if (!regsafe(&old->regs[i], &cur->regs[i], idmap))
7421 if (!stacksafe(old, cur, idmap))
7424 if (!refsafe(old, cur))
7432 static bool states_equal(struct bpf_verifier_env *env,
7433 struct bpf_verifier_state *old,
7434 struct bpf_verifier_state *cur)
7438 if (old->curframe != cur->curframe)
7441 /* Verification state from speculative execution simulation
7442 * must never prune a non-speculative execution one.
7444 if (old->speculative && !cur->speculative)
7447 if (old->active_spin_lock != cur->active_spin_lock)
7450 /* for states to be equal callsites have to be the same
7451 * and all frame states need to be equivalent
7453 for (i = 0; i <= old->curframe; i++) {
7454 if (old->frame[i]->callsite != cur->frame[i]->callsite)
7456 if (!func_states_equal(old->frame[i], cur->frame[i]))
7462 /* Return 0 if no propagation happened. Return negative error code if error
7463 * happened. Otherwise, return the propagated bit.
7465 static int propagate_liveness_reg(struct bpf_verifier_env *env,
7466 struct bpf_reg_state *reg,
7467 struct bpf_reg_state *parent_reg)
7469 u8 parent_flag = parent_reg->live & REG_LIVE_READ;
7470 u8 flag = reg->live & REG_LIVE_READ;
7473 /* When comes here, read flags of PARENT_REG or REG could be any of
7474 * REG_LIVE_READ64, REG_LIVE_READ32, REG_LIVE_NONE. There is no need
7475 * of propagation if PARENT_REG has strongest REG_LIVE_READ64.
7477 if (parent_flag == REG_LIVE_READ64 ||
7478 /* Or if there is no read flag from REG. */
7480 /* Or if the read flag from REG is the same as PARENT_REG. */
7481 parent_flag == flag)
7484 err = mark_reg_read(env, reg, parent_reg, flag);
7491 /* A write screens off any subsequent reads; but write marks come from the
7492 * straight-line code between a state and its parent. When we arrive at an
7493 * equivalent state (jump target or such) we didn't arrive by the straight-line
7494 * code, so read marks in the state must propagate to the parent regardless
7495 * of the state's write marks. That's what 'parent == state->parent' comparison
7496 * in mark_reg_read() is for.
7498 static int propagate_liveness(struct bpf_verifier_env *env,
7499 const struct bpf_verifier_state *vstate,
7500 struct bpf_verifier_state *vparent)
7502 struct bpf_reg_state *state_reg, *parent_reg;
7503 struct bpf_func_state *state, *parent;
7504 int i, frame, err = 0;
7506 if (vparent->curframe != vstate->curframe) {
7507 WARN(1, "propagate_live: parent frame %d current frame %d\n",
7508 vparent->curframe, vstate->curframe);
7511 /* Propagate read liveness of registers... */
7512 BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
7513 for (frame = 0; frame <= vstate->curframe; frame++) {
7514 parent = vparent->frame[frame];
7515 state = vstate->frame[frame];
7516 parent_reg = parent->regs;
7517 state_reg = state->regs;
7518 /* We don't need to worry about FP liveness, it's read-only */
7519 for (i = frame < vstate->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++) {
7520 err = propagate_liveness_reg(env, &state_reg[i],
7524 if (err == REG_LIVE_READ64)
7525 mark_insn_zext(env, &parent_reg[i]);
7528 /* Propagate stack slots. */
7529 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE &&
7530 i < parent->allocated_stack / BPF_REG_SIZE; i++) {
7531 parent_reg = &parent->stack[i].spilled_ptr;
7532 state_reg = &state->stack[i].spilled_ptr;
7533 err = propagate_liveness_reg(env, state_reg,
7542 /* find precise scalars in the previous equivalent state and
7543 * propagate them into the current state
7545 static int propagate_precision(struct bpf_verifier_env *env,
7546 const struct bpf_verifier_state *old)
7548 struct bpf_reg_state *state_reg;
7549 struct bpf_func_state *state;
7552 state = old->frame[old->curframe];
7553 state_reg = state->regs;
7554 for (i = 0; i < BPF_REG_FP; i++, state_reg++) {
7555 if (state_reg->type != SCALAR_VALUE ||
7556 !state_reg->precise)
7558 if (env->log.level & BPF_LOG_LEVEL2)
7559 verbose(env, "propagating r%d\n", i);
7560 err = mark_chain_precision(env, i);
7565 for (i = 0; i < state->allocated_stack / BPF_REG_SIZE; i++) {
7566 if (state->stack[i].slot_type[0] != STACK_SPILL)
7568 state_reg = &state->stack[i].spilled_ptr;
7569 if (state_reg->type != SCALAR_VALUE ||
7570 !state_reg->precise)
7572 if (env->log.level & BPF_LOG_LEVEL2)
7573 verbose(env, "propagating fp%d\n",
7574 (-i - 1) * BPF_REG_SIZE);
7575 err = mark_chain_precision_stack(env, i);
7582 static bool states_maybe_looping(struct bpf_verifier_state *old,
7583 struct bpf_verifier_state *cur)
7585 struct bpf_func_state *fold, *fcur;
7586 int i, fr = cur->curframe;
7588 if (old->curframe != fr)
7591 fold = old->frame[fr];
7592 fcur = cur->frame[fr];
7593 for (i = 0; i < MAX_BPF_REG; i++)
7594 if (memcmp(&fold->regs[i], &fcur->regs[i],
7595 offsetof(struct bpf_reg_state, parent)))
7601 static int is_state_visited(struct bpf_verifier_env *env, int insn_idx)
7603 struct bpf_verifier_state_list *new_sl;
7604 struct bpf_verifier_state_list *sl, **pprev;
7605 struct bpf_verifier_state *cur = env->cur_state, *new;
7606 int i, j, err, states_cnt = 0;
7607 bool add_new_state = env->test_state_freq ? true : false;
7609 cur->last_insn_idx = env->prev_insn_idx;
7610 if (!env->insn_aux_data[insn_idx].prune_point)
7611 /* this 'insn_idx' instruction wasn't marked, so we will not
7612 * be doing state search here
7616 /* bpf progs typically have pruning point every 4 instructions
7617 * http://vger.kernel.org/bpfconf2019.html#session-1
7618 * Do not add new state for future pruning if the verifier hasn't seen
7619 * at least 2 jumps and at least 8 instructions.
7620 * This heuristics helps decrease 'total_states' and 'peak_states' metric.
7621 * In tests that amounts to up to 50% reduction into total verifier
7622 * memory consumption and 20% verifier time speedup.
7624 if (env->jmps_processed - env->prev_jmps_processed >= 2 &&
7625 env->insn_processed - env->prev_insn_processed >= 8)
7626 add_new_state = true;
7628 pprev = explored_state(env, insn_idx);
7631 clean_live_states(env, insn_idx, cur);
7635 if (sl->state.insn_idx != insn_idx)
7637 if (sl->state.branches) {
7638 if (states_maybe_looping(&sl->state, cur) &&
7639 states_equal(env, &sl->state, cur)) {
7640 verbose_linfo(env, insn_idx, "; ");
7641 verbose(env, "infinite loop detected at insn %d\n", insn_idx);
7644 /* if the verifier is processing a loop, avoid adding new state
7645 * too often, since different loop iterations have distinct
7646 * states and may not help future pruning.
7647 * This threshold shouldn't be too low to make sure that
7648 * a loop with large bound will be rejected quickly.
7649 * The most abusive loop will be:
7651 * if r1 < 1000000 goto pc-2
7652 * 1M insn_procssed limit / 100 == 10k peak states.
7653 * This threshold shouldn't be too high either, since states
7654 * at the end of the loop are likely to be useful in pruning.
7656 if (env->jmps_processed - env->prev_jmps_processed < 20 &&
7657 env->insn_processed - env->prev_insn_processed < 100)
7658 add_new_state = false;
7661 if (states_equal(env, &sl->state, cur)) {
7663 /* reached equivalent register/stack state,
7665 * Registers read by the continuation are read by us.
7666 * If we have any write marks in env->cur_state, they
7667 * will prevent corresponding reads in the continuation
7668 * from reaching our parent (an explored_state). Our
7669 * own state will get the read marks recorded, but
7670 * they'll be immediately forgotten as we're pruning
7671 * this state and will pop a new one.
7673 err = propagate_liveness(env, &sl->state, cur);
7675 /* if previous state reached the exit with precision and
7676 * current state is equivalent to it (except precsion marks)
7677 * the precision needs to be propagated back in
7678 * the current state.
7680 err = err ? : push_jmp_history(env, cur);
7681 err = err ? : propagate_precision(env, &sl->state);
7687 /* when new state is not going to be added do not increase miss count.
7688 * Otherwise several loop iterations will remove the state
7689 * recorded earlier. The goal of these heuristics is to have
7690 * states from some iterations of the loop (some in the beginning
7691 * and some at the end) to help pruning.
7695 /* heuristic to determine whether this state is beneficial
7696 * to keep checking from state equivalence point of view.
7697 * Higher numbers increase max_states_per_insn and verification time,
7698 * but do not meaningfully decrease insn_processed.
7700 if (sl->miss_cnt > sl->hit_cnt * 3 + 3) {
7701 /* the state is unlikely to be useful. Remove it to
7702 * speed up verification
7705 if (sl->state.frame[0]->regs[0].live & REG_LIVE_DONE) {
7706 u32 br = sl->state.branches;
7709 "BUG live_done but branches_to_explore %d\n",
7711 free_verifier_state(&sl->state, false);
7715 /* cannot free this state, since parentage chain may
7716 * walk it later. Add it for free_list instead to
7717 * be freed at the end of verification
7719 sl->next = env->free_list;
7720 env->free_list = sl;
7730 if (env->max_states_per_insn < states_cnt)
7731 env->max_states_per_insn = states_cnt;
7733 if (!env->allow_ptr_leaks && states_cnt > BPF_COMPLEXITY_LIMIT_STATES)
7734 return push_jmp_history(env, cur);
7737 return push_jmp_history(env, cur);
7739 /* There were no equivalent states, remember the current one.
7740 * Technically the current state is not proven to be safe yet,
7741 * but it will either reach outer most bpf_exit (which means it's safe)
7742 * or it will be rejected. When there are no loops the verifier won't be
7743 * seeing this tuple (frame[0].callsite, frame[1].callsite, .. insn_idx)
7744 * again on the way to bpf_exit.
7745 * When looping the sl->state.branches will be > 0 and this state
7746 * will not be considered for equivalence until branches == 0.
7748 new_sl = kzalloc(sizeof(struct bpf_verifier_state_list), GFP_KERNEL);
7751 env->total_states++;
7753 env->prev_jmps_processed = env->jmps_processed;
7754 env->prev_insn_processed = env->insn_processed;
7756 /* add new state to the head of linked list */
7757 new = &new_sl->state;
7758 err = copy_verifier_state(new, cur);
7760 free_verifier_state(new, false);
7764 new->insn_idx = insn_idx;
7765 WARN_ONCE(new->branches != 1,
7766 "BUG is_state_visited:branches_to_explore=%d insn %d\n", new->branches, insn_idx);
7769 cur->first_insn_idx = insn_idx;
7770 clear_jmp_history(cur);
7771 new_sl->next = *explored_state(env, insn_idx);
7772 *explored_state(env, insn_idx) = new_sl;
7773 /* connect new state to parentage chain. Current frame needs all
7774 * registers connected. Only r6 - r9 of the callers are alive (pushed
7775 * to the stack implicitly by JITs) so in callers' frames connect just
7776 * r6 - r9 as an optimization. Callers will have r1 - r5 connected to
7777 * the state of the call instruction (with WRITTEN set), and r0 comes
7778 * from callee with its full parentage chain, anyway.
7780 /* clear write marks in current state: the writes we did are not writes
7781 * our child did, so they don't screen off its reads from us.
7782 * (There are no read marks in current state, because reads always mark
7783 * their parent and current state never has children yet. Only
7784 * explored_states can get read marks.)
7786 for (j = 0; j <= cur->curframe; j++) {
7787 for (i = j < cur->curframe ? BPF_REG_6 : 0; i < BPF_REG_FP; i++)
7788 cur->frame[j]->regs[i].parent = &new->frame[j]->regs[i];
7789 for (i = 0; i < BPF_REG_FP; i++)
7790 cur->frame[j]->regs[i].live = REG_LIVE_NONE;
7793 /* all stack frames are accessible from callee, clear them all */
7794 for (j = 0; j <= cur->curframe; j++) {
7795 struct bpf_func_state *frame = cur->frame[j];
7796 struct bpf_func_state *newframe = new->frame[j];
7798 for (i = 0; i < frame->allocated_stack / BPF_REG_SIZE; i++) {
7799 frame->stack[i].spilled_ptr.live = REG_LIVE_NONE;
7800 frame->stack[i].spilled_ptr.parent =
7801 &newframe->stack[i].spilled_ptr;
7807 /* Return true if it's OK to have the same insn return a different type. */
7808 static bool reg_type_mismatch_ok(enum bpf_reg_type type)
7813 case PTR_TO_SOCKET_OR_NULL:
7814 case PTR_TO_SOCK_COMMON:
7815 case PTR_TO_SOCK_COMMON_OR_NULL:
7816 case PTR_TO_TCP_SOCK:
7817 case PTR_TO_TCP_SOCK_OR_NULL:
7818 case PTR_TO_XDP_SOCK:
7826 /* If an instruction was previously used with particular pointer types, then we
7827 * need to be careful to avoid cases such as the below, where it may be ok
7828 * for one branch accessing the pointer, but not ok for the other branch:
7833 * R1 = some_other_valid_ptr;
7836 * R2 = *(u32 *)(R1 + 0);
7838 static bool reg_type_mismatch(enum bpf_reg_type src, enum bpf_reg_type prev)
7840 return src != prev && (!reg_type_mismatch_ok(src) ||
7841 !reg_type_mismatch_ok(prev));
7844 static int do_check(struct bpf_verifier_env *env)
7846 struct bpf_verifier_state *state = env->cur_state;
7847 struct bpf_insn *insns = env->prog->insnsi;
7848 struct bpf_reg_state *regs;
7849 int insn_cnt = env->prog->len;
7850 bool do_print_state = false;
7851 int prev_insn_idx = -1;
7854 struct bpf_insn *insn;
7858 env->prev_insn_idx = prev_insn_idx;
7859 if (env->insn_idx >= insn_cnt) {
7860 verbose(env, "invalid insn idx %d insn_cnt %d\n",
7861 env->insn_idx, insn_cnt);
7865 insn = &insns[env->insn_idx];
7866 class = BPF_CLASS(insn->code);
7868 if (++env->insn_processed > BPF_COMPLEXITY_LIMIT_INSNS) {
7870 "BPF program is too large. Processed %d insn\n",
7871 env->insn_processed);
7875 err = is_state_visited(env, env->insn_idx);
7879 /* found equivalent state, can prune the search */
7880 if (env->log.level & BPF_LOG_LEVEL) {
7882 verbose(env, "\nfrom %d to %d%s: safe\n",
7883 env->prev_insn_idx, env->insn_idx,
7884 env->cur_state->speculative ?
7885 " (speculative execution)" : "");
7887 verbose(env, "%d: safe\n", env->insn_idx);
7889 goto process_bpf_exit;
7892 if (signal_pending(current))
7898 if (env->log.level & BPF_LOG_LEVEL2 ||
7899 (env->log.level & BPF_LOG_LEVEL && do_print_state)) {
7900 if (env->log.level & BPF_LOG_LEVEL2)
7901 verbose(env, "%d:", env->insn_idx);
7903 verbose(env, "\nfrom %d to %d%s:",
7904 env->prev_insn_idx, env->insn_idx,
7905 env->cur_state->speculative ?
7906 " (speculative execution)" : "");
7907 print_verifier_state(env, state->frame[state->curframe]);
7908 do_print_state = false;
7911 if (env->log.level & BPF_LOG_LEVEL) {
7912 const struct bpf_insn_cbs cbs = {
7913 .cb_print = verbose,
7914 .private_data = env,
7917 verbose_linfo(env, env->insn_idx, "; ");
7918 verbose(env, "%d: ", env->insn_idx);
7919 print_bpf_insn(&cbs, insn, env->allow_ptr_leaks);
7922 if (bpf_prog_is_dev_bound(env->prog->aux)) {
7923 err = bpf_prog_offload_verify_insn(env, env->insn_idx,
7924 env->prev_insn_idx);
7929 regs = cur_regs(env);
7930 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
7931 prev_insn_idx = env->insn_idx;
7933 if (class == BPF_ALU || class == BPF_ALU64) {
7934 err = check_alu_op(env, insn);
7938 } else if (class == BPF_LDX) {
7939 enum bpf_reg_type *prev_src_type, src_reg_type;
7941 /* check for reserved fields is already done */
7943 /* check src operand */
7944 err = check_reg_arg(env, insn->src_reg, SRC_OP);
7948 err = check_reg_arg(env, insn->dst_reg, DST_OP_NO_MARK);
7952 src_reg_type = regs[insn->src_reg].type;
7954 /* check that memory (src_reg + off) is readable,
7955 * the state of dst_reg will be updated by this func
7957 err = check_mem_access(env, env->insn_idx, insn->src_reg,
7958 insn->off, BPF_SIZE(insn->code),
7959 BPF_READ, insn->dst_reg, false);
7963 prev_src_type = &env->insn_aux_data[env->insn_idx].ptr_type;
7965 if (*prev_src_type == NOT_INIT) {
7967 * dst_reg = *(u32 *)(src_reg + off)
7968 * save type to validate intersecting paths
7970 *prev_src_type = src_reg_type;
7972 } else if (reg_type_mismatch(src_reg_type, *prev_src_type)) {
7973 /* ABuser program is trying to use the same insn
7974 * dst_reg = *(u32*) (src_reg + off)
7975 * with different pointer types:
7976 * src_reg == ctx in one branch and
7977 * src_reg == stack|map in some other branch.
7980 verbose(env, "same insn cannot be used with different pointers\n");
7984 } else if (class == BPF_STX) {
7985 enum bpf_reg_type *prev_dst_type, dst_reg_type;
7987 if (BPF_MODE(insn->code) == BPF_XADD) {
7988 err = check_xadd(env, env->insn_idx, insn);
7995 /* check src1 operand */
7996 err = check_reg_arg(env, insn->src_reg, SRC_OP);
7999 /* check src2 operand */
8000 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
8004 dst_reg_type = regs[insn->dst_reg].type;
8006 /* check that memory (dst_reg + off) is writeable */
8007 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
8008 insn->off, BPF_SIZE(insn->code),
8009 BPF_WRITE, insn->src_reg, false);
8013 prev_dst_type = &env->insn_aux_data[env->insn_idx].ptr_type;
8015 if (*prev_dst_type == NOT_INIT) {
8016 *prev_dst_type = dst_reg_type;
8017 } else if (reg_type_mismatch(dst_reg_type, *prev_dst_type)) {
8018 verbose(env, "same insn cannot be used with different pointers\n");
8022 } else if (class == BPF_ST) {
8023 if (BPF_MODE(insn->code) != BPF_MEM ||
8024 insn->src_reg != BPF_REG_0) {
8025 verbose(env, "BPF_ST uses reserved fields\n");
8028 /* check src operand */
8029 err = check_reg_arg(env, insn->dst_reg, SRC_OP);
8033 if (is_ctx_reg(env, insn->dst_reg)) {
8034 verbose(env, "BPF_ST stores into R%d %s is not allowed\n",
8036 reg_type_str[reg_state(env, insn->dst_reg)->type]);
8040 /* check that memory (dst_reg + off) is writeable */
8041 err = check_mem_access(env, env->insn_idx, insn->dst_reg,
8042 insn->off, BPF_SIZE(insn->code),
8043 BPF_WRITE, -1, false);
8047 } else if (class == BPF_JMP || class == BPF_JMP32) {
8048 u8 opcode = BPF_OP(insn->code);
8050 env->jmps_processed++;
8051 if (opcode == BPF_CALL) {
8052 if (BPF_SRC(insn->code) != BPF_K ||
8054 (insn->src_reg != BPF_REG_0 &&
8055 insn->src_reg != BPF_PSEUDO_CALL) ||
8056 insn->dst_reg != BPF_REG_0 ||
8057 class == BPF_JMP32) {
8058 verbose(env, "BPF_CALL uses reserved fields\n");
8062 if (env->cur_state->active_spin_lock &&
8063 (insn->src_reg == BPF_PSEUDO_CALL ||
8064 insn->imm != BPF_FUNC_spin_unlock)) {
8065 verbose(env, "function calls are not allowed while holding a lock\n");
8068 if (insn->src_reg == BPF_PSEUDO_CALL)
8069 err = check_func_call(env, insn, &env->insn_idx);
8071 err = check_helper_call(env, insn->imm, env->insn_idx);
8075 } else if (opcode == BPF_JA) {
8076 if (BPF_SRC(insn->code) != BPF_K ||
8078 insn->src_reg != BPF_REG_0 ||
8079 insn->dst_reg != BPF_REG_0 ||
8080 class == BPF_JMP32) {
8081 verbose(env, "BPF_JA uses reserved fields\n");
8085 env->insn_idx += insn->off + 1;
8088 } else if (opcode == BPF_EXIT) {
8089 if (BPF_SRC(insn->code) != BPF_K ||
8091 insn->src_reg != BPF_REG_0 ||
8092 insn->dst_reg != BPF_REG_0 ||
8093 class == BPF_JMP32) {
8094 verbose(env, "BPF_EXIT uses reserved fields\n");
8098 if (env->cur_state->active_spin_lock) {
8099 verbose(env, "bpf_spin_unlock is missing\n");
8103 if (state->curframe) {
8104 /* exit from nested function */
8105 err = prepare_func_exit(env, &env->insn_idx);
8108 do_print_state = true;
8112 err = check_reference_leak(env);
8116 err = check_return_code(env);
8120 update_branch_counts(env, env->cur_state);
8121 err = pop_stack(env, &prev_insn_idx,
8128 do_print_state = true;
8132 err = check_cond_jmp_op(env, insn, &env->insn_idx);
8136 } else if (class == BPF_LD) {
8137 u8 mode = BPF_MODE(insn->code);
8139 if (mode == BPF_ABS || mode == BPF_IND) {
8140 err = check_ld_abs(env, insn);
8144 } else if (mode == BPF_IMM) {
8145 err = check_ld_imm(env, insn);
8150 env->insn_aux_data[env->insn_idx].seen = env->pass_cnt;
8152 verbose(env, "invalid BPF_LD mode\n");
8156 verbose(env, "unknown insn class %d\n", class);
8166 static int check_map_prealloc(struct bpf_map *map)
8168 return (map->map_type != BPF_MAP_TYPE_HASH &&
8169 map->map_type != BPF_MAP_TYPE_PERCPU_HASH &&
8170 map->map_type != BPF_MAP_TYPE_HASH_OF_MAPS) ||
8171 !(map->map_flags & BPF_F_NO_PREALLOC);
8174 static bool is_tracing_prog_type(enum bpf_prog_type type)
8177 case BPF_PROG_TYPE_KPROBE:
8178 case BPF_PROG_TYPE_TRACEPOINT:
8179 case BPF_PROG_TYPE_PERF_EVENT:
8180 case BPF_PROG_TYPE_RAW_TRACEPOINT:
8187 static bool is_preallocated_map(struct bpf_map *map)
8189 if (!check_map_prealloc(map))
8191 if (map->inner_map_meta && !check_map_prealloc(map->inner_map_meta))
8196 static int check_map_prog_compatibility(struct bpf_verifier_env *env,
8197 struct bpf_map *map,
8198 struct bpf_prog *prog)
8202 * Validate that trace type programs use preallocated hash maps.
8204 * For programs attached to PERF events this is mandatory as the
8205 * perf NMI can hit any arbitrary code sequence.
8207 * All other trace types using preallocated hash maps are unsafe as
8208 * well because tracepoint or kprobes can be inside locked regions
8209 * of the memory allocator or at a place where a recursion into the
8210 * memory allocator would see inconsistent state.
8212 * On RT enabled kernels run-time allocation of all trace type
8213 * programs is strictly prohibited due to lock type constraints. On
8214 * !RT kernels it is allowed for backwards compatibility reasons for
8215 * now, but warnings are emitted so developers are made aware of
8216 * the unsafety and can fix their programs before this is enforced.
8218 if (is_tracing_prog_type(prog->type) && !is_preallocated_map(map)) {
8219 if (prog->type == BPF_PROG_TYPE_PERF_EVENT) {
8220 verbose(env, "perf_event programs can only use preallocated hash map\n");
8223 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
8224 verbose(env, "trace type programs can only use preallocated hash map\n");
8227 WARN_ONCE(1, "trace type BPF program uses run-time allocation\n");
8228 verbose(env, "trace type programs with run-time allocated hash maps are unsafe. Switch to preallocated hash maps.\n");
8231 if ((is_tracing_prog_type(prog->type) ||
8232 prog->type == BPF_PROG_TYPE_SOCKET_FILTER) &&
8233 map_value_has_spin_lock(map)) {
8234 verbose(env, "tracing progs cannot use bpf_spin_lock yet\n");
8238 if ((bpf_prog_is_dev_bound(prog->aux) || bpf_map_is_dev_bound(map)) &&
8239 !bpf_offload_prog_map_match(prog, map)) {
8240 verbose(env, "offload device mismatch between prog and map\n");
8244 if (map->map_type == BPF_MAP_TYPE_STRUCT_OPS) {
8245 verbose(env, "bpf_struct_ops map cannot be used in prog\n");
8252 static bool bpf_map_is_cgroup_storage(struct bpf_map *map)
8254 return (map->map_type == BPF_MAP_TYPE_CGROUP_STORAGE ||
8255 map->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE);
8258 /* look for pseudo eBPF instructions that access map FDs and
8259 * replace them with actual map pointers
8261 static int replace_map_fd_with_map_ptr(struct bpf_verifier_env *env)
8263 struct bpf_insn *insn = env->prog->insnsi;
8264 int insn_cnt = env->prog->len;
8267 err = bpf_prog_calc_tag(env->prog);
8271 for (i = 0; i < insn_cnt; i++, insn++) {
8272 if (BPF_CLASS(insn->code) == BPF_LDX &&
8273 (BPF_MODE(insn->code) != BPF_MEM || insn->imm != 0)) {
8274 verbose(env, "BPF_LDX uses reserved fields\n");
8278 if (BPF_CLASS(insn->code) == BPF_STX &&
8279 ((BPF_MODE(insn->code) != BPF_MEM &&
8280 BPF_MODE(insn->code) != BPF_XADD) || insn->imm != 0)) {
8281 verbose(env, "BPF_STX uses reserved fields\n");
8285 if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW)) {
8286 struct bpf_insn_aux_data *aux;
8287 struct bpf_map *map;
8291 if (i == insn_cnt - 1 || insn[1].code != 0 ||
8292 insn[1].dst_reg != 0 || insn[1].src_reg != 0 ||
8294 verbose(env, "invalid bpf_ld_imm64 insn\n");
8298 if (insn[0].src_reg == 0)
8299 /* valid generic load 64-bit imm */
8302 /* In final convert_pseudo_ld_imm64() step, this is
8303 * converted into regular 64-bit imm load insn.
8305 if ((insn[0].src_reg != BPF_PSEUDO_MAP_FD &&
8306 insn[0].src_reg != BPF_PSEUDO_MAP_VALUE) ||
8307 (insn[0].src_reg == BPF_PSEUDO_MAP_FD &&
8308 insn[1].imm != 0)) {
8310 "unrecognized bpf_ld_imm64 insn\n");
8314 f = fdget(insn[0].imm);
8315 map = __bpf_map_get(f);
8317 verbose(env, "fd %d is not pointing to valid bpf_map\n",
8319 return PTR_ERR(map);
8322 err = check_map_prog_compatibility(env, map, env->prog);
8328 aux = &env->insn_aux_data[i];
8329 if (insn->src_reg == BPF_PSEUDO_MAP_FD) {
8330 addr = (unsigned long)map;
8332 u32 off = insn[1].imm;
8334 if (off >= BPF_MAX_VAR_OFF) {
8335 verbose(env, "direct value offset of %u is not allowed\n", off);
8340 if (!map->ops->map_direct_value_addr) {
8341 verbose(env, "no direct value access support for this map type\n");
8346 err = map->ops->map_direct_value_addr(map, &addr, off);
8348 verbose(env, "invalid access to map value pointer, value_size=%u off=%u\n",
8349 map->value_size, off);
8358 insn[0].imm = (u32)addr;
8359 insn[1].imm = addr >> 32;
8361 /* check whether we recorded this map already */
8362 for (j = 0; j < env->used_map_cnt; j++) {
8363 if (env->used_maps[j] == map) {
8370 if (env->used_map_cnt >= MAX_USED_MAPS) {
8375 /* hold the map. If the program is rejected by verifier,
8376 * the map will be released by release_maps() or it
8377 * will be used by the valid program until it's unloaded
8378 * and all maps are released in free_used_maps()
8382 aux->map_index = env->used_map_cnt;
8383 env->used_maps[env->used_map_cnt++] = map;
8385 if (bpf_map_is_cgroup_storage(map) &&
8386 bpf_cgroup_storage_assign(env->prog->aux, map)) {
8387 verbose(env, "only one cgroup storage of each type is allowed\n");
8399 /* Basic sanity check before we invest more work here. */
8400 if (!bpf_opcode_in_insntable(insn->code)) {
8401 verbose(env, "unknown opcode %02x\n", insn->code);
8406 /* now all pseudo BPF_LD_IMM64 instructions load valid
8407 * 'struct bpf_map *' into a register instead of user map_fd.
8408 * These pointers will be used later by verifier to validate map access.
8413 /* drop refcnt of maps used by the rejected program */
8414 static void release_maps(struct bpf_verifier_env *env)
8416 __bpf_free_used_maps(env->prog->aux, env->used_maps,
8420 /* convert pseudo BPF_LD_IMM64 into generic BPF_LD_IMM64 */
8421 static void convert_pseudo_ld_imm64(struct bpf_verifier_env *env)
8423 struct bpf_insn *insn = env->prog->insnsi;
8424 int insn_cnt = env->prog->len;
8427 for (i = 0; i < insn_cnt; i++, insn++)
8428 if (insn->code == (BPF_LD | BPF_IMM | BPF_DW))
8432 /* single env->prog->insni[off] instruction was replaced with the range
8433 * insni[off, off + cnt). Adjust corresponding insn_aux_data by copying
8434 * [0, off) and [off, end) to new locations, so the patched range stays zero
8436 static int adjust_insn_aux_data(struct bpf_verifier_env *env,
8437 struct bpf_prog *new_prog, u32 off, u32 cnt)
8439 struct bpf_insn_aux_data *new_data, *old_data = env->insn_aux_data;
8440 struct bpf_insn *insn = new_prog->insnsi;
8444 /* aux info at OFF always needs adjustment, no matter fast path
8445 * (cnt == 1) is taken or not. There is no guarantee INSN at OFF is the
8446 * original insn at old prog.
8448 old_data[off].zext_dst = insn_has_def32(env, insn + off + cnt - 1);
8452 prog_len = new_prog->len;
8453 new_data = vzalloc(array_size(prog_len,
8454 sizeof(struct bpf_insn_aux_data)));
8457 memcpy(new_data, old_data, sizeof(struct bpf_insn_aux_data) * off);
8458 memcpy(new_data + off + cnt - 1, old_data + off,
8459 sizeof(struct bpf_insn_aux_data) * (prog_len - off - cnt + 1));
8460 for (i = off; i < off + cnt - 1; i++) {
8461 new_data[i].seen = env->pass_cnt;
8462 new_data[i].zext_dst = insn_has_def32(env, insn + i);
8464 env->insn_aux_data = new_data;
8469 static void adjust_subprog_starts(struct bpf_verifier_env *env, u32 off, u32 len)
8475 /* NOTE: fake 'exit' subprog should be updated as well. */
8476 for (i = 0; i <= env->subprog_cnt; i++) {
8477 if (env->subprog_info[i].start <= off)
8479 env->subprog_info[i].start += len - 1;
8483 static struct bpf_prog *bpf_patch_insn_data(struct bpf_verifier_env *env, u32 off,
8484 const struct bpf_insn *patch, u32 len)
8486 struct bpf_prog *new_prog;
8488 new_prog = bpf_patch_insn_single(env->prog, off, patch, len);
8489 if (IS_ERR(new_prog)) {
8490 if (PTR_ERR(new_prog) == -ERANGE)
8492 "insn %d cannot be patched due to 16-bit range\n",
8493 env->insn_aux_data[off].orig_idx);
8496 if (adjust_insn_aux_data(env, new_prog, off, len))
8498 adjust_subprog_starts(env, off, len);
8502 static int adjust_subprog_starts_after_remove(struct bpf_verifier_env *env,
8507 /* find first prog starting at or after off (first to remove) */
8508 for (i = 0; i < env->subprog_cnt; i++)
8509 if (env->subprog_info[i].start >= off)
8511 /* find first prog starting at or after off + cnt (first to stay) */
8512 for (j = i; j < env->subprog_cnt; j++)
8513 if (env->subprog_info[j].start >= off + cnt)
8515 /* if j doesn't start exactly at off + cnt, we are just removing
8516 * the front of previous prog
8518 if (env->subprog_info[j].start != off + cnt)
8522 struct bpf_prog_aux *aux = env->prog->aux;
8525 /* move fake 'exit' subprog as well */
8526 move = env->subprog_cnt + 1 - j;
8528 memmove(env->subprog_info + i,
8529 env->subprog_info + j,
8530 sizeof(*env->subprog_info) * move);
8531 env->subprog_cnt -= j - i;
8533 /* remove func_info */
8534 if (aux->func_info) {
8535 move = aux->func_info_cnt - j;
8537 memmove(aux->func_info + i,
8539 sizeof(*aux->func_info) * move);
8540 aux->func_info_cnt -= j - i;
8541 /* func_info->insn_off is set after all code rewrites,
8542 * in adjust_btf_func() - no need to adjust
8546 /* convert i from "first prog to remove" to "first to adjust" */
8547 if (env->subprog_info[i].start == off)
8551 /* update fake 'exit' subprog as well */
8552 for (; i <= env->subprog_cnt; i++)
8553 env->subprog_info[i].start -= cnt;
8558 static int bpf_adj_linfo_after_remove(struct bpf_verifier_env *env, u32 off,
8561 struct bpf_prog *prog = env->prog;
8562 u32 i, l_off, l_cnt, nr_linfo;
8563 struct bpf_line_info *linfo;
8565 nr_linfo = prog->aux->nr_linfo;
8569 linfo = prog->aux->linfo;
8571 /* find first line info to remove, count lines to be removed */
8572 for (i = 0; i < nr_linfo; i++)
8573 if (linfo[i].insn_off >= off)
8578 for (; i < nr_linfo; i++)
8579 if (linfo[i].insn_off < off + cnt)
8584 /* First live insn doesn't match first live linfo, it needs to "inherit"
8585 * last removed linfo. prog is already modified, so prog->len == off
8586 * means no live instructions after (tail of the program was removed).
8588 if (prog->len != off && l_cnt &&
8589 (i == nr_linfo || linfo[i].insn_off != off + cnt)) {
8591 linfo[--i].insn_off = off + cnt;
8594 /* remove the line info which refer to the removed instructions */
8596 memmove(linfo + l_off, linfo + i,
8597 sizeof(*linfo) * (nr_linfo - i));
8599 prog->aux->nr_linfo -= l_cnt;
8600 nr_linfo = prog->aux->nr_linfo;
8603 /* pull all linfo[i].insn_off >= off + cnt in by cnt */
8604 for (i = l_off; i < nr_linfo; i++)
8605 linfo[i].insn_off -= cnt;
8607 /* fix up all subprogs (incl. 'exit') which start >= off */
8608 for (i = 0; i <= env->subprog_cnt; i++)
8609 if (env->subprog_info[i].linfo_idx > l_off) {
8610 /* program may have started in the removed region but
8611 * may not be fully removed
8613 if (env->subprog_info[i].linfo_idx >= l_off + l_cnt)
8614 env->subprog_info[i].linfo_idx -= l_cnt;
8616 env->subprog_info[i].linfo_idx = l_off;
8622 static int verifier_remove_insns(struct bpf_verifier_env *env, u32 off, u32 cnt)
8624 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
8625 unsigned int orig_prog_len = env->prog->len;
8628 if (bpf_prog_is_dev_bound(env->prog->aux))
8629 bpf_prog_offload_remove_insns(env, off, cnt);
8631 err = bpf_remove_insns(env->prog, off, cnt);
8635 err = adjust_subprog_starts_after_remove(env, off, cnt);
8639 err = bpf_adj_linfo_after_remove(env, off, cnt);
8643 memmove(aux_data + off, aux_data + off + cnt,
8644 sizeof(*aux_data) * (orig_prog_len - off - cnt));
8649 /* The verifier does more data flow analysis than llvm and will not
8650 * explore branches that are dead at run time. Malicious programs can
8651 * have dead code too. Therefore replace all dead at-run-time code
8654 * Just nops are not optimal, e.g. if they would sit at the end of the
8655 * program and through another bug we would manage to jump there, then
8656 * we'd execute beyond program memory otherwise. Returning exception
8657 * code also wouldn't work since we can have subprogs where the dead
8658 * code could be located.
8660 static void sanitize_dead_code(struct bpf_verifier_env *env)
8662 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
8663 struct bpf_insn trap = BPF_JMP_IMM(BPF_JA, 0, 0, -1);
8664 struct bpf_insn *insn = env->prog->insnsi;
8665 const int insn_cnt = env->prog->len;
8668 for (i = 0; i < insn_cnt; i++) {
8669 if (aux_data[i].seen)
8671 memcpy(insn + i, &trap, sizeof(trap));
8675 static bool insn_is_cond_jump(u8 code)
8679 if (BPF_CLASS(code) == BPF_JMP32)
8682 if (BPF_CLASS(code) != BPF_JMP)
8686 return op != BPF_JA && op != BPF_EXIT && op != BPF_CALL;
8689 static void opt_hard_wire_dead_code_branches(struct bpf_verifier_env *env)
8691 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
8692 struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
8693 struct bpf_insn *insn = env->prog->insnsi;
8694 const int insn_cnt = env->prog->len;
8697 for (i = 0; i < insn_cnt; i++, insn++) {
8698 if (!insn_is_cond_jump(insn->code))
8701 if (!aux_data[i + 1].seen)
8703 else if (!aux_data[i + 1 + insn->off].seen)
8708 if (bpf_prog_is_dev_bound(env->prog->aux))
8709 bpf_prog_offload_replace_insn(env, i, &ja);
8711 memcpy(insn, &ja, sizeof(ja));
8715 static int opt_remove_dead_code(struct bpf_verifier_env *env)
8717 struct bpf_insn_aux_data *aux_data = env->insn_aux_data;
8718 int insn_cnt = env->prog->len;
8721 for (i = 0; i < insn_cnt; i++) {
8725 while (i + j < insn_cnt && !aux_data[i + j].seen)
8730 err = verifier_remove_insns(env, i, j);
8733 insn_cnt = env->prog->len;
8739 static int opt_remove_nops(struct bpf_verifier_env *env)
8741 const struct bpf_insn ja = BPF_JMP_IMM(BPF_JA, 0, 0, 0);
8742 struct bpf_insn *insn = env->prog->insnsi;
8743 int insn_cnt = env->prog->len;
8746 for (i = 0; i < insn_cnt; i++) {
8747 if (memcmp(&insn[i], &ja, sizeof(ja)))
8750 err = verifier_remove_insns(env, i, 1);
8760 static int opt_subreg_zext_lo32_rnd_hi32(struct bpf_verifier_env *env,
8761 const union bpf_attr *attr)
8763 struct bpf_insn *patch, zext_patch[2], rnd_hi32_patch[4];
8764 struct bpf_insn_aux_data *aux = env->insn_aux_data;
8765 int i, patch_len, delta = 0, len = env->prog->len;
8766 struct bpf_insn *insns = env->prog->insnsi;
8767 struct bpf_prog *new_prog;
8770 rnd_hi32 = attr->prog_flags & BPF_F_TEST_RND_HI32;
8771 zext_patch[1] = BPF_ZEXT_REG(0);
8772 rnd_hi32_patch[1] = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, 0);
8773 rnd_hi32_patch[2] = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
8774 rnd_hi32_patch[3] = BPF_ALU64_REG(BPF_OR, 0, BPF_REG_AX);
8775 for (i = 0; i < len; i++) {
8776 int adj_idx = i + delta;
8777 struct bpf_insn insn;
8779 insn = insns[adj_idx];
8780 if (!aux[adj_idx].zext_dst) {
8788 class = BPF_CLASS(code);
8789 if (insn_no_def(&insn))
8792 /* NOTE: arg "reg" (the fourth one) is only used for
8793 * BPF_STX which has been ruled out in above
8794 * check, it is safe to pass NULL here.
8796 if (is_reg64(env, &insn, insn.dst_reg, NULL, DST_OP)) {
8797 if (class == BPF_LD &&
8798 BPF_MODE(code) == BPF_IMM)
8803 /* ctx load could be transformed into wider load. */
8804 if (class == BPF_LDX &&
8805 aux[adj_idx].ptr_type == PTR_TO_CTX)
8808 imm_rnd = get_random_int();
8809 rnd_hi32_patch[0] = insn;
8810 rnd_hi32_patch[1].imm = imm_rnd;
8811 rnd_hi32_patch[3].dst_reg = insn.dst_reg;
8812 patch = rnd_hi32_patch;
8814 goto apply_patch_buffer;
8817 if (!bpf_jit_needs_zext())
8820 zext_patch[0] = insn;
8821 zext_patch[1].dst_reg = insn.dst_reg;
8822 zext_patch[1].src_reg = insn.dst_reg;
8826 new_prog = bpf_patch_insn_data(env, adj_idx, patch, patch_len);
8829 env->prog = new_prog;
8830 insns = new_prog->insnsi;
8831 aux = env->insn_aux_data;
8832 delta += patch_len - 1;
8838 /* convert load instructions that access fields of a context type into a
8839 * sequence of instructions that access fields of the underlying structure:
8840 * struct __sk_buff -> struct sk_buff
8841 * struct bpf_sock_ops -> struct sock
8843 static int convert_ctx_accesses(struct bpf_verifier_env *env)
8845 const struct bpf_verifier_ops *ops = env->ops;
8846 int i, cnt, size, ctx_field_size, delta = 0;
8847 const int insn_cnt = env->prog->len;
8848 struct bpf_insn insn_buf[16], *insn;
8849 u32 target_size, size_default, off;
8850 struct bpf_prog *new_prog;
8851 enum bpf_access_type type;
8852 bool is_narrower_load;
8854 if (ops->gen_prologue || env->seen_direct_write) {
8855 if (!ops->gen_prologue) {
8856 verbose(env, "bpf verifier is misconfigured\n");
8859 cnt = ops->gen_prologue(insn_buf, env->seen_direct_write,
8861 if (cnt >= ARRAY_SIZE(insn_buf)) {
8862 verbose(env, "bpf verifier is misconfigured\n");
8865 new_prog = bpf_patch_insn_data(env, 0, insn_buf, cnt);
8869 env->prog = new_prog;
8874 if (bpf_prog_is_dev_bound(env->prog->aux))
8877 insn = env->prog->insnsi + delta;
8879 for (i = 0; i < insn_cnt; i++, insn++) {
8880 bpf_convert_ctx_access_t convert_ctx_access;
8882 if (insn->code == (BPF_LDX | BPF_MEM | BPF_B) ||
8883 insn->code == (BPF_LDX | BPF_MEM | BPF_H) ||
8884 insn->code == (BPF_LDX | BPF_MEM | BPF_W) ||
8885 insn->code == (BPF_LDX | BPF_MEM | BPF_DW))
8887 else if (insn->code == (BPF_STX | BPF_MEM | BPF_B) ||
8888 insn->code == (BPF_STX | BPF_MEM | BPF_H) ||
8889 insn->code == (BPF_STX | BPF_MEM | BPF_W) ||
8890 insn->code == (BPF_STX | BPF_MEM | BPF_DW))
8895 if (type == BPF_WRITE &&
8896 env->insn_aux_data[i + delta].sanitize_stack_off) {
8897 struct bpf_insn patch[] = {
8898 /* Sanitize suspicious stack slot with zero.
8899 * There are no memory dependencies for this store,
8900 * since it's only using frame pointer and immediate
8903 BPF_ST_MEM(BPF_DW, BPF_REG_FP,
8904 env->insn_aux_data[i + delta].sanitize_stack_off,
8906 /* the original STX instruction will immediately
8907 * overwrite the same stack slot with appropriate value
8912 cnt = ARRAY_SIZE(patch);
8913 new_prog = bpf_patch_insn_data(env, i + delta, patch, cnt);
8918 env->prog = new_prog;
8919 insn = new_prog->insnsi + i + delta;
8923 switch (env->insn_aux_data[i + delta].ptr_type) {
8925 if (!ops->convert_ctx_access)
8927 convert_ctx_access = ops->convert_ctx_access;
8930 case PTR_TO_SOCK_COMMON:
8931 convert_ctx_access = bpf_sock_convert_ctx_access;
8933 case PTR_TO_TCP_SOCK:
8934 convert_ctx_access = bpf_tcp_sock_convert_ctx_access;
8936 case PTR_TO_XDP_SOCK:
8937 convert_ctx_access = bpf_xdp_sock_convert_ctx_access;
8940 if (type == BPF_READ) {
8941 insn->code = BPF_LDX | BPF_PROBE_MEM |
8942 BPF_SIZE((insn)->code);
8943 env->prog->aux->num_exentries++;
8944 } else if (env->prog->type != BPF_PROG_TYPE_STRUCT_OPS) {
8945 verbose(env, "Writes through BTF pointers are not allowed\n");
8953 ctx_field_size = env->insn_aux_data[i + delta].ctx_field_size;
8954 size = BPF_LDST_BYTES(insn);
8956 /* If the read access is a narrower load of the field,
8957 * convert to a 4/8-byte load, to minimum program type specific
8958 * convert_ctx_access changes. If conversion is successful,
8959 * we will apply proper mask to the result.
8961 is_narrower_load = size < ctx_field_size;
8962 size_default = bpf_ctx_off_adjust_machine(ctx_field_size);
8964 if (is_narrower_load) {
8967 if (type == BPF_WRITE) {
8968 verbose(env, "bpf verifier narrow ctx access misconfigured\n");
8973 if (ctx_field_size == 4)
8975 else if (ctx_field_size == 8)
8978 insn->off = off & ~(size_default - 1);
8979 insn->code = BPF_LDX | BPF_MEM | size_code;
8983 cnt = convert_ctx_access(type, insn, insn_buf, env->prog,
8985 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf) ||
8986 (ctx_field_size && !target_size)) {
8987 verbose(env, "bpf verifier is misconfigured\n");
8991 if (is_narrower_load && size < target_size) {
8992 u8 shift = bpf_ctx_narrow_access_offset(
8993 off, size, size_default) * 8;
8994 if (ctx_field_size <= 4) {
8996 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_RSH,
8999 insn_buf[cnt++] = BPF_ALU32_IMM(BPF_AND, insn->dst_reg,
9000 (1 << size * 8) - 1);
9003 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_RSH,
9006 insn_buf[cnt++] = BPF_ALU64_IMM(BPF_AND, insn->dst_reg,
9007 (1ULL << size * 8) - 1);
9011 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
9017 /* keep walking new program and skip insns we just inserted */
9018 env->prog = new_prog;
9019 insn = new_prog->insnsi + i + delta;
9025 static int jit_subprogs(struct bpf_verifier_env *env)
9027 struct bpf_prog *prog = env->prog, **func, *tmp;
9028 int i, j, subprog_start, subprog_end = 0, len, subprog;
9029 struct bpf_insn *insn;
9033 if (env->subprog_cnt <= 1)
9036 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
9037 if (insn->code != (BPF_JMP | BPF_CALL) ||
9038 insn->src_reg != BPF_PSEUDO_CALL)
9040 /* Upon error here we cannot fall back to interpreter but
9041 * need a hard reject of the program. Thus -EFAULT is
9042 * propagated in any case.
9044 subprog = find_subprog(env, i + insn->imm + 1);
9046 WARN_ONCE(1, "verifier bug. No program starts at insn %d\n",
9050 /* temporarily remember subprog id inside insn instead of
9051 * aux_data, since next loop will split up all insns into funcs
9053 insn->off = subprog;
9054 /* remember original imm in case JIT fails and fallback
9055 * to interpreter will be needed
9057 env->insn_aux_data[i].call_imm = insn->imm;
9058 /* point imm to __bpf_call_base+1 from JITs point of view */
9062 err = bpf_prog_alloc_jited_linfo(prog);
9067 func = kcalloc(env->subprog_cnt, sizeof(prog), GFP_KERNEL);
9071 for (i = 0; i < env->subprog_cnt; i++) {
9072 subprog_start = subprog_end;
9073 subprog_end = env->subprog_info[i + 1].start;
9075 len = subprog_end - subprog_start;
9076 /* BPF_PROG_RUN doesn't call subprogs directly,
9077 * hence main prog stats include the runtime of subprogs.
9078 * subprogs don't have IDs and not reachable via prog_get_next_id
9079 * func[i]->aux->stats will never be accessed and stays NULL
9081 func[i] = bpf_prog_alloc_no_stats(bpf_prog_size(len), GFP_USER);
9084 memcpy(func[i]->insnsi, &prog->insnsi[subprog_start],
9085 len * sizeof(struct bpf_insn));
9086 func[i]->type = prog->type;
9088 if (bpf_prog_calc_tag(func[i]))
9090 func[i]->is_func = 1;
9091 func[i]->aux->func_idx = i;
9092 /* the btf and func_info will be freed only at prog->aux */
9093 func[i]->aux->btf = prog->aux->btf;
9094 func[i]->aux->func_info = prog->aux->func_info;
9096 /* Use bpf_prog_F_tag to indicate functions in stack traces.
9097 * Long term would need debug info to populate names
9099 func[i]->aux->name[0] = 'F';
9100 func[i]->aux->stack_depth = env->subprog_info[i].stack_depth;
9101 func[i]->jit_requested = 1;
9102 func[i]->aux->linfo = prog->aux->linfo;
9103 func[i]->aux->nr_linfo = prog->aux->nr_linfo;
9104 func[i]->aux->jited_linfo = prog->aux->jited_linfo;
9105 func[i]->aux->linfo_idx = env->subprog_info[i].linfo_idx;
9106 func[i] = bpf_int_jit_compile(func[i]);
9107 if (!func[i]->jited) {
9113 /* at this point all bpf functions were successfully JITed
9114 * now populate all bpf_calls with correct addresses and
9115 * run last pass of JIT
9117 for (i = 0; i < env->subprog_cnt; i++) {
9118 insn = func[i]->insnsi;
9119 for (j = 0; j < func[i]->len; j++, insn++) {
9120 if (insn->code != (BPF_JMP | BPF_CALL) ||
9121 insn->src_reg != BPF_PSEUDO_CALL)
9123 subprog = insn->off;
9124 insn->imm = BPF_CAST_CALL(func[subprog]->bpf_func) -
9128 /* we use the aux data to keep a list of the start addresses
9129 * of the JITed images for each function in the program
9131 * for some architectures, such as powerpc64, the imm field
9132 * might not be large enough to hold the offset of the start
9133 * address of the callee's JITed image from __bpf_call_base
9135 * in such cases, we can lookup the start address of a callee
9136 * by using its subprog id, available from the off field of
9137 * the call instruction, as an index for this list
9139 func[i]->aux->func = func;
9140 func[i]->aux->func_cnt = env->subprog_cnt;
9142 for (i = 0; i < env->subprog_cnt; i++) {
9143 old_bpf_func = func[i]->bpf_func;
9144 tmp = bpf_int_jit_compile(func[i]);
9145 if (tmp != func[i] || func[i]->bpf_func != old_bpf_func) {
9146 verbose(env, "JIT doesn't support bpf-to-bpf calls\n");
9153 /* finally lock prog and jit images for all functions and
9156 for (i = 0; i < env->subprog_cnt; i++) {
9157 bpf_prog_lock_ro(func[i]);
9158 bpf_prog_kallsyms_add(func[i]);
9161 /* Last step: make now unused interpreter insns from main
9162 * prog consistent for later dump requests, so they can
9163 * later look the same as if they were interpreted only.
9165 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
9166 if (insn->code != (BPF_JMP | BPF_CALL) ||
9167 insn->src_reg != BPF_PSEUDO_CALL)
9169 insn->off = env->insn_aux_data[i].call_imm;
9170 subprog = find_subprog(env, i + insn->off + 1);
9171 insn->imm = subprog;
9175 prog->bpf_func = func[0]->bpf_func;
9176 prog->aux->func = func;
9177 prog->aux->func_cnt = env->subprog_cnt;
9178 bpf_prog_free_unused_jited_linfo(prog);
9181 for (i = 0; i < env->subprog_cnt; i++)
9183 bpf_jit_free(func[i]);
9186 /* cleanup main prog to be interpreted */
9187 prog->jit_requested = 0;
9188 for (i = 0, insn = prog->insnsi; i < prog->len; i++, insn++) {
9189 if (insn->code != (BPF_JMP | BPF_CALL) ||
9190 insn->src_reg != BPF_PSEUDO_CALL)
9193 insn->imm = env->insn_aux_data[i].call_imm;
9195 bpf_prog_free_jited_linfo(prog);
9199 static int fixup_call_args(struct bpf_verifier_env *env)
9201 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
9202 struct bpf_prog *prog = env->prog;
9203 struct bpf_insn *insn = prog->insnsi;
9208 if (env->prog->jit_requested &&
9209 !bpf_prog_is_dev_bound(env->prog->aux)) {
9210 err = jit_subprogs(env);
9216 #ifndef CONFIG_BPF_JIT_ALWAYS_ON
9217 for (i = 0; i < prog->len; i++, insn++) {
9218 if (insn->code != (BPF_JMP | BPF_CALL) ||
9219 insn->src_reg != BPF_PSEUDO_CALL)
9221 depth = get_callee_stack_depth(env, insn, i);
9224 bpf_patch_call_args(insn, depth);
9231 /* fixup insn->imm field of bpf_call instructions
9232 * and inline eligible helpers as explicit sequence of BPF instructions
9234 * this function is called after eBPF program passed verification
9236 static int fixup_bpf_calls(struct bpf_verifier_env *env)
9238 struct bpf_prog *prog = env->prog;
9239 bool expect_blinding = bpf_jit_blinding_enabled(prog);
9240 struct bpf_insn *insn = prog->insnsi;
9241 const struct bpf_func_proto *fn;
9242 const int insn_cnt = prog->len;
9243 const struct bpf_map_ops *ops;
9244 struct bpf_insn_aux_data *aux;
9245 struct bpf_insn insn_buf[16];
9246 struct bpf_prog *new_prog;
9247 struct bpf_map *map_ptr;
9248 int i, ret, cnt, delta = 0;
9250 for (i = 0; i < insn_cnt; i++, insn++) {
9251 if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) ||
9252 insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
9253 insn->code == (BPF_ALU | BPF_MOD | BPF_X) ||
9254 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
9255 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
9256 struct bpf_insn mask_and_div[] = {
9257 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
9259 BPF_JMP_IMM(BPF_JNE, insn->src_reg, 0, 2),
9260 BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg),
9261 BPF_JMP_IMM(BPF_JA, 0, 0, 1),
9264 struct bpf_insn mask_and_mod[] = {
9265 BPF_MOV32_REG(insn->src_reg, insn->src_reg),
9266 /* Rx mod 0 -> Rx */
9267 BPF_JMP_IMM(BPF_JEQ, insn->src_reg, 0, 1),
9270 struct bpf_insn *patchlet;
9272 if (insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) ||
9273 insn->code == (BPF_ALU | BPF_DIV | BPF_X)) {
9274 patchlet = mask_and_div + (is64 ? 1 : 0);
9275 cnt = ARRAY_SIZE(mask_and_div) - (is64 ? 1 : 0);
9277 patchlet = mask_and_mod + (is64 ? 1 : 0);
9278 cnt = ARRAY_SIZE(mask_and_mod) - (is64 ? 1 : 0);
9281 new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);
9286 env->prog = prog = new_prog;
9287 insn = new_prog->insnsi + i + delta;
9291 if (BPF_CLASS(insn->code) == BPF_LD &&
9292 (BPF_MODE(insn->code) == BPF_ABS ||
9293 BPF_MODE(insn->code) == BPF_IND)) {
9294 cnt = env->ops->gen_ld_abs(insn, insn_buf);
9295 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
9296 verbose(env, "bpf verifier is misconfigured\n");
9300 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
9305 env->prog = prog = new_prog;
9306 insn = new_prog->insnsi + i + delta;
9310 if (insn->code == (BPF_ALU64 | BPF_ADD | BPF_X) ||
9311 insn->code == (BPF_ALU64 | BPF_SUB | BPF_X)) {
9312 const u8 code_add = BPF_ALU64 | BPF_ADD | BPF_X;
9313 const u8 code_sub = BPF_ALU64 | BPF_SUB | BPF_X;
9314 struct bpf_insn insn_buf[16];
9315 struct bpf_insn *patch = &insn_buf[0];
9319 aux = &env->insn_aux_data[i + delta];
9320 if (!aux->alu_state ||
9321 aux->alu_state == BPF_ALU_NON_POINTER)
9324 isneg = aux->alu_state & BPF_ALU_NEG_VALUE;
9325 issrc = (aux->alu_state & BPF_ALU_SANITIZE) ==
9326 BPF_ALU_SANITIZE_SRC;
9328 off_reg = issrc ? insn->src_reg : insn->dst_reg;
9330 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
9331 *patch++ = BPF_MOV32_IMM(BPF_REG_AX, aux->alu_limit - 1);
9332 *patch++ = BPF_ALU64_REG(BPF_SUB, BPF_REG_AX, off_reg);
9333 *patch++ = BPF_ALU64_REG(BPF_OR, BPF_REG_AX, off_reg);
9334 *patch++ = BPF_ALU64_IMM(BPF_NEG, BPF_REG_AX, 0);
9335 *patch++ = BPF_ALU64_IMM(BPF_ARSH, BPF_REG_AX, 63);
9337 *patch++ = BPF_ALU64_REG(BPF_AND, BPF_REG_AX,
9339 insn->src_reg = BPF_REG_AX;
9341 *patch++ = BPF_ALU64_REG(BPF_AND, off_reg,
9345 insn->code = insn->code == code_add ?
9346 code_sub : code_add;
9349 *patch++ = BPF_ALU64_IMM(BPF_MUL, off_reg, -1);
9350 cnt = patch - insn_buf;
9352 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
9357 env->prog = prog = new_prog;
9358 insn = new_prog->insnsi + i + delta;
9362 if (insn->code != (BPF_JMP | BPF_CALL))
9364 if (insn->src_reg == BPF_PSEUDO_CALL)
9367 if (insn->imm == BPF_FUNC_get_route_realm)
9368 prog->dst_needed = 1;
9369 if (insn->imm == BPF_FUNC_get_prandom_u32)
9370 bpf_user_rnd_init_once();
9371 if (insn->imm == BPF_FUNC_override_return)
9372 prog->kprobe_override = 1;
9373 if (insn->imm == BPF_FUNC_tail_call) {
9374 /* If we tail call into other programs, we
9375 * cannot make any assumptions since they can
9376 * be replaced dynamically during runtime in
9377 * the program array.
9379 prog->cb_access = 1;
9380 env->prog->aux->stack_depth = MAX_BPF_STACK;
9381 env->prog->aux->max_pkt_offset = MAX_PACKET_OFF;
9383 /* mark bpf_tail_call as different opcode to avoid
9384 * conditional branch in the interpeter for every normal
9385 * call and to prevent accidental JITing by JIT compiler
9386 * that doesn't support bpf_tail_call yet
9389 insn->code = BPF_JMP | BPF_TAIL_CALL;
9391 aux = &env->insn_aux_data[i + delta];
9392 if (env->allow_ptr_leaks && !expect_blinding &&
9393 prog->jit_requested &&
9394 !bpf_map_key_poisoned(aux) &&
9395 !bpf_map_ptr_poisoned(aux) &&
9396 !bpf_map_ptr_unpriv(aux)) {
9397 struct bpf_jit_poke_descriptor desc = {
9398 .reason = BPF_POKE_REASON_TAIL_CALL,
9399 .tail_call.map = BPF_MAP_PTR(aux->map_ptr_state),
9400 .tail_call.key = bpf_map_key_immediate(aux),
9403 ret = bpf_jit_add_poke_descriptor(prog, &desc);
9405 verbose(env, "adding tail call poke descriptor failed\n");
9409 insn->imm = ret + 1;
9413 if (!bpf_map_ptr_unpriv(aux))
9416 /* instead of changing every JIT dealing with tail_call
9417 * emit two extra insns:
9418 * if (index >= max_entries) goto out;
9419 * index &= array->index_mask;
9420 * to avoid out-of-bounds cpu speculation
9422 if (bpf_map_ptr_poisoned(aux)) {
9423 verbose(env, "tail_call abusing map_ptr\n");
9427 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
9428 insn_buf[0] = BPF_JMP_IMM(BPF_JGE, BPF_REG_3,
9429 map_ptr->max_entries, 2);
9430 insn_buf[1] = BPF_ALU32_IMM(BPF_AND, BPF_REG_3,
9431 container_of(map_ptr,
9434 insn_buf[2] = *insn;
9436 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt);
9441 env->prog = prog = new_prog;
9442 insn = new_prog->insnsi + i + delta;
9446 /* BPF_EMIT_CALL() assumptions in some of the map_gen_lookup
9447 * and other inlining handlers are currently limited to 64 bit
9450 if (prog->jit_requested && BITS_PER_LONG == 64 &&
9451 (insn->imm == BPF_FUNC_map_lookup_elem ||
9452 insn->imm == BPF_FUNC_map_update_elem ||
9453 insn->imm == BPF_FUNC_map_delete_elem ||
9454 insn->imm == BPF_FUNC_map_push_elem ||
9455 insn->imm == BPF_FUNC_map_pop_elem ||
9456 insn->imm == BPF_FUNC_map_peek_elem)) {
9457 aux = &env->insn_aux_data[i + delta];
9458 if (bpf_map_ptr_poisoned(aux))
9459 goto patch_call_imm;
9461 map_ptr = BPF_MAP_PTR(aux->map_ptr_state);
9463 if (insn->imm == BPF_FUNC_map_lookup_elem &&
9464 ops->map_gen_lookup) {
9465 cnt = ops->map_gen_lookup(map_ptr, insn_buf);
9466 if (cnt == 0 || cnt >= ARRAY_SIZE(insn_buf)) {
9467 verbose(env, "bpf verifier is misconfigured\n");
9471 new_prog = bpf_patch_insn_data(env, i + delta,
9477 env->prog = prog = new_prog;
9478 insn = new_prog->insnsi + i + delta;
9482 BUILD_BUG_ON(!__same_type(ops->map_lookup_elem,
9483 (void *(*)(struct bpf_map *map, void *key))NULL));
9484 BUILD_BUG_ON(!__same_type(ops->map_delete_elem,
9485 (int (*)(struct bpf_map *map, void *key))NULL));
9486 BUILD_BUG_ON(!__same_type(ops->map_update_elem,
9487 (int (*)(struct bpf_map *map, void *key, void *value,
9489 BUILD_BUG_ON(!__same_type(ops->map_push_elem,
9490 (int (*)(struct bpf_map *map, void *value,
9492 BUILD_BUG_ON(!__same_type(ops->map_pop_elem,
9493 (int (*)(struct bpf_map *map, void *value))NULL));
9494 BUILD_BUG_ON(!__same_type(ops->map_peek_elem,
9495 (int (*)(struct bpf_map *map, void *value))NULL));
9497 switch (insn->imm) {
9498 case BPF_FUNC_map_lookup_elem:
9499 insn->imm = BPF_CAST_CALL(ops->map_lookup_elem) -
9502 case BPF_FUNC_map_update_elem:
9503 insn->imm = BPF_CAST_CALL(ops->map_update_elem) -
9506 case BPF_FUNC_map_delete_elem:
9507 insn->imm = BPF_CAST_CALL(ops->map_delete_elem) -
9510 case BPF_FUNC_map_push_elem:
9511 insn->imm = BPF_CAST_CALL(ops->map_push_elem) -
9514 case BPF_FUNC_map_pop_elem:
9515 insn->imm = BPF_CAST_CALL(ops->map_pop_elem) -
9518 case BPF_FUNC_map_peek_elem:
9519 insn->imm = BPF_CAST_CALL(ops->map_peek_elem) -
9524 goto patch_call_imm;
9527 if (prog->jit_requested && BITS_PER_LONG == 64 &&
9528 insn->imm == BPF_FUNC_jiffies64) {
9529 struct bpf_insn ld_jiffies_addr[2] = {
9530 BPF_LD_IMM64(BPF_REG_0,
9531 (unsigned long)&jiffies),
9534 insn_buf[0] = ld_jiffies_addr[0];
9535 insn_buf[1] = ld_jiffies_addr[1];
9536 insn_buf[2] = BPF_LDX_MEM(BPF_DW, BPF_REG_0,
9540 new_prog = bpf_patch_insn_data(env, i + delta, insn_buf,
9546 env->prog = prog = new_prog;
9547 insn = new_prog->insnsi + i + delta;
9552 fn = env->ops->get_func_proto(insn->imm, env->prog);
9553 /* all functions that have prototype and verifier allowed
9554 * programs to call them, must be real in-kernel functions
9558 "kernel subsystem misconfigured func %s#%d\n",
9559 func_id_name(insn->imm), insn->imm);
9562 insn->imm = fn->func - __bpf_call_base;
9565 /* Since poke tab is now finalized, publish aux to tracker. */
9566 for (i = 0; i < prog->aux->size_poke_tab; i++) {
9567 map_ptr = prog->aux->poke_tab[i].tail_call.map;
9568 if (!map_ptr->ops->map_poke_track ||
9569 !map_ptr->ops->map_poke_untrack ||
9570 !map_ptr->ops->map_poke_run) {
9571 verbose(env, "bpf verifier is misconfigured\n");
9575 ret = map_ptr->ops->map_poke_track(map_ptr, prog->aux);
9577 verbose(env, "tracking tail call prog failed\n");
9585 static void free_states(struct bpf_verifier_env *env)
9587 struct bpf_verifier_state_list *sl, *sln;
9590 sl = env->free_list;
9593 free_verifier_state(&sl->state, false);
9597 env->free_list = NULL;
9599 if (!env->explored_states)
9602 for (i = 0; i < state_htab_size(env); i++) {
9603 sl = env->explored_states[i];
9607 free_verifier_state(&sl->state, false);
9611 env->explored_states[i] = NULL;
9615 /* The verifier is using insn_aux_data[] to store temporary data during
9616 * verification and to store information for passes that run after the
9617 * verification like dead code sanitization. do_check_common() for subprogram N
9618 * may analyze many other subprograms. sanitize_insn_aux_data() clears all
9619 * temporary data after do_check_common() finds that subprogram N cannot be
9620 * verified independently. pass_cnt counts the number of times
9621 * do_check_common() was run and insn->aux->seen tells the pass number
9622 * insn_aux_data was touched. These variables are compared to clear temporary
9623 * data from failed pass. For testing and experiments do_check_common() can be
9624 * run multiple times even when prior attempt to verify is unsuccessful.
9626 static void sanitize_insn_aux_data(struct bpf_verifier_env *env)
9628 struct bpf_insn *insn = env->prog->insnsi;
9629 struct bpf_insn_aux_data *aux;
9632 for (i = 0; i < env->prog->len; i++) {
9633 class = BPF_CLASS(insn[i].code);
9634 if (class != BPF_LDX && class != BPF_STX)
9636 aux = &env->insn_aux_data[i];
9637 if (aux->seen != env->pass_cnt)
9639 memset(aux, 0, offsetof(typeof(*aux), orig_idx));
9643 static int do_check_common(struct bpf_verifier_env *env, int subprog)
9645 struct bpf_verifier_state *state;
9646 struct bpf_reg_state *regs;
9649 env->prev_linfo = NULL;
9652 state = kzalloc(sizeof(struct bpf_verifier_state), GFP_KERNEL);
9655 state->curframe = 0;
9656 state->speculative = false;
9657 state->branches = 1;
9658 state->frame[0] = kzalloc(sizeof(struct bpf_func_state), GFP_KERNEL);
9659 if (!state->frame[0]) {
9663 env->cur_state = state;
9664 init_func_state(env, state->frame[0],
9665 BPF_MAIN_FUNC /* callsite */,
9669 regs = state->frame[state->curframe]->regs;
9670 if (subprog || env->prog->type == BPF_PROG_TYPE_EXT) {
9671 ret = btf_prepare_func_args(env, subprog, regs);
9674 for (i = BPF_REG_1; i <= BPF_REG_5; i++) {
9675 if (regs[i].type == PTR_TO_CTX)
9676 mark_reg_known_zero(env, regs, i);
9677 else if (regs[i].type == SCALAR_VALUE)
9678 mark_reg_unknown(env, regs, i);
9681 /* 1st arg to a function */
9682 regs[BPF_REG_1].type = PTR_TO_CTX;
9683 mark_reg_known_zero(env, regs, BPF_REG_1);
9684 ret = btf_check_func_arg_match(env, subprog, regs);
9686 /* unlikely verifier bug. abort.
9687 * ret == 0 and ret < 0 are sadly acceptable for
9688 * main() function due to backward compatibility.
9689 * Like socket filter program may be written as:
9690 * int bpf_prog(struct pt_regs *ctx)
9691 * and never dereference that ctx in the program.
9692 * 'struct pt_regs' is a type mismatch for socket
9693 * filter that should be using 'struct __sk_buff'.
9698 ret = do_check(env);
9700 /* check for NULL is necessary, since cur_state can be freed inside
9701 * do_check() under memory pressure.
9703 if (env->cur_state) {
9704 free_verifier_state(env->cur_state, true);
9705 env->cur_state = NULL;
9707 while (!pop_stack(env, NULL, NULL));
9710 /* clean aux data in case subprog was rejected */
9711 sanitize_insn_aux_data(env);
9715 /* Verify all global functions in a BPF program one by one based on their BTF.
9716 * All global functions must pass verification. Otherwise the whole program is rejected.
9727 * foo() will be verified first for R1=any_scalar_value. During verification it
9728 * will be assumed that bar() already verified successfully and call to bar()
9729 * from foo() will be checked for type match only. Later bar() will be verified
9730 * independently to check that it's safe for R1=any_scalar_value.
9732 static int do_check_subprogs(struct bpf_verifier_env *env)
9734 struct bpf_prog_aux *aux = env->prog->aux;
9737 if (!aux->func_info)
9740 for (i = 1; i < env->subprog_cnt; i++) {
9741 if (aux->func_info_aux[i].linkage != BTF_FUNC_GLOBAL)
9743 env->insn_idx = env->subprog_info[i].start;
9744 WARN_ON_ONCE(env->insn_idx == 0);
9745 ret = do_check_common(env, i);
9748 } else if (env->log.level & BPF_LOG_LEVEL) {
9750 "Func#%d is safe for any args that match its prototype\n",
9757 static int do_check_main(struct bpf_verifier_env *env)
9762 ret = do_check_common(env, 0);
9764 env->prog->aux->stack_depth = env->subprog_info[0].stack_depth;
9769 static void print_verification_stats(struct bpf_verifier_env *env)
9773 if (env->log.level & BPF_LOG_STATS) {
9774 verbose(env, "verification time %lld usec\n",
9775 div_u64(env->verification_time, 1000));
9776 verbose(env, "stack depth ");
9777 for (i = 0; i < env->subprog_cnt; i++) {
9778 u32 depth = env->subprog_info[i].stack_depth;
9780 verbose(env, "%d", depth);
9781 if (i + 1 < env->subprog_cnt)
9786 verbose(env, "processed %d insns (limit %d) max_states_per_insn %d "
9787 "total_states %d peak_states %d mark_read %d\n",
9788 env->insn_processed, BPF_COMPLEXITY_LIMIT_INSNS,
9789 env->max_states_per_insn, env->total_states,
9790 env->peak_states, env->longest_mark_read_walk);
9793 static int check_struct_ops_btf_id(struct bpf_verifier_env *env)
9795 const struct btf_type *t, *func_proto;
9796 const struct bpf_struct_ops *st_ops;
9797 const struct btf_member *member;
9798 struct bpf_prog *prog = env->prog;
9799 u32 btf_id, member_idx;
9802 btf_id = prog->aux->attach_btf_id;
9803 st_ops = bpf_struct_ops_find(btf_id);
9805 verbose(env, "attach_btf_id %u is not a supported struct\n",
9811 member_idx = prog->expected_attach_type;
9812 if (member_idx >= btf_type_vlen(t)) {
9813 verbose(env, "attach to invalid member idx %u of struct %s\n",
9814 member_idx, st_ops->name);
9818 member = &btf_type_member(t)[member_idx];
9819 mname = btf_name_by_offset(btf_vmlinux, member->name_off);
9820 func_proto = btf_type_resolve_func_ptr(btf_vmlinux, member->type,
9823 verbose(env, "attach to invalid member %s(@idx %u) of struct %s\n",
9824 mname, member_idx, st_ops->name);
9828 if (st_ops->check_member) {
9829 int err = st_ops->check_member(t, member);
9832 verbose(env, "attach to unsupported member %s of struct %s\n",
9833 mname, st_ops->name);
9838 prog->aux->attach_func_proto = func_proto;
9839 prog->aux->attach_func_name = mname;
9840 env->ops = st_ops->verifier_ops;
9844 #define SECURITY_PREFIX "security_"
9846 static int check_attach_modify_return(struct bpf_verifier_env *env)
9848 struct bpf_prog *prog = env->prog;
9849 unsigned long addr = (unsigned long) prog->aux->trampoline->func.addr;
9851 /* This is expected to be cleaned up in the future with the KRSI effort
9852 * introducing the LSM_HOOK macro for cleaning up lsm_hooks.h.
9854 if (within_error_injection_list(addr) ||
9855 !strncmp(SECURITY_PREFIX, prog->aux->attach_func_name,
9856 sizeof(SECURITY_PREFIX) - 1))
9859 verbose(env, "fmod_ret attach_btf_id %u (%s) is not modifiable\n",
9860 prog->aux->attach_btf_id, prog->aux->attach_func_name);
9865 static int check_attach_btf_id(struct bpf_verifier_env *env)
9867 struct bpf_prog *prog = env->prog;
9868 bool prog_extension = prog->type == BPF_PROG_TYPE_EXT;
9869 struct bpf_prog *tgt_prog = prog->aux->linked_prog;
9870 u32 btf_id = prog->aux->attach_btf_id;
9871 const char prefix[] = "btf_trace_";
9872 int ret = 0, subprog = -1, i;
9873 struct bpf_trampoline *tr;
9874 const struct btf_type *t;
9875 bool conservative = true;
9881 if (prog->type == BPF_PROG_TYPE_STRUCT_OPS)
9882 return check_struct_ops_btf_id(env);
9884 if (prog->type != BPF_PROG_TYPE_TRACING &&
9885 prog->type != BPF_PROG_TYPE_LSM &&
9890 verbose(env, "Tracing programs must provide btf_id\n");
9893 btf = bpf_prog_get_target_btf(prog);
9896 "FENTRY/FEXIT program can only be attached to another program annotated with BTF\n");
9899 t = btf_type_by_id(btf, btf_id);
9901 verbose(env, "attach_btf_id %u is invalid\n", btf_id);
9904 tname = btf_name_by_offset(btf, t->name_off);
9906 verbose(env, "attach_btf_id %u doesn't have a name\n", btf_id);
9910 struct bpf_prog_aux *aux = tgt_prog->aux;
9912 for (i = 0; i < aux->func_info_cnt; i++)
9913 if (aux->func_info[i].type_id == btf_id) {
9917 if (subprog == -1) {
9918 verbose(env, "Subprog %s doesn't exist\n", tname);
9921 conservative = aux->func_info_aux[subprog].unreliable;
9922 if (prog_extension) {
9925 "Cannot replace static functions\n");
9928 if (!prog->jit_requested) {
9930 "Extension programs should be JITed\n");
9933 env->ops = bpf_verifier_ops[tgt_prog->type];
9935 if (!tgt_prog->jited) {
9936 verbose(env, "Can attach to only JITed progs\n");
9939 if (tgt_prog->type == prog->type) {
9940 /* Cannot fentry/fexit another fentry/fexit program.
9941 * Cannot attach program extension to another extension.
9942 * It's ok to attach fentry/fexit to extension program.
9944 verbose(env, "Cannot recursively attach\n");
9947 if (tgt_prog->type == BPF_PROG_TYPE_TRACING &&
9949 (tgt_prog->expected_attach_type == BPF_TRACE_FENTRY ||
9950 tgt_prog->expected_attach_type == BPF_TRACE_FEXIT)) {
9951 /* Program extensions can extend all program types
9952 * except fentry/fexit. The reason is the following.
9953 * The fentry/fexit programs are used for performance
9954 * analysis, stats and can be attached to any program
9955 * type except themselves. When extension program is
9956 * replacing XDP function it is necessary to allow
9957 * performance analysis of all functions. Both original
9958 * XDP program and its program extension. Hence
9959 * attaching fentry/fexit to BPF_PROG_TYPE_EXT is
9960 * allowed. If extending of fentry/fexit was allowed it
9961 * would be possible to create long call chain
9962 * fentry->extension->fentry->extension beyond
9963 * reasonable stack size. Hence extending fentry is not
9966 verbose(env, "Cannot extend fentry/fexit\n");
9969 key = ((u64)aux->id) << 32 | btf_id;
9971 if (prog_extension) {
9972 verbose(env, "Cannot replace kernel functions\n");
9978 switch (prog->expected_attach_type) {
9979 case BPF_TRACE_RAW_TP:
9982 "Only FENTRY/FEXIT progs are attachable to another BPF prog\n");
9985 if (!btf_type_is_typedef(t)) {
9986 verbose(env, "attach_btf_id %u is not a typedef\n",
9990 if (strncmp(prefix, tname, sizeof(prefix) - 1)) {
9991 verbose(env, "attach_btf_id %u points to wrong type name %s\n",
9995 tname += sizeof(prefix) - 1;
9996 t = btf_type_by_id(btf, t->type);
9997 if (!btf_type_is_ptr(t))
9998 /* should never happen in valid vmlinux build */
10000 t = btf_type_by_id(btf, t->type);
10001 if (!btf_type_is_func_proto(t))
10002 /* should never happen in valid vmlinux build */
10005 /* remember two read only pointers that are valid for
10006 * the life time of the kernel
10008 prog->aux->attach_func_name = tname;
10009 prog->aux->attach_func_proto = t;
10010 prog->aux->attach_btf_trace = true;
10013 if (!prog_extension)
10016 case BPF_MODIFY_RETURN:
10018 case BPF_TRACE_FENTRY:
10019 case BPF_TRACE_FEXIT:
10020 prog->aux->attach_func_name = tname;
10021 if (prog->type == BPF_PROG_TYPE_LSM) {
10022 ret = bpf_lsm_verify_prog(&env->log, prog);
10027 if (!btf_type_is_func(t)) {
10028 verbose(env, "attach_btf_id %u is not a function\n",
10032 if (prog_extension &&
10033 btf_check_type_match(env, prog, btf, t))
10035 t = btf_type_by_id(btf, t->type);
10036 if (!btf_type_is_func_proto(t))
10038 tr = bpf_trampoline_lookup(key);
10041 /* t is either vmlinux type or another program's type */
10042 prog->aux->attach_func_proto = t;
10043 mutex_lock(&tr->mutex);
10044 if (tr->func.addr) {
10045 prog->aux->trampoline = tr;
10048 if (tgt_prog && conservative) {
10049 prog->aux->attach_func_proto = NULL;
10052 ret = btf_distill_func_proto(&env->log, btf, t,
10053 tname, &tr->func.model);
10058 addr = (long) tgt_prog->bpf_func;
10060 addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
10062 addr = kallsyms_lookup_name(tname);
10065 "The address of function %s cannot be found\n",
10071 tr->func.addr = (void *)addr;
10072 prog->aux->trampoline = tr;
10074 if (prog->expected_attach_type == BPF_MODIFY_RETURN)
10075 ret = check_attach_modify_return(env);
10077 mutex_unlock(&tr->mutex);
10079 bpf_trampoline_put(tr);
10084 int bpf_check(struct bpf_prog **prog, union bpf_attr *attr,
10085 union bpf_attr __user *uattr)
10087 u64 start_time = ktime_get_ns();
10088 struct bpf_verifier_env *env;
10089 struct bpf_verifier_log *log;
10090 int i, len, ret = -EINVAL;
10093 /* no program is valid */
10094 if (ARRAY_SIZE(bpf_verifier_ops) == 0)
10097 /* 'struct bpf_verifier_env' can be global, but since it's not small,
10098 * allocate/free it every time bpf_check() is called
10100 env = kzalloc(sizeof(struct bpf_verifier_env), GFP_KERNEL);
10105 len = (*prog)->len;
10106 env->insn_aux_data =
10107 vzalloc(array_size(sizeof(struct bpf_insn_aux_data), len));
10109 if (!env->insn_aux_data)
10111 for (i = 0; i < len; i++)
10112 env->insn_aux_data[i].orig_idx = i;
10114 env->ops = bpf_verifier_ops[env->prog->type];
10115 is_priv = capable(CAP_SYS_ADMIN);
10117 if (!btf_vmlinux && IS_ENABLED(CONFIG_DEBUG_INFO_BTF)) {
10118 mutex_lock(&bpf_verifier_lock);
10120 btf_vmlinux = btf_parse_vmlinux();
10121 mutex_unlock(&bpf_verifier_lock);
10124 /* grab the mutex to protect few globals used by verifier */
10126 mutex_lock(&bpf_verifier_lock);
10128 if (attr->log_level || attr->log_buf || attr->log_size) {
10129 /* user requested verbose verifier output
10130 * and supplied buffer to store the verification trace
10132 log->level = attr->log_level;
10133 log->ubuf = (char __user *) (unsigned long) attr->log_buf;
10134 log->len_total = attr->log_size;
10137 /* log attributes have to be sane */
10138 if (log->len_total < 128 || log->len_total > UINT_MAX >> 2 ||
10139 !log->level || !log->ubuf || log->level & ~BPF_LOG_MASK)
10143 if (IS_ERR(btf_vmlinux)) {
10144 /* Either gcc or pahole or kernel are broken. */
10145 verbose(env, "in-kernel BTF is malformed\n");
10146 ret = PTR_ERR(btf_vmlinux);
10147 goto skip_full_check;
10150 env->strict_alignment = !!(attr->prog_flags & BPF_F_STRICT_ALIGNMENT);
10151 if (!IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
10152 env->strict_alignment = true;
10153 if (attr->prog_flags & BPF_F_ANY_ALIGNMENT)
10154 env->strict_alignment = false;
10156 env->allow_ptr_leaks = is_priv;
10159 env->test_state_freq = attr->prog_flags & BPF_F_TEST_STATE_FREQ;
10161 ret = replace_map_fd_with_map_ptr(env);
10163 goto skip_full_check;
10165 if (bpf_prog_is_dev_bound(env->prog->aux)) {
10166 ret = bpf_prog_offload_verifier_prep(env->prog);
10168 goto skip_full_check;
10171 env->explored_states = kvcalloc(state_htab_size(env),
10172 sizeof(struct bpf_verifier_state_list *),
10175 if (!env->explored_states)
10176 goto skip_full_check;
10178 ret = check_subprogs(env);
10180 goto skip_full_check;
10182 ret = check_btf_info(env, attr, uattr);
10184 goto skip_full_check;
10186 ret = check_attach_btf_id(env);
10188 goto skip_full_check;
10190 ret = check_cfg(env);
10192 goto skip_full_check;
10194 ret = do_check_subprogs(env);
10195 ret = ret ?: do_check_main(env);
10197 if (ret == 0 && bpf_prog_is_dev_bound(env->prog->aux))
10198 ret = bpf_prog_offload_finalize(env);
10201 kvfree(env->explored_states);
10204 ret = check_max_stack_depth(env);
10206 /* instruction rewrites happen after this point */
10209 opt_hard_wire_dead_code_branches(env);
10211 ret = opt_remove_dead_code(env);
10213 ret = opt_remove_nops(env);
10216 sanitize_dead_code(env);
10220 /* program is valid, convert *(u32*)(ctx + off) accesses */
10221 ret = convert_ctx_accesses(env);
10224 ret = fixup_bpf_calls(env);
10226 /* do 32-bit optimization after insn patching has done so those patched
10227 * insns could be handled correctly.
10229 if (ret == 0 && !bpf_prog_is_dev_bound(env->prog->aux)) {
10230 ret = opt_subreg_zext_lo32_rnd_hi32(env, attr);
10231 env->prog->aux->verifier_zext = bpf_jit_needs_zext() ? !ret
10236 ret = fixup_call_args(env);
10238 env->verification_time = ktime_get_ns() - start_time;
10239 print_verification_stats(env);
10241 if (log->level && bpf_verifier_log_full(log))
10243 if (log->level && !log->ubuf) {
10245 goto err_release_maps;
10248 if (ret == 0 && env->used_map_cnt) {
10249 /* if program passed verifier, update used_maps in bpf_prog_info */
10250 env->prog->aux->used_maps = kmalloc_array(env->used_map_cnt,
10251 sizeof(env->used_maps[0]),
10254 if (!env->prog->aux->used_maps) {
10256 goto err_release_maps;
10259 memcpy(env->prog->aux->used_maps, env->used_maps,
10260 sizeof(env->used_maps[0]) * env->used_map_cnt);
10261 env->prog->aux->used_map_cnt = env->used_map_cnt;
10263 /* program is valid. Convert pseudo bpf_ld_imm64 into generic
10264 * bpf_ld_imm64 instructions
10266 convert_pseudo_ld_imm64(env);
10270 adjust_btf_func(env);
10273 if (!env->prog->aux->used_maps)
10274 /* if we didn't copy map pointers into bpf_prog_info, release
10275 * them now. Otherwise free_used_maps() will release them.
10281 mutex_unlock(&bpf_verifier_lock);
10282 vfree(env->insn_aux_data);